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_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         cpath = strjoina("/sys/fs/cgroup/", dn);
 537         if (root)
 538                 /* Also check that:
 539                  * - possible subcgroup is created at root,
 540                  * - we can modify the hierarchy.
 541                  * "Leak" cpath on stack */
 542                 cpath = strjoina(cpath, root, "/cgroup.procs");
 543
 544         if (laccess(cpath, root ? W_OK : F_OK) < 0)
 545                 return -errno;
 546
 547         return 0;
 548 }
 549
 550 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 551         int r;
 552
 553         assert(controller);
 554         assert(fs);
 555
 556         if (!cg_controller_is_valid(controller))
 557                 return -EINVAL;
 558
 559         r = cg_all_unified();
 560         if (r < 0)
 561                 return r;
 562         if (r > 0) {
 563                 /* In the unified hierarchy all controllers are considered accessible,
 564                  * except for the named hierarchies */
 565                 if (startswith(controller, "name="))
 566                         return -EOPNOTSUPP;
 567         } else {
 568                 /* Check if the specified controller is actually accessible */
 569                 r = controller_is_v1_accessible(NULL, controller);
 570                 if (r < 0)
 571                         return r;
 572         }
 573
 574         return cg_get_path(controller, path, suffix, fs);
 575 }
 576
 577 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags) {
 578         _cleanup_free_ char *fs = NULL;
 579         int r;
 580
 581         assert(path);
 582         assert(name);
 583         assert(value || size <= 0);
 584
 585         r = cg_get_path(controller, path, NULL, &fs);
 586         if (r < 0)
 587                 return r;
 588
 589         if (setxattr(fs, name, value, size, flags) < 0)
 590                 return -errno;
 591
 592         return 0;
 593 }
 594
 595 int cg_get_xattr(const char *controller, const char *path, const char *name, void *value, size_t size) {
 596         _cleanup_free_ char *fs = NULL;
 597         ssize_t n;
 598         int r;
 599
 600         assert(path);
 601         assert(name);
 602
 603         r = cg_get_path(controller, path, NULL, &fs);
 604         if (r < 0)
 605                 return r;
 606
 607         n = getxattr(fs, name, value, size);
 608         if (n < 0)
 609                 return -errno;
 610
 611         return (int) n;
 612 }
 613
 614 int cg_get_xattr_malloc(const char *controller, const char *path, const char *name, char **ret) {
 615         _cleanup_free_ char *fs = NULL;
 616         int r;
 617
 618         assert(path);
 619         assert(name);
 620
 621         r = cg_get_path(controller, path, NULL, &fs);
 622         if (r < 0)
 623                 return r;
 624
 625         r = getxattr_malloc(fs, name, ret, false);
 626         if (r < 0)
 627                 return r;
 628
 629         return r;
 630 }
 631
 632 int cg_get_xattr_bool(const char *controller, const char *path, const char *name) {
 633         _cleanup_free_ char *val = NULL;
 634         int r;
 635
 636         assert(path);
 637         assert(name);
 638
 639         r = cg_get_xattr_malloc(controller, path, name, &val);
 640         if (r < 0)
 641                 return r;
 642
 643         return parse_boolean(val);
 644 }
 645
 646 int cg_remove_xattr(const char *controller, const char *path, const char *name) {
 647         _cleanup_free_ char *fs = NULL;
 648         int r;
 649
 650         assert(path);
 651         assert(name);
 652
 653         r = cg_get_path(controller, path, NULL, &fs);
 654         if (r < 0)
 655                 return r;
 656
 657         if (removexattr(fs, name) < 0)
 658                 return -errno;
 659
 660         return 0;
 661 }
 662
 663 int cg_pid_get_path(const char *controller, pid_t pid, char **ret_path) {
 664         _cleanup_fclose_ FILE *f = NULL;
 665         const char *fs, *controller_str;
 666         int unified, r;
 667
 668         assert(pid >= 0);
 669         assert(ret_path);
 670
 671         if (controller) {
 672                 if (!cg_controller_is_valid(controller))
 673                         return -EINVAL;
 674         } else
 675                 controller = SYSTEMD_CGROUP_CONTROLLER;
 676
 677         unified = cg_unified_controller(controller);
 678         if (unified < 0)
 679                 return unified;
 680         if (unified == 0) {
 681                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 682                         controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 683                 else
 684                         controller_str = controller;
 685         }
 686
 687         fs = procfs_file_alloca(pid, "cgroup");
 688         r = fopen_unlocked(fs, "re", &f);
 689         if (r == -ENOENT)
 690                 return -ESRCH;
 691         if (r < 0)
 692                 return r;
 693
 694         for (;;) {
 695                 _cleanup_free_ char *line = NULL;
 696                 char *e;
 697
 698                 r = read_line(f, LONG_LINE_MAX, &line);
 699                 if (r < 0)
 700                         return r;
 701                 if (r == 0)
 702                         return -ENODATA;
 703
 704                 if (unified) {
 705                         e = startswith(line, "0:");
 706                         if (!e)
 707                                 continue;
 708
 709                         e = strchr(e, ':');
 710                         if (!e)
 711                                 continue;
 712                 } else {
 713                         char *l;
 714
 715                         l = strchr(line, ':');
 716                         if (!l)
 717                                 continue;
 718
 719                         l++;
 720                         e = strchr(l, ':');
 721                         if (!e)
 722                                 continue;
 723                         *e = 0;
 724
 725                         r = string_contains_word(l, ",", controller_str);
 726                         if (r < 0)
 727                                 return r;
 728                         if (r == 0)
 729                                 continue;
 730                 }
 731
 732                 char *path = strdup(e + 1);
 733                 if (!path)
 734                         return -ENOMEM;
 735
 736                 /* Truncate suffix indicating the process is a zombie */
 737                 e = endswith(path, " (deleted)");
 738                 if (e)
 739                         *e = 0;
 740
 741                 *ret_path = path;
 742                 return 0;
 743         }
 744 }
 745
 746 int cg_install_release_agent(const char *controller, const char *agent) {
 747         _cleanup_free_ char *fs = NULL, *contents = NULL;
