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