]> 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
Merge pull request #13899 from poettering/in-gid-tweak
[thirdparty/systemd.git] / src / nspawn / nspawn-oci.c
CommitLineData
de40a303
LP		 1/* SPDX-License-Identifier: LGPL-2.1+ */
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"
11#include "env-util.h"
ca78ad1d 12#include "format-util.h"
de40a303
LP		 13#include "fs-util.h"
14#include "hostname-util.h"
15#include "json.h"
16#include "missing_sched.h"
17#include "nspawn-oci.h"
18#include "path-util.h"
19#include "rlimit-util.h"
ae408d77 20#if HAVE_SECCOMP
de40a303 21#include "seccomp-util.h"
ae408d77 22#endif
de40a303
LP		 23#include "stat-util.h"
24#include "stdio-util.h"
25#include "string-util.h"
26#include "strv.h"
27#include "user-util.h"
28
29/* TODO:
30 * OCI runtime tool implementation
31 * hooks
32 *
33 * Spec issues:
34 *
35 * How is RLIM_INFINITY supposed to be encoded?
36 * configured effective caps is bullshit, as execv() corrupts it anyway
37 * pipes bind mounted is *very* different from pipes newly created, comments regarding bind mount or not are bogus
38 * annotation values structured? or string?
39 * configurable file system namespace path, but then also root path? wtf?
40 * apply sysctl inside of the container? or outside?
41 * how is unlimited pids tasks limit to be encoded?
42 * what are the defaults for caps if not specified?
43 * 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
44 * 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
45 * 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
46 * if type of mount is left out, what shall be assumed? "bind"?
47 * readonly mounts is entirely redundant?
48 * should escaping be applied when joining mount options with ","?
49 * devices cgroup support is bogus, "allow" and "deny" on the kernel level is about adding/removing entries, not about access
50 * spec needs to say that "rwm" devices cgroup combination can't be the empty string
51 * cgrouspv1 crap: kernel, kernelTCP, swapiness, disableOOMKiller, swap, devices, leafWeight
52 * general: it shouldn't leak lower level abstractions this obviously
53 * unmanagable cgroups stuff: realtimeRuntime/realtimePeriod
54 * needs to say what happense when some option is not specified, i.e. which defautls apply
55 * no architecture? no personality?
56 * seccomp example and logic is simply broken: there's no constant "SCMP_ACT_ERRNO".
57 * spec should say what to do with unknown props
58 * /bin/mount regarding NFS and FUSE required?
59 * what does terminal=false mean?
60 * sysctl inside or outside? whitelisting?
5238e957 61 * swapiness typo -> swappiness
de40a303
LP		 62 *
63 * Unsupported:
64 *
65 * apparmorProfile
66 * selinuxLabel + mountLabel
67 * hugepageLimits
68 * network
69 * rdma
70 * intelRdt
71 * swappiness, disableOOMKiller, kernel, kernelTCP, leafWeight (because it's dead, cgroupsv2 can't do it and hence systemd neither)
72 *
73 * Non-slice cgroup paths
74 * Propagation that is not slave + shared
75 * more than one uid/gid mapping, mappings with a container base != 0, or non-matching uid/gid mappings
76 * device cgroups access = false items that are not catchall
77 * 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.
78 *
79 */
80
81static int oci_unexpected(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
19130626
ZJS		 82 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
83 "Unexpected OCI element '%s' of type '%s'.", name, json_variant_type_to_string(json_variant_type(v)));
de40a303
LP		 84}
85
86static int oci_unsupported(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
19130626
ZJS		 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)));
de40a303
LP		 89}
90
91static int oci_terminal(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
92 Settings *s = userdata;
93
94 /* If not specified, or set to true, we'll default to either an interactive or a read-only
6757a013 95 * console. If specified as false, we'll forcibly move to "pipe" mode though. */
de40a303
LP		 96 s->console_mode = json_variant_boolean(v) ? _CONSOLE_MODE_INVALID : CONSOLE_PIPE;
97 return 0;
98}
99
100static int oci_console_dimension(const char *name, JsonVariant *variant, JsonDispatchFlags flags, void *userdata) {
101 unsigned *u = userdata;
102 uintmax_t k;
103
104 assert(u);
105
106 k = json_variant_unsigned(variant);
19130626
ZJS		 107 if (k == 0)
108 return json_log(variant, flags, SYNTHETIC_ERRNO(ERANGE),
109 "Console size field '%s' is too small.", strna(name));
110 if (k > USHRT_MAX) /* TIOCSWINSZ's struct winsize uses "unsigned short" for width and height */
111 return json_log(variant, flags, SYNTHETIC_ERRNO(ERANGE),
112 "Console size field '%s' is too large.", strna(name));
de40a303
LP		 113
114 *u = (unsigned) k;
115 return 0;
116}
117
118static int oci_console_size(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
119
120 static const JsonDispatch table[] = {
121 { "height", JSON_VARIANT_UNSIGNED, oci_console_dimension, offsetof(Settings, console_height), JSON_MANDATORY },
122 { "width", JSON_VARIANT_UNSIGNED, oci_console_dimension, offsetof(Settings, console_width), JSON_MANDATORY },
123 {}
124 };
125
126 return json_dispatch(v, table, oci_unexpected, flags, userdata);
127}
128
129static int oci_absolute_path(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
130 char **p = userdata;
131 const char *n;
de40a303
LP		 132
133 assert(p);
134
135 n = json_variant_string(v);
136
19130626
ZJS		 137 if (!path_is_absolute(n))
138 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
139 "Path in JSON field '%s' is not absolute: %s", strna(name), n);
de40a303 140
f1531db5 141 return free_and_strdup_warn(p, n);
de40a303
LP		 142}
143
144static int oci_env(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
145 char ***l = userdata;
146 JsonVariant *e;
147 int r;
148
149 assert(l);
150
151 JSON_VARIANT_ARRAY_FOREACH(e, v) {
152 const char *n;
153
19130626
ZJS		 154 if (!json_variant_is_string(e))
155 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
156 "Environment array contains non-string.");
de40a303
LP		 157
158 assert_se(n = json_variant_string(e));
159
19130626
ZJS		 160 if (!env_assignment_is_valid(n))
161 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
162 "Environment assignment not valid: %s", n);
de40a303
LP		 163
164 r = strv_extend(l, n);
165 if (r < 0)
166 return log_oom();
167 }
168
169 return 0;
170}
171
172static int oci_args(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
173 _cleanup_strv_free_ char **l = NULL;
174 char ***value = userdata;
175 JsonVariant *e;
176 int r;
177
178 assert(value);
179
180 JSON_VARIANT_ARRAY_FOREACH(e, v) {
181 const char *n;
182
19130626
ZJS		 183 if (!json_variant_is_string(e))
184 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
185 "Argument is not a string.");
de40a303
LP		 186
187 assert_se(n = json_variant_string(e));
188
189 r = strv_extend(&l, n);
190 if (r < 0)
191 return log_oom();
192 }
193
19130626
ZJS		 194 if (strv_isempty(l))
195 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
196 "Argument list empty, refusing.");
de40a303 197
19130626
ZJS		 198 if (isempty(l[0]))
199 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
200 "Executable name is empty, refusing.");
de40a303 201
6757a013 202 return strv_free_and_replace(*value, l);
de40a303
LP		 203}
204
205static int oci_rlimit_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
206 const char *z;
207 int t, *type = userdata;
208
209 assert_se(type);
210
211 z = startswith(json_variant_string(v), "RLIMIT_");
19130626
ZJS		 212 if (!z)
213 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
214 "rlimit entry's name does not begin with 'RLIMIT_', refusing: %s",
215 json_variant_string(v));
de40a303
LP		 216
217 t = rlimit_from_string(z);
19130626
ZJS		 218 if (t < 0)
219 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
220 "rlimit name unknown: %s", json_variant_string(v));
de40a303
LP		 221
222 *type = t;
223 return 0;
224}
225
226static int oci_rlimit_value(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
227 rlim_t z, *value = userdata;
228
229 assert(value);
230
231 if (json_variant_is_negative(v))
232 z = RLIM_INFINITY;
233 else {
19130626
ZJS		 234 if (!json_variant_is_unsigned(v))
235 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
236 "rlimits limit not unsigned, refusing.");
de40a303
LP		 237
238 z = (rlim_t) json_variant_unsigned(v);
239
19130626
ZJS		 240 if ((uintmax_t) z != json_variant_unsigned(v))
241 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
242 "rlimits limit out of range, refusing.");
de40a303
LP		 243 }
244
245 *value = z;
246 return 0;
247}
248
249static int oci_rlimits(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
250
251 Settings *s = userdata;
252 JsonVariant *e;
253 int r;
254
255 assert(s);
256
257 JSON_VARIANT_ARRAY_FOREACH(e, v) {
258
259 struct rlimit_data {
260 int type;
261 rlim_t soft;
262 rlim_t hard;
263 } data = {
264 .type = -1,