 748         const char *sc;
 749         int r;
 750
 751         assert(agent);
 752
 753         r = cg_unified_controller(controller);
 754         if (r < 0)
 755                 return r;
 756         if (r > 0) /* doesn't apply to unified hierarchy */
 757                 return -EOPNOTSUPP;
 758
 759         r = cg_get_path(controller, NULL, "release_agent", &fs);
 760         if (r < 0)
 761                 return r;
 762
 763         r = read_one_line_file(fs, &contents);
 764         if (r < 0)
 765                 return r;
 766
 767         sc = strstrip(contents);
 768         if (isempty(sc)) {
 769                 r = write_string_file(fs, agent, WRITE_STRING_FILE_DISABLE_BUFFER);
 770                 if (r < 0)
 771                         return r;
 772         } else if (!path_equal(sc, agent))
 773                 return -EEXIST;
 774
 775         fs = mfree(fs);
 776         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 777         if (r < 0)
 778                 return r;
 779
 780         contents = mfree(contents);
 781         r = read_one_line_file(fs, &contents);
 782         if (r < 0)
 783                 return r;
 784
 785         sc = strstrip(contents);
 786         if (streq(sc, "0")) {
 787                 r = write_string_file(fs, "1", WRITE_STRING_FILE_DISABLE_BUFFER);
 788                 if (r < 0)
 789                         return r;
 790
 791                 return 1;
 792         }
 793
 794         if (!streq(sc, "1"))
 795                 return -EIO;
 796
 797         return 0;
 798 }
 799
 800 int cg_uninstall_release_agent(const char *controller) {
 801         _cleanup_free_ char *fs = NULL;
 802         int r;
 803
 804         r = cg_unified_controller(controller);
 805         if (r < 0)
 806                 return r;
 807         if (r > 0) /* Doesn't apply to unified hierarchy */
 808                 return -EOPNOTSUPP;
 809
 810         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 811         if (r < 0)
 812                 return r;
 813
 814         r = write_string_file(fs, "0", WRITE_STRING_FILE_DISABLE_BUFFER);
 815         if (r < 0)
 816                 return r;
 817
 818         fs = mfree(fs);
 819
 820         r = cg_get_path(controller, NULL, "release_agent", &fs);
 821         if (r < 0)
 822                 return r;
 823
 824         r = write_string_file(fs, "", WRITE_STRING_FILE_DISABLE_BUFFER);
 825         if (r < 0)
 826                 return r;
 827
 828         return 0;
 829 }
 830
 831 int cg_is_empty(const char *controller, const char *path) {
 832         _cleanup_fclose_ FILE *f = NULL;
 833         pid_t pid;
 834         int r;
 835
 836         assert(path);
 837
 838         r = cg_enumerate_processes(controller, path, &f);
 839         if (r == -ENOENT)
 840                 return true;
 841         if (r < 0)
 842                 return r;
 843
 844         r = cg_read_pid(f, &pid);
 845         if (r < 0)
 846                 return r;
 847
 848         return r == 0;
 849 }
 850
 851 int cg_is_empty_recursive(const char *controller, const char *path) {
 852         int r;
 853
 854         assert(path);
 855
 856         /* The root cgroup is always populated */
 857         if (controller && empty_or_root(path))
 858                 return false;
 859
 860         r = cg_unified_controller(controller);
 861         if (r < 0)
 862                 return r;
 863         if (r > 0) {
 864                 _cleanup_free_ char *t = NULL;
 865
 866                 /* On the unified hierarchy we can check empty state
 867                  * via the "populated" attribute of "cgroup.events". */
 868
 869                 r = cg_read_event(controller, path, "populated", &t);
 870                 if (r == -ENOENT)
 871                         return true;
 872                 if (r < 0)
 873                         return r;
 874
 875                 return streq(t, "0");
 876         } else {
 877                 _cleanup_closedir_ DIR *d = NULL;
 878                 char *fn;
 879
 880                 r = cg_is_empty(controller, path);
 881                 if (r <= 0)
 882                         return r;
 883
 884                 r = cg_enumerate_subgroups(controller, path, &d);
 885                 if (r == -ENOENT)
 886                         return true;
 887                 if (r < 0)
 888                         return r;
 889
 890                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
 891                         _cleanup_free_ char *p = NULL;
 892
 893                         p = path_join(path, fn);
 894                         free(fn);
 895                         if (!p)
 896                                 return -ENOMEM;
 897
 898                         r = cg_is_empty_recursive(controller, p);
 899                         if (r <= 0)
 900                                 return r;
 901                 }
 902                 if (r < 0)
 903                         return r;
 904
 905                 return true;
 906         }
 907 }
 908
 909 int cg_split_spec(const char *spec, char **ret_controller, char **ret_path) {
 910         _cleanup_free_ char *controller = NULL, *path = NULL;
 911
 912         assert(spec);
 913
 914         if (*spec == '/') {
 915                 if (!path_is_normalized(spec))
 916                         return -EINVAL;
 917
 918                 if (ret_path) {
 919                         path = strdup(spec);
 920                         if (!path)
 921                                 return -ENOMEM;
 922
 923                         path_simplify(path, false);
 924                 }
 925
 926         } else {
 927                 const char *e;
 928
 929                 e = strchr(spec, ':');
 930                 if (e) {
 931                         controller = strndup(spec, e-spec);
 932                         if (!controller)
 933                                 return -ENOMEM;
 934                         if (!cg_controller_is_valid(controller))
 935                                 return -EINVAL;
 936
 937                         if (!isempty(e + 1)) {
 938                                 path = strdup(e+1);
 939                                 if (!path)
 940                                         return -ENOMEM;
 941
 942                                 if (!path_is_normalized(path) ||
 943                                     !path_is_absolute(path))
 944                                         return -EINVAL;
 945
 946                                 path_simplify(path, false);
 947                         }
 948
 949                 } else {
 950                         if (!cg_controller_is_valid(spec))
 951                                 return -EINVAL;
 952
 953                         if (ret_controller) {
 954                                 controller = strdup(spec);
 955                                 if (!controller)
 956                                         return -ENOMEM;
