src/basic/cgroup-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2 /***
   3   This file is part of systemd.
   4
   5   Copyright 2010 Lennart Poettering
   6 ***/
   7
   8 #include <dirent.h>
   9 #include <errno.h>
  10 #include <ftw.h>
  11 #include <limits.h>
  12 #include <signal.h>
  13 #include <stddef.h>
  14 #include <stdio_ext.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17 #include <sys/stat.h>
  18 #include <sys/statfs.h>
  19 #include <sys/types.h>
  20 #include <sys/xattr.h>
  21 #include <unistd.h>
  22
  23 #include "alloc-util.h"
  24 #include "cgroup-util.h"
  25 #include "def.h"
  26 #include "dirent-util.h"
  27 #include "extract-word.h"
  28 #include "fd-util.h"
  29 #include "fileio.h"
  30 #include "format-util.h"
  31 #include "fs-util.h"
  32 #include "log.h"
  33 #include "login-util.h"
  34 #include "macro.h"
  35 #include "missing.h"
  36 #include "mkdir.h"
  37 #include "parse-util.h"
  38 #include "path-util.h"
  39 #include "proc-cmdline.h"
  40 #include "process-util.h"
  41 #include "set.h"
  42 #include "special.h"
  43 #include "stat-util.h"
  44 #include "stdio-util.h"
  45 #include "string-table.h"
  46 #include "string-util.h"
  47 #include "strv.h"
  48 #include "unit-name.h"
  49 #include "user-util.h"
  50
  51 int cg_enumerate_processes(const char *controller, const char *path, FILE **_f) {
  52         _cleanup_free_ char *fs = NULL;
  53         FILE *f;
  54         int r;
  55
  56         assert(_f);
  57
  58         r = cg_get_path(controller, path, "cgroup.procs", &fs);
  59         if (r < 0)
  60                 return r;
  61
  62         f = fopen(fs, "re");
  63         if (!f)
  64                 return -errno;
  65
  66         *_f = f;
  67         return 0;
  68 }
  69
  70 int cg_read_pid(FILE *f, pid_t *_pid) {
  71         unsigned long ul;
  72
  73         /* Note that the cgroup.procs might contain duplicates! See
  74          * cgroups.txt for details. */
  75
  76         assert(f);
  77         assert(_pid);
  78
  79         errno = 0;
  80         if (fscanf(f, "%lu", &ul) != 1) {
  81
  82                 if (feof(f))
  83                         return 0;
  84
  85                 return errno > 0 ? -errno : -EIO;
  86         }
  87
  88         if (ul <= 0)
  89                 return -EIO;
  90
  91         *_pid = (pid_t) ul;
  92         return 1;
  93 }
  94
  95 int cg_read_event(
  96                 const char *controller,
  97                 const char *path,
  98                 const char *event,
  99                 char **val) {
 100
 101         _cleanup_free_ char *events = NULL, *content = NULL;
 102         char *p, *line;
 103         int r;
 104
 105         r = cg_get_path(controller, path, "cgroup.events", &events);
 106         if (r < 0)
 107                 return r;
 108
 109         r = read_full_file(events, &content, NULL);
 110         if (r < 0)
 111                 return r;
 112
 113         p = content;
 114         while ((line = strsep(&p, "\n"))) {
 115                 char *key;
 116
 117                 key = strsep(&line, " ");
 118                 if (!key || !line)
 119                         return -EINVAL;
 120
 121                 if (strcmp(key, event))
 122                         continue;
 123
 124                 *val = strdup(line);
 125                 return 0;
 126         }
 127
 128         return -ENOENT;
 129 }
 130
 131 bool cg_ns_supported(void) {
 132         static thread_local int enabled = -1;
 133
 134         if (enabled >= 0)
 135                 return enabled;
 136
 137         if (access("/proc/self/ns/cgroup", F_OK) == 0)
 138                 enabled = 1;
 139         else
 140                 enabled = 0;
 141
 142         return enabled;
 143 }
 144
 145 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **_d) {
 146         _cleanup_free_ char *fs = NULL;
 147         int r;
 148         DIR *d;
 149
 150         assert(_d);
 151
 152         /* This is not recursive! */
 153
 154         r = cg_get_path(controller, path, NULL, &fs);
 155         if (r < 0)
 156                 return r;
 157
 158         d = opendir(fs);
 159         if (!d)
 160                 return -errno;
 161
 162         *_d = d;
 163         return 0;
 164 }
 165
 166 int cg_read_subgroup(DIR *d, char **fn) {
 167         struct dirent *de;
 168
 169         assert(d);
 170         assert(fn);
 171
 172         FOREACH_DIRENT_ALL(de, d, return -errno) {
 173                 char *b;
 174
 175                 if (de->d_type != DT_DIR)
 176                         continue;
 177
 178                 if (dot_or_dot_dot(de->d_name))
 179                         continue;
 180
 181                 b = strdup(de->d_name);
 182                 if (!b)
 183                         return -ENOMEM;
 184
 185                 *fn = b;
 186                 return 1;
 187         }
 188
 189         return 0;
 190 }
 191
 192 int cg_rmdir(const char *controller, const char *path) {
 193         _cleanup_free_ char *p = NULL;
 194         int r;
 195
 196         r = cg_get_path(controller, path, NULL, &p);
 197         if (r < 0)
 198                 return r;
 199
 200         r = rmdir(p);
 201         if (r < 0 && errno != ENOENT)
 202                 return -errno;
 203
 204         r = cg_hybrid_unified();
 205         if (r < 0)
 206                 return r;
 207         if (r == 0)
 208                 return 0;
 209
 210         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 211                 r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
 212                 if (r < 0)
 213                         log_warning_errno(r, "Failed to remove compat systemd cgroup %s: %m", path);
 214         }
 215
 216         return 0;
 217 }
 218
 219 int cg_kill(
 220                 const char *controller,
 221                 const char *path,
 222                 int sig,
 223                 CGroupFlags flags,
 224                 Set *s,
 225                 cg_kill_log_func_t log_kill,
 226                 void *userdata) {
 227
 228         _cleanup_set_free_ Set *allocated_set = NULL;
 229         bool done = false;
 230         int r, ret = 0;
 231         pid_t my_pid;
 232
 233         assert(sig >= 0);
 234
 235          /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
 236           * SIGCONT on SIGKILL. */
 237         if (IN_SET(sig, SIGCONT, SIGKILL))
 238                 flags &= ~CGROUP_SIGCONT;
 239
 240         /* This goes through the tasks list and kills them all. This
 241          * is repeated until no further processes are added to the
 242          * tasks list, to properly handle forking processes */
 243
 244         if (!s) {
 245                 s = allocated_set = set_new(NULL);
 246                 if (!s)
 247                         return -ENOMEM;
 248         }
 249
 250         my_pid = getpid_cached();
 251
 252         do {
 253                 _cleanup_fclose_ FILE *f = NULL;
 254                 pid_t pid = 0;
 255                 done = true;
 256
 257                 r = cg_enumerate_processes(controller, path, &f);
 258                 if (r < 0) {
 259                         if (ret >= 0 && r != -ENOENT)
 260                                 return r;
 261
 262                         return ret;
 263                 }
 264
 265                 while ((r = cg_read_pid(f, &pid)) > 0) {
 266
 267                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 268                                 continue;
 269
 270                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 271                                 continue;
 272
 273                         if (log_kill)
 274                                 log_kill(pid, sig, userdata);
 275
 276                         /* If we haven't killed this process yet, kill
 277                          * it */
 278                         if (kill(pid, sig) < 0) {
 279                                 if (ret >= 0 && errno != ESRCH)
 280                                         ret = -errno;
 281                         } else {
 282                                 if (flags & CGROUP_SIGCONT)
 283                                         (void) kill(pid, SIGCONT);
 284
 285                                 if (ret == 0)
 286                                         ret = 1;
 287                         }
 288
 289                         done = false;
 290
 291                         r = set_put(s, PID_TO_PTR(pid));
 292                         if (r < 0) {
 293                                 if (ret >= 0)
 294                                         return r;
 295
 296                                 return ret;
 297                         }
 298                 }
 299
 300                 if (r < 0) {
 301                         if (ret >= 0)
 302                                 return r;
 303
 304                         return ret;
 305                 }
 306
 307                 /* To avoid racing against processes which fork
 308                  * quicker than we can kill them we repeat this until
 309                  * no new pids need to be killed. */
 310
 311         } while (!done);
 312
 313         return ret;
 314 }
 315
 316 int cg_kill_recursive(
 317                 const char *controller,
 318                 const char *path,
 319                 int sig,
 320                 CGroupFlags flags,
 321                 Set *s,
 322                 cg_kill_log_func_t log_kill,
 323                 void *userdata) {
 324
 325         _cleanup_set_free_ Set *allocated_set = NULL;
 326         _cleanup_closedir_ DIR *d = NULL;
 327         int r, ret;
 328         char *fn;
 329
 330         assert(path);
 331         assert(sig >= 0);
 332
 333         if (!s) {
 334                 s = allocated_set = set_new(NULL);
 335                 if (!s)
 336                         return -ENOMEM;
 337         }
 338
 339         ret = cg_kill(controller, path, sig, flags, s, log_kill, userdata);
 340
 341         r = cg_enumerate_subgroups(controller, path, &d);
 342         if (r < 0) {
 343                 if (ret >= 0 && r != -ENOENT)
 344                         return r;
 345
 346                 return ret;
 347         }
 348
 349         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 350                 _cleanup_free_ char *p = NULL;
 351
 352                 p = strjoin(path, "/", fn);
 353                 free(fn);
 354                 if (!p)
 355                         return -ENOMEM;
 356
 357                 r = cg_kill_recursive(controller, p, sig, flags, s, log_kill, userdata);
 358                 if (r != 0 && ret >= 0)
 359                         ret = r;
 360         }
 361         if (ret >= 0 && r < 0)
 362                 ret = r;
 363
 364         if (flags & CGROUP_REMOVE) {
 365                 r = cg_rmdir(controller, path);
 366                 if (r < 0 && ret >= 0 && !IN_SET(r, -ENOENT, -EBUSY))
 367                         return r;
 368         }
 369
 370         return ret;
 371 }
 372
 373 int cg_migrate(
 374                 const char *cfrom,
 375                 const char *pfrom,
 376                 const char *cto,
 377                 const char *pto,
 378                 CGroupFlags flags) {
 379
 380         bool done = false;
 381         _cleanup_set_free_ Set *s = NULL;
