Assisted Configuration of Native Image Builds

Native images are built ahead of runtime and their build relies on a static analysis of which code will be reachable. However, this analysis cannot always completely predict all usages of the Java Native Interface (JNI), Java reflection, dynamic proxy objects (java.lang.reflect.Proxy) or class path resources (Class.getResource). Undetected usages of these dynamic features need to be provided to the native-image tool in the form of configuration files.

In order to make preparing these configuration files easier and more convenient, GraalVM provides an agent that tracks all usages of dynamic features of an execution on a regular Java VM. It can be enabled on the command line of the GraalVM java command:

/path/to/graalvm/bin/java -agentlib:native-image-agent=config-output-dir=/path/to/config-dir/ ...

Note that -agentlib must be specified before a -jar option or a class name or any application parameters in the java command line.

During execution, the agent interfaces with the Java VM to intercept all calls that look up classes, methods, fields, resources or request proxy accesses. The agent then generates the files jni-config.json, reflect-config.json, proxy-config.json and resource-config.json in the specified output directory, which is /path/to/config-dir/ in the example above. The generated files are stand-alone configuration files in JSON format which contain all intercepted dynamic accesses.

It can be necessary to run the target application more than once with different inputs to trigger separate execution paths for a better coverage of dynamic accesses. The agent supports this with the config-merge-dir option which adds the intercepted accesses to an existing set of configuration files:

/path/to/graalvm/bin/java -agentlib:native-image-agent=config-merge-dir=/path/to/config-dir/ ...

If the specified target directory or configuration files in it are missing when using config-merge-dir, the agent creates them and prints a warning.

By default the agent will write the configuration files after the JVM process terminates. In addition, the agent provides the following flags to write configuration files on a periodic basis.

  • config-write-period-secs: Executes a periodic write every number of seconds as specified in this configuration. Supports only integer values greater than zero.
  • config-write-initial-delay-secs: The number of seconds before the first write is schedule for execution. Supports only integer values greater or equal to zero. Enabled only if config-write-period-secs is greater than zero.

For example:

/path/to/graalvm/bin/java -agentlib:native-image-agent=config-output-dir=/path/to/config-dir/,config-write-period-secs=300,config-write-initial-delay-secs=5 ...

It is advisable to manually review the generated configuration files. Because the agent observes only code that was executed, the resulting configurations can be missing elements that are used in other code paths. It could also make sense to simplify the generated configurations to make any future manual maintenance easier.

The generated configuration files can be supplied to the native-image tool by placing them in a META-INF/native-image/ directory on the class path, for example, in a JAR file used in the image build. This directory (or any of its subdirectories) is searched for files with the names jni-config.json, reflect-config.json, proxy-config.json and resource-config.json, which are then automatically included in the build. Not all of those files must be present. When multiple files with the same name are found, all of them are included.

Advanced Usage

Caller-based Filters

By default, the agent filters dynamic accesses which native-image supports without configuration. The filter mechanism works by identifying the Java method performing the access, also referred to as caller method, and matching its declaring class against a sequence of filter rules. The built-in filter rules exclude dynamic accesses which originate in the Java VM or in parts of the Java class library directly supported by native-image (such as java.nio) from the generated configuration files. Which item (class, method, field, resource, …) is being accessed is not relevant for filtering.

In addition to the built-in filter, custom filter files with additional rules can be specified using the caller-filter-file option, for example: -agentlib:caller-filter-file=/path/to/filter-file,config-output-dir=...

Filter files have the following structure:

{ "rules": [
    {"excludeClasses": "**"},
    {"includeClasses": "*"},
    {"excludeClasses": ""}

The rules section contains a sequence of rules. Each rule specifies either includeClasses, which means that lookups originating in matching classes will be included in the resulting configuration, or excludeClasses, which excludes lookups originating in matching classes from the configuration. Each rule defines a pattern for the set of matching classes, which can end in .* or .**: a .* ending matches all classes in a package and that package only, while a .** ending matches all classes in the package as well as in all subpackages at any depth. Without .* or .**, the rule applies only to a single class with the qualified name that matches the pattern. All rules are processed in the sequence in which they are specified, so later rules can partially or entirely override earlier ones. When multiple filter files are provided (by specifying multiple caller-filter-file options), their rules are chained together in the order in which the files are specified. The rules of the built-in caller filter are always processed first, so they can be overridden in custom filter files.

