@expo/build-tools 24.3.0 → 24.4.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@expo/build-tools",
-  "version": "24.3.0",
+  "version": "24.4.0",
   "bugs": "https://github.com/expo/eas-cli/issues",
   "license": "BUSL-1.1",
   "author": "Expo <support@expo.io>",
@@ -24,6 +24,8 @@
     "prebuild": "yarn gql",
     "build": "tsc",
     "build:record-sim": "mkdir -p bin && record_sim_bin_path=$(swift build -c release --package-path resources/record-sim --build-path resources/record-sim/.build --show-bin-path) && swift build -c release --package-path resources/record-sim --build-path resources/record-sim/.build && cp \"$record_sim_bin_path/record-sim\" bin/record-sim && chmod +x bin/record-sim",
+    "build:egress-guard": "resources/egress-guard/build.sh",
+    "test:egress-guard": "resources/egress-guard/tests/run-policy-tests.sh && resources/egress-guard/tests/run-guard-tests.sh",
     "typecheck": "tsc",
     "generate-appium-commands": "mise exec node@22.22.0 -- node scripts/generate-appium-commands.js",
     "prepack": "rimraf dist \"*.tsbuildinfo\" && yarn gql && tsc -p tsconfig.build.json",
@@ -104,5 +106,5 @@
     "typescript": "^5.5.4",
     "uuid": "^9.0.1"
   },
-  "gitHead": "66cdccb0e7c3ec47054c3f093c25b9799d60ff36"
+  "gitHead": "e7efe143366b0c908814e61dd3fb4b207600b773"
 }

package/resources/egress-guard/README.md

@@ -0,0 +1,83 @@
+# egress-guard
+
+A small dylib injected into every process the iOS Simulator launches during a
+`--egress local` session. It interposes the socket calls that open outbound
+traffic (`connect`, `connectx`, `sendto`, `sendmsg`, and the `$NOCANCEL`
+variants of the last three that libsystem exports alongside them) and refuses
+any destination that is not loopback. A sockaddr whose family was left
+`AF_UNSPEC` is classified as the family its length implies, which is how the
+kernel treats it for TCP `connect` and IPv4 sends; on a datagram socket
+`connect(AF_UNSPEC)` is left to the kernel, which dissolves the association. The system proxy and the `--egress-allow`
+forwards live on loopback, so everything that honors the proxy keeps working
+and everything that bypasses it fails with `ECONNREFUSED` in the process that
+tried, with the calling frameworks recorded.
+
+Why interposing: simulator processes share the worker's uid, so a packet
+filter cannot tell them apart from the worker's own traffic. dyld interposing
+from an inserted library reaches calls made inside Apple's own frameworks
+(CFNetwork, Network.framework), and `launchctl setenv DYLD_INSERT_LIBRARIES`
+inside the simulator makes launchd hand the library to every process it
+spawns. The simulator does not enforce library validation, which is what makes
+this possible there and nowhere else.
+
+Configuration comes from the environment the simulator's launchd provides:
+
+- `EAS_EGRESS_GUARD_LOG`: file the guard appends events to (host path).
+- `EAS_EGRESS_GUARD_MODE`: `block` (default, and what any unknown value means)
+  or `log` (observe only).
+
+Each event is one tab-separated line:
+
+```
+eas-egress-guard\t<process>\t<pid>\t<function>\t<blocked|logged>\t<peer>\t<caller>,<caller>...
+```
+
+A process lists each `(function, peer)` once, up to 128 of them. When that
+table fills it writes one more line with function `overflow` and the limit as
+the peer; further distinct destinations are still refused but not listed.
+
+Layout:
+
+- `policy.c` / `policy.h`: classification, mode, formatting, per-process
+  dedupe. Pure C, host-testable.
+- `guard.c`: the interposers and the constructor that reads the environment.
+- `check.c`: `egress-guard-check`, run inside the simulator right after the
+  guard is installed. Exits non-zero unless the library is loaded in a fresh
+  process and behaves as the mode says on `connect`, `connect$NOCANCEL` and a
+  `connect` with an `AF_UNSPEC` sockaddr; the worker fails the session on that.
