src/basic/cgroup-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2 /***
   3   This file is part of systemd.
   4
   5   Copyright 2010 Lennart Poettering
   6
   7   systemd is free software; you can redistribute it and/or modify it
   8   under the terms of the GNU Lesser General Public License as published by
   9   the Free Software Foundation; either version 2.1 of the License, or
  10   (at your option) any later version.
  11
  12   systemd is distributed in the hope that it will be useful, but
  13   WITHOUT ANY WARRANTY; without even the implied warranty of
  14   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15   Lesser General Public License for more details.
  16
  17   You should have received a copy of the GNU Lesser General Public License
  18   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  19 ***/
  20
  21 #include <dirent.h>
  22 #include <errno.h>
  23 #include <ftw.h>
  24 #include <limits.h>
  25 #include <signal.h>
  26 #include <stddef.h>
  27 #include <stdio_ext.h>
  28 #include <stdlib.h>
  29 #include <string.h>
  30 #include <sys/stat.h>
  31 #include <sys/statfs.h>
  32 #include <sys/types.h>
  33 #include <sys/xattr.h>
  34 #include <unistd.h>
  35
  36 #include "alloc-util.h"
  37 #include "cgroup-util.h"
  38 #include "def.h"
  39 #include "dirent-util.h"
  40 #include "extract-word.h"
  41 #include "fd-util.h"
  42 #include "fileio.h"
  43 #include "format-util.h"
  44 #include "fs-util.h"
  45 #include "log.h"
  46 #include "login-util.h"
  47 #include "macro.h"
  48 #include "missing.h"
  49 #include "mkdir.h"
  50 #include "parse-util.h"
  51 #include "path-util.h"
  52 #include "proc-cmdline.h"
  53 #include "process-util.h"
  54 #include "set.h"
  55 #include "special.h"
  56 #include "stat-util.h"
  57 #include "stdio-util.h"
  58 #include "string-table.h"
  59 #include "string-util.h"
  60 #include "strv.h"
  61 #include "unit-name.h"
  62 #include "user-util.h"
  63
  64 int cg_enumerate_processes(const char *controller, const char *path, FILE **_f) {
  65         _cleanup_free_ char *fs = NULL;
  66         FILE *f;
  67         int r;
  68
  69         assert(_f);
  70
  71         r = cg_get_path(controller, path, "cgroup.procs", &fs);
  72         if (r < 0)
  73                 return r;
  74
  75         f = fopen(fs, "re");
  76         if (!f)
  77                 return -errno;
  78
  79         *_f = f;
  80         return 0;
  81 }
  82
  83 int cg_read_pid(FILE *f, pid_t *_pid) {
  84         unsigned long ul;
  85
  86         /* Note that the cgroup.procs might contain duplicates! See
  87          * cgroups.txt for details. */
  88
  89         assert(f);
  90         assert(_pid);
  91
  92         errno = 0;
  93         if (fscanf(f, "%lu", &ul) != 1) {
  94
  95                 if (feof(f))
  96                         return 0;
  97
  98                 return errno > 0 ? -errno : -EIO;
  99         }
 100
 101         if (ul <= 0)
 102                 return -EIO;
 103
 104         *_pid = (pid_t) ul;
 105         return 1;
 106 }
 107
 108 int cg_read_event(
 109                 const char *controller,
 110                 const char *path,
 111                 const char *event,
 112                 char **val) {
 113
 114         _cleanup_free_ char *events = NULL, *content = NULL;
 115         char *p, *line;
 116         int r;
 117
 118         r = cg_get_path(controller, path, "cgroup.events", &events);
 119         if (r < 0)
 120                 return r;
 121
 122         r = read_full_file(events, &content, NULL);
 123         if (r < 0)
 124                 return r;
 125
 126         p = content;
 127         while ((line = strsep(&p, "\n"))) {
 128                 char *key;
 129
 130                 key = strsep(&line, " ");
 131                 if (!key || !line)
 132                         return -EINVAL;
 133
 134                 if (strcmp(key, event))
 135                         continue;
 136
 137                 *val = strdup(line);
 138                 return 0;
 139         }
 140
 141         return -ENOENT;
 142 }
 143
 144 bool cg_ns_supported(void) {
 145         static thread_local int enabled = -1;
 146
 147         if (enabled >= 0)
 148                 return enabled;
 149
 150         if (access("/proc/self/ns/cgroup", F_OK) == 0)
 151                 enabled = 1;
 152         else
 153                 enabled = 0;
 154
 155         return enabled;
 156 }
 157
 158 int cg_enumerate_subgroups(const char *controller, const char *path, DIR **_d) {
 159         _cleanup_free_ char *fs = NULL;
 160         int r;
 161         DIR *d;
 162
 163         assert(_d);
 164
 165         /* This is not recursive! */
 166
 167         r = cg_get_path(controller, path, NULL, &fs);
 168         if (r < 0)
 169                 return r;
 170
 171         d = opendir(fs);
 172         if (!d)
 173                 return -errno;
 174
 175         *_d = d;
 176         return 0;
 177 }
 178
 179 int cg_read_subgroup(DIR *d, char **fn) {
 180         struct dirent *de;
 181
 182         assert(d);
 183         assert(fn);
 184
 185         FOREACH_DIRENT_ALL(de, d, return -errno) {
 186                 char *b;
 187
 188                 if (de->d_type != DT_DIR)
 189                         continue;
 190
 191                 if (dot_or_dot_dot(de->d_name))
 192                         continue;
 193
 194                 b = strdup(de->d_name);
 195                 if (!b)
 196                         return -ENOMEM;
 197
 198                 *fn = b;
 199                 return 1;
 200         }
 201
 202         return 0;
 203 }
 204
 205 int cg_rmdir(const char *controller, const char *path) {
 206         _cleanup_free_ char *p = NULL;
 207         int r;
 208
 209         r = cg_get_path(controller, path, NULL, &p);
 210         if (r < 0)
 211                 return r;
 212
 213         r = rmdir(p);
 214         if (r < 0 && errno != ENOENT)
 215                 return -errno;
 216
 217         r = cg_hybrid_unified();
 218         if (r < 0)
 219                 return r;
 220         if (r == 0)
 221                 return 0;
 222
 223         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 224                 r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
 225                 if (r < 0)
 226                         log_warning_errno(r, "Failed to remove compat systemd cgroup %s: %m", path);
 227         }
 228
 229         return 0;
 230 }
 231
 232 int cg_kill(
 233                 const char *controller,
 234                 const char *path,
 235                 int sig,
 236                 CGroupFlags flags,
 237                 Set *s,
 238                 cg_kill_log_func_t log_kill,
 239                 void *userdata) {
 240
 241         _cleanup_set_free_ Set *allocated_set = NULL;
 242         bool done = false;
 243         int r, ret = 0;
 244         pid_t my_pid;
 245
 246         assert(sig >= 0);
 247
 248          /* Don't send SIGCONT twice. Also, SIGKILL always works even when process is suspended, hence don't send
 249           * SIGCONT on SIGKILL. */
 250         if (IN_SET(sig, SIGCONT, SIGKILL))
 251                 flags &= ~CGROUP_SIGCONT;
 252
 253         /* This goes through the tasks list and kills them all. This
 254          * is repeated until no further processes are added to the
 255          * tasks list, to properly handle forking processes */
 256
 257         if (!s) {
 258                 s = allocated_set = set_new(NULL);
 259                 if (!s)
 260                         return -ENOMEM;
 261         }
 262
 263         my_pid = getpid_cached();
 264
 265         do {
 266                 _cleanup_fclose_ FILE *f = NULL;
 267                 pid_t pid = 0;
 268                 done = true;
 269
 270                 r = cg_enumerate_processes(controller, path, &f);
 271                 if (r < 0) {
 272                         if (ret >= 0 && r != -ENOENT)
 273                                 return r;
 274
 275                         return ret;
 276                 }
 277
 278                 while ((r = cg_read_pid(f, &pid)) > 0) {
 279
 280                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 281                                 continue;
 282
 283                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 284                                 continue;
 285
 286                         if (log_kill)
 287                                 log_kill(pid, sig, userdata);
 288
 289                         /* If we haven't killed this process yet, kill
 290                          * it */
 291                         if (kill(pid, sig) < 0) {
 292                                 if (ret >= 0 && errno != ESRCH)
 293                                         ret = -errno;
 294                         } else {
 295                                 if (flags & CGROUP_SIGCONT)
 296                                         (void) kill(pid, SIGCONT);
 297
 298                                 if (ret == 0)
 299                                         ret = 1;
 300                         }
 301
 302                         done = false;
 303
 304                         r = set_put(s, PID_TO_PTR(pid));
 305                         if (r < 0) {
 306                                 if (ret >= 0)
 307                                         return r;
 308
 309                                 return ret;
 310                         }
 311                 }
 312
 313                 if (r < 0) {
 314                         if (ret >= 0)
 315                                 return r;
 316
 317                         return ret;
 318                 }
 319
 320                 /* To avoid racing against processes which fork
 321                  * quicker than we can kill them we repeat this until
 322                  * no new pids need to be killed. */
 323
 324         } while (!done);
 325
 326         return ret;
 327 }
 328
 329 int cg_kill_recursive(
 330                 const char *controller,
 331                 const char *path,
 332                 int sig,
 333                 CGroupFlags flags,
 334                 Set *s,
 335                 cg_kill_log_func_t log_kill,
 336                 void *userdata) {
 337
 338         _cleanup_set_free_ Set *allocated_set = NULL;
 339         _cleanup_closedir_ DIR *d = NULL;
 340         int r, ret;
 341         char *fn;
 342
 343         assert(path);
 344         assert(sig >= 0);
 345
 346         if (!s) {
 347                 s = allocated_set = set_new(NULL);
 348                 if (!s)
 349                         return -ENOMEM;
 350         }
 351
 352         ret = cg_kill(controller, path, sig, flags, s, log_kill, userdata);
 353
 354         r = cg_enumerate_subgroups(controller, path, &d);
 355         if (r < 0) {
 356                 if (ret >= 0 && r != -ENOENT)
 357                         return r;
 358
 359                 return ret;
 360         }
 361
 362         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 363                 _cleanup_free_ char *p = NULL;
 364
 365                 p = strjoin(path, "/", fn);
 366                 free(fn);
 367                 if (!p)
 368                         return -ENOMEM;
 369
 370                 r = cg_kill_recursive(controller, p, sig, flags, s, log_kill, userdata);
 371                 if (r != 0 && ret >= 0)
 372                         ret = r;
 373         }
 374         if (ret >= 0 && r < 0)
 375                 ret = r;
 376
 377         if (flags & CGROUP_REMOVE) {
 378                 r = cg_rmdir(controller, path);
 379                 if (r < 0 && ret >= 0 && !IN_SET(r, -ENOENT, -EBUSY))
 380                         return r;
 381         }
 382
 383         return ret;
 384 }
 385
 386 int cg_migrate(
 387                 const char *cfrom,
 388                 const char *pfrom,
 389                 const char *cto,
 390                 const char *pto,
 391                 CGroupFlags flags) {
 392
 393         bool done = false;
