src/basic/cgroup-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <errno.h>
   4 #include <ftw.h>
   5 #include <limits.h>
   6 #include <signal.h>
   7 #include <stddef.h>
   8 #include <stdlib.h>
   9 #include <sys/types.h>
  10 #include <sys/utsname.h>
  11 #include <sys/xattr.h>
  12 #include <unistd.h>
  13
  14 #include "alloc-util.h"
  15 #include "cgroup-util.h"
  16 #include "def.h"
  17 #include "dirent-util.h"
  18 #include "extract-word.h"
  19 #include "fd-util.h"
  20 #include "fileio.h"
  21 #include "format-util.h"
  22 #include "fs-util.h"
  23 #include "log.h"
  24 #include "login-util.h"
  25 #include "macro.h"
  26 #include "missing_magic.h"
  27 #include "mkdir.h"
  28 #include "parse-util.h"
  29 #include "path-util.h"
  30 #include "process-util.h"
  31 #include "set.h"
  32 #include "special.h"
  33 #include "stat-util.h"
  34 #include "stdio-util.h"
  35 #include "string-table.h"
  36 #include "string-util.h"
  37 #include "strv.h"
  38 #include "unit-name.h"
  39 #include "user-util.h"
  40 #include "xattr-util.h"
  41
  42 static int cg_enumerate_items(const char *controller, const char *path, FILE **_f, const char *item) {
  43         _cleanup_free_ char *fs = NULL;
  44         FILE *f;
  45         int r;
  46
  47         assert(_f);
  48
  49         r = cg_get_path(controller, path, item, &fs);
  50         if (r < 0)
  51                 return r;
  52
  53         f = fopen(fs, "re");
  54         if (!f)
  55                 return -errno;
  56
  57         *_f = f;
  58         return 0;
  59 }
  60
  61 int cg_enumerate_processes(const char *controller, const char *path, FILE **_f) {
  62         return cg_enumerate_items(controller, path, _f, "cgroup.procs");
  63 }
  64
  65 int cg_read_pid(FILE *f, pid_t *_pid) {
  66         unsigned long ul;
  67
  68         /* Note that the cgroup.procs might contain duplicates! See
  69          * cgroups.txt for details. */
  70
  71         assert(f);
  72         assert(_pid);
  73
  74         errno = 0;
  75         if (fscanf(f, "%lu", &ul) != 1) {
  76
  77                 if (feof(f))
  78                         return 0;
  79
  80                 return errno_or_else(EIO);
  81         }
  82
  83         if (ul <= 0)
  84                 return -EIO;
  85
  86         *_pid = (pid_t) ul;
  87         return 1;
  88 }
  89
  90 int cg_read_event(
  91                 const char *controller,
  92                 const char *path,
  93                 const char *event,
  94                 char **ret) {
  95
  96         _cleanup_free_ char *events = NULL, *content = NULL;
  97         int r;
  98
  99         r = cg_get_path(controller, path, "cgroup.events", &events);
 100         if (r < 0)
 101                 return r;
 102
 103         r = read_full_virtual_file(events, &content, NULL);
 104         if (r < 0)
 105                 return r;
 106
 107         for (const char *p = content;;) {
 108                 _cleanup_free_ char *line = NULL, *key = NULL, *val = NULL;
 109                 const char *q;
 110
 111                 r = extract_first_word(&p, &line, "\n", 0);
 112                 if (r < 0)
 113                         return r;
 114                 if (r == 0)
 115                         return -ENOENT;
 116
 117                 q = line;
 118                 r = extract_first_word(&q, &key, " ", 0);
 119                 if (r < 0)
 120                         return r;
 121                 if (r == 0)
 122                         return -EINVAL;
 123
 124                 if (!streq(key, event))
 125                         continue;
 126
 127                 val = strdup(q);
 128                 if (!val)
 129                         return -ENOMEM;
 130
 131                 *ret = TAKE_PTR(val);
 132                 return 0;
 133         }
 134 }
 135
 136 bool cg_ns_supported(void) {
 137         static thread_local int enabled = -1;
 138
 139         if (enabled >= 0)
 140                 return enabled;
 141
 142         if (access("/proc/self/ns/cgroup", F_OK) < 0) {
 143                 if (errno != ENOENT)
 144                         log_debug_errno(errno, "Failed to check whether /proc/self/ns/cgroup is available, assuming not: %m");
 145                 enabled = false;
 146         } else
 147                 enabled = true;
 148
 149         return enabled;
 150 }
 151
 152 bool cg_freezer_supported(void) {
 153         static thread_local int supported = -1;
 154
 155         if (supported >= 0)
 156                 return supported;
 157
 158         supported = cg_all_unified() > 0 && access("/sys/fs/cgroup/init.scope/cgroup.freeze", F_OK) == 0;
 159
 160         return supported;
 161 }
 162
 163 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **_d) {
 164         _cleanup_free_ char *fs = NULL;
 165         int r;
 166         DIR *d;
 167
 168         assert(_d);
 169
 170         /* This is not recursive! */
 171
 172         r = cg_get_path(controller, path, NULL, &fs);
 173         if (r < 0)
 174                 return r;
 175
 176         d = opendir(fs);
 177         if (!d)
 178                 return -errno;
 179
 180         *_d = d;
 181         return 0;
 182 }
 183
 184 int cg_read_subgroup(DIR *d, char **fn) {
 185         struct dirent *de;
 186
 187         assert(d);
 188         assert(fn);
 189
 190         FOREACH_DIRENT_ALL(de, d, return -errno) {
 191                 char *b;
 192
 193                 if (de->d_type != DT_DIR)
 194                         continue;
 195
 196                 if (dot_or_dot_dot(de->d_name))
 197                         continue;
 198
 199                 b = strdup(de->d_name);
 200                 if (!b)
 201                         return -ENOMEM;
 202
 203                 *fn = b;
 204                 return 1;
 205         }
 206
 207         return 0;
 208 }
 209
 210 int cg_rmdir(const char *controller, const char *path) {
 211         _cleanup_free_ char *p = NULL;
 212         int r;
 213
 214         r = cg_get_path(controller, path, NULL, &p);
 215         if (r < 0)
 216                 return r;
 217
 218         r = rmdir(p);
 219         if (r < 0 && errno != ENOENT)
 220                 return -errno;
 221
 222         r = cg_hybrid_unified();
 223         if (r <= 0)
 224                 return r;
 225
 226         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 227                 r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
 228                 if (r < 0)
 229                         log_warning_errno(r, "Failed to remove compat systemd cgroup %s: %m", path);
 230         }
 231
 232         return 0;
 233 }
 234
 235 static int cg_kill_items(
 236                 const char *controller,
 237                 const char *path,
 238                 int sig,
 239                 CGroupFlags flags,
 240                 Set *s,
 241                 cg_kill_log_func_t log_kill,
 242                 void *userdata,
 243                 const char *item) {
 244
 245         _cleanup_set_free_ Set *allocated_set = NULL;
 246         bool done = false;
 247         int r, ret = 0, ret_log_kill = 0;
 248         pid_t my_pid;
 249
 250         assert(sig >= 0);
 251
 252          /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
 253           * SIGCONT on SIGKILL. */
 254         if (IN_SET(sig, SIGCONT, SIGKILL))
 255                 flags &= ~CGROUP_SIGCONT;
 256
 257         /* This goes through the tasks list and kills them all. This
 258          * is repeated until no further processes are added to the
 259          * tasks list, to properly handle forking processes */
 260
 261         if (!s) {
 262                 s = allocated_set = set_new(NULL);
 263                 if (!s)
 264                         return -ENOMEM;
 265         }
 266
 267         my_pid = getpid_cached();
 268
 269         do {
 270                 _cleanup_fclose_ FILE *f = NULL;
 271                 pid_t pid = 0;
 272                 done = true;
 273
 274                 r = cg_enumerate_items(controller, path, &f, item);
 275                 if (r < 0) {
 276                         if (ret >= 0 && r != -ENOENT)
 277                                 return r;
 278
 279                         return ret;
 280                 }
 281
 282                 while ((r = cg_read_pid(f, &pid)) > 0) {
 283
 284                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 285                                 continue;
 286
 287                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 288                                 continue;
 289
 290                         if (log_kill)
 291                                 ret_log_kill = log_kill(pid, sig, userdata);
 292
 293                         /* If we haven't killed this process yet, kill
 294                          * it */
 295                         if (kill(pid, sig) < 0) {
 296                                 if (ret >= 0 && errno != ESRCH)
 297                                         ret = -errno;
 298                         } else {
 299                                 if (flags & CGROUP_SIGCONT)
 300                                         (void) kill(pid, SIGCONT);
 301
