src/basic/cgroup-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   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_accessible(const char *controller) {
 531         int r;
 532
 533         assert(controller);
 534
 535         /* Checks whether a specific controller is accessible,
 536          * i.e. its hierarchy mounted. In the unified hierarchy all
 537          * controllers are considered accessible, except for the named
 538          * hierarchies */
 539
 540         if (!cg_controller_is_valid(controller))
 541                 return -EINVAL;
 542
 543         r = cg_all_unified();
 544         if (r < 0)
 545                 return r;
 546         if (r > 0) {
 547                 /* We don't support named hierarchies if we are using
 548                  * the unified hierarchy. */
 549
 550                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 551                         return 0;
 552
 553                 if (startswith(controller, "name="))
 554                         return -EOPNOTSUPP;
 555
 556         } else {
 557                 const char *cc, *dn;
 558
 559                 dn = controller_to_dirname(controller);
 560                 cc = strjoina("/sys/fs/cgroup/", dn);
 561
 562                 if (laccess(cc, F_OK) < 0)
 563                         return -errno;
 564         }
 565
 566         return 0;
 567 }
 568
 569 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 570         int r;
 571
 572         assert(controller);
 573         assert(fs);
 574
 575         /* Check if the specified controller is actually accessible */
 576         r = controller_is_accessible(controller);
 577         if (r < 0)
 578                 return r;
 579
 580         return cg_get_path(controller, path, suffix, fs);
 581 }
 582
 583 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags) {
 584         _cleanup_free_ char *fs = NULL;
 585         int r;
 586
 587         assert(path);
 588         assert(name);
 589         assert(value || size <= 0);
 590
 591         r = cg_get_path(controller, path, NULL, &fs);
 592         if (r < 0)
 593                 return r;
 594
 595         if (setxattr(fs, name, value, size, flags) < 0)
 596                 return -errno;
 597
 598         return 0;
 599 }
 600
 601 int cg_get_xattr(const char *controller, const char *path, const char *name, void *value, size_t size) {
 602         _cleanup_free_ char *fs = NULL;
 603         ssize_t n;
 604         int r;
 605
 606         assert(path);
 607         assert(name);
 608
 609         r = cg_get_path(controller, path, NULL, &fs);
 610         if (r < 0)
 611                 return r;
 612
 613         n = getxattr(fs, name, value, size);
 614         if (n < 0)
 615                 return -errno;
 616
 617         return (int) n;
 618 }
 619
 620 int cg_get_xattr_malloc(const char *controller, const char *path, const char *name, char **ret) {
 621         _cleanup_free_ char *fs = NULL;
 622         int r;
 623
 624         assert(path);
 625         assert(name);
 626
 627         r = cg_get_path(controller, path, NULL, &fs);
 628         if (r < 0)
 629                 return r;
 630
 631         r = getxattr_malloc(fs, name, ret, false);
 632         if (r < 0)
 633                 return r;
 634
 635         return r;
 636 }
 637
 638 int cg_remove_xattr(const char *controller, const char *path, const char *name) {
 639         _cleanup_free_ char *fs = NULL;
 640         int r;
 641
 642         assert(path);
 643         assert(name);
 644
 645         r = cg_get_path(controller, path, NULL, &fs);
 646         if (r < 0)
 647                 return r;
 648
 649         if (removexattr(fs, name) < 0)
 650                 return -errno;
 651
 652         return 0;
 653 }
 654
 655 int cg_pid_get_path(const char *controller, pid_t pid, char **path) {
 656         _cleanup_fclose_ FILE *f = NULL;
 657         const char *fs, *controller_str;
 658         int unified, r;
 659         size_t cs = 0;
 660
 661         assert(path);
 662         assert(pid >= 0);
 663
 664         if (controller) {
 665                 if (!cg_controller_is_valid(controller))
 666                         return -EINVAL;
 667         } else
 668                 controller = SYSTEMD_CGROUP_CONTROLLER;
 669
 670         unified = cg_unified_controller(controller);
 671         if (unified < 0)
 672                 return unified;
 673         if (unified == 0) {
 674                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 675                         controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 676                 else
 677                         controller_str = controller;
 678
 679                 cs = strlen(controller_str);
 680         }
 681
 682         fs = procfs_file_alloca(pid, "cgroup");
 683         r = fopen_unlocked(fs, "re", &f);
 684         if (r == -ENOENT)
 685                 return -ESRCH;
 686         if (r < 0)
 687                 return r;
 688
 689         for (;;) {
 690                 _cleanup_free_ char *line = NULL;
 691                 char *e, *p;
 692
 693                 r = read_line(f, LONG_LINE_MAX, &line);
 694                 if (r < 0)
 695                         return r;
 696                 if (r == 0)
 697                         break;
 698
 699                 if (unified) {
 700                         e = startswith(line, "0:");
 701                         if (!e)
 702                                 continue;
 703
 704                         e = strchr(e, ':');
 705                         if (!e)
 706                                 continue;
 707                 } else {
 708                         char *l;
 709                         size_t k;
 710                         const char *word, *state;
 711                         bool found = false;
 712
 713                         l = strchr(line, ':');
 714                         if (!l)
 715                                 continue;
 716
 717                         l++;
 718                         e = strchr(l, ':');
 719                         if (!e)
 720                                 continue;
 721
 722                         *e = 0;
 723                         FOREACH_WORD_SEPARATOR(word, k, l, ",", state)
 724                                 if (k == cs && memcmp(word, controller_str, cs) == 0) {
 725                                         found = true;
 726                                         break;
 727                                 }
 728                         if (!found)
 729                                 continue;
 730                 }
 731
 732                 p = strdup(e + 1);
 733                 if (!p)