 957                         }
 958                 }
 959         }
 960
 961         if (ret_controller)
 962                 *ret_controller = TAKE_PTR(controller);
 963         if (ret_path)
 964                 *ret_path = TAKE_PTR(path);
 965         return 0;
 966 }
 967
 968 int cg_mangle_path(const char *path, char **result) {
 969         _cleanup_free_ char *c = NULL, *p = NULL;
 970         char *t;
 971         int r;
 972
 973         assert(path);
 974         assert(result);
 975
 976         /* First, check if it already is a filesystem path */
 977         if (path_startswith(path, "/sys/fs/cgroup")) {
 978
 979                 t = strdup(path);
 980                 if (!t)
 981                         return -ENOMEM;
 982
 983                 *result = path_simplify(t, false);
 984                 return 0;
 985         }
 986
 987         /* Otherwise, treat it as cg spec */
 988         r = cg_split_spec(path, &c, &p);
 989         if (r < 0)
 990                 return r;
 991
 992         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
 993 }
 994
 995 int cg_get_root_path(char **path) {
 996         char *p, *e;
 997         int r;
 998
 999         assert(path);
1000
1001         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1002         if (r < 0)
1003                 return r;
1004
1005         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1006         if (!e)
1007                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1008         if (!e)
1009                 e = endswith(p, "/system"); /* even more legacy */
1010         if (e)
1011                 *e = 0;
1012
1013         *path = p;
1014         return 0;
1015 }
1016
1017 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1018         _cleanup_free_ char *rt = NULL;
1019         char *p;
1020         int r;
1021
1022         assert(cgroup);
1023         assert(shifted);
1024
1025         if (!root) {
1026                 /* If the root was specified let's use that, otherwise
1027                  * let's determine it from PID 1 */
1028
1029                 r = cg_get_root_path(&rt);
1030                 if (r < 0)
1031                         return r;
1032
1033                 root = rt;
1034         }
1035
1036         p = path_startswith(cgroup, root);
1037         if (p && p > cgroup)
1038                 *shifted = p - 1;
1039         else
1040                 *shifted = cgroup;
1041
1042         return 0;
1043 }
1044
1045 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1046         _cleanup_free_ char *raw = NULL;
1047         const char *c;
1048         int r;
1049
1050         assert(pid >= 0);
1051         assert(cgroup);
1052
1053         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1054         if (r < 0)
1055                 return r;
1056
1057         r = cg_shift_path(raw, root, &c);
1058         if (r < 0)
1059                 return r;
1060
1061         if (c == raw)
1062                 *cgroup = TAKE_PTR(raw);
1063         else {
1064                 char *n;
1065
1066                 n = strdup(c);
1067                 if (!n)
1068                         return -ENOMEM;
1069
1070                 *cgroup = n;
1071         }
1072
1073         return 0;
1074 }
1075
1076 int cg_path_decode_unit(const char *cgroup, char **unit) {
1077         char *c, *s;
1078         size_t n;
1079
1080         assert(cgroup);
1081         assert(unit);
1082
1083         n = strcspn(cgroup, "/");
1084         if (n < 3)
1085                 return -ENXIO;
1086
1087         c = strndupa(cgroup, n);
1088         c = cg_unescape(c);
1089
1090         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1091                 return -ENXIO;
1092
1093         s = strdup(c);
1094         if (!s)
1095                 return -ENOMEM;
1096
1097         *unit = s;
1098         return 0;
1099 }
1100
1101 static bool valid_slice_name(const char *p, size_t n) {
1102
1103         if (!p)
1104                 return false;
1105
1106         if (n < STRLEN("x.slice"))
1107                 return false;
1108
1109         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1110                 char buf[n+1], *c;
1111
1112                 memcpy(buf, p, n);
1113                 buf[n] = 0;
1114
1115                 c = cg_unescape(buf);
1116
1117                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1118         }
1119
1120         return false;
1121 }
1122
1123 static const char *skip_slices(const char *p) {
1124         assert(p);
1125
1126         /* Skips over all slice assignments */
1127
1128         for (;;) {
1129                 size_t n;
1130
1131                 p += strspn(p, "/");
1132
1133                 n = strcspn(p, "/");
1134                 if (!valid_slice_name(p, n))
1135                         return p;
1136
1137                 p += n;
1138         }
1139 }
1140
1141 int cg_path_get_unit(const char *path, char **ret) {
1142         _cleanup_free_ char *unit = NULL;
1143         const char *e;
1144         int r;
1145
1146         assert(path);
1147         assert(ret);
1148
1149         e = skip_slices(path);
1150
1151         r = cg_path_decode_unit(e, &unit);
1152         if (r < 0)
1153                 return r;
1154
1155         /* We skipped over the slices, don't accept any now */
1156         if (endswith(unit, ".slice"))
1157                 return -ENXIO;
1158
1159         *ret = TAKE_PTR(unit);
1160         return 0;
1161 }
1162
1163 int cg_pid_get_unit(pid_t pid, char **unit) {
1164         _cleanup_free_ char *cgroup = NULL;
1165         int r;
1166
1167         assert(unit);
1168
1169         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1170         if (r < 0)
1171                 return r;
1172
1173         return cg_path_get_unit(cgroup, unit);
1174 }
1175
1176 /**
1177  * Skip session-*.scope, but require it to be there.
1178  */
1179 static const char *skip_session(const char *p) {
1180         size_t n;
1181
1182         if (isempty(p))
1183                 return NULL;
1184
1185         p += strspn(p, "/");
1186
1187         n = strcspn(p, "/");
1188         if (n < STRLEN("session-x.scope"))
1189                 return NULL;
1190
1191         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1192                 char buf[n - 8 - 6 + 1];
1193
1194                 memcpy(buf, p + 8, n - 8 - 6);
1195                 buf[n - 8 - 6] = 0;
1196
1197                 /* Note that session scopes never need unescaping,
1198                  * since they cannot conflict with the kernel's own
1199                  * names, hence we don't need to call cg_unescape()
1200                  * here. */
1201
1202                 if (!session_id_valid(buf))
1203                         return false;
1204
1205                 p += n;
1206                 p += strspn(p, "/");
1207                 return p;
1208         }
1209
1210         return NULL;
1211 }
1212
1213 /**
1214  * Skip user@*.service, but require it to be there.
