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_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_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();
 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         unified = cg_unified();
 897         if (unified < 0)
 898                 return unified;
 899         if (unified == 0) {
 900                 if (controller) {
 901                         if (!cg_controller_is_valid(controller))
 902                                 return -EINVAL;
 903                 } else
 904                         controller = SYSTEMD_CGROUP_CONTROLLER;
 905
 906                 cs = strlen(controller);
 907         }
 908
 909         fs = procfs_file_alloca(pid, "cgroup");
 910         f = fopen(fs, "re");
 911         if (!f)
 912                 return errno == ENOENT ? -ESRCH : -errno;
 913
 914         FOREACH_LINE(line, f, return -errno) {
 915                 char *e, *p;
 916
 917                 truncate_nl(line);
 918
 919                 if (unified) {
 920                         e = startswith(line, "0:");
 921                         if (!e)
 922                                 continue;
 923
 924                         e = strchr(e, ':');
 925                         if (!e)
 926                                 continue;
 927                 } else {
 928                         char *l;
 929                         size_t k;
 930                         const char *word, *state;
 931                         bool found = false;
 932
 933                         l = strchr(line, ':');
 934                         if (!l)
 935                                 continue;
 936
 937                         l++;
 938                         e = strchr(l, ':');
 939                         if (!e)
 940                                 continue;
 941
 942                         *e = 0;
 943                         FOREACH_WORD_SEPARATOR(word, k, l, ",", state) {
 944                                 if (k == cs && memcmp(word, controller, cs) == 0) {
 945                                         found = true;
 946                                         break;
 947                                 }
 948                         }
 949
 950                         if (!found)
 951                                 continue;
 952                 }
 953
 954                 p = strdup(e + 1);
 955                 if (!p)
 956                         return -ENOMEM;
 957
 958                 *path = p;
 959                 return 0;
 960         }
 961
 962         return -ENODATA;
 963 }
 964
 965 int cg_install_release_agent(const char *controller, const char *agent) {
 966         _cleanup_free_ char *fs = NULL, *contents = NULL;
 967         const char *sc;
 968         int r, unified;
 969
 970         assert(agent);
 971
 972         unified = cg_unified();
 973         if (unified < 0)
 974                 return unified;
 975         if (unified) /* doesn't apply to unified hierarchy */
 976                 return -EOPNOTSUPP;
 977
 978         r = cg_get_path(controller, NULL, "release_agent", &fs);
 979         if (r < 0)
 980                 return r;
 981
 982         r = read_one_line_file(fs, &contents);
 983         if (r < 0)
 984                 return r;
 985
 986         sc = strstrip(contents);
 987         if (isempty(sc)) {
 988                 r = write_string_file(fs, agent, 0);
 989                 if (r < 0)
 990                         return r;
 991         } else if (!path_equal(sc, agent))
 992                 return -EEXIST;
 993
 994         fs = mfree(fs);
 995         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 996         if (r < 0)
 997                 return r;
 998
 999         contents = mfree(contents);
1000         r = read_one_line_file(fs, &contents);
1001         if (r < 0)
1002                 return r;
1003
1004         sc = strstrip(contents);
1005         if (streq(sc, "0")) {
1006                 r = write_string_file(fs, "1", 0);
1007                 if (r < 0)
1008                         return r;
1009
1010                 return 1;
1011         }
1012
1013         if (!streq(sc, "1"))
1014                 return -EIO;
1015
1016         return 0;
1017 }
1018
1019 int cg_uninstall_release_agent(const char *controller) {
1020         _cleanup_free_ char *fs = NULL;
1021         int r, unified;
1022
1023         unified = cg_unified();
1024         if (unified < 0)
1025                 return unified;
1026         if (unified) /* Doesn't apply to unified hierarchy */
1027                 return -EOPNOTSUPP;
1028
1029         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
1030         if (r < 0)
1031                 return r;
1032
1033         r = write_string_file(fs, "0", 0);
1034         if (r < 0)
1035                 return r;
1036
1037         fs = mfree(fs);
1038
1039         r = cg_get_path(controller, NULL, "release_agent", &fs);
1040         if (r < 0)
1041                 return r;
1042
1043         r = write_string_file(fs, "", 0);
1044         if (r < 0)
1045                 return r;
1046
1047         return 0;
1048 }
1049
1050 int cg_is_empty(const char *controller, const char *path) {
1051         _cleanup_fclose_ FILE *f = NULL;
1052         pid_t pid;
1053         int r;
1054
1055         assert(path);
1056
1057         r = cg_enumerate_processes(controller, path, &f);
1058         if (r == -ENOENT)
1059                 return 1;
1060         if (r < 0)
1061                 return r;
1062
1063         r = cg_read_pid(f, &pid);
1064         if (r < 0)
