cgroup.c

   1 /*-*- Mode: C; c-basic-offset: 8 -*-*/
   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
  29 #include "cgroup.h"
  30 #include "log.h"
  31
  32 static int translate_error(int error, int _errno) {
  33
  34         switch (error) {
  35
  36         case ECGROUPNOTCOMPILED:
  37         case ECGROUPNOTMOUNTED:
  38         case ECGROUPNOTEXIST:
  39         case ECGROUPNOTCREATED:
  40                 return -ENOENT;
  41
  42         case ECGINVAL:
  43                 return -EINVAL;
  44
  45         case ECGROUPNOTALLOWED:
  46                 return -EPERM;
  47
  48         case ECGOTHER:
  49                 return -_errno;
  50         }
  51
  52         return -EIO;
  53 }
  54
  55 int cgroup_bonding_realize(CGroupBonding *b) {
  56         int r;
  57
  58         assert(b);
  59         assert(b->path);
  60         assert(b->controller);
  61
  62         if (b->cgroup)
  63                 return 0;
  64
  65         if (!(b->cgroup = cgroup_new_cgroup(b->path)))
  66                 return -ENOMEM;
  67
  68         if (!cgroup_add_controller(b->cgroup, b->controller)) {
  69                 r = -ENOMEM;
  70                 goto fail;
  71         }
  72
  73         if (b->inherit)
  74                 r = cgroup_create_cgroup_from_parent(b->cgroup, true);
  75         else
  76                 r = cgroup_create_cgroup(b->cgroup, true);
  77
  78         if (r != 0) {
  79                 r = translate_error(r, errno);
  80                 goto fail;
  81         }
  82
  83         return 0;
  84
  85 fail:
  86         cgroup_free(&b->cgroup);
  87         b->cgroup = NULL;
  88         return r;
  89 }
  90
  91 int cgroup_bonding_realize_list(CGroupBonding *first) {
  92         CGroupBonding *b;
  93
  94         LIST_FOREACH(by_unit, b, first) {
  95                 int r;
  96
  97                 if ((r = cgroup_bonding_realize(b)) < 0)
  98                         return r;
  99         }
 100
 101         return 0;
 102 }
 103
 104 void cgroup_bonding_free(CGroupBonding *b) {
 105         assert(b);
 106
 107         if (b->unit) {
 108                 CGroupBonding *f;
 109
 110                 LIST_REMOVE(CGroupBonding, by_unit, b->unit->meta.cgroup_bondings, b);
 111
 112                 assert_se(f = hashmap_get(b->unit->meta.manager->cgroup_bondings, b->path));
 113                 LIST_REMOVE(CGroupBonding, by_path, f, b);
 114
 115                 if (f)
 116                         hashmap_replace(b->unit->meta.manager->cgroup_bondings, b->path, f);
 117                 else
 118                         hashmap_remove(b->unit->meta.manager->cgroup_bondings, b->path);
 119         }
 120
 121         if (b->cgroup) {
 122                 if (b->only_us && b->clean_up && cgroup_bonding_is_empty(b) > 0)
 123                         cgroup_delete_cgroup_ext(b->cgroup, true);
 124
 125                 cgroup_free(&b->cgroup);
 126         }
 127
 128         free(b->controller);
 129         free(b->path);
 130         free(b);
 131 }
 132
 133 void cgroup_bonding_free_list(CGroupBonding *first) {
 134         CGroupBonding *b, *n;
 135
 136         LIST_FOREACH_SAFE(by_unit, b, n, first)
 137                 cgroup_bonding_free(b);
 138 }
 139
 140 int cgroup_bonding_install(CGroupBonding *b, pid_t pid) {
 141         int r;
 142
 143         assert(b);
 144         assert(pid >= 0);
 145
 146         if (pid == 0)
 147                 pid = getpid();
 148
 149         if (!b->cgroup)
 150                 return -ENOENT;
 151
 152         if ((r = cgroup_attach_task_pid(b->cgroup, pid)))
 153                 return translate_error(r, errno);
 154
 155         return 0;
 156 }
 157
 158 int cgroup_bonding_install_list(CGroupBonding *first, pid_t pid) {
 159         CGroupBonding *b;
 160
 161         LIST_FOREACH(by_unit, b, first) {
 162                 int r;
 163
 164                 if ((r = cgroup_bonding_install(b, pid)) < 0)
 165                         return r;
 166         }
 167
 168         return 0;
 169 }
 170
 171 int cgroup_bonding_kill(CGroupBonding *b, int sig) {
 172         int r;
 173         Set *s;
