// Copyright (c) 2012 The Chromium Authors. All rights reserved. // Use of this source code is governed by a BSD-style license that can be // found in the LICENSE file. // http://code.google.com/p/chromium/wiki/LinuxSUIDSandbox #include "sandbox/linux/suid/common/sandbox.h" #define _GNU_SOURCE #include <asm/unistd.h> #include <errno.h> #include <fcntl.h> #include <limits.h> #include <sched.h> #include <signal.h> #include <stdarg.h> #include <stdbool.h> #include <stdint.h> #include <stdio.h> #include <stdlib.h> #include <string.h> #include <sys/prctl.h> #include <sys/resource.h> #include <sys/socket.h> #include <sys/stat.h> #include <sys/time.h> #include <sys/types.h> #include <sys/vfs.h> #include <sys/wait.h> #include <unistd.h> #include "sandbox/linux/suid/common/suid_unsafe_environment_variables.h" #include "sandbox/linux/suid/linux_util.h" #include "sandbox/linux/suid/process_util.h" #if !defined(CLONE_NEWPID) #define CLONE_NEWPID 0x20000000 #endif #if !defined(CLONE_NEWNET) #define CLONE_NEWNET 0x40000000 #endif static bool DropRoot(); #define HANDLE_EINTR(x) TEMP_FAILURE_RETRY(x) static void FatalError(const char* msg, ...) __attribute__((noreturn, format(printf, 1, 2))); static void FatalError(const char* msg, ...) { va_list ap; va_start(ap, msg); vfprintf(stderr, msg, ap); fprintf(stderr, ": %s\n", strerror(errno)); fflush(stderr); va_end(ap); _exit(1); } static void ExitWithErrorSignalHandler(int signal) { const char msg[] = "\nThe setuid sandbox got signaled, exiting.\n"; if (-1 == write(2, msg, sizeof(msg) - 1)) { // Do nothing. } _exit(1); } // We will chroot() to the helper's /proc/self directory. Anything there will // not exist anymore if we make sure to wait() for the helper. // // /proc/self/fdinfo or /proc/self/fd are especially safe and will be empty // even if the helper survives as a zombie. // // There is very little reason to use fdinfo/ instead of fd/ but we are // paranoid. fdinfo/ only exists since 2.6.22 so we allow fallback to fd/ #define SAFE_DIR "/proc/self/fdinfo" #define SAFE_DIR2 "/proc/self/fd" static bool SpawnChrootHelper() { int sv[2]; if (socketpair(AF_UNIX, SOCK_STREAM, 0, sv) == -1) { perror("socketpair"); return false; } char* safedir = NULL; struct stat sdir_stat; if (!stat(SAFE_DIR, &sdir_stat) && S_ISDIR(sdir_stat.st_mode)) { safedir = SAFE_DIR; } else if (!stat(SAFE_DIR2, &sdir_stat) && S_ISDIR(sdir_stat.st_mode)) { safedir = SAFE_DIR2; } else { fprintf(stderr, "Could not find %s\n", SAFE_DIR2); return false; } const pid_t pid = syscall(__NR_clone, CLONE_FS | SIGCHLD, 0, 0, 0); if (pid == -1) { perror("clone"); close(sv[0]); close(sv[1]); return false; } if (pid == 0) { // We share our files structure with an untrusted process. As a security in // depth measure, we make sure that we can't open anything by mistake. // TODO(agl): drop CAP_SYS_RESOURCE / use SECURE_NOROOT const struct rlimit nofile = {0, 0}; if (setrlimit(RLIMIT_NOFILE, &nofile)) FatalError("Setting RLIMIT_NOFILE"); if (close(sv[1])) FatalError("close"); // wait for message char msg; ssize_t bytes; do { bytes = read(sv[0], &msg, 1); } while (bytes == -1 && errno == EINTR); if (bytes == 0) _exit(0); if (bytes != 1) FatalError("read"); // do chrooting if (msg != kMsgChrootMe) FatalError("Unknown message from sandboxed process"); // sanity check if (chdir(safedir)) FatalError("Cannot chdir into /proc/ directory"); if (chroot(safedir)) FatalError("Cannot chroot into /proc/ directory"); if (chdir("/")) FatalError("Cannot chdir to / after chroot"); const char reply = kMsgChrootSuccessful; do { bytes = write(sv[0], &reply, 1); } while (bytes == -1 && errno == EINTR); if (bytes != 1) FatalError("Writing reply"); _exit(0); // We now become a zombie. /proc/self/fd(info) is now an empty dir and we // are chrooted there. // Our (unprivileged) parent should not even be able to open "." or "/" // since they would need to pass the ptrace() check. If our parent wait() // for us, our root directory will completely disappear. } if (close(sv[0])) { close(sv[1]); perror("close"); return false; } // In the parent process, we install an environment variable containing the // number of the file descriptor. char desc_str[64]; int printed = snprintf(desc_str, sizeof(desc_str), "%u", sv[1]); if (printed < 0 || printed >= (int)sizeof(desc_str)) { fprintf(stderr, "Failed to snprintf\n"); return false; } if (setenv(kSandboxDescriptorEnvironmentVarName, desc_str, 1)) { perror("setenv"); close(sv[1]); return