xref: /linux/tools/testing/selftests/landlock/tsync_test.c (revision 67f8bc848ee31831336bd478e57d2f993551902e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Landlock tests - Enforcing the same restrictions across multiple threads
4  *
5  * Copyright © 2025 Günther Noack <gnoack3000@gmail.com>
6  */
7 
8 #define _GNU_SOURCE
9 #include <linux/landlock.h>
10 #include <pthread.h>
11 #include <signal.h>
12 #include <sys/prctl.h>
13 
14 #include "common.h"
15 
16 /* create_ruleset - Create a simple ruleset FD common to all tests */
17 static int create_ruleset(struct __test_metadata *const _metadata)
18 {
19 	struct landlock_ruleset_attr ruleset_attr = {
20 		.handled_access_fs = (LANDLOCK_ACCESS_FS_WRITE_FILE |
21 				      LANDLOCK_ACCESS_FS_TRUNCATE),
22 	};
23 	const int ruleset_fd =
24 		landlock_create_ruleset(&ruleset_attr, sizeof(ruleset_attr), 0);
25 
26 	ASSERT_LE(0, ruleset_fd)
27 	{
28 		TH_LOG("landlock_create_ruleset: %s", strerror(errno));
29 	}
30 	return ruleset_fd;
31 }
32 
33 TEST(single_threaded_success)
34 {
35 	const int ruleset_fd = create_ruleset(_metadata);
36 
37 	disable_caps(_metadata);
38 
39 	ASSERT_EQ(0, prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0));
40 	ASSERT_EQ(0, landlock_restrict_self(ruleset_fd,
41 					    LANDLOCK_RESTRICT_SELF_TSYNC));
42 
43 	EXPECT_EQ(0, close(ruleset_fd));
44 }
45 
46 static void store_no_new_privs(void *data)
47 {
48 	bool *nnp = data;
49 
50 	if (!nnp)
51 		return;
52 	*nnp = prctl(PR_GET_NO_NEW_PRIVS, 0, 0, 0, 0);
53 }
54 
55 static void *idle(void *data)
56 {
57 	pthread_cleanup_push(store_no_new_privs, data);
58 
59 	while (true)
60 		sleep(1);
61 
62 	pthread_cleanup_pop(1);
63 }
64 
65 FIXTURE(multi_threaded)
66 {
67 	int ruleset_fd;
68 };
69 
70 FIXTURE_VARIANT(multi_threaded)
71 {
72 	const __u32 restrict_flags;
73 	/* Sets no_new_privs with prctl(2) before the enforcement. */
74 	const bool prior_no_new_privs;
75 	/* Enforces the maximum number of allowed layers beforehand. */
76 	const bool max_layers;
77 	const int expected_errno;
78 	/* Expected no_new_privs state of all threads after the call. */
79 	const bool expected_no_new_privs;
80 };
81 
82 /* clang-format off */
83 FIXTURE_VARIANT_ADD(multi_threaded, success) {
84 	/* clang-format on */
85 	.restrict_flags = LANDLOCK_RESTRICT_SELF_TSYNC,
86 	.prior_no_new_privs = true,
87 	.expected_no_new_privs = true,
88 };
89 
90 /* clang-format off */
91 FIXTURE_VARIANT_ADD(multi_threaded, no_new_privs) {
92 	/* clang-format on */
93 	.restrict_flags = LANDLOCK_RESTRICT_SELF_TSYNC |
94 			  LANDLOCK_RESTRICT_SELF_NO_NEW_PRIVS,
95 	.expected_no_new_privs = true,
96 };
97 