 394         _cleanup_set_free_ Set *s = NULL;
 395         int r, ret = 0;
 396         pid_t my_pid;
 397
 398         assert(cfrom);
 399         assert(pfrom);
 400         assert(cto);
 401         assert(pto);
 402
 403         s = set_new(NULL);
 404         if (!s)
 405                 return -ENOMEM;
 406
 407         my_pid = getpid_cached();
 408
 409         do {
 410                 _cleanup_fclose_ FILE *f = NULL;
 411                 pid_t pid = 0;
 412                 done = true;
 413
 414                 r = cg_enumerate_processes(cfrom, pfrom, &f);
 415                 if (r < 0) {
 416                         if (ret >= 0 && r != -ENOENT)
 417                                 return r;
 418
 419                         return ret;
 420                 }
 421
 422                 while ((r = cg_read_pid(f, &pid)) > 0) {
 423
 424                         /* This might do weird stuff if we aren't a
 425                          * single-threaded program. However, we
 426                          * luckily know we are not */
 427                         if ((flags & CGROUP_IGNORE_SELF) && pid == my_pid)
 428                                 continue;
 429
 430                         if (set_get(s, PID_TO_PTR(pid)) == PID_TO_PTR(pid))
 431                                 continue;
 432
 433                         /* Ignore kernel threads. Since they can only
 434                          * exist in the root cgroup, we only check for
 435                          * them there. */
 436                         if (cfrom &&
 437                             (isempty(pfrom) || path_equal(pfrom, "/")) &&
 438                             is_kernel_thread(pid) > 0)
 439                                 continue;
 440
 441                         r = cg_attach(cto, pto, pid);
 442                         if (r < 0) {
 443                                 if (ret >= 0 && r != -ESRCH)
 444                                         ret = r;
 445                         } else if (ret == 0)
 446                                 ret = 1;
 447
 448                         done = false;
 449
 450                         r = set_put(s, PID_TO_PTR(pid));
 451                         if (r < 0) {
 452                                 if (ret >= 0)
 453                                         return r;
 454
 455                                 return ret;
 456                         }
 457                 }
 458
 459                 if (r < 0) {
 460                         if (ret >= 0)
 461                                 return r;
 462
 463                         return ret;
 464                 }
 465         } while (!done);
 466
 467         return ret;
 468 }
 469
 470 int cg_migrate_recursive(
 471                 const char *cfrom,
 472                 const char *pfrom,
 473                 const char *cto,
 474                 const char *pto,
 475                 CGroupFlags flags) {
 476
 477         _cleanup_closedir_ DIR *d = NULL;
 478         int r, ret = 0;
 479         char *fn;
 480
 481         assert(cfrom);
 482         assert(pfrom);
 483         assert(cto);
 484         assert(pto);
 485
 486         ret = cg_migrate(cfrom, pfrom, cto, pto, flags);
 487
 488         r = cg_enumerate_subgroups(cfrom, pfrom, &d);
 489         if (r < 0) {
 490                 if (ret >= 0 && r != -ENOENT)
 491                         return r;
 492
 493                 return ret;
 494         }
 495
 496         while ((r = cg_read_subgroup(d, &fn)) > 0) {
 497                 _cleanup_free_ char *p = NULL;
 498
 499                 p = strjoin(pfrom, "/", fn);
 500                 free(fn);
 501                 if (!p)
 502                         return -ENOMEM;
 503
 504                 r = cg_migrate_recursive(cfrom, p, cto, pto, flags);
 505                 if (r != 0 && ret >= 0)
 506                         ret = r;
 507         }
 508
 509         if (r < 0 && ret >= 0)
 510                 ret = r;
 511
 512         if (flags & CGROUP_REMOVE) {
 513                 r = cg_rmdir(cfrom, pfrom);
 514                 if (r < 0 && ret >= 0 && !IN_SET(r, -ENOENT, -EBUSY))
 515                         return r;
 516         }
 517
 518         return ret;
 519 }
 520
 521 int cg_migrate_recursive_fallback(
 522                 const char *cfrom,
 523                 const char *pfrom,
 524                 const char *cto,
 525                 const char *pto,
 526                 CGroupFlags flags) {
 527
 528         int r;
 529
 530         assert(cfrom);
 531         assert(pfrom);
 532         assert(cto);
 533         assert(pto);
 534
 535         r = cg_migrate_recursive(cfrom, pfrom, cto, pto, flags);
 536         if (r < 0) {
 537                 char prefix[strlen(pto) + 1];
 538
 539                 /* This didn't work? Then let's try all prefixes of the destination */
 540
 541                 PATH_FOREACH_PREFIX(prefix, pto) {
 542                         int q;
 543
 544                         q = cg_migrate_recursive(cfrom, pfrom, cto, prefix, flags);
 545                         if (q >= 0)
 546                                 return q;
 547                 }
 548         }
 549
 550         return r;
 551 }
 552
 553 static const char *controller_to_dirname(const char *controller) {
 554         const char *e;
 555
 556         assert(controller);
 557
 558         /* Converts a controller name to the directory name below
 559          * /sys/fs/cgroup/ we want to mount it to. Effectively, this
 560          * just cuts off the name= prefixed used for named
 561          * hierarchies, if it is specified. */
 562
 563         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 564                 if (cg_hybrid_unified() > 0)
 565                         controller = SYSTEMD_CGROUP_CONTROLLER_HYBRID;
 566                 else
 567                         controller = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
 568         }
 569
 570         e = startswith(controller, "name=");
 571         if (e)
 572                 return e;
 573
 574         return controller;
 575 }
 576
 577 static int join_path_legacy(const char *controller, const char *path, const char *suffix, char **fs) {
 578         const char *dn;
 579         char *t = NULL;
 580
 581         assert(fs);
 582         assert(controller);
 583
 584         dn = controller_to_dirname(controller);
 585
 586         if (isempty(path) && isempty(suffix))
 587                 t = strappend("/sys/fs/cgroup/", dn);
 588         else if (isempty(path))
 589                 t = strjoin("/sys/fs/cgroup/", dn, "/", suffix);
 590         else if (isempty(suffix))
 591                 t = strjoin("/sys/fs/cgroup/", dn, "/", path);
 592         else
 593                 t = strjoin("/sys/fs/cgroup/", dn, "/", path, "/", suffix);
 594         if (!t)
 595                 return -ENOMEM;
 596
 597         *fs = t;
 598         return 0;
 599 }
 600
 601 static int join_path_unified(const char *path, const char *suffix, char **fs) {
 602         char *t;
 603
 604         assert(fs);
 605
 606         if (isempty(path) && isempty(suffix))
 607                 t = strdup("/sys/fs/cgroup");
 608         else if (isempty(path))
 609                 t = strappend("/sys/fs/cgroup/", suffix);
 610         else if (isempty(suffix))
 611                 t = strappend("/sys/fs/cgroup/", path);
 612         else
 613                 t = strjoin("/sys/fs/cgroup/", path, "/", suffix);
 614         if (!t)
 615                 return -ENOMEM;
 616
 617         *fs = t;
 618         return 0;
 619 }
 620
 621 int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) {
 622         int r;
 623
 624         assert(fs);
 625
 626         if (!controller) {
 627                 char *t;
 628
 629                 /* If no controller is specified, we return the path
 630                  * *below* the controllers, without any prefix. */
 631
 632                 if (!path && !suffix)
 633                         return -EINVAL;
 634
 635                 if (!suffix)
 636                         t = strdup(path);
 637                 else if (!path)
 638                         t = strdup(suffix);
 639                 else
 640                         t = strjoin(path, "/", suffix);
 641                 if (!t)
 642                         return -ENOMEM;
 643
 644                 *fs = path_kill_slashes(t);
 645                 return 0;
 646         }
 647
 648         if (!cg_controller_is_valid(controller))
 649                 return -EINVAL;
 650
 651         r = cg_all_unified();
 652         if (r < 0)
 653                 return r;
 654         if (r > 0)
 655                 r = join_path_unified(path, suffix, fs);
 656         else
 657                 r = join_path_legacy(controller, path, suffix, fs);
 658         if (r < 0)
 659                 return r;
 660
 661         path_kill_slashes(*fs);
 662         return 0;
 663 }
 664
 665 static int controller_is_accessible(const char *controller) {
 666         int r;
 667
 668         assert(controller);
 669
 670         /* Checks whether a specific controller is accessible,
 671          * i.e. its hierarchy mounted. In the unified hierarchy all
 672          * controllers are considered accessible, except for the named
 673          * hierarchies */
 674
 675         if (!cg_controller_is_valid(controller))
 676                 return -EINVAL;
 677
 678         r = cg_all_unified();
 679         if (r < 0)
 680                 return r;
 681         if (r > 0) {
 682                 /* We don't support named hierarchies if we are using
 683                  * the unified hierarchy. */
 684
 685                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
 686                         return 0;
 687
 688                 if (startswith(controller, "name="))
 689                         return -EOPNOTSUPP;
 690
 691         } else {
 692                 const char *cc, *dn;
 693
 694                 dn = controller_to_dirname(controller);
 695                 cc = strjoina("/sys/fs/cgroup/", dn);
 696
 697                 if (laccess(cc, F_OK) < 0)
 698                         return -errno;
 699         }
 700
 701         return 0;
 702 }
 703
 704 int cg_get_path_and_check(const char *controller, const char *path, const char *suffix, char **fs) {
 705         int r;
 706
 707         assert(controller);
 708         assert(fs);
 709
 710         /* Check if the specified controller is actually accessible */
 711         r = controller_is_accessible(controller);
 712         if (r < 0)
 713                 return r;
 714
 715         return cg_get_path(controller, path, suffix, fs);
 716 }
 717
 718 static int trim_cb(const char *path, const struct stat *sb, int typeflag, struct FTW *ftwbuf) {
 719         assert(path);
 720         assert(sb);
 721         assert(ftwbuf);
 722
 723         if (typeflag != FTW_DP)
 724                 return 0;
 725
 726         if (ftwbuf->level < 1)
 727                 return 0;
 728
 729         (void) rmdir(path);
 730         return 0;
 731 }
 732
 733 int cg_trim(const char *controller, const char *path, bool delete_root) {
 734         _cleanup_free_ char *fs = NULL;
 735         int r = 0, q;
 736
 737         assert(path);
 738
 739         r = cg_get_path(controller, path, NULL, &fs);
 740         if (r < 0)
 741                 return r;
 742
 743         errno = 0;
 744         if (nftw(fs, trim_cb, 64, FTW_DEPTH|FTW_MOUNT|FTW_PHYS) != 0) {
 745                 if (errno == ENOENT)
 746                         r = 0;
 747                 else if (errno > 0)
 748                         r = -errno;
 749                 else
 750                         r = -EIO;
 751         }
 752
 753         if (delete_root) {
 754                 if (rmdir(fs) < 0 && errno != ENOENT)
 755                         return -errno;
 756         }
 757
 758         q = cg_hybrid_unified();
 759         if (q < 0)
 760                 return q;
