]> git.ipfire.org Git - thirdparty/systemd.git/blob - src/nspawn/nspawn-cgroup.c
Merge pull request #11827 from keszybz/pkgconfig-variables
[thirdparty/systemd.git] / src / nspawn / nspawn-cgroup.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
2
3 #include <sys/mount.h>
4
5 #include "alloc-util.h"
6 #include "fd-util.h"
7 #include "fileio.h"
8 #include "fs-util.h"
9 #include "mkdir.h"
10 #include "mount-util.h"
11 #include "mountpoint-util.h"
12 #include "nspawn-cgroup.h"
13 #include "nspawn-mount.h"
14 #include "path-util.h"
15 #include "rm-rf.h"
16 #include "string-util.h"
17 #include "strv.h"
18 #include "user-util.h"
19 #include "util.h"
20
21 static int chown_cgroup_path(const char *path, uid_t uid_shift) {
22 _cleanup_close_ int fd = -1;
23 const char *fn;
24
25 fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY);
26 if (fd < 0)
27 return -errno;
28
29 FOREACH_STRING(fn,
30 ".",
31 "cgroup.clone_children",
32 "cgroup.controllers",
33 "cgroup.events",
34 "cgroup.procs",
35 "cgroup.stat",
36 "cgroup.subtree_control",
37 "cgroup.threads",
38 "notify_on_release",
39 "tasks")
40 if (fchownat(fd, fn, uid_shift, uid_shift, 0) < 0)
41 log_full_errno(errno == ENOENT ? LOG_DEBUG : LOG_WARNING, errno,
42 "Failed to chown \"%s/%s\", ignoring: %m", path, fn);
43
44 return 0;
45 }
46
47 int chown_cgroup(pid_t pid, CGroupUnified unified_requested, uid_t uid_shift) {
48 _cleanup_free_ char *path = NULL, *fs = NULL;
49 int r;
50
51 r = cg_pid_get_path(NULL, pid, &path);
52 if (r < 0)
53 return log_error_errno(r, "Failed to get container cgroup path: %m");
54
55 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &fs);
56 if (r < 0)
57 return log_error_errno(r, "Failed to get file system path for container cgroup: %m");
58
59 r = chown_cgroup_path(fs, uid_shift);
60 if (r < 0)
61 return log_error_errno(r, "Failed to chown() cgroup %s: %m", fs);
62
63 if (unified_requested == CGROUP_UNIFIED_SYSTEMD || (unified_requested == CGROUP_UNIFIED_NONE && cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) > 0)) {
64 _cleanup_free_ char *lfs = NULL;
65 /* Always propagate access rights from unified to legacy controller */
66
67 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, NULL, &lfs);
68 if (r < 0)
69 return log_error_errno(r, "Failed to get file system path for container cgroup: %m");
70
71 r = chown_cgroup_path(lfs, uid_shift);
72 if (r < 0)
73 return log_error_errno(r, "Failed to chown() cgroup %s: %m", lfs);
74 }
75
76 return 0;
77 }
78
79 int sync_cgroup(pid_t pid, CGroupUnified unified_requested, uid_t uid_shift) {
80 _cleanup_free_ char *cgroup = NULL;
81 char tree[] = "/tmp/unifiedXXXXXX", pid_string[DECIMAL_STR_MAX(pid) + 1];
82 bool undo_mount = false;
83 const char *fn;
84 int r, unified_controller;
85
86 unified_controller = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
87 if (unified_controller < 0)
88 return log_error_errno(unified_controller, "Failed to determine whether the systemd hierarchy is unified: %m");
89 if ((unified_controller > 0) == (unified_requested >= CGROUP_UNIFIED_SYSTEMD))
90 return 0;
91
92 /* When the host uses the legacy cgroup setup, but the
93 * container shall use the unified hierarchy, let's make sure
94 * we copy the path from the name=systemd hierarchy into the
95 * unified hierarchy. Similar for the reverse situation. */
96
97 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
98 if (r < 0)
99 return log_error_errno(r, "Failed to get control group of " PID_FMT ": %m", pid);
100
101 /* In order to access the unified hierarchy we need to mount it */
102 if (!mkdtemp(tree))
103 return log_error_errno(errno, "Failed to generate temporary mount point for unified hierarchy: %m");
104
