src/basic/cgroup-util.c

   1 /***
   2   This file is part of systemd.
   3
   4   Copyright 2010 Lennart Poettering
   5
   6   systemd is free software; you can redistribute it and/or modify it
   7   under the terms of the GNU Lesser General Public License as published by
   8   the Free Software Foundation; either version 2.1 of the License, or
   9   (at your option) any later version.
  10
  11   systemd is distributed in the hope that it will be useful, but
  12   WITHOUT ANY WARRANTY; without even the implied warranty of
  13   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14   Lesser General Public License for more details.
  15
  16   You should have received a copy of the GNU Lesser General Public License
  17   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  18 ***/
  19
  20 #include <dirent.h>
  21 #include <errno.h>
  22 #include <ftw.h>
  23 #include <limits.h>
  24 #include <signal.h>
  25 #include <stddef.h>
  26 #include <stdlib.h>
  27 #include <string.h>
  28 #include <sys/stat.h>
  29 #include <sys/statfs.h>
  30 #include <sys/types.h>
  31 #include <unistd.h>
  32
  33 #include "alloc-util.h"
  34 #include "cgroup-util.h"
  35 #include "def.h"
  36 #include "dirent-util.h"
  37 #include "extract-word.h"
  38 #include "fd-util.h"
  39 #include "fileio.h"
  40 #include "formats-util.h"
  41 #include "fs-util.h"
  42 #include "log.h"
  43 #include "login-util.h"
  44 #include "macro.h"
  45 #include "missing.h"
  46 #include "mkdir.h"
  47 #include "parse-util.h"
  48 #include "path-util.h"
  49 #include "proc-cmdline.h"
  50 #include "process-util.h"
  51 #include "set.h"
  52 #include "special.h"
  53 #include "stat-util.h"
  54 #include "stdio-util.h"
  55 #include "string-table.h"
  56 #include "string-util.h"
  57 #include "unit-name.h"
  58 #include "user-util.h"
  59
  60 int cg_enumerate_processes(const char *controller, const char *path, FILE **_f) {
  61         _cleanup_free_ char *fs = NULL;
  62         FILE *f;
  63         int r;
  64
  65         assert(_f);
  66
  67         r = cg_get_path(controller, path, "cgroup.procs", &fs);
  68         if (r < 0)
  69                 return r;
  70
  71         f = fopen(fs, "re");
  72         if (!f)
  73                 return -errno;
  74
  75         *_f = f;
  76         return 0;
  77 }
  78
  79 int cg_read_pid(FILE *f, pid_t *_pid) {
  80         unsigned long ul;
  81
  82         /* Note that the cgroup.procs might contain duplicates! See
  83          * cgroups.txt for details. */
  84
  85         assert(f);
  86         assert(_pid);
  87
  88         errno = 0;
  89         if (fscanf(f, "%lu", &ul) != 1) {
  90
  91                 if (feof(f))
  92                         return 0;
  93
  94                 return errno > 0 ? -errno : -EIO;
  95         }
  96
  97         if (ul <= 0)
  98                 return -EIO;
  99
 100         *_pid = (pid_t) ul;
 101         return 1;
 102 }
 103
 104 int cg_read_event(const char *controller, const char *path, const char *event,
 105                   char **val)
 106 {
 107         _cleanup_free_ char *events = NULL, *content = NULL;
 108         char *p, *line;
 109         int r;
 110
 111         r = cg_get_path(controller, path, "cgroup.events", &events);
 112         if (r < 0)
 113                 return r;
 114
 115         r = read_full_file(events, &content, NULL);
 116         if (r < 0)
 117                 return r;
 118
 119         p = content;
 120         while ((line = strsep(&p, "\n"))) {
 121                 char *key;
 122
 123                 key = strsep(&line, " ");
 124                 if (!key || !line)
 125                         return -EINVAL;
 126
 127                 if (strcmp(key, event))
 128                         continue;
 129
 130                 *val = strdup(line);
 131                 return 0;
 132         }
 133
 134         return -ENOENT;
 135 }
 136
 137 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **_d) {
 138         _cleanup_free_ char *fs = NULL;
 139         int r;
 140         DIR *d;
 141
 142         assert(_d);
 143
 144         /* This is not recursive! */
 145
 146         r = cg_get_path(controller, path, NULL, &fs);
 147         if (r < 0)
 148                 return r;
 149
 150         d = opendir(fs);
 151         if (!d)
 152                 return -errno;
 153
 154         *_d = d;
 155         return 0;
 156 }
 157
 158 int cg_read_subgroup(DIR *d, char **fn) {
 159         struct dirent *de;
 160
 161         assert(d);
 162         assert(fn);
 163
 164         FOREACH_DIRENT_ALL(de, d, return -errno) {
 165                 char *b;
 166
 167                 if (de->d_type != DT_DIR)
 168                         continue;
 169
 170                 if (streq(de->d_name, ".") ||
 171                     streq(de->d_name, ".."))
 172                         continue;
 173
 174                 b = strdup(de->d_name);
 175                 if (!b)
 176                         return -ENOMEM;
 177
 178                 *fn = b;
 179                 return 1;
 180         }
 181
 182         return 0;
 183 }
 184
 185 int cg_rmdir(const char *controller, const char *path) {
 186         _cleanup_free_ char *p = NULL;
 187         int r;
 188
 189         r = cg_get_path(controller, path, NULL, &p);
 190         if (r < 0)
 191                 return r;
 192
 193         r = rmdir(p);
 194         if (r < 0 && errno != ENOENT)
 195                 return -errno;
 196
 197         return 0;
 198 }
 199
 200 int cg_kill(const char *controller, const char *path, int sig, bool sigcont, bool ignore_self, Set *s) {
 201         _cleanup_set_free_ Set *allocated_set = NULL;
 202         bool done = false;
 203         int r, ret = 0;
 204         pid_t my_pid;
 205
 206         assert(sig >= 0);
 207
 208         /* This goes through the tasks list and kills them all. This
 209          * is repeated until no further processes are added to the
 210          * tasks list, to properly handle forking processes */
 211
 212         if (!s) {
 213                 s = allocated_set = set_new(NULL);
 214                 if (!s)
 215                         return -ENOMEM;
 216         }
 217
 218         my_pid = getpid();
 219
 220         do {
 221                 _cleanup_fclose_ FILE *f = NULL;
 222                 pid_t pid = 0;
 223                 done = true;
 224
 225                 r = cg_enumerate_processes(controller, path, &f);
 226                 if (r < 0) {
 227                         if (ret >= 0 && r != -ENOENT)
 228                                 return r;
 229
 230                         return ret;
 231                 }
 232
 233                 while ((r = cg_read_pid(f, &pid)) > 0) {
 234
 235                         if (ignore_self && pid == my_pid)
 236                                 continue;
 237
 238                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 239                                 continue;
 240
 241                         /* If we haven't killed this process yet, kill
 242                          * it */
 243                         if (kill(pid, sig) < 0) {
 244                                 if (ret >= 0 && errno != ESRCH)
 245                                         ret = -errno;
 246                         } else {
 247                                 if (sigcont && sig != SIGKILL)
 248                                         (void) kill(pid, SIGCONT);
 249
 250                                 if (ret == 0)
 251                                         ret = 1;
 252                         }
 253
 254                         done = false;
 255
 256                         r = set_put(s, PID_TO_PTR(pid));
 257                         if (r < 0) {
 258                                 if (ret >= 0)
 259                                         return r;
 260
 261                                 return ret;
 262                         }
 263                 }
 264
 265                 if (r < 0) {
 266                         if (ret >= 0)
 267                                 return r;
 268
 269                         return ret;
 270                 }
 271
 272                 /* To avoid racing against processes which fork
 273                  * quicker than we can kill them we repeat this until
 274                  * no new pids need to be killed. */
 275
 276         } while (!done);
 277
 278         return ret;
 279 }
 280
 281 int cg_kill_recursive(const char *controller, const char *path, int sig, bool sigcont, bool ignore_self, bool rem, Set *s) {
 282         _cleanup_set_free_ Set *allocated_set = NULL;