 302                                 if (ret == 0) {
 303                                         if (log_kill)
 304                                                 ret = ret_log_kill;
 305                                         else
 306                                                 ret = 1;
 307                                 }
 308                         }
 309
 310                         done = false;
 311
 312                         r = set_put(s, PID_TO_PTR(pid));
 313                         if (r < 0) {
 314                                 if (ret >= 0)
 315                                         return r;
 316
 317                                 return ret;
 318                         }
 319                 }
 320
 321                 if (r < 0) {
 322                         if (ret >= 0)
 323                                 return r;
 324
 325                         return ret;
 326                 }
 327
 328                 /* To avoid racing against processes which fork
 329                  * quicker than we can kill them we repeat this until
 330                  * no new pids need to be killed. */
 331
 332         } while (!done);
 333
 334         return ret;
 335 }
 336
 337 int cg_kill(
 338                 const char *controller,
 339                 const char *path,
 340                 int sig,
 341                 CGroupFlags flags,
 342                 Set *s,
 343                 cg_kill_log_func_t log_kill,
 344                 void *userdata) {
 345         int r;
 346
 347         r = cg_kill_items(controller, path, sig, flags, s, log_kill, userdata, "cgroup.procs");
 348         if (r < 0 || sig != SIGKILL)
 349                 return r;
 350
 351         /* Only in case of killing with SIGKILL and when using cgroupsv2, kill remaining threads manually as
 352            a workaround for kernel bug. It was fixed in 5.2-rc5 (c03cd7738a83), backported to 4.19.66
 353            (4340d175b898) and 4.14.138 (feb6b123b7dd). */
 354         r = cg_unified_controller(controller);
 355         if (r <= 0)
 356                 return r;
 357
 358         return cg_kill_items(controller, path, sig, flags, s, log_kill, userdata, "cgroup.threads");
 359 }
 360
 361 int cg_kill_recursive(
 362                 const char *controller,
 363                 const char *path,
 364                 int sig,
 365                 CGroupFlags flags,
 366                 Set *s,
 367                 cg_kill_log_func_t log_kill,
 368                 void *userdata) {
 369
 370         _cleanup_set_free_ Set *allocated_set = NULL;
 371         _cleanup_closedir_ DIR *d = NULL;
 372         int r, ret;
 373         char *fn;
 374
 375         assert(path);
 376         assert(sig >= 0);
 377
 378         if (!s) {
 379                 s = allocated_set = set_new(NULL);
 380                 if (!s)
 381                         return -ENOMEM;
 382         }
 383
 384         ret = cg_kill(controller, path, sig, flags, s, log_kill, userdata);
 385
 386         r = cg_enumerate_subgroups(controller, path, &d);
 387         if (r < 0) {
 388                 if (ret >= 0 && r != -ENOENT)
 389                         return r;
 390
 391                 return ret;
 392         }
 393
 394         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 395                 _cleanup_free_ char *p = NULL;
 396
 397                 p = path_join(empty_to_root(path), fn);
 398                 free(fn);
 399                 if (!p)
 400                         return -ENOMEM;
 401
 402                 r = cg_kill_recursive(controller, p, sig, flags, s, log_kill, userdata);
 403                 if (r != 0 && ret >= 0)
 404                         ret = r;
 405         }
 406         if (ret >= 0 && r < 0)
 407                 ret = r;
 408
 409         if (flags & CGROUP_REMOVE) {
 410                 r = cg_rmdir(controller, path);
 411                 if (r < 0 && ret >= 0 && !IN_SET(r, -ENOENT, -EBUSY))
 412                         return r;
 413         }
 414
 415         return ret;
 416 }
 417
 418 static const char *controller_to_dirname(const char *controller) {
 419         const char *e;
 420
 421         assert(controller);
 422
 423         /* Converts a controller name to the directory name below
 424          * /sys/fs/cgroup/ we want to mount it to. Effectively, this
 425          * just cuts off the name= prefixed used for named
 426          * hierarchies, if it is specified. */
 427
 428         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 429                 if (cg_hybrid_unified() > 0)
 430                         controller = SYSTEMD_CGROUP_CONTROLLER_HYBRID;
 431                 else
 432                         controller = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 433         }
 434
 435         e = startswith(controller, "name=");
 436         if (e)
 437                 return e;
 438
 439         return controller;
 440 }
 441
 442 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **fs) {
 443         const char *dn;
 444         char *t = NULL;
 445
 446         assert(fs);
 447         assert(controller);
 448
 449         dn = controller_to_dirname(controller);
 450
 451         if (isempty(path) && isempty(suffix))
 452                 t = path_join("/sys/fs/cgroup", dn);
 453         else if (isempty(path))
 454                 t = path_join("/sys/fs/cgroup", dn, suffix);
 455         else if (isempty(suffix))
 456                 t = path_join("/sys/fs/cgroup", dn, path);
 457         else
 458                 t = path_join("/sys/fs/cgroup", dn, path, suffix);
 459         if (!t)
 460                 return -ENOMEM;
 461
 462         *fs = t;
 463         return 0;
 464 }
 465
 466 static int join_path_unified(const char *path, const char *suffix, char **fs) {
 467         char *t;
 468
 469         assert(fs);
 470
 471         if (isempty(path) && isempty(suffix))
 472                 t = strdup("/sys/fs/cgroup");
 473         else if (isempty(path))
 474                 t = path_join("/sys/fs/cgroup", suffix);
 475         else if (isempty(suffix))
 476                 t = path_join("/sys/fs/cgroup", path);
 477         else
 478                 t = path_join("/sys/fs/cgroup", path, suffix);
 479         if (!t)
 480                 return -ENOMEM;
 481
 482         *fs = t;
 483         return 0;
 484 }
 485
 486 int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) {
 487         int r;
 488
 489         assert(fs);
 490
 491         if (!controller) {
 492                 char *t;
 493
 494                 /* If no controller is specified, we return the path
 495                  * *below* the controllers, without any prefix. */
 496
 497                 if (!path && !suffix)
 498                         return -EINVAL;
 499
 500                 if (!suffix)
 501                         t = strdup(path);
 502                 else if (!path)
 503                         t = strdup(suffix);
 504                 else
 505                         t = path_join(path, suffix);
 506                 if (!t)
 507                         return -ENOMEM;
 508
 509                 *fs = path_simplify(t, false);
 510                 return 0;
 511         }
 512
 513         if (!cg_controller_is_valid(controller))
 514                 return -EINVAL;
 515
 516         r = cg_all_unified();
 517         if (r < 0)
 518                 return r;
 519         if (r > 0)
 520                 r = join_path_unified(path, suffix, fs);
 521         else
 522                 r = join_path_legacy(controller, path, suffix, fs);
 523         if (r < 0)
 524                 return r;
 525
 526         path_simplify(*fs, false);
 527         return 0;
 528 }
 529
 530 static int controller_is_v1_accessible(const char *root, const char *controller) {
 531         const char *cpath, *dn;
 532
 533         assert(controller);
 534
 535         dn = controller_to_dirname(controller);
 536
 537         /* If root if specified, we check that:
 538          * - possible subcgroup is created at root,
 539          * - we can modify the hierarchy. */
 540
 541         cpath = strjoina("/sys/fs/cgroup/", dn, root, root ? "/cgroup.procs" : NULL);
 542         if (laccess(cpath, root ? W_OK : F_OK) < 0)
 543                 return -errno;
 544
 545         return 0;
 546 }
 547
 548 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 549         int r;
 550
 551         assert(controller);
 552         assert(fs);
 553
 554         if (!cg_controller_is_valid(controller))
 555                 return -EINVAL;
 556
 557         r = cg_all_unified();
 558         if (r < 0)
 559                 return r;
 560         if (r > 0) {
 561                 /* In the unified hierarchy all controllers are considered accessible,
 562                  * except for the named hierarchies */
 563                 if (startswith(controller, "name="))
 564                         return -EOPNOTSUPP;
 565         } else {
 566                 /* Check if the specified controller is actually accessible */
 567                 r = controller_is_v1_accessible(NULL, controller);
 568                 if (r < 0)
 569                         return r;