 734                         return -ENOMEM;
 735
 736                 /* Truncate suffix indicating the process is a zombie */
 737                 e = endswith(p, " (deleted)");
 738                 if (e)
 739                         *e = 0;
 740
 741                 *path = p;
 742                 return 0;
 743         }
 744
 745         return -ENODATA;
 746 }
 747
 748 int cg_install_release_agent(const char *controller, const char *agent) {
 749         _cleanup_free_ char *fs = NULL, *contents = NULL;
 750         const char *sc;
 751         int r;
 752
 753         assert(agent);
 754
 755         r = cg_unified_controller(controller);
 756         if (r < 0)
 757                 return r;
 758         if (r > 0) /* doesn't apply to unified hierarchy */
 759                 return -EOPNOTSUPP;
 760
 761         r = cg_get_path(controller, NULL, "release_agent", &fs);
 762         if (r < 0)
 763                 return r;
 764
 765         r = read_one_line_file(fs, &contents);
 766         if (r < 0)
 767                 return r;
 768
 769         sc = strstrip(contents);
 770         if (isempty(sc)) {
 771                 r = write_string_file(fs, agent, WRITE_STRING_FILE_DISABLE_BUFFER);
 772                 if (r < 0)
 773                         return r;
 774         } else if (!path_equal(sc, agent))
 775                 return -EEXIST;
 776
 777         fs = mfree(fs);
 778         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 779         if (r < 0)
 780                 return r;
 781
 782         contents = mfree(contents);
 783         r = read_one_line_file(fs, &contents);
 784         if (r < 0)
 785                 return r;
 786
 787         sc = strstrip(contents);
 788         if (streq(sc, "0")) {
 789                 r = write_string_file(fs, "1", WRITE_STRING_FILE_DISABLE_BUFFER);
 790                 if (r < 0)
 791                         return r;
 792
 793                 return 1;
 794         }
 795
 796         if (!streq(sc, "1"))
 797                 return -EIO;
 798
 799         return 0;
 800 }
 801
 802 int cg_uninstall_release_agent(const char *controller) {
 803         _cleanup_free_ char *fs = NULL;
 804         int r;
 805
 806         r = cg_unified_controller(controller);
 807         if (r < 0)
 808                 return r;
 809         if (r > 0) /* Doesn't apply to unified hierarchy */
 810                 return -EOPNOTSUPP;
 811
 812         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 813         if (r < 0)
 814                 return r;
 815
 816         r = write_string_file(fs, "0", WRITE_STRING_FILE_DISABLE_BUFFER);
 817         if (r < 0)
 818                 return r;
 819
 820         fs = mfree(fs);
 821
 822         r = cg_get_path(controller, NULL, "release_agent", &fs);
 823         if (r < 0)
 824                 return r;
 825
 826         r = write_string_file(fs, "", WRITE_STRING_FILE_DISABLE_BUFFER);
 827         if (r < 0)
 828                 return r;
 829
 830         return 0;
 831 }
 832
 833 int cg_is_empty(const char *controller, const char *path) {
 834         _cleanup_fclose_ FILE *f = NULL;
 835         pid_t pid;
 836         int r;
 837
 838         assert(path);
 839
 840         r = cg_enumerate_processes(controller, path, &f);
 841         if (r == -ENOENT)
 842                 return true;
 843         if (r < 0)
 844                 return r;
 845
 846         r = cg_read_pid(f, &pid);
 847         if (r < 0)
 848                 return r;
 849
 850         return r == 0;
 851 }
 852
 853 int cg_is_empty_recursive(const char *controller, const char *path) {
 854         int r;
 855
 856         assert(path);
 857
 858         /* The root cgroup is always populated */
 859         if (controller && empty_or_root(path))
 860                 return false;
 861
 862         r = cg_unified_controller(controller);
 863         if (r < 0)
 864                 return r;
 865         if (r > 0) {
 866                 _cleanup_free_ char *t = NULL;
 867
 868                 /* On the unified hierarchy we can check empty state
 869                  * via the "populated" attribute of "cgroup.events". */
 870
 871                 r = cg_read_event(controller, path, "populated", &t);
 872                 if (r == -ENOENT)
 873                         return true;
 874                 if (r < 0)
 875                         return r;
 876
 877                 return streq(t, "0");
 878         } else {
 879                 _cleanup_closedir_ DIR *d = NULL;
 880                 char *fn;
 881
 882                 r = cg_is_empty(controller, path);
 883                 if (r <= 0)
 884                         return r;
 885
 886                 r = cg_enumerate_subgroups(controller, path, &d);
 887                 if (r == -ENOENT)
 888                         return true;
 889                 if (r < 0)
 890                         return r;
 891
 892                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
 893                         _cleanup_free_ char *p = NULL;
 894
 895                         p = path_join(path, fn);
 896                         free(fn);
 897                         if (!p)
 898                                 return -ENOMEM;
 899
 900                         r = cg_is_empty_recursive(controller, p);
 901                         if (r <= 0)
 902                                 return r;
 903                 }
 904                 if (r < 0)
 905                         return r;
 906
 907                 return true;
 908         }
 909 }
 910
 911 int cg_split_spec(const char *spec, char **ret_controller, char **ret_path) {
 912         _cleanup_free_ char *controller = NULL, *path = NULL;
 913
 914         assert(spec);
 915
 916         if (*spec == '/') {
 917                 if (!path_is_normalized(spec))
 918                         return -EINVAL;
 919
 920                 if (ret_path) {
 921                         path = strdup(spec);
 922                         if (!path)
 923                                 return -ENOMEM;
 924
 925                         path_simplify(path, false);
 926                 }
 927
 928         } else {
 929                 const char *e;
 930
 931                 e = strchr(spec, ':');
 932                 if (e) {
 933                         controller = strndup(spec, e-spec);
 934                         if (!controller)
 935                                 return -ENOMEM;
 936                         if (!cg_controller_is_valid(controller))
 937                                 return -EINVAL;
 938
 939                         if (!isempty(e + 1)) {
 940                                 path = strdup(e+1);
