1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
4 This file is part of systemd.
6 Copyright 2013 Lennart Poettering
8 systemd is free software; you can redistribute it and/or modify it
9 under the terms of the GNU Lesser General Public License as published by
10 the Free Software Foundation; either version 2.1 of the License, or
11 (at your option) any later version.
13 systemd is distributed in the hope that it will be useful, but
14 WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 Lesser General Public License for more details.
18 You should have received a copy of the GNU Lesser General Public License
19 along with systemd; If not, see <http://www.gnu.org/licenses/>.
25 #include "path-util.h"
27 #include "cgroup-util.h"
30 void cgroup_context_init(CGroupContext *c) {
33 /* Initialize everything to the kernel defaults, assuming the
34 * structure is preinitialized to 0 */
37 c->memory_limit = (uint64_t) -1;
38 c->blockio_weight = 1000;
41 void cgroup_context_free_device_allow(CGroupContext *c, CGroupDeviceAllow *a) {
45 LIST_REMOVE(device_allow, c->device_allow, a);
50 void cgroup_context_free_blockio_device_weight(CGroupContext *c, CGroupBlockIODeviceWeight *w) {
54 LIST_REMOVE(device_weights, c->blockio_device_weights, w);
59 void cgroup_context_free_blockio_device_bandwidth(CGroupContext *c, CGroupBlockIODeviceBandwidth *b) {
63 LIST_REMOVE(device_bandwidths, c->blockio_device_bandwidths, b);
68 void cgroup_context_done(CGroupContext *c) {
71 while (c->blockio_device_weights)
72 cgroup_context_free_blockio_device_weight(c, c->blockio_device_weights);
74 while (c->blockio_device_bandwidths)
75 cgroup_context_free_blockio_device_bandwidth(c, c->blockio_device_bandwidths);
77 while (c->device_allow)
78 cgroup_context_free_device_allow(c, c->device_allow);
81 void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
82 CGroupBlockIODeviceBandwidth *b;
83 CGroupBlockIODeviceWeight *w;
89 prefix = strempty(prefix);
92 "%sCPUAccounting=%s\n"
93 "%sBlockIOAccounting=%s\n"
94 "%sMemoryAccounting=%s\n"
96 "%sBlockIOWeight=%lu\n"
97 "%sMemoryLimit=%" PRIu64 "\n"
98 "%sDevicePolicy=%s\n",
99 prefix, yes_no(c->cpu_accounting),
100 prefix, yes_no(c->blockio_accounting),
101 prefix, yes_no(c->memory_accounting),
102 prefix, c->cpu_shares,
103 prefix, c->blockio_weight,
104 prefix, c->memory_limit,
105 prefix, cgroup_device_policy_to_string(c->device_policy));
107 LIST_FOREACH(device_allow, a, c->device_allow)
109 "%sDeviceAllow=%s %s%s%s\n",
112 a->r ? "r" : "", a->w ? "w" : "", a->m ? "m" : "");
114 LIST_FOREACH(device_weights, w, c->blockio_device_weights)
116 "%sBlockIODeviceWeight=%s %lu",
121 LIST_FOREACH(device_bandwidths, b, c->blockio_device_bandwidths) {
122 char buf[FORMAT_BYTES_MAX];
127 b->read ? "BlockIOReadBandwidth" : "BlockIOWriteBandwidth",
129 format_bytes(buf, sizeof(buf), b->bandwidth));
133 static int lookup_blkio_device(const char *p, dev_t *dev) {
142 log_warning("Couldn't stat device %s: %m", p);
146 if (S_ISBLK(st.st_mode))
148 else if (major(st.st_dev) != 0) {
149 /* If this is not a device node then find the block
150 * device this file is stored on */
153 /* If this is a partition, try to get the originating
155 block_get_whole_disk(*dev, dev);
157 log_warning("%s is not a block device and file system block device cannot be determined or is not local.", p);
164 static int whitelist_device(const char *path, const char *node, const char *acc) {
165 char buf[2+DECIMAL_STR_MAX(dev_t)*2+2+4];
172 if (stat(node, &st) < 0) {
173 log_warning("Couldn't stat device %s", node);
177 if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
178 log_warning("%s is not a device.", node);
184 S_ISCHR(st.st_mode) ? 'c' : 'b',
185 major(st.st_rdev), minor(st.st_rdev),
188 r = cg_set_attribute("devices", path, "devices.allow", buf);
190 log_warning("Failed to set devices.allow on %s: %s", path, strerror(-r));
195 static int whitelist_major(const char *path, const char *name, char type, const char *acc) {
196 _cleanup_fclose_ FILE *f = NULL;
203 assert(type == 'b' || type == 'c');
205 f = fopen("/proc/devices", "re");
