]>
Commit | Line | Data |
---|---|---|
db9ecf05 | 1 | /* SPDX-License-Identifier: LGPL-2.1-or-later */ |
57183d11 | 2 | |
a8fbdf54 | 3 | #include <errno.h> |
3c27973b | 4 | #include <fcntl.h> |
469830d1 | 5 | #include <linux/seccomp.h> |
a8fbdf54 | 6 | #include <stddef.h> |
469830d1 | 7 | #include <sys/mman.h> |
d347d902 | 8 | #include <sys/prctl.h> |
469830d1 | 9 | #include <sys/shm.h> |
3c27973b | 10 | #include <sys/stat.h> |
57183d11 | 11 | |
e83156c2 YW |
12 | /* include missing_syscall_def.h earlier to make __SNR_foo mapped to __NR_foo. */ |
13 | #include "missing_syscall_def.h" | |
14 | #include <seccomp.h> | |
15 | ||
469830d1 | 16 | #include "af-list.h" |
add00535 | 17 | #include "alloc-util.h" |
44aaddad | 18 | #include "env-util.h" |
d8b4d14d | 19 | #include "errno-list.h" |
a8fbdf54 | 20 | #include "macro.h" |
241b1577 | 21 | #include "namespace-util.h" |
add00535 | 22 | #include "nsflags.h" |
d8b4d14d | 23 | #include "nulstr-util.h" |
78e864e5 | 24 | #include "process-util.h" |
cf0fbc49 | 25 | #include "seccomp-util.h" |
b16bd535 | 26 | #include "set.h" |
07630cea | 27 | #include "string-util.h" |
b16bd535 | 28 | #include "strv.h" |
469830d1 | 29 | |
65976868 GDF |
30 | /* This array will be modified at runtime as seccomp_restrict_archs is called. */ |
31 | uint32_t seccomp_local_archs[] = { | |
469830d1 | 32 | |
6b000af4 | 33 | /* Note: always list the native arch we are compiled as last, so that users can deny-list seccomp(), but our own calls to it still succeed */ |
f2d9751c LP |
34 | |
35 | #if defined(__x86_64__) && defined(__ILP32__) | |
469830d1 LP |
36 | SCMP_ARCH_X86, |
37 | SCMP_ARCH_X86_64, | |
f2d9751c LP |
38 | SCMP_ARCH_X32, /* native */ |
39 | #elif defined(__x86_64__) && !defined(__ILP32__) | |
40 | SCMP_ARCH_X86, | |
469830d1 | 41 | SCMP_ARCH_X32, |
f2d9751c LP |
42 | SCMP_ARCH_X86_64, /* native */ |
43 | #elif defined(__i386__) | |
44 | SCMP_ARCH_X86, | |
45 | #elif defined(__aarch64__) | |
469830d1 | 46 | SCMP_ARCH_ARM, |
f2d9751c LP |
47 | SCMP_ARCH_AARCH64, /* native */ |
48 | #elif defined(__arm__) | |
49 | SCMP_ARCH_ARM, | |
f9d3fb6b XW |
50 | #elif defined(__loongarch_lp64) |
51 | SCMP_ARCH_LOONGARCH64, | |
f2d9751c LP |
52 | #elif defined(__mips__) && __BYTE_ORDER == __BIG_ENDIAN && _MIPS_SIM == _MIPS_SIM_ABI32 |
53 | SCMP_ARCH_MIPSEL, | |
54 | SCMP_ARCH_MIPS, /* native */ | |
55 | #elif defined(__mips__) && __BYTE_ORDER == __LITTLE_ENDIAN && _MIPS_SIM == _MIPS_SIM_ABI32 | |
469830d1 | 56 | SCMP_ARCH_MIPS, |
f2d9751c LP |
57 | SCMP_ARCH_MIPSEL, /* native */ |
58 | #elif defined(__mips__) && __BYTE_ORDER == __BIG_ENDIAN && _MIPS_SIM == _MIPS_SIM_ABI64 | |
59 | SCMP_ARCH_MIPSEL, | |
60 | SCMP_ARCH_MIPS, | |
61 | SCMP_ARCH_MIPSEL64N32, | |
469830d1 | 62 | SCMP_ARCH_MIPS64N32, |
f2d9751c LP |
63 | SCMP_ARCH_MIPSEL64, |
64 | SCMP_ARCH_MIPS64, /* native */ | |
65 | #elif defined(__mips__) && __BYTE_ORDER == __LITTLE_ENDIAN && _MIPS_SIM == _MIPS_SIM_ABI64 | |
66 | SCMP_ARCH_MIPS, | |
469830d1 | 67 | SCMP_ARCH_MIPSEL, |
f2d9751c LP |
68 | SCMP_ARCH_MIPS64N32, |
69 | SCMP_ARCH_MIPSEL64N32, | |
70 | SCMP_ARCH_MIPS64, | |
71 | SCMP_ARCH_MIPSEL64, /* native */ | |
72 | #elif defined(__mips__) && __BYTE_ORDER == __BIG_ENDIAN && _MIPS_SIM == _MIPS_SIM_NABI32 | |
73 | SCMP_ARCH_MIPSEL, | |
74 | SCMP_ARCH_MIPS, | |
469830d1 | 75 | SCMP_ARCH_MIPSEL64, |
f2d9751c | 76 | SCMP_ARCH_MIPS64, |
469830d1 | 77 | SCMP_ARCH_MIPSEL64N32, |
f2d9751c LP |
78 | SCMP_ARCH_MIPS64N32, /* native */ |
79 | #elif defined(__mips__) && __BYTE_ORDER == __LITTLE_ENDIAN && _MIPS_SIM == _MIPS_SIM_NABI32 | |
80 | SCMP_ARCH_MIPS, | |
81 | SCMP_ARCH_MIPSEL, | |
82 | SCMP_ARCH_MIPS64, | |
83 | SCMP_ARCH_MIPSEL64, | |
84 | SCMP_ARCH_MIPS64N32, | |
85 | SCMP_ARCH_MIPSEL64N32, /* native */ | |
344e6b62 SJ |
86 | #elif defined(__hppa64__) && defined(SCMP_ARCH_PARISC) && defined(SCMP_ARCH_PARISC64) |
87 | SCMP_ARCH_PARISC, | |
88 | SCMP_ARCH_PARISC64, /* native */ | |
89 | #elif defined(__hppa__) && defined(SCMP_ARCH_PARISC) | |
90 | SCMP_ARCH_PARISC, | |
f2d9751c | 91 | #elif defined(__powerpc64__) && __BYTE_ORDER == __BIG_ENDIAN |
469830d1 | 92 | SCMP_ARCH_PPC, |
469830d1 | 93 | SCMP_ARCH_PPC64LE, |
f2d9751c LP |
94 | SCMP_ARCH_PPC64, /* native */ |
95 | #elif defined(__powerpc64__) && __BYTE_ORDER == __LITTLE_ENDIAN | |
96 | SCMP_ARCH_PPC, | |
97 | SCMP_ARCH_PPC64, | |
98 | SCMP_ARCH_PPC64LE, /* native */ | |
99 | #elif defined(__powerpc__) | |
100 | SCMP_ARCH_PPC, | |
f9252236 AJ |
101 | #elif defined(__riscv) && __riscv_xlen == 64 && defined(SCMP_ARCH_RISCV64) |
102 | SCMP_ARCH_RISCV64, | |
f2d9751c LP |
103 | #elif defined(__s390x__) |
104 | SCMP_ARCH_S390, | |
105 | SCMP_ARCH_S390X, /* native */ | |
106 | #elif defined(__s390__) | |
469830d1 | 107 | SCMP_ARCH_S390, |
469830d1 | 108 | #endif |
65976868 | 109 | SECCOMP_LOCAL_ARCH_END |
469830d1 | 110 | }; |
57183d11 LP |
111 | |
112 | const char* seccomp_arch_to_string(uint32_t c) { | |
aa34055f ZJS |
113 | /* Maintain order used in <seccomp.h>. |
114 | * | |
115 | * Names used here should be the same as those used for ConditionArchitecture=, | |
116 | * except for "subarchitectures" like x32. */ | |
57183d11 | 117 | |
79893116 | 118 | switch (c) { |
aa34055f | 119 | case SCMP_ARCH_NATIVE: |
57183d11 | 120 | return "native"; |
aa34055f | 121 | case SCMP_ARCH_X86: |
57183d11 | 122 | return "x86"; |
aa34055f | 123 | case SCMP_ARCH_X86_64: |
57183d11 | 124 | return "x86-64"; |
aa34055f | 125 | case SCMP_ARCH_X32: |
57183d11 | 126 | return "x32"; |
aa34055f | 127 | case SCMP_ARCH_ARM: |
57183d11 | 128 | return "arm"; |
aa34055f ZJS |
129 | case SCMP_ARCH_AARCH64: |
130 | return "arm64"; | |
f9d3fb6b XW |
131 | #ifdef SCMP_ARCH_LOONGARCH64 |
132 | case SCMP_ARCH_LOONGARCH64: | |
133 | return "loongarch64"; | |
134 | #endif | |
aa34055f ZJS |
135 | case SCMP_ARCH_MIPS: |
136 | return "mips"; | |
137 | case SCMP_ARCH_MIPS64: | |
138 | return "mips64"; | |
139 | case SCMP_ARCH_MIPS64N32: | |
140 | return "mips64-n32"; | |
141 | case SCMP_ARCH_MIPSEL: | |
142 | return "mips-le"; | |
143 | case SCMP_ARCH_MIPSEL64: | |
144 | return "mips64-le"; | |
145 | case SCMP_ARCH_MIPSEL64N32: | |
146 | return "mips64-le-n32"; | |
344e6b62 SJ |
147 | #ifdef SCMP_ARCH_PARISC |
148 | case SCMP_ARCH_PARISC: | |
149 | return "parisc"; | |
150 | #endif | |
151 | #ifdef SCMP_ARCH_PARISC64 | |
152 | case SCMP_ARCH_PARISC64: | |
153 | return "parisc64"; | |
154 | #endif | |
aa34055f ZJS |
155 | case SCMP_ARCH_PPC: |
156 | return "ppc"; | |
157 | case SCMP_ARCH_PPC64: | |
158 | return "ppc64"; | |
159 | case SCMP_ARCH_PPC64LE: | |
160 | return "ppc64-le"; | |
f9252236 AJ |
161 | #ifdef SCMP_ARCH_RISCV64 |
162 | case SCMP_ARCH_RISCV64: | |
163 | return "riscv64"; | |
164 | #endif | |
aa34055f | 165 | case SCMP_ARCH_S390: |
6abfd303 | 166 | return "s390"; |
aa34055f | 167 | case SCMP_ARCH_S390X: |
6abfd303 | 168 | return "s390x"; |
aa34055f ZJS |
169 | default: |
170 | return NULL; | |
171 | } | |
57183d11 LP |
172 | } |
173 | ||
174 | int seccomp_arch_from_string(const char *n, uint32_t *ret) { | |
175 | if (!n) | |
176 | return -EINVAL; | |
177 | ||
178 | assert(ret); | |
179 | ||
180 | if (streq(n, "native")) | |
181 | *ret = SCMP_ARCH_NATIVE; | |
182 | else if (streq(n, "x86")) | |
183 | *ret = SCMP_ARCH_X86; | |
184 | else if (streq(n, "x86-64")) | |
185 | *ret = SCMP_ARCH_X86_64; | |
186 | else if (streq(n, "x32")) | |
187 | *ret = SCMP_ARCH_X32; | |
188 | else if (streq(n, "arm")) | |
189 | *ret = SCMP_ARCH_ARM; | |
aa34055f ZJS |
190 | else if (streq(n, "arm64")) |
191 | *ret = SCMP_ARCH_AARCH64; | |
f9d3fb6b XW |
192 | #ifdef SCMP_ARCH_LOONGARCH64 |
193 | else if (streq(n, "loongarch64")) | |
194 | *ret = SCMP_ARCH_LOONGARCH64; | |
195 | #endif | |
aa34055f ZJS |
196 | else if (streq(n, "mips")) |
197 | *ret = SCMP_ARCH_MIPS; | |
198 | else if (streq(n, "mips64")) | |
199 | *ret = SCMP_ARCH_MIPS64; | |
200 | else if (streq(n, "mips64-n32")) | |
201 | *ret = SCMP_ARCH_MIPS64N32; | |
202 | else if (streq(n, "mips-le")) | |
203 | *ret = SCMP_ARCH_MIPSEL; | |
204 | else if (streq(n, "mips64-le")) | |
205 | *ret = SCMP_ARCH_MIPSEL64; | |
206 | else if (streq(n, "mips64-le-n32")) | |
207 | *ret = SCMP_ARCH_MIPSEL64N32; | |
344e6b62 SJ |
208 | #ifdef SCMP_ARCH_PARISC |
209 | else if (streq(n, "parisc")) | |
210 | *ret = SCMP_ARCH_PARISC; | |
211 | #endif | |
212 | #ifdef SCMP_ARCH_PARISC64 | |
213 | else if (streq(n, "parisc64")) | |
214 | *ret = SCMP_ARCH_PARISC64; | |
215 | #endif | |
aa34055f ZJS |
216 | else if (streq(n, "ppc")) |
217 | *ret = SCMP_ARCH_PPC; | |
218 | else if (streq(n, "ppc64")) | |
219 | *ret = SCMP_ARCH_PPC64; | |
220 | else if (streq(n, "ppc64-le")) | |
221 | *ret = SCMP_ARCH_PPC64LE; | |
f9252236 AJ |
222 | #ifdef SCMP_ARCH_RISCV64 |
223 | else if (streq(n, "riscv64")) | |
224 | *ret = SCMP_ARCH_RISCV64; | |
225 | #endif | |
6abfd303 HB |
226 | else if (streq(n, "s390")) |
227 | *ret = SCMP_ARCH_S390; | |
228 | else if (streq(n, "s390x")) | |
229 | *ret = SCMP_ARCH_S390X; | |
57183d11 LP |
230 | else |
231 | return -EINVAL; | |
232 | ||
233 | return 0; | |
234 | } | |
e9642be2 | 235 | |
469830d1 | 236 | int seccomp_init_for_arch(scmp_filter_ctx *ret, uint32_t arch, uint32_t default_action) { |
b4eaa6cc | 237 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; |
8d7b0c8f LP |
238 | int r; |
239 | ||
469830d1 LP |
240 | /* Much like seccomp_init(), but initializes the filter for one specific architecture only, without affecting |
241 | * any others. Also, turns off the NNP fiddling. */ | |
8d7b0c8f LP |
242 | |
243 | seccomp = seccomp_init(default_action); | |
244 | if (!seccomp) | |
245 | return -ENOMEM; | |
246 | ||
469830d1 LP |
247 | if (arch != SCMP_ARCH_NATIVE && |
248 | arch != seccomp_arch_native()) { | |
249 | ||
1b52793d | 250 | r = seccomp_arch_remove(seccomp, seccomp_arch_native()); |
469830d1 | 251 | if (r < 0) |
b4eaa6cc | 252 | return r; |
469830d1 | 253 | |
1b52793d | 254 | r = seccomp_arch_add(seccomp, arch); |
469830d1 | 255 | if (r < 0) |
b4eaa6cc | 256 | return r; |
469830d1 LP |
257 | |
258 | assert(seccomp_arch_exist(seccomp, arch) >= 0); | |
259 | assert(seccomp_arch_exist(seccomp, SCMP_ARCH_NATIVE) == -EEXIST); | |
260 | assert(seccomp_arch_exist(seccomp, seccomp_arch_native()) == -EEXIST); | |
261 | } else { | |
262 | assert(seccomp_arch_exist(seccomp, SCMP_ARCH_NATIVE) >= 0); | |
263 | assert(seccomp_arch_exist(seccomp, seccomp_arch_native()) >= 0); | |
264 | } | |
265 | ||
266 | r = seccomp_attr_set(seccomp, SCMP_FLTATR_ACT_BADARCH, SCMP_ACT_ALLOW); | |
8d7b0c8f | 267 | if (r < 0) |
b4eaa6cc | 268 | return r; |
8d7b0c8f LP |
269 | |
270 | r = seccomp_attr_set(seccomp, SCMP_FLTATR_CTL_NNP, 0); | |
271 | if (r < 0) | |
b4eaa6cc | 272 | return r; |
8d7b0c8f | 273 | |
44aaddad SD |
274 | #if SCMP_VER_MAJOR >= 3 || (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR >= 4) |
275 | if (getenv_bool("SYSTEMD_LOG_SECCOMP") > 0) { | |
276 | r = seccomp_attr_set(seccomp, SCMP_FLTATR_CTL_LOG, 1); | |
277 | if (r < 0) | |
278 | log_debug_errno(r, "Failed to enable seccomp event logging: %m"); | |
279 | } | |
280 | #endif | |
281 | ||
b4eaa6cc | 282 | *ret = TAKE_PTR(seccomp); |
8d7b0c8f | 283 | return 0; |
8d7b0c8f LP |
284 | } |
285 | ||
d347d902 | 286 | static bool is_basic_seccomp_available(void) { |
4d5bd50a | 287 | return prctl(PR_GET_SECCOMP, 0, 0, 0, 0) >= 0; |
d347d902 FS |
288 | } |
289 | ||
290 | static bool is_seccomp_filter_available(void) { | |
4d5bd50a LP |