 174         bool done;
 175         bool killed = false;
 176
 177         assert(b);
 178         assert(sig > 0);
 179
 180         if (!b->only_us)
 181                 return -EAGAIN;
 182
 183         if (!(s = set_new(trivial_hash_func, trivial_compare_func)))
 184                 return -ENOMEM;
 185
 186         do {
 187                 void *iterator;
 188                 pid_t pid;
 189
 190                 done = true;
 191
 192                 if ((r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid)) != 0) {
 193                         if (r == ECGEOF) {
 194                                 r = 0;
 195                                 goto kill_done;
 196                         } else {
 197                                 r = translate_error(r, errno);
 198                                 break;
 199                         }
 200                 }
 201
 202                 for (;;) {
 203                         if (set_get(s, INT_TO_PTR(pid)) != INT_TO_PTR(pid)) {
 204
 205                                 /* If we haven't killed this process
 206                                  * yet, kill it */
 207
 208                                 if (kill(pid, sig) < 0 && errno != ESRCH) {
 209                                         r = -errno;
 210                                         break;
 211                                 }
 212
 213                                 killed = true;
 214                                 done = false;
 215
 216                                 if ((r = set_put(s, INT_TO_PTR(pid))) < 0)
 217                                     break;
 218                         }
 219
 220                         if ((r = cgroup_get_task_next(&iterator, &pid)) != 0) {
 221
 222                                 if (r == ECGEOF)
 223                                         r = 0;
 224                                 else
 225                                         r = translate_error(r, errno);
 226
 227                                 break;
 228                         }
 229                 }
 230
 231         kill_done:
 232                 assert_se(cgroup_get_task_end(&iterator) == 0);
 233
 234                 /* To avoid racing against processes which fork
 235                  * quicker than we can kill them we repeat this until
 236                  * no new pids need to be killed. */
 237
 238         } while (!done && r >= 0);
 239
 240         set_free(s);
 241
 242         if (r < 0)
 243                 return r;
 244
 245         return killed ? 0 : -ESRCH;
 246 }
 247
 248 int cgroup_bonding_kill_list(CGroupBonding *first, int sig) {
 249         CGroupBonding *b;
 250         int r = -EAGAIN;
 251
 252         LIST_FOREACH(by_unit, b, first) {
 253                 if ((r = cgroup_bonding_kill(b, sig)) < 0) {
 254                         if (r == -EAGAIN || -ESRCH)
 255                                 continue;
 256
 257                         return r;
 258                 }
 259
 260                 return 0;
 261         }
 262
 263         return r;
 264 }
 265
 266 /* Returns 1 if the group is empty, 0 if it is not, -EAGAIN if we
 267  * cannot know */
 268 int cgroup_bonding_is_empty(CGroupBonding *b) {
 269         void *iterator;
 270         pid_t pid;
 271         int r;
 272
 273         assert(b);
 274
 275         r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid);
 276
 277         if (r == 0 || r == ECGEOF)
 278                 cgroup_get_task_end(&iterator);
 279
 280         /* Hmm, no PID in this group? Then it is definitely empty */
 281         if (r == ECGEOF)
 282                 return 1;
 283
 284         /* Some error? Let's return it */
 285         if (r != 0)
 286                 return translate_error(r, errno);
 287
 288         /* It's not empty, and we are the only user, then it is
 289          * definitely not empty */
 290         if (b->only_us)
 291                 return 0;
 292
 293         /* There are PIDs in the group but we aren't the only users,
 294          * hence we cannot say */
 295         return -EAGAIN;
 296 }
 297
 298 int cgroup_bonding_is_empty_list(CGroupBonding *first) {
