?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
1:255.16-alt1
[systemd_ALT.git] / src / nspawn / nspawn-oci.c
blob8f1ac7ccaddfae49d9b358da8b8be30b3482ff3a
1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
2
3 #include <linux/oom.h>
4
5 #include "bus-util.h"
6 #include "cap-list.h"
7 #include "cpu-set-util.h"
8 #include "device-util.h"
9 #include "devnum-util.h"
10 #include "env-util.h"
11 #include "format-util.h"
12 #include "fs-util.h"
13 #include "hostname-util.h"
14 #include "json.h"
15 #include "missing_sched.h"
16 #include "nspawn-oci.h"
17 #include "path-util.h"
18 #include "rlimit-util.h"
19 #include "seccomp-util.h"
20 #include "stdio-util.h"
21 #include "string-util.h"
22 #include "strv.h"
23 #include "user-util.h"
24
25 /* TODO:
26 * OCI runtime tool implementation
27 * hooks
28 *
29 * Spec issues:
30 *
31 * How is RLIM_INFINITY supposed to be encoded?
32 * configured effective caps is bullshit, as execv() corrupts it anyway
33 * pipes bind mounted is *very* different from pipes newly created, comments regarding bind mount or not are bogus
34 * annotation values structured? or string?
35 * configurable file system namespace path, but then also root path? wtf?
36 * apply sysctl inside of the container? or outside?
37 * how is unlimited pids tasks limit to be encoded?
38 * what are the defaults for caps if not specified?
39 * 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
40 * 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
41 * 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
42 * if type of mount is left out, what shall be assumed? "bind"?
43 * readonly mounts is entirely redundant?
44 * should escaping be applied when joining mount options with ","?
45 * devices cgroup support is bogus, "allow" and "deny" on the kernel level is about adding/removing entries, not about access
46 * spec needs to say that "rwm" devices cgroup combination can't be the empty string
47 * cgrouspv1 crap: kernel, kernelTCP, swappiness, disableOOMKiller, swap, devices, leafWeight
48 * general: it shouldn't leak lower level abstractions this obviously
49 * unmanagable cgroups stuff: realtimeRuntime/realtimePeriod
50 * needs to say what happense when some option is not specified, i.e. which defaults apply
51 * no architecture? no personality?
52 * seccomp example and logic is simply broken: there's no constant "SCMP_ACT_ERRNO".
53 * spec should say what to do with unknown props
54 * /bin/mount regarding NFS and FUSE required?
55 * what does terminal=false mean?
56 * sysctl inside or outside? allow-listing?
57 * swapiness typo -> swappiness
58 *
59 * Unsupported:
60 *
61 * apparmorProfile
62 * selinuxLabel + mountLabel
63 * hugepageLimits
64 * network
65 * rdma
66 * intelRdt
67 * swappiness, disableOOMKiller, kernel, kernelTCP, leafWeight (because it's dead, cgroupsv2 can't do it and hence systemd neither)
68 *
69 * Non-slice cgroup paths
70 * Propagation that is not slave + shared
71 * more than one uid/gid mapping, mappings with a container base != 0, or non-matching uid/gid mappings
72 * device cgroups access = false items that are not catchall
73 * 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.
74 *
75 */
76
77 static int oci_unexpected(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
78 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
79 "Unexpected OCI element '%s' of type '%s'.", name, json_variant_type_to_string(json_variant_type(v)));
80 }
81
82 static int oci_dispatch(JsonVariant *v, const JsonDispatch table[], JsonDispatchFlags flags, void *userdata) {
83 return json_dispatch_full(v, table, oci_unexpected, flags, userdata, /* reterr_bad_field= */ NULL);
84 }
85
86 static int oci_unsupported(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
87 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
88 "Unsupported OCI element '%s' of type '%s'.", name, json_variant_type_to_string(json_variant_type(v)));
89 }
90
91 static int oci_terminal(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
92 Settings *s = ASSERT_PTR(userdata);
93
94 /* If not specified, or set to true, we'll default to either an interactive or a read-only
95 * console. If specified as false, we'll forcibly move to "pipe" mode though. */
96 s->console_mode = json_variant_boolean(v) ? _CONSOLE_MODE_INVALID : CONSOLE_PIPE;
97 return 0;
98 }
99
100 static int oci_console_dimension(const char *name, JsonVariant *variant, JsonDispatchFlags flags, void *userdata) {
101 unsigned *u = ASSERT_PTR(userdata);
102 uint64_t k;
103
104 k = json_variant_unsigned(variant);
105 if (k == 0)
106 return json_log(variant, flags, SYNTHETIC_ERRNO(ERANGE),
107 "Console size field '%s' is too small.", strna(name));
108 if (k > USHRT_MAX) /* TIOCSWINSZ's struct winsize uses "unsigned short" for width and height */
109 return json_log(variant, flags, SYNTHETIC_ERRNO(ERANGE),
110 "Console size field '%s' is too large.", strna(name));
111
112 *u = (unsigned) k;
113 return 0;
114 }
115
116 static int oci_console_size(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
117 Settings *s = ASSERT_PTR(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 oci_dispatch(v, table, flags, s);
126 }
127
128 static int oci_absolute_path(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
129 char **p = ASSERT_PTR(userdata);
130 const char *n;
131
132 n = json_variant_string(v);
133
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);
137
138 return free_and_strdup_warn(p, n);
139 }
140
141 static int oci_env(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
142 char ***l = ASSERT_PTR(userdata);
143 JsonVariant *e;
144 int r;
145
146 JSON_VARIANT_ARRAY_FOREACH(e, v) {
147 const char *n;
148
149 if (!json_variant_is_string(e))
150 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
151 "Environment array contains non-string.");
152
153 assert_se(n = json_variant_string(e));
154
155 if (!env_assignment_is_valid(n))
156 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
157 "Environment assignment not valid: %s", n);
158
159 r = strv_extend(l, n);
160 if (r < 0)
161 return log_oom();
162 }
163
164 return 0;
165 }
166
167 static int oci_args(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
168 _cleanup_strv_free_ char **l = NULL;
169 char ***value = ASSERT_PTR(userdata);
170 int r;
171
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");
175
176 if (strv_isempty(l))
177 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
178 "Argument list empty, refusing.");
179
180 if (isempty(l[0]))
181 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
182 "Executable name is empty, refusing.");
183
184 return strv_free_and_replace(*value, l);
185 }
186
187 static int oci_rlimit_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
188 const char *z;
189 int *type = ASSERT_PTR(userdata);
190 int t;
191
192 z = startswith(json_variant_string(v), "RLIMIT_");
193 if (!z)
194 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
195 "rlimit entry's name does not begin with 'RLIMIT_', refusing: %s",
196 json_variant_string(v));
197
198 t = rlimit_from_string(z);
199 if (t < 0)
200 return json_log(v, flags, t,
201 "rlimit name unknown: %s", json_variant_string(v));
202
203 *type = t;
204 return 0;
205 }
206
207 static int oci_rlimit_value(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
208 rlim_t *value = ASSERT_PTR(userdata);
209 rlim_t z;
210
211 if (json_variant_is_negative(v))
212 z = RLIM_INFINITY;
213 else {
214 if (!json_variant_is_unsigned(v))
215 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
216 "rlimits limit not unsigned, refusing.");
217
218 z = (rlim_t) json_variant_unsigned(v);
219
220 if ((uint64_t) z != json_variant_unsigned(v))
221 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
222 "rlimits limit out of range, refusing.");
223 }
224
225 *value = z;
226 return 0;
227 }
228
229 static int oci_rlimits(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
230 Settings *s = ASSERT_PTR(userdata);
231 JsonVariant *e;
232 int r;
233
234 JSON_VARIANT_ARRAY_FOREACH(e, v) {
235
236 struct rlimit_data {
237 int type;
238 rlim_t soft;
239 rlim_t hard;
240 } data = {
241 .type = -1,
242 .soft = RLIM_INFINITY,
243 .hard = RLIM_INFINITY,
244 };
245
246 static const JsonDispatch table[] = {
247 { "soft", JSON_VARIANT_NUMBER, oci_rlimit_value, offsetof(struct rlimit_data, soft), JSON_MANDATORY },
248 { "hard", JSON_VARIANT_NUMBER, oci_rlimit_value, offsetof(struct rlimit_data, hard), JSON_MANDATORY },