 941                                 if (!path)
 942                                         return -ENOMEM;
 943
 944                                 if (!path_is_normalized(path) ||
 945                                     !path_is_absolute(path))
 946                                         return -EINVAL;
 947
 948                                 path_simplify(path, false);
 949                         }
 950
 951                 } else {
 952                         if (!cg_controller_is_valid(spec))
 953                                 return -EINVAL;
 954
 955                         if (ret_controller) {
 956                                 controller = strdup(spec);
 957                                 if (!controller)
 958                                         return -ENOMEM;
 959                         }
 960                 }
 961         }
 962
 963         if (ret_controller)
 964                 *ret_controller = TAKE_PTR(controller);
 965         if (ret_path)
 966                 *ret_path = TAKE_PTR(path);
 967         return 0;
 968 }
 969
 970 int cg_mangle_path(const char *path, char **result) {
 971         _cleanup_free_ char *c = NULL, *p = NULL;
 972         char *t;
 973         int r;
 974
 975         assert(path);
 976         assert(result);
 977
 978         /* First, check if it already is a filesystem path */
 979         if (path_startswith(path, "/sys/fs/cgroup")) {
 980
 981                 t = strdup(path);
 982                 if (!t)
 983                         return -ENOMEM;
 984
 985                 *result = path_simplify(t, false);
 986                 return 0;
 987         }
 988
 989         /* Otherwise, treat it as cg spec */
 990         r = cg_split_spec(path, &c, &p);
 991         if (r < 0)
 992                 return r;
 993
 994         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
 995 }
 996
 997 int cg_get_root_path(char **path) {
 998         char *p, *e;
 999         int r;
1000
1001         assert(path);
1002
1003         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1004         if (r < 0)
1005                 return r;
1006
1007         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1008         if (!e)
1009                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1010         if (!e)
1011                 e = endswith(p, "/system"); /* even more legacy */
1012         if (e)
1013                 *e = 0;
1014
1015         *path = p;
1016         return 0;
1017 }
1018
1019 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1020         _cleanup_free_ char *rt = NULL;
1021         char *p;
1022         int r;
1023
1024         assert(cgroup);
1025         assert(shifted);
1026
1027         if (!root) {
1028                 /* If the root was specified let's use that, otherwise
1029                  * let's determine it from PID 1 */
1030
1031                 r = cg_get_root_path(&rt);
1032                 if (r < 0)
1033                         return r;
1034
1035                 root = rt;
1036         }
1037
1038         p = path_startswith(cgroup, root);
1039         if (p && p > cgroup)
1040                 *shifted = p - 1;
1041         else
1042                 *shifted = cgroup;
1043
1044         return 0;
1045 }
1046
1047 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1048         _cleanup_free_ char *raw = NULL;
1049         const char *c;
1050         int r;
1051
1052         assert(pid >= 0);
1053         assert(cgroup);
1054
1055         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1056         if (r < 0)
1057                 return r;
1058
1059         r = cg_shift_path(raw, root, &c);
1060         if (r < 0)
1061                 return r;
1062
1063         if (c == raw)
1064                 *cgroup = TAKE_PTR(raw);
1065         else {
1066                 char *n;
1067
1068                 n = strdup(c);
1069                 if (!n)
1070                         return -ENOMEM;
1071
1072                 *cgroup = n;
1073         }
1074
1075         return 0;
1076 }
1077
1078 int cg_path_decode_unit(const char *cgroup, char **unit) {
1079         char *c, *s;
1080         size_t n;
1081
1082         assert(cgroup);
1083         assert(unit);
1084
1085         n = strcspn(cgroup, "/");
1086         if (n < 3)
1087                 return -ENXIO;
1088
1089         c = strndupa(cgroup, n);
1090         c = cg_unescape(c);
1091
1092         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1093                 return -ENXIO;
1094
1095         s = strdup(c);
1096         if (!s)
1097                 return -ENOMEM;
1098
1099         *unit = s;
1100         return 0;
1101 }
1102
1103 static bool valid_slice_name(const char *p, size_t n) {
1104
1105         if (!p)
1106                 return false;
1107
1108         if (n < STRLEN("x.slice"))
1109                 return false;
1110
1111         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1112                 char buf[n+1], *c;
1113
1114                 memcpy(buf, p, n);
1115                 buf[n] = 0;
1116
1117                 c = cg_unescape(buf);
1118
1119                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1120         }
1121
1122         return false;
1123 }
1124
1125 static const char *skip_slices(const char *p) {
1126         assert(p);
1127
1128         /* Skips over all slice assignments */
1129
1130         for (;;) {
1131                 size_t n;
1132
1133                 p += strspn(p, "/");
1134
1135                 n = strcspn(p, "/");
1136                 if (!valid_slice_name(p, n))
1137                         return p;
1138
1139                 p += n;
1140         }
1141 }
1142
1143 int cg_path_get_unit(const char *path, char **ret) {
1144         const char *e;
1145         char *unit;
1146         int r;
1147
1148         assert(path);
1149         assert(ret);
1150
1151         e = skip_slices(path);
1152
1153         r = cg_path_decode_unit(e, &unit);
1154         if (r < 0)
1155                 return r;
1156
1157         /* We skipped over the slices, don't accept any now */
1158         if (endswith(unit, ".slice")) {
1159                 free(unit);
1160                 return -ENXIO;
1161         }
1162
1163         *ret = unit;
1164         return 0;
1165 }
1166
1167 int cg_pid_get_unit(pid_t pid, char **unit) {
1168         _cleanup_free_ char *cgroup = NULL;
1169         int r;
1170
1171         assert(unit);
1172
1173         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1174         if (r < 0)
1175                 return r;
1176
1177         return cg_path_get_unit(cgroup, unit);
1178 }
1179
1180 /**
1181  * Skip session-*.scope, but require it to be there.
