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_shares)
 692                                 shares = cgroup_context_cpu_shares(c, state);
 693                         else if (has_weight) {
 694                                 uint64_t weight = cgroup_context_cpu_weight(c, state);
 695
 696                                 shares = cgroup_cpu_weight_to_shares(weight);
 697
 698                                 log_cgroup_compat(u, "Applying [Startup]CpuWeight %" PRIu64 " as [Startup]CpuShares %" PRIu64 " on %s",
 699                                                   weight, shares, path);
 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_blockio)
 793                                 weight = cgroup_context_blkio_weight(c, state);
 794                         else if (has_io) {
 795                                 uint64_t io_weight = cgroup_context_io_weight(c, state);
 796
 797                                 weight = cgroup_weight_io_to_blkio(cgroup_context_io_weight(c, state));
 798
 799                                 log_cgroup_compat(u, "Applying [Startup]IOWeight %" PRIu64 " as [Startup]BlockIOWeight %" PRIu64,
 800                                                   io_weight, weight);
 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_blockio) {
 811                                 CGroupBlockIODeviceWeight *w;
 812
 813                                 /* FIXME: no way to reset this list */
 814                                 LIST_FOREACH(device_weights, w, c->blockio_device_weights)
 815                                         cgroup_apply_blkio_device_weight(u, w->path, w->weight);
 816                         } else if (has_io) {
 817                                 CGroupIODeviceWeight *w;
 818
 819                                 /* FIXME: no way to reset this list */
 820                                 LIST_FOREACH(device_weights, w, c->io_device_weights) {
 821                                         weight = cgroup_weight_io_to_blkio(w->weight);
 822
 823                                         log_cgroup_compat(u, "Applying IODeviceWeight %" PRIu64 " as BlockIODeviceWeight %" PRIu64 " for %s",
 824                                                           w->weight, weight, w->path);
 825
 826                                         cgroup_apply_blkio_device_weight(u, w->path, weight);
 827                                 }
 828                         }
 829                 }
 830
 831                 /* Apply limits and free ones without config. */
 832                 if (has_blockio) {
 833                         CGroupBlockIODeviceBandwidth *b, *next;
 834
 835                         LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths) {
 836                                 if (!cgroup_apply_blkio_device_limit(u, b->path, b->rbps, b->wbps))
 837                                         cgroup_context_free_blockio_device_bandwidth(c, b);
 838                         }
 839                 } else if (has_io) {
 840                         CGroupIODeviceLimit *l, *next;
 841
 842                         LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
 843                                 log_cgroup_compat(u, "Applying IO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as BlockIO{Read|Write}BandwidthMax for %s",
 844                                                   l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX], l->path);
 845
 846                                 if (!cgroup_apply_blkio_device_limit(u, l->path, l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX]))
 847                                         cgroup_context_free_io_device_limit(c, l);
 848                         }
 849                 }
 850         }
 851
 852         if ((mask & CGROUP_MASK_MEMORY) && !is_root) {
 853                 if (cg_all_unified() > 0) {
 854                         uint64_t max = c->memory_max;
 855                         uint64_t swap_max = c->memory_swap_max;
 856
 857                         if (cgroup_context_has_unified_memory_config(c)) {
 858                                 max = c->memory_max;
 859                                 swap_max = c->memory_swap_max;
 860                         } else {
 861                                 max = c->memory_limit;
 862
 863                                 if (max != CGROUP_LIMIT_MAX)
 864                                         log_cgroup_compat(u, "Applying MemoryLimit %" PRIu64 " as MemoryMax", max);
 865                         }
 866
 867                         cgroup_apply_unified_memory_limit(u, "memory.low", c->memory_low);
 868                         cgroup_apply_unified_memory_limit(u, "memory.high", c->memory_high);
 869                         cgroup_apply_unified_memory_limit(u, "memory.max", max);
 870                         cgroup_apply_unified_memory_limit(u, "memory.swap.max", swap_max);
 871                 } else {
 872                         char buf[DECIMAL_STR_MAX(uint64_t) + 1];
 873                         uint64_t val = c->memory_limit;
 874
 875                         if (val == CGROUP_LIMIT_MAX) {
 876                                 val = c->memory_max;
 877
 878                                 if (val != CGROUP_LIMIT_MAX)
 879                                         log_cgroup_compat(u, "Applying MemoryMax %" PRIi64 " as MemoryLimit", c->memory_max);
 880                         }
 881
 882                         if (val == CGROUP_LIMIT_MAX)
 883                                 strncpy(buf, "-1\n", sizeof(buf));
 884                         else
 885                                 xsprintf(buf, "%" PRIu64 "\n", val);
 886
 887                         r = cg_set_attribute("memory", path, "memory.limit_in_bytes", buf);
 888                         if (r < 0)
 889                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 890                                               "Failed to set memory.limit_in_bytes: %m");
 891                 }
 892         }
 893
 894         if ((mask & CGROUP_MASK_DEVICES) && !is_root) {
 895                 CGroupDeviceAllow *a;
 896
 897                 /* Changing the devices list of a populated cgroup
 898                  * might result in EINVAL, hence ignore EINVAL
 899                  * here. */
 900
 901                 if (c->device_allow || c->device_policy != CGROUP_AUTO)
 902                         r = cg_set_attribute("devices", path, "devices.deny", "a");
 903                 else
 904                         r = cg_set_attribute("devices", path, "devices.allow", "a");
 905                 if (r < 0)
 906                         log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 907                                       "Failed to reset devices.list: %m");
 908
 909                 if (c->device_policy == CGROUP_CLOSED ||