249 { "type", JSON_VARIANT_STRING, oci_rlimit_type, offsetof(struct rlimit_data, type), JSON_MANDATORY },
250 {}
251 };
252
253 r = oci_dispatch(e, table, flags, &data);
254 if (r < 0)
255 return r;
256
257 assert(data.type >= 0);
258 assert(data.type < _RLIMIT_MAX);
259
260 if (s->rlimit[data.type])
261 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
262 "rlimits array contains duplicate entry, refusing.");
263
264 s->rlimit[data.type] = new(struct rlimit, 1);
265 if (!s->rlimit[data.type])
266 return log_oom();
267
268 *s->rlimit[data.type] = (struct rlimit) {
269 .rlim_cur = data.soft,
270 .rlim_max = data.hard,
271 };
272
273 }
274 return 0;
275 }
276
277 static int oci_capability_array(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
278 uint64_t *mask = ASSERT_PTR(userdata);
279 uint64_t m = 0;
280 JsonVariant *e;
281
282 JSON_VARIANT_ARRAY_FOREACH(e, v) {
283 const char *n;
284 int cap;
285
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.");
289
290 assert_se(n = json_variant_string(e));
291
292 cap = capability_from_name(n);
293 if (cap < 0)
294 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
295 "Unknown capability: %s", n);
296
297 m |= UINT64_C(1) << cap;
298 }
299
300 if (*mask == UINT64_MAX)
301 *mask = m;
302 else
303 *mask |= m;
304
305 return 0;
306 }
307
308 static 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
319 Settings *s = ASSERT_PTR(userdata);
320 int r;
321
322 r = oci_dispatch(v, table, flags, &s->full_capabilities);
323 if (r < 0)
324 return r;
325
326 if (s->full_capabilities.bounding != UINT64_MAX) {
327 s->capability = s->full_capabilities.bounding;
328 s->drop_capability = ~s->full_capabilities.bounding;
329 }
330
331 return 0;
332 }
333
334 static int oci_oom_score_adj(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
335 Settings *s = ASSERT_PTR(userdata);
336 int64_t k;
337
338 k = json_variant_integer(v);
339 if (k < OOM_SCORE_ADJ_MIN || k > OOM_SCORE_ADJ_MAX)
340 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
341 "oomScoreAdj value out of range: %" PRIi64, k);
342
343 s->oom_score_adjust = (int) k;
344 s->oom_score_adjust_set = true;
345
346 return 0;
347 }
348
349 static int oci_uid_gid(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
350 uid_t *uid = ASSERT_PTR(userdata);
351 uid_t u;
352 uint64_t k;
353
354 assert_cc(sizeof(uid_t) == sizeof(gid_t));
355
356 k = json_variant_unsigned(v);
357 u = (uid_t) k;
358 if ((uint64_t) u != k)
359 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
360 "UID/GID out of range: %" PRIu64, k);
361
362 if (!uid_is_valid(u))
363 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
364 "UID/GID is not valid: " UID_FMT, u);
365
366 *uid = u;
367 return 0;
368 }
369
370 static int oci_supplementary_gids(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
371 Settings *s = ASSERT_PTR(userdata);
372 JsonVariant *e;
373 int r;
374
375 JSON_VARIANT_ARRAY_FOREACH(e, v) {
376 gid_t gid, *a;
377
378 if (!json_variant_is_unsigned(e))
379 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
380 "Supplementary GID entry is not a UID.");
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
397 static int oci_user(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
398
399 static const JsonDispatch table[] = {
400 { "uid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(Settings, uid), JSON_MANDATORY },
401 { "gid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(Settings, gid), JSON_MANDATORY },
402 { "additionalGids", JSON_VARIANT_ARRAY, oci_supplementary_gids, 0, 0 },
403 {}
404 };
405
406 return oci_dispatch(v, table, flags, userdata);
407 }
408
409 static int oci_process(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
410
411 static const JsonDispatch table[] = {
412 { "terminal", JSON_VARIANT_BOOLEAN, oci_terminal, 0, 0 },
413 { "consoleSize", JSON_VARIANT_OBJECT, oci_console_size, 0, 0 },
414 { "cwd", JSON_VARIANT_STRING, oci_absolute_path, offsetof(Settings, working_directory), 0 },
415 { "env", JSON_VARIANT_ARRAY, oci_env, offsetof(Settings, environment), 0 },
416 { "args", JSON_VARIANT_ARRAY, oci_args, offsetof(Settings, parameters), 0 },
417 { "rlimits", JSON_VARIANT_ARRAY, oci_rlimits, 0, 0 },
418 { "apparmorProfile", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
419 { "capabilities", JSON_VARIANT_OBJECT, oci_capabilities, 0, 0 },
420 { "noNewPrivileges", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(Settings, no_new_privileges), 0 },
421 { "oomScoreAdj", JSON_VARIANT_INTEGER, oci_oom_score_adj, 0, 0 },
422 { "selinuxLabel", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
423 { "user", JSON_VARIANT_OBJECT, oci_user, 0, 0 },
424 {}
425 };
426
427 return oci_dispatch(v, table, flags, userdata);
428 }
429
430 static int oci_root(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
431 Settings *s = ASSERT_PTR(userdata);
432 int r;
433
434 static const JsonDispatch table[] = {
435 { "path", JSON_VARIANT_STRING, json_dispatch_string, offsetof(Settings, root) },
436 { "readonly", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(Settings, read_only) },
437 {}
438 };
439
440 r = oci_dispatch(v, table, flags, s);
441 if (r < 0)
442 return r;
443
444 if (s->root && !path_is_absolute(s->root)) {
445 char *joined;
446
447 joined = path_join(s->bundle, s->root);
448 if (!joined)
449 return log_oom();
450
451 free_and_replace(s->root, joined);
452 }
453
454 return 0;
455 }
456
457 static int oci_hostname(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
458 Settings *s = ASSERT_PTR(userdata);
459 const char *n;
460
461 assert_se(n = json_variant_string(v));
462
463 if (!hostname_is_valid(n, 0))
464 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
465 "Hostname string is not a valid hostname: %s", n);
466
467 return free_and_strdup_warn(&s->hostname, n);
468 }
469
470 static bool oci_exclude_mount(const char *path) {
471
472 /* Returns "true" for all mounts we insist to mount on our own, and hence ignore the OCI data. */
473
474 if (PATH_IN_SET(path,
475 "/dev",
476 "/dev/mqueue",
477 "/dev/pts",
478 "/dev/shm",
479 "/proc",
480 "/proc/acpi",
481 "/proc/apm",
482 "/proc/asound",
483 "/proc/bus",
484 "/proc/fs",
485 "/proc/irq",
486 "/proc/kallsyms",
487 "/proc/kcore",
488 "/proc/keys",
489 "/proc/scsi",
490 "/proc/sys",
491 "/proc/sys/net",
492 "/proc/sysrq-trigger",
493 "/proc/timer_list",
494 "/run",
495 "/sys",
496 "/sys",
497 "/sys/fs/selinux",
498 "/tmp"))
499 return true;
500
501 /* Similar, skip the whole /sys/fs/cgroups subtree */
502 if (path_startswith(path, "/sys/fs/cgroup"))
503 return true;
504
505 return false;
506 }
507
508 typedef struct oci_mount_data {
509 char *destination;
510 char *source;
511 char *type;
512 char **options;
513 } oci_mount_data;
514
515 static void oci_mount_data_done(oci_mount_data *data) {
516 assert(data);
517
518 free(data->destination);
519 free(data->source);
520 free(data->type);
521 strv_free(data->options);
522 }
523
524 static int oci_mounts(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
525 Settings *s = ASSERT_PTR(userdata);
526 JsonVariant *e;
527 int r;
528
529 JSON_VARIANT_ARRAY_FOREACH(e, v) {
530 static const JsonDispatch table[] = {
531 { "destination", JSON_VARIANT_STRING, oci_absolute_path, offsetof(oci_mount_data, destination), JSON_MANDATORY },
532 { "source", JSON_VARIANT_STRING, json_dispatch_string, offsetof(oci_mount_data, source), 0 },
533 { "options", JSON_VARIANT_ARRAY, json_dispatch_strv, offsetof(oci_mount_data, options), 0, },
534 { "type", JSON_VARIANT_STRING, json_dispatch_string, offsetof(oci_mount_data, type), 0 },
535 {}
536 };
537
538 _cleanup_free_ char *joined_options = NULL;
539 _cleanup_(oci_mount_data_done) oci_mount_data data = {};
540 CustomMount *m;
541
542 r = oci_dispatch(e, table, flags, &data);
543 if (r < 0)
544 return r;
545
546 if (!path_is_absolute(data.destination))
547 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
548 "Mount destination not an absolute path: %s", data.destination);
549
550 if (oci_exclude_mount(data.destination))
551 continue;
552
553 if (data.options) {
554 joined_options = strv_join(data.options, ",");
555 if (!joined_options)
556 return log_oom();
557 }
558
559 if (!data.type || streq(data.type, "bind")) {
560 if (data.source && !path_is_absolute(data.source)) {
561 char *joined;
562
563 joined = path_join(s->bundle, data.source);
