]> git.ipfire.org Git - thirdparty/systemd.git/blame - src/nspawn/nspawn-oci.c
projects / thirdparty / systemd.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
tree-wide: use ASSERT_PTR more
[thirdparty/systemd.git] / src / nspawn / nspawn-oci.c
CommitLineData
db9ecf05 1/* SPDX-License-Identifier: LGPL-2.1-or-later */
de40a303 2
de40a303 3#include <linux/oom.h>
ae408d77 4#if HAVE_SECCOMP
de40a303 5#include <seccomp.h>
ae408d77 6#endif
de40a303
LP		 7
8#include "bus-util.h"
9#include "cap-list.h"
10#include "cpu-set-util.h"
ca822829 11#include "device-util.h"
7176f06c 12#include "devnum-util.h"
de40a303 13#include "env-util.h"
ca78ad1d 14#include "format-util.h"
de40a303
LP		 15#include "fs-util.h"
16#include "hostname-util.h"
17#include "json.h"
18#include "missing_sched.h"
19#include "nspawn-oci.h"
20#include "path-util.h"
21#include "rlimit-util.h"
ae408d77 22#if HAVE_SECCOMP
de40a303 23#include "seccomp-util.h"
ae408d77 24#endif
de40a303
LP		 25#include "stdio-util.h"
26#include "string-util.h"
27#include "strv.h"
28#include "user-util.h"
29
30/* TODO:
31 * OCI runtime tool implementation
32 * hooks
33 *
34 * Spec issues:
35 *
36 * How is RLIM_INFINITY supposed to be encoded?
37 * configured effective caps is bullshit, as execv() corrupts it anyway
38 * pipes bind mounted is *very* different from pipes newly created, comments regarding bind mount or not are bogus
39 * annotation values structured? or string?
40 * configurable file system namespace path, but then also root path? wtf?
41 * apply sysctl inside of the container? or outside?
42 * how is unlimited pids tasks limit to be encoded?
43 * what are the defaults for caps if not specified?
44 * what are the default uid/gid mappings if one is missing but the other set, or when user ns is on but no namespace configured
45 * the source field of "mounts" is really weird, as it cannot realistically be relative to the bundle, since we never know if that's what the fs wants
46 * spec contradicts itself on the mount "type" field, as the example uses "bind" as type, but it's not listed in /proc/filesystem, and is something made up by /bin/mount
47 * if type of mount is left out, what shall be assumed? "bind"?
48 * readonly mounts is entirely redundant?
49 * should escaping be applied when joining mount options with ","?
50 * devices cgroup support is bogus, "allow" and "deny" on the kernel level is about adding/removing entries, not about access
51 * spec needs to say that "rwm" devices cgroup combination can't be the empty string
ba669952 52 * cgrouspv1 crap: kernel, kernelTCP, swappiness, disableOOMKiller, swap, devices, leafWeight
de40a303
LP		 53 * general: it shouldn't leak lower level abstractions this obviously
54 * unmanagable cgroups stuff: realtimeRuntime/realtimePeriod
86b52a39 55 * needs to say what happense when some option is not specified, i.e. which defaults apply
de40a303
LP		 56 * no architecture? no personality?
57 * seccomp example and logic is simply broken: there's no constant "SCMP_ACT_ERRNO".
58 * spec should say what to do with unknown props
59 * /bin/mount regarding NFS and FUSE required?
60 * what does terminal=false mean?
6b000af4 61 * sysctl inside or outside? allow-listing?
5238e957 62 * swapiness typo -> swappiness
de40a303
LP		 63 *
64 * Unsupported:
65 *
66 * apparmorProfile
67 * selinuxLabel + mountLabel
68 * hugepageLimits
69 * network
70 * rdma
71 * intelRdt
72 * swappiness, disableOOMKiller, kernel, kernelTCP, leafWeight (because it's dead, cgroupsv2 can't do it and hence systemd neither)
73 *
74 * Non-slice cgroup paths
75 * Propagation that is not slave + shared
76 * more than one uid/gid mapping, mappings with a container base != 0, or non-matching uid/gid mappings
77 * device cgroups access = false items that are not catchall
78 * device cgroups matches where minor is specified, but major isn't. similar where major is specified but char/block is not. also, any match that only has a type set that has less than "rwm" set. also, any entry that has none of rwm set.
79 *
80 */
81
82static int oci_unexpected(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
19130626
ZJS		 83 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
84 "Unexpected OCI element '%s' of type '%s'.", name, json_variant_type_to_string(json_variant_type(v)));
de40a303
LP		 85}
86
87static int oci_unsupported(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
19130626
ZJS		 88 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
89 "Unsupported OCI element '%s' of type '%s'.", name, json_variant_type_to_string(json_variant_type(v)));
de40a303
LP		 90}
91
92static int oci_terminal(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
93 Settings *s = userdata;
94
95 /* If not specified, or set to true, we'll default to either an interactive or a read-only
6757a013 96 * console. If specified as false, we'll forcibly move to "pipe" mode though. */
de40a303
LP		 97 s->console_mode = json_variant_boolean(v) ? _CONSOLE_MODE_INVALID : CONSOLE_PIPE;
98 return 0;
99}
100
101static int oci_console_dimension(const char *name, JsonVariant *variant, JsonDispatchFlags flags, void *userdata) {
99534007 102 unsigned *u = ASSERT_PTR(userdata);
718ca772 103 uint64_t k;
de40a303 104
de40a303 105 k = json_variant_unsigned(variant);
19130626
ZJS		 106 if (k == 0)
107 return json_log(variant, flags, SYNTHETIC_ERRNO(ERANGE),
108 "Console size field '%s' is too small.", strna(name));
109 if (k > USHRT_MAX) /* TIOCSWINSZ's struct winsize uses "unsigned short" for width and height */
110 return json_log(variant, flags, SYNTHETIC_ERRNO(ERANGE),
111 "Console size field '%s' is too large.", strna(name));
de40a303
LP		 112
113 *u = (unsigned) k;
114 return 0;
115}
116
117static int oci_console_size(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
118
119 static const JsonDispatch table[] = {
120 { "height", JSON_VARIANT_UNSIGNED, oci_console_dimension, offsetof(Settings, console_height), JSON_MANDATORY },
121 { "width", JSON_VARIANT_UNSIGNED, oci_console_dimension, offsetof(Settings, console_width), JSON_MANDATORY },
122 {}
123 };
124
125 return json_dispatch(v, table, oci_unexpected, flags, userdata);
126}
127
128static int oci_absolute_path(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 129 char **p = ASSERT_PTR(userdata);
de40a303 130 const char *n;
de40a303 131
de40a303
LP		 132 n = json_variant_string(v);
133
19130626
ZJS		 134 if (!path_is_absolute(n))
135 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
136 "Path in JSON field '%s' is not absolute: %s", strna(name), n);
de40a303 137
f1531db5 138 return free_and_strdup_warn(p, n);
de40a303
LP		 139}
140
141static int oci_env(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 142 char ***l = ASSERT_PTR(userdata);
de40a303
LP		 143 JsonVariant *e;
144 int r;
145
de40a303
LP		 146 JSON_VARIANT_ARRAY_FOREACH(e, v) {
147 const char *n;
148
19130626
ZJS		 149 if (!json_variant_is_string(e))
150 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
151 "Environment array contains non-string.");
de40a303
LP		 152
153 assert_se(n = json_variant_string(e));
154
19130626
ZJS		 155 if (!env_assignment_is_valid(n))
156 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
157 "Environment assignment not valid: %s", n);
de40a303
LP		 158
159 r = strv_extend(l, n);
160 if (r < 0)
161 return log_oom();
162 }
163
164 return 0;
165}
166
167static int oci_args(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
168 _cleanup_strv_free_ char **l = NULL;
99534007 169 char ***value = ASSERT_PTR(userdata);
de40a303
LP		 170 int r;
171
a7f8c9ce
LP		 172 r = json_variant_strv(v, &l);
173 if (r < 0)
174 return json_log(v, flags, r, "Cannot parse arguments as list of strings: %m");
de40a303 175
19130626
ZJS		 176 if (strv_isempty(l))
177 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
178 "Argument list empty, refusing.");
de40a303 179
19130626
ZJS		 180 if (isempty(l[0]))
181 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
182 "Executable name is empty, refusing.");
de40a303 183
6757a013 184 return strv_free_and_replace(*value, l);
de40a303
LP		 185}
186
187static int oci_rlimit_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
188 const char *z;
189 int t, *type = userdata;
190
191 assert_se(type);
192
193 z = startswith(json_variant_string(v), "RLIMIT_");
19130626
ZJS		 194 if (!z)
195 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
196 "rlimit entry's name does not begin with 'RLIMIT_', refusing: %s",
197 json_variant_string(v));
de40a303
LP		 198
199 t = rlimit_from_string(z);
19130626 200 if (t < 0)
7211c853 201 return json_log(v, flags, t,
19130626 202 "rlimit name unknown: %s", json_variant_string(v));
de40a303
LP		 203
204 *type = t;
205 return 0;
206}
207
208static int oci_rlimit_value(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 209 rlim_t z, *value = ASSERT_PTR(userdata);
de40a303
LP		 210
211 if (json_variant_is_negative(v))
212 z = RLIM_INFINITY;
213 else {
19130626
ZJS		 214 if (!json_variant_is_unsigned(v))
215 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
216 "rlimits limit not unsigned, refusing.");
de40a303
LP		 217
218 z = (rlim_t) json_variant_unsigned(v);
219
718ca772 220 if ((uint64_t) z != json_variant_unsigned(v))
19130626