 299         CGroupBonding *b;
 300
 301         LIST_FOREACH(by_unit, b, first) {
 302                 int r;
 303
 304                 if ((r = cgroup_bonding_is_empty(b)) < 0) {
 305                         /* If this returned -EAGAIN, then we don't know if the
 306                          * group is empty, so let's see if another group can
 307                          * tell us */
 308
 309                         if (r != -EAGAIN)
 310                                 return r;
 311                 } else
 312                         return r;
 313         }
 314
 315         return -EAGAIN;
 316 }
 317
 318 static int install_release_agent(Manager *m, const char *mount_point) {
 319         char *p, *c, *sc;
 320         int r;
 321
 322         assert(m);
 323         assert(mount_point);
 324
 325         if (asprintf(&p, "%s/release_agent", mount_point) < 0)
 326                 return -ENOMEM;
 327
 328         if ((r = read_one_line_file(p, &c)) < 0) {
 329                 free(p);
 330                 return r;
 331         }
 332
 333         sc = strstrip(c);
 334
 335         if (sc[0] == 0) {
 336                 if ((r = write_one_line_file(p, CGROUP_AGENT_PATH "\n" )) < 0) {
 337                         free(p);
 338                         free(c);
 339                         return r;
 340                 }
 341         } else if (!streq(sc, CGROUP_AGENT_PATH)) {
 342                 free(p);
 343                 free(c);
 344                 return -EEXIST;
 345         }
 346
 347         free(c);
 348         free(p);
 349
 350         if (asprintf(&p, "%s/notify_on_release", mount_point) < 0)
 351                 return -ENOMEM;
 352
 353         if ((r = read_one_line_file(p, &c)) < 0) {
 354                 free(p);
 355                 return r;
 356         }
 357
 358         sc = strstrip(c);
 359
 360         if (streq(sc, "0")) {
 361                 if ((r = write_one_line_file(p, "1\n")) < 0) {
 362                         free(p);
 363                         free(c);
 364                         return r;
 365                 }
 366         } else if (!streq(sc, "1")) {
 367                 free(p);
 368                 free(c);
 369                 return -EIO;
 370         }
 371
 372         free(p);
 373         free(c);
 374
 375         return 0;
 376 }
 377
 378 static int create_hierarchy_cgroup(Manager *m) {
 379         struct cgroup *cg;
 380         int r;
 381
 382         assert(m);
 383
 384         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
 385                 return -ENOMEM;
 386
 387         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
 388                 r = -ENOMEM;
 389                 goto finish;
 390         }
 391
 392         if ((r = cgroup_create_cgroup(cg, true)) != 0) {
 393                 log_error("Failed to create cgroup hierarchy group: %s", cgroup_strerror(r));
 394                 r = translate_error(r, errno);
 395                 goto finish;
 396         }
 397
 398         if ((r = cgroup_attach_task(cg)) != 0) {
 399                 log_error("Failed to add ourselves to hierarchy group: %s", cgroup_strerror(r));
 400                 r = translate_error(r, errno);
 401                 goto finish;
 402         }
 403
 404         r = 0;
 405
 406 finish:
 407         cgroup_free(&cg);
 408         return r;
 409 }
 410
 411 int manager_setup_cgroup(Manager *m) {
 412         char *mp, *cp;
 413         int r;
 414         pid_t pid;
 415         char suffix[32];
 416
 417         assert(m);
 418
 419         if ((r = cgroup_init()) != 0) {
 420                 log_error("Failed to initialize libcg: %s", cgroup_strerror(r));
 421                 return translate_error(r, errno);
 422         }
 423
 424         free(m->cgroup_controller);
 425         if (!(m->cgroup_controller = strdup("debug")))
 426                 return -ENOMEM;
 427
 428         if ((r = cgroup_get_subsys_mount_point(m->cgroup_controller, &mp)))
 429                 return translate_error(r, errno);
 430
 431         pid = getpid();
 432
 433         if ((r = cgroup_get_current_controller_path(pid, m->cgroup_controller, &cp))) {