 283         _cleanup_closedir_ DIR *d = NULL;
 284         int r, ret;
 285         char *fn;
 286
 287         assert(path);
 288         assert(sig >= 0);
 289
 290         if (!s) {
 291                 s = allocated_set = set_new(NULL);
 292                 if (!s)
 293                         return -ENOMEM;
 294         }
 295
 296         ret = cg_kill(controller, path, sig, sigcont, ignore_self, s);
 297
 298         r = cg_enumerate_subgroups(controller, path, &d);
 299         if (r < 0) {
 300                 if (ret >= 0 && r != -ENOENT)
 301                         return r;
 302
 303                 return ret;
 304         }
 305
 306         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 307                 _cleanup_free_ char *p = NULL;
 308
 309                 p = strjoin(path, "/", fn, NULL);
 310                 free(fn);
 311                 if (!p)
 312                         return -ENOMEM;
 313
 314                 r = cg_kill_recursive(controller, p, sig, sigcont, ignore_self, rem, s);
 315                 if (r != 0 && ret >= 0)
 316                         ret = r;
 317         }
 318
 319         if (ret >= 0 && r < 0)
 320                 ret = r;
 321
 322         if (rem) {
 323                 r = cg_rmdir(controller, path);
 324                 if (r < 0 && ret >= 0 && r != -ENOENT && r != -EBUSY)
 325                         return r;
 326         }
 327
 328         return ret;
 329 }
 330
 331 int cg_migrate(const char *cfrom, const char *pfrom, const char *cto, const char *pto, bool ignore_self) {
 332         bool done = false;
 333         _cleanup_set_free_ Set *s = NULL;
 334         int r, ret = 0;
 335         pid_t my_pid;
 336
 337         assert(cfrom);
 338         assert(pfrom);
 339         assert(cto);
 340         assert(pto);
 341
 342         s = set_new(NULL);
 343         if (!s)
 344                 return -ENOMEM;
 345
 346         my_pid = getpid();
 347
 348         do {
 349                 _cleanup_fclose_ FILE *f = NULL;
 350                 pid_t pid = 0;
 351                 done = true;
 352
 353                 r = cg_enumerate_processes(cfrom, pfrom, &f);
 354                 if (r < 0) {
 355                         if (ret >= 0 && r != -ENOENT)
 356                                 return r;
 357
 358                         return ret;
 359                 }
 360
 361                 while ((r = cg_read_pid(f, &pid)) > 0) {
 362
 363                         /* This might do weird stuff if we aren't a
 364                          * single-threaded program. However, we
 365                          * luckily know we are not */
 366                         if (ignore_self && pid == my_pid)
 367                                 continue;
 368
 369                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 370                                 continue;
 371
 372                         /* Ignore kernel threads. Since they can only
 373                          * exist in the root cgroup, we only check for
 374                          * them there. */
 375                         if (cfrom &&
 376                             (isempty(pfrom) || path_equal(pfrom, "/")) &&
 377                             is_kernel_thread(pid) > 0)
 378                                 continue;
 379
 380                         r = cg_attach(cto, pto, pid);
 381                         if (r < 0) {
 382                                 if (ret >= 0 && r != -ESRCH)
 383                                         ret = r;
 384                         } else if (ret == 0)
 385                                 ret = 1;
 386
 387                         done = false;
 388
 389                         r = set_put(s, PID_TO_PTR(pid));
 390                         if (r < 0) {
 391                                 if (ret >= 0)
 392                                         return r;
 393
 394                                 return ret;
 395                         }
 396                 }
 397
 398                 if (r < 0) {
 399                         if (ret >= 0)
 400                                 return r;
 401
 402                         return ret;
 403                 }
 404         } while (!done);
 405
 406         return ret;
 407 }
 408
 409 int cg_migrate_recursive(
 410                 const char *cfrom,
 411                 const char *pfrom,
 412                 const char *cto,
 413                 const char *pto,
 414                 bool ignore_self,
 415                 bool rem) {
 416
 417         _cleanup_closedir_ DIR *d = NULL;
 418         int r, ret = 0;
 419         char *fn;
 420
 421         assert(cfrom);
 422         assert(pfrom);
 423         assert(cto);
 424         assert(pto);
 425
 426         ret = cg_migrate(cfrom, pfrom, cto, pto, ignore_self);
 427
 428         r = cg_enumerate_subgroups(cfrom, pfrom, &d);
 429         if (r < 0) {
 430                 if (ret >= 0 && r != -ENOENT)
 431                         return r;
 432
 433                 return ret;
 434         }
 435
 436         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 437                 _cleanup_free_ char *p = NULL;
 438
 439                 p = strjoin(pfrom, "/", fn, NULL);
 440                 free(fn);
 441                 if (!p)
 442                         return -ENOMEM;
 443
 444                 r = cg_migrate_recursive(cfrom, p, cto, pto, ignore_self, rem);
 445                 if (r != 0 && ret >= 0)
 446                         ret = r;
 447         }
 448
 449         if (r < 0 && ret >= 0)
 450                 ret = r;
 451
 452         if (rem) {
 453                 r = cg_rmdir(cfrom, pfrom);
 454                 if (r < 0 && ret >= 0 && r != -ENOENT && r != -EBUSY)
 455                         return r;
 456         }
 457
 458         return ret;
 459 }
 460
 461 int cg_migrate_recursive_fallback(
 462                 const char *cfrom,
 463                 const char *pfrom,
 464                 const char *cto,
 465                 const char *pto,
 466                 bool ignore_self,
 467                 bool rem) {
 468
 469         int r;
 470
 471         assert(cfrom);
 472         assert(pfrom);
 473         assert(cto);
 474         assert(pto);
 475
 476         r = cg_migrate_recursive(cfrom, pfrom, cto, pto, ignore_self, rem);
 477         if (r < 0) {
 478                 char prefix[strlen(pto) + 1];
 479
 480                 /* This didn't work? Then let's try all prefixes of the destination */
 481
 482                 PATH_FOREACH_PREFIX(prefix, pto) {
 483                         int q;
 484
 485                         q = cg_migrate_recursive(cfrom, pfrom, cto, prefix, ignore_self, rem);
 486                         if (q >= 0)
 487                                 return q;
 488                 }
 489         }
 490
 491         return r;
 492 }
 493
 494 static const char *controller_to_dirname(const char *controller) {
 495         const char *e;
 496
 497         assert(controller);
 498
 499         /* Converts a controller name to the directory name below
 500          * /sys/fs/cgroup/ we want to mount it to. Effectively, this
 501          * just cuts off the name= prefixed used for named
 502          * hierarchies, if it is specified. */
 503
 504         e = startswith(controller, "name=");
 505         if (e)
 506                 return e;
 507
 508         return controller;
 509 }
 510
 511 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **fs) {
 512         const char *dn;
 513         char *t = NULL;
 514
 515         assert(fs);
 516         assert(controller);
 517
 518         dn = controller_to_dirname(controller);
 519
 520         if (isempty(path) && isempty(suffix))
 521                 t = strappend("/sys/fs/cgroup/", dn);
 522         else if (isempty(path))
 523                 t = strjoin("/sys/fs/cgroup/", dn, "/", suffix, NULL);
 524         else if (isempty(suffix))
 525                 t = strjoin("/sys/fs/cgroup/", dn, "/", path, NULL);
 526         else
 527                 t = strjoin("/sys/fs/cgroup/", dn, "/", path, "/", suffix, NULL);
 528         if (!t)
 529                 return -ENOMEM;
 530
 531         *fs = t;
 532         return 0;
 533 }
 534
 535 static int join_path_unified(const char *path, const char *suffix, char **fs) {
 536         char *t;
 537
 538         assert(fs);
 539
 540         if (isempty(path) && isempty(suffix))
 541                 t = strdup("/sys/fs/cgroup");
 542         else if (isempty(path))
 543                 t = strappend("/sys/fs/cgroup/", suffix);