1182  */
1183 static const char *skip_session(const char *p) {
1184         size_t n;
1185
1186         if (isempty(p))
1187                 return NULL;
1188
1189         p += strspn(p, "/");
1190
1191         n = strcspn(p, "/");
1192         if (n < STRLEN("session-x.scope"))
1193                 return NULL;
1194
1195         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1196                 char buf[n - 8 - 6 + 1];
1197
1198                 memcpy(buf, p + 8, n - 8 - 6);
1199                 buf[n - 8 - 6] = 0;
1200
1201                 /* Note that session scopes never need unescaping,
1202                  * since they cannot conflict with the kernel's own
1203                  * names, hence we don't need to call cg_unescape()
1204                  * here. */
1205
1206                 if (!session_id_valid(buf))
1207                         return false;
1208
1209                 p += n;
1210                 p += strspn(p, "/");
1211                 return p;
1212         }
1213
1214         return NULL;
1215 }
1216
1217 /**
1218  * Skip user@*.service, but require it to be there.
1219  */
1220 static const char *skip_user_manager(const char *p) {
1221         size_t n;
1222
1223         if (isempty(p))
1224                 return NULL;
1225
1226         p += strspn(p, "/");
1227
1228         n = strcspn(p, "/");
1229         if (n < STRLEN("user@x.service"))
1230                 return NULL;
1231
1232         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1233                 char buf[n - 5 - 8 + 1];
1234
1235                 memcpy(buf, p + 5, n - 5 - 8);
1236                 buf[n - 5 - 8] = 0;
1237
1238                 /* Note that user manager services never need unescaping,
1239                  * since they cannot conflict with the kernel's own
1240                  * names, hence we don't need to call cg_unescape()
1241                  * here. */
1242
1243                 if (parse_uid(buf, NULL) < 0)
1244                         return NULL;
1245
1246                 p += n;
1247                 p += strspn(p, "/");
1248
1249                 return p;
1250         }
1251
1252         return NULL;
1253 }
1254
1255 static const char *skip_user_prefix(const char *path) {
1256         const char *e, *t;
1257
1258         assert(path);
1259
1260         /* Skip slices, if there are any */
1261         e = skip_slices(path);
1262
1263         /* Skip the user manager, if it's in the path now... */
1264         t = skip_user_manager(e);
1265         if (t)
1266                 return t;
1267
1268         /* Alternatively skip the user session if it is in the path... */
1269         return skip_session(e);
1270 }
1271
1272 int cg_path_get_user_unit(const char *path, char **ret) {
1273         const char *t;
1274
1275         assert(path);
1276         assert(ret);
1277
1278         t = skip_user_prefix(path);
1279         if (!t)
1280                 return -ENXIO;
1281
1282         /* And from here on it looks pretty much the same as for a system unit, hence let's use the same
1283          * parser. */
1284         return cg_path_get_unit(t, ret);
1285 }
1286
1287 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1288         _cleanup_free_ char *cgroup = NULL;
1289         int r;
1290
1291         assert(unit);
1292
1293         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1294         if (r < 0)
1295                 return r;
1296
1297         return cg_path_get_user_unit(cgroup, unit);
1298 }
1299
1300 int cg_path_get_machine_name(const char *path, char **machine) {
1301         _cleanup_free_ char *u = NULL;
1302         const char *sl;
1303         int r;
1304
1305         r = cg_path_get_unit(path, &u);
1306         if (r < 0)
1307                 return r;
1308
1309         sl = strjoina("/run/systemd/machines/unit:", u);
1310         return readlink_malloc(sl, machine);
1311 }
1312
1313 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1314         _cleanup_free_ char *cgroup = NULL;
1315         int r;
1316
1317         assert(machine);
1318
1319         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1320         if (r < 0)
1321                 return r;
1322
1323         return cg_path_get_machine_name(cgroup, machine);
1324 }
1325
1326 int cg_path_get_session(const char *path, char **session) {
1327         _cleanup_free_ char *unit = NULL;
1328         char *start, *end;
1329         int r;
1330
1331         assert(path);
1332
1333         r = cg_path_get_unit(path, &unit);
1334         if (r < 0)
1335                 return r;
1336
1337         start = startswith(unit, "session-");
1338         if (!start)
1339                 return -ENXIO;
1340         end = endswith(start, ".scope");
1341         if (!end)
1342                 return -ENXIO;
1343
1344         *end = 0;
1345         if (!session_id_valid(start))
1346                 return -ENXIO;
1347
1348         if (session) {
1349                 char *rr;
1350
1351                 rr = strdup(start);
1352                 if (!rr)
1353                         return -ENOMEM;
1354
1355                 *session = rr;
1356         }
1357
1358         return 0;
1359 }
1360
1361 int cg_pid_get_session(pid_t pid, char **session) {
1362         _cleanup_free_ char *cgroup = NULL;
1363         int r;
1364
1365         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1366         if (r < 0)
1367                 return r;
1368
1369         return cg_path_get_session(cgroup, session);
1370 }
1371
1372 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1373         _cleanup_free_ char *slice = NULL;
1374         char *start, *end;
1375         int r;
1376
1377         assert(path);
1378
1379         r = cg_path_get_slice(path, &slice);
1380         if (r < 0)
1381                 return r;
1382
1383         start = startswith(slice, "user-");
1384         if (!start)
1385                 return -ENXIO;
1386         end = endswith(start, ".slice");
1387         if (!end)
1388                 return -ENXIO;
1389
1390         *end = 0;
1391         if (parse_uid(start, uid) < 0)
1392                 return -ENXIO;
1393
1394         return 0;
1395 }
1396
1397 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1398         _cleanup_free_ char *cgroup = NULL;
1399         int r;
1400
