Skip to content

README.md

Latest commit

 

History

History
79 lines (63 loc) · 3.59 KB

README.md

File metadata and controls

79 lines (63 loc) · 3.59 KB
Raw

Fastjson 1.2.83 modern-JDK file-descriptor lab

This is an isolated, marker-only reconstruction of the Linux file-descriptor continuation for Fastjson 1.2.83. It models the application shape that matters for the claim:

  • Spring Boot 3.2.0, launched as an executable jar with its default embedded Tomcat and default LaunchedClassLoader;
  • Eclipse Temurin JDK 17.0.19+10;
  • Fastjson 1.2.83 with AutoType left at its default (false);
  • a final DTO containing List<Object>; and
  • the literal sink final BoundEnvelope parsed = JSON.parseObject(body, BoundEnvelope.class).

The generated classes cannot run commands. The success class only sets the fixed JVM property FASTJSON_MODERN_FD_MARKER and prints a fixed token. The target and artifact server run on an internal Docker network. Compose requests a loopback-only target-port publication; Docker Desktop may decline to publish that port for an internal-only network, so the scripts deliberately drive the endpoint from the artifact container. Both services run unprivileged with read-only root filesystems and dropped capabilities.

Mechanism under test

One JSON body contains an ordered array. Its first @type causes the Boot/JDK URL loader to retrieve an inert jar through a jar:http: resource name. The stage-one class has no @JSONType, no static initializer, and cannot be defined under the requested URL-shaped binary name. Fastjson 1.2.83's unresolved Exception handling lets parsing continue.

Later elements probe jar:file:/proc/self/fd/N!/fdN/Exception.class. The jar contains one marker-only class for each candidate descriptor. Each class-file internal name exactly matches its URL-shaped @type. If a candidate descriptor is the still-open JDK jar cache, the resource scan sees @JSONType, Fastjson loads the class, and its static initializer sets the fixed marker property.

Run

From this directory:

./scripts/static-safety-check.sh
./scripts/run-positive.sh
./scripts/run-controls.sh

Both scripts build their own Compose project (fj-modern-fd), write complete responses/logs/inspection data beneath evidence/, issue requests over the internal network, and remove only containers and networks belonging to that project. Set MODERN_FD_PROJECT_NAME to a unique value for an isolated parallel run, for example:

MODERN_FD_PROJECT_NAME=fj-modern-fd-review ./scripts/run-positive.sh

Set KEEP_LAB=1 to leave the final case running for manual inspection. If the local Docker engine honors the requested publication, the manual endpoint is http://127.0.0.1:18080/parse; otherwise use Compose exec from a lab service and keep the network internal.

The positive script fails closed unless the exact pinned Java/Fastjson/loader facts, body and value counts, marker, retained JAR cache, and exactly two artifact GETs are present. The control script verifies those same runtime facts, the exact per-case GET counts and parse outcomes, and the expected SafeMode exception. Both Maven builds use a fixed output timestamp so identical source builds produce byte-identical application and builder JARs. Generated evidence and Maven targets are intentionally ignored; SOURCE_SHA256SUMS.txt seals the 15 tracked lab source and documentation files.

The controls are:

  1. ordinary DTO body (no @type);
  2. descriptor candidates without the HTTP seed;
  3. HTTP seed followed only by deliberately impossible descriptors; and
  4. the complete one-body payload with Fastjson SafeMode enabled.

These are lab reproduction tools, not a production scanner. See the repository scanner documentation for passive exposure detection and safe canary guidance.