1065                 return r;
1066
1067         return r == 0;
1068 }
1069
1070 int cg_is_empty_recursive(const char *controller, const char *path) {
1071         int unified, r;
1072
1073         assert(path);
1074
1075         /* The root cgroup is always populated */
1076         if (controller && (isempty(path) || path_equal(path, "/")))
1077                 return false;
1078
1079         unified = cg_unified();
1080         if (unified < 0)
1081                 return unified;
1082
1083         if (unified > 0) {
1084                 _cleanup_free_ char *t = NULL;
1085
1086                 /* On the unified hierarchy we can check empty state
1087                  * via the "populated" attribute of "cgroup.events". */
1088
1089                 r = cg_read_event(controller, path, "populated", &t);
1090                 if (r < 0)
1091                         return r;
1092
1093                 return streq(t, "0");
1094         } else {
1095                 _cleanup_closedir_ DIR *d = NULL;
1096                 char *fn;
1097
1098                 r = cg_is_empty(controller, path);
1099                 if (r <= 0)
1100                         return r;
1101
1102                 r = cg_enumerate_subgroups(controller, path, &d);
1103                 if (r == -ENOENT)
1104                         return 1;
1105                 if (r < 0)
1106                         return r;
1107
1108                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1109                         _cleanup_free_ char *p = NULL;
1110
1111                         p = strjoin(path, "/", fn, NULL);
1112                         free(fn);
1113                         if (!p)
1114                                 return -ENOMEM;
1115
1116                         r = cg_is_empty_recursive(controller, p);
1117                         if (r <= 0)
1118                                 return r;
1119                 }
1120                 if (r < 0)
1121                         return r;
1122
1123                 return true;
1124         }
1125 }
1126
1127 int cg_split_spec(const char *spec, char **controller, char **path) {
1128         char *t = NULL, *u = NULL;
1129         const char *e;
1130
1131         assert(spec);
1132
1133         if (*spec == '/') {
1134                 if (!path_is_safe(spec))
1135                         return -EINVAL;
1136
1137                 if (path) {
1138                         t = strdup(spec);
1139                         if (!t)
1140                                 return -ENOMEM;
1141
1142                         *path = path_kill_slashes(t);
1143                 }
1144
1145                 if (controller)
1146                         *controller = NULL;
1147
1148                 return 0;
1149         }
1150
1151         e = strchr(spec, ':');
1152         if (!e) {
1153                 if (!cg_controller_is_valid(spec))
1154                         return -EINVAL;
1155
1156                 if (controller) {
1157                         t = strdup(spec);
1158                         if (!t)
1159                                 return -ENOMEM;
1160
1161                         *controller = t;
1162                 }
1163
1164                 if (path)
1165                         *path = NULL;
1166
1167                 return 0;
1168         }
1169
1170         t = strndup(spec, e-spec);
1171         if (!t)
1172                 return -ENOMEM;
1173         if (!cg_controller_is_valid(t)) {
1174                 free(t);
1175                 return -EINVAL;
1176         }
1177
1178         if (isempty(e+1))
1179                 u = NULL;
1180         else {
1181                 u = strdup(e+1);
1182                 if (!u) {
1183                         free(t);
1184                         return -ENOMEM;
1185                 }
1186
1187                 if (!path_is_safe(u) ||
1188                     !path_is_absolute(u)) {
1189                         free(t);
1190                         free(u);
1191                         return -EINVAL;
1192                 }
1193
1194                 path_kill_slashes(u);
1195         }
1196
1197         if (controller)
1198                 *controller = t;
1199         else
1200                 free(t);
1201
1202         if (path)
1203                 *path = u;
1204         else
1205                 free(u);
1206
1207         return 0;
1208 }
1209
1210 int cg_mangle_path(const char *path, char **result) {
1211         _cleanup_free_ char *c = NULL, *p = NULL;
1212         char *t;
1213         int r;
1214
1215         assert(path);
1216         assert(result);
1217
1218         /* First, check if it already is a filesystem path */
1219         if (path_startswith(path, "/sys/fs/cgroup")) {
1220
1221                 t = strdup(path);
1222                 if (!t)
1223                         return -ENOMEM;
1224
1225                 *result = path_kill_slashes(t);
1226                 return 0;
1227         }
1228
1229         /* Otherwise, treat it as cg spec */
1230         r = cg_split_spec(path, &c, &p);
1231         if (r < 0)
1232                 return r;
1233
1234         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
1235 }
1236
1237 int cg_get_root_path(char **path) {
1238         char *p, *e;
1239         int r;
1240
1241         assert(path);
1242
1243         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1244         if (r < 0)
1245                 return r;
1246
1247         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1248         if (!e)