564 if (!joined)
565 return log_oom();
566
567 free_and_replace(data.source, joined);
568 }
569
570 data.type = mfree(data.type);
571
572 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_BIND);
573 } else
574 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_ARBITRARY);
575 if (!m)
576 return log_oom();
577
578 m->destination = TAKE_PTR(data.destination);
579 m->source = TAKE_PTR(data.source);
580 m->options = TAKE_PTR(joined_options);
581 m->type_argument = TAKE_PTR(data.type);
582 }
583
584 return 0;
585 }
586
587 static int oci_namespace_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
588 unsigned long *nsflags = ASSERT_PTR(userdata);
589 const char *n;
590
591 assert_se(n = json_variant_string(v));
592
593 /* We don't use namespace_flags_from_string() here, as the OCI spec uses slightly different names than the
594 * kernel here. */
595 if (streq(n, "pid"))
596 *nsflags = CLONE_NEWPID;
597 else if (streq(n, "network"))
598 *nsflags = CLONE_NEWNET;
599 else if (streq(n, "mount"))
600 *nsflags = CLONE_NEWNS;
601 else if (streq(n, "ipc"))
602 *nsflags = CLONE_NEWIPC;
603 else if (streq(n, "uts"))
604 *nsflags = CLONE_NEWUTS;
605 else if (streq(n, "user"))
606 *nsflags = CLONE_NEWUSER;
607 else if (streq(n, "cgroup"))
608 *nsflags = CLONE_NEWCGROUP;
609 else
610 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
611 "Unknown namespace type, refusing: %s", n);
612
613 return 0;
614 }
615
616 struct namespace_data {
617 unsigned long type;
618 char *path;
619 };
620
621 static void namespace_data_done(struct namespace_data *data) {
622 assert(data);
623
624 free(data->path);
625 }
626
627 static int oci_namespaces(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
628 Settings *s = ASSERT_PTR(userdata);
629 unsigned long n = 0;
630 JsonVariant *e;
631 int r;
632
633 JSON_VARIANT_ARRAY_FOREACH(e, v) {
634 _cleanup_(namespace_data_done) struct namespace_data data = {};
635
636 static const JsonDispatch table[] = {
637 { "type", JSON_VARIANT_STRING, oci_namespace_type, offsetof(struct namespace_data, type), JSON_MANDATORY },
638 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(struct namespace_data, path), 0 },
639 {}
640 };
641
642 r = oci_dispatch(e, table, flags, &data);
643 if (r < 0)
644 return r;
645
646 if (data.path) {
647 if (data.type != CLONE_NEWNET)
648 return json_log(e, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
649 "Specifying namespace path for non-network namespace is not supported.");
650
651 if (s->network_namespace_path)
652 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
653 "Network namespace path specified more than once, refusing.");
654
655 free_and_replace(s->network_namespace_path, data.path);
656 }
657
658 if (FLAGS_SET(n, data.type))
659 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
660 "Duplicate namespace specification, refusing.");
661
662 n |= data.type;
663 }
664
665 if (!FLAGS_SET(n, CLONE_NEWNS))
666 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
667 "Containers without a mount namespace aren't supported.");
668
669 s->private_network = FLAGS_SET(n, CLONE_NEWNET);
670 s->userns_mode = FLAGS_SET(n, CLONE_NEWUSER) ? USER_NAMESPACE_FIXED : USER_NAMESPACE_NO;
671 s->use_cgns = FLAGS_SET(n, CLONE_NEWCGROUP);
672
673 s->clone_ns_flags = n & (CLONE_NEWIPC|CLONE_NEWPID|CLONE_NEWUTS);
674
675 return 0;
676 }
677
678 static int oci_uid_gid_range(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
679 uid_t *uid = ASSERT_PTR(userdata);
680 uid_t u;
681 uint64_t k;
682
683 assert_cc(sizeof(uid_t) == sizeof(gid_t));
684
685 /* This is very much like oci_uid_gid(), except the checks are a bit different, as this is a UID range rather
686 * than a specific UID, and hence UID_INVALID has no special significance. OTOH a range of zero makes no
687 * sense. */
688
689 k = json_variant_unsigned(v);
690 u = (uid_t) k;
691 if ((uint64_t) u != k)
692 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
693 "UID/GID out of range: %" PRIu64, k);
694 if (u == 0)
695 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
696 "UID/GID range can't be zero.");
697
698 *uid = u;
699 return 0;
700 }
701
702 static int oci_uid_gid_mappings(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
703 struct mapping_data {
704 uid_t host_id;
705 uid_t container_id;
706 uid_t range;
707 } data = {
708 .host_id = UID_INVALID,
709 .container_id = UID_INVALID,
710 .range = 0,
711 };
712
713 static const JsonDispatch table[] = {
714 { "containerID", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(struct mapping_data, container_id), JSON_MANDATORY },
715 { "hostID", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(struct mapping_data, host_id), JSON_MANDATORY },
716 { "size", JSON_VARIANT_UNSIGNED, oci_uid_gid_range, offsetof(struct mapping_data, range), JSON_MANDATORY },
717 {}
718 };
719
720 Settings *s = ASSERT_PTR(userdata);
721 JsonVariant *e;
722 int r;
723
724 if (json_variant_elements(v) == 0)
725 return 0;
726
727 if (json_variant_elements(v) > 1)
728 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
729 "UID/GID mappings with more than one entry are not supported.");
730
731 assert_se(e = json_variant_by_index(v, 0));
732
733 r = oci_dispatch(e, table, flags, &data);
734 if (r < 0)
735 return r;
736
737 if (data.host_id + data.range < data.host_id ||
738 data.container_id + data.range < data.container_id)
739 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
740 "UID/GID range goes beyond UID/GID validity range, refusing.");
741
742 if (data.container_id != 0)
743 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
744 "UID/GID mappings with a non-zero container base are not supported.");
745
746 if (data.range < 0x10000)
747 json_log(v, flags|JSON_WARNING, 0,
748 "UID/GID mapping with less than 65536 UID/GIDS set up, you are looking for trouble.");
749
750 if (s->uid_range != UID_INVALID &&
751 (s->uid_shift != data.host_id || s->uid_range != data.range))
752 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
753 "Non-matching UID and GID mappings are not supported.");
754
755 s->uid_shift = data.host_id;
756 s->uid_range = data.range;
757
758 return 0;
759 }
760
761 static int oci_device_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
762 mode_t *mode = ASSERT_PTR(userdata);
763 const char *t;
764
765 assert_se(t = json_variant_string(v));
766
767 if (STR_IN_SET(t, "c", "u"))
768 *mode = (*mode & ~S_IFMT) | S_IFCHR;
769 else if (streq(t, "b"))
770 *mode = (*mode & ~S_IFMT) | S_IFBLK;
771 else if (streq(t, "p"))
772 *mode = (*mode & ~S_IFMT) | S_IFIFO;
773 else
774 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
775 "Unknown device type: %s", t);
776
777 return 0;
778 }
779
780 static int oci_device_major(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
781 unsigned *u = ASSERT_PTR(userdata);
782 uint64_t k;
783
784 k = json_variant_unsigned(v);
785 if (!DEVICE_MAJOR_VALID(k))
786 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
787 "Device major %" PRIu64 " out of range.", k);
788
789 *u = (unsigned) k;
790 return 0;
791 }
792
793 static int oci_device_minor(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
794 unsigned *u = ASSERT_PTR(userdata);
795 uint64_t k;
796
797 k = json_variant_unsigned(v);
798 if (!DEVICE_MINOR_VALID(k))
799 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
800 "Device minor %" PRIu64 " out of range.", k);
801
802 *u = (unsigned) k;
803 return 0;
804 }
805
806 static int oci_device_file_mode(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
807 mode_t *mode = ASSERT_PTR(userdata);
808 mode_t m;
809 uint64_t k;
810
811 k = json_variant_unsigned(v);
812 m = (mode_t) k;
813
814 if ((m & ~07777) != 0 || (uint64_t) m != k)
815 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
816 "fileMode out of range, refusing.");
817
818 *mode = (*mode & ~07777) | m;
819 return 0;
820 }
821
822 static int oci_devices(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
823 Settings *s = ASSERT_PTR(userdata);
824 JsonVariant *e;
825 int r;
826
827 JSON_VARIANT_ARRAY_FOREACH(e, v) {
828
829 static const JsonDispatch table[] = {
830 { "type", JSON_VARIANT_STRING, oci_device_type, offsetof(DeviceNode, mode), JSON_MANDATORY },
831 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(DeviceNode, path), JSON_MANDATORY },
832 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(DeviceNode, major), 0 },