 761         if (q > 0 && streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 762                 q = cg_trim(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, delete_root);
 763                 if (q < 0)
 764                         log_warning_errno(q, "Failed to trim compat systemd cgroup %s: %m", path);
 765         }
 766
 767         return r;
 768 }
 769
 770 int cg_create(const char *controller, const char *path) {
 771         _cleanup_free_ char *fs = NULL;
 772         int r;
 773
 774         r = cg_get_path_and_check(controller, path, NULL, &fs);
 775         if (r < 0)
 776                 return r;
 777
 778         r = mkdir_parents(fs, 0755);
 779         if (r < 0)
 780                 return r;
 781
 782         r = mkdir_errno_wrapper(fs, 0755);
 783         if (r == -EEXIST)
 784                 return 0;
 785         if (r < 0)
 786                 return r;
 787
 788         r = cg_hybrid_unified();
 789         if (r < 0)
 790                 return r;
 791
 792         if (r > 0 && streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 793                 r = cg_create(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path);
 794                 if (r < 0)
 795                         log_warning_errno(r, "Failed to create compat systemd cgroup %s: %m", path);
 796         }
 797
 798         return 1;
 799 }
 800
 801 int cg_create_and_attach(const char *controller, const char *path, pid_t pid) {
 802         int r, q;
 803
 804         assert(pid >= 0);
 805
 806         r = cg_create(controller, path);
 807         if (r < 0)
 808                 return r;
 809
 810         q = cg_attach(controller, path, pid);
 811         if (q < 0)
 812                 return q;
 813
 814         /* This does not remove the cgroup on failure */
 815         return r;
 816 }
 817
 818 int cg_attach(const char *controller, const char *path, pid_t pid) {
 819         _cleanup_free_ char *fs = NULL;
 820         char c[DECIMAL_STR_MAX(pid_t) + 2];
 821         int r;
 822
 823         assert(path);
 824         assert(pid >= 0);
 825
 826         r = cg_get_path_and_check(controller, path, "cgroup.procs", &fs);
 827         if (r < 0)
 828                 return r;
 829
 830         if (pid == 0)
 831                 pid = getpid_cached();
 832
 833         xsprintf(c, PID_FMT "\n", pid);
 834
 835         r = write_string_file(fs, c, 0);
 836         if (r < 0)
 837                 return r;
 838
 839         r = cg_hybrid_unified();
 840         if (r < 0)
 841                 return r;
 842
 843         if (r > 0 && streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 844                 r = cg_attach(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, pid);
 845                 if (r < 0)
 846                         log_warning_errno(r, "Failed to attach "PID_FMT" to compat systemd cgroup %s: %m", pid, path);
 847         }
 848
 849         return 0;
 850 }
 851
 852 int cg_attach_fallback(const char *controller, const char *path, pid_t pid) {
 853         int r;
 854
 855         assert(controller);
 856         assert(path);
 857         assert(pid >= 0);
 858
 859         r = cg_attach(controller, path, pid);
 860         if (r < 0) {
 861                 char prefix[strlen(path) + 1];
 862
 863                 /* This didn't work? Then let's try all prefixes of
 864                  * the destination */
 865
 866                 PATH_FOREACH_PREFIX(prefix, path) {
 867                         int q;
 868
 869                         q = cg_attach(controller, prefix, pid);
 870                         if (q >= 0)
 871                                 return q;
 872                 }
 873         }
 874
 875         return r;
 876 }
 877
 878 int cg_set_access(
 879                 const char *controller,
 880                 const char *path,
 881                 uid_t uid,
 882                 gid_t gid) {
 883
 884         struct Attribute {
 885                 const char *name;
 886                 bool fatal;
 887         };
 888
 889         /* cgroupsv1, aka legacy/non-unified */
 890         static const struct Attribute legacy_attributes[] = {
 891                 { "cgroup.procs",           true  },
 892                 { "tasks",                  false },
 893                 { "cgroup.clone_children",  false },
 894                 {},
 895         };
 896
 897         /* cgroupsv2, aka unified */
 898         static const struct Attribute unified_attributes[] = {
 899                 { "cgroup.procs",           true  },
 900                 { "cgroup.subtree_control", true  },
 901                 { "cgroup.threads",         false },
 902                 {},
 903         };
 904
 905         static const struct Attribute* const attributes[] = {
 906                 [false] = legacy_attributes,
 907                 [true]  = unified_attributes,
 908         };
 909
 910         _cleanup_free_ char *fs = NULL;
 911         const struct Attribute *i;
 912         int r, unified;
 913
 914         assert(path);
 915
 916         if (uid == UID_INVALID && gid == GID_INVALID)
 917                 return 0;
 918
 919         unified = cg_unified_controller(controller);
 920         if (unified < 0)
 921                 return unified;
 922
 923         /* Configure access to the cgroup itself */
 924         r = cg_get_path(controller, path, NULL, &fs);
 925         if (r < 0)
 926                 return r;
 927
 928         r = chmod_and_chown(fs, 0755, uid, gid);
 929         if (r < 0)
 930                 return r;
 931
 932         /* Configure access to the cgroup's attributes */
 933         for (i = attributes[unified]; i->name; i++) {
 934                 fs = mfree(fs);
 935
 936                 r = cg_get_path(controller, path, i->name, &fs);
 937                 if (r < 0)
 938                         return r;
 939
 940                 r = chmod_and_chown(fs, 0644, uid, gid);
 941                 if (r < 0) {
 942                         if (i->fatal)
 943                                 return r;
 944
 945                         log_debug_errno(r, "Failed to set access on cgroup %s, ignoring: %m", fs);
 946                 }
 947         }
 948
 949         if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) {
 950                 r = cg_hybrid_unified();
 951                 if (r < 0)
 952                         return r;
 953                 if (r > 0) {
 954                         /* Always propagate access mode from unified to legacy controller */
 955                         r = cg_set_access(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, uid, gid);
 956                         if (r < 0)
 957                                 log_debug_errno(r, "Failed to set access on compatibility systemd cgroup %s, ignoring: %m", path);
 958                 }
 959         }
 960
 961         return 0;
 962 }
 963
 964 int cg_set_xattr(const char *controller, const char *path, const char *name, const void *value, size_t size, int flags) {
 965         _cleanup_free_ char *fs = NULL;
 966         int r;
 967
 968         assert(path);
 969         assert(name);
 970         assert(value || size <= 0);
 971
 972         r = cg_get_path(controller, path, NULL, &fs);
 973         if (r < 0)
 974                 return r;
 975
 976         if (setxattr(fs, name, value, size, flags) < 0)
 977                 return -errno;
 978
 979         return 0;
 980 }
 981
 982 int cg_get_xattr(const char *controller, const char *path, const char *name, void *value, size_t size) {
 983         _cleanup_free_ char *fs = NULL;
 984         ssize_t n;
 985         int r;
 986
 987         assert(path);
 988         assert(name);
 989
 990         r = cg_get_path(controller, path, NULL, &fs);
 991         if (r < 0)
 992                 return r;
 993
 994         n = getxattr(fs, name, value, size);
 995         if (n < 0)
 996                 return -errno;
 997
 998         return (int) n;
 999 }
1000
1001 int cg_pid_get_path(const char *controller, pid_t pid, char **path) {
1002         _cleanup_fclose_ FILE *f = NULL;
1003         char line[LINE_MAX];
1004         const char *fs, *controller_str;
1005         size_t cs = 0;
1006         int unified;
1007
1008         assert(path);
1009         assert(pid >= 0);
1010
1011         if (controller) {
1012                 if (!cg_controller_is_valid(controller))
1013                         return -EINVAL;
1014         } else
1015                 controller = SYSTEMD_CGROUP_CONTROLLER;
1016
1017         unified = cg_unified_controller(controller);
1018         if (unified < 0)
1019                 return unified;
1020         if (unified == 0) {
1021                 if (streq(controller, SYSTEMD_CGROUP_CONTROLLER))
1022                         controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY;
1023                 else
1024                         controller_str = controller;
1025
1026                 cs = strlen(controller_str);
1027         }
1028
1029         fs = procfs_file_alloca(pid, "cgroup");
1030         f = fopen(fs, "re");
1031         if (!f)
1032                 return errno == ENOENT ? -ESRCH : -errno;
1033
1034         (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
1035
1036         FOREACH_LINE(line, f, return -errno) {
1037                 char *e, *p;
1038
1039                 truncate_nl(line);
1040
1041                 if (unified) {
1042                         e = startswith(line, "0:");
1043                         if (!e)
1044                                 continue;
1045
1046                         e = strchr(e, ':');
1047                         if (!e)
1048                                 continue;
1049                 } else {
1050                         char *l;
1051                         size_t k;
1052                         const char *word, *state;
1053                         bool found = false;
1054
1055                         l = strchr(line, ':');
1056                         if (!l)
1057                                 continue;
1058
1059                         l++;
1060                         e = strchr(l, ':');
1061                         if (!e)
1062                                 continue;
1063
1064                         *e = 0;
1065                         FOREACH_WORD_SEPARATOR(word, k, l, ",", state)
1066                                 if (k == cs && memcmp(word, controller_str, cs) == 0) {
1067                                         found = true;
1068                                         break;
1069                                 }
1070                         if (!found)
1071                                 continue;
1072                 }
1073
1074                 p = strdup(e + 1);
1075                 if (!p)
1076                         return -ENOMEM;
1077
1078                 /* Truncate suffix indicating the process is a zombie */
1079                 e = endswith(p, " (deleted)");
1080                 if (e)
1081                         *e = 0;
1082
1083                 *path = p;
1084                 return 0;
1085         }
1086
1087         return -ENODATA;
1088 }
1089
1090 int cg_install_release_agent(const char *controller, const char *agent) {
1091         _cleanup_free_ char *fs = NULL, *contents = NULL;
1092         const char *sc;
1093         int r;
1094
1095         assert(agent);
1096
1097         r = cg_unified_controller(controller);
1098         if (r < 0)
1099                 return r;
1100         if (r > 0) /* doesn't apply to unified hierarchy */
1101                 return -EOPNOTSUPP;
1102
1103         r = cg_get_path(controller, NULL, "release_agent", &fs);
1104         if (r < 0)
1105                 return r;
1106
1107         r = read_one_line_file(fs, &contents);
1108         if (r < 0)
1109                 return r;
1110
1111         sc = strstrip(contents);
1112         if (isempty(sc)) {
1113                 r = write_string_file(fs, agent, 0);
1114                 if (r < 0)
1115                         return r;
1116         } else if (!path_equal(sc, agent))
1117                 return -EEXIST;
1118
1119         fs = mfree(fs);