 910                     (c->device_policy == CGROUP_AUTO && c->device_allow)) {
 911                         static const char auto_devices[] =
 912                                 "/dev/null\0" "rwm\0"
 913                                 "/dev/zero\0" "rwm\0"
 914                                 "/dev/full\0" "rwm\0"
 915                                 "/dev/random\0" "rwm\0"
 916                                 "/dev/urandom\0" "rwm\0"
 917                                 "/dev/tty\0" "rwm\0"
 918                                 "/dev/pts/ptmx\0" "rw\0" /* /dev/pts/ptmx may not be duplicated, but accessed */
 919                                 /* Allow /run/elogind/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
 920                                 /* Allow /run/systemd/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
 921                                 "/run/systemd/inaccessible/chr\0" "rwm\0"
 922                                 "/run/systemd/inaccessible/blk\0" "rwm\0";
 923
 924                         const char *x, *y;
 925
 926                         NULSTR_FOREACH_PAIR(x, y, auto_devices)
 927                                 whitelist_device(path, x, y);
 928
 929                         whitelist_major(path, "pts", 'c', "rw");
 930                         whitelist_major(path, "kdbus", 'c', "rw");
 931                         whitelist_major(path, "kdbus/*", 'c', "rw");
 932                 }
 933
 934                 LIST_FOREACH(device_allow, a, c->device_allow) {
 935                         char acc[4];
 936                         unsigned k = 0;
 937
 938                         if (a->r)
 939                                 acc[k++] = 'r';
 940                         if (a->w)
 941                                 acc[k++] = 'w';
 942                         if (a->m)
 943                                 acc[k++] = 'm';
 944
 945                         if (k == 0)
 946                                 continue;
 947
 948                         acc[k++] = 0;
 949
 950                         if (startswith(a->path, "/dev/"))
 951                                 whitelist_device(path, a->path, acc);
 952                         else if (startswith(a->path, "block-"))
 953                                 whitelist_major(path, a->path + 6, 'b', acc);
 954                         else if (startswith(a->path, "char-"))
 955                                 whitelist_major(path, a->path + 5, 'c', acc);
 956                         else
 957                                 log_unit_debug(u, "Ignoring device %s while writing cgroup attribute.", a->path);
 958                 }
 959         }
 960
 961         if ((mask & CGROUP_MASK_PIDS) && !is_root) {
 962
 963                 if (c->tasks_max != CGROUP_LIMIT_MAX) {
 964                         char buf[DECIMAL_STR_MAX(uint64_t) + 2];
 965
 966                         sprintf(buf, "%" PRIu64 "\n", c->tasks_max);
 967                         r = cg_set_attribute("pids", path, "pids.max", buf);
 968                 } else
 969                         r = cg_set_attribute("pids", path, "pids.max", "max");
 970
 971                 if (r < 0)
 972                         log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 973                                       "Failed to set pids.max: %m");
 974         }
 975 }
 976
 977 CGroupMask cgroup_context_get_mask(CGroupContext *c) {
 978         CGroupMask mask = 0;
 979
 980         /* Figure out which controllers we need */
 981
 982         if (c->cpu_accounting ||
 983             cgroup_context_has_cpu_weight(c) ||
 984             cgroup_context_has_cpu_shares(c) ||
 985             c->cpu_quota_per_sec_usec != USEC_INFINITY)
 986                 mask |= CGROUP_MASK_CPUACCT | CGROUP_MASK_CPU;
 987
 988         if (cgroup_context_has_io_config(c) || cgroup_context_has_blockio_config(c))
 989                 mask |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
 990
 991         if (c->memory_accounting ||
 992             c->memory_limit != CGROUP_LIMIT_MAX ||
 993             cgroup_context_has_unified_memory_config(c))
 994                 mask |= CGROUP_MASK_MEMORY;
 995
 996         if (c->device_allow ||
 997             c->device_policy != CGROUP_AUTO)
 998                 mask |= CGROUP_MASK_DEVICES;
 999
1000         if (c->tasks_accounting ||
1001             c->tasks_max != (uint64_t) -1)
1002                 mask |= CGROUP_MASK_PIDS;
1003
1004         return mask;
1005 }
1006
1007 CGroupMask unit_get_own_mask(Unit *u) {
1008         CGroupContext *c;
1009
1010         /* Returns the mask of controllers the unit needs for itself */
1011
1012         c = unit_get_cgroup_context(u);
1013         if (!c)
1014                 return 0;
1015
1016         /* If delegation is turned on, then turn on all cgroups,
1017          * unless we are on the legacy hierarchy and the process we
1018          * fork into it is known to drop privileges, and hence
1019          * shouldn't get access to the controllers.
1020          *
1021          * Note that on the unified hierarchy it is safe to delegate
1022          * controllers to unprivileged services. */
1023
1024         if (c->delegate) {
1025                 ExecContext *e;
1026
1027                 e = unit_get_exec_context(u);
1028                 if (!e ||
1029                     exec_context_maintains_privileges(e) ||
1030                     cg_all_unified() > 0)
1031                         return _CGROUP_MASK_ALL;
1032         }
1033
1034         return cgroup_context_get_mask(c);
1035 }
1036
1037 CGroupMask unit_get_members_mask(Unit *u) {
1038         assert(u);
1039
1040         /* Returns the mask of controllers all of the unit's children
1041          * require, merged */
1042
1043         if (u->cgroup_members_mask_valid)
1044                 return u->cgroup_members_mask;
1045
1046         u->cgroup_members_mask = 0;
1047
1048         if (u->type == UNIT_SLICE) {
1049                 Unit *member;
1050                 Iterator i;
1051
1052                 SET_FOREACH(member, u->dependencies[UNIT_BEFORE], i) {
1053
1054                         if (member == u)
1055                                 continue;
1056
1057                         if (UNIT_DEREF(member->slice) != u)
1058                                 continue;
1059
1060                         u->cgroup_members_mask |=
1061                                 unit_get_own_mask(member) |
1062                                 unit_get_members_mask(member);
1063                 }
1064         }
1065
1066         u->cgroup_members_mask_valid = true;
1067         return u->cgroup_members_mask;
1068 }
1069
1070 CGroupMask unit_get_siblings_mask(Unit *u) {
1071         assert(u);
1072
1073         /* Returns the mask of controllers all of the unit's siblings
1074          * require, i.e. the members mask of the unit's parent slice
