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 bool cg_ns_supported(void) {
 138         static thread_local int enabled = -1;
 139
 140         if (enabled >= 0)
 141                 return enabled;
 142
 143         if (access("/proc/self/ns/cgroup", F_OK) == 0)
 144                 enabled = 1;
 145         else
 146                 enabled = 0;
 147
 148         return enabled;
 149 }
 150
 151 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **_d) {
 152         _cleanup_free_ char *fs = NULL;
 153         int r;
 154         DIR *d;
 155
 156         assert(_d);
 157
 158         /* This is not recursive! */
 159
 160         r = cg_get_path(controller, path, NULL, &fs);
 161         if (r < 0)
 162                 return r;
 163
 164         d = opendir(fs);
 165         if (!d)
 166                 return -errno;
 167
 168         *_d = d;
 169         return 0;
 170 }
 171
 172 int cg_read_subgroup(DIR *d, char **fn) {
 173         struct dirent *de;
 174
 175         assert(d);
 176         assert(fn);
 177
 178         FOREACH_DIRENT_ALL(de, d, return -errno) {
 179                 char *b;
 180
 181                 if (de->d_type != DT_DIR)
 182                         continue;
 183
 184                 if (streq(de->d_name, ".") ||
 185                     streq(de->d_name, ".."))
 186                         continue;
 187
 188                 b = strdup(de->d_name);
 189                 if (!b)
 190                         return -ENOMEM;
 191
 192                 *fn = b;
 193                 return 1;
 194         }
 195
 196         return 0;
 197 }
 198
 199 int cg_rmdir(const char *controller, const char *path) {
 200         _cleanup_free_ char *p = NULL;
 201         int r;
 202
 203         r = cg_get_path(controller, path, NULL, &p);
 204         if (r < 0)
 205                 return r;
 206
 207         r = rmdir(p);
 208         if (r < 0 && errno != ENOENT)
 209                 return -errno;
 210
 211         return 0;
 212 }
 213
 214 int cg_kill(
 215                 const char *controller,
 216                 const char *path,
 217                 int sig,
 218                 CGroupFlags flags,
 219                 Set *s,
 220                 cg_kill_log_func_t log_kill,
 221                 void *userdata) {
 222
 223         _cleanup_set_free_ Set *allocated_set = NULL;
 224         bool done = false;
 225         int r, ret = 0;
 226         pid_t my_pid;
 227
 228         assert(sig >= 0);
 229
 230          /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
 231           * SIGCONT on SIGKILL. */
 232         if (IN_SET(sig, SIGCONT, SIGKILL))
 233                 flags &= ~CGROUP_SIGCONT;
 234
 235         /* This goes through the tasks list and kills them all. This
 236          * is repeated until no further processes are added to the
 237          * tasks list, to properly handle forking processes */
 238
 239         if (!s) {
 240                 s = allocated_set = set_new(NULL);
 241                 if (!s)
 242                         return -ENOMEM;
 243         }
 244
 245         my_pid = getpid();
 246
 247         do {
 248                 _cleanup_fclose_ FILE *f = NULL;
 249                 pid_t pid = 0;
 250                 done = true;
 251
 252                 r = cg_enumerate_processes(controller, path, &f);
 253                 if (r < 0) {
 254                         if (ret >= 0 && r != -ENOENT)
 255                                 return r;
 256
 257                         return ret;
 258                 }
 259
 260                 while ((r = cg_read_pid(f, &pid)) > 0) {
 261
 262                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 263                                 continue;
 264
 265                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 266                                 continue;
 267
 268                         if (log_kill)
 269                                 log_kill(pid, sig, userdata);
 270
 271                         /* If we haven't killed this process yet, kill
 272                          * it */
 273                         if (kill(pid, sig) < 0) {
 274                                 if (ret >= 0 && errno != ESRCH)
 275                                         ret = -errno;
 276                         } else {
 277                                 if (flags & CGROUP_SIGCONT)
 278                                         (void) kill(pid, SIGCONT);
 279
 280                                 if (ret == 0)
 281                                         ret = 1;
 282                         }
 283
 284                         done = false;
 285
 286                         r = set_put(s, PID_TO_PTR(pid));
 287                         if (r < 0) {
 288                                 if (ret >= 0)
 289                                         return r;
 290
 291                                 return ret;
 292                         }
 293                 }
 294
 295                 if (r < 0) {
 296                         if (ret >= 0)
 297                                 return r;
 298
 299                         return ret;
 300                 }
 301
 302                 /* To avoid racing against processes which fork
 303                  * quicker than we can kill them we repeat this until
 304                  * no new pids need to be killed. */
 305
 306         } while (!done);
 307
 308         return ret;
 309 }
 310
 311 int cg_kill_recursive(
 312                 const char *controller,
 313                 const char *path,
 314                 int sig,
 315                 CGroupFlags flags,
 316                 Set *s,
 317                 cg_kill_log_func_t log_kill,
 318                 void *userdata) {
 319
 320         _cleanup_set_free_ Set *allocated_set = NULL;
 321         _cleanup_closedir_ DIR *d = NULL;
 322         int r, ret;
 323         char *fn;
 324
 325         assert(path);
 326         assert(sig >= 0);
 327
 328         if (!s) {
 329                 s = allocated_set = set_new(NULL);
 330                 if (!s)
 331                         return -ENOMEM;
 332         }
 333
 334         ret = cg_kill(controller, path, sig, flags, s, log_kill, userdata);
 335
 336         r = cg_enumerate_subgroups(controller, path, &d);
 337         if (r < 0) {
 338                 if (ret >= 0 && r != -ENOENT)
 339                         return r;
 340
 341                 return ret;
 342         }
 343
 344         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 345                 _cleanup_free_ char *p = NULL;
 346
 347                 p = strjoin(path, "/", fn, NULL);
 348                 free(fn);
 349                 if (!p)
 350                         return -ENOMEM;
 351
 352                 r = cg_kill_recursive(controller, p, sig, flags, s, log_kill, userdata);
 353                 if (r != 0 && ret >= 0)
 354                         ret = r;
 355         }
 356         if (ret >= 0 && r < 0)
 357                 ret = r;
 358
 359         if (flags & CGROUP_REMOVE) {
 360                 r = cg_rmdir(controller, path);
 361                 if (r < 0 && ret >= 0 && r != -ENOENT && r != -EBUSY)
 362                         return r;
 363         }
 364
 365         return ret;
 366 }
 367
 368 int cg_migrate(
 369                 const char *cfrom,
 370                 const char *pfrom,
 371                 const char *cto,
 372                 const char *pto,
 373                 CGroupFlags flags) {
 374
 375         bool done = false;
 376         _cleanup_set_free_ Set *s = NULL;
 377         int r, ret = 0;
 378         pid_t my_pid;
 379
 380         assert(cfrom);
 381         assert(pfrom);
 382         assert(cto);
 383         assert(pto);
 384
 385         s = set_new(NULL);
 386         if (!s)
 387                 return -ENOMEM;
 388
 389         my_pid = getpid();
 390
 391         do {
 392                 _cleanup_fclose_ FILE *f = NULL;
 393                 pid_t pid = 0;
 394                 done = true;
 395
 396                 r = cg_enumerate_processes(cfrom, pfrom, &f);
 397                 if (r < 0) {
 398                         if (ret >= 0 && r != -ENOENT)
 399                                 return r;
 400
 401                         return ret;
 402                 }
 403
 404                 while ((r = cg_read_pid(f, &pid)) > 0) {
 405
 406                         /* This might do weird stuff if we aren't a
 407                          * single-threaded program. However, we
 408                          * luckily know we are not */
 409                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 410                                 continue;
 411
 412                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 413                                 continue;
 414
 415                         /* Ignore kernel threads. Since they can only
 416                          * exist in the root cgroup, we only check for
 417                          * them there. */
 418                         if (cfrom &&
 419                             (isempty(pfrom) || path_equal(pfrom, "/")) &&
 420                             is_kernel_thread(pid) > 0)
 421                                 continue;
 422
 423                         r = cg_attach(cto, pto, pid);
 424                         if (r < 0) {