265 .soft = RLIM_INFINITY,
266 .hard = RLIM_INFINITY,
267 };
268
269 static const JsonDispatch table[] = {
270 { "soft", JSON_VARIANT_NUMBER, oci_rlimit_value, offsetof(struct rlimit_data, soft), JSON_MANDATORY },
271 { "hard", JSON_VARIANT_NUMBER, oci_rlimit_value, offsetof(struct rlimit_data, hard), JSON_MANDATORY },
272 { "type", JSON_VARIANT_STRING, oci_rlimit_type, offsetof(struct rlimit_data, type), JSON_MANDATORY },
273 {}
274 };
275
de40a303
LP		 276 r = json_dispatch(e, table, oci_unexpected, flags, &data);
277 if (r < 0)
278 return r;
279
280 assert(data.type >= 0);
281 assert(data.type < _RLIMIT_MAX);
282
19130626
ZJS		 283 if (s->rlimit[data.type])
284 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
285 "rlimits array contains duplicate entry, refusing.");
de40a303
LP		 286
287 s->rlimit[data.type] = new(struct rlimit, 1);
288 if (!s->rlimit[data.type])
289 return log_oom();
290
291 *s->rlimit[data.type] = (struct rlimit) {
292 .rlim_cur = data.soft,
293 .rlim_max = data.hard,
294 };
295
296 }
297 return 0;
298}
299
300static int oci_capability_array(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
301 uint64_t *mask = userdata, m = 0;
302 JsonVariant *e;
303
304 JSON_VARIANT_ARRAY_FOREACH(e, v) {
305 const char *n;
306 int cap;
307
19130626
ZJS		 308 if (!json_variant_is_string(e))
309 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
310 "Entry in capabilities array is not a string.");
de40a303
LP		 311
312 assert_se(n = json_variant_string(e));
313
314 cap = capability_from_name(n);
19130626
ZJS		 315 if (cap < 0)
316 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
317 "Unknown capability: %s", n);
de40a303
LP		 318
319 m |= UINT64_C(1) << cap;
320 }
321
322 if (*mask == (uint64_t) -1)
323 *mask = m;
324 else
325 *mask |= m;
326
327 return 0;
328}
329
330static int oci_capabilities(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
331
332 static const JsonDispatch table[] = {
333 { "effective", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, effective) },
334 { "bounding", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, bounding) },
335 { "inheritable", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, inheritable) },
336 { "permitted", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, permitted) },
337 { "ambient", JSON_VARIANT_ARRAY, oci_capability_array, offsetof(CapabilityQuintet, ambient) },
338 {}
339 };
340
341 Settings *s = userdata;
342 int r;
343
344 assert(s);
345
346 r = json_dispatch(v, table, oci_unexpected, flags, &s->full_capabilities);
347 if (r < 0)
348 return r;
349
350 if (s->full_capabilities.bounding != (uint64_t) -1) {
351 s->capability = s->full_capabilities.bounding;
352 s->drop_capability = ~s->full_capabilities.bounding;
353 }
354
355 return 0;
356}
357
358static int oci_oom_score_adj(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
359 Settings *s = userdata;
360 intmax_t k;
361
362 assert(s);
363
364 k = json_variant_integer(v);
19130626
ZJS		 365 if (k < OOM_SCORE_ADJ_MIN || k > OOM_SCORE_ADJ_MAX)
366 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
367 "oomScoreAdj value out of range: %ji", k);
de40a303
LP		 368
369 s->oom_score_adjust = (int) k;
370 s->oom_score_adjust_set = true;
371
372 return 0;
373}
374
375static int oci_uid_gid(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
376 uid_t *uid = userdata, u;
377 uintmax_t k;
378
379 assert(uid);
380 assert_cc(sizeof(uid_t) == sizeof(gid_t));
381
382 k = json_variant_unsigned(v);
383 u = (uid_t) k;
19130626
ZJS		 384 if ((uintmax_t) u != k)
385 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
386 "UID/GID out of range: %ji", k);
de40a303 387
19130626
ZJS		 388 if (!uid_is_valid(u))
389 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
390 "UID/GID is not valid: " UID_FMT, u);
de40a303
LP		 391
392 *uid = u;
393 return 0;
394}
395
396static int oci_supplementary_gids(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
397 Settings *s = userdata;
398 JsonVariant *e;
399 int r;
400
401 assert(s);
402
403 JSON_VARIANT_ARRAY_FOREACH(e, v) {
404 gid_t gid, *a;
405
19130626
ZJS		 406 if (!json_variant_is_unsigned(e))
407 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
408 "Supplementary GID entry is not a UID.");
de40a303
LP		 409
410 r = oci_uid_gid(name, e, flags, &gid);
411 if (r < 0)
412 return r;
413
414 a = reallocarray(s->supplementary_gids, s->n_supplementary_gids + 1, sizeof(gid_t));
415 if (!a)
416 return log_oom();
417
418 s->supplementary_gids = a;
419 s->supplementary_gids[s->n_supplementary_gids++] = gid;
420 }
421
422 return 0;
423}
424
425static int oci_user(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
426 static const JsonDispatch table[] = {
427 { "uid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(Settings, uid), JSON_MANDATORY },
428 { "gid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(Settings, gid), JSON_MANDATORY },
429 { "additionalGids", JSON_VARIANT_ARRAY, oci_supplementary_gids, 0, 0 },
430 {}
431 };
432
433 return json_dispatch(v, table, oci_unexpected, flags, userdata);
434}
435
436static int oci_process(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
437
438 static const JsonDispatch table[] = {
439 { "terminal", JSON_VARIANT_BOOLEAN, oci_terminal, 0, 0 },
440 { "consoleSize", JSON_VARIANT_OBJECT, oci_console_size, 0, 0 },
441 { "cwd", JSON_VARIANT_STRING, oci_absolute_path, offsetof(Settings, working_directory), 0 },
442 { "env", JSON_VARIANT_ARRAY, oci_env, offsetof(Settings, environment), 0 },
443 { "args", JSON_VARIANT_ARRAY, oci_args, offsetof(Settings, parameters), 0 },
444 { "rlimits", JSON_VARIANT_ARRAY, oci_rlimits, 0, 0 },
445 { "apparmorProfile", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
446 { "capabilities", JSON_VARIANT_OBJECT, oci_capabilities, 0, 0 },
447 { "noNewPrivileges", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(Settings, no_new_privileges), 0 },
448 { "oomScoreAdj", JSON_VARIANT_INTEGER, oci_oom_score_adj, 0, 0 },
449 { "selinuxLabel", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
450 { "user", JSON_VARIANT_OBJECT, oci_user, 0, 0 },
451 {}
452 };
453
454 return json_dispatch(v, table, oci_unexpected, flags, userdata);
455}
456
457static int oci_root(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
458
459 static const JsonDispatch table[] = {
460 { "path", JSON_VARIANT_STRING, json_dispatch_string, offsetof(Settings, root) },
461 { "readonly", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(Settings, read_only) },
462 {}
463 };
464
465 return json_dispatch(v, table, oci_unexpected, flags, userdata);
466}
467
468static int oci_hostname(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
469 Settings *s = userdata;
470 const char *n;
471
472 assert(s);
473
474 assert_se(n = json_variant_string(v));
475
19130626
ZJS		 476 if (!hostname_is_valid(n, false))
477 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
478 "Hostname string is not a valid hostname: %s", n);
de40a303 479
f1531db5 480 return free_and_strdup_warn(&s->hostname, n);
de40a303
LP		 481}
482
483static bool oci_exclude_mount(const char *path) {
484
485 /* Returns "true" for all mounts we insist to mount on our own, and hence ignore the OCI data. */
486
487 if (PATH_IN_SET(path,
488 "/dev",
489 "/dev/mqueue",
490 "/dev/pts",
491 "/dev/shm",
492 "/proc",
493 "/proc/acpi",
494 "/proc/apm",
495 "/proc/asound",
496 "/proc/bus",
497 "/proc/fs",
498 "/proc/irq",
499 "/proc/kallsyms",
500 "/proc/kcore",
501 "/proc/keys",
502 "/proc/scsi",
503 "/proc/sys",
504 "/proc/sys/net",
505 "/proc/sysrq-trigger",
506 "/proc/timer_list",
507 "/run",
508 "/sys",
509 "/sys",
510 "/sys/fs/selinux",
511 "/tmp"))
512 return true;
513
514 /* Similar, skip the whole /sys/fs/cgroups subtree */
515 if (path_startswith(path, "/sys/fs/cgroup"))
516 return true;
517
518 return false;
519}
520
b2e07b1a
ZJS		 521typedef struct oci_mount_data {
522 char *destination;
523 char *source;
524 char *type;
525 char **options;
526} oci_mount_data;
527
528static void cleanup_oci_mount_data(oci_mount_data *data) {
529 free(data->destination);
530 free(data->source);
531 strv_free(data->options);
532 free(data->type);
533}
534
de40a303
LP		 535static int oci_mounts(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
536 Settings *s = userdata;
537 JsonVariant *e;
538 int r;
539
540 assert(s);
541
542 JSON_VARIANT_ARRAY_FOREACH(e, v) {
de40a303 543 static const JsonDispatch table[] = {
b2e07b1a
ZJS		 544 { "destination", JSON_VARIANT_STRING, oci_absolute_path, offsetof(oci_mount_data, destination), JSON_MANDATORY },
545 { "source", JSON_VARIANT_STRING, json_dispatch_string, offsetof(oci_mount_data, source), 0 },
546 { "options", JSON_VARIANT_ARRAY, json_dispatch_strv, offsetof(oci_mount_data, options), 0, },
547 { "type", JSON_VARIANT_STRING, json_dispatch_string, offsetof(oci_mount_data, type), 0 },
de40a303
LP		 548 {}
549 };
550
551 _cleanup_free_ char *joined_options = NULL;