 382         int r, ret = 0;
 383         pid_t my_pid;
 384
 385         assert(cfrom);
 386         assert(pfrom);
 387         assert(cto);
 388         assert(pto);
 389
 390         s = set_new(NULL);
 391         if (!s)
 392                 return -ENOMEM;
 393
 394         my_pid = getpid_cached();
 395
 396         do {
 397                 _cleanup_fclose_ FILE *f = NULL;
 398                 pid_t pid = 0;
 399                 done = true;
 400
 401                 r = cg_enumerate_processes(cfrom, pfrom, &f);
 402                 if (r < 0) {
 403                         if (ret >= 0 && r != -ENOENT)
 404                                 return r;
 405
 406                         return ret;
 407                 }
 408
 409                 while ((r = cg_read_pid(f, &pid)) > 0) {
 410
 411                         /* This might do weird stuff if we aren't a
 412                          * single-threaded program. However, we
 413                          * luckily know we are not */
 414                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 415                                 continue;
 416
 417                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 418                                 continue;
 419
 420                         /* Ignore kernel threads. Since they can only
 421                          * exist in the root cgroup, we only check for
 422                          * them there. */
 423                         if (cfrom &&
 424                             (isempty(pfrom) || path_equal(pfrom, "/")) &&
 425                             is_kernel_thread(pid) > 0)
 426                                 continue;
 427
 428                         r = cg_attach(cto, pto, pid);
 429                         if (r < 0) {
 430                                 if (ret >= 0 && r != -ESRCH)
 431                                         ret = r;
 432                         } else if (ret == 0)
 433                                 ret = 1;
 434
 435                         done = false;
 436
 437                         r = set_put(s, PID_TO_PTR(pid));
 438                         if (r < 0) {
 439                                 if (ret >= 0)
 440                                         return r;
 441
 442                                 return ret;
 443                         }
 444                 }
 445
 446                 if (r < 0) {
 447                         if (ret >= 0)
 448                                 return r;
 449
 450                         return ret;
 451                 }
 452         } while (!done);
 453
 454         return ret;
 455 }
 456
 457 int cg_migrate_recursive(
 458                 const char *cfrom,
 459                 const char *pfrom,
 460                 const char *cto,
 461                 const char *pto,
 462                 CGroupFlags flags) {
 463
 464         _cleanup_closedir_ DIR *d = NULL;
 465         int r, ret = 0;
 466         char *fn;
 467
 468         assert(cfrom);
 469         assert(pfrom);
 470         assert(cto);
 471         assert(pto);
 472
 473         ret = cg_migrate(cfrom, pfrom, cto, pto, flags);
 474
 475         r = cg_enumerate_subgroups(cfrom, pfrom, &d);
 476         if (r < 0) {
 477                 if (ret >= 0 && r != -ENOENT)
 478                         return r;
 479
 480                 return ret;
 481         }
 482
 483         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 484                 _cleanup_free_ char *p = NULL;
 485
 486                 p = strjoin(pfrom, "/", fn);
 487                 free(fn);
 488                 if (!p)
 489                         return -ENOMEM;
 490
 491                 r = cg_migrate_recursive(cfrom, p, cto, pto, flags);
 492                 if (r != 0 && ret >= 0)
 493                         ret = r;
 494         }
 495
 496         if (r < 0 && ret >= 0)
 497                 ret = r;
 498
 499         if (flags & CGROUP_REMOVE) {
 500                 r = cg_rmdir(cfrom, pfrom);
 501                 if (r < 0 && ret >= 0 && !IN_SET(r, -ENOENT, -EBUSY))
 502                         return r;
 503         }
 504
 505         return ret;
 506 }
 507
 508 int cg_migrate_recursive_fallback(
 509                 const char *cfrom,
 510                 const char *pfrom,
 511                 const char *cto,
 512                 const char *pto,
 513                 CGroupFlags flags) {
 514
 515         int r;
 516
 517         assert(cfrom);
 518         assert(pfrom);
 519         assert(cto);
 520         assert(pto);
 521
 522         r = cg_migrate_recursive(cfrom, pfrom, cto, pto, flags);
 523         if (r < 0) {
 524                 char prefix[strlen(pto) + 1];
 525
 526                 /* This didn't work? Then let's try all prefixes of the destination */
 527
 528                 PATH_FOREACH_PREFIX(prefix, pto) {
 529                         int q;
 530
 531                         q = cg_migrate_recursive(cfrom, pfrom, cto, prefix, flags);
 532                         if (q >= 0)
 533                                 return q;
 534                 }
 535         }
 536
 537         return r;
 538 }
 539
 540 static const char *controller_to_dirname(const char *controller) {
 541         const char *e;
 542
 543         assert(controller);
 544
 545         /* Converts a controller name to the directory name below
 546          * /sys/fs/cgroup/ we want to mount it to. Effectively, this
 547          * just cuts off the name= prefixed used for named
 548          * hierarchies, if it is specified. */
 549
 550         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 551                 if (cg_hybrid_unified() > 0)
 552                         controller = SYSTEMD_CGROUP_CONTROLLER_HYBRID;
 553                 else
 554                         controller = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 555         }
 556
 557         e = startswith(controller, "name=");
 558         if (e)
 559                 return e;
 560
 561         return controller;
 562 }
 563
 564 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **fs) {
 565         const char *dn;
 566         char *t = NULL;
 567
 568         assert(fs);
 569         assert(controller);
 570
 571         dn = controller_to_dirname(controller);
 572
 573         if (isempty(path) && isempty(suffix))
 574                 t = strappend("/sys/fs/cgroup/", dn);
 575         else if (isempty(path))
 576                 t = strjoin("/sys/fs/cgroup/", dn, "/", suffix);
 577         else if (isempty(suffix))
 578                 t = strjoin("/sys/fs/cgroup/", dn, "/", path);
 579         else
 580                 t = strjoin("/sys/fs/cgroup/", dn, "/", path, "/", suffix);
 581         if (!t)
 582                 return -ENOMEM;
 583
 584         *fs = t;
 585         return 0;
 586 }
 587
 588 static int join_path_unified(const char *path, const char *suffix, char **fs) {
 589         char *t;
 590
 591         assert(fs);
 592
 593         if (isempty(path) && isempty(suffix))
 594                 t = strdup("/sys/fs/cgroup");
 595         else if (isempty(path))
 596                 t = strappend("/sys/fs/cgroup/", suffix);
 597         else if (isempty(suffix))
 598                 t = strappend("/sys/fs/cgroup/", path);
 599         else
 600                 t = strjoin("/sys/fs/cgroup/", path, "/", suffix);
 601         if (!t)
 602                 return -ENOMEM;
 603
 604         *fs = t;
 605         return 0;
 606 }
 607
 608 int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) {
 609         int r;
 610
 611         assert(fs);
 612
 613         if (!controller) {
 614                 char *t;
 615
 616                 /* If no controller is specified, we return the path
 617                  * *below* the controllers, without any prefix. */
 618
 619                 if (!path && !suffix)
 620                         return -EINVAL;
 621
 622                 if (!suffix)
 623                         t = strdup(path);
 624                 else if (!path)
 625                         t = strdup(suffix);
 626                 else
 627                         t = strjoin(path, "/", suffix);
 628                 if (!t)
 629                         return -ENOMEM;
 630
 631                 *fs = path_kill_slashes(t);
 632                 return 0;
 633         }
 634
 635         if (!cg_controller_is_valid(controller))
 636                 return -EINVAL;
 637
 638         r = cg_all_unified();
 639         if (r < 0)
 640                 return r;
 641         if (r > 0)
 642                 r = join_path_unified(path, suffix, fs);
 643         else
 644                 r = join_path_legacy(controller, path, suffix, fs);
 645         if (r < 0)
 646                 return r;
 647
 648         path_kill_slashes(*fs);
 649         return 0;
 650 }
 651
 652 static int controller_is_accessible(const char *controller) {
 653         int r;
 654
 655         assert(controller);
 656
 657         /* Checks whether a specific controller is accessible,
 658          * i.e. its hierarchy mounted. In the unified hierarchy all
 659          * controllers are considered accessible, except for the named
 660          * hierarchies */
 661
 662         if (!cg_controller_is_valid(controller))
 663                 return -EINVAL;
 664
 665         r = cg_all_unified();
 666         if (r < 0)
 667                 return r;
 668         if (r > 0) {
 669                 /* We don't support named hierarchies if we are using
 670                  * the unified hierarchy. */
 671
 672                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 673                         return 0;
 674
 675                 if (startswith(controller, "name="))
 676                         return -EOPNOTSUPP;
 677
 678         } else {
 679                 const char *cc, *dn;
 680
 681                 dn = controller_to_dirname(controller);
 682                 cc = strjoina("/sys/fs/cgroup/", dn);
 683
 684                 if (laccess(cc, F_OK) < 0)
 685                         return -errno;
 686         }
 687
 688         return 0;
 689 }
 690
 691 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 692         int r;
 693
 694         assert(controller);
 695         assert(fs);
 696
 697         /* Check if the specified controller is actually accessible */
 698         r = controller_is_accessible(controller);
 699         if (r < 0)
 700                 return r;
 701
 702         return cg_get_path(controller, path, suffix, fs);
 703 }
 704
 705 static int trim_cb(const char *path, const struct stat *sb, int typeflag, struct FTW *ftwbuf) {
 706         assert(path);
 707         assert(sb);
 708         assert(ftwbuf);
 709
 710         if (typeflag != FTW_DP)
 711                 return 0;
 712
 713         if (ftwbuf->level < 1)
 714                 return 0;
 715
 716         (void) rmdir(path);
 717         return 0;
 718 }
 719
 720 int cg_trim(const char *controller, const char *path, bool delete_root) {
 721         _cleanup_free_ char *fs = NULL;
 722         int r = 0, q;
 723
 724         assert(path);
 725
 726         r = cg_get_path(controller, path, NULL, &fs);
 727         if (r < 0)
 728                 return r;
 729
 730         errno = 0;
 731         if (nftw(fs, trim_cb, 64, FTW_DEPTH|FTW_MOUNT|FTW_PHYS) != 0) {