 425                                 if (ret >= 0 && r != -ESRCH)
 426                                         ret = r;
 427                         } else if (ret == 0)
 428                                 ret = 1;
 429
 430                         done = false;
 431
 432                         r = set_put(s, PID_TO_PTR(pid));
 433                         if (r < 0) {
 434                                 if (ret >= 0)
 435                                         return r;
 436
 437                                 return ret;
 438                         }
 439                 }
 440
 441                 if (r < 0) {
 442                         if (ret >= 0)
 443                                 return r;
 444
 445                         return ret;
 446                 }
 447         } while (!done);
 448
 449         return ret;
 450 }
 451
 452 int cg_migrate_recursive(
 453                 const char *cfrom,
 454                 const char *pfrom,
 455                 const char *cto,
 456                 const char *pto,
 457                 CGroupFlags flags) {
 458
 459         _cleanup_closedir_ DIR *d = NULL;
 460         int r, ret = 0;
 461         char *fn;
 462
 463         assert(cfrom);
 464         assert(pfrom);
 465         assert(cto);
 466         assert(pto);
 467
 468         ret = cg_migrate(cfrom, pfrom, cto, pto, flags);
 469
 470         r = cg_enumerate_subgroups(cfrom, pfrom, &d);
 471         if (r < 0) {
 472                 if (ret >= 0 && r != -ENOENT)
 473                         return r;
 474
 475                 return ret;
 476         }
 477
 478         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 479                 _cleanup_free_ char *p = NULL;
 480
 481                 p = strjoin(pfrom, "/", fn, NULL);
 482                 free(fn);
 483                 if (!p)
 484                         return -ENOMEM;
 485
 486                 r = cg_migrate_recursive(cfrom, p, cto, pto, flags);
 487                 if (r != 0 && ret >= 0)
 488                         ret = r;
 489         }
 490
 491         if (r < 0 && ret >= 0)
 492                 ret = r;
 493
 494         if (flags & CGROUP_REMOVE) {
 495                 r = cg_rmdir(cfrom, pfrom);
 496                 if (r < 0 && ret >= 0 && r != -ENOENT && r != -EBUSY)
 497                         return r;
 498         }
 499
 500         return ret;
 501 }
 502
 503 int cg_migrate_recursive_fallback(
 504                 const char *cfrom,
 505                 const char *pfrom,
 506                 const char *cto,
 507                 const char *pto,
 508                 CGroupFlags flags) {
 509
 510         int r;
 511
 512         assert(cfrom);
 513         assert(pfrom);
 514         assert(cto);
 515         assert(pto);
 516
 517         r = cg_migrate_recursive(cfrom, pfrom, cto, pto, flags);
 518         if (r < 0) {
 519                 char prefix[strlen(pto) + 1];
 520
 521                 /* This didn't work? Then let's try all prefixes of the destination */
 522
 523                 PATH_FOREACH_PREFIX(prefix, pto) {
 524                         int q;
 525
 526                         q = cg_migrate_recursive(cfrom, pfrom, cto, prefix, flags);
 527                         if (q >= 0)
 528                                 return q;
 529                 }
 530         }
 531
 532         return r;
 533 }
 534
 535 static const char *controller_to_dirname(const char *controller) {
 536         const char *e;
 537
 538         assert(controller);
 539
 540         /* Converts a controller name to the directory name below
 541          * /sys/fs/cgroup/ we want to mount it to. Effectively, this
 542          * just cuts off the name= prefixed used for named
 543          * hierarchies, if it is specified. */
 544
 545         e = startswith(controller, "name=");
 546         if (e)
 547                 return e;
 548
 549         return controller;
 550 }
 551
 552 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **fs) {
 553         const char *dn;
 554         char *t = NULL;
 555
 556         assert(fs);
 557         assert(controller);
 558
 559         dn = controller_to_dirname(controller);
 560
 561         if (isempty(path) && isempty(suffix))
 562                 t = strappend("/sys/fs/cgroup/", dn);
 563         else if (isempty(path))
 564                 t = strjoin("/sys/fs/cgroup/", dn, "/", suffix, NULL);
 565         else if (isempty(suffix))
 566                 t = strjoin("/sys/fs/cgroup/", dn, "/", path, NULL);
 567         else
 568                 t = strjoin("/sys/fs/cgroup/", dn, "/", path, "/", suffix, NULL);
 569         if (!t)
 570                 return -ENOMEM;
 571
 572         *fs = t;
 573         return 0;
 574 }
 575
 576 static int join_path_unified(const char *path, const char *suffix, char **fs) {
 577         char *t;
 578
 579         assert(fs);
 580
 581         if (isempty(path) && isempty(suffix))
 582                 t = strdup("/sys/fs/cgroup");
 583         else if (isempty(path))
 584                 t = strappend("/sys/fs/cgroup/", suffix);
 585         else if (isempty(suffix))
 586                 t = strappend("/sys/fs/cgroup/", path);
 587         else
 588                 t = strjoin("/sys/fs/cgroup/", path, "/", suffix, NULL);
 589         if (!t)
 590                 return -ENOMEM;
 591
 592         *fs = t;
 593         return 0;
 594 }
 595
 596 int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) {
 597         int unified, r;
 598
 599         assert(fs);
 600
 601         if (!controller) {
 602                 char *t;
 603
 604                 /* If no controller is specified, we return the path
 605                  * *below* the controllers, without any prefix. */
 606
 607                 if (!path && !suffix)
 608                         return -EINVAL;
 609
 610                 if (!suffix)
 611                         t = strdup(path);
 612                 else if (!path)
 613                         t = strdup(suffix);
 614                 else
 615                         t = strjoin(path, "/", suffix, NULL);
 616                 if (!t)
 617                         return -ENOMEM;
 618
 619                 *fs = path_kill_slashes(t);
 620                 return 0;
 621         }
 622
 623         if (!cg_controller_is_valid(controller))
 624                 return -EINVAL;
 625
 626         unified = cg_all_unified();
 627         if (unified < 0)
 628                 return unified;
 629
 630         if (unified > 0)
 631                 r = join_path_unified(path, suffix, fs);
 632         else
 633                 r = join_path_legacy(controller, path, suffix, fs);
 634         if (r < 0)
 635                 return r;
 636
 637         path_kill_slashes(*fs);
 638         return 0;
 639 }
 640
 641 static int controller_is_accessible(const char *controller) {
 642         int unified;
 643
 644         assert(controller);
 645
 646         /* Checks whether a specific controller is accessible,
 647          * i.e. its hierarchy mounted. In the unified hierarchy all
 648          * controllers are considered accessible, except for the named
 649          * hierarchies */
 650
 651         if (!cg_controller_is_valid(controller))
 652                 return -EINVAL;
 653
 654         unified = cg_all_unified();
 655         if (unified < 0)
 656                 return unified;
 657         if (unified > 0) {
 658                 /* We don't support named hierarchies if we are using
 659                  * the unified hierarchy. */
 660
 661                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 662                         return 0;
 663
 664                 if (startswith(controller, "name="))
 665                         return -EOPNOTSUPP;
 666
 667         } else {
 668                 const char *cc, *dn;
 669
 670                 dn = controller_to_dirname(controller);
 671                 cc = strjoina("/sys/fs/cgroup/", dn);
 672
 673                 if (laccess(cc, F_OK) < 0)
 674                         return -errno;
 675         }
 676
 677         return 0;
 678 }
 679
 680 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 681         int r;
 682
 683         assert(controller);
 684         assert(fs);
 685
 686         /* Check if the specified controller is actually accessible */
 687         r = controller_is_accessible(controller);
 688         if (r < 0)
 689                 return r;
 690
 691         return cg_get_path(controller, path, suffix, fs);
 692 }
 693
 694 static int trim_cb(const char *path, const struct stat *sb, int typeflag, struct FTW *ftwbuf) {
 695         assert(path);
 696         assert(sb);
 697         assert(ftwbuf);
 698
 699         if (typeflag != FTW_DP)
 700                 return 0;
 701
 702         if (ftwbuf->level < 1)
 703                 return 0;
 704
 705         (void) rmdir(path);
 706         return 0;
 707 }
 708
 709 int cg_trim(const char *controller, const char *path, bool delete_root) {
 710         _cleanup_free_ char *fs = NULL;
 711         int r = 0;
 712
 713         assert(path);
 714
 715         r = cg_get_path(controller, path, NULL, &fs);
 716         if (r < 0)
 717                 return r;
 718
 719         errno = 0;
 720         if (nftw(fs, trim_cb, 64, FTW_DEPTH|FTW_MOUNT|FTW_PHYS) != 0) {
 721                 if (errno == ENOENT)
