Avoid executing ELF files directly

.clang-tidy

@@ -0,0 +1,2 @@
+Checks: '-clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,-clang-analyzer-security.insecureAPI.strcpy,-clang-analyzer-valist.Uninitialized'
+

.github/workflows/ci.yml

@@ -0,0 +1,28 @@
+name: CI
+
+on:
+  push:
+    branches:
+      - '*'
+  pull_request:
+
+jobs:
+  check:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: Homebrew/actions/setup-homebrew@master
+      - run: brew install clang-format
+      - run: make
+      - run: make check
+      - run: make unit-test
+
+  actionlint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4 
+      - name: Download actionlint
+        id: get_actionlint
+        run: bash <(curl https://raw.githubusercontent.com/rhysd/actionlint/main/scripts/download-actionlint.bash)
+      - name: Check workflow files
+        run: ${{ steps.get_actionlint.outputs.executable }} -color

.gitignore

@@ -1,3 +1,6 @@
 *.so
 *.o
 *-actual
+*.swo
+*.swp
+test-binary

Makefile

@@ -1,20 +1,35 @@
-TERMUX_PREFIX := /data/data/com.termux/files/usr
 TERMUX_BASE_DIR := /data/data/com.termux/files
-CFLAGS += -Wall -Wextra -Werror -Oz
+CFLAGS += -Wall -Wextra -Werror -Wshadow -O2
+C_SOURCE := termux-exec.c exec-variants.c
+CLANG_FORMAT := clang-format --sort-includes --style="{ColumnLimit: 120}" $(C_SOURCE)
+CLANG_TIDY := clang-tidy
 
-libtermux-exec.so: termux-exec.c
-	$(CC) $(CFLAGS) $(LDFLAGS) termux-exec.c -DTERMUX_PREFIX=\"$(TERMUX_PREFIX)\" -DTERMUX_BASE_DIR=\"$(TERMUX_BASE_DIR)\" -shared -fPIC -o libtermux-exec.so
+libtermux-exec.so: $(C_SOURCE)
+	$(CC) $(CFLAGS) $(LDFLAGS) $(C_SOURCE) -DTERMUX_PREFIX=\"$(TERMUX_PREFIX)\" -DTERMUX_BASE_DIR=\"$(TERMUX_BASE_DIR)\" -shared -fPIC -o libtermux-exec.so
+
+clean:
+	rm -f libtermux-exec.so tests/*-actual test-binary
 
 install: libtermux-exec.so
 	install libtermux-exec.so $(DESTDIR)$(PREFIX)/lib/libtermux-exec.so
 
 uninstall:
 	rm -f $(DESTDIR)$(PREFIX)/lib/libtermux-exec.so
 
-test: libtermux-exec.so
+on-device-tests: libtermux-exec.so
 	@LD_PRELOAD=${CURDIR}/libtermux-exec.so ./run-tests.sh
 
-clean:
-	rm -f libtermux-exec.so tests/*-actual
+format:
+	$(CLANG_FORMAT) -i $(C_SOURCE)
+
+check:
+	$(CLANG_FORMAT) --dry-run $(C_SOURCE)
+	$(CLANG_TIDY) -warnings-as-errors='*' $(C_SOURCE) --
+
+test-binary: $(C_SOURCE)
+	$(CC) $(CFLAGS) $(LDFLAGS) $(C_SOURCE) -g -fsanitize=address -fno-omit-frame-pointer -DUNIT_TEST=1 -o test-binary
+
+unit-test: test-binary
+	./test-binary
 
-.PHONY: clean install test uninstall
+.PHONY: clean install uninstall test format check-format test

README.md

@@ -1,15 +1,74 @@
 # termux-exec
-A `execve()` wrapper to fix problem with shebangs when running in Termux.
+A `execve()` wrapper to fix two problems with exec-ing files in Termux.
 