1249                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1250         if (!e)
1251                 e = endswith(p, "/system"); /* even more legacy */
1252         if (e)
1253                 *e = 0;
1254
1255         *path = p;
1256         return 0;
1257 }
1258
1259 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1260         _cleanup_free_ char *rt = NULL;
1261         char *p;
1262         int r;
1263
1264         assert(cgroup);
1265         assert(shifted);
1266
1267         if (!root) {
1268                 /* If the root was specified let's use that, otherwise
1269                  * let's determine it from PID 1 */
1270
1271                 r = cg_get_root_path(&rt);
1272                 if (r < 0)
1273                         return r;
1274
1275                 root = rt;
1276         }
1277
1278         p = path_startswith(cgroup, root);
1279         if (p && p > cgroup)
1280                 *shifted = p - 1;
1281         else
1282                 *shifted = cgroup;
1283
1284         return 0;
1285 }
1286
1287 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1288         _cleanup_free_ char *raw = NULL;
1289         const char *c;
1290         int r;
1291
1292         assert(pid >= 0);
1293         assert(cgroup);
1294
1295         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1296         if (r < 0)
1297                 return r;
1298
1299         r = cg_shift_path(raw, root, &c);
1300         if (r < 0)
1301                 return r;
1302
1303         if (c == raw) {
1304                 *cgroup = raw;
1305                 raw = NULL;
1306         } else {
1307                 char *n;
1308
1309                 n = strdup(c);
1310                 if (!n)
1311                         return -ENOMEM;
1312
1313                 *cgroup = n;
1314         }
1315
1316         return 0;
1317 }
1318
1319 int cg_path_decode_unit(const char *cgroup, char **unit) {
1320         char *c, *s;
1321         size_t n;
1322
1323         assert(cgroup);
1324         assert(unit);
1325
1326         n = strcspn(cgroup, "/");
1327         if (n < 3)
1328                 return -ENXIO;
1329
1330         c = strndupa(cgroup, n);
1331         c = cg_unescape(c);
1332
1333         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1334                 return -ENXIO;
1335
1336         s = strdup(c);
1337         if (!s)
1338                 return -ENOMEM;
1339
1340         *unit = s;
1341         return 0;
1342 }
1343
1344 static bool valid_slice_name(const char *p, size_t n) {
1345
1346         if (!p)
1347                 return false;
1348
1349         if (n < strlen("x.slice"))
1350                 return false;
1351
1352         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1353                 char buf[n+1], *c;
1354
1355                 memcpy(buf, p, n);
1356                 buf[n] = 0;
1357
1358                 c = cg_unescape(buf);
1359
1360                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1361         }
1362
1363         return false;
1364 }
1365
1366 static const char *skip_slices(const char *p) {
1367         assert(p);
1368
1369         /* Skips over all slice assignments */
1370
1371         for (;;) {
1372                 size_t n;
1373
1374                 p += strspn(p, "/");
1375
1376                 n = strcspn(p, "/");
1377                 if (!valid_slice_name(p, n))
1378                         return p;
1379
1380                 p += n;
1381         }
1382 }
1383
1384 int cg_path_get_unit(const char *path, char **ret) {
1385         const char *e;
1386         char *unit;
1387         int r;
1388
1389         assert(path);
1390         assert(ret);
1391
1392         e = skip_slices(path);
1393
1394         r = cg_path_decode_unit(e, &unit);
1395         if (r < 0)
1396                 return r;
1397
1398         /* We skipped over the slices, don't accept any now */
1399         if (endswith(unit, ".slice")) {
1400                 free(unit);
1401                 return -ENXIO;
1402         }
1403
1404         *ret = unit;
1405         return 0;
1406 }
1407
1408 int cg_pid_get_unit(pid_t pid, char **unit) {
1409         _cleanup_free_ char *cgroup = NULL;
1410         int r;
1411
1412         assert(unit);
1413
1414         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1415         if (r < 0)
1416                 return r;
1417
1418         return cg_path_get_unit(cgroup, unit);
1419 }
1420
1421 /**
1422  * Skip session-*.scope, but require it to be there.
1423  */
1424 static const char *skip_session(const char *p) {
1425         size_t n;
1426
1427         if (isempty(p))
1428                 return NULL;
1429
1430         p += strspn(p, "/");
1431
1432         n = strcspn(p, "/");
1433         if (n < strlen("session-x.scope"))
1434                 return NULL;
1435
1436         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1437                 char buf[n - 8 - 6 + 1];
1438
1439                 memcpy(buf, p + 8, n - 8 - 6);
1440                 buf[n - 8 - 6] = 0;
1441
1442                 /* Note that session scopes never need unescaping,
1443                  * since they cannot conflict with the kernel's own
1444                  * names, hence we don't need to call cg_unescape()
1445                  * here. */
1446
1447                 if (!session_id_valid(buf))
1448                         return false;
1449
1450                 p += n;
1451                 p += strspn(p, "/");
1452                 return p;
1453         }
1454
1455         return NULL;
1456 }
1457
1458 /**
1459  * Skip user@*.service, but require it to be there.