 722                         r = 0;
 723                 else if (errno > 0)
 724                         r = -errno;
 725                 else
 726                         r = -EIO;
 727         }
 728
 729         if (delete_root) {
 730                 if (rmdir(fs) < 0 && errno != ENOENT)
 731                         return -errno;
 732         }
 733
 734         return r;
 735 }
 736
 737 int cg_create(const char *controller, const char *path) {
 738         _cleanup_free_ char *fs = NULL;
 739         int r;
 740
 741         r = cg_get_path_and_check(controller, path, NULL, &fs);
 742         if (r < 0)
 743                 return r;
 744
 745         r = mkdir_parents(fs, 0755);
 746         if (r < 0)
 747                 return r;
 748
 749         if (mkdir(fs, 0755) < 0) {
 750
 751                 if (errno == EEXIST)
 752                         return 0;
 753
 754                 return -errno;
 755         }
 756
 757         return 1;
 758 }
 759
 760 int cg_create_and_attach(const char *controller, const char *path, pid_t pid) {
 761         int r, q;
 762
 763         assert(pid >= 0);
 764
 765         r = cg_create(controller, path);
 766         if (r < 0)
 767                 return r;
 768
 769         q = cg_attach(controller, path, pid);
 770         if (q < 0)
 771                 return q;
 772
 773         /* This does not remove the cgroup on failure */
 774         return r;
 775 }
 776
 777 int cg_attach(const char *controller, const char *path, pid_t pid) {
 778         _cleanup_free_ char *fs = NULL;
 779         char c[DECIMAL_STR_MAX(pid_t) + 2];
 780         int r;
 781
 782         assert(path);
 783         assert(pid >= 0);
 784
 785         r = cg_get_path_and_check(controller, path, "cgroup.procs", &fs);
 786         if (r < 0)
 787                 return r;
 788
 789         if (pid == 0)
 790                 pid = getpid();
 791
 792         xsprintf(c, PID_FMT "\n", pid);
 793
 794         return write_string_file(fs, c, 0);
 795 }
 796
 797 int cg_attach_fallback(const char *controller, const char *path, pid_t pid) {
 798         int r;
 799
 800         assert(controller);
 801         assert(path);
 802         assert(pid >= 0);
 803
 804         r = cg_attach(controller, path, pid);
 805         if (r < 0) {
 806                 char prefix[strlen(path) + 1];
 807
 808                 /* This didn't work? Then let's try all prefixes of
 809                  * the destination */
 810
 811                 PATH_FOREACH_PREFIX(prefix, path) {
 812                         int q;
 813
 814                         q = cg_attach(controller, prefix, pid);
 815                         if (q >= 0)
 816                                 return q;
 817                 }
 818         }
 819
 820         return r;
 821 }
 822
 823 int cg_set_group_access(
 824                 const char *controller,
 825                 const char *path,
 826                 mode_t mode,
 827                 uid_t uid,
 828                 gid_t gid) {
 829
 830         _cleanup_free_ char *fs = NULL;
 831         int r;
 832
 833         if (mode == MODE_INVALID && uid == UID_INVALID && gid == GID_INVALID)
 834                 return 0;
 835
 836         if (mode != MODE_INVALID)
 837                 mode &= 0777;
 838
 839         r = cg_get_path(controller, path, NULL, &fs);
 840         if (r < 0)
 841                 return r;
 842
 843         return chmod_and_chown(fs, mode, uid, gid);
 844 }
 845
 846 int cg_set_task_access(
 847                 const char *controller,
 848                 const char *path,
 849                 mode_t mode,
 850                 uid_t uid,
 851                 gid_t gid) {
 852
 853         _cleanup_free_ char *fs = NULL, *procs = NULL;
 854         int r, unified;
 855
 856         assert(path);
 857
 858         if (mode == MODE_INVALID && uid == UID_INVALID && gid == GID_INVALID)
 859                 return 0;
 860
 861         if (mode != MODE_INVALID)
 862                 mode &= 0666;
 863
 864         r = cg_get_path(controller, path, "cgroup.procs", &fs);
 865         if (r < 0)
 866                 return r;
 867
 868         r = chmod_and_chown(fs, mode, uid, gid);
 869         if (r < 0)
 870                 return r;
 871
 872         unified = cg_unified(controller);
 873         if (unified < 0)
 874                 return unified;
 875         if (unified)
 876                 return 0;
 877
 878         /* Compatibility, Always keep values for "tasks" in sync with
 879          * "cgroup.procs" */
 880         if (cg_get_path(controller, path, "tasks", &procs) >= 0)
 881                 (void) chmod_and_chown(procs, mode, uid, gid);
 882
 883         return 0;
 884 }
 885
 886 int cg_pid_get_path(const char *controller, pid_t pid, char **path) {
 887         _cleanup_fclose_ FILE *f = NULL;
 888         char line[LINE_MAX];
 889         const char *fs;
 890         size_t cs = 0;
 891         int unified;
 892
 893         assert(path);
 894         assert(pid >= 0);
 895
 896         if (controller) {
 897                 if (!cg_controller_is_valid(controller))
 898                         return -EINVAL;
 899         } else
 900                 controller = SYSTEMD_CGROUP_CONTROLLER;
 901
 902         unified = cg_unified(controller);
 903         if (unified < 0)
 904                 return unified;
 905         if (unified == 0)
 906                 cs = strlen(controller);
 907
 908         fs = procfs_file_alloca(pid, "cgroup");
 909         f = fopen(fs, "re");
 910         if (!f)
 911                 return errno == ENOENT ? -ESRCH : -errno;
 912
 913         FOREACH_LINE(line, f, return -errno) {
 914                 char *e, *p;
 915
 916                 truncate_nl(line);
 917
 918                 if (unified) {
 919                         e = startswith(line, "0:");
 920                         if (!e)
 921                                 continue;
 922
 923                         e = strchr(e, ':');
 924                         if (!e)
 925                                 continue;
 926                 } else {
 927                         char *l;
 928                         size_t k;
 929                         const char *word, *state;
 930                         bool found = false;
 931
 932                         l = strchr(line, ':');
 933                         if (!l)
 934                                 continue;
 935
 936                         l++;
 937                         e = strchr(l, ':');
 938                         if (!e)
 939                                 continue;
 940
 941                         *e = 0;
 942                         FOREACH_WORD_SEPARATOR(word, k, l, ",", state) {
 943                                 if (k == cs && memcmp(word, controller, cs) == 0) {
 944                                         found = true;
 945                                         break;
 946                                 }
 947                         }
 948
 949                         if (!found)
 950                                 continue;
 951                 }
 952
 953                 p = strdup(e + 1);
 954                 if (!p)
 955                         return -ENOMEM;
 956
 957                 *path = p;
 958                 return 0;
 959         }
 960
 961         return -ENODATA;
 962 }
 963
 964 int cg_install_release_agent(const char *controller, const char *agent) {
 965         _cleanup_free_ char *fs = NULL, *contents = NULL;
 966         const char *sc;
 967         int r, unified;
 968
 969         assert(agent);
 970
 971         unified = cg_unified(controller);
 972         if (unified < 0)
 973                 return unified;
 974         if (unified) /* doesn't apply to unified hierarchy */
 975                 return -EOPNOTSUPP;
 976
 977         r = cg_get_path(controller, NULL, "release_agent", &fs);
 978         if (r < 0)
 979                 return r;
 980
 981         r = read_one_line_file(fs, &contents);
 982         if (r < 0)
 983                 return r;
 984
 985         sc = strstrip(contents);
 986         if (isempty(sc)) {
 987                 r = write_string_file(fs, agent, 0);
 988                 if (r < 0)
 989                         return r;
 990         } else if (!path_equal(sc, agent))
 991                 return -EEXIST;
 992
 993         fs = mfree(fs);
 994         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 995         if (r < 0)
 996                 return r;
 997
 998         contents = mfree(contents);
 999         r = read_one_line_file(fs, &contents);
1000         if (r < 0)
1001                 return r;
1002
1003         sc = strstrip(contents);
1004         if (streq(sc, "0")) {
1005                 r = write_string_file(fs, "1", 0);
1006                 if (r < 0)
1007                         return r;
1008
1009                 return 1;
1010         }
1011
1012         if (!streq(sc, "1"))
1013                 return -EIO;
1014
1015         return 0;
1016 }
1017
1018 int cg_uninstall_release_agent(const char *controller) {
1019         _cleanup_free_ char *fs = NULL;
1020         int r, unified;
1021
1022         unified = cg_unified(controller);
1023         if (unified < 0)
1024                 return unified;
1025         if (unified) /* Doesn't apply to unified hierarchy */