 544         else if (isempty(suffix))
 545                 t = strappend("/sys/fs/cgroup/", path);
 546         else
 547                 t = strjoin("/sys/fs/cgroup/", path, "/", suffix, NULL);
 548         if (!t)
 549                 return -ENOMEM;
 550
 551         *fs = t;
 552         return 0;
 553 }
 554
 555 int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) {
 556         int unified, r;
 557
 558         assert(fs);
 559
 560         if (!controller) {
 561                 char *t;
 562
 563                 /* If no controller is specified, we return the path
 564                  * *below* the controllers, without any prefix. */
 565
 566                 if (!path && !suffix)
 567                         return -EINVAL;
 568
 569                 if (!suffix)
 570                         t = strdup(path);
 571                 else if (!path)
 572                         t = strdup(suffix);
 573                 else
 574                         t = strjoin(path, "/", suffix, NULL);
 575                 if (!t)
 576                         return -ENOMEM;
 577
 578                 *fs = path_kill_slashes(t);
 579                 return 0;
 580         }
 581
 582         if (!cg_controller_is_valid(controller))
 583                 return -EINVAL;
 584
 585         unified = cg_unified();
 586         if (unified < 0)
 587                 return unified;
 588
 589         if (unified > 0)
 590                 r = join_path_unified(path, suffix, fs);
 591         else
 592                 r = join_path_legacy(controller, path, suffix, fs);
 593         if (r < 0)
 594                 return r;
 595
 596         path_kill_slashes(*fs);
 597         return 0;
 598 }
 599
 600 static int controller_is_accessible(const char *controller) {
 601         int unified;
 602
 603         assert(controller);
 604
 605         /* Checks whether a specific controller is accessible,
 606          * i.e. its hierarchy mounted. In the unified hierarchy all
 607          * controllers are considered accessible, except for the named
 608          * hierarchies */
 609
 610         if (!cg_controller_is_valid(controller))
 611                 return -EINVAL;
 612
 613         unified = cg_unified();
 614         if (unified < 0)
 615                 return unified;
 616         if (unified > 0) {
 617                 /* We don't support named hierarchies if we are using
 618                  * the unified hierarchy. */
 619
 620                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 621                         return 0;
 622
 623                 if (startswith(controller, "name="))
 624                         return -EOPNOTSUPP;
 625
 626         } else {
 627                 const char *cc, *dn;
 628
 629                 dn = controller_to_dirname(controller);
 630                 cc = strjoina("/sys/fs/cgroup/", dn);
 631
 632                 if (laccess(cc, F_OK) < 0)
 633                         return -errno;
 634         }
 635
 636         return 0;
 637 }
 638
 639 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 640         int r;
 641
 642         assert(controller);
 643         assert(fs);
 644
 645         /* Check if the specified controller is actually accessible */
 646         r = controller_is_accessible(controller);
 647         if (r < 0)
 648                 return r;
 649
 650         return cg_get_path(controller, path, suffix, fs);
 651 }
 652
 653 static int trim_cb(const char *path, const struct stat *sb, int typeflag, struct FTW *ftwbuf) {
 654         assert(path);
 655         assert(sb);
 656         assert(ftwbuf);
 657
 658         if (typeflag != FTW_DP)
 659                 return 0;
 660
 661         if (ftwbuf->level < 1)
 662                 return 0;
 663
 664         (void) rmdir(path);
 665         return 0;
 666 }
 667
 668 int cg_trim(const char *controller, const char *path, bool delete_root) {
 669         _cleanup_free_ char *fs = NULL;
 670         int r = 0;
 671
 672         assert(path);
 673
 674         r = cg_get_path(controller, path, NULL, &fs);
 675         if (r < 0)
 676                 return r;
 677
 678         errno = 0;
 679         if (nftw(fs, trim_cb, 64, FTW_DEPTH|FTW_MOUNT|FTW_PHYS) != 0) {
 680                 if (errno == ENOENT)
 681                         r = 0;
 682                 else if (errno > 0)
 683                         r = -errno;
 684                 else
 685                         r = -EIO;
 686         }
 687
 688         if (delete_root) {
 689                 if (rmdir(fs) < 0 && errno != ENOENT)
 690                         return -errno;
 691         }
 692
 693         return r;
 694 }
 695
 696 int cg_create(const char *controller, const char *path) {
 697         _cleanup_free_ char *fs = NULL;
 698         int r;
 699
 700         r = cg_get_path_and_check(controller, path, NULL, &fs);
 701         if (r < 0)
 702                 return r;
 703
 704         r = mkdir_parents(fs, 0755);
 705         if (r < 0)
 706                 return r;
 707
 708         if (mkdir(fs, 0755) < 0) {
 709
 710                 if (errno == EEXIST)
 711                         return 0;
 712
 713                 return -errno;
 714         }
 715
 716         return 1;
 717 }
 718
 719 int cg_create_and_attach(const char *controller, const char *path, pid_t pid) {
 720         int r, q;
 721
 722         assert(pid >= 0);
 723
 724         r = cg_create(controller, path);
 725         if (r < 0)
 726                 return r;
 727
 728         q = cg_attach(controller, path, pid);
 729         if (q < 0)
 730                 return q;
 731
 732         /* This does not remove the cgroup on failure */
 733         return r;
 734 }
 735
 736 int cg_attach(const char *controller, const char *path, pid_t pid) {
 737         _cleanup_free_ char *fs = NULL;
 738         char c[DECIMAL_STR_MAX(pid_t) + 2];
 739         int r;
 740
 741         assert(path);
 742         assert(pid >= 0);
 743
 744         r = cg_get_path_and_check(controller, path, "cgroup.procs", &fs);
 745         if (r < 0)
 746                 return r;
 747
 748         if (pid == 0)
 749                 pid = getpid();
 750
 751         xsprintf(c, PID_FMT "\n", pid);
 752
 753         return write_string_file(fs, c, 0);
 754 }
 755
 756 int cg_attach_fallback(const char *controller, const char *path, pid_t pid) {
 757         int r;
 758
 759         assert(controller);
 760         assert(path);
 761         assert(pid >= 0);
 762
 763         r = cg_attach(controller, path, pid);
 764         if (r < 0) {
 765                 char prefix[strlen(path) + 1];
 766
 767                 /* This didn't work? Then let's try all prefixes of
 768                  * the destination */
 769
 770                 PATH_FOREACH_PREFIX(prefix, path) {
 771                         int q;
 772
 773                         q = cg_attach(controller, prefix, pid);
 774                         if (q >= 0)
 775                                 return q;
 776                 }
 777         }
 778
 779         return r;
 780 }
 781
 782 int cg_set_group_access(
 783                 const char *controller,
 784                 const char *path,
 785                 mode_t mode,
 786                 uid_t uid,
 787                 gid_t gid) {
 788
 789         _cleanup_free_ char *fs = NULL;
 790         int r;
 791
 792         if (mode == MODE_INVALID && uid == UID_INVALID && gid == GID_INVALID)
 793                 return 0;
 794
 795         if (mode != MODE_INVALID)
 796                 mode &= 0777;
 797
 798         r = cg_get_path(controller, path, NULL, &fs);
 799         if (r < 0)
 800                 return r;
 801
 802         return chmod_and_chown(fs, mode, uid, gid);
 803 }
 804
 805 int cg_set_task_access(
 806                 const char *controller,
 807                 const char *path,
 808                 mode_t mode,
 809                 uid_t uid,
 810                 gid_t gid) {
 811
 812         _cleanup_free_ char *fs = NULL, *procs = NULL;
 813         int r, unified;
 814
 815         assert(path);
 816
 817         if (mode == MODE_INVALID && uid == UID_INVALID && gid == GID_INVALID)
 818                 return 0;
 819
 820         if (mode != MODE_INVALID)
 821                 mode &= 0666;
 822
 823         r = cg_get_path(controller, path, "cgroup.procs", &fs);
 824         if (r < 0)
 825                 return r;
 826
 827         r = chmod_and_chown(fs, mode, uid, gid);
 828         if (r < 0)
 829                 return r;
 830
 831         unified = cg_unified();