1401         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1402         if (r < 0)
1403                 return r;
1404
1405         return cg_path_get_owner_uid(cgroup, uid);
1406 }
1407
1408 int cg_path_get_slice(const char *p, char **slice) {
1409         const char *e = NULL;
1410
1411         assert(p);
1412         assert(slice);
1413
1414         /* Finds the right-most slice unit from the beginning, but
1415          * stops before we come to the first non-slice unit. */
1416
1417         for (;;) {
1418                 size_t n;
1419
1420                 p += strspn(p, "/");
1421
1422                 n = strcspn(p, "/");
1423                 if (!valid_slice_name(p, n)) {
1424
1425                         if (!e) {
1426                                 char *s;
1427
1428                                 s = strdup(SPECIAL_ROOT_SLICE);
1429                                 if (!s)
1430                                         return -ENOMEM;
1431
1432                                 *slice = s;
1433                                 return 0;
1434                         }
1435
1436                         return cg_path_decode_unit(e, slice);
1437                 }
1438
1439                 e = p;
1440                 p += n;
1441         }
1442 }
1443
1444 int cg_pid_get_slice(pid_t pid, char **slice) {
1445         _cleanup_free_ char *cgroup = NULL;
1446         int r;
1447
1448         assert(slice);
1449
1450         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1451         if (r < 0)
1452                 return r;
1453
1454         return cg_path_get_slice(cgroup, slice);
1455 }
1456
1457 int cg_path_get_user_slice(const char *p, char **slice) {
1458         const char *t;
1459         assert(p);
1460         assert(slice);
1461
1462         t = skip_user_prefix(p);
1463         if (!t)
1464                 return -ENXIO;
1465
1466         /* And now it looks pretty much the same as for a system
1467          * slice, so let's just use the same parser from here on. */
1468         return cg_path_get_slice(t, slice);
1469 }
1470
1471 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1472         _cleanup_free_ char *cgroup = NULL;
1473         int r;
1474
1475         assert(slice);
1476
1477         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1478         if (r < 0)
1479                 return r;
1480
1481         return cg_path_get_user_slice(cgroup, slice);
1482 }
1483
1484 char *cg_escape(const char *p) {
1485         bool need_prefix = false;
1486
1487         /* This implements very minimal escaping for names to be used
1488          * as file names in the cgroup tree: any name which might
1489          * conflict with a kernel name or is prefixed with '_' is
1490          * prefixed with a '_'. That way, when reading cgroup names it
1491          * is sufficient to remove a single prefixing underscore if
1492          * there is one. */
1493
1494         /* The return value of this function (unlike cg_unescape())
1495          * needs free()! */
1496
1497         if (IN_SET(p[0], 0, '_', '.') ||
1498             STR_IN_SET(p, "notify_on_release", "release_agent", "tasks") ||
1499             startswith(p, "cgroup."))
1500                 need_prefix = true;
1501         else {
1502                 const char *dot;
1503
1504                 dot = strrchr(p, '.');
1505                 if (dot) {
1506                         CGroupController c;
1507                         size_t l = dot - p;
1508
1509                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1510                                 const char *n;
1511
1512                                 n = cgroup_controller_to_string(c);
1513
1514                                 if (l != strlen(n))
1515                                         continue;
1516
1517                                 if (memcmp(p, n, l) != 0)
1518                                         continue;
1519
1520                                 need_prefix = true;
1521                                 break;
1522                         }
1523                 }
1524         }
1525
1526         if (need_prefix)
1527                 return strjoin("_", p);
1528
1529         return strdup(p);
1530 }
1531
1532 char *cg_unescape(const char *p) {
1533         assert(p);
1534
1535         /* The return value of this function (unlike cg_escape())
1536          * doesn't need free()! */
1537
1538         if (p[0] == '_')
1539                 return (char*) p+1;
1540
1541         return (char*) p;
1542 }
1543
1544 #define CONTROLLER_VALID                        \
1545         DIGITS LETTERS                          \
1546         "_"
1547
1548 bool cg_controller_is_valid(const char *p) {
1549         const char *t, *s;
1550
1551         if (!p)
1552                 return false;
1553
1554         if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1555                 return true;
1556
1557         s = startswith(p, "name=");
1558         if (s)
1559                 p = s;
1560
1561         if (IN_SET(*p, 0, '_'))
1562                 return false;
1563
1564         for (t = p; *t; t++)
1565                 if (!strchr(CONTROLLER_VALID, *t))
1566                         return false;
1567
1568         if (t - p > FILENAME_MAX)
1569                 return false;
1570
1571         return true;
1572 }
1573
1574 int cg_slice_to_path(const char *unit, char **ret) {
1575         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1576         const char *dash;
1577         int r;
1578
1579         assert(unit);
1580         assert(ret);
1581
1582         if (streq(unit, SPECIAL_ROOT_SLICE)) {
1583                 char *x;
1584
1585                 x = strdup("");
1586                 if (!x)
1587                         return -ENOMEM;
1588                 *ret = x;
1589                 return 0;
1590         }
1591
1592         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1593                 return -EINVAL;
1594
1595         if (!endswith(unit, ".slice"))
1596                 return -EINVAL;
1597
1598         r = unit_name_to_prefix(unit, &p);
1599         if (r < 0)
1600                 return r;
1601
1602         dash = strchr(p, '-');
1603
1604         /* Don't allow initial dashes */
1605         if (dash == p)
1606                 return -EINVAL;
1607
1608         while (dash) {
