src/basic/cgroup-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <errno.h>
   4 #include <limits.h>
   5 #include <signal.h>
   6 #include <stddef.h>
   7 #include <stdlib.h>
   8 #include <sys/types.h>
   9 #include <sys/utsname.h>
  10 #include <sys/xattr.h>
  11 #include <unistd.h>
  12
  13 #include "alloc-util.h"
  14 #include "cgroup-util.h"
  15 #include "constants.h"
  16 #include "dirent-util.h"
  17 #include "extract-word.h"
  18 #include "fd-util.h"
  19 #include "fileio.h"
  20 #include "format-util.h"
  21 #include "fs-util.h"
  22 #include "log.h"
  23 #include "login-util.h"
  24 #include "macro.h"
  25 #include "missing_fs.h"
  26 #include "missing_magic.h"
  27 #include "missing_threads.h"
  28 #include "mkdir.h"
  29 #include "parse-util.h"
  30 #include "path-util.h"
  31 #include "process-util.h"
  32 #include "set.h"
  33 #include "special.h"
  34 #include "stat-util.h"
  35 #include "stdio-util.h"
  36 #include "string-table.h"
  37 #include "string-util.h"
  38 #include "strv.h"
  39 #include "unit-name.h"
  40 #include "user-util.h"
  41 #include "xattr-util.h"
  42
  43 int cg_path_open(const char *controller, const char *path) {
  44         _cleanup_free_ char *fs = NULL;
  45         int r;
  46
  47         r = cg_get_path(controller, path, /* item=*/ NULL, &fs);
  48         if (r < 0)
  49                 return r;
  50
  51         return RET_NERRNO(open(fs, O_DIRECTORY|O_CLOEXEC));
  52 }
  53
  54 int cg_cgroupid_open(int cgroupfs_fd, uint64_t id) {
  55         _cleanup_close_ int fsfd = -EBADF;
  56
  57         if (cgroupfs_fd < 0) {
  58                 fsfd = open("/sys/fs/cgroup", O_CLOEXEC|O_DIRECTORY);
  59                 if (fsfd < 0)
  60                         return -errno;
  61
  62                 cgroupfs_fd = fsfd;
  63         }
  64
  65         cg_file_handle fh = CG_FILE_HANDLE_INIT;
  66         CG_FILE_HANDLE_CGROUPID(fh) = id;
  67
  68         int fd = open_by_handle_at(cgroupfs_fd, &fh.file_handle, O_DIRECTORY|O_CLOEXEC);
  69         if (fd < 0)
  70                 return -errno;
  71
  72         return fd;
  73 }
  74
  75 static int cg_enumerate_items(const char *controller, const char *path, FILE **ret, const char *item) {
  76         _cleanup_free_ char *fs = NULL;
  77         FILE *f;
  78         int r;
  79
  80         assert(ret);
  81
  82         r = cg_get_path(controller, path, item, &fs);
  83         if (r < 0)
  84                 return r;
  85
  86         f = fopen(fs, "re");
  87         if (!f)
  88                 return -errno;
  89
  90         *ret = f;
  91         return 0;
  92 }
  93
  94 int cg_enumerate_processes(const char *controller, const char *path, FILE **ret) {
  95         return cg_enumerate_items(controller, path, ret, "cgroup.procs");
  96 }
  97
  98 int cg_read_pid(FILE *f, pid_t *ret, CGroupFlags flags) {
  99         unsigned long ul;
 100
 101         /* Note that the cgroup.procs might contain duplicates! See cgroups.txt for details. */
 102
 103         assert(f);
 104         assert(ret);
 105
 106         for (;;) {
 107                 errno = 0;
 108                 if (fscanf(f, "%lu", &ul) != 1) {
 109
 110                         if (feof(f)) {
 111                                 *ret = 0;
 112                                 return 0;
 113                         }
 114
 115                         return errno_or_else(EIO);
 116                 }
 117
 118                 if (ul > PID_T_MAX)
 119                         return -EIO;
 120
 121                 /* In some circumstances (e.g. WSL), cgroups might contain unmappable PIDs from other
 122                  * contexts. These show up as zeros, and depending on the caller, can either be plain
 123                  * skipped over, or returned as-is. */
 124                 if (ul == 0 && !FLAGS_SET(flags, CGROUP_DONT_SKIP_UNMAPPED))
 125                         continue;
 126
 127                 *ret = (pid_t) ul;
 128                 return 1;
 129         }
 130 }
 131
 132 int cg_read_pidref(FILE *f, PidRef *ret, CGroupFlags flags) {
 133         int r;
 134
 135         assert(f);
 136         assert(ret);
 137
 138         for (;;) {
 139                 pid_t pid;
 140
 141                 r = cg_read_pid(f, &pid, flags);
 142                 if (r < 0)
 143                         return log_debug_errno(r, "Failed to read pid from cgroup item: %m");
 144                 if (r == 0) {
 145                         *ret = PIDREF_NULL;
 146                         return 0;
 147                 }
 148
 149                 if (pid == 0)
 150                         return -EREMOTE;
 151
 152                 if (FLAGS_SET(flags, CGROUP_NO_PIDFD)) {
 153                         *ret = PIDREF_MAKE_FROM_PID(pid);
 154                         return 1;
 155                 }
 156
 157                 r = pidref_set_pid(ret, pid);
 158                 if (r >= 0)
 159                         return 1;
 160                 if (r != -ESRCH)
 161                         return r;
 162
 163                 /* ESRCH → gone by now? just skip over it, read the next */
 164         }
 165 }
 166
 167 int cg_read_event(
 168                 const char *controller,
 169                 const char *path,
 170                 const char *event,
 171                 char **ret) {
 172
 173         _cleanup_free_ char *events = NULL, *content = NULL;
 174         int r;
 175
 176         r = cg_get_path(controller, path, "cgroup.events", &events);
 177         if (r < 0)
 178                 return r;
 179
 180         r = read_full_virtual_file(events, &content, NULL);
 181         if (r < 0)
 182                 return r;
 183
 184         for (const char *p = content;;) {
 185                 _cleanup_free_ char *line = NULL, *key = NULL;
 186                 const char *q;
 187
 188                 r = extract_first_word(&p, &line, "\n", 0);
 189                 if (r < 0)
 190                         return r;
 191                 if (r == 0)
 192                         return -ENOENT;
 193
 194                 q = line;
 195                 r = extract_first_word(&q, &key, " ", 0);
 196                 if (r < 0)
 197                         return r;
 198                 if (r == 0)
 199                         return -EINVAL;
 200
 201                 if (!streq(key, event))
 202                         continue;
 203
 204                 return strdup_to(ret, q);
 205         }
 206 }
 207
 208 bool cg_ns_supported(void) {
 209         static thread_local int enabled = -1;
 210
 211         if (enabled >= 0)
 212                 return enabled;
 213
 214         if (access("/proc/self/ns/cgroup", F_OK) < 0) {
 215                 if (errno != ENOENT)
 216                         log_debug_errno(errno, "Failed to check whether /proc/self/ns/cgroup is available, assuming not: %m");
 217                 enabled = false;
 218         } else
 219                 enabled = true;
 220
 221         return enabled;
 222 }
 223
 224 bool cg_freezer_supported(void) {
 225         static thread_local int supported = -1;
 226
 227         if (supported >= 0)
 228                 return supported;
 229
 230         supported = cg_all_unified() > 0 && access("/sys/fs/cgroup/init.scope/cgroup.freeze", F_OK) == 0;
 231
 232         return supported;
 233 }
 234
 235 bool cg_kill_supported(void) {
 236         static thread_local int supported = -1;
 237
 238         if (supported >= 0)
 239                 return supported;
 240
 241         if (cg_all_unified() <= 0)
 242                 supported = false;
 243         else if (access("/sys/fs/cgroup/init.scope/cgroup.kill", F_OK) < 0) {
 244                 if (errno != ENOENT)
 245                         log_debug_errno(errno, "Failed to check if cgroup.kill is available, assuming not: %m");
 246                 supported = false;
 247         } else
 248                 supported = true;
 249
 250         return supported;
 251 }
 252
 253 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **ret) {
 254         _cleanup_free_ char *fs = NULL;
 255         DIR *d;
 256         int r;
 257
 258         assert(ret);
 259
 260         /* This is not recursive! */
 261
 262         r = cg_get_path(controller, path, NULL, &fs);
 263         if (r < 0)
 264                 return r;
 265
 266         d = opendir(fs);
 267         if (!d)
 268                 return -errno;
 269
 270         *ret = d;
 271         return 0;
 272 }
 273
 274 int cg_read_subgroup(DIR *d, char **ret) {
 275         assert(d);
 276         assert(ret);
 277
 278         FOREACH_DIRENT_ALL(de, d, return -errno) {
 279                 if (de->d_type != DT_DIR)
 280                         continue;
 281
 282                 if (dot_or_dot_dot(de->d_name))
 283                         continue;
 284
 285                 return strdup_to_full(ret, de->d_name);
 286         }
 287
 288         *ret = NULL;
 289         return 0;
 290 }
 291
 292 int cg_rmdir(const char *controller, const char *path) {
 293         _cleanup_free_ char *p = NULL;
 294         int r;
 295
 296         r = cg_get_path(controller, path, NULL, &p);
 297         if (r < 0)
 298                 return r;
 299
 300         r = rmdir(p);
 301         if (r < 0 && errno != ENOENT)
 302                 return -errno;
 303
 304         r = cg_hybrid_unified();
 305         if (r <= 0)
 306                 return r;
 307
 308         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 309                 r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
 310                 if (r < 0)
 311                         log_warning_errno(r, "Failed to remove compat systemd cgroup %s: %m", path);
 312         }
 313
 314         return 0;
 315 }
 316
 317 static int cg_kill_items(
 318                 const char *path,
 319                 int sig,
 320                 CGroupFlags flags,
 321                 Set *s,
 322                 cg_kill_log_func_t log_kill,
 323                 void *userdata,
 324                 const char *item) {
 325
 326         _cleanup_set_free_ Set *allocated_set = NULL;
 327         bool done = false;
 328         int r, ret = 0, ret_log_kill = 0;
 329
 330         assert(sig >= 0);
 331
 332          /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
 333           * SIGCONT on SIGKILL. */
 334         if (IN_SET(sig, SIGCONT, SIGKILL))
 335                 flags &= ~CGROUP_SIGCONT;
 336
 337         /* This goes through the tasks list and kills them all. This
 338          * is repeated until no further processes are added to the
 339          * tasks list, to properly handle forking processes */
 340
 341         if (!s) {
 342                 s = allocated_set = set_new(NULL);
 343                 if (!s)
 344                         return -ENOMEM;
 345         }
 346
 347         do {
 348                 _cleanup_fclose_ FILE *f = NULL;
 349                 done = true;
 350