105 if (unified_controller > 0)
106 r = mount_verbose(LOG_ERR, "cgroup", tree, "cgroup",
107 MS_NOSUID|MS_NOEXEC|MS_NODEV, "none,name=systemd,xattr");
108 else
109 r = mount_verbose(LOG_ERR, "cgroup", tree, "cgroup2",
110 MS_NOSUID|MS_NOEXEC|MS_NODEV, NULL);
111 if (r < 0)
112 goto finish;
113
114 undo_mount = true;
115
116 /* If nspawn dies abruptly the cgroup hierarchy created below
117 * its unit isn't cleaned up. So, let's remove it
118 * https://github.com/systemd/systemd/pull/4223#issuecomment-252519810 */
119 fn = strjoina(tree, cgroup);
120 (void) rm_rf(fn, REMOVE_ROOT|REMOVE_ONLY_DIRECTORIES);
121
122 fn = strjoina(tree, cgroup, "/cgroup.procs");
123 (void) mkdir_parents(fn, 0755);
124
125 sprintf(pid_string, PID_FMT, pid);
126 r = write_string_file(fn, pid_string, WRITE_STRING_FILE_DISABLE_BUFFER);
127 if (r < 0) {
128 log_error_errno(r, "Failed to move process: %m");
129 goto finish;
130 }
131
132 fn = strjoina(tree, cgroup);
133 r = chown_cgroup_path(fn, uid_shift);
134 if (r < 0)
135 log_error_errno(r, "Failed to chown() cgroup %s: %m", fn);
136 finish:
137 if (undo_mount)
138 (void) umount_verbose(tree);
139
140 (void) rmdir(tree);
141 return r;
142 }
143
144 int create_subcgroup(pid_t pid, bool keep_unit, CGroupUnified unified_requested) {
145 _cleanup_free_ char *cgroup = NULL;
146 CGroupMask supported;
147 const char *payload;
148 int r;
149
150 assert(pid > 1);
151
152 /* In the unified hierarchy inner nodes may only contain subgroups, but not processes. Hence, if we running in
153 * the unified hierarchy and the container does the same, and we did not create a scope unit for the container
154 * move us and the container into two separate subcgroups.
155 *
156 * Moreover, container payloads such as systemd try to manage the cgroup they run in in full (i.e. including
157 * its attributes), while the host systemd will only delegate cgroups for children of the cgroup created for a
158 * delegation unit, instead of the cgroup itself. This means, if we'd pass on the cgroup allocated from the
159 * host systemd directly to the payload, the host and payload systemd might fight for the cgroup
160 * attributes. Hence, let's insert an intermediary cgroup to cover that case too.
161 *
162 * Note that we only bother with the main hierarchy here, not with any secondary ones. On the unified setup
163 * that's fine because there's only one hiearchy anyway and controllers are enabled directly on it. On the
164 * legacy setup, this is fine too, since delegation of controllers is generally not safe there, hence we won't
165 * do it. */
166
167 r = cg_mask_supported(&supported);
168 if (r < 0)
169 return log_error_errno(r, "Failed to determine supported controllers: %m");
170
171 if (keep_unit)
172 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 0, &cgroup);
173 else
174 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
175 if (r < 0)
176 return log_error_errno(r, "Failed to get our control group: %m");
177
178 payload = strjoina(cgroup, "/payload");
179 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, payload, pid);
180 if (r < 0)
181 return log_error_errno(r, "Failed to create %s subcgroup: %m", payload);
182
183 if (keep_unit) {
184 const char *supervisor;
185
186 supervisor = strjoina(cgroup, "/supervisor");
187 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, supervisor, 0);
188 if (r < 0)
189 return log_error_errno(r, "Failed to create %s subcgroup: %m", supervisor);
190 }
191
192 /* Try to enable as many controllers as possible for the new payload. */
193 (void) cg_enable_everywhere(supported, supported, cgroup, NULL);
194 return 0;
195 }
196
197 /* Retrieve existing subsystems. This function is called in a new cgroup
198 * namespace.
