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         log_debug("Killing processes from process group %s:%s", b->controller, b->path);
 187
 188         do {
 189                 void *iterator;
 190                 pid_t pid;
 191
 192                 done = true;
 193
 194                 if ((r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid)) != 0) {
 195                         if (r == ECGEOF) {
 196                                 r = 0;
 197                                 goto kill_done;
 198                         } else {
 199                                 r = translate_error(r, errno);
 200                                 break;
 201                         }
 202                 }
 203
 204                 for (;;) {
 205                         if (set_get(s, INT_TO_PTR(pid)) != INT_TO_PTR(pid)) {
 206
 207                                 /* If we haven't killed this process
 208                                  * yet, kill it */
 209
 210                                 if (kill(pid, sig) < 0 && errno != ESRCH) {
 211                                         r = -errno;
 212                                         break;
 213                                 }
 214
 215                                 killed = true;
 216                                 done = false;
 217
 218                                 if ((r = set_put(s, INT_TO_PTR(pid))) < 0)
 219                                     break;
 220                         }
 221
 222                         if ((r = cgroup_get_task_next(&iterator, &pid)) != 0) {
 223
 224                                 if (r == ECGEOF)
 225                                         r = 0;
 226                                 else
 227                                         r = translate_error(r, errno);
 228
 229                                 break;
 230                         }
 231                 }
 232
 233         kill_done:
 234                 assert_se(cgroup_get_task_end(&iterator) == 0);
 235
 236                 /* To avoid racing against processes which fork
 237                  * quicker than we can kill them we repeat this until
 238                  * no new pids need to be killed. */
 239
 240         } while (!done && r >= 0);
 241
 242         set_free(s);
 243
 244         if (r < 0)
 245                 return r;
 246
 247         return killed ? 0 : -ESRCH;
 248 }
 249
 250 int cgroup_bonding_kill_list(CGroupBonding *first, int sig) {
 251         CGroupBonding *b;
 252         int r = -EAGAIN;
 253
 254         LIST_FOREACH(by_unit, b, first) {
 255                 if ((r = cgroup_bonding_kill(b, sig)) < 0) {
 256                         if (r == -EAGAIN || -ESRCH)
 257                                 continue;
 258
 259                         return r;
 260                 }
 261
 262                 return 0;
 263         }
 264
 265         return r;
 266 }
 267
 268 /* Returns 1 if the group is empty, 0 if it is not, -EAGAIN if we
 269  * cannot know */
 270 int cgroup_bonding_is_empty(CGroupBonding *b) {
 271         void *iterator;
 272         pid_t pid;
 273         int r;
 274
 275         assert(b);
 276
 277         r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid);
 278
 279         if (r == 0 || r == ECGEOF)
 280                 cgroup_get_task_end(&iterator);
 281
 282         /* Hmm, no PID in this group? Then it is definitely empty */
 283         if (r == ECGEOF)
 284                 return 1;
 285
 286         /* Some error? Let's return it */
 287         if (r != 0)
 288                 return translate_error(r, errno);
 289
 290         /* It's not empty, and we are the only user, then it is
 291          * definitely not empty */
 292         if (b->only_us)
 293                 return 0;
 294
 295         /* There are PIDs in the group but we aren't the only users,
 296          * hence we cannot say */
 297         return -EAGAIN;
 298 }
 299
 300 int cgroup_bonding_is_empty_list(CGroupBonding *first) {
 301         CGroupBonding *b;
 302
 303         LIST_FOREACH(by_unit, b, first) {
 304                 int r;
 305
 306                 if ((r = cgroup_bonding_is_empty(b)) < 0) {
 307                         /* If this returned -EAGAIN, then we don't know if the
 308                          * group is empty, so let's see if another group can
 309                          * tell us */
 310
 311                         if (r != -EAGAIN)
 312                                 return r;
 313                 } else
 314                         return r;
 315         }
 316
 317         return -EAGAIN;
 318 }
 319
 320 static int install_release_agent(Manager *m, const char *mount_point) {
 321         char *p, *c, *sc;
 322         int r;
 323
 324         assert(m);
 325         assert(mount_point);
 326
 327         if (asprintf(&p, "%s/release_agent", mount_point) < 0)
 328                 return -ENOMEM;
 329
 330         if ((r = read_one_line_file(p, &c)) < 0) {
 331                 free(p);
 332                 return r;
 333         }
 334
 335         sc = strstrip(c);
 336
 337         if (sc[0] == 0) {
 338                 if ((r = write_one_line_file(p, CGROUP_AGENT_PATH "\n" )) < 0) {
 339                         free(p);
 340                         free(c);
 341                         return r;
 342                 }
 343         } else if (!streq(sc, CGROUP_AGENT_PATH)) {
 344                 free(p);
 345                 free(c);
 346                 return -EEXIST;
 347         }
 348
 349         free(c);
 350         free(p);
 351
 352         if (asprintf(&p, "%s/notify_on_release", mount_point) < 0)
 353                 return -ENOMEM;
 354
 355         if ((r = read_one_line_file(p, &c)) < 0) {
 356                 free(p);
 357                 return r;
 358         }
 359
 360         sc = strstrip(c);
 361
 362         if (streq(sc, "0")) {
 363                 if ((r = write_one_line_file(p, "1\n")) < 0) {
 364                         free(p);
 365                         free(c);
 366                         return r;
 367                 }
 368         } else if (!streq(sc, "1")) {