+- `tests/policy_test.c`: host unit tests, `tests/run-policy-tests.sh`.
+- `tests/guard_test.c`, `tests/guard_insert_test.c`: host tests of the
+  interposers themselves, `tests/run-guard-tests.sh`. The first compiles
+  `guard.c` in and forces the lock schedules (a signal handler making a socket
+  call inside the critical section, a fork while another thread holds the lock,
+  contention); the second inserts the built library like the simulator does and
+  checks descriptor reuse, the `$NOCANCEL` symbols, `AF_UNSPEC` shapes and the
+  UDP disconnect. Nothing in them opens a socket to a non-loopback address.
+- `tests/nettest.swift`: probe binary the simulator end-to-end test runs inside
+  a device; exercises URLSession, Network.framework, BSD TCP and UDP, DNS.
+- `build.sh`: universal simulator dylib into `packages/build-tools/bin/`.
+
+Installation order matters. `simctl` forwards every `SIMCTL_CHILD_`-prefixed
+variable of its own environment to the process it starts, and for `simctl
+boot` that process is the simulator's launchd, so the worker boots with the
+guard and proxy variables in that form and every process of the boot inherits
+them (measured: 176 of 176). `launchctl setenv` after boot is kept for a
+device that was already booted, but by then the boot's own processes have
+started without it. It also cannot change a variable the boot already carried:
+a device booted with one guard configuration keeps it until it is shut down,
+which is why the worker boots devices itself and every configuration in the
+end-to-end test gets its own boot. The self-check runs once boot completes and is followed
+by a coverage report from `lsof`, listing any process without the library. Both the dylib and the check binary are built by
+`packages/worker/package.sh` for the iOS worker tarball and by the
+`test-egress-guard` EAS workflow, not committed.
+
+Failure semantics: a missing library, a failed `launchctl setenv`, or a failed
+self-check fails the session, since a `--egress local` session without the
+guard would silently leak. Only an unwritable event log is a warning, because
+refusals still happen and only the reporting is lost. Recording is best-effort
+in the process too: the event log is opened per event rather than held open
+(a process that closes descriptors it does not own, as `launchd_sim` does,
+would otherwise make the guard write into whatever reused the number), and a
+dedupe slot that cannot be taken within a bounded number of attempts (a signal
+handler re-entering the guard, a fork child inheriting a held lock) skips the
+event rather than blocking or aborting the process.

package/resources/egress-guard/build.sh

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+# Builds the guard as a universal (arm64 + x86_64) iOS Simulator dylib into
+# packages/build-tools/bin/egress-guard.dylib. Requires Xcode.
+set -euo pipefail
+cd "$(dirname "$0")"
+# Output directory; defaults to packages/build-tools/bin.
+bin_dir="${1:-../../bin}"
+mkdir -p "$bin_dir"
+out=$(mktemp -d)
+sdk=$(xcrun --sdk iphonesimulator --show-sdk-path)
+for arch in arm64 x86_64; do
+  xcrun --sdk iphonesimulator clang \
+    -std=c11 -Wall -Wextra -Werror -O2 \
+    -target "$arch-apple-ios15.0-simulator" -isysroot "$sdk" \
+    -dynamiclib -install_name @rpath/egress-guard.dylib \
+    -o "$out/egress-guard-$arch.dylib" policy.c guard.c
+done
+lipo -create -output "$bin_dir/egress-guard.dylib" "$out/egress-guard-arm64.dylib" "$out/egress-guard-x86_64.dylib"
+lipo -info "$bin_dir/egress-guard.dylib"
+
+# The self-check the worker runs inside the simulator after installing the guard.