1215  */
1216 static const char *skip_user_manager(const char *p) {
1217         size_t n;
1218
1219         if (isempty(p))
1220                 return NULL;
1221
1222         p += strspn(p, "/");
1223
1224         n = strcspn(p, "/");
1225         if (n < STRLEN("user@x.service"))
1226                 return NULL;
1227
1228         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1229                 char buf[n - 5 - 8 + 1];
1230
1231                 memcpy(buf, p + 5, n - 5 - 8);
1232                 buf[n - 5 - 8] = 0;
1233
1234                 /* Note that user manager services never need unescaping,
1235                  * since they cannot conflict with the kernel's own
1236                  * names, hence we don't need to call cg_unescape()
1237                  * here. */
1238
1239                 if (parse_uid(buf, NULL) < 0)
1240                         return NULL;
1241
1242                 p += n;
1243                 p += strspn(p, "/");
1244
1245                 return p;
1246         }
1247
1248         return NULL;
1249 }
1250
1251 static const char *skip_user_prefix(const char *path) {
1252         const char *e, *t;
1253
1254         assert(path);
1255
1256         /* Skip slices, if there are any */
1257         e = skip_slices(path);
1258
1259         /* Skip the user manager, if it's in the path now... */
1260         t = skip_user_manager(e);
1261         if (t)
1262                 return t;
1263
1264         /* Alternatively skip the user session if it is in the path... */
1265         return skip_session(e);
1266 }
1267
1268 int cg_path_get_user_unit(const char *path, char **ret) {
1269         const char *t;
1270
1271         assert(path);
1272         assert(ret);
1273
1274         t = skip_user_prefix(path);
1275         if (!t)
1276                 return -ENXIO;
1277
1278         /* And from here on it looks pretty much the same as for a system unit, hence let's use the same
1279          * parser. */
1280         return cg_path_get_unit(t, ret);
1281 }
1282
1283 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1284         _cleanup_free_ char *cgroup = NULL;
1285         int r;
1286
1287         assert(unit);
1288
1289         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1290         if (r < 0)
1291                 return r;
1292
1293         return cg_path_get_user_unit(cgroup, unit);
1294 }
1295
1296 int cg_path_get_machine_name(const char *path, char **machine) {
1297         _cleanup_free_ char *u = NULL;
1298         const char *sl;
1299         int r;
1300
1301         r = cg_path_get_unit(path, &u);
1302         if (r < 0)
1303                 return r;
1304
1305         sl = strjoina("/run/systemd/machines/unit:", u);
1306         return readlink_malloc(sl, machine);
1307 }
1308
1309 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1310         _cleanup_free_ char *cgroup = NULL;
1311         int r;
1312
1313         assert(machine);
1314
1315         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1316         if (r < 0)
1317                 return r;
1318
1319         return cg_path_get_machine_name(cgroup, machine);
1320 }
1321
1322 int cg_path_get_session(const char *path, char **session) {
1323         _cleanup_free_ char *unit = NULL;
1324         char *start, *end;
1325         int r;
1326
1327         assert(path);
1328
1329         r = cg_path_get_unit(path, &unit);
1330         if (r < 0)
1331                 return r;
1332
1333         start = startswith(unit, "session-");
1334         if (!start)
1335                 return -ENXIO;
1336         end = endswith(start, ".scope");
1337         if (!end)
1338                 return -ENXIO;
1339
1340         *end = 0;
1341         if (!session_id_valid(start))
1342                 return -ENXIO;
1343
1344         if (session) {
1345                 char *rr;
1346
1347                 rr = strdup(start);
1348                 if (!rr)
1349                         return -ENOMEM;
1350
1351                 *session = rr;
1352         }
1353
1354         return 0;
1355 }
1356
1357 int cg_pid_get_session(pid_t pid, char **session) {
1358         _cleanup_free_ char *cgroup = NULL;
1359         int r;
1360
1361         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1362         if (r < 0)
1363                 return r;
1364
1365         return cg_path_get_session(cgroup, session);
1366 }
1367
1368 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1369         _cleanup_free_ char *slice = NULL;
1370         char *start, *end;
1371         int r;
1372
1373         assert(path);
1374
1375         r = cg_path_get_slice(path, &slice);
1376         if (r < 0)
1377                 return r;
1378
1379         start = startswith(slice, "user-");
1380         if (!start)
1381                 return -ENXIO;
1382         end = endswith(start, ".slice");
1383         if (!end)
1384                 return -ENXIO;
1385
1386         *end = 0;
1387         if (parse_uid(start, uid) < 0)
1388                 return -ENXIO;
1389
1390         return 0;
1391 }
1392
1393 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1394         _cleanup_free_ char *cgroup = NULL;
1395         int r;
1396
1397         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1398         if (r < 0)
1399                 return r;
1400
1401         return cg_path_get_owner_uid(cgroup, uid);
1402 }
1403
1404 int cg_path_get_slice(const char *p, char **slice) {
1405         const char *e = NULL;
1406
1407         assert(p);
1408         assert(slice);
1409
1410         /* Finds the right-most slice unit from the beginning, but
1411          * stops before we come to the first non-slice unit. */
1412
1413         for (;;) {
1414                 size_t n;
1415
1416                 p += strspn(p, "/");
1417
1418                 n = strcspn(p, "/");
1419                 if (!valid_slice_name(p, n)) {
1420
1421                         if (!e) {
1422                                 char *s;
1423
1424                                 s = strdup(SPECIAL_ROOT_SLICE);
1425                                 if (!s)
1426                                         return -ENOMEM;
1427
1428                                 *slice = s;
1429                                 return 0;
1430                         }
1431
1432                         return cg_path_decode_unit(e, slice);
