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         free(b->controller);
 122         free(b->path);
 123
 124         if (b->cgroup) {
 125
 126                 if (b->only_us && b->clean_up)
 127                         cgroup_delete_cgroup(b->cgroup, true);
 128
 129                 cgroup_free(&b->cgroup);
 130         }
 131
 132         free(b);
 133 }
 134
 135 void cgroup_bonding_free_list(CGroupBonding *first) {
 136         CGroupBonding *b, *n;
 137
 138         LIST_FOREACH_SAFE(by_unit, b, n, first)
 139                 cgroup_bonding_free(b);
 140 }
 141
 142 int cgroup_bonding_install(CGroupBonding *b, pid_t pid) {
 143         int r;
 144
 145         assert(b);
 146         assert(pid >= 0);
 147
 148         if (pid == 0)
 149                 pid = getpid();
 150
 151         if (!b->cgroup)
 152                 return -ENOENT;
 153
 154         if ((r = cgroup_attach_task_pid(b->cgroup, pid)))
 155                 return translate_error(r, errno);
 156
 157         return 0;
 158 }
 159
 160 int cgroup_bonding_install_list(CGroupBonding *first, pid_t pid) {
 161         CGroupBonding *b;
 162
 163         LIST_FOREACH(by_unit, b, first) {
 164                 int r;
 165
 166                 if ((r = cgroup_bonding_install(b, pid)) < 0)
 167                         return r;
 168         }
 169
 170         return 0;
 171 }
 172
 173 int cgroup_bonding_kill(CGroupBonding *b, int sig) {
 174         int r;
 175         Set *s;
 176         bool done;
 177         bool killed = false;
 178
 179         assert(b);
 180         assert(sig > 0);
 181
 182         if (!b->only_us)
 183                 return -EAGAIN;
 184
 185         if (!(s = set_new(trivial_hash_func, trivial_compare_func)))
 186                 return -ENOMEM;
 187
 188         log_debug("Killing processes from process group %s:%s", b->controller, b->path);
 189
 190         do {
 191                 void *iterator;
 192                 pid_t pid;
 193
 194                 done = true;
 195
 196                 if ((r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid)) != 0) {
 197                         if (r == ECGEOF) {
 198                                 r = 0;
 199                                 goto kill_done;
 200                         } else {
 201                                 r = translate_error(r, errno);
 202                                 break;
 203                         }
 204                 }
 205
 206                 for (;;) {
 207                         if (set_get(s, INT_TO_PTR(pid)) != INT_TO_PTR(pid)) {
 208
 209                                 /* If we haven't killed this process
 210                                  * yet, kill it */
 211
 212                                 if (kill(pid, sig) < 0 && errno != ESRCH) {
 213                                         r = -errno;
 214                                         break;
 215                                 }
 216
 217                                 killed = true;
 218                                 done = false;
 219
 220                                 if ((r = set_put(s, INT_TO_PTR(pid))) < 0)
 221                                     break;
 222                         }
 223
 224                         if ((r = cgroup_get_task_next(&iterator, &pid)) != 0) {
 225
 226                                 if (r == ECGEOF)
 227                                         r = 0;
 228                                 else
 229                                         r = translate_error(r, errno);
 230
 231                                 break;
 232                         }
 233                 }
 234
 235         kill_done:
 236                 assert_se(cgroup_get_task_end(&iterator) == 0);
 237
 238                 /* To avoid racing against processes which fork
 239                  * quicker than we can kill them we repeat this until
 240                  * no new pids need to be killed. */
 241
 242         } while (!done && r >= 0);
 243
 244         set_free(s);
 245
 246         if (r < 0)
 247                 return r;
 248
 249         return killed ? 0 : -ESRCH;
 250 }
 251
 252 int cgroup_bonding_kill_list(CGroupBonding *first, int sig) {
 253         CGroupBonding *b;
 254         int r = -EAGAIN;
 255
 256         LIST_FOREACH(by_unit, b, first) {
 257                 if ((r = cgroup_bonding_kill(b, sig)) < 0) {
 258                         if (r == -EAGAIN || -ESRCH)