1075          * if there is one. */
1076
1077         if (UNIT_ISSET(u->slice))
1078                 return unit_get_members_mask(UNIT_DEREF(u->slice));
1079
1080         return unit_get_own_mask(u) | unit_get_members_mask(u);
1081 }
1082
1083 CGroupMask unit_get_subtree_mask(Unit *u) {
1084
1085         /* Returns the mask of this subtree, meaning of the group
1086          * itself and its children. */
1087
1088         return unit_get_own_mask(u) | unit_get_members_mask(u);
1089 }
1090
1091 CGroupMask unit_get_target_mask(Unit *u) {
1092         CGroupMask mask;
1093
1094         /* This returns the cgroup mask of all controllers to enable
1095          * for a specific cgroup, i.e. everything it needs itself,
1096          * plus all that its children need, plus all that its siblings
1097          * need. This is primarily useful on the legacy cgroup
1098          * hierarchy, where we need to duplicate each cgroup in each
1099          * hierarchy that shall be enabled for it. */
1100
1101         mask = unit_get_own_mask(u) | unit_get_members_mask(u) | unit_get_siblings_mask(u);
1102         mask &= u->manager->cgroup_supported;
1103
1104         return mask;
1105 }
1106
1107 CGroupMask unit_get_enable_mask(Unit *u) {
1108         CGroupMask mask;
1109
1110         /* This returns the cgroup mask of all controllers to enable
1111          * for the children of a specific cgroup. This is primarily
1112          * useful for the unified cgroup hierarchy, where each cgroup
1113          * controls which controllers are enabled for its children. */
1114
1115         mask = unit_get_members_mask(u);
1116         mask &= u->manager->cgroup_supported;
1117
1118         return mask;
1119 }
1120
1121 /* Recurse from a unit up through its containing slices, propagating
1122  * mask bits upward. A unit is also member of itself. */
1123 void unit_update_cgroup_members_masks(Unit *u) {
1124         CGroupMask m;
1125         bool more;
1126
1127         assert(u);
1128
1129         /* Calculate subtree mask */
1130         m = unit_get_subtree_mask(u);
1131
1132         /* See if anything changed from the previous invocation. If
1133          * not, we're done. */
1134         if (u->cgroup_subtree_mask_valid && m == u->cgroup_subtree_mask)
1135                 return;
1136
1137         more =
1138                 u->cgroup_subtree_mask_valid &&
1139                 ((m & ~u->cgroup_subtree_mask) != 0) &&
1140                 ((~m & u->cgroup_subtree_mask) == 0);
1141
1142         u->cgroup_subtree_mask = m;
1143         u->cgroup_subtree_mask_valid = true;
1144
1145         if (UNIT_ISSET(u->slice)) {
1146                 Unit *s = UNIT_DEREF(u->slice);
1147
1148                 if (more)
1149                         /* There's more set now than before. We
1150                          * propagate the new mask to the parent's mask
1151                          * (not caring if it actually was valid or
1152                          * not). */
1153
1154                         s->cgroup_members_mask |= m;
1155
1156                 else
1157                         /* There's less set now than before (or we
1158                          * don't know), we need to recalculate
1159                          * everything, so let's invalidate the
1160                          * parent's members mask */
1161
1162                         s->cgroup_members_mask_valid = false;
1163
1164                 /* And now make sure that this change also hits our
1165                  * grandparents */
1166                 unit_update_cgroup_members_masks(s);
1167         }
1168 }
1169
1170 static const char *migrate_callback(CGroupMask mask, void *userdata) {
1171         Unit *u = userdata;
1172
1173         assert(mask != 0);
1174         assert(u);
1175
1176         while (u) {
1177                 if (u->cgroup_path &&
1178                     u->cgroup_realized &&
1179                     (u->cgroup_realized_mask & mask) == mask)
1180                         return u->cgroup_path;
1181
1182                 u = UNIT_DEREF(u->slice);
1183         }
1184
1185         return NULL;
1186 }
1187
1188 char *unit_default_cgroup_path(Unit *u) {
1189         _cleanup_free_ char *escaped = NULL, *slice = NULL;
1190         int r;
1191
1192         assert(u);
1193
1194         if (unit_has_name(u, SPECIAL_ROOT_SLICE))
1195                 return strdup(u->manager->cgroup_root);
1196
1197         if (UNIT_ISSET(u->slice) && !unit_has_name(UNIT_DEREF(u->slice), SPECIAL_ROOT_SLICE)) {
1198                 r = cg_slice_to_path(UNIT_DEREF(u->slice)->id, &slice);
1199                 if (r < 0)
1200                         return NULL;
1201         }
1202
1203         escaped = cg_escape(u->id);
1204         if (!escaped)
1205                 return NULL;
1206
1207         if (slice)
1208                 return strjoin(u->manager->cgroup_root, "/", slice, "/", escaped, NULL);
1209         else
1210                 return strjoin(u->manager->cgroup_root, "/", escaped, NULL);
1211 }
1212
1213 int unit_set_cgroup_path(Unit *u, const char *path) {
1214         _cleanup_free_ char *p = NULL;
1215         int r;
1216
1217         assert(u);
1218
1219         if (path) {
1220                 p = strdup(path);
1221                 if (!p)
1222                         return -ENOMEM;
1223         } else
1224                 p = NULL;
1225
1226         if (streq_ptr(u->cgroup_path, p))
1227                 return 0;
1228
1229         if (p) {
1230                 r = hashmap_put(u->manager->cgroup_unit, p, u);
1231                 if (r < 0)
1232                         return r;
1233         }
1234
1235         unit_release_cgroup(u);
1236
1237         u->cgroup_path = p;
1238         p = NULL;
1239
1240         return 1;
1241 }
1242
1243 int unit_watch_cgroup(Unit *u) {
1244         _cleanup_free_ char *events = NULL;
1245         int r;
1246
1247         assert(u);
1248
1249         if (!u->cgroup_path)
1250                 return 0;
1251
1252         if (u->cgroup_inotify_wd >= 0)
1253                 return 0;
1254
1255         /* Only applies to the unified hierarchy */
1256         r = cg_unified(SYSTEMD_CGROUP_CONTROLLER);
1257         if (r < 0)
1258                 return log_unit_error_errno(u, r, "Failed detect whether the unified hierarchy is used: %m");
1259         if (r == 0)
1260                 return 0;
1261
1262         /* Don't watch the root slice, it's pointless. */
1263         if (unit_has_name(u, SPECIAL_ROOT_SLICE))
1264                 return 0;
1265
1266         r = hashmap_ensure_allocated(&u->manager->cgroup_inotify_wd_unit, &trivial_hash_ops);
1267         if (r < 0)
1268                 return log_oom();
1269
1270         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, "cgroup.events", &events);
1271         if (r < 0)
1272                 return log_oom();
1273