+for arch in arm64 x86_64; do
+  xcrun --sdk iphonesimulator clang \
+    -std=c11 -Wall -Wextra -Werror -O2 \
+    -target "$arch-apple-ios15.0-simulator" -isysroot "$sdk" \
+    -o "$out/egress-guard-check-$arch" check.c
+done
+lipo -create -output "$bin_dir/egress-guard-check" "$out/egress-guard-check-arm64" "$out/egress-guard-check-x86_64"
+chmod +x "$bin_dir/egress-guard-check"
+lipo -info "$bin_dir/egress-guard-check"

package/resources/egress-guard/check.c

@@ -0,0 +1,112 @@
+// Self-check run inside a simulator right after the guard is installed:
+// proves the guard library is loaded in a freshly spawned process and that it
+// behaves as the requested mode says, on every connect entry point the guard
+// covers. Exit 0 on success; 2 when the library is not loaded; 3 when it is
+// loaded but did not behave; 1 on usage errors.
+//
+//   egress-guard-check --mode block|log
+#include <arpa/inet.h>
+#include <dlfcn.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <mach-o/dyld.h>
+#include <netinet/in.h>
+#include <stdio.h>
+#include <string.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+// TEST-NET-1 (RFC 5737): never routed, so without the guard a non-blocking
+// connect reports EINPROGRESS and nothing ever answers.
+#define PROBE_ADDRESS "192.0.2.1"
+#define PROBE_PORT 9
+
+typedef int (*connect_fn)(int, const struct sockaddr *, socklen_t);
+
+static int guard_loaded(void) {
+  for (uint32_t i = 0; i < _dyld_image_count(); i++) {
+    const char *name = _dyld_get_image_name(i);
+    if (name != NULL && strstr(name, "egress-guard.dylib") != NULL) {
+      return 1;
+    }
+  }
+  return 0;
+}
+
+// One non-blocking connect to the probe address through `entry`, with the
+// sockaddr family the caller wants. Returns the connect result; errno is
+// stored in *err.
+static int probe(connect_fn entry, sa_family_t family, int *err) {
+  int fd = socket(AF_INET, SOCK_STREAM, 0);
+  if (fd < 0) {
+    *err = errno;
+    return -2;
+  }
+  fcntl(fd, F_SETFL, fcntl(fd, F_GETFL, 0) | O_NONBLOCK);
+  struct sockaddr_in to;
+  memset(&to, 0, sizeof to);
+  to.sin_family = family;
+  to.sin_len = sizeof to;
+  to.sin_port = htons(PROBE_PORT);
+  inet_pton(AF_INET, PROBE_ADDRESS, &to.sin_addr);
+  int rc = entry(fd, (struct sockaddr *)&to, sizeof to);
+  *err = errno;
+  close(fd);
+  return rc;
+}
+
+int main(int argc, char **argv) {
+  const char *mode = "block";
+  if (argc == 3 && strcmp(argv[1], "--mode") == 0) {
+    mode = argv[2];
+  } else if (argc != 1) {
+    fprintf(stderr, "usage: egress-guard-check [--mode block|log]\n");
+    return 1;
+  }
+  if (!guard_loaded()) {
+    printf("egress-guard-check: the guard library is not loaded in this process\n");
+    return 2;
+  }
+  connect_fn connect_nocancel = (connect_fn)dlsym(RTLD_DEFAULT, "connect$NOCANCEL");
+  if (connect_nocancel == NULL) {
+    printf("egress-guard-check: connect$NOCANCEL is not exported in this process\n");
+    return 3;
+  }
+
+  // The plain call, the non-cancelable variant, and a sockaddr whose family
+  // was left AF_UNSPEC, which the kernel connects as IPv4.