 351                 r = cg_enumerate_items(SYSTEMD_CGROUP_CONTROLLER, path, &f, item);
 352                 if (r == -ENOENT)
 353                         break;
 354                 if (r < 0)
 355                         return RET_GATHER(ret, log_debug_errno(r, "Failed to enumerate cgroup items: %m"));
 356
 357                 for (;;) {
 358                         _cleanup_(pidref_done) PidRef pidref = PIDREF_NULL;
 359
 360                         r = cg_read_pidref(f, &pidref, flags);
 361                         if (r < 0)
 362                                 return RET_GATHER(ret, log_debug_errno(r, "Failed to read pidref from cgroup '%s': %m", path));
 363                         if (r == 0)
 364                                 break;
 365
 366                         if ((flags & CGROUP_IGNORE_SELF) && pidref_is_self(&pidref))
 367                                 continue;
 368
 369                         if (set_get(s, PID_TO_PTR(pidref.pid)) == PID_TO_PTR(pidref.pid))
 370                                 continue;
 371
 372                         if (log_kill)
 373                                 ret_log_kill = log_kill(&pidref, sig, userdata);
 374
 375                         /* If we haven't killed this process yet, kill it */
 376                         r = pidref_kill(&pidref, sig);
 377                         if (r < 0 && r != -ESRCH)
 378                                 RET_GATHER(ret, log_debug_errno(r, "Failed to kill process with pid " PID_FMT " from cgroup '%s': %m", pidref.pid, path));
 379                         if (r >= 0) {
 380                                 if (flags & CGROUP_SIGCONT)
 381                                         (void) pidref_kill(&pidref, SIGCONT);
 382
 383                                 if (ret == 0) {
 384                                         if (log_kill)
 385                                                 ret = ret_log_kill;
 386                                         else
 387                                                 ret = 1;
 388                                 }
 389                         }
 390
 391                         done = false;
 392
 393                         r = set_put(s, PID_TO_PTR(pidref.pid));
 394                         if (r < 0)
 395                                 return RET_GATHER(ret, r);
 396                 }
 397
 398                 /* To avoid racing against processes which fork quicker than we can kill them, we repeat this
 399                  * until no new pids need to be killed. */
 400
 401         } while (!done);
 402
 403         return ret;
 404 }
 405
 406 int cg_kill(
 407                 const char *path,
 408                 int sig,
 409                 CGroupFlags flags,
 410                 Set *s,
 411                 cg_kill_log_func_t log_kill,
 412                 void *userdata) {
 413
 414         int r, ret;
 415
 416         r = cg_kill_items(path, sig, flags, s, log_kill, userdata, "cgroup.procs");
 417         if (r < 0)
 418                 log_debug_errno(r, "Failed to kill processes in cgroup '%s' item cgroup.procs: %m", path);
 419         if (r < 0 || sig != SIGKILL)
 420                 return r;
 421
 422         ret = r;
 423
 424         /* Only in case of killing with SIGKILL and when using cgroupsv2, kill remaining threads manually as
 425            a workaround for kernel bug. It was fixed in 5.2-rc5 (c03cd7738a83), backported to 4.19.66
 426            (4340d175b898) and 4.14.138 (feb6b123b7dd). */
 427         r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
 428         if (r < 0)
 429                 return r;
 430         if (r == 0)
 431                 return ret;
 432
 433         /* Opening pidfds for non thread group leaders only works from 6.9 onwards with PIDFD_THREAD. On
 434          * older kernels or without PIDFD_THREAD pidfd_open() fails with EINVAL. Since we might read non
 435          * thread group leader IDs from cgroup.threads, we set CGROUP_NO_PIDFD to avoid trying open pidfd's
 436          * for them and instead use the regular pid. */
 437         r = cg_kill_items(path, sig, flags|CGROUP_NO_PIDFD, s, log_kill, userdata, "cgroup.threads");
 438         if (r < 0)
 439                 return log_debug_errno(r, "Failed to kill processes in cgroup '%s' item cgroup.threads: %m", path);
 440
 441         return r > 0 || ret > 0;
 442 }
 443
 444 int cg_kill_kernel_sigkill(const char *path) {
 445         /* Kills the cgroup at `path` directly by writing to its cgroup.kill file.  This sends SIGKILL to all
 446          * processes in the cgroup and has the advantage of being completely atomic, unlike cg_kill_items(). */
 447
 448         _cleanup_free_ char *killfile = NULL;
 449         int r;
 450
 451         assert(path);
 452
 453         if (!cg_kill_supported())
 454                 return -EOPNOTSUPP;
 455
 456         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, "cgroup.kill", &killfile);
 457         if (r < 0)
 458                 return r;
 459
 460         r = write_string_file(killfile, "1", WRITE_STRING_FILE_DISABLE_BUFFER);
 461         if (r < 0)
 462                 return log_debug_errno(r, "Failed to write to cgroup.kill for cgroup '%s': %m", path);
 463
 464         return 0;
 465 }
 466
 467 int cg_kill_recursive(
 468                 const char *path,
 469                 int sig,
 470                 CGroupFlags flags,
 471                 Set *s,
 472                 cg_kill_log_func_t log_kill,
 473                 void *userdata) {
 474
 475         int r, ret;
 476
 477         assert(path);
 478         assert(sig >= 0);
 479
 480         if (sig == SIGKILL && cg_kill_supported() &&
 481             !FLAGS_SET(flags, CGROUP_IGNORE_SELF) && !s && !log_kill)
 482                 /* ignore CGROUP_SIGCONT, since this is a no-op alongside SIGKILL */
 483                 ret = cg_kill_kernel_sigkill(path);
 484         else {
 485                 _cleanup_set_free_ Set *allocated_set = NULL;
 486                 _cleanup_closedir_ DIR *d = NULL;
 487
 488                 if (!s) {
 489                         s = allocated_set = set_new(NULL);
 490                         if (!s)
 491                                 return -ENOMEM;
 492                 }
 493
 494                 ret = cg_kill(path, sig, flags, s, log_kill, userdata);
 495
 496                 r = cg_enumerate_subgroups(SYSTEMD_CGROUP_CONTROLLER, path, &d);
 497                 if (r < 0) {
 498                         if (r != -ENOENT)
 499                                 RET_GATHER(ret, log_debug_errno(r, "Failed to enumerate cgroup '%s' subgroups: %m", path));
 500
 501                         return ret;
 502                 }
 503
 504                 for (;;) {
 505                         _cleanup_free_ char *fn = NULL, *p = NULL;
 506
 507                         r = cg_read_subgroup(d, &fn);
 508                         if (r < 0) {
 509                                 RET_GATHER(ret, log_debug_errno(r, "Failed to read subgroup from cgroup '%s': %m", path));
 510                                 break;
 511                         }
 512                         if (r == 0)
 513                                 break;
 514
 515                         p = path_join(empty_to_root(path), fn);
 516                         if (!p)
 517                                 return -ENOMEM;
 518
 519                         r = cg_kill_recursive(p, sig, flags, s, log_kill, userdata);
 520                         if (r < 0)
 521                                 log_debug_errno(r, "Failed to recursively kill processes in cgroup '%s': %m", p);
 522                         if (r != 0 && ret >= 0)
 523                                 ret = r;
 524                 }
 525         }
 526
 527         if (FLAGS_SET(flags, CGROUP_REMOVE)) {
 528                 r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER, path);
 529                 if (!IN_SET(r, -ENOENT, -EBUSY))
 530                         RET_GATHER(ret, log_debug_errno(r, "Failed to remove cgroup '%s': %m", path));
 531         }
 532
 533         return ret;
 534 }
 535
 536 static const char *controller_to_dirname(const char *controller) {
 537         assert(controller);
 538
 539         /* Converts a controller name to the directory name below /sys/fs/cgroup/ we want to mount it
 540          * to. Effectively, this just cuts off the name= prefixed used for named hierarchies, if it is
 541          * specified. */
 542
 543         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 544                 if (cg_hybrid_unified() > 0)
 545                         controller = SYSTEMD_CGROUP_CONTROLLER_HYBRID;
 546                 else
 547                         controller = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 548         }
 549
 550         return startswith(controller, "name=") ?: controller;
 551 }
 552
 553 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **ret) {
 554         const char *dn;
 555         char *t = NULL;
 556
 557         assert(ret);
 558         assert(controller);
 559
 560         dn = controller_to_dirname(controller);
 561
 562         if (isempty(path) && isempty(suffix))
 563                 t = path_join("/sys/fs/cgroup", dn);
 564         else if (isempty(path))
 565                 t = path_join("/sys/fs/cgroup", dn, suffix);
 566         else if (isempty(suffix))
 567                 t = path_join("/sys/fs/cgroup", dn, path);
 568         else
 569                 t = path_join("/sys/fs/cgroup", dn, path, suffix);
 570         if (!t)
 571                 return -ENOMEM;
 572
 573         *ret = t;
 574         return 0;
 575 }
 576
 577 static int join_path_unified(const char *path, const char *suffix, char **ret) {
 578         char *t;
 579
 580         assert(ret);
 581
 582         if (isempty(path) && isempty(suffix))
 583                 t = strdup("/sys/fs/cgroup");
 584         else if (isempty(path))
 585                 t = path_join("/sys/fs/cgroup", suffix);
 586         else if (isempty(suffix))
 587                 t = path_join("/sys/fs/cgroup", path);
 588         else
 589                 t = path_join("/sys/fs/cgroup", path, suffix);
 590         if (!t)
 591                 return -ENOMEM;
 592
 593         *ret = t;
 594         return 0;
 595 }
 596
 597 int cg_get_path(const char *controller, const char *path, const char *suffix, char **ret) {
 598         int r;
 599
 600         assert(ret);
 601
 602         if (!controller) {
 603                 char *t;
 604
 605                 /* If no controller is specified, we return the path *below* the controllers, without any
 606                  * prefix. */
 607
 608                 if (isempty(path) && isempty(suffix))
 609                         return -EINVAL;
 610
 611                 if (isempty(suffix))
 612                         t = strdup(path);
 613                 else if (isempty(path))
 614                         t = strdup(suffix);
 615                 else