1274         u->cgroup_inotify_wd = inotify_add_watch(u->manager->cgroup_inotify_fd, events, IN_MODIFY);
1275         if (u->cgroup_inotify_wd < 0) {
1276
1277                 if (errno == ENOENT) /* If the directory is already
1278                                       * gone we don't need to track
1279                                       * it, so this is not an error */
1280                         return 0;
1281
1282                 return log_unit_error_errno(u, errno, "Failed to add inotify watch descriptor for control group %s: %m", u->cgroup_path);
1283         }
1284
1285         r = hashmap_put(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd), u);
1286         if (r < 0)
1287                 return log_unit_error_errno(u, r, "Failed to add inotify watch descriptor to hash map: %m");
1288
1289         return 0;
1290 }
1291
1292 static int unit_create_cgroup(
1293                 Unit *u,
1294                 CGroupMask target_mask,
1295                 CGroupMask enable_mask) {
1296
1297         CGroupContext *c;
1298         int r;
1299
1300         assert(u);
1301
1302         c = unit_get_cgroup_context(u);
1303         if (!c)
1304                 return 0;
1305
1306         if (!u->cgroup_path) {
1307                 _cleanup_free_ char *path = NULL;
1308
1309                 path = unit_default_cgroup_path(u);
1310                 if (!path)
1311                         return log_oom();
1312
1313                 r = unit_set_cgroup_path(u, path);
1314                 if (r == -EEXIST)
1315                         return log_unit_error_errno(u, r, "Control group %s exists already.", path);
1316                 if (r < 0)
1317                         return log_unit_error_errno(u, r, "Failed to set unit's control group path to %s: %m", path);
1318         }
1319
1320         /* First, create our own group */
1321         r = cg_create_everywhere(u->manager->cgroup_supported, target_mask, u->cgroup_path);
1322         if (r < 0)
1323                 return log_unit_error_errno(u, r, "Failed to create cgroup %s: %m", u->cgroup_path);
1324
1325         /* Start watching it */
1326         (void) unit_watch_cgroup(u);
1327
1328         /* Enable all controllers we need */
1329         r = cg_enable_everywhere(u->manager->cgroup_supported, enable_mask, u->cgroup_path);
1330         if (r < 0)
1331                 log_unit_warning_errno(u, r, "Failed to enable controllers on cgroup %s, ignoring: %m", u->cgroup_path);
1332
1333         /* Keep track that this is now realized */
1334         u->cgroup_realized = true;
1335         u->cgroup_realized_mask = target_mask;
1336         u->cgroup_enabled_mask = enable_mask;
1337
1338         if (u->type != UNIT_SLICE && !c->delegate) {
1339
1340                 /* Then, possibly move things over, but not if
1341                  * subgroups may contain processes, which is the case
1342                  * for slice and delegation units. */
1343                 r = cg_migrate_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->cgroup_path, migrate_callback, u);
1344                 if (r < 0)
1345                         log_unit_warning_errno(u, r, "Failed to migrate cgroup from to %s, ignoring: %m", u->cgroup_path);
1346         }
1347
1348         return 0;
1349 }
1350
1351 int unit_attach_pids_to_cgroup(Unit *u) {
1352         int r;
1353         assert(u);
1354
1355         r = unit_realize_cgroup(u);
1356         if (r < 0)
1357                 return r;
1358
1359         r = cg_attach_many_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->pids, migrate_callback, u);
1360         if (r < 0)
1361                 return r;
1362
1363         return 0;
1364 }
1365
1366 static void cgroup_xattr_apply(Unit *u) {
1367         char ids[SD_ID128_STRING_MAX];
1368         int r;
1369
1370         assert(u);
1371
1372         if (!MANAGER_IS_SYSTEM(u->manager))
1373                 return;
1374
1375         if (sd_id128_is_null(u->invocation_id))
1376                 return;
1377
1378         r = cg_set_xattr(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path,
1379                          "trusted.invocation_id",
1380                          sd_id128_to_string(u->invocation_id, ids), 32,
1381                          0);
1382         if (r < 0)
1383                 log_unit_warning_errno(u, r, "Failed to set invocation ID on control group %s, ignoring: %m", u->cgroup_path);
1384 }
1385
1386 static bool unit_has_mask_realized(Unit *u, CGroupMask target_mask, CGroupMask enable_mask) {
1387         assert(u);
1388
1389         return u->cgroup_realized && u->cgroup_realized_mask == target_mask && u->cgroup_enabled_mask == enable_mask;
1390 }
1391
1392 /* Check if necessary controllers and attributes for a unit are in place.
1393  *
1394  * If so, do nothing.
1395  * If not, create paths, move processes over, and set attributes.
1396  *
1397  * Returns 0 on success and < 0 on failure. */
1398 static int unit_realize_cgroup_now(Unit *u, ManagerState state) {
1399         CGroupMask target_mask, enable_mask;
1400         int r;
1401
1402         assert(u);
1403
1404         if (u->in_cgroup_queue) {
1405                 LIST_REMOVE(cgroup_queue, u->manager->cgroup_queue, u);
1406                 u->in_cgroup_queue = false;
1407         }
1408
1409         target_mask = unit_get_target_mask(u);
1410         enable_mask = unit_get_enable_mask(u);
1411
1412         if (unit_has_mask_realized(u, target_mask, enable_mask))
1413                 return 0;
1414
1415         /* First, realize parents */
1416         if (UNIT_ISSET(u->slice)) {
1417                 r = unit_realize_cgroup_now(UNIT_DEREF(u->slice), state);
1418                 if (r < 0)
1419                         return r;
1420         }
1421
1422         /* And then do the real work */
1423         r = unit_create_cgroup(u, target_mask, enable_mask);
1424         if (r < 0)
1425                 return r;
1426
1427         /* Finally, apply the necessary attributes. */
1428         cgroup_context_apply(u, target_mask, state);
1429         cgroup_xattr_apply(u);
1430
1431         return 0;
1432 }
1433
1434 static void unit_add_to_cgroup_queue(Unit *u) {
1435
1436         if (u->in_cgroup_queue)
1437                 return;
1438
1439         LIST_PREPEND(cgroup_queue, u->manager->cgroup_queue, u);
1440         u->in_cgroup_queue = true;
1441 }
1442
1443 unsigned manager_dispatch_cgroup_queue(Manager *m) {
1444         ManagerState state;
1445         unsigned n = 0;
1446         Unit *i;
1447         int r;
1448
1449         state = manager_state(m);
1450
1451         while ((i = m->cgroup_queue)) {
1452                 assert(i->in_cgroup_queue);
1453
1454                 r = unit_realize_cgroup_now(i, state);
1455                 if (r < 0)
1456                         log_warning_errno(r, "Failed to realize cgroups for queued unit %s, ignoring: %m", i->id);