In the example above, the first rule excludes lookups originating in all classes from package and from all of its subpackages (and their subpackages, etc.) from the generated configuration. In the next rule however, lookups from those classes that are directly in package are included again. Finally, lookups from the HostedHelper class is excluded again. Each of these rules partially overrides the previous ones. For example, if the rules were in the reverse order, the exclusion of** would be the last rule and would override all other rules.

For testing purposes, the built-in filter for Java class library lookups can be disabled by adding the no-builtin-caller-filter option, but the resulting configuration files are generally unsuitable for a native image build. Similarly, the built-in filter for Java VM-internal accesses based on heuristics can be disabled with no-builtin-heuristic-filter and will also generally lead to less usable configuration files. For example: -agentlib:native-image-agent=no-builtin-caller-filter,no-builtin-heuristic-filter,config-output-dir=...

Access Filters

Unlike the caller-based filters described above, which filter dynamic accesses based on where they originate from, access filters apply to the target of the access. Therefore, access filters enable directly excluding packages and classes (and their members) from the generated configuration.

By default, all accessed classes (which also pass the caller-based filters and the built-in filters) are included in the generated configuration. Using the access-filter-file option, a custom filter file that follows the file structure described above can be added. The option can be specified more than once to add multiple filter files and can be combined with the other filter options. For example: -agentlib:access-filter-file=/path/to/access-filter-file,caller-filter-file=/path/to/caller-filter-file,config-output-dir=...

Specifying Configuration Files as native-image Arguments

A directory containing configuration files that is not part of the class path can be specified to native-image via -H:ConfigurationFileDirectories=/path/to/config-dir/. This directory must directly contain all four files jni-config.json, reflect-config.json, proxy-config.json and resource-config.json. A directory with the same four configuration files that is on the class path, but not in META-INF/native-image/, can be provided via -H:ConfigurationResourceRoots=path/to/resources/. Both -H:ConfigurationFileDirectories and -H:ConfigurationResourceRoots can also take a comma-separated list of directories.

Injecting the agent via the process environment

Altering the java command line to inject the agent can prove to be difficult if the Java process is launched by an application or script file or if Java is even embedded in an existing process. In that case, it is also possible to inject the agent via the JAVA_TOOL_OPTIONS environment variable. This environment variable can be picked up by multiple Java processes which run at the same time, in which case each agent must write to a separate output directory with config-output-dir. (The next section describes how to merge sets of configuration files.) In order to use separate paths with a single global JAVA_TOOL_OPTIONS variable, the agent’s output path options support placeholders:

export JAVA_TOOL_OPTIONS="java -agentlib:native-image-agent=config-output-dir=/path/to/config-output-dir-{pid}-{datetime}/"

The {pid} placeholder is replaced with the process identifier, while {datetime} is replaced with the system date and time in UTC, formatted according to ISO 8601. For the above example, the resulting path could be: /path/to/config-output-dir-31415-20181231T235950Z/.

The Configuration Tool

When using the agent in multiple processes at the same time as described in the previous section, config-output-dir is a safe option, but results in multiple sets of configuration files. The native-image-configure tool can be used to merge these configuration files. This tool must first be built with:

native-image --tool:native-image-configure

Then, the tool can be used to merge sets of configuration files as follows:

native-image-configure generate --input-dir=/path/to/config-dir-0/ --input-dir=/path/to/config-dir-1/ --output-dir=/path/to/merged-config-dir/

This command reads one set of configuration files from /path/to/config-dir-0/ and another from /path/to/config-dir-1/ and then writes a set of configuration files that contains both of their information to /path/to/merged-config-dir/.

An arbitrary number of --input-dir arguments with sets of configuration files can be specified. See native-image-configure help for all options.

Trace Files

In the examples above, native-image-agent has been used to both keep track of the dynamic accesses in a Java VM and then to generate a set of configuration files from them. However, for a better understanding of the execution, the agent can also write a trace file in JSON format that contains each individual access:

/path/to/graalvm/bin/java -agentlib:native-image-agent=trace-output=/path/to/trace-file.json ...

The native-image-configure tool can transform trace files to configuration files that can be used in native image builds. The following command reads and processes trace-file.json and generates a set of configuration files in directory /path/to/config-dir/:

native-image-configure generate --trace-input=/path/to/trace-file.json --output-dir=/path/to/config-dir/


Although the agent is distributed with Graal VM, it uses the Java VM Tool Interface (JVMTI) and can potentially be used with other Java VMs that support JVMTI. In this case, it is necessary to provide the absolute path of the agent:

/path/to/some/java -agentpath:/path/to/graalvm/jre/lib/amd64/<options> ...