 616                         t = path_join(path, suffix);
 617                 if (!t)
 618                         return -ENOMEM;
 619
 620                 *ret = path_simplify(t);
 621                 return 0;
 622         }
 623
 624         if (!cg_controller_is_valid(controller))
 625                 return -EINVAL;
 626
 627         r = cg_all_unified();
 628         if (r < 0)
 629                 return r;
 630         if (r > 0)
 631                 r = join_path_unified(path, suffix, ret);
 632         else
 633                 r = join_path_legacy(controller, path, suffix, ret);
 634         if (r < 0)
 635                 return r;
 636
 637         path_simplify(*ret);
 638         return 0;
 639 }
 640
 641 static int controller_is_v1_accessible(const char *root, const char *controller) {
 642         const char *cpath, *dn;
 643
 644         assert(controller);
 645
 646         dn = controller_to_dirname(controller);
 647
 648         /* If root if specified, we check that:
 649          * - possible subcgroup is created at root,
 650          * - we can modify the hierarchy. */
 651
 652         cpath = strjoina("/sys/fs/cgroup/", dn, root, root ? "/cgroup.procs" : NULL);
 653         return laccess(cpath, root ? W_OK : F_OK);
 654 }
 655
 656 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **ret) {
 657         int r;
 658
 659         assert(controller);
 660         assert(ret);
 661
 662         if (!cg_controller_is_valid(controller))
 663                 return -EINVAL;
 664
 665         r = cg_all_unified();
 666         if (r < 0)
 667                 return r;
 668         if (r > 0) {
 669                 /* In the unified hierarchy all controllers are considered accessible,
 670                  * except for the named hierarchies */
 671                 if (startswith(controller, "name="))
 672                         return -EOPNOTSUPP;
 673         } else {
 674                 /* Check if the specified controller is actually accessible */
 675                 r = controller_is_v1_accessible(NULL, controller);
 676                 if (r < 0)
 677                         return r;
 678         }
 679
 680         return cg_get_path(controller, path, suffix, ret);
 681 }
 682
 683 int cg_set_xattr(const char *path, const char *name, const void *value, size_t size, int flags) {
 684         _cleanup_free_ char *fs = NULL;
 685         int r;
 686
 687         assert(path);
 688         assert(name);
 689         assert(value || size <= 0);
 690
 691         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &fs);
 692         if (r < 0)
 693                 return r;
 694
 695         return RET_NERRNO(setxattr(fs, name, value, size, flags));
 696 }
 697
 698 int cg_get_xattr(const char *path, const char *name, void *value, size_t size) {
 699         _cleanup_free_ char *fs = NULL;
 700         ssize_t n;
 701         int r;
 702
 703         assert(path);
 704         assert(name);
 705
 706         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &fs);
 707         if (r < 0)
 708                 return r;
 709
 710         n = getxattr(fs, name, value, size);
 711         if (n < 0)
 712                 return -errno;
 713
 714         return (int) n;
 715 }
 716
 717 int cg_get_xattr_malloc(const char *path, const char *name, char **ret) {
 718         _cleanup_free_ char *fs = NULL;
 719         int r;
 720
 721         assert(path);
 722         assert(name);
 723
 724         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &fs);
 725         if (r < 0)
 726                 return r;
 727
 728         return lgetxattr_malloc(fs, name, ret);
 729 }
 730
 731 int cg_get_xattr_bool(const char *path, const char *name) {
 732         _cleanup_free_ char *fs = NULL;
 733         int r;
 734
 735         assert(path);
 736         assert(name);
 737
 738         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &fs);
 739         if (r < 0)
 740                 return r;
 741
 742         return getxattr_at_bool(AT_FDCWD, fs, name, /* flags= */ 0);
 743 }
 744
 745 int cg_remove_xattr(const char *path, const char *name) {
 746         _cleanup_free_ char *fs = NULL;
 747         int r;
 748
 749         assert(path);
 750         assert(name);
 751
 752         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &fs);
 753         if (r < 0)
 754                 return r;
 755
 756         return RET_NERRNO(removexattr(fs, name));
 757 }
 758
 759 int cg_pid_get_path(const char *controller, pid_t pid, char **ret_path) {
 760         _cleanup_fclose_ FILE *f = NULL;
 761         const char *fs, *controller_str = NULL;  /* avoid false maybe-uninitialized warning */
 762         int unified, r;
 763
 764         assert(pid >= 0);
 765         assert(ret_path);
 766
 767         if (controller) {
 768                 if (!cg_controller_is_valid(controller))
 769                         return -EINVAL;
 770         } else
 771                 controller = SYSTEMD_CGROUP_CONTROLLER;
 772
 773         unified = cg_unified_controller(controller);
 774         if (unified < 0)
 775                 return unified;
 776         if (unified == 0) {
 777                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 778                         controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 779                 else
 780                         controller_str = controller;
 781         }
 782
 783         fs = procfs_file_alloca(pid, "cgroup");
 784         r = fopen_unlocked(fs, "re", &f);
 785         if (r == -ENOENT)
 786                 return -ESRCH;
 787         if (r < 0)
 788                 return r;
 789
 790         for (;;) {
 791                 _cleanup_free_ char *line = NULL;
 792                 char *e;
 793
 794                 r = read_line(f, LONG_LINE_MAX, &line);
 795                 if (r < 0)
 796                         return r;
 797                 if (r == 0)
 798                         return -ENODATA;
 799
 800                 if (unified) {
 801                         e = startswith(line, "0:");
 802                         if (!e)
 803                                 continue;
 804
 805                         e = strchr(e, ':');
 806                         if (!e)
 807                                 continue;
 808                 } else {
 809                         char *l;
 810
 811                         l = strchr(line, ':');
 812                         if (!l)
 813                                 continue;
 814
 815                         l++;
 816                         e = strchr(l, ':');
 817                         if (!e)
 818                                 continue;
 819                         *e = 0;
 820
 821                         assert(controller_str);
 822                         r = string_contains_word(l, ",", controller_str);
 823                         if (r < 0)
 824                                 return r;
 825                         if (r == 0)
 826                                 continue;
 827                 }
 828
 829                 char *path = strdup(e + 1);
 830                 if (!path)
 831                         return -ENOMEM;
 832
 833                 /* Truncate suffix indicating the process is a zombie */
 834                 e = endswith(path, " (deleted)");
 835                 if (e)
 836                         *e = 0;
 837
 838                 *ret_path = path;
 839                 return 0;
 840         }
 841 }
 842
 843 int cg_pidref_get_path(const char *controller, const PidRef *pidref, char **ret_path) {
 844         _cleanup_free_ char *path = NULL;
 845         int r;
 846
 847         assert(ret_path);
 848
 849         if (!pidref_is_set(pidref))
 850                 return -ESRCH;
 851
 852         r = cg_pid_get_path(controller, pidref->pid, &path);
 853         if (r < 0)
 854                 return r;
 855
 856         /* Before we return the path, make sure the procfs entry for this pid still matches the pidref */
 857         r = pidref_verify(pidref);
 858         if (r < 0)
 859                 return r;
 860
 861         *ret_path = TAKE_PTR(path);
 862         return 0;
 863 }
 864
 865 int cg_install_release_agent(const char *controller, const char *agent) {
 866         _cleanup_free_ char *fs = NULL, *contents = NULL;
 867         const char *sc;
 868         int r;
 869
 870         assert(agent);
 871
 872         r = cg_unified_controller(controller);
 873         if (r < 0)
 874                 return r;
 875         if (r > 0) /* doesn't apply to unified hierarchy */
 876                 return -EOPNOTSUPP;
 877
 878         r = cg_get_path(controller, NULL, "release_agent", &fs);
 879         if (r < 0)
 880                 return r;
 881
 882         r = read_one_line_file(fs, &contents);
 883         if (r < 0)
 884                 return r;
 885
 886         sc = strstrip(contents);
 887         if (isempty(sc)) {
 888                 r = write_string_file(fs, agent, WRITE_STRING_FILE_DISABLE_BUFFER);
 889                 if (r < 0)
 890                         return r;
 891         } else if (!path_equal(sc, agent))
 892                 return -EEXIST;
 893
 894         fs = mfree(fs);
 895         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 896         if (r < 0)
 897                 return r;
 898
 899         contents = mfree(contents);
 900         r = read_one_line_file(fs, &contents);
 901         if (r < 0)
 902                 return r;
 903
 904         sc = strstrip(contents);
 905         if (streq(sc, "0")) {
 906                 r = write_string_file(fs, "1", WRITE_STRING_FILE_DISABLE_BUFFER);
 907                 if (r < 0)
 908                         return r;
 909
 910                 return 1;
 911         }
 912
 913         if (!streq(sc, "1"))
 914                 return -EIO;
 915
 916         return 0;
 917 }
 918
 919 int cg_uninstall_release_agent(const char *controller) {
 920         _cleanup_free_ char *fs = NULL;
 921         int r;
 922
 923         r = cg_unified_controller(controller);
 924         if (r < 0)
 925                 return r;
 926         if (r > 0) /* Doesn't apply to unified hierarchy */
 927                 return -EOPNOTSUPP;
 928
 929         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
 930         if (r < 0)
 931                 return r;
 932
 933         r = write_string_file(fs, "0", WRITE_STRING_FILE_DISABLE_BUFFER);
 934         if (r < 0)
 935                 return r;
 936
 937         fs = mfree(fs);
 938
 939         r = cg_get_path(controller, NULL, "release_agent", &fs);
 940         if (r < 0)
 941                 return r;
 942
 943         r = write_string_file(fs, "", WRITE_STRING_FILE_DISABLE_BUFFER);
 944         if (r < 0)
 945                 return r;
 946
 947         return 0;
 948 }
 949
 950 int cg_is_empty(const char *controller, const char *path) {
 951         _cleanup_fclose_ FILE *f = NULL;
 952         pid_t pid;
 953         int r;
 954