552 CustomMount *m;
b2e07b1a 553 _cleanup_(cleanup_oci_mount_data) oci_mount_data data = {};
de40a303
LP		 554
555 r = json_dispatch(e, table, oci_unexpected, flags, &data);
556 if (r < 0)
b2e07b1a 557 return r;
de40a303 558
b2e07b1a
ZJS		 559 if (!path_is_absolute(data.destination))
560 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
561 "Mount destination not an absolute path: %s", data.destination);
de40a303
LP		 562
563 if (oci_exclude_mount(data.destination))
b2e07b1a 564 continue;
de40a303
LP		 565
566 if (data.options) {
567 joined_options = strv_join(data.options, ",");
b2e07b1a
ZJS		 568 if (!joined_options)
569 return log_oom();
de40a303
LP		 570 }
571
572 if (!data.type || streq(data.type, "bind")) {
b1f13b0e 573 if (data.source && !path_is_absolute(data.source)) {
de40a303
LP		 574 char *joined;
575
576 joined = path_join(s->bundle, data.source);
b2e07b1a
ZJS		 577 if (!joined)
578 return log_oom();
de40a303
LP		 579
580 free_and_replace(data.source, joined);
581 }
582
583 data.type = mfree(data.type);
584
585 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_BIND);
586 } else
587 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_ARBITRARY);
b2e07b1a
ZJS		 588 if (!m)
589 return log_oom();
de40a303
LP		 590
591 m->destination = TAKE_PTR(data.destination);
592 m->source = TAKE_PTR(data.source);
593 m->options = TAKE_PTR(joined_options);
594 m->type_argument = TAKE_PTR(data.type);
de40a303
LP		 595 }
596
597 return 0;
598}
599
600static int oci_namespace_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
601 unsigned long *nsflags = userdata;
602 const char *n;
603
604 assert(nsflags);
605 assert_se(n = json_variant_string(v));
606
607 /* We don't use namespace_flags_from_string() here, as the OCI spec uses slightly different names than the
608 * kernel here. */
609 if (streq(n, "pid"))
610 *nsflags = CLONE_NEWPID;
611 else if (streq(n, "network"))
612 *nsflags = CLONE_NEWNET;
613 else if (streq(n, "mount"))
614 *nsflags = CLONE_NEWNS;
615 else if (streq(n, "ipc"))
616 *nsflags = CLONE_NEWIPC;
617 else if (streq(n, "uts"))
618 *nsflags = CLONE_NEWUTS;
619 else if (streq(n, "user"))
620 *nsflags = CLONE_NEWUSER;
621 else if (streq(n, "cgroup"))
622 *nsflags = CLONE_NEWCGROUP;
19130626
ZJS		 623 else
624 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
625 "Unknown cgroup type, refusing: %s", n);
de40a303
LP		 626
627 return 0;
628}
629
630static int oci_namespaces(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
631 Settings *s = userdata;
632 unsigned long n = 0;
633 JsonVariant *e;
634 int r;
635
636 assert_se(s);
637
638 JSON_VARIANT_ARRAY_FOREACH(e, v) {
639
640 struct namespace_data {
641 unsigned long type;
642 char *path;
643 } data = {};
644
645 static const JsonDispatch table[] = {
646 { "type", JSON_VARIANT_STRING, oci_namespace_type, offsetof(struct namespace_data, type), JSON_MANDATORY },
647 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(struct namespace_data, path), 0 },
648 {}
649 };
650
651 r = json_dispatch(e, table, oci_unexpected, flags, &data);
652 if (r < 0) {
653 free(data.path);
654 return r;
655 }
656
657 if (data.path) {
658 if (data.type != CLONE_NEWNET) {
de40a303 659 free(data.path);
19130626
ZJS		 660 return json_log(e, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
661 "Specifying namespace path for non-network namespace is not supported.");
de40a303
LP		 662 }
663
664 if (s->network_namespace_path) {
de40a303 665 free(data.path);
19130626
ZJS		 666 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
667 "Network namespace path specified more than once, refusing.");
de40a303
LP		 668 }
669
670 free(s->network_namespace_path);
671 s->network_namespace_path = data.path;
672 }
673
674 if (FLAGS_SET(n, data.type)) {
19130626
ZJS		 675 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
676 "Duplicate namespace specification, refusing.");
de40a303
LP		 677 }
678
679 n |= data.type;
680 }
681
19130626
ZJS		 682 if (!FLAGS_SET(n, CLONE_NEWNS))
683 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
684 "Containers without file system namespace aren't supported.");
de40a303
LP		 685
686 s->private_network = FLAGS_SET(n, CLONE_NEWNET);
687 s->userns_mode = FLAGS_SET(n, CLONE_NEWUSER) ? USER_NAMESPACE_FIXED : USER_NAMESPACE_NO;
688 s->use_cgns = FLAGS_SET(n, CLONE_NEWCGROUP);
689
690 s->clone_ns_flags = n & (CLONE_NEWIPC|CLONE_NEWPID|CLONE_NEWUTS);
691
692 return 0;
693}
694
695static int oci_uid_gid_range(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
696 uid_t *uid = userdata, u;
697 uintmax_t k;
698
699 assert(uid);
700 assert_cc(sizeof(uid_t) == sizeof(gid_t));
701
702 /* This is very much like oci_uid_gid(), except the checks are a bit different, as this is a UID range rather
703 * than a specific UID, and hence (uid_t) -1 has no special significance. OTOH a range of zero makes no
704 * sense. */
705
706 k = json_variant_unsigned(v);
707 u = (uid_t) k;
19130626
ZJS		 708 if ((uintmax_t) u != k)
709 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
710 "UID/GID out of range: %ji", k);
711 if (u == 0)
712 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
713 "UID/GID range can't be zero.");
de40a303
LP		 714
715 *uid = u;
716 return 0;
717}
718
719static int oci_uid_gid_mappings(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
720 struct mapping_data {
721 uid_t host_id;
722 uid_t container_id;
723 uid_t range;
724 } data = {
725 .host_id = UID_INVALID,
726 .container_id = UID_INVALID,
727 .range = 0,
728 };
729
730 static const JsonDispatch table[] = {
731 { "containerID", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(struct mapping_data, container_id), JSON_MANDATORY },
732 { "hostID", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(struct mapping_data, host_id), JSON_MANDATORY },
733 { "size", JSON_VARIANT_UNSIGNED, oci_uid_gid_range, offsetof(struct mapping_data, range), JSON_MANDATORY },
734 {}
735 };
736
737 Settings *s = userdata;
738 JsonVariant *e;
739 int r;
740
741 assert(s);
742
743 if (json_variant_elements(v) == 0)
744 return 0;
745
19130626
ZJS		 746 if (json_variant_elements(v) > 1)
747 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
748 "UID/GID mappings with more than one entry are not supported.");
de40a303
LP		 749
750 assert_se(e = json_variant_by_index(v, 0));
751
752 r = json_dispatch(e, table, oci_unexpected, flags, &data);
753 if (r < 0)
754 return r;
755
756 if (data.host_id + data.range < data.host_id ||
19130626
ZJS		 757 data.container_id + data.range < data.container_id)
758 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
759 "UID/GID range goes beyond UID/GID validity range, refusing.");
de40a303 760
19130626
ZJS		 761 if (data.container_id != 0)
762 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
763 "UID/GID mappings with a non-zero container base are not supported.");
de40a303
LP		 764
765 if (data.range < 0x10000)
19130626
ZJS		 766 json_log(v, flags|JSON_WARNING, 0,
767 "UID/GID mapping with less than 65536 UID/GIDS set up, you are looking for trouble.");
de40a303
LP		 768
769 if (s->uid_range != UID_INVALID &&
19130626
ZJS		 770 (s->uid_shift != data.host_id || s->uid_range != data.range))
771 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
772 "Non-matching UID and GID mappings are not supported.");
de40a303
LP		 773
774 s->uid_shift = data.host_id;
775 s->uid_range = data.range;
776
777 return 0;
778}
779
780static int oci_device_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
781 mode_t *mode = userdata;
782 const char *t;
783
784 assert(mode);
785 assert_se(t = json_variant_string(v));
786
787 if (STR_IN_SET(t, "c", "u"))
788 *mode = (*mode & ~S_IFMT) | S_IFCHR;
789 else if (streq(t, "b"))
790 *mode = (*mode & ~S_IFMT) | S_IFBLK;
791 else if (streq(t, "p"))
792 *mode = (*mode & ~S_IFMT) | S_IFIFO;
19130626
ZJS		 793 else
794 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
795 "Unknown device type: %s", t);
de40a303
LP		 796
797 return 0;
798}
799
800static int oci_device_major(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
801 unsigned *u = userdata;
802 uintmax_t k;
803
804 assert_se(u);
805
806 k = json_variant_unsigned(v);
19130626
ZJS		 807 if (!DEVICE_MAJOR_VALID(k))
808 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
809 "Device major %ji out of range.", k);
de40a303
LP		 810
811 *u = (unsigned) k;
812 return 0;
813}
814
815static int oci_device_minor(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
816 unsigned *u = userdata;
817 uintmax_t k;
818
819 assert_se(u);
820
821 k = json_variant_unsigned(v);
19130626
ZJS		 822 if (!DEVICE_MINOR_VALID(k))
823 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
824 "Device minor %ji out of range.", k);
de40a303
LP		 825
826 *u = (unsigned) k;
827 return 0;
828}
829
830static int oci_device_file_mode(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
831 mode_t *mode = userdata, m;
832 uintmax_t k;
833
834 assert(mode);
835
