src/cgroup.c

   1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
   2
   3 /***
   4   This file is part of systemd.
   5
   6   Copyright 2010 Lennart Poettering
   7
   8   systemd is free software; you can redistribute it and/or modify it
   9   under the terms of the GNU General Public License as published by
  10   the Free Software Foundation; either version 2 of the License, or
  11   (at your option) any later version.
  12
  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   General Public License for more details.
  17
  18   You should have received a copy of the GNU General Public License
  19   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  20 ***/
  21
  22 #include <errno.h>
  23 #include <assert.h>
  24 #include <unistd.h>
  25 #include <sys/types.h>
  26 #include <signal.h>
  27 #include <sys/mount.h>
  28 #include <fcntl.h>
  29
  30 #include "cgroup.h"
  31 #include "cgroup-util.h"
  32 #include "log.h"
  33
  34 int cgroup_bonding_realize(CGroupBonding *b) {
  35         int r;
  36
  37         assert(b);
  38         assert(b->path);
  39         assert(b->controller);
  40
  41         if (b->realized)
  42                 return 0;
  43
  44         if ((r = cg_create(b->controller, b->path)) < 0)
  45                 return r;
  46
  47         b->realized = true;
  48
  49         if (b->ours)
  50                 cg_trim(b->controller, b->path, false);
  51
  52         return 0;
  53 }
  54
  55 int cgroup_bonding_realize_list(CGroupBonding *first) {
  56         CGroupBonding *b;
  57         int r;
  58
  59         LIST_FOREACH(by_unit, b, first)
  60                 if ((r = cgroup_bonding_realize(b)) < 0 && b->essential)
  61                         return r;
  62
  63         return 0;
  64 }
  65
  66 void cgroup_bonding_free(CGroupBonding *b) {
  67         assert(b);
  68
  69         if (b->unit) {
  70                 CGroupBonding *f;
  71
  72                 LIST_REMOVE(CGroupBonding, by_unit, b->unit->meta.cgroup_bondings, b);
  73
  74                 if (streq(b->controller, SYSTEMD_CGROUP_CONTROLLER)) {
  75                         assert_se(f = hashmap_get(b->unit->meta.manager->cgroup_bondings, b->path));
  76                         LIST_REMOVE(CGroupBonding, by_path, f, b);
  77
  78                         if (f)
  79                                 hashmap_replace(b->unit->meta.manager->cgroup_bondings, b->path, f);
  80                         else
  81                                 hashmap_remove(b->unit->meta.manager->cgroup_bondings, b->path);
  82                 }
  83         }
  84
  85         if (b->realized && b->ours) {
  86
  87                 if (cgroup_bonding_is_empty(b) > 0)
  88                         cg_delete(b->controller, b->path);
  89                 else
  90                         cg_trim(b->controller, b->path, false);
  91         }
  92
  93         free(b->controller);
  94         free(b->path);
  95         free(b);
  96 }
  97
  98 void cgroup_bonding_free_list(CGroupBonding *first) {
  99         CGroupBonding *b, *n;
 100
 101         LIST_FOREACH_SAFE(by_unit, b, n, first)
 102                 cgroup_bonding_free(b);
 103 }
 104
 105 void cgroup_bonding_trim(CGroupBonding *b, bool delete_root) {
 106         assert(b);
 107
 108         if (b->realized && b->ours)
 109                 cg_trim(b->controller, b->path, delete_root);
 110 }
 111
 112 void cgroup_bonding_trim_list(CGroupBonding *first, bool delete_root) {
 113         CGroupBonding *b;
 114
 115         LIST_FOREACH(by_unit, b, first)
 116                 cgroup_bonding_trim(b, delete_root);
 117 }
 118
 119 int cgroup_bonding_install(CGroupBonding *b, pid_t pid) {
 120         int r;
 121
 122         assert(b);
 123         assert(pid >= 0);
 124
 125         if ((r = cg_create_and_attach(b->controller, b->path, pid)) < 0)
 126                 return r;
 127
 128         b->realized = true;
 129         return 0;
 130 }
 131
 132 int cgroup_bonding_install_list(CGroupBonding *first, pid_t pid) {
 133         CGroupBonding *b;
 134         int r;
 135
 136         LIST_FOREACH(by_unit, b, first)
 137                 if ((r = cgroup_bonding_install(b, pid)) < 0 && b->essential)
 138                         return r;
 139
 140         return 0;
 141 }
 142
 143 int cgroup_bonding_kill(CGroupBonding *b, int sig, Set *s) {
 144         assert(b);
 145         assert(sig >= 0);
 146
 147         /* Don't kill cgroups that aren't ours */
 148         if (!b->realized || !b->ours)
 149                 return 0;
 150
 151         return cg_kill_recursive(b->controller, b->path, sig, true, false, s);
 152 }
 153
 154 int cgroup_bonding_kill_list(CGroupBonding *first, int sig, Set *s) {
 155         CGroupBonding *b;
 156         Set *allocated_set = NULL;