 570         }
 571
 572         return cg_get_path(controller, path, suffix, fs);
 573 }
 574
 575 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags) {
 576         _cleanup_free_ char *fs = NULL;
 577         int r;
 578
 579         assert(path);
 580         assert(name);
 581         assert(value || size <= 0);
 582
 583         r = cg_get_path(controller, path, NULL, &fs);
 584         if (r < 0)
 585                 return r;
 586
 587         if (setxattr(fs, name, value, size, flags) < 0)
 588                 return -errno;
 589
 590         return 0;
 591 }
 592
 593 int cg_get_xattr(const char *controller, const char *path, const char *name, void *value, size_t size) {
 594         _cleanup_free_ char *fs = NULL;
 595         ssize_t n;
 596         int r;
 597
 598         assert(path);
 599         assert(name);
 600
 601         r = cg_get_path(controller, path, NULL, &fs);
 602         if (r < 0)
 603                 return r;
 604
 605         n = getxattr(fs, name, value, size);
 606         if (n < 0)
 607                 return -errno;
 608
 609         return (int) n;
 610 }
 611
 612 int cg_get_xattr_malloc(const char *controller, const char *path, const char *name, char **ret) {
 613         _cleanup_free_ char *fs = NULL;
 614         int r;
 615
 616         assert(path);
 617         assert(name);
 618
 619         r = cg_get_path(controller, path, NULL, &fs);
 620         if (r < 0)
 621                 return r;
 622
 623         r = getxattr_malloc(fs, name, ret, false);
 624         if (r < 0)
 625                 return r;
 626
 627         return r;
 628 }
 629
 630 int cg_get_xattr_bool(const char *controller, const char *path, const char *name) {
 631         _cleanup_free_ char *val = NULL;
 632         int r;
 633
 634         assert(path);
 635         assert(name);
 636
 637         r = cg_get_xattr_malloc(controller, path, name, &val);
 638         if (r < 0)
 639                 return r;
 640
 641         return parse_boolean(val);
 642 }
 643
 644 int cg_remove_xattr(const char *controller, const char *path, const char *name) {
 645         _cleanup_free_ char *fs = NULL;
 646         int r;
 647
 648         assert(path);
 649         assert(name);
 650
 651         r = cg_get_path(controller, path, NULL, &fs);
 652         if (r < 0)
 653                 return r;
 654
 655         if (removexattr(fs, name) < 0)
 656                 return -errno;
 657
 658         return 0;
 659 }
 660
 661 int cg_pid_get_path(const char *controller, pid_t pid, char **ret_path) {
 662         _cleanup_fclose_ FILE *f = NULL;
 663         const char *fs, *controller_str;
 664         int unified, r;
 665
 666         assert(pid >= 0);
 667         assert(ret_path);
 668
 669         if (controller) {
 670                 if (!cg_controller_is_valid(controller))
 671                         return -EINVAL;
 672         } else
 673                 controller = SYSTEMD_CGROUP_CONTROLLER;
 674
 675         unified = cg_unified_controller(controller);
 676         if (unified < 0)
 677                 return unified;
 678         if (unified == 0) {
 679                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 680                         controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 681                 else
 682                         controller_str = controller;
 683         }
 684
 685         fs = procfs_file_alloca(pid, "cgroup");
 686         r = fopen_unlocked(fs, "re", &f);
 687         if (r == -ENOENT)
 688                 return -ESRCH;
 689         if (r < 0)
 690                 return r;
 691
 692         for (;;) {
 693                 _cleanup_free_ char *line = NULL;
 694                 char *e;
 695
 696                 r = read_line(f, LONG_LINE_MAX, &line);
 697                 if (r < 0)
 698                         return r;
 699                 if (r == 0)
 700                         return -ENODATA;
 701
 702                 if (unified) {
 703                         e = startswith(line, "0:");
 704                         if (!e)
 705                                 continue;
 706
 707                         e = strchr(e, ':');
 708                         if (!e)
 709                                 continue;
 710                 } else {
 711                         char *l;
 712
 713                         l = strchr(line, ':');
 714                         if (!l)
 715                                 continue;
 716
 717                         l++;
 718                         e = strchr(l, ':');
 719                         if (!e)
 720                                 continue;
 721                         *e = 0;
 722
 723                         r = string_contains_word(l, ",", controller_str);
 724                         if (r < 0)
 725                                 return r;
 726                         if (r == 0)
 727                                 continue;
 728                 }
 729
 730                 char *path = strdup(e + 1);
 731                 if (!path)
 732                         return -ENOMEM;
 733
 734                 /* Truncate suffix indicating the process is a zombie */
 735                 e = endswith(path, " (deleted)");
 736                 if (e)
 737                         *e = 0;
 738
 739                 *ret_path = path;
 740                 return 0;
 741         }
 742 }
 743
 744 int cg_install_release_agent(const char *controller, const char *agent) {
 745         _cleanup_free_ char *fs = NULL, *contents = NULL;
 746         const char *sc;
 747         int r;
 748
 749         assert(agent);
 750
 751         r = cg_unified_controller(controller);
 752         if (r < 0)
 753                 return r;
 754         if (r > 0) /* doesn't apply to unified hierarchy */
 755                 return -EOPNOTSUPP;
 756
 757         r = cg_get_path(controller, NULL, "release_agent", &fs);
 758         if (r < 0)
 759                 return r;
 760
 761         r = read_one_line_file(fs, &contents);
 762         if (r < 0)
 763                 return r;
 764
 765         sc = strstrip(contents);
 766         if (isempty(sc)) {
 767                 r = write_string_file(fs, agent, WRITE_STRING_FILE_DISABLE_BUFFER);
 768                 if (r < 0)
 769                         return r;
 770         } else if (!path_equal(sc, agent))
 771                 return -EEXIST;
 772
 773         fs = mfree(fs);
 774         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 775         if (r < 0)
 776                 return r;
 777
 778         contents = mfree(contents);
 779         r = read_one_line_file(fs, &contents);
 780         if (r < 0)
 781                 return r;
 782
 783         sc = strstrip(contents);
 784         if (streq(sc, "0")) {
 785                 r = write_string_file(fs, "1", WRITE_STRING_FILE_DISABLE_BUFFER);
 786                 if (r < 0)
 787                         return r;
 788
 789                 return 1;
 790         }
 791
 792         if (!streq(sc, "1"))
 793                 return -EIO;
 794
 795         return 0;
 796 }
 797
 798 int cg_uninstall_release_agent(const char *controller) {
 799         _cleanup_free_ char *fs = NULL;
 800         int r;
 801
 802         r = cg_unified_controller(controller);
 803         if (r < 0)
 804                 return r;
 805         if (r > 0) /* Doesn't apply to unified hierarchy */
 806                 return -EOPNOTSUPP;
 807
 808         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 809         if (r < 0)
 810                 return r;
 811
 812         r = write_string_file(fs, "0", WRITE_STRING_FILE_DISABLE_BUFFER);
 813         if (r < 0)
 814                 return r;
 815
 816         fs = mfree(fs);
 817
 818         r = cg_get_path(controller, NULL, "release_agent", &fs);
 819         if (r < 0)
 820                 return r;
 821
 822         r = write_string_file(fs, "", WRITE_STRING_FILE_DISABLE_BUFFER);
 823         if (r < 0)
 824                 return r;
 825
 826         return 0;
 827 }
 828
 829 int cg_is_empty(const char *controller, const char *path) {
 830         _cleanup_fclose_ FILE *f = NULL;
 831         pid_t pid;
 832         int r;
 833
 834         assert(path);
 835
 836         r = cg_enumerate_processes(controller, path, &f);
 837         if (r == -ENOENT)
 838                 return true;
 839         if (r < 0)
 840                 return r;
 841
 842         r = cg_read_pid(f, &pid);
 843         if (r < 0)
 844                 return r;
 845
 846         return r == 0;
 847 }
 848
 849 int cg_is_empty_recursive(const char *controller, const char *path) {
 850         int r;
 851
 852         assert(path);
 853
 854         /* The root cgroup is always populated */
 855         if (controller && empty_or_root(path))
 856                 return false;
 857
 858         r = cg_unified_controller(controller);
 859         if (r < 0)
 860                 return r;
 861         if (r > 0) {
 862                 _cleanup_free_ char *t = NULL;
 863
 864                 /* On the unified hierarchy we can check empty state
 865                  * via the "populated" attribute of "cgroup.events". */
 866