833 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(DeviceNode, minor), 0 },
834 { "fileMode", JSON_VARIANT_UNSIGNED, oci_device_file_mode, offsetof(DeviceNode, mode), 0 },
835 { "uid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(DeviceNode, uid), 0 },
836 { "gid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(DeviceNode, gid), 0 },
837 {}
838 };
839
840 DeviceNode *node, *nodes;
841
842 nodes = reallocarray(s->extra_nodes, s->n_extra_nodes + 1, sizeof(DeviceNode));
843 if (!nodes)
844 return log_oom();
845
846 s->extra_nodes = nodes;
847
848 node = nodes + s->n_extra_nodes;
849 *node = (DeviceNode) {
850 .uid = UID_INVALID,
851 .gid = GID_INVALID,
852 .major = UINT_MAX,
853 .minor = UINT_MAX,
854 .mode = 0644,
855 };
856
857 r = oci_dispatch(e, table, flags, node);
858 if (r < 0)
859 goto fail_element;
860
861 if (S_ISCHR(node->mode) || S_ISBLK(node->mode)) {
862 _cleanup_free_ char *path = NULL;
863
864 if (node->major == UINT_MAX || node->minor == UINT_MAX) {
865 r = json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
866 "Major/minor required when device node is device node");
867 goto fail_element;
868 }
869
870 /* Suppress a couple of implicit device nodes */
871 r = devname_from_devnum(node->mode, makedev(node->major, node->minor), &path);
872 if (r < 0)
873 json_log(e, flags|JSON_DEBUG, r, "Failed to resolve device node %u:%u, ignoring: %m", node->major, node->minor);
874 else {
875 if (PATH_IN_SET(path,
876 "/dev/null",
877 "/dev/zero",
878 "/dev/full",
879 "/dev/random",
880 "/dev/urandom",
881 "/dev/tty",
882 "/dev/net/tun",
883 "/dev/ptmx",
884 "/dev/pts/ptmx",
885 "/dev/console")) {
886
887 json_log(e, flags|JSON_DEBUG, 0, "Ignoring devices item for device '%s', as it is implicitly created anyway.", path);
888 free(node->path);
889 continue;
890 }
891 }
892 }
893
894 s->n_extra_nodes++;
895 continue;
896
897 fail_element:
898 free(node->path);
899 return r;
900 }
901
902 return 0;
903 }
904
905 static int oci_cgroups_path(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
906 _cleanup_free_ char *slice = NULL, *backwards = NULL;
907 Settings *s = ASSERT_PTR(userdata);
908 const char *p;
909 int r;
910
911 assert_se(p = json_variant_string(v));
912
913 r = cg_path_get_slice(p, &slice);
914 if (r < 0)
915 return json_log(v, flags, r, "Couldn't derive slice unit name from path '%s': %m", p);
916
917 r = cg_slice_to_path(slice, &backwards);
918 if (r < 0)
919 return json_log(v, flags, r, "Couldn't convert slice unit name '%s' back to path: %m", slice);
920
921 if (!path_equal(backwards, p))
922 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
923 "Control group path '%s' does not refer to slice unit, refusing.", p);
924
925 free_and_replace(s->slice, slice);
926 return 0;
927 }
928
929 static int oci_cgroup_device_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
930 mode_t *mode = ASSERT_PTR(userdata);
931 const char *n;
932
933 assert_se(n = json_variant_string(v));
934
935 if (streq(n, "c"))
936 *mode = S_IFCHR;
937 else if (streq(n, "b"))
938 *mode = S_IFBLK;
939 else
940 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
941 "Control group device type unknown: %s", n);
942
943 return 0;
944 }
945
946 struct device_data {
947 bool allow;
948 bool r;
949 bool w;
950 bool m;
951 mode_t type;
952 unsigned major;
953 unsigned minor;
954 };
955
956 static int oci_cgroup_device_access(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
957 struct device_data *d = ASSERT_PTR(userdata);
958 bool r = false, w = false, m = false;
959 const char *s;
960 size_t i;
961
962 assert_se(s = json_variant_string(v));
963
964 for (i = 0; s[i]; i++)
965 if (s[i] == 'r')
966 r = true;
967 else if (s[i] == 'w')
968 w = true;
969 else if (s[i] == 'm')
970 m = true;
971 else
972 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
973 "Unknown device access character '%c'.", s[i]);
974
975 d->r = r;
976 d->w = w;
977 d->m = m;
978
979 return 0;
980 }
981
982 static int oci_cgroup_devices(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
983 _cleanup_free_ struct device_data *list = NULL;
984 Settings *s = ASSERT_PTR(userdata);
985 size_t n_list = 0, i;
986 bool noop = false;
987 JsonVariant *e;
988 int r;
989
990 JSON_VARIANT_ARRAY_FOREACH(e, v) {
991
992 struct device_data data = {
993 .major = UINT_MAX,
994 .minor = UINT_MAX,
995 }, *a;
996
997 static const JsonDispatch table[] = {
998 { "allow", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(struct device_data, allow), JSON_MANDATORY },
999 { "type", JSON_VARIANT_STRING, oci_cgroup_device_type, offsetof(struct device_data, type), 0 },
1000 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), 0 },
1001 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), 0 },
1002 { "access", JSON_VARIANT_STRING, oci_cgroup_device_access, 0, 0 },
1003 {}
1004 };
1005
1006 r = oci_dispatch(e, table, flags, &data);
1007 if (r < 0)
1008 return r;
1009
1010 if (!data.allow) {
1011 /* The fact that OCI allows 'deny' entries makes really no sense, as 'allow'
1012 * vs. 'deny' for the devices cgroup controller is really not about allow-listing and
1013 * deny-listing but about adding and removing entries from the allow list. Since we
1014 * always start out with an empty allow list we hence ignore the whole thing, as
1015 * removing entries which don't exist make no sense. We'll log about this, since this
1016 * is really borked in the spec, with one exception: the entry that's supposed to
1017 * drop the kernel's default we ignore silently */
1018
1019 if (!data.r || !data.w || !data.m || data.type != 0 || data.major != UINT_MAX || data.minor != UINT_MAX)
1020 json_log(v, flags|JSON_WARNING, 0, "Devices cgroup allow list with arbitrary 'allow' entries not supported, ignoring.");
1021
1022 /* We ignore the 'deny' entry as for us that's implied */
1023 continue;
1024 }
1025
1026 if (!data.r && !data.w && !data.m) {
1027 json_log(v, flags|LOG_WARNING, 0, "Device cgroup allow list entry with no effect found, ignoring.");
1028 continue;
1029 }
1030
1031 if (data.minor != UINT_MAX && data.major == UINT_MAX)
1032 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
1033 "Device cgroup allow list entries with minors but no majors not supported.");
1034
1035 if (data.major != UINT_MAX && data.type == 0)
1036 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
1037 "Device cgroup allow list entries with majors but no device node type not supported.");
1038
1039 if (data.type == 0) {
1040 if (data.r && data.w && data.m) /* a catchall allow list entry means we are looking at a noop */
1041 noop = true;
1042 else
1043 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
1044 "Device cgroup allow list entries with no type not supported.");
1045 }
1046
1047 a = reallocarray(list, n_list + 1, sizeof(struct device_data));
1048 if (!a)
1049 return log_oom();
1050
1051 list = a;
1052 list[n_list++] = data;
1053 }
1054
1055 if (noop)
1056 return 0;
1057
1058 r = settings_allocate_properties(s);
1059 if (r < 0)
1060 return r;
1061
1062 r = sd_bus_message_open_container(s->properties, 'r', "sv");
1063 if (r < 0)
1064 return bus_log_create_error(r);
1065
1066 r = sd_bus_message_append(s->properties, "s", "DeviceAllow");
1067 if (r < 0)
1068 return bus_log_create_error(r);
1069
1070 r = sd_bus_message_open_container(s->properties, 'v', "a(ss)");
1071 if (r < 0)
1072 return bus_log_create_error(r);
1073
1074 r = sd_bus_message_open_container(s->properties, 'a', "(ss)");
1075 if (r < 0)
1076 return bus_log_create_error(r);
1077
1078 for (i = 0; i < n_list; i++) {
1079 _cleanup_free_ char *pattern = NULL;
1080 char access[4];
1081 size_t n = 0;
1082
1083 if (list[i].minor == UINT_MAX) {
1084 const char *t;
1085
1086 if (list[i].type == S_IFBLK)
1087 t = "block";
1088 else {
1089 assert(list[i].type == S_IFCHR);
1090 t = "char";
1091 }
1092
1093 if (list[i].major == UINT_MAX) {
1094 pattern = strjoin(t, "-*");
1095 if (!pattern)
1096 return log_oom();
1097 } else {
1098 if (asprintf(&pattern, "%s-%u", t, list[i].major) < 0)
1099 return log_oom();
1100 }
1101
1102 } else {
1103 assert(list[i].major != UINT_MAX); /* If a minor is specified, then a major also needs to be specified */
1104
1105 r = device_path_make_major_minor(list[i].type, makedev(list[i].major, list[i].minor), &pattern);
1106 if (r < 0)
1107 return log_oom();
1108 }
1109
1110 if (list[i].r)
1111 access[n++] = 'r';
1112 if (list[i].w)
1113 access[n++] = 'w';
1114 if (list[i].m)