false; } // We also install an environment variable containing the pid of the child char helper_pid_str[64]; printed = snprintf(helper_pid_str, sizeof(helper_pid_str), "%u", pid); if (printed < 0 || printed >= (int)sizeof(helper_pid_str)) { fprintf(stderr, "Failed to snprintf\n"); return false; } if (setenv(kSandboxHelperPidEnvironmentVarName, helper_pid_str, 1)) { perror("setenv"); close(sv[1]); return false; } return true; } // Block until child_pid exits, then exit. Try to preserve the exit code. static void WaitForChildAndExit(pid_t child_pid) { int exit_code = -1; siginfo_t reaped_child_info; // Don't "Core" on SIGABRT. SIGABRT is sent by the Chrome OS session manager // when things are hanging. // Here, the current process is going to waitid() and _exit(), so there is no // point in generating a crash report. The child process is the one // blocking us. if (signal(SIGABRT, ExitWithErrorSignalHandler) == SIG_ERR) { FatalError("Failed to change signal handler"); } int wait_ret = HANDLE_EINTR(waitid(P_PID, child_pid, &reaped_child_info, WEXITED)); if (!wait_ret && reaped_child_info.si_pid == child_pid) { if (reaped_child_info.si_code == CLD_EXITED) { exit_code = reaped_child_info.si_status; } else { // Exit with code 0 if the child got signaled. exit_code = 0; } } _exit(exit_code); } static bool MoveToNewNamespaces() { // These are the sets of flags which we'll try, in order. const int kCloneExtraFlags[] = {CLONE_NEWPID | CLONE_NEWNET, CLONE_NEWPID, }; // We need to close kZygoteIdFd before the child can continue. We use this // socketpair to tell the child when to continue; int sync_fds[2]; if (socketpair(AF_UNIX, SOCK_STREAM, 0, sync_fds)) { FatalError("Failed to create a socketpair"); } for (size_t i = 0; i < sizeof(kCloneExtraFlags) / sizeof(kCloneExtraFlags[0]); i++) { pid_t pid = syscall(__NR_clone, SIGCHLD | kCloneExtraFlags[i], 0, 0, 0); const int clone_errno = errno; if (pid > 0) { if (!DropRoot()) { FatalError("Could not drop privileges"); } else { if (close(sync_fds[0]) || shutdown(sync_fds[1], SHUT_RD)) FatalError("Could not close socketpair"); // The kZygoteIdFd needs to be closed in the parent before // Zygote gets started. if (close(kZygoteIdFd)) FatalError("close"); // Tell our child to continue if (HANDLE_EINTR(send(sync_fds[1], "C", 1, MSG_NOSIGNAL)) != 1) FatalError("send"); if (close(sync_fds[1])) FatalError("close"); // We want to keep a full process tree and we don't want our childs to // be reparented to (the outer PID namespace) init. So we wait for it. WaitForChildAndExit(pid); } // NOTREACHED FatalError("Not reached"); } if (pid == 0) { if (close(sync_fds[1]) || shutdown(sync_fds[0], SHUT_WR)) FatalError("Could not close socketpair"); // Wait for the parent to confirm it closed kZygoteIdFd before we // continue char should_continue; if (HANDLE_EINTR(read(sync_fds[0], &should_continue, 1)) != 1) FatalError("Read on socketpair"); if (close(sync_fds[0])) FatalError("close"); if (kCloneExtraFlags[i] & CLONE_NEWPID) { setenv(kSandboxPIDNSEnvironmentVarName, "", 1 /* overwrite */); } else { unsetenv(kSandboxPIDNSEnvironmentVarName); } if (kCloneExtraFlags[i] & CLONE_NEWNET) { setenv(kSandboxNETNSEnvironmentVarName, "", 1 /* overwrite */); } else { unsetenv(kSandboxNETNSEnvironmentVarName); } break; } // If EINVAL then the system doesn't support the requested flags, so // continue to try a different set. // On any other errno value the system *does* support these flags but // something went wrong, hence we bail with an error message rather then // provide less security. if (errno != EINVAL) { fprintf(stderr, "Failed to move to new namespace:"); if (kCloneExtraFlags[i] & CLONE_NEWPID) { fprintf(stderr, " PID namespaces supported,"); } if (kCloneExtraFlags[i] & CLONE_NEWNET) { fprintf(stderr, " Network namespace supported,"); } fprintf(stderr, " but failed: errno = %s\n", strerror(clone_errno)); return false; } } // If the system doesn't support NEWPID then we carry on anyway. return true; } static bool DropRoot() { if (prctl(PR_SET_DUMPABLE, 0, 0, 0, 0)) { perror("prctl(PR_SET_DUMPABLE)"); return false; } if (prctl(PR_GET_DUMPABLE, 0, 0, 0, 0)) { perror("Still dumpable after prctl(PR_SET_DUMPABLE)"); return false; } gid_t rgid, egid, sgid; if (getresgid(&rgid, &egid, &sgid)) { perror("getresgid"); return false; } if (setresgid(rgid, rgid, rgid)) { perror("setresgid"); return false; } uid_t ruid, euid, suid; if (getresuid(&ruid, &euid, &suid)) { perror("getresuid"); return