 955         assert(path);
 956
 957         r = cg_enumerate_processes(controller, path, &f);
 958         if (r == -ENOENT)
 959                 return true;
 960         if (r < 0)
 961                 return r;
 962
 963         r = cg_read_pid(f, &pid, CGROUP_DONT_SKIP_UNMAPPED);
 964         if (r < 0)
 965                 return r;
 966
 967         return r == 0;
 968 }
 969
 970 int cg_is_empty_recursive(const char *controller, const char *path) {
 971         int r;
 972
 973         assert(path);
 974
 975         /* The root cgroup is always populated */
 976         if (controller && empty_or_root(path))
 977                 return false;
 978
 979         r = cg_unified_controller(controller);
 980         if (r < 0)
 981                 return r;
 982         if (r > 0) {
 983                 _cleanup_free_ char *t = NULL;
 984
 985                 /* On the unified hierarchy we can check empty state
 986                  * via the "populated" attribute of "cgroup.events". */
 987
 988                 r = cg_read_event(controller, path, "populated", &t);
 989                 if (r == -ENOENT)
 990                         return true;
 991                 if (r < 0)
 992                         return r;
 993
 994                 return streq(t, "0");
 995         } else {
 996                 _cleanup_closedir_ DIR *d = NULL;
 997                 char *fn;
 998
 999                 r = cg_is_empty(controller, path);
1000                 if (r <= 0)
1001                         return r;
1002
1003                 r = cg_enumerate_subgroups(controller, path, &d);
1004                 if (r == -ENOENT)
1005                         return true;
1006                 if (r < 0)
1007                         return r;
1008
1009                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1010                         _cleanup_free_ char *p = NULL;
1011
1012                         p = path_join(path, fn);
1013                         free(fn);
1014                         if (!p)
1015                                 return -ENOMEM;
1016
1017                         r = cg_is_empty_recursive(controller, p);
1018                         if (r <= 0)
1019                                 return r;
1020                 }
1021                 if (r < 0)
1022                         return r;
1023
1024                 return true;
1025         }
1026 }
1027
1028 int cg_split_spec(const char *spec, char **ret_controller, char **ret_path) {
1029         _cleanup_free_ char *controller = NULL, *path = NULL;
1030         int r;
1031
1032         assert(spec);
1033
1034         if (*spec == '/') {
1035                 if (!path_is_normalized(spec))
1036                         return -EINVAL;
1037
1038                 if (ret_path) {
1039                         r = path_simplify_alloc(spec, &path);
1040                         if (r < 0)
1041                                 return r;
1042                 }
1043
1044         } else {
1045                 const char *e;
1046
1047                 e = strchr(spec, ':');
1048                 if (e) {
1049                         controller = strndup(spec, e-spec);
1050                         if (!controller)
1051                                 return -ENOMEM;
1052                         if (!cg_controller_is_valid(controller))
1053                                 return -EINVAL;
1054
1055                         if (!isempty(e + 1)) {
1056                                 path = strdup(e+1);
1057                                 if (!path)
1058                                         return -ENOMEM;
1059
1060                                 if (!path_is_normalized(path) ||
1061                                     !path_is_absolute(path))
1062                                         return -EINVAL;
1063
1064                                 path_simplify(path);
1065                         }
1066
1067                 } else {
1068                         if (!cg_controller_is_valid(spec))
1069                                 return -EINVAL;
1070
1071                         if (ret_controller) {
1072                                 controller = strdup(spec);
1073                                 if (!controller)
1074                                         return -ENOMEM;
1075                         }
1076                 }
1077         }
1078
1079         if (ret_controller)
1080                 *ret_controller = TAKE_PTR(controller);
1081         if (ret_path)
1082                 *ret_path = TAKE_PTR(path);
1083         return 0;
1084 }
1085
1086 int cg_mangle_path(const char *path, char **ret) {
1087         _cleanup_free_ char *c = NULL, *p = NULL;
1088         int r;
1089
1090         assert(path);
1091         assert(ret);
1092
1093         /* First, check if it already is a filesystem path */
1094         if (path_startswith(path, "/sys/fs/cgroup"))
1095                 return path_simplify_alloc(path, ret);
1096
1097         /* Otherwise, treat it as cg spec */
1098         r = cg_split_spec(path, &c, &p);
1099         if (r < 0)
1100                 return r;
1101
1102         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, ret);
1103 }
1104
1105 int cg_get_root_path(char **ret_path) {
1106         char *p, *e;
1107         int r;
1108
1109         assert(ret_path);
1110
1111         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1112         if (r < 0)
1113                 return r;
1114
1115         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1116         if (!e)
1117                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1118         if (!e)
1119                 e = endswith(p, "/system"); /* even more legacy */
1120         if (e)
1121                 *e = 0;
1122
1123         *ret_path = p;
1124         return 0;
1125 }
1126
1127 int cg_shift_path(const char *cgroup, const char *root, const char **ret_shifted) {
1128         _cleanup_free_ char *rt = NULL;
1129         char *p;
1130         int r;
1131
1132         assert(cgroup);
1133         assert(ret_shifted);
1134
1135         if (!root) {
1136                 /* If the root was specified let's use that, otherwise
1137                  * let's determine it from PID 1 */
1138
1139                 r = cg_get_root_path(&rt);
1140                 if (r < 0)
1141                         return r;
1142
1143                 root = rt;
1144         }
1145
1146         p = path_startswith(cgroup, root);
1147         if (p && p > cgroup)
1148                 *ret_shifted = p - 1;
1149         else
1150                 *ret_shifted = cgroup;
1151
1152         return 0;
1153 }
1154
1155 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **ret_cgroup) {
1156         _cleanup_free_ char *raw = NULL;
1157         const char *c;
1158         int r;
1159
1160         assert(pid >= 0);
1161         assert(ret_cgroup);
1162
1163         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1164         if (r < 0)
1165                 return r;
1166
1167         r = cg_shift_path(raw, root, &c);
1168         if (r < 0)
1169                 return r;
1170
1171         if (c == raw) {
1172                 *ret_cgroup = TAKE_PTR(raw);
1173                 return 0;
1174         }
1175
1176         return strdup_to(ret_cgroup, c);
1177 }
1178
1179 int cg_path_decode_unit(const char *cgroup, char **ret_unit) {
1180         assert(cgroup);
1181         assert(ret_unit);
1182
1183         size_t n = strcspn(cgroup, "/");
1184         if (n < 3)
1185                 return -ENXIO;
1186
1187         char *c = strndupa_safe(cgroup, n);
1188         c = cg_unescape(c);
1189
1190         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1191                 return -ENXIO;
1192
1193         return strdup_to(ret_unit, c);
1194 }
1195
1196 static bool valid_slice_name(const char *p, size_t n) {
1197
1198         if (!p)
1199                 return false;
1200
1201         if (n < STRLEN("x.slice"))
1202                 return false;
1203
1204         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1205                 char buf[n+1], *c;
1206
1207                 memcpy(buf, p, n);
1208                 buf[n] = 0;
1209
1210                 c = cg_unescape(buf);
1211
1212                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1213         }
1214
1215         return false;
1216 }
1217
1218 static const char *skip_slices(const char *p) {
1219         assert(p);
1220
1221         /* Skips over all slice assignments */
1222
1223         for (;;) {
1224                 size_t n;
1225
1226                 p += strspn(p, "/");
1227
1228                 n = strcspn(p, "/");
1229                 if (!valid_slice_name(p, n))
1230                         return p;
1231
1232                 p += n;
1233         }
1234 }
1235
1236 int cg_path_get_unit(const char *path, char **ret) {
1237         _cleanup_free_ char *unit = NULL;
1238         const char *e;
1239         int r;
1240
1241         assert(path);
1242         assert(ret);
1243
1244         e = skip_slices(path);
1245
1246         r = cg_path_decode_unit(e, &unit);
1247         if (r < 0)
1248                 return r;
1249
1250         /* We skipped over the slices, don't accept any now */
1251         if (endswith(unit, ".slice"))
1252                 return -ENXIO;
1253
1254         *ret = TAKE_PTR(unit);
1255         return 0;
1256 }
1257
1258 int cg_path_get_unit_path(const char *path, char **ret) {
1259         _cleanup_free_ char *path_copy = NULL;
1260         char *unit_name;
1261
1262         assert(path);
1263         assert(ret);
1264
1265         path_copy = strdup(path);
1266         if (!path_copy)
1267                 return -ENOMEM;
1268
1269         unit_name = (char *)skip_slices(path_copy);
1270         unit_name[strcspn(unit_name, "/")] = 0;
1271
1272         if (!unit_name_is_valid(cg_unescape(unit_name), UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1273                 return -ENXIO;
1274
1275         *ret = TAKE_PTR(path_copy);
1276
1277         return 0;
1278 }
1279
1280 int cg_pid_get_unit(pid_t pid, char **ret_unit) {
1281         _cleanup_free_ char *cgroup = NULL;
1282         int r;
1283
1284         assert(ret_unit);
1285
1286         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1287         if (r < 0)
1288                 return r;
1289
1290         return cg_path_get_unit(cgroup, ret_unit);
1291 }
1292
1293 int cg_pidref_get_unit(const PidRef *pidref, char **ret) {
1294         _cleanup_free_ char *unit = NULL;
1295         int r;
1296
1297         assert(ret);
1298
1299         if (!pidref_is_set(pidref))
1300                 return -ESRCH;
1301
1302         r = cg_pid_get_unit(pidref->pid, &unit);
1303         if (r < 0)
1304                 return r;
1305
1306         r = pidref_verify(pidref);
1307         if (r < 0)
1308                 return r;
1309
1310         *ret = TAKE_PTR(unit);
1311         return 0;
1312 }
1313
1314 /**
1315  * Skip session-*.scope, but require it to be there.