199 */
200 static int get_process_controllers(Set **ret) {
201 _cleanup_set_free_free_ Set *controllers = NULL;
202 _cleanup_fclose_ FILE *f = NULL;
203 int r;
204
205 assert(ret);
206
207 controllers = set_new(&string_hash_ops);
208 if (!controllers)
209 return -ENOMEM;
210
211 f = fopen("/proc/self/cgroup", "re");
212 if (!f)
213 return errno == ENOENT ? -ESRCH : -errno;
214
215 for (;;) {
216 _cleanup_free_ char *line = NULL;
217 char *e, *l;
218
219 r = read_line(f, LONG_LINE_MAX, &line);
220 if (r < 0)
221 return r;
222 if (r == 0)
223 break;
224
225 l = strchr(line, ':');
226 if (!l)
227 continue;
228
229 l++;
230 e = strchr(l, ':');
231 if (!e)
232 continue;
233
234 *e = 0;
235
236 if (STR_IN_SET(l, "", "name=systemd", "name=unified"))
237 continue;
238
239 r = set_put_strdup(controllers, l);
240 if (r < 0)
241 return r;
242 }
243
244 *ret = TAKE_PTR(controllers);
245
246 return 0;
247 }
248
249 static int mount_legacy_cgroup_hierarchy(
250 const char *dest,
251 const char *controller,
252 const char *hierarchy,
253 bool read_only) {
254
255 const char *to, *fstype, *opts;
256 int r;
257
258 to = strjoina(strempty(dest), "/sys/fs/cgroup/", hierarchy);
259
260 r = path_is_mount_point(to, dest, 0);
261 if (r < 0 && r != -ENOENT)
262 return log_error_errno(r, "Failed to determine if %s is mounted already: %m", to);
263 if (r > 0)
264 return 0;
265
266 mkdir_p(to, 0755);
267
268 /* The superblock mount options of the mount point need to be
269 * identical to the hosts', and hence writable... */
270 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER_HYBRID)) {
271 fstype = "cgroup2";
272 opts = NULL;
273 } else if (streq(controller, SYSTEMD_CGROUP_CONTROLLER_LEGACY)) {
274 fstype = "cgroup";
275 opts = "none,name=systemd,xattr";
276 } else {
277 fstype = "cgroup";
278 opts = controller;
279 }
280
281 r = mount_verbose(LOG_ERR, "cgroup", to, fstype, MS_NOSUID|MS_NOEXEC|MS_NODEV, opts);
282 if (r < 0)
283 return r;
284
285 /* ... hence let's only make the bind mount read-only, not the superblock. */
286 if (read_only) {
287 r = mount_verbose(LOG_ERR, NULL, to, NULL,
288 MS_BIND|MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, NULL);
289 if (r < 0)
290 return r;
291 }
292
293 return 1;
294 }
295
296 /* Mount a legacy cgroup hierarchy when cgroup namespaces are supported. */
297 static int mount_legacy_cgns_supported(
298 const char *dest,
299 CGroupUnified unified_requested,
300 bool userns,
301 uid_t uid_shift,
302 uid_t uid_range,
303 const char *selinux_apifs_context) {
304
305 _cleanup_set_free_free_ Set *controllers = NULL;
306 const char *cgroup_root = "/sys/fs/cgroup", *c;
307 int r;
308
309 (void) mkdir_p(cgroup_root, 0755);
310
311 /* Mount a tmpfs to /sys/fs/cgroup if it's not mounted there yet. */
312 r = path_is_mount_point(cgroup_root, dest, AT_SYMLINK_FOLLOW);
313 if (r < 0)
314 return log_error_errno(r, "Failed to determine if /sys/fs/cgroup is already mounted: %m");
315 if (r == 0) {
316 _cleanup_free_ char *options = NULL;
317
318 /* When cgroup namespaces are enabled and user namespaces are
319 * used then the mount of the cgroupfs is done *inside* the new
320 * user namespace. We're root in the new user namespace and the
321 * kernel will happily translate our uid/gid to the correct
322 * uid/gid as seen from e.g. /proc/1/mountinfo. So we simply
323 * pass uid 0 and not uid_shift to tmpfs_patch_options().
324 */
325 r = tmpfs_patch_options("mode=755", 0, selinux_apifs_context, &options);
326 if (r < 0)
327 return log_oom();
328
329 r = mount_verbose(LOG_ERR, "tmpfs", cgroup_root, "tmpfs",
330 MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME, options);
331 if (r < 0)
332 return r;
333 }
334
335 r = cg_all_unified();
336 if (r < 0)
337 return r;
338 if (r > 0)
339 goto skip_controllers;
340
341 r = get_process_controllers(&controllers);
342 if (r < 0)
343 return log_error_errno(r, "Failed to determine cgroup controllers: %m");
344
345 for (;;) {
346 _cleanup_free_ const char *controller = NULL;
347
348 controller = set_steal_first(controllers);
349 if (!controller)
350 break;
351
352 r = mount_legacy_cgroup_hierarchy("", controller, controller, !userns);
353 if (r < 0)
354 return r;
355
356 /* When multiple hierarchies are co-mounted, make their
357 * constituting individual hierarchies a symlink to the
358 * co-mount.
