@@ -26,6 +26,10 @@
2626using namespace std;
2727
2828#define err_exit(format, ...) { fprintf(stderr, format ": %s\n", ##__VA_ARGS__, strerror(errno)); exit(EXIT_FAILURE); }
29+static int child_proc(const char *rootdir, const char *nixdir, uint8_t clear_env, list<struct DirMapping> dirMappings, list<struct SetEnv> envMappings, const char *executable, char * const new_argv[]);
30+
31+volatile uint8_t child_died = 0;
32+int child_pid = 0;
2933
3034static void usage(const char *pname) {
3135 fprintf(stderr, "Usage: %s -n <nixpath> -- <command>\n", pname);
@@ -33,6 +37,7 @@ static void usage(const char *pname) {
3337 fprintf(stderr, "\t-m <src>:<src>\tmap src on the host to dest in the sandbox\n");
3438 fprintf(stderr, "\t-d\tdelete all default dir mappings, may break things\n");
3539 fprintf(stderr, "\t-p <var>\tpreserve the value of a variable across the -c clear\n");
40+ fprintf(stderr, "\t-e\tadd an /escape-hatch to the sandbox, and run (outside the sandbox) any strings written to it\n");
3641
3742 exit(EXIT_FAILURE);
3843}
@@ -91,8 +96,15 @@ struct DirMapping parseMapping(string input) {
9196 return (struct DirMapping){ src, dest };
9297}
9398
99+static void handle_child_death(int signo, siginfo_t *info, void *context) {
100+ if ( (child_pid == 0) || (info->si_pid == child_pid) ) {
101+ child_died = 1;
102+ }
103+}
104+
94105int main(int argc, char *argv[]) {
95106 uint8_t clear_env = 0;
107+ uint8_t enable_escape_hatch = 0;
96108 char *nixdir = NULL;
97109 list<struct DirMapping> dirMappings;
98110 list<struct SetEnv> envMappings;
@@ -105,14 +117,20 @@ int main(int argc, char *argv[]) {
105117 x("run");
106118 x("tmp");
107119 x("var");
120+ x("etc");
121+ x("usr");
122+ x("home");
123+ x("root");
108124#undef x
109125
110126 int opt;
111− while ((opt = getopt(argc, argv, "cn:m:dp:")) != -1) {
127+ while ((opt = getopt(argc, argv, "cen:m:dp:")) != -1) {
112128 switch (opt) {
113129 case 'c':
114130 clear_env = 1;
115− puts("clearing env");
131+ break;
132+ case 'e':
133+ enable_escape_hatch = 1;
116134 break;
117135 case 'n':
118136 // determine absolute directory for nix dir
@@ -161,13 +179,74 @@ int main(int argc, char *argv[]) {
161179 err_exit("mkdtemp(%s)", template_);
162180 }
163181
182+ int unrace[2];
183+
184+ if (pipe(unrace)) {
185+ err_exit("pipe()");
186+ }
187+
188+ struct sigaction handle_child;
189+ handle_child.sa_sigaction = handle_child_death;
190+ handle_child.sa_flags = SA_SIGINFO;
191+
192+ struct sigaction old_handler;
193+ if (sigaction(SIGCHLD, &handle_child, &old_handler)) {
194+ err_exit("sigaction()");
195+ }
196+
197+ int child;
198+ child = child_pid = fork();
199+ if (child < 0) {
200+ err_exit("fork()");
201+ } else if (child == 0) {
202+ sigaction(SIGCHLD, &old_handler, NULL);
203+ close(unrace[1]);
204+ char buf[10];
205+ read(unrace[0], buf, 10);
206+ close(unrace[0]);
207+ return child_proc(rootdir, nixdir, clear_env, dirMappings, envMappings, argv[optind], argv + optind);
208+ } else {
209+ close(unrace[0]);
210+ char fifopath[PATH_MAX];
211+ if (enable_escape_hatch) {
212+ snprintf(fifopath, PATH_MAX, "%s/escape-hatch", rootdir);
213+ mkfifo(fifopath, 0600);
214+ }
215+ close(unrace[1]);
216+ if (enable_escape_hatch) {
217+ char buffer[1024];
218+ while (!child_died) {
219+ int fd = open(fifopath, O_RDONLY);
220+ if (fd < 0) {
221+ if (errno == EINTR) continue;
222+ fprintf(stderr, "error opening escape-hatch: %s\n", strerror(errno));
223+ continue;
224+ }
225+ int size = read(fd, buffer, 1024);
226+ buffer[size] = 0;
227+ system(buffer);
228+ close(fd);
229+ }
230+ }
231+ int status;
232+ int ret = waitpid(child, &status, 0);
233+ return WEXITSTATUS(status);
234+ }
235+}
236+
237+static int child_proc(const char *rootdir, const char *nixdir, uint8_t clear_env, list<struct DirMapping> dirMappings, list<struct SetEnv> envMappings, const char *executable, char * const new_argv[]) {
