src/core/cgroup.c

   1 /***
   2   This file is part of systemd.
   3
   4   Copyright 2013 Lennart Poettering
   5
   6   systemd is free software; you can redistribute it and/or modify it
   7   under the terms of the GNU Lesser General Public License as published by
   8   the Free Software Foundation; either version 2.1 of the License, or
   9   (at your option) any later version.
  10
  11   systemd is distributed in the hope that it will be useful, but
  12   WITHOUT ANY WARRANTY; without even the implied warranty of
  13   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14   Lesser General Public License for more details.
  15
  16   You should have received a copy of the GNU Lesser General Public License
  17   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  18 ***/
  19
  20 #include <fcntl.h>
  21 #include <fnmatch.h>
  22
  23 #include "alloc-util.h"
  24 #include "cgroup-util.h"
  25 #include "cgroup.h"
  26 #include "fd-util.h"
  27 #include "fileio.h"
  28 #include "fs-util.h"
  29 #include "parse-util.h"
  30 #include "path-util.h"
  31 #include "process-util.h"
  32 //#include "special.h"
  33 #include "string-table.h"
  34 #include "string-util.h"
  35 #include "stdio-util.h"
  36
  37 #define CGROUP_CPU_QUOTA_PERIOD_USEC ((usec_t) 100 * USEC_PER_MSEC)
  38
  39 #if 0 /// UNNEEDED by elogind
  40 static void cgroup_compat_warn(void) {
  41         static bool cgroup_compat_warned = false;
  42
  43         if (cgroup_compat_warned)
  44                 return;
  45
  46         log_warning("cgroup compatibility translation between legacy and unified hierarchy settings activated. See cgroup-compat debug messages for details.");
  47         cgroup_compat_warned = true;
  48 }
  49
  50 #define log_cgroup_compat(unit, fmt, ...) do {                                  \
  51                 cgroup_compat_warn();                                           \
  52                 log_unit_debug(unit, "cgroup-compat: " fmt, ##__VA_ARGS__);     \
  53         } while (false)
  54
  55 void cgroup_context_init(CGroupContext *c) {
  56         assert(c);
  57
  58         /* Initialize everything to the kernel defaults, assuming the
  59          * structure is preinitialized to 0 */
  60
  61         c->cpu_weight = CGROUP_WEIGHT_INVALID;
  62         c->startup_cpu_weight = CGROUP_WEIGHT_INVALID;
  63         c->cpu_quota_per_sec_usec = USEC_INFINITY;
  64
  65         c->cpu_shares = CGROUP_CPU_SHARES_INVALID;
  66         c->startup_cpu_shares = CGROUP_CPU_SHARES_INVALID;
  67
  68         c->memory_high = CGROUP_LIMIT_MAX;
  69         c->memory_max = CGROUP_LIMIT_MAX;
  70         c->memory_swap_max = CGROUP_LIMIT_MAX;
  71
  72         c->memory_limit = CGROUP_LIMIT_MAX;
  73
  74         c->io_weight = CGROUP_WEIGHT_INVALID;
  75         c->startup_io_weight = CGROUP_WEIGHT_INVALID;
  76
  77         c->blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID;
  78         c->startup_blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID;
  79
  80         c->tasks_max = (uint64_t) -1;
  81 }
  82
  83 void cgroup_context_free_device_allow(CGroupContext *c, CGroupDeviceAllow *a) {
  84         assert(c);
  85         assert(a);
  86
  87         LIST_REMOVE(device_allow, c->device_allow, a);
  88         free(a->path);
  89         free(a);
  90 }
  91
  92 void cgroup_context_free_io_device_weight(CGroupContext *c, CGroupIODeviceWeight *w) {
  93         assert(c);
  94         assert(w);
  95
  96         LIST_REMOVE(device_weights, c->io_device_weights, w);
  97         free(w->path);
  98         free(w);
  99 }
 100
 101 void cgroup_context_free_io_device_limit(CGroupContext *c, CGroupIODeviceLimit *l) {
 102         assert(c);
 103         assert(l);
 104
 105         LIST_REMOVE(device_limits, c->io_device_limits, l);
 106         free(l->path);
 107         free(l);
 108 }
 109
 110 void cgroup_context_free_blockio_device_weight(CGroupContext *c, CGroupBlockIODeviceWeight *w) {
 111         assert(c);
 112         assert(w);
 113
 114         LIST_REMOVE(device_weights, c->blockio_device_weights, w);
 115         free(w->path);
 116         free(w);
 117 }
 118
 119 void cgroup_context_free_blockio_device_bandwidth(CGroupContext *c, CGroupBlockIODeviceBandwidth *b) {
 120         assert(c);
 121         assert(b);
 122
 123         LIST_REMOVE(device_bandwidths, c->blockio_device_bandwidths, b);
 124         free(b->path);
 125         free(b);
 126 }
 127
 128 void cgroup_context_done(CGroupContext *c) {
 129         assert(c);
 130
 131         while (c->io_device_weights)
 132                 cgroup_context_free_io_device_weight(c, c->io_device_weights);
 133
 134         while (c->io_device_limits)
 135                 cgroup_context_free_io_device_limit(c, c->io_device_limits);
 136
 137         while (c->blockio_device_weights)
 138                 cgroup_context_free_blockio_device_weight(c, c->blockio_device_weights);
 139
 140         while (c->blockio_device_bandwidths)
 141                 cgroup_context_free_blockio_device_bandwidth(c, c->blockio_device_bandwidths);
 142
 143         while (c->device_allow)
 144                 cgroup_context_free_device_allow(c, c->device_allow);
 145 }
 146
 147 void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
 148         CGroupIODeviceLimit *il;
 149         CGroupIODeviceWeight *iw;
 150         CGroupBlockIODeviceBandwidth *b;
 151         CGroupBlockIODeviceWeight *w;
 152         CGroupDeviceAllow *a;
 153         char u[FORMAT_TIMESPAN_MAX];
 154
 155         assert(c);
 156         assert(f);
 157
 158         prefix = strempty(prefix);
 159
 160         fprintf(f,
 161                 "%sCPUAccounting=%s\n"
 162                 "%sIOAccounting=%s\n"
 163                 "%sBlockIOAccounting=%s\n"
 164                 "%sMemoryAccounting=%s\n"
 165                 "%sTasksAccounting=%s\n"
 166                 "%sCPUWeight=%" PRIu64 "\n"
 167                 "%sStartupCPUWeight=%" PRIu64 "\n"
 168                 "%sCPUShares=%" PRIu64 "\n"
 169                 "%sStartupCPUShares=%" PRIu64 "\n"
 170                 "%sCPUQuotaPerSecSec=%s\n"
 171                 "%sIOWeight=%" PRIu64 "\n"
 172                 "%sStartupIOWeight=%" PRIu64 "\n"
 173                 "%sBlockIOWeight=%" PRIu64 "\n"
 174                 "%sStartupBlockIOWeight=%" PRIu64 "\n"
 175                 "%sMemoryLow=%" PRIu64 "\n"
 176                 "%sMemoryHigh=%" PRIu64 "\n"
 177                 "%sMemoryMax=%" PRIu64 "\n"
 178                 "%sMemorySwapMax=%" PRIu64 "\n"
 179                 "%sMemoryLimit=%" PRIu64 "\n"
 180                 "%sTasksMax=%" PRIu64 "\n"
 181                 "%sDevicePolicy=%s\n"
 182                 "%sDelegate=%s\n",
 183                 prefix, yes_no(c->cpu_accounting),
 184                 prefix, yes_no(c->io_accounting),
 185                 prefix, yes_no(c->blockio_accounting),
 186                 prefix, yes_no(c->memory_accounting),
 187                 prefix, yes_no(c->tasks_accounting),
 188                 prefix, c->cpu_weight,
 189                 prefix, c->startup_cpu_weight,
 190                 prefix, c->cpu_shares,
 191                 prefix, c->startup_cpu_shares,
 192                 prefix, format_timespan(u, sizeof(u), c->cpu_quota_per_sec_usec, 1),
 193                 prefix, c->io_weight,
 194                 prefix, c->startup_io_weight,
 195                 prefix, c->blockio_weight,
 196                 prefix, c->startup_blockio_weight,
 197                 prefix, c->memory_low,
 198                 prefix, c->memory_high,
 199                 prefix, c->memory_max,
 200                 prefix, c->memory_swap_max,
 201                 prefix, c->memory_limit,
 202                 prefix, c->tasks_max,
 203                 prefix, cgroup_device_policy_to_string(c->device_policy),
 204                 prefix, yes_no(c->delegate));
 205
 206         LIST_FOREACH(device_allow, a, c->device_allow)
 207                 fprintf(f,
 208                         "%sDeviceAllow=%s %s%s%s\n",
 209                         prefix,
 210                         a->path,
 211                         a->r ? "r" : "", a->w ? "w" : "", a->m ? "m" : "");
 212
 213         LIST_FOREACH(device_weights, iw, c->io_device_weights)
 214                 fprintf(f,
 215                         "%sIODeviceWeight=%s %" PRIu64,
 216                         prefix,
 217                         iw->path,
 218                         iw->weight);
 219
 220         LIST_FOREACH(device_limits, il, c->io_device_limits) {
 221                 char buf[FORMAT_BYTES_MAX];
 222                 CGroupIOLimitType type;
 223
 224                 for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++)
 225                         if (il->limits[type] != cgroup_io_limit_defaults[type])
 226                                 fprintf(f,
 227                                         "%s%s=%s %s\n",
 228                                         prefix,
 229                                         cgroup_io_limit_type_to_string(type),
 230                                         il->path,
 231                                         format_bytes(buf, sizeof(buf), il->limits[type]));
 232         }
 233
 234         LIST_FOREACH(device_weights, w, c->blockio_device_weights)
 235                 fprintf(f,
 236                         "%sBlockIODeviceWeight=%s %" PRIu64,
 237                         prefix,
 238                         w->path,
 239                         w->weight);
 240
 241         LIST_FOREACH(device_bandwidths, b, c->blockio_device_bandwidths) {
 242                 char buf[FORMAT_BYTES_MAX];
 243
 244                 if (b->rbps != CGROUP_LIMIT_MAX)
 245                         fprintf(f,
 246                                 "%sBlockIOReadBandwidth=%s %s\n",
 247                                 prefix,
 248                                 b->path,
 249                                 format_bytes(buf, sizeof(buf), b->rbps));
 250                 if (b->wbps != CGROUP_LIMIT_MAX)
 251                         fprintf(f,
 252                                 "%sBlockIOWriteBandwidth=%s %s\n",
 253                                 prefix,
 254                                 b->path,
 255                                 format_bytes(buf, sizeof(buf), b->wbps));
 256         }
 257 }
 258
 259 static int lookup_block_device(const char *p, dev_t *dev) {
 260         struct stat st;
 261         int r;
 262
 263         assert(p);
 264         assert(dev);
 265
 266         r = stat(p, &st);
 267         if (r < 0)
 268                 return log_warning_errno(errno, "Couldn't stat device %s: %m", p);
 269
 270         if (S_ISBLK(st.st_mode))
 271                 *dev = st.st_rdev;
 272         else if (major(st.st_dev) != 0) {
 273                 /* If this is not a device node then find the block
 274                  * device this file is stored on */
 275                 *dev = st.st_dev;
 276
 277                 /* If this is a partition, try to get the originating
 278                  * block device */
 279                 block_get_whole_disk(*dev, dev);
 280         } else {
 281                 log_warning("%s is not a block device and file system block device cannot be determined or is not local.", p);
 282                 return -ENODEV;
 283         }
 284
 285         return 0;
 286 }
 287
 288 static int whitelist_device(const char *path, const char *node, const char *acc) {
 289         char buf[2+DECIMAL_STR_MAX(dev_t)*2+2+4];