-# Problem
+# Problem 1: Cannot execute files not part of the APK
+Android 10 started blocking executing files under the app data directory, as
+that is a [W^X](https://en.wikipedia.org/wiki/W%5EX) violation - files should be either
+writeable or executable, but not both. Resources:
+
+- [Google Android issue](https://issuetracker.google.com/issues/128554619)
+- [Termux: No more exec from data folder on targetAPI >= Android Q](https://github.com/termux/termux-app/issues/1072)
+- [Termux: Revisit the Android W^X problem](https://github.com/termux/termux-app/issues/2155)
+
+While there is merit in that general principle, this prevents using Termux and Android
+as a general computing device, where it should be possible for users to create executable
+scripts and binaries.
+
+# Solution 1: Cannot execute files not part of the APK
+Create an `execve()` wrapper that instead of exec-ing an ELF file directly, executes
+`/system/bin/linker64 /path/to/elf`. Explanation follows below.
+
+On Linux, the kernel is normally responsible for loading both the executable and the
+[dynamic linker](https://en.wikipedia.org/wiki/Dynamic_linker). The executable is invoked
+by filename with `execve()`. The kernel loads the executable into the process, and looks
+for a `PT_INTERP` entry in its ELF Program Headers; this specifies the filename of the
+dynamic linker (`/system/bin/linker64` for 64-bit Android). This entry exists for
+dynamically linked executables.
+
+There is another way to load the two ELF objects: the dynamic linker can be invoked directly
+with `execve()`. If passed the filename of an executable, the dynamic linker will load the
+executable itself. So, instead of executing `path/to/mybinary`, it's possible to execute
+`/system/bin/linker64 /absolute/path/to/mybinary` (the linker needs an absolute path).
+
+This is what `termux-exec` does to circumvent the block on executing files in the data
+directory - the kernel sees only `/system/bin/linker64` being executed.
+
+This also means that we need to extract shebangs. So for example, a call to execute:
+
+```sh
+./path/to/myscript.sh <args>
+```
+
+where the script has a `#!/path/to/interpreter` shebang, is replaced with:
+
+```sh
+/system/bin/linker64 /path/to/interpreter ./path/to/myscript.sh <args>
+```
+
+Implications:
+
+- It's important that `LD_PRELOAD` is kept - see e.g. [this change in sshd](https://github.com/termux/termux-packages/pull/18069).
+We could also consider patching this exec interception into the build process of termux packages, so `LD_PRELOAD` would not be necessary for packages built by the termux-packages repository.
+
+- The executable will be `/system/bin/linker64`. So some programs that inspects the executable name (on itself or other programs) needs to be changed. See [this llvm driver change](https://github.com/termux/termux-packages/pull/18074) and [this pgrep/pkill change](https://github.com/termux/termux-packages/pull/18075).
+
+**NOTE**: The above example used `/system/bin/linker64` - on 32-bit systems, the corresponding
+path is `/system/bin/linker`.
+
+**NOTE**: While this circumvents the technical restriction, it still might be considered
+violating [Google Play policy](https://support.google.com/googleplay/android-developer/answer/9888379).
+So this workaround is not guaranteed to enable Play store distribution of Termux - but it's
+worth an attempt, and regardless of Play store distribution, updating the targetSdk is necessary.
+
+# Problem 2: Shebang paths
 A lot of Linux software is written with the assumption that `/bin/sh`, `/usr/bin/env`
 and similar file exists. This is not the case on Android where neither `/bin/` nor `/usr/`
 exists.
 
 When building packages for Termux those hard-coded assumptions are patched away - but this
 does not help with installing scripts and programs from other sources than Termux packages.
 
-# Solution
+# Solution 2: Shebang paths
 Create an `execve()` wrapper that rewrites calls to execute files under `/bin/` and `/usr/bin`
 into the matching Termux executables under `$PREFIX/bin/` and inject that into processes
 using `LD_PRELOAD`.
@@ -22,3 +81,6 @@ using `LD_PRELOAD`.
 # Where is LD_PRELOAD set?
 The `$PREFIX/bin/login` program which is used to create new Termux sessions checks for
 `$PREFIX/lib/libtermux-exec.so` and if so sets up `LD_PRELOAD` before launching the login shell.
+
+Soon, when making a switch to target Android 10+, this will be setup by the Termux app even before
+launching any process, as `LD_PRELOAD` will be necessary for anything non-system to execute.