 732                 if (errno == ENOENT)
 733                         r = 0;
 734                 else if (errno > 0)
 735                         r = -errno;
 736                 else
 737                         r = -EIO;
 738         }
 739
 740         if (delete_root) {
 741                 if (rmdir(fs) < 0 && errno != ENOENT)
 742                         return -errno;
 743         }
 744
 745         q = cg_hybrid_unified();
 746         if (q < 0)
 747                 return q;
 748         if (q > 0 && streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 749                 q = cg_trim(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, delete_root);
 750                 if (q < 0)
 751                         log_warning_errno(q, "Failed to trim compat systemd cgroup %s: %m", path);
 752         }
 753
 754         return r;
 755 }
 756
 757 int cg_create(const char *controller, const char *path) {
 758         _cleanup_free_ char *fs = NULL;
 759         int r;
 760
 761         r = cg_get_path_and_check(controller, path, NULL, &fs);
 762         if (r < 0)
 763                 return r;
 764
 765         r = mkdir_parents(fs, 0755);
 766         if (r < 0)
 767                 return r;
 768
 769         r = mkdir_errno_wrapper(fs, 0755);
 770         if (r == -EEXIST)
 771                 return 0;
 772         if (r < 0)
 773                 return r;
 774
 775         r = cg_hybrid_unified();
 776         if (r < 0)
 777                 return r;
 778
 779         if (r > 0 && streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 780                 r = cg_create(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
 781                 if (r < 0)
 782                         log_warning_errno(r, "Failed to create compat systemd cgroup %s: %m", path);
 783         }
 784
 785         return 1;
 786 }
 787
 788 int cg_create_and_attach(const char *controller, const char *path, pid_t pid) {
 789         int r, q;
 790
 791         assert(pid >= 0);
 792
 793         r = cg_create(controller, path);
 794         if (r < 0)
 795                 return r;
 796
 797         q = cg_attach(controller, path, pid);
 798         if (q < 0)
 799                 return q;
 800
 801         /* This does not remove the cgroup on failure */
 802         return r;
 803 }
 804
 805 int cg_attach(const char *controller, const char *path, pid_t pid) {
 806         _cleanup_free_ char *fs = NULL;
 807         char c[DECIMAL_STR_MAX(pid_t) + 2];
 808         int r;
 809
 810         assert(path);
 811         assert(pid >= 0);
 812
 813         r = cg_get_path_and_check(controller, path, "cgroup.procs", &fs);
 814         if (r < 0)
 815                 return r;
 816
 817         if (pid == 0)
 818                 pid = getpid_cached();
 819
 820         xsprintf(c, PID_FMT "\n", pid);
 821
 822         r = write_string_file(fs, c, 0);
 823         if (r < 0)
 824                 return r;
 825
 826         r = cg_hybrid_unified();
 827         if (r < 0)
 828                 return r;
 829
 830         if (r > 0 && streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 831                 r = cg_attach(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, pid);
 832                 if (r < 0)
 833                         log_warning_errno(r, "Failed to attach "PID_FMT" to compat systemd cgroup %s: %m", pid, path);
 834         }
 835
 836         return 0;
 837 }
 838
 839 int cg_attach_fallback(const char *controller, const char *path, pid_t pid) {
 840         int r;
 841
 842         assert(controller);
 843         assert(path);
 844         assert(pid >= 0);
 845
 846         r = cg_attach(controller, path, pid);
 847         if (r < 0) {
 848                 char prefix[strlen(path) + 1];
 849
 850                 /* This didn't work? Then let's try all prefixes of
 851                  * the destination */
 852
 853                 PATH_FOREACH_PREFIX(prefix, path) {
 854                         int q;
 855
 856                         q = cg_attach(controller, prefix, pid);
 857                         if (q >= 0)
 858                                 return q;
 859                 }
 860         }
 861
 862         return r;
 863 }
 864
 865 int cg_set_access(
 866                 const char *controller,
 867                 const char *path,
 868                 uid_t uid,
 869                 gid_t gid) {
 870
 871         struct Attribute {
 872                 const char *name;
 873                 bool fatal;
 874         };
 875
 876         /* cgroupsv1, aka legacy/non-unified */
 877         static const struct Attribute legacy_attributes[] = {
 878                 { "cgroup.procs",           true  },
 879                 { "tasks",                  false },
 880                 { "cgroup.clone_children",  false },
 881                 {},
 882         };
 883
 884         /* cgroupsv2, aka unified */
 885         static const struct Attribute unified_attributes[] = {
 886                 { "cgroup.procs",           true  },
 887                 { "cgroup.subtree_control", true  },
 888                 { "cgroup.threads",         false },
 889                 {},
 890         };
 891
 892         static const struct Attribute* const attributes[] = {
 893                 [false] = legacy_attributes,
 894                 [true]  = unified_attributes,
 895         };
 896
 897         _cleanup_free_ char *fs = NULL;
 898         const struct Attribute *i;
 899         int r, unified;
 900
 901         assert(path);
 902
 903         if (uid == UID_INVALID && gid == GID_INVALID)
 904                 return 0;
 905
 906         unified = cg_unified_controller(controller);
 907         if (unified < 0)
 908                 return unified;
 909
 910         /* Configure access to the cgroup itself */
 911         r = cg_get_path(controller, path, NULL, &fs);
 912         if (r < 0)
 913                 return r;
 914
 915         r = chmod_and_chown(fs, 0755, uid, gid);
 916         if (r < 0)
 917                 return r;
 918
 919         /* Configure access to the cgroup's attributes */
 920         for (i = attributes[unified]; i->name; i++) {
 921                 fs = mfree(fs);
 922
 923                 r = cg_get_path(controller, path, i->name, &fs);
 924                 if (r < 0)
 925                         return r;
 926
 927                 r = chmod_and_chown(fs, 0644, uid, gid);
 928                 if (r < 0) {
 929                         if (i->fatal)
 930                                 return r;
 931
 932                         log_debug_errno(r, "Failed to set access on cgroup %s, ignoring: %m", fs);
 933                 }
 934         }
 935
 936         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 937                 r = cg_hybrid_unified();
 938                 if (r < 0)
 939                         return r;
 940                 if (r > 0) {
 941                         /* Always propagate access mode from unified to legacy controller */
 942                         r = cg_set_access(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, uid, gid);
 943                         if (r < 0)
 944                                 log_debug_errno(r, "Failed to set access on compatibility systemd cgroup %s, ignoring: %m", path);
 945                 }
 946         }
 947
 948         return 0;
 949 }
 950
 951 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags) {
 952         _cleanup_free_ char *fs = NULL;
 953         int r;
 954
 955         assert(path);
 956         assert(name);
 957         assert(value || size <= 0);
 958
 959         r = cg_get_path(controller, path, NULL, &fs);
 960         if (r < 0)
 961                 return r;
 962
 963         if (setxattr(fs, name, value, size, flags) < 0)
 964                 return -errno;
 965
 966         return 0;
 967 }
 968
 969 int cg_get_xattr(const char *controller, const char *path, const char *name, void *value, size_t size) {
 970         _cleanup_free_ char *fs = NULL;
 971         ssize_t n;
 972         int r;
 973
 974         assert(path);
 975         assert(name);
 976
 977         r = cg_get_path(controller, path, NULL, &fs);
 978         if (r < 0)
 979                 return r;
 980
 981         n = getxattr(fs, name, value, size);
 982         if (n < 0)
 983                 return -errno;
 984
 985         return (int) n;
 986 }
 987
 988 int cg_pid_get_path(const char *controller, pid_t pid, char **path) {
 989         _cleanup_fclose_ FILE *f = NULL;
 990         char line[LINE_MAX];
 991         const char *fs, *controller_str;
 992         size_t cs = 0;
 993         int unified;
 994
 995         assert(path);
 996         assert(pid >= 0);
 997
 998         if (controller) {
 999                 if (!cg_controller_is_valid(controller))
1000                         return -EINVAL;
1001         } else
1002                 controller = SYSTEMD_CGROUP_CONTROLLER;
1003
1004         unified = cg_unified_controller(controller);
1005         if (unified < 0)
1006                 return unified;
1007         if (unified == 0) {
1008                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
1009                         controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
1010                 else
1011                         controller_str = controller;
1012
1013                 cs = strlen(controller_str);
1014         }
1015
1016         fs = procfs_file_alloca(pid, "cgroup");
1017         f = fopen(fs, "re");
1018         if (!f)
1019                 return errno == ENOENT ? -ESRCH : -errno;
1020
1021         (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
1022
1023         FOREACH_LINE(line, f, return -errno) {
1024                 char *e, *p;
1025
1026                 truncate_nl(line);
1027
1028                 if (unified) {
1029                         e = startswith(line, "0:");
1030                         if (!e)
1031                                 continue;
1032
1033                         e = strchr(e, ':');
1034                         if (!e)
1035                                 continue;
1036                 } else {
1037                         char *l;
1038                         size_t k;
1039                         const char *word, *state;
1040                         bool found = false;
1041
1042                         l = strchr(line, ':');
1043                         if (!l)
1044                                 continue;
1045
1046                         l++;
1047                         e = strchr(l, ':');
1048                         if (!e)
1049                                 continue;
1050
1051                         *e = 0;
1052                         FOREACH_WORD_SEPARATOR(word, k, l, ",", state)
1053                                 if (k == cs && memcmp(word, controller_str, cs) == 0) {
1054                                         found = true;
1055                                         break;
1056                                 }
1057                         if (!found)
1058                                 continue;
1059                 }
1060
1061                 p = strdup(e + 1);
1062                 if (!p)
1063                         return -ENOMEM;
1064
1065                 /* Truncate suffix indicating the process is a zombie */
1066                 e = endswith(p, " (deleted)");