+  const struct {
+    const char *name;
+    connect_fn entry;
+    sa_family_t family;
+  } probes[] = {
+      {"connect", connect, AF_INET},
+      {"connect$NOCANCEL", connect_nocancel, AF_INET},
+      {"connect with AF_UNSPEC", connect, AF_UNSPEC},
+  };
+  int block = strcmp(mode, "block") == 0;
+  for (size_t i = 0; i < sizeof probes / sizeof probes[0]; i++) {
+    int err = 0;
+    int rc = probe(probes[i].entry, probes[i].family, &err);
+    if (rc == -2) {
+      printf("egress-guard-check: socket() failed: %s\n", strerror(err));
+      return 3;
+    }
+    if (block && !(rc == -1 && err == ECONNREFUSED)) {
+      printf("egress-guard-check: guard loaded but a non-loopback %s was not refused (rc=%d, errno=%d %s)\n",
+             probes[i].name, rc, err, strerror(err));
+      return 3;
+    }
+    if (!block && !(rc == 0 || (rc == -1 && err == EINPROGRESS))) {
+      printf("egress-guard-check: guard loaded in log mode but %s failed (errno=%d %s)\n",
+             probes[i].name, err, strerror(err));
+      return 3;
+    }
+  }
+  if (block) {
+    printf("egress-guard-check: guard loaded; non-loopback connections are refused\n");
+  } else {
+    printf("egress-guard-check: guard loaded in log mode; connections pass through\n");
+  }
+  return 0;
+}

package/resources/egress-guard/guard.c

@@ -0,0 +1,274 @@
+// The local egress guard: interposes the calls that open outbound traffic in
+// every simulator process it is inserted into, refuses non-loopback
+// destinations in block mode, and records one event per destination per
+// process. See README.md and policy.h.
+#include "policy.h"
+
+#include <dlfcn.h>
+#include <errno.h>
+#include <execinfo.h>
+#include <fcntl.h>
+#include <os/lock.h>
+#include <sched.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/socket.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#define EG_LOG_ENV "EAS_EGRESS_GUARD_LOG"
+#define EG_MODE_ENV "EAS_EGRESS_GUARD_MODE"
+#define EG_MAX_CALLERS 6
+#define EG_CALLER_LENGTH 64
+#define EG_LOCK_ATTEMPTS 64
+
+static int eg_initialized = 0;
+static void *eg_own_image = NULL;
+static char eg_log_path[1024];
+static eg_mode_t eg_mode = EG_MODE_BLOCK;
+static eg_seen_t eg_seen;
+static int eg_overflow_reported = 0;
+static os_unfair_lock eg_lock = OS_UNFAIR_LOCK_INIT;
+
+static void eg_init(void) {
+  if (eg_initialized) {
+    return;
+  }
+  eg_mode = eg_parse_mode(getenv(EG_MODE_ENV));
+  Dl_info self;
+  if (dladdr((const void *)eg_init, &self)) {
+    eg_own_image = self.dli_fbase;
+  }
+  const char *log_path = getenv(EG_LOG_ENV);
+  // The path is kept rather than an open descriptor: a process that closes
+  // descriptors it does not know about (launchd_sim does) would otherwise
+  // leave the guard writing into whatever file reused the number.
+  if (log_path != NULL && log_path[0] != '\0' && strlen(log_path) < sizeof eg_log_path) {
+    strcpy(eg_log_path, log_path);
+  }
+  eg_initialized = 1;
+}
+
+__attribute__((constructor)) static void eg_constructor(void) { eg_init(); }
+
+// Image names of the frames above the guard's own, innermost first, with
+// consecutive repeats collapsed: "Network,CFNetwork,MyApp".
+__attribute__((noinline)) static int eg_collect_callers(char names[EG_MAX_CALLERS][EG_CALLER_LENGTH]) {
+  void *frames[EG_MAX_CALLERS + 6];
+  int frame_count = backtrace(frames, EG_MAX_CALLERS + 6);
+  int count = 0;
+  for (int i = 1; i < frame_count && count < EG_MAX_CALLERS; i++) {
+    Dl_info info;
+    const char *image = "?";
+    if (dladdr(frames[i], &info)) {
+      if (info.dli_fbase == eg_own_image) {
+        continue;
+      }
+      if (info.dli_fname != NULL) {
+        const char *slash = strrchr(info.dli_fname, '/');
+        image = slash ? slash + 1 : info.dli_fname;
+      }
+    }
+    if (count > 0 && strcmp(names[count - 1], image) == 0) {
+      continue;
+    }
+    strncpy(names[count], image, EG_CALLER_LENGTH - 1);
+    names[count][EG_CALLER_LENGTH - 1] = '\0';
+    count++;
+  }
+  return count;
+}
+
+// A lock held by this same thread (a signal handler making a socket call
+// while the interrupted call held it) or by a thread that did not survive
+// fork() never becomes available and os_unfair_lock_lock aborts the process
+// in both cases. Give up on recording the event instead; refusal does not
+// depend on it. Ordinary contention over the microsecond-long critical
+// section resolves within the first attempts.