 832         if (unified < 0)
 833                 return unified;
 834         if (unified)
 835                 return 0;
 836
 837         /* Compatibility, Always keep values for "tasks" in sync with
 838          * "cgroup.procs" */
 839         if (cg_get_path(controller, path, "tasks", &procs) >= 0)
 840                 (void) chmod_and_chown(procs, mode, uid, gid);
 841
 842         return 0;
 843 }
 844
 845 int cg_pid_get_path(const char *controller, pid_t pid, char **path) {
 846         _cleanup_fclose_ FILE *f = NULL;
 847         char line[LINE_MAX];
 848         const char *fs;
 849         size_t cs = 0;
 850         int unified;
 851
 852         assert(path);
 853         assert(pid >= 0);
 854
 855         unified = cg_unified();
 856         if (unified < 0)
 857                 return unified;
 858         if (unified == 0) {
 859                 if (controller) {
 860                         if (!cg_controller_is_valid(controller))
 861                                 return -EINVAL;
 862                 } else
 863                         controller = SYSTEMD_CGROUP_CONTROLLER;
 864
 865                 cs = strlen(controller);
 866         }
 867
 868         fs = procfs_file_alloca(pid, "cgroup");
 869         f = fopen(fs, "re");
 870         if (!f)
 871                 return errno == ENOENT ? -ESRCH : -errno;
 872
 873         FOREACH_LINE(line, f, return -errno) {
 874                 char *e, *p;
 875
 876                 truncate_nl(line);
 877
 878                 if (unified) {
 879                         e = startswith(line, "0:");
 880                         if (!e)
 881                                 continue;
 882
 883                         e = strchr(e, ':');
 884                         if (!e)
 885                                 continue;
 886                 } else {
 887                         char *l;
 888                         size_t k;
 889                         const char *word, *state;
 890                         bool found = false;
 891
 892                         l = strchr(line, ':');
 893                         if (!l)
 894                                 continue;
 895
 896                         l++;
 897                         e = strchr(l, ':');
 898                         if (!e)
 899                                 continue;
 900
 901                         *e = 0;
 902                         FOREACH_WORD_SEPARATOR(word, k, l, ",", state) {
 903                                 if (k == cs && memcmp(word, controller, cs) == 0) {
 904                                         found = true;
 905                                         break;
 906                                 }
 907                         }
 908
 909                         if (!found)
 910                                 continue;
 911                 }
 912
 913                 p = strdup(e + 1);
 914                 if (!p)
 915                         return -ENOMEM;
 916
 917                 *path = p;
 918                 return 0;
 919         }
 920
 921         return -ENODATA;
 922 }
 923
 924 int cg_install_release_agent(const char *controller, const char *agent) {
 925         _cleanup_free_ char *fs = NULL, *contents = NULL;
 926         const char *sc;
 927         int r, unified;
 928
 929         assert(agent);
 930
 931         unified = cg_unified();
 932         if (unified < 0)
 933                 return unified;
 934         if (unified) /* doesn't apply to unified hierarchy */
 935                 return -EOPNOTSUPP;
 936
 937         r = cg_get_path(controller, NULL, "release_agent", &fs);
 938         if (r < 0)
 939                 return r;
 940
 941         r = read_one_line_file(fs, &contents);
 942         if (r < 0)
 943                 return r;
 944
 945         sc = strstrip(contents);
 946         if (isempty(sc)) {
 947                 r = write_string_file(fs, agent, 0);
 948                 if (r < 0)
 949                         return r;
 950         } else if (!path_equal(sc, agent))
 951                 return -EEXIST;
 952
 953         fs = mfree(fs);
 954         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 955         if (r < 0)
 956                 return r;
 957
 958         contents = mfree(contents);
 959         r = read_one_line_file(fs, &contents);
 960         if (r < 0)
 961                 return r;
 962
 963         sc = strstrip(contents);
 964         if (streq(sc, "0")) {
 965                 r = write_string_file(fs, "1", 0);
 966                 if (r < 0)
 967                         return r;
 968
 969                 return 1;
 970         }
 971
 972         if (!streq(sc, "1"))
 973                 return -EIO;
 974
 975         return 0;
 976 }
 977
 978 int cg_uninstall_release_agent(const char *controller) {
 979         _cleanup_free_ char *fs = NULL;
 980         int r, unified;
 981
 982         unified = cg_unified();
 983         if (unified < 0)
 984                 return unified;
 985         if (unified) /* Doesn't apply to unified hierarchy */
 986                 return -EOPNOTSUPP;
 987
 988         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 989         if (r < 0)
 990                 return r;
 991
 992         r = write_string_file(fs, "0", 0);
 993         if (r < 0)
 994                 return r;
 995
 996         fs = mfree(fs);
 997
 998         r = cg_get_path(controller, NULL, "release_agent", &fs);
 999         if (r < 0)
1000                 return r;
1001
1002         r = write_string_file(fs, "", 0);
1003         if (r < 0)
1004                 return r;
1005
1006         return 0;
1007 }
1008
1009 int cg_is_empty(const char *controller, const char *path) {
1010         _cleanup_fclose_ FILE *f = NULL;
1011         pid_t pid;
1012         int r;
1013
1014         assert(path);
1015
1016         r = cg_enumerate_processes(controller, path, &f);
1017         if (r == -ENOENT)
1018                 return 1;
1019         if (r < 0)
1020                 return r;
1021
1022         r = cg_read_pid(f, &pid);
1023         if (r < 0)
1024                 return r;
1025
1026         return r == 0;
1027 }
1028
1029 int cg_is_empty_recursive(const char *controller, const char *path) {
1030         int unified, r;
1031
1032         assert(path);
1033
1034         /* The root cgroup is always populated */
1035         if (controller && (isempty(path) || path_equal(path, "/")))
1036                 return false;
1037
1038         unified = cg_unified();
1039         if (unified < 0)
1040                 return unified;
1041
1042         if (unified > 0) {
1043                 _cleanup_free_ char *t = NULL;
1044
1045                 /* On the unified hierarchy we can check empty state
1046                  * via the "populated" attribute of "cgroup.events". */
1047
1048                 r = cg_read_event(controller, path, "populated", &t);
1049                 if (r < 0)
1050                         return r;
1051
1052                 return streq(t, "0");
1053         } else {
1054                 _cleanup_closedir_ DIR *d = NULL;
1055                 char *fn;
1056
1057                 r = cg_is_empty(controller, path);
1058                 if (r <= 0)
1059                         return r;
1060
1061                 r = cg_enumerate_subgroups(controller, path, &d);
1062                 if (r == -ENOENT)
1063                         return 1;
1064                 if (r < 0)
1065                         return r;
1066
1067                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1068                         _cleanup_free_ char *p = NULL;
1069
1070                         p = strjoin(path, "/", fn, NULL);
1071                         free(fn);
1072                         if (!p)
1073                                 return -ENOMEM;
1074
1075                         r = cg_is_empty_recursive(controller, p);
1076                         if (r <= 0)
1077                                 return r;
1078                 }
1079                 if (r < 0)
1080                         return r;
1081
1082                 return true;
1083         }
1084 }
1085
1086 int cg_split_spec(const char *spec, char **controller, char **path) {
1087         char *t = NULL, *u = NULL;
1088         const char *e;
1089
1090         assert(spec);
1091
1092         if (*spec == '/') {
1093                 if (!path_is_safe(spec))
1094                         return -EINVAL;
1095
1096                 if (path) {
1097                         t = strdup(spec);
1098                         if (!t)
1099                                 return -ENOMEM;
1100