98 /* clang-format off */
99 FIXTURE_VARIANT_ADD(multi_threaded, no_new_privs_max_layers) {
100 	/* clang-format on */
101 	.restrict_flags = LANDLOCK_RESTRICT_SELF_TSYNC |
102 			  LANDLOCK_RESTRICT_SELF_NO_NEW_PRIVS,
103 	.max_layers = true,
104 	.expected_errno = E2BIG,
105 	.expected_no_new_privs = false,
106 };
107 
108 FIXTURE_SETUP(multi_threaded)
109 {
110 	self->ruleset_fd = create_ruleset(_metadata);
111 
112 	if (variant->max_layers) {
113 		/* Enforces the maximum number of allowed layers. */
114 		for (int i = 0; i < LANDLOCK_MAX_NUM_LAYERS; i++)
115 			ASSERT_EQ(0,
116 				  landlock_restrict_self(self->ruleset_fd, 0));
117 	}
118 
119 	disable_caps(_metadata);
120 }
121 
122 FIXTURE_TEARDOWN(multi_threaded)
123 {
124 	EXPECT_EQ(0, close(self->ruleset_fd));
125 }
126 
127 TEST_F(multi_threaded, restrict)
128 {
129 	pthread_t t1, t2;
130 	bool no_new_privs1, no_new_privs2;
131 
132 	ASSERT_EQ(0, pthread_create(&t1, NULL, idle, &no_new_privs1));
133 	ASSERT_EQ(0, pthread_create(&t2, NULL, idle, &no_new_privs2));
134 
135 	if (variant->prior_no_new_privs) {
136 		ASSERT_EQ(0, prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0));
137 	} else {
138 		/* No prior prctl(2) PR_SET_NO_NEW_PRIVS call. */
139 		ASSERT_EQ(0, prctl(PR_GET_NO_NEW_PRIVS, 0, 0, 0, 0));
140 	}
141 
142 	if (variant->expected_errno) {
143 		EXPECT_EQ(-1, landlock_restrict_self(self->ruleset_fd,
144 						     variant->restrict_flags));
145 		EXPECT_EQ(variant->expected_errno, errno);
146 	} else {
147 		EXPECT_EQ(0, landlock_restrict_self(self->ruleset_fd,
148 						    variant->restrict_flags));
149 	}
150 
151 	/* Checks the no_new_privs state of the calling thread. */
152 	EXPECT_EQ(variant->expected_no_new_privs,
153 		  prctl(PR_GET_NO_NEW_PRIVS, 0, 0, 0, 0));
154 
155 	ASSERT_EQ(0, pthread_cancel(t1));
156 	ASSERT_EQ(0, pthread_cancel(t2));
157 	ASSERT_EQ(0, pthread_join(t1, NULL));
158 	ASSERT_EQ(0, pthread_join(t2, NULL));
159 
160 	/* Checks the no_new_privs state of the sibling threads. */
161 	EXPECT_EQ(variant->expected_no_new_privs, no_new_privs1);
162 	EXPECT_EQ(variant->expected_no_new_privs, no_new_privs2);
163 }
164 
165 TEST(multi_threaded_success_despite_diverging_domains)
166 {
167 	pthread_t t1, t2;
168 	const int ruleset_fd = create_ruleset(_metadata);
169 
170 	disable_caps(_metadata);
171 
172 	ASSERT_EQ(0, prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0));
173 
174 	ASSERT_EQ(0, pthread_create(&t1, NULL, idle, NULL));
175 	ASSERT_EQ(0, pthread_create(&t2, NULL, idle, NULL));
176 
177 	/*
178 	 * The main thread enforces a ruleset,
179 	 * thereby bringing the threads' Landlock domains out of sync.