ZJS		 221 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
222 "rlimits limit out of range, refusing.");
de40a303
LP		 223 }
224
225 *value = z;
226 return 0;
227}
228
229static int oci_rlimits(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
230
99534007 231 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 232 JsonVariant *e;
233 int r;
234
de40a303
LP		 235 JSON_VARIANT_ARRAY_FOREACH(e, v) {
236
237 struct rlimit_data {
238 int type;
239 rlim_t soft;
240 rlim_t hard;
241 } data = {
242 .type = -1,
243 .soft = RLIM_INFINITY,
244 .hard = RLIM_INFINITY,
245 };
246
247 static const JsonDispatch table[] = {
248 { "soft", JSON_VARIANT_NUMBER, oci_rlimit_value, offsetof(struct rlimit_data, soft), JSON_MANDATORY },
249 { "hard", JSON_VARIANT_NUMBER, oci_rlimit_value, offsetof(struct rlimit_data, hard), JSON_MANDATORY },
250 { "type", JSON_VARIANT_STRING, oci_rlimit_type, offsetof(struct rlimit_data, type), JSON_MANDATORY },
251 {}
252 };
253
de40a303
LP		 254 r = json_dispatch(e, table, oci_unexpected, flags, &data);
255 if (r < 0)
256 return r;
257
258 assert(data.type >= 0);
259 assert(data.type < _RLIMIT_MAX);
260
19130626
ZJS		 261 if (s->rlimit[data.type])
262 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
263 "rlimits array contains duplicate entry, refusing.");
de40a303
LP		 264
265 s->rlimit[data.type] = new(struct rlimit, 1);
266 if (!s->rlimit[data.type])
267 return log_oom();
268
269 *s->rlimit[data.type] = (struct rlimit) {
270 .rlim_cur = data.soft,
271 .rlim_max = data.hard,
272 };
273
274 }
275 return 0;
276}
277
278static int oci_capability_array(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
279 uint64_t *mask = userdata, m = 0;
280 JsonVariant *e;
281
282 JSON_VARIANT_ARRAY_FOREACH(e, v) {
283 const char *n;
284 int cap;
285
19130626
ZJS		 286 if (!json_variant_is_string(e))
287 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
288 "Entry in capabilities array is not a string.");
de40a303
LP		 289
290 assert_se(n = json_variant_string(e));
291
292 cap = capability_from_name(n);
19130626
ZJS		 293 if (cap < 0)
294 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
295 "Unknown capability: %s", n);
de40a303
LP		 296
297 m |= UINT64_C(1) << cap;
298 }
299
f5fbe71d 300 if (*mask == UINT64_MAX)
de40a303
LP		 301 *mask = m;
302 else
303 *mask |= m;
304
305 return 0;
306}
307
308static int oci_capabilities(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
309
310 static const JsonDispatch table[] = {
311 { "effective", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, effective) },
312 { "bounding", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, bounding) },
313 { "inheritable", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, inheritable) },
314 { "permitted", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, permitted) },
315 { "ambient", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, ambient) },
316 {}
317 };
318
99534007 319 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 320 int r;
321
de40a303
LP		 322 r = json_dispatch(v, table, oci_unexpected, flags, &s->full_capabilities);
323 if (r < 0)
324 return r;
325
f5fbe71d 326 if (s->full_capabilities.bounding != UINT64_MAX) {
de40a303
LP		 327 s->capability = s->full_capabilities.bounding;
328 s->drop_capability = ~s->full_capabilities.bounding;
329 }
330
331 return 0;
332}
333
334static int oci_oom_score_adj(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 335 Settings *s = ASSERT_PTR(userdata);
718ca772 336 int64_t k;
de40a303 337
de40a303 338 k = json_variant_integer(v);
19130626
ZJS		 339 if (k < OOM_SCORE_ADJ_MIN || k > OOM_SCORE_ADJ_MAX)
340 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
c0f86d66 341 "oomScoreAdj value out of range: %" PRIi64, k);
de40a303
LP		 342
343 s->oom_score_adjust = (int) k;
344 s->oom_score_adjust_set = true;
345
346 return 0;
347}
348
349static int oci_uid_gid(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
350 uid_t *uid = userdata, u;
718ca772 351 uint64_t k;
de40a303
LP		 352
353 assert(uid);
354 assert_cc(sizeof(uid_t) == sizeof(gid_t));
355
356 k = json_variant_unsigned(v);
357 u = (uid_t) k;
718ca772 358 if ((uint64_t) u != k)
19130626 359 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
c0f86d66 360 "UID/GID out of range: %" PRIu64, k);
de40a303 361
19130626
ZJS		 362 if (!uid_is_valid(u))
363 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
364 "UID/GID is not valid: " UID_FMT, u);
de40a303
LP		 365
366 *uid = u;
367 return 0;
368}
369
370static int oci_supplementary_gids(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 371 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 372 JsonVariant *e;
373 int r;
374
de40a303
LP		 375 JSON_VARIANT_ARRAY_FOREACH(e, v) {
376 gid_t gid, *a;
377
19130626
ZJS		 378 if (!json_variant_is_unsigned(e))
379 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
380 "Supplementary GID entry is not a UID.");
de40a303
LP		 381
382 r = oci_uid_gid(name, e, flags, &gid);
383 if (r < 0)
384 return r;
385
386 a = reallocarray(s->supplementary_gids, s->n_supplementary_gids + 1, sizeof(gid_t));
387 if (!a)
388 return log_oom();
389
390 s->supplementary_gids = a;
391 s->supplementary_gids[s->n_supplementary_gids++] = gid;
392 }
393
394 return 0;
395}
396
397static int oci_user(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
398 static const JsonDispatch table[] = {
399 { "uid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(Settings, uid), JSON_MANDATORY },
400 { "gid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(Settings, gid), JSON_MANDATORY },
401 { "additionalGids", JSON_VARIANT_ARRAY, oci_supplementary_gids, 0, 0 },
402 {}
403 };
404
405 return json_dispatch(v, table, oci_unexpected, flags, userdata);
406}
407
408static int oci_process(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
409
410 static const JsonDispatch table[] = {
411 { "terminal", JSON_VARIANT_BOOLEAN, oci_terminal, 0, 0 },
412 { "consoleSize", JSON_VARIANT_OBJECT, oci_console_size, 0, 0 },
413 { "cwd", JSON_VARIANT_STRING, oci_absolute_path, offsetof(Settings, working_directory), 0 },
414 { "env", JSON_VARIANT_ARRAY, oci_env, offsetof(Settings, environment), 0 },
415 { "args", JSON_VARIANT_ARRAY, oci_args, offsetof(Settings, parameters), 0 },
416 { "rlimits", JSON_VARIANT_ARRAY, oci_rlimits, 0, 0 },
417 { "apparmorProfile", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
418 { "capabilities", JSON_VARIANT_OBJECT, oci_capabilities, 0, 0 },
419 { "noNewPrivileges", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(Settings, no_new_privileges), 0 },
420 { "oomScoreAdj", JSON_VARIANT_INTEGER, oci_oom_score_adj, 0, 0 },
421 { "selinuxLabel", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
422 { "user", JSON_VARIANT_OBJECT, oci_user, 0, 0 },
423 {}
424 };
425
426 return json_dispatch(v, table, oci_unexpected, flags, userdata);
427}
428
429static int oci_root(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
089cd8b3
AP		 430 Settings *s = userdata;
431 int r;
de40a303
LP		 432
433 static const JsonDispatch table[] = {
434 { "path", JSON_VARIANT_STRING, json_dispatch_string, offsetof(Settings, root) },
435 { "readonly", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(Settings, read_only) },
436 {}
437 };
438
089cd8b3
AP		 439 r = json_dispatch(v, table, oci_unexpected, flags, s);
440 if (r < 0)
441 return r;
442
443 if (s->root && !path_is_absolute(s->root)) {
444 char *joined;
445
446 joined = path_join(s->bundle, s->root);
447 if (!joined)
448 return log_oom();
449
450 free_and_replace(s->root, joined);
451 }
452
453 return 0;
de40a303
LP		 454}
455
456static int oci_hostname(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 457 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 458 const char *n;
459
de40a303
LP		 460 assert_se(n = json_variant_string(v));
461
52ef5dd7 462 if (!hostname_is_valid(n, 0))
19130626
ZJS		 463 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
464 "Hostname string is not a valid hostname: %s", n);
de40a303 465
f1531db5 466 return free_and_strdup_warn(&s->hostname, n);
de40a303
LP		 467}
468
469static bool oci_exclude_mount(const char *path) {
470
471 /* Returns "true" for all mounts we insist to mount on our own, and hence ignore the OCI data. */
472
473 if (PATH_IN_SET(path,
474 "/dev",
475 "/dev/mqueue",
476 "/dev/pts",
477 "/dev/shm",
478 "/proc",
479 "/proc/acpi",
480 "/proc/apm",
481 "/proc/asound",
482 "/proc/bus",
483 "/proc/fs",
484 "/proc/irq",
485 "/proc/kallsyms",
486 "/proc/kcore",
487 "/proc/keys",
488 "/proc/scsi",
489 "/proc/sys",
490 "/proc/sys/net",
491 "/proc/sysrq-trigger",
492 "/proc/timer_list",
493 "/run",
494 "/sys",
495 "/sys",
496 "/sys/fs/selinux",
497 "/tmp"))
498 return true;
499
500 /* Similar, skip the whole /sys/fs/cgroups subtree */
501 if (path_startswith(path, "/sys/fs/cgroup"))
502 return true;
503
504 return false;
505}
506
b2e07b1a
ZJS		 507typedef struct oci_mount_data {
508 char *destination;
509 char *source;
510 char *type;
511 char **options;
512} oci_mount_data;
513
514static void cleanup_oci_mount_data(oci_mount_data *data) {
515 free(data->destination);
516 free(data->source);
517 strv_free(data->options);
518 free(data->type);
519}
520