1067                 if (e)
1068                         *e = 0;
1069
1070                 *path = p;
1071                 return 0;
1072         }
1073
1074         return -ENODATA;
1075 }
1076
1077 int cg_install_release_agent(const char *controller, const char *agent) {
1078         _cleanup_free_ char *fs = NULL, *contents = NULL;
1079         const char *sc;
1080         int r;
1081
1082         assert(agent);
1083
1084         r = cg_unified_controller(controller);
1085         if (r < 0)
1086                 return r;
1087         if (r > 0) /* doesn't apply to unified hierarchy */
1088                 return -EOPNOTSUPP;
1089
1090         r = cg_get_path(controller, NULL, "release_agent", &fs);
1091         if (r < 0)
1092                 return r;
1093
1094         r = read_one_line_file(fs, &contents);
1095         if (r < 0)
1096                 return r;
1097
1098         sc = strstrip(contents);
1099         if (isempty(sc)) {
1100                 r = write_string_file(fs, agent, 0);
1101                 if (r < 0)
1102                         return r;
1103         } else if (!path_equal(sc, agent))
1104                 return -EEXIST;
1105
1106         fs = mfree(fs);
1107         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
1108         if (r < 0)
1109                 return r;
1110
1111         contents = mfree(contents);
1112         r = read_one_line_file(fs, &contents);
1113         if (r < 0)
1114                 return r;
1115
1116         sc = strstrip(contents);
1117         if (streq(sc, "0")) {
1118                 r = write_string_file(fs, "1", 0);
1119                 if (r < 0)
1120                         return r;
1121
1122                 return 1;
1123         }
1124
1125         if (!streq(sc, "1"))
1126                 return -EIO;
1127
1128         return 0;
1129 }
1130
1131 int cg_uninstall_release_agent(const char *controller) {
1132         _cleanup_free_ char *fs = NULL;
1133         int r;
1134
1135         r = cg_unified_controller(controller);
1136         if (r < 0)
1137                 return r;
1138         if (r > 0) /* Doesn't apply to unified hierarchy */
1139                 return -EOPNOTSUPP;
1140
1141         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
1142         if (r < 0)
1143                 return r;
1144
1145         r = write_string_file(fs, "0", 0);
1146         if (r < 0)
1147                 return r;
1148
1149         fs = mfree(fs);
1150
1151         r = cg_get_path(controller, NULL, "release_agent", &fs);
1152         if (r < 0)
1153                 return r;
1154
1155         r = write_string_file(fs, "", 0);
1156         if (r < 0)
1157                 return r;
1158
1159         return 0;
1160 }
1161
1162 int cg_is_empty(const char *controller, const char *path) {
1163         _cleanup_fclose_ FILE *f = NULL;
1164         pid_t pid;
1165         int r;
1166
1167         assert(path);
1168
1169         r = cg_enumerate_processes(controller, path, &f);
1170         if (r == -ENOENT)
1171                 return 1;
1172         if (r < 0)
1173                 return r;
1174
1175         r = cg_read_pid(f, &pid);
1176         if (r < 0)
1177                 return r;
1178
1179         return r == 0;
1180 }
1181
1182 int cg_is_empty_recursive(const char *controller, const char *path) {
1183         int r;
1184
1185         assert(path);
1186
1187         /* The root cgroup is always populated */
1188         if (controller && (isempty(path) || path_equal(path, "/")))
1189                 return false;
1190
1191         r = cg_unified_controller(controller);
1192         if (r < 0)
1193                 return r;
1194         if (r > 0) {
1195                 _cleanup_free_ char *t = NULL;
1196
1197                 /* On the unified hierarchy we can check empty state
1198                  * via the "populated" attribute of "cgroup.events". */
1199
1200                 r = cg_read_event(controller, path, "populated", &t);
1201                 if (r < 0)
1202                         return r;
1203
1204                 return streq(t, "0");
1205         } else {
1206                 _cleanup_closedir_ DIR *d = NULL;
1207                 char *fn;
1208
1209                 r = cg_is_empty(controller, path);
1210                 if (r <= 0)
1211                         return r;
1212
1213                 r = cg_enumerate_subgroups(controller, path, &d);
1214                 if (r == -ENOENT)
1215                         return 1;
1216                 if (r < 0)
1217                         return r;
1218
1219                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1220                         _cleanup_free_ char *p = NULL;
1221
1222                         p = strjoin(path, "/", fn);
1223                         free(fn);
1224                         if (!p)
1225                                 return -ENOMEM;
1226
1227                         r = cg_is_empty_recursive(controller, p);
1228                         if (r <= 0)
1229                                 return r;
1230                 }
1231                 if (r < 0)
1232                         return r;
1233
1234                 return true;
1235         }
1236 }
1237
1238 int cg_split_spec(const char *spec, char **controller, char **path) {
1239         char *t = NULL, *u = NULL;
1240         const char *e;
1241
1242         assert(spec);
1243
1244         if (*spec == '/') {
1245                 if (!path_is_normalized(spec))
1246                         return -EINVAL;
1247
1248                 if (path) {
1249                         t = strdup(spec);
1250                         if (!t)
1251                                 return -ENOMEM;
1252
1253                         *path = path_kill_slashes(t);
1254                 }
1255
1256                 if (controller)
1257                         *controller = NULL;
1258
1259                 return 0;
1260         }
1261
1262         e = strchr(spec, ':');
1263         if (!e) {
1264                 if (!cg_controller_is_valid(spec))
1265                         return -EINVAL;
1266
1267                 if (controller) {
1268                         t = strdup(spec);
1269                         if (!t)
1270                                 return -ENOMEM;
1271
1272                         *controller = t;
1273                 }
1274
1275                 if (path)
1276                         *path = NULL;
1277
1278                 return 0;
1279         }
1280
1281         t = strndup(spec, e-spec);
1282         if (!t)
1283                 return -ENOMEM;
1284         if (!cg_controller_is_valid(t)) {
1285                 free(t);
1286                 return -EINVAL;
1287         }
1288
1289         if (isempty(e+1))
1290                 u = NULL;
1291         else {
1292                 u = strdup(e+1);
1293                 if (!u) {
1294                         free(t);
1295                         return -ENOMEM;
1296                 }
1297
1298                 if (!path_is_normalized(u) ||
1299                     !path_is_absolute(u)) {
1300                         free(t);
1301                         free(u);
1302                         return -EINVAL;
1303                 }
1304
1305                 path_kill_slashes(u);
1306         }
1307
1308         if (controller)
1309                 *controller = t;
1310         else
1311                 free(t);
1312
1313         if (path)
1314                 *path = u;
1315         else
1316                 free(u);
1317
1318         return 0;
1319 }
1320
1321 int cg_mangle_path(const char *path, char **result) {
1322         _cleanup_free_ char *c = NULL, *p = NULL;
1323         char *t;
1324         int r;
1325
1326         assert(path);
1327         assert(result);
1328
1329         /* First, check if it already is a filesystem path */
1330         if (path_startswith(path, "/sys/fs/cgroup")) {
1331
1332                 t = strdup(path);
1333                 if (!t)
1334                         return -ENOMEM;
1335
1336                 *result = path_kill_slashes(t);
1337                 return 0;
1338         }
1339
1340         /* Otherwise, treat it as cg spec */
1341         r = cg_split_spec(path, &c, &p);
1342         if (r < 0)
1343                 return r;
1344
1345         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
1346 }
1347
1348 int cg_get_root_path(char **path) {
1349         char *p, *e;
1350         int r;
1351
1352         assert(path);
1353
1354         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1355         if (r < 0)
1356                 return r;
1357
1358         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1359         if (!e)
1360                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1361         if (!e)
1362                 e = endswith(p, "/system"); /* even more legacy */
1363         if (e)
1364                 *e = 0;
1365
1366         *path = p;
1367         return 0;
1368 }
1369
1370 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1371         _cleanup_free_ char *rt = NULL;
1372         char *p;
1373         int r;
1374
1375         assert(cgroup);
1376         assert(shifted);
1377
1378         if (!root) {
1379                 /* If the root was specified let's use that, otherwise
1380                  * let's determine it from PID 1 */
1381
1382                 r = cg_get_root_path(&rt);
1383                 if (r < 0)
1384                         return r;
1385
1386                 root = rt;
1387         }
1388
1389         p = path_startswith(cgroup, root);
1390         if (p && p > cgroup)
1391                 *shifted = p - 1;
1392         else
1393                 *shifted = cgroup;
1394
1395         return 0;
1396 }
1397
1398 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1399         _cleanup_free_ char *raw = NULL;
1400         const char *c;
1401         int r;
1402
1403         assert(pid >= 0);
1404         assert(cgroup);
1405
1406         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1407         if (r < 0)
1408                 return r;
1409
1410         r = cg_shift_path(raw, root, &c);
1411         if (r < 0)
1412                 return r;
1413
1414         if (c == raw)
1415                 *cgroup = TAKE_PTR(raw);
1416         else {
1417                 char *n;
1418
1419                 n = strdup(c);
1420                 if (!n)
1421                         return -ENOMEM;
1422
1423                 *cgroup = n;
1424         }
1425
1426         return 0;
1427 }
1428
1429 int cg_path_decode_unit(const char *cgroup, char **unit) {
1430         char *c, *s;
1431         size_t n;
1432
1433         assert(cgroup);
1434         assert(unit);
1435
1436         n = strcspn(cgroup, "/");
1437         if (n < 3)
1438                 return -ENXIO;
1439
1440         c = strndupa(cgroup, n);
1441         c = cg_unescape(c);
1442
1443         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1444                 return -ENXIO;
1445
1446         s = strdup(c);
1447         if (!s)
1448                 return -ENOMEM;
1449
1450         *unit = s;
1451         return 0;
1452 }
1453
1454 static bool valid_slice_name(const char *p, size_t n) {
1455
1456         if (!p)
1457                 return false;
1458
1459         if (n < STRLEN("x.slice"))
1460                 return false;
1461