1120         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
1121         if (r < 0)
1122                 return r;
1123
1124         contents = mfree(contents);
1125         r = read_one_line_file(fs, &contents);
1126         if (r < 0)
1127                 return r;
1128
1129         sc = strstrip(contents);
1130         if (streq(sc, "0")) {
1131                 r = write_string_file(fs, "1", 0);
1132                 if (r < 0)
1133                         return r;
1134
1135                 return 1;
1136         }
1137
1138         if (!streq(sc, "1"))
1139                 return -EIO;
1140
1141         return 0;
1142 }
1143
1144 int cg_uninstall_release_agent(const char *controller) {
1145         _cleanup_free_ char *fs = NULL;
1146         int r;
1147
1148         r = cg_unified_controller(controller);
1149         if (r < 0)
1150                 return r;
1151         if (r > 0) /* Doesn't apply to unified hierarchy */
1152                 return -EOPNOTSUPP;
1153
1154         r = cg_get_path(controller, NULL, "notify_on_release", &fs);
1155         if (r < 0)
1156                 return r;
1157
1158         r = write_string_file(fs, "0", 0);
1159         if (r < 0)
1160                 return r;
1161
1162         fs = mfree(fs);
1163
1164         r = cg_get_path(controller, NULL, "release_agent", &fs);
1165         if (r < 0)
1166                 return r;
1167
1168         r = write_string_file(fs, "", 0);
1169         if (r < 0)
1170                 return r;
1171
1172         return 0;
1173 }
1174
1175 int cg_is_empty(const char *controller, const char *path) {
1176         _cleanup_fclose_ FILE *f = NULL;
1177         pid_t pid;
1178         int r;
1179
1180         assert(path);
1181
1182         r = cg_enumerate_processes(controller, path, &f);
1183         if (r == -ENOENT)
1184                 return 1;
1185         if (r < 0)
1186                 return r;
1187
1188         r = cg_read_pid(f, &pid);
1189         if (r < 0)
1190                 return r;
1191
1192         return r == 0;
1193 }
1194
1195 int cg_is_empty_recursive(const char *controller, const char *path) {
1196         int r;
1197
1198         assert(path);
1199
1200         /* The root cgroup is always populated */
1201         if (controller && (isempty(path) || path_equal(path, "/")))
1202                 return false;
1203
1204         r = cg_unified_controller(controller);
1205         if (r < 0)
1206                 return r;
1207         if (r > 0) {
1208                 _cleanup_free_ char *t = NULL;
1209
1210                 /* On the unified hierarchy we can check empty state
1211                  * via the "populated" attribute of "cgroup.events". */
1212
1213                 r = cg_read_event(controller, path, "populated", &t);
1214                 if (r < 0)
1215                         return r;
1216
1217                 return streq(t, "0");
1218         } else {
1219                 _cleanup_closedir_ DIR *d = NULL;
1220                 char *fn;
1221
1222                 r = cg_is_empty(controller, path);
1223                 if (r <= 0)
1224                         return r;
1225
1226                 r = cg_enumerate_subgroups(controller, path, &d);
1227                 if (r == -ENOENT)
1228                         return 1;
1229                 if (r < 0)
1230                         return r;
1231
1232                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1233                         _cleanup_free_ char *p = NULL;
1234
1235                         p = strjoin(path, "/", fn);
1236                         free(fn);
1237                         if (!p)
1238                                 return -ENOMEM;
1239
1240                         r = cg_is_empty_recursive(controller, p);
1241                         if (r <= 0)
1242                                 return r;
1243                 }
1244                 if (r < 0)
1245                         return r;
1246
1247                 return true;
1248         }
1249 }
1250
1251 int cg_split_spec(const char *spec, char **controller, char **path) {
1252         char *t = NULL, *u = NULL;
1253         const char *e;
1254
1255         assert(spec);
1256
1257         if (*spec == '/') {
1258                 if (!path_is_normalized(spec))
1259                         return -EINVAL;
1260
1261                 if (path) {
1262                         t = strdup(spec);
1263                         if (!t)
1264                                 return -ENOMEM;
1265
1266                         *path = path_kill_slashes(t);
1267                 }
1268
1269                 if (controller)
1270                         *controller = NULL;
1271
1272                 return 0;
1273         }
1274
1275         e = strchr(spec, ':');
1276         if (!e) {
1277                 if (!cg_controller_is_valid(spec))
1278                         return -EINVAL;
1279
1280                 if (controller) {
1281                         t = strdup(spec);
1282                         if (!t)
1283                                 return -ENOMEM;
1284
1285                         *controller = t;
1286                 }
1287
1288                 if (path)
1289                         *path = NULL;
1290
1291                 return 0;
1292         }
1293
1294         t = strndup(spec, e-spec);
1295         if (!t)
1296                 return -ENOMEM;
1297         if (!cg_controller_is_valid(t)) {
1298                 free(t);
1299                 return -EINVAL;
1300         }
1301
1302         if (isempty(e+1))
1303                 u = NULL;
1304         else {
1305                 u = strdup(e+1);
1306                 if (!u) {
1307                         free(t);
1308                         return -ENOMEM;
1309                 }
1310
1311                 if (!path_is_normalized(u) ||
1312                     !path_is_absolute(u)) {
1313                         free(t);
1314                         free(u);
1315                         return -EINVAL;
1316                 }
1317
1318                 path_kill_slashes(u);
1319         }
1320
1321         if (controller)
1322                 *controller = t;
1323         else
1324                 free(t);
1325
1326         if (path)
1327                 *path = u;
1328         else
1329                 free(u);
1330
1331         return 0;
1332 }
1333
1334 int cg_mangle_path(const char *path, char **result) {
1335         _cleanup_free_ char *c = NULL, *p = NULL;
1336         char *t;
1337         int r;
1338
1339         assert(path);
1340         assert(result);
1341
1342         /* First, check if it already is a filesystem path */
1343         if (path_startswith(path, "/sys/fs/cgroup")) {
1344
1345                 t = strdup(path);
1346                 if (!t)
1347                         return -ENOMEM;
1348
1349                 *result = path_kill_slashes(t);
1350                 return 0;
1351         }
1352
1353         /* Otherwise, treat it as cg spec */
1354         r = cg_split_spec(path, &c, &p);
1355         if (r < 0)
1356                 return r;
1357
1358         return cg_get_path(c ?: SYSTEMD_CGROUP_CONTROLLER, p ?: "/", NULL, result);
1359 }
1360
1361 int cg_get_root_path(char **path) {
1362         char *p, *e;
1363         int r;
1364
1365         assert(path);
1366
1367         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 1, &p);
1368         if (r < 0)
1369                 return r;
1370
1371         e = endswith(p, "/" SPECIAL_INIT_SCOPE);
1372         if (!e)
1373                 e = endswith(p, "/" SPECIAL_SYSTEM_SLICE); /* legacy */
1374         if (!e)
1375                 e = endswith(p, "/system"); /* even more legacy */
1376         if (e)
1377                 *e = 0;
1378
1379         *path = p;
1380         return 0;
1381 }
1382
1383 int cg_shift_path(const char *cgroup, const char *root, const char **shifted) {
1384         _cleanup_free_ char *rt = NULL;
1385         char *p;
1386         int r;
1387
1388         assert(cgroup);
1389         assert(shifted);
1390
1391         if (!root) {
1392                 /* If the root was specified let's use that, otherwise
1393                  * let's determine it from PID 1 */
1394
1395                 r = cg_get_root_path(&rt);
1396                 if (r < 0)
1397                         return r;
1398
1399                 root = rt;
1400         }
1401
1402         p = path_startswith(cgroup, root);
1403         if (p && p > cgroup)
1404                 *shifted = p - 1;
1405         else
1406                 *shifted = cgroup;
1407
1408         return 0;
1409 }
1410
1411 int cg_pid_get_path_shifted(pid_t pid, const char *root, char **cgroup) {
1412         _cleanup_free_ char *raw = NULL;
1413         const char *c;
1414         int r;
1415
1416         assert(pid >= 0);
1417         assert(cgroup);
1418
1419         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &raw);
1420         if (r < 0)
1421                 return r;
1422
1423         r = cg_shift_path(raw, root, &c);
1424         if (r < 0)
1425                 return r;
1426
1427         if (c == raw)
1428                 *cgroup = TAKE_PTR(raw);
1429         else {
1430                 char *n;
1431
1432                 n = strdup(c);
1433                 if (!n)
1434                         return -ENOMEM;
1435
1436                 *cgroup = n;
1437         }
1438
1439         return 0;
1440 }
1441
1442 int cg_path_decode_unit(const char *cgroup, char **unit) {
1443         char *c, *s;
1444         size_t n;
1445
1446         assert(cgroup);
1447         assert(unit);
1448
1449         n = strcspn(cgroup, "/");
1450         if (n < 3)
1451                 return -ENXIO;
1452
1453         c = strndupa(cgroup, n);
1454         c = cg_unescape(c);
1455
1456         if (!unit_name_is_valid(c, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE))
1457                 return -ENXIO;
1458
1459         s = strdup(c);
1460         if (!s)
1461                 return -ENOMEM;
1462
1463         *unit = s;
1464         return 0;
1465 }
1466
1467 static bool valid_slice_name(const char *p, size_t n) {
1468
1469         if (!p)
1470                 return false;
1471
1472         if (n < STRLEN("x.slice"))
1473                 return false;
1474
1475         if (memcmp(p + n - 6, ".slice", 6) == 0) {
1476                 char buf[n+1], *c;
1477
1478                 memcpy(buf, p, n);
1479                 buf[n] = 0;
1480
1481                 c = cg_unescape(buf);
1482
1483                 return unit_name_is_valid(c, UNIT_NAME_PLAIN);
1484         }
1485
1486         return false;
1487 }
1488
1489 static const char *skip_slices(const char *p) {
1490         assert(p);
1491
1492         /* Skips over all slice assignments */
1493
1494         for (;;) {
1495                 size_t n;
1496
1497                 p += strspn(p, "/");
1498
1499                 n = strcspn(p, "/");
1500                 if (!valid_slice_name(p, n))
1501                         return p;
1502
1503                 p += n;
1504         }
1505 }
1506
1507 int cg_path_get_unit(const char *path, char **ret) {
1508         const char *e;
1509         char *unit;
1510         int r;
1511
1512         assert(path);
1513         assert(ret);
1514
1515         e = skip_slices(path);
1516
1517         r = cg_path_decode_unit(e, &unit);
1518         if (r < 0)
1519                 return r;
1520
1521         /* We skipped over the slices, don't accept any now */
1522         if (endswith(unit, ".slice")) {
1523                 free(unit);
1524                 return -ENXIO;
1525         }
1526
1527         *ret = unit;
1528         return 0;
1529 }
1530
1531 int cg_pid_get_unit(pid_t pid, char **unit) {
1532         _cleanup_free_ char *cgroup = NULL;
1533         int r;
1534
1535         assert(unit);
1536
1537         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1538         if (r < 0)
1539                 return r;
1540
1541         return cg_path_get_unit(cgroup, unit);
1542 }
1543
1544 /**
1545  * Skip session-*.scope, but require it to be there.