1457
1458                 n++;
1459         }
1460
1461         return n;
1462 }
1463
1464 static void unit_queue_siblings(Unit *u) {
1465         Unit *slice;
1466
1467         /* This adds the siblings of the specified unit and the
1468          * siblings of all parent units to the cgroup queue. (But
1469          * neither the specified unit itself nor the parents.) */
1470
1471         while ((slice = UNIT_DEREF(u->slice))) {
1472                 Iterator i;
1473                 Unit *m;
1474
1475                 SET_FOREACH(m, slice->dependencies[UNIT_BEFORE], i) {
1476                         if (m == u)
1477                                 continue;
1478
1479                         /* Skip units that have a dependency on the slice
1480                          * but aren't actually in it. */
1481                         if (UNIT_DEREF(m->slice) != slice)
1482                                 continue;
1483
1484                         /* No point in doing cgroup application for units
1485                          * without active processes. */
1486                         if (UNIT_IS_INACTIVE_OR_FAILED(unit_active_state(m)))
1487                                 continue;
1488
1489                         /* If the unit doesn't need any new controllers
1490                          * and has current ones realized, it doesn't need
1491                          * any changes. */
1492                         if (unit_has_mask_realized(m, unit_get_target_mask(m), unit_get_enable_mask(m)))
1493                                 continue;
1494
1495                         unit_add_to_cgroup_queue(m);
1496                 }
1497
1498                 u = slice;
1499         }
1500 }
1501
1502 int unit_realize_cgroup(Unit *u) {
1503         assert(u);
1504
1505         if (!UNIT_HAS_CGROUP_CONTEXT(u))
1506                 return 0;
1507
1508         /* So, here's the deal: when realizing the cgroups for this
1509          * unit, we need to first create all parents, but there's more
1510          * actually: for the weight-based controllers we also need to
1511          * make sure that all our siblings (i.e. units that are in the
1512          * same slice as we are) have cgroups, too. Otherwise, things
1513          * would become very uneven as each of their processes would
1514          * get as much resources as all our group together. This call
1515          * will synchronously create the parent cgroups, but will
1516          * defer work on the siblings to the next event loop
1517          * iteration. */
1518
1519         /* Add all sibling slices to the cgroup queue. */
1520         unit_queue_siblings(u);
1521
1522         /* And realize this one now (and apply the values) */
1523         return unit_realize_cgroup_now(u, manager_state(u->manager));
1524 }
1525
1526 void unit_release_cgroup(Unit *u) {
1527         assert(u);
1528
1529         /* Forgets all cgroup details for this cgroup */
1530
1531         if (u->cgroup_path) {
1532                 (void) hashmap_remove(u->manager->cgroup_unit, u->cgroup_path);
1533                 u->cgroup_path = mfree(u->cgroup_path);
1534         }
1535
1536         if (u->cgroup_inotify_wd >= 0) {
1537                 if (inotify_rm_watch(u->manager->cgroup_inotify_fd, u->cgroup_inotify_wd) < 0)
1538                         log_unit_debug_errno(u, errno, "Failed to remove cgroup inotify watch %i for %s, ignoring", u->cgroup_inotify_wd, u->id);
1539
1540                 (void) hashmap_remove(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd));
1541                 u->cgroup_inotify_wd = -1;
1542         }
1543 }
1544
1545 void unit_prune_cgroup(Unit *u) {
1546         int r;
1547         bool is_root_slice;
1548
1549         assert(u);
1550
1551         /* Removes the cgroup, if empty and possible, and stops watching it. */
1552
1553         if (!u->cgroup_path)
1554                 return;
1555
1556         (void) unit_get_cpu_usage(u, NULL); /* Cache the last CPU usage value before we destroy the cgroup */
1557
1558         is_root_slice = unit_has_name(u, SPECIAL_ROOT_SLICE);
1559
1560         r = cg_trim_everywhere(u->manager->cgroup_supported, u->cgroup_path, !is_root_slice);
1561         if (r < 0) {
1562                 log_unit_debug_errno(u, r, "Failed to destroy cgroup %s, ignoring: %m", u->cgroup_path);
1563                 return;
1564         }
1565
1566         if (is_root_slice)
1567                 return;
1568
1569         unit_release_cgroup(u);
1570
1571         u->cgroup_realized = false;
1572         u->cgroup_realized_mask = 0;
1573         u->cgroup_enabled_mask = 0;
1574 }
1575
1576 int unit_search_main_pid(Unit *u, pid_t *ret) {
1577         _cleanup_fclose_ FILE *f = NULL;
1578         pid_t pid = 0, npid, mypid;
1579         int r;
1580
1581         assert(u);
1582         assert(ret);
1583
1584         if (!u->cgroup_path)
1585                 return -ENXIO;
1586
1587         r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, &f);
1588         if (r < 0)
1589                 return r;
1590
1591         mypid = getpid();
1592         while (cg_read_pid(f, &npid) > 0)  {
1593                 pid_t ppid;
1594
1595                 if (npid == pid)
1596                         continue;
1597
1598                 /* Ignore processes that aren't our kids */
1599                 if (get_process_ppid(npid, &ppid) >= 0 && ppid != mypid)
1600                         continue;
1601
1602                 if (pid != 0)
1603                         /* Dang, there's more than one daemonized PID
1604                         in this group, so we don't know what process
1605                         is the main process. */
1606
1607                         return -ENODATA;
1608
1609                 pid = npid;
1610         }
1611
1612         *ret = pid;
1613         return 0;
1614 }
1615
1616 static int unit_watch_pids_in_path(Unit *u, const char *path) {
1617         _cleanup_closedir_ DIR *d = NULL;
1618         _cleanup_fclose_ FILE *f = NULL;
1619         int ret = 0, r;
1620
1621         assert(u);
1622         assert(path);
1623
1624         r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, path, &f);
1625         if (r < 0)
1626                 ret = r;
1627         else {
1628                 pid_t pid;
1629
1630                 while ((r = cg_read_pid(f, &pid)) > 0) {
1631                         r = unit_watch_pid(u, pid);
1632                         if (r < 0 && ret >= 0)
1633                                 ret = r;
1634                 }
1635
1636                 if (r < 0 && ret >= 0)
1637                         ret = r;
1638         }
1639
1640         r = cg_enumerate_subgroups(SYSTEMD_CGROUP_CONTROLLER, path, &d);
1641         if (r < 0) {
1642                 if (ret >= 0)
1643                         ret = r;
1644         } else {
1645                 char *fn;