de40a303 521static int oci_mounts(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 522 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 523 JsonVariant *e;
524 int r;
525
de40a303 526 JSON_VARIANT_ARRAY_FOREACH(e, v) {
de40a303 527 static const JsonDispatch table[] = {
b2e07b1a
ZJS		 528 { "destination", JSON_VARIANT_STRING, oci_absolute_path, offsetof(oci_mount_data, destination), JSON_MANDATORY },
529 { "source", JSON_VARIANT_STRING, json_dispatch_string, offsetof(oci_mount_data, source), 0 },
530 { "options", JSON_VARIANT_ARRAY, json_dispatch_strv, offsetof(oci_mount_data, options), 0, },
531 { "type", JSON_VARIANT_STRING, json_dispatch_string, offsetof(oci_mount_data, type), 0 },
de40a303
LP		 532 {}
533 };
534
535 _cleanup_free_ char *joined_options = NULL;
536 CustomMount *m;
b2e07b1a 537 _cleanup_(cleanup_oci_mount_data) oci_mount_data data = {};
de40a303
LP		 538
539 r = json_dispatch(e, table, oci_unexpected, flags, &data);
540 if (r < 0)
b2e07b1a 541 return r;
de40a303 542
b2e07b1a
ZJS		 543 if (!path_is_absolute(data.destination))
544 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
545 "Mount destination not an absolute path: %s", data.destination);
de40a303
LP		 546
547 if (oci_exclude_mount(data.destination))
b2e07b1a 548 continue;
de40a303
LP		 549
550 if (data.options) {
551 joined_options = strv_join(data.options, ",");
b2e07b1a
ZJS		 552 if (!joined_options)
553 return log_oom();
de40a303
LP		 554 }
555
556 if (!data.type || streq(data.type, "bind")) {
b1f13b0e 557 if (data.source && !path_is_absolute(data.source)) {
de40a303
LP		 558 char *joined;
559
560 joined = path_join(s->bundle, data.source);
b2e07b1a
ZJS		 561 if (!joined)
562 return log_oom();
de40a303
LP		 563
564 free_and_replace(data.source, joined);
565 }
566
567 data.type = mfree(data.type);
568
569 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_BIND);
570 } else
571 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_ARBITRARY);
b2e07b1a
ZJS		 572 if (!m)
573 return log_oom();
de40a303
LP		 574
575 m->destination = TAKE_PTR(data.destination);
576 m->source = TAKE_PTR(data.source);
577 m->options = TAKE_PTR(joined_options);
578 m->type_argument = TAKE_PTR(data.type);
de40a303
LP		 579 }
580
581 return 0;
582}
583
584static int oci_namespace_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 585 unsigned long *nsflags = ASSERT_PTR(userdata);
de40a303
LP		 586 const char *n;
587
de40a303
LP		 588 assert_se(n = json_variant_string(v));
589
590 /* We don't use namespace_flags_from_string() here, as the OCI spec uses slightly different names than the
591 * kernel here. */
592 if (streq(n, "pid"))
593 *nsflags = CLONE_NEWPID;
594 else if (streq(n, "network"))
595 *nsflags = CLONE_NEWNET;
596 else if (streq(n, "mount"))
597 *nsflags = CLONE_NEWNS;
598 else if (streq(n, "ipc"))
599 *nsflags = CLONE_NEWIPC;
600 else if (streq(n, "uts"))
601 *nsflags = CLONE_NEWUTS;
602 else if (streq(n, "user"))
603 *nsflags = CLONE_NEWUSER;
604 else if (streq(n, "cgroup"))
605 *nsflags = CLONE_NEWCGROUP;
19130626
ZJS		 606 else
607 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
608 "Unknown cgroup type, refusing: %s", n);
de40a303
LP		 609
610 return 0;
611}
612
613static int oci_namespaces(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
614 Settings *s = userdata;
615 unsigned long n = 0;
616 JsonVariant *e;
617 int r;
618
619 assert_se(s);
620
621 JSON_VARIANT_ARRAY_FOREACH(e, v) {
622
623 struct namespace_data {
624 unsigned long type;
625 char *path;
626 } data = {};
627
628 static const JsonDispatch table[] = {
629 { "type", JSON_VARIANT_STRING, oci_namespace_type, offsetof(struct namespace_data, type), JSON_MANDATORY },
630 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(struct namespace_data, path), 0 },
631 {}
632 };
633
634 r = json_dispatch(e, table, oci_unexpected, flags, &data);
635 if (r < 0) {
636 free(data.path);
637 return r;
638 }
639
640 if (data.path) {
641 if (data.type != CLONE_NEWNET) {
de40a303 642 free(data.path);
19130626
ZJS		 643 return json_log(e, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
644 "Specifying namespace path for non-network namespace is not supported.");
de40a303
LP		 645 }
646
647 if (s->network_namespace_path) {
de40a303 648 free(data.path);
19130626
ZJS		 649 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
650 "Network namespace path specified more than once, refusing.");
de40a303
LP		 651 }
652
653 free(s->network_namespace_path);
654 s->network_namespace_path = data.path;
655 }
656
d7a0f1f4 657 if (FLAGS_SET(n, data.type))
19130626
ZJS		 658 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
659 "Duplicate namespace specification, refusing.");
de40a303
LP		 660
661 n |= data.type;
662 }
663
19130626
ZJS		 664 if (!FLAGS_SET(n, CLONE_NEWNS))
665 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
666 "Containers without file system namespace aren't supported.");
de40a303
LP		 667
668 s->private_network = FLAGS_SET(n, CLONE_NEWNET);
669 s->userns_mode = FLAGS_SET(n, CLONE_NEWUSER) ? USER_NAMESPACE_FIXED : USER_NAMESPACE_NO;
670 s->use_cgns = FLAGS_SET(n, CLONE_NEWCGROUP);
671
672 s->clone_ns_flags = n & (CLONE_NEWIPC|CLONE_NEWPID|CLONE_NEWUTS);
673
674 return 0;
675}
676
677static int oci_uid_gid_range(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
678 uid_t *uid = userdata, u;
718ca772 679 uint64_t k;
de40a303
LP		 680
681 assert(uid);
682 assert_cc(sizeof(uid_t) == sizeof(gid_t));
683
684 /* This is very much like oci_uid_gid(), except the checks are a bit different, as this is a UID range rather
f5fbe71d 685 * than a specific UID, and hence UID_INVALID has no special significance. OTOH a range of zero makes no
de40a303
LP		 686 * sense. */
687
688 k = json_variant_unsigned(v);
689 u = (uid_t) k;
718ca772 690 if ((uint64_t) u != k)
19130626 691 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
c0f86d66 692 "UID/GID out of range: %" PRIu64, k);
19130626
ZJS		 693 if (u == 0)
694 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
695 "UID/GID range can't be zero.");
de40a303
LP		 696
697 *uid = u;
698 return 0;
699}
700
701static int oci_uid_gid_mappings(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
702 struct mapping_data {
703 uid_t host_id;
704 uid_t container_id;
705 uid_t range;
706 } data = {
707 .host_id = UID_INVALID,
708 .container_id = UID_INVALID,
709 .range = 0,
710 };
711
712 static const JsonDispatch table[] = {
713 { "containerID", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(struct mapping_data, container_id), JSON_MANDATORY },
714 { "hostID", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(struct mapping_data, host_id), JSON_MANDATORY },
715 { "size", JSON_VARIANT_UNSIGNED, oci_uid_gid_range, offsetof(struct mapping_data, range), JSON_MANDATORY },
716 {}
717 };
718
99534007 719 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 720 JsonVariant *e;
721 int r;
722
de40a303
LP		 723 if (json_variant_elements(v) == 0)
724 return 0;
725
19130626
ZJS		 726 if (json_variant_elements(v) > 1)
727 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
728 "UID/GID mappings with more than one entry are not supported.");
de40a303
LP		 729
730 assert_se(e = json_variant_by_index(v, 0));
731
732 r = json_dispatch(e, table, oci_unexpected, flags, &data);
733 if (r < 0)
734 return r;
735
736 if (data.host_id + data.range < data.host_id ||
19130626
ZJS		 737 data.container_id + data.range < data.container_id)
738 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
739 "UID/GID range goes beyond UID/GID validity range, refusing.");
de40a303 740
19130626
ZJS		 741 if (data.container_id != 0)
742 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
743 "UID/GID mappings with a non-zero container base are not supported.");
de40a303
LP		 744
745 if (data.range < 0x10000)
19130626
ZJS		 746 json_log(v, flags|JSON_WARNING, 0,
747 "UID/GID mapping with less than 65536 UID/GIDS set up, you are looking for trouble.");
de40a303
LP		 748
749 if (s->uid_range != UID_INVALID &&
19130626
ZJS		 750 (s->uid_shift != data.host_id || s->uid_range != data.range))
751 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
752 "Non-matching UID and GID mappings are not supported.");
de40a303
LP		 753
754 s->uid_shift = data.host_id;
755 s->uid_range = data.range;
756
757 return 0;
758}
759
760static int oci_device_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 761 mode_t *mode = ASSERT_PTR(userdata);
de40a303
LP		 762 const char *t;
763
de40a303
LP		 764 assert_se(t = json_variant_string(v));
765
766 if (STR_IN_SET(t, "c", "u"))
767 *mode = (*mode & ~S_IFMT) | S_IFCHR;
768 else if (streq(t, "b"))
769 *mode = (*mode & ~S_IFMT) | S_IFBLK;
770 else if (streq(t, "p"))
771 *mode = (*mode & ~S_IFMT) | S_IFIFO;
19130626
ZJS		 772 else
773 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
774 "Unknown device type: %s", t);
de40a303
LP		 775
776 return 0;
777}
778
779static int oci_device_major(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
780 unsigned *u = userdata;
718ca772 781 uint64_t k;
de40a303
LP		 782
783 assert_se(u);
784
785 k = json_variant_unsigned(v);
19130626
ZJS		 786 if (!DEVICE_MAJOR_VALID(k))
787 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
c0f86d66 788 "Device major %" PRIu64 " out of range.", k);
de40a303
LP		 789
790 *u = (unsigned) k;
