src/core/cgroup.c

   1 /***
   2   This file is part of systemd.
   3
   4   Copyright 2013 Lennart Poettering
   5
   6   systemd is free software; you can redistribute it and/or modify it
   7   under the terms of the GNU Lesser General Public License as published by
   8   the Free Software Foundation; either version 2.1 of the License, or
   9   (at your option) any later version.
  10
  11   systemd is distributed in the hope that it will be useful, but
  12   WITHOUT ANY WARRANTY; without even the implied warranty of
  13   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14   Lesser General Public License for more details.
  15
  16   You should have received a copy of the GNU Lesser General Public License
  17   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  18 ***/
  19
  20 #include <fcntl.h>
  21 #include <fnmatch.h>
  22
  23 #include "alloc-util.h"
  24 #include "cgroup-util.h"
  25 #include "cgroup.h"
  26 #include "fd-util.h"
  27 #include "fileio.h"
  28 #include "fs-util.h"
  29 #include "parse-util.h"
  30 #include "path-util.h"
  31 #include "process-util.h"
  32 //#include "special.h"
  33 #include "string-table.h"
  34 #include "string-util.h"
  35 #include "stdio-util.h"
  36
  37 #define CGROUP_CPU_QUOTA_PERIOD_USEC ((usec_t) 100 * USEC_PER_MSEC)
  38
  39 #if 0 /// UNNEEDED by elogind
  40 static void cgroup_compat_warn(void) {
  41         static bool cgroup_compat_warned = false;
  42
  43         if (cgroup_compat_warned)
  44                 return;
  45
  46         log_warning("cgroup compatibility translation between legacy and unified hierarchy settings activated. See cgroup-compat debug messages for details.");
  47         cgroup_compat_warned = true;
  48 }
  49
  50 #define log_cgroup_compat(unit, fmt, ...) do {                                  \
  51                 cgroup_compat_warn();                                           \
  52                 log_unit_debug(unit, "cgroup-compat: " fmt, ##__VA_ARGS__);     \
  53         } while (false)
  54
  55 void cgroup_context_init(CGroupContext *c) {
  56         assert(c);
  57
  58         /* Initialize everything to the kernel defaults, assuming the
  59          * structure is preinitialized to 0 */
  60
  61         c->cpu_weight = CGROUP_WEIGHT_INVALID;
  62         c->startup_cpu_weight = CGROUP_WEIGHT_INVALID;
  63         c->cpu_quota_per_sec_usec = USEC_INFINITY;
  64
  65         c->cpu_shares = CGROUP_CPU_SHARES_INVALID;
  66         c->startup_cpu_shares = CGROUP_CPU_SHARES_INVALID;
  67
  68         c->memory_high = CGROUP_LIMIT_MAX;
  69         c->memory_max = CGROUP_LIMIT_MAX;
  70         c->memory_swap_max = CGROUP_LIMIT_MAX;
  71
  72         c->memory_limit = CGROUP_LIMIT_MAX;
  73
  74         c->io_weight = CGROUP_WEIGHT_INVALID;
  75         c->startup_io_weight = CGROUP_WEIGHT_INVALID;
  76
  77         c->blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID;
  78         c->startup_blockio_weight = CGROUP_BLKIO_WEIGHT_INVALID;
  79
  80         c->tasks_max = (uint64_t) -1;
  81 }
  82
  83 void cgroup_context_free_device_allow(CGroupContext *c, CGroupDeviceAllow *a) {
  84         assert(c);
  85         assert(a);
  86
  87         LIST_REMOVE(device_allow, c->device_allow, a);
  88         free(a->path);
  89         free(a);
  90 }
  91
  92 void cgroup_context_free_io_device_weight(CGroupContext *c, CGroupIODeviceWeight *w) {
  93         assert(c);
  94         assert(w);
  95
  96         LIST_REMOVE(device_weights, c->io_device_weights, w);
  97         free(w->path);
  98         free(w);
  99 }
 100
 101 void cgroup_context_free_io_device_limit(CGroupContext *c, CGroupIODeviceLimit *l) {
 102         assert(c);
 103         assert(l);
 104
 105         LIST_REMOVE(device_limits, c->io_device_limits, l);
 106         free(l->path);
 107         free(l);
 108 }
 109
 110 void cgroup_context_free_blockio_device_weight(CGroupContext *c, CGroupBlockIODeviceWeight *w) {
 111         assert(c);
 112         assert(w);
 113
 114         LIST_REMOVE(device_weights, c->blockio_device_weights, w);
 115         free(w->path);
 116         free(w);
 117 }
 118
 119 void cgroup_context_free_blockio_device_bandwidth(CGroupContext *c, CGroupBlockIODeviceBandwidth *b) {
 120         assert(c);
 121         assert(b);
 122
 123         LIST_REMOVE(device_bandwidths, c->blockio_device_bandwidths, b);
 124         free(b->path);
 125         free(b);
 126 }
 127
 128 void cgroup_context_done(CGroupContext *c) {
 129         assert(c);
 130
 131         while (c->io_device_weights)
 132                 cgroup_context_free_io_device_weight(c, c->io_device_weights);
 133
 134         while (c->io_device_limits)
 135                 cgroup_context_free_io_device_limit(c, c->io_device_limits);
 136
 137         while (c->blockio_device_weights)
 138                 cgroup_context_free_blockio_device_weight(c, c->blockio_device_weights);
 139
 140         while (c->blockio_device_bandwidths)
 141                 cgroup_context_free_blockio_device_bandwidth(c, c->blockio_device_bandwidths);
 142
 143         while (c->device_allow)
 144                 cgroup_context_free_device_allow(c, c->device_allow);
 145 }
 146
 147 void cgroup_context_dump(CGroupContext *c, FILE* f, const char *prefix) {
 148         CGroupIODeviceLimit *il;
 149         CGroupIODeviceWeight *iw;
 150         CGroupBlockIODeviceBandwidth *b;
 151         CGroupBlockIODeviceWeight *w;
 152         CGroupDeviceAllow *a;
 153         char u[FORMAT_TIMESPAN_MAX];
 154
 155         assert(c);
 156         assert(f);
 157
 158         prefix = strempty(prefix);
 159
 160         fprintf(f,
 161                 "%sCPUAccounting=%s\n"
 162                 "%sIOAccounting=%s\n"
 163                 "%sBlockIOAccounting=%s\n"
 164                 "%sMemoryAccounting=%s\n"
 165                 "%sTasksAccounting=%s\n"
 166                 "%sCPUWeight=%" PRIu64 "\n"
 167                 "%sStartupCPUWeight=%" PRIu64 "\n"
 168                 "%sCPUShares=%" PRIu64 "\n"
 169                 "%sStartupCPUShares=%" PRIu64 "\n"
 170                 "%sCPUQuotaPerSecSec=%s\n"
 171                 "%sIOWeight=%" PRIu64 "\n"
 172                 "%sStartupIOWeight=%" PRIu64 "\n"
 173                 "%sBlockIOWeight=%" PRIu64 "\n"
 174                 "%sStartupBlockIOWeight=%" PRIu64 "\n"
 175                 "%sMemoryLow=%" PRIu64 "\n"
 176                 "%sMemoryHigh=%" PRIu64 "\n"
 177                 "%sMemoryMax=%" PRIu64 "\n"
 178                 "%sMemorySwapMax=%" PRIu64 "\n"
 179                 "%sMemoryLimit=%" PRIu64 "\n"
 180                 "%sTasksMax=%" PRIu64 "\n"
 181                 "%sDevicePolicy=%s\n"
 182                 "%sDelegate=%s\n",
 183                 prefix, yes_no(c->cpu_accounting),
 184                 prefix, yes_no(c->io_accounting),
 185                 prefix, yes_no(c->blockio_accounting),
 186                 prefix, yes_no(c->memory_accounting),
 187                 prefix, yes_no(c->tasks_accounting),
 188                 prefix, c->cpu_weight,
 189                 prefix, c->startup_cpu_weight,
 190                 prefix, c->cpu_shares,
 191                 prefix, c->startup_cpu_shares,
 192                 prefix, format_timespan(u, sizeof(u), c->cpu_quota_per_sec_usec, 1),
 193                 prefix, c->io_weight,
 194                 prefix, c->startup_io_weight,
 195                 prefix, c->blockio_weight,
 196                 prefix, c->startup_blockio_weight,
 197                 prefix, c->memory_low,
 198                 prefix, c->memory_high,
 199                 prefix, c->memory_max,
 200                 prefix, c->memory_swap_max,
 201                 prefix, c->memory_limit,
 202                 prefix, c->tasks_max,
 203                 prefix, cgroup_device_policy_to_string(c->device_policy),
 204                 prefix, yes_no(c->delegate));
 205
 206         LIST_FOREACH(device_allow, a, c->device_allow)
 207                 fprintf(f,
 208                         "%sDeviceAllow=%s %s%s%s\n",
 209                         prefix,
 210                         a->path,
 211                         a->r ? "r" : "", a->w ? "w" : "", a->m ? "m" : "");
 212
 213         LIST_FOREACH(device_weights, iw, c->io_device_weights)
 214                 fprintf(f,
 215                         "%sIODeviceWeight=%s %" PRIu64,
 216                         prefix,
 217                         iw->path,
 218                         iw->weight);
 219
 220         LIST_FOREACH(device_limits, il, c->io_device_limits) {
 221                 char buf[FORMAT_BYTES_MAX];
 222                 CGroupIOLimitType type;
 223
 224                 for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++)
 225                         if (il->limits[type] != cgroup_io_limit_defaults[type])
 226                                 fprintf(f,
 227                                         "%s%s=%s %s\n",
 228                                         prefix,
 229                                         cgroup_io_limit_type_to_string(type),
 230                                         il->path,
 231                                         format_bytes(buf, sizeof(buf), il->limits[type]));
 232         }
 233
 234         LIST_FOREACH(device_weights, w, c->blockio_device_weights)
 235                 fprintf(f,
 236                         "%sBlockIODeviceWeight=%s %" PRIu64,
 237                         prefix,
 238                         w->path,
 239                         w->weight);
 240
 241         LIST_FOREACH(device_bandwidths, b, c->blockio_device_bandwidths) {
 242                 char buf[FORMAT_BYTES_MAX];
 243
 244                 if (b->rbps != CGROUP_LIMIT_MAX)
 245                         fprintf(f,
 246                                 "%sBlockIOReadBandwidth=%s %s\n",
 247                                 prefix,
 248                                 b->path,
 249                                 format_bytes(buf, sizeof(buf), b->rbps));
 250                 if (b->wbps != CGROUP_LIMIT_MAX)
 251                         fprintf(f,
 252                                 "%sBlockIOWriteBandwidth=%s %s\n",
 253                                 prefix,
 254                                 b->path,
 255                                 format_bytes(buf, sizeof(buf), b->wbps));
 256         }
 257 }
 258
 259 static int lookup_block_device(const char *p, dev_t *dev) {
 260         struct stat st;
 261         int r;
 262
 263         assert(p);
 264         assert(dev);
 265
 266         r = stat(p, &st);
 267         if (r < 0)
 268                 return log_warning_errno(errno, "Couldn't stat device %s: %m", p);
 269
 270         if (S_ISBLK(st.st_mode))
 271                 *dev = st.st_rdev;
 272         else if (major(st.st_dev) != 0) {
 273                 /* If this is not a device node then find the block
 274                  * device this file is stored on */
 275                 *dev = st.st_dev;
 276
 277                 /* If this is a partition, try to get the originating
 278                  * block device */
 279                 block_get_whole_disk(*dev, dev);
 280         } else {
 281                 log_warning("%s is not a block device and file system block device cannot be determined or is not local.", p);
 282                 return -ENODEV;
 283         }
 284
 285         return 0;
 286 }
 287
 288 static int whitelist_device(const char *path, const char *node, const char *acc) {
 289         char buf[2+DECIMAL_STR_MAX(dev_t)*2+2+4];