1609                 _cleanup_free_ char *escaped = NULL;
1610                 char n[dash - p + sizeof(".slice")];
1611
1612 #if HAS_FEATURE_MEMORY_SANITIZER
1613                 /* msan doesn't instrument stpncpy, so it thinks
1614                  * n is later used uninitialized:
1615                  * https://github.com/google/sanitizers/issues/926
1616                  */
1617                 zero(n);
1618 #endif
1619
1620                 /* Don't allow trailing or double dashes */
1621                 if (IN_SET(dash[1], 0, '-'))
1622                         return -EINVAL;
1623
1624                 strcpy(stpncpy(n, p, dash - p), ".slice");
1625                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1626                         return -EINVAL;
1627
1628                 escaped = cg_escape(n);
1629                 if (!escaped)
1630                         return -ENOMEM;
1631
1632                 if (!strextend(&s, escaped, "/", NULL))
1633                         return -ENOMEM;
1634
1635                 dash = strchr(dash+1, '-');
1636         }
1637
1638         e = cg_escape(unit);
1639         if (!e)
1640                 return -ENOMEM;
1641
1642         if (!strextend(&s, e, NULL))
1643                 return -ENOMEM;
1644
1645         *ret = TAKE_PTR(s);
1646
1647         return 0;
1648 }
1649
1650 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1651         _cleanup_free_ char *p = NULL;
1652         int r;
1653
1654         r = cg_get_path(controller, path, attribute, &p);
1655         if (r < 0)
1656                 return r;
1657
1658         return write_string_file(p, value, WRITE_STRING_FILE_DISABLE_BUFFER);
1659 }
1660
1661 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1662         _cleanup_free_ char *p = NULL;
1663         int r;
1664
1665         r = cg_get_path(controller, path, attribute, &p);
1666         if (r < 0)
1667                 return r;
1668
1669         return read_one_line_file(p, ret);
1670 }
1671
1672 int cg_get_attribute_as_uint64(const char *controller, const char *path, const char *attribute, uint64_t *ret) {
1673         _cleanup_free_ char *value = NULL;
1674         uint64_t v;
1675         int r;
1676
1677         assert(ret);
1678
1679         r = cg_get_attribute(controller, path, attribute, &value);
1680         if (r == -ENOENT)
1681                 return -ENODATA;
1682         if (r < 0)
1683                 return r;
1684
1685         if (streq(value, "max")) {
1686                 *ret = CGROUP_LIMIT_MAX;
1687                 return 0;
1688         }
1689
1690         r = safe_atou64(value, &v);
1691         if (r < 0)
1692                 return r;
1693
1694         *ret = v;
1695         return 0;
1696 }
1697
1698 int cg_get_keyed_attribute_full(
1699                 const char *controller,
1700                 const char *path,
1701                 const char *attribute,
1702                 char **keys,
1703                 char **ret_values,
1704                 CGroupKeyMode mode) {
1705
1706         _cleanup_free_ char *filename = NULL, *contents = NULL;
1707         const char *p;
1708         size_t n, i, n_done = 0;
1709         char **v;
1710         int r;
1711
1712         /* Reads one or more fields of a cgroup v2 keyed attribute file. The 'keys' parameter should be an strv with
1713          * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
1714          * entries as 'keys'. On success each entry will be set to the value of the matching key.
1715          *
1716          * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. If mode
1717          * is set to GG_KEY_MODE_GRACEFUL we ignore missing keys and return those that were parsed successfully. */
1718
1719         r = cg_get_path(controller, path, attribute, &filename);
1720         if (r < 0)
1721                 return r;
1722
1723         r = read_full_file(filename, &contents, NULL);
1724         if (r < 0)
1725                 return r;
1726
1727         n = strv_length(keys);
1728         if (n == 0) /* No keys to retrieve? That's easy, we are done then */
1729                 return 0;
1730
1731         /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
1732         v = newa0(char*, n);
1733
1734         for (p = contents; *p;) {
1735                 const char *w = NULL;
1736
1737                 for (i = 0; i < n; i++)
1738                         if (!v[i]) {
1739                                 w = first_word(p, keys[i]);
1740                                 if (w)
1741                                         break;
1742                         }
1743
1744                 if (w) {
1745                         size_t l;
1746
1747                         l = strcspn(w, NEWLINE);
1748                         v[i] = strndup(w, l);
1749                         if (!v[i]) {
1750                                 r = -ENOMEM;
1751                                 goto fail;
1752                         }
1753
1754                         n_done++;
1755                         if (n_done >= n)
1756                                 goto done;
1757
1758                         p = w + l;
1759                 } else
1760                         p += strcspn(p, NEWLINE);
1761
1762                 p += strspn(p, NEWLINE);
1763         }
1764
1765         if (mode & CG_KEY_MODE_GRACEFUL)
1766                 goto done;
1767
1768         r = -ENXIO;
1769
1770 fail:
1771         for (i = 0; i < n; i++)
1772                 free(v[i]);
1773
1774         return r;
1775
1776 done:
1777         memcpy(ret_values, v, sizeof(char*) * n);
1778         if (mode & CG_KEY_MODE_GRACEFUL)
1779                 return n_done;
1780
1781         return 0;
1782 }
1783
1784 int cg_mask_to_string(CGroupMask mask, char **ret) {
1785         _cleanup_free_ char *s = NULL;
1786         size_t n = 0, allocated = 0;
1787         bool space = false;
1788         CGroupController c;
1789
1790         assert(ret);
1791
1792         if (mask == 0) {
1793                 *ret = NULL;
1794                 return 0;
1795         }
1796
1797         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1798                 const char *k;
1799                 size_t l;
1800
1801                 if (!FLAGS_SET(mask, CGROUP_CONTROLLER_TO_MASK(c)))
1802                         continue;