1460  */
1461 static const char *skip_user_manager(const char *p) {
1462         size_t n;
1463
1464         if (isempty(p))
1465                 return NULL;
1466
1467         p += strspn(p, "/");
1468
1469         n = strcspn(p, "/");
1470         if (n < strlen("user@x.service"))
1471                 return NULL;
1472
1473         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1474                 char buf[n - 5 - 8 + 1];
1475
1476                 memcpy(buf, p + 5, n - 5 - 8);
1477                 buf[n - 5 - 8] = 0;
1478
1479                 /* Note that user manager services never need unescaping,
1480                  * since they cannot conflict with the kernel's own
1481                  * names, hence we don't need to call cg_unescape()
1482                  * here. */
1483
1484                 if (parse_uid(buf, NULL) < 0)
1485                         return NULL;
1486
1487                 p += n;
1488                 p += strspn(p, "/");
1489
1490                 return p;
1491         }
1492
1493         return NULL;
1494 }
1495
1496 static const char *skip_user_prefix(const char *path) {
1497         const char *e, *t;
1498
1499         assert(path);
1500
1501         /* Skip slices, if there are any */
1502         e = skip_slices(path);
1503
1504         /* Skip the user manager, if it's in the path now... */
1505         t = skip_user_manager(e);
1506         if (t)
1507                 return t;
1508
1509         /* Alternatively skip the user session if it is in the path... */
1510         return skip_session(e);
1511 }
1512
1513 int cg_path_get_user_unit(const char *path, char **ret) {
1514         const char *t;
1515
1516         assert(path);
1517         assert(ret);
1518
1519         t = skip_user_prefix(path);
1520         if (!t)
1521                 return -ENXIO;
1522
1523         /* And from here on it looks pretty much the same as for a
1524          * system unit, hence let's use the same parser from here
1525          * on. */
1526         return cg_path_get_unit(t, ret);
1527 }
1528
1529 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1530         _cleanup_free_ char *cgroup = NULL;
1531         int r;
1532
1533         assert(unit);
1534
1535         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1536         if (r < 0)
1537                 return r;
1538
1539         return cg_path_get_user_unit(cgroup, unit);
1540 }
1541
1542 int cg_path_get_machine_name(const char *path, char **machine) {
1543         _cleanup_free_ char *u = NULL;
1544         const char *sl;
1545         int r;
1546
1547         r = cg_path_get_unit(path, &u);
1548         if (r < 0)
1549                 return r;
1550
1551         sl = strjoina("/run/systemd/machines/unit:", u);
1552         return readlink_malloc(sl, machine);
1553 }
1554
1555 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1556         _cleanup_free_ char *cgroup = NULL;
1557         int r;
1558
1559         assert(machine);
1560
1561         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1562         if (r < 0)
1563                 return r;
1564
1565         return cg_path_get_machine_name(cgroup, machine);
1566 }
1567
1568 int cg_path_get_session(const char *path, char **session) {
1569         _cleanup_free_ char *unit = NULL;
1570         char *start, *end;
1571         int r;
1572
1573         assert(path);
1574
1575         r = cg_path_get_unit(path, &unit);
1576         if (r < 0)
1577                 return r;
1578
1579         start = startswith(unit, "session-");
1580         if (!start)
1581                 return -ENXIO;
1582         end = endswith(start, ".scope");
1583         if (!end)
1584                 return -ENXIO;
1585
1586         *end = 0;
1587         if (!session_id_valid(start))
1588                 return -ENXIO;
1589
1590         if (session) {
1591                 char *rr;
1592
1593                 rr = strdup(start);
1594                 if (!rr)
1595                         return -ENOMEM;
1596
1597                 *session = rr;
1598         }
1599
1600         return 0;
1601 }
1602
1603 int cg_pid_get_session(pid_t pid, char **session) {
1604         _cleanup_free_ char *cgroup = NULL;
1605         int r;
1606
1607         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1608         if (r < 0)
1609                 return r;
1610
1611         return cg_path_get_session(cgroup, session);
1612 }
1613
1614 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1615         _cleanup_free_ char *slice = NULL;
1616         char *start, *end;
1617         int r;
1618
1619         assert(path);
1620
1621         r = cg_path_get_slice(path, &slice);
1622         if (r < 0)
1623                 return r;
1624
1625         start = startswith(slice, "user-");
1626         if (!start)
1627                 return -ENXIO;
1628         end = endswith(start, ".slice");
1629         if (!end)
1630                 return -ENXIO;
1631
1632         *end = 0;
1633         if (parse_uid(start, uid) < 0)
1634                 return -ENXIO;
1635
1636         return 0;
1637 }
1638
1639 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1640         _cleanup_free_ char *cgroup = NULL;
1641         int r;
1642
1643         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1644         if (r < 0)
1645                 return r;
1646
1647         return cg_path_get_owner_uid(cgroup, uid);
1648 }
1649
1650 int cg_path_get_slice(const char *p, char **slice) {
1651         const char *e = NULL;
1652
1653         assert(p);
1654         assert(slice);
1655
1656         /* Finds the right-most slice unit from the beginning, but
1657          * stops before we come to the first non-slice unit. */
1658
1659         for (;;) {
1660                 size_t n;
1661
1662                 p += strspn(p, "/");
1663
1664                 n = strcspn(p, "/");