1462         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1463                 char buf[n+1], *c;
1464
1465                 memcpy(buf, p, n);
1466                 buf[n] = 0;
1467
1468                 c = cg_unescape(buf);
1469
1470                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1471         }
1472
1473         return false;
1474 }
1475
1476 static const char *skip_slices(const char *p) {
1477         assert(p);
1478
1479         /* Skips over all slice assignments */
1480
1481         for (;;) {
1482                 size_t n;
1483
1484                 p += strspn(p, "/");
1485
1486                 n = strcspn(p, "/");
1487                 if (!valid_slice_name(p, n))
1488                         return p;
1489
1490                 p += n;
1491         }
1492 }
1493
1494 int cg_path_get_unit(const char *path, char **ret) {
1495         const char *e;
1496         char *unit;
1497         int r;
1498
1499         assert(path);
1500         assert(ret);
1501
1502         e = skip_slices(path);
1503
1504         r = cg_path_decode_unit(e, &unit);
1505         if (r < 0)
1506                 return r;
1507
1508         /* We skipped over the slices, don't accept any now */
1509         if (endswith(unit, ".slice")) {
1510                 free(unit);
1511                 return -ENXIO;
1512         }
1513
1514         *ret = unit;
1515         return 0;
1516 }
1517
1518 int cg_pid_get_unit(pid_t pid, char **unit) {
1519         _cleanup_free_ char *cgroup = NULL;
1520         int r;
1521
1522         assert(unit);
1523
1524         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1525         if (r < 0)
1526                 return r;
1527
1528         return cg_path_get_unit(cgroup, unit);
1529 }
1530
1531 /**
1532  * Skip session-*.scope, but require it to be there.
1533  */
1534 static const char *skip_session(const char *p) {
1535         size_t n;
1536
1537         if (isempty(p))
1538                 return NULL;
1539
1540         p += strspn(p, "/");
1541
1542         n = strcspn(p, "/");
1543         if (n < STRLEN("session-x.scope"))
1544                 return NULL;
1545
1546         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1547                 char buf[n - 8 - 6 + 1];
1548
1549                 memcpy(buf, p + 8, n - 8 - 6);
1550                 buf[n - 8 - 6] = 0;
1551
1552                 /* Note that session scopes never need unescaping,
1553                  * since they cannot conflict with the kernel's own
1554                  * names, hence we don't need to call cg_unescape()
1555                  * here. */
1556
1557                 if (!session_id_valid(buf))
1558                         return false;
1559
1560                 p += n;
1561                 p += strspn(p, "/");
1562                 return p;
1563         }
1564
1565         return NULL;
1566 }
1567
1568 /**
1569  * Skip user@*.service, but require it to be there.
1570  */
1571 static const char *skip_user_manager(const char *p) {
1572         size_t n;
1573
1574         if (isempty(p))
1575                 return NULL;
1576
1577         p += strspn(p, "/");
1578
1579         n = strcspn(p, "/");
1580         if (n < STRLEN("user@x.service"))
1581                 return NULL;
1582
1583         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1584                 char buf[n - 5 - 8 + 1];
1585
1586                 memcpy(buf, p + 5, n - 5 - 8);
1587                 buf[n - 5 - 8] = 0;
1588
1589                 /* Note that user manager services never need unescaping,
1590                  * since they cannot conflict with the kernel's own
1591                  * names, hence we don't need to call cg_unescape()
1592                  * here. */
1593
1594                 if (parse_uid(buf, NULL) < 0)
1595                         return NULL;
1596
1597                 p += n;
1598                 p += strspn(p, "/");
1599
1600                 return p;
1601         }
1602
1603         return NULL;
1604 }
1605
1606 static const char *skip_user_prefix(const char *path) {
1607         const char *e, *t;
1608
1609         assert(path);
1610
1611         /* Skip slices, if there are any */
1612         e = skip_slices(path);
1613
1614         /* Skip the user manager, if it's in the path now... */
1615         t = skip_user_manager(e);
1616         if (t)
1617                 return t;
1618
1619         /* Alternatively skip the user session if it is in the path... */
1620         return skip_session(e);
1621 }
1622
1623 int cg_path_get_user_unit(const char *path, char **ret) {
1624         const char *t;
1625
1626         assert(path);
1627         assert(ret);
1628
1629         t = skip_user_prefix(path);
1630         if (!t)
1631                 return -ENXIO;
1632
1633         /* And from here on it looks pretty much the same as for a
1634          * system unit, hence let's use the same parser from here
1635          * on. */
1636         return cg_path_get_unit(t, ret);
1637 }
1638
1639 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1640         _cleanup_free_ char *cgroup = NULL;
1641         int r;
1642
1643         assert(unit);
1644
1645         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1646         if (r < 0)
1647                 return r;
1648
1649         return cg_path_get_user_unit(cgroup, unit);
1650 }
1651
1652 int cg_path_get_machine_name(const char *path, char **machine) {
1653         _cleanup_free_ char *u = NULL;
1654         const char *sl;
1655         int r;
1656
1657         r = cg_path_get_unit(path, &u);
1658         if (r < 0)
1659                 return r;
1660
1661         sl = strjoina("/run/systemd/machines/unit:", u);
1662         return readlink_malloc(sl, machine);
1663 }
1664
1665 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1666         _cleanup_free_ char *cgroup = NULL;
1667         int r;
1668
1669         assert(machine);
1670
1671         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1672         if (r < 0)
1673                 return r;
1674
1675         return cg_path_get_machine_name(cgroup, machine);
1676 }
1677
1678 int cg_path_get_session(const char *path, char **session) {
1679         _cleanup_free_ char *unit = NULL;
1680         char *start, *end;
1681         int r;
1682
1683         assert(path);
1684
1685         r = cg_path_get_unit(path, &unit);
1686         if (r < 0)
1687                 return r;
1688
1689         start = startswith(unit, "session-");
1690         if (!start)
1691                 return -ENXIO;
1692         end = endswith(start, ".scope");
1693         if (!end)
1694                 return -ENXIO;
1695
1696         *end = 0;
1697         if (!session_id_valid(start))
1698                 return -ENXIO;
1699
1700         if (session) {
1701                 char *rr;
1702
1703                 rr = strdup(start);
1704                 if (!rr)
1705                         return -ENOMEM;
1706
1707                 *session = rr;
1708         }
1709
1710         return 0;
1711 }
1712
1713 int cg_pid_get_session(pid_t pid, char **session) {
1714         _cleanup_free_ char *cgroup = NULL;
1715         int r;
1716
1717         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1718         if (r < 0)
1719                 return r;
1720
1721         return cg_path_get_session(cgroup, session);
1722 }
1723
1724 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1725         _cleanup_free_ char *slice = NULL;
1726         char *start, *end;
1727         int r;
1728
1729         assert(path);
1730
1731         r = cg_path_get_slice(path, &slice);
1732         if (r < 0)
1733                 return r;
1734
1735         start = startswith(slice, "user-");
1736         if (!start)
1737                 return -ENXIO;
1738         end = endswith(start, ".slice");
1739         if (!end)
1740                 return -ENXIO;
1741
1742         *end = 0;
1743         if (parse_uid(start, uid) < 0)
1744                 return -ENXIO;
1745
1746         return 0;
1747 }
1748
1749 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1750         _cleanup_free_ char *cgroup = NULL;
1751         int r;
1752
1753         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1754         if (r < 0)
1755                 return r;
1756
1757         return cg_path_get_owner_uid(cgroup, uid);
1758 }
1759
1760 int cg_path_get_slice(const char *p, char **slice) {
1761         const char *e = NULL;
1762
1763         assert(p);
1764         assert(slice);
1765
1766         /* Finds the right-most slice unit from the beginning, but
1767          * stops before we come to the first non-slice unit. */
1768
1769         for (;;) {
1770                 size_t n;
1771
1772                 p += strspn(p, "/");
1773
1774                 n = strcspn(p, "/");
1775                 if (!valid_slice_name(p, n)) {
1776
1777                         if (!e) {
1778                                 char *s;
1779
1780                                 s = strdup(SPECIAL_ROOT_SLICE);
1781                                 if (!s)
1782                                         return -ENOMEM;
1783
1784                                 *slice = s;
1785                                 return 0;
1786                         }
1787
1788                         return cg_path_decode_unit(e, slice);
1789                 }
1790
1791                 e = p;
1792                 p += n;
1793         }
1794 }
1795
1796 int cg_pid_get_slice(pid_t pid, char **slice) {
1797         _cleanup_free_ char *cgroup = NULL;
1798         int r;
1799
1800         assert(slice);
1801
1802         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1803         if (r < 0)
1804                 return r;
1805
1806         return cg_path_get_slice(cgroup, slice);
1807 }
1808
1809 int cg_path_get_user_slice(const char *p, char **slice) {
1810         const char *t;
1811         assert(p);
1812         assert(slice);
1813
1814         t = skip_user_prefix(p);
1815         if (!t)
1816                 return -ENXIO;
1817
1818         /* And now it looks pretty much the same as for a system
1819          * slice, so let's just use the same parser from here on. */
1820         return cg_path_get_slice(t, slice);
1821 }
1822
1823 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1824         _cleanup_free_ char *cgroup = NULL;
1825         int r;
1826
1827         assert(slice);
1828
1829         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1830         if (r < 0)
1831                 return r;
1832
1833         return cg_path_get_user_slice(cgroup, slice);
1834 }
1835
1836 char *cg_escape(const char *p) {
1837         bool need_prefix = false;
1838
1839         /* This implements very minimal escaping for names to be used
1840          * as file names in the cgroup tree: any name which might
1841          * conflict with a kernel name or is prefixed with '_' is
1842          * prefixed with a '_'. That way, when reading cgroup names it
1843          * is sufficient to remove a single prefixing underscore if
1844          * there is one. */
1845
1846         /* The return value of this function (unlike cg_unescape())
1847          * needs free()! */
1848
1849         if (IN_SET(p[0], 0, '_', '.') ||
1850             streq(p, "notify_on_release") ||
1851             streq(p, "release_agent") ||
1852             streq(p, "tasks") ||
1853             startswith(p, "cgroup."))