1433                 }
1434
1435                 e = p;
1436                 p += n;
1437         }
1438 }
1439
1440 int cg_pid_get_slice(pid_t pid, char **slice) {
1441         _cleanup_free_ char *cgroup = NULL;
1442         int r;
1443
1444         assert(slice);
1445
1446         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1447         if (r < 0)
1448                 return r;
1449
1450         return cg_path_get_slice(cgroup, slice);
1451 }
1452
1453 int cg_path_get_user_slice(const char *p, char **slice) {
1454         const char *t;
1455         assert(p);
1456         assert(slice);
1457
1458         t = skip_user_prefix(p);
1459         if (!t)
1460                 return -ENXIO;
1461
1462         /* And now it looks pretty much the same as for a system
1463          * slice, so let's just use the same parser from here on. */
1464         return cg_path_get_slice(t, slice);
1465 }
1466
1467 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1468         _cleanup_free_ char *cgroup = NULL;
1469         int r;
1470
1471         assert(slice);
1472
1473         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1474         if (r < 0)
1475                 return r;
1476
1477         return cg_path_get_user_slice(cgroup, slice);
1478 }
1479
1480 char *cg_escape(const char *p) {
1481         bool need_prefix = false;
1482
1483         /* This implements very minimal escaping for names to be used
1484          * as file names in the cgroup tree: any name which might
1485          * conflict with a kernel name or is prefixed with '_' is
1486          * prefixed with a '_'. That way, when reading cgroup names it
1487          * is sufficient to remove a single prefixing underscore if
1488          * there is one. */
1489
1490         /* The return value of this function (unlike cg_unescape())
1491          * needs free()! */
1492
1493         if (IN_SET(p[0], 0, '_', '.') ||
1494             STR_IN_SET(p, "notify_on_release", "release_agent", "tasks") ||
1495             startswith(p, "cgroup."))
1496                 need_prefix = true;
1497         else {
1498                 const char *dot;
1499
1500                 dot = strrchr(p, '.');
1501                 if (dot) {
1502                         CGroupController c;
1503                         size_t l = dot - p;
1504
1505                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1506                                 const char *n;
1507
1508                                 n = cgroup_controller_to_string(c);
1509
1510                                 if (l != strlen(n))
1511                                         continue;
1512
1513                                 if (memcmp(p, n, l) != 0)
1514                                         continue;
1515
1516                                 need_prefix = true;
1517                                 break;
1518                         }
1519                 }
1520         }
1521
1522         if (need_prefix)
1523                 return strjoin("_", p);
1524
1525         return strdup(p);
1526 }
1527
1528 char *cg_unescape(const char *p) {
1529         assert(p);
1530
1531         /* The return value of this function (unlike cg_escape())
1532          * doesn't need free()! */
1533
1534         if (p[0] == '_')
1535                 return (char*) p+1;
1536
1537         return (char*) p;
1538 }
1539
1540 #define CONTROLLER_VALID                        \
1541         DIGITS LETTERS                          \
1542         "_"
1543
1544 bool cg_controller_is_valid(const char *p) {
1545         const char *t, *s;
1546
1547         if (!p)
1548                 return false;
1549
1550         if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1551                 return true;
1552
1553         s = startswith(p, "name=");
1554         if (s)
1555                 p = s;
1556
1557         if (IN_SET(*p, 0, '_'))
1558                 return false;
1559
1560         for (t = p; *t; t++)
1561                 if (!strchr(CONTROLLER_VALID, *t))
1562                         return false;
1563
1564         if (t - p > NAME_MAX)
1565                 return false;
1566
1567         return true;
1568 }
1569
1570 int cg_slice_to_path(const char *unit, char **ret) {
1571         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1572         const char *dash;
1573         int r;
1574
1575         assert(unit);
1576         assert(ret);
1577
1578         if (streq(unit, SPECIAL_ROOT_SLICE)) {
1579                 char *x;
1580
1581                 x = strdup("");
1582                 if (!x)
1583                         return -ENOMEM;
1584                 *ret = x;
1585                 return 0;
1586         }
1587
1588         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1589                 return -EINVAL;
1590
1591         if (!endswith(unit, ".slice"))
1592                 return -EINVAL;
1593
1594         r = unit_name_to_prefix(unit, &p);
1595         if (r < 0)
1596                 return r;
1597
1598         dash = strchr(p, '-');
1599
1600         /* Don't allow initial dashes */
1601         if (dash == p)
1602                 return -EINVAL;
1603
1604         while (dash) {
1605                 _cleanup_free_ char *escaped = NULL;
1606                 char n[dash - p + sizeof(".slice")];
1607
1608 #if HAS_FEATURE_MEMORY_SANITIZER
1609                 /* msan doesn't instrument stpncpy, so it thinks
1610                  * n is later used uninitialized:
1611                  * https://github.com/google/sanitizers/issues/926
1612                  */
1613                 zero(n);
1614 #endif
1615
1616                 /* Don't allow trailing or double dashes */
1617                 if (IN_SET(dash[1], 0, '-'))
1618                         return -EINVAL;
1619
1620                 strcpy(stpncpy(n, p, dash - p), ".slice");
1621                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1622                         return -EINVAL;
1623
1624                 escaped = cg_escape(n);
1625                 if (!escaped)
1626                         return -ENOMEM;
1627
1628                 if (!strextend(&s, escaped, "/"))
1629                         return -ENOMEM;
1630
1631                 dash = strchr(dash+1, '-');
1632         }
1633
1634         e = cg_escape(unit);
1635         if (!e)
1636                 return -ENOMEM;