 290         struct stat st;
 291         bool ignore_notfound;
 292         int r;
 293
 294         assert(path);
 295         assert(acc);
 296
 297         if (node[0] == '-') {
 298                 /* Non-existent paths starting with "-" must be silently ignored */
 299                 node++;
 300                 ignore_notfound = true;
 301         } else
 302                 ignore_notfound = false;
 303
 304         if (stat(node, &st) < 0) {
 305                 if (errno == ENOENT && ignore_notfound)
 306                         return 0;
 307
 308                 return log_warning_errno(errno, "Couldn't stat device %s: %m", node);
 309         }
 310
 311         if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
 312                 log_warning("%s is not a device.", node);
 313                 return -ENODEV;
 314         }
 315
 316         sprintf(buf,
 317                 "%c %u:%u %s",
 318                 S_ISCHR(st.st_mode) ? 'c' : 'b',
 319                 major(st.st_rdev), minor(st.st_rdev),
 320                 acc);
 321
 322         r = cg_set_attribute("devices", path, "devices.allow", buf);
 323         if (r < 0)
 324                 log_full_errno(IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 325                                "Failed to set devices.allow on %s: %m", path);
 326
 327         return r;
 328 }
 329
 330 static int whitelist_major(const char *path, const char *name, char type, const char *acc) {
 331         _cleanup_fclose_ FILE *f = NULL;
 332         char line[LINE_MAX];
 333         bool good = false;
 334         int r;
 335
 336         assert(path);
 337         assert(acc);
 338         assert(type == 'b' || type == 'c');
 339
 340         f = fopen("/proc/devices", "re");
 341         if (!f)
 342                 return log_warning_errno(errno, "Cannot open /proc/devices to resolve %s (%c): %m", name, type);
 343
 344         FOREACH_LINE(line, f, goto fail) {
 345                 char buf[2+DECIMAL_STR_MAX(unsigned)+3+4], *p, *w;
 346                 unsigned maj;
 347
 348                 truncate_nl(line);
 349
 350                 if (type == 'c' && streq(line, "Character devices:")) {
 351                         good = true;
 352                         continue;
 353                 }
 354
 355                 if (type == 'b' && streq(line, "Block devices:")) {
 356                         good = true;
 357                         continue;
 358                 }
 359
 360                 if (isempty(line)) {
 361                         good = false;
 362                         continue;
 363                 }
 364
 365                 if (!good)
 366                         continue;
 367
 368                 p = strstrip(line);
 369
 370                 w = strpbrk(p, WHITESPACE);
 371                 if (!w)
 372                         continue;
 373                 *w = 0;
 374
 375                 r = safe_atou(p, &maj);
 376                 if (r < 0)
 377                         continue;
 378                 if (maj <= 0)
 379                         continue;
 380
 381                 w++;
 382                 w += strspn(w, WHITESPACE);
 383
 384                 if (fnmatch(name, w, 0) != 0)
 385                         continue;
 386
 387                 sprintf(buf,
 388                         "%c %u:* %s",
 389                         type,
 390                         maj,
 391                         acc);
 392
 393                 r = cg_set_attribute("devices", path, "devices.allow", buf);
 394                 if (r < 0)
 395                         log_full_errno(IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 396                                        "Failed to set devices.allow on %s: %m", path);
 397         }
 398
 399         return 0;
 400
 401 fail:
 402         return log_warning_errno(errno, "Failed to read /proc/devices: %m");
 403 }
 404
 405 static bool cgroup_context_has_cpu_weight(CGroupContext *c) {
 406         return c->cpu_weight != CGROUP_WEIGHT_INVALID ||
 407                 c->startup_cpu_weight != CGROUP_WEIGHT_INVALID;
 408 }
 409
 410 static bool cgroup_context_has_cpu_shares(CGroupContext *c) {
 411         return c->cpu_shares != CGROUP_CPU_SHARES_INVALID ||
 412                 c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID;
 413 }
 414
 415 static uint64_t cgroup_context_cpu_weight(CGroupContext *c, ManagerState state) {
 416         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 417             c->startup_cpu_weight != CGROUP_WEIGHT_INVALID)
 418                 return c->startup_cpu_weight;
 419         else if (c->cpu_weight != CGROUP_WEIGHT_INVALID)
 420                 return c->cpu_weight;
 421         else
 422                 return CGROUP_WEIGHT_DEFAULT;
 423 }
 424
 425 static uint64_t cgroup_context_cpu_shares(CGroupContext *c, ManagerState state) {
 426         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 427             c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID)
 428                 return c->startup_cpu_shares;
 429         else if (c->cpu_shares != CGROUP_CPU_SHARES_INVALID)
 430                 return c->cpu_shares;
 431         else
 432                 return CGROUP_CPU_SHARES_DEFAULT;
 433 }
 434
 435 static void cgroup_apply_unified_cpu_config(Unit *u, uint64_t weight, uint64_t quota) {
 436         char buf[MAX(DECIMAL_STR_MAX(uint64_t) + 1, (DECIMAL_STR_MAX(usec_t) + 1) * 2)];
 437         int r;
 438
 439         xsprintf(buf, "%" PRIu64 "\n", weight);
 440         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.weight", buf);
 441         if (r < 0)
 442                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 443                               "Failed to set cpu.weight: %m");
 444
 445         if (quota != USEC_INFINITY)
 446                 xsprintf(buf, USEC_FMT " " USEC_FMT "\n",
 447                          quota * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC, CGROUP_CPU_QUOTA_PERIOD_USEC);
 448         else
 449                 xsprintf(buf, "max " USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
 450
 451         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.max", buf);
 452
 453         if (r < 0)
 454                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 455                               "Failed to set cpu.max: %m");
 456 }
 457
 458 static void cgroup_apply_legacy_cpu_config(Unit *u, uint64_t shares, uint64_t quota) {
 459         char buf[MAX(DECIMAL_STR_MAX(uint64_t), DECIMAL_STR_MAX(usec_t)) + 1];
 460         int r;
 461
 462         xsprintf(buf, "%" PRIu64 "\n", shares);
 463         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.shares", buf);
 464         if (r < 0)
 465                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 466                               "Failed to set cpu.shares: %m");
 467
 468         xsprintf(buf, USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
 469         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_period_us", buf);
 470         if (r < 0)
 471                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 472                               "Failed to set cpu.cfs_period_us: %m");
 473
 474         if (quota != USEC_INFINITY) {
 475                 xsprintf(buf, USEC_FMT "\n", quota * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC);
 476                 r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_quota_us", buf);
 477         } else
 478                 r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_quota_us", "-1");
 479         if (r < 0)
 480                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 481                               "Failed to set cpu.cfs_quota_us: %m");
 482 }
 483
 484 static uint64_t cgroup_cpu_shares_to_weight(uint64_t shares) {
 485         return CLAMP(shares * CGROUP_WEIGHT_DEFAULT / CGROUP_CPU_SHARES_DEFAULT,
 486                      CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
 487 }
 488
 489 static uint64_t cgroup_cpu_weight_to_shares(uint64_t weight) {
 490         return CLAMP(weight * CGROUP_CPU_SHARES_DEFAULT / CGROUP_WEIGHT_DEFAULT,
 491                      CGROUP_CPU_SHARES_MIN, CGROUP_CPU_SHARES_MAX);
 492 }
 493
 494 static bool cgroup_context_has_io_config(CGroupContext *c) {
 495         return c->io_accounting ||
 496                 c->io_weight != CGROUP_WEIGHT_INVALID ||
 497                 c->startup_io_weight != CGROUP_WEIGHT_INVALID ||
 498                 c->io_device_weights ||
 499                 c->io_device_limits;
 500 }
 501
 502 static bool cgroup_context_has_blockio_config(CGroupContext *c) {
 503         return c->blockio_accounting ||
 504                 c->blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID ||
 505                 c->startup_blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID ||
 506                 c->blockio_device_weights ||
 507                 c->blockio_device_bandwidths;
 508 }
 509
 510 static uint64_t cgroup_context_io_weight(CGroupContext *c, ManagerState state) {
 511         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 512             c->startup_io_weight != CGROUP_WEIGHT_INVALID)
 513                 return c->startup_io_weight;
 514         else if (c->io_weight != CGROUP_WEIGHT_INVALID)
 515                 return c->io_weight;
 516         else
 517                 return CGROUP_WEIGHT_DEFAULT;
 518 }
 519
 520 static uint64_t cgroup_context_blkio_weight(CGroupContext *c, ManagerState state) {
 521         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 522             c->startup_blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID)
 523                 return c->startup_blockio_weight;
 524         else if (c->blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID)
 525                 return c->blockio_weight;
 526         else
 527                 return CGROUP_BLKIO_WEIGHT_DEFAULT;
 528 }
 529
 530 static uint64_t cgroup_weight_blkio_to_io(uint64_t blkio_weight) {
 531         return CLAMP(blkio_weight * CGROUP_WEIGHT_DEFAULT / CGROUP_BLKIO_WEIGHT_DEFAULT,
 532                      CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
 533 }
 534
 535 static uint64_t cgroup_weight_io_to_blkio(uint64_t io_weight) {
 536         return CLAMP(io_weight * CGROUP_BLKIO_WEIGHT_DEFAULT / CGROUP_WEIGHT_DEFAULT,
 537                      CGROUP_BLKIO_WEIGHT_MIN, CGROUP_BLKIO_WEIGHT_MAX);
 538 }
 539
 540 static void cgroup_apply_io_device_weight(Unit *u, const char *dev_path, uint64_t io_weight) {
 541         char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
 542         dev_t dev;
 543         int r;
 544
 545         r = lookup_block_device(dev_path, &dev);
 546         if (r < 0)
 547                 return;
 548
 549         xsprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), io_weight);
 550         r = cg_set_attribute("io", u->cgroup_path, "io.weight", buf);
 551         if (r < 0)
 552                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 553                               "Failed to set io.weight: %m");
 554 }
 555
 556 static void cgroup_apply_blkio_device_weight(Unit *u, const char *dev_path, uint64_t blkio_weight) {
 557         char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
 558         dev_t dev;
 559         int r;
 560
 561         r = lookup_block_device(dev_path, &dev);
 562         if (r < 0)
 563                 return;
 564
 565         xsprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), blkio_weight);