false; } if (setresuid(ruid, ruid, ruid)) { perror("setresuid"); return false; } return true; } static bool SetupChildEnvironment() { unsigned i; // ld.so may have cleared several environment variables because we are SUID. // However, the child process might need them so zygote_host_linux.cc saves a // copy in SANDBOX_$x. This is safe because we have dropped root by this // point, so we can only exec a binary with the permissions of the user who // ran us in the first place. for (i = 0; kSUIDUnsafeEnvironmentVariables[i]; ++i) { const char* const envvar = kSUIDUnsafeEnvironmentVariables[i]; char* const saved_envvar = SandboxSavedEnvironmentVariable(envvar); if (!saved_envvar) return false; const char* const value = getenv(saved_envvar); if (value) { setenv(envvar, value, 1 /* overwrite */); unsetenv(saved_envvar); } free(saved_envvar); } return true; } bool CheckAndExportApiVersion() { // Check the environment to see if a specific API version was requested. // assume version 0 if none. long api_number = -1; char* api_string = getenv(kSandboxEnvironmentApiRequest); if (!api_string) { api_number = 0; } else { errno = 0; char* endptr = NULL; api_number = strtol(api_string, &endptr, 10); if (!endptr || *endptr || errno != 0) return false; } // Warn only for now. if (api_number != kSUIDSandboxApiNumber) { fprintf( stderr, "The setuid sandbox provides API version %ld, " "but you need %ld\n" "Please read " "https://code.google.com/p/chromium/wiki/LinuxSUIDSandboxDevelopment." "\n\n", kSUIDSandboxApiNumber, api_number); } // Export our version so that the sandboxed process can verify it did not // use an old sandbox. char version_string[64]; snprintf( version_string, sizeof(version_string), "%ld", kSUIDSandboxApiNumber); if (setenv(kSandboxEnvironmentApiProvides, version_string, 1)) { perror("setenv"); return false; } return true; } int main(int argc, char** argv) { if (argc <= 1) { if (argc <= 0) { return 1; } fprintf(stderr, "Usage: %s <renderer process> <args...>\n", argv[0]); return 1; } // Allow someone to query our API version if (argc == 2 && 0 == strcmp(argv[1], kSuidSandboxGetApiSwitch)) { printf("%ld\n", kSUIDSandboxApiNumber); return 0; } // In the SUID sandbox, if we succeed in calling MoveToNewNamespaces() // below, then the zygote and all the renderers are in an alternate PID // namespace and do not know their real PIDs. As such, they report the wrong // PIDs to the task manager. // // To fix this, when the zygote spawns a new renderer, it gives the renderer // a dummy socket, which has a unique inode number. Then it asks the sandbox // host to find the PID of the process holding that fd by searching /proc. // // Since the zygote and renderers are all spawned by this setuid executable, // their entries in /proc are owned by root and only readable by root. In // order to search /proc for the fd we want, this setuid executable has to // double as a helper and perform the search. The code block below does this // when you call it with --find-inode INODE_NUMBER. if (argc == 3 && (0 == strcmp(argv[1], kFindInodeSwitch))) { pid_t pid; char* endptr = NULL; errno = 0; ino_t inode = strtoull(argv[2], &endptr, 10); if (inode == ULLONG_MAX || !endptr || *endptr || errno != 0) return 1; if (!FindProcessHoldingSocket(&pid, inode)) return 1; printf("%d\n", pid); return 0; } // Likewise, we cannot adjust /proc/pid/oom_adj for sandboxed renderers // because those files are owned by root. So we need another helper here. if (argc == 4 && (0 == strcmp(argv[1], kAdjustOOMScoreSwitch))) { char* endptr = NULL; long score; errno = 0; unsigned long pid_ul = strtoul(argv[2], &endptr, 10); if (pid_ul == ULONG_MAX || !endptr || *endptr || errno != 0) return 1; pid_t pid = pid_ul; endptr = NULL; errno = 0; score = strtol(argv[3], &endptr, 10); if (score == LONG_MAX || score == LONG_MIN || !endptr || *endptr || errno != 0) { return 1; } return AdjustOOMScore(pid, score); } // Protect the core setuid sandbox functionality with an API version if (!CheckAndExportApiVersion()) { return 1; } if (geteuid() != 0) { fprintf(stderr, "The setuid sandbox is not running as root. Common causes:\n" " * An unprivileged process using ptrace on it, like a debugger.\n" " * A parent process set prctl(PR_SET_NO_NEW_PRIVS, ...)\n"); } if (!MoveToNewNamespaces()) return 1; if (!SpawnChrootHelper()) return 1; if (!DropRoot()) return 1; if (!SetupChildEnvironment()) return 1; execv(argv[1], &argv[1]); FatalError("execv failed"); return 1; }