207 log_warning("Cannot open /proc/devices to resolve %s (%c): %m", name, type);
211 FOREACH_LINE(line, f, goto fail) {
212 char buf[2+DECIMAL_STR_MAX(unsigned)+3+4], *p, *w;
217 if (type == 'c' && streq(line, "Character devices:")) {
222 if (type == 'b' && streq(line, "Block devices:")) {
237 w = strpbrk(p, WHITESPACE);
242 r = safe_atou(p, &maj);
249 w += strspn(w, WHITESPACE);
251 if (fnmatch(name, w, 0) != 0)
260 r = cg_set_attribute("devices", path, "devices.allow", buf);
262 log_warning("Failed to set devices.allow on %s: %s", path, strerror(-r));
268 log_warning("Failed to read /proc/devices: %m");
272 void cgroup_context_apply(CGroupContext *c, CGroupControllerMask mask, const char *path) {
282 /* Some cgroup attributes are not support on the root cgroup,
283 * hence silently ignore */
284 is_root = isempty(path) || path_equal(path, "/");
286 if ((mask & CGROUP_CPU) && !is_root) {
287 char buf[DECIMAL_STR_MAX(unsigned long) + 1];
289 sprintf(buf, "%lu\n", c->cpu_shares);
290 r = cg_set_attribute("cpu", path, "cpu.shares", buf);
292 log_warning("Failed to set cpu.shares on %s: %s", path, strerror(-r));
295 if (mask & CGROUP_BLKIO) {
296 char buf[MAX3(DECIMAL_STR_MAX(unsigned long)+1,
297 DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(unsigned long)*1,
298 DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1)];
299 CGroupBlockIODeviceWeight *w;
300 CGroupBlockIODeviceBandwidth *b;
303 sprintf(buf, "%lu\n", c->blockio_weight);
304 r = cg_set_attribute("blkio", path, "blkio.weight", buf);
306 log_warning("Failed to set blkio.weight on %s: %s", path, strerror(-r));
308 /* FIXME: no way to reset this list */
309 LIST_FOREACH(device_weights, w, c->blockio_device_weights) {
312 r = lookup_blkio_device(w->path, &dev);
316 sprintf(buf, "%u:%u %lu", major(dev), minor(dev), w->weight);
317 r = cg_set_attribute("blkio", path, "blkio.weight_device", buf);
319 log_error("Failed to set blkio.weight_device on %s: %s", path, strerror(-r));
323 /* FIXME: no way to reset this list */
324 LIST_FOREACH(device_bandwidths, b, c->blockio_device_bandwidths) {
328 r = lookup_blkio_device(b->path, &dev);
332 a = b->read ? "blkio.throttle.read_bps_device" : "blkio.throttle.write_bps_device";
334 sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), b->bandwidth);
335 r = cg_set_attribute("blkio", path, a, buf);
337 log_error("Failed to set %s on %s: %s", a, path, strerror(-r));
341 if (mask & CGROUP_MEMORY) {
342 if (c->memory_limit != (uint64_t) -1) {
343 char buf[DECIMAL_STR_MAX(uint64_t) + 1];
345 sprintf(buf, "%" PRIu64 "\n", c->memory_limit);
346 r = cg_set_attribute("memory", path, "memory.limit_in_bytes", buf);
348 r = cg_set_attribute("memory", path, "memory.limit_in_bytes", "-1");
351 log_error("Failed to set memory.limit_in_bytes on %s: %s", path, strerror(-r));
354 if ((mask & CGROUP_DEVICE) && !is_root) {
355 CGroupDeviceAllow *a;
357 if (c->device_allow || c->device_policy != CGROUP_AUTO)
358 r = cg_set_attribute("devices", path, "devices.deny", "a");
360 r = cg_set_attribute("devices", path, "devices.allow", "a");
362 log_warning("Failed to reset devices.list on %s: %s", path, strerror(-r));
364 if (c->device_policy == CGROUP_CLOSED ||
365 (c->device_policy == CGROUP_AUTO && c->device_allow)) {
366 static const char auto_devices[] =
370 "/dev/random\0" "rw\0"
371 "/dev/urandom\0" "rw\0";
375 NULSTR_FOREACH_PAIR(x, y, auto_devices)
376 whitelist_device(path, x, y);
379 LIST_FOREACH(device_allow, a, c->device_allow) {
395 if (startswith(a->path, "/dev/"))
396 whitelist_device(path, a->path, acc);
397 else if (startswith(a->path, "block-"))
398 whitelist_major(path, a->path + 6, 'b', acc);
399 else if (startswith(a->path, "char-"))
400 whitelist_major(path, a->path + 5, 'c', acc);
402 log_debug("Ignoring device %s while writing cgroup attribute.", a->path);
407 CGroupControllerMask cgroup_context_get_mask(CGroupContext *c) {
408 CGroupControllerMask mask = 0;
410 /* Figure out which controllers we need */
412 if (c->cpu_accounting || c->cpu_shares != 1024)
413 mask |= CGROUP_CPUACCT | CGROUP_CPU;
415 if (c->blockio_accounting ||