 434                 free(mp);
 435                 return translate_error(r, errno);
 436         }
 437
 438         snprintf(suffix, sizeof(suffix), "/systemd-%u", (unsigned) pid);
 439         char_array_0(suffix);
 440
 441         free(m->cgroup_hierarchy);
 442
 443         if (endswith(cp, suffix))
 444                 /* We probably got reexecuted and can continue to use our root cgroup */
 445                 m->cgroup_hierarchy = cp;
 446         else {
 447                 /* We need a new root cgroup */
 448
 449                 m->cgroup_hierarchy = NULL;
 450                 r = asprintf(&m->cgroup_hierarchy, "%s%s", streq(cp, "/") ? "" : cp, suffix);
 451                 free(cp);
 452
 453                 if (r < 0) {
 454                         free(mp);
 455                         return -ENOMEM;
 456                 }
 457         }
 458
 459         log_info("Using cgroup controller <%s>, hierarchy mounted at <%s>, using root group <%s>.",
 460                  m->cgroup_controller,
 461                  mp,
 462                  m->cgroup_hierarchy);
 463
 464         if ((r = install_release_agent(m, mp)) < 0)
 465                 log_warning("Failed to install release agent, ignoring: %s", strerror(-r));
 466         else
 467                 log_info("Installed release agent, or already installed.");
 468
 469         free(mp);
 470
 471         if ((r = create_hierarchy_cgroup(m)) < 0)
 472                 log_error("Failed to create root cgroup hierarchy: %s", strerror(-r));
 473         else
 474                 log_info("Created root group.");
 475
 476         return r;
 477 }
 478
 479 int manager_shutdown_cgroup(Manager *m, bool delete) {
 480         struct cgroup *cg;
 481         int r;
 482
 483         assert(m);
 484
 485         if (!m->cgroup_hierarchy)
 486                 return 0;
 487
 488         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
 489                 return -ENOMEM;
 490
 491         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
 492                 r = -ENOMEM;
 493                 goto finish;
 494         }
 495
 496         /* Often enough we won't be able to delete the cgroup we
 497          * ourselves are in, hence ignore all errors here */
 498         if (delete)
 499                 cgroup_delete_cgroup_ext(cg, CGFLAG_DELETE_IGNORE_MIGRATION|CGFLAG_DELETE_RECURSIVE);
 500         r = 0;
 501
 502 finish:
 503         cgroup_free(&cg);
 504         return r;
 505
 506 }
 507
 508 int cgroup_notify_empty(Manager *m, const char *group) {
 509         CGroupBonding *l, *b;
 510
 511         assert(m);
 512         assert(group);
 513
 514         if (!(l = hashmap_get(m->cgroup_bondings, group)))
 515                 return 0;
 516
 517         LIST_FOREACH(by_path, b, l) {
 518                 int t;
 519
 520                 if (!b->unit)
 521                         continue;
 522
 523                 if ((t = cgroup_bonding_is_empty_list(b)) < 0) {
 524
 525                         /* If we don't know, we don't know */
 526                         if (t != -EAGAIN)
 527                                 log_warning("Failed to check whether cgroup is empty: %s", strerror(errno));
 528
 529                         continue;
 530                 }
 531
 532                 if (t > 0)
 533                         if (UNIT_VTABLE(b->unit)->cgroup_notify_empty)
 534                                 UNIT_VTABLE(b->unit)->cgroup_notify_empty(b->unit);
 535         }
 536
 537         return 0;
 538 }
 539
 540 CGroupBonding *cgroup_bonding_find_list(CGroupBonding *first, const char *controller) {
 541         CGroupBonding *b;
 542
 543         assert(controller);
 544
 545         LIST_FOREACH(by_unit, b, first)
 546                 if (streq(b->controller, controller))
 547                         return b;
 548
 549         return NULL;
 550 }
 551
 552 char *cgroup_bonding_to_string(CGroupBonding *b) {
 553         char *r;
 554
 555         assert(b);
 556
 557         if (asprintf(&r, "%s:%s", b->controller, b->path) < 0)
 558                 return NULL;
 559
 560         return r;
 561 }