 566         r = cg_set_attribute("blkio", u->cgroup_path, "blkio.weight_device", buf);
 567         if (r < 0)
 568                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 569                               "Failed to set blkio.weight_device: %m");
 570 }
 571
 572 static unsigned cgroup_apply_io_device_limit(Unit *u, const char *dev_path, uint64_t *limits) {
 573         char limit_bufs[_CGROUP_IO_LIMIT_TYPE_MAX][DECIMAL_STR_MAX(uint64_t)];
 574         char buf[DECIMAL_STR_MAX(dev_t)*2+2+(6+DECIMAL_STR_MAX(uint64_t)+1)*4];
 575         CGroupIOLimitType type;
 576         dev_t dev;
 577         unsigned n = 0;
 578         int r;
 579
 580         r = lookup_block_device(dev_path, &dev);
 581         if (r < 0)
 582                 return 0;
 583
 584         for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++) {
 585                 if (limits[type] != cgroup_io_limit_defaults[type]) {
 586                         xsprintf(limit_bufs[type], "%" PRIu64, limits[type]);
 587                         n++;
 588                 } else {
 589                         xsprintf(limit_bufs[type], "%s", limits[type] == CGROUP_LIMIT_MAX ? "max" : "0");
 590                 }
 591         }
 592
 593         xsprintf(buf, "%u:%u rbps=%s wbps=%s riops=%s wiops=%s\n", major(dev), minor(dev),
 594                  limit_bufs[CGROUP_IO_RBPS_MAX], limit_bufs[CGROUP_IO_WBPS_MAX],
 595                  limit_bufs[CGROUP_IO_RIOPS_MAX], limit_bufs[CGROUP_IO_WIOPS_MAX]);
 596         r = cg_set_attribute("io", u->cgroup_path, "io.max", buf);
 597         if (r < 0)
 598                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 599                               "Failed to set io.max: %m");
 600         return n;
 601 }
 602
 603 static unsigned cgroup_apply_blkio_device_limit(Unit *u, const char *dev_path, uint64_t rbps, uint64_t wbps) {
 604         char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
 605         dev_t dev;
 606         unsigned n = 0;
 607         int r;
 608
 609         r = lookup_block_device(dev_path, &dev);
 610         if (r < 0)
 611                 return 0;
 612
 613         if (rbps != CGROUP_LIMIT_MAX)
 614                 n++;
 615         sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), rbps);
 616         r = cg_set_attribute("blkio", u->cgroup_path, "blkio.throttle.read_bps_device", buf);
 617         if (r < 0)
 618                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 619                               "Failed to set blkio.throttle.read_bps_device: %m");
 620
 621         if (wbps != CGROUP_LIMIT_MAX)
 622                 n++;
 623         sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), wbps);
 624         r = cg_set_attribute("blkio", u->cgroup_path, "blkio.throttle.write_bps_device", buf);
 625         if (r < 0)
 626                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 627                               "Failed to set blkio.throttle.write_bps_device: %m");
 628
 629         return n;
 630 }
 631
 632 static bool cgroup_context_has_unified_memory_config(CGroupContext *c) {
 633         return c->memory_low > 0 || c->memory_high != CGROUP_LIMIT_MAX || c->memory_max != CGROUP_LIMIT_MAX || c->memory_swap_max != CGROUP_LIMIT_MAX;
 634 }
 635
 636 static void cgroup_apply_unified_memory_limit(Unit *u, const char *file, uint64_t v) {
 637         char buf[DECIMAL_STR_MAX(uint64_t) + 1] = "max";
 638         int r;
 639
 640         if (v != CGROUP_LIMIT_MAX)
 641                 xsprintf(buf, "%" PRIu64 "\n", v);
 642
 643         r = cg_set_attribute("memory", u->cgroup_path, file, buf);
 644         if (r < 0)
 645                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 646                               "Failed to set %s: %m", file);
 647 }
 648
 649 static void cgroup_context_apply(Unit *u, CGroupMask mask, ManagerState state) {
 650         const char *path;
 651         CGroupContext *c;
 652         bool is_root;
 653         int r;
 654
 655         assert(u);
 656
 657         c = unit_get_cgroup_context(u);
 658         path = u->cgroup_path;
 659
 660         assert(c);
 661         assert(path);
 662
 663         if (mask == 0)
 664                 return;
 665
 666         /* Some cgroup attributes are not supported on the root cgroup,
 667          * hence silently ignore */
 668         is_root = isempty(path) || path_equal(path, "/");
 669         if (is_root)
 670                 /* Make sure we don't try to display messages with an empty path. */
 671                 path = "/";
 672
 673         /* We generally ignore errors caused by read-only mounted
 674          * cgroup trees (assuming we are running in a container then),
 675          * and missing cgroups, i.e. EROFS and ENOENT. */
 676
 677         if ((mask & CGROUP_MASK_CPU) && !is_root) {
 678                 bool has_weight = cgroup_context_has_cpu_weight(c);
 679                 bool has_shares = cgroup_context_has_cpu_shares(c);
 680
 681                 if (cg_all_unified() > 0) {
 682                         uint64_t weight;
 683
 684                         if (has_weight)
 685                                 weight = cgroup_context_cpu_weight(c, state);
 686                         else if (has_shares) {
 687                                 uint64_t shares = cgroup_context_cpu_shares(c, state);
 688
 689                                 weight = cgroup_cpu_shares_to_weight(shares);
 690
 691                                 log_cgroup_compat(u, "Applying [Startup]CpuShares %" PRIu64 " as [Startup]CpuWeight %" PRIu64 " on %s",
 692                                                   shares, weight, path);
 693                         } else
 694                                 weight = CGROUP_WEIGHT_DEFAULT;
 695
 696                         cgroup_apply_unified_cpu_config(u, weight, c->cpu_quota_per_sec_usec);
 697                 } else {
 698                         uint64_t shares;
 699
 700                         if (has_weight) {
 701                                 uint64_t weight = cgroup_context_cpu_weight(c, state);
 702
 703                                 shares = cgroup_cpu_weight_to_shares(weight);
 704
 705                                 log_cgroup_compat(u, "Applying [Startup]CpuWeight %" PRIu64 " as [Startup]CpuShares %" PRIu64 " on %s",
 706                                                   weight, shares, path);
 707                         } else if (has_shares)
 708                                 shares = cgroup_context_cpu_shares(c, state);
 709                         else
 710                                 shares = CGROUP_CPU_SHARES_DEFAULT;
 711
 712                         cgroup_apply_legacy_cpu_config(u, shares, c->cpu_quota_per_sec_usec);
 713                 }
 714         }
 715
 716         if (mask & CGROUP_MASK_IO) {
 717                 bool has_io = cgroup_context_has_io_config(c);
 718                 bool has_blockio = cgroup_context_has_blockio_config(c);
 719
 720                 if (!is_root) {
 721                         char buf[8+DECIMAL_STR_MAX(uint64_t)+1];
 722                         uint64_t weight;
 723
 724                         if (has_io)
 725                                 weight = cgroup_context_io_weight(c, state);
 726                         else if (has_blockio) {
 727                                 uint64_t blkio_weight = cgroup_context_blkio_weight(c, state);
 728
 729                                 weight = cgroup_weight_blkio_to_io(blkio_weight);
 730
 731                                 log_cgroup_compat(u, "Applying [Startup]BlockIOWeight %" PRIu64 " as [Startup]IOWeight %" PRIu64,
 732                                                   blkio_weight, weight);
 733                         } else
 734                                 weight = CGROUP_WEIGHT_DEFAULT;
 735
 736                         xsprintf(buf, "default %" PRIu64 "\n", weight);
 737                         r = cg_set_attribute("io", path, "io.weight", buf);
 738                         if (r < 0)
 739                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 740                                               "Failed to set io.weight: %m");
 741
 742                         if (has_io) {
 743                                 CGroupIODeviceWeight *w;
 744
 745                                 /* FIXME: no way to reset this list */
 746                                 LIST_FOREACH(device_weights, w, c->io_device_weights)
 747                                         cgroup_apply_io_device_weight(u, w->path, w->weight);
 748                         } else if (has_blockio) {
 749                                 CGroupBlockIODeviceWeight *w;
 750
 751                                 /* FIXME: no way to reset this list */
 752                                 LIST_FOREACH(device_weights, w, c->blockio_device_weights) {
 753                                         weight = cgroup_weight_blkio_to_io(w->weight);
 754
 755                                         log_cgroup_compat(u, "Applying BlockIODeviceWeight %" PRIu64 " as IODeviceWeight %" PRIu64 " for %s",
 756                                                           w->weight, weight, w->path);
 757
 758                                         cgroup_apply_io_device_weight(u, w->path, weight);
 759                                 }
 760                         }
 761                 }
 762
 763                 /* Apply limits and free ones without config. */
 764                 if (has_io) {
 765                         CGroupIODeviceLimit *l, *next;
 766
 767                         LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
 768                                 if (!cgroup_apply_io_device_limit(u, l->path, l->limits))
 769                                         cgroup_context_free_io_device_limit(c, l);
 770                         }
 771                 } else if (has_blockio) {
 772                         CGroupBlockIODeviceBandwidth *b, *next;
 773
 774                         LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths) {
 775                                 uint64_t limits[_CGROUP_IO_LIMIT_TYPE_MAX];
 776                                 CGroupIOLimitType type;
 777
 778                                 for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++)
 779                                         limits[type] = cgroup_io_limit_defaults[type];
 780
 781                                 limits[CGROUP_IO_RBPS_MAX] = b->rbps;
 782                                 limits[CGROUP_IO_WBPS_MAX] = b->wbps;
 783
 784                                 log_cgroup_compat(u, "Applying BlockIO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as IO{Read|Write}BandwidthMax for %s",
 785                                                   b->rbps, b->wbps, b->path);
 786
 787                                 if (!cgroup_apply_io_device_limit(u, b->path, limits))
 788                                         cgroup_context_free_blockio_device_bandwidth(c, b);
 789                         }
 790                 }
 791         }
 792
 793         if (mask & CGROUP_MASK_BLKIO) {
 794                 bool has_io = cgroup_context_has_io_config(c);
 795                 bool has_blockio = cgroup_context_has_blockio_config(c);
 796
 797                 if (!is_root) {
 798                         char buf[DECIMAL_STR_MAX(uint64_t)+1];
 799                         uint64_t weight;
 800
 801                         if (has_io) {