791 return 0;
792}
793
794static int oci_device_minor(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
795 unsigned *u = userdata;
718ca772 796 uint64_t k;
de40a303
LP		 797
798 assert_se(u);
799
800 k = json_variant_unsigned(v);
19130626
ZJS		 801 if (!DEVICE_MINOR_VALID(k))
802 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
c0f86d66 803 "Device minor %" PRIu64 " out of range.", k);
de40a303
LP		 804
805 *u = (unsigned) k;
806 return 0;
807}
808
809static int oci_device_file_mode(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
810 mode_t *mode = userdata, m;
718ca772 811 uint64_t k;
de40a303
LP		 812
813 assert(mode);
814
815 k = json_variant_unsigned(v);
816 m = (mode_t) k;
817
718ca772 818 if ((m & ~07777) != 0 || (uint64_t) m != k)
19130626
ZJS		 819 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
820 "fileMode out of range, refusing.");
de40a303
LP		 821
822 *mode = m;
823 return 0;
824}
825
826static int oci_devices(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 827 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 828 JsonVariant *e;
829 int r;
830
de40a303
LP		 831 JSON_VARIANT_ARRAY_FOREACH(e, v) {
832
833 static const JsonDispatch table[] = {
834 { "type", JSON_VARIANT_STRING, oci_device_type, offsetof(DeviceNode, mode), JSON_MANDATORY },
835 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(DeviceNode, path), JSON_MANDATORY },
836 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(DeviceNode, major), 0 },
837 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(DeviceNode, minor), 0 },
838 { "fileMode", JSON_VARIANT_UNSIGNED, oci_device_file_mode, offsetof(DeviceNode, mode), 0 },
839 { "uid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(DeviceNode, uid), 0 },
840 { "gid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(DeviceNode, gid), 0 },
841 {}
842 };
843
844 DeviceNode *node, *nodes;
845
846 nodes = reallocarray(s->extra_nodes, s->n_extra_nodes + 1, sizeof(DeviceNode));
847 if (!nodes)
848 return log_oom();
849
850 s->extra_nodes = nodes;
851
852 node = nodes + s->n_extra_nodes;
853 *node = (DeviceNode) {
854 .uid = UID_INVALID,
855 .gid = GID_INVALID,
f5fbe71d
YW		 856 .major = UINT_MAX,
857 .minor = UINT_MAX,
de40a303
LP		 858 .mode = 0644,
859 };
860
861 r = json_dispatch(e, table, oci_unexpected, flags, node);
862 if (r < 0)
863 goto fail_element;
864
865 if (S_ISCHR(node->mode) || S_ISBLK(node->mode)) {
866 _cleanup_free_ char *path = NULL;
867
f5fbe71d 868 if (node->major == UINT_MAX || node->minor == UINT_MAX) {
19130626
ZJS		 869 r = json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
870 "Major/minor required when device node is device node");
de40a303
LP		 871 goto fail_element;
872 }
873
874 /* Suppress a couple of implicit device nodes */
4fe46c34 875 r = devname_from_devnum(node->mode, makedev(node->major, node->minor), &path);
de40a303
LP		 876 if (r < 0)
877 json_log(e, flags|JSON_DEBUG, 0, "Failed to resolve device node %u:%u, ignoring: %m", node->major, node->minor);
878 else {
879 if (PATH_IN_SET(path,
880 "/dev/null",
881 "/dev/zero",
882 "/dev/full",
883 "/dev/random",
884 "/dev/urandom",
885 "/dev/tty",
886 "/dev/net/tun",
887 "/dev/ptmx",
888 "/dev/pts/ptmx",
889 "/dev/console")) {
890
891 json_log(e, flags|JSON_DEBUG, 0, "Ignoring devices item for device '%s', as it is implicitly created anyway.", path);
892 free(node->path);
893 continue;
894 }
895 }
896 }
897
898 s->n_extra_nodes++;
899 continue;
900
901 fail_element:
902 free(node->path);
903 return r;
904 }
905
906 return 0;
907}
908
909static int oci_cgroups_path(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
910 _cleanup_free_ char *slice = NULL, *backwards = NULL;
99534007 911 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 912 const char *p;
913 int r;
914
de40a303
LP		 915 assert_se(p = json_variant_string(v));
916
917 r = cg_path_get_slice(p, &slice);
918 if (r < 0)
919 return json_log(v, flags, r, "Couldn't derive slice unit name from path '%s': %m", p);
920
921 r = cg_slice_to_path(slice, &backwards);
922 if (r < 0)
923 return json_log(v, flags, r, "Couldn't convert slice unit name '%s' back to path: %m", slice);
924
19130626
ZJS		 925 if (!path_equal(backwards, p))
926 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
927 "Control group path '%s' does not refer to slice unit, refusing.", p);
de40a303
LP		 928
929 free_and_replace(s->slice, slice);
930 return 0;
931}
932
933static int oci_cgroup_device_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
934 mode_t *mode = userdata;
935 const char *n;
936
937 assert_se(n = json_variant_string(v));
938
939 if (streq(n, "c"))
940 *mode = S_IFCHR;
941 else if (streq(n, "b"))
942 *mode = S_IFBLK;
19130626
ZJS		 943 else
944 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
945 "Control group device type unknown: %s", n);
de40a303
LP		 946
947 return 0;
948}
949
950struct device_data {
951 bool allow;
952 bool r;
953 bool w;
954 bool m;
955 mode_t type;
956 unsigned major;
957 unsigned minor;
958};
959
960static int oci_cgroup_device_access(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
961 struct device_data *d = userdata;
962 bool r = false, w = false, m = false;
963 const char *s;
964 size_t i;
965
966 assert_se(s = json_variant_string(v));
967
19130626 968 for (i = 0; s[i]; i++)
de40a303
LP		 969 if (s[i] == 'r')
970 r = true;
971 else if (s[i] == 'w')
972 w = true;
973 else if (s[i] == 'm')
974 m = true;
19130626
ZJS		 975 else
976 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
977 "Unknown device access character '%c'.", s[i]);
de40a303
LP		 978
979 d->r = r;
980 d->w = w;
981 d->m = m;
982
983 return 0;
984}
985
986static int oci_cgroup_devices(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
987
988 _cleanup_free_ struct device_data *list = NULL;
99534007 989 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 990 size_t n_list = 0, i;
991 bool noop = false;
992 JsonVariant *e;
993 int r;
994
de40a303
LP		 995 JSON_VARIANT_ARRAY_FOREACH(e, v) {
996
997 struct device_data data = {
f5fbe71d
YW		 998 .major = UINT_MAX,
999 .minor = UINT_MAX,
de40a303
LP		 1000 }, *a;
1001
1002 static const JsonDispatch table[] = {
1003 { "allow", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(struct device_data, allow), JSON_MANDATORY },
1004 { "type", JSON_VARIANT_STRING, oci_cgroup_device_type, offsetof(struct device_data, type), 0 },
1005 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), 0 },
1006 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), 0 },
1007 { "access", JSON_VARIANT_STRING, oci_cgroup_device_access, 0, 0 },
1008 {}
1009 };
1010
1011 r = json_dispatch(e, table, oci_unexpected, flags, &data);
1012 if (r < 0)
1013 return r;
1014
1015 if (!data.allow) {
6b000af4
LP		 1016 /* The fact that OCI allows 'deny' entries makes really no sense, as 'allow'
1017 * vs. 'deny' for the devices cgroup controller is really not about allow-listing and
1018 * deny-listing but about adding and removing entries from the allow list. Since we
1019 * always start out with an empty allow list we hence ignore the whole thing, as
1020 * removing entries which don't exist make no sense. We'll log about this, since this
1021 * is really borked in the spec, with one exception: the entry that's supposed to
1022 * drop the kernel's default we ignore silently */
de40a303 1023
f5fbe71d 1024 if (!data.r || !data.w || !data.m || data.type != 0 || data.major != UINT_MAX || data.minor != UINT_MAX)
6b000af4 1025 json_log(v, flags|JSON_WARNING, 0, "Devices cgroup allow list with arbitrary 'allow' entries not supported, ignoring.");
de40a303
LP		 1026
1027 /* We ignore the 'deny' entry as for us that's implied */
1028 continue;
1029 }
1030
1031 if (!data.r && !data.w && !data.m) {
6b000af4 1032 json_log(v, flags|LOG_WARNING, 0, "Device cgroup allow list entry with no effect found, ignoring.");
de40a303
LP		 1033 continue;
1034 }
1035
f5fbe71d 1036 if (data.minor != UINT_MAX && data.major == UINT_MAX)
19130626 1037 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
6b000af4 1038 "Device cgroup allow list entries with minors but no majors not supported.");
de40a303 1039
f5fbe71d 1040 if (data.major != UINT_MAX && data.type == 0)
19130626 1041 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
6b000af4 1042 "Device cgroup allow list entries with majors but no device node type not supported.");
de40a303
LP		 1043
1044 if (data.type == 0) {
6b000af4 1045 if (data.r && data.w && data.m) /* a catchall allow list entry means we are looking at a noop */
de40a303 1046 noop = true;
19130626
ZJS		 1047 else
1048 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
6b000af4 1049 "Device cgroup allow list entries with no type not supported.");
de40a303
LP		 1050 }
1051
1052 a = reallocarray(list, n_list + 1, sizeof(struct device_data));
1053 if (!a)
1054 return log_oom();
1055
1056 list = a;
1057 list[n_list++] = data;
1058 }
1059
1060 if (noop)
1061 return 0;
1062
1063 r = settings_allocate_properties(s);
1064 if (r < 0)
1065 return r;
1066
1067 r = sd_bus_message_open_container(s->properties, 'r', "sv");
1068 if (r < 0)
1069 return bus_log_create_error(r);
1070
1071 r = sd_bus_message_append(s->properties, "s", "DeviceAllow");
1072 if (r < 0)
1073 return bus_log_create_error(r);