1026                 return -EOPNOTSUPP;
1027
1028         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
1029         if (r < 0)
1030                 return r;
1031
1032         r = write_string_file(fs, "0", 0);
1033         if (r < 0)
1034                 return r;
1035
1036         fs = mfree(fs);
1037
1038         r = cg_get_path(controller, NULL, "release_agent", &fs);
1039         if (r < 0)
1040                 return r;
1041
1042         r = write_string_file(fs, "", 0);
1043         if (r < 0)
1044                 return r;
1045
1046         return 0;
1047 }
1048
1049 int cg_is_empty(const char *controller, const char *path) {
1050         _cleanup_fclose_ FILE *f = NULL;
1051         pid_t pid;
1052         int r;
1053
1054         assert(path);
1055
1056         r = cg_enumerate_processes(controller, path, &f);
1057         if (r == -ENOENT)
1058                 return 1;
1059         if (r < 0)
1060                 return r;
1061
1062         r = cg_read_pid(f, &pid);
1063         if (r < 0)
1064                 return r;
1065
1066         return r == 0;
1067 }
1068
1069 int cg_is_empty_recursive(const char *controller, const char *path) {
1070         int unified, r;
1071
1072         assert(path);
1073
1074         /* The root cgroup is always populated */
1075         if (controller && (isempty(path) || path_equal(path, "/")))
1076                 return false;
1077
1078         unified = cg_unified(controller);
1079         if (unified < 0)
1080                 return unified;
1081
1082         if (unified > 0) {
1083                 _cleanup_free_ char *t = NULL;
1084
1085                 /* On the unified hierarchy we can check empty state
1086                  * via the "populated" attribute of "cgroup.events". */
1087
1088                 r = cg_read_event(controller, path, "populated", &t);
1089                 if (r < 0)
1090                         return r;
1091
1092                 return streq(t, "0");
1093         } else {
1094                 _cleanup_closedir_ DIR *d = NULL;
1095                 char *fn;
1096
1097                 r = cg_is_empty(controller, path);
1098                 if (r <= 0)
1099                         return r;
1100
1101                 r = cg_enumerate_subgroups(controller, path, &d);
1102                 if (r == -ENOENT)
1103                         return 1;
1104                 if (r < 0)
1105                         return r;
1106
1107                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1108                         _cleanup_free_ char *p = NULL;
1109
1110                         p = strjoin(path, "/", fn, NULL);
1111                         free(fn);
1112                         if (!p)
1113                                 return -ENOMEM;
1114
1115                         r = cg_is_empty_recursive(controller, p);
1116                         if (r <= 0)
1117                                 return r;
1118                 }
1119                 if (r < 0)
1120                         return r;
1121
1122                 return true;
1123         }
1124 }
1125
1126 int cg_split_spec(const char *spec, char **controller, char **path) {
1127         char *t = NULL, *u = NULL;
1128         const char *e;
1129
1130         assert(spec);
1131
1132         if (*spec == '/') {
1133                 if (!path_is_safe(spec))
1134                         return -EINVAL;
1135
1136                 if (path) {
1137                         t = strdup(spec);
1138                         if (!t)
1139                                 return -ENOMEM;
1140
1141                         *path = path_kill_slashes(t);
1142                 }
1143
1144                 if (controller)
1145                         *controller = NULL;
1146
1147                 return 0;
1148         }
1149
1150         e = strchr(spec, ':');
1151         if (!e) {
1152                 if (!cg_controller_is_valid(spec))
1153                         return -EINVAL;
1154
1155                 if (controller) {
1156                         t = strdup(spec);
1157                         if (!t)
1158                                 return -ENOMEM;
1159
1160                         *controller = t;
1161                 }
1162
1163                 if (path)
1164                         *path = NULL;
1165
1166                 return 0;
1167         }
1168
1169         t = strndup(spec, e-spec);
1170         if (!t)
1171                 return -ENOMEM;
1172         if (!cg_controller_is_valid(t)) {
1173                 free(t);
1174                 return -EINVAL;
1175         }
1176
1177         if (isempty(e+1))
1178                 u = NULL;
1179         else {
1180                 u = strdup(e+1);
1181                 if (!u) {
1182                         free(t);
1183                         return -ENOMEM;
1184                 }
1185
1186                 if (!path_is_safe(u) ||
1187                     !path_is_absolute(u)) {
1188                         free(t);
1189                         free(u);
1190                         return -EINVAL;
1191                 }
1192
1193                 path_kill_slashes(u);
1194         }
1195
1196         if (controller)
1197                 *controller = t;
1198         else
1199                 free(t);
1200
1201         if (path)
1202                 *path = u;
1203         else
1204                 free(u);
1205
1206         return 0;
1207 }
1208
1209 int cg_mangle_path(const char *path, char **result) {
1210         _cleanup_free_ char *c = NULL, *p = NULL;
1211         char *t;
1212         int r;
1213
1214         assert(path);
1215         assert(result);
1216
1217         /* First, check if it already is a filesystem path */
1218         if (path_startswith(path, "/sys/fs/cgroup")) {
1219
1220                 t = strdup(path);
1221                 if (!t)
1222                         return -ENOMEM;
1223
1224                 *result = path_kill_slashes(t);
1225                 return 0;
1226         }
1227
1228         /* Otherwise, treat it as cg spec */
1229         r = cg_split_spec(path, &c, &p);
1230         if (r < 0)
1231                 return r;
1232
1233         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
1234 }
1235
1236 int cg_get_root_path(char **path) {
1237         char *p, *e;
1238         int r;
1239
1240         assert(path);
1241
1242         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1243         if (r < 0)
1244                 return r;
1245
1246         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1247         if (!e)
1248                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1249         if (!e)
1250                 e = endswith(p, "/system"); /* even more legacy */
1251         if (e)
1252                 *e = 0;
1253
1254         *path = p;
1255         return 0;
1256 }
1257
1258 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1259         _cleanup_free_ char *rt = NULL;
1260         char *p;
1261         int r;
1262
1263         assert(cgroup);
1264         assert(shifted);
1265
1266         if (!root) {
1267                 /* If the root was specified let's use that, otherwise
1268                  * let's determine it from PID 1 */
1269
1270                 r = cg_get_root_path(&rt);
1271                 if (r < 0)
1272                         return r;
1273
1274                 root = rt;
1275         }
1276
1277         p = path_startswith(cgroup, root);
1278         if (p && p > cgroup)
1279                 *shifted = p - 1;
1280         else
1281                 *shifted = cgroup;
1282
1283         return 0;
1284 }
1285
1286 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1287         _cleanup_free_ char *raw = NULL;
1288         const char *c;
1289         int r;
1290
1291         assert(pid >= 0);
1292         assert(cgroup);
1293
1294         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1295         if (r < 0)
1296                 return r;
1297
1298         r = cg_shift_path(raw, root, &c);
1299         if (r < 0)
1300                 return r;
1301
1302         if (c == raw) {
1303                 *cgroup = raw;
1304                 raw = NULL;
1305         } else {
1306                 char *n;
1307
1308                 n = strdup(c);
1309                 if (!n)
1310                         return -ENOMEM;
1311
1312                 *cgroup = n;
1313         }
1314
1315         return 0;
1316 }
1317
1318 int cg_path_decode_unit(const char *cgroup, char **unit) {
1319         char *c, *s;
1320         size_t n;
1321
1322         assert(cgroup);
1323         assert(unit);
1324
1325         n = strcspn(cgroup, "/");
1326         if (n < 3)
1327                 return -ENXIO;
1328
1329         c = strndupa(cgroup, n);
1330         c = cg_unescape(c);
1331
1332         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1333                 return -ENXIO;
1334
1335         s = strdup(c);
1336         if (!s)
1337                 return -ENOMEM;
1338
1339         *unit = s;
1340         return 0;
1341 }
1342
1343 static bool valid_slice_name(const char *p, size_t n) {
1344
1345         if (!p)
1346                 return false;
1347
1348         if (n < strlen("x.slice"))
1349                 return false;
1350
1351         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1352                 char buf[n+1], *c;
1353
1354                 memcpy(buf, p, n);