1115 access[n++] = 'm';
1116 access[n] = 0;
1117
1118 assert(n > 0);
1119
1120 r = sd_bus_message_append(s->properties, "(ss)", pattern, access);
1121 if (r < 0)
1122 return bus_log_create_error(r);
1123 }
1124
1125 r = sd_bus_message_close_container(s->properties);
1126 if (r < 0)
1127 return bus_log_create_error(r);
1128
1129 r = sd_bus_message_close_container(s->properties);
1130 if (r < 0)
1131 return bus_log_create_error(r);
1132
1133 r = sd_bus_message_close_container(s->properties);
1134 if (r < 0)
1135 return bus_log_create_error(r);
1136
1137 return 0;
1138 }
1139
1140 static int oci_cgroup_memory_limit(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1141 uint64_t *m = ASSERT_PTR(userdata);
1142 uint64_t k;
1143
1144 if (json_variant_is_negative(v)) {
1145 *m = UINT64_MAX;
1146 return 0;
1147 }
1148
1149 if (!json_variant_is_unsigned(v))
1150 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1151 "Memory limit is not an unsigned integer");
1152
1153 k = json_variant_unsigned(v);
1154 if (k >= UINT64_MAX)
1155 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1156 "Memory limit too large: %" PRIu64, k);
1157
1158 *m = (uint64_t) k;
1159 return 0;
1160 }
1161
1162 static int oci_cgroup_memory(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1163
1164 struct memory_data {
1165 uint64_t limit;
1166 uint64_t reservation;
1167 uint64_t swap;
1168 } data = {
1169 .limit = UINT64_MAX,
1170 .reservation = UINT64_MAX,
1171 .swap = UINT64_MAX,
1172 };
1173
1174 static const JsonDispatch table[] = {
1175 { "limit", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, limit), 0 },
1176 { "reservation", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, reservation), 0 },
1177 { "swap", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, swap), 0 },
1178 { "kernel", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1179 { "kernelTCP", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1180 { "swapiness", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1181 { "disableOOMKiller", JSON_VARIANT_BOOLEAN, oci_unsupported, 0, JSON_PERMISSIVE },
1182 {}
1183 };
1184
1185 Settings *s = ASSERT_PTR(userdata);
1186 int r;
1187
1188 r = oci_dispatch(v, table, flags, &data);
1189 if (r < 0)
1190 return r;
1191
1192 if (data.swap != UINT64_MAX) {
1193 if (data.limit == UINT64_MAX)
1194 json_log(v, flags|LOG_WARNING, 0, "swap limit without memory limit is not supported, ignoring.");
1195 else if (data.swap < data.limit)
1196 json_log(v, flags|LOG_WARNING, 0, "swap limit is below memory limit, ignoring.");
1197 else {
1198 r = settings_allocate_properties(s);
1199 if (r < 0)
1200 return r;
1201
1202 r = sd_bus_message_append(s->properties, "(sv)", "MemorySwapMax", "t", data.swap - data.limit);
1203 if (r < 0)
1204 return bus_log_create_error(r);
1205 }
1206 }
1207
1208 if (data.limit != UINT64_MAX) {
1209 r = settings_allocate_properties(s);
1210 if (r < 0)
1211 return r;
1212
1213 r = sd_bus_message_append(s->properties, "(sv)", "MemoryMax", "t", data.limit);
1214 if (r < 0)
1215 return bus_log_create_error(r);
1216 }
1217
1218 if (data.reservation != UINT64_MAX) {
1219 r = settings_allocate_properties(s);
1220 if (r < 0)
1221 return r;
1222
1223 r = sd_bus_message_append(s->properties, "(sv)", "MemoryLow", "t", data.reservation);
1224 if (r < 0)
1225 return bus_log_create_error(r);
1226 }
1227
1228 return 0;
1229 }
1230
1231 struct cpu_data {
1232 uint64_t shares;
1233 uint64_t quota;
1234 uint64_t period;
1235 CPUSet cpu_set;
1236 };
1237
1238 static int oci_cgroup_cpu_shares(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1239 uint64_t *u = ASSERT_PTR(userdata);
1240 uint64_t k;
1241
1242 k = json_variant_unsigned(v);
1243 if (k < CGROUP_CPU_SHARES_MIN || k > CGROUP_CPU_SHARES_MAX)
1244 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1245 "shares value out of range.");
1246
1247 *u = (uint64_t) k;
1248 return 0;
1249 }
1250
1251 static int oci_cgroup_cpu_quota(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1252 uint64_t *u = ASSERT_PTR(userdata);
1253 uint64_t k;
1254
1255 k = json_variant_unsigned(v);
1256 if (k <= 0 || k >= UINT64_MAX)
1257 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1258 "period/quota value out of range.");
1259
1260 *u = (uint64_t) k;
1261 return 0;
1262 }
1263
1264 static int oci_cgroup_cpu_cpus(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1265 struct cpu_data *data = ASSERT_PTR(userdata);
1266 CPUSet set;
1267 const char *n;
1268 int r;
1269
1270 assert_se(n = json_variant_string(v));
1271
1272 r = parse_cpu_set(n, &set);
1273 if (r < 0)
1274 return json_log(v, flags, r, "Failed to parse CPU set specification: %s", n);
1275
1276 cpu_set_reset(&data->cpu_set);
1277 data->cpu_set = set;
1278
1279 return 0;
1280 }
1281
1282 static int oci_cgroup_cpu(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1283
1284 static const JsonDispatch table[] = {
1285 { "shares", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_shares, offsetof(struct cpu_data, shares), 0 },
1286 { "quota", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_quota, offsetof(struct cpu_data, quota), 0 },
1287 { "period", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_quota, offsetof(struct cpu_data, period), 0 },
1288 { "realtimeRuntime", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, 0 },
1289 { "realtimePeriod", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, 0 },
1290 { "cpus", JSON_VARIANT_STRING, oci_cgroup_cpu_cpus, 0, 0 },
1291 { "mems", JSON_VARIANT_STRING, oci_unsupported, 0, 0 },
1292 {}
1293 };
1294
1295 struct cpu_data data = {
1296 .shares = UINT64_MAX,
1297 .quota = UINT64_MAX,
1298 .period = UINT64_MAX,
1299 };
1300
1301 Settings *s = ASSERT_PTR(userdata);
1302 int r;
1303
1304 r = oci_dispatch(v, table, flags, &data);
1305 if (r < 0) {
1306 cpu_set_reset(&data.cpu_set);
1307 return r;
1308 }
1309
1310 cpu_set_reset(&s->cpu_set);
1311 s->cpu_set = data.cpu_set;
1312
1313 if (data.shares != UINT64_MAX) {
1314 r = settings_allocate_properties(s);
1315 if (r < 0)
1316 return r;
1317
1318 r = sd_bus_message_append(s->properties, "(sv)", "CPUShares", "t", data.shares);
1319 if (r < 0)
1320 return bus_log_create_error(r);
1321 }
1322
1323 if (data.quota != UINT64_MAX && data.period != UINT64_MAX) {
1324 r = settings_allocate_properties(s);
1325 if (r < 0)
1326 return r;
1327
1328 r = sd_bus_message_append(s->properties, "(sv)", "CPUQuotaPerSecUSec", "t", (uint64_t) (data.quota * USEC_PER_SEC / data.period));
1329 if (r < 0)
1330 return bus_log_create_error(r);
1331
1332 } else if ((data.quota != UINT64_MAX) != (data.period != UINT64_MAX))
1333 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1334 "CPU quota and period not used together.");
1335
1336 return 0;
1337 }
1338
1339 static int oci_cgroup_block_io_weight(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1340 Settings *s = ASSERT_PTR(userdata);
1341 uint64_t k;
1342 int r;
1343
1344 k = json_variant_unsigned(v);
1345 if (k < CGROUP_BLKIO_WEIGHT_MIN || k > CGROUP_BLKIO_WEIGHT_MAX)
1346 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1347 "Block I/O weight out of range.");
1348
1349 r = settings_allocate_properties(s);
1350 if (r < 0)
1351 return r;
1352
1353 r = sd_bus_message_append(s->properties, "(sv)", "BlockIOWeight", "t", (uint64_t) k);
1354 if (r < 0)
1355 return bus_log_create_error(r);
1356
1357 return 0;
1358 }
1359
1360 static int oci_cgroup_block_io_weight_device(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1361 Settings *s = ASSERT_PTR(userdata);
1362 JsonVariant *e;
1363 int r;
1364
1365 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1366 struct device_data {
1367 unsigned major;
1368 unsigned minor;
1369 uint64_t weight;
1370 } data = {
1371 .major = UINT_MAX,
1372 .minor = UINT_MAX,
1373 .weight = UINT64_MAX,
1374 };
1375
1376 static const JsonDispatch table[] = {
1377 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), JSON_MANDATORY },
1378 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), JSON_MANDATORY },
1379 { "weight", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct device_data, weight), 0 },
1380 { "leafWeight", JSON_VARIANT_INTEGER, oci_unsupported, 0, JSON_PERMISSIVE },
1381 {}
1382 };
1383
1384 _cleanup_free_ char *path = NULL;
1385
1386 r = oci_dispatch(e, table, flags, &data);
1387 if (r < 0)
1388 return r;
1389
1390 if (data.weight == UINT64_MAX)
1391 continue;
1392