836 k = json_variant_unsigned(v);
837 m = (mode_t) k;
838
19130626
ZJS		 839 if ((m & ~07777) != 0 || (uintmax_t) m != k)
840 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
841 "fileMode out of range, refusing.");
de40a303
LP		 842
843 *mode = m;
844 return 0;
845}
846
847static int oci_devices(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
848 Settings *s = userdata;
849 JsonVariant *e;
850 int r;
851
852 assert(s);
853
854 JSON_VARIANT_ARRAY_FOREACH(e, v) {
855
856 static const JsonDispatch table[] = {
857 { "type", JSON_VARIANT_STRING, oci_device_type, offsetof(DeviceNode, mode), JSON_MANDATORY },
858 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(DeviceNode, path), JSON_MANDATORY },
859 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(DeviceNode, major), 0 },
860 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(DeviceNode, minor), 0 },
861 { "fileMode", JSON_VARIANT_UNSIGNED, oci_device_file_mode, offsetof(DeviceNode, mode), 0 },
862 { "uid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(DeviceNode, uid), 0 },
863 { "gid", JSON_VARIANT_UNSIGNED, oci_uid_gid, offsetof(DeviceNode, gid), 0 },
864 {}
865 };
866
867 DeviceNode *node, *nodes;
868
869 nodes = reallocarray(s->extra_nodes, s->n_extra_nodes + 1, sizeof(DeviceNode));
870 if (!nodes)
871 return log_oom();
872
873 s->extra_nodes = nodes;
874
875 node = nodes + s->n_extra_nodes;
876 *node = (DeviceNode) {
877 .uid = UID_INVALID,
878 .gid = GID_INVALID,
879 .major = (unsigned) -1,
880 .minor = (unsigned) -1,
881 .mode = 0644,
882 };
883
884 r = json_dispatch(e, table, oci_unexpected, flags, node);
885 if (r < 0)
886 goto fail_element;
887
888 if (S_ISCHR(node->mode) || S_ISBLK(node->mode)) {
889 _cleanup_free_ char *path = NULL;
890
891 if (node->major == (unsigned) -1 || node->minor == (unsigned) -1) {
19130626
ZJS		 892 r = json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
893 "Major/minor required when device node is device node");
de40a303
LP		 894 goto fail_element;
895 }
896
897 /* Suppress a couple of implicit device nodes */
898 r = device_path_make_canonical(node->mode, makedev(node->major, node->minor), &path);
899 if (r < 0)
900 json_log(e, flags|JSON_DEBUG, 0, "Failed to resolve device node %u:%u, ignoring: %m", node->major, node->minor);
901 else {
902 if (PATH_IN_SET(path,
903 "/dev/null",
904 "/dev/zero",
905 "/dev/full",
906 "/dev/random",
907 "/dev/urandom",
908 "/dev/tty",
909 "/dev/net/tun",
910 "/dev/ptmx",
911 "/dev/pts/ptmx",
912 "/dev/console")) {
913
914 json_log(e, flags|JSON_DEBUG, 0, "Ignoring devices item for device '%s', as it is implicitly created anyway.", path);
915 free(node->path);
916 continue;
917 }
918 }
919 }
920
921 s->n_extra_nodes++;
922 continue;
923
924 fail_element:
925 free(node->path);
926 return r;
927 }
928
929 return 0;
930}
931
932static int oci_cgroups_path(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
933 _cleanup_free_ char *slice = NULL, *backwards = NULL;
934 Settings *s = userdata;
935 const char *p;
936 int r;
937
938 assert(s);
939
940 assert_se(p = json_variant_string(v));
941
942 r = cg_path_get_slice(p, &slice);
943 if (r < 0)
944 return json_log(v, flags, r, "Couldn't derive slice unit name from path '%s': %m", p);
945
946 r = cg_slice_to_path(slice, &backwards);
947 if (r < 0)
948 return json_log(v, flags, r, "Couldn't convert slice unit name '%s' back to path: %m", slice);
949
19130626
ZJS		 950 if (!path_equal(backwards, p))
951 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
952 "Control group path '%s' does not refer to slice unit, refusing.", p);
de40a303
LP		 953
954 free_and_replace(s->slice, slice);
955 return 0;
956}
957
958static int oci_cgroup_device_type(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
959 mode_t *mode = userdata;
960 const char *n;
961
962 assert_se(n = json_variant_string(v));
963
964 if (streq(n, "c"))
965 *mode = S_IFCHR;
966 else if (streq(n, "b"))
967 *mode = S_IFBLK;
19130626
ZJS		 968 else
969 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
970 "Control group device type unknown: %s", n);
de40a303
LP		 971
972 return 0;
973}
974
975struct device_data {
976 bool allow;
977 bool r;
978 bool w;
979 bool m;
980 mode_t type;
981 unsigned major;
982 unsigned minor;
983};
984
985static int oci_cgroup_device_access(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
986 struct device_data *d = userdata;
987 bool r = false, w = false, m = false;
988 const char *s;
989 size_t i;
990
991 assert_se(s = json_variant_string(v));
992
19130626 993 for (i = 0; s[i]; i++)
de40a303
LP		 994 if (s[i] == 'r')
995 r = true;
996 else if (s[i] == 'w')
997 w = true;
998 else if (s[i] == 'm')
999 m = true;
19130626
ZJS		 1000 else
1001 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1002 "Unknown device access character '%c'.", s[i]);
de40a303
LP		 1003
1004 d->r = r;
1005 d->w = w;
1006 d->m = m;
1007
1008 return 0;
1009}
1010
1011static int oci_cgroup_devices(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1012
1013 _cleanup_free_ struct device_data *list = NULL;
1014 Settings *s = userdata;
1015 size_t n_list = 0, i;
1016 bool noop = false;
1017 JsonVariant *e;
1018 int r;
1019
1020 assert(s);
1021
1022 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1023
1024 struct device_data data = {
1025 .major = (unsigned) -1,
1026 .minor = (unsigned) -1,
1027 }, *a;
1028
1029 static const JsonDispatch table[] = {
1030 { "allow", JSON_VARIANT_BOOLEAN, json_dispatch_boolean, offsetof(struct device_data, allow), JSON_MANDATORY },
1031 { "type", JSON_VARIANT_STRING, oci_cgroup_device_type, offsetof(struct device_data, type), 0 },
1032 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), 0 },
1033 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), 0 },
1034 { "access", JSON_VARIANT_STRING, oci_cgroup_device_access, 0, 0 },
1035 {}
1036 };
1037
1038 r = json_dispatch(e, table, oci_unexpected, flags, &data);
1039 if (r < 0)
1040 return r;
1041
1042 if (!data.allow) {
1043 /* The fact that OCI allows 'deny' entries makes really no sense, as 'allow' vs. 'deny' for the
1044 * devices cgroup controller is really not about whitelisting and blacklisting but about adding
1045 * and removing entries from the whitelist. Since we always start out with an empty whitelist
1046 * we hence ignore the whole thing, as removing entries which don't exist make no sense. We'll
1047 * log about this, since this is really borked in the spec, with one exception: the entry
1048 * that's supposed to drop the kernel's default we ignore silently */
1049
1050 if (!data.r || !data.w || !data.m || data.type != 0 || data.major != (unsigned) -1 || data.minor != (unsigned) -1)
1051 json_log(v, flags|JSON_WARNING, 0, "Devices cgroup whitelist with arbitrary 'allow' entries not supported, ignoring.");
1052
1053 /* We ignore the 'deny' entry as for us that's implied */
1054 continue;
1055 }
1056
1057 if (!data.r && !data.w && !data.m) {
1058 json_log(v, flags|LOG_WARNING, 0, "Device cgroup whitelist entry with no effect found, ignoring.");
1059 continue;
1060 }
1061
19130626
ZJS		 1062 if (data.minor != (unsigned) -1 && data.major == (unsigned) -1)
1063 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
1064 "Device cgroup whitelist entries with minors but no majors not supported.");
de40a303 1065
19130626
ZJS		 1066 if (data.major != (unsigned) -1 && data.type == 0)
1067 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
1068 "Device cgroup whitelist entries with majors but no device node type not supported.");
de40a303
LP		 1069
1070 if (data.type == 0) {
1071 if (data.r && data.w && data.m) /* a catchall whitelist entry means we are looking at a noop */
1072 noop = true;
19130626
ZJS		 1073 else
1074 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP),
1075 "Device cgroup whitelist entries with no type not supported.");
de40a303
LP		 1076 }
1077
1078 a = reallocarray(list, n_list + 1, sizeof(struct device_data));
1079 if (!a)
1080 return log_oom();
1081
1082 list = a;
1083 list[n_list++] = data;
1084 }
1085
1086 if (noop)
1087 return 0;
1088
1089 r = settings_allocate_properties(s);
1090 if (r < 0)
1091 return r;
1092
1093 r = sd_bus_message_open_container(s->properties, 'r', "sv");
1094 if (r < 0)
1095 return bus_log_create_error(r);
1096
1097 r = sd_bus_message_append(s->properties, "s", "DeviceAllow");
1098 if (r < 0)
1099 return bus_log_create_error(r);
1100
1101 r = sd_bus_message_open_container(s->properties, 'v', "a(ss)");
1102 if (r < 0)
1103 return bus_log_create_error(r);
1104
1105 r = sd_bus_message_open_container(s->properties, 'a', "(ss)");
1106 if (r < 0)
1107 return bus_log_create_error(r);
1108
1109 for (i = 0; i < n_list; i++) {
1110 _cleanup_free_ char *pattern = NULL;
1111 char access[4];
1112 size_t n = 0;
1113
1114 if (list[i].minor == (unsigned) -1) {
1115 const char *t;
1116
1117 if (list[i].type == S_IFBLK)
1118 t = "block";
1119 else {
1120 assert(list[i].type == S_IFCHR);