 290         struct stat st;
 291         bool ignore_notfound;
 292         int r;
 293
 294         assert(path);
 295         assert(acc);
 296
 297         if (node[0] == '-') {
 298                 /* Non-existent paths starting with "-" must be silently ignored */
 299                 node++;
 300                 ignore_notfound = true;
 301         } else
 302                 ignore_notfound = false;
 303
 304         if (stat(node, &st) < 0) {
 305                 if (errno == ENOENT && ignore_notfound)
 306                         return 0;
 307
 308                 return log_warning_errno(errno, "Couldn't stat device %s: %m", node);
 309         }
 310
 311         if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
 312                 log_warning("%s is not a device.", node);
 313                 return -ENODEV;
 314         }
 315
 316         sprintf(buf,
 317                 "%c %u:%u %s",
 318                 S_ISCHR(st.st_mode) ? 'c' : 'b',
 319                 major(st.st_rdev), minor(st.st_rdev),
 320                 acc);
 321
 322         r = cg_set_attribute("devices", path, "devices.allow", buf);
 323         if (r < 0)
 324                 log_full_errno(IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 325                                "Failed to set devices.allow on %s: %m", path);
 326
 327         return r;
 328 }
 329
 330 static int whitelist_major(const char *path, const char *name, char type, const char *acc) {
 331         _cleanup_fclose_ FILE *f = NULL;
 332         char line[LINE_MAX];
 333         bool good = false;
 334         int r;
 335
 336         assert(path);
 337         assert(acc);
 338         assert(type == 'b' || type == 'c');
 339
 340         f = fopen("/proc/devices", "re");
 341         if (!f)
 342                 return log_warning_errno(errno, "Cannot open /proc/devices to resolve %s (%c): %m", name, type);
 343
 344         FOREACH_LINE(line, f, goto fail) {
 345                 char buf[2+DECIMAL_STR_MAX(unsigned)+3+4], *p, *w;
 346                 unsigned maj;
 347
 348                 truncate_nl(line);
 349
 350                 if (type == 'c' && streq(line, "Character devices:")) {
 351                         good = true;
 352                         continue;
 353                 }
 354
 355                 if (type == 'b' && streq(line, "Block devices:")) {
 356                         good = true;
 357                         continue;
 358                 }
 359
 360                 if (isempty(line)) {
 361                         good = false;
 362                         continue;
 363                 }
 364
 365                 if (!good)
 366                         continue;
 367
 368                 p = strstrip(line);
 369
 370                 w = strpbrk(p, WHITESPACE);
 371                 if (!w)
 372                         continue;
 373                 *w = 0;
 374
 375                 r = safe_atou(p, &maj);
 376                 if (r < 0)
 377                         continue;
 378                 if (maj <= 0)
 379                         continue;
 380
 381                 w++;
 382                 w += strspn(w, WHITESPACE);
 383
 384                 if (fnmatch(name, w, 0) != 0)
 385                         continue;
 386
 387                 sprintf(buf,
 388                         "%c %u:* %s",
 389                         type,
 390                         maj,
 391                         acc);
 392
 393                 r = cg_set_attribute("devices", path, "devices.allow", buf);
 394                 if (r < 0)
 395                         log_full_errno(IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 396                                        "Failed to set devices.allow on %s: %m", path);
 397         }
 398
 399         return 0;
 400
 401 fail:
 402         log_warning_errno(errno, "Failed to read /proc/devices: %m");
 403         return -errno;
 404 }
 405
 406 static bool cgroup_context_has_cpu_weight(CGroupContext *c) {
 407         return c->cpu_weight != CGROUP_WEIGHT_INVALID ||
 408                 c->startup_cpu_weight != CGROUP_WEIGHT_INVALID;
 409 }
 410
 411 static bool cgroup_context_has_cpu_shares(CGroupContext *c) {
 412         return c->cpu_shares != CGROUP_CPU_SHARES_INVALID ||
 413                 c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID;
 414 }
 415
 416 static uint64_t cgroup_context_cpu_weight(CGroupContext *c, ManagerState state) {
 417         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 418             c->startup_cpu_weight != CGROUP_WEIGHT_INVALID)
 419                 return c->startup_cpu_weight;
 420         else if (c->cpu_weight != CGROUP_WEIGHT_INVALID)
 421                 return c->cpu_weight;
 422         else
 423                 return CGROUP_WEIGHT_DEFAULT;
 424 }
 425
 426 static uint64_t cgroup_context_cpu_shares(CGroupContext *c, ManagerState state) {
 427         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 428             c->startup_cpu_shares != CGROUP_CPU_SHARES_INVALID)
 429                 return c->startup_cpu_shares;
 430         else if (c->cpu_shares != CGROUP_CPU_SHARES_INVALID)
 431                 return c->cpu_shares;
 432         else
 433                 return CGROUP_CPU_SHARES_DEFAULT;
 434 }
 435
 436 static void cgroup_apply_unified_cpu_config(Unit *u, uint64_t weight, uint64_t quota) {
 437         char buf[MAX(DECIMAL_STR_MAX(uint64_t) + 1, (DECIMAL_STR_MAX(usec_t) + 1) * 2)];
 438         int r;
 439
 440         xsprintf(buf, "%" PRIu64 "\n", weight);
 441         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.weight", buf);
 442         if (r < 0)
 443                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 444                               "Failed to set cpu.weight: %m");
 445
 446         if (quota != USEC_INFINITY)
 447                 xsprintf(buf, USEC_FMT " " USEC_FMT "\n",
 448                          quota * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC, CGROUP_CPU_QUOTA_PERIOD_USEC);
 449         else
 450                 xsprintf(buf, "max " USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
 451
 452         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.max", buf);
 453
 454         if (r < 0)
 455                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 456                               "Failed to set cpu.max: %m");
 457 }
 458
 459 static void cgroup_apply_legacy_cpu_config(Unit *u, uint64_t shares, uint64_t quota) {
 460         char buf[MAX(DECIMAL_STR_MAX(uint64_t), DECIMAL_STR_MAX(usec_t)) + 1];
 461         int r;
 462
 463         xsprintf(buf, "%" PRIu64 "\n", shares);
 464         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.shares", buf);
 465         if (r < 0)
 466                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 467                               "Failed to set cpu.shares: %m");
 468
 469         xsprintf(buf, USEC_FMT "\n", CGROUP_CPU_QUOTA_PERIOD_USEC);
 470         r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_period_us", buf);
 471         if (r < 0)
 472                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 473                               "Failed to set cpu.cfs_period_us: %m");
 474
 475         if (quota != USEC_INFINITY) {
 476                 xsprintf(buf, USEC_FMT "\n", quota * CGROUP_CPU_QUOTA_PERIOD_USEC / USEC_PER_SEC);
 477                 r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_quota_us", buf);
 478         } else
 479                 r = cg_set_attribute("cpu", u->cgroup_path, "cpu.cfs_quota_us", "-1");
 480         if (r < 0)
 481                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 482                               "Failed to set cpu.cfs_quota_us: %m");
 483 }
 484
 485 static uint64_t cgroup_cpu_shares_to_weight(uint64_t shares) {
 486         return CLAMP(shares * CGROUP_WEIGHT_DEFAULT / CGROUP_CPU_SHARES_DEFAULT,
 487                      CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
 488 }
 489
 490 static uint64_t cgroup_cpu_weight_to_shares(uint64_t weight) {
 491         return CLAMP(weight * CGROUP_CPU_SHARES_DEFAULT / CGROUP_WEIGHT_DEFAULT,
 492                      CGROUP_CPU_SHARES_MIN, CGROUP_CPU_SHARES_MAX);
 493 }
 494
 495 static bool cgroup_context_has_io_config(CGroupContext *c) {
 496         return c->io_accounting ||
 497                 c->io_weight != CGROUP_WEIGHT_INVALID ||
 498                 c->startup_io_weight != CGROUP_WEIGHT_INVALID ||
 499                 c->io_device_weights ||
 500                 c->io_device_limits;
 501 }
 502
 503 static bool cgroup_context_has_blockio_config(CGroupContext *c) {
 504         return c->blockio_accounting ||
 505                 c->blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID ||
 506                 c->startup_blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID ||
 507                 c->blockio_device_weights ||
 508                 c->blockio_device_bandwidths;
 509 }
 510
 511 static uint64_t cgroup_context_io_weight(CGroupContext *c, ManagerState state) {
 512         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 513             c->startup_io_weight != CGROUP_WEIGHT_INVALID)
 514                 return c->startup_io_weight;
 515         else if (c->io_weight != CGROUP_WEIGHT_INVALID)
 516                 return c->io_weight;
 517         else
 518                 return CGROUP_WEIGHT_DEFAULT;
 519 }
 520
 521 static uint64_t cgroup_context_blkio_weight(CGroupContext *c, ManagerState state) {
 522         if (IN_SET(state, MANAGER_STARTING, MANAGER_INITIALIZING) &&
 523             c->startup_blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID)
 524                 return c->startup_blockio_weight;
 525         else if (c->blockio_weight != CGROUP_BLKIO_WEIGHT_INVALID)
 526                 return c->blockio_weight;
 527         else
 528                 return CGROUP_BLKIO_WEIGHT_DEFAULT;
 529 }
 530
 531 static uint64_t cgroup_weight_blkio_to_io(uint64_t blkio_weight) {
 532         return CLAMP(blkio_weight * CGROUP_WEIGHT_DEFAULT / CGROUP_BLKIO_WEIGHT_DEFAULT,
 533                      CGROUP_WEIGHT_MIN, CGROUP_WEIGHT_MAX);
 534 }
 535
 536 static uint64_t cgroup_weight_io_to_blkio(uint64_t io_weight) {
 537         return CLAMP(io_weight * CGROUP_BLKIO_WEIGHT_DEFAULT / CGROUP_WEIGHT_DEFAULT,
 538                      CGROUP_BLKIO_WEIGHT_MIN, CGROUP_BLKIO_WEIGHT_MAX);
 539 }
 540
 541 static void cgroup_apply_io_device_weight(Unit *u, const char *dev_path, uint64_t io_weight) {
 542         char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
 543         dev_t dev;
 544         int r;
 545
 546         r = lookup_block_device(dev_path, &dev);
 547         if (r < 0)
 548                 return;
 549
 550         xsprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), io_weight);
 551         r = cg_set_attribute("io", u->cgroup_path, "io.weight", buf);
 552         if (r < 0)
 553                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 554                               "Failed to set io.weight: %m");
 555 }
 556
 557 static void cgroup_apply_blkio_device_weight(Unit *u, const char *dev_path, uint64_t blkio_weight) {
 558         char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
 559         dev_t dev;
 560         int r;
 561
 562         r = lookup_block_device(dev_path, &dev);
 563         if (r < 0)
 564                 return;
 565
 566         xsprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), blkio_weight);