416 c->blockio_weight != 1000 ||
417 c->blockio_device_weights ||
418 c->blockio_device_bandwidths)
419 mask |= CGROUP_BLKIO;
421 if (c->memory_accounting ||
422 c->memory_limit != (uint64_t) -1)
423 mask |= CGROUP_MEMORY;
425 if (c->device_allow || c->device_policy != CGROUP_AUTO)
426 mask |= CGROUP_DEVICE;
431 CGroupControllerMask unit_get_cgroup_mask(Unit *u) {
434 c = unit_get_cgroup_context(u);
438 return cgroup_context_get_mask(c);
441 CGroupControllerMask unit_get_members_mask(Unit *u) {
444 if (u->cgroup_members_mask_valid)
445 return u->cgroup_members_mask;
447 u->cgroup_members_mask = 0;
449 if (u->type == UNIT_SLICE) {
453 SET_FOREACH(member, u->dependencies[UNIT_BEFORE], i) {
458 if (UNIT_DEREF(member->slice) != u)
461 u->cgroup_members_mask |=
462 unit_get_cgroup_mask(member) |
463 unit_get_members_mask(member);
467 u->cgroup_members_mask_valid = true;
468 return u->cgroup_members_mask;
471 CGroupControllerMask unit_get_siblings_mask(Unit *u) {
472 CGroupControllerMask m;
476 if (UNIT_ISSET(u->slice))
477 m = unit_get_members_mask(UNIT_DEREF(u->slice));
479 m = unit_get_cgroup_mask(u) | unit_get_members_mask(u);
481 /* Sibling propagation is only relevant for weight-based
482 * controllers, so let's mask out everything else */
483 return m & (CGROUP_CPU|CGROUP_BLKIO|CGROUP_CPUACCT);
486 CGroupControllerMask unit_get_target_mask(Unit *u) {
487 CGroupControllerMask mask;
489 mask = unit_get_cgroup_mask(u) | unit_get_members_mask(u) | unit_get_siblings_mask(u);
490 mask &= u->manager->cgroup_supported;
495 /* Recurse from a unit up through its containing slices, propagating
496 * mask bits upward. A unit is also member of itself. */
497 void unit_update_cgroup_members_masks(Unit *u) {
498 CGroupControllerMask m;
503 /* Calculate subtree mask */
504 m = unit_get_cgroup_mask(u) | unit_get_members_mask(u);
506 /* See if anything changed from the previous invocation. If
507 * not, we're done. */
508 if (u->cgroup_subtree_mask_valid && m == u->cgroup_subtree_mask)
512 u->cgroup_subtree_mask_valid &&
513 ((m & ~u->cgroup_subtree_mask) != 0) &&
514 ((~m & u->cgroup_subtree_mask) == 0);
516 u->cgroup_subtree_mask = m;
517 u->cgroup_subtree_mask_valid = true;
519 if (UNIT_ISSET(u->slice)) {
520 Unit *s = UNIT_DEREF(u->slice);
523 /* There's more set now than before. We
524 * propagate the new mask to the parent's mask
525 * (not caring if it actually was valid or
528 s->cgroup_members_mask |= m;
531 /* There's less set now than before (or we
532 * don't know), we need to recalculate
533 * everything, so let's invalidate the
534 * parent's members mask */
536 s->cgroup_members_mask_valid = false;
538 /* And now make sure that this change also hits our
540 unit_update_cgroup_members_masks(s);
544 static const char *migrate_callback(CGroupControllerMask mask, void *userdata) {
551 if (u->cgroup_path &&
552 u->cgroup_realized &&
553 (u->cgroup_realized_mask & mask) == mask)
554 return u->cgroup_path;
556 u = UNIT_DEREF(u->slice);
562 static int unit_create_cgroups(Unit *u, CGroupControllerMask mask) {
563 _cleanup_free_ char *path = NULL;
568 path = unit_default_cgroup_path(u);
572 r = hashmap_put(u->manager->cgroup_unit, path, u);
574 log_error(r == -EEXIST ? "cgroup %s exists already: %s" : "hashmap_put failed for %s: %s", path, strerror(-r));
578 u->cgroup_path = path;
582 /* First, create our own group */
583 r = cg_create_everywhere(u->manager->cgroup_supported, mask, u->cgroup_path);
585 log_error("Failed to create cgroup %s: %s", u->cgroup_path, strerror(-r));
589 /* Keep track that this is now realized */
590 u->cgroup_realized = true;
591 u->cgroup_realized_mask = mask;
593 /* Then, possibly move things over */
594 r = cg_migrate_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->cgroup_path, migrate_callback, u);
596 log_warning("Failed to migrate cgroup from to %s: %s", u->cgroup_path, strerror(-r));
601 static bool unit_has_mask_realized(Unit *u, CGroupControllerMask mask) {
604 return u->cgroup_realized && u->cgroup_realized_mask == mask;
607 /* Check if necessary controllers and attributes for a unit are in place.