 867                 r = cg_read_event(controller, path, "populated", &t);
 868                 if (r == -ENOENT)
 869                         return true;
 870                 if (r < 0)
 871                         return r;
 872
 873                 return streq(t, "0");
 874         } else {
 875                 _cleanup_closedir_ DIR *d = NULL;
 876                 char *fn;
 877
 878                 r = cg_is_empty(controller, path);
 879                 if (r <= 0)
 880                         return r;
 881
 882                 r = cg_enumerate_subgroups(controller, path, &d);
 883                 if (r == -ENOENT)
 884                         return true;
 885                 if (r < 0)
 886                         return r;
 887
 888                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
 889                         _cleanup_free_ char *p = NULL;
 890
 891                         p = path_join(path, fn);
 892                         free(fn);
 893                         if (!p)
 894                                 return -ENOMEM;
 895
 896                         r = cg_is_empty_recursive(controller, p);
 897                         if (r <= 0)
 898                                 return r;
 899                 }
 900                 if (r < 0)
 901                         return r;
 902
 903                 return true;
 904         }
 905 }
 906
 907 int cg_split_spec(const char *spec, char **ret_controller, char **ret_path) {
 908         _cleanup_free_ char *controller = NULL, *path = NULL;
 909
 910         assert(spec);
 911
 912         if (*spec == '/') {
 913                 if (!path_is_normalized(spec))
 914                         return -EINVAL;
 915
 916                 if (ret_path) {
 917                         path = strdup(spec);
 918                         if (!path)
 919                                 return -ENOMEM;
 920
 921                         path_simplify(path, false);
 922                 }
 923
 924         } else {
 925                 const char *e;
 926
 927                 e = strchr(spec, ':');
 928                 if (e) {
 929                         controller = strndup(spec, e-spec);
 930                         if (!controller)
 931                                 return -ENOMEM;
 932                         if (!cg_controller_is_valid(controller))
 933                                 return -EINVAL;
 934
 935                         if (!isempty(e + 1)) {
 936                                 path = strdup(e+1);
 937                                 if (!path)
 938                                         return -ENOMEM;
 939
 940                                 if (!path_is_normalized(path) ||
 941                                     !path_is_absolute(path))
 942                                         return -EINVAL;
 943
 944                                 path_simplify(path, false);
 945                         }
 946
 947                 } else {
 948                         if (!cg_controller_is_valid(spec))
 949                                 return -EINVAL;
 950
 951                         if (ret_controller) {
 952                                 controller = strdup(spec);
 953                                 if (!controller)
 954                                         return -ENOMEM;
 955                         }
 956                 }
 957         }
 958
 959         if (ret_controller)
 960                 *ret_controller = TAKE_PTR(controller);
 961         if (ret_path)
 962                 *ret_path = TAKE_PTR(path);
 963         return 0;
 964 }
 965
 966 int cg_mangle_path(const char *path, char **result) {
 967         _cleanup_free_ char *c = NULL, *p = NULL;
 968         char *t;
 969         int r;
 970
 971         assert(path);
 972         assert(result);
 973
 974         /* First, check if it already is a filesystem path */
 975         if (path_startswith(path, "/sys/fs/cgroup")) {
 976
 977                 t = strdup(path);
 978                 if (!t)
 979                         return -ENOMEM;
 980
 981                 *result = path_simplify(t, false);
 982                 return 0;
 983         }
 984
 985         /* Otherwise, treat it as cg spec */
 986         r = cg_split_spec(path, &c, &p);
 987         if (r < 0)
 988                 return r;
 989
 990         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
 991 }
 992
 993 int cg_get_root_path(char **path) {
 994         char *p, *e;
 995         int r;
 996
 997         assert(path);
 998
 999         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1000         if (r < 0)
1001                 return r;
1002
1003         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1004         if (!e)
1005                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1006         if (!e)
1007                 e = endswith(p, "/system"); /* even more legacy */
1008         if (e)
1009                 *e = 0;
1010
1011         *path = p;
1012         return 0;
1013 }
1014
1015 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1016         _cleanup_free_ char *rt = NULL;
1017         char *p;
1018         int r;
1019
1020         assert(cgroup);
1021         assert(shifted);
1022
1023         if (!root) {
1024                 /* If the root was specified let's use that, otherwise
1025                  * let's determine it from PID 1 */
1026
1027                 r = cg_get_root_path(&rt);
1028                 if (r < 0)
1029                         return r;
1030
1031                 root = rt;
1032         }
1033
1034         p = path_startswith(cgroup, root);
1035         if (p && p > cgroup)
1036                 *shifted = p - 1;
1037         else
1038                 *shifted = cgroup;
1039
1040         return 0;
1041 }
1042
1043 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1044         _cleanup_free_ char *raw = NULL;
1045         const char *c;
1046         int r;
1047
1048         assert(pid >= 0);
1049         assert(cgroup);
1050
1051         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1052         if (r < 0)
1053                 return r;
1054
1055         r = cg_shift_path(raw, root, &c);
1056         if (r < 0)
1057                 return r;
1058
1059         if (c == raw)
1060                 *cgroup = TAKE_PTR(raw);
1061         else {
1062                 char *n;
1063
1064                 n = strdup(c);
1065                 if (!n)
1066                         return -ENOMEM;
1067
1068                 *cgroup = n;
1069         }
1070
1071         return 0;
1072 }
1073
1074 int cg_path_decode_unit(const char *cgroup, char **unit) {
1075         char *c, *s;
1076         size_t n;
1077
1078         assert(cgroup);
1079         assert(unit);
1080
1081         n = strcspn(cgroup, "/");
1082         if (n < 3)
1083                 return -ENXIO;
1084
1085         c = strndupa(cgroup, n);
1086         c = cg_unescape(c);
1087
1088         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1089                 return -ENXIO;
1090
1091         s = strdup(c);
1092         if (!s)
1093                 return -ENOMEM;
1094
1095         *unit = s;
1096         return 0;
1097 }
1098
1099 static bool valid_slice_name(const char *p, size_t n) {
1100
1101         if (!p)
1102                 return false;
1103
1104         if (n < STRLEN("x.slice"))
1105                 return false;
1106
1107         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1108                 char buf[n+1], *c;
1109
1110                 memcpy(buf, p, n);
1111                 buf[n] = 0;
1112
1113                 c = cg_unescape(buf);
1114
1115                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1116         }
1117
1118         return false;
1119 }
1120
1121 static const char *skip_slices(const char *p) {
1122         assert(p);
1123
1124         /* Skips over all slice assignments */
1125
1126         for (;;) {
1127                 size_t n;
1128
1129                 p += strspn(p, "/");
1130
1131                 n = strcspn(p, "/");
1132                 if (!valid_slice_name(p, n))
1133                         return p;
1134
1135                 p += n;
1136         }
1137 }
1138
1139 int cg_path_get_unit(const char *path, char **ret) {
1140         _cleanup_free_ char *unit = NULL;
1141         const char *e;
1142         int r;
1143
1144         assert(path);
1145         assert(ret);
1146
1147         e = skip_slices(path);
1148
1149         r = cg_path_decode_unit(e, &unit);
1150         if (r < 0)
1151                 return r;
1152
1153         /* We skipped over the slices, don't accept any now */
1154         if (endswith(unit, ".slice"))
1155                 return -ENXIO;
1156
1157         *ret = TAKE_PTR(unit);
1158         return 0;
1159 }
1160
1161 int cg_pid_get_unit(pid_t pid, char **unit) {
1162         _cleanup_free_ char *cgroup = NULL;
1163         int r;
1164
1165         assert(unit);
1166
1167         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1168         if (r < 0)
1169                 return r;
1170
1171         return cg_path_get_unit(cgroup, unit);
1172 }
1173
1174 /**
1175  * Skip session-*.scope, but require it to be there.