 802                                 uint64_t io_weight = cgroup_context_io_weight(c, state);
 803
 804                                 weight = cgroup_weight_io_to_blkio(cgroup_context_io_weight(c, state));
 805
 806                                 log_cgroup_compat(u, "Applying [Startup]IOWeight %" PRIu64 " as [Startup]BlockIOWeight %" PRIu64,
 807                                                   io_weight, weight);
 808                         } else if (has_blockio)
 809                                 weight = cgroup_context_blkio_weight(c, state);
 810                         else
 811                                 weight = CGROUP_BLKIO_WEIGHT_DEFAULT;
 812
 813                         xsprintf(buf, "%" PRIu64 "\n", weight);
 814                         r = cg_set_attribute("blkio", path, "blkio.weight", buf);
 815                         if (r < 0)
 816                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 817                                               "Failed to set blkio.weight: %m");
 818
 819                         if (has_io) {
 820                                 CGroupIODeviceWeight *w;
 821
 822                                 /* FIXME: no way to reset this list */
 823                                 LIST_FOREACH(device_weights, w, c->io_device_weights) {
 824                                         weight = cgroup_weight_io_to_blkio(w->weight);
 825
 826                                         log_cgroup_compat(u, "Applying IODeviceWeight %" PRIu64 " as BlockIODeviceWeight %" PRIu64 " for %s",
 827                                                           w->weight, weight, w->path);
 828
 829                                         cgroup_apply_blkio_device_weight(u, w->path, weight);
 830                                 }
 831                         } else if (has_blockio) {
 832                                 CGroupBlockIODeviceWeight *w;
 833
 834                                 /* FIXME: no way to reset this list */
 835                                 LIST_FOREACH(device_weights, w, c->blockio_device_weights)
 836                                         cgroup_apply_blkio_device_weight(u, w->path, w->weight);
 837                         }
 838                 }
 839
 840                 /* Apply limits and free ones without config. */
 841                 if (has_io) {
 842                         CGroupIODeviceLimit *l, *next;
 843
 844                         LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
 845                                 log_cgroup_compat(u, "Applying IO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as BlockIO{Read|Write}BandwidthMax for %s",
 846                                                   l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX], l->path);
 847
 848                                 if (!cgroup_apply_blkio_device_limit(u, l->path, l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX]))
 849                                         cgroup_context_free_io_device_limit(c, l);
 850                         }
 851                 } else if (has_blockio) {
 852                         CGroupBlockIODeviceBandwidth *b, *next;
 853
 854                         LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths)
 855                                 if (!cgroup_apply_blkio_device_limit(u, b->path, b->rbps, b->wbps))
 856                                         cgroup_context_free_blockio_device_bandwidth(c, b);
 857                 }
 858         }
 859
 860         if ((mask & CGROUP_MASK_MEMORY) && !is_root) {
 861                 if (cg_all_unified() > 0) {
 862                         uint64_t max, swap_max = CGROUP_LIMIT_MAX;
 863
 864                         if (cgroup_context_has_unified_memory_config(c)) {
 865                                 max = c->memory_max;
 866                                 swap_max = c->memory_swap_max;
 867                         } else {
 868                                 max = c->memory_limit;
 869
 870                                 if (max != CGROUP_LIMIT_MAX)
 871                                         log_cgroup_compat(u, "Applying MemoryLimit %" PRIu64 " as MemoryMax", max);
 872                         }
 873
 874                         cgroup_apply_unified_memory_limit(u, "memory.low", c->memory_low);
 875                         cgroup_apply_unified_memory_limit(u, "memory.high", c->memory_high);
 876                         cgroup_apply_unified_memory_limit(u, "memory.max", max);
 877                         cgroup_apply_unified_memory_limit(u, "memory.swap.max", swap_max);
 878                 } else {
 879                         char buf[DECIMAL_STR_MAX(uint64_t) + 1];
 880                         uint64_t val;
 881
 882                         if (cgroup_context_has_unified_memory_config(c)) {
 883                                 val = c->memory_max;
 884                                 log_cgroup_compat(u, "Applying MemoryMax %" PRIi64 " as MemoryLimit", val);
 885                         } else
 886                                 val = c->memory_limit;
 887
 888                         if (val == CGROUP_LIMIT_MAX)
 889                                 strncpy(buf, "-1\n", sizeof(buf));
 890                         else
 891                                 xsprintf(buf, "%" PRIu64 "\n", val);
 892
 893                         r = cg_set_attribute("memory", path, "memory.limit_in_bytes", buf);
 894                         if (r < 0)
 895                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 896                                               "Failed to set memory.limit_in_bytes: %m");
 897                 }
 898         }
 899
 900         if ((mask & CGROUP_MASK_DEVICES) && !is_root) {
 901                 CGroupDeviceAllow *a;
 902
 903                 /* Changing the devices list of a populated cgroup
 904                  * might result in EINVAL, hence ignore EINVAL
 905                  * here. */
 906
 907                 if (c->device_allow || c->device_policy != CGROUP_AUTO)
 908                         r = cg_set_attribute("devices", path, "devices.deny", "a");
 909                 else
 910                         r = cg_set_attribute("devices", path, "devices.allow", "a");
 911                 if (r < 0)
 912                         log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 913                                       "Failed to reset devices.list: %m");
 914
 915                 if (c->device_policy == CGROUP_CLOSED ||
 916                     (c->device_policy == CGROUP_AUTO && c->device_allow)) {
 917                         static const char auto_devices[] =
 918                                 "/dev/null\0" "rwm\0"
 919                                 "/dev/zero\0" "rwm\0"
 920                                 "/dev/full\0" "rwm\0"
 921                                 "/dev/random\0" "rwm\0"
 922                                 "/dev/urandom\0" "rwm\0"
 923                                 "/dev/tty\0" "rwm\0"
 924                                 "/dev/pts/ptmx\0" "rw\0" /* /dev/pts/ptmx may not be duplicated, but accessed */
 925                                 /* Allow /run/elogind/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
 926                                 /* Allow /run/systemd/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
 927                                 "-/run/systemd/inaccessible/chr\0" "rwm\0"
 928                                 "-/run/systemd/inaccessible/blk\0" "rwm\0";
 929
 930                         const char *x, *y;
 931
 932                         NULSTR_FOREACH_PAIR(x, y, auto_devices)
 933                                 whitelist_device(path, x, y);
 934
 935                         whitelist_major(path, "pts", 'c', "rw");
 936                         whitelist_major(path, "kdbus", 'c', "rw");
 937                         whitelist_major(path, "kdbus/*", 'c', "rw");
 938                 }
 939
 940                 LIST_FOREACH(device_allow, a, c->device_allow) {
 941                         char acc[4], *val;
 942                         unsigned k = 0;
 943
 944                         if (a->r)
 945                                 acc[k++] = 'r';
 946                         if (a->w)
 947                                 acc[k++] = 'w';
 948                         if (a->m)
 949                                 acc[k++] = 'm';
 950
 951                         if (k == 0)
 952                                 continue;
 953
 954                         acc[k++] = 0;
 955
 956                         if (startswith(a->path, "/dev/"))
 957                                 whitelist_device(path, a->path, acc);
 958                         else if ((val = startswith(a->path, "block-")))
 959                                 whitelist_major(path, val, 'b', acc);
 960                         else if ((val = startswith(a->path, "char-")))
 961                                 whitelist_major(path, val, 'c', acc);
 962                         else
 963                                 log_unit_debug(u, "Ignoring device %s while writing cgroup attribute.", a->path);
 964                 }
 965         }
 966
 967         if ((mask & CGROUP_MASK_PIDS) && !is_root) {
 968
 969                 if (c->tasks_max != CGROUP_LIMIT_MAX) {
 970                         char buf[DECIMAL_STR_MAX(uint64_t) + 2];
 971
 972                         sprintf(buf, "%" PRIu64 "\n", c->tasks_max);
 973                         r = cg_set_attribute("pids", path, "pids.max", buf);
 974                 } else
 975                         r = cg_set_attribute("pids", path, "pids.max", "max");
 976
 977                 if (r < 0)
 978                         log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 979                                       "Failed to set pids.max: %m");
 980         }
 981 }
 982
 983 CGroupMask cgroup_context_get_mask(CGroupContext *c) {
 984         CGroupMask mask = 0;
 985
 986         /* Figure out which controllers we need */
 987
 988         if (c->cpu_accounting ||
 989             cgroup_context_has_cpu_weight(c) ||
 990             cgroup_context_has_cpu_shares(c) ||
 991             c->cpu_quota_per_sec_usec != USEC_INFINITY)
 992                 mask |= CGROUP_MASK_CPUACCT | CGROUP_MASK_CPU;
 993
 994         if (cgroup_context_has_io_config(c) || cgroup_context_has_blockio_config(c))
 995                 mask |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
 996
 997         if (c->memory_accounting ||
 998             c->memory_limit != CGROUP_LIMIT_MAX ||
 999             cgroup_context_has_unified_memory_config(c))
1000                 mask |= CGROUP_MASK_MEMORY;
1001
1002         if (c->device_allow ||
1003             c->device_policy != CGROUP_AUTO)
1004                 mask |= CGROUP_MASK_DEVICES;
1005
1006         if (c->tasks_accounting ||
1007             c->tasks_max != (uint64_t) -1)
1008                 mask |= CGROUP_MASK_PIDS;
1009
1010         return mask;
1011 }
1012
1013 CGroupMask unit_get_own_mask(Unit *u) {
1014         CGroupContext *c;
1015
1016         /* Returns the mask of controllers the unit needs for itself */
1017
1018         c = unit_get_cgroup_context(u);
1019         if (!c)
1020                 return 0;
1021
1022         /* If delegation is turned on, then turn on all cgroups,
1023          * unless we are on the legacy hierarchy and the process we
1024          * fork into it is known to drop privileges, and hence
1025          * shouldn't get access to the controllers.
