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