 567         r = cg_set_attribute("blkio", u->cgroup_path, "blkio.weight_device", buf);
 568         if (r < 0)
 569                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 570                               "Failed to set blkio.weight_device: %m");
 571 }
 572
 573 static unsigned cgroup_apply_io_device_limit(Unit *u, const char *dev_path, uint64_t *limits) {
 574         char limit_bufs[_CGROUP_IO_LIMIT_TYPE_MAX][DECIMAL_STR_MAX(uint64_t)];
 575         char buf[DECIMAL_STR_MAX(dev_t)*2+2+(6+DECIMAL_STR_MAX(uint64_t)+1)*4];
 576         CGroupIOLimitType type;
 577         dev_t dev;
 578         unsigned n = 0;
 579         int r;
 580
 581         r = lookup_block_device(dev_path, &dev);
 582         if (r < 0)
 583                 return 0;
 584
 585         for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++) {
 586                 if (limits[type] != cgroup_io_limit_defaults[type]) {
 587                         xsprintf(limit_bufs[type], "%" PRIu64, limits[type]);
 588                         n++;
 589                 } else {
 590                         xsprintf(limit_bufs[type], "%s", limits[type] == CGROUP_LIMIT_MAX ? "max" : "0");
 591                 }
 592         }
 593
 594         xsprintf(buf, "%u:%u rbps=%s wbps=%s riops=%s wiops=%s\n", major(dev), minor(dev),
 595                  limit_bufs[CGROUP_IO_RBPS_MAX], limit_bufs[CGROUP_IO_WBPS_MAX],
 596                  limit_bufs[CGROUP_IO_RIOPS_MAX], limit_bufs[CGROUP_IO_WIOPS_MAX]);
 597         r = cg_set_attribute("io", u->cgroup_path, "io.max", buf);
 598         if (r < 0)
 599                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 600                               "Failed to set io.max: %m");
 601         return n;
 602 }
 603
 604 static unsigned cgroup_apply_blkio_device_limit(Unit *u, const char *dev_path, uint64_t rbps, uint64_t wbps) {
 605         char buf[DECIMAL_STR_MAX(dev_t)*2+2+DECIMAL_STR_MAX(uint64_t)+1];
 606         dev_t dev;
 607         unsigned n = 0;
 608         int r;
 609
 610         r = lookup_block_device(dev_path, &dev);
 611         if (r < 0)
 612                 return 0;
 613
 614         if (rbps != CGROUP_LIMIT_MAX)
 615                 n++;
 616         sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), rbps);
 617         r = cg_set_attribute("blkio", u->cgroup_path, "blkio.throttle.read_bps_device", 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 blkio.throttle.read_bps_device: %m");
 621
 622         if (wbps != CGROUP_LIMIT_MAX)
 623                 n++;
 624         sprintf(buf, "%u:%u %" PRIu64 "\n", major(dev), minor(dev), wbps);
 625         r = cg_set_attribute("blkio", u->cgroup_path, "blkio.throttle.write_bps_device", buf);
 626         if (r < 0)
 627                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 628                               "Failed to set blkio.throttle.write_bps_device: %m");
 629
 630         return n;
 631 }
 632
 633 static bool cgroup_context_has_unified_memory_config(CGroupContext *c) {
 634         return c->memory_low > 0 || c->memory_high != CGROUP_LIMIT_MAX || c->memory_max != CGROUP_LIMIT_MAX || c->memory_swap_max != CGROUP_LIMIT_MAX;
 635 }
 636
 637 static void cgroup_apply_unified_memory_limit(Unit *u, const char *file, uint64_t v) {
 638         char buf[DECIMAL_STR_MAX(uint64_t) + 1] = "max";
 639         int r;
 640
 641         if (v != CGROUP_LIMIT_MAX)
 642                 xsprintf(buf, "%" PRIu64 "\n", v);
 643
 644         r = cg_set_attribute("memory", u->cgroup_path, file, buf);
 645         if (r < 0)
 646                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 647                               "Failed to set %s: %m", file);
 648 }
 649
 650 static void cgroup_context_apply(Unit *u, CGroupMask mask, ManagerState state) {
 651         const char *path;
 652         CGroupContext *c;
 653         bool is_root;
 654         int r;
 655
 656         assert(u);
 657
 658         c = unit_get_cgroup_context(u);
 659         path = u->cgroup_path;
 660
 661         assert(c);
 662         assert(path);
 663
 664         if (mask == 0)
 665                 return;
 666
 667         /* Some cgroup attributes are not supported on the root cgroup,
 668          * hence silently ignore */
 669         is_root = isempty(path) || path_equal(path, "/");
 670         if (is_root)
 671                 /* Make sure we don't try to display messages with an empty path. */
 672                 path = "/";
 673
 674         /* We generally ignore errors caused by read-only mounted
 675          * cgroup trees (assuming we are running in a container then),
 676          * and missing cgroups, i.e. EROFS and ENOENT. */
 677
 678         if ((mask & CGROUP_MASK_CPU) && !is_root) {
 679                 bool has_weight = cgroup_context_has_cpu_weight(c);
 680                 bool has_shares = cgroup_context_has_cpu_shares(c);
 681
 682                 if (cg_all_unified() > 0) {
 683                         uint64_t weight;
 684
 685                         if (has_weight)
 686                                 weight = cgroup_context_cpu_weight(c, state);
 687                         else if (has_shares) {
 688                                 uint64_t shares = cgroup_context_cpu_shares(c, state);
 689
 690                                 weight = cgroup_cpu_shares_to_weight(shares);
 691
 692                                 log_cgroup_compat(u, "Applying [Startup]CpuShares %" PRIu64 " as [Startup]CpuWeight %" PRIu64 " on %s",
 693                                                   shares, weight, path);
 694                         } else
 695                                 weight = CGROUP_WEIGHT_DEFAULT;
 696
 697                         cgroup_apply_unified_cpu_config(u, weight, c->cpu_quota_per_sec_usec);
 698                 } else {
 699                         uint64_t shares;
 700
 701                         if (has_weight) {
 702                                 uint64_t weight = cgroup_context_cpu_weight(c, state);
 703
 704                                 shares = cgroup_cpu_weight_to_shares(weight);
 705
 706                                 log_cgroup_compat(u, "Applying [Startup]CpuWeight %" PRIu64 " as [Startup]CpuShares %" PRIu64 " on %s",
 707                                                   weight, shares, path);
 708                         } else if (has_shares)
 709                                 shares = cgroup_context_cpu_shares(c, state);
 710                         else
 711                                 shares = CGROUP_CPU_SHARES_DEFAULT;
 712
 713                         cgroup_apply_legacy_cpu_config(u, shares, c->cpu_quota_per_sec_usec);
 714                 }
 715         }
 716
 717         if (mask & CGROUP_MASK_IO) {
 718                 bool has_io = cgroup_context_has_io_config(c);
 719                 bool has_blockio = cgroup_context_has_blockio_config(c);
 720
 721                 if (!is_root) {
 722                         char buf[8+DECIMAL_STR_MAX(uint64_t)+1];
 723                         uint64_t weight;
 724
 725                         if (has_io)
 726                                 weight = cgroup_context_io_weight(c, state);
 727                         else if (has_blockio) {
 728                                 uint64_t blkio_weight = cgroup_context_blkio_weight(c, state);
 729
 730                                 weight = cgroup_weight_blkio_to_io(blkio_weight);
 731
 732                                 log_cgroup_compat(u, "Applying [Startup]BlockIOWeight %" PRIu64 " as [Startup]IOWeight %" PRIu64,
 733                                                   blkio_weight, weight);
 734                         } else
 735                                 weight = CGROUP_WEIGHT_DEFAULT;
 736
 737                         xsprintf(buf, "default %" PRIu64 "\n", weight);
 738                         r = cg_set_attribute("io", path, "io.weight", buf);
 739                         if (r < 0)
 740                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 741                                               "Failed to set io.weight: %m");
 742
 743                         if (has_io) {
 744                                 CGroupIODeviceWeight *w;
 745
 746                                 /* FIXME: no way to reset this list */
 747                                 LIST_FOREACH(device_weights, w, c->io_device_weights)
 748                                         cgroup_apply_io_device_weight(u, w->path, w->weight);
 749                         } else if (has_blockio) {
 750                                 CGroupBlockIODeviceWeight *w;
 751
 752                                 /* FIXME: no way to reset this list */
 753                                 LIST_FOREACH(device_weights, w, c->blockio_device_weights) {
 754                                         weight = cgroup_weight_blkio_to_io(w->weight);
 755
 756                                         log_cgroup_compat(u, "Applying BlockIODeviceWeight %" PRIu64 " as IODeviceWeight %" PRIu64 " for %s",
 757                                                           w->weight, weight, w->path);
 758
 759                                         cgroup_apply_io_device_weight(u, w->path, weight);
 760                                 }
 761                         }
 762                 }
 763
 764                 /* Apply limits and free ones without config. */
 765                 if (has_io) {
 766                         CGroupIODeviceLimit *l, *next;
 767
 768                         LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
 769                                 if (!cgroup_apply_io_device_limit(u, l->path, l->limits))
 770                                         cgroup_context_free_io_device_limit(c, l);
 771                         }
 772                 } else if (has_blockio) {
 773                         CGroupBlockIODeviceBandwidth *b, *next;
 774
 775                         LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths) {
 776                                 uint64_t limits[_CGROUP_IO_LIMIT_TYPE_MAX];
 777                                 CGroupIOLimitType type;
 778
 779                                 for (type = 0; type < _CGROUP_IO_LIMIT_TYPE_MAX; type++)
 780                                         limits[type] = cgroup_io_limit_defaults[type];
 781
 782                                 limits[CGROUP_IO_RBPS_MAX] = b->rbps;
 783                                 limits[CGROUP_IO_WBPS_MAX] = b->wbps;
 784
 785                                 log_cgroup_compat(u, "Applying BlockIO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as IO{Read|Write}BandwidthMax for %s",
 786                                                   b->rbps, b->wbps, b->path);
 787
 788                                 if (!cgroup_apply_io_device_limit(u, b->path, limits))
 789                                         cgroup_context_free_blockio_device_bandwidth(c, b);
 790                         }
 791                 }
 792         }
 793
 794         if (mask & CGROUP_MASK_BLKIO) {
 795                 bool has_io = cgroup_context_has_io_config(c);
 796                 bool has_blockio = cgroup_context_has_blockio_config(c);
 797
 798                 if (!is_root) {
 799                         char buf[DECIMAL_STR_MAX(uint64_t)+1];
 800                         uint64_t weight;
 801
 802                         if (has_io) {
 803                                 uint64_t io_weight = cgroup_context_io_weight(c, state);
 804