1316  */
1317 static const char *skip_session(const char *p) {
1318         size_t n;
1319
1320         if (isempty(p))
1321                 return NULL;
1322
1323         p += strspn(p, "/");
1324
1325         n = strcspn(p, "/");
1326         if (n < STRLEN("session-x.scope"))
1327                 return NULL;
1328
1329         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1330                 char buf[n - 8 - 6 + 1];
1331
1332                 memcpy(buf, p + 8, n - 8 - 6);
1333                 buf[n - 8 - 6] = 0;
1334
1335                 /* Note that session scopes never need unescaping,
1336                  * since they cannot conflict with the kernel's own
1337                  * names, hence we don't need to call cg_unescape()
1338                  * here. */
1339
1340                 if (!session_id_valid(buf))
1341                         return NULL;
1342
1343                 p += n;
1344                 p += strspn(p, "/");
1345                 return p;
1346         }
1347
1348         return NULL;
1349 }
1350
1351 /**
1352  * Skip user@*.service, but require it to be there.
1353  */
1354 static const char *skip_user_manager(const char *p) {
1355         size_t n;
1356
1357         if (isempty(p))
1358                 return NULL;
1359
1360         p += strspn(p, "/");
1361
1362         n = strcspn(p, "/");
1363         if (n < STRLEN("user@x.service"))
1364                 return NULL;
1365
1366         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1367                 char buf[n - 5 - 8 + 1];
1368
1369                 memcpy(buf, p + 5, n - 5 - 8);
1370                 buf[n - 5 - 8] = 0;
1371
1372                 /* Note that user manager services never need unescaping,
1373                  * since they cannot conflict with the kernel's own
1374                  * names, hence we don't need to call cg_unescape()
1375                  * here. */
1376
1377                 if (parse_uid(buf, NULL) < 0)
1378                         return NULL;
1379
1380                 p += n;
1381                 p += strspn(p, "/");
1382
1383                 return p;
1384         }
1385
1386         return NULL;
1387 }
1388
1389 static const char *skip_user_prefix(const char *path) {
1390         const char *e, *t;
1391
1392         assert(path);
1393
1394         /* Skip slices, if there are any */
1395         e = skip_slices(path);
1396
1397         /* Skip the user manager, if it's in the path now... */
1398         t = skip_user_manager(e);
1399         if (t)
1400                 return t;
1401
1402         /* Alternatively skip the user session if it is in the path... */
1403         return skip_session(e);
1404 }
1405
1406 int cg_path_get_user_unit(const char *path, char **ret) {
1407         const char *t;
1408
1409         assert(path);
1410         assert(ret);
1411
1412         t = skip_user_prefix(path);
1413         if (!t)
1414                 return -ENXIO;
1415
1416         /* And from here on it looks pretty much the same as for a system unit, hence let's use the same
1417          * parser. */
1418         return cg_path_get_unit(t, ret);
1419 }
1420
1421 int cg_pid_get_user_unit(pid_t pid, char **ret_unit) {
1422         _cleanup_free_ char *cgroup = NULL;
1423         int r;
1424
1425         assert(ret_unit);
1426
1427         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1428         if (r < 0)
1429                 return r;
1430
1431         return cg_path_get_user_unit(cgroup, ret_unit);
1432 }
1433
1434 int cg_path_get_machine_name(const char *path, char **ret_machine) {
1435         _cleanup_free_ char *u = NULL;
1436         const char *sl;
1437         int r;
1438
1439         r = cg_path_get_unit(path, &u);
1440         if (r < 0)
1441                 return r;
1442
1443         sl = strjoina("/run/systemd/machines/unit:", u);
1444         return readlink_malloc(sl, ret_machine);
1445 }
1446
1447 int cg_pid_get_machine_name(pid_t pid, char **ret_machine) {
1448         _cleanup_free_ char *cgroup = NULL;
1449         int r;
1450
1451         assert(ret_machine);
1452
1453         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1454         if (r < 0)
1455                 return r;
1456
1457         return cg_path_get_machine_name(cgroup, ret_machine);
1458 }
1459
1460 int cg_path_get_cgroupid(const char *path, uint64_t *ret) {
1461         cg_file_handle fh = CG_FILE_HANDLE_INIT;
1462         int mnt_id;
1463
1464         assert(path);
1465         assert(ret);
1466
1467         /* This is cgroupfs so we know the size of the handle, thus no need to loop around like
1468          * name_to_handle_at_loop() does in mountpoint-util.c */
1469         if (name_to_handle_at(AT_FDCWD, path, &fh.file_handle, &mnt_id, 0) < 0)
1470                 return -errno;
1471
1472         *ret = CG_FILE_HANDLE_CGROUPID(fh);
1473         return 0;
1474 }
1475
1476 int cg_fd_get_cgroupid(int fd, uint64_t *ret) {
1477         cg_file_handle fh = CG_FILE_HANDLE_INIT;
1478         int mnt_id = -1;
1479
1480         assert(fd >= 0);
1481         assert(ret);
1482
1483         if (name_to_handle_at(fd, "", &fh.file_handle, &mnt_id, AT_EMPTY_PATH) < 0)
1484                 return -errno;
1485
1486         *ret = CG_FILE_HANDLE_CGROUPID(fh);
1487         return 0;
1488 }
1489
1490 int cg_path_get_session(const char *path, char **ret_session) {
1491         _cleanup_free_ char *unit = NULL;
1492         char *start, *end;
1493         int r;
1494
1495         assert(path);
1496
1497         r = cg_path_get_unit(path, &unit);
1498         if (r < 0)
1499                 return r;
1500
1501         start = startswith(unit, "session-");
1502         if (!start)
1503                 return -ENXIO;
1504         end = endswith(start, ".scope");
1505         if (!end)
1506                 return -ENXIO;
1507
1508         *end = 0;
1509         if (!session_id_valid(start))
1510                 return -ENXIO;
1511
1512         if (!ret_session)
1513                 return 0;
1514
1515         return strdup_to(ret_session, start);
1516 }
1517
1518 int cg_pid_get_session(pid_t pid, char **ret_session) {
1519         _cleanup_free_ char *cgroup = NULL;
1520         int r;
1521
1522         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1523         if (r < 0)
1524                 return r;
1525
1526         return cg_path_get_session(cgroup, ret_session);
1527 }
1528
1529 int cg_path_get_owner_uid(const char *path, uid_t *ret_uid) {
1530         _cleanup_free_ char *slice = NULL;
1531         char *start, *end;
1532         int r;
1533
1534         assert(path);
1535
1536         r = cg_path_get_slice(path, &slice);
1537         if (r < 0)
1538                 return r;
1539
1540         start = startswith(slice, "user-");
1541         if (!start)
1542                 return -ENXIO;
1543
1544         end = endswith(start, ".slice");
1545         if (!end)
1546                 return -ENXIO;
1547
1548         *end = 0;
1549         if (parse_uid(start, ret_uid) < 0)
1550                 return -ENXIO;
1551
1552         return 0;
1553 }
1554
1555 int cg_pid_get_owner_uid(pid_t pid, uid_t *ret_uid) {
1556         _cleanup_free_ char *cgroup = NULL;
1557         int r;
1558
1559         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1560         if (r < 0)
1561                 return r;
1562
1563         return cg_path_get_owner_uid(cgroup, ret_uid);
1564 }
1565
1566 int cg_path_get_slice(const char *p, char **ret_slice) {
1567         const char *e = NULL;
1568
1569         assert(p);
1570         assert(ret_slice);
1571
1572         /* Finds the right-most slice unit from the beginning, but stops before we come to
1573          * the first non-slice unit. */
1574
1575         for (;;) {
1576                 const char *s;
1577                 int n;
1578
1579                 n = path_find_first_component(&p, /* accept_dot_dot = */ false, &s);
1580                 if (n < 0)
1581                         return n;
1582                 if (!valid_slice_name(s, n))
1583                         break;
1584
1585                 e = s;
1586         }
1587
1588         if (e)
1589                 return cg_path_decode_unit(e, ret_slice);
1590
1591         return strdup_to(ret_slice, SPECIAL_ROOT_SLICE);
1592 }
1593
1594 int cg_pid_get_slice(pid_t pid, char **ret_slice) {
1595         _cleanup_free_ char *cgroup = NULL;
1596         int r;
1597
1598         assert(ret_slice);
1599
1600         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1601         if (r < 0)
1602                 return r;
1603
1604         return cg_path_get_slice(cgroup, ret_slice);
1605 }
1606
1607 int cg_path_get_user_slice(const char *p, char **ret_slice) {
1608         const char *t;
1609         assert(p);
1610         assert(ret_slice);
1611
1612         t = skip_user_prefix(p);
1613         if (!t)
1614                 return -ENXIO;
1615
1616         /* And now it looks pretty much the same as for a system slice, so let's just use the same parser
1617          * from here on. */
1618         return cg_path_get_slice(t, ret_slice);
1619 }
1620
1621 int cg_pid_get_user_slice(pid_t pid, char **ret_slice) {
1622         _cleanup_free_ char *cgroup = NULL;
1623         int r;
1624
1625         assert(ret_slice);
1626
1627         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1628         if (r < 0)
1629                 return r;
1630
1631         return cg_path_get_user_slice(cgroup, ret_slice);
1632 }
1633
1634 bool cg_needs_escape(const char *p) {
1635
1636         /* Checks if the specified path is a valid cgroup name by our rules, or if it must be escaped. Note
1637          * that we consider escaped cgroup names invalid here, as they need to be escaped a second time if
1638          * they shall be used. Also note that various names cannot be made valid by escaping even if we
1639          * return true here (because too long, or contain the forbidden character "/"). */
1640
1641         if (!filename_is_valid(p))
1642                 return true;
1643
1644         if (IN_SET(p[0], '_', '.'))