1121 t = "char";
1122 }
1123
1124 if (list[i].major == (unsigned) -1) {
1125 pattern = strjoin(t, "-*");
1126 if (!pattern)
1127 return log_oom();
1128 } else {
1129 if (asprintf(&pattern, "%s-%u", t, list[i].major) < 0)
1130 return log_oom();
1131 }
1132
1133 } else {
1134 assert(list[i].major != (unsigned) -1); /* If a minor is specified, then a major also needs to be specified */
1135
1136 r = device_path_make_major_minor(list[i].type, makedev(list[i].major, list[i].minor), &pattern);
1137 if (r < 0)
1138 return log_oom();
1139 }
1140
1141 if (list[i].r)
1142 access[n++] = 'r';
1143 if (list[i].w)
1144 access[n++] = 'w';
1145 if (list[i].m)
1146 access[n++] = 'm';
1147 access[n] = 0;
1148
1149 assert(n > 0);
1150
1151 r = sd_bus_message_append(s->properties, "(ss)", pattern, access);
1152 if (r < 0)
1153 return bus_log_create_error(r);
1154 }
1155
1156 r = sd_bus_message_close_container(s->properties);
1157 if (r < 0)
1158 return bus_log_create_error(r);
1159
1160 r = sd_bus_message_close_container(s->properties);
1161 if (r < 0)
1162 return bus_log_create_error(r);
1163
1164 r = sd_bus_message_close_container(s->properties);
1165 if (r < 0)
1166 return bus_log_create_error(r);
1167
1168 return 0;
1169}
1170
1171static int oci_cgroup_memory_limit(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1172 uint64_t *m = userdata;
1173 uintmax_t k;
1174
1175 assert(m);
1176
1177 if (json_variant_is_negative(v)) {
1178 *m = UINT64_MAX;
1179 return 0;
1180 }
1181
19130626
ZJS		 1182 if (!json_variant_is_unsigned(v))
1183 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1184 "Memory limit is not an unsigned integer");
de40a303
LP		 1185
1186 k = json_variant_unsigned(v);
19130626
ZJS		 1187 if (k >= UINT64_MAX)
1188 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1189 "Memory limit too large: %ji", k);
de40a303
LP		 1190
1191 *m = (uint64_t) k;
1192 return 0;
1193}
1194
1195static int oci_cgroup_memory(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1196
1197 struct memory_data {
1198 uint64_t limit;
1199 uint64_t reservation;
1200 uint64_t swap;
1201 } data = {
1202 .limit = UINT64_MAX,
1203 .reservation = UINT64_MAX,
1204 .swap = UINT64_MAX,
1205 };
1206
1207 static const JsonDispatch table[] = {
1208 { "limit", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, limit), 0 },
1209 { "reservation", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, reservation), 0 },
1210 { "swap", JSON_VARIANT_NUMBER, oci_cgroup_memory_limit, offsetof(struct memory_data, swap), 0 },
1211 { "kernel", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1212 { "kernelTCP", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1213 { "swapiness", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1214 { "disableOOMKiller", JSON_VARIANT_NUMBER, oci_unsupported, 0, JSON_PERMISSIVE },
1215 {}
1216 };
1217
1218 Settings *s = userdata;
1219 int r;
1220
1221 r = json_dispatch(v, table, oci_unexpected, flags, &data);
1222 if (r < 0)
1223 return r;
1224
1225 if (data.swap != UINT64_MAX) {
1226 if (data.limit == UINT64_MAX)
1227 json_log(v, flags|LOG_WARNING, 0, "swap limit without memory limit is not supported, ignoring.");
1228 else if (data.swap < data.limit)
1229 json_log(v, flags|LOG_WARNING, 0, "swap limit is below memory limit, ignoring.");
1230 else {
1231 r = settings_allocate_properties(s);
1232 if (r < 0)
1233 return r;
1234
1235 r = sd_bus_message_append(s->properties, "(sv)", "MemorySwapMax", "t", data.swap - data.limit);
1236 if (r < 0)
1237 return bus_log_create_error(r);
1238 }
1239 }
1240
1241 if (data.limit != UINT64_MAX) {
1242 r = settings_allocate_properties(s);
1243 if (r < 0)
1244 return r;
1245
1246 r = sd_bus_message_append(s->properties, "(sv)", "MemoryMax", "t", data.limit);
1247 if (r < 0)
1248 return bus_log_create_error(r);
1249 }
1250
1251 if (data.reservation != UINT64_MAX) {
1252 r = settings_allocate_properties(s);
1253 if (r < 0)
1254 return r;
1255
1256 r = sd_bus_message_append(s->properties, "(sv)", "MemoryLow", "t", data.reservation);
1257 if (r < 0)
1258 return bus_log_create_error(r);
1259 }
1260
1261 return 0;
1262}
1263
1264struct cpu_data {
1265 uint64_t shares;
1266 uint64_t quota;
1267 uint64_t period;
0985c7c4 1268 CPUSet cpu_set;
de40a303
LP		 1269};
1270
1271static int oci_cgroup_cpu_shares(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1272 uint64_t *u = userdata;
1273 uintmax_t k;
1274
1275 assert(u);
1276
1277 k = json_variant_unsigned(v);
19130626
ZJS		 1278 if (k < CGROUP_CPU_SHARES_MIN || k > CGROUP_CPU_SHARES_MAX)
1279 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1280 "shares value out of range.");
de40a303
LP		 1281
1282 *u = (uint64_t) k;
1283 return 0;
1284}
1285
1286static int oci_cgroup_cpu_quota(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1287 uint64_t *u = userdata;
1288 uintmax_t k;
1289
1290 assert(u);
1291
1292 k = json_variant_unsigned(v);
19130626
ZJS		 1293 if (k <= 0 || k >= UINT64_MAX)
1294 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1295 "period/quota value out of range.");
de40a303
LP		 1296
1297 *u = (uint64_t) k;
1298 return 0;
1299}
1300
1301static int oci_cgroup_cpu_cpus(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1302 struct cpu_data *data = userdata;
0985c7c4 1303 CPUSet set;
de40a303 1304 const char *n;
0985c7c4 1305 int r;
de40a303
LP		 1306
1307 assert(data);
1308
1309 assert_se(n = json_variant_string(v));
1310
0985c7c4
ZJS		 1311 r = parse_cpu_set(n, &set);
1312 if (r < 0)
1313 return json_log(v, flags, r, "Failed to parse CPU set specification: %s", n);
de40a303 1314
0985c7c4
ZJS		 1315 cpu_set_reset(&data->cpu_set);
1316 data->cpu_set = set;
de40a303
LP		 1317
1318 return 0;
1319}
1320
1321static int oci_cgroup_cpu(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1322
1323 static const JsonDispatch table[] = {
1324 { "shares", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_shares, offsetof(struct cpu_data, shares), 0 },
1325 { "quota", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_quota, offsetof(struct cpu_data, quota), 0 },
1326 { "period", JSON_VARIANT_UNSIGNED, oci_cgroup_cpu_quota, offsetof(struct cpu_data, period), 0 },
1327 { "realtimeRuntime", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, 0 },
1328 { "realtimePeriod", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, 0 },
1329 { "cpus", JSON_VARIANT_STRING, oci_cgroup_cpu_cpus, 0, 0 },
1330 { "mems", JSON_VARIANT_STRING, oci_unsupported, 0, 0 },
1331 {}
1332 };
1333
1334 struct cpu_data data = {
1335 .shares = UINT64_MAX,
1336 .quota = UINT64_MAX,
1337 .period = UINT64_MAX,
1338 };
1339
1340 Settings *s = userdata;
1341 int r;
1342
1343 r = json_dispatch(v, table, oci_unexpected, flags, &data);
1344 if (r < 0) {
0985c7c4 1345 cpu_set_reset(&data.cpu_set);
de40a303
LP		 1346 return r;
1347 }
1348
0985c7c4
ZJS		 1349 cpu_set_reset(&s->cpu_set);
1350 s->cpu_set = data.cpu_set;
de40a303
LP		 1351
1352 if (data.shares != UINT64_MAX) {
1353 r = settings_allocate_properties(s);
1354 if (r < 0)
1355 return r;
1356
1357 r = sd_bus_message_append(s->properties, "(sv)", "CPUShares", "t", data.shares);
1358 if (r < 0)
1359 return bus_log_create_error(r);
1360 }
1361
1362 if (data.quota != UINT64_MAX && data.period != UINT64_MAX) {
1363 r = settings_allocate_properties(s);
1364 if (r < 0)
1365 return r;
1366
1367 r = sd_bus_message_append(s->properties, "(sv)", "CPUQuotaPerSecUSec", "t", (uint64_t) (data.quota * USEC_PER_SEC / data.period));
1368 if (r < 0)
1369 return bus_log_create_error(r);
1370
19130626
ZJS		 1371 } else if ((data.quota != UINT64_MAX) != (data.period != UINT64_MAX))
1372 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1373 "CPU quota and period not used together.");
de40a303
LP		 1374
1375 return 0;
1376}
1377
1378static int oci_cgroup_block_io_weight(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1379 Settings *s = userdata;
1380 uintmax_t k;
1381 int r;
1382
1383 assert(s);
1384
1385 k = json_variant_unsigned(v);
19130626
ZJS		 1386 if (k < CGROUP_BLKIO_WEIGHT_MIN || k > CGROUP_BLKIO_WEIGHT_MAX)
1387 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1388 "Block I/O weight out of range.");
de40a303
LP		 1389
1390 r = settings_allocate_properties(s);
1391 if (r < 0)
1392 return r;
1393
1394 r = sd_bus_message_append(s->properties, "(sv)", "BlockIOWeight", "t", (uint64_t) k);
1395 if (r < 0)
1396 return bus_log_create_error(r);
1397
1398 return 0;
1399}
1400
1401static int oci_cgroup_block_io_weight_device(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1402 Settings *s = userdata;
1403 JsonVariant *e;
1404 int r;
1405
1406 assert(s);
1407
1408 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1409 struct device_data {
1410 unsigned major;
1411 unsigned minor;
1412 uintmax_t weight;
1413 } data = {
1414 .major = (unsigned) -1,
1415 .minor = (unsigned) -1,
1416 .weight = UINTMAX_MAX,
1417 };
1418
1419 static const JsonDispatch table[] = {
