Fix: Make it impossible to load dynamically generated shared libraries

Author: hoodmane

src/pyodide/helpers.bzl

@@ -239,6 +239,17 @@ _REPLACEMENTS = [
         "var tableBase=metadata.tableSize?wasmTable.length:0;" +
         "Module.snapshotDebug && console.log('loadWebAssemblyModule', libName, memoryBase, tableBase);",
     ],
+    [
+        "function loadLibData(){",
+        """
+        function loadLibData(){
+            var f = findLibraryFS(libName, flags.rpath);
+            var libData = Module.patched_loadLibData(Module, f);
+            return flags.loadAsync ? Promise.resolve(libData) : libData;
+        }
+        function loadLibData1(){
+        """,
+    ],
 ]
 
 def _python_bundle(version, *, pyodide_asm_wasm = None, pyodide_asm_js = None, python_stdlib_zip = None, emscripten_setup_override = None):

src/pyodide/internal/pool/builtin_wrappers.ts

@@ -226,12 +226,9 @@ function prepareStackTrace(
       return [false, funcName];
     }
     return [
-      [
-        'loadModule',
-        'convertJsFunctionToWasm',
-        'generate',
-        'getPyEMCountArgsPtr',
-      ].includes(funcName),
+      ['convertJsFunctionToWasm', 'generate', 'getPyEMCountArgsPtr'].includes(
+        funcName
+      ),
       funcName,
     ];
   } catch (e) {

src/pyodide/internal/python.ts

@@ -38,6 +38,7 @@ import {
 import { loadPackages } from 'pyodide-internal:loadPackage';
 import { default as MetadataReader } from 'pyodide-internal:runtime-generated/metadata';
 import { TRANSITIVE_REQUIREMENTS } from 'pyodide-internal:metadata';
+import { getTrustedReadFunc } from 'pyodide-internal:readOnlyFS';
 
 /**
  * After running `instantiateEmscriptenModule` but before calling into any C
@@ -206,6 +207,24 @@ export function clearSignals(Module: Module): void {
   }
 }
 
+function patched_loadLibData(Module: Module, path: string): WebAssembly.Module {
+  const { node } = Module.FS.lookupPath(path);
+  // Get the trusted read function from our private Map, not from the node
+  // or filesystem object (which could have been tampered with by user code)
+  const trustedRead = getTrustedReadFunc(node);
+  if (!trustedRead) {
+    throw new Error(
+      'Can only load shared libraries from read only file systems.'
+    );
+  }
+  const stat = node.node_ops.getattr(node);
+  const buffer = new Uint8Array(stat.size);
+  // Create a minimal stream object and read using trusted read function
+  const stream = { node, position: 0 };
+  trustedRead(stream, buffer, 0, stat.size, 0);
+  return UnsafeEval.newWasmModule(buffer);
+}
+
 export function loadPyodide(
   isWorkerd: boolean,
   lockfile: PackageLock,
@@ -216,6 +235,7 @@ export function loadPyodide(
     const Module = enterJaegerSpan('instantiate_emscripten', () =>
       SetupEmscripten.getModule()
     );
+    Module.patched_loadLibData = patched_loadLibData;
     Module.API.config.jsglobals = globalThis;
     if (isWorkerd) {
       Module.API.config.indexURL = indexURL;

src/pyodide/internal/readOnlyFS.ts

@@ -1,3 +1,18 @@
+type ReadFn<Info> = FSStreamOps<Info>['read'];
+
+// When we load shared libraries we need to ensure they come from a read only file system.
+
+// Map to store the original trusted read function for each read-only filesystem. We store the
+// function itself to prevent attacks where user code modifies stream_ops.read after filesystem
+// creation and tricks us into loading a dynamically generated so file.
+const TRUSTED_READ_FUNCS: Map<object, ReadFn<any>> = new Map();
+
+export function getTrustedReadFunc<Info>(
+  node: FSNode<Info>
+): ReadFn<Info> | undefined {
+  return TRUSTED_READ_FUNCS.get(node.mount.type);
+}
+
 export function createReadonlyFS<Info>(
   FSOps: FSOps<Info>,
   Module: Module
@@ -77,5 +92,8 @@ export function createReadonlyFS<Info>(
       },
     },
   };
+  // Register this filesystem as read-only and store its trusted read function so we can load so
+  // files from it.
+  TRUSTED_READ_FUNCS.set(ReadOnlyFS, ReadOnlyFS.stream_ops.read);
   return ReadOnlyFS;
 }

src/pyodide/types/emscripten.d.ts

@@ -135,4 +135,5 @@ interface Module {
   Py_EmscriptenSignalBuffer: Uint8Array;
   _Py_EMSCRIPTEN_SIGNAL_HANDLING: number;
   ___memory_base: WebAssembly.Global<'i32'>;
+  patched_loadLibData: (Module: Module, path: string) => WebAssembly.Module;
 }

src/pyodide/types/filesystem.d.ts

@@ -18,17 +18,31 @@ interface TarFSInfo {
 declare type MetadataDirInfo = Map<string, MetadataFSInfo>;
 declare type MetadataFSInfo = MetadataDirInfo | number; // file infos are numbers and dir infos are maps
 