1665                 if (!valid_slice_name(p, n)) {
1666
1667                         if (!e) {
1668                                 char *s;
1669
1670                                 s = strdup("-.slice");
1671                                 if (!s)
1672                                         return -ENOMEM;
1673
1674                                 *slice = s;
1675                                 return 0;
1676                         }
1677
1678                         return cg_path_decode_unit(e, slice);
1679                 }
1680
1681                 e = p;
1682                 p += n;
1683         }
1684 }
1685
1686 int cg_pid_get_slice(pid_t pid, char **slice) {
1687         _cleanup_free_ char *cgroup = NULL;
1688         int r;
1689
1690         assert(slice);
1691
1692         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1693         if (r < 0)
1694                 return r;
1695
1696         return cg_path_get_slice(cgroup, slice);
1697 }
1698
1699 int cg_path_get_user_slice(const char *p, char **slice) {
1700         const char *t;
1701         assert(p);
1702         assert(slice);
1703
1704         t = skip_user_prefix(p);
1705         if (!t)
1706                 return -ENXIO;
1707
1708         /* And now it looks pretty much the same as for a system
1709          * slice, so let's just use the same parser from here on. */
1710         return cg_path_get_slice(t, slice);
1711 }
1712
1713 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1714         _cleanup_free_ char *cgroup = NULL;
1715         int r;
1716
1717         assert(slice);
1718
1719         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1720         if (r < 0)
1721                 return r;
1722
1723         return cg_path_get_user_slice(cgroup, slice);
1724 }
1725
1726 char *cg_escape(const char *p) {
1727         bool need_prefix = false;
1728
1729         /* This implements very minimal escaping for names to be used
1730          * as file names in the cgroup tree: any name which might
1731          * conflict with a kernel name or is prefixed with '_' is
1732          * prefixed with a '_'. That way, when reading cgroup names it
1733          * is sufficient to remove a single prefixing underscore if
1734          * there is one. */
1735
1736         /* The return value of this function (unlike cg_unescape())
1737          * needs free()! */
1738
1739         if (p[0] == 0 ||
1740             p[0] == '_' ||
1741             p[0] == '.' ||
1742             streq(p, "notify_on_release") ||
1743             streq(p, "release_agent") ||
1744             streq(p, "tasks") ||
1745             startswith(p, "cgroup."))
1746                 need_prefix = true;
1747         else {
1748                 const char *dot;
1749
1750                 dot = strrchr(p, '.');
1751                 if (dot) {
1752                         CGroupController c;
1753                         size_t l = dot - p;
1754
1755                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1756                                 const char *n;
1757
1758                                 n = cgroup_controller_to_string(c);
1759
1760                                 if (l != strlen(n))
1761                                         continue;
1762
1763                                 if (memcmp(p, n, l) != 0)
1764                                         continue;
1765
1766                                 need_prefix = true;
1767                                 break;
1768                         }
1769                 }
1770         }
1771
1772         if (need_prefix)
1773                 return strappend("_", p);
1774
1775         return strdup(p);
1776 }
1777
1778 char *cg_unescape(const char *p) {
1779         assert(p);
1780
1781         /* The return value of this function (unlike cg_escape())
1782          * doesn't need free()! */
1783
1784         if (p[0] == '_')
1785                 return (char*) p+1;
1786
1787         return (char*) p;
1788 }
1789
1790 #define CONTROLLER_VALID                        \
1791         DIGITS LETTERS                          \
1792         "_"
1793
1794 bool cg_controller_is_valid(const char *p) {
1795         const char *t, *s;
1796
1797         if (!p)
1798                 return false;
1799
1800         s = startswith(p, "name=");
1801         if (s)
1802                 p = s;
1803
1804         if (*p == 0 || *p == '_')
1805                 return false;
1806
1807         for (t = p; *t; t++)
1808                 if (!strchr(CONTROLLER_VALID, *t))
1809                         return false;
1810
1811         if (t - p > FILENAME_MAX)
1812                 return false;
1813
1814         return true;
1815 }
1816
1817 int cg_slice_to_path(const char *unit, char **ret) {
1818         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1819         const char *dash;
1820         int r;
1821
1822         assert(unit);
1823         assert(ret);
1824
1825         if (streq(unit, "-.slice")) {
1826                 char *x;
1827
1828                 x = strdup("");
1829                 if (!x)
1830                         return -ENOMEM;
1831                 *ret = x;
1832                 return 0;
1833         }
1834
1835         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1836                 return -EINVAL;
1837
1838         if (!endswith(unit, ".slice"))
1839                 return -EINVAL;
1840
1841         r = unit_name_to_prefix(unit, &p);
1842         if (r < 0)
1843                 return r;
1844
1845         dash = strchr(p, '-');
1846
1847         /* Don't allow initial dashes */
1848         if (dash == p)
1849                 return -EINVAL;
1850
1851         while (dash) {
1852                 _cleanup_free_ char *escaped = NULL;
1853                 char n[dash - p + sizeof(".slice")];
1854
1855                 /* Don't allow trailing or double dashes */