1645                 return true;
1646
1647         if (STR_IN_SET(p, "notify_on_release", "release_agent", "tasks"))
1648                 return true;
1649
1650         if (startswith(p, "cgroup."))
1651                 return true;
1652
1653         for (CGroupController c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1654                 const char *q;
1655
1656                 q = startswith(p, cgroup_controller_to_string(c));
1657                 if (!q)
1658                         continue;
1659
1660                 if (q[0] == '.')
1661                         return true;
1662         }
1663
1664         return false;
1665 }
1666
1667 int cg_escape(const char *p, char **ret) {
1668         _cleanup_free_ char *n = NULL;
1669
1670         /* This implements very minimal escaping for names to be used as file names in the cgroup tree: any
1671          * name which might conflict with a kernel name or is prefixed with '_' is prefixed with a '_'. That
1672          * way, when reading cgroup names it is sufficient to remove a single prefixing underscore if there
1673          * is one. */
1674
1675         /* The return value of this function (unlike cg_unescape()) needs free()! */
1676
1677         if (cg_needs_escape(p)) {
1678                 n = strjoin("_", p);
1679                 if (!n)
1680                         return -ENOMEM;
1681
1682                 if (!filename_is_valid(n)) /* became invalid due to the prefixing? Or contained things like a slash that cannot be fixed by prefixing? */
1683                         return -EINVAL;
1684         } else {
1685                 n = strdup(p);
1686                 if (!n)
1687                         return -ENOMEM;
1688         }
1689
1690         *ret = TAKE_PTR(n);
1691         return 0;
1692 }
1693
1694 char *cg_unescape(const char *p) {
1695         assert(p);
1696
1697         /* The return value of this function (unlike cg_escape())
1698          * doesn't need free()! */
1699
1700         if (p[0] == '_')
1701                 return (char*) p+1;
1702
1703         return (char*) p;
1704 }
1705
1706 #define CONTROLLER_VALID                        \
1707         DIGITS LETTERS                          \
1708         "_"
1709
1710 bool cg_controller_is_valid(const char *p) {
1711         const char *t, *s;
1712
1713         if (!p)
1714                 return false;
1715
1716         if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1717                 return true;
1718
1719         s = startswith(p, "name=");
1720         if (s)
1721                 p = s;
1722
1723         if (IN_SET(*p, 0, '_'))
1724                 return false;
1725
1726         for (t = p; *t; t++)
1727                 if (!strchr(CONTROLLER_VALID, *t))
1728                         return false;
1729
1730         if (t - p > NAME_MAX)
1731                 return false;
1732
1733         return true;
1734 }
1735
1736 int cg_slice_to_path(const char *unit, char **ret) {
1737         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1738         const char *dash;
1739         int r;
1740
1741         assert(unit);
1742         assert(ret);
1743
1744         if (streq(unit, SPECIAL_ROOT_SLICE))
1745                 return strdup_to(ret, "");
1746
1747         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1748                 return -EINVAL;
1749
1750         if (!endswith(unit, ".slice"))
1751                 return -EINVAL;
1752
1753         r = unit_name_to_prefix(unit, &p);
1754         if (r < 0)
1755                 return r;
1756
1757         dash = strchr(p, '-');
1758
1759         /* Don't allow initial dashes */
1760         if (dash == p)
1761                 return -EINVAL;
1762
1763         while (dash) {
1764                 _cleanup_free_ char *escaped = NULL;
1765                 char n[dash - p + sizeof(".slice")];
1766
1767 #if HAS_FEATURE_MEMORY_SANITIZER
1768                 /* msan doesn't instrument stpncpy, so it thinks
1769                  * n is later used uninitialized:
1770                  * https://github.com/google/sanitizers/issues/926
1771                  */
1772                 zero(n);
1773 #endif
1774
1775                 /* Don't allow trailing or double dashes */
1776                 if (IN_SET(dash[1], 0, '-'))
1777                         return -EINVAL;
1778
1779                 strcpy(stpncpy(n, p, dash - p), ".slice");
1780                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1781                         return -EINVAL;
1782
1783                 r = cg_escape(n, &escaped);
1784                 if (r < 0)
1785                         return r;
1786
1787                 if (!strextend(&s, escaped, "/"))
1788                         return -ENOMEM;
1789
1790                 dash = strchr(dash+1, '-');
1791         }
1792
1793         r = cg_escape(unit, &e);
1794         if (r < 0)
1795                 return r;
1796
1797         if (!strextend(&s, e))
1798                 return -ENOMEM;
1799
1800         *ret = TAKE_PTR(s);
1801         return 0;
1802 }
1803
1804 int cg_is_threaded(const char *path) {
1805         _cleanup_free_ char *fs = NULL, *contents = NULL;
1806         _cleanup_strv_free_ char **v = NULL;
1807         int r;
1808
1809         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, "cgroup.type", &fs);
1810         if (r < 0)
1811                 return r;
1812
1813         r = read_full_virtual_file(fs, &contents, NULL);
1814         if (r == -ENOENT)
1815                 return false; /* Assume no. */
1816         if (r < 0)
1817                 return r;
1818
1819         v = strv_split(contents, NULL);
1820         if (!v)
1821                 return -ENOMEM;
1822
1823         /* If the cgroup is in the threaded mode, it contains "threaded".