610 * If not, create paths, move processes over, and set attributes.
612 * Returns 0 on success and < 0 on failure. */
613 static int unit_realize_cgroup_now(Unit *u) {
614 CGroupControllerMask mask;
619 if (u->in_cgroup_queue) {
620 LIST_REMOVE(cgroup_queue, u->manager->cgroup_queue, u);
621 u->in_cgroup_queue = false;
624 mask = unit_get_target_mask(u);
626 if (unit_has_mask_realized(u, mask))
629 /* First, realize parents */
630 if (UNIT_ISSET(u->slice)) {
631 r = unit_realize_cgroup_now(UNIT_DEREF(u->slice));
636 /* And then do the real work */
637 r = unit_create_cgroups(u, mask);
641 /* Finally, apply the necessary attributes. */
642 cgroup_context_apply(unit_get_cgroup_context(u), mask, u->cgroup_path);
647 static void unit_add_to_cgroup_queue(Unit *u) {
649 if (u->in_cgroup_queue)
652 LIST_PREPEND(cgroup_queue, u->manager->cgroup_queue, u);
653 u->in_cgroup_queue = true;
656 unsigned manager_dispatch_cgroup_queue(Manager *m) {
661 while ((i = m->cgroup_queue)) {
662 assert(i->in_cgroup_queue);
664 r = unit_realize_cgroup_now(i);
666 log_warning("Failed to realize cgroups for queued unit %s: %s", i->id, strerror(-r));
674 static void unit_queue_siblings(Unit *u) {
677 /* This adds the siblings of the specified unit and the
678 * siblings of all parent units to the cgroup queue. (But
679 * neither the specified unit itself nor the parents.) */
681 while ((slice = UNIT_DEREF(u->slice))) {
685 SET_FOREACH(m, slice->dependencies[UNIT_BEFORE], i) {
689 /* Skip units that have a dependency on the slice
690 * but aren't actually in it. */
691 if (UNIT_DEREF(m->slice) != slice)
694 /* No point in doing cgroup application for units
695 * without active processes. */
696 if (UNIT_IS_INACTIVE_OR_FAILED(unit_active_state(m)))
699 /* If the unit doesn't need any new controllers
700 * and has current ones realized, it doesn't need
702 if (unit_has_mask_realized(m, unit_get_target_mask(m)))
705 unit_add_to_cgroup_queue(m);
712 int unit_realize_cgroup(Unit *u) {
717 c = unit_get_cgroup_context(u);
721 /* So, here's the deal: when realizing the cgroups for this
722 * unit, we need to first create all parents, but there's more
723 * actually: for the weight-based controllers we also need to
724 * make sure that all our siblings (i.e. units that are in the
725 * same slice as we are) have cgroups, too. Otherwise, things
726 * would become very uneven as each of their processes would
727 * get as much resources as all our group together. This call
728 * will synchronously create the parent cgroups, but will
729 * defer work on the siblings to the next event loop
732 /* Add all sibling slices to the cgroup queue. */
733 unit_queue_siblings(u);
735 /* And realize this one now (and apply the values) */
736 return unit_realize_cgroup_now(u);
739 void unit_destroy_cgroup(Unit *u) {
747 r = cg_trim_everywhere(u->manager->cgroup_supported, u->cgroup_path, !unit_has_name(u, SPECIAL_ROOT_SLICE));
749 log_debug("Failed to destroy cgroup %s: %s", u->cgroup_path, strerror(-r));
751 hashmap_remove(u->manager->cgroup_unit, u->cgroup_path);
753 free(u->cgroup_path);
754 u->cgroup_path = NULL;
755 u->cgroup_realized = false;
756 u->cgroup_realized_mask = 0;
760 pid_t unit_search_main_pid(Unit *u) {
761 _cleanup_fclose_ FILE *f = NULL;
762 pid_t pid = 0, npid, mypid;
769 if (cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, &f) < 0)
773 while (cg_read_pid(f, &npid) > 0) {
779 /* Ignore processes that aren't our kids */
780 if (get_parent_of_pid(npid, &ppid) >= 0 && ppid != mypid)
784 /* Dang, there's more than one daemonized PID
785 in this group, so we don't know what process
786 is the main process. */