1355                 buf[n] = 0;
1356
1357                 c = cg_unescape(buf);
1358
1359                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1360         }
1361
1362         return false;
1363 }
1364
1365 static const char *skip_slices(const char *p) {
1366         assert(p);
1367
1368         /* Skips over all slice assignments */
1369
1370         for (;;) {
1371                 size_t n;
1372
1373                 p += strspn(p, "/");
1374
1375                 n = strcspn(p, "/");
1376                 if (!valid_slice_name(p, n))
1377                         return p;
1378
1379                 p += n;
1380         }
1381 }
1382
1383 int cg_path_get_unit(const char *path, char **ret) {
1384         const char *e;
1385         char *unit;
1386         int r;
1387
1388         assert(path);
1389         assert(ret);
1390
1391         e = skip_slices(path);
1392
1393         r = cg_path_decode_unit(e, &unit);
1394         if (r < 0)
1395                 return r;
1396
1397         /* We skipped over the slices, don't accept any now */
1398         if (endswith(unit, ".slice")) {
1399                 free(unit);
1400                 return -ENXIO;
1401         }
1402
1403         *ret = unit;
1404         return 0;
1405 }
1406
1407 int cg_pid_get_unit(pid_t pid, char **unit) {
1408         _cleanup_free_ char *cgroup = NULL;
1409         int r;
1410
1411         assert(unit);
1412
1413         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1414         if (r < 0)
1415                 return r;
1416
1417         return cg_path_get_unit(cgroup, unit);
1418 }
1419
1420 /**
1421  * Skip session-*.scope, but require it to be there.
1422  */
1423 static const char *skip_session(const char *p) {
1424         size_t n;
1425
1426         if (isempty(p))
1427                 return NULL;
1428
1429         p += strspn(p, "/");
1430
1431         n = strcspn(p, "/");
1432         if (n < strlen("session-x.scope"))
1433                 return NULL;
1434
1435         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1436                 char buf[n - 8 - 6 + 1];
1437
1438                 memcpy(buf, p + 8, n - 8 - 6);
1439                 buf[n - 8 - 6] = 0;
1440
1441                 /* Note that session scopes never need unescaping,
1442                  * since they cannot conflict with the kernel's own
1443                  * names, hence we don't need to call cg_unescape()
1444                  * here. */
1445
1446                 if (!session_id_valid(buf))
1447                         return false;
1448
1449                 p += n;
1450                 p += strspn(p, "/");
1451                 return p;
1452         }
1453
1454         return NULL;
1455 }
1456
1457 /**
1458  * Skip user@*.service, but require it to be there.
1459  */
1460 static const char *skip_user_manager(const char *p) {
1461         size_t n;
1462
1463         if (isempty(p))
1464                 return NULL;
1465
1466         p += strspn(p, "/");
1467
1468         n = strcspn(p, "/");
1469         if (n < strlen("user@x.service"))
1470                 return NULL;
1471
1472         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1473                 char buf[n - 5 - 8 + 1];
1474
1475                 memcpy(buf, p + 5, n - 5 - 8);
1476                 buf[n - 5 - 8] = 0;
1477
1478                 /* Note that user manager services never need unescaping,
1479                  * since they cannot conflict with the kernel's own
1480                  * names, hence we don't need to call cg_unescape()
1481                  * here. */
1482
1483                 if (parse_uid(buf, NULL) < 0)
1484                         return NULL;
1485
1486                 p += n;
1487                 p += strspn(p, "/");
1488
1489                 return p;
1490         }
1491
1492         return NULL;
1493 }
1494
1495 static const char *skip_user_prefix(const char *path) {
1496         const char *e, *t;
1497
1498         assert(path);
1499
1500         /* Skip slices, if there are any */
1501         e = skip_slices(path);
1502
1503         /* Skip the user manager, if it's in the path now... */
1504         t = skip_user_manager(e);
1505         if (t)
1506                 return t;
1507
1508         /* Alternatively skip the user session if it is in the path... */
1509         return skip_session(e);
1510 }
1511
1512 int cg_path_get_user_unit(const char *path, char **ret) {
1513         const char *t;
1514
1515         assert(path);
1516         assert(ret);
1517
1518         t = skip_user_prefix(path);
1519         if (!t)
1520                 return -ENXIO;
1521
1522         /* And from here on it looks pretty much the same as for a
1523          * system unit, hence let's use the same parser from here
1524          * on. */
1525         return cg_path_get_unit(t, ret);
1526 }
1527
1528 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1529         _cleanup_free_ char *cgroup = NULL;
1530         int r;
1531
1532         assert(unit);
1533
1534         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1535         if (r < 0)
1536                 return r;
1537
1538         return cg_path_get_user_unit(cgroup, unit);
1539 }
1540
1541 int cg_path_get_machine_name(const char *path, char **machine) {
1542         _cleanup_free_ char *u = NULL;
1543         const char *sl;
1544         int r;
1545
1546         r = cg_path_get_unit(path, &u);
1547         if (r < 0)
1548                 return r;
1549
1550         sl = strjoina("/run/systemd/machines/unit:", u);
1551         return readlink_malloc(sl, machine);
1552 }
1553
1554 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1555         _cleanup_free_ char *cgroup = NULL;
1556         int r;
1557
1558         assert(machine);
1559
1560         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1561         if (r < 0)
1562                 return r;
1563
1564         return cg_path_get_machine_name(cgroup, machine);
1565 }
1566
1567 int cg_path_get_session(const char *path, char **session) {
1568         _cleanup_free_ char *unit = NULL;
1569         char *start, *end;
1570         int r;
1571
1572         assert(path);
1573
1574         r = cg_path_get_unit(path, &unit);
1575         if (r < 0)
1576                 return r;
1577
1578         start = startswith(unit, "session-");
1579         if (!start)
1580                 return -ENXIO;
1581         end = endswith(start, ".scope");
1582         if (!end)
1583                 return -ENXIO;
1584
1585         *end = 0;
1586         if (!session_id_valid(start))
1587                 return -ENXIO;
1588
1589         if (session) {
1590                 char *rr;
1591
1592                 rr = strdup(start);
1593                 if (!rr)
1594                         return -ENOMEM;
1595
1596                 *session = rr;
1597         }
1598
1599         return 0;
1600 }
1601
1602 int cg_pid_get_session(pid_t pid, char **session) {
1603         _cleanup_free_ char *cgroup = NULL;
1604         int r;
1605
1606         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1607         if (r < 0)
1608                 return r;
1609
1610         return cg_path_get_session(cgroup, session);
1611 }
1612
1613 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1614         _cleanup_free_ char *slice = NULL;
1615         char *start, *end;
1616         int r;
1617
1618         assert(path);
1619
1620         r = cg_path_get_slice(path, &slice);
1621         if (r < 0)
1622                 return r;
1623
1624         start = startswith(slice, "user-");
1625         if (!start)
1626                 return -ENXIO;
1627         end = endswith(start, ".slice");
1628         if (!end)
1629                 return -ENXIO;
1630
1631         *end = 0;
1632         if (parse_uid(start, uid) < 0)
1633                 return -ENXIO;
1634
1635         return 0;
1636 }
1637
1638 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1639         _cleanup_free_ char *cgroup = NULL;
1640         int r;
1641
1642         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1643         if (r < 0)
1644                 return r;
1645
1646         return cg_path_get_owner_uid(cgroup, uid);
1647 }
1648
1649 int cg_path_get_slice(const char *p, char **slice) {
1650         const char *e = NULL;
1651
1652         assert(p);
1653         assert(slice);
1654
1655         /* Finds the right-most slice unit from the beginning, but
1656          * stops before we come to the first non-slice unit. */
1657
1658         for (;;) {
1659                 size_t n;
1660
1661                 p += strspn(p, "/");
1662
1663                 n = strcspn(p, "/");
1664                 if (!valid_slice_name(p, n)) {
1665
1666                         if (!e) {
1667                                 char *s;
1668
1669                                 s = strdup("-.slice");
1670                                 if (!s)
1671                                         return -ENOMEM;
1672
1673                                 *slice = s;
1674                                 return 0;
1675                         }
1676
1677                         return cg_path_decode_unit(e, slice);
1678                 }
1679
1680                 e = p;
1681                 p += n;
1682         }
1683 }
1684
1685 int cg_pid_get_slice(pid_t pid, char **slice) {
1686         _cleanup_free_ char *cgroup = NULL;