1546  */
1547 static const char *skip_session(const char *p) {
1548         size_t n;
1549
1550         if (isempty(p))
1551                 return NULL;
1552
1553         p += strspn(p, "/");
1554
1555         n = strcspn(p, "/");
1556         if (n < STRLEN("session-x.scope"))
1557                 return NULL;
1558
1559         if (memcmp(p, "session-", 8) == 0 && memcmp(p + n - 6, ".scope", 6) == 0) {
1560                 char buf[n - 8 - 6 + 1];
1561
1562                 memcpy(buf, p + 8, n - 8 - 6);
1563                 buf[n - 8 - 6] = 0;
1564
1565                 /* Note that session scopes never need unescaping,
1566                  * since they cannot conflict with the kernel's own
1567                  * names, hence we don't need to call cg_unescape()
1568                  * here. */
1569
1570                 if (!session_id_valid(buf))
1571                         return false;
1572
1573                 p += n;
1574                 p += strspn(p, "/");
1575                 return p;
1576         }
1577
1578         return NULL;
1579 }
1580
1581 /**
1582  * Skip user@*.service, but require it to be there.
1583  */
1584 static const char *skip_user_manager(const char *p) {
1585         size_t n;
1586
1587         if (isempty(p))
1588                 return NULL;
1589
1590         p += strspn(p, "/");
1591
1592         n = strcspn(p, "/");
1593         if (n < STRLEN("user@x.service"))
1594                 return NULL;
1595
1596         if (memcmp(p, "user@", 5) == 0 && memcmp(p + n - 8, ".service", 8) == 0) {
1597                 char buf[n - 5 - 8 + 1];
1598
1599                 memcpy(buf, p + 5, n - 5 - 8);
1600                 buf[n - 5 - 8] = 0;
1601
1602                 /* Note that user manager services never need unescaping,
1603                  * since they cannot conflict with the kernel's own
1604                  * names, hence we don't need to call cg_unescape()
1605                  * here. */
1606
1607                 if (parse_uid(buf, NULL) < 0)
1608                         return NULL;
1609
1610                 p += n;
1611                 p += strspn(p, "/");
1612
1613                 return p;
1614         }
1615
1616         return NULL;
1617 }
1618
1619 static const char *skip_user_prefix(const char *path) {
1620         const char *e, *t;
1621
1622         assert(path);
1623
1624         /* Skip slices, if there are any */
1625         e = skip_slices(path);
1626
1627         /* Skip the user manager, if it's in the path now... */
1628         t = skip_user_manager(e);
1629         if (t)
1630                 return t;
1631
1632         /* Alternatively skip the user session if it is in the path... */
1633         return skip_session(e);
1634 }
1635
1636 int cg_path_get_user_unit(const char *path, char **ret) {
1637         const char *t;
1638
1639         assert(path);
1640         assert(ret);
1641
1642         t = skip_user_prefix(path);
1643         if (!t)
1644                 return -ENXIO;
1645
1646         /* And from here on it looks pretty much the same as for a
1647          * system unit, hence let's use the same parser from here
1648          * on. */
1649         return cg_path_get_unit(t, ret);
1650 }
1651
1652 int cg_pid_get_user_unit(pid_t pid, char **unit) {
1653         _cleanup_free_ char *cgroup = NULL;
1654         int r;
1655
1656         assert(unit);
1657
1658         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1659         if (r < 0)
1660                 return r;
1661
1662         return cg_path_get_user_unit(cgroup, unit);
1663 }
1664
1665 int cg_path_get_machine_name(const char *path, char **machine) {
1666         _cleanup_free_ char *u = NULL;
1667         const char *sl;
1668         int r;
1669
1670         r = cg_path_get_unit(path, &u);
1671         if (r < 0)
1672                 return r;
1673
1674         sl = strjoina("/run/systemd/machines/unit:", u);
1675         return readlink_malloc(sl, machine);
1676 }
1677
1678 int cg_pid_get_machine_name(pid_t pid, char **machine) {
1679         _cleanup_free_ char *cgroup = NULL;
1680         int r;
1681
1682         assert(machine);
1683
1684         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1685         if (r < 0)
1686                 return r;
1687
1688         return cg_path_get_machine_name(cgroup, machine);
1689 }
1690
1691 int cg_path_get_session(const char *path, char **session) {
1692         _cleanup_free_ char *unit = NULL;
1693         char *start, *end;
1694         int r;
1695
1696         assert(path);
1697
1698         r = cg_path_get_unit(path, &unit);
1699         if (r < 0)
1700                 return r;
1701
1702         start = startswith(unit, "session-");
1703         if (!start)
1704                 return -ENXIO;
1705         end = endswith(start, ".scope");
1706         if (!end)
1707                 return -ENXIO;
1708
1709         *end = 0;
1710         if (!session_id_valid(start))
1711                 return -ENXIO;
1712
1713         if (session) {
1714                 char *rr;
1715
1716                 rr = strdup(start);
1717                 if (!rr)
1718                         return -ENOMEM;
1719
1720                 *session = rr;
1721         }
1722
1723         return 0;
1724 }
1725
1726 int cg_pid_get_session(pid_t pid, char **session) {
1727         _cleanup_free_ char *cgroup = NULL;
1728         int r;
1729
1730         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1731         if (r < 0)
1732                 return r;
1733
1734         return cg_path_get_session(cgroup, session);
1735 }
1736
1737 int cg_path_get_owner_uid(const char *path, uid_t *uid) {
1738         _cleanup_free_ char *slice = NULL;
1739         char *start, *end;
1740         int r;
1741
1742         assert(path);
1743
1744         r = cg_path_get_slice(path, &slice);
1745         if (r < 0)
1746                 return r;
1747
1748         start = startswith(slice, "user-");
1749         if (!start)
1750                 return -ENXIO;
1751         end = endswith(start, ".slice");
1752         if (!end)
1753                 return -ENXIO;
1754
1755         *end = 0;
1756         if (parse_uid(start, uid) < 0)
1757                 return -ENXIO;
1758
1759         return 0;
1760 }
1761
1762 int cg_pid_get_owner_uid(pid_t pid, uid_t *uid) {
1763         _cleanup_free_ char *cgroup = NULL;
1764         int r;
1765
1766         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1767         if (r < 0)
1768                 return r;
1769
1770         return cg_path_get_owner_uid(cgroup, uid);
1771 }
1772
1773 int cg_path_get_slice(const char *p, char **slice) {
1774         const char *e = NULL;
1775
1776         assert(p);
1777         assert(slice);
1778
1779         /* Finds the right-most slice unit from the beginning, but
1780          * stops before we come to the first non-slice unit. */
1781
1782         for (;;) {
1783                 size_t n;
1784
1785                 p += strspn(p, "/");
1786
1787                 n = strcspn(p, "/");
1788                 if (!valid_slice_name(p, n)) {
1789
1790                         if (!e) {
1791                                 char *s;
1792
1793                                 s = strdup(SPECIAL_ROOT_SLICE);
1794                                 if (!s)
1795                                         return -ENOMEM;
1796
1797                                 *slice = s;
1798                                 return 0;
1799                         }
1800
1801                         return cg_path_decode_unit(e, slice);
1802                 }
1803
1804                 e = p;
1805                 p += n;
1806         }
1807 }
1808
1809 int cg_pid_get_slice(pid_t pid, char **slice) {
1810         _cleanup_free_ char *cgroup = NULL;
1811         int r;
1812
1813         assert(slice);
1814
1815         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1816         if (r < 0)
1817                 return r;
1818
1819         return cg_path_get_slice(cgroup, slice);
1820 }
1821
1822 int cg_path_get_user_slice(const char *p, char **slice) {
1823         const char *t;
1824         assert(p);
1825         assert(slice);
1826
1827         t = skip_user_prefix(p);
1828         if (!t)
1829                 return -ENXIO;
1830
1831         /* And now it looks pretty much the same as for a system
1832          * slice, so let's just use the same parser from here on. */
1833         return cg_path_get_slice(t, slice);
1834 }
1835
1836 int cg_pid_get_user_slice(pid_t pid, char **slice) {
1837         _cleanup_free_ char *cgroup = NULL;
1838         int r;
1839
1840         assert(slice);
1841
1842         r = cg_pid_get_path_shifted(pid, NULL, &cgroup);
1843         if (r < 0)
1844                 return r;
1845
1846         return cg_path_get_user_slice(cgroup, slice);
1847 }
1848
1849 char *cg_escape(const char *p) {
1850         bool need_prefix = false;
1851
1852         /* This implements very minimal escaping for names to be used
1853          * as file names in the cgroup tree: any name which might
1854          * conflict with a kernel name or is prefixed with '_' is
1855          * prefixed with a '_'. That way, when reading cgroup names it
1856          * is sufficient to remove a single prefixing underscore if
1857          * there is one. */
1858
1859         /* The return value of this function (unlike cg_unescape())
1860          * needs free()! */
1861
1862         if (IN_SET(p[0], 0, '_', '.') ||
1863             streq(p, "notify_on_release") ||
1864             streq(p, "release_agent") ||
1865             streq(p, "tasks") ||
1866             startswith(p, "cgroup."))