1854                 need_prefix = true;
1855         else {
1856                 const char *dot;
1857
1858                 dot = strrchr(p, '.');
1859                 if (dot) {
1860                         CGroupController c;
1861                         size_t l = dot - p;
1862
1863                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1864                                 const char *n;
1865
1866                                 n = cgroup_controller_to_string(c);
1867
1868                                 if (l != strlen(n))
1869                                         continue;
1870
1871                                 if (memcmp(p, n, l) != 0)
1872                                         continue;
1873
1874                                 need_prefix = true;
1875                                 break;
1876                         }
1877                 }
1878         }
1879
1880         if (need_prefix)
1881                 return strappend("_", p);
1882
1883         return strdup(p);
1884 }
1885
1886 char *cg_unescape(const char *p) {
1887         assert(p);
1888
1889         /* The return value of this function (unlike cg_escape())
1890          * doesn't need free()! */
1891
1892         if (p[0] == '_')
1893                 return (char*) p+1;
1894
1895         return (char*) p;
1896 }
1897
1898 #define CONTROLLER_VALID                        \
1899         DIGITS LETTERS                          \
1900         "_"
1901
1902 bool cg_controller_is_valid(const char *p) {
1903         const char *t, *s;
1904
1905         if (!p)
1906                 return false;
1907
1908         if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1909                 return true;
1910
1911         s = startswith(p, "name=");
1912         if (s)
1913                 p = s;
1914
1915         if (IN_SET(*p, 0, '_'))
1916                 return false;
1917
1918         for (t = p; *t; t++)
1919                 if (!strchr(CONTROLLER_VALID, *t))
1920                         return false;
1921
1922         if (t - p > FILENAME_MAX)
1923                 return false;
1924
1925         return true;
1926 }
1927
1928 int cg_slice_to_path(const char *unit, char **ret) {
1929         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1930         const char *dash;
1931         int r;
1932
1933         assert(unit);
1934         assert(ret);
1935
1936         if (streq(unit, SPECIAL_ROOT_SLICE)) {
1937                 char *x;
1938
1939                 x = strdup("");
1940                 if (!x)
1941                         return -ENOMEM;
1942                 *ret = x;
1943                 return 0;
1944         }
1945
1946         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1947                 return -EINVAL;
1948
1949         if (!endswith(unit, ".slice"))
1950                 return -EINVAL;
1951
1952         r = unit_name_to_prefix(unit, &p);
1953         if (r < 0)
1954                 return r;
1955
1956         dash = strchr(p, '-');
1957
1958         /* Don't allow initial dashes */
1959         if (dash == p)
1960                 return -EINVAL;
1961
1962         while (dash) {
1963                 _cleanup_free_ char *escaped = NULL;
1964                 char n[dash - p + sizeof(".slice")];
1965
1966 #if HAS_FEATURE_MEMORY_SANITIZER
1967                 /* msan doesn't instrument stpncpy, so it thinks
1968                  * n is later used unitialized:
1969                  * https://github.com/google/sanitizers/issues/926
1970                  */
1971                 zero(n);
1972 #endif
1973
1974                 /* Don't allow trailing or double dashes */
1975                 if (IN_SET(dash[1], 0, '-'))
1976                         return -EINVAL;
1977
1978                 strcpy(stpncpy(n, p, dash - p), ".slice");
1979                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1980                         return -EINVAL;
1981
1982                 escaped = cg_escape(n);
1983                 if (!escaped)
1984                         return -ENOMEM;
1985
1986                 if (!strextend(&s, escaped, "/", NULL))
1987                         return -ENOMEM;
1988
1989                 dash = strchr(dash+1, '-');
1990         }
1991
1992         e = cg_escape(unit);
1993         if (!e)
1994                 return -ENOMEM;
1995
1996         if (!strextend(&s, e, NULL))
1997                 return -ENOMEM;
1998
1999         *ret = TAKE_PTR(s);
2000
2001         return 0;
2002 }
2003
2004 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
2005         _cleanup_free_ char *p = NULL;
2006         int r;
2007
2008         r = cg_get_path(controller, path, attribute, &p);
2009         if (r < 0)
2010                 return r;
2011
2012         return write_string_file(p, value, 0);
2013 }
2014
2015 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
2016         _cleanup_free_ char *p = NULL;
2017         int r;
2018
2019         r = cg_get_path(controller, path, attribute, &p);
2020         if (r < 0)
2021                 return r;
2022
2023         return read_one_line_file(p, ret);
2024 }
2025
2026 int cg_get_keyed_attribute(
2027                 const char *controller,
2028                 const char *path,
2029                 const char *attribute,
2030                 char **keys,
2031                 char **ret_values) {
2032
2033         _cleanup_free_ char *filename = NULL, *contents = NULL;
2034         const char *p;
2035         size_t n, i, n_done = 0;
2036         char **v;
2037         int r;
2038
2039         /* Reads one or more fields of a cgroupsv2 keyed attribute file. The 'keys' parameter should be an strv with
2040          * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
2041          * entries as 'keys'. On success each entry will be set to the value of the matching key.
