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