1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * memfd GUP test-case
4  * This tests memfd interactions with get_user_pages(). We require the
5  * fuse_mnt.c program to provide a fake direct-IO FUSE mount-point for us. This
6  * file-system delays _all_ reads by 1s and forces direct-IO. This means, any
7  * read() on files in that file-system will pin the receive-buffer pages for at
8  * least 1s via get_user_pages().
9  *
10  * We use this trick to race ADD_SEALS against a write on a memfd object. The
11  * ADD_SEALS must fail if the memfd pages are still pinned. Note that we use
12  * the read() syscall with our memory-mapped memfd object as receive buffer to
13  * force the kernel to write into our memfd object.
14  */
15 
16 #define _GNU_SOURCE
17 #define __EXPORTED_HEADERS__
18 
19 #include <errno.h>
20 #include <inttypes.h>
21 #include <limits.h>
22 #include <linux/falloc.h>
23 #include <fcntl.h>
24 #include <linux/memfd.h>
25 #include <sched.h>
26 #include <stdio.h>
27 #include <stdlib.h>
28 #include <signal.h>
29 #include <string.h>
30 #include <sys/mman.h>
31 #include <sys/stat.h>
32 #include <sys/syscall.h>
33 #include <sys/wait.h>
34 #include <unistd.h>
35 
36 #include "common.h"
37 
38 #define MFD_DEF_SIZE 8192
39 #define STACK_SIZE 65536
40 
41 static size_t mfd_def_size = MFD_DEF_SIZE;
42 
43 static int mfd_assert_new(const char *name, loff_t sz, unsigned int flags)
44 {
45 	int r, fd;
46 
47 	fd = sys_memfd_create(name, flags);
48 	if (fd < 0) {
49 		printf("memfd_create(\"%s\", %u) failed: %m\n",
50 		       name, flags);
51 		abort();
52 	}
53 
54 	r = ftruncate(fd, sz);
55 	if (r < 0) {
56 		printf("ftruncate(%llu) failed: %m\n", (unsigned long long)sz);
57 		abort();
58 	}
59 
60 	return fd;
61 }
62 
63 static __u64 mfd_assert_get_seals(int fd)
64 {
65 	long r;
66 
67 	r = fcntl(fd, F_GET_SEALS);
68 	if (r < 0) {
69 		printf("GET_SEALS(%d) failed: %m\n", fd);
70 		abort();
71 	}
72 
73 	return r;
74 }
75 
76 static void mfd_assert_has_seals(int fd, __u64 seals)
77 {
78 	__u64 s;
79 
80 	s = mfd_assert_get_seals(fd);
81 	if (s != seals) {
82 		printf("%llu != %llu = GET_SEALS(%d)\n",
83 		       (unsigned long long)seals, (unsigned long long)s, fd);
84 		abort();
85 	}
86 }
87 
88 static void mfd_assert_add_seals(int fd, __u64 seals)
89 {
90 	long r;
91 	__u64 s;
92 
93 	s = mfd_assert_get_seals(fd);
94 	r = fcntl(fd, F_ADD_SEALS, seals);
95 	if (r < 0) {
96 		printf("ADD_SEALS(%d, %llu -> %llu) failed: %m\n",
97 		       fd, (unsigned long long)s, (unsigned long long)seals);
98 		abort();
99 	}
100 }
101 
102 static int mfd_busy_add_seals(int fd, __u64 seals)
103 {
104 	long r;
105 	__u64 s;
106 
107 	r = fcntl(fd, F_GET_SEALS);
108 	if (r < 0)
109 		s = 0;
110 	else
111 		s = r;
112 
113 	r = fcntl(fd, F_ADD_SEALS, seals);
114 	if (r < 0 && errno != EBUSY) {
115 		printf("ADD_SEALS(%d, %llu -> %llu) didn't fail as expected with EBUSY: %m\n",
116 		       fd, (unsigned long long)s, (unsigned long long)seals);
117 		abort();
118 	}
119 
120 	return r;
121 }
122 
123 static void *mfd_assert_mmap_shared(int fd)
124 {
125 	void *p;
126 
127 	p = mmap(NULL,
128 		 mfd_def_size,
129 		 PROT_READ | PROT_WRITE,
130 		 MAP_SHARED,
131 		 fd,
132 		 0);
133 	if (p == MAP_FAILED) {
134 		printf("mmap() failed: %m\n");
135 		abort();
136 	}
137 
138 	return p;
139 }
140 
141 static void *mfd_assert_mmap_private(int fd)
142 {
143 	void *p;
144 
145 	p = mmap(NULL,
146 		 mfd_def_size,
147 		 PROT_READ | PROT_WRITE,
148 		 MAP_PRIVATE,
149 		 fd,
150 		 0);
151 	if (p == MAP_FAILED) {
152 		printf("mmap() failed: %m\n");
153 		abort();
154 	}
155 
156 	return p;
157 }
158 
159 static int global_mfd = -1;
160 static void *global_p = NULL;
161 
162 static int sealing_thread_fn(void *arg)
163 {
164 	int sig, r;
165 
166 	/*
167 	 * This thread first waits 200ms so any pending operation in the parent
168 	 * is correctly started. After that, it tries to seal @global_mfd as
169 	 * SEAL_WRITE. This _must_ fail as the parent thread has a read() into
170 	 * that memory mapped object still ongoing.
171 	 * We then wait one more second and try sealing again. This time it
172 	 * must succeed as there shouldn't be anyone else pinning the pages.