1176  */
1177 static const char *skip_session(const char *p) {
1178         size_t n;
1179
1180         if (isempty(p))
1181                 return NULL;
1182
1183         p += strspn(p, "/");
1184
1185         n = strcspn(p, "/");
1186         if (n < STRLEN("session-x.scope"))
1187                 return NULL;
1188
1189         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1190                 char buf[n - 8 - 6 + 1];
1191
1192                 memcpy(buf, p + 8, n - 8 - 6);
1193                 buf[n - 8 - 6] = 0;
1194
1195                 /* Note that session scopes never need unescaping,
1196                  * since they cannot conflict with the kernel's own
1197                  * names, hence we don't need to call cg_unescape()
1198                  * here. */
1199
1200                 if (!session_id_valid(buf))
1201                         return false;
1202
1203                 p += n;
1204                 p += strspn(p, "/");
1205                 return p;
1206         }
1207
1208         return NULL;
1209 }
1210
1211 /**
1212  * Skip user@*.service, but require it to be there.
1213  */
1214 static const char *skip_user_manager(const char *p) {
1215         size_t n;
1216
1217         if (isempty(p))
1218                 return NULL;
1219
1220         p += strspn(p, "/");
1221
1222         n = strcspn(p, "/");
1223         if (n < STRLEN("user@x.service"))
1224                 return NULL;
1225
1226         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1227                 char buf[n - 5 - 8 + 1];
1228
1229                 memcpy(buf, p + 5, n - 5 - 8);
1230                 buf[n - 5 - 8] = 0;
1231
1232                 /* Note that user manager services never need unescaping,
1233                  * since they cannot conflict with the kernel's own
1234                  * names, hence we don't need to call cg_unescape()
1235                  * here. */
1236
1237                 if (parse_uid(buf, NULL) < 0)
1238                         return NULL;
1239
1240                 p += n;
1241                 p += strspn(p, "/");
1242
1243                 return p;
1244         }
1245
1246         return NULL;
1247 }
1248
1249 static const char *skip_user_prefix(const char *path) {
1250         const char *e, *t;
1251
1252         assert(path);
1253
1254         /* Skip slices, if there are any */
1255         e = skip_slices(path);
1256
1257         /* Skip the user manager, if it's in the path now... */
1258         t = skip_user_manager(e);
1259         if (t)
1260                 return t;
1261
1262         /* Alternatively skip the user session if it is in the path... */
1263         return skip_session(e);
1264 }
1265
1266 int cg_path_get_user_unit(const char *path, char **ret) {
1267         const char *t;
1268
1269         assert(path);
1270         assert(ret);
1271
1272         t = skip_user_prefix(path);
1273         if (!t)
1274                 return -ENXIO;
1275
1276         /* And from here on it looks pretty much the same as for a system unit, hence let's use the same
1277          * parser. */
1278         return cg_path_get_unit(t, ret);
1279 }
1280
1281 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1282         _cleanup_free_ char *cgroup = NULL;
1283         int r;
1284
1285         assert(unit);
1286
1287         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1288         if (r < 0)
1289                 return r;
1290
1291         return cg_path_get_user_unit(cgroup, unit);
1292 }
1293
1294 int cg_path_get_machine_name(const char *path, char **machine) {
1295         _cleanup_free_ char *u = NULL;
1296         const char *sl;
1297         int r;
1298
1299         r = cg_path_get_unit(path, &u);
1300         if (r < 0)
1301                 return r;
1302
1303         sl = strjoina("/run/systemd/machines/unit:", u);
1304         return readlink_malloc(sl, machine);
1305 }
1306
1307 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1308         _cleanup_free_ char *cgroup = NULL;
1309         int r;
1310
1311         assert(machine);
1312
1313         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1314         if (r < 0)
1315                 return r;
1316
1317         return cg_path_get_machine_name(cgroup, machine);
1318 }
1319
1320 int cg_path_get_session(const char *path, char **session) {
1321         _cleanup_free_ char *unit = NULL;
1322         char *start, *end;
1323         int r;
1324
1325         assert(path);
1326
1327         r = cg_path_get_unit(path, &unit);
1328         if (r < 0)
1329                 return r;
1330
1331         start = startswith(unit, "session-");
1332         if (!start)
1333                 return -ENXIO;
1334         end = endswith(start, ".scope");
1335         if (!end)
1336                 return -ENXIO;
1337
1338         *end = 0;
1339         if (!session_id_valid(start))
1340                 return -ENXIO;
1341
1342         if (session) {
1343                 char *rr;
1344
1345                 rr = strdup(start);
1346                 if (!rr)
1347                         return -ENOMEM;
1348
1349                 *session = rr;
1350         }
1351
1352         return 0;
1353 }
1354
1355 int cg_pid_get_session(pid_t pid, char **session) {
1356         _cleanup_free_ char *cgroup = NULL;
1357         int r;
1358
1359         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1360         if (r < 0)
1361                 return r;
1362
1363         return cg_path_get_session(cgroup, session);
1364 }
1365
1366 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1367         _cleanup_free_ char *slice = NULL;
1368         char *start, *end;
1369         int r;
1370
1371         assert(path);
1372
1373         r = cg_path_get_slice(path, &slice);
1374         if (r < 0)
1375                 return r;
1376
1377         start = startswith(slice, "user-");
1378         if (!start)
1379                 return -ENXIO;
1380         end = endswith(start, ".slice");
1381         if (!end)
1382                 return -ENXIO;
1383
1384         *end = 0;
1385         if (parse_uid(start, uid) < 0)
1386                 return -ENXIO;
1387
1388         return 0;
1389 }
1390
1391 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1392         _cleanup_free_ char *cgroup = NULL;
1393         int r;
1394
1395         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1396         if (r < 0)
1397                 return r;
1398
1399         return cg_path_get_owner_uid(cgroup, uid);
1400 }
1401
1402 int cg_path_get_slice(const char *p, char **slice) {
1403         const char *e = NULL;
1404
1405         assert(p);
1406         assert(slice);
1407
1408         /* Finds the right-most slice unit from the beginning, but
1409          * stops before we come to the first non-slice unit. */
1410
1411         for (;;) {
1412                 size_t n;
1413
1414                 p += strspn(p, "/");
1415
1416                 n = strcspn(p, "/");
1417                 if (!valid_slice_name(p, n)) {
1418
1419                         if (!e) {
1420                                 char *s;
1421
1422                                 s = strdup(SPECIAL_ROOT_SLICE);
1423                                 if (!s)
1424                                         return -ENOMEM;
1425
1426                                 *slice = s;
1427                                 return 0;
1428                         }
1429
1430                         return cg_path_decode_unit(e, slice);
1431                 }
1432
1433                 e = p;
1434                 p += n;
1435         }
1436 }
1437
1438 int cg_pid_get_slice(pid_t pid, char **slice) {
1439         _cleanup_free_ char *cgroup = NULL;
1440         int r;
1441
1442         assert(slice);
1443
1444         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1445         if (r < 0)
1446                 return r;
1447
1448         return cg_path_get_slice(cgroup, slice);
1449 }
1450
1451 int cg_path_get_user_slice(const char *p, char **slice) {
1452         const char *t;
1453         assert(p);
1454         assert(slice);
1455
1456         t = skip_user_prefix(p);
1457         if (!t)
1458                 return -ENXIO;
1459
1460         /* And now it looks pretty much the same as for a system
1461          * slice, so let's just use the same parser from here on. */
1462         return cg_path_get_slice(t, slice);
1463 }
1464
1465 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1466         _cleanup_free_ char *cgroup = NULL;
1467         int r;
1468
1469         assert(slice);
1470
1471         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1472         if (r < 0)
1473                 return r;
1474
1475         return cg_path_get_user_slice(cgroup, slice);
1476 }
1477
1478 char *cg_escape(const char *p) {
1479         bool need_prefix = false;
1480
1481         /* This implements very minimal escaping for names to be used
1482          * as file names in the cgroup tree: any name which might
1483          * conflict with a kernel name or is prefixed with '_' is
1484          * prefixed with a '_'. That way, when reading cgroup names it
1485          * is sufficient to remove a single prefixing underscore if
1486          * there is one. */
1487
1488         /* The return value of this function (unlike cg_unescape())
1489          * needs free()! */
1490
1491         if (IN_SET(p[0], 0, '_', '.') ||
1492             STR_IN_SET(p, "notify_on_release", "release_agent", "tasks") ||
1493             startswith(p, "cgroup."))