+static int eg_lock_bounded(void) {
+  for (int i = 0; i < EG_LOCK_ATTEMPTS; i++) {
+    if (os_unfair_lock_trylock(&eg_lock)) {
+      return 1;
+    }
+    sched_yield();
+  }
+  return 0;
+}
+
+// One line in the event log. O_APPEND keeps whole lines intact across
+// processes writing concurrently; the log is opened per event, at most
+// EG_SEEN_CAPACITY + 1 times per process.
+__attribute__((noinline)) static void eg_write_event(const char *function, const char *action,
+                                                     const char *peer, int with_callers) {
+  if (eg_log_path[0] == '\0') {
+    return;
+  }
+  char names[EG_MAX_CALLERS][EG_CALLER_LENGTH];
+  const char *callers[EG_MAX_CALLERS];
+  int caller_count = with_callers ? eg_collect_callers(names) : 0;
+  for (int i = 0; i < caller_count; i++) {
+    callers[i] = names[i];
+  }
+  char line[1024];
+  int n = eg_format_event(line, sizeof line, getprogname(), getpid(), function, action, peer,
+                          callers, caller_count);
+  if (n <= 0) {
+    return;
+  }
+  int fd = open(eg_log_path, O_WRONLY | O_APPEND | O_CREAT | O_CLOEXEC, 0644);
+  if (fd >= 0) {
+    (void)write(fd, line, (size_t)n);
+    close(fd);
+  }
+}
+
+// Returns 1 when the call must be refused.
+__attribute__((noinline)) static int eg_handle(const char *function, const struct sockaddr *sa,
+                                              socklen_t len) {
+  eg_class_t cls = eg_classify(sa, len);
+  if (cls != EG_REMOTE) {
+    return 0;
+  }
+  eg_init();
+  int deny = eg_should_deny(eg_mode, cls);
+
+  char peer[96];
+  if (eg_format_peer(sa, len, peer, sizeof peer) != 0) {
+    strncpy(peer, "?", sizeof peer);
+  }
+  char key[EG_SEEN_KEY_LENGTH];
+  snprintf(key, sizeof key, "%s %s", function, peer);
+  int fresh = 0;
+  int overflowed = 0;
+  if (eg_lock_bounded()) {
+    fresh = eg_seen_insert(&eg_seen, key);
+    if (!fresh && eg_seen.overflow > 0 && !eg_overflow_reported) {
+      eg_overflow_reported = 1;
+      overflowed = 1;
+    }
+    os_unfair_lock_unlock(&eg_lock);
+  }
+
+  const char *action = deny ? "blocked" : "logged";
+  if (fresh) {
+    eg_write_event(function, action, peer, 1);
+  } else if (overflowed) {
+    // Once per process: the table is full, so further distinct destinations
+    // are still refused but no longer listed.
+    char limit[64];
+    snprintf(limit, sizeof limit, "%d distinct destinations", EG_SEEN_CAPACITY);
+    eg_write_event(EG_OVERFLOW_FUNCTION, action, limit, 0);
+  }
+  return deny;
+}
+
+// connect() with an AF_UNSPEC address on a datagram socket dissolves the
+// association and sends nothing; the kernel answers EAFNOSUPPORT. Leave that
+// to the kernel. On a stream socket the kernel connects to the address the
+// length implies, which eg_handle classifies.
+static int eg_dissolves_association(int fd, const struct sockaddr *sa, socklen_t len) {
+  if (sa == NULL || len < 2 || sa->sa_family != AF_UNSPEC) {
+    return 0;
+  }
+  int type = 0;
+  socklen_t type_len = sizeof type;
+  return getsockopt(fd, SOL_SOCKET, SO_TYPE, &type, &type_len) == 0 && type == SOCK_DGRAM;
+}
+
+typedef int (*eg_connect_fn)(int, const struct sockaddr *, socklen_t);
+typedef ssize_t (*eg_sendto_fn)(int, const void *, size_t, int, const struct sockaddr *,
+                                socklen_t);
+typedef ssize_t (*eg_sendmsg_fn)(int, const struct msghdr *, int);
+
+// The non-cancelable variants libsystem_kernel exports next to the cancelable
+// ones; dyld interposing is per symbol, so each needs its own entry.