1026          *
1027          * Note that on the unified hierarchy it is safe to delegate
1028          * controllers to unprivileged services. */
1029
1030         if (c->delegate) {
1031                 ExecContext *e;
1032
1033                 e = unit_get_exec_context(u);
1034                 if (!e ||
1035                     exec_context_maintains_privileges(e) ||
1036                     cg_all_unified() > 0)
1037                         return _CGROUP_MASK_ALL;
1038         }
1039
1040         return cgroup_context_get_mask(c);
1041 }
1042
1043 CGroupMask unit_get_members_mask(Unit *u) {
1044         assert(u);
1045
1046         /* Returns the mask of controllers all of the unit's children
1047          * require, merged */
1048
1049         if (u->cgroup_members_mask_valid)
1050                 return u->cgroup_members_mask;
1051
1052         u->cgroup_members_mask = 0;
1053
1054         if (u->type == UNIT_SLICE) {
1055                 Unit *member;
1056                 Iterator i;
1057
1058                 SET_FOREACH(member, u->dependencies[UNIT_BEFORE], i) {
1059
1060                         if (member == u)
1061                                 continue;
1062
1063                         if (UNIT_DEREF(member->slice) != u)
1064                                 continue;
1065
1066                         u->cgroup_members_mask |=
1067                                 unit_get_own_mask(member) |
1068                                 unit_get_members_mask(member);
1069                 }
1070         }
1071
1072         u->cgroup_members_mask_valid = true;
1073         return u->cgroup_members_mask;
1074 }
1075
1076 CGroupMask unit_get_siblings_mask(Unit *u) {
1077         assert(u);
1078
1079         /* Returns the mask of controllers all of the unit's siblings
1080          * require, i.e. the members mask of the unit's parent slice
1081          * if there is one. */
1082
1083         if (UNIT_ISSET(u->slice))
1084                 return unit_get_members_mask(UNIT_DEREF(u->slice));
1085
1086         return unit_get_own_mask(u) | unit_get_members_mask(u);
1087 }
1088
1089 CGroupMask unit_get_subtree_mask(Unit *u) {
1090
1091         /* Returns the mask of this subtree, meaning of the group
1092          * itself and its children. */
1093
1094         return unit_get_own_mask(u) | unit_get_members_mask(u);
1095 }
1096
1097 CGroupMask unit_get_target_mask(Unit *u) {
1098         CGroupMask mask;
1099
1100         /* This returns the cgroup mask of all controllers to enable
1101          * for a specific cgroup, i.e. everything it needs itself,
1102          * plus all that its children need, plus all that its siblings
1103          * need. This is primarily useful on the legacy cgroup
1104          * hierarchy, where we need to duplicate each cgroup in each
1105          * hierarchy that shall be enabled for it. */
1106
1107         mask = unit_get_own_mask(u) | unit_get_members_mask(u) | unit_get_siblings_mask(u);
1108         mask &= u->manager->cgroup_supported;
1109
1110         return mask;
1111 }
1112
1113 CGroupMask unit_get_enable_mask(Unit *u) {
1114         CGroupMask mask;
1115
1116         /* This returns the cgroup mask of all controllers to enable
1117          * for the children of a specific cgroup. This is primarily
1118          * useful for the unified cgroup hierarchy, where each cgroup
1119          * controls which controllers are enabled for its children. */
1120
1121         mask = unit_get_members_mask(u);
1122         mask &= u->manager->cgroup_supported;
1123
1124         return mask;
1125 }
1126
1127 /* Recurse from a unit up through its containing slices, propagating
1128  * mask bits upward. A unit is also member of itself. */
1129 void unit_update_cgroup_members_masks(Unit *u) {
1130         CGroupMask m;
1131         bool more;
1132
1133         assert(u);
1134
1135         /* Calculate subtree mask */
1136         m = unit_get_subtree_mask(u);
1137
1138         /* See if anything changed from the previous invocation. If
1139          * not, we're done. */
1140         if (u->cgroup_subtree_mask_valid && m == u->cgroup_subtree_mask)
1141                 return;
1142
1143         more =
1144                 u->cgroup_subtree_mask_valid &&
1145                 ((m & ~u->cgroup_subtree_mask) != 0) &&
1146                 ((~m & u->cgroup_subtree_mask) == 0);
1147
1148         u->cgroup_subtree_mask = m;
1149         u->cgroup_subtree_mask_valid = true;
1150
1151         if (UNIT_ISSET(u->slice)) {
1152                 Unit *s = UNIT_DEREF(u->slice);
1153
1154                 if (more)
1155                         /* There's more set now than before. We
1156                          * propagate the new mask to the parent's mask
1157                          * (not caring if it actually was valid or
1158                          * not). */
1159
1160                         s->cgroup_members_mask |= m;
1161
1162                 else
1163                         /* There's less set now than before (or we
1164                          * don't know), we need to recalculate
1165                          * everything, so let's invalidate the
1166                          * parent's members mask */
1167
1168                         s->cgroup_members_mask_valid = false;
1169
1170                 /* And now make sure that this change also hits our
1171                  * grandparents */
1172                 unit_update_cgroup_members_masks(s);
1173         }
1174 }
1175
1176 static const char *migrate_callback(CGroupMask mask, void *userdata) {
1177         Unit *u = userdata;
1178
1179         assert(mask != 0);
1180         assert(u);
1181
1182         while (u) {
1183                 if (u->cgroup_path &&
1184                     u->cgroup_realized &&
1185                     (u->cgroup_realized_mask & mask) == mask)
1186                         return u->cgroup_path;
1187
1188                 u = UNIT_DEREF(u->slice);
1189         }
1190
1191         return NULL;
1192 }
1193
1194 char *unit_default_cgroup_path(Unit *u) {
1195         _cleanup_free_ char *escaped = NULL, *slice = NULL;
1196         int r;
1197
1198         assert(u);
1199
1200         if (unit_has_name(u, SPECIAL_ROOT_SLICE))
1201                 return strdup(u->manager->cgroup_root);
1202
1203         if (UNIT_ISSET(u->slice) && !unit_has_name(UNIT_DEREF(u->slice), SPECIAL_ROOT_SLICE)) {
1204                 r = cg_slice_to_path(UNIT_DEREF(u->slice)->id, &slice);
1205                 if (r < 0)
1206                         return NULL;
1207         }
1208
1209         escaped = cg_escape(u->id);
1210         if (!escaped)
1211                 return NULL;
1212
1213         if (slice)
1214                 return strjoin(u->manager->cgroup_root, "/", slice, "/",
1215                                escaped);
1216         else
1217                 return strjoin(u->manager->cgroup_root, "/", escaped);
1218 }
1219
1220 int unit_set_cgroup_path(Unit *u, const char *path) {
1221         _cleanup_free_ char *p = NULL;
1222         int r;
1223
1224         assert(u);
1225
1226         if (path) {
1227                 p = strdup(path);
1228                 if (!p)
1229                         return -ENOMEM;
1230         } else
1231                 p = NULL;
1232
1233         if (streq_ptr(u->cgroup_path, p))
1234                 return 0;
1235
1236         if (p) {
1237                 r = hashmap_put(u->manager->cgroup_unit, p, u);
1238                 if (r < 0)
1239                         return r;
1240         }
1241
1242         unit_release_cgroup(u);
1243
1244         u->cgroup_path = p;
1245         p = NULL;
1246
1247         return 1;
1248 }
1249
1250 int unit_watch_cgroup(Unit *u) {
1251         _cleanup_free_ char *events = NULL;
1252         int r;
1253
1254         assert(u);
1255
1256         if (!u->cgroup_path)
1257                 return 0;
1258
1259         if (u->cgroup_inotify_wd >= 0)
1260                 return 0;
1261
1262         /* Only applies to the unified hierarchy */
1263         r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1264         if (r < 0)
1265                 return log_error_errno(r, "Failed to determine whether the name=systemd hierarchy is unified: %m");
1266         if (r == 0)
1267                 return 0;
1268
1269         /* Don't watch the root slice, it's pointless. */
1270         if (unit_has_name(u, SPECIAL_ROOT_SLICE))
1271                 return 0;
1272
1273         r = hashmap_ensure_allocated(&u->manager->cgroup_inotify_wd_unit, &trivial_hash_ops);
1274         if (r < 0)
1275                 return log_oom();
1276
1277         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, "cgroup.events", &events);
1278         if (r < 0)
1279                 return log_oom();
1280
1281         u->cgroup_inotify_wd = inotify_add_watch(u->manager->cgroup_inotify_fd, events, IN_MODIFY);
1282         if (u->cgroup_inotify_wd < 0) {
1283
1284                 if (errno == ENOENT) /* If the directory is already
1285                                       * gone we don't need to track
1286                                       * it, so this is not an error */
1287                         return 0;
1288
1289                 return log_unit_error_errno(u, errno, "Failed to add inotify watch descriptor for control group %s: %m", u->cgroup_path);
1290         }
1291
1292         r = hashmap_put(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd), u);
1293         if (r < 0)
1294                 return log_unit_error_errno(u, r, "Failed to add inotify watch descriptor to hash map: %m");
1295
1296         return 0;
1297 }
1298
1299 static int unit_create_cgroup(
1300                 Unit *u,
1301                 CGroupMask target_mask,
1302                 CGroupMask enable_mask) {
1303
1304         CGroupContext *c;
1305         int r;
1306
1307         assert(u);
1308
1309         c = unit_get_cgroup_context(u);
1310         if (!c)
1311                 return 0;
1312
1313         if (!u->cgroup_path) {
1314                 _cleanup_free_ char *path = NULL;
1315
1316                 path = unit_default_cgroup_path(u);
1317                 if (!path)
1318                         return log_oom();
1319
1320                 r = unit_set_cgroup_path(u, path);
1321                 if (r == -EEXIST)
1322                         return log_unit_error_errno(u, r, "Control group %s exists already.", path);
1323                 if (r < 0)
1324                         return log_unit_error_errno(u, r, "Failed to set unit's control group path to %s: %m", path);
1325         }
1326
1327         /* First, create our own group */
1328         r = cg_create_everywhere(u->manager->cgroup_supported, target_mask, u->cgroup_path);
1329         if (r < 0)
1330                 return log_unit_error_errno(u, r, "Failed to create cgroup %s: %m", u->cgroup_path);
1331
1332         /* Start watching it */
1333         (void) unit_watch_cgroup(u);
1334
1335         /* Enable all controllers we need */
1336         r = cg_enable_everywhere(u->manager->cgroup_supported, enable_mask, u->cgroup_path);
1337         if (r < 0)
1338                 log_unit_warning_errno(u, r, "Failed to enable controllers on cgroup %s, ignoring: %m", u->cgroup_path);
1339
1340         /* Keep track that this is now realized */
1341         u->cgroup_realized = true;
1342         u->cgroup_realized_mask = target_mask;
1343         u->cgroup_enabled_mask = enable_mask;
1344
1345         if (u->type != UNIT_SLICE && !c->delegate) {
1346
1347                 /* Then, possibly move things over, but not if
1348                  * subgroups may contain processes, which is the case
1349                  * for slice and delegation units. */
1350                 r = cg_migrate_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->cgroup_path, migrate_callback, u);
1351                 if (r < 0)
1352                         log_unit_warning_errno(u, r, "Failed to migrate cgroup from to %s, ignoring: %m", u->cgroup_path);
1353         }
1354
1355         return 0;
1356 }
1357
1358 int unit_attach_pids_to_cgroup(Unit *u) {
1359         int r;
1360         assert(u);
1361
1362         r = unit_realize_cgroup(u);
1363         if (r < 0)
1364                 return r;
1365
1366         r = cg_attach_many_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->pids, migrate_callback, u);
1367         if (r < 0)
1368                 return r;
1369
1370         return 0;
1371 }
1372
1373 static void cgroup_xattr_apply(Unit *u) {
1374         char ids[SD_ID128_STRING_MAX];
1375         int r;
1376
1377         assert(u);
1378
1379         if (!MANAGER_IS_SYSTEM(u->manager))
1380                 return;
1381
1382         if (sd_id128_is_null(u->invocation_id))
1383                 return;
1384
1385         r = cg_set_xattr(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path,
1386                          "trusted.invocation_id",
1387                          sd_id128_to_string(u->invocation_id, ids), 32,
1388                          0);
1389         if (r < 0)
1390                 log_unit_warning_errno(u, r, "Failed to set invocation ID on control group %s, ignoring: %m", u->cgroup_path);
1391 }
1392
1393 static bool unit_has_mask_realized(Unit *u, CGroupMask target_mask, CGroupMask enable_mask) {
1394         assert(u);
1395
1396         return u->cgroup_realized && u->cgroup_realized_mask == target_mask && u->cgroup_enabled_mask == enable_mask;
1397 }
1398
1399 /* Check if necessary controllers and attributes for a unit are in place.