1494                 need_prefix = true;
1495         else {
1496                 const char *dot;
1497
1498                 dot = strrchr(p, '.');
1499                 if (dot) {
1500                         CGroupController c;
1501                         size_t l = dot - p;
1502
1503                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1504                                 const char *n;
1505
1506                                 n = cgroup_controller_to_string(c);
1507
1508                                 if (l != strlen(n))
1509                                         continue;
1510
1511                                 if (memcmp(p, n, l) != 0)
1512                                         continue;
1513
1514                                 need_prefix = true;
1515                                 break;
1516                         }
1517                 }
1518         }
1519
1520         if (need_prefix)
1521                 return strjoin("_", p);
1522
1523         return strdup(p);
1524 }
1525
1526 char *cg_unescape(const char *p) {
1527         assert(p);
1528
1529         /* The return value of this function (unlike cg_escape())
1530          * doesn't need free()! */
1531
1532         if (p[0] == '_')
1533                 return (char*) p+1;
1534
1535         return (char*) p;
1536 }
1537
1538 #define CONTROLLER_VALID                        \
1539         DIGITS LETTERS                          \
1540         "_"
1541
1542 bool cg_controller_is_valid(const char *p) {
1543         const char *t, *s;
1544
1545         if (!p)
1546                 return false;
1547
1548         if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1549                 return true;
1550
1551         s = startswith(p, "name=");
1552         if (s)
1553                 p = s;
1554
1555         if (IN_SET(*p, 0, '_'))
1556                 return false;
1557
1558         for (t = p; *t; t++)
1559                 if (!strchr(CONTROLLER_VALID, *t))
1560                         return false;
1561
1562         if (t - p > NAME_MAX)
1563                 return false;
1564
1565         return true;
1566 }
1567
1568 int cg_slice_to_path(const char *unit, char **ret) {
1569         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1570         const char *dash;
1571         int r;
1572
1573         assert(unit);
1574         assert(ret);
1575
1576         if (streq(unit, SPECIAL_ROOT_SLICE)) {
1577                 char *x;
1578
1579                 x = strdup("");
1580                 if (!x)
1581                         return -ENOMEM;
1582                 *ret = x;
1583                 return 0;
1584         }
1585
1586         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1587                 return -EINVAL;
1588
1589         if (!endswith(unit, ".slice"))
1590                 return -EINVAL;
1591
1592         r = unit_name_to_prefix(unit, &p);
1593         if (r < 0)
1594                 return r;
1595
1596         dash = strchr(p, '-');
1597
1598         /* Don't allow initial dashes */
1599         if (dash == p)
1600                 return -EINVAL;
1601
1602         while (dash) {
1603                 _cleanup_free_ char *escaped = NULL;
1604                 char n[dash - p + sizeof(".slice")];
1605
1606 #if HAS_FEATURE_MEMORY_SANITIZER
1607                 /* msan doesn't instrument stpncpy, so it thinks
1608                  * n is later used uninitialized:
1609                  * https://github.com/google/sanitizers/issues/926
1610                  */
1611                 zero(n);
1612 #endif
1613
1614                 /* Don't allow trailing or double dashes */
1615                 if (IN_SET(dash[1], 0, '-'))
1616                         return -EINVAL;
1617
1618                 strcpy(stpncpy(n, p, dash - p), ".slice");
1619                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1620                         return -EINVAL;
1621
1622                 escaped = cg_escape(n);
1623                 if (!escaped)
1624                         return -ENOMEM;
1625
1626                 if (!strextend(&s, escaped, "/"))
1627                         return -ENOMEM;
1628
1629                 dash = strchr(dash+1, '-');
1630         }
1631
1632         e = cg_escape(unit);
1633         if (!e)
1634                 return -ENOMEM;
1635
1636         if (!strextend(&s, e))
1637                 return -ENOMEM;
1638
1639         *ret = TAKE_PTR(s);
1640
1641         return 0;
1642 }
1643
1644 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1645         _cleanup_free_ char *p = NULL;
1646         int r;
1647
1648         r = cg_get_path(controller, path, attribute, &p);
1649         if (r < 0)
1650                 return r;
1651
1652         return write_string_file(p, value, WRITE_STRING_FILE_DISABLE_BUFFER);
1653 }
1654
1655 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1656         _cleanup_free_ char *p = NULL;
1657         int r;
1658
1659         r = cg_get_path(controller, path, attribute, &p);
1660         if (r < 0)
1661                 return r;
1662
1663         return read_one_line_file(p, ret);
1664 }
1665
1666 int cg_get_attribute_as_uint64(const char *controller, const char *path, const char *attribute, uint64_t *ret) {
1667         _cleanup_free_ char *value = NULL;
1668         uint64_t v;
1669         int r;
1670
1671         assert(ret);
1672
1673         r = cg_get_attribute(controller, path, attribute, &value);
1674         if (r == -ENOENT)
1675                 return -ENODATA;
1676         if (r < 0)
1677                 return r;
1678
1679         if (streq(value, "max")) {
1680                 *ret = CGROUP_LIMIT_MAX;
1681                 return 0;
1682         }
1683
1684         r = safe_atou64(value, &v);
1685         if (r < 0)
1686                 return r;
1687
1688         *ret = v;
1689         return 0;
1690 }
1691
1692 int cg_get_attribute_as_bool(const char *controller, const char *path, const char *attribute, bool *ret) {
1693         _cleanup_free_ char *value = NULL;
1694         int r;
1695
1696         assert(ret);
1697
1698         r = cg_get_attribute(controller, path, attribute, &value);
1699         if (r == -ENOENT)
1700                 return -ENODATA;
1701         if (r < 0)
1702                 return r;
1703
1704         r = parse_boolean(value);
1705         if (r < 0)
1706                 return r;
1707
1708         *ret = r;
1709         return 0;
1710 }
1711
1712 int cg_get_owner(const char *controller, const char *path, uid_t *ret_uid) {
1713         _cleanup_free_ char *f = NULL;
1714         struct stat stats;
1715         int r;
1716
1717         assert(ret_uid);
1718
1719         r = cg_get_path(controller, path, NULL, &f);
1720         if (r < 0)
1721                 return r;
1722
1723         r = stat(f, &stats);
1724         if (r < 0)
1725                 return -errno;
1726
1727         *ret_uid = stats.st_uid;
1728         return 0;
1729 }
1730
1731 int cg_get_keyed_attribute_full(
1732                 const char *controller,
1733                 const char *path,
1734                 const char *attribute,
1735                 char **keys,
1736                 char **ret_values,
1737                 CGroupKeyMode mode) {
1738
1739         _cleanup_free_ char *filename = NULL, *contents = NULL;
1740         const char *p;
1741         size_t n, i, n_done = 0;
1742         char **v;
1743         int r;
1744
1745         /* Reads one or more fields of a cgroup v2 keyed attribute file. The 'keys' parameter should be an strv with
1746          * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
1747          * entries as 'keys'. On success each entry will be set to the value of the matching key.