291 | return prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, NULL, 0, 0) < 0 && |
292 | errno == EFAULT; | |
d347d902 FS |
293 | } |
294 | ||
83f12b27 | 295 | bool is_seccomp_available(void) { |
83f12b27 | 296 | static int cached_enabled = -1; |
4d5bd50a | 297 | |
ce8f6d47 LP |
298 | if (cached_enabled < 0) { |
299 | int b; | |
300 | ||
efb9b3ba | 301 | b = secure_getenv_bool("SYSTEMD_SECCOMP"); |
ce8f6d47 LP |
302 | if (b != 0) { |
303 | if (b < 0 && b != -ENXIO) /* ENXIO: env var unset */ | |
304 | log_debug_errno(b, "Failed to parse $SYSTEMD_SECCOMP value, ignoring."); | |
305 | ||
306 | cached_enabled = | |
307 | is_basic_seccomp_available() && | |
308 | is_seccomp_filter_available(); | |
309 | } else | |
310 | cached_enabled = false; | |
311 | } | |
4d5bd50a | 312 | |
83f12b27 FS |
313 | return cached_enabled; |
314 | } | |
315 | ||
8130926d | 316 | const SyscallFilterSet syscall_filter_sets[_SYSCALL_FILTER_SET_MAX] = { |
40eb6a80 | 317 | [SYSCALL_FILTER_SET_DEFAULT] = { |
40eb6a80 | 318 | .name = "@default", |
d5efc18b | 319 | .help = "System calls that are always permitted", |
40eb6a80 | 320 | .value = |
5f02870a | 321 | "arch_prctl\0" /* Used during platform-specific initialization by ld-linux.so. */ |
5abede32 | 322 | "brk\0" |
8e24b1d2 | 323 | "cacheflush\0" |
40eb6a80 | 324 | "clock_getres\0" |
6ca67710 | 325 | "clock_getres_time64\0" |
40eb6a80 | 326 | "clock_gettime\0" |
6ca67710 | 327 | "clock_gettime64\0" |
40eb6a80 | 328 | "clock_nanosleep\0" |
6ca67710 | 329 | "clock_nanosleep_time64\0" |
40eb6a80 ZJS |
330 | "execve\0" |
331 | "exit\0" | |
332 | "exit_group\0" | |
e41b0f42 | 333 | "futex\0" |
6ca67710 | 334 | "futex_time64\0" |
76e86b8d | 335 | "futex_waitv\0" |
e41b0f42 LP |
336 | "get_robust_list\0" |
337 | "get_thread_area\0" | |
09d3020b DH |
338 | "getegid\0" |
339 | "getegid32\0" | |
340 | "geteuid\0" | |
341 | "geteuid32\0" | |
342 | "getgid\0" | |
343 | "getgid32\0" | |
344 | "getgroups\0" | |
345 | "getgroups32\0" | |
346 | "getpgid\0" | |
347 | "getpgrp\0" | |
348 | "getpid\0" | |
349 | "getppid\0" | |
14f4b1b5 | 350 | "getrandom\0" |
09d3020b DH |
351 | "getresgid\0" |
352 | "getresgid32\0" | |
353 | "getresuid\0" | |
354 | "getresuid32\0" | |
40eb6a80 | 355 | "getrlimit\0" /* make sure processes can query stack size and such */ |
09d3020b DH |
356 | "getsid\0" |
357 | "gettid\0" | |
40eb6a80 | 358 | "gettimeofday\0" |
09d3020b DH |
359 | "getuid\0" |
360 | "getuid32\0" | |
e41b0f42 | 361 | "membarrier\0" |
5abede32 LP |
362 | "mmap\0" |
363 | "mmap2\0" | |
47286254 | 364 | "mprotect\0" |
11b9105d | 365 | "munmap\0" |
40eb6a80 ZJS |
366 | "nanosleep\0" |
367 | "pause\0" | |
4c3a9176 | 368 | "prlimit64\0" |
e41b0f42 | 369 | "restart_syscall\0" |
09925036 | 370 | "riscv_flush_icache\0" |
ca15fc48 | 371 | "riscv_hwprobe\0" |
6fee3be0 | 372 | "rseq\0" |
40eb6a80 | 373 | "rt_sigreturn\0" |
7df660e4 | 374 | "sched_getaffinity\0" |
8f44de08 | 375 | "sched_yield\0" |
e41b0f42 LP |
376 | "set_robust_list\0" |
377 | "set_thread_area\0" | |
378 | "set_tid_address\0" | |
ce5faeac | 379 | "set_tls\0" |
40eb6a80 ZJS |
380 | "sigreturn\0" |
381 | "time\0" | |
4c3a9176 | 382 | "ugetrlimit\0" |
40eb6a80 | 383 | }, |
44898c53 LP |
384 | [SYSCALL_FILTER_SET_AIO] = { |
385 | .name = "@aio", | |
386 | .help = "Asynchronous IO", | |
387 | .value = | |
388 | "io_cancel\0" | |
389 | "io_destroy\0" | |
390 | "io_getevents\0" | |
a05cfe23 | 391 | "io_pgetevents\0" |
6ca67710 | 392 | "io_pgetevents_time64\0" |
44898c53 LP |
393 | "io_setup\0" |
394 | "io_submit\0" | |
9e486265 LP |
395 | "io_uring_enter\0" |
396 | "io_uring_register\0" | |
397 | "io_uring_setup\0" | |
44898c53 | 398 | }, |
133ddbbe | 399 | [SYSCALL_FILTER_SET_BASIC_IO] = { |
133ddbbe | 400 | .name = "@basic-io", |
d5efc18b | 401 | .help = "Basic IO", |
133ddbbe | 402 | .value = |
648a0ed0 | 403 | "_llseek\0" |
133ddbbe | 404 | "close\0" |
6ea0d25c | 405 | "close_range\0" |
648a0ed0 | 406 | "dup\0" |
133ddbbe LP |
407 | "dup2\0" |
408 | "dup3\0" | |
133ddbbe LP |
409 | "lseek\0" |
410 | "pread64\0" | |
411 | "preadv\0" | |
44898c53 | 412 | "preadv2\0" |
133ddbbe LP |
413 | "pwrite64\0" |
414 | "pwritev\0" | |
44898c53 | 415 | "pwritev2\0" |
133ddbbe LP |
416 | "read\0" |
417 | "readv\0" | |
418 | "write\0" | |
419 | "writev\0" | |
420 | }, | |
44898c53 LP |
421 | [SYSCALL_FILTER_SET_CHOWN] = { |
422 | .name = "@chown", | |
423 | .help = "Change ownership of files and directories", | |
424 | .value = | |
425 | "chown\0" | |
426 | "chown32\0" | |
427 | "fchown\0" | |
428 | "fchown32\0" | |
429 | "fchownat\0" | |
430 | "lchown\0" | |
431 | "lchown32\0" | |
432 | }, | |
8130926d | 433 | [SYSCALL_FILTER_SET_CLOCK] = { |
8130926d | 434 | .name = "@clock", |
d5efc18b | 435 | .help = "Change the system time", |
201c1cc2 TM |
436 | .value = |
437 | "adjtimex\0" | |
1f9ac68b | 438 | "clock_adjtime\0" |
6ca67710 | 439 | "clock_adjtime64\0" |
1f9ac68b | 440 | "clock_settime\0" |
6ca67710 | 441 | "clock_settime64\0" |
201c1cc2 | 442 | "settimeofday\0" |
8130926d LP |
443 | }, |
444 | [SYSCALL_FILTER_SET_CPU_EMULATION] = { | |
8130926d | 445 | .name = "@cpu-emulation", |
d5efc18b | 446 | .help = "System calls for CPU emulation functionality", |
1f9ac68b LP |
447 | .value = |
448 | "modify_ldt\0" | |
449 | "subpage_prot\0" | |
450 | "switch_endian\0" | |
451 | "vm86\0" | |
452 | "vm86old\0" | |
8130926d LP |
453 | }, |
454 | [SYSCALL_FILTER_SET_DEBUG] = { | |
8130926d | 455 | .name = "@debug", |
d5efc18b | 456 | .help = "Debugging, performance monitoring and tracing functionality", |
1f9ac68b LP |
457 | .value = |
458 | "lookup_dcookie\0" | |
459 | "perf_event_open\0" | |
8270e3d8 | 460 | "pidfd_getfd\0" |
1f9ac68b LP |
461 | "ptrace\0" |
462 | "rtas\0" | |
463 | "s390_runtime_instr\0" | |
464 | "sys_debug_setcontext\0" | |
8130926d | 465 | }, |
1a1b13c9 LP |
466 | [SYSCALL_FILTER_SET_FILE_SYSTEM] = { |
467 | .name = "@file-system", | |
468 | .help = "File system operations", | |
469 | .value = | |
470 | "access\0" | |
471 | "chdir\0" | |
472 | "chmod\0" | |
473 | "close\0" | |
474 | "creat\0" | |
475 | "faccessat\0" | |
bcf08acb | 476 | "faccessat2\0" |
1a1b13c9 LP |
477 | "fallocate\0" |
478 | "fchdir\0" | |
479 | "fchmod\0" | |
480 | "fchmodat\0" | |
6e10405a | 481 | "fchmodat2\0" |
1a1b13c9 | 482 | "fcntl\0" |
ceaa6aa7 | 483 | "fcntl64\0" |
1a1b13c9 LP |
484 | "fgetxattr\0" |
485 | "flistxattr\0" | |
ceaa6aa7 | 486 | "fremovexattr\0" |
1a1b13c9 | 487 | "fsetxattr\0" |
1a1b13c9 | 488 | "fstat\0" |
ceaa6aa7 | 489 | "fstat64\0" |
1a1b13c9 | 490 | "fstatat64\0" |
1a1b13c9 | 491 | "fstatfs\0" |
ceaa6aa7 | 492 | "fstatfs64\0" |
1a1b13c9 | 493 | "ftruncate\0" |
ceaa6aa7 | 494 | "ftruncate64\0" |
1a1b13c9 LP |
495 | "futimesat\0" |
496 | "getcwd\0" | |
1a1b13c9 | 497 | "getdents\0" |
ceaa6aa7 | 498 | "getdents64\0" |
1a1b13c9 LP |
499 | "getxattr\0" |
500 | "inotify_add_watch\0" | |
ceaa6aa7 | 501 | "inotify_init\0" |
1a1b13c9 LP |
502 | "inotify_init1\0" |
503 | "inotify_rm_watch\0" | |
504 | "lgetxattr\0" | |
505 | "link\0" | |
506 | "linkat\0" | |
507 | "listxattr\0" | |
508 | "llistxattr\0" | |
509 | "lremovexattr\0" | |
510 | "lsetxattr\0" | |
1a1b13c9 | 511 | "lstat\0" |
ceaa6aa7 | 512 | "lstat64\0" |
1a1b13c9 LP |
513 | "mkdir\0" |
514 | "mkdirat\0" | |
515 | "mknod\0" | |
516 | "mknodat\0" | |
1a1b13c9 | 517 | "newfstatat\0" |
ceaa6aa7 LP |
518 | "oldfstat\0" |
519 | "oldlstat\0" | |
520 | "oldstat\0" | |
1a1b13c9 LP |
521 | "open\0" |
522 | "openat\0" | |
8270e3d8 | 523 | "openat2\0" |
1a1b13c9 LP |
524 | "readlink\0" |
525 | "readlinkat\0" | |
526 | "removexattr\0" | |
527 | "rename\0" | |
1a1b13c9 | 528 | "renameat\0" |
ceaa6aa7 | 529 | "renameat2\0" |
1a1b13c9 LP |
530 | "rmdir\0" |
531 | "setxattr\0" | |
1a1b13c9 | 532 | "stat\0" |
ceaa6aa7 | 533 | "stat64\0" |
1a1b13c9 | 534 | "statfs\0" |
ceaa6aa7 | 535 | "statfs64\0" |
a4135a74 | 536 | "statx\0" |
1a1b13c9 LP |
537 | "symlink\0" |
538 | "symlinkat\0" | |
1a1b13c9 | 539 | "truncate\0" |
ceaa6aa7 | 540 | "truncate64\0" |
1a1b13c9 LP |
541 | "unlink\0" |
542 | "unlinkat\0" | |
ceaa6aa7 | 543 | "utime\0" |
1a1b13c9 | 544 | "utimensat\0" |
6ca67710 | 545 | "utimensat_time64\0" |
1a1b13c9 LP |
546 | "utimes\0" |
547 | }, | |
8130926d | 548 | [SYSCALL_FILTER_SET_IO_EVENT] = { |
8130926d | 549 | .name = "@io-event", |
d5efc18b | 550 | .help = "Event loop system calls", |
201c1cc2 TM |
551 | .value = |
552 | "_newselect\0" | |
201c1cc2 | 553 | "epoll_create\0" |
215728ff | 554 | "epoll_create1\0" |
201c1cc2 TM |
555 | "epoll_ctl\0" |
556 | "epoll_ctl_old\0" | |
557 | "epoll_pwait\0" | |
34254e59 | 558 | "epoll_pwait2\0" |
201c1cc2 TM |
559 | "epoll_wait\0" |
560 | "epoll_wait_old\0" | |
201c1cc2 | 561 | "eventfd\0" |
215728ff | 562 | "eventfd2\0" |
201c1cc2 TM |
563 | "poll\0" |
564 | "ppoll\0" | |
6ca67710 | 565 | "ppoll_time64\0" |
201c1cc2 | 566 | "pselect6\0" |
6ca67710 | 567 | "pselect6_time64\0" |
201c1cc2 | 568 | "select\0" |
8130926d LP |
569 | }, |
570 | [SYSCALL_FILTER_SET_IPC] = { | |
8130926d | 571 | .name = "@ipc", |
d5efc18b ZJS |
572 | .help = "SysV IPC, POSIX Message Queues or other IPC", |
573 | .value = | |
574 | "ipc\0" | |
cd5bfd7e | 575 | "memfd_create\0" |
201c1cc2 TM |
576 | "mq_getsetattr\0" |
577 | "mq_notify\0" | |
578 | "mq_open\0" | |
579 | "mq_timedreceive\0" | |
6ca67710 | 580 | "mq_timedreceive_time64\0" |
201c1cc2 | 581 | "mq_timedsend\0" |
6ca67710 | 582 | "mq_timedsend_time64\0" |
201c1cc2 TM |
583 | "mq_unlink\0" |
584 | "msgctl\0" | |
585 | "msgget\0" | |
586 | "msgrcv\0" | |
587 | "msgsnd\0" | |
cd5bfd7e | 588 | "pipe\0" |
215728ff | 589 | "pipe2\0" |
34254e59 | 590 | "process_madvise\0" |
201c1cc2 TM |
591 | "process_vm_readv\0" |
592 | "process_vm_writev\0" | |
593 | "semctl\0" | |
594 | "semget\0" | |
595 | "semop\0" | |
596 | "semtimedop\0" | |
6ca67710 | 597 | "semtimedop_time64\0" |
201c1cc2 TM |
598 | "shmat\0" |
599 | "shmctl\0" | |
600 | "shmdt\0" | |
601 | "shmget\0" | |
8130926d LP |
602 | }, |
603 | [SYSCALL_FILTER_SET_KEYRING] = { | |
8130926d | 604 | .name = "@keyring", |
d5efc18b | 605 | .help = "Kernel keyring access", |
1f9ac68b LP |
606 | .value = |
607 | "add_key\0" | |
608 | "keyctl\0" | |
609 | "request_key\0" | |
8130926d | 610 | }, |
cd0ddf6f LP |
611 | [SYSCALL_FILTER_SET_MEMLOCK] = { |
612 | .name = "@memlock", | |
613 | .help = "Memory locking control", | |
614 | .value = | |
615 | "mlock\0" | |
616 | "mlock2\0" | |
617 | "mlockall\0" | |
618 | "munlock\0" | |
619 | "munlockall\0" | |
620 | }, | |
8130926d | 621 | [SYSCALL_FILTER_SET_MODULE] = { |
8130926d | 622 | .name = "@module", |
d5efc18b | 623 | .help = "Loading and unloading of kernel modules", |
201c1cc2 | 624 | .value = |
201c1cc2 TM |
625 | "delete_module\0" |
626 | "finit_module\0" | |
627 | "init_module\0" | |
8130926d LP |
628 | }, |
629 | [SYSCALL_FILTER_SET_MOUNT] = { | |
8130926d | 630 | .name = "@mount", |
d5efc18b | 631 | .help = "Mounting and unmounting of file systems", |
201c1cc2 TM |
632 | .value = |
633 | "chroot\0" | |
9e486265 LP |
634 | "fsconfig\0" |
635 | "fsmount\0" | |
636 | "fsopen\0" | |
637 | "fspick\0" | |
201c1cc2 | 638 | "mount\0" |
34254e59 | 639 | "mount_setattr\0" |
9e486265 LP |
640 | "move_mount\0" |
641 | "open_tree\0" | |
201c1cc2 | 642 | "pivot_root\0" |
201c1cc2 | 643 | "umount\0" |
215728ff | 644 | "umount2\0" |
8130926d LP |
645 | }, |
646 | [SYSCALL_FILTER_SET_NETWORK_IO] = { | |
8130926d | 647 | .name = "@network-io", |
d5efc18b | 648 | .help = "Network or Unix socket IO, should not be needed if not network facing", |
201c1cc2 | 649 | .value = |