1400  *
1401  * If so, do nothing.
1402  * If not, create paths, move processes over, and set attributes.
1403  *
1404  * Returns 0 on success and < 0 on failure. */
1405 static int unit_realize_cgroup_now(Unit *u, ManagerState state) {
1406         CGroupMask target_mask, enable_mask;
1407         int r;
1408
1409         assert(u);
1410
1411         if (u->in_cgroup_queue) {
1412                 LIST_REMOVE(cgroup_queue, u->manager->cgroup_queue, u);
1413                 u->in_cgroup_queue = false;
1414         }
1415
1416         target_mask = unit_get_target_mask(u);
1417         enable_mask = unit_get_enable_mask(u);
1418
1419         if (unit_has_mask_realized(u, target_mask, enable_mask))
1420                 return 0;
1421
1422         /* First, realize parents */
1423         if (UNIT_ISSET(u->slice)) {
1424                 r = unit_realize_cgroup_now(UNIT_DEREF(u->slice), state);
1425                 if (r < 0)
1426                         return r;
1427         }
1428
1429         /* And then do the real work */
1430         r = unit_create_cgroup(u, target_mask, enable_mask);
1431         if (r < 0)
1432                 return r;
1433
1434         /* Finally, apply the necessary attributes. */
1435         cgroup_context_apply(u, target_mask, state);
1436         cgroup_xattr_apply(u);
1437
1438         return 0;
1439 }
1440
1441 static void unit_add_to_cgroup_queue(Unit *u) {
1442
1443         if (u->in_cgroup_queue)
1444                 return;
1445
1446         LIST_PREPEND(cgroup_queue, u->manager->cgroup_queue, u);
1447         u->in_cgroup_queue = true;
1448 }
1449
1450 unsigned manager_dispatch_cgroup_queue(Manager *m) {
1451         ManagerState state;
1452         unsigned n = 0;
1453         Unit *i;
1454         int r;
1455
1456         state = manager_state(m);
1457
1458         while ((i = m->cgroup_queue)) {
1459                 assert(i->in_cgroup_queue);
1460
1461                 r = unit_realize_cgroup_now(i, state);
1462                 if (r < 0)
1463                         log_warning_errno(r, "Failed to realize cgroups for queued unit %s, ignoring: %m", i->id);
1464
1465                 n++;
1466         }
1467
1468         return n;
1469 }
1470
1471 static void unit_queue_siblings(Unit *u) {
1472         Unit *slice;
1473
1474         /* This adds the siblings of the specified unit and the
1475          * siblings of all parent units to the cgroup queue. (But
1476          * neither the specified unit itself nor the parents.) */
1477
1478         while ((slice = UNIT_DEREF(u->slice))) {
1479                 Iterator i;
1480                 Unit *m;
1481
1482                 SET_FOREACH(m, slice->dependencies[UNIT_BEFORE], i) {
1483                         if (m == u)
1484                                 continue;
1485
1486                         /* Skip units that have a dependency on the slice
1487                          * but aren't actually in it. */
1488                         if (UNIT_DEREF(m->slice) != slice)
1489                                 continue;
1490
1491                         /* No point in doing cgroup application for units
1492                          * without active processes. */
1493                         if (UNIT_IS_INACTIVE_OR_FAILED(unit_active_state(m)))
1494                                 continue;
1495
1496                         /* If the unit doesn't need any new controllers
1497                          * and has current ones realized, it doesn't need
1498                          * any changes. */
1499                         if (unit_has_mask_realized(m, unit_get_target_mask(m), unit_get_enable_mask(m)))
1500                                 continue;
1501
1502                         unit_add_to_cgroup_queue(m);
1503                 }
1504
1505                 u = slice;
1506         }
1507 }
1508
1509 int unit_realize_cgroup(Unit *u) {
1510         assert(u);
1511
1512         if (!UNIT_HAS_CGROUP_CONTEXT(u))
1513                 return 0;
1514
1515         /* So, here's the deal: when realizing the cgroups for this
1516          * unit, we need to first create all parents, but there's more
1517          * actually: for the weight-based controllers we also need to
1518          * make sure that all our siblings (i.e. units that are in the
1519          * same slice as we are) have cgroups, too. Otherwise, things
1520          * would become very uneven as each of their processes would
1521          * get as much resources as all our group together. This call
1522          * will synchronously create the parent cgroups, but will
1523          * defer work on the siblings to the next event loop
1524          * iteration. */
1525
1526         /* Add all sibling slices to the cgroup queue. */
1527         unit_queue_siblings(u);
1528
1529         /* And realize this one now (and apply the values) */
1530         return unit_realize_cgroup_now(u, manager_state(u->manager));
1531 }
1532
1533 void unit_release_cgroup(Unit *u) {
1534         assert(u);
1535
1536         /* Forgets all cgroup details for this cgroup */
1537
1538         if (u->cgroup_path) {
1539                 (void) hashmap_remove(u->manager->cgroup_unit, u->cgroup_path);
1540                 u->cgroup_path = mfree(u->cgroup_path);
1541         }
1542
1543         if (u->cgroup_inotify_wd >= 0) {
1544                 if (inotify_rm_watch(u->manager->cgroup_inotify_fd, u->cgroup_inotify_wd) < 0)
1545                         log_unit_debug_errno(u, errno, "Failed to remove cgroup inotify watch %i for %s, ignoring", u->cgroup_inotify_wd, u->id);
1546
1547                 (void) hashmap_remove(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd));
1548                 u->cgroup_inotify_wd = -1;
1549         }
1550 }
1551
1552 void unit_prune_cgroup(Unit *u) {
1553         int r;
1554         bool is_root_slice;
1555
1556         assert(u);
1557
1558         /* Removes the cgroup, if empty and possible, and stops watching it. */
1559
1560         if (!u->cgroup_path)
1561                 return;
1562
1563         (void) unit_get_cpu_usage(u, NULL); /* Cache the last CPU usage value before we destroy the cgroup */
1564
1565         is_root_slice = unit_has_name(u, SPECIAL_ROOT_SLICE);
1566
1567         r = cg_trim_everywhere(u->manager->cgroup_supported, u->cgroup_path, !is_root_slice);
1568         if (r < 0) {
1569                 log_unit_debug_errno(u, r, "Failed to destroy cgroup %s, ignoring: %m", u->cgroup_path);
1570                 return;
1571         }
1572
1573         if (is_root_slice)
1574                 return;
1575
1576         unit_release_cgroup(u);
1577
1578         u->cgroup_realized = false;
1579         u->cgroup_realized_mask = 0;
1580         u->cgroup_enabled_mask = 0;
1581 }
1582
1583 int unit_search_main_pid(Unit *u, pid_t *ret) {
1584         _cleanup_fclose_ FILE *f = NULL;
1585         pid_t pid = 0, npid, mypid;
1586         int r;
1587
1588         assert(u);
1589         assert(ret);
1590
1591         if (!u->cgroup_path)
1592                 return -ENXIO;
1593
1594         r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, &f);
1595         if (r < 0)
1596                 return r;
1597
1598         mypid = getpid();
1599         while (cg_read_pid(f, &npid) > 0)  {
1600                 pid_t ppid;
1601
1602                 if (npid == pid)
1603                         continue;
1604
1605                 /* Ignore processes that aren't our kids */
1606                 if (get_process_ppid(npid, &ppid) >= 0 && ppid != mypid)
1607                         continue;
1608
1609                 if (pid != 0)
1610                         /* Dang, there's more than one daemonized PID
1611                         in this group, so we don't know what process
1612                         is the main process. */
1613
1614                         return -ENODATA;
1615
1616                 pid = npid;
1617         }
1618
1619         *ret = pid;
1620         return 0;
1621 }
1622
1623 static int unit_watch_pids_in_path(Unit *u, const char *path) {
1624         _cleanup_closedir_ DIR *d = NULL;
1625         _cleanup_fclose_ FILE *f = NULL;
1626         int ret = 0, r;
1627
1628         assert(u);
1629         assert(path);
1630
1631         r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, path, &f);
1632         if (r < 0)
1633                 ret = r;
1634         else {
1635                 pid_t pid;
1636
1637                 while ((r = cg_read_pid(f, &pid)) > 0) {
1638                         r = unit_watch_pid(u, pid);
1639                         if (r < 0 && ret >= 0)
1640                                 ret = r;
1641                 }
1642
1643                 if (r < 0 && ret >= 0)
1644                         ret = r;
1645         }
1646
1647         r = cg_enumerate_subgroups(SYSTEMD_CGROUP_CONTROLLER, path, &d);
1648         if (r < 0) {
1649                 if (ret >= 0)
1650                         ret = r;
1651         } else {
1652                 char *fn;
1653
1654                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1655                         _cleanup_free_ char *p = NULL;
1656
1657                         p = strjoin(path, "/", fn);
1658                         free(fn);
1659
1660                         if (!p)
1661                                 return -ENOMEM;
1662
1663                         r = unit_watch_pids_in_path(u, p);
1664                         if (r < 0 && ret >= 0)
1665                                 ret = r;
1666                 }
1667
1668                 if (r < 0 && ret >= 0)
1669                         ret = r;
1670         }
1671
1672         return ret;
1673 }
1674
1675 int unit_watch_all_pids(Unit *u) {
1676         int r;
1677
1678         assert(u);
1679
1680         /* Adds all PIDs from our cgroup to the set of PIDs we
1681          * watch. This is a fallback logic for cases where we do not
1682          * get reliable cgroup empty notifications: we try to use
1683          * SIGCHLD as replacement. */
1684
1685         if (!u->cgroup_path)
1686                 return -ENOENT;
1687
1688         r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1689         if (r < 0)
1690                 return r;
1691         if (r > 0) /* On unified we can use proper notifications */
1692                 return 0;
1693
1694         return unit_watch_pids_in_path(u, u->cgroup_path);
1695 }
1696
1697 int unit_notify_cgroup_empty(Unit *u) {
1698         int r;
1699
1700         assert(u);
1701
1702         if (!u->cgroup_path)
1703                 return 0;
1704
1705         r = cg_is_empty_recursive(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path);
1706         if (r <= 0)
1707                 return r;
1708
1709         unit_add_to_gc_queue(u);
1710
1711         if (UNIT_VTABLE(u)->notify_cgroup_empty)
1712                 UNIT_VTABLE(u)->notify_cgroup_empty(u);
1713
1714         return 0;
1715 }
1716