 157         int ret = -EAGAIN, r;
 158
 159         if (!s)
 160                 if (!(s = allocated_set = set_new(trivial_hash_func, trivial_compare_func)))
 161                         return -ENOMEM;
 162
 163         LIST_FOREACH(by_unit, b, first) {
 164                 if ((r = cgroup_bonding_kill(b, sig, s)) < 0) {
 165                         if (r == -EAGAIN || r == -ESRCH)
 166                                 continue;
 167
 168                         ret = r;
 169                         goto finish;
 170                 }
 171
 172                 if (ret < 0 || r > 0)
 173                         ret = r;
 174         }
 175
 176 finish:
 177         if (allocated_set)
 178                 set_free(allocated_set);
 179
 180         return ret;
 181 }
 182
 183 /* Returns 1 if the group is empty, 0 if it is not, -EAGAIN if we
 184  * cannot know */
 185 int cgroup_bonding_is_empty(CGroupBonding *b) {
 186         int r;
 187
 188         assert(b);
 189
 190         if ((r = cg_is_empty_recursive(b->controller, b->path, true)) < 0)
 191                 return r;
 192
 193         /* If it is empty it is empty */
 194         if (r > 0)
 195                 return 1;
 196
 197         /* It's not only us using this cgroup, so we just don't know */
 198         return b->ours ? 0 : -EAGAIN;
 199 }
 200
 201 int cgroup_bonding_is_empty_list(CGroupBonding *first) {
 202         CGroupBonding *b;
 203
 204         LIST_FOREACH(by_unit, b, first) {
 205                 int r;
 206
 207                 if ((r = cgroup_bonding_is_empty(b)) < 0) {
 208                         /* If this returned -EAGAIN, then we don't know if the
 209                          * group is empty, so let's see if another group can
 210                          * tell us */
 211
 212                         if (r != -EAGAIN)
 213                                 return r;
 214                 } else
 215                         return r;
 216         }
 217
 218         return -EAGAIN;
 219 }
 220
 221 int manager_setup_cgroup(Manager *m) {
 222         char *current = NULL, *path = NULL;
 223         int r;
 224         char suffix[32];
 225
 226         assert(m);
 227
 228         /* 1. Determine hierarchy */
 229         if ((r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &current)) < 0)
 230                 goto finish;
 231
 232         if (m->running_as == MANAGER_SYSTEM)
 233                 strcpy(suffix, "/system");
 234         else {
 235                 snprintf(suffix, sizeof(suffix), "/systemd-%lu", (unsigned long) getpid());
 236                 char_array_0(suffix);
 237         }
 238
 239         free(m->cgroup_hierarchy);
 240         if (endswith(current, suffix)) {
 241                 /* We probably got reexecuted and can continue to use our root cgroup */
 242                 m->cgroup_hierarchy = current;
 243                 current = NULL;
 244
 245         } else {
 246                 /* We need a new root cgroup */
 247                 m->cgroup_hierarchy = NULL;
 248                 if (asprintf(&m->cgroup_hierarchy, "%s%s", streq(current, "/") ? "" : current, suffix) < 0) {
 249                         r = -ENOMEM;
 250                         goto finish;
 251                 }
 252         }
 253
 254         /* 2. Show data */
 255         if ((r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_hierarchy, NULL, &path)) < 0)
 256                 goto finish;
 257
 258         log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER ". File system hierarchy is at %s.", path);
 259
 260         /* 3. Install agent */
 261         if ((r = cg_install_release_agent(SYSTEMD_CGROUP_CONTROLLER, SYSTEMD_CGROUP_AGENT_PATH)) < 0)
 262                 log_warning("Failed to install release agent, ignoring: %s", strerror(-r));
 263         else if (r > 0)
 264                 log_debug("Installed release agent.");
 265         else
 266                 log_debug("Release agent already installed.");
 267
 268         /* 4. Realize the group */
 269         if ((r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_hierarchy, 0)) < 0) {
 270                 log_error("Failed to create root cgroup hierarchy: %s", strerror(-r));
 271                 goto finish;
 272         }
 273
 274         /* 5. And pin it, so that it cannot be unmounted */
 275         if (m->pin_cgroupfs_fd >= 0)
 276                 close_nointr_nofail(m->pin_cgroupfs_fd);
 277
 278         if ((m->pin_cgroupfs_fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_NOCTTY|O_NONBLOCK)) < 0) {
 279                 r = -errno;
 280                 goto finish;
 281         }
 282
 283         log_debug("Created root group.");
 284
 285 finish:
 286         free(current);
 287         free(path);
 288
 289         return r;