1867                 need_prefix = true;
1868         else {
1869                 const char *dot;
1870
1871                 dot = strrchr(p, '.');
1872                 if (dot) {
1873                         CGroupController c;
1874                         size_t l = dot - p;
1875
1876                         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
1877                                 const char *n;
1878
1879                                 n = cgroup_controller_to_string(c);
1880
1881                                 if (l != strlen(n))
1882                                         continue;
1883
1884                                 if (memcmp(p, n, l) != 0)
1885                                         continue;
1886
1887                                 need_prefix = true;
1888                                 break;
1889                         }
1890                 }
1891         }
1892
1893         if (need_prefix)
1894                 return strappend("_", p);
1895
1896         return strdup(p);
1897 }
1898
1899 char *cg_unescape(const char *p) {
1900         assert(p);
1901
1902         /* The return value of this function (unlike cg_escape())
1903          * doesn't need free()! */
1904
1905         if (p[0] == '_')
1906                 return (char*) p+1;
1907
1908         return (char*) p;
1909 }
1910
1911 #define CONTROLLER_VALID                        \
1912         DIGITS LETTERS                          \
1913         "_"
1914
1915 bool cg_controller_is_valid(const char *p) {
1916         const char *t, *s;
1917
1918         if (!p)
1919                 return false;
1920
1921         if (streq(p, SYSTEMD_CGROUP_CONTROLLER))
1922                 return true;
1923
1924         s = startswith(p, "name=");
1925         if (s)
1926                 p = s;
1927
1928         if (IN_SET(*p, 0, '_'))
1929                 return false;
1930
1931         for (t = p; *t; t++)
1932                 if (!strchr(CONTROLLER_VALID, *t))
1933                         return false;
1934
1935         if (t - p > FILENAME_MAX)
1936                 return false;
1937
1938         return true;
1939 }
1940
1941 int cg_slice_to_path(const char *unit, char **ret) {
1942         _cleanup_free_ char *p = NULL, *s = NULL, *e = NULL;
1943         const char *dash;
1944         int r;
1945
1946         assert(unit);
1947         assert(ret);
1948
1949         if (streq(unit, SPECIAL_ROOT_SLICE)) {
1950                 char *x;
1951
1952                 x = strdup("");
1953                 if (!x)
1954                         return -ENOMEM;
1955                 *ret = x;
1956                 return 0;
1957         }
1958
1959         if (!unit_name_is_valid(unit, UNIT_NAME_PLAIN))
1960                 return -EINVAL;
1961
1962         if (!endswith(unit, ".slice"))
1963                 return -EINVAL;
1964
1965         r = unit_name_to_prefix(unit, &p);
1966         if (r < 0)
1967                 return r;
1968
1969         dash = strchr(p, '-');
1970
1971         /* Don't allow initial dashes */
1972         if (dash == p)
1973                 return -EINVAL;
1974
1975         while (dash) {
1976                 _cleanup_free_ char *escaped = NULL;
1977                 char n[dash - p + sizeof(".slice")];
1978
1979 #if HAS_FEATURE_MEMORY_SANITIZER
1980                 /* msan doesn't instrument stpncpy, so it thinks
1981                  * n is later used unitialized:
1982                  * https://github.com/google/sanitizers/issues/926
1983                  */
1984                 zero(n);
1985 #endif
1986
1987                 /* Don't allow trailing or double dashes */
1988                 if (IN_SET(dash[1], 0, '-'))
1989                         return -EINVAL;
1990
1991                 strcpy(stpncpy(n, p, dash - p), ".slice");
1992                 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN))
1993                         return -EINVAL;
1994
1995                 escaped = cg_escape(n);
1996                 if (!escaped)
1997                         return -ENOMEM;
1998
1999                 if (!strextend(&s, escaped, "/", NULL))
2000                         return -ENOMEM;
2001
2002                 dash = strchr(dash+1, '-');
2003         }
2004
2005         e = cg_escape(unit);
2006         if (!e)
2007                 return -ENOMEM;
2008
2009         if (!strextend(&s, e, NULL))
2010                 return -ENOMEM;
2011
2012         *ret = TAKE_PTR(s);
2013
2014         return 0;
2015 }
2016
2017 int cg_set_attribute(const char *controller, const char *path, const char *attribute, const char *value) {
2018         _cleanup_free_ char *p = NULL;
2019         int r;
2020
2021         r = cg_get_path(controller, path, attribute, &p);
2022         if (r < 0)
2023                 return r;
2024
2025         return write_string_file(p, value, 0);
2026 }
2027
2028 int cg_get_attribute(const char *controller, const char *path, const char *attribute, char **ret) {
2029         _cleanup_free_ char *p = NULL;
2030         int r;
2031
2032         r = cg_get_path(controller, path, attribute, &p);
2033         if (r < 0)
2034                 return r;
2035
2036         return read_one_line_file(p, ret);
2037 }
2038
2039 int cg_get_keyed_attribute(
2040                 const char *controller,
2041                 const char *path,
2042                 const char *attribute,
2043                 char **keys,
2044                 char **ret_values) {
2045
2046         _cleanup_free_ char *filename = NULL, *contents = NULL;
2047         const char *p;
2048         size_t n, i, n_done = 0;
2049         char **v;
2050         int r;
2051
2052         /* Reads one or more fields of a cgroupsv2 keyed attribute file. The 'keys' parameter should be an strv with
2053          * all keys to retrieve. The 'ret_values' parameter should be passed as string size with the same number of
2054          * entries as 'keys'. On success each entry will be set to the value of the matching key.
2055          *
2056          * If the attribute file doesn't exist at all returns ENOENT, if any key is not found returns ENXIO. */
2057
2058         r = cg_get_path(controller, path, attribute, &filename);
2059         if (r < 0)
2060                 return r;
2061
2062         r = read_full_file(filename, &contents, NULL);
2063         if (r < 0)
2064                 return r;
2065
2066         n = strv_length(keys);
2067         if (n == 0) /* No keys to retrieve? That's easy, we are done then */
2068                 return 0;
2069
2070         /* Let's build this up in a temporary array for now in order not to clobber the return parameter on failure */
2071         v = newa0(char*, n);
2072
2073         for (p = contents; *p;) {
2074                 const char *w = NULL;
2075
2076                 for (i = 0; i < n; i++)
2077                         if (!v[i]) {
2078                                 w = first_word(p, keys[i]);
2079                                 if (w)
2080                                         break;
2081                         }
2082
2083                 if (w) {
2084                         size_t l;
2085
2086                         l = strcspn(w, NEWLINE);
2087                         v[i] = strndup(w, l);
2088                         if (!v[i]) {
2089                                 r = -ENOMEM;
2090                                 goto fail;
2091                         }
2092
2093                         n_done++;
2094                         if (n_done >= n)
2095                                 goto done;
2096
2097                         p = w + l;
2098                 } else
2099                         p += strcspn(p, NEWLINE);
2100
2101                 p += strspn(p, NEWLINE);
2102         }
2103
2104         r = -ENXIO;
2105
2106 fail:
2107         for (i = 0; i < n; i++)
2108                 free(v[i]);
2109
2110         return r;
2111
2112 done:
2113         memcpy(ret_values, v, sizeof(char*) * n);
2114         return 0;
2115
2116 }
2117
2118 int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) {
2119         CGroupController c;
2120         int r;
2121
2122         /* This one will create a cgroup in our private tree, but also
2123          * duplicate it in the trees specified in mask, and remove it
2124          * in all others */
2125
2126         /* First create the cgroup in our own hierarchy. */
2127         r = cg_create(SYSTEMD_CGROUP_CONTROLLER, path);
2128         if (r < 0)
2129                 return r;
2130
2131         /* If we are in the unified hierarchy, we are done now */
2132         r = cg_all_unified();
2133         if (r < 0)
2134                 return r;
2135         if (r > 0)
2136                 return 0;
2137
2138         /* Otherwise, do the same in the other hierarchies */
2139         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2140                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2141                 const char *n;
2142
2143                 n = cgroup_controller_to_string(c);
2144
2145                 if (mask & bit)
2146                         (void) cg_create(n, path);
2147                 else if (supported & bit)
2148                         (void) cg_trim(n, path, true);
2149         }
2150
2151         return 0;
2152 }
2153
2154 int cg_attach_everywhere(CGroupMask supported, const char *path, pid_t pid, cg_migrate_callback_t path_callback, void *userdata) {
2155         CGroupController c;
2156         int r;
2157
2158         r = cg_attach(SYSTEMD_CGROUP_CONTROLLER, path, pid);
2159         if (r < 0)
2160                 return r;
2161
2162         r = cg_all_unified();
2163         if (r < 0)
2164                 return r;
2165         if (r > 0)
2166                 return 0;
2167
2168         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2169                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2170                 const char *p = NULL;
2171
2172                 if (!(supported & bit))
2173                         continue;