1074
1075 r = sd_bus_message_open_container(s->properties, 'v', "a(ss)");
1076 if (r < 0)
1077 return bus_log_create_error(r);
1078
1079 r = sd_bus_message_open_container(s->properties, 'a', "(ss)");
1080 if (r < 0)
1081 return bus_log_create_error(r);
1082
1083 for (i = 0; i < n_list; i++) {
1084 _cleanup_free_ char *pattern = NULL;
1085 char access[4];
1086 size_t n = 0;
1087
f5fbe71d 1088 if (list[i].minor == UINT_MAX) {
de40a303
LP		 1089 const char *t;
1090
1091 if (list[i].type == S_IFBLK)
1092 t = "block";
1093 else {
1094 assert(list[i].type == S_IFCHR);
1095 t = "char";
1096 }
1097
f5fbe71d 1098 if (list[i].major == UINT_MAX) {
de40a303
LP		 1099 pattern = strjoin(t, "-*");
1100 if (!pattern)
1101 return log_oom();
1102 } else {
1103 if (asprintf(&pattern, "%s-%u", t, list[i].major) < 0)
1104 return log_oom();
1105 }
1106
1107 } else {
f5fbe71d 1108 assert(list[i].major != UINT_MAX); /* If a minor is specified, then a major also needs to be specified */
de40a303
LP		 1109
1110 r = device_path_make_major_minor(list[i].type, makedev(list[i].major, list[i].minor), &pattern);
1111 if (r < 0)
1112 return log_oom();
1113 }
1114
1115 if (list[i].r)
1116 access[n++] = 'r';
1117 if (list[i].w)
1118 access[n++] = 'w';
1119 if (list[i].m)
1120 access[n++] = 'm';
1121 access[n] = 0;
1122
1123 assert(n > 0);
1124
1125 r = sd_bus_message_append(s->properties, "(ss)", pattern, access);
1126 if (r < 0)
1127 return bus_log_create_error(r);
1128 }
1129
1130 r = sd_bus_message_close_container(s->properties);
1131 if (r < 0)
1132 return bus_log_create_error(r);
1133
1134 r = sd_bus_message_close_container(s->properties);
1135 if (r < 0)
1136 return bus_log_create_error(r);
1137
1138 r = sd_bus_message_close_container(s->properties);
1139 if (r < 0)
1140 return bus_log_create_error(r);
1141
1142 return 0;
1143}
1144
1145static int oci_cgroup_memory_limit(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1146 uint64_t *m = ASSERT_PTR(userdata);
718ca772 1147 uint64_t k;
de40a303 1148
de40a303
LP		 1149 if (json_variant_is_negative(v)) {
1150 *m = UINT64_MAX;
1151 return 0;
1152 }
1153
19130626
ZJS		 1154 if (!json_variant_is_unsigned(v))
1155 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1156 "Memory limit is not an unsigned integer");
de40a303
LP		 1157
1158 k = json_variant_unsigned(v);
19130626
ZJS		 1159 if (k >= UINT64_MAX)
1160 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
c0f86d66 1161 "Memory limit too large: %" PRIu64, k);
de40a303
LP		 1162
1163 *m = (uint64_t) k;
1164 return 0;
1165}
1166
1167static int oci_cgroup_memory(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1168
1169 struct memory_data {
1170 uint64_t limit;
1171 uint64_t reservation;
1172 uint64_t swap;
1173 } data = {
1174 .limit = UINT64_MAX,
1175 .reservation = UINT64_MAX,
1176 .swap = UINT64_MAX,
1177 };
1178
1179 static const JsonDispatch table[] = {
1180 { "limit", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, limit), 0 },
1181 { "reservation", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, reservation), 0 },
1182 { "swap", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, swap), 0 },
1183 { "kernel", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1184 { "kernelTCP", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1185 { "swapiness", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1186 { "disableOOMKiller", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1187 {}
1188 };
1189
1190 Settings *s = userdata;
1191 int r;
1192
1193 r = json_dispatch(v, table, oci_unexpected, flags, &data);
1194 if (r < 0)
1195 return r;
1196
1197 if (data.swap != UINT64_MAX) {
1198 if (data.limit == UINT64_MAX)
1199 json_log(v, flags|LOG_WARNING, 0, "swap limit without memory limit is not supported, ignoring.");
1200 else if (data.swap < data.limit)
1201 json_log(v, flags|LOG_WARNING, 0, "swap limit is below memory limit, ignoring.");
1202 else {
1203 r = settings_allocate_properties(s);
1204 if (r < 0)
1205 return r;
1206
1207 r = sd_bus_message_append(s->properties, "(sv)", "MemorySwapMax", "t", data.swap - data.limit);
1208 if (r < 0)
1209 return bus_log_create_error(r);
1210 }
1211 }
1212
1213 if (data.limit != UINT64_MAX) {
1214 r = settings_allocate_properties(s);
1215 if (r < 0)
1216 return r;
1217
1218 r = sd_bus_message_append(s->properties, "(sv)", "MemoryMax", "t", data.limit);
1219 if (r < 0)
1220 return bus_log_create_error(r);
1221 }
1222
1223 if (data.reservation != UINT64_MAX) {
1224 r = settings_allocate_properties(s);
1225 if (r < 0)
1226 return r;
1227
1228 r = sd_bus_message_append(s->properties, "(sv)", "MemoryLow", "t", data.reservation);
1229 if (r < 0)
1230 return bus_log_create_error(r);
1231 }
1232
1233 return 0;
1234}
1235
1236struct cpu_data {
1237 uint64_t shares;
1238 uint64_t quota;
1239 uint64_t period;
0985c7c4 1240 CPUSet cpu_set;
de40a303
LP		 1241};
1242
1243static int oci_cgroup_cpu_shares(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1244 uint64_t *u = ASSERT_PTR(userdata);
718ca772 1245 uint64_t k;
de40a303 1246
de40a303 1247 k = json_variant_unsigned(v);
19130626
ZJS		 1248 if (k < CGROUP_CPU_SHARES_MIN || k > CGROUP_CPU_SHARES_MAX)
1249 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1250 "shares value out of range.");
de40a303
LP		 1251
1252 *u = (uint64_t) k;
1253 return 0;
1254}
1255
1256static int oci_cgroup_cpu_quota(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1257 uint64_t *u = ASSERT_PTR(userdata);
718ca772 1258 uint64_t k;
de40a303 1259
de40a303 1260 k = json_variant_unsigned(v);
19130626
ZJS		 1261 if (k <= 0 || k >= UINT64_MAX)
1262 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1263 "period/quota value out of range.");
de40a303
LP		 1264
1265 *u = (uint64_t) k;
1266 return 0;
1267}
1268
1269static int oci_cgroup_cpu_cpus(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1270 struct cpu_data *data = ASSERT_PTR(userdata);
0985c7c4 1271 CPUSet set;
de40a303 1272 const char *n;
0985c7c4 1273 int r;
de40a303 1274
de40a303
LP		 1275 assert_se(n = json_variant_string(v));
1276
0985c7c4
ZJS		 1277 r = parse_cpu_set(n, &set);
1278 if (r < 0)
1279 return json_log(v, flags, r, "Failed to parse CPU set specification: %s", n);
de40a303 1280
0985c7c4
ZJS		 1281 cpu_set_reset(&data->cpu_set);
1282 data->cpu_set = set;
de40a303
LP		 1283
1284 return 0;
1285}
1286
1287static int oci_cgroup_cpu(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1288
1289 static const JsonDispatch table[] = {
1290 { "shares", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_shares, offsetof(struct cpu_data, shares), 0 },
1291 { "quota", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_quota, offsetof(struct cpu_data, quota), 0 },
1292 { "period", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_quota, offsetof(struct cpu_data, period), 0 },
1293 { "realtimeRuntime", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, 0 },
1294 { "realtimePeriod", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, 0 },
1295 { "cpus", JSON_VARIANT_STRING, oci_cgroup_cpu_cpus, 0, 0 },
1296 { "mems", JSON_VARIANT_STRING, oci_unsupported, 0, 0 },
1297 {}
1298 };
1299
1300 struct cpu_data data = {
1301 .shares = UINT64_MAX,
1302 .quota = UINT64_MAX,
1303 .period = UINT64_MAX,
1304 };
1305
1306 Settings *s = userdata;
1307 int r;
1308
1309 r = json_dispatch(v, table, oci_unexpected, flags, &data);
1310 if (r < 0) {
0985c7c4 1311 cpu_set_reset(&data.cpu_set);
de40a303
LP		 1312 return r;
1313 }
1314
0985c7c4
ZJS		 1315 cpu_set_reset(&s->cpu_set);
1316 s->cpu_set = data.cpu_set;
de40a303
LP		 1317
1318 if (data.shares != UINT64_MAX) {
1319 r = settings_allocate_properties(s);
1320 if (r < 0)
1321 return r;
1322
1323 r = sd_bus_message_append(s->properties, "(sv)", "CPUShares", "t", data.shares);
1324 if (r < 0)
1325 return bus_log_create_error(r);
1326 }
1327
1328 if (data.quota != UINT64_MAX && data.period != UINT64_MAX) {
1329 r = settings_allocate_properties(s);
1330 if (r < 0)
1331 return r;
1332
1333 r = sd_bus_message_append(s->properties, "(sv)", "CPUQuotaPerSecUSec", "t", (uint64_t) (data.quota * USEC_PER_SEC / data.period));
1334 if (r < 0)
1335 return bus_log_create_error(r);
1336
19130626
ZJS		 1337 } else if ((data.quota != UINT64_MAX) != (data.period != UINT64_MAX))
1338 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1339 "CPU quota and period not used together.");
de40a303
LP		 1340
1341 return 0;
1342}
1343
1344static int oci_cgroup_block_io_weight(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1345 Settings *s = ASSERT_PTR(userdata);
718ca772 1346 uint64_t k;
de40a303
LP		 1347 int r;
1348
de40a303 1349 k = json_variant_unsigned(v);
19130626
ZJS		 1350 if (k < CGROUP_BLKIO_WEIGHT_MIN || k > CGROUP_BLKIO_WEIGHT_MAX)
1351 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1352 "Block I/O weight out of range.");
de40a303
LP		 1353
1354 r = settings_allocate_properties(s);
1355 if (r < 0)
1356 return r;
1357
1358 r = sd_bus_message_append(s->properties, "(sv)", "BlockIOWeight", "t", (uint64_t) k);
1359 if (r < 0)
1360 return bus_log_create_error(r);
1361
1362 return 0;
1363}
1364
1365static int oci_cgroup_block_io_weight_device(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1366 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 1367 JsonVariant *e;