 805                                 weight = cgroup_weight_io_to_blkio(cgroup_context_io_weight(c, state));
 806
 807                                 log_cgroup_compat(u, "Applying [Startup]IOWeight %" PRIu64 " as [Startup]BlockIOWeight %" PRIu64,
 808                                                   io_weight, weight);
 809                         } else if (has_blockio)
 810                                 weight = cgroup_context_blkio_weight(c, state);
 811                         else
 812                                 weight = CGROUP_BLKIO_WEIGHT_DEFAULT;
 813
 814                         xsprintf(buf, "%" PRIu64 "\n", weight);
 815                         r = cg_set_attribute("blkio", path, "blkio.weight", buf);
 816                         if (r < 0)
 817                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 818                                               "Failed to set blkio.weight: %m");
 819
 820                         if (has_io) {
 821                                 CGroupIODeviceWeight *w;
 822
 823                                 /* FIXME: no way to reset this list */
 824                                 LIST_FOREACH(device_weights, w, c->io_device_weights) {
 825                                         weight = cgroup_weight_io_to_blkio(w->weight);
 826
 827                                         log_cgroup_compat(u, "Applying IODeviceWeight %" PRIu64 " as BlockIODeviceWeight %" PRIu64 " for %s",
 828                                                           w->weight, weight, w->path);
 829
 830                                         cgroup_apply_blkio_device_weight(u, w->path, weight);
 831                                 }
 832                         } else if (has_blockio) {
 833                                 CGroupBlockIODeviceWeight *w;
 834
 835                                 /* FIXME: no way to reset this list */
 836                                 LIST_FOREACH(device_weights, w, c->blockio_device_weights)
 837                                         cgroup_apply_blkio_device_weight(u, w->path, w->weight);
 838                         }
 839                 }
 840
 841                 /* Apply limits and free ones without config. */
 842                 if (has_io) {
 843                         CGroupIODeviceLimit *l, *next;
 844
 845                         LIST_FOREACH_SAFE(device_limits, l, next, c->io_device_limits) {
 846                                 log_cgroup_compat(u, "Applying IO{Read|Write}Bandwidth %" PRIu64 " %" PRIu64 " as BlockIO{Read|Write}BandwidthMax for %s",
 847                                                   l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX], l->path);
 848
 849                                 if (!cgroup_apply_blkio_device_limit(u, l->path, l->limits[CGROUP_IO_RBPS_MAX], l->limits[CGROUP_IO_WBPS_MAX]))
 850                                         cgroup_context_free_io_device_limit(c, l);
 851                         }
 852                 } else if (has_blockio) {
 853                         CGroupBlockIODeviceBandwidth *b, *next;
 854
 855                         LIST_FOREACH_SAFE(device_bandwidths, b, next, c->blockio_device_bandwidths)
 856                                 if (!cgroup_apply_blkio_device_limit(u, b->path, b->rbps, b->wbps))
 857                                         cgroup_context_free_blockio_device_bandwidth(c, b);
 858                 }
 859         }
 860
 861         if ((mask & CGROUP_MASK_MEMORY) && !is_root) {
 862                 if (cg_all_unified() > 0) {
 863                         uint64_t max, swap_max = CGROUP_LIMIT_MAX;
 864
 865                         if (cgroup_context_has_unified_memory_config(c)) {
 866                                 max = c->memory_max;
 867                                 swap_max = c->memory_swap_max;
 868                         } else {
 869                                 max = c->memory_limit;
 870
 871                                 if (max != CGROUP_LIMIT_MAX)
 872                                         log_cgroup_compat(u, "Applying MemoryLimit %" PRIu64 " as MemoryMax", max);
 873                         }
 874
 875                         cgroup_apply_unified_memory_limit(u, "memory.low", c->memory_low);
 876                         cgroup_apply_unified_memory_limit(u, "memory.high", c->memory_high);
 877                         cgroup_apply_unified_memory_limit(u, "memory.max", max);
 878                         cgroup_apply_unified_memory_limit(u, "memory.swap.max", swap_max);
 879                 } else {
 880                         char buf[DECIMAL_STR_MAX(uint64_t) + 1];
 881                         uint64_t val;
 882
 883                         if (cgroup_context_has_unified_memory_config(c)) {
 884                                 val = c->memory_max;
 885                                 log_cgroup_compat(u, "Applying MemoryMax %" PRIi64 " as MemoryLimit", val);
 886                         } else
 887                                 val = c->memory_limit;
 888
 889                         if (val == CGROUP_LIMIT_MAX)
 890                                 strncpy(buf, "-1\n", sizeof(buf));
 891                         else
 892                                 xsprintf(buf, "%" PRIu64 "\n", val);
 893
 894                         r = cg_set_attribute("memory", path, "memory.limit_in_bytes", buf);
 895                         if (r < 0)
 896                                 log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 897                                               "Failed to set memory.limit_in_bytes: %m");
 898                 }
 899         }
 900
 901         if ((mask & CGROUP_MASK_DEVICES) && !is_root) {
 902                 CGroupDeviceAllow *a;
 903
 904                 /* Changing the devices list of a populated cgroup
 905                  * might result in EINVAL, hence ignore EINVAL
 906                  * here. */
 907
 908                 if (c->device_allow || c->device_policy != CGROUP_AUTO)
 909                         r = cg_set_attribute("devices", path, "devices.deny", "a");
 910                 else
 911                         r = cg_set_attribute("devices", path, "devices.allow", "a");
 912                 if (r < 0)
 913                         log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EINVAL, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 914                                       "Failed to reset devices.list: %m");
 915
 916                 if (c->device_policy == CGROUP_CLOSED ||
 917                     (c->device_policy == CGROUP_AUTO && c->device_allow)) {
 918                         static const char auto_devices[] =
 919                                 "/dev/null\0" "rwm\0"
 920                                 "/dev/zero\0" "rwm\0"
 921                                 "/dev/full\0" "rwm\0"
 922                                 "/dev/random\0" "rwm\0"
 923                                 "/dev/urandom\0" "rwm\0"
 924                                 "/dev/tty\0" "rwm\0"
 925                                 "/dev/pts/ptmx\0" "rw\0" /* /dev/pts/ptmx may not be duplicated, but accessed */
 926                                 /* Allow /run/elogind/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
 927                                 /* Allow /run/systemd/inaccessible/{chr,blk} devices for mapping InaccessiblePaths */
 928                                 "-/run/systemd/inaccessible/chr\0" "rwm\0"
 929                                 "-/run/systemd/inaccessible/blk\0" "rwm\0";
 930
 931                         const char *x, *y;
 932
 933                         NULSTR_FOREACH_PAIR(x, y, auto_devices)
 934                                 whitelist_device(path, x, y);
 935
 936                         whitelist_major(path, "pts", 'c', "rw");
 937                         whitelist_major(path, "kdbus", 'c', "rw");
 938                         whitelist_major(path, "kdbus/*", 'c', "rw");
 939                 }
 940
 941                 LIST_FOREACH(device_allow, a, c->device_allow) {
 942                         char acc[4], *val;
 943                         unsigned k = 0;
 944
 945                         if (a->r)
 946                                 acc[k++] = 'r';
 947                         if (a->w)
 948                                 acc[k++] = 'w';
 949                         if (a->m)
 950                                 acc[k++] = 'm';
 951
 952                         if (k == 0)
 953                                 continue;
 954
 955                         acc[k++] = 0;
 956
 957                         if (startswith(a->path, "/dev/"))
 958                                 whitelist_device(path, a->path, acc);
 959                         else if ((val = startswith(a->path, "block-")))
 960                                 whitelist_major(path, val, 'b', acc);
 961                         else if ((val = startswith(a->path, "char-")))
 962                                 whitelist_major(path, val, 'c', acc);
 963                         else
 964                                 log_unit_debug(u, "Ignoring device %s while writing cgroup attribute.", a->path);
 965                 }
 966         }
 967
 968         if ((mask & CGROUP_MASK_PIDS) && !is_root) {
 969
 970                 if (c->tasks_max != CGROUP_LIMIT_MAX) {
 971                         char buf[DECIMAL_STR_MAX(uint64_t) + 2];
 972
 973                         sprintf(buf, "%" PRIu64 "\n", c->tasks_max);
 974                         r = cg_set_attribute("pids", path, "pids.max", buf);
 975                 } else
 976                         r = cg_set_attribute("pids", path, "pids.max", "max");
 977
 978                 if (r < 0)
 979                         log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
 980                                       "Failed to set pids.max: %m");
 981         }
 982 }
 983
 984 CGroupMask cgroup_context_get_mask(CGroupContext *c) {
 985         CGroupMask mask = 0;
 986
 987         /* Figure out which controllers we need */
 988
 989         if (c->cpu_accounting ||
 990             cgroup_context_has_cpu_weight(c) ||
 991             cgroup_context_has_cpu_shares(c) ||
 992             c->cpu_quota_per_sec_usec != USEC_INFINITY)
 993                 mask |= CGROUP_MASK_CPUACCT | CGROUP_MASK_CPU;
 994
 995         if (cgroup_context_has_io_config(c) || cgroup_context_has_blockio_config(c))
 996                 mask |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
 997
 998         if (c->memory_accounting ||
 999             c->memory_limit != CGROUP_LIMIT_MAX ||
1000             cgroup_context_has_unified_memory_config(c))
1001                 mask |= CGROUP_MASK_MEMORY;
1002
1003         if (c->device_allow ||
1004             c->device_policy != CGROUP_AUTO)
1005                 mask |= CGROUP_MASK_DEVICES;
1006
1007         if (c->tasks_accounting ||
1008             c->tasks_max != (uint64_t) -1)
1009                 mask |= CGROUP_MASK_PIDS;
1010
1011         return mask;
1012 }
1013
1014 CGroupMask unit_get_own_mask(Unit *u) {
1015         CGroupContext *c;
1016
1017         /* Returns the mask of controllers the unit needs for itself */
1018
1019         c = unit_get_cgroup_context(u);
1020         if (!c)
1021                 return 0;
1022
1023         /* If delegation is turned on, then turn on all cgroups,
1024          * unless we are on the legacy hierarchy and the process we
1025          * fork into it is known to drop privileges, and hence
1026          * shouldn't get access to the controllers.
