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