180 	 */
181 	EXPECT_EQ(0, landlock_restrict_self(ruleset_fd, 0));
182 
183 	/* Still, TSYNC succeeds, bringing the threads in sync again. */
184 	EXPECT_EQ(0, landlock_restrict_self(ruleset_fd,
185 					    LANDLOCK_RESTRICT_SELF_TSYNC));
186 
187 	ASSERT_EQ(0, pthread_cancel(t1));
188 	ASSERT_EQ(0, pthread_cancel(t2));
189 	ASSERT_EQ(0, pthread_join(t1, NULL));
190 	ASSERT_EQ(0, pthread_join(t2, NULL));
191 	EXPECT_EQ(0, close(ruleset_fd));
192 }
193 
194 struct thread_restrict_data {
195 	pthread_t t;
196 	int ruleset_fd;
197 	int result;
198 };
199 
200 static void *thread_restrict(void *data)
201 {
202 	struct thread_restrict_data *d = data;
203 
204 	d->result = landlock_restrict_self(d->ruleset_fd,
205 					   LANDLOCK_RESTRICT_SELF_TSYNC);
206 	return NULL;
207 }
208 
209 TEST(competing_enablement)
210 {
211 	const int ruleset_fd = create_ruleset(_metadata);
212 	struct thread_restrict_data d[] = {
213 		{ .ruleset_fd = ruleset_fd },
214 		{ .ruleset_fd = ruleset_fd },
215 	};
216 
217 	disable_caps(_metadata);
218 
219 	ASSERT_EQ(0, prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0));
220 	ASSERT_EQ(0, pthread_create(&d[0].t, NULL, thread_restrict, &d[0]));
221 	ASSERT_EQ(0, pthread_create(&d[1].t, NULL, thread_restrict, &d[1]));
222 
223 	/* Wait for threads to finish. */
224 	ASSERT_EQ(0, pthread_join(d[0].t, NULL));
225 	ASSERT_EQ(0, pthread_join(d[1].t, NULL));
226 
227 	/* Expect that both succeeded. */
228 	EXPECT_EQ(0, d[0].result);
229 	EXPECT_EQ(0, d[1].result);
230 
231 	EXPECT_EQ(0, close(ruleset_fd));
232 }
233 
234 static void signal_nop_handler(int sig)
235 {
236 }
237 
238 struct signaler_data {
239 	pthread_t target;
240 	volatile bool stop;
241 };
242 
243 static void *signaler_thread(void *data)
244 {
245 	struct signaler_data *sd = data;
246 
247 	while (!sd->stop)
248 		pthread_kill(sd->target, SIGUSR1);
249 
250 	return NULL;
251 }
252 
253 /*
254  * Number of idle sibling threads.  This must be large enough that even on
255  * machines with many cores, the sibling threads cannot all complete their
256  * credential preparation in a single parallel wave, otherwise the signaler
257  * thread has no window to interrupt wait_for_completion_interruptible().
258  * 200 threads on a 64-core machine yields ~3 serialized waves, giving the
259  * tight signal loop enough time to land an interruption.
260  */
261 #define NUM_IDLE_THREADS 200
262 
263 /*
264  * Exercises the tsync interruption and cancellation paths in tsync.c.
265  *
266  * When a signal interrupts the calling thread while it waits for sibling
267  * threads to finish their credential preparation
268  * (wait_for_completion_interruptible in landlock_restrict_sibling_threads),
269  * the kernel sets ERESTARTNOINTR, cancels queued task works that have not
270  * started yet (cancel_tsync_works), then waits for the remaining works to
271  * finish.  On the error return, syscalls.c aborts the prepared credentials.
272  * The kernel automatically restarts the syscall, so userspace sees success.
273  */
274 TEST(tsync_interrupt)
275 {
276 	size_t i;
277 	pthread_t threads[NUM_IDLE_THREADS];
278 	pthread_t signaler;
279 	struct signaler_data sd;
280 	struct sigaction sa = {};
281 	const int ruleset_fd = create_ruleset(_metadata);
282 
283 	disable_caps(_metadata);
284 
285 	/* Install a no-op SIGUSR1 handler so the signal does not kill us. */
286 	sa.sa_handler = signal_nop_handler;
287 	sigemptyset(&sa.sa_mask);
288 	ASSERT_EQ(0, sigaction(SIGUSR1, &sa, NULL));
289 
290 	ASSERT_EQ(0, prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0));
291 
292 	for (i = 0; i < NUM_IDLE_THREADS; i++)
293 		ASSERT_EQ(0, pthread_create(&threads[i], NULL, idle, NULL));
294 
295 	/*
296 	 * Start a signaler thread that continuously sends SIGUSR1 to the
297 	 * calling thread.  This maximizes the chance of interrupting
298 	 * wait_for_completion_interruptible() in the kernel's tsync path.
299 	 */
300 	sd.target = pthread_self();
301 	sd.stop = false;
302 	ASSERT_EQ(0, pthread_create(&signaler, NULL, signaler_thread, &sd));
303 
304 	/*
305 	 * The syscall may be interrupted and transparently restarted by the
306 	 * kernel (ERESTARTNOINTR).  From userspace, it should always succeed.