2174
2175                 if (path_callback)
2176                         p = path_callback(bit, userdata);
2177
2178                 if (!p)
2179                         p = path;
2180
2181                 (void) cg_attach_fallback(cgroup_controller_to_string(c), p, pid);
2182         }
2183
2184         return 0;
2185 }
2186
2187 int cg_attach_many_everywhere(CGroupMask supported, const char *path, Set* pids, cg_migrate_callback_t path_callback, void *userdata) {
2188         Iterator i;
2189         void *pidp;
2190         int r = 0;
2191
2192         SET_FOREACH(pidp, pids, i) {
2193                 pid_t pid = PTR_TO_PID(pidp);
2194                 int q;
2195
2196                 q = cg_attach_everywhere(supported, path, pid, path_callback, userdata);
2197                 if (q < 0 && r >= 0)
2198                         r = q;
2199         }
2200
2201         return r;
2202 }
2203
2204 int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to, cg_migrate_callback_t to_callback, void *userdata) {
2205         CGroupController c;
2206         int r = 0, q;
2207
2208         if (!path_equal(from, to))  {
2209                 r = cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER, from, SYSTEMD_CGROUP_CONTROLLER, to, CGROUP_REMOVE);
2210                 if (r < 0)
2211                         return r;
2212         }
2213
2214         q = cg_all_unified();
2215         if (q < 0)
2216                 return q;
2217         if (q > 0)
2218                 return r;
2219
2220         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2221                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2222                 const char *p = NULL;
2223
2224                 if (!(supported & bit))
2225                         continue;
2226
2227                 if (to_callback)
2228                         p = to_callback(bit, userdata);
2229
2230                 if (!p)
2231                         p = to;
2232
2233                 (void) cg_migrate_recursive_fallback(SYSTEMD_CGROUP_CONTROLLER, to, cgroup_controller_to_string(c), p, 0);
2234         }
2235
2236         return 0;
2237 }
2238
2239 int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) {
2240         CGroupController c;
2241         int r, q;
2242
2243         r = cg_trim(SYSTEMD_CGROUP_CONTROLLER, path, delete_root);
2244         if (r < 0)
2245                 return r;
2246
2247         q = cg_all_unified();
2248         if (q < 0)
2249                 return q;
2250         if (q > 0)
2251                 return r;
2252
2253         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2254                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2255
2256                 if (!(supported & bit))
2257                         continue;
2258
2259                 (void) cg_trim(cgroup_controller_to_string(c), path, delete_root);
2260         }
2261
2262         return 0;
2263 }
2264
2265 int cg_mask_to_string(CGroupMask mask, char **ret) {
2266         _cleanup_free_ char *s = NULL;
2267         size_t n = 0, allocated = 0;
2268         bool space = false;
2269         CGroupController c;
2270
2271         assert(ret);
2272
2273         if (mask == 0) {
2274                 *ret = NULL;
2275                 return 0;
2276         }
2277
2278         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2279                 const char *k;
2280                 size_t l;
2281
2282                 if (!(mask & CGROUP_CONTROLLER_TO_MASK(c)))
2283                         continue;
2284
2285                 k = cgroup_controller_to_string(c);
2286                 l = strlen(k);
2287
2288                 if (!GREEDY_REALLOC(s, allocated, n + space + l + 1))
2289                         return -ENOMEM;
2290
2291                 if (space)
2292                         s[n] = ' ';
2293                 memcpy(s + n + space, k, l);
2294                 n += space + l;
2295
2296                 space = true;
2297         }
2298
2299         assert(s);
2300
2301         s[n] = 0;
2302         *ret = TAKE_PTR(s);
2303
2304         return 0;
2305 }
2306
2307 int cg_mask_from_string(const char *value, CGroupMask *mask) {
2308         assert(mask);
2309         assert(value);
2310
2311         for (;;) {
2312                 _cleanup_free_ char *n = NULL;
2313                 CGroupController v;
2314                 int r;
2315
2316                 r = extract_first_word(&value, &n, NULL, 0);
2317                 if (r < 0)
2318                         return r;
2319                 if (r == 0)
2320                         break;
2321
2322                 v = cgroup_controller_from_string(n);
2323                 if (v < 0)
2324                         continue;
2325
2326                 *mask |= CGROUP_CONTROLLER_TO_MASK(v);
2327         }
2328         return 0;
2329 }
2330
2331 int cg_mask_supported(CGroupMask *ret) {
2332         CGroupMask mask = 0;
2333         int r;
2334
2335         /* Determines the mask of supported cgroup controllers. Only
2336          * includes controllers we can make sense of and that are
2337          * actually accessible. */
2338
2339         r = cg_all_unified();
2340         if (r < 0)
2341                 return r;
2342         if (r > 0) {
2343                 _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL;
2344
2345                 /* In the unified hierarchy we can read the supported
2346                  * and accessible controllers from a the top-level
2347                  * cgroup attribute */
2348
2349                 r = cg_get_root_path(&root);
2350                 if (r < 0)
2351                         return r;
2352
2353                 r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, root, "cgroup.controllers", &path);
2354                 if (r < 0)
2355                         return r;
2356
2357                 r = read_one_line_file(path, &controllers);
2358                 if (r < 0)
2359                         return r;
2360
2361                 r = cg_mask_from_string(controllers, &mask);
2362                 if (r < 0)
2363                         return r;
2364
2365                 /* Currently, we support the cpu, memory, io and pids
2366                  * controller in the unified hierarchy, mask
2367                  * everything else off. */
2368                 mask &= CGROUP_MASK_CPU | CGROUP_MASK_MEMORY | CGROUP_MASK_IO | CGROUP_MASK_PIDS;
2369
2370         } else {
2371                 CGroupController c;
2372
2373                 /* In the legacy hierarchy, we check whether which
2374                  * hierarchies are mounted. */
2375
2376                 for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2377                         const char *n;
2378
2379                         n = cgroup_controller_to_string(c);
2380                         if (controller_is_accessible(n) >= 0)
2381                                 mask |= CGROUP_CONTROLLER_TO_MASK(c);
2382                 }
2383         }
2384
2385         *ret = mask;
2386         return 0;
2387 }
2388
2389 int cg_kernel_controllers(Set **ret) {
2390         _cleanup_set_free_free_ Set *controllers = NULL;
2391         _cleanup_fclose_ FILE *f = NULL;
2392         int r;
2393
2394         assert(ret);
2395
2396         /* Determines the full list of kernel-known controllers. Might
2397          * include controllers we don't actually support, arbitrary
2398          * named hierarchies and controllers that aren't currently
2399          * accessible (because not mounted). */
2400
2401         controllers = set_new(&string_hash_ops);
2402         if (!controllers)
2403                 return -ENOMEM;
2404
2405         f = fopen("/proc/cgroups", "re");
2406         if (!f) {
2407                 if (errno == ENOENT) {
2408                         *ret = NULL;
2409                         return 0;
2410                 }
2411
2412                 return -errno;
2413         }
2414
2415         (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
2416
2417         /* Ignore the header line */
2418         (void) read_line(f, (size_t) -1, NULL);
2419
2420         for (;;) {
2421                 char *controller;
2422                 int enabled = 0;
2423
2424                 errno = 0;
2425                 if (fscanf(f, "%ms %*i %*i %i", &controller, &enabled) != 2) {
2426
2427                         if (feof(f))
2428                                 break;
2429
2430                         if (ferror(f) && errno > 0)
2431                                 return -errno;
2432
2433                         return -EBADMSG;
2434                 }
2435
2436                 if (!enabled) {
2437                         free(controller);
2438                         continue;
2439                 }
2440
2441                 if (!cg_controller_is_valid(controller)) {
2442                         free(controller);
2443                         return -EBADMSG;
2444                 }
2445
2446                 r = set_consume(controllers, controller);
2447                 if (r < 0)
2448                         return r;
2449         }
2450
2451         *ret = TAKE_PTR(controllers);
2452
2453         return 0;
2454 }
2455
2456 static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN;
2457
2458 /* The hybrid mode was initially implemented in v232 and simply mounted cgroup v2 on /sys/fs/cgroup/systemd.  This
2459  * unfortunately broke other tools (such as docker) which expected the v1 "name=systemd" hierarchy on
2460  * /sys/fs/cgroup/systemd.  From v233 and on, the hybrid mode mountnbs v2 on /sys/fs/cgroup/unified and maintains
2461  * "name=systemd" hierarchy on /sys/fs/cgroup/systemd for compatibility with other tools.
2462  *
2463  * To keep live upgrade working, we detect and support v232 layout.  When v232 layout is detected, to keep cgroup v2
2464  * process management but disable the compat dual layout, we return %true on
2465  * cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) and %false on cg_hybrid_unified().