1393 if (data.weight < CGROUP_BLKIO_WEIGHT_MIN || data.weight > CGROUP_BLKIO_WEIGHT_MAX)
1394 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1395 "Block I/O device weight out of range.");
1396
1397 r = device_path_make_major_minor(S_IFBLK, makedev(data.major, data.minor), &path);
1398 if (r < 0)
1399 return json_log(v, flags, r, "Failed to build device path: %m");
1400
1401 r = settings_allocate_properties(s);
1402 if (r < 0)
1403 return r;
1404
1405 r = sd_bus_message_append(s->properties, "(sv)", "BlockIODeviceWeight", "a(st)", 1, path, (uint64_t) data.weight);
1406 if (r < 0)
1407 return bus_log_create_error(r);
1408 }
1409
1410 return 0;
1411 }
1412
1413 static int oci_cgroup_block_io_throttle(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1414 Settings *s = ASSERT_PTR(userdata);
1415 const char *pname;
1416 JsonVariant *e;
1417 int r;
1418
1419 pname = streq(name, "throttleReadBpsDevice") ? "IOReadBandwidthMax" :
1420 streq(name, "throttleWriteBpsDevice") ? "IOWriteBandwidthMax" :
1421 streq(name, "throttleReadIOPSDevice") ? "IOReadIOPSMax" :
1422 "IOWriteIOPSMax";
1423
1424 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1425 struct device_data {
1426 unsigned major;
1427 unsigned minor;
1428 uint64_t rate;
1429 } data = {
1430 .major = UINT_MAX,
1431 .minor = UINT_MAX,
1432 };
1433
1434 static const JsonDispatch table[] = {
1435 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), JSON_MANDATORY },
1436 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), JSON_MANDATORY },
1437 { "rate", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct device_data, rate), JSON_MANDATORY },
1438 {}
1439 };
1440
1441 _cleanup_free_ char *path = NULL;
1442
1443 r = oci_dispatch(e, table, flags, &data);
1444 if (r < 0)
1445 return r;
1446
1447 if (data.rate >= UINT64_MAX)
1448 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1449 "Block I/O device rate out of range.");
1450
1451 r = device_path_make_major_minor(S_IFBLK, makedev(data.major, data.minor), &path);
1452 if (r < 0)
1453 return json_log(v, flags, r, "Failed to build device path: %m");
1454
1455 r = settings_allocate_properties(s);
1456 if (r < 0)
1457 return r;
1458
1459 r = sd_bus_message_append(s->properties, "(sv)", pname, "a(st)", 1, path, (uint64_t) data.rate);
1460 if (r < 0)
1461 return bus_log_create_error(r);
1462 }
1463
1464 return 0;
1465 }
1466
1467 static int oci_cgroup_block_io(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1468
1469 static const JsonDispatch table[] = {
1470 { "weight", JSON_VARIANT_UNSIGNED, oci_cgroup_block_io_weight, 0, 0 },
1471 { "leafWeight", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, JSON_PERMISSIVE },
1472 { "weightDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_weight_device, 0, 0 },
1473 { "throttleReadBpsDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1474 { "throttleWriteBpsDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1475 { "throttleReadIOPSDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1476 { "throttleWriteIOPSDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1477 {}
1478 };
1479
1480 return oci_dispatch(v, table, flags, userdata);
1481 }
1482
1483 static int oci_cgroup_pids(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1484
1485 static const JsonDispatch table[] = {
1486 { "limit", JSON_VARIANT_NUMBER, json_dispatch_variant, 0, JSON_MANDATORY },
1487 {}
1488 };
1489
1490 _cleanup_(json_variant_unrefp) JsonVariant *k = NULL;
1491 Settings *s = ASSERT_PTR(userdata);
1492 uint64_t m;
1493 int r;
1494
1495 r = oci_dispatch(v, table, flags, &k);
1496 if (r < 0)
1497 return r;
1498
1499 if (json_variant_is_negative(k))
1500 m = UINT64_MAX;
1501 else {
1502 if (!json_variant_is_unsigned(k))
1503 return json_log(k, flags, SYNTHETIC_ERRNO(EINVAL),
1504 "pids limit not unsigned integer, refusing.");
1505
1506 m = (uint64_t) json_variant_unsigned(k);
1507
1508 if ((uint64_t) m != json_variant_unsigned(k))
1509 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1510 "pids limit out of range, refusing.");
1511 }
1512
1513 r = settings_allocate_properties(s);
1514 if (r < 0)
1515 return r;
1516
1517 r = sd_bus_message_append(s->properties, "(sv)", "TasksMax", "t", m);
1518 if (r < 0)
1519 return bus_log_create_error(r);
1520
1521 return 0;
1522 }
1523
1524 static int oci_resources(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1525
1526 static const JsonDispatch table[] = {
1527 { "devices", JSON_VARIANT_ARRAY, oci_cgroup_devices, 0, 0 },
1528 { "memory", JSON_VARIANT_OBJECT, oci_cgroup_memory, 0, 0 },
1529 { "cpu", JSON_VARIANT_OBJECT, oci_cgroup_cpu, 0, 0 },
1530 { "blockIO", JSON_VARIANT_OBJECT, oci_cgroup_block_io, 0, 0 },
1531 { "hugepageLimits", JSON_VARIANT_ARRAY, oci_unsupported, 0, 0 },
1532 { "network", JSON_VARIANT_OBJECT, oci_unsupported, 0, 0 },
1533 { "pids", JSON_VARIANT_OBJECT, oci_cgroup_pids, 0, 0 },
1534 { "rdma", JSON_VARIANT_OBJECT, oci_unsupported, 0, 0 },
1535 {}
1536 };
1537
1538 return oci_dispatch(v, table, flags, userdata);
1539 }
1540
1541 static bool sysctl_key_valid(const char *s) {
1542 bool dot = true;
1543
1544 /* Note that we are a bit stricter here than in systemd-sysctl, as that inherited semantics from the old sysctl
1545 * tool, which were really weird (as it swaps / and . in both ways) */
1546
1547 if (isempty(s))
1548 return false;
1549
1550 for (; *s; s++) {
1551
1552 if (*s <= ' ' || *s >= 127)
1553 return false;
1554 if (*s == '/')
1555 return false;
1556 if (*s == '.') {
1557
1558 if (dot) /* Don't allow two dots next to each other (or at the beginning) */
1559 return false;
1560
1561 dot = true;
1562 } else
1563 dot = false;
1564 }
1565
1566 if (dot) /* don't allow a dot at the end */
1567 return false;
1568
1569 return true;
1570 }
1571
1572 static int oci_sysctl(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1573 Settings *s = ASSERT_PTR(userdata);
1574 JsonVariant *w;
1575 const char *k;
1576 int r;
1577
1578 JSON_VARIANT_OBJECT_FOREACH(k, w, v) {
1579 const char *m;
1580
1581 if (!json_variant_is_string(w))
1582 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1583 "sysctl parameter is not a string, refusing.");
1584
1585 assert_se(m = json_variant_string(w));
1586
1587 if (!sysctl_key_valid(k))
1588 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1589 "sysctl key invalid, refusing: %s", k);
1590
1591 r = strv_extend_strv(&s->sysctl, STRV_MAKE(k, m), false);
1592 if (r < 0)
1593 return log_oom();
1594 }
1595
1596 return 0;
1597 }
1598
1599 #if HAVE_SECCOMP
1600 static int oci_seccomp_action_from_string(const char *name, uint32_t *ret) {
1601
1602 static const struct {
1603 const char *name;
1604 uint32_t action;
1605 } table[] = {
1606 { "SCMP_ACT_ALLOW", SCMP_ACT_ALLOW },
1607 { "SCMP_ACT_ERRNO", SCMP_ACT_ERRNO(EPERM) }, /* the OCI spec doesn't document the error, but it appears EPERM is supposed to be used */
1608 { "SCMP_ACT_KILL", SCMP_ACT_KILL },
1609 #ifdef SCMP_ACT_KILL_PROCESS
1610 { "SCMP_ACT_KILL_PROCESS", SCMP_ACT_KILL_PROCESS },
1611 #endif
1612 #ifdef SCMP_ACT_KILL_THREAD
1613 { "SCMP_ACT_KILL_THREAD", SCMP_ACT_KILL_THREAD },
1614 #endif
1615 #ifdef SCMP_ACT_LOG
1616 { "SCMP_ACT_LOG", SCMP_ACT_LOG },
1617 #endif
1618 { "SCMP_ACT_TRAP", SCMP_ACT_TRAP },
1619
1620 /* We don't support SCMP_ACT_TRACE because that requires a tracer, and that doesn't really make sense
1621 * here */
1622 };
1623
1624 size_t i;
1625
1626 for (i = 0; i < ELEMENTSOF(table); i++)
1627 if (streq_ptr(name, table[i].name)) {
1628 *ret = table[i].action;
1629 return 0;
1630 }
1631
1632 return -EINVAL;
1633 }
1634
1635 static int oci_seccomp_arch_from_string(const char *name, uint32_t *ret) {
1636
1637 static const struct {
1638 const char *name;
1639 uint32_t arch;
1640 } table[] = {
1641 { "SCMP_ARCH_AARCH64", SCMP_ARCH_AARCH64 },
1642 { "SCMP_ARCH_ARM", SCMP_ARCH_ARM },
1643 #ifdef SCMP_ARCH_LOONGARCH64
1644 { "SCMP_ARCH_LOONGARCH64", SCMP_ARCH_LOONGARCH64 },
1645 #endif
1646 { "SCMP_ARCH_MIPS", SCMP_ARCH_MIPS },
1647 { "SCMP_ARCH_MIPS64", SCMP_ARCH_MIPS64 },
1648 { "SCMP_ARCH_MIPS64N32", SCMP_ARCH_MIPS64N32 },
1649 { "SCMP_ARCH_MIPSEL", SCMP_ARCH_MIPSEL },
1650 { "SCMP_ARCH_MIPSEL64", SCMP_ARCH_MIPSEL64 },
1651 { "SCMP_ARCH_MIPSEL64N32", SCMP_ARCH_MIPSEL64N32 },
1652 { "SCMP_ARCH_NATIVE", SCMP_ARCH_NATIVE },
1653 #ifdef SCMP_ARCH_PARISC
1654 { "SCMP_ARCH_PARISC", SCMP_ARCH_PARISC },
1655 #endif
1656 #ifdef SCMP_ARCH_PARISC64