1637
1638         if (!strextend(&s, e))
1639                 return -ENOMEM;
1640
1641         *ret = TAKE_PTR(s);
1642
1643         return 0;
1644 }
1645
1646 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1647         _cleanup_free_ char *p = NULL;
1648         int r;
1649
1650         r = cg_get_path(controller, path, attribute, &p);
1651         if (r < 0)
1652                 return r;
1653
1654         return write_string_file(p, value, WRITE_STRING_FILE_DISABLE_BUFFER);
1655 }
1656
1657 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1658         _cleanup_free_ char *p = NULL;
1659         int r;
1660
1661         r = cg_get_path(controller, path, attribute, &p);
1662         if (r < 0)
1663                 return r;
1664
1665         return read_one_line_file(p, ret);
1666 }
1667
1668 int cg_get_attribute_as_uint64(const char *controller, const char *path, const char *attribute, uint64_t *ret) {
1669         _cleanup_free_ char *value = NULL;
1670         uint64_t v;
1671         int r;
1672
1673         assert(ret);
1674
1675         r = cg_get_attribute(controller, path, attribute, &value);
1676         if (r == -ENOENT)
1677                 return -ENODATA;
1678         if (r < 0)
1679                 return r;
1680
1681         if (streq(value, "max")) {
1682                 *ret = CGROUP_LIMIT_MAX;
1683                 return 0;
1684         }
1685
1686         r = safe_atou64(value, &v);
1687         if (r < 0)
1688                 return r;
1689
1690         *ret = v;
1691         return 0;
1692 }
1693
1694 int cg_get_attribute_as_bool(const char *controller, const char *path, const char *attribute, bool *ret) {
1695         _cleanup_free_ char *value = NULL;
1696         int r;
1697
1698         assert(ret);
1699
1700         r = cg_get_attribute(controller, path, attribute, &value);
1701         if (r == -ENOENT)
1702                 return -ENODATA;
1703         if (r < 0)
1704                 return r;
1705
1706         r = parse_boolean(value);
1707         if (r < 0)
1708                 return r;
1709
1710         *ret = r;
1711         return 0;
1712 }
1713
1714 int cg_get_owner(const char *controller, const char *path, uid_t *ret_uid) {
1715         _cleanup_free_ char *f = NULL;
1716         struct stat stats;
1717         int r;
1718
1719         assert(ret_uid);
1720
1721         r = cg_get_path(controller, path, NULL, &f);
1722         if (r < 0)
1723                 return r;
1724
1725         r = stat(f, &stats);
1726         if (r < 0)
1727                 return -errno;
1728
1729         *ret_uid = stats.st_uid;
1730         return 0;
1731 }
1732
1733 int cg_get_keyed_attribute_full(
1734                 const char *controller,
1735                 const char *path,
1736                 const char *attribute,
1737                 char **keys,
1738                 char **ret_values,
1739                 CGroupKeyMode mode) {
1740
1741         _cleanup_free_ char *filename = NULL, *contents = NULL;
1742         const char *p;
1743         size_t n, i, n_done = 0;
1744         char **v;
1745         int r;
1746
1747         /* Reads one or more fields of a cgroup v2 keyed attribute file. The 'keys' parameter should be an strv with
1748          * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
1749          * entries as 'keys'. On success each entry will be set to the value of the matching key.
1750          *
1751          * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. If mode
1752          * is set to GG_KEY_MODE_GRACEFUL we ignore missing keys and return those that were parsed successfully. */
1753
1754         r = cg_get_path(controller, path, attribute, &filename);
1755         if (r < 0)
1756                 return r;
1757
1758         r = read_full_file(filename, &contents, NULL);
1759         if (r < 0)
1760                 return r;
1761
1762         n = strv_length(keys);
1763         if (n == 0) /* No keys to retrieve? That's easy, we are done then */
1764                 return 0;
1765
1766         /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
1767         v = newa0(char*, n);
1768
1769         for (p = contents; *p;) {
1770                 const char *w = NULL;
1771
1772                 for (i = 0; i < n; i++)
1773                         if (!v[i]) {
1774                                 w = first_word(p, keys[i]);
1775                                 if (w)
1776                                         break;
1777                         }
1778
1779                 if (w) {
1780                         size_t l;
1781
1782                         l = strcspn(w, NEWLINE);
1783                         v[i] = strndup(w, l);
1784                         if (!v[i]) {
1785                                 r = -ENOMEM;
1786                                 goto fail;
1787                         }
1788
1789                         n_done++;
1790                         if (n_done >= n)
1791                                 goto done;
1792
1793                         p = w + l;
1794                 } else
1795                         p += strcspn(p, NEWLINE);
1796
1797                 p += strspn(p, NEWLINE);
1798         }
1799
1800         if (mode & CG_KEY_MODE_GRACEFUL)
1801                 goto done;
1802
1803         r = -ENXIO;
1804
1805 fail:
1806         for (i = 0; i < n; i++)
1807                 free(v[i]);
1808
1809         return r;
1810
1811 done:
1812         memcpy(ret_values, v, sizeof(char*) * n);
1813         if (mode & CG_KEY_MODE_GRACEFUL)
1814                 return n_done;
1815
1816         return 0;
1817 }
1818
1819 int cg_mask_to_string(CGroupMask mask, char **ret) {
1820         _cleanup_free_ char *s = NULL;
1821         size_t n = 0, allocated = 0;
1822         bool space = false;
1823         CGroupController c;
1824
1825         assert(ret);
1826
1827         if (mask == 0) {
1828                 *ret = NULL;
1829                 return 0;
1830         }
1831
1832         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1833                 const char *k;
1834                 size_t l;
1835
1836                 if (!FLAGS_SET(mask, CGROUP_CONTROLLER_TO_MASK(c)))
1837                         continue;
1838
1839                 k = cgroup_controller_to_string(c);
1840                 l = strlen(k);
1841
1842                 if (!GREEDY_REALLOC(s, allocated, n + space + l + 1))
