Jelly is a static analyzer for performing

• call graphs construction,
• library usage pattern matching, and
• vulnerability exposure analysis

for JavaScript (and TypeScript) programs that use the Node.js platform.

The analyzer design is based on ideas from JAM [1], TAPIR [2] and ACG [3]. Its core is a flow-insensitive control-flow and points-to analysis that uses a hybrid of field-based and allocation-site abstraction, together with access paths for tracking library usage. It models the main parts of the ECMAScript language and standard library (intentionally not fully soundly!), and not (yet) the Node.js standard library.

[1] Benjamin Barslev Nielsen, Martin Toldam Torp, Anders Møller: Modular call graph construction for security scanning of Node.js applications. ISSTA 2021: 29-41

[2] Anders Møller, Benjamin Barslev Nielsen, Martin Toldam Torp: Detecting locations in JavaScript programs affected by breaking library changes. Proc. ACM Program. Lang. 4(OOPSLA): 187:1-187:25 (2020)

[3] Asger Feldthaus, Max Schäfer, Manu Sridharan, Julian Dolby, Frank Tip: Efficient construction of approximate call graphs for JavaScript IDE services. ICSE 2013: 752-761

npm install -g @cs-au-dk/jelly

Other options are described below at How to build.

See the full usage:

jelly --help

When running the Jelly static analyzer, one or more entry files are given as input. Directories are expanded (using heuristics to skip certain files and directories, see files.ts). All files reachable from entry files are analyzed, except if option --ignore-dependencies is used, in which case only entry files are analyzed, and only files within the base directory (auto-detected or specified using option --basedir or -b) are included.

As an example, generate a call graph for the winston package and all its dependencies, both in JSON format and for HTML visualization:

jelly -j cg.json -m cg.html node_modules/winston -b node_modules

Viewing cg.html in a browser:

To set the heap limit, prefix commands by, for example:

NODE_OPTIONS=--max-old-space-size=4096

Note that analyzing with all dependencies (i.e., not using --ignore-dependencies) can take a long time. The options --max-rounds or --timeout can be used to terminate the analysis early to provide partial (unsound) results.

Install dependencies:

npm install

Compile TypeScript code:

npm run build

After compilation, Jelly can be run like this:

node lib/main.js

Build binary executables (optional), placed in dist/:

sudo npm install -g pkg
npm run pkg

Build Docker image (including support for dynamic call graph construction):

npm run build-docker

Run Jelly in Docker with the directory specified as first argument as current working directory:

./bin/jelly-docker . tests/helloworld/app.js --callgraph-html cg.html

Jelly can be run in server-mode as an alternative to the command-line interface:

jelly-server

or

node lib/server.js

See also the instructions above for how to build binary executables.

Requests to the server are sent on stdin using the JSON format described in typings/ipc.d.ts. Responses are returned (asynchronously) on stdout with the two-line header (including the empty line)

Content-Length: <bytes>

with \r\n linebreaks.

Jelly supports dynamic call graph construction via NodeProf, which can be used for measuring recall (or unsoundness) of the static analysis.

Install NodeProf (see also the information about Docker above):

sudo dnf install g++ libstdc++-static
mkdir -p ~/tools; cd ~/tools
git clone --depth 1 --branch 6.0.4 https://github.com/graalvm/mx.git
export PATH=$PATH:$HOME/tools/mx
mx -y fetch-jdk --java-distribution labsjdk-ce-17
export JAVA_HOME=$HOME/.mx/jdks/labsjdk-ce-17-jvmci-22.2-b01
git clone --depth 1 https://github.com/Haiyang-Sun/nodeprof.js.git
cd nodeprof.js
mx sforceimports
mx --dy /compiler build

As an example, run tests/micro/classes.js or tests/helloworld/app.js with instrumentation for call graph construction:

export GRAAL_HOME=$HOME/tools/graal/sdk/latest_graalvm_home

jelly tests/micro/classes.js -d cg.json
jelly tests/helloworld/app.js -d cg.json

Extra arguments to the JavaScript program can be added after --.

It is also possible to run npm test with instrumentation:

jelly --npm-test tests/mochatest -d cg.json

Another approach is to add $JELLY_HOME/lib/bin/node to PATH and set JELLY_OUT, for example to run Mocha directly:

cd tests/mochatest
PATH=$JELLY_HOME/lib/bin:$PATH JELLY_OUT=cg.json node_modules/.bin/mocha

where JELLY_HOME is the home directory of Jelly. This results in a file cg.json-<PID> for each instrumented file that is executed.

Call graphs (generated either statically or dynamically) can be compared for precision and recall:

jelly --compare-callgraphs cg1.json cg2.json

Compile TypeScript code in watch mode:

npm run build-watch

Install as scripts (jelly and jelly-server) for development:

sudo npm link

Install dependencies for tests:

npm run tests-install

Run all tests:

npm test

Run individual tests (specified by regex), for example:

npm test -- -t tests/helloworld

Differential testing can be used to test if updated code results in lower recall than the previous version by comparing the dataflow graph and call graphs of the two versions.

Run the following command to test the testing framework:

TAG=<tag> npm run differential -- -t tiny

where <tag> is the git tag of the previous version you want to compare to.

Then run the following commands to start full test:

TAG=<tag> npm run differential

During the test, the old version of Jelly will be installed in tests/node_modules/jelly-previous and test packages will be installed in tmp/packages.