201c1cc2 | 650 | "accept\0" |
215728ff | 651 | "accept4\0" |
201c1cc2 TM |
652 | "bind\0" |
653 | "connect\0" | |
654 | "getpeername\0" | |
655 | "getsockname\0" | |
656 | "getsockopt\0" | |
657 | "listen\0" | |
658 | "recv\0" | |
659 | "recvfrom\0" | |
660 | "recvmmsg\0" | |
6ca67710 | 661 | "recvmmsg_time64\0" |
201c1cc2 TM |
662 | "recvmsg\0" |
663 | "send\0" | |
664 | "sendmmsg\0" | |
665 | "sendmsg\0" | |
666 | "sendto\0" | |
667 | "setsockopt\0" | |
668 | "shutdown\0" | |
669 | "socket\0" | |
670 | "socketcall\0" | |
671 | "socketpair\0" | |
8130926d LP |
672 | }, |
673 | [SYSCALL_FILTER_SET_OBSOLETE] = { | |
d5efc18b | 674 | /* some unknown even to libseccomp */ |
8130926d | 675 | .name = "@obsolete", |
d5efc18b | 676 | .help = "Unusual, obsolete or unimplemented system calls", |
201c1cc2 TM |
677 | .value = |
678 | "_sysctl\0" | |
679 | "afs_syscall\0" | |
802fa07a | 680 | "bdflush\0" |
201c1cc2 | 681 | "break\0" |
1f9ac68b | 682 | "create_module\0" |
201c1cc2 TM |
683 | "ftime\0" |
684 | "get_kernel_syms\0" | |
201c1cc2 TM |
685 | "getpmsg\0" |
686 | "gtty\0" | |
7e0c3b8f | 687 | "idle\0" |
201c1cc2 | 688 | "lock\0" |
201c1cc2 | 689 | "mpx\0" |
201c1cc2 TM |
690 | "prof\0" |
691 | "profil\0" | |
201c1cc2 TM |
692 | "putpmsg\0" |
693 | "query_module\0" | |
201c1cc2 TM |
694 | "security\0" |
695 | "sgetmask\0" | |
696 | "ssetmask\0" | |
ae5e9bf4 | 697 | "stime\0" |
201c1cc2 | 698 | "stty\0" |
1f9ac68b | 699 | "sysfs\0" |
201c1cc2 TM |
700 | "tuxcall\0" |
701 | "ulimit\0" | |
702 | "uselib\0" | |
1f9ac68b | 703 | "ustat\0" |
201c1cc2 | 704 | "vserver\0" |
8130926d | 705 | }, |
9493b168 ZJS |
706 | [SYSCALL_FILTER_SET_PKEY] = { |
707 | .name = "@pkey", | |
708 | .help = "System calls used for memory protection keys", | |
709 | .value = | |
710 | "pkey_alloc\0" | |
711 | "pkey_free\0" | |
712 | "pkey_mprotect\0" | |
713 | }, | |
8130926d | 714 | [SYSCALL_FILTER_SET_PRIVILEGED] = { |
8130926d | 715 | .name = "@privileged", |
d5efc18b | 716 | .help = "All system calls which need super-user capabilities", |
201c1cc2 | 717 | .value = |
44898c53 | 718 | "@chown\0" |
201c1cc2 TM |
719 | "@clock\0" |
720 | "@module\0" | |
721 | "@raw-io\0" | |
af0f047b LP |
722 | "@reboot\0" |
723 | "@swap\0" | |
215728ff | 724 | "_sysctl\0" |
201c1cc2 | 725 | "acct\0" |
201c1cc2 | 726 | "bpf\0" |
1f9ac68b | 727 | "capset\0" |
201c1cc2 | 728 | "chroot\0" |
a05cfe23 | 729 | "fanotify_init\0" |
9e486265 | 730 | "fanotify_mark\0" |
201c1cc2 | 731 | "nfsservctl\0" |
a05cfe23 | 732 | "open_by_handle_at\0" |
201c1cc2 TM |
733 | "pivot_root\0" |
734 | "quotactl\0" | |
76e86b8d | 735 | "quotactl_fd\0" |
201c1cc2 | 736 | "setdomainname\0" |
201c1cc2 | 737 | "setfsuid\0" |
215728ff | 738 | "setfsuid32\0" |
201c1cc2 | 739 | "setgroups\0" |
215728ff | 740 | "setgroups32\0" |
201c1cc2 | 741 | "sethostname\0" |
201c1cc2 | 742 | "setresuid\0" |
215728ff | 743 | "setresuid32\0" |
201c1cc2 | 744 | "setreuid\0" |
215728ff | 745 | "setreuid32\0" |
e05ee49b | 746 | "setuid\0" /* We list the explicit system calls here, as @setuid also includes setgid() which is not necessarily privileged */ |
215728ff | 747 | "setuid32\0" |
201c1cc2 | 748 | "vhangup\0" |
8130926d LP |
749 | }, |
750 | [SYSCALL_FILTER_SET_PROCESS] = { | |
8130926d | 751 | .name = "@process", |
7b121df6 | 752 | .help = "Process control, execution, namespacing operations", |
201c1cc2 | 753 | .value = |
09d3020b | 754 | "capget\0" /* Able to query arbitrary processes */ |
201c1cc2 | 755 | "clone\0" |
c5503601 ZJS |
756 | /* ia64 as the only architecture has clone2, a replacement for clone, but ia64 doesn't |
757 | * implement seccomp, so we don't need to list it at all. C.f. | |
758 | * acce2f71779c54086962fefce3833d886c655f62 in the kernel. */ | |
9e486265 | 759 | "clone3\0" |
201c1cc2 TM |
760 | "execveat\0" |
761 | "fork\0" | |
b887d2eb | 762 | "getrusage\0" |
201c1cc2 | 763 | "kill\0" |
9e486265 | 764 | "pidfd_open\0" |
46fcf95d | 765 | "pidfd_send_signal\0" |
201c1cc2 | 766 | "prctl\0" |
b887d2eb LP |
767 | "rt_sigqueueinfo\0" |
768 | "rt_tgsigqueueinfo\0" | |
201c1cc2 | 769 | "setns\0" |
a9518dc3 | 770 | "swapcontext\0" /* Some archs e.g. powerpc32 are using it to do userspace context switches */ |
201c1cc2 | 771 | "tgkill\0" |
b887d2eb | 772 | "times\0" |
201c1cc2 TM |
773 | "tkill\0" |
774 | "unshare\0" | |
775 | "vfork\0" | |
b887d2eb LP |
776 | "wait4\0" |
777 | "waitid\0" | |
778 | "waitpid\0" | |
8130926d LP |
779 | }, |
780 | [SYSCALL_FILTER_SET_RAW_IO] = { | |
8130926d | 781 | .name = "@raw-io", |
d5efc18b | 782 | .help = "Raw I/O port access", |
201c1cc2 TM |
783 | .value = |
784 | "ioperm\0" | |
785 | "iopl\0" | |
1f9ac68b | 786 | "pciconfig_iobase\0" |
201c1cc2 TM |
787 | "pciconfig_read\0" |
788 | "pciconfig_write\0" | |
789 | "s390_pci_mmio_read\0" | |
790 | "s390_pci_mmio_write\0" | |
8130926d | 791 | }, |
bd2ab3f4 LP |
792 | [SYSCALL_FILTER_SET_REBOOT] = { |
793 | .name = "@reboot", | |
794 | .help = "Reboot and reboot preparation/kexec", | |
795 | .value = | |
bd2ab3f4 | 796 | "kexec_file_load\0" |
e59608fa | 797 | "kexec_load\0" |
bd2ab3f4 LP |
798 | "reboot\0" |
799 | }, | |
133ddbbe | 800 | [SYSCALL_FILTER_SET_RESOURCES] = { |
133ddbbe | 801 | .name = "@resources", |
58a8f68b | 802 | .help = "Alter resource settings", |
133ddbbe | 803 | .value = |
0963c053 LP |
804 | "ioprio_set\0" |
805 | "mbind\0" | |
806 | "migrate_pages\0" | |
807 | "move_pages\0" | |
808 | "nice\0" | |
0963c053 LP |
809 | "sched_setaffinity\0" |
810 | "sched_setattr\0" | |
133ddbbe LP |
811 | "sched_setparam\0" |
812 | "sched_setscheduler\0" | |
0963c053 | 813 | "set_mempolicy\0" |
76e86b8d | 814 | "set_mempolicy_home_node\0" |
133ddbbe LP |
815 | "setpriority\0" |
816 | "setrlimit\0" | |
133ddbbe | 817 | }, |
d12632a8 LP |
818 | [SYSCALL_FILTER_SET_SANDBOX] = { |
819 | .name = "@sandbox", | |
820 | .help = "Sandbox functionality", | |
821 | .value = | |
822 | "landlock_add_rule\0" | |
823 | "landlock_create_ruleset\0" | |
824 | "landlock_restrict_self\0" | |
825 | "seccomp\0" | |
826 | }, | |
6eaaeee9 LP |
827 | [SYSCALL_FILTER_SET_SETUID] = { |
828 | .name = "@setuid", | |
829 | .help = "Operations for changing user/group credentials", | |
830 | .value = | |
6eaaeee9 | 831 | "setgid\0" |
215728ff | 832 | "setgid32\0" |
6eaaeee9 | 833 | "setgroups\0" |
215728ff | 834 | "setgroups32\0" |
6eaaeee9 | 835 | "setregid\0" |
215728ff | 836 | "setregid32\0" |
6eaaeee9 | 837 | "setresgid\0" |
215728ff | 838 | "setresgid32\0" |
6eaaeee9 | 839 | "setresuid\0" |
215728ff | 840 | "setresuid32\0" |
6eaaeee9 | 841 | "setreuid\0" |
215728ff | 842 | "setreuid32\0" |
6eaaeee9 | 843 | "setuid\0" |
215728ff | 844 | "setuid32\0" |
6eaaeee9 | 845 | }, |
cd0ddf6f LP |
846 | [SYSCALL_FILTER_SET_SIGNAL] = { |
847 | .name = "@signal", | |
848 | .help = "Process signal handling", | |
849 | .value = | |
850 | "rt_sigaction\0" | |
851 | "rt_sigpending\0" | |
852 | "rt_sigprocmask\0" | |
853 | "rt_sigsuspend\0" | |
854 | "rt_sigtimedwait\0" | |
6ca67710 | 855 | "rt_sigtimedwait_time64\0" |
cd0ddf6f LP |
856 | "sigaction\0" |
857 | "sigaltstack\0" | |
858 | "signal\0" | |
859 | "signalfd\0" | |
860 | "signalfd4\0" | |
861 | "sigpending\0" | |
862 | "sigprocmask\0" | |
863 | "sigsuspend\0" | |
864 | }, | |
bd2ab3f4 LP |
865 | [SYSCALL_FILTER_SET_SWAP] = { |
866 | .name = "@swap", | |
867 | .help = "Enable/disable swap devices", | |
868 | .value = | |
869 | "swapoff\0" | |
870 | "swapon\0" | |
871 | }, | |
44898c53 LP |
872 | [SYSCALL_FILTER_SET_SYNC] = { |
873 | .name = "@sync", | |
874 | .help = "Synchronize files and memory to storage", | |
875 | .value = | |
876 | "fdatasync\0" | |
877 | "fsync\0" | |
878 | "msync\0" | |
879 | "sync\0" | |
880 | "sync_file_range\0" | |
a8fb09f5 | 881 | "sync_file_range2\0" |
44898c53 LP |
882 | "syncfs\0" |
883 | }, | |
70526841 LP |
884 | [SYSCALL_FILTER_SET_SYSTEM_SERVICE] = { |
885 | .name = "@system-service", | |
886 | .help = "General system service operations", | |
887 | .value = | |
888 | "@aio\0" | |
889 | "@basic-io\0" | |
890 | "@chown\0" | |
891 | "@default\0" | |
892 | "@file-system\0" | |
893 | "@io-event\0" | |
894 | "@ipc\0" | |
895 | "@keyring\0" | |
896 | "@memlock\0" | |
897 | "@network-io\0" | |
898 | "@process\0" | |
899 | "@resources\0" | |
900 | "@setuid\0" | |
901 | "@signal\0" | |
902 | "@sync\0" | |
903 | "@timer\0" | |
26b682e8 | 904 | "arm_fadvise64_64\0" |
70526841 LP |
905 | "capget\0" |
906 | "capset\0" | |
907 | "copy_file_range\0" | |
908 | "fadvise64\0" | |
909 | "fadvise64_64\0" | |
910 | "flock\0" | |
911 | "get_mempolicy\0" | |
912 | "getcpu\0" | |
913 | "getpriority\0" | |
70526841 LP |
914 | "ioctl\0" |
915 | "ioprio_get\0" | |
916 | "kcmp\0" | |
917 | "madvise\0" | |
70526841 LP |
918 | "mremap\0" |
919 | "name_to_handle_at\0" | |
920 | "oldolduname\0" | |
921 | "olduname\0" | |
922 | "personality\0" | |
923 | "readahead\0" | |
924 | "readdir\0" | |
925 | "remap_file_pages\0" | |
926 | "sched_get_priority_max\0" | |
927 | "sched_get_priority_min\0" | |
70526841 LP |
928 | "sched_getattr\0" |
929 | "sched_getparam\0" | |
930 | "sched_getscheduler\0" | |
931 | "sched_rr_get_interval\0" | |
6ca67710 | 932 | "sched_rr_get_interval_time64\0" |
70526841 LP |
933 | "sched_yield\0" |
934 | "sendfile\0" | |
935 | "sendfile64\0" | |
936 | "setfsgid\0" | |
937 | "setfsgid32\0" | |
938 | "setfsuid\0" | |
939 | "setfsuid32\0" | |
940 | "setpgid\0" | |
941 | "setsid\0" | |
942 | "splice\0" | |
943 | "sysinfo\0" | |
944 | "tee\0" | |
945 | "umask\0" | |
946 | "uname\0" | |
947 | "userfaultfd\0" | |
948 | "vmsplice\0" | |
949 | }, | |
cd0ddf6f LP |
950 | [SYSCALL_FILTER_SET_TIMER] = { |
951 | .name = "@timer", | |
952 | .help = "Schedule operations by time", | |
953 | .value = | |
954 | "alarm\0" | |
955 | "getitimer\0" | |
956 | "setitimer\0" | |
957 | "timer_create\0" | |
958 | "timer_delete\0" | |
959 | "timer_getoverrun\0" | |
960 | "timer_gettime\0" | |
6ca67710 | 961 | "timer_gettime64\0" |
cd0ddf6f | 962 | "timer_settime\0" |
6ca67710 | 963 | "timer_settime64\0" |
cd0ddf6f LP |
964 | "timerfd_create\0" |
965 | "timerfd_gettime\0" | |
6ca67710 | 966 | "timerfd_gettime64\0" |
cd0ddf6f | 967 | "timerfd_settime\0" |
6ca67710 | 968 | "timerfd_settime64\0" |
cd0ddf6f LP |
969 | "times\0" |
970 | }, | |
95aac012 ZJS |
971 | [SYSCALL_FILTER_SET_KNOWN] = { |
972 | .name = "@known", | |
973 | .help = "All known syscalls declared in the kernel", | |
974 | .value = | |
6d6a0854 | 975 | "@obsolete\0" |
95aac012 ZJS |
976 | #include "syscall-list.h" |
977 | }, | |
201c1cc2 | 978 | }; |
8130926d LP |
979 | |
980 | const SyscallFilterSet *syscall_filter_set_find(const char *name) { | |
8130926d LP |
981 | if (isempty(name) || name[0] != '@') |
982 | return NULL; | |
983 | ||
077e8fc0 | 984 | for (unsigned i = 0; i < _SYSCALL_FILTER_SET_MAX; i++) |
8130926d LP |
985 | if (streq(syscall_filter_sets[i].name, name)) |
986 | return syscall_filter_sets + i; | |
987 | ||
988 | return NULL; | |
989 | } | |
990 | ||
000c0520 ZJS |
991 | static int add_syscall_filter_set( |
992 | scmp_filter_ctx seccomp, | |
993 | const SyscallFilterSet *set, | |
994 | uint32_t action, | |
995 | char **exclude, | |
996 | bool log_missing, | |
997 | char ***added); | |
998 | ||
999 | int seccomp_add_syscall_filter_item( | |
1000 | scmp_filter_ctx *seccomp, | |
1001 | const char *name, | |
1002 | uint32_t action, | |
1003 | char **exclude, | |
1004 | bool log_missing, | |
1005 | char ***added) { | |
69b1b241 LP |
1006 | |
1007 | assert(seccomp); | |