2042          *
2043          * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. */
2044
2045         r = cg_get_path(controller, path, attribute, &filename);
2046         if (r < 0)
2047                 return r;
2048
2049         r = read_full_file(filename, &contents, NULL);
2050         if (r < 0)
2051                 return r;
2052
2053         n = strv_length(keys);
2054         if (n == 0) /* No keys to retrieve? That's easy, we are done then */
2055                 return 0;
2056
2057         /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
2058         v = newa0(char*, n);
2059
2060         for (p = contents; *p;) {
2061                 const char *w = NULL;
2062
2063                 for (i = 0; i < n; i++)
2064                         if (!v[i]) {
2065                                 w = first_word(p, keys[i]);
2066                                 if (w)
2067                                         break;
2068                         }
2069
2070                 if (w) {
2071                         size_t l;
2072
2073                         l = strcspn(w, NEWLINE);
2074                         v[i] = strndup(w, l);
2075                         if (!v[i]) {
2076                                 r = -ENOMEM;
2077                                 goto fail;
2078                         }
2079
2080                         n_done++;
2081                         if (n_done >= n)
2082                                 goto done;
2083
2084                         p = w + l;
2085                 } else
2086                         p += strcspn(p, NEWLINE);
2087
2088                 p += strspn(p, NEWLINE);
2089         }
2090
2091         r = -ENXIO;
2092
2093 fail:
2094         for (i = 0; i < n; i++)
2095                 free(v[i]);
2096
2097         return r;
2098
2099 done:
2100         memcpy(ret_values, v, sizeof(char*) * n);
2101         return 0;
2102
2103 }
2104
2105 int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) {
2106         CGroupController c;
2107         int r;
2108
2109         /* This one will create a cgroup in our private tree, but also
2110          * duplicate it in the trees specified in mask, and remove it
2111          * in all others */
2112
2113         /* First create the cgroup in our own hierarchy. */
2114         r = cg_create(SYSTEMD_CGROUP_CONTROLLER, path);
2115         if (r < 0)
2116                 return r;
2117
2118         /* If we are in the unified hierarchy, we are done now */
2119         r = cg_all_unified();
2120         if (r < 0)
2121                 return r;
2122         if (r > 0)
2123                 return 0;
2124
2125         /* Otherwise, do the same in the other hierarchies */
2126         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2127                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2128                 const char *n;
2129
2130                 n = cgroup_controller_to_string(c);
2131
2132                 if (mask & bit)
2133                         (void) cg_create(n, path);
2134                 else if (supported & bit)
2135                         (void) cg_trim(n, path, true);
2136         }
2137
2138         return 0;
2139 }
2140
2141 int cg_attach_everywhere(CGroupMask supported, const char *path, pid_t pid, cg_migrate_callback_t path_callback, void *userdata) {
2142         CGroupController c;
2143         int r;
2144
2145         r = cg_attach(SYSTEMD_CGROUP_CONTROLLER, path, pid);
2146         if (r < 0)
2147                 return r;
2148
2149         r = cg_all_unified();
2150         if (r < 0)
2151                 return r;
2152         if (r > 0)
2153                 return 0;
2154
2155         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2156                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2157                 const char *p = NULL;
2158
2159                 if (!(supported & bit))
2160                         continue;
2161
2162                 if (path_callback)
2163                         p = path_callback(bit, userdata);
2164
2165                 if (!p)
2166                         p = path;
2167
2168                 (void) cg_attach_fallback(cgroup_controller_to_string(c), p, pid);
2169         }
2170
2171         return 0;
2172 }
2173
2174 int cg_attach_many_everywhere(CGroupMask supported, const char *path, Set* pids, cg_migrate_callback_t path_callback, void *userdata) {
2175         Iterator i;
2176         void *pidp;
2177         int r = 0;
2178
2179         SET_FOREACH(pidp, pids, i) {
2180                 pid_t pid = PTR_TO_PID(pidp);
2181                 int q;
2182
2183                 q = cg_attach_everywhere(supported, path, pid, path_callback, userdata);
2184                 if (q < 0 && r >= 0)
2185                         r = q;
2186         }
2187
2188         return r;
2189 }
2190
2191 int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to, cg_migrate_callback_t to_callback, void *userdata) {
2192         CGroupController c;
2193         int r = 0, q;
2194
2195         if (!path_equal(from, to))  {
2196                 r = cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER, from, SYSTEMD_CGROUP_CONTROLLER, to, CGROUP_REMOVE);
2197                 if (r < 0)
2198                         return r;
2199         }
2200
2201         q = cg_all_unified();
2202         if (q < 0)
2203                 return q;
2204         if (q > 0)
2205                 return r;
2206
2207         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2208                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2209                 const char *p = NULL;
2210
2211                 if (!(supported & bit))
2212                         continue;
2213
2214                 if (to_callback)
2215                         p = to_callback(bit, userdata);
2216
2217                 if (!p)
2218                         p = to;
2219
2220                 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER, to, cgroup_controller_to_string(c), p, 0);
2221         }
2222
2223         return 0;
2224 }
2225
2226 int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) {
2227         CGroupController c;
2228         int r, q;
2229
2230         r = cg_trim(SYSTEMD_CGROUP_CONTROLLER, path, delete_root);
2231         if (r < 0)
2232                 return r;
2233
2234         q = cg_all_unified();
2235         if (q < 0)
2236                 return q;
2237         if (q > 0)
2238                 return r;
2239
2240         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2241                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2242
2243                 if (!(supported & bit))
2244                         continue;
2245
2246                 (void) cg_trim(cgroup_controller_to_string(c), path, delete_root);
2247         }
2248
2249         return 0;
2250 }
2251
2252 int cg_mask_to_string(CGroupMask mask, char **ret) {
2253         _cleanup_free_ char *s = NULL;
2254         size_t n = 0, allocated = 0;
2255         bool space = false;
2256         CGroupController c;
2257
2258         assert(ret);
2259
2260         if (mask == 0) {
2261                 *ret = NULL;
2262                 return 0;
2263         }
2264
2265         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2266                 const char *k;
2267                 size_t l;
2268
2269                 if (!(mask & CGROUP_CONTROLLER_TO_MASK(c)))
2270                         continue;
2271
2272                 k = cgroup_controller_to_string(c);
2273                 l = strlen(k);
2274
2275                 if (!GREEDY_REALLOC(s, allocated, n + space + l + 1))
2276                         return -ENOMEM;
2277
2278                 if (space)
2279                         s[n] = ' ';
2280                 memcpy(s + n + space, k, l);
2281                 n += space + l;
2282
2283                 space = true;
2284         }
2285
2286         assert(s);
2287
2288         s[n] = 0;
2289         *ret = TAKE_PTR(s);
2290
2291         return 0;
2292 }
2293
2294 int cg_mask_from_string(const char *value, CGroupMask *mask) {
2295         assert(mask);
2296         assert(value);
2297
2298         for (;;) {
2299                 _cleanup_free_ char *n = NULL;
2300                 CGroupController v;
2301                 int r;
2302
2303                 r = extract_first_word(&value, &n, NULL, 0);
2304                 if (r < 0)
2305                         return r;
2306                 if (r == 0)
2307                         break;
2308
2309                 v = cgroup_controller_from_string(n);
2310                 if (v < 0)
2311                         continue;
2312
2313                 *mask |= CGROUP_CONTROLLER_TO_MASK(v);
2314         }
2315         return 0;
2316 }
2317
2318 int cg_mask_supported(CGroupMask *ret) {
2319         CGroupMask mask = 0;
2320         int r;
2321
2322         /* Determines the mask of supported cgroup controllers. Only
2323          * includes controllers we can make sense of and that are
2324          * actually accessible. */
2325
2326         r = cg_all_unified();
2327         if (r < 0)
2328                 return r;
2329         if (r > 0) {
2330                 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
2331
2332                 /* In the unified hierarchy we can read the supported
2333                  * and accessible controllers from a the top-level
2334                  * cgroup attribute */
2335
2336                 r = cg_get_root_path(&root);
2337                 if (r < 0)
2338                         return r;
2339
2340                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
2341                 if (r < 0)
2342                         return r;
2343
2344                 r = read_one_line_file(path, &controllers);
2345                 if (r < 0)
2346                         return r;
2347
2348                 r = cg_mask_from_string(controllers, &mask);
2349                 if (r < 0)
2350                         return r;
2351
2352                 /* Currently, we support the cpu, memory, io and pids
2353                  * controller in the unified hierarchy, mask
2354                  * everything else off. */
2355                 mask &= CGROUP_MASK_CPU | CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
2356
2357         } else {
2358                 CGroupController c;
2359
2360                 /* In the legacy hierarchy, we check whether which
2361                  * hierarchies are mounted. */
2362
2363                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2364                         const char *n;
2365
2366                         n = cgroup_controller_to_string(c);
2367                         if (controller_is_accessible(n) >= 0)
2368                                 mask |= CGROUP_CONTROLLER_TO_MASK(c);
2369                 }
2370         }
2371
2372         *ret = mask;
2373         return 0;
2374 }
2375
2376 int cg_kernel_controllers(Set **ret) {
2377         _cleanup_set_free_free_ Set *controllers = NULL;
2378         _cleanup_fclose_ FILE *f = NULL;
2379         int r;
2380
2381         assert(ret);
2382
2383         /* Determines the full list of kernel-known controllers. Might
2384          * include controllers we don't actually support, arbitrary
2385          * named hierarchies and controllers that aren't currently
2386          * accessible (because not mounted). */
2387
2388         controllers = set_new(&string_hash_ops);
2389         if (!controllers)
2390                 return -ENOMEM;
2391
2392         f = fopen("/proc/cgroups", "re");
2393         if (!f) {
2394                 if (errno == ENOENT) {
2395                         *ret = NULL;
2396                         return 0;
2397                 }
2398
2399                 return -errno;
2400         }
2401
2402         (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
2403
2404         /* Ignore the header line */
2405         (void) read_line(f, (size_t) -1, NULL);
2406
2407         for (;;) {
2408                 char *controller;
2409                 int enabled = 0;
2410
2411                 errno = 0;
2412                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
2413
2414                         if (feof(f))
2415                                 break;
2416
2417                         if (ferror(f) && errno > 0)
2418                                 return -errno;
2419
2420                         return -EBADMSG;
2421                 }
2422
2423                 if (!enabled) {
2424                         free(controller);
2425                         continue;
2426                 }
2427
2428                 if (!cg_controller_is_valid(controller)) {
2429                         free(controller);
2430                         return -EBADMSG;
2431                 }
2432
2433                 r = set_consume(controllers, controller);
2434                 if (r < 0)
2435                         return r;
2436         }
2437
2438         *ret = TAKE_PTR(controllers);
2439
2440         return 0;
2441 }
2442
2443 static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2444
2445 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup v2 on /sys/fs/cgroup/systemd.  This
2446  * unfortunately broke other tools (such as docker) which expected the v1 "name=systemd" hierarchy on
2447  * /sys/fs/cgroup/systemd.  From v233 and on, the hybrid mode mountnbs v2 on /sys/fs/cgroup/unified and maintains
2448  * "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility with other tools.
2449  *
2450  * To keep live upgrade working, we detect and support v232 layout.  When v232 layout is detected, to keep cgroup v2
2451  * process management but disable the compat dual layout, we return %true on
2452  * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and %false on cg_hybrid_unified().