1101                         *path = path_kill_slashes(t);
1102                 }
1103
1104                 if (controller)
1105                         *controller = NULL;
1106
1107                 return 0;
1108         }
1109
1110         e = strchr(spec, ':');
1111         if (!e) {
1112                 if (!cg_controller_is_valid(spec))
1113                         return -EINVAL;
1114
1115                 if (controller) {
1116                         t = strdup(spec);
1117                         if (!t)
1118                                 return -ENOMEM;
1119
1120                         *controller = t;
1121                 }
1122
1123                 if (path)
1124                         *path = NULL;
1125
1126                 return 0;
1127         }
1128
1129         t = strndup(spec, e-spec);
1130         if (!t)
1131                 return -ENOMEM;
1132         if (!cg_controller_is_valid(t)) {
1133                 free(t);
1134                 return -EINVAL;
1135         }
1136
1137         if (isempty(e+1))
1138                 u = NULL;
1139         else {
1140                 u = strdup(e+1);
1141                 if (!u) {
1142                         free(t);
1143                         return -ENOMEM;
1144                 }
1145
1146                 if (!path_is_safe(u) ||
1147                     !path_is_absolute(u)) {
1148                         free(t);
1149                         free(u);
1150                         return -EINVAL;
1151                 }
1152
1153                 path_kill_slashes(u);
1154         }
1155
1156         if (controller)
1157                 *controller = t;
1158         else
1159                 free(t);
1160
1161         if (path)
1162                 *path = u;
1163         else
1164                 free(u);
1165
1166         return 0;
1167 }
1168
1169 int cg_mangle_path(const char *path, char **result) {
1170         _cleanup_free_ char *c = NULL, *p = NULL;
1171         char *t;
1172         int r;
1173
1174         assert(path);
1175         assert(result);
1176
1177         /* First, check if it already is a filesystem path */
1178         if (path_startswith(path, "/sys/fs/cgroup")) {
1179
1180                 t = strdup(path);
1181                 if (!t)
1182                         return -ENOMEM;
1183
1184                 *result = path_kill_slashes(t);
1185                 return 0;
1186         }
1187
1188         /* Otherwise, treat it as cg spec */
1189         r = cg_split_spec(path, &c, &p);
1190         if (r < 0)
1191                 return r;
1192
1193         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
1194 }
1195
1196 int cg_get_root_path(char **path) {
1197         char *p, *e;
1198         int r;
1199
1200         assert(path);
1201
1202         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1203         if (r < 0)
1204                 return r;
1205
1206         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1207         if (!e)
1208                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1209         if (!e)
1210                 e = endswith(p, "/system"); /* even more legacy */
1211         if (e)
1212                 *e = 0;
1213
1214         *path = p;
1215         return 0;
1216 }
1217
1218 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1219         _cleanup_free_ char *rt = NULL;
1220         char *p;
1221         int r;
1222
1223         assert(cgroup);
1224         assert(shifted);
1225
1226         if (!root) {
1227                 /* If the root was specified let's use that, otherwise
1228                  * let's determine it from PID 1 */
1229
1230                 r = cg_get_root_path(&rt);
1231                 if (r < 0)
1232                         return r;
1233
1234                 root = rt;
1235         }
1236
1237         p = path_startswith(cgroup, root);
1238         if (p && p > cgroup)
1239                 *shifted = p - 1;
1240         else
1241                 *shifted = cgroup;
1242
1243         return 0;
1244 }
1245
1246 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1247         _cleanup_free_ char *raw = NULL;
1248         const char *c;
1249         int r;
1250
1251         assert(pid >= 0);
1252         assert(cgroup);
1253
1254         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1255         if (r < 0)
1256                 return r;
1257
1258         r = cg_shift_path(raw, root, &c);
1259         if (r < 0)
1260                 return r;
1261
1262         if (c == raw) {
1263                 *cgroup = raw;
1264                 raw = NULL;
1265         } else {
1266                 char *n;
1267
1268                 n = strdup(c);
1269                 if (!n)
1270                         return -ENOMEM;
1271
1272                 *cgroup = n;
1273         }
1274
1275         return 0;
1276 }
1277
1278 int cg_path_decode_unit(const char *cgroup, char **unit) {
1279         char *c, *s;
1280         size_t n;
1281
1282         assert(cgroup);
1283         assert(unit);
1284
1285         n = strcspn(cgroup, "/");
1286         if (n < 3)
1287                 return -ENXIO;
1288
1289         c = strndupa(cgroup, n);
1290         c = cg_unescape(c);
1291
1292         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1293                 return -ENXIO;
1294
1295         s = strdup(c);
1296         if (!s)
1297                 return -ENOMEM;
1298
1299         *unit = s;
1300         return 0;
1301 }
1302
1303 static bool valid_slice_name(const char *p, size_t n) {
1304
1305         if (!p)
1306                 return false;
1307
1308         if (n < strlen("x.slice"))
1309                 return false;
1310
1311         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1312                 char buf[n+1], *c;
1313
1314                 memcpy(buf, p, n);
1315                 buf[n] = 0;
1316
1317                 c = cg_unescape(buf);
1318
1319                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1320         }
1321
1322         return false;
1323 }
1324
1325 static const char *skip_slices(const char *p) {
1326         assert(p);
1327
1328         /* Skips over all slice assignments */
1329
1330         for (;;) {
1331                 size_t n;
1332
1333                 p += strspn(p, "/");
1334
1335                 n = strcspn(p, "/");
1336                 if (!valid_slice_name(p, n))
1337                         return p;
1338
1339                 p += n;
1340         }
1341 }
1342
1343 int cg_path_get_unit(const char *path, char **ret) {
1344         const char *e;
1345         char *unit;
1346         int r;
1347
1348         assert(path);
1349         assert(ret);
1350
1351         e = skip_slices(path);
1352
1353         r = cg_path_decode_unit(e, &unit);
1354         if (r < 0)
1355                 return r;
1356
1357         /* We skipped over the slices, don't accept any now */
1358         if (endswith(unit, ".slice")) {
1359                 free(unit);
1360                 return -ENXIO;
1361         }
1362
1363         *ret = unit;
1364         return 0;
1365 }
1366
1367 int cg_pid_get_unit(pid_t pid, char **unit) {
1368         _cleanup_free_ char *cgroup = NULL;
1369         int r;
1370
1371         assert(unit);
1372
1373         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1374         if (r < 0)
1375                 return r;
1376
1377         return cg_path_get_unit(cgroup, unit);
1378 }
1379
1380 /**
1381  * Skip session-*.scope, but require it to be there.
1382  */
1383 static const char *skip_session(const char *p) {
1384         size_t n;
1385
1386         if (isempty(p))
1387                 return NULL;
1388
1389         p += strspn(p, "/");
1390
1391         n = strcspn(p, "/");
1392         if (n < strlen("session-x.scope"))
1393                 return NULL;
1394
1395         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1396                 char buf[n - 8 - 6 + 1];
1397
1398                 memcpy(buf, p + 8, n - 8 - 6);
1399                 buf[n - 8 - 6] = 0;
1400
1401                 /* Note that session scopes never need unescaping,
1402                  * since they cannot conflict with the kernel's own
1403                  * names, hence we don't need to call cg_unescape()
1404                  * here. */
1405
1406                 if (!session_id_valid(buf))
1407                         return false;
1408
1409                 p += n;
1410                 p += strspn(p, "/");
1411                 return p;
1412         }
1413
1414         return NULL;
1415 }
1416
1417 /**
1418  * Skip user@*.service, but require it to be there.