1008 | assert(name); | |
1009 | ||
960e4569 LP |
1010 | if (strv_contains(exclude, name)) |
1011 | return 0; | |
1012 | ||
000c0520 ZJS |
1013 | /* Any syscalls that are handled are added to the *added strv. The pointer |
1014 | * must be either NULL or point to a valid pre-initialized possibly-empty strv. */ | |
1015 | ||
69b1b241 LP |
1016 | if (name[0] == '@') { |
1017 | const SyscallFilterSet *other; | |
1018 | ||
1019 | other = syscall_filter_set_find(name); | |
baaa35ad ZJS |
1020 | if (!other) |
1021 | return log_debug_errno(SYNTHETIC_ERRNO(EINVAL), | |
1022 | "Filter set %s is not known!", | |
1023 | name); | |
69b1b241 | 1024 | |
000c0520 | 1025 | return add_syscall_filter_set(seccomp, other, action, exclude, log_missing, added); |
b54f36c6 | 1026 | |
69b1b241 | 1027 | } else { |
b54f36c6 | 1028 | int id, r; |
69b1b241 LP |
1029 | |
1030 | id = seccomp_syscall_resolve_name(name); | |
cff7bff8 | 1031 | if (id == __NR_SCMP_ERROR) { |
b54f36c6 ZJS |
1032 | if (log_missing) |
1033 | log_debug("System call %s is not known, ignoring.", name); | |
ff217dc3 | 1034 | return 0; |
cff7bff8 | 1035 | } |
69b1b241 LP |
1036 | |
1037 | r = seccomp_rule_add_exact(seccomp, action, id, 0); | |
b54f36c6 | 1038 | if (r < 0) { |
69b1b241 | 1039 | /* If the system call is not known on this architecture, then that's fine, let's ignore it */ |
7e86bd73 ZJS |
1040 | bool ignore = r == -EDOM; |
1041 | ||
1042 | if (!ignore || log_missing) | |
1043 | log_debug_errno(r, "Failed to add rule for system call %s() / %d%s: %m", | |
1044 | name, id, ignore ? ", ignoring" : ""); | |
1045 | if (!ignore) | |
1046 | return r; | |
b54f36c6 | 1047 | } |
69b1b241 | 1048 | |
000c0520 ZJS |
1049 | if (added) { |
1050 | r = strv_extend(added, name); | |
1051 | if (r < 0) | |
1052 | return r; | |
1053 | } | |
1054 | ||
b54f36c6 ZJS |
1055 | return 0; |
1056 | } | |
69b1b241 LP |
1057 | } |
1058 | ||
000c0520 | 1059 | static int add_syscall_filter_set( |
469830d1 | 1060 | scmp_filter_ctx seccomp, |
469830d1 | 1061 | const SyscallFilterSet *set, |
960e4569 | 1062 | uint32_t action, |
b54f36c6 | 1063 | char **exclude, |
000c0520 ZJS |
1064 | bool log_missing, |
1065 | char ***added) { | |
469830d1 | 1066 | |
8130926d LP |
1067 | int r; |
1068 | ||
000c0520 ZJS |
1069 | /* Any syscalls that are handled are added to the *added strv. It needs to be initialized. */ |
1070 | ||
8130926d LP |
1071 | assert(seccomp); |
1072 | assert(set); | |
1073 | ||
1074 | NULSTR_FOREACH(sys, set->value) { | |
000c0520 | 1075 | r = seccomp_add_syscall_filter_item(seccomp, sys, action, exclude, log_missing, added); |
69b1b241 LP |
1076 | if (r < 0) |
1077 | return r; | |
469830d1 LP |
1078 | } |
1079 | ||
1080 | return 0; | |
1081 | } | |
1082 | ||
03c0730f YW |
1083 | static uint32_t override_default_action(uint32_t default_action) { |
1084 | /* When the requested filter is an allow-list, and the default action is something critical, we | |
1085 | * install ENOSYS as the default action, but it will only apply to syscalls which are not in the | |
1086 | * @known set. */ | |
1087 | ||
1088 | if (default_action == SCMP_ACT_ALLOW) | |
1089 | return default_action; | |
1090 | ||
1091 | #ifdef SCMP_ACT_LOG | |
1092 | if (default_action == SCMP_ACT_LOG) | |
1093 | return default_action; | |
1094 | #endif | |
1095 | ||
1096 | return SCMP_ACT_ERRNO(ENOSYS); | |
1097 | } | |
1098 | ||
b54f36c6 | 1099 | int seccomp_load_syscall_filter_set(uint32_t default_action, const SyscallFilterSet *set, uint32_t action, bool log_missing) { |
03c0730f | 1100 | uint32_t arch, default_action_override; |
469830d1 LP |
1101 | int r; |
1102 | ||
1103 | assert(set); | |
1104 | ||
1105 | /* The one-stop solution: allocate a seccomp object, add the specified filter to it, and apply it. Once for | |
a90db619 | 1106 | * each local arch. */ |
469830d1 | 1107 | |
78b2ad7d YW |
1108 | default_action_override = override_default_action(default_action); |
1109 | ||
469830d1 LP |
1110 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { |
1111 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
78b2ad7d | 1112 | _cleanup_strv_free_ char **added = NULL; |
469830d1 | 1113 | |
30868c1c | 1114 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
469830d1 | 1115 | |
78b2ad7d | 1116 | r = seccomp_init_for_arch(&seccomp, arch, default_action_override); |
8130926d LP |
1117 | if (r < 0) |
1118 | return r; | |
469830d1 | 1119 | |
78b2ad7d | 1120 | r = add_syscall_filter_set(seccomp, set, action, NULL, log_missing, &added); |
7e86bd73 ZJS |
1121 | if (r < 0) |
1122 | return log_debug_errno(r, "Failed to add filter set: %m"); | |
469830d1 | 1123 | |
78b2ad7d YW |
1124 | if (default_action != default_action_override) |
1125 | NULSTR_FOREACH(name, syscall_filter_sets[SYSCALL_FILTER_SET_KNOWN].value) { | |
1126 | int id; | |
1127 | ||
1128 | id = seccomp_syscall_resolve_name(name); | |
1129 | if (id < 0) | |
1130 | continue; | |
1131 | ||
1132 | /* Ignore the syscall if it was already handled above */ | |
1133 | if (strv_contains(added, name)) | |
1134 | continue; | |
1135 | ||
1136 | r = seccomp_rule_add_exact(seccomp, default_action, id, 0); | |
1137 | if (r < 0 && r != -EDOM) /* EDOM means that the syscall is not available for arch */ | |
1138 | return log_debug_errno(r, "Failed to add rule for system call %s() / %d: %m", | |
1139 | name, id); | |
1140 | } | |
1141 | ||
1142 | #if (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR >= 5) || SCMP_VER_MAJOR > 2 | |
1143 | /* We have a large filter here, so let's turn on the binary tree mode if possible. */ | |
1144 | r = seccomp_attr_set(seccomp, SCMP_FLTATR_CTL_OPTIMIZE, 2); | |
1145 | if (r < 0) | |
1146 | log_warning_errno(r, "Failed to set SCMP_FLTATR_CTL_OPTIMIZE, ignoring: %m"); | |
1147 | #endif | |
1148 | ||
469830d1 | 1149 | r = seccomp_load(seccomp); |
3c098014 ZJS |
1150 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1151 | return r; | |
1152 | if (r < 0) | |
1153 | log_debug_errno(r, "Failed to install filter set for architecture %s, skipping: %m", | |
1154 | seccomp_arch_to_string(arch)); | |
8130926d LP |
1155 | } |
1156 | ||
1157 | return 0; | |
1158 | } | |
a3be2849 | 1159 | |
1862b310 | 1160 | int seccomp_load_syscall_filter_set_raw(uint32_t default_action, Hashmap* filter, uint32_t action, bool log_missing) { |
03c0730f | 1161 | uint32_t arch, default_action_override; |
a3be2849 LP |
1162 | int r; |
1163 | ||
1862b310 YW |
1164 | /* Similar to seccomp_load_syscall_filter_set(), but takes a raw Hashmap* of syscalls, instead |
1165 | * of a SyscallFilterSet* table. */ | |
a3be2849 | 1166 | |
1862b310 | 1167 | if (hashmap_isempty(filter) && default_action == SCMP_ACT_ALLOW) |
469830d1 | 1168 | return 0; |
a3be2849 | 1169 | |
03c0730f YW |
1170 | default_action_override = override_default_action(default_action); |
1171 | ||
469830d1 LP |
1172 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { |
1173 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
b54f36c6 | 1174 | void *syscall_id, *val; |
a3be2849 | 1175 | |
30868c1c | 1176 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
a3be2849 | 1177 | |
03c0730f | 1178 | r = seccomp_init_for_arch(&seccomp, arch, default_action_override); |
469830d1 LP |
1179 | if (r < 0) |
1180 | return r; | |
a3be2849 | 1181 | |
1862b310 | 1182 | HASHMAP_FOREACH_KEY(val, syscall_id, filter) { |
8cfa775f | 1183 | uint32_t a = action; |
b54f36c6 ZJS |
1184 | int id = PTR_TO_INT(syscall_id) - 1; |
1185 | int error = PTR_TO_INT(val); | |
8cfa775f | 1186 | |
005bfaf1 TM |
1187 | if (error == SECCOMP_ERROR_NUMBER_KILL) |
1188 | a = scmp_act_kill_process(); | |
9df2cdd8 TM |
1189 | #ifdef SCMP_ACT_LOG |
1190 | else if (action == SCMP_ACT_LOG) | |
1191 | a = SCMP_ACT_LOG; | |
1192 | #endif | |
68acc1af | 1193 | else if (error >= 0) |
b54f36c6 | 1194 | a = SCMP_ACT_ERRNO(error); |
8cfa775f | 1195 | |
b54f36c6 | 1196 | r = seccomp_rule_add_exact(seccomp, a, id, 0); |
469830d1 | 1197 | if (r < 0) { |
1862b310 YW |
1198 | /* If the system call is not known on this architecture, then that's |
1199 | * fine, let's ignore it */ | |
469830d1 | 1200 | _cleanup_free_ char *n = NULL; |
7e86bd73 | 1201 | bool ignore; |
469830d1 | 1202 | |
b54f36c6 | 1203 | n = seccomp_syscall_resolve_num_arch(SCMP_ARCH_NATIVE, id); |
7e86bd73 ZJS |
1204 | ignore = r == -EDOM; |
1205 | if (!ignore || log_missing) | |
1206 | log_debug_errno(r, "Failed to add rule for system call %s() / %d%s: %m", | |
1207 | strna(n), id, ignore ? ", ignoring" : ""); | |
1208 | if (!ignore) | |
1209 | return r; | |
469830d1 LP |
1210 | } |
1211 | } | |
1212 | ||
03c0730f YW |
1213 | if (default_action != default_action_override) |
1214 | NULSTR_FOREACH(name, syscall_filter_sets[SYSCALL_FILTER_SET_KNOWN].value) { | |
1215 | int id; | |
2331c02d | 1216 | |
03c0730f YW |
1217 | id = seccomp_syscall_resolve_name(name); |
1218 | if (id < 0) | |
1219 | continue; | |
2331c02d | 1220 | |
03c0730f YW |
1221 | /* Ignore the syscall if it was already handled above */ |
1222 | if (hashmap_contains(filter, INT_TO_PTR(id + 1))) | |
1223 | continue; | |
2331c02d | 1224 | |
03c0730f YW |
1225 | r = seccomp_rule_add_exact(seccomp, default_action, id, 0); |
1226 | if (r < 0 && r != -EDOM) /* EDOM means that the syscall is not available for arch */ | |
1227 | return log_debug_errno(r, "Failed to add rule for system call %s() / %d: %m", | |
1228 | name, id); | |
1229 | } | |
2331c02d | 1230 | |
e6c5386d ZJS |
1231 | #if (SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR >= 5) || SCMP_VER_MAJOR > 2 |
1232 | /* We have a large filter here, so let's turn on the binary tree mode if possible. */ | |
1233 | r = seccomp_attr_set(seccomp, SCMP_FLTATR_CTL_OPTIMIZE, 2); | |
1234 | if (r < 0) | |
1235 | log_warning_errno(r, "Failed to set SCMP_FLTATR_CTL_OPTIMIZE, ignoring: %m"); | |
1236 | #endif | |
1237 | ||
469830d1 | 1238 | r = seccomp_load(seccomp); |
3c098014 ZJS |
1239 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1240 | return r; | |
1241 | if (r < 0) | |
a52765a5 | 1242 | log_debug_errno(r, "Failed to install system call filter for architecture %s, skipping: %m", |
1862b310 | 1243 | seccomp_arch_to_string(arch)); |
469830d1 LP |
1244 | } |
1245 | ||
1246 | return 0; | |
add00535 LP |
1247 | } |
1248 | ||
58f6ab44 | 1249 | int seccomp_parse_syscall_filter( |
898748d8 YW |
1250 | const char *name, |
1251 | int errno_num, | |
1252 | Hashmap *filter, | |
13d92c63 | 1253 | SeccompParseFlags flags, |
898748d8 YW |
1254 | const char *unit, |
1255 | const char *filename, | |
1256 | unsigned line) { | |
1257 | ||
1258 | int r; | |
1259 | ||
1260 | assert(name); | |
1261 | assert(filter); | |
1262 | ||
084a46d7 YW |
1263 | if (!FLAGS_SET(flags, SECCOMP_PARSE_INVERT) && errno_num >= 0) |
1264 | return -EINVAL; | |
1265 | ||
898748d8 YW |
1266 | if (name[0] == '@') { |
1267 | const SyscallFilterSet *set; | |
898748d8 YW |
1268 | |
1269 | set = syscall_filter_set_find(name); | |
1270 | if (!set) { | |
9e29ee40 | 1271 | if (!FLAGS_SET(flags, SECCOMP_PARSE_PERMISSIVE)) |
898748d8 | 1272 | return -EINVAL; |
13d92c63 | 1273 | |
9e29ee40 | 1274 | log_syntax(unit, FLAGS_SET(flags, SECCOMP_PARSE_LOG) ? LOG_WARNING : LOG_DEBUG, filename, line, 0, |
13d92c63 LP |
1275 | "Unknown system call group, ignoring: %s", name); |
1276 | return 0; | |
898748d8 YW |
1277 | } |
1278 | ||
1279 | NULSTR_FOREACH(i, set->value) { | |
3c098014 ZJS |
1280 | /* Call ourselves again, for the group to parse. Note that we downgrade logging here |