1657 { "SCMP_ARCH_PARISC64", SCMP_ARCH_PARISC64 },
1658 #endif
1659 { "SCMP_ARCH_PPC", SCMP_ARCH_PPC },
1660 { "SCMP_ARCH_PPC64", SCMP_ARCH_PPC64 },
1661 { "SCMP_ARCH_PPC64LE", SCMP_ARCH_PPC64LE },
1662 #ifdef SCMP_ARCH_RISCV64
1663 { "SCMP_ARCH_RISCV64", SCMP_ARCH_RISCV64 },
1664 #endif
1665 { "SCMP_ARCH_S390", SCMP_ARCH_S390 },
1666 { "SCMP_ARCH_S390X", SCMP_ARCH_S390X },
1667 { "SCMP_ARCH_X32", SCMP_ARCH_X32 },
1668 { "SCMP_ARCH_X86", SCMP_ARCH_X86 },
1669 { "SCMP_ARCH_X86_64", SCMP_ARCH_X86_64 },
1670 };
1671
1672 size_t i;
1673
1674 for (i = 0; i < ELEMENTSOF(table); i++)
1675 if (streq_ptr(table[i].name, name)) {
1676 *ret = table[i].arch;
1677 return 0;
1678 }
1679
1680 return -EINVAL;
1681 }
1682
1683 static int oci_seccomp_compare_from_string(const char *name, enum scmp_compare *ret) {
1684
1685 static const struct {
1686 const char *name;
1687 enum scmp_compare op;
1688 } table[] = {
1689 { "SCMP_CMP_NE", SCMP_CMP_NE },
1690 { "SCMP_CMP_LT", SCMP_CMP_LT },
1691 { "SCMP_CMP_LE", SCMP_CMP_LE },
1692 { "SCMP_CMP_EQ", SCMP_CMP_EQ },
1693 { "SCMP_CMP_GE", SCMP_CMP_GE },
1694 { "SCMP_CMP_GT", SCMP_CMP_GT },
1695 { "SCMP_CMP_MASKED_EQ", SCMP_CMP_MASKED_EQ },
1696 };
1697
1698 size_t i;
1699
1700 for (i = 0; i < ELEMENTSOF(table); i++)
1701 if (streq_ptr(table[i].name, name)) {
1702 *ret = table[i].op;
1703 return 0;
1704 }
1705
1706 return -EINVAL;
1707 }
1708
1709 static int oci_seccomp_archs(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1710 scmp_filter_ctx *sc = ASSERT_PTR(userdata);
1711 JsonVariant *e;
1712 int r;
1713
1714 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1715 uint32_t a;
1716
1717 if (!json_variant_is_string(e))
1718 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
1719 "Architecture entry is not a string");
1720
1721 r = oci_seccomp_arch_from_string(json_variant_string(e), &a);
1722 if (r < 0)
1723 return json_log(e, flags, r, "Unknown architecture: %s", json_variant_string(e));
1724
1725 r = seccomp_arch_add(sc, a);
1726 if (r == -EEXIST)
1727 continue;
1728 if (r < 0)
1729 return json_log(e, flags, r, "Failed to add architecture to seccomp filter: %m");
1730 }
1731
1732 return 0;
1733 }
1734
1735 struct syscall_rule {
1736 char **names;
1737 uint32_t action;
1738 struct scmp_arg_cmp *arguments;
1739 size_t n_arguments;
1740 };
1741
1742 static void syscall_rule_done(struct syscall_rule *rule) {
1743 assert(rule);
1744
1745 strv_free(rule->names);
1746 free(rule->arguments);
1747 };
1748
1749 static int oci_seccomp_action(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1750 uint32_t *action = ASSERT_PTR(userdata);
1751 int r;
1752
1753 r = oci_seccomp_action_from_string(json_variant_string(v), action);
1754 if (r < 0)
1755 return json_log(v, flags, r, "Unknown system call action '%s': %m", json_variant_string(v));
1756
1757 return 0;
1758 }
1759
1760 static int oci_seccomp_op(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1761 enum scmp_compare *op = ASSERT_PTR(userdata);
1762 int r;
1763
1764 r = oci_seccomp_compare_from_string(json_variant_string(v), op);
1765 if (r < 0)
1766 return json_log(v, flags, r, "Unknown seccomp operator '%s': %m", json_variant_string(v));
1767
1768 return 0;
1769 }
1770
1771 static int oci_seccomp_args(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1772 struct syscall_rule *rule = ASSERT_PTR(userdata);
1773 JsonVariant *e;
1774 int r;
1775
1776 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1777 static const struct JsonDispatch table[] = {
1778 { "index", JSON_VARIANT_UNSIGNED, json_dispatch_uint32, offsetof(struct scmp_arg_cmp, arg), JSON_MANDATORY },
1779 { "value", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct scmp_arg_cmp, datum_a), JSON_MANDATORY },
1780 { "valueTwo", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct scmp_arg_cmp, datum_b), 0 },
1781 { "op", JSON_VARIANT_STRING, oci_seccomp_op, offsetof(struct scmp_arg_cmp, op), JSON_MANDATORY },
1782 {},
1783 };
1784
1785 struct scmp_arg_cmp *a, *p;
1786 int expected;
1787
1788 a = reallocarray(rule->arguments, rule->n_arguments + 1, sizeof(struct syscall_rule));
1789 if (!a)
1790 return log_oom();
1791
1792 rule->arguments = a;
1793 p = rule->arguments + rule->n_arguments;
1794
1795 *p = (struct scmp_arg_cmp) {
1796 .arg = 0,
1797 .datum_a = 0,
1798 .datum_b = 0,
1799 .op = 0,
1800 };
1801
1802 r = oci_dispatch(e, table, flags, p);
1803 if (r < 0)
1804 return r;
1805
1806 expected = p->op == SCMP_CMP_MASKED_EQ ? 4 : 3;
1807 if (r != expected)
1808 json_log(e, flags|JSON_WARNING, 0, "Wrong number of system call arguments for JSON data, ignoring.");
1809
1810 /* Note that we are a bit sloppy here and do not insist that SCMP_CMP_MASKED_EQ gets two datum values,
1811 * and the other only one. That's because buildah for example by default calls things with
1812 * SCMP_CMP_MASKED_EQ but only one argument. We use 0 when the value is not specified. */
1813
1814 rule->n_arguments++;
1815 }
1816
1817 return 0;
1818 }
1819
1820 static int oci_seccomp_syscalls(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1821 scmp_filter_ctx *sc = ASSERT_PTR(userdata);
1822 JsonVariant *e;
1823 int r;
1824
1825 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1826 static const JsonDispatch table[] = {
1827 { "names", JSON_VARIANT_ARRAY, json_dispatch_strv, offsetof(struct syscall_rule, names), JSON_MANDATORY },
1828 { "action", JSON_VARIANT_STRING, oci_seccomp_action, offsetof(struct syscall_rule, action), JSON_MANDATORY },
1829 { "args", JSON_VARIANT_ARRAY, oci_seccomp_args, 0, 0 },
1830 {}
1831 };
1832 _cleanup_(syscall_rule_done) struct syscall_rule rule = {
1833 .action = UINT32_MAX,
1834 };
1835
1836 r = oci_dispatch(e, table, flags, &rule);
1837 if (r < 0)
1838 return r;
1839
1840 if (strv_isempty(rule.names)) {
1841 json_log(e, flags, 0, "System call name list is empty.");
1842 return -EINVAL;
1843 }
1844
1845 STRV_FOREACH(i, rule.names) {
1846 int nr;
1847
1848 nr = seccomp_syscall_resolve_name(*i);
1849 if (nr == __NR_SCMP_ERROR) {
1850 log_debug("Unknown syscall %s, skipping.", *i);
1851 continue;
1852 }
1853
1854 r = seccomp_rule_add_array(sc, rule.action, nr, rule.n_arguments, rule.arguments);
1855 if (r < 0)
1856 return r;
1857 }
1858 }
1859
1860 return 0;
1861 }
1862 #endif
1863
1864 static int oci_seccomp(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1865
1866 #if HAVE_SECCOMP
1867 static const JsonDispatch table[] = {
1868 { "defaultAction", JSON_VARIANT_STRING, NULL, 0, JSON_MANDATORY },
1869 { "architectures", JSON_VARIANT_ARRAY, oci_seccomp_archs, 0, 0 },
1870 { "syscalls", JSON_VARIANT_ARRAY, oci_seccomp_syscalls, 0, 0 },
1871 {}
1872 };
1873
1874 _cleanup_(seccomp_releasep) scmp_filter_ctx sc = NULL;
1875 Settings *s = ASSERT_PTR(userdata);
1876 JsonVariant *def;
1877 uint32_t d;
1878 int r;
1879
1880 def = json_variant_by_key(v, "defaultAction");
1881 if (!def)
1882 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL), "defaultAction element missing.");
1883
1884 if (!json_variant_is_string(def))
1885 return json_log(def, flags, SYNTHETIC_ERRNO(EINVAL), "defaultAction is not a string.");
1886
1887 r = oci_seccomp_action_from_string(json_variant_string(def), &d);
1888 if (r < 0)
1889 return json_log(def, flags, r, "Unknown default action: %s", json_variant_string(def));
1890
1891 sc = seccomp_init(d);
1892 if (!sc)
1893 return json_log(v, flags, SYNTHETIC_ERRNO(ENOMEM), "Couldn't allocate seccomp object.");
1894
1895 r = oci_dispatch(v, table, flags, sc);
1896 if (r < 0)
1897 return r;
1898
1899 seccomp_release(s->seccomp);
1900 s->seccomp = TAKE_PTR(sc);
1901 return 0;
1902 #else
1903 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP), "libseccomp support not enabled, can't parse seccomp object.");
1904 #endif
1905 }
1906
1907 static int oci_rootfs_propagation(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1908 const char *s;
1909
1910 s = json_variant_string(v);
1911
1912 if (streq(s, "shared"))
1913 return 0;
1914
1915 json_log(v, flags|JSON_DEBUG, 0, "Ignoring rootfsPropagation setting '%s'.", s);
1916 return 0;
1917 }
1918
1919 static int oci_masked_paths(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1920 Settings *s = ASSERT_PTR(userdata);
1921 JsonVariant *e;
1922
1923 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1924 _cleanup_free_ char *destination = NULL;
1925 CustomMount *m;
1926 const char *p;
1927
1928 if (!json_variant_is_string(e))