1419  */
1420 static const char *skip_user_manager(const char *p) {
1421         size_t n;
1422
1423         if (isempty(p))
1424                 return NULL;
1425
1426         p += strspn(p, "/");
1427
1428         n = strcspn(p, "/");
1429         if (n < strlen("user@x.service"))
1430                 return NULL;
1431
1432         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1433                 char buf[n - 5 - 8 + 1];
1434
1435                 memcpy(buf, p + 5, n - 5 - 8);
1436                 buf[n - 5 - 8] = 0;
1437
1438                 /* Note that user manager services never need unescaping,
1439                  * since they cannot conflict with the kernel's own
1440                  * names, hence we don't need to call cg_unescape()
1441                  * here. */
1442
1443                 if (parse_uid(buf, NULL) < 0)
1444                         return NULL;
1445
1446                 p += n;
1447                 p += strspn(p, "/");
1448
1449                 return p;
1450         }
1451
1452         return NULL;
1453 }
1454
1455 static const char *skip_user_prefix(const char *path) {
1456         const char *e, *t;
1457
1458         assert(path);
1459
1460         /* Skip slices, if there are any */
1461         e = skip_slices(path);
1462
1463         /* Skip the user manager, if it's in the path now... */
1464         t = skip_user_manager(e);
1465         if (t)
1466                 return t;
1467
1468         /* Alternatively skip the user session if it is in the path... */
1469         return skip_session(e);
1470 }
1471
1472 int cg_path_get_user_unit(const char *path, char **ret) {
1473         const char *t;
1474
1475         assert(path);
1476         assert(ret);
1477
1478         t = skip_user_prefix(path);
1479         if (!t)
1480                 return -ENXIO;
1481
1482         /* And from here on it looks pretty much the same as for a
1483          * system unit, hence let's use the same parser from here
1484          * on. */
1485         return cg_path_get_unit(t, ret);
1486 }
1487
1488 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1489         _cleanup_free_ char *cgroup = NULL;
1490         int r;
1491
1492         assert(unit);
1493
1494         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1495         if (r < 0)
1496                 return r;
1497
1498         return cg_path_get_user_unit(cgroup, unit);
1499 }
1500
1501 int cg_path_get_machine_name(const char *path, char **machine) {
1502         _cleanup_free_ char *u = NULL;
1503         const char *sl;
1504         int r;
1505
1506         r = cg_path_get_unit(path, &u);
1507         if (r < 0)
1508                 return r;
1509
1510         sl = strjoina("/run/systemd/machines/unit:", u);
1511         return readlink_malloc(sl, machine);
1512 }
1513
1514 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1515         _cleanup_free_ char *cgroup = NULL;
1516         int r;
1517
1518         assert(machine);
1519
1520         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1521         if (r < 0)
1522                 return r;
1523
1524         return cg_path_get_machine_name(cgroup, machine);
1525 }
1526
1527 int cg_path_get_session(const char *path, char **session) {
1528         _cleanup_free_ char *unit = NULL;
1529         char *start, *end;
1530         int r;
1531
1532         assert(path);
1533
1534         r = cg_path_get_unit(path, &unit);
1535         if (r < 0)
1536                 return r;
1537
1538         start = startswith(unit, "session-");
1539         if (!start)
1540                 return -ENXIO;
1541         end = endswith(start, ".scope");
1542         if (!end)
1543                 return -ENXIO;
1544
1545         *end = 0;
1546         if (!session_id_valid(start))
1547                 return -ENXIO;
1548
1549         if (session) {
1550                 char *rr;
1551
1552                 rr = strdup(start);
1553                 if (!rr)
1554                         return -ENOMEM;
1555
1556                 *session = rr;
1557         }
1558
1559         return 0;
1560 }
1561
1562 int cg_pid_get_session(pid_t pid, char **session) {
1563         _cleanup_free_ char *cgroup = NULL;
1564         int r;
1565
1566         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1567         if (r < 0)
1568                 return r;
1569
1570         return cg_path_get_session(cgroup, session);
1571 }
1572
1573 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1574         _cleanup_free_ char *slice = NULL;
1575         char *start, *end;
1576         int r;
1577
1578         assert(path);
1579
1580         r = cg_path_get_slice(path, &slice);
1581         if (r < 0)
1582                 return r;
1583
1584         start = startswith(slice, "user-");
1585         if (!start)
1586                 return -ENXIO;
1587         end = endswith(start, ".slice");
1588         if (!end)
1589                 return -ENXIO;
1590
1591         *end = 0;
1592         if (parse_uid(start, uid) < 0)
1593                 return -ENXIO;
1594
1595         return 0;
1596 }
1597
1598 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1599         _cleanup_free_ char *cgroup = NULL;
1600         int r;
1601
1602         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1603         if (r < 0)
1604                 return r;
1605
1606         return cg_path_get_owner_uid(cgroup, uid);
1607 }
1608
1609 int cg_path_get_slice(const char *p, char **slice) {
1610         const char *e = NULL;
1611
1612         assert(p);
1613         assert(slice);
1614
1615         /* Finds the right-most slice unit from the beginning, but
1616          * stops before we come to the first non-slice unit. */
1617
1618         for (;;) {
1619                 size_t n;
1620
1621                 p += strspn(p, "/");
1622
1623                 n = strcspn(p, "/");
1624                 if (!valid_slice_name(p, n)) {
1625
1626                         if (!e) {
1627                                 char *s;
1628
1629                                 s = strdup("-.slice");
1630                                 if (!s)
1631                                         return -ENOMEM;
1632
1633                                 *slice = s;
1634                                 return 0;
1635                         }
1636
1637                         return cg_path_decode_unit(e, slice);
1638                 }
1639
1640                 e = p;
1641                 p += n;
1642         }
1643 }
1644
1645 int cg_pid_get_slice(pid_t pid, char **slice) {
1646         _cleanup_free_ char *cgroup = NULL;
1647         int r;
1648
1649         assert(slice);
1650
1651         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1652         if (r < 0)
1653                 return r;
1654
1655         return cg_path_get_slice(cgroup, slice);
1656 }
1657
1658 int cg_path_get_user_slice(const char *p, char **slice) {
1659         const char *t;
1660         assert(p);
1661         assert(slice);
1662
1663         t = skip_user_prefix(p);
1664         if (!t)
1665                 return -ENXIO;
1666
1667         /* And now it looks pretty much the same as for a system
1668          * slice, so let's just use the same parser from here on. */
1669         return cg_path_get_slice(t, slice);
1670 }
1671
1672 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1673         _cleanup_free_ char *cgroup = NULL;
1674         int r;
1675
1676         assert(slice);
1677
1678         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1679         if (r < 0)
1680                 return r;
1681
1682         return cg_path_get_user_slice(cgroup, slice);
1683 }
1684
1685 char *cg_escape(const char *p) {
1686         bool need_prefix = false;
1687
1688         /* This implements very minimal escaping for names to be used
1689          * as file names in the cgroup tree: any name which might
1690          * conflict with a kernel name or is prefixed with '_' is
1691          * prefixed with a '_'. That way, when reading cgroup names it
1692          * is sufficient to remove a single prefixing underscore if
1693          * there is one. */
1694
1695         /* The return value of this function (unlike cg_unescape())
1696          * needs free()! */
1697
1698         if (p[0] == 0 ||
1699             p[0] == '_' ||
1700             p[0] == '.' ||
1701             streq(p, "notify_on_release") ||
1702             streq(p, "release_agent") ||
1703             streq(p, "tasks") ||
1704             startswith(p, "cgroup."))