 259                                 continue;
 260
 261                         return r;
 262                 }
 263
 264                 return 0;
 265         }
 266
 267         return r;
 268 }
 269
 270 /* Returns 1 if the group is empty, 0 if it is not, -EAGAIN if we
 271  * cannot know */
 272 int cgroup_bonding_is_empty(CGroupBonding *b) {
 273         void *iterator;
 274         pid_t pid;
 275         int r;
 276
 277         assert(b);
 278
 279         r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid);
 280
 281         if (r == 0 || r == ECGEOF)
 282                 cgroup_get_task_end(&iterator);
 283
 284         /* Hmm, no PID in this group? Then it is definitely empty */
 285         if (r == ECGEOF)
 286                 return 1;
 287
 288         /* Some error? Let's return it */
 289         if (r != 0)
 290                 return translate_error(r, errno);
 291
 292         /* It's not empty, and we are the only user, then it is
 293          * definitely not empty */
 294         if (b->only_us)
 295                 return 0;
 296
 297         /* There are PIDs in the group but we aren't the only users,
 298          * hence we cannot say */
 299         return -EAGAIN;
 300 }
 301
 302 int cgroup_bonding_is_empty_list(CGroupBonding *first) {
 303         CGroupBonding *b;
 304
 305         LIST_FOREACH(by_unit, b, first) {
 306                 int r;
 307
 308                 if ((r = cgroup_bonding_is_empty(b)) < 0) {
 309                         /* If this returned -EAGAIN, then we don't know if the
 310                          * group is empty, so let's see if another group can
 311                          * tell us */
 312
 313                         if (r != -EAGAIN)
 314                                 return r;
 315                 } else
 316                         return r;
 317         }
 318
 319         return -EAGAIN;
 320 }
 321
 322 static int install_release_agent(Manager *m, const char *mount_point) {
 323         char *p, *c, *sc;
 324         int r;
 325
 326         assert(m);
 327         assert(mount_point);
 328
 329         if (asprintf(&p, "%s/release_agent", mount_point) < 0)
 330                 return -ENOMEM;
 331
 332         if ((r = read_one_line_file(p, &c)) < 0) {
 333                 free(p);
 334                 return r;
 335         }
 336
 337         sc = strstrip(c);
 338
 339         if (sc[0] == 0) {
 340                 if ((r = write_one_line_file(p, CGROUP_AGENT_PATH "\n" )) < 0) {
 341                         free(p);
 342                         free(c);
 343                         return r;
 344                 }
 345         } else if (!streq(sc, CGROUP_AGENT_PATH)) {
 346                 free(p);
 347                 free(c);
 348                 return -EEXIST;
 349         }
 350
 351         free(c);
 352         free(p);
 353
 354         if (asprintf(&p, "%s/notify_on_release", mount_point) < 0)
 355                 return -ENOMEM;
 356
 357         if ((r = read_one_line_file(p, &c)) < 0) {
 358                 free(p);
 359                 return r;
 360         }
 361
 362         sc = strstrip(c);
 363
 364         if (streq(sc, "0")) {
 365                 if ((r = write_one_line_file(p, "1\n")) < 0) {
 366                         free(p);
 367                         free(c);
 368                         return r;
 369                 }
 370         } else if (!streq(sc, "1")) {
 371                 free(p);
 372                 free(c);
 373                 return -EIO;
 374         }
 375
 376         free(p);
 377         free(c);
 378
 379         return 0;
 380 }
 381
 382 static int create_hierarchy_cgroup(Manager *m) {
 383         struct cgroup *cg;
 384         int r;
 385
 386         assert(m);
 387
 388         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
 389                 return -ENOMEM;
 390
 391         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
 392                 r = -ENOMEM;
 393                 goto finish;
 394         }
 395
 396         if ((r = cgroup_create_cgroup(cg, true)) != 0) {
 397                 log_error("Failed to create cgroup hierarchy group: %s", cgroup_strerror(r));
 398                 r = translate_error(r, errno);
 399                 goto finish;
 400         }
 401
 402         if ((r = cgroup_attach_task(cg)) != 0) {
 403                 log_error("Failed to add ourselves to hierarchy group: %s", cgroup_strerror(r));