1717 static int on_cgroup_inotify_event(sd_event_source *s, int fd, uint32_t revents, void *userdata) {
1718         Manager *m = userdata;
1719
1720         assert(s);
1721         assert(fd >= 0);
1722         assert(m);
1723
1724         for (;;) {
1725                 union inotify_event_buffer buffer;
1726                 struct inotify_event *e;
1727                 ssize_t l;
1728
1729                 l = read(fd, &buffer, sizeof(buffer));
1730                 if (l < 0) {
1731                         if (errno == EINTR || errno == EAGAIN)
1732                                 return 0;
1733
1734                         return log_error_errno(errno, "Failed to read control group inotify events: %m");
1735                 }
1736
1737                 FOREACH_INOTIFY_EVENT(e, buffer, l) {
1738                         Unit *u;
1739
1740                         if (e->wd < 0)
1741                                 /* Queue overflow has no watch descriptor */
1742                                 continue;
1743
1744                         if (e->mask & IN_IGNORED)
1745                                 /* The watch was just removed */
1746                                 continue;
1747
1748                         u = hashmap_get(m->cgroup_inotify_wd_unit, INT_TO_PTR(e->wd));
1749                         if (!u) /* Not that inotify might deliver
1750                                  * events for a watch even after it
1751                                  * was removed, because it was queued
1752                                  * before the removal. Let's ignore
1753                                  * this here safely. */
1754                                 continue;
1755
1756                         (void) unit_notify_cgroup_empty(u);
1757                 }
1758         }
1759 }
1760 #endif // 0
1761
1762 int manager_setup_cgroup(Manager *m) {
1763         _cleanup_free_ char *path = NULL;
1764         CGroupController c;
1765         int r, all_unified;
1766         char *e;
1767
1768         assert(m);
1769
1770         /* 1. Determine hierarchy */
1771         m->cgroup_root = mfree(m->cgroup_root);
1772         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 0, &m->cgroup_root);
1773         if (r < 0)
1774                 return log_error_errno(r, "Cannot determine cgroup we are running in: %m");
1775
1776 #if 0 /// elogind does not support systemd scopes and slices
1777         /* Chop off the init scope, if we are already located in it */
1778         e = endswith(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
1779
1780         /* LEGACY: Also chop off the system slice if we are in
1781          * it. This is to support live upgrades from older systemd
1782          * versions where PID 1 was moved there. Also see
1783          * cg_get_root_path(). */
1784         if (!e && MANAGER_IS_SYSTEM(m)) {
1785                 e = endswith(m->cgroup_root, "/" SPECIAL_SYSTEM_SLICE);
1786                 if (!e)
1787                         e = endswith(m->cgroup_root, "/system"); /* even more legacy */
1788         }
1789         if (e)
1790                 *e = 0;
1791 #endif // 0
1792
1793         /* And make sure to store away the root value without trailing
1794          * slash, even for the root dir, so that we can easily prepend
1795          * it everywhere. */
1796         while ((e = endswith(m->cgroup_root, "/")))
1797                 *e = 0;
1798         log_debug_elogind("Cgroup Controller \"%s\" -> root \"%s\"",
1799                           SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root);
1800
1801         /* 2. Show data */
1802         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, NULL, &path);
1803         if (r < 0)
1804                 return log_error_errno(r, "Cannot find cgroup mount point: %m");
1805
1806         r = cg_unified_flush();
1807         if (r < 0)
1808                 return log_error_errno(r, "Couldn't determine if we are running in the unified hierarchy: %m");
1809
1810         all_unified = cg_all_unified();
1811         if (r < 0)
1812                 return log_error_errno(r, "Couldn't determine whether we are in all unified mode: %m");
1813         if (r > 0)
1814                 log_debug("Unified cgroup hierarchy is located at %s.", path);
1815         else {
1816                 r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1817                 if (r < 0)
1818                         return log_error_errno(r, "Failed to determine whether systemd's own controller is in unified mode: %m");
1819                 if (r > 0)
1820                         log_debug("Unified cgroup hierarchy is located at %s. Controllers are on legacy hierarchies.", path);
1821                 else
1822                         log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER_LEGACY ". File system hierarchy is at %s.", path);
1823         }
1824
1825         if (!m->test_run) {
1826                 const char *scope_path;
1827
1828                 /* 3. Install agent */
1829                 if (cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) > 0) {
1830
1831                         /* In the unified hierarchy we can get
1832                          * cgroup empty notifications via inotify. */
1833
1834 #if 0 /// elogind does not support the unified hierarchy, yet.
1835                         m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
1836                         safe_close(m->cgroup_inotify_fd);
1837
1838                         m->cgroup_inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
1839                         if (m->cgroup_inotify_fd < 0)
1840                                 return log_error_errno(errno, "Failed to create control group inotify object: %m");
1841
1842                         r = sd_event_add_io(m->event, &m->cgroup_inotify_event_source, m->cgroup_inotify_fd, EPOLLIN, on_cgroup_inotify_event, m);
1843                         if (r < 0)
1844                                 return log_error_errno(r, "Failed to watch control group inotify object: %m");
1845
1846                         /* Process cgroup empty notifications early, but after service notifications and SIGCHLD. Also
1847                          * see handling of cgroup agent notifications, for the classic cgroup hierarchy support. */
1848                         r = sd_event_source_set_priority(m->cgroup_inotify_event_source, SD_EVENT_PRIORITY_NORMAL-5);
1849                         if (r < 0)
1850                                 return log_error_errno(r, "Failed to set priority of inotify event source: %m");
1851
1852                         (void) sd_event_source_set_description(m->cgroup_inotify_event_source, "cgroup-inotify");
1853
1854 #else
1855                         return log_error_errno(EOPNOTSUPP, "Unified cgroup hierarchy not supported: %m");
1856 #endif // 0
1857                 } else if (MANAGER_IS_SYSTEM(m)) {
1858
1859                         /* On the legacy hierarchy we only get
1860                          * notifications via cgroup agents. (Which
1861                          * isn't really reliable, since it does not
1862                          * generate events when control groups with
1863                          * children run empty. */
1864
1865                         r = cg_install_release_agent(SYSTEMD_CGROUP_CONTROLLER, SYSTEMD_CGROUP_AGENT_PATH);
1866                         if (r < 0)
1867                                 log_warning_errno(r, "Failed to install release agent, ignoring: %m");
1868                         else if (r > 0)
1869                                 log_debug("Installed release agent.");
1870                         else if (r == 0)
1871                                 log_debug("Release agent already installed.");
1872                 }
1873
1874 #if 0 /// elogind is not meant to run in systemd init scope
1875                 /* 4. Make sure we are in the special "init.scope" unit in the root slice. */
1876                 scope_path = strjoina(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
1877                 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1878 #else
1879                 if (streq(SYSTEMD_CGROUP_CONTROLLER, "name=elogind"))
1880                         // we are our own cgroup controller
1881                         scope_path = strjoina("");
1882                 else if (streq(m->cgroup_root, "/elogind"))
1883                         // root already is our cgroup
1884                         scope_path = strjoina(m->cgroup_root);
1885                 else
1886                         // we have to create our own group
1887                         scope_path = strjoina(m->cgroup_root, "/elogind");
1888                 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1889 #endif // 0
1890                 if (r < 0)
1891                         return log_error_errno(r, "Failed to create %s control group: %m", scope_path);
1892                 log_debug_elogind("Created control group \"%s\"", scope_path);
1893
1894                 /* also, move all other userspace processes remaining
1895                  * in the root cgroup into that scope. */
1896                 r = cg_migrate(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1897                 if (r < 0)
1898                         log_warning_errno(r, "Couldn't move remaining userspace processes, ignoring: %m");
1899
1900                 /* 5. And pin it, so that it cannot be unmounted */
1901                 safe_close(m->pin_cgroupfs_fd);
1902                 m->pin_cgroupfs_fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_NOCTTY|O_NONBLOCK);
1903                 if (m->pin_cgroupfs_fd < 0)
1904                         return log_error_errno(errno, "Failed to open pin file: %m");
1905
1906                 /* 6.  Always enable hierarchical support if it exists... */
1907                 if (!all_unified)
1908                         (void) cg_set_attribute("memory", "/", "memory.use_hierarchy", "1");
1909         }
1910
1911         /* 7. Figure out which controllers are supported */
1912         r = cg_mask_supported(&m->cgroup_supported);
1913         if (r < 0)
1914                 return log_error_errno(r, "Failed to determine supported controllers: %m");
1915
1916         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++)
1917                 log_debug("Controller '%s' supported: %s", cgroup_controller_to_string(c), yes_no(m->cgroup_supported & CGROUP_CONTROLLER_TO_MASK(c)));
1918
1919         return 0;
1920 }
1921
1922 void manager_shutdown_cgroup(Manager *m, bool delete) {
1923         assert(m);
1924
1925         /* We can't really delete the group, since we are in it. But
1926          * let's trim it. */
1927         if (delete && m->cgroup_root)
1928                 (void) cg_trim(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, false);
1929
1930 #if 0 /// elogind does not support the unified hierarchy, yet.