exec-variants.c

@@ -0,0 +1,146 @@
+#define _GNU_SOURCE
+#include <errno.h>
+#include <paths.h>
+#include <stdarg.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+enum { ExecL, ExecLE, ExecLP };
+
+static int __exec_as_script(const char *buf, char *const *argv, char *const *envp) {
+  size_t arg_count = 1;
+  while (argv[arg_count] != NULL)
+    ++arg_count;
+
+  const char *script_argv[arg_count + 2];
+  script_argv[0] = "sh";
+  script_argv[1] = buf;
+  memcpy(script_argv + 2, argv + 1, arg_count * sizeof(char *));
+  return execve("/data/data/com.termux/files/usr/bin/sh", (char **const)script_argv, envp);
+}
+
+int execv(const char *name, char *const *argv) { return execve(name, argv, environ); }
+
+int execvp(const char *name, char *const *argv) { return execvpe(name, argv, environ); }
+
+int execvpe(const char *name, char *const *argv, char *const *envp) {
+  // if (name == NULL || *name == '\0') { errno = ENOENT; return -1; }
+
+  // If it's an absolute or relative path name, it's easy.
+  if (strchr(name, '/') && execve(name, argv, envp) == -1) {
+    if (errno == ENOEXEC)
+      return __exec_as_script(name, argv, envp);
+    return -1;
+  }
+
+  // Get the path we're searching.
+  const char *path = getenv("PATH");
+  if (path == NULL)
+    path = _PATH_DEFPATH;
+
+  // Make a writable copy.
+  size_t len = strlen(path) + 1;
+  char writable_path[len];
+  memcpy(writable_path, path, len);
+
+  bool saw_EACCES = false;
+
+  // Try each element of $PATH in turn...
+  char *strsep_buf = writable_path;
+  const char *dir;
+  while ((dir = strsep(&strsep_buf, ":"))) {
+    // It's a shell path: double, leading and trailing colons
+    // mean the current directory.
+    if (*dir == '\0')
+      dir = ".";
+
+    size_t dir_len = strlen(dir);
+    size_t name_len = strlen(name);
+
+    char buf[dir_len + 1 + name_len + 1];
+    mempcpy(mempcpy(mempcpy(buf, dir, dir_len), "/", 1), name, name_len + 1);
+
+    execve(buf, argv, envp);
+    switch (errno) {
+    case EISDIR:
+    case ELOOP:
+    case ENAMETOOLONG:
+    case ENOENT:
+    case ENOTDIR:
+      break;
+    case ENOEXEC:
+      return __exec_as_script(buf, argv, envp);
+      return -1;
+    case EACCES:
+      saw_EACCES = true;
+      break;
+    default:
+      return -1;
+    }
+  }
+  if (saw_EACCES)
+    errno = EACCES;
+  return -1;
+}
+
+static int __execl(int variant, const char *name, const char *argv0, va_list ap) {
+  // Count the arguments.
+  va_list count_ap;
+  va_copy(count_ap, ap);
+  size_t n = 1;
+  while (va_arg(count_ap, char *) != NULL) {
+    ++n;
+  }
+  va_end(count_ap);
+
+  // Construct the new argv.
+  char *argv[n + 1];
+  argv[0] = (char *)argv0;
+  n = 1;
+  while ((argv[n] = va_arg(ap, char *)) != NULL) {
+    ++n;
+  }
+
+  // Collect the argp too.
+  char **argp = (variant == ExecLE) ? va_arg(ap, char **) : environ;
+
+  va_end(ap);
+
+  return (variant == ExecLP) ? execvp(name, argv) : execve(name, argv, argp);
+}
+
+int execl(const char *name, const char *arg, ...) {
+  va_list ap;
+  va_start(ap, arg);
+  int result = __execl(ExecL, name, arg, ap);
+  va_end(ap);
+  return result;
+}
+
+int execle(const char *name, const char *arg, ...) {
+  va_list ap;
+  va_start(ap, arg);
+  int result = __execl(ExecLE, name, arg, ap);
+  va_end(ap);
+  return result;
+}
+
+int execlp(const char *name, const char *arg, ...) {
+  va_list ap;
+  va_start(ap, arg);
+  int result = __execl(ExecLP, name, arg, ap);
+  va_end(ap);
+  return result;
+}
+
+int fexecve(int fd, char *const *argv, char *const *envp) {
+  char buf[40];
+  snprintf(buf, sizeof(buf), "/proc/self/fd/%d", fd);
+  execve(buf, argv, envp);
+  if (errno == ENOENT)
+    errno = EBADF;
+  return -1;
+}