1281 | * (i.e. take away the SECCOMP_PARSE_LOG flag) since any issues in the group table | |
1282 | * are our own problem, not a problem in user configuration data and we shouldn't | |
1283 | * pretend otherwise by complaining about them. */ | |
58f6ab44 | 1284 | r = seccomp_parse_syscall_filter(i, errno_num, filter, flags &~ SECCOMP_PARSE_LOG, unit, filename, line); |
898748d8 YW |
1285 | if (r < 0) |
1286 | return r; | |
1287 | } | |
1288 | } else { | |
1289 | int id; | |
1290 | ||
1291 | id = seccomp_syscall_resolve_name(name); | |
1292 | if (id == __NR_SCMP_ERROR) { | |
9e29ee40 | 1293 | if (!FLAGS_SET(flags, SECCOMP_PARSE_PERMISSIVE)) |
898748d8 | 1294 | return -EINVAL; |
13d92c63 | 1295 | |
9e29ee40 | 1296 | log_syntax(unit, FLAGS_SET(flags, SECCOMP_PARSE_LOG) ? LOG_WARNING : LOG_DEBUG, filename, line, 0, |
309a4212 | 1297 | "System call %s is not known, ignoring.", name); |
13d92c63 | 1298 | return 0; |
898748d8 YW |
1299 | } |
1300 | ||
3c098014 ZJS |
1301 | /* If we previously wanted to forbid a syscall and now we want to allow it, then remove it |
1302 | * from the list. The entries in allow-list with non-negative error value will be handled | |
1303 | * with SCMP_ACT_ERRNO() instead of the default action. */ | |
68acc1af YW |
1304 | if (!FLAGS_SET(flags, SECCOMP_PARSE_INVERT) == FLAGS_SET(flags, SECCOMP_PARSE_ALLOW_LIST) || |
1305 | (FLAGS_SET(flags, SECCOMP_PARSE_INVERT | SECCOMP_PARSE_ALLOW_LIST) && errno_num >= 0)) { | |
898748d8 YW |
1306 | r = hashmap_put(filter, INT_TO_PTR(id + 1), INT_TO_PTR(errno_num)); |
1307 | if (r < 0) | |
851ee70a LW |
1308 | switch (r) { |
1309 | case -ENOMEM: | |
9e29ee40 | 1310 | return FLAGS_SET(flags, SECCOMP_PARSE_LOG) ? log_oom() : -ENOMEM; |
851ee70a | 1311 | case -EEXIST: |
9d7fe7c6 LW |
1312 | assert_se(hashmap_update(filter, INT_TO_PTR(id + 1), INT_TO_PTR(errno_num)) == 0); |
1313 | break; | |
851ee70a LW |
1314 | default: |
1315 | return r; | |
1316 | } | |
898748d8 YW |
1317 | } else |
1318 | (void) hashmap_remove(filter, INT_TO_PTR(id + 1)); | |
1319 | } | |
1320 | ||
1321 | return 0; | |
1322 | } | |
1323 | ||
add00535 | 1324 | int seccomp_restrict_namespaces(unsigned long retain) { |
469830d1 | 1325 | uint32_t arch; |
add00535 LP |
1326 | int r; |
1327 | ||
f1d34068 | 1328 | if (DEBUG_LOGGING) { |
add00535 LP |
1329 | _cleanup_free_ char *s = NULL; |
1330 | ||
86c2a9f1 | 1331 | (void) namespace_flags_to_string(retain, &s); |
add00535 LP |
1332 | log_debug("Restricting namespace to: %s.", strna(s)); |
1333 | } | |
1334 | ||
1335 | /* NOOP? */ | |
d7a0f1f4 | 1336 | if (FLAGS_SET(retain, NAMESPACE_FLAGS_ALL)) |
add00535 LP |
1337 | return 0; |
1338 | ||
469830d1 LP |
1339 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { |
1340 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
add00535 | 1341 | |
30868c1c | 1342 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
469830d1 LP |
1343 | |
1344 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); | |
1345 | if (r < 0) | |
1346 | return r; | |
1347 | ||
30193fe8 ZJS |
1348 | /* We cannot filter on individual flags to clone3(), and we need to disable the |
1349 | * syscall altogether. ENOSYS is used instead of EPERM, so that glibc and other | |
1350 | * users shall fall back to clone(), as if on an older kernel. | |
1351 | * | |
1352 | * C.f. https://github.com/flatpak/flatpak/commit/a10f52a7565c549612c92b8e736a6698a53db330, | |
1353 | * https://github.com/moby/moby/issues/42680. */ | |
1354 | ||
1355 | r = seccomp_rule_add_exact( | |
1356 | seccomp, | |
1357 | SCMP_ACT_ERRNO(ENOSYS), | |
1358 | SCMP_SYS(clone3), | |
1359 | 0); | |
1360 | if (r < 0) | |
3c098014 ZJS |
1361 | log_debug_errno(r, "Failed to add clone3() rule for architecture %s, ignoring: %m", |
1362 | seccomp_arch_to_string(arch)); | |
30193fe8 | 1363 | |
469830d1 | 1364 | if ((retain & NAMESPACE_FLAGS_ALL) == 0) |
3c098014 ZJS |
1365 | /* If every single kind of namespace shall be prohibited, then let's block the whole |
1366 | * setns() syscall altogether. */ | |
469830d1 LP |
1367 | r = seccomp_rule_add_exact( |
1368 | seccomp, | |
1369 | SCMP_ACT_ERRNO(EPERM), | |
1370 | SCMP_SYS(setns), | |
1371 | 0); | |
1372 | else | |
3c098014 ZJS |
1373 | /* Otherwise, block only the invocations with the appropriate flags in the loop |
1374 | * below, but also the special invocation with a zero flags argument, right here. */ | |
469830d1 LP |
1375 | r = seccomp_rule_add_exact( |
1376 | seccomp, | |
1377 | SCMP_ACT_ERRNO(EPERM), | |
1378 | SCMP_SYS(setns), | |
1379 | 1, | |
1380 | SCMP_A1(SCMP_CMP_EQ, 0)); | |
1381 | if (r < 0) { | |
3c098014 ZJS |
1382 | log_debug_errno(r, "Failed to add setns() rule for architecture %s, skipping: %m", |
1383 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1384 | continue; |
1385 | } | |
1386 | ||
241b1577 | 1387 | for (unsigned i = 0; namespace_info[i].proc_name; i++) { |
469830d1 LP |
1388 | unsigned long f; |
1389 | ||
241b1577 | 1390 | f = namespace_info[i].clone_flag; |
d7a0f1f4 | 1391 | if (FLAGS_SET(retain, f)) { |
241b1577 | 1392 | log_debug("Permitting %s.", namespace_info[i].proc_name); |
469830d1 LP |
1393 | continue; |
1394 | } | |
1395 | ||
30868c1c | 1396 | log_trace("Blocking %s.", namespace_info[i].proc_name); |
469830d1 LP |
1397 | |
1398 | r = seccomp_rule_add_exact( | |
1399 | seccomp, | |
1400 | SCMP_ACT_ERRNO(EPERM), | |
1401 | SCMP_SYS(unshare), | |
1402 | 1, | |
1403 | SCMP_A0(SCMP_CMP_MASKED_EQ, f, f)); | |
1404 | if (r < 0) { | |
3c098014 ZJS |
1405 | log_debug_errno(r, "Failed to add unshare() rule for architecture %s, skipping: %m", |
1406 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1407 | break; |
1408 | } | |
1409 | ||
511ceb1f ZJS |
1410 | /* On s390/s390x the first two parameters to clone are switched */ |
1411 | if (!IN_SET(arch, SCMP_ARCH_S390, SCMP_ARCH_S390X)) | |
ae9d60ce LP |
1412 | r = seccomp_rule_add_exact( |
1413 | seccomp, | |
1414 | SCMP_ACT_ERRNO(EPERM), | |
1415 | SCMP_SYS(clone), | |
1416 | 1, | |
1417 | SCMP_A0(SCMP_CMP_MASKED_EQ, f, f)); | |
1418 | else | |
1419 | r = seccomp_rule_add_exact( | |
1420 | seccomp, | |
1421 | SCMP_ACT_ERRNO(EPERM), | |
1422 | SCMP_SYS(clone), | |
1423 | 1, | |
1424 | SCMP_A1(SCMP_CMP_MASKED_EQ, f, f)); | |
469830d1 | 1425 | if (r < 0) { |
3c098014 ZJS |
1426 | log_debug_errno(r, "Failed to add clone() rule for architecture %s, skipping: %m", |
1427 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1428 | break; |
1429 | } | |
1430 | ||
1431 | if ((retain & NAMESPACE_FLAGS_ALL) != 0) { | |
1432 | r = seccomp_rule_add_exact( | |
1433 | seccomp, | |
1434 | SCMP_ACT_ERRNO(EPERM), | |
1435 | SCMP_SYS(setns), | |
1436 | 1, | |
1437 | SCMP_A1(SCMP_CMP_MASKED_EQ, f, f)); | |
1438 | if (r < 0) { | |
3c098014 ZJS |
1439 | log_debug_errno(r, "Failed to add setns() rule for architecture %s, skipping: %m", |
1440 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1441 | break; |
1442 | } | |
1443 | } | |
1444 | } | |
1445 | if (r < 0) | |
1446 | continue; | |
1447 | ||
1448 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
1449 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1450 | return r; | |
1451 | if (r < 0) | |
1452 | log_debug_errno(r, "Failed to install namespace restriction rules for architecture %s, skipping: %m", | |
1453 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1454 | } |
1455 | ||
1456 | return 0; | |
1457 | } | |
1458 | ||
1459 | int seccomp_protect_sysctl(void) { | |
1460 | uint32_t arch; | |
1461 | int r; | |
1462 | ||
1463 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
1464 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
1465 | ||
30868c1c | 1466 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
469830d1 | 1467 | |
f9252236 AJ |
1468 | if (IN_SET(arch, |
1469 | SCMP_ARCH_AARCH64, | |
f9d3fb6b XW |
1470 | #ifdef SCMP_ARCH_LOONGARCH64 |
1471 | SCMP_ARCH_LOONGARCH64, | |
1472 | #endif | |
f9252236 AJ |
1473 | #ifdef SCMP_ARCH_RISCV64 |
1474 | SCMP_ARCH_RISCV64, | |
1475 | #endif | |
1476 | SCMP_ARCH_X32 | |
1477 | )) | |
2e64e8f4 ZJS |
1478 | /* No _sysctl syscall */ |
1479 | continue; | |
1480 | ||
469830d1 LP |
1481 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); |
1482 | if (r < 0) | |
1483 | return r; | |
1484 | ||
1485 | r = seccomp_rule_add_exact( | |
add00535 LP |
1486 | seccomp, |
1487 | SCMP_ACT_ERRNO(EPERM), | |
469830d1 | 1488 | SCMP_SYS(_sysctl), |
add00535 | 1489 | 0); |
469830d1 | 1490 | if (r < 0) { |
3c098014 ZJS |
1491 | log_debug_errno(r, "Failed to add _sysctl() rule for architecture %s, skipping: %m", |
1492 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1493 | continue; |
1494 | } | |
1495 | ||
1496 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
1497 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1498 | return r; | |
1499 | if (r < 0) | |
1500 | log_debug_errno(r, "Failed to install sysctl protection rules for architecture %s, skipping: %m", | |
1501 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1502 | } |
1503 | ||
1504 | return 0; | |
1505 | } | |
1506 | ||
620dbdd2 KK |
1507 | int seccomp_protect_syslog(void) { |
1508 | uint32_t arch; | |
1509 | int r; | |
1510 | ||
1511 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
1512 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
1513 | ||
1514 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); | |
1515 | if (r < 0) | |
1516 | return r; | |
1517 | ||
1518 | r = seccomp_rule_add_exact( | |
1519 | seccomp, | |
1520 | SCMP_ACT_ERRNO(EPERM), | |
1521 | SCMP_SYS(syslog), | |
1522 | 0); | |
1523 | ||
1524 | if (r < 0) { | |
1525 | log_debug_errno(r, "Failed to add syslog() rule for architecture %s, skipping %m", seccomp_arch_to_string(arch)); | |
1526 | continue; | |
1527 | } | |
1528 | ||
1529 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
1530 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1531 | return r; | |
1532 | if (r < 0) | |
1533 | log_debug_errno(r, "Failed to install syslog protection rules for architecture %s, skipping %m", | |
1534 | seccomp_arch_to_string(arch)); | |
620dbdd2 KK |
1535 | } |
1536 | ||
1537 | return 0; | |
1538 | } | |
1539 | ||
6b000af4 | 1540 | int seccomp_restrict_address_families(Set *address_families, bool allow_list) { |
469830d1 LP |
1541 | uint32_t arch; |
1542 | int r; | |
1543 | ||
1544 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
1545 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
9606bc4b | 1546 | bool supported; |
469830d1 | 1547 | |
30868c1c | 1548 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
469830d1 | 1549 | |
9606bc4b LP |
1550 | switch (arch) { |
1551 | ||
1552 | case SCMP_ARCH_X86_64: | |
1553 | case SCMP_ARCH_X32: | |
1554 | case SCMP_ARCH_ARM: | |
1555 | case SCMP_ARCH_AARCH64: | |
f9d3fb6b XW |
1556 | #ifdef SCMP_ARCH_LOONGARCH64 |
1557 | case SCMP_ARCH_LOONGARCH64: | |
1558 | #endif | |
f5aeac14 JC |
1559 | case SCMP_ARCH_MIPSEL64N32: |
1560 | case SCMP_ARCH_MIPS64N32: | |
1561 | case SCMP_ARCH_MIPSEL64: | |
1562 | case SCMP_ARCH_MIPS64: | |
f9252236 AJ |
1563 | #ifdef SCMP_ARCH_RISCV64 |
1564 | case SCMP_ARCH_RISCV64: | |
1565 | #endif | |
9606bc4b LP |
1566 | /* These we know we support (i.e. are the ones that do not use socketcall()) */ |
1567 | supported = true; | |
1568 | break; | |
1569 | ||
9606bc4b LP |
1570 | case SCMP_ARCH_S390: |
1571 | case SCMP_ARCH_S390X: | |
da1921a5 | 1572 | case SCMP_ARCH_X86: |
f5aeac14 JC |
1573 | case SCMP_ARCH_MIPSEL: |
1574 | case SCMP_ARCH_MIPS: | |
344e6b62 SJ |
1575 | #ifdef SCMP_ARCH_PARISC |
1576 | case SCMP_ARCH_PARISC: | |