359 */
360 c = controller;
361 for (;;) {
362 _cleanup_free_ char *target = NULL, *tok = NULL;
363
364 r = extract_first_word(&c, &tok, ",", 0);
365 if (r < 0)
366 return log_error_errno(r, "Failed to extract co-mounted cgroup controller: %m");
367 if (r == 0)
368 break;
369
370 if (streq(controller, tok))
371 break;
372
373 target = prefix_root("/sys/fs/cgroup/", tok);
374 if (!target)
375 return log_oom();
376
377 r = symlink_idempotent(controller, target, false);
378 if (r == -EINVAL)
379 return log_error_errno(r, "Invalid existing symlink for combined hierarchy: %m");
380 if (r < 0)
381 return log_error_errno(r, "Failed to create symlink for combined hierarchy: %m");
382 }
383 }
384
385 skip_controllers:
386 if (unified_requested >= CGROUP_UNIFIED_SYSTEMD) {
387 r = mount_legacy_cgroup_hierarchy("", SYSTEMD_CGROUP_CONTROLLER_HYBRID, "unified", false);
388 if (r < 0)
389 return r;
390 }
391
392 r = mount_legacy_cgroup_hierarchy("", SYSTEMD_CGROUP_CONTROLLER_LEGACY, "systemd", false);
393 if (r < 0)
394 return r;
395
396 if (!userns)
397 return mount_verbose(LOG_ERR, NULL, cgroup_root, NULL,
398 MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME|MS_RDONLY, "mode=755");
399
400 return 0;
401 }
402
403 /* Mount legacy cgroup hierarchy when cgroup namespaces are unsupported. */
404 static int mount_legacy_cgns_unsupported(
405 const char *dest,
406 CGroupUnified unified_requested,
407 bool userns,
408 uid_t uid_shift,
409 uid_t uid_range,
410 const char *selinux_apifs_context) {
411
412 _cleanup_set_free_free_ Set *controllers = NULL;
413 const char *cgroup_root;
414 int r;
415
416 cgroup_root = prefix_roota(dest, "/sys/fs/cgroup");
417
418 (void) mkdir_p(cgroup_root, 0755);
419
420 /* Mount a tmpfs to /sys/fs/cgroup if it's not mounted there yet. */
421 r = path_is_mount_point(cgroup_root, dest, AT_SYMLINK_FOLLOW);
422 if (r < 0)
423 return log_error_errno(r, "Failed to determine if /sys/fs/cgroup is already mounted: %m");
424 if (r == 0) {
425 _cleanup_free_ char *options = NULL;
426
427 r = tmpfs_patch_options("mode=755", uid_shift == 0 ? UID_INVALID : uid_shift, selinux_apifs_context, &options);
428 if (r < 0)
429 return log_oom();
430
431 r = mount_verbose(LOG_ERR, "tmpfs", cgroup_root, "tmpfs",
432 MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME, options);
433 if (r < 0)
434 return r;
435 }
436
437 r = cg_all_unified();
438 if (r < 0)
439 return r;
440 if (r > 0)
441 goto skip_controllers;
442
443 r = cg_kernel_controllers(&controllers);
444 if (r < 0)
445 return log_error_errno(r, "Failed to determine cgroup controllers: %m");
446
447 for (;;) {
448 _cleanup_free_ char *controller = NULL, *origin = NULL, *combined = NULL;
449
450 controller = set_steal_first(controllers);
451 if (!controller)
452 break;
453
454 origin = prefix_root("/sys/fs/cgroup/", controller);
455 if (!origin)
456 return log_oom();
457
458 r = readlink_malloc(origin, &combined);
459 if (r == -EINVAL) {
460 /* Not a symbolic link, but directly a single cgroup hierarchy */
461
462 r = mount_legacy_cgroup_hierarchy(dest, controller, controller, true);
463 if (r < 0)
464 return r;
465
466 } else if (r < 0)
467 return log_error_errno(r, "Failed to read link %s: %m", origin);
468 else {
469 _cleanup_free_ char *target = NULL;
470
471 target = prefix_root(dest, origin);
472 if (!target)
473 return log_oom();
474
475 /* A symbolic link, a combination of controllers in one hierarchy */
476
477 if (!filename_is_valid(combined)) {
478 log_warning("Ignoring invalid combined hierarchy %s.", combined);
479 continue;
480 }
481
482 r = mount_legacy_cgroup_hierarchy(dest, combined, combined, true);
483 if (r < 0)
484 return r;
485
486 r = symlink_idempotent(combined, target, false);
487 if (r == -EINVAL)