1748          *
1749          * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. If mode
1750          * is set to GG_KEY_MODE_GRACEFUL we ignore missing keys and return those that were parsed successfully. */
1751
1752         r = cg_get_path(controller, path, attribute, &filename);
1753         if (r < 0)
1754                 return r;
1755
1756         r = read_full_file(filename, &contents, NULL);
1757         if (r < 0)
1758                 return r;
1759
1760         n = strv_length(keys);
1761         if (n == 0) /* No keys to retrieve? That's easy, we are done then */
1762                 return 0;
1763
1764         /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
1765         v = newa0(char*, n);
1766
1767         for (p = contents; *p;) {
1768                 const char *w = NULL;
1769
1770                 for (i = 0; i < n; i++)
1771                         if (!v[i]) {
1772                                 w = first_word(p, keys[i]);
1773                                 if (w)
1774                                         break;
1775                         }
1776
1777                 if (w) {
1778                         size_t l;
1779
1780                         l = strcspn(w, NEWLINE);
1781                         v[i] = strndup(w, l);
1782                         if (!v[i]) {
1783                                 r = -ENOMEM;
1784                                 goto fail;
1785                         }
1786
1787                         n_done++;
1788                         if (n_done >= n)
1789                                 goto done;
1790
1791                         p = w + l;
1792                 } else
1793                         p += strcspn(p, NEWLINE);
1794
1795                 p += strspn(p, NEWLINE);
1796         }
1797
1798         if (mode & CG_KEY_MODE_GRACEFUL)
1799                 goto done;
1800
1801         r = -ENXIO;
1802
1803 fail:
1804         for (i = 0; i < n; i++)
1805                 free(v[i]);
1806
1807         return r;
1808
1809 done:
1810         memcpy(ret_values, v, sizeof(char*) * n);
1811         if (mode & CG_KEY_MODE_GRACEFUL)
1812                 return n_done;
1813
1814         return 0;
1815 }
1816
1817 int cg_mask_to_string(CGroupMask mask, char **ret) {
1818         _cleanup_free_ char *s = NULL;
1819         size_t n = 0, allocated = 0;
1820         bool space = false;
1821         CGroupController c;
1822
1823         assert(ret);
1824
1825         if (mask == 0) {
1826                 *ret = NULL;
1827                 return 0;
1828         }
1829
1830         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1831                 const char *k;
1832                 size_t l;
1833
1834                 if (!FLAGS_SET(mask, CGROUP_CONTROLLER_TO_MASK(c)))
1835                         continue;
1836
1837                 k = cgroup_controller_to_string(c);
1838                 l = strlen(k);
1839
1840                 if (!GREEDY_REALLOC(s, allocated, n + space + l + 1))
1841                         return -ENOMEM;
1842
1843                 if (space)
1844                         s[n] = ' ';
1845                 memcpy(s + n + space, k, l);
1846                 n += space + l;
1847
1848                 space = true;
1849         }
1850
1851         assert(s);
1852
1853         s[n] = 0;
1854         *ret = TAKE_PTR(s);
1855
1856         return 0;
1857 }
1858
1859 int cg_mask_from_string(const char *value, CGroupMask *ret) {
1860         CGroupMask m = 0;
1861
1862         assert(ret);
1863         assert(value);
1864
1865         for (;;) {
1866                 _cleanup_free_ char *n = NULL;
1867                 CGroupController v;
1868                 int r;
1869
1870                 r = extract_first_word(&value, &n, NULL, 0);
1871                 if (r < 0)
1872                         return r;
1873                 if (r == 0)
1874                         break;
1875
1876                 v = cgroup_controller_from_string(n);
1877                 if (v < 0)
1878                         continue;
1879
1880                 m |= CGROUP_CONTROLLER_TO_MASK(v);
1881         }
1882
1883         *ret = m;
1884         return 0;
1885 }
1886
1887 int cg_mask_supported_subtree(const char *root, CGroupMask *ret) {
1888         CGroupMask mask;
1889         int r;
1890
1891         /* Determines the mask of supported cgroup controllers. Only includes controllers we can make sense of and that
1892          * are actually accessible. Only covers real controllers, i.e. not the CGROUP_CONTROLLER_BPF_xyz
1893          * pseudo-controllers. */
1894
1895         r = cg_all_unified();
1896         if (r < 0)
1897                 return r;
1898         if (r > 0) {
1899                 _cleanup_free_ char *controllers = NULL, *path = NULL;
1900
1901                 /* In the unified hierarchy we can read the supported and accessible controllers from
1902                  * the top-level cgroup attribute */
1903
1904                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
1905                 if (r < 0)
1906                         return r;
1907
1908                 r = read_one_line_file(path, &controllers);
1909                 if (r < 0)
1910                         return r;
1911
1912                 r = cg_mask_from_string(controllers, &mask);
1913                 if (r < 0)
1914                         return r;
1915
1916                 /* Mask controllers that are not supported in unified hierarchy. */
1917                 mask &= CGROUP_MASK_V2;
1918
1919         } else {
1920                 CGroupController c;
1921
1922                 /* In the legacy hierarchy, we check which hierarchies are accessible. */
1923
1924                 mask = 0;
1925                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1926                         CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1927                         const char *n;
1928
1929                         if (!FLAGS_SET(CGROUP_MASK_V1, bit))
1930                                 continue;
1931
1932                         n = cgroup_controller_to_string(c);
1933                         if (controller_is_v1_accessible(root, n) >= 0)
1934                                 mask |= bit;
1935                 }
1936         }
1937
1938         *ret = mask;
1939         return 0;
1940 }
1941
1942 int cg_mask_supported(CGroupMask *ret) {
1943         _cleanup_free_ char *root = NULL;
1944         int r;
1945
1946         r = cg_get_root_path(&root);
1947         if (r < 0)
1948                 return r;
1949
1950         return cg_mask_supported_subtree(root, ret);
1951 }
1952
1953 int cg_kernel_controllers(Set **ret) {
1954         _cleanup_set_free_free_ Set *controllers = NULL;
1955         _cleanup_fclose_ FILE *f = NULL;
1956         int r;
1957
1958         assert(ret);
1959
1960         /* Determines the full list of kernel-known controllers. Might include controllers we don't actually support
1961          * and controllers that aren't currently accessible (because not mounted). This does not include "name="
1962          * pseudo-controllers. */
1963
1964         controllers = set_new(&string_hash_ops);
1965         if (!controllers)
1966                 return -ENOMEM;
1967
1968         r = fopen_unlocked("/proc/cgroups", "re", &f);
1969         if (r == -ENOENT) {
1970                 *ret = NULL;
1971                 return 0;
1972         }
1973         if (r < 0)
1974                 return r;
1975
1976         /* Ignore the header line */
1977         (void) read_line(f, SIZE_MAX, NULL);
1978
1979         for (;;) {
1980                 char *controller;
1981                 int enabled = 0;
1982
1983                 errno = 0;
1984                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
1985
1986                         if (feof(f))
1987                                 break;
1988
1989                         if (ferror(f))
1990                                 return errno_or_else(EIO);
1991
1992                         return -EBADMSG;
1993                 }
1994
1995                 if (!enabled) {
1996                         free(controller);
1997                         continue;
1998                 }
1999
2000                 if (!cg_controller_is_valid(controller)) {
2001                         free(controller);
2002                         return -EBADMSG;
2003                 }
2004
2005                 r = set_consume(controllers, controller);
2006                 if (r < 0)
2007                         return r;
2008         }
2009
2010         *ret = TAKE_PTR(controllers);
2011
2012         return 0;
2013 }
2014
2015 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup2 on
2016  * /sys/fs/cgroup/systemd. This unfortunately broke other tools (such as docker) which expected the v1
2017  * "name=systemd" hierarchy on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mounts v2 on
2018  * /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility
2019  * with other tools.