1646
1647                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1648                         _cleanup_free_ char *p = NULL;
1649
1650                         p = strjoin(path, "/", fn, NULL);
1651                         free(fn);
1652
1653                         if (!p)
1654                                 return -ENOMEM;
1655
1656                         r = unit_watch_pids_in_path(u, p);
1657                         if (r < 0 && ret >= 0)
1658                                 ret = r;
1659                 }
1660
1661                 if (r < 0 && ret >= 0)
1662                         ret = r;
1663         }
1664
1665         return ret;
1666 }
1667
1668 int unit_watch_all_pids(Unit *u) {
1669         assert(u);
1670
1671         /* Adds all PIDs from our cgroup to the set of PIDs we
1672          * watch. This is a fallback logic for cases where we do not
1673          * get reliable cgroup empty notifications: we try to use
1674          * SIGCHLD as replacement. */
1675
1676         if (!u->cgroup_path)
1677                 return -ENOENT;
1678
1679         if (cg_unified(SYSTEMD_CGROUP_CONTROLLER) > 0) /* On unified we can use proper notifications */
1680                 return 0;
1681
1682         return unit_watch_pids_in_path(u, u->cgroup_path);
1683 }
1684
1685 int unit_notify_cgroup_empty(Unit *u) {
1686         int r;
1687
1688         assert(u);
1689
1690         if (!u->cgroup_path)
1691                 return 0;
1692
1693         r = cg_is_empty_recursive(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path);
1694         if (r <= 0)
1695                 return r;
1696
1697         unit_add_to_gc_queue(u);
1698
1699         if (UNIT_VTABLE(u)->notify_cgroup_empty)
1700                 UNIT_VTABLE(u)->notify_cgroup_empty(u);
1701
1702         return 0;
1703 }
1704
1705 static int on_cgroup_inotify_event(sd_event_source *s, int fd, uint32_t revents, void *userdata) {
1706         Manager *m = userdata;
1707
1708         assert(s);
1709         assert(fd >= 0);
1710         assert(m);
1711
1712         for (;;) {
1713                 union inotify_event_buffer buffer;
1714                 struct inotify_event *e;
1715                 ssize_t l;
1716
1717                 l = read(fd, &buffer, sizeof(buffer));
1718                 if (l < 0) {
1719                         if (errno == EINTR || errno == EAGAIN)
1720                                 return 0;
1721
1722                         return log_error_errno(errno, "Failed to read control group inotify events: %m");
1723                 }
1724
1725                 FOREACH_INOTIFY_EVENT(e, buffer, l) {
1726                         Unit *u;
1727
1728                         if (e->wd < 0)
1729                                 /* Queue overflow has no watch descriptor */
1730                                 continue;
1731
1732                         if (e->mask & IN_IGNORED)
1733                                 /* The watch was just removed */
1734                                 continue;
1735
1736                         u = hashmap_get(m->cgroup_inotify_wd_unit, INT_TO_PTR(e->wd));
1737                         if (!u) /* Not that inotify might deliver
1738                                  * events for a watch even after it
1739                                  * was removed, because it was queued
1740                                  * before the removal. Let's ignore
1741                                  * this here safely. */
1742                                 continue;
1743
1744                         (void) unit_notify_cgroup_empty(u);
1745                 }
1746         }
1747 }
1748 #endif // 0
1749
1750 int manager_setup_cgroup(Manager *m) {
1751         _cleanup_free_ char *path = NULL;
1752         CGroupController c;
1753         int r, all_unified, systemd_unified;
1754         char *e;
1755
1756         assert(m);
1757
1758         /* 1. Determine hierarchy */
1759         m->cgroup_root = mfree(m->cgroup_root);
1760         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 0, &m->cgroup_root);
1761         if (r < 0)
1762                 return log_error_errno(r, "Cannot determine cgroup we are running in: %m");
1763
1764 #if 0 /// elogind does not support systemd scopes and slices
1765         /* Chop off the init scope, if we are already located in it */
1766         e = endswith(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
1767
1768         /* LEGACY: Also chop off the system slice if we are in
1769          * it. This is to support live upgrades from older systemd
1770          * versions where PID 1 was moved there. Also see
1771          * cg_get_root_path(). */
1772         if (!e && MANAGER_IS_SYSTEM(m)) {
1773                 e = endswith(m->cgroup_root, "/" SPECIAL_SYSTEM_SLICE);
1774                 if (!e)
1775                         e = endswith(m->cgroup_root, "/system"); /* even more legacy */
1776         }
1777         if (e)
1778                 *e = 0;
1779 #endif // 0
1780
1781         /* And make sure to store away the root value without trailing
1782          * slash, even for the root dir, so that we can easily prepend
1783          * it everywhere. */
1784         while ((e = endswith(m->cgroup_root, "/")))
1785                 *e = 0;
1786         log_debug_elogind("Cgroup Controller \"%s\" -> root \"%s\"",
1787                           SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root);
1788
1789         /* 2. Show data */
1790         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, NULL, &path);
1791         if (r < 0)
1792                 return log_error_errno(r, "Cannot find cgroup mount point: %m");
1793
1794         all_unified = cg_all_unified();
1795         systemd_unified = cg_unified(SYSTEMD_CGROUP_CONTROLLER);
1796
1797         if (all_unified < 0 || systemd_unified < 0)
1798                 return log_error_errno(all_unified < 0 ? all_unified : systemd_unified,
1799                                        "Couldn't determine if we are running in the unified hierarchy: %m");
1800
1801         if (all_unified > 0)
1802                 log_debug("Unified cgroup hierarchy is located at %s.", path);
1803         else if (systemd_unified > 0)
1804                 log_debug("Unified cgroup hierarchy is located at %s. Controllers are on legacy hierarchies.", path);
1805         else
1806                 log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER ". File system hierarchy is at %s.", path);
1807
1808         if (!m->test_run) {
1809                 const char *scope_path;
1810
1811                 /* 3. Install agent */
1812                 if (systemd_unified) {
1813
1814                         /* In the unified hierarchy we can get
1815                          * cgroup empty notifications via inotify. */
1816
1817 #if 0 /// elogind does not support the unified hierarchy, yet.