1929 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1930 "Path is not a string, refusing.");
1931
1932 assert_se(p = json_variant_string(e));
1933
1934 if (!path_is_absolute(p))
1935 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1936 "Path is not absolute, refusing: %s", p);
1937
1938 if (oci_exclude_mount(p))
1939 continue;
1940
1941 destination = strdup(p);
1942 if (!destination)
1943 return log_oom();
1944
1945 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_INACCESSIBLE);
1946 if (!m)
1947 return log_oom();
1948
1949 m->destination = TAKE_PTR(destination);
1950
1951 /* The spec doesn't say this, but apparently pre-existing implementations are lenient towards
1952 * non-existing paths to mask. Let's hence be too. */
1953 m->graceful = true;
1954 }
1955
1956 return 0;
1957 }
1958
1959 static int oci_readonly_paths(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1960 Settings *s = ASSERT_PTR(userdata);
1961 JsonVariant *e;
1962
1963 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1964 _cleanup_free_ char *source = NULL, *destination = NULL;
1965 CustomMount *m;
1966 const char *p;
1967
1968 if (!json_variant_is_string(e))
1969 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1970 "Path is not a string, refusing.");
1971
1972 assert_se(p = json_variant_string(e));
1973
1974 if (!path_is_absolute(p))
1975 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1976 "Path is not absolute, refusing: %s", p);
1977
1978 if (oci_exclude_mount(p))
1979 continue;
1980
1981 source = strjoin("+", p);
1982 if (!source)
1983 return log_oom();
1984
1985 destination = strdup(p);
1986 if (!destination)
1987 return log_oom();
1988
1989 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_BIND);
1990 if (!m)
1991 return log_oom();
1992
1993 m->source = TAKE_PTR(source);
1994 m->destination = TAKE_PTR(destination);
1995 m->read_only = true;
1996 }
1997
1998 return 0;
1999 }
2000
2001 static int oci_linux(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2002
2003 static const JsonDispatch table[] = {
2004 { "namespaces", JSON_VARIANT_ARRAY, oci_namespaces, 0, 0 },
2005 { "uidMappings", JSON_VARIANT_ARRAY, oci_uid_gid_mappings, 0, 0 },
2006 { "gidMappings", JSON_VARIANT_ARRAY, oci_uid_gid_mappings, 0, 0 },
2007 { "devices", JSON_VARIANT_ARRAY, oci_devices, 0, 0 },
2008 { "cgroupsPath", JSON_VARIANT_STRING, oci_cgroups_path, 0, 0 },
2009 { "resources", JSON_VARIANT_OBJECT, oci_resources, 0, 0 },
2010 { "intelRdt", JSON_VARIANT_OBJECT, oci_unsupported, 0, JSON_PERMISSIVE },
2011 { "sysctl", JSON_VARIANT_OBJECT, oci_sysctl, 0, 0 },
2012 { "seccomp", JSON_VARIANT_OBJECT, oci_seccomp, 0, 0 },
2013 { "rootfsPropagation", JSON_VARIANT_STRING, oci_rootfs_propagation, 0, 0 },
2014 { "maskedPaths", JSON_VARIANT_ARRAY, oci_masked_paths, 0, 0 },
2015 { "readonlyPaths", JSON_VARIANT_ARRAY, oci_readonly_paths, 0, 0 },
2016 { "mountLabel", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
2017 {}
2018 };
2019
2020 return oci_dispatch(v, table, flags, userdata);
2021 }
2022
2023 static int oci_hook_timeout(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2024 usec_t *u = ASSERT_PTR(userdata);
2025 uint64_t k;
2026
2027 k = json_variant_unsigned(v);
2028 if (k == 0 || k > (UINT64_MAX-1)/USEC_PER_SEC)
2029 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
2030 "Hook timeout value out of range.");
2031
2032 *u = k * USEC_PER_SEC;
2033 return 0;
2034 }
2035
2036 static int oci_hooks_array(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2037 Settings *s = ASSERT_PTR(userdata);
2038 JsonVariant *e;
2039 int r;
2040
2041 JSON_VARIANT_ARRAY_FOREACH(e, v) {
2042
2043 static const JsonDispatch table[] = {
2044 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(OciHook, path), JSON_MANDATORY },
2045 { "args", JSON_VARIANT_ARRAY, oci_args, offsetof(OciHook, args), 0 },
2046 { "env", JSON_VARIANT_ARRAY, oci_env, offsetof(OciHook, env), 0 },
2047 { "timeout", JSON_VARIANT_UNSIGNED, oci_hook_timeout, offsetof(OciHook, timeout), 0 },
2048 {}
2049 };
2050
2051 OciHook *a, **array, *new_item;
2052 size_t *n_array;
2053
2054 if (streq(name, "prestart")) {
2055 array = &s->oci_hooks_prestart;
2056 n_array = &s->n_oci_hooks_prestart;
2057 } else if (streq(name, "poststart")) {
2058 array = &s->oci_hooks_poststart;
2059 n_array = &s->n_oci_hooks_poststart;
2060 } else {
2061 assert(streq(name, "poststop"));
2062 array = &s->oci_hooks_poststop;
2063 n_array = &s->n_oci_hooks_poststop;
2064 }
2065
2066 a = reallocarray(*array, *n_array + 1, sizeof(OciHook));
2067 if (!a)
2068 return log_oom();
2069
2070 *array = a;
2071 new_item = a + *n_array;
2072
2073 *new_item = (OciHook) {
2074 .timeout = USEC_INFINITY,
2075 };
2076
2077 r = oci_dispatch(e, table, flags, new_item);
2078 if (r < 0) {
2079 free(new_item->path);
2080 strv_free(new_item->args);
2081 strv_free(new_item->env);
2082 return r;
2083 }
2084
2085 (*n_array) ++;
2086 }
2087
2088 return 0;
2089 }
2090
2091 static int oci_hooks(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2092
2093 static const JsonDispatch table[] = {
2094 { "prestart", JSON_VARIANT_ARRAY, oci_hooks_array, 0, 0 },
2095 { "poststart", JSON_VARIANT_ARRAY, oci_hooks_array, 0, 0 },
2096 { "poststop", JSON_VARIANT_ARRAY, oci_hooks_array, 0, 0 },
2097 {}
2098 };
2099
2100 return oci_dispatch(v, table, flags, userdata);
2101 }
2102
2103 static int oci_annotations(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2104 JsonVariant *w;
2105 const char *k;
2106
2107 JSON_VARIANT_OBJECT_FOREACH(k, w, v) {
2108
2109 if (isempty(k))
2110 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
2111 "Annotation with empty key, refusing.");
2112
2113 if (!json_variant_is_string(w))
2114 return json_log(w, flags, SYNTHETIC_ERRNO(EINVAL),
2115 "Annotation has non-string value, refusing.");
2116
2117 json_log(w, flags|JSON_DEBUG, 0, "Ignoring annotation '%s' with value '%s'.", k, json_variant_string(w));
2118 }
2119
2120 return 0;
2121 }
2122
2123 int oci_load(FILE *f, const char *bundle, Settings **ret) {
2124
2125 static const JsonDispatch table[] = {
2126 { "ociVersion", JSON_VARIANT_STRING, NULL, 0, JSON_MANDATORY },
2127 { "process", JSON_VARIANT_OBJECT, oci_process, 0, 0 },
2128 { "root", JSON_VARIANT_OBJECT, oci_root, 0, 0 },
2129 { "hostname", JSON_VARIANT_STRING, oci_hostname, 0, 0 },
2130 { "mounts", JSON_VARIANT_ARRAY, oci_mounts, 0, 0 },
2131 { "linux", JSON_VARIANT_OBJECT, oci_linux, 0, 0 },
2132 { "hooks", JSON_VARIANT_OBJECT, oci_hooks, 0, 0 },
2133 { "annotations", JSON_VARIANT_OBJECT, oci_annotations, 0, 0 },
2134 {}
2135 };
2136
2137 _cleanup_(json_variant_unrefp) JsonVariant *oci = NULL;
2138 _cleanup_(settings_freep) Settings *s = NULL;
2139 unsigned line = 0, column = 0;
2140 JsonVariant *v;
2141 const char *path;
2142 int r;
2143
2144 assert_se(bundle);
2145
2146 path = strjoina(bundle, "/config.json");
2147
2148 r = json_parse_file(f, path, 0, &oci, &line, &column);
2149 if (r < 0) {
2150 if (line != 0 && column != 0)
2151 return log_error_errno(r, "Failed to parse '%s' at %u:%u: %m", path, line, column);
2152 else
2153 return log_error_errno(r, "Failed to parse '%s': %m", path);
2154 }
2155
2156 v = json_variant_by_key(oci, "ociVersion");
2157 if (!v)
2158 return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
2159 "JSON file '%s' is not an OCI bundle configuration file. Refusing.",
2160 path);
2161 if (!streq_ptr(json_variant_string(v), "1.0.0"))
2162 return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
2163 "OCI bundle version not supported: %s",
2164 strna(json_variant_string(v)));
2165
2166 // {
2167 // _cleanup_free_ char *formatted = NULL;
2168 // assert_se(json_variant_format(oci, JSON_FORMAT_PRETTY|JSON_FORMAT_COLOR, &formatted) >= 0);
2169 // fputs(formatted, stdout);
2170 // }
2171
2172 s = settings_new();
2173 if (!s)
2174 return log_oom();
2175
2176 s->start_mode = START_PID1;
2177 s->resolv_conf = RESOLV_CONF_OFF;
2178 s->link_journal = LINK_NO;
2179 s->timezone = TIMEZONE_OFF;
2180
2181 s->bundle = strdup(bundle);
2182 if (!s->bundle)
2183 return log_oom();
2184
2185 r = oci_dispatch(oci, table, 0, s);
2186 if (r < 0)
2187 return r;
2188
2189 if (s->properties) {
2190 r = sd_bus_message_seal(s->properties, 0, 0);
2191 if (r < 0)
2192 return log_error_errno(r, "Cannot seal properties bus message: %m");
2193 }
2194
2195 *ret = TAKE_PTR(s);
2196 return 0;
2197 }
A System and Session Manager (ALT Linux package)