1687         int r;
1688
1689         assert(slice);
1690
1691         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1692         if (r < 0)
1693                 return r;
1694
1695         return cg_path_get_slice(cgroup, slice);
1696 }
1697
1698 int cg_path_get_user_slice(const char *p, char **slice) {
1699         const char *t;
1700         assert(p);
1701         assert(slice);
1702
1703         t = skip_user_prefix(p);
1704         if (!t)
1705                 return -ENXIO;
1706
1707         /* And now it looks pretty much the same as for a system
1708          * slice, so let's just use the same parser from here on. */
1709         return cg_path_get_slice(t, slice);
1710 }
1711
1712 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1713         _cleanup_free_ char *cgroup = NULL;
1714         int r;
1715
1716         assert(slice);
1717
1718         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1719         if (r < 0)
1720                 return r;
1721
1722         return cg_path_get_user_slice(cgroup, slice);
1723 }
1724
1725 char *cg_escape(const char *p) {
1726         bool need_prefix = false;
1727
1728         /* This implements very minimal escaping for names to be used
1729          * as file names in the cgroup tree: any name which might
1730          * conflict with a kernel name or is prefixed with '_' is
1731          * prefixed with a '_'. That way, when reading cgroup names it
1732          * is sufficient to remove a single prefixing underscore if
1733          * there is one. */
1734
1735         /* The return value of this function (unlike cg_unescape())
1736          * needs free()! */
1737
1738         if (p[0] == 0 ||
1739             p[0] == '_' ||
1740             p[0] == '.' ||
1741             streq(p, "notify_on_release") ||
1742             streq(p, "release_agent") ||
1743             streq(p, "tasks") ||
1744             startswith(p, "cgroup."))
1745                 need_prefix = true;
1746         else {
1747                 const char *dot;
1748
1749                 dot = strrchr(p, '.');
1750                 if (dot) {
1751                         CGroupController c;
1752                         size_t l = dot - p;
1753
1754                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1755                                 const char *n;
1756
1757                                 n = cgroup_controller_to_string(c);
1758
1759                                 if (l != strlen(n))
1760                                         continue;
1761
1762                                 if (memcmp(p, n, l) != 0)
1763                                         continue;
1764
1765                                 need_prefix = true;
1766                                 break;
1767                         }
1768                 }
1769         }
1770
1771         if (need_prefix)
1772                 return strappend("_", p);
1773
1774         return strdup(p);
1775 }
1776
1777 char *cg_unescape(const char *p) {
1778         assert(p);
1779
1780         /* The return value of this function (unlike cg_escape())
1781          * doesn't need free()! */
1782
1783         if (p[0] == '_')
1784                 return (char*) p+1;
1785
1786         return (char*) p;
1787 }
1788
1789 #define CONTROLLER_VALID                        \
1790         DIGITS LETTERS                          \
1791         "_"
1792
1793 bool cg_controller_is_valid(const char *p) {
1794         const char *t, *s;
1795
1796         if (!p)
1797                 return false;
1798
1799         s = startswith(p, "name=");
1800         if (s)
1801                 p = s;
1802
1803         if (*p == 0 || *p == '_')
1804                 return false;
1805
1806         for (t = p; *t; t++)
1807                 if (!strchr(CONTROLLER_VALID, *t))
1808                         return false;
1809
1810         if (t - p > FILENAME_MAX)
1811                 return false;
1812
1813         return true;
1814 }
1815
1816 int cg_slice_to_path(const char *unit, char **ret) {
1817         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1818         const char *dash;
1819         int r;
1820
1821         assert(unit);
1822         assert(ret);
1823
1824         if (streq(unit, "-.slice")) {
1825                 char *x;
1826
1827                 x = strdup("");
1828                 if (!x)
1829                         return -ENOMEM;
1830                 *ret = x;
1831                 return 0;
1832         }
1833
1834         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1835                 return -EINVAL;
1836
1837         if (!endswith(unit, ".slice"))
1838                 return -EINVAL;
1839
1840         r = unit_name_to_prefix(unit, &p);
1841         if (r < 0)
1842                 return r;
1843
1844         dash = strchr(p, '-');
1845
1846         /* Don't allow initial dashes */
1847         if (dash == p)
1848                 return -EINVAL;
1849
1850         while (dash) {
1851                 _cleanup_free_ char *escaped = NULL;
1852                 char n[dash - p + sizeof(".slice")];
1853
1854                 /* Don't allow trailing or double dashes */
1855                 if (dash[1] == 0 || dash[1] == '-')
1856                         return -EINVAL;
1857
1858                 strcpy(stpncpy(n, p, dash - p), ".slice");
1859                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1860                         return -EINVAL;
1861
1862                 escaped = cg_escape(n);
1863                 if (!escaped)
1864                         return -ENOMEM;
1865
1866                 if (!strextend(&s, escaped, "/", NULL))
1867                         return -ENOMEM;
1868
1869                 dash = strchr(dash+1, '-');
1870         }
1871
1872         e = cg_escape(unit);
1873         if (!e)
1874                 return -ENOMEM;
1875
1876         if (!strextend(&s, e, NULL))
1877                 return -ENOMEM;
1878
1879         *ret = s;
1880         s = NULL;
1881
1882         return 0;
1883 }
1884
1885 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1886         _cleanup_free_ char *p = NULL;
1887         int r;
1888
1889         r = cg_get_path(controller, path, attribute, &p);
1890         if (r < 0)
1891                 return r;
1892
1893         return write_string_file(p, value, 0);
1894 }
1895
1896 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1897         _cleanup_free_ char *p = NULL;
1898         int r;
1899
1900         r = cg_get_path(controller, path, attribute, &p);
1901         if (r < 0)
1902                 return r;
1903
1904         return read_one_line_file(p, ret);
1905 }
1906
1907 int cg_get_keyed_attribute(const char *controller, const char *path, const char *attribute, const char **keys, char **values) {
1908         _cleanup_free_ char *filename = NULL, *content = NULL;
1909         char *line, *p;
1910         int i, r;
1911
1912         for (i = 0; keys[i]; i++)
1913                 values[i] = NULL;
1914
1915         r = cg_get_path(controller, path, attribute, &filename);
1916         if (r < 0)
1917                 return r;
1918
1919         r = read_full_file(filename, &content, NULL);
1920         if (r < 0)
1921                 return r;
1922
1923         p = content;
1924         while ((line = strsep(&p, "\n"))) {
1925                 char *key;
1926
1927                 key = strsep(&line, " ");
1928
1929                 for (i = 0; keys[i]; i++) {
1930                         if (streq(key, keys[i])) {
1931                                 values[i] = strdup(line);
1932                                 break;
1933                         }
1934                 }
1935         }
1936
1937         for (i = 0; keys[i]; i++) {
1938                 if (!values[i]) {
1939                         for (i = 0; keys[i]; i++) {
1940                                 free(values[i]);
1941                                 values[i] = NULL;
1942                         }
1943                         return -ENOENT;
1944                 }
1945         }
1946
1947         return 0;
1948 }
1949
1950 int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) {
1951         CGroupController c;
1952         int r, unified;
1953
1954         /* This one will create a cgroup in our private tree, but also
1955          * duplicate it in the trees specified in mask, and remove it
1956          * in all others */
1957
1958         /* First create the cgroup in our own hierarchy. */
1959         r = cg_create(SYSTEMD_CGROUP_CONTROLLER, path);
1960         if (r < 0)
1961                 return r;
1962
1963         /* If we are in the unified hierarchy, we are done now */
1964         unified = cg_all_unified();
1965         if (unified < 0)
1966                 return unified;
1967         if (unified > 0)
1968                 return 0;
1969
1970         /* Otherwise, do the same in the other hierarchies */
1971         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1972                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1973                 const char *n;
1974
1975                 n = cgroup_controller_to_string(c);
1976
1977                 if (mask & bit)
1978                         (void) cg_create(n, path);
1979                 else if (supported & bit)
1980                         (void) cg_trim(n, path, true);
1981         }
1982