488 return log_error_errno(r, "Invalid existing symlink for combined hierarchy: %m");
489 if (r < 0)
490 return log_error_errno(r, "Failed to create symlink for combined hierarchy: %m");
491 }
492 }
493
494 skip_controllers:
495 if (unified_requested >= CGROUP_UNIFIED_SYSTEMD) {
496 r = mount_legacy_cgroup_hierarchy(dest, SYSTEMD_CGROUP_CONTROLLER_HYBRID, "unified", false);
497 if (r < 0)
498 return r;
499 }
500
501 r = mount_legacy_cgroup_hierarchy(dest, SYSTEMD_CGROUP_CONTROLLER_LEGACY, "systemd", false);
502 if (r < 0)
503 return r;
504
505 return mount_verbose(LOG_ERR, NULL, cgroup_root, NULL,
506 MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME|MS_RDONLY, "mode=755");
507 }
508
509 static int mount_unified_cgroups(const char *dest) {
510 const char *p;
511 int r;
512
513 assert(dest);
514
515 p = prefix_roota(dest, "/sys/fs/cgroup");
516
517 (void) mkdir_p(p, 0755);
518
519 r = path_is_mount_point(p, dest, AT_SYMLINK_FOLLOW);
520 if (r < 0)
521 return log_error_errno(r, "Failed to determine if %s is mounted already: %m", p);
522 if (r > 0) {
523 p = prefix_roota(dest, "/sys/fs/cgroup/cgroup.procs");
524 if (access(p, F_OK) >= 0)
525 return 0;
526 if (errno != ENOENT)
527 return log_error_errno(errno, "Failed to determine if mount point %s contains the unified cgroup hierarchy: %m", p);
528
529 return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
530 "%s is already mounted but not a unified cgroup hierarchy. Refusing.", p);
531 }
532
533 return mount_verbose(LOG_ERR, "cgroup", p, "cgroup2", MS_NOSUID|MS_NOEXEC|MS_NODEV, NULL);
534 }
535
536 int mount_cgroups(
537 const char *dest,
538 CGroupUnified unified_requested,
539 bool userns,
540 uid_t uid_shift,
541 uid_t uid_range,
542 const char *selinux_apifs_context,
543 bool use_cgns) {
544
545 if (unified_requested >= CGROUP_UNIFIED_ALL)
546 return mount_unified_cgroups(dest);
547 if (use_cgns)
548 return mount_legacy_cgns_supported(dest, unified_requested, userns, uid_shift, uid_range, selinux_apifs_context);
549
550 return mount_legacy_cgns_unsupported(dest, unified_requested, userns, uid_shift, uid_range, selinux_apifs_context);
551 }
552
553 static int mount_systemd_cgroup_writable_one(const char *root, const char *own) {
554 int r;
555
556 assert(root);
557 assert(own);
558
559 /* Make our own cgroup a (writable) bind mount */
560 r = mount_verbose(LOG_ERR, own, own, NULL, MS_BIND, NULL);
561 if (r < 0)
562 return r;
563
564 /* And then remount the systemd cgroup root read-only */
565 return mount_verbose(LOG_ERR, NULL, root, NULL,
566 MS_BIND|MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, NULL);
567 }
568
569 int mount_systemd_cgroup_writable(
570 const char *dest,
571 CGroupUnified unified_requested) {
572
573 _cleanup_free_ char *own_cgroup_path = NULL;
574 const char *root, *own;
575 int r;
576
577 assert(dest);
578
579 r = cg_pid_get_path(NULL, 0, &own_cgroup_path);
580 if (r < 0)
581 return log_error_errno(r, "Failed to determine our own cgroup path: %m");
582
583 /* If we are living in the top-level, then there's nothing to do... */
584 if (path_equal(own_cgroup_path, "/"))
585 return 0;
586
587 if (unified_requested >= CGROUP_UNIFIED_ALL) {
588
589 root = prefix_roota(dest, "/sys/fs/cgroup");
590 own = strjoina(root, own_cgroup_path);
591
592 } else {
593
594 if (unified_requested >= CGROUP_UNIFIED_SYSTEMD) {
595 root = prefix_roota(dest, "/sys/fs/cgroup/unified");
596 own = strjoina(root, own_cgroup_path);
597
598 r = mount_systemd_cgroup_writable_one(root, own);
599 if (r < 0)
600 return r;
601 }
602
603 root = prefix_roota(dest, "/sys/fs/cgroup/systemd");
604 own = strjoina(root, own_cgroup_path);
605 }
606
607 return mount_systemd_cgroup_writable_one(root, own);
608 }