2020  *
2021  * To keep live upgrade working, we detect and support v232 layout. When v232 layout is detected, to keep
2022  * cgroup v2 process management but disable the compat dual layout, we return true on
2023  * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and false on cg_hybrid_unified().
2024  */
2025 static thread_local bool unified_systemd_v232;
2026
2027 int cg_unified_cached(bool flush) {
2028         static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2029
2030         struct statfs fs;
2031
2032         /* Checks if we support the unified hierarchy. Returns an
2033          * error when the cgroup hierarchies aren't mounted yet or we
2034          * have any other trouble determining if the unified hierarchy
2035          * is supported. */
2036
2037         if (flush)
2038                 unified_cache = CGROUP_UNIFIED_UNKNOWN;
2039         else if (unified_cache >= CGROUP_UNIFIED_NONE)
2040                 return unified_cache;
2041
2042         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2043                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2044
2045         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2046                 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2047                 unified_cache = CGROUP_UNIFIED_ALL;
2048         } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2049                 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2050                     F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2051                         log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2052                         unified_cache = CGROUP_UNIFIED_SYSTEMD;
2053                         unified_systemd_v232 = false;
2054                 } else {
2055                         if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0) {
2056                                 if (errno == ENOENT) {
2057                                         /* Some other software may have set up /sys/fs/cgroup in a configuration we do not recognize. */
2058                                         log_debug_errno(errno, "Unsupported cgroupsv1 setup detected: name=systemd hierarchy not found.");
2059                                         return -ENOMEDIUM;
2060                                 }
2061                                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2062                         }
2063
2064                         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2065                                 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2066                                 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2067                                 unified_systemd_v232 = true;
2068                         } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2069                                 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2070                                 unified_cache = CGROUP_UNIFIED_NONE;
2071                         } else {
2072                                 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2073                                           (unsigned long long) fs.f_type);
2074                                 unified_cache = CGROUP_UNIFIED_NONE;
2075                         }
2076                 }
2077         } else if (F_TYPE_EQUAL(fs.f_type, SYSFS_MAGIC)) {
2078                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2079                                        "No filesystem is currently mounted on /sys/fs/cgroup.");
2080         } else
2081                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2082                                        "Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2083                                        (unsigned long long)fs.f_type);
2084
2085         return unified_cache;
2086 }
2087
2088 int cg_unified_controller(const char *controller) {
2089         int r;
2090
2091         r = cg_unified_cached(false);
2092         if (r < 0)
2093                 return r;
2094
2095         if (r == CGROUP_UNIFIED_NONE)
2096                 return false;
2097
2098         if (r >= CGROUP_UNIFIED_ALL)
2099                 return true;
2100
2101         return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2102 }
2103
2104 int cg_all_unified(void) {
2105         int r;
2106
2107         r = cg_unified_cached(false);
2108         if (r < 0)
2109                 return r;
2110
2111         return r >= CGROUP_UNIFIED_ALL;
2112 }
2113
2114 int cg_hybrid_unified(void) {
2115         int r;
2116
2117         r = cg_unified_cached(false);
2118         if (r < 0)
2119                 return r;
2120
2121         return r == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2122 }
2123
2124 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2125         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2126         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2127         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2128         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2129 };
2130
2131 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2132         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2133         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2134         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2135         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2136 };
2137
2138 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2139
2140 bool is_cgroup_fs(const struct statfs *s) {
2141         return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2142                is_fs_type(s, CGROUP2_SUPER_MAGIC);
2143 }
2144
2145 bool fd_is_cgroup_fs(int fd) {
2146         struct statfs s;
2147
2148         if (fstatfs(fd, &s) < 0)
2149                 return -errno;
2150
2151         return is_cgroup_fs(&s);
2152 }
2153
2154 static const char *const cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2155         [CGROUP_CONTROLLER_CPU] = "cpu",
2156         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2157         [CGROUP_CONTROLLER_CPUSET] = "cpuset",
2158         [CGROUP_CONTROLLER_IO] = "io",
2159         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2160         [CGROUP_CONTROLLER_MEMORY] = "memory",
2161         [CGROUP_CONTROLLER_DEVICES] = "devices",
2162         [CGROUP_CONTROLLER_PIDS] = "pids",
2163         [CGROUP_CONTROLLER_BPF_FIREWALL] = "bpf-firewall",
2164         [CGROUP_CONTROLLER_BPF_DEVICES] = "bpf-devices",
2165 };
2166
2167 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);
2168
2169 CGroupMask get_cpu_accounting_mask(void) {
2170         static CGroupMask needed_mask = (CGroupMask) -1;
2171
2172         /* On kernel ≥4.15 with unified hierarchy, cpu.stat's usage_usec is
2173          * provided externally from the CPU controller, which means we don't
2174          * need to enable the CPU controller just to get metrics. This is good,
2175          * because enabling the CPU controller comes at a minor performance
2176          * hit, especially when it's propagated deep into large hierarchies.
2177          * There's also no separate CPU accounting controller available within
2178          * a unified hierarchy.
2179          *
2180          * This combination of factors results in the desired cgroup mask to
2181          * enable for CPU accounting varying as follows:
2182          *
2183          *                   ╔═════════════════════╤═════════════════════╗
2184          *                   ║     Linux ≥4.15     │     Linux <4.15     ║
2185          *   ╔═══════════════╬═════════════════════╪═════════════════════╣
2186          *   ║ Unified       ║ nothing             │ CGROUP_MASK_CPU     ║
2187          *   ╟───────────────╫─────────────────────┼─────────────────────╢
2188          *   ║ Hybrid/Legacy ║ CGROUP_MASK_CPUACCT │ CGROUP_MASK_CPUACCT ║
2189          *   ╚═══════════════╩═════════════════════╧═════════════════════╝
2190          *
2191          * We check kernel version here instead of manually checking whether
2192          * cpu.stat is present for every cgroup, as that check in itself would
2193          * already be fairly expensive.
2194          *
2195          * Kernels where this patch has been backported will therefore have the
2196          * CPU controller enabled unnecessarily. This is more expensive than
2197          * necessary, but harmless. ☺️
2198          */
2199
2200         if (needed_mask == (CGroupMask) -1) {
2201                 if (cg_all_unified()) {
2202                         struct utsname u;
2203                         assert_se(uname(&u) >= 0);
2204
2205                         if (strverscmp_improved(u.release, "4.15") < 0)
2206                                 needed_mask = CGROUP_MASK_CPU;
2207                         else
2208                                 needed_mask = 0;
2209                 } else
2210                         needed_mask = CGROUP_MASK_CPUACCT;
2211         }
2212
2213         return needed_mask;
2214 }
2215
2216 bool cpu_accounting_is_cheap(void) {
2217         return get_cpu_accounting_mask() == 0;
2218 }
2219
2220 static const char* const managed_oom_mode_table[_MANAGED_OOM_MODE_MAX] = {
2221         [MANAGED_OOM_AUTO] = "auto",
2222         [MANAGED_OOM_KILL] = "kill",
2223 };
2224
2225 DEFINE_STRING_TABLE_LOOKUP(managed_oom_mode, ManagedOOMMode);
2226
2227 static const char* const managed_oom_preference_table[_MANAGED_OOM_PREFERENCE_MAX] = {
2228         [MANAGED_OOM_PREFERENCE_NONE] = "none",
2229         [MANAGED_OOM_PREFERENCE_AVOID] = "avoid",
2230         [MANAGED_OOM_PREFERENCE_OMIT] = "omit",
2231 };
2232
2233 DEFINE_STRING_TABLE_LOOKUP(managed_oom_preference, ManagedOOMPreference);