1983         return 0;
1984 }
1985
1986 int cg_attach_everywhere(CGroupMask supported, const char *path, pid_t pid, cg_migrate_callback_t path_callback, void *userdata) {
1987         CGroupController c;
1988         int r, unified;
1989
1990         r = cg_attach(SYSTEMD_CGROUP_CONTROLLER, path, pid);
1991         if (r < 0)
1992                 return r;
1993
1994         unified = cg_all_unified();
1995         if (unified < 0)
1996                 return unified;
1997         if (unified > 0)
1998                 return 0;
1999
2000         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2001                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2002                 const char *p = NULL;
2003
2004                 if (!(supported & bit))
2005                         continue;
2006
2007                 if (path_callback)
2008                         p = path_callback(bit, userdata);
2009
2010                 if (!p)
2011                         p = path;
2012
2013                 (void) cg_attach_fallback(cgroup_controller_to_string(c), p, pid);
2014         }
2015
2016         return 0;
2017 }
2018
2019 int cg_attach_many_everywhere(CGroupMask supported, const char *path, Set* pids, cg_migrate_callback_t path_callback, void *userdata) {
2020         Iterator i;
2021         void *pidp;
2022         int r = 0;
2023
2024         SET_FOREACH(pidp, pids, i) {
2025                 pid_t pid = PTR_TO_PID(pidp);
2026                 int q;
2027
2028                 q = cg_attach_everywhere(supported, path, pid, path_callback, userdata);
2029                 if (q < 0 && r >= 0)
2030                         r = q;
2031         }
2032
2033         return r;
2034 }
2035
2036 int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to, cg_migrate_callback_t to_callback, void *userdata) {
2037         CGroupController c;
2038         int r = 0, unified;
2039
2040         if (!path_equal(from, to))  {
2041                 r = cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER, from, SYSTEMD_CGROUP_CONTROLLER, to, CGROUP_REMOVE);
2042                 if (r < 0)
2043                         return r;
2044         }
2045
2046         unified = cg_all_unified();
2047         if (unified < 0)
2048                 return unified;
2049         if (unified > 0)
2050                 return r;
2051
2052         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2053                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2054                 const char *p = NULL;
2055
2056                 if (!(supported & bit))
2057                         continue;
2058
2059                 if (to_callback)
2060                         p = to_callback(bit, userdata);
2061
2062                 if (!p)
2063                         p = to;
2064
2065                 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER, to, cgroup_controller_to_string(c), p, 0);
2066         }
2067
2068         return 0;
2069 }
2070
2071 int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) {
2072         CGroupController c;
2073         int r, unified;
2074
2075         r = cg_trim(SYSTEMD_CGROUP_CONTROLLER, path, delete_root);
2076         if (r < 0)
2077                 return r;
2078
2079         unified = cg_all_unified();
2080         if (unified < 0)
2081                 return unified;
2082         if (unified > 0)
2083                 return r;
2084
2085         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2086                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2087
2088                 if (!(supported & bit))
2089                         continue;
2090
2091                 (void) cg_trim(cgroup_controller_to_string(c), path, delete_root);
2092         }
2093
2094         return 0;
2095 }
2096
2097 int cg_mask_supported(CGroupMask *ret) {
2098         CGroupMask mask = 0;
2099         int r, unified;
2100
2101         /* Determines the mask of supported cgroup controllers. Only
2102          * includes controllers we can make sense of and that are
2103          * actually accessible. */
2104
2105         unified = cg_all_unified();
2106         if (unified < 0)
2107                 return unified;
2108         if (unified > 0) {
2109                 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
2110                 const char *c;
2111
2112                 /* In the unified hierarchy we can read the supported
2113                  * and accessible controllers from a the top-level
2114                  * cgroup attribute */
2115
2116                 r = cg_get_root_path(&root);
2117                 if (r < 0)
2118                         return r;
2119
2120                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
2121                 if (r < 0)
2122                         return r;
2123
2124                 r = read_one_line_file(path, &controllers);
2125                 if (r < 0)
2126                         return r;
2127
2128                 c = controllers;
2129                 for (;;) {
2130                         _cleanup_free_ char *n = NULL;
2131                         CGroupController v;
2132
2133                         r = extract_first_word(&c, &n, NULL, 0);
2134                         if (r < 0)
2135                                 return r;
2136                         if (r == 0)
2137                                 break;
2138
2139                         v = cgroup_controller_from_string(n);
2140                         if (v < 0)
2141                                 continue;
2142
2143                         mask |= CGROUP_CONTROLLER_TO_MASK(v);
2144                 }
2145
2146                 /* Currently, we support the cpu, memory, io and pids
2147                  * controller in the unified hierarchy, mask
2148                  * everything else off. */
2149                 mask &= CGROUP_MASK_CPU | CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
2150
2151         } else {
2152                 CGroupController c;
2153
2154                 /* In the legacy hierarchy, we check whether which
2155                  * hierarchies are mounted. */
2156
2157                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2158                         const char *n;
2159
2160                         n = cgroup_controller_to_string(c);
2161                         if (controller_is_accessible(n) >= 0)
2162                                 mask |= CGROUP_CONTROLLER_TO_MASK(c);
2163                 }
2164         }
2165
2166         *ret = mask;
2167         return 0;
2168 }
2169
2170 int cg_kernel_controllers(Set *controllers) {
2171         _cleanup_fclose_ FILE *f = NULL;
2172         char buf[LINE_MAX];
2173         int r;
2174
2175         assert(controllers);
2176
2177         /* Determines the full list of kernel-known controllers. Might
2178          * include controllers we don't actually support, arbitrary
2179          * named hierarchies and controllers that aren't currently
2180          * accessible (because not mounted). */
2181
2182         f = fopen("/proc/cgroups", "re");
2183         if (!f) {
2184                 if (errno == ENOENT)
2185                         return 0;
2186                 return -errno;
2187         }
2188
2189         /* Ignore the header line */
2190         (void) fgets(buf, sizeof(buf), f);
2191
2192         for (;;) {
2193                 char *controller;
2194                 int enabled = 0;
2195
2196                 errno = 0;
2197                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
2198
2199                         if (feof(f))
2200                                 break;
2201
2202                         if (ferror(f) && errno > 0)
2203                                 return -errno;
2204
2205                         return -EBADMSG;
2206                 }
2207
2208                 if (!enabled) {
2209                         free(controller);
2210                         continue;
2211                 }
2212
2213                 if (!cg_controller_is_valid(controller)) {
2214                         free(controller);
2215                         return -EBADMSG;
2216                 }
2217
2218                 r = set_consume(controllers, controller);
2219                 if (r < 0)
2220                         return r;
2221         }
2222
2223         return 0;
2224 }
2225
2226 static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2227
2228 static int cg_update_unified(void) {
2229
2230         struct statfs fs;
2231
2232         /* Checks if we support the unified hierarchy. Returns an
2233          * error when the cgroup hierarchies aren't mounted yet or we
2234          * have any other trouble determining if the unified hierarchy
2235          * is supported. */
2236
2237         if (unified_cache >= CGROUP_UNIFIED_NONE)
2238                 return 0;
2239
2240         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2241                 return -errno;
2242
2243         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC))
2244                 unified_cache = CGROUP_UNIFIED_ALL;
2245         else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2246                 if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0)
2247                         return -errno;
2248
2249                 unified_cache = F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC) ?