1705                 need_prefix = true;
1706         else {
1707                 const char *dot;
1708
1709                 dot = strrchr(p, '.');
1710                 if (dot) {
1711                         CGroupController c;
1712                         size_t l = dot - p;
1713
1714                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1715                                 const char *n;
1716
1717                                 n = cgroup_controller_to_string(c);
1718
1719                                 if (l != strlen(n))
1720                                         continue;
1721
1722                                 if (memcmp(p, n, l) != 0)
1723                                         continue;
1724
1725                                 need_prefix = true;
1726                                 break;
1727                         }
1728                 }
1729         }
1730
1731         if (need_prefix)
1732                 return strappend("_", p);
1733
1734         return strdup(p);
1735 }
1736
1737 char *cg_unescape(const char *p) {
1738         assert(p);
1739
1740         /* The return value of this function (unlike cg_escape())
1741          * doesn't need free()! */
1742
1743         if (p[0] == '_')
1744                 return (char*) p+1;
1745
1746         return (char*) p;
1747 }
1748
1749 #define CONTROLLER_VALID                        \
1750         DIGITS LETTERS                          \
1751         "_"
1752
1753 bool cg_controller_is_valid(const char *p) {
1754         const char *t, *s;
1755
1756         if (!p)
1757                 return false;
1758
1759         s = startswith(p, "name=");
1760         if (s)
1761                 p = s;
1762
1763         if (*p == 0 || *p == '_')
1764                 return false;
1765
1766         for (t = p; *t; t++)
1767                 if (!strchr(CONTROLLER_VALID, *t))
1768                         return false;
1769
1770         if (t - p > FILENAME_MAX)
1771                 return false;
1772
1773         return true;
1774 }
1775
1776 int cg_slice_to_path(const char *unit, char **ret) {
1777         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1778         const char *dash;
1779         int r;
1780
1781         assert(unit);
1782         assert(ret);
1783
1784         if (streq(unit, "-.slice")) {
1785                 char *x;
1786
1787                 x = strdup("");
1788                 if (!x)
1789                         return -ENOMEM;
1790                 *ret = x;
1791                 return 0;
1792         }
1793
1794         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1795                 return -EINVAL;
1796
1797         if (!endswith(unit, ".slice"))
1798                 return -EINVAL;
1799
1800         r = unit_name_to_prefix(unit, &p);
1801         if (r < 0)
1802                 return r;
1803
1804         dash = strchr(p, '-');
1805
1806         /* Don't allow initial dashes */
1807         if (dash == p)
1808                 return -EINVAL;
1809
1810         while (dash) {
1811                 _cleanup_free_ char *escaped = NULL;
1812                 char n[dash - p + sizeof(".slice")];
1813
1814                 /* Don't allow trailing or double dashes */
1815                 if (dash[1] == 0 || dash[1] == '-')
1816                         return -EINVAL;
1817
1818                 strcpy(stpncpy(n, p, dash - p), ".slice");
1819                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1820                         return -EINVAL;
1821
1822                 escaped = cg_escape(n);
1823                 if (!escaped)
1824                         return -ENOMEM;
1825
1826                 if (!strextend(&s, escaped, "/", NULL))
1827                         return -ENOMEM;
1828
1829                 dash = strchr(dash+1, '-');
1830         }
1831
1832         e = cg_escape(unit);
1833         if (!e)
1834                 return -ENOMEM;
1835
1836         if (!strextend(&s, e, NULL))
1837                 return -ENOMEM;
1838
1839         *ret = s;
1840         s = NULL;
1841
1842         return 0;
1843 }
1844
1845 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1846         _cleanup_free_ char *p = NULL;
1847         int r;
1848
1849         r = cg_get_path(controller, path, attribute, &p);
1850         if (r < 0)
1851                 return r;
1852
1853         return write_string_file(p, value, 0);
1854 }
1855
1856 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1857         _cleanup_free_ char *p = NULL;
1858         int r;
1859
1860         r = cg_get_path(controller, path, attribute, &p);
1861         if (r < 0)
1862                 return r;
1863
1864         return read_one_line_file(p, ret);
1865 }
1866
1867 int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) {
1868         CGroupController c;
1869         int r, unified;
1870
1871         /* This one will create a cgroup in our private tree, but also
1872          * duplicate it in the trees specified in mask, and remove it
1873          * in all others */
1874
1875         /* First create the cgroup in our own hierarchy. */
1876         r = cg_create(SYSTEMD_CGROUP_CONTROLLER, path);
1877         if (r < 0)
1878                 return r;
1879
1880         /* If we are in the unified hierarchy, we are done now */
1881         unified = cg_unified();
1882         if (unified < 0)
1883                 return unified;
1884         if (unified > 0)
1885                 return 0;
1886
1887         /* Otherwise, do the same in the other hierarchies */
1888         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1889                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1890                 const char *n;
1891
1892                 n = cgroup_controller_to_string(c);
1893
1894                 if (mask & bit)
1895                         (void) cg_create(n, path);
1896                 else if (supported & bit)
1897                         (void) cg_trim(n, path, true);
1898         }
1899
1900         return 0;
1901 }
1902
1903 int cg_attach_everywhere(CGroupMask supported, const char *path, pid_t pid, cg_migrate_callback_t path_callback, void *userdata) {
1904         CGroupController c;
1905         int r, unified;
1906
1907         r = cg_attach(SYSTEMD_CGROUP_CONTROLLER, path, pid);
1908         if (r < 0)
1909                 return r;
1910
1911         unified = cg_unified();
1912         if (unified < 0)
1913                 return unified;
1914         if (unified > 0)
1915                 return 0;
1916
1917         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1918                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1919                 const char *p = NULL;
1920
1921                 if (!(supported & bit))
1922                         continue;
1923
1924                 if (path_callback)
1925                         p = path_callback(bit, userdata);
1926
1927                 if (!p)
1928                         p = path;
1929
1930                 (void) cg_attach_fallback(cgroup_controller_to_string(c), p, pid);
1931         }
1932
1933         return 0;
1934 }
1935
1936 int cg_attach_many_everywhere(CGroupMask supported, const char *path, Set* pids, cg_migrate_callback_t path_callback, void *userdata) {
1937         Iterator i;
1938         void *pidp;
1939         int r = 0;
1940
1941         SET_FOREACH(pidp, pids, i) {
1942                 pid_t pid = PTR_TO_PID(pidp);
1943                 int q;
1944
1945                 q = cg_attach_everywhere(supported, path, pid, path_callback, userdata);
1946                 if (q < 0 && r >= 0)
1947                         r = q;
1948         }
1949
1950         return r;
1951 }
1952
1953 int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to, cg_migrate_callback_t to_callback, void *userdata) {
1954         CGroupController c;
1955         int r = 0, unified;
1956
1957         if (!path_equal(from, to))  {
1958                 r = cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER, from, SYSTEMD_CGROUP_CONTROLLER, to, false, true);
1959                 if (r < 0)
1960                         return r;
1961         }
1962
1963         unified = cg_unified();
1964         if (unified < 0)
1965                 return unified;
1966         if (unified > 0)
1967                 return r;
1968
1969         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1970                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1971                 const char *p = NULL;
1972
1973                 if (!(supported & bit))
1974                         continue;
1975
1976                 if (to_callback)
1977                         p = to_callback(bit, userdata);
1978
1979                 if (!p)
1980                         p = to;
1981
1982                 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER, to, cgroup_controller_to_string(c), p, false, false);
1983         }
1984
1985         return 0;
1986 }
1987
1988 int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) {
1989         CGroupController c;
1990         int r, unified;
1991
1992         r = cg_trim(SYSTEMD_CGROUP_CONTROLLER, path, delete_root);
1993         if (r < 0)
1994                 return r;
1995
1996         unified = cg_unified();
1997         if (unified < 0)
1998                 return unified;
1999         if (unified > 0)
2000                 return r;
2001
2002         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2003                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2004
2005                 if (!(supported & bit))
2006                         continue;
2007
2008                 (void) cg_trim(cgroup_controller_to_string(c), path, delete_root);
2009         }
2010
2011         return 0;
2012 }
2013
2014 int cg_mask_supported(CGroupMask *ret) {
2015         CGroupMask mask = 0;
2016         int r, unified;
2017
2018         /* Determines the mask of supported cgroup controllers. Only
2019          * includes controllers we can make sense of and that are
2020          * actually accessible. */
2021
2022         unified = cg_unified();
2023         if (unified < 0)
2024                 return unified;
2025         if (unified > 0) {
2026                 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
2027                 const char *c;