1368 int r;
1369
de40a303
LP		 1370 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1371 struct device_data {
1372 unsigned major;
1373 unsigned minor;
718ca772 1374 uint64_t weight;
de40a303 1375 } data = {
f5fbe71d
YW		 1376 .major = UINT_MAX,
1377 .minor = UINT_MAX,
718ca772 1378 .weight = UINT64_MAX,
de40a303
LP		 1379 };
1380
1381 static const JsonDispatch table[] = {
718ca772
ZJS		 1382 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), JSON_MANDATORY },
1383 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), JSON_MANDATORY },
1384 { "weight", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct device_data, weight), 0 },
1385 { "leafWeight", JSON_VARIANT_INTEGER, oci_unsupported, 0, JSON_PERMISSIVE },
de40a303
LP		 1386 {}
1387 };
1388
1389 _cleanup_free_ char *path = NULL;
1390
1391 r = json_dispatch(e, table, oci_unexpected, flags, &data);
1392 if (r < 0)
1393 return r;
1394
718ca772 1395 if (data.weight == UINT64_MAX)
de40a303
LP		 1396 continue;
1397
19130626
ZJS		 1398 if (data.weight < CGROUP_BLKIO_WEIGHT_MIN || data.weight > CGROUP_BLKIO_WEIGHT_MAX)
1399 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1400 "Block I/O device weight out of range.");
de40a303
LP		 1401
1402 r = device_path_make_major_minor(S_IFBLK, makedev(data.major, data.minor), &path);
1403 if (r < 0)
1404 return json_log(v, flags, r, "Failed to build device path: %m");
1405
1406 r = settings_allocate_properties(s);
1407 if (r < 0)
1408 return r;
1409
1410 r = sd_bus_message_append(s->properties, "(sv)", "BlockIODeviceWeight", "a(st)", 1, path, (uint64_t) data.weight);
1411 if (r < 0)
1412 return bus_log_create_error(r);
1413 }
1414
1415 return 0;
1416}
1417
1418static int oci_cgroup_block_io_throttle(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1419 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 1420 const char *pname;
1421 JsonVariant *e;
1422 int r;
1423
de40a303
LP		 1424 pname = streq(name, "throttleReadBpsDevice") ? "IOReadBandwidthMax" :
1425 streq(name, "throttleWriteBpsDevice") ? "IOWriteBandwidthMax" :
1426 streq(name, "throttleReadIOPSDevice") ? "IOReadIOPSMax" :
1427 "IOWriteIOPSMax";
1428
1429 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1430 struct device_data {
1431 unsigned major;
1432 unsigned minor;
718ca772 1433 uint64_t rate;
de40a303 1434 } data = {
f5fbe71d
YW		 1435 .major = UINT_MAX,
1436 .minor = UINT_MAX,
de40a303
LP		 1437 };
1438
1439 static const JsonDispatch table[] = {
718ca772
ZJS		 1440 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), JSON_MANDATORY },
1441 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), JSON_MANDATORY },
1442 { "rate", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct device_data, rate), JSON_MANDATORY },
de40a303
LP		 1443 {}
1444 };
1445
1446 _cleanup_free_ char *path = NULL;
1447
1448 r = json_dispatch(e, table, oci_unexpected, flags, &data);
1449 if (r < 0)
1450 return r;
1451
19130626
ZJS		 1452 if (data.rate >= UINT64_MAX)
1453 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1454 "Block I/O device rate out of range.");
de40a303
LP		 1455
1456 r = device_path_make_major_minor(S_IFBLK, makedev(data.major, data.minor), &path);
1457 if (r < 0)
1458 return json_log(v, flags, r, "Failed to build device path: %m");
1459
1460 r = settings_allocate_properties(s);
1461 if (r < 0)
1462 return r;
1463
1464 r = sd_bus_message_append(s->properties, "(sv)", pname, "a(st)", 1, path, (uint64_t) data.rate);
1465 if (r < 0)
1466 return bus_log_create_error(r);
1467 }
1468
1469 return 0;
1470}
1471
1472static int oci_cgroup_block_io(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1473
1474 static const JsonDispatch table[] = {
1475 { "weight", JSON_VARIANT_UNSIGNED, oci_cgroup_block_io_weight, 0, 0 },
1476 { "leafWeight", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, JSON_PERMISSIVE },
1477 { "weightDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_weight_device, 0, 0 },
1478 { "throttleReadBpsDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1479 { "throttleWriteBpsDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1480 { "throttleReadIOPSDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1481 { "throttleWriteIOPSDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1482 {}
1483 };
1484
1485 return json_dispatch(v, table, oci_unexpected, flags, userdata);
1486}
1487
1488static int oci_cgroup_pids(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1489
1490 static const JsonDispatch table[] = {
1491 { "limit", JSON_VARIANT_NUMBER, json_dispatch_variant, 0, JSON_MANDATORY },
1492 {}
1493 };
1494
1495 _cleanup_(json_variant_unrefp) JsonVariant *k = NULL;
99534007 1496 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 1497 uint64_t m;
1498 int r;
1499
de40a303
LP		 1500 r = json_dispatch(v, table, oci_unexpected, flags, &k);
1501 if (r < 0)
1502 return r;
1503
1504 if (json_variant_is_negative(k))
1505 m = UINT64_MAX;
1506 else {
19130626
ZJS		 1507 if (!json_variant_is_unsigned(k))
1508 return json_log(k, flags, SYNTHETIC_ERRNO(EINVAL),
1509 "pids limit not unsigned integer, refusing.");
de40a303
LP		 1510
1511 m = (uint64_t) json_variant_unsigned(k);
1512
718ca772 1513 if ((uint64_t) m != json_variant_unsigned(k))
19130626
ZJS		 1514 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1515 "pids limit out of range, refusing.");
de40a303
LP		 1516 }
1517
1518 r = settings_allocate_properties(s);
1519 if (r < 0)
1520 return r;
1521
1522 r = sd_bus_message_append(s->properties, "(sv)", "TasksMax", "t", m);
1523 if (r < 0)
1524 return bus_log_create_error(r);
1525
1526 return 0;
1527}
1528
1529static int oci_resources(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1530
1531 static const JsonDispatch table[] = {
1532 { "devices", JSON_VARIANT_ARRAY, oci_cgroup_devices, 0, 0 },
1533 { "memory", JSON_VARIANT_OBJECT, oci_cgroup_memory, 0, 0 },
1534 { "cpu", JSON_VARIANT_OBJECT, oci_cgroup_cpu, 0, 0 },
1535 { "blockIO", JSON_VARIANT_OBJECT, oci_cgroup_block_io, 0, 0 },
1536 { "hugepageLimits", JSON_VARIANT_ARRAY, oci_unsupported, 0, 0 },
1537 { "network", JSON_VARIANT_OBJECT, oci_unsupported, 0, 0 },
1538 { "pids", JSON_VARIANT_OBJECT, oci_cgroup_pids, 0, 0 },
1539 { "rdma", JSON_VARIANT_OBJECT, oci_unsupported, 0, 0 },
1540 {}
1541 };
1542
1543 return json_dispatch(v, table, oci_unexpected, flags, userdata);
1544}
1545
1546static bool sysctl_key_valid(const char *s) {
1547 bool dot = true;
1548
1549 /* Note that we are a bit stricter here than in systemd-sysctl, as that inherited semantics from the old sysctl
1550 * tool, which were really weird (as it swaps / and . in both ways) */
1551
1552 if (isempty(s))
1553 return false;
1554
1555 for (; *s; s++) {
1556
1557 if (*s <= ' ' || *s >= 127)
1558 return false;
1559 if (*s == '/')
1560 return false;
1561 if (*s == '.') {
1562
1563 if (dot) /* Don't allow two dots next to each other (or at the beginning) */
1564 return false;
1565
1566 dot = true;
1567 } else
1568 dot = false;
1569 }
1570
1571 if (dot) /* don't allow a dot at the end */
1572 return false;
1573
1574 return true;
1575}
1576
1577static int oci_sysctl(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1578 Settings *s = ASSERT_PTR(userdata);
33d60b8d
LP		 1579 JsonVariant *w;
1580 const char *k;
de40a303
LP		 1581 int r;
1582
de40a303 1583 JSON_VARIANT_OBJECT_FOREACH(k, w, v) {
33d60b8d 1584 const char *m;
de40a303 1585
19130626
ZJS		 1586 if (!json_variant_is_string(w))
1587 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1588 "sysctl parameter is not a string, refusing.");
de40a303 1589
de40a303
LP		 1590 assert_se(m = json_variant_string(w));
1591
33d60b8d 1592 if (sysctl_key_valid(k))
19130626 1593 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
33d60b8d 1594 "sysctl key invalid, refusing: %s", k);
de40a303 1595
33d60b8d 1596 r = strv_extend_strv(&s->sysctl, STRV_MAKE(k, m), false);
de40a303
LP		 1597 if (r < 0)
1598 return log_oom();
1599 }
1600
1601 return 0;
1602}
1603
ae408d77 1604#if HAVE_SECCOMP
de40a303
LP		 1605static int oci_seccomp_action_from_string(const char *name, uint32_t *ret) {
1606
1607 static const struct {
1608 const char *name;
1609 uint32_t action;
1610 } table[] = {
f9a3d8e2
LP		 1611 { "SCMP_ACT_ALLOW", SCMP_ACT_ALLOW },
1612 { "SCMP_ACT_ERRNO", SCMP_ACT_ERRNO(EPERM) }, /* the OCI spec doesn't document the error, but it appears EPERM is supposed to be used */
1613 { "SCMP_ACT_KILL", SCMP_ACT_KILL },
1614#ifdef SCMP_ACT_KILL_PROCESS
1615 { "SCMP_ACT_KILL_PROCESS", SCMP_ACT_KILL_PROCESS },
1616#endif
1617#ifdef SCMP_ACT_KILL_THREAD
1618 { "SCMP_ACT_KILL_THREAD", SCMP_ACT_KILL_THREAD },
1619#endif
de40a303 1620#ifdef SCMP_ACT_LOG
f9a3d8e2 1621 { "SCMP_ACT_LOG", SCMP_ACT_LOG },
de40a303 1622#endif
f9a3d8e2 1623 { "SCMP_ACT_TRAP", SCMP_ACT_TRAP },
de40a303
LP		 1624
1625 /* We don't support SCMP_ACT_TRACE because that requires a tracer, and that doesn't really make sense
1626 * here */
1627 };
1628
1629 size_t i;
1630
1631 for (i = 0; i < ELEMENTSOF(table); i++)
1632 if (streq_ptr(name, table[i].name)) {
1633 *ret = table[i].action;
1634 return 0;
1635 }
1636
1637 return -EINVAL;
1638}
1639
1640static int oci_seccomp_arch_from_string(const char *name, uint32_t *ret) {
1641
1642 static const struct {