2453  */
2454 static thread_local bool unified_systemd_v232;
2455
2456 static int cg_unified_update(void) {
2457
2458         struct statfs fs;
2459
2460         /* Checks if we support the unified hierarchy. Returns an
2461          * error when the cgroup hierarchies aren't mounted yet or we
2462          * have any other trouble determining if the unified hierarchy
2463          * is supported. */
2464
2465         if (unified_cache >= CGROUP_UNIFIED_NONE)
2466                 return 0;
2467
2468         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2469                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2470
2471         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2472                 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2473                 unified_cache = CGROUP_UNIFIED_ALL;
2474         } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2475                 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2476                     F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2477                         log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2478                         unified_cache = CGROUP_UNIFIED_SYSTEMD;
2479                         unified_systemd_v232 = false;
2480                 } else {
2481                         if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0)
2482                                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2483
2484                         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2485                                 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2486                                 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2487                                 unified_systemd_v232 = true;
2488                         } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2489                                 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2490                                 unified_cache = CGROUP_UNIFIED_NONE;
2491                         } else {
2492                                 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2493                                           (unsigned long long) fs.f_type);
2494                                 unified_cache = CGROUP_UNIFIED_NONE;
2495                         }
2496                 }
2497         } else {
2498                 log_debug("Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2499                           (unsigned long long) fs.f_type);
2500                 return -ENOMEDIUM;
2501         }
2502
2503         return 0;
2504 }
2505
2506 int cg_unified_controller(const char *controller) {
2507         int r;
2508
2509         r = cg_unified_update();
2510         if (r < 0)
2511                 return r;
2512
2513         if (unified_cache == CGROUP_UNIFIED_NONE)
2514                 return false;
2515
2516         if (unified_cache >= CGROUP_UNIFIED_ALL)
2517                 return true;
2518
2519         return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2520 }
2521
2522 int cg_all_unified(void) {
2523         int r;
2524
2525         r = cg_unified_update();
2526         if (r < 0)
2527                 return r;
2528
2529         return unified_cache >= CGROUP_UNIFIED_ALL;
2530 }
2531
2532 int cg_hybrid_unified(void) {
2533         int r;
2534
2535         r = cg_unified_update();
2536         if (r < 0)
2537                 return r;
2538
2539         return unified_cache == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2540 }
2541
2542 int cg_unified_flush(void) {
2543         unified_cache = CGROUP_UNIFIED_UNKNOWN;
2544
2545         return cg_unified_update();
2546 }
2547
2548 int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) {
2549         _cleanup_fclose_ FILE *f = NULL;
2550         _cleanup_free_ char *fs = NULL;
2551         CGroupController c;
2552         int r;
2553
2554         assert(p);
2555
2556         if (supported == 0)
2557                 return 0;
2558
2559         r = cg_all_unified();
2560         if (r < 0)
2561                 return r;
2562         if (r == 0) /* on the legacy hiearchy there's no joining of controllers defined */
2563                 return 0;
2564
2565         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, p, "cgroup.subtree_control", &fs);
2566         if (r < 0)
2567                 return r;
2568
2569         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2570                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2571                 const char *n;
2572
2573                 if (!(supported & bit))
2574                         continue;
2575
2576                 n = cgroup_controller_to_string(c);
2577                 {
2578                         char s[1 + strlen(n) + 1];
2579
2580                         s[0] = mask & bit ? '+' : '-';
2581                         strcpy(s + 1, n);
2582
2583                         if (!f) {
2584                                 f = fopen(fs, "we");
2585                                 if (!f) {
2586                                         log_debug_errno(errno, "Failed to open cgroup.subtree_control file of %s: %m", p);
2587                                         break;
2588                                 }
2589                         }
2590
2591                         r = write_string_stream(f, s, 0);
2592                         if (r < 0)
2593                                 log_debug_errno(r, "Failed to enable controller %s for %s (%s): %m", n, p, fs);
2594                 }
2595         }
2596
2597         return 0;
2598 }
2599
2600 bool cg_is_unified_wanted(void) {
2601         static thread_local int wanted = -1;
2602         int r;
2603         bool b;
2604         const bool is_default = DEFAULT_HIERARCHY == CGROUP_UNIFIED_ALL;
2605
2606         /* If we have a cached value, return that. */
2607         if (wanted >= 0)
2608                 return wanted;
2609
2610         /* If the hierarchy is already mounted, then follow whatever
2611          * was chosen for it. */
2612         if (cg_unified_flush() >= 0)
2613                 return (wanted = unified_cache >= CGROUP_UNIFIED_ALL);
2614
2615         /* Otherwise, let's see what the kernel command line has to say.
2616          * Since checking is expensive, cache a non-error result. */
2617         r = proc_cmdline_get_bool("systemd.unified_cgroup_hierarchy", &b);
2618
2619         return (wanted = r > 0 ? b : is_default);
2620 }
2621
2622 bool cg_is_legacy_wanted(void) {
2623         static thread_local int wanted = -1;
2624
2625         /* If we have a cached value, return that. */
2626         if (wanted >= 0)
2627                 return wanted;
2628
2629         /* Check if we have cgroups2 already mounted. */
2630         if (cg_unified_flush() >= 0 &&
2631             unified_cache == CGROUP_UNIFIED_ALL)
2632                 return (wanted = false);
2633
2634         /* Otherwise, assume that at least partial legacy is wanted,
2635          * since cgroups2 should already be mounted at this point. */
2636         return (wanted = true);
2637 }
2638
2639 bool cg_is_hybrid_wanted(void) {
2640         static thread_local int wanted = -1;
2641         int r;
2642         bool b;
2643         const bool is_default = DEFAULT_HIERARCHY >= CGROUP_UNIFIED_SYSTEMD;
2644         /* We default to true if the default is "hybrid", obviously,
2645          * but also when the default is "unified", because if we get
2646          * called, it means that unified hierarchy was not mounted. */
2647
2648         /* If we have a cached value, return that. */
2649         if (wanted >= 0)
2650                 return wanted;
2651
2652         /* If the hierarchy is already mounted, then follow whatever
2653          * was chosen for it. */
2654         if (cg_unified_flush() >= 0 &&
2655             unified_cache == CGROUP_UNIFIED_ALL)
2656                 return (wanted = false);
2657
2658         /* Otherwise, let's see what the kernel command line has to say.
2659          * Since checking is expensive, cache a non-error result. */
2660         r = proc_cmdline_get_bool("systemd.legacy_systemd_cgroup_controller", &b);
2661
2662         /* The meaning of the kernel option is reversed wrt. to the return value
2663          * of this function, hence the negation. */
2664         return (wanted = r > 0 ? !b : is_default);
2665 }
2666
2667 int cg_weight_parse(const char *s, uint64_t *ret) {
2668         uint64_t u;
2669         int r;
2670
2671         if (isempty(s)) {
2672                 *ret = CGROUP_WEIGHT_INVALID;
2673                 return 0;
2674         }
2675
2676         r = safe_atou64(s, &u);
2677         if (r < 0)
2678                 return r;
2679
2680         if (u < CGROUP_WEIGHT_MIN || u > CGROUP_WEIGHT_MAX)
2681                 return -ERANGE;
2682
2683         *ret = u;
2684         return 0;
2685 }
2686
2687 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2688         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2689         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2690         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2691         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2692 };
2693
2694 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2695         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2696         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2697         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2698         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2699 };
2700
2701 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2702
2703 int cg_cpu_shares_parse(const char *s, uint64_t *ret) {
2704         uint64_t u;
2705         int r;
2706
2707         if (isempty(s)) {
2708                 *ret = CGROUP_CPU_SHARES_INVALID;
2709                 return 0;
2710         }
2711
2712         r = safe_atou64(s, &u);
2713         if (r < 0)
2714                 return r;
2715
2716         if (u < CGROUP_CPU_SHARES_MIN || u > CGROUP_CPU_SHARES_MAX)
2717                 return -ERANGE;
2718
2719         *ret = u;
2720         return 0;
2721 }
2722
2723 int cg_blkio_weight_parse(const char *s, uint64_t *ret) {
2724         uint64_t u;
2725         int r;
2726
2727         if (isempty(s)) {
2728                 *ret = CGROUP_BLKIO_WEIGHT_INVALID;
2729                 return 0;
2730         }
2731
2732         r = safe_atou64(s, &u);
2733         if (r < 0)
2734                 return r;
2735
2736         if (u < CGROUP_BLKIO_WEIGHT_MIN || u > CGROUP_BLKIO_WEIGHT_MAX)
2737                 return -ERANGE;
2738
2739         *ret = u;
2740         return 0;
2741 }
2742
2743 bool is_cgroup_fs(const struct statfs *s) {
2744         return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2745                is_fs_type(s, CGROUP2_SUPER_MAGIC);
2746 }
2747
2748 bool fd_is_cgroup_fs(int fd) {
2749         struct statfs s;
2750
2751         if (fstatfs(fd, &s) < 0)
2752                 return -errno;
2753
2754         return is_cgroup_fs(&s);
2755 }
2756
2757 static const char *cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2758         [CGROUP_CONTROLLER_CPU] = "cpu",
2759         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2760         [CGROUP_CONTROLLER_IO] = "io",
2761         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2762         [CGROUP_CONTROLLER_MEMORY] = "memory",
2763         [CGROUP_CONTROLLER_DEVICES] = "devices",
2764         [CGROUP_CONTROLLER_PIDS] = "pids",
2765 };
2766
2767 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);