1027          *
1028          * Note that on the unified hierarchy it is safe to delegate
1029          * controllers to unprivileged services. */
1030
1031         if (c->delegate) {
1032                 ExecContext *e;
1033
1034                 e = unit_get_exec_context(u);
1035                 if (!e ||
1036                     exec_context_maintains_privileges(e) ||
1037                     cg_all_unified() > 0)
1038                         return _CGROUP_MASK_ALL;
1039         }
1040
1041         return cgroup_context_get_mask(c);
1042 }
1043
1044 CGroupMask unit_get_members_mask(Unit *u) {
1045         assert(u);
1046
1047         /* Returns the mask of controllers all of the unit's children
1048          * require, merged */
1049
1050         if (u->cgroup_members_mask_valid)
1051                 return u->cgroup_members_mask;
1052
1053         u->cgroup_members_mask = 0;
1054
1055         if (u->type == UNIT_SLICE) {
1056                 Unit *member;
1057                 Iterator i;
1058
1059                 SET_FOREACH(member, u->dependencies[UNIT_BEFORE], i) {
1060
1061                         if (member == u)
1062                                 continue;
1063
1064                         if (UNIT_DEREF(member->slice) != u)
1065                                 continue;
1066
1067                         u->cgroup_members_mask |=
1068                                 unit_get_own_mask(member) |
1069                                 unit_get_members_mask(member);
1070                 }
1071         }
1072
1073         u->cgroup_members_mask_valid = true;
1074         return u->cgroup_members_mask;
1075 }
1076
1077 CGroupMask unit_get_siblings_mask(Unit *u) {
1078         assert(u);
1079
1080         /* Returns the mask of controllers all of the unit's siblings
1081          * require, i.e. the members mask of the unit's parent slice
1082          * if there is one. */
1083
1084         if (UNIT_ISSET(u->slice))
1085                 return unit_get_members_mask(UNIT_DEREF(u->slice));
1086
1087         return unit_get_own_mask(u) | unit_get_members_mask(u);
1088 }
1089
1090 CGroupMask unit_get_subtree_mask(Unit *u) {
1091
1092         /* Returns the mask of this subtree, meaning of the group
1093          * itself and its children. */
1094
1095         return unit_get_own_mask(u) | unit_get_members_mask(u);
1096 }
1097
1098 CGroupMask unit_get_target_mask(Unit *u) {
1099         CGroupMask mask;
1100
1101         /* This returns the cgroup mask of all controllers to enable
1102          * for a specific cgroup, i.e. everything it needs itself,
1103          * plus all that its children need, plus all that its siblings
1104          * need. This is primarily useful on the legacy cgroup
1105          * hierarchy, where we need to duplicate each cgroup in each
1106          * hierarchy that shall be enabled for it. */
1107
1108         mask = unit_get_own_mask(u) | unit_get_members_mask(u) | unit_get_siblings_mask(u);
1109         mask &= u->manager->cgroup_supported;
1110
1111         return mask;
1112 }
1113
1114 CGroupMask unit_get_enable_mask(Unit *u) {
1115         CGroupMask mask;
1116
1117         /* This returns the cgroup mask of all controllers to enable
1118          * for the children of a specific cgroup. This is primarily
1119          * useful for the unified cgroup hierarchy, where each cgroup
1120          * controls which controllers are enabled for its children. */
1121
1122         mask = unit_get_members_mask(u);
1123         mask &= u->manager->cgroup_supported;
1124
1125         return mask;
1126 }
1127
1128 /* Recurse from a unit up through its containing slices, propagating
1129  * mask bits upward. A unit is also member of itself. */
1130 void unit_update_cgroup_members_masks(Unit *u) {
1131         CGroupMask m;
1132         bool more;
1133
1134         assert(u);
1135
1136         /* Calculate subtree mask */
1137         m = unit_get_subtree_mask(u);
1138
1139         /* See if anything changed from the previous invocation. If
1140          * not, we're done. */
1141         if (u->cgroup_subtree_mask_valid && m == u->cgroup_subtree_mask)
1142                 return;
1143
1144         more =
1145                 u->cgroup_subtree_mask_valid &&
1146                 ((m & ~u->cgroup_subtree_mask) != 0) &&
1147                 ((~m & u->cgroup_subtree_mask) == 0);
1148
1149         u->cgroup_subtree_mask = m;
1150         u->cgroup_subtree_mask_valid = true;
1151
1152         if (UNIT_ISSET(u->slice)) {
1153                 Unit *s = UNIT_DEREF(u->slice);
1154
1155                 if (more)
1156                         /* There's more set now than before. We
1157                          * propagate the new mask to the parent's mask
1158                          * (not caring if it actually was valid or
1159                          * not). */
1160
1161                         s->cgroup_members_mask |= m;
1162
1163                 else
1164                         /* There's less set now than before (or we
1165                          * don't know), we need to recalculate
1166                          * everything, so let's invalidate the
1167                          * parent's members mask */
1168
1169                         s->cgroup_members_mask_valid = false;
1170
1171                 /* And now make sure that this change also hits our
1172                  * grandparents */
1173                 unit_update_cgroup_members_masks(s);
1174         }
1175 }
1176
1177 static const char *migrate_callback(CGroupMask mask, void *userdata) {
1178         Unit *u = userdata;
1179
1180         assert(mask != 0);
1181         assert(u);
1182
1183         while (u) {
1184                 if (u->cgroup_path &&
1185                     u->cgroup_realized &&
1186                     (u->cgroup_realized_mask & mask) == mask)
1187                         return u->cgroup_path;
1188
1189                 u = UNIT_DEREF(u->slice);
1190         }
1191
1192         return NULL;
1193 }
1194
1195 char *unit_default_cgroup_path(Unit *u) {
1196         _cleanup_free_ char *escaped = NULL, *slice = NULL;
1197         int r;
1198
1199         assert(u);
1200
1201         if (unit_has_name(u, SPECIAL_ROOT_SLICE))
1202                 return strdup(u->manager->cgroup_root);
1203
1204         if (UNIT_ISSET(u->slice) && !unit_has_name(UNIT_DEREF(u->slice), SPECIAL_ROOT_SLICE)) {
1205                 r = cg_slice_to_path(UNIT_DEREF(u->slice)->id, &slice);
1206                 if (r < 0)
1207                         return NULL;
1208         }
1209
1210         escaped = cg_escape(u->id);
1211         if (!escaped)
1212                 return NULL;
1213
1214         if (slice)
1215                 return strjoin(u->manager->cgroup_root, "/", slice, "/",
1216                                escaped);
1217         else
1218                 return strjoin(u->manager->cgroup_root, "/", escaped);
1219 }
1220
1221 int unit_set_cgroup_path(Unit *u, const char *path) {
1222         _cleanup_free_ char *p = NULL;
1223         int r;
1224
1225         assert(u);
1226
1227         if (path) {
1228                 p = strdup(path);
1229                 if (!p)
1230                         return -ENOMEM;
1231         } else
1232                 p = NULL;
1233
1234         if (streq_ptr(u->cgroup_path, p))
1235                 return 0;
1236
1237         if (p) {
1238                 r = hashmap_put(u->manager->cgroup_unit, p, u);
1239                 if (r < 0)
1240                         return r;
1241         }
1242
1243         unit_release_cgroup(u);
1244
1245         u->cgroup_path = p;
1246         p = NULL;
1247
1248         return 1;
1249 }
1250
1251 int unit_watch_cgroup(Unit *u) {
1252         _cleanup_free_ char *events = NULL;
1253         int r;
1254
1255         assert(u);
1256
1257         if (!u->cgroup_path)
1258                 return 0;
1259
1260         if (u->cgroup_inotify_wd >= 0)
1261                 return 0;
1262
1263         /* Only applies to the unified hierarchy */
1264         r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1265         if (r < 0)
1266                 return log_error_errno(r, "Failed to determine whether the name=systemd hierarchy is unified: %m");
1267         if (r == 0)
1268                 return 0;
1269
1270         /* Don't watch the root slice, it's pointless. */
1271         if (unit_has_name(u, SPECIAL_ROOT_SLICE))
1272                 return 0;
1273
1274         r = hashmap_ensure_allocated(&u->manager->cgroup_inotify_wd_unit, &trivial_hash_ops);
1275         if (r < 0)
1276                 return log_oom();
1277
1278         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, "cgroup.events", &events);
1279         if (r < 0)
1280                 return log_oom();
1281
1282         u->cgroup_inotify_wd = inotify_add_watch(u->manager->cgroup_inotify_fd, events, IN_MODIFY);
1283         if (u->cgroup_inotify_wd < 0) {
1284
1285                 if (errno == ENOENT) /* If the directory is already
1286                                       * gone we don't need to track
1287                                       * it, so this is not an error */
1288                         return 0;
1289
1290                 return log_unit_error_errno(u, errno, "Failed to add inotify watch descriptor for control group %s: %m", u->cgroup_path);
1291         }
1292
1293         r = hashmap_put(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd), u);
1294         if (r < 0)
1295                 return log_unit_error_errno(u, r, "Failed to add inotify watch descriptor to hash map: %m");
1296
1297         return 0;
1298 }
1299
1300 static int unit_create_cgroup(
1301                 Unit *u,
1302                 CGroupMask target_mask,
1303                 CGroupMask enable_mask) {
1304
1305         CGroupContext *c;
1306         int r;
1307
1308         assert(u);
1309
1310         c = unit_get_cgroup_context(u);
1311         if (!c)
1312                 return 0;
1313
1314         if (!u->cgroup_path) {
1315                 _cleanup_free_ char *path = NULL;
1316
1317                 path = unit_default_cgroup_path(u);
1318                 if (!path)
1319                         return log_oom();
1320
1321                 r = unit_set_cgroup_path(u, path);
1322                 if (r == -EEXIST)
1323                         return log_unit_error_errno(u, r, "Control group %s exists already.", path);
1324                 if (r < 0)
1325                         return log_unit_error_errno(u, r, "Failed to set unit's control group path to %s: %m", path);
1326         }
1327
1328         /* First, create our own group */
1329         r = cg_create_everywhere(u->manager->cgroup_supported, target_mask, u->cgroup_path);
1330         if (r < 0)
1331                 return log_unit_error_errno(u, r, "Failed to create cgroup %s: %m", u->cgroup_path);
1332
1333         /* Start watching it */
1334         (void) unit_watch_cgroup(u);
1335
1336         /* Enable all controllers we need */
1337         r = cg_enable_everywhere(u->manager->cgroup_supported, enable_mask, u->cgroup_path);
1338         if (r < 0)
1339                 log_unit_warning_errno(u, r, "Failed to enable controllers on cgroup %s, ignoring: %m", u->cgroup_path);
1340
1341         /* Keep track that this is now realized */
1342         u->cgroup_realized = true;
1343         u->cgroup_realized_mask = target_mask;
1344         u->cgroup_enabled_mask = enable_mask;
1345
1346         if (u->type != UNIT_SLICE && !c->delegate) {
1347
1348                 /* Then, possibly move things over, but not if
1349                  * subgroups may contain processes, which is the case
1350                  * for slice and delegation units. */
1351                 r = cg_migrate_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->cgroup_path, migrate_callback, u);
1352                 if (r < 0)
1353                         log_unit_warning_errno(u, r, "Failed to migrate cgroup from to %s, ignoring: %m", u->cgroup_path);
1354         }
1355
1356         return 0;
1357 }
1358
1359 int unit_attach_pids_to_cgroup(Unit *u) {
1360         int r;
1361         assert(u);
1362
1363         r = unit_realize_cgroup(u);
1364         if (r < 0)
1365                 return r;
1366
1367         r = cg_attach_many_everywhere(u->manager->cgroup_supported, u->cgroup_path, u->pids, migrate_callback, u);
1368         if (r < 0)
1369                 return r;
1370
1371         return 0;
1372 }
1373
1374 static void cgroup_xattr_apply(Unit *u) {
1375         char ids[SD_ID128_STRING_MAX];
1376         int r;
1377
1378         assert(u);
1379
1380         if (!MANAGER_IS_SYSTEM(u->manager))
1381                 return;
1382
1383         if (sd_id128_is_null(u->invocation_id))
1384                 return;
1385
1386         r = cg_set_xattr(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path,
1387                          "trusted.invocation_id",
1388                          sd_id128_to_string(u->invocation_id, ids), 32,
1389                          0);
1390         if (r < 0)
1391                 log_unit_warning_errno(u, r, "Failed to set invocation ID on control group %s, ignoring: %m", u->cgroup_path);
1392 }
1393
1394 static bool unit_has_mask_realized(Unit *u, CGroupMask target_mask, CGroupMask enable_mask) {
1395         assert(u);
1396
1397         return u->cgroup_realized && u->cgroup_realized_mask == target_mask && u->cgroup_enabled_mask == enable_mask;
1398 }
1399
1400 /* Check if necessary controllers and attributes for a unit are in place.