797 int manager_setup_cgroup(Manager *m) {
798 _cleanup_free_ char *path = NULL;
804 /* 0. Be nice to Ingo Molnar #628004 */
805 if (path_is_mount_point("/sys/fs/cgroup/systemd", false) <= 0) {
806 log_warning("No control group support available, not creating root group.");
810 /* 1. Determine hierarchy */
811 free(m->cgroup_root);
812 m->cgroup_root = NULL;
814 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 0, &m->cgroup_root);
816 log_error("Cannot determine cgroup we are running in: %s", strerror(-r));
820 /* LEGACY: Already in /system.slice? If so, let's cut this
821 * off. This is to support live upgrades from older systemd
822 * versions where PID 1 was moved there. */
823 if (m->running_as == SYSTEMD_SYSTEM) {
824 e = endswith(m->cgroup_root, "/" SPECIAL_SYSTEM_SLICE);
826 e = endswith(m->cgroup_root, "/system");
831 /* And make sure to store away the root value without trailing
832 * slash, even for the root dir, so that we can easily prepend
834 if (streq(m->cgroup_root, "/"))
835 m->cgroup_root[0] = 0;
838 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, NULL, &path);
840 log_error("Cannot find cgroup mount point: %s", strerror(-r));
844 log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER ". File system hierarchy is at %s.", path);
846 /* 3. Install agent */
847 if (m->running_as == SYSTEMD_SYSTEM) {
848 r = cg_install_release_agent(SYSTEMD_CGROUP_CONTROLLER, SYSTEMD_CGROUP_AGENT_PATH);
850 log_warning("Failed to install release agent, ignoring: %s", strerror(-r));
852 log_debug("Installed release agent.");
854 log_debug("Release agent already installed.");
857 /* 4. Make sure we are in the root cgroup */
858 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, 0);
860 log_error("Failed to create root cgroup hierarchy: %s", strerror(-r));
864 /* 5. And pin it, so that it cannot be unmounted */
865 if (m->pin_cgroupfs_fd >= 0)
866 close_nointr_nofail(m->pin_cgroupfs_fd);
868 m->pin_cgroupfs_fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_NOCTTY|O_NONBLOCK);
870 log_error("Failed to open pin file: %m");
874 /* 6. Figure out which controllers are supported */
875 m->cgroup_supported = cg_mask_supported();
877 /* 7. Always enable hierarchial support if it exists... */
878 cg_set_attribute("memory", "/", "memory.use_hierarchy", "1");
883 void manager_shutdown_cgroup(Manager *m, bool delete) {
886 /* We can't really delete the group, since we are in it. But
888 if (delete && m->cgroup_root)
889 cg_trim(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, false);
891 if (m->pin_cgroupfs_fd >= 0) {
892 close_nointr_nofail(m->pin_cgroupfs_fd);
893 m->pin_cgroupfs_fd = -1;
896 free(m->cgroup_root);
897 m->cgroup_root = NULL;
900 Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
907 u = hashmap_get(m->cgroup_unit, cgroup);
921 u = hashmap_get(m->cgroup_unit, p);
927 Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
928 _cleanup_free_ char *cgroup = NULL;
936 r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
940 return manager_get_unit_by_cgroup(m, cgroup);
943 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
950 u = manager_get_unit_by_cgroup(m, cgroup);
952 r = cg_is_empty_recursive(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, true);
954 if (UNIT_VTABLE(u)->notify_cgroup_empty)
955 UNIT_VTABLE(u)->notify_cgroup_empty(u);
957 unit_add_to_gc_queue(u);
964 static const char* const cgroup_device_policy_table[_CGROUP_DEVICE_POLICY_MAX] = {
965 [CGROUP_AUTO] = "auto",
966 [CGROUP_CLOSED] = "closed",
967 [CGROUP_STRICT] = "strict",
970 DEFINE_STRING_TABLE_LOOKUP(cgroup_device_policy, CGroupDevicePolicy);