1420 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), JSON_MANDATORY },
1421 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), JSON_MANDATORY },
1422 { "weight", JSON_VARIANT_UNSIGNED, json_dispatch_unsigned, offsetof(struct device_data, weight), 0 },
1423 { "leafWeight", JSON_VARIANT_INTEGER, oci_unsupported, 0, JSON_PERMISSIVE },
1424 {}
1425 };
1426
1427 _cleanup_free_ char *path = NULL;
1428
1429 r = json_dispatch(e, table, oci_unexpected, flags, &data);
1430 if (r < 0)
1431 return r;
1432
1433 if (data.weight == UINTMAX_MAX)
1434 continue;
1435
19130626
ZJS		 1436 if (data.weight < CGROUP_BLKIO_WEIGHT_MIN || data.weight > CGROUP_BLKIO_WEIGHT_MAX)
1437 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1438 "Block I/O device weight out of range.");
de40a303
LP		 1439
1440 r = device_path_make_major_minor(S_IFBLK, makedev(data.major, data.minor), &path);
1441 if (r < 0)
1442 return json_log(v, flags, r, "Failed to build device path: %m");
1443
1444 r = settings_allocate_properties(s);
1445 if (r < 0)
1446 return r;
1447
1448 r = sd_bus_message_append(s->properties, "(sv)", "BlockIODeviceWeight", "a(st)", 1, path, (uint64_t) data.weight);
1449 if (r < 0)
1450 return bus_log_create_error(r);
1451 }
1452
1453 return 0;
1454}
1455
1456static int oci_cgroup_block_io_throttle(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1457 Settings *s = userdata;
1458 const char *pname;
1459 JsonVariant *e;
1460 int r;
1461
1462 assert(s);
1463
1464 pname = streq(name, "throttleReadBpsDevice") ? "IOReadBandwidthMax" :
1465 streq(name, "throttleWriteBpsDevice") ? "IOWriteBandwidthMax" :
1466 streq(name, "throttleReadIOPSDevice") ? "IOReadIOPSMax" :
1467 "IOWriteIOPSMax";
1468
1469 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1470 struct device_data {
1471 unsigned major;
1472 unsigned minor;
1473 uintmax_t rate;
1474 } data = {
1475 .major = (unsigned) -1,
1476 .minor = (unsigned) -1,
1477 };
1478
1479 static const JsonDispatch table[] = {
1480 { "major", JSON_VARIANT_UNSIGNED, oci_device_major, offsetof(struct device_data, major), JSON_MANDATORY },
1481 { "minor", JSON_VARIANT_UNSIGNED, oci_device_minor, offsetof(struct device_data, minor), JSON_MANDATORY },
1482 { "rate", JSON_VARIANT_UNSIGNED, json_dispatch_unsigned, offsetof(struct device_data, rate), JSON_MANDATORY },
1483 {}
1484 };
1485
1486 _cleanup_free_ char *path = NULL;
1487
1488 r = json_dispatch(e, table, oci_unexpected, flags, &data);
1489 if (r < 0)
1490 return r;
1491
19130626
ZJS		 1492 if (data.rate >= UINT64_MAX)
1493 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
1494 "Block I/O device rate out of range.");
de40a303
LP		 1495
1496 r = device_path_make_major_minor(S_IFBLK, makedev(data.major, data.minor), &path);
1497 if (r < 0)
1498 return json_log(v, flags, r, "Failed to build device path: %m");
1499
1500 r = settings_allocate_properties(s);
1501 if (r < 0)
1502 return r;
1503
1504 r = sd_bus_message_append(s->properties, "(sv)", pname, "a(st)", 1, path, (uint64_t) data.rate);
1505 if (r < 0)
1506 return bus_log_create_error(r);
1507 }
1508
1509 return 0;
1510}
1511
1512static int oci_cgroup_block_io(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1513
1514 static const JsonDispatch table[] = {
1515 { "weight", JSON_VARIANT_UNSIGNED, oci_cgroup_block_io_weight, 0, 0 },
1516 { "leafWeight", JSON_VARIANT_UNSIGNED, oci_unsupported, 0, JSON_PERMISSIVE },
1517 { "weightDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_weight_device, 0, 0 },
1518 { "throttleReadBpsDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1519 { "throttleWriteBpsDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1520 { "throttleReadIOPSDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1521 { "throttleWriteIOPSDevice", JSON_VARIANT_ARRAY, oci_cgroup_block_io_throttle, 0, 0 },
1522 {}
1523 };
1524
1525 return json_dispatch(v, table, oci_unexpected, flags, userdata);
1526}
1527
1528static int oci_cgroup_pids(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1529
1530 static const JsonDispatch table[] = {
1531 { "limit", JSON_VARIANT_NUMBER, json_dispatch_variant, 0, JSON_MANDATORY },
1532 {}
1533 };
1534
1535 _cleanup_(json_variant_unrefp) JsonVariant *k = NULL;
1536 Settings *s = userdata;
1537 uint64_t m;
1538 int r;
1539
1540 assert(s);
1541
1542 r = json_dispatch(v, table, oci_unexpected, flags, &k);
1543 if (r < 0)
1544 return r;
1545
1546 if (json_variant_is_negative(k))
1547 m = UINT64_MAX;
1548 else {
19130626
ZJS		 1549 if (!json_variant_is_unsigned(k))
1550 return json_log(k, flags, SYNTHETIC_ERRNO(EINVAL),
1551 "pids limit not unsigned integer, refusing.");
de40a303
LP		 1552
1553 m = (uint64_t) json_variant_unsigned(k);
1554
19130626
ZJS		 1555 if ((uintmax_t) m != json_variant_unsigned(k))
1556 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1557 "pids limit out of range, refusing.");
de40a303
LP		 1558 }
1559
1560 r = settings_allocate_properties(s);
1561 if (r < 0)
1562 return r;
1563
1564 r = sd_bus_message_append(s->properties, "(sv)", "TasksMax", "t", m);
1565 if (r < 0)
1566 return bus_log_create_error(r);
1567
1568 return 0;
1569}
1570
1571static int oci_resources(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1572
1573 static const JsonDispatch table[] = {
1574 { "devices", JSON_VARIANT_ARRAY, oci_cgroup_devices, 0, 0 },
1575 { "memory", JSON_VARIANT_OBJECT, oci_cgroup_memory, 0, 0 },
1576 { "cpu", JSON_VARIANT_OBJECT, oci_cgroup_cpu, 0, 0 },
1577 { "blockIO", JSON_VARIANT_OBJECT, oci_cgroup_block_io, 0, 0 },
1578 { "hugepageLimits", JSON_VARIANT_ARRAY, oci_unsupported, 0, 0 },
1579 { "network", JSON_VARIANT_OBJECT, oci_unsupported, 0, 0 },
1580 { "pids", JSON_VARIANT_OBJECT, oci_cgroup_pids, 0, 0 },
1581 { "rdma", JSON_VARIANT_OBJECT, oci_unsupported, 0, 0 },
1582 {}
1583 };
1584
1585 return json_dispatch(v, table, oci_unexpected, flags, userdata);
1586}
1587
1588static bool sysctl_key_valid(const char *s) {
1589 bool dot = true;
1590
1591 /* Note that we are a bit stricter here than in systemd-sysctl, as that inherited semantics from the old sysctl
1592 * tool, which were really weird (as it swaps / and . in both ways) */
1593
1594 if (isempty(s))
1595 return false;
1596
1597 for (; *s; s++) {
1598
1599 if (*s <= ' ' || *s >= 127)
1600 return false;
1601 if (*s == '/')
1602 return false;
1603 if (*s == '.') {
1604
1605 if (dot) /* Don't allow two dots next to each other (or at the beginning) */
1606 return false;
1607
1608 dot = true;
1609 } else
1610 dot = false;
1611 }
1612
1613 if (dot) /* don't allow a dot at the end */
1614 return false;
1615
1616 return true;
1617}
1618
1619static int oci_sysctl(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1620 Settings *s = userdata;
33d60b8d
LP		 1621 JsonVariant *w;
1622 const char *k;
de40a303
LP		 1623 int r;
1624
1625 assert(s);
1626
1627 JSON_VARIANT_OBJECT_FOREACH(k, w, v) {
33d60b8d 1628 const char *m;
de40a303 1629
19130626
ZJS		 1630 if (!json_variant_is_string(w))
1631 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1632 "sysctl parameter is not a string, refusing.");
de40a303 1633
de40a303
LP		 1634 assert_se(m = json_variant_string(w));
1635
33d60b8d 1636 if (sysctl_key_valid(k))
19130626 1637 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
33d60b8d 1638 "sysctl key invalid, refusing: %s", k);
de40a303 1639
33d60b8d 1640 r = strv_extend_strv(&s->sysctl, STRV_MAKE(k, m), false);
de40a303
LP		 1641 if (r < 0)
1642 return log_oom();
1643 }
1644
1645 return 0;
1646}
1647
ae408d77 1648#if HAVE_SECCOMP
de40a303
LP		 1649static int oci_seccomp_action_from_string(const char *name, uint32_t *ret) {
1650
1651 static const struct {
1652 const char *name;
1653 uint32_t action;
1654 } table[] = {
f9a3d8e2
LP		 1655 { "SCMP_ACT_ALLOW", SCMP_ACT_ALLOW },
1656 { "SCMP_ACT_ERRNO", SCMP_ACT_ERRNO(EPERM) }, /* the OCI spec doesn't document the error, but it appears EPERM is supposed to be used */
1657 { "SCMP_ACT_KILL", SCMP_ACT_KILL },
1658#ifdef SCMP_ACT_KILL_PROCESS
1659 { "SCMP_ACT_KILL_PROCESS", SCMP_ACT_KILL_PROCESS },
1660#endif
1661#ifdef SCMP_ACT_KILL_THREAD
1662 { "SCMP_ACT_KILL_THREAD", SCMP_ACT_KILL_THREAD },
1663#endif
de40a303 1664#ifdef SCMP_ACT_LOG
f9a3d8e2 1665 { "SCMP_ACT_LOG", SCMP_ACT_LOG },
de40a303 1666#endif
f9a3d8e2 1667 { "SCMP_ACT_TRAP", SCMP_ACT_TRAP },
de40a303
LP		 1668
1669 /* We don't support SCMP_ACT_TRACE because that requires a tracer, and that doesn't really make sense
1670 * here */
1671 };
1672
1673 size_t i;
1674
1675 for (i = 0; i < ELEMENTSOF(table); i++)
1676 if (streq_ptr(name, table[i].name)) {
1677 *ret = table[i].action;
1678 return 0;
1679 }
1680
1681 return -EINVAL;
1682}
1683
1684static int oci_seccomp_arch_from_string(const char *name, uint32_t *ret) {
1685
1686 static const struct {
1687 const char *name;
1688 uint32_t arch;
1689 } table[] = {
1690 { "SCMP_ARCH_AARCH64", SCMP_ARCH_AARCH64 },
1691 { "SCMP_ARCH_ARM", SCMP_ARCH_ARM },
1692 { "SCMP_ARCH_MIPS", SCMP_ARCH_MIPS },