 404                 r = translate_error(r, errno);
 405                 goto finish;
 406         }
 407
 408         r = 0;
 409
 410 finish:
 411         cgroup_free(&cg);
 412         return r;
 413 }
 414
 415 int manager_setup_cgroup(Manager *m) {
 416         char *mp, *cp;
 417         int r;
 418         pid_t pid;
 419
 420         assert(m);
 421
 422         if ((r = cgroup_init()) != 0) {
 423                 log_error("Failed to initialize libcg: %s", cgroup_strerror(r));
 424                 return translate_error(r, errno);
 425         }
 426
 427         free(m->cgroup_controller);
 428         if (!(m->cgroup_controller = strdup("debug")))
 429                 return -ENOMEM;
 430
 431         if ((r = cgroup_get_subsys_mount_point(m->cgroup_controller, &mp)))
 432                 return translate_error(r, errno);
 433
 434         pid = getpid();
 435
 436         if ((r = cgroup_get_current_controller_path(pid, m->cgroup_controller, &cp))) {
 437                 free(mp);
 438                 return translate_error(r, errno);
 439         }
 440
 441         free(m->cgroup_hierarchy);
 442         m->cgroup_hierarchy = NULL;
 443         if (asprintf(&m->cgroup_hierarchy, "%s/systemd-%llu", strcmp(cp, "/") == 0 ? "" : cp, (unsigned long long) pid) < 0) {
 444                 free(cp);
 445                 free(mp);
 446                 return -ENOMEM;
 447         }
 448
 449         log_info("Using cgroup controller <%s>, hierarchy mounted at <%s>, using root group <%s>.",
 450                  m->cgroup_controller,
 451                  mp,
 452                  m->cgroup_hierarchy);
 453
 454         if ((r = install_release_agent(m, mp)) < 0)
 455                 log_warning("Failed to install release agent, ignoring: %s", strerror(-r));
 456         else
 457                 log_info("Installed release agent, or already installed.");
 458
 459         free(mp);
 460         free(cp);
 461
 462         if ((r = create_hierarchy_cgroup(m)) < 0)
 463                 log_error("Failed to create root cgroup hierarchy: %s", strerror(-r));
 464         else
 465                 log_info("Created root group.");
 466
 467         return r;
 468 }
 469
 470 int manager_shutdown_cgroup(Manager *m) {
 471         struct cgroup *cg;
 472         int r;
 473
 474         assert(m);
 475
 476         if (!m->cgroup_hierarchy)
 477                 return 0;
 478
 479         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
 480                 return -ENOMEM;
 481
 482         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
 483                 r = -ENOMEM;
 484                 goto finish;
 485         }
 486
 487         if ((r = cgroup_delete_cgroup_ext(cg, CGFLAG_DELETE_IGNORE_MIGRATION|CGFLAG_DELETE_RECURSIVE)) != 0) {
 488                 log_error("Failed to delete cgroup hierarchy group: %s", cgroup_strerror(r));
 489                 r = translate_error(r, errno);
 490                 goto finish;
 491         }
 492         r = 0;
 493
 494 finish:
 495         cgroup_free(&cg);
 496         return r;
 497
 498 }
 499
 500 int cgroup_notify_empty(Manager *m, const char *group) {
 501         CGroupBonding *l, *b;
 502
 503         assert(m);
 504         assert(group);
 505
 506         if (!(l = hashmap_get(m->cgroup_bondings, group)))
 507                 return 0;
 508
 509         LIST_FOREACH(by_path, b, l) {
 510                 int t;
 511
 512                 if (!b->unit)
 513                         continue;
 514
 515                 if ((t = cgroup_bonding_is_empty_list(b)) < 0) {
 516
 517                         /* If we don't know, we don't know */
 518                         if (t != -EAGAIN)
 519                                 log_warning("Failed to check whether cgroup is empty: %s", strerror(errno));
 520
 521                         continue;
 522                 }
 523
 524                 if (t > 0)
 525                         if (UNIT_VTABLE(b->unit)->cgroup_notify_empty)
 526                                 UNIT_VTABLE(b->unit)->cgroup_notify_empty(b->unit);
 527         }
 528
 529         return 0;
 530 }
 531
 532 CGroupBonding *cgroup_bonding_find_list(CGroupBonding *first, const char *controller) {
 533         CGroupBonding *b;
 534
 535         assert(controller);
 536
 537         LIST_FOREACH(by_unit, b, first)
 538                 if (streq(b->controller, controller))
 539                         return b;
 540
 541         return NULL;
 542 }