1401  *
1402  * If so, do nothing.
1403  * If not, create paths, move processes over, and set attributes.
1404  *
1405  * Returns 0 on success and < 0 on failure. */
1406 static int unit_realize_cgroup_now(Unit *u, ManagerState state) {
1407         CGroupMask target_mask, enable_mask;
1408         int r;
1409
1410         assert(u);
1411
1412         if (u->in_cgroup_queue) {
1413                 LIST_REMOVE(cgroup_queue, u->manager->cgroup_queue, u);
1414                 u->in_cgroup_queue = false;
1415         }
1416
1417         target_mask = unit_get_target_mask(u);
1418         enable_mask = unit_get_enable_mask(u);
1419
1420         if (unit_has_mask_realized(u, target_mask, enable_mask))
1421                 return 0;
1422
1423         /* First, realize parents */
1424         if (UNIT_ISSET(u->slice)) {
1425                 r = unit_realize_cgroup_now(UNIT_DEREF(u->slice), state);
1426                 if (r < 0)
1427                         return r;
1428         }
1429
1430         /* And then do the real work */
1431         r = unit_create_cgroup(u, target_mask, enable_mask);
1432         if (r < 0)
1433                 return r;
1434
1435         /* Finally, apply the necessary attributes. */
1436         cgroup_context_apply(u, target_mask, state);
1437         cgroup_xattr_apply(u);
1438
1439         return 0;
1440 }
1441
1442 static void unit_add_to_cgroup_queue(Unit *u) {
1443
1444         if (u->in_cgroup_queue)
1445                 return;
1446
1447         LIST_PREPEND(cgroup_queue, u->manager->cgroup_queue, u);
1448         u->in_cgroup_queue = true;
1449 }
1450
1451 unsigned manager_dispatch_cgroup_queue(Manager *m) {
1452         ManagerState state;
1453         unsigned n = 0;
1454         Unit *i;
1455         int r;
1456
1457         state = manager_state(m);
1458
1459         while ((i = m->cgroup_queue)) {
1460                 assert(i->in_cgroup_queue);
1461
1462                 r = unit_realize_cgroup_now(i, state);
1463                 if (r < 0)
1464                         log_warning_errno(r, "Failed to realize cgroups for queued unit %s, ignoring: %m", i->id);
1465
1466                 n++;
1467         }
1468
1469         return n;
1470 }
1471
1472 static void unit_queue_siblings(Unit *u) {
1473         Unit *slice;
1474
1475         /* This adds the siblings of the specified unit and the
1476          * siblings of all parent units to the cgroup queue. (But
1477          * neither the specified unit itself nor the parents.) */
1478
1479         while ((slice = UNIT_DEREF(u->slice))) {
1480                 Iterator i;
1481                 Unit *m;
1482
1483                 SET_FOREACH(m, slice->dependencies[UNIT_BEFORE], i) {
1484                         if (m == u)
1485                                 continue;
1486
1487                         /* Skip units that have a dependency on the slice
1488                          * but aren't actually in it. */
1489                         if (UNIT_DEREF(m->slice) != slice)
1490                                 continue;
1491
1492                         /* No point in doing cgroup application for units
1493                          * without active processes. */
1494                         if (UNIT_IS_INACTIVE_OR_FAILED(unit_active_state(m)))
1495                                 continue;
1496
1497                         /* If the unit doesn't need any new controllers
1498                          * and has current ones realized, it doesn't need
1499                          * any changes. */
1500                         if (unit_has_mask_realized(m, unit_get_target_mask(m), unit_get_enable_mask(m)))
1501                                 continue;
1502
1503                         unit_add_to_cgroup_queue(m);
1504                 }
1505
1506                 u = slice;
1507         }
1508 }
1509
1510 int unit_realize_cgroup(Unit *u) {
1511         assert(u);
1512
1513         if (!UNIT_HAS_CGROUP_CONTEXT(u))
1514                 return 0;
1515
1516         /* So, here's the deal: when realizing the cgroups for this
1517          * unit, we need to first create all parents, but there's more
1518          * actually: for the weight-based controllers we also need to
1519          * make sure that all our siblings (i.e. units that are in the
1520          * same slice as we are) have cgroups, too. Otherwise, things
1521          * would become very uneven as each of their processes would
1522          * get as much resources as all our group together. This call
1523          * will synchronously create the parent cgroups, but will
1524          * defer work on the siblings to the next event loop
1525          * iteration. */
1526
1527         /* Add all sibling slices to the cgroup queue. */
1528         unit_queue_siblings(u);
1529
1530         /* And realize this one now (and apply the values) */
1531         return unit_realize_cgroup_now(u, manager_state(u->manager));
1532 }
1533
1534 void unit_release_cgroup(Unit *u) {
1535         assert(u);
1536
1537         /* Forgets all cgroup details for this cgroup */
1538
1539         if (u->cgroup_path) {
1540                 (void) hashmap_remove(u->manager->cgroup_unit, u->cgroup_path);
1541                 u->cgroup_path = mfree(u->cgroup_path);
1542         }
1543
1544         if (u->cgroup_inotify_wd >= 0) {
1545                 if (inotify_rm_watch(u->manager->cgroup_inotify_fd, u->cgroup_inotify_wd) < 0)
1546                         log_unit_debug_errno(u, errno, "Failed to remove cgroup inotify watch %i for %s, ignoring", u->cgroup_inotify_wd, u->id);
1547
1548                 (void) hashmap_remove(u->manager->cgroup_inotify_wd_unit, INT_TO_PTR(u->cgroup_inotify_wd));
1549                 u->cgroup_inotify_wd = -1;
1550         }
1551 }
1552
1553 void unit_prune_cgroup(Unit *u) {
1554         int r;
1555         bool is_root_slice;
1556
1557         assert(u);
1558
1559         /* Removes the cgroup, if empty and possible, and stops watching it. */
1560
1561         if (!u->cgroup_path)
1562                 return;
1563
1564         (void) unit_get_cpu_usage(u, NULL); /* Cache the last CPU usage value before we destroy the cgroup */
1565
1566         is_root_slice = unit_has_name(u, SPECIAL_ROOT_SLICE);
1567
1568         r = cg_trim_everywhere(u->manager->cgroup_supported, u->cgroup_path, !is_root_slice);
1569         if (r < 0) {
1570                 log_unit_debug_errno(u, r, "Failed to destroy cgroup %s, ignoring: %m", u->cgroup_path);
1571                 return;
1572         }
1573
1574         if (is_root_slice)
1575                 return;
1576
1577         unit_release_cgroup(u);
1578
1579         u->cgroup_realized = false;
1580         u->cgroup_realized_mask = 0;
1581         u->cgroup_enabled_mask = 0;
1582 }
1583
1584 int unit_search_main_pid(Unit *u, pid_t *ret) {
1585         _cleanup_fclose_ FILE *f = NULL;
1586         pid_t pid = 0, npid, mypid;
1587         int r;
1588
1589         assert(u);
1590         assert(ret);
1591
1592         if (!u->cgroup_path)
1593                 return -ENXIO;
1594
1595         r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, &f);
1596         if (r < 0)
1597                 return r;
1598
1599         mypid = getpid();
1600         while (cg_read_pid(f, &npid) > 0)  {
1601                 pid_t ppid;
1602
1603                 if (npid == pid)
1604                         continue;
1605
1606                 /* Ignore processes that aren't our kids */
1607                 if (get_process_ppid(npid, &ppid) >= 0 && ppid != mypid)
1608                         continue;
1609
1610                 if (pid != 0)
1611                         /* Dang, there's more than one daemonized PID
1612                         in this group, so we don't know what process
1613                         is the main process. */
1614
1615                         return -ENODATA;
1616
1617                 pid = npid;
1618         }
1619
1620         *ret = pid;
1621         return 0;
1622 }
1623
1624 static int unit_watch_pids_in_path(Unit *u, const char *path) {
1625         _cleanup_closedir_ DIR *d = NULL;
1626         _cleanup_fclose_ FILE *f = NULL;
1627         int ret = 0, r;
1628
1629         assert(u);
1630         assert(path);
1631
1632         r = cg_enumerate_processes(SYSTEMD_CGROUP_CONTROLLER, path, &f);
1633         if (r < 0)
1634                 ret = r;
1635         else {
1636                 pid_t pid;
1637
1638                 while ((r = cg_read_pid(f, &pid)) > 0) {
1639                         r = unit_watch_pid(u, pid);
1640                         if (r < 0 && ret >= 0)
1641                                 ret = r;
1642                 }
1643
1644                 if (r < 0 && ret >= 0)
1645                         ret = r;
1646         }
1647
1648         r = cg_enumerate_subgroups(SYSTEMD_CGROUP_CONTROLLER, path, &d);
1649         if (r < 0) {
1650                 if (ret >= 0)
1651                         ret = r;
1652         } else {
1653                 char *fn;
1654
1655                 while ((r = cg_read_subgroup(d, &fn)) > 0) {
1656                         _cleanup_free_ char *p = NULL;
1657
1658                         p = strjoin(path, "/", fn);
1659                         free(fn);
1660
1661                         if (!p)
1662                                 return -ENOMEM;
1663
1664                         r = unit_watch_pids_in_path(u, p);
1665                         if (r < 0 && ret >= 0)
1666                                 ret = r;
1667                 }
1668
1669                 if (r < 0 && ret >= 0)
1670                         ret = r;
1671         }
1672
1673         return ret;
1674 }
1675
1676 int unit_watch_all_pids(Unit *u) {
1677         int r;
1678
1679         assert(u);
1680
1681         /* Adds all PIDs from our cgroup to the set of PIDs we
1682          * watch. This is a fallback logic for cases where we do not
1683          * get reliable cgroup empty notifications: we try to use
1684          * SIGCHLD as replacement. */
1685
1686         if (!u->cgroup_path)
1687                 return -ENOENT;
1688
1689         r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1690         if (r < 0)
1691                 return r;
1692         if (r > 0) /* On unified we can use proper notifications */
1693                 return 0;
1694
1695         return unit_watch_pids_in_path(u, u->cgroup_path);
1696 }
1697
1698 int unit_notify_cgroup_empty(Unit *u) {
1699         int r;
1700
1701         assert(u);
1702
1703         if (!u->cgroup_path)
1704                 return 0;
1705
1706         r = cg_is_empty_recursive(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path);
1707         if (r <= 0)
1708                 return r;
1709
1710         unit_add_to_gc_queue(u);
1711
1712         if (UNIT_VTABLE(u)->notify_cgroup_empty)
1713                 UNIT_VTABLE(u)->notify_cgroup_empty(u);
1714
1715         return 0;
1716 }
1717