2028
2029                 /* In the unified hierarchy we can read the supported
2030                  * and accessible controllers from a the top-level
2031                  * cgroup attribute */
2032
2033                 r = cg_get_root_path(&root);
2034                 if (r < 0)
2035                         return r;
2036
2037                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
2038                 if (r < 0)
2039                         return r;
2040
2041                 r = read_one_line_file(path, &controllers);
2042                 if (r < 0)
2043                         return r;
2044
2045                 c = controllers;
2046                 for (;;) {
2047                         _cleanup_free_ char *n = NULL;
2048                         CGroupController v;
2049
2050                         r = extract_first_word(&c, &n, NULL, 0);
2051                         if (r < 0)
2052                                 return r;
2053                         if (r == 0)
2054                                 break;
2055
2056                         v = cgroup_controller_from_string(n);
2057                         if (v < 0)
2058                                 continue;
2059
2060                         mask |= CGROUP_CONTROLLER_TO_MASK(v);
2061                 }
2062
2063                 /* Currently, we only support the memory, io and pids
2064                  * controller in the unified hierarchy, mask
2065                  * everything else off. */
2066                 mask &= CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
2067
2068         } else {
2069                 CGroupController c;
2070
2071                 /* In the legacy hierarchy, we check whether which
2072                  * hierarchies are mounted. */
2073
2074                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2075                         const char *n;
2076
2077                         n = cgroup_controller_to_string(c);
2078                         if (controller_is_accessible(n) >= 0)
2079                                 mask |= CGROUP_CONTROLLER_TO_MASK(c);
2080                 }
2081         }
2082
2083         *ret = mask;
2084         return 0;
2085 }
2086
2087 int cg_kernel_controllers(Set *controllers) {
2088         _cleanup_fclose_ FILE *f = NULL;
2089         char buf[LINE_MAX];
2090         int r;
2091
2092         assert(controllers);
2093
2094         /* Determines the full list of kernel-known controllers. Might
2095          * include controllers we don't actually support, arbitrary
2096          * named hierarchies and controllers that aren't currently
2097          * accessible (because not mounted). */
2098
2099         f = fopen("/proc/cgroups", "re");
2100         if (!f) {
2101                 if (errno == ENOENT)
2102                         return 0;
2103                 return -errno;
2104         }
2105
2106         /* Ignore the header line */
2107         (void) fgets(buf, sizeof(buf), f);
2108
2109         for (;;) {
2110                 char *controller;
2111                 int enabled = 0;
2112
2113                 errno = 0;
2114                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
2115
2116                         if (feof(f))
2117                                 break;
2118
2119                         if (ferror(f) && errno > 0)
2120                                 return -errno;
2121
2122                         return -EBADMSG;
2123                 }
2124
2125                 if (!enabled) {
2126                         free(controller);
2127                         continue;
2128                 }
2129
2130                 if (!cg_controller_is_valid(controller)) {
2131                         free(controller);
2132                         return -EBADMSG;
2133                 }
2134
2135                 r = set_consume(controllers, controller);
2136                 if (r < 0)
2137                         return r;
2138         }
2139
2140         return 0;
2141 }
2142
2143 static thread_local int unified_cache = -1;
2144
2145 int cg_unified(void) {
2146         struct statfs fs;
2147
2148         /* Checks if we support the unified hierarchy. Returns an
2149          * error when the cgroup hierarchies aren't mounted yet or we
2150          * have any other trouble determining if the unified hierarchy
2151          * is supported. */
2152
2153         if (unified_cache >= 0)
2154                 return unified_cache;
2155
2156         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2157                 return -errno;
2158
2159         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC))
2160                 unified_cache = true;
2161         else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC))
2162                 unified_cache = false;
2163         else
2164                 return -ENOMEDIUM;
2165
2166         return unified_cache;
2167 }
2168
2169 void cg_unified_flush(void) {
2170         unified_cache = -1;
2171 }
2172
2173 int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) {
2174         _cleanup_free_ char *fs = NULL;
2175         CGroupController c;
2176         int r, unified;
2177
2178         assert(p);
2179
2180         if (supported == 0)
2181                 return 0;
2182
2183         unified = cg_unified();
2184         if (unified < 0)
2185                 return unified;
2186         if (!unified) /* on the legacy hiearchy there's no joining of controllers defined */
2187                 return 0;
2188
2189         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, p, "cgroup.subtree_control", &fs);
2190         if (r < 0)
2191                 return r;
2192
2193         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2194                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2195                 const char *n;
2196
2197                 if (!(supported & bit))
2198                         continue;
2199
2200                 n = cgroup_controller_to_string(c);
2201                 {
2202                         char s[1 + strlen(n) + 1];
2203
2204                         s[0] = mask & bit ? '+' : '-';
2205                         strcpy(s + 1, n);
2206
2207                         r = write_string_file(fs, s, 0);
2208                         if (r < 0)
2209                                 log_debug_errno(r, "Failed to enable controller %s for %s (%s): %m", n, p, fs);
2210                 }
2211         }
2212
2213         return 0;
2214 }
2215
2216 bool cg_is_unified_wanted(void) {
2217         static thread_local int wanted = -1;
2218         int r, unified;
2219
2220         /* If the hierarchy is already mounted, then follow whatever
2221          * was chosen for it. */
2222         unified = cg_unified();
2223         if (unified >= 0)
2224                 return unified;
2225
2226         /* Otherwise, let's see what the kernel command line has to
2227          * say. Since checking that is expensive, let's cache the
2228          * result. */
2229         if (wanted >= 0)
2230                 return wanted;
2231
2232         r = get_proc_cmdline_key("systemd.unified_cgroup_hierarchy", NULL);
2233         if (r > 0)
2234                 return (wanted = true);
2235         else {
2236                 _cleanup_free_ char *value = NULL;
2237
2238                 r = get_proc_cmdline_key("systemd.unified_cgroup_hierarchy=", &value);
2239                 if (r < 0)
2240                         return false;
2241                 if (r == 0)
2242                         return (wanted = false);
2243
2244                 return (wanted = parse_boolean(value) > 0);
2245         }
2246 }
2247
2248 bool cg_is_legacy_wanted(void) {
2249         return !cg_is_unified_wanted();
2250 }
2251
2252 int cg_weight_parse(const char *s, uint64_t *ret) {
2253         uint64_t u;
2254         int r;
2255
2256         if (isempty(s)) {
2257                 *ret = CGROUP_WEIGHT_INVALID;
2258                 return 0;
2259         }
2260
2261         r = safe_atou64(s, &u);
2262         if (r < 0)
2263                 return r;
2264
2265         if (u < CGROUP_WEIGHT_MIN || u > CGROUP_WEIGHT_MAX)
2266                 return -ERANGE;
2267
2268         *ret = u;
2269         return 0;
2270 }
2271
2272 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2273         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2274         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2275 };
2276
2277 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2278         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2279         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2280 };
2281
2282 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2283
2284 int cg_cpu_shares_parse(const char *s, uint64_t *ret) {
2285         uint64_t u;
2286         int r;
2287
2288         if (isempty(s)) {
2289                 *ret = CGROUP_CPU_SHARES_INVALID;
2290                 return 0;
2291         }
2292
2293         r = safe_atou64(s, &u);
2294         if (r < 0)
2295                 return r;
2296
2297         if (u < CGROUP_CPU_SHARES_MIN || u > CGROUP_CPU_SHARES_MAX)
2298                 return -ERANGE;
2299
2300         *ret = u;
2301         return 0;
2302 }
2303
2304 int cg_blkio_weight_parse(const char *s, uint64_t *ret) {
2305         uint64_t u;
2306         int r;
2307
2308         if (isempty(s)) {
2309                 *ret = CGROUP_BLKIO_WEIGHT_INVALID;
2310                 return 0;
2311         }
2312
2313         r = safe_atou64(s, &u);
2314         if (r < 0)
2315                 return r;
2316
2317         if (u < CGROUP_BLKIO_WEIGHT_MIN || u > CGROUP_BLKIO_WEIGHT_MAX)
2318                 return -ERANGE;
2319
2320         *ret = u;
2321         return 0;
2322 }
2323
2324 static const char *cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2325         [CGROUP_CONTROLLER_CPU] = "cpu",
2326         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2327         [CGROUP_CONTROLLER_IO] = "io",
2328         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2329         [CGROUP_CONTROLLER_MEMORY] = "memory",
2330         [CGROUP_CONTROLLER_DEVICES] = "devices",
2331         [CGROUP_CONTROLLER_PIDS] = "pids",
2332 };
2333
2334 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);