1577 | #endif | |
1578 | #ifdef SCMP_ARCH_PARISC64 | |
1579 | case SCMP_ARCH_PARISC64: | |
1580 | #endif | |
d5923e38 ZJS |
1581 | case SCMP_ARCH_PPC: |
1582 | case SCMP_ARCH_PPC64: | |
1583 | case SCMP_ARCH_PPC64LE: | |
9606bc4b LP |
1584 | default: |
1585 | /* These we either know we don't support (i.e. are the ones that do use socketcall()), or we | |
1586 | * don't know */ | |
1587 | supported = false; | |
1588 | break; | |
1589 | } | |
1590 | ||
1591 | if (!supported) | |
1592 | continue; | |
1593 | ||
469830d1 LP |
1594 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); |
1595 | if (r < 0) | |
1596 | return r; | |
1597 | ||
6b000af4 | 1598 | if (allow_list) { |
077e8fc0 | 1599 | int first = 0, last = 0; |
469830d1 LP |
1600 | void *afp; |
1601 | ||
6b000af4 LP |
1602 | /* If this is an allow list, we first block the address families that are out of |
1603 | * range and then everything that is not in the set. First, we find the lowest and | |
1604 | * highest address family in the set. */ | |
469830d1 | 1605 | |
90e74a66 | 1606 | SET_FOREACH(afp, address_families) { |
077e8fc0 | 1607 | int af = PTR_TO_INT(afp); |
469830d1 LP |
1608 | |
1609 | if (af <= 0 || af >= af_max()) | |
1610 | continue; | |
1611 | ||
1612 | if (first == 0 || af < first) | |
1613 | first = af; | |
1614 | ||
1615 | if (last == 0 || af > last) | |
1616 | last = af; | |
1617 | } | |
1618 | ||
1619 | assert((first == 0) == (last == 0)); | |
1620 | ||
1621 | if (first == 0) { | |
1622 | ||
1623 | /* No entries in the valid range, block everything */ | |
1624 | r = seccomp_rule_add_exact( | |
1625 | seccomp, | |
1626 | SCMP_ACT_ERRNO(EAFNOSUPPORT), | |
1627 | SCMP_SYS(socket), | |
1628 | 0); | |
1629 | if (r < 0) { | |
3c098014 ZJS |
1630 | log_debug_errno(r, "Failed to add socket() rule for architecture %s, skipping: %m", |
1631 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1632 | continue; |
1633 | } | |
1634 | ||
1635 | } else { | |
1636 | ||
1637 | /* Block everything below the first entry */ | |
1638 | r = seccomp_rule_add_exact( | |
1639 | seccomp, | |
1640 | SCMP_ACT_ERRNO(EAFNOSUPPORT), | |
1641 | SCMP_SYS(socket), | |
1642 | 1, | |
1643 | SCMP_A0(SCMP_CMP_LT, first)); | |
1644 | if (r < 0) { | |
3c098014 ZJS |
1645 | log_debug_errno(r, "Failed to add socket() rule for architecture %s, skipping: %m", |
1646 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1647 | continue; |
1648 | } | |
1649 | ||
1650 | /* Block everything above the last entry */ | |
1651 | r = seccomp_rule_add_exact( | |
1652 | seccomp, | |
1653 | SCMP_ACT_ERRNO(EAFNOSUPPORT), | |
1654 | SCMP_SYS(socket), | |
1655 | 1, | |
1656 | SCMP_A0(SCMP_CMP_GT, last)); | |
1657 | if (r < 0) { | |
3c098014 ZJS |
1658 | log_debug_errno(r, "Failed to add socket() rule for architecture %s, skipping: %m", |
1659 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1660 | continue; |
1661 | } | |
1662 | ||
1663 | /* Block everything between the first and last entry */ | |
077e8fc0 | 1664 | for (int af = 1; af < af_max(); af++) { |
469830d1 LP |
1665 | |
1666 | if (set_contains(address_families, INT_TO_PTR(af))) | |
1667 | continue; | |
1668 | ||
1669 | r = seccomp_rule_add_exact( | |
1670 | seccomp, | |
1671 | SCMP_ACT_ERRNO(EAFNOSUPPORT), | |
1672 | SCMP_SYS(socket), | |
1673 | 1, | |
1674 | SCMP_A0(SCMP_CMP_EQ, af)); | |
1675 | if (r < 0) | |
1676 | break; | |
1677 | } | |
469830d1 | 1678 | if (r < 0) { |
3c098014 ZJS |
1679 | log_debug_errno(r, "Failed to add socket() rule for architecture %s, skipping: %m", |
1680 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1681 | continue; |
1682 | } | |
1683 | } | |
1684 | ||
1685 | } else { | |
1686 | void *af; | |
1687 | ||
6b000af4 LP |
1688 | /* If this is a deny list, then generate one rule for each address family that are |
1689 | * then combined in OR checks. */ | |
469830d1 | 1690 | |
90e74a66 | 1691 | SET_FOREACH(af, address_families) { |
469830d1 LP |
1692 | r = seccomp_rule_add_exact( |
1693 | seccomp, | |
1694 | SCMP_ACT_ERRNO(EAFNOSUPPORT), | |
1695 | SCMP_SYS(socket), | |
1696 | 1, | |
1697 | SCMP_A0(SCMP_CMP_EQ, PTR_TO_INT(af))); | |
1698 | if (r < 0) | |
1699 | break; | |
1700 | } | |
469830d1 | 1701 | if (r < 0) { |
3c098014 ZJS |
1702 | log_debug_errno(r, "Failed to add socket() rule for architecture %s, skipping: %m", |
1703 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1704 | continue; |
1705 | } | |
1706 | } | |
1707 | ||
1708 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
1709 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1710 | return r; | |
1711 | if (r < 0) | |
1712 | log_debug_errno(r, "Failed to install socket family rules for architecture %s, skipping: %m", | |
1713 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1714 | } |
1715 | ||
1716 | return 0; | |
1717 | } | |
1718 | ||
a9002749 | 1719 | int seccomp_restrict_realtime_full(int error_code) { |
469830d1 LP |
1720 | static const int permitted_policies[] = { |
1721 | SCHED_OTHER, | |
1722 | SCHED_BATCH, | |
1723 | SCHED_IDLE, | |
1724 | }; | |
1725 | ||
1726 | int r, max_policy = 0; | |
1727 | uint32_t arch; | |
1728 | unsigned i; | |
1729 | ||
a9002749 YW |
1730 | assert(error_code > 0); |
1731 | ||
469830d1 LP |
1732 | /* Determine the highest policy constant we want to allow */ |
1733 | for (i = 0; i < ELEMENTSOF(permitted_policies); i++) | |
1734 | if (permitted_policies[i] > max_policy) | |
1735 | max_policy = permitted_policies[i]; | |
1736 | ||
1737 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
1738 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
1739 | int p; | |
1740 | ||
30868c1c | 1741 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
469830d1 LP |
1742 | |
1743 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); | |
1744 | if (r < 0) | |
1745 | return r; | |
1746 | ||
1747 | /* Go through all policies with lower values than that, and block them -- unless they appear in the | |
6b000af4 | 1748 | * allow list. */ |
469830d1 LP |
1749 | for (p = 0; p < max_policy; p++) { |
1750 | bool good = false; | |
1751 | ||
6b000af4 | 1752 | /* Check if this is in the allow list. */ |
469830d1 LP |
1753 | for (i = 0; i < ELEMENTSOF(permitted_policies); i++) |
1754 | if (permitted_policies[i] == p) { | |
1755 | good = true; | |
1756 | break; | |
1757 | } | |
1758 | ||
1759 | if (good) | |
1760 | continue; | |
1761 | ||
1762 | /* Deny this policy */ | |
1763 | r = seccomp_rule_add_exact( | |
1764 | seccomp, | |
a9002749 | 1765 | SCMP_ACT_ERRNO(error_code), |
469830d1 LP |
1766 | SCMP_SYS(sched_setscheduler), |
1767 | 1, | |
1768 | SCMP_A1(SCMP_CMP_EQ, p)); | |
1769 | if (r < 0) { | |
3c098014 ZJS |
1770 | log_debug_errno(r, "Failed to add scheduler rule for architecture %s, skipping: %m", |
1771 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1772 | continue; |
1773 | } | |
1774 | } | |
1775 | ||
6b000af4 LP |
1776 | /* Deny-list all other policies, i.e. the ones with higher values. Note that all comparisons |
1777 | * are unsigned here, hence no need no check for < 0 values. */ | |
469830d1 | 1778 | r = seccomp_rule_add_exact( |
add00535 | 1779 | seccomp, |
a9002749 | 1780 | SCMP_ACT_ERRNO(error_code), |
469830d1 | 1781 | SCMP_SYS(sched_setscheduler), |
add00535 | 1782 | 1, |
469830d1 LP |
1783 | SCMP_A1(SCMP_CMP_GT, max_policy)); |
1784 | if (r < 0) { | |
3c098014 ZJS |
1785 | log_debug_errno(r, "Failed to add scheduler rule for architecture %s, skipping: %m", |
1786 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1787 | continue; |
1788 | } | |
add00535 | 1789 | |
469830d1 | 1790 | r = seccomp_load(seccomp); |
3c098014 ZJS |
1791 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1792 | return r; | |
1793 | if (r < 0) | |
1794 | log_debug_errno(r, "Failed to install realtime protection rules for architecture %s, skipping: %m", | |
1795 | seccomp_arch_to_string(arch)); | |
469830d1 LP |
1796 | } |
1797 | ||
1798 | return 0; | |
1799 | } | |
1800 | ||
6dc66688 ZJS |
1801 | static int add_seccomp_syscall_filter(scmp_filter_ctx seccomp, |
1802 | uint32_t arch, | |
1803 | int nr, | |
14cb109d | 1804 | unsigned arg_cnt, |
6dc66688 ZJS |
1805 | const struct scmp_arg_cmp arg) { |
1806 | int r; | |
1807 | ||
1808 | r = seccomp_rule_add_exact(seccomp, SCMP_ACT_ERRNO(EPERM), nr, arg_cnt, arg); | |
1809 | if (r < 0) { | |
1810 | _cleanup_free_ char *n = NULL; | |
1811 | ||
1812 | n = seccomp_syscall_resolve_num_arch(arch, nr); | |
1813 | log_debug_errno(r, "Failed to add %s() rule for architecture %s, skipping: %m", | |
1814 | strna(n), | |
1815 | seccomp_arch_to_string(arch)); | |
1816 | } | |
1817 | ||
1818 | return r; | |
1819 | } | |
1820 | ||
2a8d6e63 | 1821 | /* For known architectures, check that syscalls are indeed defined or not. */ |
f9d3fb6b | 1822 | #if defined(__x86_64__) || defined(__arm__) || defined(__aarch64__) || defined(__loongarch_lp64) || (defined(__riscv) && __riscv_xlen == 64) |
2a8d6e63 ZJS |
1823 | assert_cc(SCMP_SYS(shmget) > 0); |
1824 | assert_cc(SCMP_SYS(shmat) > 0); | |
1825 | assert_cc(SCMP_SYS(shmdt) > 0); | |
2a8d6e63 | 1826 | #endif |
6dc66688 | 1827 | |
469830d1 LP |
1828 | int seccomp_memory_deny_write_execute(void) { |
1829 | uint32_t arch; | |
b069c2a3 | 1830 | unsigned loaded = 0; |
469830d1 LP |
1831 | |
1832 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
1833 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
b069c2a3 | 1834 | int filter_syscall = 0, block_syscall = 0, shmat_syscall = 0, r; |
add00535 | 1835 | |
30868c1c | 1836 | log_trace("Operating on architecture: %s", seccomp_arch_to_string(arch)); |
469830d1 | 1837 | |
8a50cf69 LP |
1838 | switch (arch) { |
1839 | ||
bed4668d CE |
1840 | /* Note that on some architectures shmat() isn't available, and the call is multiplexed through ipc(). |
1841 | * We ignore that here, which means there's still a way to get writable/executable | |
344e6b62 SJ |
1842 | * memory, if an IPC key is mapped like this. That's a pity, but no total loss. |
1843 | * | |
1844 | * Also, PARISC isn't here right now because it still needs executable memory, but work is in progress | |
1845 | * on that front (kernel work done in 5.18). | |
1846 | */ | |
bed4668d | 1847 | |
8a50cf69 | 1848 | case SCMP_ARCH_X86: |
57311925 | 1849 | case SCMP_ARCH_S390: |
8a50cf69 LP |
1850 | filter_syscall = SCMP_SYS(mmap2); |
1851 | block_syscall = SCMP_SYS(mmap); | |
bed4668d | 1852 | /* shmat multiplexed, see above */ |
2a8d6e63 ZJS |
1853 | break; |
1854 | ||
63d00dfb | 1855 | case SCMP_ARCH_PPC: |
2a8d6e63 ZJS |
1856 | case SCMP_ARCH_PPC64: |
1857 | case SCMP_ARCH_PPC64LE: | |
bed4668d | 1858 | case SCMP_ARCH_S390X: |
2a8d6e63 | 1859 | filter_syscall = SCMP_SYS(mmap); |
bed4668d | 1860 | /* shmat multiplexed, see above */ |
8a50cf69 LP |
1861 | break; |
1862 | ||
4278d1f5 ZJS |
1863 | case SCMP_ARCH_ARM: |
1864 | filter_syscall = SCMP_SYS(mmap2); /* arm has only mmap2 */ | |
1865 | shmat_syscall = SCMP_SYS(shmat); | |
1866 | break; | |
1867 | ||
8a50cf69 LP |
1868 | case SCMP_ARCH_X86_64: |
1869 | case SCMP_ARCH_X32: | |
79873bc8 | 1870 | case SCMP_ARCH_AARCH64: |
f9d3fb6b XW |
1871 | #ifdef SCMP_ARCH_LOONGARCH64 |
1872 | case SCMP_ARCH_LOONGARCH64: | |
1873 | #endif | |
f9252236 AJ |
1874 | #ifdef SCMP_ARCH_RISCV64 |
1875 | case SCMP_ARCH_RISCV64: | |
1876 | #endif | |
f9d3fb6b | 1877 | filter_syscall = SCMP_SYS(mmap); /* amd64, x32, arm64, loongarch64 and riscv64 have only mmap */ |
8a50cf69 LP |
1878 | shmat_syscall = SCMP_SYS(shmat); |
1879 | break; | |
1880 | ||