1718 static int on_cgroup_inotify_event(sd_event_source *s, int fd, uint32_t revents, void *userdata) {
1719         Manager *m = userdata;
1720
1721         assert(s);
1722         assert(fd >= 0);
1723         assert(m);
1724
1725         for (;;) {
1726                 union inotify_event_buffer buffer;
1727                 struct inotify_event *e;
1728                 ssize_t l;
1729
1730                 l = read(fd, &buffer, sizeof(buffer));
1731                 if (l < 0) {
1732                         if (errno == EINTR || errno == EAGAIN)
1733                                 return 0;
1734
1735                         return log_error_errno(errno, "Failed to read control group inotify events: %m");
1736                 }
1737
1738                 FOREACH_INOTIFY_EVENT(e, buffer, l) {
1739                         Unit *u;
1740
1741                         if (e->wd < 0)
1742                                 /* Queue overflow has no watch descriptor */
1743                                 continue;
1744
1745                         if (e->mask & IN_IGNORED)
1746                                 /* The watch was just removed */
1747                                 continue;
1748
1749                         u = hashmap_get(m->cgroup_inotify_wd_unit, INT_TO_PTR(e->wd));
1750                         if (!u) /* Not that inotify might deliver
1751                                  * events for a watch even after it
1752                                  * was removed, because it was queued
1753                                  * before the removal. Let's ignore
1754                                  * this here safely. */
1755                                 continue;
1756
1757                         (void) unit_notify_cgroup_empty(u);
1758                 }
1759         }
1760 }
1761 #endif // 0
1762
1763 int manager_setup_cgroup(Manager *m) {
1764         _cleanup_free_ char *path = NULL;
1765         CGroupController c;
1766         int r, all_unified;
1767         char *e;
1768
1769         assert(m);
1770
1771         /* 1. Determine hierarchy */
1772         m->cgroup_root = mfree(m->cgroup_root);
1773         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, 0, &m->cgroup_root);
1774         if (r < 0)
1775                 return log_error_errno(r, "Cannot determine cgroup we are running in: %m");
1776
1777 #if 0 /// elogind does not support systemd scopes and slices
1778         /* Chop off the init scope, if we are already located in it */
1779         e = endswith(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
1780
1781         /* LEGACY: Also chop off the system slice if we are in
1782          * it. This is to support live upgrades from older systemd
1783          * versions where PID 1 was moved there. Also see
1784          * cg_get_root_path(). */
1785         if (!e && MANAGER_IS_SYSTEM(m)) {
1786                 e = endswith(m->cgroup_root, "/" SPECIAL_SYSTEM_SLICE);
1787                 if (!e)
1788                         e = endswith(m->cgroup_root, "/system"); /* even more legacy */
1789         }
1790         if (e)
1791                 *e = 0;
1792 #endif // 0
1793
1794         /* And make sure to store away the root value without trailing
1795          * slash, even for the root dir, so that we can easily prepend
1796          * it everywhere. */
1797         while ((e = endswith(m->cgroup_root, "/")))
1798                 *e = 0;
1799         log_debug_elogind("Cgroup Controller \"%s\" -> root \"%s\"",
1800                           SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root);
1801
1802         /* 2. Show data */
1803         r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, NULL, &path);
1804         if (r < 0)
1805                 return log_error_errno(r, "Cannot find cgroup mount point: %m");
1806
1807         r = cg_unified_flush();
1808         if (r < 0)
1809                 return log_error_errno(r, "Couldn't determine if we are running in the unified hierarchy: %m");
1810
1811         all_unified = cg_all_unified();
1812         if (r < 0)
1813                 return log_error_errno(r, "Couldn't determine whether we are in all unified mode: %m");
1814         if (r > 0)
1815                 log_debug("Unified cgroup hierarchy is located at %s.", path);
1816         else {
1817                 r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1818                 if (r < 0)
1819                         return log_error_errno(r, "Failed to determine whether systemd's own controller is in unified mode: %m");
1820                 if (r > 0)
1821                         log_debug("Unified cgroup hierarchy is located at %s. Controllers are on legacy hierarchies.", path);
1822                 else
1823                         log_debug("Using cgroup controller " SYSTEMD_CGROUP_CONTROLLER_LEGACY ". File system hierarchy is at %s.", path);
1824         }
1825
1826         if (!m->test_run) {
1827                 const char *scope_path;
1828
1829                 /* 3. Install agent */
1830                 if (cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER) > 0) {
1831
1832                         /* In the unified hierarchy we can get
1833                          * cgroup empty notifications via inotify. */
1834
1835 #if 0 /// elogind does not support the unified hierarchy, yet.
1836                         m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
1837                         safe_close(m->cgroup_inotify_fd);
1838
1839                         m->cgroup_inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
1840                         if (m->cgroup_inotify_fd < 0)
1841                                 return log_error_errno(errno, "Failed to create control group inotify object: %m");
1842
1843                         r = sd_event_add_io(m->event, &m->cgroup_inotify_event_source, m->cgroup_inotify_fd, EPOLLIN, on_cgroup_inotify_event, m);
1844                         if (r < 0)
1845                                 return log_error_errno(r, "Failed to watch control group inotify object: %m");
1846
1847                         /* Process cgroup empty notifications early, but after service notifications and SIGCHLD. Also
1848                          * see handling of cgroup agent notifications, for the classic cgroup hierarchy support. */
1849                         r = sd_event_source_set_priority(m->cgroup_inotify_event_source, SD_EVENT_PRIORITY_NORMAL-5);
1850                         if (r < 0)
1851                                 return log_error_errno(r, "Failed to set priority of inotify event source: %m");
1852
1853                         (void) sd_event_source_set_description(m->cgroup_inotify_event_source, "cgroup-inotify");
1854
1855 #else
1856                         return log_error_errno(EOPNOTSUPP, "Unified cgroup hierarchy not supported: %m");
1857 #endif // 0
1858                 } else if (MANAGER_IS_SYSTEM(m)) {
1859
1860                         /* On the legacy hierarchy we only get
1861                          * notifications via cgroup agents. (Which
1862                          * isn't really reliable, since it does not
1863                          * generate events when control groups with
1864                          * children run empty. */
1865
1866                         r = cg_install_release_agent(SYSTEMD_CGROUP_CONTROLLER, SYSTEMD_CGROUP_AGENT_PATH);
1867                         if (r < 0)
1868                                 log_warning_errno(r, "Failed to install release agent, ignoring: %m");
1869                         else if (r > 0)
1870                                 log_debug("Installed release agent.");
1871                         else if (r == 0)
1872                                 log_debug("Release agent already installed.");
1873                 }
1874
1875 #if 0 /// elogind is not meant to run in systemd init scope
1876                 /* 4. Make sure we are in the special "init.scope" unit in the root slice. */
1877                 scope_path = strjoina(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
1878                 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1879 #else
1880                 if (streq(SYSTEMD_CGROUP_CONTROLLER, "name=elogind"))
1881                         // we are our own cgroup controller
1882                         scope_path = strjoina("");
1883                 else if (streq(m->cgroup_root, "/elogind"))
1884                         // root already is our cgroup
1885                         scope_path = strjoina(m->cgroup_root);
1886                 else
1887                         // we have to create our own group
1888                         scope_path = strjoina(m->cgroup_root, "/elogind");
1889                 r = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1890 #endif // 0
1891                 if (r < 0)
1892                         return log_error_errno(r, "Failed to create %s control group: %m", scope_path);
1893                 log_debug_elogind("Created control group \"%s\"", scope_path);
1894
1895                 /* also, move all other userspace processes remaining
1896                  * in the root cgroup into that scope. */
1897                 r = cg_migrate(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, SYSTEMD_CGROUP_CONTROLLER, scope_path, 0);
1898                 if (r < 0)
1899                         log_warning_errno(r, "Couldn't move remaining userspace processes, ignoring: %m");
1900
1901                 /* 5. And pin it, so that it cannot be unmounted */
1902                 safe_close(m->pin_cgroupfs_fd);
1903                 m->pin_cgroupfs_fd = open(path, O_RDONLY|O_CLOEXEC|O_DIRECTORY|O_NOCTTY|O_NONBLOCK);
1904                 if (m->pin_cgroupfs_fd < 0)
1905                         return log_error_errno(errno, "Failed to open pin file: %m");
1906
1907                 /* 6.  Always enable hierarchical support if it exists... */
1908                 if (!all_unified)
1909                         (void) cg_set_attribute("memory", "/", "memory.use_hierarchy", "1");
1910         }
1911
1912         /* 7. Figure out which controllers are supported */
1913         r = cg_mask_supported(&m->cgroup_supported);
1914         if (r < 0)
1915                 return log_error_errno(r, "Failed to determine supported controllers: %m");
1916
1917         for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++)
1918                 log_debug("Controller '%s' supported: %s", cgroup_controller_to_string(c), yes_no(m->cgroup_supported & CGROUP_CONTROLLER_TO_MASK(c)));
1919
1920         return 0;
1921 }
1922
1923 void manager_shutdown_cgroup(Manager *m, bool delete) {
1924         assert(m);
1925
1926         /* We can't really delete the group, since we are in it. But
1927          * let's trim it. */
1928         if (delete && m->cgroup_root)
1929                 (void) cg_trim(SYSTEMD_CGROUP_CONTROLLER, m->cgroup_root, false);
1930
1931 #if 0 /// elogind does not support the unified hierarchy, yet.