1643 const char *name;
1644 uint32_t arch;
1645 } table[] = {
1646 { "SCMP_ARCH_AARCH64", SCMP_ARCH_AARCH64 },
1647 { "SCMP_ARCH_ARM", SCMP_ARCH_ARM },
1648 { "SCMP_ARCH_MIPS", SCMP_ARCH_MIPS },
1649 { "SCMP_ARCH_MIPS64", SCMP_ARCH_MIPS64 },
1650 { "SCMP_ARCH_MIPS64N32", SCMP_ARCH_MIPS64N32 },
1651 { "SCMP_ARCH_MIPSEL", SCMP_ARCH_MIPSEL },
1652 { "SCMP_ARCH_MIPSEL64", SCMP_ARCH_MIPSEL64 },
1653 { "SCMP_ARCH_MIPSEL64N32", SCMP_ARCH_MIPSEL64N32 },
1654 { "SCMP_ARCH_NATIVE", SCMP_ARCH_NATIVE },
1655#ifdef SCMP_ARCH_PARISC
1656 { "SCMP_ARCH_PARISC", SCMP_ARCH_PARISC },
1657#endif
1658#ifdef SCMP_ARCH_PARISC64
1659 { "SCMP_ARCH_PARISC64", SCMP_ARCH_PARISC64 },
1660#endif
1661 { "SCMP_ARCH_PPC", SCMP_ARCH_PPC },
1662 { "SCMP_ARCH_PPC64", SCMP_ARCH_PPC64 },
1663 { "SCMP_ARCH_PPC64LE", SCMP_ARCH_PPC64LE },
f9252236
AJ		 1664#ifdef SCMP_ARCH_RISCV64
1665 { "SCMP_ARCH_RISCV64", SCMP_ARCH_RISCV64 },
1666#endif
de40a303
LP		 1667 { "SCMP_ARCH_S390", SCMP_ARCH_S390 },
1668 { "SCMP_ARCH_S390X", SCMP_ARCH_S390X },
1669 { "SCMP_ARCH_X32", SCMP_ARCH_X32 },
1670 { "SCMP_ARCH_X86", SCMP_ARCH_X86 },
1671 { "SCMP_ARCH_X86_64", SCMP_ARCH_X86_64 },
1672 };
1673
1674 size_t i;
1675
1676 for (i = 0; i < ELEMENTSOF(table); i++)
1677 if (streq_ptr(table[i].name, name)) {
1678 *ret = table[i].arch;
1679 return 0;
1680 }
1681
1682 return -EINVAL;
1683}
1684
1685static int oci_seccomp_compare_from_string(const char *name, enum scmp_compare *ret) {
1686
1687 static const struct {
1688 const char *name;
1689 enum scmp_compare op;
1690 } table[] = {
1691 { "SCMP_CMP_NE", SCMP_CMP_NE },
1692 { "SCMP_CMP_LT", SCMP_CMP_LT },
1693 { "SCMP_CMP_LE", SCMP_CMP_LE },
1694 { "SCMP_CMP_EQ", SCMP_CMP_EQ },
1695 { "SCMP_CMP_GE", SCMP_CMP_GE },
1696 { "SCMP_CMP_GT", SCMP_CMP_GT },
1697 { "SCMP_CMP_MASKED_EQ", SCMP_CMP_MASKED_EQ },
1698 };
1699
1700 size_t i;
1701
1702 for (i = 0; i < ELEMENTSOF(table); i++)
1703 if (streq_ptr(table[i].name, name)) {
1704 *ret = table[i].op;
1705 return 0;
1706 }
1707
1708 return -EINVAL;
1709}
1710
1711static int oci_seccomp_archs(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1712 scmp_filter_ctx *sc = ASSERT_PTR(userdata);
de40a303
LP		 1713 JsonVariant *e;
1714 int r;
1715
de40a303
LP		 1716 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1717 uint32_t a;
1718
19130626
ZJS		 1719 if (!json_variant_is_string(e))
1720 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
1721 "Architecture entry is not a string");
de40a303
LP		 1722
1723 r = oci_seccomp_arch_from_string(json_variant_string(e), &a);
1724 if (r < 0)
1725 return json_log(e, flags, r, "Unknown architecture: %s", json_variant_string(e));
1726
1727 r = seccomp_arch_add(sc, a);
1728 if (r == -EEXIST)
1729 continue;
1730 if (r < 0)
1731 return json_log(e, flags, r, "Failed to add architecture to seccomp filter: %m");
1732 }
1733
1734 return 0;
1735}
1736
1737struct syscall_rule {
1738 char **names;
1739 uint32_t action;
1740 struct scmp_arg_cmp *arguments;
1741 size_t n_arguments;
1742};
1743
1744static void syscall_rule_free(struct syscall_rule *rule) {
1745 assert(rule);
1746
1747 strv_free(rule->names);
1748 free(rule->arguments);
1749};
1750
1751static int oci_seccomp_action(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1752 uint32_t *action = ASSERT_PTR(userdata);
de40a303
LP		 1753 int r;
1754
de40a303
LP		 1755 r = oci_seccomp_action_from_string(json_variant_string(v), action);
1756 if (r < 0)
1757 return json_log(v, flags, r, "Unknown system call action '%s': %m", json_variant_string(v));
1758
1759 return 0;
1760}
1761
1762static int oci_seccomp_op(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1763 enum scmp_compare *op = ASSERT_PTR(userdata);
de40a303
LP		 1764 int r;
1765
de40a303
LP		 1766 r = oci_seccomp_compare_from_string(json_variant_string(v), op);
1767 if (r < 0)
1768 return json_log(v, flags, r, "Unknown seccomp operator '%s': %m", json_variant_string(v));
1769
1770 return 0;
1771}
1772
1773static int oci_seccomp_args(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1774 struct syscall_rule *rule = ASSERT_PTR(userdata);
de40a303
LP		 1775 JsonVariant *e;
1776 int r;
1777
de40a303
LP		 1778 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1779 static const struct JsonDispatch table[] = {
1780 { "index", JSON_VARIANT_UNSIGNED, json_dispatch_uint32, offsetof(struct scmp_arg_cmp, arg), JSON_MANDATORY },
1781 { "value", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct scmp_arg_cmp, datum_a), JSON_MANDATORY },
1782 { "valueTwo", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct scmp_arg_cmp, datum_b), 0 },
1783 { "op", JSON_VARIANT_STRING, oci_seccomp_op, offsetof(struct scmp_arg_cmp, op), JSON_MANDATORY },
1784 {},
1785 };
1786
1787 struct scmp_arg_cmp *a, *p;
1788 int expected;
1789
1790 a = reallocarray(rule->arguments, rule->n_arguments + 1, sizeof(struct syscall_rule));
1791 if (!a)
1792 return log_oom();
1793
1794 rule->arguments = a;
1795 p = rule->arguments + rule->n_arguments;
1796
1797 *p = (struct scmp_arg_cmp) {
1798 .arg = 0,
1799 .datum_a = 0,
1800 .datum_b = 0,
1801 .op = 0,
1802 };
1803
1804 r = json_dispatch(e, table, oci_unexpected, flags, p);
1805 if (r < 0)
1806 return r;
1807
1808 expected = p->op == SCMP_CMP_MASKED_EQ ? 4 : 3;
1809 if (r != expected)
1810 json_log(e, flags|JSON_WARNING, 0, "Wrong number of system call arguments for JSON data data, ignoring.");
1811
1812 /* Note that we are a bit sloppy here and do not insist that SCMP_CMP_MASKED_EQ gets two datum values,
1813 * and the other only one. That's because buildah for example by default calls things with
1814 * SCMP_CMP_MASKED_EQ but only one argument. We use 0 when the value is not specified. */
1815
1816 rule->n_arguments++;
1817 }
1818
1819 return 0;
1820}
1821
1822static int oci_seccomp_syscalls(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1823 scmp_filter_ctx *sc = ASSERT_PTR(userdata);
de40a303
LP		 1824 JsonVariant *e;
1825 int r;
1826
de40a303
LP		 1827 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1828 static const JsonDispatch table[] = {
1829 { "names", JSON_VARIANT_ARRAY, json_dispatch_strv, offsetof(struct syscall_rule, names), JSON_MANDATORY },
1830 { "action", JSON_VARIANT_STRING, oci_seccomp_action, offsetof(struct syscall_rule, action), JSON_MANDATORY },
1831 { "args", JSON_VARIANT_ARRAY, oci_seccomp_args, 0, 0 },
1832 };
1833 struct syscall_rule rule = {
f5fbe71d 1834 .action = UINT32_MAX,
de40a303 1835 };
de40a303
LP		 1836
1837 r = json_dispatch(e, table, oci_unexpected, flags, &rule);
1838 if (r < 0)
1839 goto fail_rule;
1840
1841 if (strv_isempty(rule.names)) {
1842 json_log(e, flags, 0, "System call name list is empty.");
1843 r = -EINVAL;
1844 goto fail_rule;
1845 }
1846
1847 STRV_FOREACH(i, rule.names) {
1848 int nr;
1849
1850 nr = seccomp_syscall_resolve_name(*i);
1851 if (nr == __NR_SCMP_ERROR) {
1852 log_debug("Unknown syscall %s, skipping.", *i);
1853 continue;
1854 }
1855
1856 r = seccomp_rule_add_array(sc, rule.action, nr, rule.n_arguments, rule.arguments);
1857 if (r < 0)
1858 goto fail_rule;
1859 }
1860
1861 syscall_rule_free(&rule);
1862 continue;
1863
1864 fail_rule:
1865 syscall_rule_free(&rule);
1866 return r;
1867 }
1868
1869 return 0;
1870}
ae408d77 1871#endif
de40a303
LP		 1872
1873static int oci_seccomp(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1874
ae408d77 1875#if HAVE_SECCOMP
de40a303
LP		 1876 static const JsonDispatch table[] = {
1877 { "defaultAction", JSON_VARIANT_STRING, NULL, 0, JSON_MANDATORY },
1878 { "architectures", JSON_VARIANT_ARRAY, oci_seccomp_archs, 0, 0 },
1879 { "syscalls", JSON_VARIANT_ARRAY, oci_seccomp_syscalls, 0, 0 },
1880 {}
1881 };
1882
1883 _cleanup_(seccomp_releasep) scmp_filter_ctx sc = NULL;
99534007 1884 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 1885 JsonVariant *def;
1886 uint32_t d;
1887 int r;
1888
de40a303 1889 def = json_variant_by_key(v, "defaultAction");
19130626
ZJS		 1890 if (!def)
1891 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL), "defaultAction element missing.");
de40a303 1892
19130626
ZJS		 1893 if (!json_variant_is_string(def))
1894 return json_log(def, flags, SYNTHETIC_ERRNO(EINVAL), "defaultAction is not a string.");
de40a303
LP		 1895
1896 r = oci_seccomp_action_from_string(json_variant_string(def), &d);
1897 if (r < 0)
1898 return json_log(def, flags, r, "Unknown default action: %s", json_variant_string(def));
1899
1900 sc = seccomp_init(d);
19130626 1901 if (!sc)
ae408d77 1902 return json_log(v, flags, SYNTHETIC_ERRNO(ENOMEM), "Couldn't allocate seccomp object.");
de40a303
LP		 1903
1904 r = json_dispatch(v, table, oci_unexpected, flags, sc);
1905 if (r < 0)
1906 return r;
1907
1908 seccomp_release(s->seccomp);
1909 s->seccomp = TAKE_PTR(sc);
de40a303 1910 return 0;
ae408d77
LP		 1911#else
1912 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP), "libseccomp support not enabled, can't parse seccomp object.");
1913#endif
de40a303
LP		 1914}
1915
1916static int oci_rootfs_propagation(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1917 const char *s;
1918
1919 s = json_variant_string(v);
1920
1921 if (streq(s, "shared"))
1922 return 0;
1923
1924 json_log(v, flags|JSON_DEBUG, 0, "Ignoring rootfsPropagation setting '%s'.", s);
1925 return 0;
1926}
1927