1803
1804                 k = cgroup_controller_to_string(c);
1805                 l = strlen(k);
1806
1807                 if (!GREEDY_REALLOC(s, allocated, n + space + l + 1))
1808                         return -ENOMEM;
1809
1810                 if (space)
1811                         s[n] = ' ';
1812                 memcpy(s + n + space, k, l);
1813                 n += space + l;
1814
1815                 space = true;
1816         }
1817
1818         assert(s);
1819
1820         s[n] = 0;
1821         *ret = TAKE_PTR(s);
1822
1823         return 0;
1824 }
1825
1826 int cg_mask_from_string(const char *value, CGroupMask *ret) {
1827         CGroupMask m = 0;
1828
1829         assert(ret);
1830         assert(value);
1831
1832         for (;;) {
1833                 _cleanup_free_ char *n = NULL;
1834                 CGroupController v;
1835                 int r;
1836
1837                 r = extract_first_word(&value, &n, NULL, 0);
1838                 if (r < 0)
1839                         return r;
1840                 if (r == 0)
1841                         break;
1842
1843                 v = cgroup_controller_from_string(n);
1844                 if (v < 0)
1845                         continue;
1846
1847                 m |= CGROUP_CONTROLLER_TO_MASK(v);
1848         }
1849
1850         *ret = m;
1851         return 0;
1852 }
1853
1854 int cg_mask_supported(CGroupMask *ret) {
1855         CGroupMask mask;
1856         int r;
1857
1858         /* Determines the mask of supported cgroup controllers. Only includes controllers we can make sense of and that
1859          * are actually accessible. Only covers real controllers, i.e. not the CGROUP_CONTROLLER_BPF_xyz
1860          * pseudo-controllers. */
1861
1862         r = cg_all_unified();
1863         if (r < 0)
1864                 return r;
1865         if (r > 0) {
1866                 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
1867
1868                 /* In the unified hierarchy we can read the supported
1869                  * and accessible controllers from a the top-level
1870                  * cgroup attribute */
1871
1872                 r = cg_get_root_path(&root);
1873                 if (r < 0)
1874                         return r;
1875
1876                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
1877                 if (r < 0)
1878                         return r;
1879
1880                 r = read_one_line_file(path, &controllers);
1881                 if (r < 0)
1882                         return r;
1883
1884                 r = cg_mask_from_string(controllers, &mask);
1885                 if (r < 0)
1886                         return r;
1887
1888                 /* Mask controllers that are not supported in unified hierarchy. */
1889                 mask &= CGROUP_MASK_V2;
1890
1891         } else {
1892                 CGroupController c;
1893
1894                 /* In the legacy hierarchy, we check which hierarchies are mounted. */
1895
1896                 mask = 0;
1897                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1898                         CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
1899                         const char *n;
1900
1901                         if (!FLAGS_SET(CGROUP_MASK_V1, bit))
1902                                 continue;
1903
1904                         n = cgroup_controller_to_string(c);
1905                         if (controller_is_accessible(n) >= 0)
1906                                 mask |= bit;
1907                 }
1908         }
1909
1910         *ret = mask;
1911         return 0;
1912 }
1913
1914 int cg_kernel_controllers(Set **ret) {
1915         _cleanup_set_free_free_ Set *controllers = NULL;
1916         _cleanup_fclose_ FILE *f = NULL;
1917         int r;
1918
1919         assert(ret);
1920
1921         /* Determines the full list of kernel-known controllers. Might include controllers we don't actually support
1922          * and controllers that aren't currently accessible (because not mounted). This does not include "name="
1923          * pseudo-controllers. */
1924
1925         controllers = set_new(&string_hash_ops);
1926         if (!controllers)
1927                 return -ENOMEM;
1928
1929         r = fopen_unlocked("/proc/cgroups", "re", &f);
1930         if (r == -ENOENT) {
1931                 *ret = NULL;
1932                 return 0;
1933         }
1934         if (r < 0)
1935                 return r;
1936
1937         /* Ignore the header line */
1938         (void) read_line(f, (size_t) -1, NULL);
1939
1940         for (;;) {
1941                 char *controller;
1942                 int enabled = 0;
1943
1944                 errno = 0;
1945                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
1946
1947                         if (feof(f))
1948                                 break;
1949
1950                         if (ferror(f))
1951                                 return errno_or_else(EIO);
1952
1953                         return -EBADMSG;
1954                 }
1955
1956                 if (!enabled) {
1957                         free(controller);
1958                         continue;
1959                 }
1960
1961                 if (!cg_controller_is_valid(controller)) {
1962                         free(controller);
1963                         return -EBADMSG;
1964                 }
1965
1966                 r = set_consume(controllers, controller);
1967                 if (r < 0)
1968                         return r;
1969         }
1970
1971         *ret = TAKE_PTR(controllers);
1972
1973         return 0;
1974 }
1975
1976 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup2 on
1977  * /sys/fs/cgroup/systemd. This unfortunately broke other tools (such as docker) which expected the v1
1978  * "name=systemd" hierarchy on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mounts v2 on
1979  * /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility
1980  * with other tools.
1981  *
1982  * To keep live upgrade working, we detect and support v232 layout. When v232 layout is detected, to keep
1983  * cgroup v2 process management but disable the compat dual layout, we return true on
1984  * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and false on cg_hybrid_unified().