1824          * If one of the parents or siblings is in the threaded mode, it may contain "invalid". */
1825         return strv_contains(v, "threaded") || strv_contains(v, "invalid");
1826 }
1827
1828 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
1829         _cleanup_free_ char *p = NULL;
1830         int r;
1831
1832         r = cg_get_path(controller, path, attribute, &p);
1833         if (r < 0)
1834                 return r;
1835
1836         return write_string_file(p, value, WRITE_STRING_FILE_DISABLE_BUFFER);
1837 }
1838
1839 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
1840         _cleanup_free_ char *p = NULL;
1841         int r;
1842
1843         r = cg_get_path(controller, path, attribute, &p);
1844         if (r < 0)
1845                 return r;
1846
1847         return read_one_line_file(p, ret);
1848 }
1849
1850 int cg_get_attribute_as_uint64(const char *controller, const char *path, const char *attribute, uint64_t *ret) {
1851         _cleanup_free_ char *value = NULL;
1852         uint64_t v;
1853         int r;
1854
1855         assert(ret);
1856
1857         r = cg_get_attribute(controller, path, attribute, &value);
1858         if (r == -ENOENT)
1859                 return -ENODATA;
1860         if (r < 0)
1861                 return r;
1862
1863         if (streq(value, "max")) {
1864                 *ret = CGROUP_LIMIT_MAX;
1865                 return 0;
1866         }
1867
1868         r = safe_atou64(value, &v);
1869         if (r < 0)
1870                 return r;
1871
1872         *ret = v;
1873         return 0;
1874 }
1875
1876 int cg_get_attribute_as_bool(const char *controller, const char *path, const char *attribute, bool *ret) {
1877         _cleanup_free_ char *value = NULL;
1878         int r;
1879
1880         assert(ret);
1881
1882         r = cg_get_attribute(controller, path, attribute, &value);
1883         if (r == -ENOENT)
1884                 return -ENODATA;
1885         if (r < 0)
1886                 return r;
1887
1888         r = parse_boolean(value);
1889         if (r < 0)
1890                 return r;
1891
1892         *ret = r;
1893         return 0;
1894 }
1895
1896 int cg_get_owner(const char *path, uid_t *ret_uid) {
1897         _cleanup_free_ char *f = NULL;
1898         struct stat stats;
1899         int r;
1900
1901         assert(ret_uid);
1902
1903         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, path, NULL, &f);
1904         if (r < 0)
1905                 return r;
1906
1907         if (stat(f, &stats) < 0)
1908                 return -errno;
1909
1910         r = stat_verify_directory(&stats);
1911         if (r < 0)
1912                 return r;
1913
1914         *ret_uid = stats.st_uid;
1915         return 0;
1916 }
1917
1918 int cg_get_keyed_attribute_full(
1919                 const char *controller,
1920                 const char *path,
1921                 const char *attribute,
1922                 char **keys,
1923                 char **ret_values,
1924                 CGroupKeyMode mode) {
1925
1926         _cleanup_free_ char *filename = NULL, *contents = NULL;
1927         const char *p;
1928         size_t n, i, n_done = 0;
1929         char **v;
1930         int r;
1931
1932         /* Reads one or more fields of a cgroup v2 keyed attribute file. The 'keys' parameter should be an strv with
1933          * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
1934          * entries as 'keys'. On success each entry will be set to the value of the matching key.
1935          *
1936          * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. If mode
1937          * is set to GG_KEY_MODE_GRACEFUL we ignore missing keys and return those that were parsed successfully. */
1938
1939         r = cg_get_path(controller, path, attribute, &filename);
1940         if (r < 0)
1941                 return r;
1942
1943         r = read_full_file(filename, &contents, NULL);
1944         if (r < 0)
1945                 return r;
1946
1947         n = strv_length(keys);
1948         if (n == 0) /* No keys to retrieve? That's easy, we are done then */
1949                 return 0;
1950
1951         /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
1952         v = newa0(char*, n);
1953
1954         for (p = contents; *p;) {
1955                 const char *w = NULL;
1956
1957                 for (i = 0; i < n; i++)
1958                         if (!v[i]) {
1959                                 w = first_word(p, keys[i]);
1960                                 if (w)
1961                                         break;
1962                         }
1963
1964                 if (w) {
1965                         size_t l;
1966
1967                         l = strcspn(w, NEWLINE);
1968                         v[i] = strndup(w, l);
1969                         if (!v[i]) {
1970                                 r = -ENOMEM;
1971                                 goto fail;
1972                         }
1973
1974                         n_done++;
1975                         if (n_done >= n)
1976                                 goto done;
1977
1978                         p = w + l;
1979                 } else
1980                         p += strcspn(p, NEWLINE);
1981
1982                 p += strspn(p, NEWLINE);
1983         }
1984
1985         if (mode & CG_KEY_MODE_GRACEFUL)
1986                 goto done;
1987
1988         r = -ENXIO;
1989
1990 fail:
1991         free_many_charp(v, n);
1992         return r;
1993
1994 done:
1995         memcpy(ret_values, v, sizeof(char*) * n);
1996         if (mode & CG_KEY_MODE_GRACEFUL)
1997                 return n_done;
1998
1999         return 0;
2000 }
2001
2002 int cg_mask_to_string(CGroupMask mask, char **ret) {
2003         _cleanup_free_ char *s = NULL;
2004         bool space = false;
2005         CGroupController c;
2006         size_t n = 0;
2007
2008         assert(ret);
2009
2010         if (mask == 0) {
2011                 *ret = NULL;
2012                 return 0;
2013         }
2014
2015         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2016                 const char *k;
2017                 size_t l;
2018
2019                 if (!FLAGS_SET(mask, CGROUP_CONTROLLER_TO_MASK(c)))
2020                         continue;
2021
2022                 k = cgroup_controller_to_string(c);
2023                 l = strlen(k);
2024
2025                 if (!GREEDY_REALLOC(s, n + space + l + 1))
2026                         return -ENOMEM;
2027
2028                 if (space)
2029                         s[n] = ' ';
2030                 memcpy(s + n + space, k, l);
2031                 n += space + l;
2032
2033                 space = true;
2034         }
2035
2036         assert(s);
2037
2038         s[n] = 0;
2039         *ret = TAKE_PTR(s);
2040
2041         return 0;
2042 }
2043
2044 int cg_mask_from_string(const char *value, CGroupMask *ret) {
2045         CGroupMask m = 0;
2046
2047         assert(ret);
2048         assert(value);
2049
2050         for (;;) {
2051                 _cleanup_free_ char *n = NULL;
2052                 CGroupController v;
2053                 int r;
2054
2055                 r = extract_first_word(&value, &n, NULL, 0);
2056                 if (r < 0)
2057                         return r;
2058                 if (r == 0)
2059                         break;
2060
2061                 v = cgroup_controller_from_string(n);
2062                 if (v < 0)
2063                         continue;
2064
2065                 m |= CGROUP_CONTROLLER_TO_MASK(v);
2066         }
2067
2068         *ret = m;
2069         return 0;
2070 }
2071
2072 int cg_mask_supported_subtree(const char *root, CGroupMask *ret) {
2073         CGroupMask mask;
2074         int r;
2075
2076         /* Determines the mask of supported cgroup controllers. Only includes controllers we can make sense of and that
2077          * are actually accessible. Only covers real controllers, i.e. not the CGROUP_CONTROLLER_BPF_xyz
2078          * pseudo-controllers. */
2079
2080         r = cg_all_unified();
2081         if (r < 0)
2082                 return r;
2083         if (r > 0) {
2084                 _cleanup_free_ char *controllers = NULL, *path = NULL;
2085
2086                 /* In the unified hierarchy we can read the supported and accessible controllers from
2087                  * the top-level cgroup attribute */
2088
2089                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
2090                 if (r < 0)
2091                         return r;
2092
2093                 r = read_one_line_file(path, &controllers);
2094                 if (r < 0)
2095                         return r;
2096
2097                 r = cg_mask_from_string(controllers, &mask);
2098                 if (r < 0)
2099                         return r;
2100
2101                 /* Mask controllers that are not supported in unified hierarchy. */
2102                 mask &= CGROUP_MASK_V2;
2103
2104         } else {
2105                 CGroupController c;
2106
2107                 /* In the legacy hierarchy, we check which hierarchies are accessible. */
2108
2109                 mask = 0;
2110                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2111                         CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2112                         const char *n;
2113
2114                         if (!FLAGS_SET(CGROUP_MASK_V1, bit))
2115                                 continue;
2116
2117                         n = cgroup_controller_to_string(c);
2118                         if (controller_is_v1_accessible(root, n) >= 0)
2119                                 mask |= bit;
2120                 }
2121         }
2122
2123         *ret = mask;
2124         return 0;
2125 }
2126
2127 int cg_mask_supported(CGroupMask *ret) {
2128         _cleanup_free_ char *root = NULL;
2129         int r;
2130
2131         r = cg_get_root_path(&root);
2132         if (r < 0)
2133                 return r;
2134
2135         return cg_mask_supported_subtree(root, ret);
2136 }
2137
2138 int cg_kernel_controllers(Set **ret) {
2139         _cleanup_set_free_ Set *controllers = NULL;
2140         _cleanup_fclose_ FILE *f = NULL;
2141         int r;
2142
2143         assert(ret);
2144
2145         /* Determines the full list of kernel-known controllers. Might include controllers we don't actually support
2146          * and controllers that aren't currently accessible (because not mounted). This does not include "name="
2147          * pseudo-controllers. */
2148
2149         r = fopen_unlocked("/proc/cgroups", "re", &f);
2150         if (r == -ENOENT) {
2151                 *ret = NULL;
2152                 return 0;
2153         }
2154         if (r < 0)
2155                 return r;
2156
2157         /* Ignore the header line */
2158         (void) read_line(f, SIZE_MAX, NULL);
2159
2160         for (;;) {
2161                 _cleanup_free_ char *controller = NULL;
2162                 int enabled = 0;
2163
2164                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
2165
2166                         if (ferror(f))
2167                                 return -errno;
2168
2169                         if (feof(f))
2170                                 break;
2171
2172                         return -EBADMSG;
2173                 }
2174
2175                 if (!enabled)
2176                         continue;
2177
2178                 if (!cg_controller_is_valid(controller))
2179                         return -EBADMSG;
2180
2181                 r = set_ensure_consume(&controllers, &string_hash_ops_free, TAKE_PTR(controller));
2182                 if (r < 0)
2183                         return r;
2184         }
2185
2186         *ret = TAKE_PTR(controllers);
2187
2188         return 0;
2189 }
2190
2191 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup2 on
2192  * /sys/fs/cgroup/systemd. This unfortunately broke other tools (such as docker) which expected the v1
2193  * "name=systemd" hierarchy on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mounts v2 on
2194  * /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility
2195  * with other tools.