1928static int oci_masked_paths(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1929 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 1930 JsonVariant *e;
1931
de40a303
LP		 1932 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1933 _cleanup_free_ char *destination = NULL;
1934 CustomMount *m;
1935 const char *p;
1936
19130626
ZJS		 1937 if (!json_variant_is_string(e))
1938 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1939 "Path is not a string, refusing.");
de40a303
LP		 1940
1941 assert_se(p = json_variant_string(e));
1942
19130626
ZJS		 1943 if (!path_is_absolute(p))
1944 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
ad337e55 1945 "Path is not absolute, refusing: %s", p);
de40a303
LP		 1946
1947 if (oci_exclude_mount(p))
1948 continue;
1949
1950 destination = strdup(p);
1951 if (!destination)
1952 return log_oom();
1953
1954 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_INACCESSIBLE);
1955 if (!m)
1956 return log_oom();
1957
1958 m->destination = TAKE_PTR(destination);
1959
1960 /* The spec doesn't say this, but apparently pre-existing implementations are lenient towards
1961 * non-existing paths to mask. Let's hence be too. */
1962 m->graceful = true;
1963 }
1964
1965 return 0;
1966}
1967
1968static int oci_readonly_paths(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 1969 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 1970 JsonVariant *e;
1971
de40a303
LP		 1972 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1973 _cleanup_free_ char *source = NULL, *destination = NULL;
1974 CustomMount *m;
1975 const char *p;
1976
19130626
ZJS		 1977 if (!json_variant_is_string(e))
1978 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1979 "Path is not a string, refusing.");
de40a303
LP		 1980
1981 assert_se(p = json_variant_string(e));
1982
19130626
ZJS		 1983 if (!path_is_absolute(p))
1984 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
ad337e55 1985 "Path is not absolute, refusing: %s", p);
de40a303
LP		 1986
1987 if (oci_exclude_mount(p))
1988 continue;
1989
1990 source = strjoin("+", p);
1991 if (!source)
1992 return log_oom();
1993
1994 destination = strdup(p);
1995 if (!destination)
1996 return log_oom();
1997
1998 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_BIND);
1999 if (!m)
2000 return log_oom();
2001
2002 m->source = TAKE_PTR(source);
2003 m->destination = TAKE_PTR(destination);
2004 m->read_only = true;
2005 }
2006
2007 return 0;
2008}
2009
2010static int oci_linux(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2011
2012 static const JsonDispatch table[] = {
2013 { "namespaces", JSON_VARIANT_ARRAY, oci_namespaces, 0, 0 },
2014 { "uidMappings", JSON_VARIANT_ARRAY, oci_uid_gid_mappings, 0, 0 },
2015 { "gidMappings", JSON_VARIANT_ARRAY, oci_uid_gid_mappings, 0, 0 },
2016 { "devices", JSON_VARIANT_ARRAY, oci_devices, 0, 0 },
2017 { "cgroupsPath", JSON_VARIANT_STRING, oci_cgroups_path, 0, 0 },
2018 { "resources", JSON_VARIANT_OBJECT, oci_resources, 0, 0 },
2019 { "intelRdt", JSON_VARIANT_OBJECT, oci_unsupported, 0, JSON_PERMISSIVE },
2020 { "sysctl", JSON_VARIANT_OBJECT, oci_sysctl, 0, 0 },
2021 { "seccomp", JSON_VARIANT_OBJECT, oci_seccomp, 0, 0 },
2022 { "rootfsPropagation", JSON_VARIANT_STRING, oci_rootfs_propagation, 0, 0 },
2023 { "maskedPaths", JSON_VARIANT_ARRAY, oci_masked_paths, 0, 0 },
2024 { "readonlyPaths", JSON_VARIANT_ARRAY, oci_readonly_paths, 0, 0 },
2025 { "mountLabel", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
2026 {}
2027 };
2028
2029 return json_dispatch(v, table, oci_unexpected, flags, userdata);
2030}
2031
2032static int oci_hook_timeout(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2033 usec_t *u = userdata;
718ca772 2034 uint64_t k;
de40a303
LP		 2035
2036 k = json_variant_unsigned(v);
fa8b675a
ZJS		 2037 if (k == 0 || k > (UINT64_MAX-1)/USEC_PER_SEC)
2038 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
2039 "Hook timeout value out of range.");
de40a303
LP		 2040
2041 *u = k * USEC_PER_SEC;
2042 return 0;
2043}
2044
2045static int oci_hooks_array(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
99534007 2046 Settings *s = ASSERT_PTR(userdata);
de40a303
LP		 2047 JsonVariant *e;
2048 int r;
2049
de40a303
LP		 2050 JSON_VARIANT_ARRAY_FOREACH(e, v) {
2051
2052 static const JsonDispatch table[] = {
2053 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(OciHook, path), JSON_MANDATORY },
2054 { "args", JSON_VARIANT_ARRAY, oci_args, offsetof(OciHook, args), 0 },
2055 { "env", JSON_VARIANT_ARRAY, oci_env, offsetof(OciHook, env), 0 },
2056 { "timeout", JSON_VARIANT_UNSIGNED, oci_hook_timeout, offsetof(OciHook, timeout), 0 },
2057 {}
2058 };
2059
2060 OciHook *a, **array, *new_item;
2061 size_t *n_array;
2062
2063 if (streq(name, "prestart")) {
2064 array = &s->oci_hooks_prestart;
2065 n_array = &s->n_oci_hooks_prestart;
2066 } else if (streq(name, "poststart")) {
2067 array = &s->oci_hooks_poststart;
2068 n_array = &s->n_oci_hooks_poststart;
2069 } else {
2070 assert(streq(name, "poststop"));
2071 array = &s->oci_hooks_poststop;
2072 n_array = &s->n_oci_hooks_poststop;
2073 }
2074
2075 a = reallocarray(*array, *n_array + 1, sizeof(OciHook));
2076 if (!a)
2077 return log_oom();
2078
2079 *array = a;
2080 new_item = a + *n_array;
2081
2082 *new_item = (OciHook) {
2083 .timeout = USEC_INFINITY,
2084 };
2085
2086 r = json_dispatch(e, table, oci_unexpected, flags, userdata);
2087 if (r < 0) {
2088 free(new_item->path);
2089 strv_free(new_item->args);
2090 strv_free(new_item->env);
2091 return r;
2092 }
2093
2094 (*n_array) ++;
2095 }
2096
2097 return 0;
2098}
2099
2100static int oci_hooks(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2101
2102 static const JsonDispatch table[] = {
2103 { "prestart", JSON_VARIANT_OBJECT, oci_hooks_array, 0, 0 },
2104 { "poststart", JSON_VARIANT_OBJECT, oci_hooks_array, 0, 0 },
2105 { "poststop", JSON_VARIANT_OBJECT, oci_hooks_array, 0, 0 },
2106 {}
2107 };
2108
2109 return json_dispatch(v, table, oci_unexpected, flags, userdata);
2110}
2111
2112static int oci_annotations(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
33d60b8d
LP		 2113 JsonVariant *w;
2114 const char *k;
de40a303
LP		 2115
2116 JSON_VARIANT_OBJECT_FOREACH(k, w, v) {
de40a303 2117
33d60b8d
LP		 2118 if (isempty(k))
2119 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
19130626 2120 "Annotation with empty key, refusing.");
de40a303 2121
19130626
ZJS		 2122 if (!json_variant_is_string(w))
2123 return json_log(w, flags, SYNTHETIC_ERRNO(EINVAL),
2124 "Annotation has non-string value, refusing.");
de40a303 2125
33d60b8d 2126 json_log(w, flags|JSON_DEBUG, 0, "Ignoring annotation '%s' with value '%s'.", k, json_variant_string(w));
de40a303
LP		 2127 }
2128
2129 return 0;
2130}
2131
2132int oci_load(FILE *f, const char *bundle, Settings **ret) {
2133
2134 static const JsonDispatch table[] = {
2135 { "ociVersion", JSON_VARIANT_STRING, NULL, 0, JSON_MANDATORY },
2136 { "process", JSON_VARIANT_OBJECT, oci_process, 0, 0 },
2137 { "root", JSON_VARIANT_OBJECT, oci_root, 0, 0 },
2138 { "hostname", JSON_VARIANT_STRING, oci_hostname, 0, 0 },
2139 { "mounts", JSON_VARIANT_ARRAY, oci_mounts, 0, 0 },
2140 { "linux", JSON_VARIANT_OBJECT, oci_linux, 0, 0 },
2141 { "hooks", JSON_VARIANT_OBJECT, oci_hooks, 0, 0 },
2142 { "annotations", JSON_VARIANT_OBJECT, oci_annotations, 0, 0 },
2143 {}
2144 };
2145
2146 _cleanup_(json_variant_unrefp) JsonVariant *oci = NULL;
2147 _cleanup_(settings_freep) Settings *s = NULL;
2148 unsigned line = 0, column = 0;
2149 JsonVariant *v;
2150 const char *path;
2151 int r;
2152
2153 assert_se(bundle);
2154
2155 path = strjoina(bundle, "/config.json");
2156
d642f640 2157 r = json_parse_file(f, path, 0, &oci, &line, &column);
de40a303
LP		 2158 if (r < 0) {
2159 if (line != 0 && column != 0)
2160 return log_error_errno(r, "Failed to parse '%s' at %u:%u: %m", path, line, column);
2161 else
2162 return log_error_errno(r, "Failed to parse '%s': %m", path);
2163 }
2164
2165 v = json_variant_by_key(oci, "ociVersion");
d7a0f1f4
FS		 2166 if (!v)
2167 return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
2168 "JSON file '%s' is not an OCI bundle configuration file. Refusing.",
2169 path);
2170 if (!streq_ptr(json_variant_string(v), "1.0.0"))
2171 return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
2172 "OCI bundle version not supported: %s",
2173 strna(json_variant_string(v)));
de40a303
LP		 2174
2175 // {
2176 // _cleanup_free_ char *formatted = NULL;
2177 // assert_se(json_variant_format(oci, JSON_FORMAT_PRETTY|JSON_FORMAT_COLOR, &formatted) >= 0);
2178 // fputs(formatted, stdout);
2179 // }
2180
2181 s = settings_new();
2182 if (!s)
2183 return log_oom();
2184
2185 s->start_mode = START_PID1;
2186 s->resolv_conf = RESOLV_CONF_OFF;
2187 s->link_journal = LINK_NO;
2188 s->timezone = TIMEZONE_OFF;
2189
2190 s->bundle = strdup(bundle);
2191 if (!s->bundle)
2192 return log_oom();
2193
2194 r = json_dispatch(oci, table, oci_unexpected, 0, s);
2195 if (r < 0)
2196 return r;
2197
2198 if (s->properties) {
2199 r = sd_bus_message_seal(s->properties, 0, 0);
2200 if (r < 0)
2201 return log_error_errno(r, "Cannot seal properties bus message: %m");
2202 }
2203
2204 *ret = TAKE_PTR(s);
2205 return 0;
2206}
Unnamed repository; edit this file 'description' to name the repository.