1985  */
1986 static thread_local bool unified_systemd_v232;
1987
1988 int cg_unified_cached(bool flush) {
1989         static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
1990
1991         struct statfs fs;
1992
1993         /* Checks if we support the unified hierarchy. Returns an
1994          * error when the cgroup hierarchies aren't mounted yet or we
1995          * have any other trouble determining if the unified hierarchy
1996          * is supported. */
1997
1998         if (flush)
1999                 unified_cache = CGROUP_UNIFIED_UNKNOWN;
2000         else if (unified_cache >= CGROUP_UNIFIED_NONE)
2001                 return unified_cache;
2002
2003         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2004                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2005
2006         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2007                 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2008                 unified_cache = CGROUP_UNIFIED_ALL;
2009         } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2010                 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2011                     F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2012                         log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2013                         unified_cache = CGROUP_UNIFIED_SYSTEMD;
2014                         unified_systemd_v232 = false;
2015                 } else {
2016                         if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0)
2017                                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2018
2019                         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2020                                 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2021                                 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2022                                 unified_systemd_v232 = true;
2023                         } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2024                                 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2025                                 unified_cache = CGROUP_UNIFIED_NONE;
2026                         } else {
2027                                 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2028                                           (unsigned long long) fs.f_type);
2029                                 unified_cache = CGROUP_UNIFIED_NONE;
2030                         }
2031                 }
2032         } else if (F_TYPE_EQUAL(fs.f_type, SYSFS_MAGIC)) {
2033                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2034                                        "No filesystem is currently mounted on /sys/fs/cgroup.");
2035         } else
2036                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2037                                        "Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2038                                        (unsigned long long)fs.f_type);
2039
2040         return unified_cache;
2041 }
2042
2043 int cg_unified_controller(const char *controller) {
2044         int r;
2045
2046         r = cg_unified_cached(false);
2047         if (r < 0)
2048                 return r;
2049
2050         if (r == CGROUP_UNIFIED_NONE)
2051                 return false;
2052
2053         if (r >= CGROUP_UNIFIED_ALL)
2054                 return true;
2055
2056         return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2057 }
2058
2059 int cg_all_unified(void) {
2060         int r;
2061
2062         r = cg_unified_cached(false);
2063         if (r < 0)
2064                 return r;
2065
2066         return r >= CGROUP_UNIFIED_ALL;
2067 }
2068
2069 int cg_hybrid_unified(void) {
2070         int r;
2071
2072         r = cg_unified_cached(false);
2073         if (r < 0)
2074                 return r;
2075
2076         return r == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2077 }
2078
2079 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2080         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2081         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2082         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2083         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2084 };
2085
2086 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2087         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2088         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2089         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2090         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2091 };
2092
2093 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2094
2095 bool is_cgroup_fs(const struct statfs *s) {
2096         return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2097                is_fs_type(s, CGROUP2_SUPER_MAGIC);
2098 }
2099
2100 bool fd_is_cgroup_fs(int fd) {
2101         struct statfs s;
2102
2103         if (fstatfs(fd, &s) < 0)
2104                 return -errno;
2105
2106         return is_cgroup_fs(&s);
2107 }
2108
2109 static const char *const cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2110         [CGROUP_CONTROLLER_CPU] = "cpu",
2111         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2112         [CGROUP_CONTROLLER_CPUSET] = "cpuset",
2113         [CGROUP_CONTROLLER_IO] = "io",
2114         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2115         [CGROUP_CONTROLLER_MEMORY] = "memory",
2116         [CGROUP_CONTROLLER_DEVICES] = "devices",
2117         [CGROUP_CONTROLLER_PIDS] = "pids",
2118         [CGROUP_CONTROLLER_BPF_FIREWALL] = "bpf-firewall",
2119         [CGROUP_CONTROLLER_BPF_DEVICES] = "bpf-devices",
2120 };
2121
2122 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);
2123
2124 CGroupMask get_cpu_accounting_mask(void) {
2125         static CGroupMask needed_mask = (CGroupMask) -1;
2126
2127         /* On kernel ≥4.15 with unified hierarchy, cpu.stat's usage_usec is
2128          * provided externally from the CPU controller, which means we don't
2129          * need to enable the CPU controller just to get metrics. This is good,
2130          * because enabling the CPU controller comes at a minor performance
2131          * hit, especially when it's propagated deep into large hierarchies.
2132          * There's also no separate CPU accounting controller available within
2133          * a unified hierarchy.
2134          *
2135          * This combination of factors results in the desired cgroup mask to
2136          * enable for CPU accounting varying as follows:
2137          *
2138          *                   ╔═════════════════════╤═════════════════════╗
2139          *                   ║     Linux ≥4.15     │     Linux <4.15     ║
2140          *   ╔═══════════════╬═════════════════════╪═════════════════════╣
2141          *   ║ Unified       ║ nothing             │ CGROUP_MASK_CPU     ║
2142          *   ╟───────────────╫─────────────────────┼─────────────────────╢
2143          *   ║ Hybrid/Legacy ║ CGROUP_MASK_CPUACCT │ CGROUP_MASK_CPUACCT ║
2144          *   ╚═══════════════╩═════════════════════╧═════════════════════╝
2145          *
2146          * We check kernel version here instead of manually checking whether
2147          * cpu.stat is present for every cgroup, as that check in itself would
2148          * already be fairly expensive.
2149          *
2150          * Kernels where this patch has been backported will therefore have the
2151          * CPU controller enabled unnecessarily. This is more expensive than
2152          * necessary, but harmless. ☺️
2153          */
2154
2155         if (needed_mask == (CGroupMask) -1) {
2156                 if (cg_all_unified()) {
2157                         struct utsname u;
2158                         assert_se(uname(&u) >= 0);
2159
2160                         if (str_verscmp(u.release, "4.15") < 0)
2161                                 needed_mask = CGROUP_MASK_CPU;
2162                         else
2163                                 needed_mask = 0;
2164                 } else
2165                         needed_mask = CGROUP_MASK_CPUACCT;
2166         }
2167
2168         return needed_mask;
2169 }
2170
2171 bool cpu_accounting_is_cheap(void) {
2172         return get_cpu_accounting_mask() == 0;
2173 }