2196  *
2197  * To keep live upgrade working, we detect and support v232 layout. When v232 layout is detected, to keep
2198  * cgroup v2 process management but disable the compat dual layout, we return true on
2199  * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and false on cg_hybrid_unified().
2200  */
2201 static thread_local bool unified_systemd_v232;
2202
2203 int cg_unified_cached(bool flush) {
2204         static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2205
2206         struct statfs fs;
2207
2208         /* Checks if we support the unified hierarchy. Returns an
2209          * error when the cgroup hierarchies aren't mounted yet or we
2210          * have any other trouble determining if the unified hierarchy
2211          * is supported. */
2212
2213         if (flush)
2214                 unified_cache = CGROUP_UNIFIED_UNKNOWN;
2215         else if (unified_cache >= CGROUP_UNIFIED_NONE)
2216                 return unified_cache;
2217
2218         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2219                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2220
2221         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2222                 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2223                 unified_cache = CGROUP_UNIFIED_ALL;
2224         } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2225                 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2226                     F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2227                         log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2228                         unified_cache = CGROUP_UNIFIED_SYSTEMD;
2229                         unified_systemd_v232 = false;
2230                 } else {
2231                         if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0) {
2232                                 if (errno == ENOENT) {
2233                                         /* Some other software may have set up /sys/fs/cgroup in a configuration we do not recognize. */
2234                                         log_debug_errno(errno, "Unsupported cgroupsv1 setup detected: name=systemd hierarchy not found.");
2235                                         return -ENOMEDIUM;
2236                                 }
2237                                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2238                         }
2239
2240                         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2241                                 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2242                                 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2243                                 unified_systemd_v232 = true;
2244                         } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2245                                 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2246                                 unified_cache = CGROUP_UNIFIED_NONE;
2247                         } else {
2248                                 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2249                                           (unsigned long long) fs.f_type);
2250                                 unified_cache = CGROUP_UNIFIED_NONE;
2251                         }
2252                 }
2253         } else if (F_TYPE_EQUAL(fs.f_type, SYSFS_MAGIC)) {
2254                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2255                                        "No filesystem is currently mounted on /sys/fs/cgroup.");
2256         } else
2257                 return log_debug_errno(SYNTHETIC_ERRNO(ENOMEDIUM),
2258                                        "Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2259                                        (unsigned long long)fs.f_type);
2260
2261         return unified_cache;
2262 }
2263
2264 int cg_unified_controller(const char *controller) {
2265         int r;
2266
2267         r = cg_unified_cached(false);
2268         if (r < 0)
2269                 return r;
2270
2271         if (r == CGROUP_UNIFIED_NONE)
2272                 return false;
2273
2274         if (r >= CGROUP_UNIFIED_ALL)
2275                 return true;
2276
2277         return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2278 }
2279
2280 int cg_all_unified(void) {
2281         int r;
2282
2283         r = cg_unified_cached(false);
2284         if (r < 0)
2285                 return r;
2286
2287         return r >= CGROUP_UNIFIED_ALL;
2288 }
2289
2290 int cg_hybrid_unified(void) {
2291         int r;
2292
2293         r = cg_unified_cached(false);
2294         if (r < 0)
2295                 return r;
2296
2297         return r == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2298 }
2299
2300 int cg_is_delegated(const char *path) {
2301         int r;
2302
2303         assert(path);
2304
2305         r = cg_get_xattr_bool(path, "trusted.delegate");
2306         if (!ERRNO_IS_NEG_XATTR_ABSENT(r))
2307                 return r;
2308
2309         /* If the trusted xattr isn't set (preferred), then check the untrusted one. Under the assumption
2310          * that whoever is trusted enough to own the cgroup, is also trusted enough to decide if it is
2311          * delegated or not this should be safe. */
2312         r = cg_get_xattr_bool(path, "user.delegate");
2313         return ERRNO_IS_NEG_XATTR_ABSENT(r) ? false : r;
2314 }
2315
2316 int cg_is_delegated_fd(int fd) {
2317         int r;
2318
2319         assert(fd >= 0);
2320
2321         r = getxattr_at_bool(fd, /* path= */ NULL, "trusted.delegate", /* flags= */ 0);
2322         if (!ERRNO_IS_NEG_XATTR_ABSENT(r))
2323                 return r;
2324
2325         r = getxattr_at_bool(fd, /* path= */ NULL, "user.delegate", /* flags= */ 0);
2326         return ERRNO_IS_NEG_XATTR_ABSENT(r) ? false : r;
2327 }
2328
2329 int cg_has_coredump_receive(const char *path) {
2330         int r;
2331
2332         assert(path);
2333
2334         r = cg_get_xattr_bool(path, "user.coredump_receive");
2335         if (ERRNO_IS_NEG_XATTR_ABSENT(r))
2336                 return false;
2337
2338         return r;
2339 }
2340
2341 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2342         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2343         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2344         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2345         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2346 };
2347
2348 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2349         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2350         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2351         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2352         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2353 };
2354
2355 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2356
2357 bool is_cgroup_fs(const struct statfs *s) {
2358         return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2359                is_fs_type(s, CGROUP2_SUPER_MAGIC);
2360 }
2361
2362 bool fd_is_cgroup_fs(int fd) {
2363         struct statfs s;
2364
2365         if (fstatfs(fd, &s) < 0)
2366                 return -errno;
2367
2368         return is_cgroup_fs(&s);
2369 }
2370
2371 static const char *const cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2372         [CGROUP_CONTROLLER_CPU] = "cpu",
2373         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2374         [CGROUP_CONTROLLER_CPUSET] = "cpuset",
2375         [CGROUP_CONTROLLER_IO] = "io",
2376         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2377         [CGROUP_CONTROLLER_MEMORY] = "memory",
2378         [CGROUP_CONTROLLER_DEVICES] = "devices",
2379         [CGROUP_CONTROLLER_PIDS] = "pids",
2380         [CGROUP_CONTROLLER_BPF_FIREWALL] = "bpf-firewall",
2381         [CGROUP_CONTROLLER_BPF_DEVICES] = "bpf-devices",
2382         [CGROUP_CONTROLLER_BPF_FOREIGN] = "bpf-foreign",
2383         [CGROUP_CONTROLLER_BPF_SOCKET_BIND] = "bpf-socket-bind",
2384         [CGROUP_CONTROLLER_BPF_RESTRICT_NETWORK_INTERFACES] = "bpf-restrict-network-interfaces",
2385 };
2386
2387 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);
2388
2389 CGroupMask get_cpu_accounting_mask(void) {
2390         static CGroupMask needed_mask = (CGroupMask) -1;
2391
2392         /* On kernel ≥4.15 with unified hierarchy, cpu.stat's usage_usec is
2393          * provided externally from the CPU controller, which means we don't
2394          * need to enable the CPU controller just to get metrics. This is good,
2395          * because enabling the CPU controller comes at a minor performance
2396          * hit, especially when it's propagated deep into large hierarchies.
2397          * There's also no separate CPU accounting controller available within
2398          * a unified hierarchy.
2399          *
2400          * This combination of factors results in the desired cgroup mask to
2401          * enable for CPU accounting varying as follows:
2402          *
2403          *                   ╔═════════════════════╤═════════════════════╗
2404          *                   ║     Linux ≥4.15     │     Linux <4.15     ║
2405          *   ╔═══════════════╬═════════════════════╪═════════════════════╣
2406          *   ║ Unified       ║ nothing             │ CGROUP_MASK_CPU     ║
2407          *   ╟───────────────╫─────────────────────┼─────────────────────╢
2408          *   ║ Hybrid/Legacy ║ CGROUP_MASK_CPUACCT │ CGROUP_MASK_CPUACCT ║
2409          *   ╚═══════════════╩═════════════════════╧═════════════════════╝
2410          *
2411          * We check kernel version here instead of manually checking whether
2412          * cpu.stat is present for every cgroup, as that check in itself would
2413          * already be fairly expensive.
2414          *
2415          * Kernels where this patch has been backported will therefore have the
2416          * CPU controller enabled unnecessarily. This is more expensive than
2417          * necessary, but harmless. ☺️
2418          */
2419
2420         if (needed_mask == (CGroupMask) -1) {
2421                 if (cg_all_unified()) {
2422                         struct utsname u;
2423                         assert_se(uname(&u) >= 0);
2424
2425                         if (strverscmp_improved(u.release, "4.15") < 0)
2426                                 needed_mask = CGROUP_MASK_CPU;
2427                         else
2428                                 needed_mask = 0;
2429                 } else
2430                         needed_mask = CGROUP_MASK_CPUACCT;
2431         }
2432
2433         return needed_mask;
2434 }
2435
2436 bool cpu_accounting_is_cheap(void) {
2437         return get_cpu_accounting_mask() == 0;
2438 }
2439
2440 static const char* const managed_oom_mode_table[_MANAGED_OOM_MODE_MAX] = {
2441         [MANAGED_OOM_AUTO] = "auto",
2442         [MANAGED_OOM_KILL] = "kill",
2443 };
2444
2445 DEFINE_STRING_TABLE_LOOKUP(managed_oom_mode, ManagedOOMMode);
2446
2447 static const char* const managed_oom_preference_table[_MANAGED_OOM_PREFERENCE_MAX] = {
2448         [MANAGED_OOM_PREFERENCE_NONE] = "none",
2449         [MANAGED_OOM_PREFERENCE_AVOID] = "avoid",
2450         [MANAGED_OOM_PREFERENCE_OMIT] = "omit",
2451 };
2452
2453 DEFINE_STRING_TABLE_LOOKUP(managed_oom_preference, ManagedOOMPreference);