1693 { "SCMP_ARCH_MIPS64", SCMP_ARCH_MIPS64 },
1694 { "SCMP_ARCH_MIPS64N32", SCMP_ARCH_MIPS64N32 },
1695 { "SCMP_ARCH_MIPSEL", SCMP_ARCH_MIPSEL },
1696 { "SCMP_ARCH_MIPSEL64", SCMP_ARCH_MIPSEL64 },
1697 { "SCMP_ARCH_MIPSEL64N32", SCMP_ARCH_MIPSEL64N32 },
1698 { "SCMP_ARCH_NATIVE", SCMP_ARCH_NATIVE },
1699#ifdef SCMP_ARCH_PARISC
1700 { "SCMP_ARCH_PARISC", SCMP_ARCH_PARISC },
1701#endif
1702#ifdef SCMP_ARCH_PARISC64
1703 { "SCMP_ARCH_PARISC64", SCMP_ARCH_PARISC64 },
1704#endif
1705 { "SCMP_ARCH_PPC", SCMP_ARCH_PPC },
1706 { "SCMP_ARCH_PPC64", SCMP_ARCH_PPC64 },
1707 { "SCMP_ARCH_PPC64LE", SCMP_ARCH_PPC64LE },
1708 { "SCMP_ARCH_S390", SCMP_ARCH_S390 },
1709 { "SCMP_ARCH_S390X", SCMP_ARCH_S390X },
1710 { "SCMP_ARCH_X32", SCMP_ARCH_X32 },
1711 { "SCMP_ARCH_X86", SCMP_ARCH_X86 },
1712 { "SCMP_ARCH_X86_64", SCMP_ARCH_X86_64 },
1713 };
1714
1715 size_t i;
1716
1717 for (i = 0; i < ELEMENTSOF(table); i++)
1718 if (streq_ptr(table[i].name, name)) {
1719 *ret = table[i].arch;
1720 return 0;
1721 }
1722
1723 return -EINVAL;
1724}
1725
1726static int oci_seccomp_compare_from_string(const char *name, enum scmp_compare *ret) {
1727
1728 static const struct {
1729 const char *name;
1730 enum scmp_compare op;
1731 } table[] = {
1732 { "SCMP_CMP_NE", SCMP_CMP_NE },
1733 { "SCMP_CMP_LT", SCMP_CMP_LT },
1734 { "SCMP_CMP_LE", SCMP_CMP_LE },
1735 { "SCMP_CMP_EQ", SCMP_CMP_EQ },
1736 { "SCMP_CMP_GE", SCMP_CMP_GE },
1737 { "SCMP_CMP_GT", SCMP_CMP_GT },
1738 { "SCMP_CMP_MASKED_EQ", SCMP_CMP_MASKED_EQ },
1739 };
1740
1741 size_t i;
1742
1743 for (i = 0; i < ELEMENTSOF(table); i++)
1744 if (streq_ptr(table[i].name, name)) {
1745 *ret = table[i].op;
1746 return 0;
1747 }
1748
1749 return -EINVAL;
1750}
1751
1752static int oci_seccomp_archs(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1753 scmp_filter_ctx *sc = userdata;
1754 JsonVariant *e;
1755 int r;
1756
1757 assert(sc);
1758
1759 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1760 uint32_t a;
1761
19130626
ZJS		 1762 if (!json_variant_is_string(e))
1763 return json_log(e, flags, SYNTHETIC_ERRNO(EINVAL),
1764 "Architecture entry is not a string");
de40a303
LP		 1765
1766 r = oci_seccomp_arch_from_string(json_variant_string(e), &a);
1767 if (r < 0)
1768 return json_log(e, flags, r, "Unknown architecture: %s", json_variant_string(e));
1769
1770 r = seccomp_arch_add(sc, a);
1771 if (r == -EEXIST)
1772 continue;
1773 if (r < 0)
1774 return json_log(e, flags, r, "Failed to add architecture to seccomp filter: %m");
1775 }
1776
1777 return 0;
1778}
1779
1780struct syscall_rule {
1781 char **names;
1782 uint32_t action;
1783 struct scmp_arg_cmp *arguments;
1784 size_t n_arguments;
1785};
1786
1787static void syscall_rule_free(struct syscall_rule *rule) {
1788 assert(rule);
1789
1790 strv_free(rule->names);
1791 free(rule->arguments);
1792};
1793
1794static int oci_seccomp_action(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1795 uint32_t *action = userdata;
1796 int r;
1797
1798 assert(action);
1799
1800 r = oci_seccomp_action_from_string(json_variant_string(v), action);
1801 if (r < 0)
1802 return json_log(v, flags, r, "Unknown system call action '%s': %m", json_variant_string(v));
1803
1804 return 0;
1805}
1806
1807static int oci_seccomp_op(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1808 enum scmp_compare *op = userdata;
1809 int r;
1810
1811 assert(op);
1812
1813 r = oci_seccomp_compare_from_string(json_variant_string(v), op);
1814 if (r < 0)
1815 return json_log(v, flags, r, "Unknown seccomp operator '%s': %m", json_variant_string(v));
1816
1817 return 0;
1818}
1819
1820static int oci_seccomp_args(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1821 struct syscall_rule *rule = userdata;
1822 JsonVariant *e;
1823 int r;
1824
1825 assert(rule);
1826
1827 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1828 static const struct JsonDispatch table[] = {
1829 { "index", JSON_VARIANT_UNSIGNED, json_dispatch_uint32, offsetof(struct scmp_arg_cmp, arg), JSON_MANDATORY },
1830 { "value", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct scmp_arg_cmp, datum_a), JSON_MANDATORY },
1831 { "valueTwo", JSON_VARIANT_UNSIGNED, json_dispatch_uint64, offsetof(struct scmp_arg_cmp, datum_b), 0 },
1832 { "op", JSON_VARIANT_STRING, oci_seccomp_op, offsetof(struct scmp_arg_cmp, op), JSON_MANDATORY },
1833 {},
1834 };
1835
1836 struct scmp_arg_cmp *a, *p;
1837 int expected;
1838
1839 a = reallocarray(rule->arguments, rule->n_arguments + 1, sizeof(struct syscall_rule));
1840 if (!a)
1841 return log_oom();
1842
1843 rule->arguments = a;
1844 p = rule->arguments + rule->n_arguments;
1845
1846 *p = (struct scmp_arg_cmp) {
1847 .arg = 0,
1848 .datum_a = 0,
1849 .datum_b = 0,
1850 .op = 0,
1851 };
1852
1853 r = json_dispatch(e, table, oci_unexpected, flags, p);
1854 if (r < 0)
1855 return r;
1856
1857 expected = p->op == SCMP_CMP_MASKED_EQ ? 4 : 3;
1858 if (r != expected)
1859 json_log(e, flags|JSON_WARNING, 0, "Wrong number of system call arguments for JSON data data, ignoring.");
1860
1861 /* Note that we are a bit sloppy here and do not insist that SCMP_CMP_MASKED_EQ gets two datum values,
1862 * and the other only one. That's because buildah for example by default calls things with
1863 * SCMP_CMP_MASKED_EQ but only one argument. We use 0 when the value is not specified. */
1864
1865 rule->n_arguments++;
1866 }
1867
1868 return 0;
1869}
1870
1871static int oci_seccomp_syscalls(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1872 scmp_filter_ctx *sc = userdata;
1873 JsonVariant *e;
1874 int r;
1875
1876 assert(sc);
1877
1878 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1879 static const JsonDispatch table[] = {
1880 { "names", JSON_VARIANT_ARRAY, json_dispatch_strv, offsetof(struct syscall_rule, names), JSON_MANDATORY },
1881 { "action", JSON_VARIANT_STRING, oci_seccomp_action, offsetof(struct syscall_rule, action), JSON_MANDATORY },
1882 { "args", JSON_VARIANT_ARRAY, oci_seccomp_args, 0, 0 },
1883 };
1884 struct syscall_rule rule = {
1885 .action = (uint32_t) -1,
1886 };
1887 char **i;
1888
1889 r = json_dispatch(e, table, oci_unexpected, flags, &rule);
1890 if (r < 0)
1891 goto fail_rule;
1892
1893 if (strv_isempty(rule.names)) {
1894 json_log(e, flags, 0, "System call name list is empty.");
1895 r = -EINVAL;
1896 goto fail_rule;
1897 }
1898
1899 STRV_FOREACH(i, rule.names) {
1900 int nr;
1901
1902 nr = seccomp_syscall_resolve_name(*i);
1903 if (nr == __NR_SCMP_ERROR) {
1904 log_debug("Unknown syscall %s, skipping.", *i);
1905 continue;
1906 }
1907
1908 r = seccomp_rule_add_array(sc, rule.action, nr, rule.n_arguments, rule.arguments);
1909 if (r < 0)
1910 goto fail_rule;
1911 }
1912
1913 syscall_rule_free(&rule);
1914 continue;
1915
1916 fail_rule:
1917 syscall_rule_free(&rule);
1918 return r;
1919 }
1920
1921 return 0;
1922}
ae408d77 1923#endif
de40a303
LP		 1924
1925static int oci_seccomp(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1926
ae408d77 1927#if HAVE_SECCOMP
de40a303
LP		 1928 static const JsonDispatch table[] = {
1929 { "defaultAction", JSON_VARIANT_STRING, NULL, 0, JSON_MANDATORY },
1930 { "architectures", JSON_VARIANT_ARRAY, oci_seccomp_archs, 0, 0 },
1931 { "syscalls", JSON_VARIANT_ARRAY, oci_seccomp_syscalls, 0, 0 },
1932 {}
1933 };
1934
1935 _cleanup_(seccomp_releasep) scmp_filter_ctx sc = NULL;
1936 Settings *s = userdata;
1937 JsonVariant *def;
1938 uint32_t d;
1939 int r;
1940
1941 assert(s);
1942
1943 def = json_variant_by_key(v, "defaultAction");
19130626
ZJS		 1944 if (!def)
1945 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL), "defaultAction element missing.");
de40a303 1946
19130626
ZJS		 1947 if (!json_variant_is_string(def))
1948 return json_log(def, flags, SYNTHETIC_ERRNO(EINVAL), "defaultAction is not a string.");
de40a303
LP		 1949
1950 r = oci_seccomp_action_from_string(json_variant_string(def), &d);
1951 if (r < 0)
1952 return json_log(def, flags, r, "Unknown default action: %s", json_variant_string(def));
1953
1954 sc = seccomp_init(d);
19130626 1955 if (!sc)
ae408d77 1956 return json_log(v, flags, SYNTHETIC_ERRNO(ENOMEM), "Couldn't allocate seccomp object.");
de40a303
LP		 1957
1958 r = json_dispatch(v, table, oci_unexpected, flags, sc);
1959 if (r < 0)
1960 return r;
1961
1962 seccomp_release(s->seccomp);
1963 s->seccomp = TAKE_PTR(sc);
de40a303 1964 return 0;
ae408d77
LP		 1965#else
1966 return json_log(v, flags, SYNTHETIC_ERRNO(EOPNOTSUPP), "libseccomp support not enabled, can't parse seccomp object.");
1967#endif
de40a303
LP		 1968}
1969
1970static int oci_rootfs_propagation(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1971 const char *s;
1972
1973 s = json_variant_string(v);
1974
1975 if (streq(s, "shared"))
1976 return 0;
1977
1978 json_log(v, flags|JSON_DEBUG, 0, "Ignoring rootfsPropagation setting '%s'.", s);
1979 return 0;
1980}
1981
1982static int oci_masked_paths(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
1983 Settings *s = userdata;
1984 JsonVariant *e;