1818                         m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
1819                         safe_close(m->cgroup_inotify_fd);
1820
1821                         m->cgroup_inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
1822                         if (m->cgroup_inotify_fd < 0)
1823                                 return log_error_errno(errno, "Failed to create control group inotify object: %m");
1824
1825                         r = sd_event_add_io(m->event, &m->cgroup_inotify_event_source, m->cgroup_inotify_fd, EPOLLIN, on_cgroup_inotify_event, m);
1826                         if (r < 0)
1827                                 return log_error_errno(r, "Failed to watch control group inotify object: %m");
1828
1829                         /* Process cgroup empty notifications early, but after service notifications and SIGCHLD. Also
1830                          * see handling of cgroup agent notifications, for the classic cgroup hierarchy support. */
1831                         r = sd_event_source_set_priority(m->cgroup_inotify_event_source, SD_EVENT_PRIORITY_NORMAL-5);
1832                         if (r < 0)
1833                                 return log_error_errno(r, "Failed to set priority of inotify event source: %m");
1834
1835                         (void) sd_event_source_set_description(m->cgroup_inotify_event_source, "cgroup-inotify");
1836
1837 #else
1838                         return log_error_errno(EOPNOTSUPP, "Unified cgroup hierarchy not supported: %m");
1839 #endif // 0
1840                 } else if (MANAGER_IS_SYSTEM(m)) {
1841
1842                         /* On the legacy hierarchy we only get
1843                          * notifications via cgroup agents. (Which
1844                          * isn't really reliable, since it does not
1845                          * generate events when control groups with
1846                          * children run empty. */
1847
1848                         r = cg_install_release_agent(SYSTEMD_CGROUP_CONTROLLER, SYSTEMD_CGROUP_AGENT_PATH);
1849                         if (r < 0)
1850                                 log_warning_errno(r, "Failed to install release agent, ignoring: %m");
1851                         else if (r > 0)
1852                                 log_debug("Installed release agent.");
1853                         else if (r == 0)
1854                                 log_debug("Release agent already installed.");
1855                 }
1856
1857 #if 0 /// elogind is not meant to run in systemd init scope
1858                 /* 4. Make sure we are in the special "init.scope" unit in the root slice. */
1859                 scope_path = strjoina(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
1860                 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1861 #else
1862                 if (streq(SYSTEMD_CGROUP_CONTROLLER, "name=elogind"))
1863                         // we are our own cgroup controller
1864                         scope_path = strjoina("");
1865                 else if (streq(m->cgroup_root, "/elogind"))
1866                         // root already is our cgroup
1867                         scope_path = strjoina(m->cgroup_root);
1868                 else
1869                         // we have to create our own group
1870                         scope_path = strjoina(m->cgroup_root, "/elogind");
1871                 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1872 #endif // 0
1873                 if (r < 0)
1874                         return log_error_errno(r, "Failed to create %s control group: %m", scope_path);
1875                 log_debug_elogind("Created control group \"%s\"", scope_path);
1876
1877                 /* also, move all other userspace processes remaining
1878                  * in the root cgroup into that scope. */
1879                 r = cg_migrate(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1880                 if (r < 0)
1881                         log_warning_errno(r, "Couldn't move remaining userspace processes, ignoring: %m");
1882
1883                 /* 5. And pin it, so that it cannot be unmounted */
1884                 safe_close(m->pin_cgroupfs_fd);
1885                 m->pin_cgroupfs_fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_NOCTTY|O_NONBLOCK);
1886                 if (m->pin_cgroupfs_fd < 0)
1887                         return log_error_errno(errno, "Failed to open pin file: %m");
1888
1889                 /* 6.  Always enable hierarchical support if it exists... */
1890                 if (!all_unified)
1891                         (void) cg_set_attribute("memory", "/", "memory.use_hierarchy", "1");
1892         }
1893
1894         /* 7. Figure out which controllers are supported */
1895         r = cg_mask_supported(&m->cgroup_supported);
1896         if (r < 0)
1897                 return log_error_errno(r, "Failed to determine supported controllers: %m");
1898
1899         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++)
1900                 log_debug("Controller '%s' supported: %s", cgroup_controller_to_string(c), yes_no(m->cgroup_supported & CGROUP_CONTROLLER_TO_MASK(c)));
1901
1902         return 0;
1903 }
1904
1905 void manager_shutdown_cgroup(Manager *m, bool delete) {
1906         assert(m);
1907
1908         /* We can't really delete the group, since we are in it. But
1909          * let's trim it. */
1910         if (delete && m->cgroup_root)
1911                 (void) cg_trim(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, false);
1912
1913 #if 0 /// elogind does not support the unified hierarchy, yet.
1914         m->cgroup_inotify_wd_unit = hashmap_free(m->cgroup_inotify_wd_unit);
1915
1916         m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
1917         m->cgroup_inotify_fd = safe_close(m->cgroup_inotify_fd);
1918 #endif // 0
1919
1920         m->pin_cgroupfs_fd = safe_close(m->pin_cgroupfs_fd);
1921
1922         m->cgroup_root = mfree(m->cgroup_root);
1923 }
1924
1925 #if 0 /// UNNEEDED by elogind
1926 Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
1927         char *p;
1928         Unit *u;
1929
1930         assert(m);
1931         assert(cgroup);
1932
1933         u = hashmap_get(m->cgroup_unit, cgroup);
1934         if (u)
1935                 return u;
1936
1937         p = strdupa(cgroup);
1938         for (;;) {
1939                 char *e;
1940
1941                 e = strrchr(p, '/');
1942                 if (!e || e == p)
1943                         return hashmap_get(m->cgroup_unit, SPECIAL_ROOT_SLICE);
1944
1945                 *e = 0;
1946
1947                 u = hashmap_get(m->cgroup_unit, p);
1948                 if (u)
1949                         return u;
1950         }
1951 }
1952
1953 Unit *manager_get_unit_by_pid_cgroup(Manager *m, pid_t pid) {
1954         _cleanup_free_ char *cgroup = NULL;
1955         int r;
1956
1957         assert(m);
1958
1959         if (pid <= 0)
1960                 return NULL;
1961
1962         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
1963         if (r < 0)
1964                 return NULL;
1965
1966         return manager_get_unit_by_cgroup(m, cgroup);
1967 }
1968
1969 Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
1970         Unit *u;
1971
1972         assert(m);
1973
1974         if (pid <= 0)
1975                 return NULL;
1976
1977         if (pid == 1)
1978                 return hashmap_get(m->units, SPECIAL_INIT_SCOPE);