+extern int eg_connect_nocancel_original(int, const struct sockaddr *, socklen_t) __asm__(
+    "_connect$NOCANCEL");
+extern ssize_t eg_sendto_nocancel_original(int, const void *, size_t, int,
+                                           const struct sockaddr *, socklen_t) __asm__(
+    "_sendto$NOCANCEL");
+extern ssize_t eg_sendmsg_nocancel_original(int, const struct msghdr *, int) __asm__(
+    "_sendmsg$NOCANCEL");
+
+static int eg_connect_through(eg_connect_fn original, int fd, const struct sockaddr *sa,
+                              socklen_t len) {
+  if (!eg_dissolves_association(fd, sa, len) && eg_handle("connect", sa, len)) {
+    errno = ECONNREFUSED;
+    return -1;
+  }
+  return original(fd, sa, len);
+}
+
+static int eg_connect(int fd, const struct sockaddr *sa, socklen_t len) {
+  return eg_connect_through(connect, fd, sa, len);
+}
+
+static int eg_connect_nocancel(int fd, const struct sockaddr *sa, socklen_t len) {
+  return eg_connect_through(eg_connect_nocancel_original, fd, sa, len);
+}
+
+static int eg_connectx(int s, const sa_endpoints_t *endpoints, sae_associd_t associd,
+                       unsigned int flags, const struct iovec *iov, unsigned int iovcnt,
+                       size_t *len, sae_connid_t *connid) {
+  if (endpoints != NULL && endpoints->sae_dstaddr != NULL &&
+      !eg_dissolves_association(s, endpoints->sae_dstaddr, endpoints->sae_dstaddrlen) &&
+      eg_handle("connectx", endpoints->sae_dstaddr, endpoints->sae_dstaddrlen)) {
+    errno = ECONNREFUSED;
+    return -1;
+  }
+  return connectx(s, endpoints, associd, flags, iov, iovcnt, len, connid);
+}
+
+static ssize_t eg_sendto_through(eg_sendto_fn original, int fd, const void *buf, size_t n,
+                                 int flags, const struct sockaddr *sa, socklen_t len) {
+  if (sa != NULL && eg_handle("sendto", sa, len)) {
+    errno = ECONNREFUSED;
+    return -1;
+  }
+  return original(fd, buf, n, flags, sa, len);
+}
+
+static ssize_t eg_sendto(int fd, const void *buf, size_t n, int flags, const struct sockaddr *sa,
+                         socklen_t len) {
+  return eg_sendto_through(sendto, fd, buf, n, flags, sa, len);
+}
+
+static ssize_t eg_sendto_nocancel(int fd, const void *buf, size_t n, int flags,
+                                  const struct sockaddr *sa, socklen_t len) {
+  return eg_sendto_through(eg_sendto_nocancel_original, fd, buf, n, flags, sa, len);
+}
+
+static ssize_t eg_sendmsg_through(eg_sendmsg_fn original, int fd, const struct msghdr *msg,
+                                  int flags) {
+  if (msg != NULL && msg->msg_name != NULL && msg->msg_namelen > 0 &&
+      eg_handle("sendmsg", (const struct sockaddr *)msg->msg_name, msg->msg_namelen)) {
+    errno = ECONNREFUSED;
+    return -1;
+  }
+  return original(fd, msg, flags);
+}
+
+static ssize_t eg_sendmsg(int fd, const struct msghdr *msg, int flags) {
+  return eg_sendmsg_through(sendmsg, fd, msg, flags);
+}
+
+static ssize_t eg_sendmsg_nocancel(int fd, const struct msghdr *msg, int flags) {
+  return eg_sendmsg_through(eg_sendmsg_nocancel_original, fd, msg, flags);
+}
+
+// dyld applies these to every image in the process, including the shared
+// cache, which is how calls made inside CFNetwork and Network.framework are
+// caught.