1985
1986 assert(s);
1987
1988 JSON_VARIANT_ARRAY_FOREACH(e, v) {
1989 _cleanup_free_ char *destination = NULL;
1990 CustomMount *m;
1991 const char *p;
1992
19130626
ZJS		 1993 if (!json_variant_is_string(e))
1994 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
1995 "Path is not a string, refusing.");
de40a303
LP		 1996
1997 assert_se(p = json_variant_string(e));
1998
19130626
ZJS		 1999 if (!path_is_absolute(p))
2000 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
2001 "Path is not not absolute, refusing: %s", p);
de40a303
LP		 2002
2003 if (oci_exclude_mount(p))
2004 continue;
2005
2006 destination = strdup(p);
2007 if (!destination)
2008 return log_oom();
2009
2010 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_INACCESSIBLE);
2011 if (!m)
2012 return log_oom();
2013
2014 m->destination = TAKE_PTR(destination);
2015
2016 /* The spec doesn't say this, but apparently pre-existing implementations are lenient towards
2017 * non-existing paths to mask. Let's hence be too. */
2018 m->graceful = true;
2019 }
2020
2021 return 0;
2022}
2023
2024static int oci_readonly_paths(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2025 Settings *s = userdata;
2026 JsonVariant *e;
2027
2028 assert(s);
2029
2030 JSON_VARIANT_ARRAY_FOREACH(e, v) {
2031 _cleanup_free_ char *source = NULL, *destination = NULL;
2032 CustomMount *m;
2033 const char *p;
2034
19130626
ZJS		 2035 if (!json_variant_is_string(e))
2036 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
2037 "Path is not a string, refusing.");
de40a303
LP		 2038
2039 assert_se(p = json_variant_string(e));
2040
19130626
ZJS		 2041 if (!path_is_absolute(p))
2042 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
2043 "Path is not not absolute, refusing: %s", p);
de40a303
LP		 2044
2045 if (oci_exclude_mount(p))
2046 continue;
2047
2048 source = strjoin("+", p);
2049 if (!source)
2050 return log_oom();
2051
2052 destination = strdup(p);
2053 if (!destination)
2054 return log_oom();
2055
2056 m = custom_mount_add(&s->custom_mounts, &s->n_custom_mounts, CUSTOM_MOUNT_BIND);
2057 if (!m)
2058 return log_oom();
2059
2060 m->source = TAKE_PTR(source);
2061 m->destination = TAKE_PTR(destination);
2062 m->read_only = true;
2063 }
2064
2065 return 0;
2066}
2067
2068static int oci_linux(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2069
2070 static const JsonDispatch table[] = {
2071 { "namespaces", JSON_VARIANT_ARRAY, oci_namespaces, 0, 0 },
2072 { "uidMappings", JSON_VARIANT_ARRAY, oci_uid_gid_mappings, 0, 0 },
2073 { "gidMappings", JSON_VARIANT_ARRAY, oci_uid_gid_mappings, 0, 0 },
2074 { "devices", JSON_VARIANT_ARRAY, oci_devices, 0, 0 },
2075 { "cgroupsPath", JSON_VARIANT_STRING, oci_cgroups_path, 0, 0 },
2076 { "resources", JSON_VARIANT_OBJECT, oci_resources, 0, 0 },
2077 { "intelRdt", JSON_VARIANT_OBJECT, oci_unsupported, 0, JSON_PERMISSIVE },
2078 { "sysctl", JSON_VARIANT_OBJECT, oci_sysctl, 0, 0 },
2079 { "seccomp", JSON_VARIANT_OBJECT, oci_seccomp, 0, 0 },
2080 { "rootfsPropagation", JSON_VARIANT_STRING, oci_rootfs_propagation, 0, 0 },
2081 { "maskedPaths", JSON_VARIANT_ARRAY, oci_masked_paths, 0, 0 },
2082 { "readonlyPaths", JSON_VARIANT_ARRAY, oci_readonly_paths, 0, 0 },
2083 { "mountLabel", JSON_VARIANT_STRING, oci_unsupported, 0, JSON_PERMISSIVE },
2084 {}
2085 };
2086
2087 return json_dispatch(v, table, oci_unexpected, flags, userdata);
2088}
2089
2090static int oci_hook_timeout(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2091 usec_t *u = userdata;
2092 uintmax_t k;
2093
2094 k = json_variant_unsigned(v);
fa8b675a
ZJS		 2095 if (k == 0 || k > (UINT64_MAX-1)/USEC_PER_SEC)
2096 return json_log(v, flags, SYNTHETIC_ERRNO(ERANGE),
2097 "Hook timeout value out of range.");
de40a303
LP		 2098
2099 *u = k * USEC_PER_SEC;
2100 return 0;
2101}
2102
2103static int oci_hooks_array(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2104 Settings *s = userdata;
2105 JsonVariant *e;
2106 int r;
2107
2108 assert(s);
2109
2110 JSON_VARIANT_ARRAY_FOREACH(e, v) {
2111
2112 static const JsonDispatch table[] = {
2113 { "path", JSON_VARIANT_STRING, oci_absolute_path, offsetof(OciHook, path), JSON_MANDATORY },
2114 { "args", JSON_VARIANT_ARRAY, oci_args, offsetof(OciHook, args), 0 },
2115 { "env", JSON_VARIANT_ARRAY, oci_env, offsetof(OciHook, env), 0 },
2116 { "timeout", JSON_VARIANT_UNSIGNED, oci_hook_timeout, offsetof(OciHook, timeout), 0 },
2117 {}
2118 };
2119
2120 OciHook *a, **array, *new_item;
2121 size_t *n_array;
2122
2123 if (streq(name, "prestart")) {
2124 array = &s->oci_hooks_prestart;
2125 n_array = &s->n_oci_hooks_prestart;
2126 } else if (streq(name, "poststart")) {
2127 array = &s->oci_hooks_poststart;
2128 n_array = &s->n_oci_hooks_poststart;
2129 } else {
2130 assert(streq(name, "poststop"));
2131 array = &s->oci_hooks_poststop;
2132 n_array = &s->n_oci_hooks_poststop;
2133 }
2134
2135 a = reallocarray(*array, *n_array + 1, sizeof(OciHook));
2136 if (!a)
2137 return log_oom();
2138
2139 *array = a;
2140 new_item = a + *n_array;
2141
2142 *new_item = (OciHook) {
2143 .timeout = USEC_INFINITY,
2144 };
2145
2146 r = json_dispatch(e, table, oci_unexpected, flags, userdata);
2147 if (r < 0) {
2148 free(new_item->path);
2149 strv_free(new_item->args);
2150 strv_free(new_item->env);
2151 return r;
2152 }
2153
2154 (*n_array) ++;
2155 }
2156
2157 return 0;
2158}
2159
2160static int oci_hooks(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
2161
2162 static const JsonDispatch table[] = {
2163 { "prestart", JSON_VARIANT_OBJECT, oci_hooks_array, 0, 0 },
2164 { "poststart", JSON_VARIANT_OBJECT, oci_hooks_array, 0, 0 },
2165 { "poststop", JSON_VARIANT_OBJECT, oci_hooks_array, 0, 0 },
2166 {}
2167 };
2168
2169 return json_dispatch(v, table, oci_unexpected, flags, userdata);
2170}
2171
2172static int oci_annotations(const char *name, JsonVariant *v, JsonDispatchFlags flags, void *userdata) {
33d60b8d
LP		 2173 JsonVariant *w;
2174 const char *k;
de40a303
LP		 2175
2176 JSON_VARIANT_OBJECT_FOREACH(k, w, v) {
de40a303 2177
33d60b8d
LP		 2178 if (isempty(k))
2179 return json_log(v, flags, SYNTHETIC_ERRNO(EINVAL),
19130626 2180 "Annotation with empty key, refusing.");
de40a303 2181
19130626
ZJS		 2182 if (!json_variant_is_string(w))
2183 return json_log(w, flags, SYNTHETIC_ERRNO(EINVAL),
2184 "Annotation has non-string value, refusing.");
de40a303 2185
33d60b8d 2186 json_log(w, flags|JSON_DEBUG, 0, "Ignoring annotation '%s' with value '%s'.", k, json_variant_string(w));
de40a303
LP		 2187 }
2188
2189 return 0;
2190}
2191
2192int oci_load(FILE *f, const char *bundle, Settings **ret) {
2193
2194 static const JsonDispatch table[] = {
2195 { "ociVersion", JSON_VARIANT_STRING, NULL, 0, JSON_MANDATORY },
2196 { "process", JSON_VARIANT_OBJECT, oci_process, 0, 0 },
2197 { "root", JSON_VARIANT_OBJECT, oci_root, 0, 0 },
2198 { "hostname", JSON_VARIANT_STRING, oci_hostname, 0, 0 },
2199 { "mounts", JSON_VARIANT_ARRAY, oci_mounts, 0, 0 },
2200 { "linux", JSON_VARIANT_OBJECT, oci_linux, 0, 0 },
2201 { "hooks", JSON_VARIANT_OBJECT, oci_hooks, 0, 0 },
2202 { "annotations", JSON_VARIANT_OBJECT, oci_annotations, 0, 0 },
2203 {}
2204 };
2205
2206 _cleanup_(json_variant_unrefp) JsonVariant *oci = NULL;
2207 _cleanup_(settings_freep) Settings *s = NULL;
2208 unsigned line = 0, column = 0;
2209 JsonVariant *v;
2210 const char *path;
2211 int r;
2212
2213 assert_se(bundle);
2214
2215 path = strjoina(bundle, "/config.json");
2216
2217 r = json_parse_file(f, path, &oci, &line, &column);
2218 if (r < 0) {
2219 if (line != 0 && column != 0)
2220 return log_error_errno(r, "Failed to parse '%s' at %u:%u: %m", path, line, column);
2221 else
2222 return log_error_errno(r, "Failed to parse '%s': %m", path);
2223 }
2224
2225 v = json_variant_by_key(oci, "ociVersion");
2226 if (!v) {
2227 log_error("JSON file '%s' is not an OCI bundle configuration file. Refusing.", path);
2228 return -EINVAL;
2229 }
2230 if (!streq_ptr(json_variant_string(v), "1.0.0")) {
2231 log_error("OCI bundle version not supported: %s", strna(json_variant_string(v)));
2232 return -EINVAL;
2233 }
2234
2235 // {
2236 // _cleanup_free_ char *formatted = NULL;
2237 // assert_se(json_variant_format(oci, JSON_FORMAT_PRETTY|JSON_FORMAT_COLOR, &formatted) >= 0);
2238 // fputs(formatted, stdout);
2239 // }
2240
2241 s = settings_new();
2242 if (!s)
2243 return log_oom();
2244
2245 s->start_mode = START_PID1;
2246 s->resolv_conf = RESOLV_CONF_OFF;
2247 s->link_journal = LINK_NO;
2248 s->timezone = TIMEZONE_OFF;
2249
2250 s->bundle = strdup(bundle);
2251 if (!s->bundle)
2252 return log_oom();
2253
2254 r = json_dispatch(oci, table, oci_unexpected, 0, s);
2255 if (r < 0)
2256 return r;
2257
2258 if (s->properties) {
2259 r = sd_bus_message_seal(s->properties, 0, 0);
2260 if (r < 0)
2261 return log_error_errno(r, "Cannot seal properties bus message: %m");
2262 }
2263
2264 *ret = TAKE_PTR(s);
2265 return 0;
2266}
Unnamed repository; edit this file 'description' to name the repository.