1843                         return -ENOMEM;
1844
1845                 if (space)
1846                         s[n] = ' ';
1847                 memcpy(s + n + space, k, l);
1848                 n += space + l;
1849
1850                 space = true;
1851         }
1852
1853         assert(s);
1854
1855         s[n] = 0;
1856         *ret = TAKE_PTR(s);
1857
1858         return 0;
1859 }
1860
1861 int cg_mask_from_string(const char *value, CGroupMask *ret) {
1862         CGroupMask m = 0;
1863
1864         assert(ret);
1865         assert(value);
1866
1867         for (;;) {
1868                 _cleanup_free_ char *n = NULL;
1869                 CGroupController v;
1870                 int r;
1871
1872                 r = extract_first_word(&value, &n, NULL, 0);
1873                 if (r < 0)
1874                         return r;
1875                 if (r == 0)
1876                         break;
1877
1878                 v = cgroup_controller_from_string(n);
1879                 if (v < 0)
1880                         continue;
1881
1882                 m |= CGROUP_CONTROLLER_TO_MASK(v);
1883         }
1884
1885         *ret = m;
1886         return 0;
1887 }
1888
1889 int cg_mask_supported_subtree(const char *root, CGroupMask *ret) {
1890         CGroupMask mask;
1891         int r;
1892
1893         /* Determines the mask of supported cgroup controllers. Only includes controllers we can make sense of and that
1894          * are actually accessible. Only covers real controllers, i.e. not the CGROUP_CONTROLLER_BPF_xyz
1895          * pseudo-controllers. */
1896
1897         r = cg_all_unified();
1898         if (r < 0)
1899                 return r;
1900         if (r > 0) {
1901                 _cleanup_free_ char *controllers = NULL, *path = NULL;
1902
1903                 /* In the unified hierarchy we can read the supported and accessible controllers from
1904                  * the top-level cgroup attribute */
1905
1906                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
1907                 if (r < 0)
1908                         return r;
1909
1910                 r = read_one_line_file(path, &controllers);
1911                 if (r < 0)
1912                         return r;
1913
1914                 r = cg_mask_from_string(controllers, &mask);
1915                 if (r < 0)
1916                         return r;
1917
1918                 /* Mask controllers that are not supported in unified hierarchy. */
1919                 mask &= CGROUP_MASK_V2;
1920
1921         } else {
1922                 CGroupController c;
1923
1924                 /* In the legacy hierarchy, we check which hierarchies are accessible. */
1925
1926                 mask = 0;
1927                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1928                         CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1929                         const char *n;
1930
1931                         if (!FLAGS_SET(CGROUP_MASK_V1, bit))
1932                                 continue;
1933
1934                         n = cgroup_controller_to_string(c);
1935                         if (controller_is_v1_accessible(root, n) >= 0)
1936                                 mask |= bit;
1937                 }
1938         }
1939
1940         *ret = mask;
1941         return 0;
1942 }
1943
1944 int cg_mask_supported(CGroupMask *ret) {
1945         _cleanup_free_ char *root = NULL;
1946         int r;
1947
1948         r = cg_get_root_path(&root);
1949         if (r < 0)
1950                 return r;
1951
1952         return cg_mask_supported_subtree(root, ret);
1953 }
1954
1955 int cg_kernel_controllers(Set **ret) {
1956         _cleanup_set_free_free_ Set *controllers = NULL;
1957         _cleanup_fclose_ FILE *f = NULL;
1958         int r;
1959
1960         assert(ret);
1961
1962         /* Determines the full list of kernel-known controllers. Might include controllers we don't actually support
1963          * and controllers that aren't currently accessible (because not mounted). This does not include "name="
1964          * pseudo-controllers. */
1965
1966         controllers = set_new(&string_hash_ops);
1967         if (!controllers)
1968                 return -ENOMEM;
1969
1970         r = fopen_unlocked("/proc/cgroups", "re", &f);
1971         if (r == -ENOENT) {
1972                 *ret = NULL;
1973                 return 0;
1974         }
1975         if (r < 0)
1976                 return r;
1977
1978         /* Ignore the header line */
1979         (void) read_line(f, SIZE_MAX, NULL);
1980
1981         for (;;) {
1982                 char *controller;
1983                 int enabled = 0;
1984
1985                 errno = 0;
1986                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
1987
1988                         if (feof(f))
1989                                 break;
1990
1991                         if (ferror(f))
1992                                 return errno_or_else(EIO);
1993
1994                         return -EBADMSG;
1995                 }
1996
1997                 if (!enabled) {
1998                         free(controller);
1999                         continue;
2000                 }
2001
2002                 if (!cg_controller_is_valid(controller)) {
2003                         free(controller);
2004                         return -EBADMSG;
2005                 }
2006
2007                 r = set_consume(controllers, controller);
2008                 if (r < 0)
2009                         return r;
2010         }
2011
2012         *ret = TAKE_PTR(controllers);
2013
2014         return 0;
2015 }
2016
2017 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup2 on
2018  * /sys/fs/cgroup/systemd. This unfortunately broke other tools (such as docker) which expected the v1
2019  * "name=systemd" hierarchy on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mounts v2 on
2020  * /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility
2021  * with other tools.
2022  *
2023  * To keep live upgrade working, we detect and support v232 layout. When v232 layout is detected, to keep
2024  * cgroup v2 process management but disable the compat dual layout, we return true on
2025  * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and false on cg_hybrid_unified().