1856                 if (dash[1] == 0 || dash[1] == '-')
1857                         return -EINVAL;
1858
1859                 strcpy(stpncpy(n, p, dash - p), ".slice");
1860                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1861                         return -EINVAL;
1862
1863                 escaped = cg_escape(n);
1864                 if (!escaped)
1865                         return -ENOMEM;
1866
1867                 if (!strextend(&s, escaped, "/", NULL))
1868                         return -ENOMEM;
1869
1870                 dash = strchr(dash+1, '-');
1871         }
1872
1873         e = cg_escape(unit);
1874         if (!e)
1875                 return -ENOMEM;
1876
1877         if (!strextend(&s, e, NULL))
1878                 return -ENOMEM;
1879
1880         *ret = s;
1881         s = NULL;
1882
1883         return 0;
1884 }
1885
1886 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1887         _cleanup_free_ char *p = NULL;
1888         int r;
1889
1890         r = cg_get_path(controller, path, attribute, &p);
1891         if (r < 0)
1892                 return r;
1893
1894         return write_string_file(p, value, 0);
1895 }
1896
1897 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1898         _cleanup_free_ char *p = NULL;
1899         int r;
1900
1901         r = cg_get_path(controller, path, attribute, &p);
1902         if (r < 0)
1903                 return r;
1904
1905         return read_one_line_file(p, ret);
1906 }
1907
1908 int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) {
1909         CGroupController c;
1910         int r, unified;
1911
1912         /* This one will create a cgroup in our private tree, but also
1913          * duplicate it in the trees specified in mask, and remove it
1914          * in all others */
1915
1916         /* First create the cgroup in our own hierarchy. */
1917         r = cg_create(SYSTEMD_CGROUP_CONTROLLER, path);
1918         if (r < 0)
1919                 return r;
1920
1921         /* If we are in the unified hierarchy, we are done now */
1922         unified = cg_unified();
1923         if (unified < 0)
1924                 return unified;
1925         if (unified > 0)
1926                 return 0;
1927
1928         /* Otherwise, do the same in the other hierarchies */
1929         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1930                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1931                 const char *n;
1932
1933                 n = cgroup_controller_to_string(c);
1934
1935                 if (mask & bit)
1936                         (void) cg_create(n, path);
1937                 else if (supported & bit)
1938                         (void) cg_trim(n, path, true);
1939         }
1940
1941         return 0;
1942 }
1943
1944 int cg_attach_everywhere(CGroupMask supported, const char *path, pid_t pid, cg_migrate_callback_t path_callback, void *userdata) {
1945         CGroupController c;
1946         int r, unified;
1947
1948         r = cg_attach(SYSTEMD_CGROUP_CONTROLLER, path, pid);
1949         if (r < 0)
1950                 return r;
1951
1952         unified = cg_unified();
1953         if (unified < 0)
1954                 return unified;
1955         if (unified > 0)
1956                 return 0;
1957
1958         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1959                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1960                 const char *p = NULL;
1961
1962                 if (!(supported & bit))
1963                         continue;
1964
1965                 if (path_callback)
1966                         p = path_callback(bit, userdata);
1967
1968                 if (!p)
1969                         p = path;
1970
1971                 (void) cg_attach_fallback(cgroup_controller_to_string(c), p, pid);
1972         }
1973
1974         return 0;
1975 }
1976
1977 int cg_attach_many_everywhere(CGroupMask supported, const char *path, Set* pids, cg_migrate_callback_t path_callback, void *userdata) {
1978         Iterator i;
1979         void *pidp;
1980         int r = 0;
1981
1982         SET_FOREACH(pidp, pids, i) {
1983                 pid_t pid = PTR_TO_PID(pidp);
1984                 int q;
1985
1986                 q = cg_attach_everywhere(supported, path, pid, path_callback, userdata);
1987                 if (q < 0 && r >= 0)
1988                         r = q;
1989         }
1990
1991         return r;
1992 }
1993
1994 int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to, cg_migrate_callback_t to_callback, void *userdata) {
1995         CGroupController c;
1996         int r = 0, unified;
1997
1998         if (!path_equal(from, to))  {
1999                 r = cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER, from, SYSTEMD_CGROUP_CONTROLLER, to, CGROUP_REMOVE);
2000                 if (r < 0)
2001                         return r;
2002         }
2003
2004         unified = cg_unified();
2005         if (unified < 0)
2006                 return unified;
2007         if (unified > 0)
2008                 return r;
2009
2010         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2011                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2012                 const char *p = NULL;
2013
2014                 if (!(supported & bit))
2015                         continue;
2016
2017                 if (to_callback)
2018                         p = to_callback(bit, userdata);
2019
2020                 if (!p)
2021                         p = to;
2022
2023                 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER, to, cgroup_controller_to_string(c), p, 0);
2024         }
2025
2026         return 0;
2027 }
2028
2029 int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) {
2030         CGroupController c;
2031         int r, unified;
2032