1881 | /* Please add more definitions here, if you port systemd to other architectures! */ | |
1882 | ||
f9d3fb6b | 1883 | #if !defined(__i386__) && !defined(__x86_64__) && !defined(__hppa__) && !defined(__hppa64__) && !defined(__powerpc__) && !defined(__powerpc64__) && !defined(__arm__) && !defined(__aarch64__) && !defined(__s390__) && !defined(__s390x__) && !(defined(__riscv) && __riscv_xlen == 64) && !defined(__loongarch_lp64) |
8a50cf69 LP |
1884 | #warning "Consider adding the right mmap() syscall definitions here!" |
1885 | #endif | |
1886 | } | |
1887 | ||
1888 | /* Can't filter mmap() on this arch, then skip it */ | |
1889 | if (filter_syscall == 0) | |
1890 | continue; | |
1891 | ||
469830d1 LP |
1892 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); |
1893 | if (r < 0) | |
1894 | return r; | |
1895 | ||
6dc66688 ZJS |
1896 | r = add_seccomp_syscall_filter(seccomp, arch, filter_syscall, |
1897 | 1, | |
1898 | SCMP_A2(SCMP_CMP_MASKED_EQ, PROT_EXEC|PROT_WRITE, PROT_EXEC|PROT_WRITE)); | |
1899 | if (r < 0) | |
1900 | continue; | |
8a50cf69 LP |
1901 | |
1902 | if (block_syscall != 0) { | |
6dc66688 ZJS |
1903 | r = add_seccomp_syscall_filter(seccomp, arch, block_syscall, 0, (const struct scmp_arg_cmp){} ); |
1904 | if (r < 0) | |
8a50cf69 | 1905 | continue; |
add00535 | 1906 | } |
a3be2849 | 1907 | |
6dc66688 ZJS |
1908 | r = add_seccomp_syscall_filter(seccomp, arch, SCMP_SYS(mprotect), |
1909 | 1, | |
b835eeb4 ZJS |
1910 | SCMP_A2(SCMP_CMP_MASKED_EQ, PROT_EXEC, PROT_EXEC)); |
1911 | if (r < 0) | |
1912 | continue; | |
1913 | ||
1914 | r = add_seccomp_syscall_filter(seccomp, arch, SCMP_SYS(pkey_mprotect), | |
1915 | 1, | |
6dc66688 ZJS |
1916 | SCMP_A2(SCMP_CMP_MASKED_EQ, PROT_EXEC, PROT_EXEC)); |
1917 | if (r < 0) | |
469830d1 | 1918 | continue; |
add00535 | 1919 | |
67fb5f33 | 1920 | if (shmat_syscall > 0) { |
5ef3ed97 | 1921 | r = add_seccomp_syscall_filter(seccomp, arch, shmat_syscall, |
6dc66688 ZJS |
1922 | 1, |
1923 | SCMP_A2(SCMP_CMP_MASKED_EQ, SHM_EXEC, SHM_EXEC)); | |
1924 | if (r < 0) | |
8a50cf69 | 1925 | continue; |
469830d1 LP |
1926 | } |
1927 | ||
1928 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
1929 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
1930 | return r; | |
1931 | if (r < 0) | |
b069c2a3 ZJS |
1932 | log_debug_errno(r, "Failed to install MemoryDenyWriteExecute= rule for architecture %s, skipping: %m", |
1933 | seccomp_arch_to_string(arch)); | |
903659e7 | 1934 | loaded++; |
469830d1 | 1935 | } |
add00535 | 1936 | |
903659e7 | 1937 | if (loaded == 0) |
b069c2a3 | 1938 | log_debug("Failed to install any seccomp rules for MemoryDenyWriteExecute=."); |
903659e7 CE |
1939 | |
1940 | return loaded; | |
469830d1 LP |
1941 | } |
1942 | ||
1943 | int seccomp_restrict_archs(Set *archs) { | |
1944 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
469830d1 | 1945 | int r; |
65976868 | 1946 | bool blocked_new = false; |
469830d1 LP |
1947 | |
1948 | /* This installs a filter with no rules, but that restricts the system call architectures to the specified | |
2428aaf8 AJ |
1949 | * list. |
1950 | * | |
1951 | * There are some qualifications. However the most important use is to stop processes from bypassing | |
1952 | * system call restrictions, in case they used a broader (multiplexing) syscall which is only available | |
1953 | * in a non-native architecture. There are no holes in this use case, at least so far. */ | |
469830d1 | 1954 | |
2428aaf8 AJ |
1955 | /* Note libseccomp includes our "native" (current) architecture in the filter by default. |
1956 | * We do not remove it. For example, our callers expect to be able to call execve() afterwards | |
1957 | * to run a program with the restrictions applied. */ | |
469830d1 LP |
1958 | seccomp = seccomp_init(SCMP_ACT_ALLOW); |
1959 | if (!seccomp) | |
1960 | return -ENOMEM; | |
1961 | ||
65976868 GDF |
1962 | for (unsigned i = 0; seccomp_local_archs[i] != SECCOMP_LOCAL_ARCH_END; ++i) { |
1963 | uint32_t arch = seccomp_local_archs[i]; | |
2428aaf8 | 1964 | |
f833df38 BB |
1965 | /* See above comment, our "native" architecture is never blocked. */ |
1966 | if (arch == seccomp_arch_native()) | |
1967 | continue; | |
1968 | ||
65976868 GDF |
1969 | /* That architecture might have already been blocked by a previous call to seccomp_restrict_archs. */ |
1970 | if (arch == SECCOMP_LOCAL_ARCH_BLOCKED) | |
1971 | continue; | |
2428aaf8 | 1972 | |
65976868 | 1973 | bool block = !set_contains(archs, UINT32_TO_PTR(arch + 1)); |
2428aaf8 | 1974 | |
65976868 GDF |
1975 | /* The vdso for x32 assumes that x86-64 syscalls are available. Let's allow them, since x32 |
1976 | * x32 syscalls should basically match x86-64 for everything except the pointer type. | |
1977 | * The important thing is that you can block the old 32-bit x86 syscalls. | |
1978 | * https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=850047 */ | |
1979 | if (block && arch == SCMP_ARCH_X86_64 && seccomp_arch_native() == SCMP_ARCH_X32) | |
1980 | block = !set_contains(archs, UINT32_TO_PTR(SCMP_ARCH_X32 + 1)); | |
1981 | ||
1982 | if (block) { | |
1983 | seccomp_local_archs[i] = SECCOMP_LOCAL_ARCH_BLOCKED; | |
1984 | blocked_new = true; | |
1985 | } else { | |
1986 | r = seccomp_arch_add(seccomp, arch); | |
1987 | if (r < 0 && r != -EEXIST) | |
1988 | return r; | |
1989 | } | |
add00535 LP |
1990 | } |
1991 | ||
65976868 GDF |
1992 | /* All architectures that will be blocked by the seccomp program were |
1993 | * already blocked. */ | |
1994 | if (!blocked_new) | |
1995 | return 0; | |
1996 | ||
469830d1 LP |
1997 | r = seccomp_attr_set(seccomp, SCMP_FLTATR_CTL_NNP, 0); |
1998 | if (r < 0) | |
1999 | return r; | |
add00535 | 2000 | |
1c6af69b | 2001 | r = seccomp_load(seccomp); |
3c098014 ZJS |
2002 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
2003 | return r; | |
2004 | if (r < 0) | |
1c6af69b LP |
2005 | log_debug_errno(r, "Failed to restrict system call architectures, skipping: %m"); |
2006 | ||
2007 | return 0; | |
a3be2849 | 2008 | } |
b16bd535 | 2009 | |
de7fef4b ZJS |
2010 | int parse_syscall_archs(char **l, Set **ret_archs) { |
2011 | _cleanup_set_free_ Set *archs = NULL; | |
b16bd535 YW |
2012 | int r; |
2013 | ||
2014 | assert(l); | |
de7fef4b | 2015 | assert(ret_archs); |
b16bd535 YW |
2016 | |
2017 | STRV_FOREACH(s, l) { | |
2018 | uint32_t a; | |
2019 | ||
2020 | r = seccomp_arch_from_string(*s, &a); | |
2021 | if (r < 0) | |
2022 | return -EINVAL; | |
2023 | ||
de7fef4b | 2024 | r = set_ensure_put(&archs, NULL, UINT32_TO_PTR(a + 1)); |
b16bd535 YW |
2025 | if (r < 0) |
2026 | return -ENOMEM; | |
2027 | } | |
2028 | ||
de7fef4b | 2029 | *ret_archs = TAKE_PTR(archs); |
b16bd535 YW |
2030 | return 0; |
2031 | } | |
165a31c0 | 2032 | |
8cfa775f | 2033 | int seccomp_filter_set_add(Hashmap *filter, bool add, const SyscallFilterSet *set) { |
165a31c0 LP |
2034 | int r; |
2035 | ||
2036 | assert(set); | |
2037 | ||
2038 | NULSTR_FOREACH(i, set->value) { | |
2039 | ||
2040 | if (i[0] == '@') { | |
2041 | const SyscallFilterSet *more; | |
2042 | ||
2043 | more = syscall_filter_set_find(i); | |
2044 | if (!more) | |
2045 | return -ENXIO; | |
2046 | ||
165a31c0 LP |
2047 | r = seccomp_filter_set_add(filter, add, more); |
2048 | if (r < 0) | |
2049 | return r; | |
2050 | } else { | |
2051 | int id; | |
2052 | ||
2053 | id = seccomp_syscall_resolve_name(i); | |
ff217dc3 | 2054 | if (id == __NR_SCMP_ERROR) { |
309a4212 | 2055 | log_debug("System call %s is not known, ignoring.", i); |
ff217dc3 LP |
2056 | continue; |
2057 | } | |
165a31c0 LP |
2058 | |
2059 | if (add) { | |
8cfa775f | 2060 | r = hashmap_put(filter, INT_TO_PTR(id + 1), INT_TO_PTR(-1)); |
165a31c0 LP |
2061 | if (r < 0) |
2062 | return r; | |
2063 | } else | |
8cfa775f | 2064 | (void) hashmap_remove(filter, INT_TO_PTR(id + 1)); |
165a31c0 LP |
2065 | } |
2066 | } | |
2067 | ||
2068 | return 0; | |
2069 | } | |
78e864e5 TM |
2070 | |
2071 | int seccomp_lock_personality(unsigned long personality) { | |
72eafe71 | 2072 | uint32_t arch; |
78e864e5 TM |
2073 | int r; |
2074 | ||
72eafe71 LP |
2075 | if (personality >= PERSONALITY_INVALID) |
2076 | return -EINVAL; | |
78e864e5 | 2077 | |
72eafe71 LP |
2078 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { |
2079 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
78e864e5 | 2080 | |
72eafe71 LP |
2081 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); |
2082 | if (r < 0) | |
2083 | return r; | |
2084 | ||
2085 | r = seccomp_rule_add_exact( | |
2086 | seccomp, | |
2087 | SCMP_ACT_ERRNO(EPERM), | |
2088 | SCMP_SYS(personality), | |
2089 | 1, | |
2090 | SCMP_A0(SCMP_CMP_NE, personality)); | |
448ac526 | 2091 | if (r < 0) { |
3c098014 ZJS |
2092 | log_debug_errno(r, "Failed to add scheduler rule for architecture %s, skipping: %m", |
2093 | seccomp_arch_to_string(arch)); | |
448ac526 LP |
2094 | continue; |
2095 | } | |
72eafe71 LP |
2096 | |
2097 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
2098 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
2099 | return r; | |
2100 | if (r < 0) | |
2101 | log_debug_errno(r, "Failed to enable personality lock for architecture %s, skipping: %m", | |
2102 | seccomp_arch_to_string(arch)); | |
72eafe71 LP |
2103 | } |
2104 | ||
2105 | return 0; | |
78e864e5 | 2106 | } |
aecd5ac6 TM |
2107 | |
2108 | int seccomp_protect_hostname(void) { | |
2109 | uint32_t arch; | |
2110 | int r; | |
2111 | ||
2112 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
2113 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
2114 | ||
2115 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); | |
2116 | if (r < 0) | |
2117 | return r; | |
2118 | ||
2119 | r = seccomp_rule_add_exact( | |
2120 | seccomp, | |
2121 | SCMP_ACT_ERRNO(EPERM), | |
2122 | SCMP_SYS(sethostname), | |
2123 | 0); | |
9e6e543c | 2124 | if (r < 0) { |
3c098014 ZJS |
2125 | log_debug_errno(r, "Failed to add sethostname() rule for architecture %s, skipping: %m", |
2126 | seccomp_arch_to_string(arch)); | |
aecd5ac6 | 2127 | continue; |
9e6e543c | 2128 | } |
aecd5ac6 TM |
2129 | |
2130 | r = seccomp_rule_add_exact( | |
2131 | seccomp, | |
2132 | SCMP_ACT_ERRNO(EPERM), | |
2133 | SCMP_SYS(setdomainname), | |
2134 | 0); | |
9e6e543c | 2135 | if (r < 0) { |
3c098014 ZJS |
2136 | log_debug_errno(r, "Failed to add setdomainname() rule for architecture %s, skipping: %m", |
2137 | seccomp_arch_to_string(arch)); | |
aecd5ac6 | 2138 | continue; |
9e6e543c | 2139 | } |
aecd5ac6 TM |
2140 | |
2141 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
2142 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
2143 | return r; | |
2144 | if (r < 0) | |
2145 | log_debug_errno(r, "Failed to apply hostname restrictions for architecture %s, skipping: %m", | |
2146 | seccomp_arch_to_string(arch)); | |
aecd5ac6 TM |
2147 | } |
2148 | ||
2149 | return 0; | |
2150 | } | |
3c27973b | 2151 | |
da4dc9a6 ZJS |
2152 | static int seccomp_restrict_sxid(scmp_filter_ctx seccomp, mode_t m) { |
2153 | /* Checks the mode_t parameter of the following system calls: | |
2154 | * | |
8b45281d | 2155 | * → chmod() + fchmod() + fchmodat() + fchmodat2() |
da4dc9a6 ZJS |
2156 | * → open() + creat() + openat() |
2157 | * → mkdir() + mkdirat() | |
2158 | * → mknod() + mknodat() | |
2159 | * | |
2160 | * Returns error if *everything* failed, and 0 otherwise. | |
2161 | */ | |
6d95e7d9 | 2162 | int r; |
da4dc9a6 ZJS |
2163 | bool any = false; |
2164 | ||
2165 | r = seccomp_rule_add_exact( | |
2166 | seccomp, | |
2167 | SCMP_ACT_ERRNO(EPERM), | |