164238 // get uid, gid before going to new namespace
165239 uid_t uid = getuid();
166240 gid_t gid = getgid();
167241
168242 // "unshare" into new namespace
169243 if (unshare(CLONE_NEWNS | CLONE_NEWUSER) < 0) {
170− err_exit("unshare()");
244+ if (errno == EPERM) {
245+ fputs("Run the following to enable unprivileged namespace use:\nsudo bash -c \"sysctl -w kernel.unprivileged_userns_clone=1 ; echo kernel.unprivileged_userns_clone=1 > /etc/sysctl.d/nix-user-chroot.conf\"\n\n", stderr);
246+ exit(EXIT_FAILURE);
247+ } else {
248+ err_exit("unshare()");
249+ }
171250 }
172251
173252 // add necessary system stuff to rootdir namespace
@@ -177,57 +256,52 @@ int main(int argc, char *argv[]) {
177256 add_path(m.src, m.dest, rootdir);
178257 }
179258
180− char path_buf[PATH_MAX];
181− snprintf(path_buf, sizeof(path_buf), "%s/tmp", rootdir);
182− mkdir(path_buf, ~0);
183− snprintf(path_buf, sizeof(path_buf), "%s/var", rootdir);
184−
185− // make sure nixdir exists
186− struct stat statbuf2;
187− if (stat(nixdir, &statbuf2) < 0) {
188− err_exit("stat(%s)", nixdir);
189− }
259+ // make sure nixdir exists
260+ struct stat statbuf2;
261+ if (stat(nixdir, &statbuf2) < 0) {
262+ err_exit("stat(%s)", nixdir);
263+ }
190264
191− // mount /nix to new namespace
192− snprintf(path_buf, sizeof(path_buf), "%s/nix", rootdir);
193− mkdir(path_buf, statbuf2.st_mode & ~S_IFMT);
194− if (mount(nixdir, path_buf, "none", MS_BIND | MS_REC, NULL) < 0) {
195− err_exit("mount(%s, %s)", nixdir, path_buf);
196− }
265+ char path_buf[PATH_MAX];
266+ // mount /nix to new namespace
267+ snprintf(path_buf, sizeof(path_buf), "%s/nix", rootdir);
268+ mkdir(path_buf, statbuf2.st_mode & ~S_IFMT);
269+ if (mount(nixdir, path_buf, "none", MS_BIND | MS_REC, NULL) < 0) {
270+ err_exit("mount(%s, %s)", nixdir, path_buf);
271+ }
197272
198− // fixes issue #1 where writing to /proc/self/gid_map fails
199− // see user_namespaces(7) for more documentation
200− int fd_setgroups = open("/proc/self/setgroups", O_WRONLY);
201− if (fd_setgroups > 0) {
202− write(fd_setgroups, "deny", 4);
203− }
273+ // fixes issue #1 where writing to /proc/self/gid_map fails
274+ // see user_namespaces(7) for more documentation
275+ int fd_setgroups = open("/proc/self/setgroups", O_WRONLY);
276+ if (fd_setgroups > 0) {
277+ write(fd_setgroups, "deny", 4);
278+ close(fd_setgroups);
279+ }
204280
205− // map the original uid/gid in the new ns
206− char map_buf[1024];
207− snprintf(map_buf, sizeof(map_buf), "%d %d 1", uid, uid);
208− update_map(map_buf, "/proc/self/uid_map");
209− snprintf(map_buf, sizeof(map_buf), "%d %d 1", gid, gid);
210− update_map(map_buf, "/proc/self/gid_map");
281+ // map the original uid/gid in the new ns
282+ char map_buf[1024];
283+ snprintf(map_buf, sizeof(map_buf), "%d %d 1", uid, uid);
284+ update_map(map_buf, "/proc/self/uid_map");
285+ snprintf(map_buf, sizeof(map_buf), "%d %d 1", gid, gid);
286+ update_map(map_buf, "/proc/self/gid_map");
211287
212− // chroot to rootdir
213− if (chroot(rootdir) < 0) {
214− err_exit("chroot(%s)", rootdir);
215− }
288+ // chroot to rootdir
289+ if (chroot(rootdir) < 0) {
290+ err_exit("chroot(%s)", rootdir);
291+ }
216292
217− chdir("/");
293+ chdir("/");
218294
219295 if (clear_env) clearenv();
220− setenv("NIX_PATH", NIX_PATH, 1);
221− setenv("NIX_SSL_CERT_FILE", NIX_SSL_CERT_FILE, 1);
222296 setenv("PATH", ENV_PATH, 1);
223297
224298 for (list<struct SetEnv>::iterator it = envMappings.begin();
225− it != envMappings.end(); ++it) {
299+ it != envMappings.end(); ++it) {
226300 struct SetEnv e = *it;
227301 setenv(e.key.c_str(), e.value.c_str(), 1);
228302 }
229303
230304 // execute the command
231− execvp(argv[optind], argv + optind);
232− err_exit("execvp(%s)", argv[optind]);
305+ execvp(executable, new_argv);
306+ err_exit("execvp(%s)", executable);
233307}