2033         r = cg_trim(SYSTEMD_CGROUP_CONTROLLER, path, delete_root);
2034         if (r < 0)
2035                 return r;
2036
2037         unified = cg_unified();
2038         if (unified < 0)
2039                 return unified;
2040         if (unified > 0)
2041                 return r;
2042
2043         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2044                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2045
2046                 if (!(supported & bit))
2047                         continue;
2048
2049                 (void) cg_trim(cgroup_controller_to_string(c), path, delete_root);
2050         }
2051
2052         return 0;
2053 }
2054
2055 int cg_mask_supported(CGroupMask *ret) {
2056         CGroupMask mask = 0;
2057         int r, unified;
2058
2059         /* Determines the mask of supported cgroup controllers. Only
2060          * includes controllers we can make sense of and that are
2061          * actually accessible. */
2062
2063         unified = cg_unified();
2064         if (unified < 0)
2065                 return unified;
2066         if (unified > 0) {
2067                 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
2068                 const char *c;
2069
2070                 /* In the unified hierarchy we can read the supported
2071                  * and accessible controllers from a the top-level
2072                  * cgroup attribute */
2073
2074                 r = cg_get_root_path(&root);
2075                 if (r < 0)
2076                         return r;
2077
2078                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
2079                 if (r < 0)
2080                         return r;
2081
2082                 r = read_one_line_file(path, &controllers);
2083                 if (r < 0)
2084                         return r;
2085
2086                 c = controllers;
2087                 for (;;) {
2088                         _cleanup_free_ char *n = NULL;
2089                         CGroupController v;
2090
2091                         r = extract_first_word(&c, &n, NULL, 0);
2092                         if (r < 0)
2093                                 return r;
2094                         if (r == 0)
2095                                 break;
2096
2097                         v = cgroup_controller_from_string(n);
2098                         if (v < 0)
2099                                 continue;
2100
2101                         mask |= CGROUP_CONTROLLER_TO_MASK(v);
2102                 }
2103
2104                 /* Currently, we only support the memory, io and pids
2105                  * controller in the unified hierarchy, mask
2106                  * everything else off. */
2107                 mask &= CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
2108
2109         } else {
2110                 CGroupController c;
2111
2112                 /* In the legacy hierarchy, we check whether which
2113                  * hierarchies are mounted. */
2114
2115                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2116                         const char *n;
2117
2118                         n = cgroup_controller_to_string(c);
2119                         if (controller_is_accessible(n) >= 0)
2120                                 mask |= CGROUP_CONTROLLER_TO_MASK(c);
2121                 }
2122         }
2123
2124         *ret = mask;
2125         return 0;
2126 }
2127
2128 int cg_kernel_controllers(Set *controllers) {
2129         _cleanup_fclose_ FILE *f = NULL;
2130         char buf[LINE_MAX];
2131         int r;
2132
2133         assert(controllers);
2134
2135         /* Determines the full list of kernel-known controllers. Might
2136          * include controllers we don't actually support, arbitrary
2137          * named hierarchies and controllers that aren't currently
2138          * accessible (because not mounted). */
2139
2140         f = fopen("/proc/cgroups", "re");
2141         if (!f) {
2142                 if (errno == ENOENT)
2143                         return 0;
2144                 return -errno;
2145         }
2146
2147         /* Ignore the header line */
2148         (void) fgets(buf, sizeof(buf), f);
2149
2150         for (;;) {
2151                 char *controller;
2152                 int enabled = 0;
2153
2154                 errno = 0;
2155                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
2156
2157                         if (feof(f))
2158                                 break;
2159
2160                         if (ferror(f) && errno > 0)
2161                                 return -errno;
2162
2163                         return -EBADMSG;
2164                 }
2165
2166                 if (!enabled) {
2167                         free(controller);
2168                         continue;
2169                 }
2170
2171                 if (!cg_controller_is_valid(controller)) {
2172                         free(controller);
2173                         return -EBADMSG;
2174                 }
2175
2176                 r = set_consume(controllers, controller);
2177                 if (r < 0)
2178                         return r;
2179         }
2180
2181         return 0;
2182 }
2183
2184 static thread_local int unified_cache = -1;
2185
2186 int cg_unified(void) {
2187         struct statfs fs;
2188
2189         /* Checks if we support the unified hierarchy. Returns an
2190          * error when the cgroup hierarchies aren't mounted yet or we
2191          * have any other trouble determining if the unified hierarchy
2192          * is supported. */
2193
2194         if (unified_cache >= 0)
2195                 return unified_cache;
2196
2197         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2198                 return -errno;
2199
2200         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC))
2201                 unified_cache = true;