1932         m->cgroup_inotify_wd_unit = hashmap_free(m->cgroup_inotify_wd_unit);
1933
1934         m->cgroup_inotify_event_source = sd_event_source_unref(m->cgroup_inotify_event_source);
1935         m->cgroup_inotify_fd = safe_close(m->cgroup_inotify_fd);
1936 #endif // 0
1937
1938         m->pin_cgroupfs_fd = safe_close(m->pin_cgroupfs_fd);
1939
1940         m->cgroup_root = mfree(m->cgroup_root);
1941 }
1942
1943 #if 0 /// UNNEEDED by elogind
1944 Unit* manager_get_unit_by_cgroup(Manager *m, const char *cgroup) {
1945         char *p;
1946         Unit *u;
1947
1948         assert(m);
1949         assert(cgroup);
1950
1951         u = hashmap_get(m->cgroup_unit, cgroup);
1952         if (u)
1953                 return u;
1954
1955         p = strdupa(cgroup);
1956         for (;;) {
1957                 char *e;
1958
1959                 e = strrchr(p, '/');
1960                 if (!e || e == p)
1961                         return hashmap_get(m->cgroup_unit, SPECIAL_ROOT_SLICE);
1962
1963                 *e = 0;
1964
1965                 u = hashmap_get(m->cgroup_unit, p);
1966                 if (u)
1967                         return u;
1968         }
1969 }
1970
1971 Unit *manager_get_unit_by_pid_cgroup(Manager *m, pid_t pid) {
1972         _cleanup_free_ char *cgroup = NULL;
1973         int r;
1974
1975         assert(m);
1976
1977         if (pid <= 0)
1978                 return NULL;
1979
1980         r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
1981         if (r < 0)
1982                 return NULL;
1983
1984         return manager_get_unit_by_cgroup(m, cgroup);
1985 }
1986
1987 Unit *manager_get_unit_by_pid(Manager *m, pid_t pid) {
1988         Unit *u;
1989
1990         assert(m);
1991
1992         if (pid <= 0)
1993                 return NULL;
1994
1995         if (pid == 1)
1996                 return hashmap_get(m->units, SPECIAL_INIT_SCOPE);
1997
1998         u = hashmap_get(m->watch_pids1, PID_TO_PTR(pid));
1999         if (u)
2000                 return u;
2001
2002         u = hashmap_get(m->watch_pids2, PID_TO_PTR(pid));
2003         if (u)
2004                 return u;
2005
2006         return manager_get_unit_by_pid_cgroup(m, pid);
2007 }
2008 #endif // 0
2009
2010 #if 0 /// elogind must substitute this with its own variant
2011 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
2012         Unit *u;
2013
2014         assert(m);
2015         assert(cgroup);
2016
2017         log_debug("Got cgroup empty notification for: %s", cgroup);
2018
2019         u = manager_get_unit_by_cgroup(m, cgroup);
2020         if (!u)
2021                 return 0;
2022
2023         return unit_notify_cgroup_empty(u);
2024 }
2025 #else
2026 int manager_notify_cgroup_empty(Manager *m, const char *cgroup) {
2027         Session *s;
2028
2029         assert(m);
2030         assert(cgroup);
2031
2032         log_debug("Got cgroup empty notification for: %s", cgroup);
2033
2034         s = hashmap_get(m->sessions, cgroup);
2035
2036         if (s) {
2037                 session_finalize(s);
2038                 session_free(s);
2039         } else
2040                 log_warning("Session not found: %s", cgroup);
2041
2042         return 0;
2043 }
2044 #endif // 0
2045 #if 0 /// UNNEEDED by elogind
2046 int unit_get_memory_current(Unit *u, uint64_t *ret) {
2047         _cleanup_free_ char *v = NULL;
2048         int r;
2049
2050         assert(u);
2051         assert(ret);
2052
2053         if (!u->cgroup_path)
2054                 return -ENODATA;
2055
2056         if ((u->cgroup_realized_mask & CGROUP_MASK_MEMORY) == 0)
2057                 return -ENODATA;
2058
2059         r = cg_all_unified();
2060         if (r < 0)
2061                 return r;
2062         if (r > 0)
2063                 r = cg_get_attribute("memory", u->cgroup_path, "memory.current", &v);
2064         else
2065                 r = cg_get_attribute("memory", u->cgroup_path, "memory.usage_in_bytes", &v);
2066         if (r == -ENOENT)
2067                 return -ENODATA;
2068         if (r < 0)
2069                 return r;
2070
2071         return safe_atou64(v, ret);
2072 }
2073
2074 int unit_get_tasks_current(Unit *u, uint64_t *ret) {
2075         _cleanup_free_ char *v = NULL;
2076         int r;
2077
2078         assert(u);
2079         assert(ret);
2080
2081         if (!u->cgroup_path)
2082                 return -ENODATA;
2083
2084         if ((u->cgroup_realized_mask & CGROUP_MASK_PIDS) == 0)
2085                 return -ENODATA;
2086
2087         r = cg_get_attribute("pids", u->cgroup_path, "pids.current", &v);
2088         if (r == -ENOENT)
2089                 return -ENODATA;
2090         if (r < 0)
2091                 return r;
2092
2093         return safe_atou64(v, ret);
2094 }
2095
2096 static int unit_get_cpu_usage_raw(Unit *u, nsec_t *ret) {
2097         _cleanup_free_ char *v = NULL;
2098         uint64_t ns;
2099         int r;
2100
2101         assert(u);
2102         assert(ret);
2103
2104         if (!u->cgroup_path)
2105                 return -ENODATA;
2106
2107         r = cg_all_unified();
2108         if (r < 0)
2109                 return r;
2110         if (r > 0) {
2111                 const char *keys[] = { "usage_usec", NULL };
2112                 _cleanup_free_ char *val = NULL;
2113                 uint64_t us;
2114
2115                 if ((u->cgroup_realized_mask & CGROUP_MASK_CPU) == 0)
2116                         return -ENODATA;
2117
2118                 r = cg_get_keyed_attribute("cpu", u->cgroup_path, "cpu.stat", keys, &val);
2119                 if (r < 0)
2120                         return r;
2121
2122                 r = safe_atou64(val, &us);
2123                 if (r < 0)
2124                         return r;
2125
2126                 ns = us * NSEC_PER_USEC;
2127         } else {
2128                 if ((u->cgroup_realized_mask & CGROUP_MASK_CPUACCT) == 0)
2129                         return -ENODATA;
2130
2131                 r = cg_get_attribute("cpuacct", u->cgroup_path, "cpuacct.usage", &v);
2132                 if (r == -ENOENT)
2133                         return -ENODATA;
2134                 if (r < 0)
2135                         return r;
2136
2137                 r = safe_atou64(v, &ns);
2138                 if (r < 0)
2139                         return r;
2140         }
2141
2142         *ret = ns;
2143         return 0;
2144 }
2145
2146 int unit_get_cpu_usage(Unit *u, nsec_t *ret) {
2147         nsec_t ns;
2148         int r;
2149
2150         assert(u);
2151
2152         /* Retrieve the current CPU usage counter. This will subtract the CPU counter taken when the unit was
2153          * started. If the cgroup has been removed already, returns the last cached value. To cache the value, simply
2154          * call this function with a NULL return value. */
2155
2156         r = unit_get_cpu_usage_raw(u, &ns);
2157         if (r == -ENODATA && u->cpu_usage_last != NSEC_INFINITY) {
2158                 /* If we can't get the CPU usage anymore (because the cgroup was already removed, for example), use our
2159                  * cached value. */
2160
2161                 if (ret)
2162                         *ret = u->cpu_usage_last;
2163                 return 0;
2164         }
2165         if (r < 0)
2166                 return r;
2167
2168         if (ns > u->cpu_usage_base)
2169                 ns -= u->cpu_usage_base;
2170         else
2171                 ns = 0;
2172
2173         u->cpu_usage_last = ns;
2174         if (ret)
2175                 *ret = ns;
2176
2177         return 0;
2178 }
2179
2180 int unit_reset_cpu_usage(Unit *u) {
2181         nsec_t ns;
2182         int r;
2183
2184         assert(u);
2185
2186         u->cpu_usage_last = NSEC_INFINITY;
2187
2188         r = unit_get_cpu_usage_raw(u, &ns);
2189         if (r < 0) {
2190                 u->cpu_usage_base = 0;
2191                 return r;
2192         }
2193
2194         u->cpu_usage_base = ns;
2195         return 0;
2196 }
2197
2198 bool unit_cgroup_delegate(Unit *u) {
2199         CGroupContext *c;
2200
2201         assert(u);
2202
2203         c = unit_get_cgroup_context(u);
2204         if (!c)
2205                 return false;
2206
2207         return c->delegate;
2208 }
2209
2210 void unit_invalidate_cgroup(Unit *u, CGroupMask m) {
2211         assert(u);
2212
2213         if (!UNIT_HAS_CGROUP_CONTEXT(u))
2214                 return;
2215
2216         if (m == 0)
2217                 return;
2218
2219         /* always invalidate compat pairs together */
2220         if (m & (CGROUP_MASK_IO | CGROUP_MASK_BLKIO))
2221                 m |= CGROUP_MASK_IO | CGROUP_MASK_BLKIO;
2222
2223         if ((u->cgroup_realized_mask & m) == 0)
2224                 return;
2225
2226         u->cgroup_realized_mask &= ~m;
2227         unit_add_to_cgroup_queue(u);
2228 }
2229
2230 void manager_invalidate_startup_units(Manager *m) {
2231         Iterator i;
2232         Unit *u;
2233
2234         assert(m);
2235
2236         SET_FOREACH(u, m->startup_units, i)
2237                 unit_invalidate_cgroup(u, CGROUP_MASK_CPU|CGROUP_MASK_IO|CGROUP_MASK_BLKIO);
2238 }
2239
2240 static const char* const cgroup_device_policy_table[_CGROUP_DEVICE_POLICY_MAX] = {
2241         [CGROUP_AUTO] = "auto",
2242         [CGROUP_CLOSED] = "closed",
2243         [CGROUP_STRICT] = "strict",
2244 };
2245
2246 DEFINE_STRING_TABLE_LOOKUP(cgroup_device_policy, CGroupDevicePolicy);
2247 #endif // 0