1979
1980         u = hashmap_get(m->watch_pids1, PID_TO_PTR(pid));
1981         if (u)
1982                 return u;
1983
1984         u = hashmap_get(m->watch_pids2, PID_TO_PTR(pid));
1985         if (u)
1986                 return u;
1987
1988         return manager_get_unit_by_pid_cgroup(m, pid);
1989 }
1990 #endif // 0
1991
1992 #if 0 /// elogind must substitute this with its own variant
1993 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
1994         Unit *u;
1995
1996         assert(m);
1997         assert(cgroup);
1998
1999         log_debug("Got cgroup empty notification for: %s", cgroup);
2000
2001         u = manager_get_unit_by_cgroup(m, cgroup);
2002         if (!u)
2003                 return 0;
2004
2005         return unit_notify_cgroup_empty(u);
2006 }
2007 #else
2008 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
2009         Session *s;
2010
2011         assert(m);
2012         assert(cgroup);
2013
2014         log_debug("Got cgroup empty notification for: %s", cgroup);
2015
2016         s = hashmap_get(m->sessions, cgroup);
2017
2018         if (s) {
2019                 session_finalize(s);
2020                 session_free(s);
2021         } else
2022                 log_warning("Session not found: %s", cgroup);
2023
2024         return 0;
2025 }
2026 #endif // 0
2027 #if 0 /// UNNEEDED by elogind
2028 int unit_get_memory_current(Unit *u, uint64_t *ret) {
2029         _cleanup_free_ char *v = NULL;
2030         int r;
2031
2032         assert(u);
2033         assert(ret);
2034
2035         if (!u->cgroup_path)
2036                 return -ENODATA;
2037
2038         if ((u->cgroup_realized_mask & CGROUP_MASK_MEMORY) == 0)
2039                 return -ENODATA;
2040
2041         if (cg_all_unified() <= 0)
2042                 r = cg_get_attribute("memory", u->cgroup_path, "memory.usage_in_bytes", &v);
2043         else
2044                 r = cg_get_attribute("memory", u->cgroup_path, "memory.current", &v);
2045         if (r == -ENOENT)
2046                 return -ENODATA;
2047         if (r < 0)
2048                 return r;
2049
2050         return safe_atou64(v, ret);
2051 }
2052
2053 int unit_get_tasks_current(Unit *u, uint64_t *ret) {
2054         _cleanup_free_ char *v = NULL;
2055         int r;
2056
2057         assert(u);
2058         assert(ret);
2059
2060         if (!u->cgroup_path)
2061                 return -ENODATA;
2062
2063         if ((u->cgroup_realized_mask & CGROUP_MASK_PIDS) == 0)
2064                 return -ENODATA;
2065
2066         r = cg_get_attribute("pids", u->cgroup_path, "pids.current", &v);
2067         if (r == -ENOENT)
2068                 return -ENODATA;
2069         if (r < 0)
2070                 return r;
2071
2072         return safe_atou64(v, ret);
2073 }
2074
2075 static int unit_get_cpu_usage_raw(Unit *u, nsec_t *ret) {
2076         _cleanup_free_ char *v = NULL;
2077         uint64_t ns;
2078         int r;
2079
2080         assert(u);
2081         assert(ret);
2082
2083         if (!u->cgroup_path)
2084                 return -ENODATA;
2085
2086         if (cg_all_unified() > 0) {
2087                 const char *keys[] = { "usage_usec", NULL };
2088                 _cleanup_free_ char *val = NULL;
2089                 uint64_t us;
2090
2091                 if ((u->cgroup_realized_mask & CGROUP_MASK_CPU) == 0)
2092                         return -ENODATA;
2093
2094                 r = cg_get_keyed_attribute("cpu", u->cgroup_path, "cpu.stat", keys, &val);
2095                 if (r < 0)
2096                         return r;
2097
2098                 r = safe_atou64(val, &us);
2099                 if (r < 0)
2100                         return r;
2101
2102                 ns = us * NSEC_PER_USEC;
2103         } else {
2104                 if ((u->cgroup_realized_mask & CGROUP_MASK_CPUACCT) == 0)
2105                         return -ENODATA;
2106
2107                 r = cg_get_attribute("cpuacct", u->cgroup_path, "cpuacct.usage", &v);
2108                 if (r == -ENOENT)
2109                         return -ENODATA;
2110                 if (r < 0)
2111                         return r;
2112
2113                 r = safe_atou64(v, &ns);
2114                 if (r < 0)
2115                         return r;
2116         }
2117
2118         *ret = ns;
2119         return 0;
2120 }
2121
2122 int unit_get_cpu_usage(Unit *u, nsec_t *ret) {
2123         nsec_t ns;
2124         int r;
2125
2126         assert(u);
2127
2128         /* Retrieve the current CPU usage counter. This will subtract the CPU counter taken when the unit was
2129          * started. If the cgroup has been removed already, returns the last cached value. To cache the value, simply
2130          * call this function with a NULL return value. */
2131
2132         r = unit_get_cpu_usage_raw(u, &ns);
2133         if (r == -ENODATA && u->cpu_usage_last != NSEC_INFINITY) {
2134                 /* If we can't get the CPU usage anymore (because the cgroup was already removed, for example), use our
2135                  * cached value. */
2136
2137                 if (ret)
2138                         *ret = u->cpu_usage_last;
2139                 return 0;
2140         }
2141         if (r < 0)
2142                 return r;
2143
2144         if (ns > u->cpu_usage_base)
2145                 ns -= u->cpu_usage_base;
2146         else
2147                 ns = 0;
2148
2149         u->cpu_usage_last = ns;
2150         if (ret)
2151                 *ret = ns;
2152
2153         return 0;
2154 }
2155
2156 int unit_reset_cpu_usage(Unit *u) {
2157         nsec_t ns;
2158         int r;
2159
2160         assert(u);
2161
2162         u->cpu_usage_last = NSEC_INFINITY;
2163
2164         r = unit_get_cpu_usage_raw(u, &ns);
2165         if (r < 0) {
2166                 u->cpu_usage_base = 0;
2167                 return r;
2168         }
2169
2170         u->cpu_usage_base = ns;
2171         return 0;
2172 }
2173
2174 bool unit_cgroup_delegate(Unit *u) {
2175         CGroupContext *c;
2176
2177         assert(u);
2178
2179         c = unit_get_cgroup_context(u);
2180         if (!c)
2181                 return false;
2182
2183         return c->delegate;
2184 }
2185
2186 void unit_invalidate_cgroup(Unit *u, CGroupMask m) {
2187         assert(u);
2188
2189         if (!UNIT_HAS_CGROUP_CONTEXT(u))
2190                 return;
2191
2192         if (m == 0)
2193                 return;
2194
2195         /* always invalidate compat pairs together */
2196         if (m & (CGROUP_MASK_IO | CGROUP_MASK_BLKIO))
2197                 m |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
2198
2199         if ((u->cgroup_realized_mask & m) == 0)
2200                 return;
2201
2202         u->cgroup_realized_mask &= ~m;
2203         unit_add_to_cgroup_queue(u);
2204 }
2205
2206 void manager_invalidate_startup_units(Manager *m) {
2207         Iterator i;
2208         Unit *u;
2209
2210         assert(m);
2211
2212         SET_FOREACH(u, m->startup_units, i)
2213                 unit_invalidate_cgroup(u, CGROUP_MASK_CPU|CGROUP_MASK_IO|CGROUP_MASK_BLKIO);
2214 }
2215
2216 static const char* const cgroup_device_policy_table[_CGROUP_DEVICE_POLICY_MAX] = {
2217         [CGROUP_AUTO] = "auto",
2218         [CGROUP_CLOSED] = "closed",
2219         [CGROUP_STRICT] = "strict",
2220 };
2221
2222 DEFINE_STRING_TABLE_LOOKUP(cgroup_device_policy, CGroupDevicePolicy);
2223 #endif // 0