2250                         CGROUP_UNIFIED_SYSTEMD : CGROUP_UNIFIED_NONE;
2251         } else
2252                 return -ENOMEDIUM;
2253
2254         return 0;
2255 }
2256
2257 int cg_unified(const char *controller) {
2258
2259         int r;
2260
2261         r = cg_update_unified();
2262         if (r < 0)
2263                 return r;
2264
2265         if (streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER))
2266                 return unified_cache >= CGROUP_UNIFIED_SYSTEMD;
2267         else
2268                 return unified_cache >= CGROUP_UNIFIED_ALL;
2269 }
2270
2271 int cg_all_unified(void) {
2272
2273         return cg_unified(NULL);
2274 }
2275
2276 void cg_unified_flush(void) {
2277         unified_cache = CGROUP_UNIFIED_UNKNOWN;
2278 }
2279
2280 int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) {
2281         _cleanup_free_ char *fs = NULL;
2282         CGroupController c;
2283         int r, unified;
2284
2285         assert(p);
2286
2287         if (supported == 0)
2288                 return 0;
2289
2290         unified = cg_all_unified();
2291         if (unified < 0)
2292                 return unified;
2293         if (!unified) /* on the legacy hiearchy there's no joining of controllers defined */
2294                 return 0;
2295
2296         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, p, "cgroup.subtree_control", &fs);
2297         if (r < 0)
2298                 return r;
2299
2300         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2301                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2302                 const char *n;
2303
2304                 if (!(supported & bit))
2305                         continue;
2306
2307                 n = cgroup_controller_to_string(c);
2308                 {
2309                         char s[1 + strlen(n) + 1];
2310
2311                         s[0] = mask & bit ? '+' : '-';
2312                         strcpy(s + 1, n);
2313
2314                         r = write_string_file(fs, s, 0);
2315                         if (r < 0)
2316                                 log_debug_errno(r, "Failed to enable controller %s for %s (%s): %m", n, p, fs);
2317                 }
2318         }
2319
2320         return 0;
2321 }
2322
2323 bool cg_is_unified_wanted(void) {
2324         static thread_local int wanted = -1;
2325         int r, unified;
2326
2327         /* If the hierarchy is already mounted, then follow whatever
2328          * was chosen for it. */
2329         unified = cg_all_unified();
2330         if (unified >= 0)
2331                 return unified;
2332
2333         /* Otherwise, let's see what the kernel command line has to
2334          * say. Since checking that is expensive, let's cache the
2335          * result. */
2336         if (wanted >= 0)
2337                 return wanted;
2338
2339         r = get_proc_cmdline_key("systemd.unified_cgroup_hierarchy", NULL);
2340         if (r > 0)
2341                 return (wanted = true);
2342         else {
2343                 _cleanup_free_ char *value = NULL;
2344
2345                 r = get_proc_cmdline_key("systemd.unified_cgroup_hierarchy=", &value);
2346                 if (r < 0)
2347                         return false;
2348                 if (r == 0)
2349                         return (wanted = false);
2350
2351                 return (wanted = parse_boolean(value) > 0);
2352         }
2353 }
2354
2355 bool cg_is_legacy_wanted(void) {
2356         return !cg_is_unified_wanted();
2357 }
2358
2359 bool cg_is_unified_systemd_controller_wanted(void) {
2360         static thread_local int wanted = -1;
2361         int r, unified;
2362
2363         /* If the unified hierarchy is requested in full, no need to
2364          * bother with this. */
2365         if (cg_is_unified_wanted())
2366                 return 0;
2367
2368         /* If the hierarchy is already mounted, then follow whatever
2369          * was chosen for it. */
2370         unified = cg_unified(SYSTEMD_CGROUP_CONTROLLER);
2371         if (unified >= 0)
2372                 return unified;
2373
2374         /* Otherwise, let's see what the kernel command line has to
2375          * say. Since checking that is expensive, let's cache the
2376          * result. */
2377         if (wanted >= 0)
2378                 return wanted;
2379
2380         r = get_proc_cmdline_key("systemd.legacy_systemd_cgroup_controller", NULL);
2381         if (r > 0)
2382                 wanted = false;
2383         else {
2384                 _cleanup_free_ char *value = NULL;
2385
2386                 r = get_proc_cmdline_key("systemd.legacy_systemd_cgroup_controller=", &value);
2387                 if (r < 0)
2388                         return true;
2389
2390                 if (r == 0)
2391                         wanted = true;
2392                 else
2393                         wanted = parse_boolean(value) <= 0;
2394         }
2395
2396         return wanted;
2397 }
2398
2399 bool cg_is_legacy_systemd_controller_wanted(void) {
2400         return cg_is_legacy_wanted() && !cg_is_unified_systemd_controller_wanted();
2401 }
2402
2403 int cg_weight_parse(const char *s, uint64_t *ret) {
2404         uint64_t u;
2405         int r;
2406
2407         if (isempty(s)) {
2408                 *ret = CGROUP_WEIGHT_INVALID;
2409                 return 0;
2410         }
2411
2412         r = safe_atou64(s, &u);
2413         if (r < 0)
2414                 return r;
2415
2416         if (u < CGROUP_WEIGHT_MIN || u > CGROUP_WEIGHT_MAX)
2417                 return -ERANGE;
2418
2419         *ret = u;
2420         return 0;
2421 }
2422
2423 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2424         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2425         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2426         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2427         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2428 };
2429
2430 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2431         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2432         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2433         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2434         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2435 };
2436
2437 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2438
2439 int cg_cpu_shares_parse(const char *s, uint64_t *ret) {
2440         uint64_t u;
2441         int r;
2442
2443         if (isempty(s)) {
2444                 *ret = CGROUP_CPU_SHARES_INVALID;
2445                 return 0;
2446         }
2447
2448         r = safe_atou64(s, &u);
2449         if (r < 0)
2450                 return r;
2451
2452         if (u < CGROUP_CPU_SHARES_MIN || u > CGROUP_CPU_SHARES_MAX)
2453                 return -ERANGE;
2454
2455         *ret = u;
2456         return 0;
2457 }
2458
2459 int cg_blkio_weight_parse(const char *s, uint64_t *ret) {
2460         uint64_t u;
2461         int r;
2462
2463         if (isempty(s)) {
2464                 *ret = CGROUP_BLKIO_WEIGHT_INVALID;
2465                 return 0;
2466         }
2467
2468         r = safe_atou64(s, &u);
2469         if (r < 0)
2470                 return r;
2471
2472         if (u < CGROUP_BLKIO_WEIGHT_MIN || u > CGROUP_BLKIO_WEIGHT_MAX)
2473                 return -ERANGE;
2474
2475         *ret = u;
2476         return 0;
2477 }
2478
2479 static const char *cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2480         [CGROUP_CONTROLLER_CPU] = "cpu",
2481         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2482         [CGROUP_CONTROLLER_IO] = "io",
2483         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2484         [CGROUP_CONTROLLER_MEMORY] = "memory",
2485         [CGROUP_CONTROLLER_DEVICES] = "devices",
2486         [CGROUP_CONTROLLER_PIDS] = "pids",
2487 };
2488
2489 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);