2168 | SCMP_SYS(chmod), | |
2169 | 1, | |
2170 | SCMP_A1(SCMP_CMP_MASKED_EQ, m, m)); | |
2171 | if (r < 0) | |
2172 | log_debug_errno(r, "Failed to add filter for chmod: %m"); | |
2173 | else | |
2174 | any = true; | |
2175 | ||
2176 | r = seccomp_rule_add_exact( | |
2177 | seccomp, | |
2178 | SCMP_ACT_ERRNO(EPERM), | |
2179 | SCMP_SYS(fchmod), | |
2180 | 1, | |
2181 | SCMP_A1(SCMP_CMP_MASKED_EQ, m, m)); | |
2182 | if (r < 0) | |
2183 | log_debug_errno(r, "Failed to add filter for fchmod: %m"); | |
2184 | else | |
2185 | any = true; | |
2186 | ||
2187 | r = seccomp_rule_add_exact( | |
2188 | seccomp, | |
2189 | SCMP_ACT_ERRNO(EPERM), | |
2190 | SCMP_SYS(fchmodat), | |
2191 | 1, | |
2192 | SCMP_A2(SCMP_CMP_MASKED_EQ, m, m)); | |
2193 | if (r < 0) | |
2194 | log_debug_errno(r, "Failed to add filter for fchmodat: %m"); | |
2195 | else | |
2196 | any = true; | |
2197 | ||
8b45281d AM |
2198 | #if defined(__SNR_fchmodat2) |
2199 | r = seccomp_rule_add_exact( | |
2200 | seccomp, | |
2201 | SCMP_ACT_ERRNO(EPERM), | |
2202 | SCMP_SYS(fchmodat2), | |
2203 | 1, | |
2204 | SCMP_A2(SCMP_CMP_MASKED_EQ, m, m)); | |
2205 | #else | |
2206 | /* It looks like this libseccomp does not know about fchmodat2(). | |
2207 | * Pretend the fchmodat2() system call is not supported at all, | |
2208 | * regardless of the kernel version. */ | |
2209 | r = seccomp_rule_add_exact( | |
2210 | seccomp, | |
2211 | SCMP_ACT_ERRNO(ENOSYS), | |
2212 | __NR_fchmodat2, | |
2213 | 0); | |
2214 | #endif | |
2215 | if (r < 0) | |
2216 | log_debug_errno(r, "Failed to add filter for fchmodat2: %m"); | |
2217 | else | |
2218 | any = true; | |
2219 | ||
da4dc9a6 ZJS |
2220 | r = seccomp_rule_add_exact( |
2221 | seccomp, | |
2222 | SCMP_ACT_ERRNO(EPERM), | |
2223 | SCMP_SYS(mkdir), | |
2224 | 1, | |
2225 | SCMP_A1(SCMP_CMP_MASKED_EQ, m, m)); | |
2226 | if (r < 0) | |
2227 | log_debug_errno(r, "Failed to add filter for mkdir: %m"); | |
2228 | else | |
2229 | any = true; | |
2230 | ||
2231 | r = seccomp_rule_add_exact( | |
2232 | seccomp, | |
2233 | SCMP_ACT_ERRNO(EPERM), | |
2234 | SCMP_SYS(mkdirat), | |
2235 | 1, | |
2236 | SCMP_A2(SCMP_CMP_MASKED_EQ, m, m)); | |
2237 | if (r < 0) | |
2238 | log_debug_errno(r, "Failed to add filter for mkdirat: %m"); | |
2239 | else | |
2240 | any = true; | |
2241 | ||
2242 | r = seccomp_rule_add_exact( | |
2243 | seccomp, | |
2244 | SCMP_ACT_ERRNO(EPERM), | |
2245 | SCMP_SYS(mknod), | |
2246 | 1, | |
2247 | SCMP_A1(SCMP_CMP_MASKED_EQ, m, m)); | |
2248 | if (r < 0) | |
2249 | log_debug_errno(r, "Failed to add filter for mknod: %m"); | |
2250 | else | |
2251 | any = true; | |
2252 | ||
2253 | r = seccomp_rule_add_exact( | |
2254 | seccomp, | |
2255 | SCMP_ACT_ERRNO(EPERM), | |
2256 | SCMP_SYS(mknodat), | |
2257 | 1, | |
2258 | SCMP_A2(SCMP_CMP_MASKED_EQ, m, m)); | |
2259 | if (r < 0) | |
2260 | log_debug_errno(r, "Failed to add filter for mknodat: %m"); | |
2261 | else | |
2262 | any = true; | |
2263 | ||
da4dc9a6 ZJS |
2264 | r = seccomp_rule_add_exact( |
2265 | seccomp, | |
2266 | SCMP_ACT_ERRNO(EPERM), | |
2267 | SCMP_SYS(open), | |
2268 | 2, | |
2269 | SCMP_A1(SCMP_CMP_MASKED_EQ, O_CREAT, O_CREAT), | |
2270 | SCMP_A2(SCMP_CMP_MASKED_EQ, m, m)); | |
2271 | if (r < 0) | |
2272 | log_debug_errno(r, "Failed to add filter for open: %m"); | |
2273 | else | |
2274 | any = true; | |
da4dc9a6 ZJS |
2275 | |
2276 | r = seccomp_rule_add_exact( | |
2277 | seccomp, | |
2278 | SCMP_ACT_ERRNO(EPERM), | |
2279 | SCMP_SYS(openat), | |
2280 | 2, | |
2281 | SCMP_A2(SCMP_CMP_MASKED_EQ, O_CREAT, O_CREAT), | |
2282 | SCMP_A3(SCMP_CMP_MASKED_EQ, m, m)); | |
2283 | if (r < 0) | |
2284 | log_debug_errno(r, "Failed to add filter for openat: %m"); | |
2285 | else | |
2286 | any = true; | |
2287 | ||
ecc04067 LP |
2288 | #if defined(__SNR_openat2) |
2289 | /* The new openat2() system call can't be filtered sensibly, since it moves the flags parameter into | |
2290 | * an indirect structure. Let's block it entirely for now. That should be a reasonably OK thing to do | |
2291 | * for now, since openat2() is very new and code generally needs fallback logic anyway to be | |
57353d29 MG |
2292 | * compatible with kernels that are not absolutely recent. We would normally return EPERM for a |
2293 | * policy check, but this isn't strictly a policy check. Instead, we return ENOSYS to force programs | |
2294 | * to call open() or openat() instead. We can properly enforce policy for those functions. */ | |
ecc04067 LP |
2295 | r = seccomp_rule_add_exact( |
2296 | seccomp, | |
57353d29 | 2297 | SCMP_ACT_ERRNO(ENOSYS), |
ecc04067 LP |
2298 | SCMP_SYS(openat2), |
2299 | 0); | |
2300 | if (r < 0) | |
2301 | log_debug_errno(r, "Failed to add filter for openat2: %m"); | |
2302 | else | |
2303 | any = true; | |
2304 | #endif | |
2305 | ||
da4dc9a6 ZJS |
2306 | r = seccomp_rule_add_exact( |
2307 | seccomp, | |
2308 | SCMP_ACT_ERRNO(EPERM), | |
2309 | SCMP_SYS(creat), | |
2310 | 1, | |
2311 | SCMP_A1(SCMP_CMP_MASKED_EQ, m, m)); | |
2312 | if (r < 0) | |
2313 | log_debug_errno(r, "Failed to add filter for creat: %m"); | |
2314 | else | |
2315 | any = true; | |
2316 | ||
2317 | return any ? 0 : r; | |
2318 | } | |
2319 | ||
3c27973b LP |
2320 | int seccomp_restrict_suid_sgid(void) { |
2321 | uint32_t arch; | |
da4dc9a6 | 2322 | int r, k; |
3c27973b LP |
2323 | |
2324 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
2325 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
2326 | ||
2327 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); | |
2328 | if (r < 0) | |
2329 | return r; | |
2330 | ||
da4dc9a6 ZJS |
2331 | r = seccomp_restrict_sxid(seccomp, S_ISUID); |
2332 | if (r < 0) | |
3c098014 ZJS |
2333 | log_debug_errno(r, "Failed to add suid rule for architecture %s, ignoring: %m", |
2334 | seccomp_arch_to_string(arch)); | |
3c27973b | 2335 | |
da4dc9a6 ZJS |
2336 | k = seccomp_restrict_sxid(seccomp, S_ISGID); |
2337 | if (k < 0) | |
a539314a | 2338 | log_debug_errno(k, "Failed to add sgid rule for architecture %s, ignoring: %m", |
3c098014 | 2339 | seccomp_arch_to_string(arch)); |
3c27973b | 2340 | |
da4dc9a6 | 2341 | if (r < 0 && k < 0) |
3c27973b | 2342 | continue; |
3c27973b LP |
2343 | |
2344 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
2345 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
2346 | return r; | |
2347 | if (r < 0) | |
2348 | log_debug_errno(r, "Failed to apply suid/sgid restrictions for architecture %s, skipping: %m", | |
2349 | seccomp_arch_to_string(arch)); | |
3c27973b LP |
2350 | } |
2351 | ||
2352 | return 0; | |
2353 | } | |
915fb324 LP |
2354 | |
2355 | uint32_t scmp_act_kill_process(void) { | |
2356 | ||
2357 | /* Returns SCMP_ACT_KILL_PROCESS if it's supported, and SCMP_ACT_KILL_THREAD otherwise. We never | |
2358 | * actually want to use SCMP_ACT_KILL_THREAD as its semantics are nuts (killing arbitrary threads of | |
2359 | * a program is just a bad idea), but on old kernels/old libseccomp it is all we have, and at least | |
2360 | * for single-threaded apps does the right thing. */ | |
2361 | ||
2362 | #ifdef SCMP_ACT_KILL_PROCESS | |
2363 | if (seccomp_api_get() >= 3) | |
2364 | return SCMP_ACT_KILL_PROCESS; | |
2365 | #endif | |
2366 | ||
2367 | return SCMP_ACT_KILL; /* same as SCMP_ACT_KILL_THREAD */ | |
2368 | } | |
22eadc28 YW |
2369 | |
2370 | int parse_syscall_and_errno(const char *in, char **name, int *error) { | |
2371 | _cleanup_free_ char *n = NULL; | |
2372 | char *p; | |
2373 | int e = -1; | |
2374 | ||
2375 | assert(in); | |
2376 | assert(name); | |
2377 | assert(error); | |
2378 | ||
2379 | /* | |
2380 | * This parse "syscall:errno" like "uname:EILSEQ", "@sync:255". | |
2381 | * If errno is omitted, then error is set to -1. | |
2382 | * Empty syscall name is not allowed. | |
2383 | * Here, we do not check that the syscall name is valid or not. | |
2384 | */ | |
2385 | ||
2386 | p = strchr(in, ':'); | |
2387 | if (p) { | |
2388 | e = seccomp_parse_errno_or_action(p + 1); | |
2389 | if (e < 0) | |
2390 | return e; | |
2391 | ||
2392 | n = strndup(in, p - in); | |
2393 | } else | |
2394 | n = strdup(in); | |
2395 | ||
2396 | if (!n) | |
2397 | return -ENOMEM; | |
2398 | ||
2399 | if (isempty(n)) | |
2400 | return -EINVAL; | |
2401 | ||
2402 | *error = e; | |
2403 | *name = TAKE_PTR(n); | |
2404 | ||
2405 | return 0; | |
2406 | } | |
4a4654e0 LP |
2407 | |
2408 | static int block_open_flag(scmp_filter_ctx seccomp, int flag) { | |
2409 | bool any = false; | |
2410 | int r; | |
2411 | ||
2412 | /* Blocks open() with the specified flag, where flag is O_SYNC or so. This makes these calls return | |
2413 | * EINVAL, in the hope the client code will retry without O_SYNC then. */ | |
2414 | ||
4a4654e0 LP |
2415 | r = seccomp_rule_add_exact( |
2416 | seccomp, | |
2417 | SCMP_ACT_ERRNO(EINVAL), | |
2418 | SCMP_SYS(open), | |
2419 | 1, | |
2420 | SCMP_A1(SCMP_CMP_MASKED_EQ, flag, flag)); | |
2421 | if (r < 0) | |
2422 | log_debug_errno(r, "Failed to add filter for open: %m"); | |
2423 | else | |
2424 | any = true; | |
4a4654e0 LP |
2425 | |
2426 | r = seccomp_rule_add_exact( | |
2427 | seccomp, | |
2428 | SCMP_ACT_ERRNO(EINVAL), | |
2429 | SCMP_SYS(openat), | |
2430 | 1, | |
2431 | SCMP_A2(SCMP_CMP_MASKED_EQ, flag, flag)); | |
2432 | if (r < 0) | |
2433 | log_debug_errno(r, "Failed to add filter for openat: %m"); | |
2434 | else | |
2435 | any = true; | |
2436 | ||
2437 | #if defined(__SNR_openat2) | |
2438 | /* The new openat2() system call can't be filtered sensibly, see above. */ | |
2439 | r = seccomp_rule_add_exact( | |
2440 | seccomp, | |
2441 | SCMP_ACT_ERRNO(ENOSYS), | |
2442 | SCMP_SYS(openat2), | |
2443 | 0); | |
2444 | if (r < 0) | |
2445 | log_debug_errno(r, "Failed to add filter for openat2: %m"); | |
2446 | else | |
2447 | any = true; | |
2448 | #endif | |
2449 | ||
2450 | return any ? 0 : r; | |
2451 | } | |
2452 | ||
2453 | int seccomp_suppress_sync(void) { | |
2454 | uint32_t arch; | |
2455 | int r; | |
2456 | ||
2457 | /* This is mostly identical to SystemCallFilter=~@sync:0, but simpler to use, and separately | |
2458 | * manageable, and also masks O_SYNC/O_DSYNC */ | |
2459 | ||
2460 | SECCOMP_FOREACH_LOCAL_ARCH(arch) { | |
2461 | _cleanup_(seccomp_releasep) scmp_filter_ctx seccomp = NULL; | |
4a4654e0 LP |
2462 | |
2463 | r = seccomp_init_for_arch(&seccomp, arch, SCMP_ACT_ALLOW); | |
2464 | if (r < 0) | |
2465 | return r; | |
2466 | ||
2467 | NULSTR_FOREACH(c, syscall_filter_sets[SYSCALL_FILTER_SET_SYNC].value) { | |
2468 | int id; | |
2469 | ||
2470 | id = seccomp_syscall_resolve_name(c); | |
2471 | if (id == __NR_SCMP_ERROR) { | |
2472 | log_debug("System call %s is not known, ignoring.", c); | |
2473 | continue; | |
2474 | } | |
2475 | ||
2476 | r = seccomp_rule_add_exact( | |
2477 | seccomp, | |
2478 | SCMP_ACT_ERRNO(0), /* success → we want this to be a NOP after all */ | |
2479 | id, | |
2480 | 0); | |
2481 | if (r < 0) | |
2482 | log_debug_errno(r, "Failed to add filter for system call %s, ignoring: %m", c); | |
2483 | } | |
2484 | ||
2485 | (void) block_open_flag(seccomp, O_SYNC); | |
2486 | #if O_DSYNC != O_SYNC | |
2487 | (void) block_open_flag(seccomp, O_DSYNC); | |
2488 | #endif | |
2489 | ||
2490 | r = seccomp_load(seccomp); | |
3c098014 ZJS |
2491 | if (ERRNO_IS_NEG_SECCOMP_FATAL(r)) |
2492 | return r; | |
2493 | if (r < 0) | |
2494 | log_debug_errno(r, "Failed to apply sync() suppression for architecture %s, skipping: %m", | |
2495 | seccomp_arch_to_string(arch)); | |
4a4654e0 LP |
2496 | } |
2497 | ||
2498 | return 0; | |
2499 | } |