307 	 */
308 	EXPECT_EQ(0, landlock_restrict_self(ruleset_fd,
309 					    LANDLOCK_RESTRICT_SELF_TSYNC));
310 
311 	sd.stop = true;
312 	ASSERT_EQ(0, pthread_join(signaler, NULL));
313 
314 	for (i = 0; i < NUM_IDLE_THREADS; i++) {
315 		ASSERT_EQ(0, pthread_cancel(threads[i]));
316 		ASSERT_EQ(0, pthread_join(threads[i], NULL));
317 	}
318 
319 	EXPECT_EQ(0, close(ruleset_fd));
320 }
321 
322 /* clang-format off */
323 FIXTURE(tsync_without_ruleset) {};
324 /* clang-format on */
325 
326 FIXTURE_VARIANT(tsync_without_ruleset)
327 {
328 	const __u32 flags;
329 	const int expected_errno;
330 };
331 
332 /* clang-format off */
333 FIXTURE_VARIANT_ADD(tsync_without_ruleset, tsync_only) {
334 	/* clang-format on */
335 	.flags = LANDLOCK_RESTRICT_SELF_TSYNC,
336 	.expected_errno = EBADF,
337 };
338 
339 /* clang-format off */
340 FIXTURE_VARIANT_ADD(tsync_without_ruleset, subdomains_off_same_exec_off) {
341 	/* clang-format on */
342 	.flags = LANDLOCK_RESTRICT_SELF_LOG_SUBDOMAINS_OFF |
343 		 LANDLOCK_RESTRICT_SELF_LOG_SAME_EXEC_OFF |
344 		 LANDLOCK_RESTRICT_SELF_TSYNC,
345 	.expected_errno = EBADF,
346 };
347 
348 /* clang-format off */
349 FIXTURE_VARIANT_ADD(tsync_without_ruleset, subdomains_off_new_exec_on) {
350 	/* clang-format on */
351 	.flags = LANDLOCK_RESTRICT_SELF_LOG_SUBDOMAINS_OFF |
352 		 LANDLOCK_RESTRICT_SELF_LOG_NEW_EXEC_ON |
353 		 LANDLOCK_RESTRICT_SELF_TSYNC,
354 	.expected_errno = EBADF,
355 };
356 
357 /* clang-format off */
358 FIXTURE_VARIANT_ADD(tsync_without_ruleset, all_flags) {
359 	/* clang-format on */
360 	.flags = LANDLOCK_RESTRICT_SELF_LOG_SAME_EXEC_OFF |
361 		 LANDLOCK_RESTRICT_SELF_LOG_NEW_EXEC_ON |
362 		 LANDLOCK_RESTRICT_SELF_LOG_SUBDOMAINS_OFF |
363 		 LANDLOCK_RESTRICT_SELF_TSYNC,
364 	.expected_errno = EBADF,
365 };
366 
367 /* clang-format off */
368 FIXTURE_VARIANT_ADD(tsync_without_ruleset, subdomains_off) {
369 	/* clang-format on */
370 	.flags = LANDLOCK_RESTRICT_SELF_LOG_SUBDOMAINS_OFF |
371 		 LANDLOCK_RESTRICT_SELF_TSYNC,
372 	.expected_errno = 0,
373 };
374 
375 FIXTURE_SETUP(tsync_without_ruleset)
376 {
377 	disable_caps(_metadata);
378 }
379 
380 FIXTURE_TEARDOWN(tsync_without_ruleset)
381 {
382 }
383 
384 TEST_F(tsync_without_ruleset, check)
385 {
386 	int ret;
387 
388 	ASSERT_EQ(0, prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0));
389 
390 	ret = landlock_restrict_self(-1, variant->flags);
391 	if (variant->expected_errno) {
392 		EXPECT_EQ(-1, ret);
393 		EXPECT_EQ(variant->expected_errno, errno);
394 	} else {
395 		EXPECT_EQ(0, ret);
396 	}
397 }
398 
399 TEST_HARNESS_MAIN
400