2026  */
2027 static thread_local bool unified_systemd_v232;
2028
2029 int cg_unified_cached(bool flush) {
2030         static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2031
2032         struct statfs fs;
2033
2034         /* Checks if we support the unified hierarchy. Returns an
2035          * error when the cgroup hierarchies aren't mounted yet or we
2036          * have any other trouble determining if the unified hierarchy
2037          * is supported. */
2038
2039         if (flush)
2040                 unified_cache = CGROUP_UNIFIED_UNKNOWN;
2041         else if (unified_cache >= CGROUP_UNIFIED_NONE)
2042                 return unified_cache;
2043
2044         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2045                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2046
2047         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2048                 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2049                 unified_cache = CGROUP_UNIFIED_ALL;
2050         } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2051                 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2052                     F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2053                         log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2054                         unified_cache = CGROUP_UNIFIED_SYSTEMD;
2055                         unified_systemd_v232 = false;
2056                 } else {
2057                         if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0)
2058                                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2059
2060                         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2061                                 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2062                                 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2063                                 unified_systemd_v232 = true;
2064                         } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2065                                 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2066                                 unified_cache = CGROUP_UNIFIED_NONE;
2067                         } else {
2068                                 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2069                                           (unsigned long long) fs.f_type);
2070                                 unified_cache = CGROUP_UNIFIED_NONE;
2071                         }
2072                 }
2073         } else if (F_TYPE_EQUAL(fs.f_type, SYSFS_MAGIC)) {
2074                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2075                                        "No filesystem is currently mounted on /sys/fs/cgroup.");
2076         } else
2077                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2078                                        "Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2079                                        (unsigned long long)fs.f_type);
2080
2081         return unified_cache;
2082 }
2083
2084 int cg_unified_controller(const char *controller) {
2085         int r;
2086
2087         r = cg_unified_cached(false);
2088         if (r < 0)
2089                 return r;
2090
2091         if (r == CGROUP_UNIFIED_NONE)
2092                 return false;
2093
2094         if (r >= CGROUP_UNIFIED_ALL)
2095                 return true;
2096
2097         return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2098 }
2099
2100 int cg_all_unified(void) {
2101         int r;
2102
2103         r = cg_unified_cached(false);
2104         if (r < 0)
2105                 return r;
2106
2107         return r >= CGROUP_UNIFIED_ALL;
2108 }
2109
2110 int cg_hybrid_unified(void) {
2111         int r;
2112
2113         r = cg_unified_cached(false);
2114         if (r < 0)
2115                 return r;
2116
2117         return r == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2118 }
2119
2120 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2121         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2122         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2123         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2124         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2125 };
2126
2127 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2128         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2129         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2130         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2131         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2132 };
2133
2134 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2135
2136 bool is_cgroup_fs(const struct statfs *s) {
2137         return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2138                is_fs_type(s, CGROUP2_SUPER_MAGIC);
2139 }
2140
2141 bool fd_is_cgroup_fs(int fd) {
2142         struct statfs s;
2143
2144         if (fstatfs(fd, &s) < 0)
2145                 return -errno;
2146
2147         return is_cgroup_fs(&s);
2148 }
2149
2150 static const char *const cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2151         [CGROUP_CONTROLLER_CPU] = "cpu",
2152         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2153         [CGROUP_CONTROLLER_CPUSET] = "cpuset",
2154         [CGROUP_CONTROLLER_IO] = "io",
2155         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2156         [CGROUP_CONTROLLER_MEMORY] = "memory",
2157         [CGROUP_CONTROLLER_DEVICES] = "devices",
2158         [CGROUP_CONTROLLER_PIDS] = "pids",
2159         [CGROUP_CONTROLLER_BPF_FIREWALL] = "bpf-firewall",
2160         [CGROUP_CONTROLLER_BPF_DEVICES] = "bpf-devices",
2161 };
2162
2163 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);
2164
2165 CGroupMask get_cpu_accounting_mask(void) {
2166         static CGroupMask needed_mask = (CGroupMask) -1;
2167
2168         /* On kernel ≥4.15 with unified hierarchy, cpu.stat's usage_usec is
2169          * provided externally from the CPU controller, which means we don't
2170          * need to enable the CPU controller just to get metrics. This is good,
2171          * because enabling the CPU controller comes at a minor performance
2172          * hit, especially when it's propagated deep into large hierarchies.
2173          * There's also no separate CPU accounting controller available within
2174          * a unified hierarchy.
2175          *
2176          * This combination of factors results in the desired cgroup mask to
2177          * enable for CPU accounting varying as follows:
2178          *
2179          *                   ╔═════════════════════╤═════════════════════╗
2180          *                   ║     Linux ≥4.15     │     Linux <4.15     ║
2181          *   ╔═══════════════╬═════════════════════╪═════════════════════╣
2182          *   ║ Unified       ║ nothing             │ CGROUP_MASK_CPU     ║
2183          *   ╟───────────────╫─────────────────────┼─────────────────────╢
2184          *   ║ Hybrid/Legacy ║ CGROUP_MASK_CPUACCT │ CGROUP_MASK_CPUACCT ║
2185          *   ╚═══════════════╩═════════════════════╧═════════════════════╝
2186          *
2187          * We check kernel version here instead of manually checking whether
2188          * cpu.stat is present for every cgroup, as that check in itself would
2189          * already be fairly expensive.
2190          *
2191          * Kernels where this patch has been backported will therefore have the
2192          * CPU controller enabled unnecessarily. This is more expensive than
2193          * necessary, but harmless. ☺️
2194          */
2195
2196         if (needed_mask == (CGroupMask) -1) {
2197                 if (cg_all_unified()) {
2198                         struct utsname u;
2199                         assert_se(uname(&u) >= 0);
2200
2201                         if (strverscmp_improved(u.release, "4.15") < 0)
2202                                 needed_mask = CGROUP_MASK_CPU;
2203                         else
2204                                 needed_mask = 0;
2205                 } else
2206                         needed_mask = CGROUP_MASK_CPUACCT;
2207         }
2208
2209         return needed_mask;
2210 }
2211
2212 bool cpu_accounting_is_cheap(void) {
2213         return get_cpu_accounting_mask() == 0;
2214 }
2215
2216 static const char* const managed_oom_mode_table[_MANAGED_OOM_MODE_MAX] = {
2217         [MANAGED_OOM_AUTO] = "auto",
2218         [MANAGED_OOM_KILL] = "kill",
2219 };
2220
2221 DEFINE_STRING_TABLE_LOOKUP(managed_oom_mode, ManagedOOMMode);
2222
2223 static const char* const managed_oom_preference_table[_MANAGED_OOM_PREFERENCE_MAX] = {
2224         [MANAGED_OOM_PREFERENCE_NONE] = "none",
2225         [MANAGED_OOM_PREFERENCE_AVOID] = "avoid",
2226         [MANAGED_OOM_PREFERENCE_OMIT] = "omit",
2227 };
2228
2229 DEFINE_STRING_TABLE_LOOKUP(managed_oom_preference, ManagedOOMPreference);