1931         m->cgroup_inotify_wd_unit = hashmap_free(m->cgroup_inotify_wd_unit);
1932
1933         m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
1934         m->cgroup_inotify_fd = safe_close(m->cgroup_inotify_fd);
1935 #endif // 0
1936
1937         m->pin_cgroupfs_fd = safe_close(m->pin_cgroupfs_fd);
1938
1939         m->cgroup_root = mfree(m->cgroup_root);
1940 }
1941
1942 #if 0 /// UNNEEDED by elogind
1943 Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
1944         char *p;
1945         Unit *u;
1946
1947         assert(m);
1948         assert(cgroup);
1949
1950         u = hashmap_get(m->cgroup_unit, cgroup);
1951         if (u)
1952                 return u;
1953
1954         p = strdupa(cgroup);
1955         for (;;) {
1956                 char *e;
1957
1958                 e = strrchr(p, '/');
1959                 if (!e || e == p)
1960                         return hashmap_get(m->cgroup_unit, SPECIAL_ROOT_SLICE);
1961
1962                 *e = 0;
1963
1964                 u = hashmap_get(m->cgroup_unit, p);
1965                 if (u)
1966                         return u;
1967         }
1968 }
1969
1970 Unit *manager_get_unit_by_pid_cgroup(Manager *m, pid_t pid) {
1971         _cleanup_free_ char *cgroup = NULL;
1972         int r;
1973
1974         assert(m);
1975
1976         if (pid <= 0)
1977                 return NULL;
1978
1979         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
1980         if (r < 0)
1981                 return NULL;
1982
1983         return manager_get_unit_by_cgroup(m, cgroup);
1984 }
1985
1986 Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
1987         Unit *u;
1988
1989         assert(m);
1990
1991         if (pid <= 0)
1992                 return NULL;
1993
1994         if (pid == 1)
1995                 return hashmap_get(m->units, SPECIAL_INIT_SCOPE);
1996
1997         u = hashmap_get(m->watch_pids1, PID_TO_PTR(pid));
1998         if (u)
1999                 return u;
2000
2001         u = hashmap_get(m->watch_pids2, PID_TO_PTR(pid));
2002         if (u)
2003                 return u;
2004
2005         return manager_get_unit_by_pid_cgroup(m, pid);
2006 }
2007 #endif // 0
2008
2009 #if 0 /// elogind must substitute this with its own variant
2010 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
2011         Unit *u;
2012
2013         assert(m);
2014         assert(cgroup);
2015
2016         log_debug("Got cgroup empty notification for: %s", cgroup);
2017
2018         u = manager_get_unit_by_cgroup(m, cgroup);
2019         if (!u)
2020                 return 0;
2021
2022         return unit_notify_cgroup_empty(u);
2023 }
2024 #else
2025 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
2026         Session *s;
2027
2028         assert(m);
2029         assert(cgroup);
2030
2031         log_debug("Got cgroup empty notification for: %s", cgroup);
2032
2033         s = hashmap_get(m->sessions, cgroup);
2034
2035         if (s) {
2036                 session_finalize(s);
2037                 session_free(s);
2038         } else
2039                 log_warning("Session not found: %s", cgroup);
2040
2041         return 0;
2042 }
2043 #endif // 0
2044 #if 0 /// UNNEEDED by elogind
2045 int unit_get_memory_current(Unit *u, uint64_t *ret) {
2046         _cleanup_free_ char *v = NULL;
2047         int r;
2048
2049         assert(u);
2050         assert(ret);
2051
2052         if (!u->cgroup_path)
2053                 return -ENODATA;
2054
2055         if ((u->cgroup_realized_mask & CGROUP_MASK_MEMORY) == 0)
2056                 return -ENODATA;
2057
2058         r = cg_all_unified();
2059         if (r < 0)
2060                 return r;
2061         if (r > 0)
2062                 r = cg_get_attribute("memory", u->cgroup_path, "memory.current", &v);
2063         else
2064                 r = cg_get_attribute("memory", u->cgroup_path, "memory.usage_in_bytes", &v);
2065         if (r == -ENOENT)
2066                 return -ENODATA;
2067         if (r < 0)
2068                 return r;
2069
2070         return safe_atou64(v, ret);
2071 }
2072
2073 int unit_get_tasks_current(Unit *u, uint64_t *ret) {
2074         _cleanup_free_ char *v = NULL;
2075         int r;
2076
2077         assert(u);
2078         assert(ret);
2079
2080         if (!u->cgroup_path)
2081                 return -ENODATA;
2082
2083         if ((u->cgroup_realized_mask & CGROUP_MASK_PIDS) == 0)
2084                 return -ENODATA;
2085
2086         r = cg_get_attribute("pids", u->cgroup_path, "pids.current", &v);
2087         if (r == -ENOENT)
2088                 return -ENODATA;
2089         if (r < 0)
2090                 return r;
2091
2092         return safe_atou64(v, ret);
2093 }
2094
2095 static int unit_get_cpu_usage_raw(Unit *u, nsec_t *ret) {
2096         _cleanup_free_ char *v = NULL;
2097         uint64_t ns;
2098         int r;
2099
2100         assert(u);
2101         assert(ret);
2102
2103         if (!u->cgroup_path)
2104                 return -ENODATA;
2105
2106         r = cg_all_unified();
2107         if (r < 0)
2108                 return r;
2109         if (r > 0) {
2110                 const char *keys[] = { "usage_usec", NULL };
2111                 _cleanup_free_ char *val = NULL;
2112                 uint64_t us;
2113
2114                 if ((u->cgroup_realized_mask & CGROUP_MASK_CPU) == 0)
2115                         return -ENODATA;
2116
2117                 r = cg_get_keyed_attribute("cpu", u->cgroup_path, "cpu.stat", keys, &val);
2118                 if (r < 0)
2119                         return r;
2120
2121                 r = safe_atou64(val, &us);
2122                 if (r < 0)
2123                         return r;
2124
2125                 ns = us * NSEC_PER_USEC;
2126         } else {
2127                 if ((u->cgroup_realized_mask & CGROUP_MASK_CPUACCT) == 0)
2128                         return -ENODATA;
2129
2130                 r = cg_get_attribute("cpuacct", u->cgroup_path, "cpuacct.usage", &v);
2131                 if (r == -ENOENT)
2132                         return -ENODATA;
2133                 if (r < 0)
2134                         return r;
2135
2136                 r = safe_atou64(v, &ns);
2137                 if (r < 0)
2138                         return r;
2139         }
2140
2141         *ret = ns;
2142         return 0;
2143 }
2144
2145 int unit_get_cpu_usage(Unit *u, nsec_t *ret) {
2146         nsec_t ns;
2147         int r;
2148
2149         assert(u);
2150
2151         /* Retrieve the current CPU usage counter. This will subtract the CPU counter taken when the unit was
2152          * started. If the cgroup has been removed already, returns the last cached value. To cache the value, simply
2153          * call this function with a NULL return value. */
2154
2155         r = unit_get_cpu_usage_raw(u, &ns);
2156         if (r == -ENODATA && u->cpu_usage_last != NSEC_INFINITY) {
2157                 /* If we can't get the CPU usage anymore (because the cgroup was already removed, for example), use our
2158                  * cached value. */
2159
2160                 if (ret)
2161                         *ret = u->cpu_usage_last;
2162                 return 0;
2163         }
2164         if (r < 0)
2165                 return r;
2166
2167         if (ns > u->cpu_usage_base)
2168                 ns -= u->cpu_usage_base;
2169         else
2170                 ns = 0;
2171
2172         u->cpu_usage_last = ns;
2173         if (ret)
2174                 *ret = ns;
2175
2176         return 0;
2177 }
2178
2179 int unit_reset_cpu_usage(Unit *u) {
2180         nsec_t ns;
2181         int r;
2182
2183         assert(u);
2184
2185         u->cpu_usage_last = NSEC_INFINITY;
2186
2187         r = unit_get_cpu_usage_raw(u, &ns);
2188         if (r < 0) {
2189                 u->cpu_usage_base = 0;
2190                 return r;
2191         }
2192
2193         u->cpu_usage_base = ns;
2194         return 0;
2195 }
2196
2197 bool unit_cgroup_delegate(Unit *u) {
2198         CGroupContext *c;
2199
2200         assert(u);
2201
2202         c = unit_get_cgroup_context(u);
2203         if (!c)
2204                 return false;
2205
2206         return c->delegate;
2207 }
2208
2209 void unit_invalidate_cgroup(Unit *u, CGroupMask m) {
2210         assert(u);
2211
2212         if (!UNIT_HAS_CGROUP_CONTEXT(u))
2213                 return;
2214
2215         if (m == 0)
2216                 return;
2217
2218         /* always invalidate compat pairs together */
2219         if (m & (CGROUP_MASK_IO | CGROUP_MASK_BLKIO))
2220                 m |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
2221
2222         if ((u->cgroup_realized_mask & m) == 0)
2223                 return;
2224
2225         u->cgroup_realized_mask &= ~m;
2226         unit_add_to_cgroup_queue(u);
2227 }
2228
2229 void manager_invalidate_startup_units(Manager *m) {
2230         Iterator i;
2231         Unit *u;
2232
2233         assert(m);
2234
2235         SET_FOREACH(u, m->startup_units, i)
2236                 unit_invalidate_cgroup(u, CGROUP_MASK_CPU|CGROUP_MASK_IO|CGROUP_MASK_BLKIO);
2237 }
2238
2239 static const char* const cgroup_device_policy_table[_CGROUP_DEVICE_POLICY_MAX] = {
2240         [CGROUP_AUTO] = "auto",
2241         [CGROUP_CLOSED] = "closed",
2242         [CGROUP_STRICT] = "strict",
2243 };
2244
2245 DEFINE_STRING_TABLE_LOOKUP(cgroup_device_policy, CGroupDevicePolicy);
2246 #endif // 0