2466  */
2467 static thread_local bool unified_systemd_v232;
2468
2469 static int cg_unified_update(void) {
2470
2471         struct statfs fs;
2472
2473         /* Checks if we support the unified hierarchy. Returns an
2474          * error when the cgroup hierarchies aren't mounted yet or we
2475          * have any other trouble determining if the unified hierarchy
2476          * is supported. */
2477
2478         if (unified_cache >= CGROUP_UNIFIED_NONE)
2479                 return 0;
2480
2481         if (statfs("/sys/fs/cgroup/", &fs) < 0)
2482                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/\") failed: %m");
2483
2484         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2485                 log_debug("Found cgroup2 on /sys/fs/cgroup/, full unified hierarchy");
2486                 unified_cache = CGROUP_UNIFIED_ALL;
2487         } else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) {
2488                 if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 &&
2489                     F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2490                         log_debug("Found cgroup2 on /sys/fs/cgroup/unified, unified hierarchy for systemd controller");
2491                         unified_cache = CGROUP_UNIFIED_SYSTEMD;
2492                         unified_systemd_v232 = false;
2493                 } else {
2494                         if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0)
2495                                 return log_debug_errno(errno, "statfs(\"/sys/fs/cgroup/systemd\" failed: %m");
2496
2497                         if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) {
2498                                 log_debug("Found cgroup2 on /sys/fs/cgroup/systemd, unified hierarchy for systemd controller (v232 variant)");
2499                                 unified_cache = CGROUP_UNIFIED_SYSTEMD;
2500                                 unified_systemd_v232 = true;
2501                         } else if (F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) {
2502                                 log_debug("Found cgroup on /sys/fs/cgroup/systemd, legacy hierarchy");
2503                                 unified_cache = CGROUP_UNIFIED_NONE;
2504                         } else {
2505                                 log_debug("Unexpected filesystem type %llx mounted on /sys/fs/cgroup/systemd, assuming legacy hierarchy",
2506                                           (unsigned long long) fs.f_type);
2507                                 unified_cache = CGROUP_UNIFIED_NONE;
2508                         }
2509                 }
2510         } else {
2511                 log_debug("Unknown filesystem type %llx mounted on /sys/fs/cgroup.",
2512                           (unsigned long long) fs.f_type);
2513                 return -ENOMEDIUM;
2514         }
2515
2516         return 0;
2517 }
2518
2519 int cg_unified_controller(const char *controller) {
2520         int r;
2521
2522         r = cg_unified_update();
2523         if (r < 0)
2524                 return r;
2525
2526         if (unified_cache == CGROUP_UNIFIED_NONE)
2527                 return false;
2528
2529         if (unified_cache >= CGROUP_UNIFIED_ALL)
2530                 return true;
2531
2532         return streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER);
2533 }
2534
2535 int cg_all_unified(void) {
2536         int r;
2537
2538         r = cg_unified_update();
2539         if (r < 0)
2540                 return r;
2541
2542         return unified_cache >= CGROUP_UNIFIED_ALL;
2543 }
2544
2545 int cg_hybrid_unified(void) {
2546         int r;
2547
2548         r = cg_unified_update();
2549         if (r < 0)
2550                 return r;
2551
2552         return unified_cache == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232;
2553 }
2554
2555 int cg_unified_flush(void) {
2556         unified_cache = CGROUP_UNIFIED_UNKNOWN;
2557
2558         return cg_unified_update();
2559 }
2560
2561 int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) {
2562         _cleanup_fclose_ FILE *f = NULL;
2563         _cleanup_free_ char *fs = NULL;
2564         CGroupController c;
2565         int r;
2566
2567         assert(p);
2568
2569         if (supported == 0)
2570                 return 0;
2571
2572         r = cg_all_unified();
2573         if (r < 0)
2574                 return r;
2575         if (r == 0) /* on the legacy hiearchy there's no joining of controllers defined */
2576                 return 0;
2577
2578         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, p, "cgroup.subtree_control", &fs);
2579         if (r < 0)
2580                 return r;
2581
2582         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) {
2583                 CGroupMask bit = CGROUP_CONTROLLER_TO_MASK(c);
2584                 const char *n;
2585
2586                 if (!(supported & bit))
2587                         continue;
2588
2589                 n = cgroup_controller_to_string(c);
2590                 {
2591                         char s[1 + strlen(n) + 1];
2592
2593                         s[0] = mask & bit ? '+' : '-';
2594                         strcpy(s + 1, n);
2595
2596                         if (!f) {
2597                                 f = fopen(fs, "we");
2598                                 if (!f) {
2599                                         log_debug_errno(errno, "Failed to open cgroup.subtree_control file of %s: %m", p);
2600                                         break;
2601                                 }
2602                         }
2603
2604                         r = write_string_stream(f, s, 0);
2605                         if (r < 0)
2606                                 log_debug_errno(r, "Failed to enable controller %s for %s (%s): %m", n, p, fs);
2607                 }
2608         }
2609
2610         return 0;
2611 }
2612
2613 bool cg_is_unified_wanted(void) {
2614         static thread_local int wanted = -1;
2615         int r;
2616         bool b;
2617         const bool is_default = DEFAULT_HIERARCHY == CGROUP_UNIFIED_ALL;
2618
2619         /* If we have a cached value, return that. */
2620         if (wanted >= 0)
2621                 return wanted;
2622
2623         /* If the hierarchy is already mounted, then follow whatever
2624          * was chosen for it. */
2625         if (cg_unified_flush() >= 0)
2626                 return (wanted = unified_cache >= CGROUP_UNIFIED_ALL);
2627
2628         /* Otherwise, let's see what the kernel command line has to say.
2629          * Since checking is expensive, cache a non-error result. */
2630         r = proc_cmdline_get_bool("systemd.unified_cgroup_hierarchy", &b);
2631
2632         return (wanted = r > 0 ? b : is_default);
2633 }
2634
2635 bool cg_is_legacy_wanted(void) {
2636         static thread_local int wanted = -1;
2637
2638         /* If we have a cached value, return that. */
2639         if (wanted >= 0)
2640                 return wanted;
2641
2642         /* Check if we have cgroups2 already mounted. */
2643         if (cg_unified_flush() >= 0 &&
2644             unified_cache == CGROUP_UNIFIED_ALL)
2645                 return (wanted = false);
2646
2647         /* Otherwise, assume that at least partial legacy is wanted,
2648          * since cgroups2 should already be mounted at this point. */
2649         return (wanted = true);
2650 }
2651
2652 bool cg_is_hybrid_wanted(void) {
2653         static thread_local int wanted = -1;
2654         int r;
2655         bool b;
2656         const bool is_default = DEFAULT_HIERARCHY >= CGROUP_UNIFIED_SYSTEMD;
2657         /* We default to true if the default is "hybrid", obviously,
2658          * but also when the default is "unified", because if we get
2659          * called, it means that unified hierarchy was not mounted. */
2660
2661         /* If we have a cached value, return that. */
2662         if (wanted >= 0)
2663                 return wanted;
2664
2665         /* If the hierarchy is already mounted, then follow whatever
2666          * was chosen for it. */
2667         if (cg_unified_flush() >= 0 &&
2668             unified_cache == CGROUP_UNIFIED_ALL)
2669                 return (wanted = false);
2670
2671         /* Otherwise, let's see what the kernel command line has to say.
2672          * Since checking is expensive, cache a non-error result. */
2673         r = proc_cmdline_get_bool("systemd.legacy_systemd_cgroup_controller", &b);
2674
2675         /* The meaning of the kernel option is reversed wrt. to the return value
2676          * of this function, hence the negation. */
2677         return (wanted = r > 0 ? !b : is_default);
2678 }
2679
2680 int cg_weight_parse(const char *s, uint64_t *ret) {
2681         uint64_t u;
2682         int r;
2683
2684         if (isempty(s)) {
2685                 *ret = CGROUP_WEIGHT_INVALID;
2686                 return 0;
2687         }
2688
2689         r = safe_atou64(s, &u);
2690         if (r < 0)
2691                 return r;
2692
2693         if (u < CGROUP_WEIGHT_MIN || u > CGROUP_WEIGHT_MAX)
2694                 return -ERANGE;
2695
2696         *ret = u;
2697         return 0;
2698 }
2699
2700 const uint64_t cgroup_io_limit_defaults[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2701         [CGROUP_IO_RBPS_MAX]    = CGROUP_LIMIT_MAX,
2702         [CGROUP_IO_WBPS_MAX]    = CGROUP_LIMIT_MAX,
2703         [CGROUP_IO_RIOPS_MAX]   = CGROUP_LIMIT_MAX,
2704         [CGROUP_IO_WIOPS_MAX]   = CGROUP_LIMIT_MAX,
2705 };
2706
2707 static const char* const cgroup_io_limit_type_table[_CGROUP_IO_LIMIT_TYPE_MAX] = {
2708         [CGROUP_IO_RBPS_MAX]    = "IOReadBandwidthMax",
2709         [CGROUP_IO_WBPS_MAX]    = "IOWriteBandwidthMax",
2710         [CGROUP_IO_RIOPS_MAX]   = "IOReadIOPSMax",
2711         [CGROUP_IO_WIOPS_MAX]   = "IOWriteIOPSMax",
2712 };
2713
2714 DEFINE_STRING_TABLE_LOOKUP(cgroup_io_limit_type, CGroupIOLimitType);
2715
2716 int cg_cpu_shares_parse(const char *s, uint64_t *ret) {
2717         uint64_t u;
2718         int r;
2719
2720         if (isempty(s)) {
2721                 *ret = CGROUP_CPU_SHARES_INVALID;
2722                 return 0;
2723         }
2724
2725         r = safe_atou64(s, &u);
2726         if (r < 0)
2727                 return r;
2728
2729         if (u < CGROUP_CPU_SHARES_MIN || u > CGROUP_CPU_SHARES_MAX)
2730                 return -ERANGE;
2731
2732         *ret = u;
2733         return 0;
2734 }
2735
2736 int cg_blkio_weight_parse(const char *s, uint64_t *ret) {
2737         uint64_t u;
2738         int r;
2739
2740         if (isempty(s)) {
2741                 *ret = CGROUP_BLKIO_WEIGHT_INVALID;
2742                 return 0;
2743         }
2744
2745         r = safe_atou64(s, &u);
2746         if (r < 0)
2747                 return r;
2748
2749         if (u < CGROUP_BLKIO_WEIGHT_MIN || u > CGROUP_BLKIO_WEIGHT_MAX)
2750                 return -ERANGE;
2751
2752         *ret = u;
2753         return 0;
2754 }
2755
2756 bool is_cgroup_fs(const struct statfs *s) {
2757         return is_fs_type(s, CGROUP_SUPER_MAGIC) ||
2758                is_fs_type(s, CGROUP2_SUPER_MAGIC);
2759 }
2760
2761 bool fd_is_cgroup_fs(int fd) {
2762         struct statfs s;
2763
2764         if (fstatfs(fd, &s) < 0)
2765                 return -errno;
2766
2767         return is_cgroup_fs(&s);
2768 }
2769
2770 static const char *cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
2771         [CGROUP_CONTROLLER_CPU] = "cpu",
2772         [CGROUP_CONTROLLER_CPUACCT] = "cpuacct",
2773         [CGROUP_CONTROLLER_IO] = "io",
2774         [CGROUP_CONTROLLER_BLKIO] = "blkio",
2775         [CGROUP_CONTROLLER_MEMORY] = "memory",
2776         [CGROUP_CONTROLLER_DEVICES] = "devices",
2777         [CGROUP_CONTROLLER_PIDS] = "pids",
2778 };
2779
2780 DEFINE_STRING_TABLE_LOOKUP(cgroup_controller, CGroupController);