+__attribute__((used)) static const struct {
+  const void *replacement;
+  const void *original;
+} eg_interposers[] __attribute__((section("__DATA,__interpose"))) = {
+    {(const void *)eg_connect, (const void *)connect},
+    {(const void *)eg_connect_nocancel, (const void *)eg_connect_nocancel_original},
+    {(const void *)eg_connectx, (const void *)connectx},
+    {(const void *)eg_sendto, (const void *)sendto},
+    {(const void *)eg_sendto_nocancel, (const void *)eg_sendto_nocancel_original},
+    {(const void *)eg_sendmsg, (const void *)sendmsg},
+    {(const void *)eg_sendmsg_nocancel, (const void *)eg_sendmsg_nocancel_original},
+};

package/resources/egress-guard/policy.c

@@ -0,0 +1,164 @@
+#include "policy.h"
+
+#include <arpa/inet.h>
+#include <netinet/in.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+
+static int is_v4_local(uint32_t host_order) {
+  return (host_order >> 24) == 127 || host_order == 0;
+}
+
+// XNU accepts a sockaddr whose family is AF_UNSPEC as the family its length
+// implies: TCP connect() translates it on both socket families and IPv4
+// sendto()/sendmsg() deliver with it. Classify it the way the kernel uses it.
+static sa_family_t eg_family(const struct sockaddr *sa, socklen_t len) {
+  if (sa->sa_family == AF_UNSPEC) {
+    if (len == sizeof(struct sockaddr_in)) {
+      return AF_INET;
+    }
+    if (len == sizeof(struct sockaddr_in6)) {
+      return AF_INET6;
+    }
+  }
+  return sa->sa_family;
+}
+
+eg_class_t eg_classify(const struct sockaddr *sa, socklen_t len) {
+  if (sa == NULL || len < 2) {
+    return EG_PASSTHROUGH;
+  }
+  sa_family_t family = eg_family(sa, len);
+  if (family == AF_INET) {
+    if (len < sizeof(struct sockaddr_in)) {
+      return EG_PASSTHROUGH;
+    }
+    const struct sockaddr_in *in = (const struct sockaddr_in *)sa;
+    return is_v4_local(ntohl(in->sin_addr.s_addr)) ? EG_LOOPBACK : EG_REMOTE;
+  }
+  if (family == AF_INET6) {
+    if (len < sizeof(struct sockaddr_in6)) {
+      return EG_PASSTHROUGH;
+    }
+    const struct in6_addr *a6 = &((const struct sockaddr_in6 *)sa)->sin6_addr;
+    if (IN6_IS_ADDR_LOOPBACK(a6) || IN6_IS_ADDR_UNSPECIFIED(a6)) {
+      return EG_LOOPBACK;
+    }
+    if (IN6_IS_ADDR_V4MAPPED(a6)) {
+      uint32_t v4;
+      memcpy(&v4, &a6->s6_addr[12], sizeof v4);
+      return is_v4_local(ntohl(v4)) ? EG_LOOPBACK : EG_REMOTE;
+    }
+    return EG_REMOTE;
+  }
+  return EG_PASSTHROUGH;
+}
+
+eg_mode_t eg_parse_mode(const char *value) {
+  if (value != NULL && strcmp(value, "log") == 0) {
+    return EG_MODE_LOG;
+  }
+  return EG_MODE_BLOCK;
+}
+
+int eg_should_deny(eg_mode_t mode, eg_class_t cls) {
+  return mode == EG_MODE_BLOCK && cls == EG_REMOTE;
+}
+
+int eg_format_peer(const struct sockaddr *sa, socklen_t len, char *out, size_t n) {
+  char ip[INET6_ADDRSTRLEN];
+  int written;
+  if (sa == NULL || len < 2) {
+    return -1;
+  }
+  sa_family_t family = eg_family(sa, len);
+  if (family == AF_INET && len >= sizeof(struct sockaddr_in)) {