2202         else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC))
2203                 unified_cache = false;
2204         else
2205                 return -ENOMEDIUM;
2206
2207         return unified_cache;
2208 }
2209
2210 void cg_unified_flush(void) {
2211         unified_cache = -1;
2212 }
2213
2214 int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) {
2215         _cleanup_free_ char *fs = NULL;
2216         CGroupController c;
2217         int r, unified;
2218
2219         assert(p);
2220
2221         if (supported == 0)
2222                 return 0;
2223
2224         unified = cg_unified();
2225         if (unified < 0)
2226                 return unified;
2227         if (!unified) /* on the legacy hiearchy there's no joining of controllers defined */
2228                 return 0;
2229
2230         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, p, "cgroup.subtree_control", &fs);
2231         if (r < 0)
2232                 return r;
2233
2234         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2235                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2236                 const char *n;
2237
2238                 if (!(supported & bit))
2239                         continue;
2240
2241                 n = cgroup_controller_to_string(c);
2242                 {
2243                         char s[1 + strlen(n) + 1];
2244
2245                         s[0] = mask & bit ? '+' : '-';
2246                         strcpy(s + 1, n);
2247
2248                         r = write_string_file(fs, s, 0);
2249                         if (r < 0)
2250                                 log_debug_errno(r, "Failed to enable controller %s for %s (%s): %m", n, p, fs);
2251                 }
2252         }
2253
2254         return 0;
2255 }
2256
2257 bool cg_is_unified_wanted(void) {
2258         static thread_local int wanted = -1;
2259         int r, unified;
2260
2261         /* If the hierarchy is already mounted, then follow whatever
2262          * was chosen for it. */
2263         unified = cg_unified();
2264         if (unified >= 0)
2265                 return unified;
2266
2267         /* Otherwise, let's see what the kernel command line has to
2268          * say. Since checking that is expensive, let's cache the
2269          * result. */
2270         if (wanted >= 0)
2271                 return wanted;
2272
2273         r = get_proc_cmdline_key("systemd.unified_cgroup_hierarchy", NULL);
2274         if (r > 0)
2275                 return (wanted = true);
2276         else {
2277                 _cleanup_free_ char *value = NULL;
2278
2279                 r = get_proc_cmdline_key("systemd.unified_cgroup_hierarchy=", &value);
2280                 if (r < 0)
2281                         return false;
2282                 if (r == 0)
2283                         return (wanted = false);
2284
2285                 return (wanted = parse_boolean(value) > 0);
2286         }
2287 }
2288
2289 bool cg_is_legacy_wanted(void) {
2290         return !cg_is_unified_wanted();
2291 }
2292
2293 int cg_weight_parse(const char *s, uint64_t *ret) {
2294         uint64_t u;
2295         int r;
2296
2297         if (isempty(s)) {
2298                 *ret = CGROUP_WEIGHT_INVALID;
2299                 return 0;
2300         }
2301
2302         r = safe_atou64(s, &u);
2303         if (r < 0)
2304                 return r;
2305
2306         if (u < CGROUP_WEIGHT_MIN || u > CGROUP_WEIGHT_MAX)
2307                 return -ERANGE;
2308
2309         *ret = u;
2310         return 0;
2311 }
2312
2313 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2314         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2315         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2316         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2317         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2318 };
2319
2320 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2321         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2322         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2323         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2324         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2325 };
2326
2327 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2328
2329 int cg_cpu_shares_parse(const char *s, uint64_t *ret) {
2330         uint64_t u;
2331         int r;
2332
2333         if (isempty(s)) {
2334                 *ret = CGROUP_CPU_SHARES_INVALID;
2335                 return 0;
2336         }
2337
2338         r = safe_atou64(s, &u);
2339         if (r < 0)
2340                 return r;
2341
2342         if (u < CGROUP_CPU_SHARES_MIN || u > CGROUP_CPU_SHARES_MAX)
2343                 return -ERANGE;
2344
2345         *ret = u;
2346         return 0;
2347 }
2348
2349 int cg_blkio_weight_parse(const char *s, uint64_t *ret) {
2350         uint64_t u;
2351         int r;
2352
2353         if (isempty(s)) {
2354                 *ret = CGROUP_BLKIO_WEIGHT_INVALID;
2355                 return 0;
2356         }
2357
2358         r = safe_atou64(s, &u);
2359         if (r < 0)
2360                 return r;
2361
2362         if (u < CGROUP_BLKIO_WEIGHT_MIN || u > CGROUP_BLKIO_WEIGHT_MAX)
2363                 return -ERANGE;
2364
2365         *ret = u;
2366         return 0;
2367 }
2368
2369 static const char *cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2370         [CGROUP_CONTROLLER_CPU] = "cpu",
2371         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2372         [CGROUP_CONTROLLER_IO] = "io",
2373         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2374         [CGROUP_CONTROLLER_MEMORY] = "memory",
2375         [CGROUP_CONTROLLER_DEVICES] = "devices",
2376         [CGROUP_CONTROLLER_PIDS] = "pids",
2377 };
2378
2379 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);