 369                 free(p);
 370                 free(c);
 371                 return -EIO;
 372         }
 373
 374         free(p);
 375         free(c);
 376
 377         return 0;
 378 }
 379
 380 static int create_hierarchy_cgroup(Manager *m) {
 381         struct cgroup *cg;
 382         int r;
 383
 384         assert(m);
 385
 386         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
 387                 return -ENOMEM;
 388
 389         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
 390                 r = -ENOMEM;
 391                 goto finish;
 392         }
 393
 394         if ((r = cgroup_create_cgroup(cg, true)) != 0) {
 395                 log_error("Failed to create cgroup hierarchy group: %s", cgroup_strerror(r));
 396                 r = translate_error(r, errno);
 397                 goto finish;
 398         }
 399
 400         if ((r = cgroup_attach_task(cg)) != 0) {
 401                 log_error("Failed to add ourselves to hierarchy group: %s", cgroup_strerror(r));
 402                 r = translate_error(r, errno);
 403                 goto finish;
 404         }
 405
 406         r = 0;
 407
 408 finish:
 409         cgroup_free(&cg);
 410         return r;
 411 }
 412
 413 int manager_setup_cgroup(Manager *m) {
 414         char *mp, *cp;
 415         int r;
 416         pid_t pid;
 417         char suffix[32];
 418
 419         assert(m);
 420
 421         if ((r = cgroup_init()) != 0) {
 422                 log_error("Failed to initialize libcg: %s", cgroup_strerror(r));
 423                 return translate_error(r, errno);
 424         }
 425
 426         free(m->cgroup_controller);
 427         if (!(m->cgroup_controller = strdup("debug")))
 428                 return -ENOMEM;
 429
 430         if ((r = cgroup_get_subsys_mount_point(m->cgroup_controller, &mp)))
 431                 return translate_error(r, errno);
 432
 433         pid = getpid();
 434
 435         if ((r = cgroup_get_current_controller_path(pid, m->cgroup_controller, &cp))) {
 436                 free(mp);
 437                 return translate_error(r, errno);
 438         }
 439
 440         snprintf(suffix, sizeof(suffix), "/systemd-%u", (unsigned) pid);
 441         char_array_0(suffix);
 442
 443         free(m->cgroup_hierarchy);
 444
 445         if (endswith(cp, suffix))
 446                 /* We probably got reexecuted and can continue to use our root cgroup */
 447                 m->cgroup_hierarchy = cp;
 448         else {
 449                 /* We need a new root cgroup */
 450
 451                 m->cgroup_hierarchy = NULL;
 452                 r = asprintf(&m->cgroup_hierarchy, "%s%s", streq(cp, "/") ? "" : cp, suffix);
 453                 free(cp);
 454
 455                 if (r < 0) {
 456                         free(mp);
 457                         return -ENOMEM;
 458                 }
 459         }
 460
 461         log_info("Using cgroup controller <%s>, hierarchy mounted at <%s>, using root group <%s>.",
 462                  m->cgroup_controller,
 463                  mp,
 464                  m->cgroup_hierarchy);
 465
 466         if ((r = install_release_agent(m, mp)) < 0)
 467                 log_warning("Failed to install release agent, ignoring: %s", strerror(-r));
 468         else
 469                 log_info("Installed release agent, or already installed.");
 470
 471         free(mp);
 472
 473         if ((r = create_hierarchy_cgroup(m)) < 0)
 474                 log_error("Failed to create root cgroup hierarchy: %s", strerror(-r));
 475         else
 476                 log_info("Created root group.");
 477
 478         return r;
 479 }
 480
 481 int manager_shutdown_cgroup(Manager *m, bool delete) {
 482         struct cgroup *cg;
 483         int r;
 484
 485         assert(m);
 486
 487         if (!m->cgroup_hierarchy)
 488                 return 0;
 489
 490         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
 491                 return -ENOMEM;
 492
 493         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
 494                 r = -ENOMEM;
 495                 goto finish;
 496         }
 497
 498         /* Often enough we won't be able to delete the cgroup we
 499          * ourselves are in, hence ignore all errors here */
 500         if (delete)
 501                 cgroup_delete_cgroup_ext(cg, CGFLAG_DELETE_IGNORE_MIGRATION|CGFLAG_DELETE_RECURSIVE);
 502         r = 0;
 503
 504 finish:
 505         cgroup_free(&cg);
 506         return r;
 507
 508 }
 509
 510 int cgroup_notify_empty(Manager *m, const char *group) {
 511         CGroupBonding *l, *b;
 512
 513         assert(m);
 514         assert(group);
 515
 516         if (!(l = hashmap_get(m->cgroup_bondings, group)))
 517                 return 0;
 518
 519         LIST_FOREACH(by_path, b, l) {
 520                 int t;
 521
 522                 if (!b->unit)
 523                         continue;
 524
 525                 if ((t = cgroup_bonding_is_empty_list(b)) < 0) {
 526
 527                         /* If we don't know, we don't know */
 528                         if (t != -EAGAIN)
 529                                 log_warning("Failed to check whether cgroup is empty: %s", strerror(errno));
 530
 531                         continue;
 532                 }
 533
 534                 if (t > 0)
 535                         if (UNIT_VTABLE(b->unit)->cgroup_notify_empty)
 536                                 UNIT_VTABLE(b->unit)->cgroup_notify_empty(b->unit);
 537         }
 538
 539         return 0;
 540 }
 541
 542 CGroupBonding *cgroup_bonding_find_list(CGroupBonding *first, const char *controller) {
 543         CGroupBonding *b;
 544
 545         assert(controller);
 546
 547         LIST_FOREACH(by_unit, b, first)
 548                 if (streq(b->controller, controller))
 549                         return b;
 550
 551         return NULL;
 552 }
 553
 554 char *cgroup_bonding_to_string(CGroupBonding *b) {
 555         char *r;
 556
 557         assert(b);
 558
 559         if (asprintf(&r, "%s:%s", b->controller, b->path) < 0)
 560                 return NULL;
 561
 562         return r;
 563 }