 290 }
 291
 292 void manager_shutdown_cgroup(Manager *m, bool delete) {
 293         assert(m);
 294
 295         if (delete && m->cgroup_hierarchy)
 296                 cg_delete(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_hierarchy);
 297
 298         if (m->pin_cgroupfs_fd >= 0) {
 299                 close_nointr_nofail(m->pin_cgroupfs_fd);
 300                 m->pin_cgroupfs_fd = -1;
 301         }
 302
 303         free(m->cgroup_hierarchy);
 304         m->cgroup_hierarchy = NULL;
 305 }
 306
 307 int cgroup_notify_empty(Manager *m, const char *group) {
 308         CGroupBonding *l, *b;
 309
 310         assert(m);
 311         assert(group);
 312
 313         if (!(l = hashmap_get(m->cgroup_bondings, group)))
 314                 return 0;
 315
 316         LIST_FOREACH(by_path, b, l) {
 317                 int t;
 318
 319                 if (!b->unit)
 320                         continue;
 321
 322                 if ((t = cgroup_bonding_is_empty_list(b)) < 0) {
 323
 324                         /* If we don't know, we don't know */
 325                         if (t != -EAGAIN)
 326                                 log_warning("Failed to check whether cgroup is empty: %s", strerror(errno));
 327
 328                         continue;
 329                 }
 330
 331                 if (t > 0)
 332                         if (UNIT_VTABLE(b->unit)->cgroup_notify_empty)
 333                                 UNIT_VTABLE(b->unit)->cgroup_notify_empty(b->unit);
 334         }
 335
 336         return 0;
 337 }
 338
 339 Unit* cgroup_unit_by_pid(Manager *m, pid_t pid) {
 340         CGroupBonding *l, *b;
 341         char *group = NULL;
 342
 343         assert(m);
 344
 345         if (pid <= 1)
 346                 return NULL;
 347
 348         if (cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, pid, &group) < 0)
 349                 return NULL;
 350
 351         l = hashmap_get(m->cgroup_bondings, group);
 352
 353         if (!l) {
 354                 char *slash;
 355
 356                 while ((slash = strrchr(group, '/'))) {
 357                         if (slash == group)
 358                                 break;
 359
 360                         *slash = 0;
 361
 362                         if ((l = hashmap_get(m->cgroup_bondings, group)))
 363                                 break;
 364                 }
 365         }
 366
 367         free(group);
 368
 369         LIST_FOREACH(by_path, b, l) {
 370
 371                 if (!b->unit)
 372                         continue;
 373
 374                 if (b->ours)
 375                         return b->unit;
 376         }
 377
 378         return NULL;
 379 }
 380
 381 CGroupBonding *cgroup_bonding_find_list(CGroupBonding *first, const char *controller) {
 382         CGroupBonding *b;
 383
 384         assert(controller);
 385
 386         LIST_FOREACH(by_unit, b, first)
 387                 if (streq(b->controller, controller))
 388                         return b;
 389
 390         return NULL;
 391 }
 392
 393 char *cgroup_bonding_to_string(CGroupBonding *b) {
 394         char *r;
 395
 396         assert(b);
 397
 398         if (asprintf(&r, "%s:%s", b->controller, b->path) < 0)
 399                 return NULL;
 400
 401         return r;
 402 }
 403
 404 pid_t cgroup_bonding_search_main_pid(CGroupBonding *b) {
 405         FILE *f;
 406         pid_t pid = 0, npid;
 407
 408         assert(b);
 409
 410         if (!b->ours)
 411                 return 0;
 412
 413         if (cg_enumerate_processes(b->controller, b->path, &f) < 0)
 414                 return 0;
 415
 416         while (cg_read_pid(f, &npid) > 0)  {
 417
 418                 if (npid == pid)
 419                         continue;
 420
 421                 if (pid != 0) {
 422                         /* Dang, there's more than one PID in this
 423                          * group, so we don't know what process is the
 424                          * main process. */
 425                         pid = 0;
 426                         break;
 427                 }
 428
 429                 pid = npid;
 430         }
 431
 432         fclose(f);
 433
 434         return pid;
 435 }
 436
 437 pid_t cgroup_bonding_search_main_pid_list(CGroupBonding *first) {
 438         CGroupBonding *b;
 439         pid_t pid;
 440
 441         /* Try to find a main pid from this cgroup, but checking if
 442          * there's only one PID in the cgroup and returning it. Later
 443          * on we might want to add additional, smarter heuristics
 444          * here. */
 445
 446         LIST_FOREACH(by_unit, b, first)
 447                 if ((pid = cgroup_bonding_search_main_pid(b)) != 0)
 448                         return pid;
 449
 450         return 0;
 451
 452 }