+    const struct sockaddr_in *in = (const struct sockaddr_in *)sa;
+    if (inet_ntop(AF_INET, &in->sin_addr, ip, sizeof ip) == NULL) {
+      return -1;
+    }
+    written = snprintf(out, n, "%s:%d", ip, ntohs(in->sin_port));
+  } else if (family == AF_INET6 && len >= sizeof(struct sockaddr_in6)) {
+    const struct sockaddr_in6 *in6 = (const struct sockaddr_in6 *)sa;
+    if (inet_ntop(AF_INET6, &in6->sin6_addr, ip, sizeof ip) == NULL) {
+      return -1;
+    }
+    written = snprintf(out, n, "[%s]:%d", ip, ntohs(in6->sin6_port));
+  } else {
+    return -1;
+  }
+  return (written < 0 || (size_t)written >= n) ? -1 : 0;
+}
+
+int eg_seen_insert(eg_seen_t *seen, const char *key) {
+  for (int i = 0; i < seen->count; i++) {
+    if (strncmp(seen->keys[i], key, EG_SEEN_KEY_LENGTH - 1) == 0) {
+      return 0;
+    }
+  }
+  if (seen->count >= EG_SEEN_CAPACITY) {
+    seen->overflow++;
+    return 0;
+  }
+  strncpy(seen->keys[seen->count], key, EG_SEEN_KEY_LENGTH - 1);
+  seen->keys[seen->count][EG_SEEN_KEY_LENGTH - 1] = '\0';
+  seen->count++;
+  return 1;
+}
+
+// Append `text` to out[*pos], replacing field separators; returns 0 when it fit.
+static int append_field(char *out, size_t n, size_t *pos, const char *text) {
+  for (const char *c = text; *c; c++) {
+    if (*pos + 1 >= n) {
+      return -1;
+    }
+    out[(*pos)++] = (*c == '\t' || *c == '\n' || *c == '\r') ? ' ' : *c;
+  }
+  return 0;
+}
+
+static int append_raw(char *out, size_t n, size_t *pos, const char *text) {
+  size_t len = strlen(text);
+  if (*pos + len + 1 > n) {
+    return -1;
+  }
+  memcpy(out + *pos, text, len);
+  *pos += len;
+  return 0;
+}
+
+int eg_format_event(char *out, size_t n, const char *progname, int pid, const char *function,
+                    const char *action, const char *peer, const char *const *callers,
+                    int caller_count) {
+  size_t pos = 0;
+  char pid_text[16];
+  snprintf(pid_text, sizeof pid_text, "%d", pid);
+  if (append_raw(out, n, &pos, EG_EVENT_PREFIX "\t") != 0 ||
+      append_field(out, n, &pos, progname ? progname : "?") != 0 ||
+      append_raw(out, n, &pos, "\t") != 0 || append_raw(out, n, &pos, pid_text) != 0 ||
+      append_raw(out, n, &pos, "\t") != 0 || append_field(out, n, &pos, function) != 0 ||
+      append_raw(out, n, &pos, "\t") != 0 || append_field(out, n, &pos, action) != 0 ||
+      append_raw(out, n, &pos, "\t") != 0 || append_field(out, n, &pos, peer) != 0 ||
+      append_raw(out, n, &pos, "\t") != 0) {
+    return -1;
+  }
+  for (int i = 0; i < caller_count && callers != NULL; i++) {
+    if (i > 0 && append_raw(out, n, &pos, ",") != 0) {
+      return -1;
+    }
+    // Commas separate callers, so they cannot appear inside one.
+    for (const char *c = callers[i]; *c; c++) {
+      char ch = (*c == ',' || *c == '\t' || *c == '\n' || *c == '\r') ? ' ' : *c;
+      if (pos + 1 >= n) {
+        return -1;
+      }
+      out[pos++] = ch;
+    }
+  }
+  if (append_raw(out, n, &pos, "\n") != 0) {
+    return -1;
+  }
+  out[pos] = '\0';
+  return (int)pos;
+}