+interface FSLookupResult<Info> {
+  node: FSNode<Info>;
+}
+
 interface FS {
   mkdir: (dirname: string) => void;
   mkdirTree: (dirname: string) => void;
   writeFile: (fname: string, contents: Uint8Array, options: object) => void;
-  readFile: (fname: string) => Uint8Array;
+  readFile: (fname: string, options?: { encoding?: string }) => Uint8Array;
   mount(fs: object, options: { info?: any }, path: string): void;
   createNode<Info>(
     parent: FSNode<Info> | null,
     name: string,
     mode: number
   ): FSNode<Info>;
+  lookupPath<Info>(path: string): FSLookupResult<Info>;
+  open<Info>(nodeOrPath: FSNode<Info> | string, flags?: number): FSStream<Info>;
+  read<Info>(
+    stream: FSStream<Info>,
+    buffer: Uint8Array,
+    offset: number,
+    length: number,
+    position: number
+  ): number;
+  close<Info>(stream: FSStream<Info>): void;
   isFile: (mode: number) => boolean;
   readdir: (path: string) => string[];
   genericErrors: { 44: Error };
@@ -86,13 +100,18 @@ interface FSStreamOps<Info> {
   ) => number;
 }
 
+interface FSMount {
+  type: EmscriptenFS<any>;
+}
+
 interface FSNode<Info> {
   id: number;
   usedBytes: number;
   mode: number;
   modtime: number;
   node_ops: FSNodeOps<Info>;
   stream_ops: FSStreamOps<Info>;
+  mount: FSMount;
   info: Info;
   contentsOffset?: number | undefined;
   tree?: MetadataDirInfo;

src/workerd/server/tests/python/pytest/pytest.wd-test

@@ -9,6 +9,7 @@ const unitTests :Workerd.Config = (
           (name = "tests/test_env.py", pythonModule = embed "pytest/tests/test_env.py"),
           (name = "tests/test_fs.py", pythonModule = embed "pytest/tests/test_fs.py"),
           (name = "tests/test_import_from_javascript.py", pythonModule = embed "pytest/tests/test_import_from_javascript.py"),
+          (name = "tests/test_dynlib_loading.py", pythonModule = embed "pytest/tests/test_dynlib_loading.py"),
           %PYTHON_VENDORED_MODULES%
         ],
         compatibilityFlags = [

src/workerd/server/tests/python/pytest/tests/test_dynlib_loading.py

@@ -0,0 +1,36 @@
+from pathlib import Path
+
+import pytest
+
+
+def use(x):
+    pass
+
+
+def test_dynlib_loading(tmp_path, monkeypatch):
+    # fmt: off
+    Path(tmp_path / "a.so").write_bytes(
+        bytes(
+            [
+                0x00, 0x61, 0x73, 0x6d, 0x01, 0x00, 0x00, 0x00, 0x00, 0x0f, 0x08, 0x64,
+                0x79, 0x6c, 0x69, 0x6e, 0x6b, 0x2e, 0x30, 0x01, 0x04, 0x00, 0x00, 0x00,
+                0x00, 0x01, 0x04, 0x01, 0x60, 0x00, 0x00, 0x02, 0x38, 0x03, 0x03, 0x65,
+                0x6e, 0x76, 0x06, 0x6d, 0x65, 0x6d, 0x6f, 0x72, 0x79, 0x02, 0x00, 0x00,
+                0x03, 0x65, 0x6e, 0x76, 0x0d, 0x5f, 0x5f, 0x6d, 0x65, 0x6d, 0x6f, 0x72,
+                0x79, 0x5f, 0x62, 0x61, 0x73, 0x65, 0x03, 0x7f, 0x00, 0x03, 0x65, 0x6e,
+                0x76, 0x0c, 0x5f, 0x5f, 0x74, 0x61, 0x62, 0x6c, 0x65, 0x5f, 0x62, 0x61,
+                0x73, 0x65, 0x03, 0x7f, 0x00, 0x03, 0x02, 0x01, 0x00, 0x07, 0x15, 0x01,
+                0x11, 0x5f, 0x5f, 0x77, 0x61, 0x73, 0x6d, 0x5f, 0x63, 0x61, 0x6c, 0x6c,
+                0x5f, 0x63, 0x74, 0x6f, 0x72, 0x73, 0x00, 0x00, 0x0a, 0x04, 0x01, 0x02,
+                0x00, 0x0b
+            ]
+        )
+    )
+    # fmt: on
+    monkeypatch.syspath_prepend(tmp_path)
+    with pytest.raises(
+        ImportError, match="Can only load shared libraries from read only file systems"
+    ):
+        import a
+
+        use(a)