173 	 */
174 
175 	/* wait 200ms for FUSE-request to be active */
176 	usleep(200000);
177 
178 	/* unmount mapping before sealing to avoid i_mmap_writable failures */
179 	munmap(global_p, mfd_def_size);
180 
181 	/* Try sealing the global file; expect EBUSY or success. Current
182 	 * kernels will never succeed, but in the future, kernels might
183 	 * implement page-replacements or other fancy ways to avoid racing
184 	 * writes. */
185 	r = mfd_busy_add_seals(global_mfd, F_SEAL_WRITE);
186 	if (r >= 0) {
187 		printf("HURRAY! This kernel fixed GUP races!\n");
188 	} else {
189 		/* wait 1s more so the FUSE-request is done */
190 		sleep(1);
191 
192 		/* try sealing the global file again */
193 		mfd_assert_add_seals(global_mfd, F_SEAL_WRITE);
194 	}
195 
196 	return 0;
197 }
198 
199 static pid_t spawn_sealing_thread(void)
200 {
201 	uint8_t *stack;
202 	pid_t pid;
203 
204 	stack = malloc(STACK_SIZE);
205 	if (!stack) {
206 		printf("malloc(STACK_SIZE) failed: %m\n");
207 		abort();
208 	}
209 
210 	pid = clone(sealing_thread_fn,
211 		    stack + STACK_SIZE,
212 		    SIGCHLD | CLONE_FILES | CLONE_FS | CLONE_VM,
213 		    NULL);
214 	if (pid < 0) {
215 		printf("clone() failed: %m\n");
216 		abort();
217 	}
218 
219 	return pid;
220 }
221 
222 static void join_sealing_thread(pid_t pid)
223 {
224 	waitpid(pid, NULL, 0);
225 }
226 
227 int main(int argc, char **argv)
228 {
229 	char *zero;
230 	int fd, mfd, r;
231 	void *p;
232 	int was_sealed;
233 	pid_t pid;
234 
235 	if (argc < 2) {
236 		printf("error: please pass path to file in fuse_mnt mount-point\n");
237 		abort();
238 	}
239 
240 	if (argc >= 3) {
241 		if (!strcmp(argv[2], "hugetlbfs")) {
242 			unsigned long hpage_size = default_huge_page_size();
243 
244 			if (!hpage_size) {
245 				printf("Unable to determine huge page size\n");
246 				abort();
247 			}
248 
249 			hugetlbfs_test = 1;
250 			mfd_def_size = hpage_size * 2;
251 		} else {
252 			printf("Unknown option: %s\n", argv[2]);
253 			abort();
254 		}
255 	}
256 
257 	zero = calloc(sizeof(*zero), mfd_def_size);
258 
259 	/* open FUSE memfd file for GUP testing */
260 	printf("opening: %s\n", argv[1]);
261 	fd = open(argv[1], O_RDONLY | O_CLOEXEC);
262 	if (fd < 0) {
263 		printf("cannot open(\"%s\"): %m\n", argv[1]);
264 		abort();
265 	}
266 
267 	/* create new memfd-object */
268 	mfd = mfd_assert_new("kern_memfd_fuse",
269 			     mfd_def_size,
270 			     MFD_CLOEXEC | MFD_ALLOW_SEALING);
271 
272 	/* mmap memfd-object for writing */
273 	p = mfd_assert_mmap_shared(mfd);
274 
275 	/* pass mfd+mapping to a separate sealing-thread which tries to seal
276 	 * the memfd objects with SEAL_WRITE while we write into it */
277 	global_mfd = mfd;
278 	global_p = p;
279 	pid = spawn_sealing_thread();
280 
281 	/* Use read() on the FUSE file to read into our memory-mapped memfd
282 	 * object. This races the other thread which tries to seal the
283 	 * memfd-object.
284 	 * If @fd is on the memfd-fake-FUSE-FS, the read() is delayed by 1s.
285 	 * This guarantees that the receive-buffer is pinned for 1s until the
286 	 * data is written into it. The racing ADD_SEALS should thus fail as
287 	 * the pages are still pinned. */
288 	r = read(fd, p, mfd_def_size);
289 	if (r < 0) {
290 		printf("read() failed: %m\n");
291 		abort();
292 	} else if (!r) {
293 		printf("unexpected EOF on read()\n");
294 		abort();
295 	}
296 
297 	was_sealed = mfd_assert_get_seals(mfd) & F_SEAL_WRITE;
298 
299 	/* Wait for sealing-thread to finish and verify that it
300 	 * successfully sealed the file after the second try. */
301 	join_sealing_thread(pid);
302 	mfd_assert_has_seals(mfd, F_SEAL_WRITE);
303 
304 	/* *IF* the memfd-object was sealed at the time our read() returned,
305 	 * then the kernel did a page-replacement or canceled the read() (or
306 	 * whatever magic it did..). In that case, the memfd object is still
307 	 * all zero.
308 	 * In case the memfd-object was *not* sealed, the read() was successfull
309 	 * and the memfd object must *not* be all zero.
310 	 * Note that in real scenarios, there might be a mixture of both, but
311 	 * in this test-cases, we have explicit 200ms delays which should be
312 	 * enough to avoid any in-flight writes. */
313 
314 	p = mfd_assert_mmap_private(mfd);
315 	if (was_sealed && memcmp(p, zero, mfd_def_size)) {
316 		printf("memfd sealed during read() but data not discarded\n");
317 		abort();
318 	} else if (!was_sealed && !memcmp(p, zero, mfd_def_size)) {
319 		printf("memfd sealed after read() but data discarded\n");
320 		abort();
321 	}
322 
323 	close(mfd);
324 	close(fd);
325 
326 	printf("fuse: DONE\n");
327 	free(zero);
328 
329 	return 0;
330 }
331