chiark / gitweb /
service: optionally, call setsid() on services
[elogind.git] / cgroup.c
1 /*-*- Mode: C; c-basic-offset: 8 -*-*/
2
3 /***
4   This file is part of systemd.
5
6   Copyright 2010 Lennart Poettering
7
8   systemd is free software; you can redistribute it and/or modify it
9   under the terms of the GNU General Public License as published by
10   the Free Software Foundation; either version 2 of the License, or
11   (at your option) any later version.
12
13   systemd is distributed in the hope that it will be useful, but
14   WITHOUT ANY WARRANTY; without even the implied warranty of
15   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16   General Public License for more details.
17
18   You should have received a copy of the GNU General Public License
19   along with systemd; If not, see <http://www.gnu.org/licenses/>.
20 ***/
21
22 #include <errno.h>
23 #include <assert.h>
24 #include <unistd.h>
25 #include <sys/types.h>
26 #include <signal.h>
27 #include <sys/mount.h>
28
29 #include "cgroup.h"
30 #include "log.h"
31
32 static int translate_error(int error, int _errno) {
33
34         switch (error) {
35
36         case ECGROUPNOTCOMPILED:
37         case ECGROUPNOTMOUNTED:
38         case ECGROUPNOTEXIST:
39         case ECGROUPNOTCREATED:
40                 return -ENOENT;
41
42         case ECGINVAL:
43                 return -EINVAL;
44
45         case ECGROUPNOTALLOWED:
46                 return -EPERM;
47
48         case ECGOTHER:
49                 return -_errno;
50         }
51
52         return -EIO;
53 }
54
55 int cgroup_bonding_realize(CGroupBonding *b) {
56         int r;
57
58         assert(b);
59         assert(b->path);
60         assert(b->controller);
61
62         if (b->cgroup)
63                 return 0;
64
65         if (!(b->cgroup = cgroup_new_cgroup(b->path)))
66                 return -ENOMEM;
67
68         if (!cgroup_add_controller(b->cgroup, b->controller)) {
69                 r = -ENOMEM;
70                 goto fail;
71         }
72
73         if (b->inherit)
74                 r = cgroup_create_cgroup_from_parent(b->cgroup, true);
75         else
76                 r = cgroup_create_cgroup(b->cgroup, true);
77
78         if (r != 0) {
79                 r = translate_error(r, errno);
80                 goto fail;
81         }
82
83         return 0;
84
85 fail:
86         cgroup_free(&b->cgroup);
87         b->cgroup = NULL;
88         return r;
89 }
90
91 int cgroup_bonding_realize_list(CGroupBonding *first) {
92         CGroupBonding *b;
93
94         LIST_FOREACH(by_unit, b, first) {
95                 int r;
96
97                 if ((r = cgroup_bonding_realize(b)) < 0)
98                         return r;
99         }
100
101         return 0;
102 }
103
104 void cgroup_bonding_free(CGroupBonding *b) {
105         assert(b);
106
107         if (b->unit) {
108                 CGroupBonding *f;
109
110                 LIST_REMOVE(CGroupBonding, by_unit, b->unit->meta.cgroup_bondings, b);
111
112                 assert_se(f = hashmap_get(b->unit->meta.manager->cgroup_bondings, b->path));
113                 LIST_REMOVE(CGroupBonding, by_path, f, b);
114
115                 if (f)
116                         hashmap_replace(b->unit->meta.manager->cgroup_bondings, b->path, f);
117                 else
118                         hashmap_remove(b->unit->meta.manager->cgroup_bondings, b->path);
119         }
120
121         free(b->controller);
122         free(b->path);
123
124         if (b->cgroup) {
125
126                 if (b->only_us && b->clean_up)
127                         cgroup_delete_cgroup(b->cgroup, true);
128
129                 cgroup_free(&b->cgroup);
130         }
131
132         free(b);
133 }
134
135 void cgroup_bonding_free_list(CGroupBonding *first) {
136         CGroupBonding *b, *n;
137
138         LIST_FOREACH_SAFE(by_unit, b, n, first)
139                 cgroup_bonding_free(b);
140 }
141
142 int cgroup_bonding_install(CGroupBonding *b, pid_t pid) {
143         int r;
144
145         assert(b);
146         assert(pid >= 0);
147
148         if (pid == 0)
149                 pid = getpid();
150
151         if (!b->cgroup)
152                 return -ENOENT;
153
154         if ((r = cgroup_attach_task_pid(b->cgroup, pid)))
155                 return translate_error(r, errno);
156
157         return 0;
158 }
159
160 int cgroup_bonding_install_list(CGroupBonding *first, pid_t pid) {
161         CGroupBonding *b;
162
163         LIST_FOREACH(by_unit, b, first) {
164                 int r;
165
166                 if ((r = cgroup_bonding_install(b, pid)) < 0)
167                         return r;
168         }
169
170         return 0;
171 }
172
173 int cgroup_bonding_kill(CGroupBonding *b, int sig) {
174         int r;
175         Set *s;
176         bool done;
177
178         assert(b);
179         assert(sig > 0);
180
181         if (!(s = set_new(trivial_hash_func, trivial_compare_func)))
182                 return -ENOMEM;
183
184         do {
185                 void *iterator;
186                 pid_t pid;
187
188                 done = true;
189
190                 if ((r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid)) != 0) {
191                         if (r == ECGEOF) {
192                                 r = 0;
193                                 goto kill_done;
194                         } else {
195                                 r = translate_error(r, errno);
196                                 break;
197                         }
198                 }
199
200                 for (;;) {
201                         if (set_get(s, INT_TO_PTR(pid)) != INT_TO_PTR(pid)) {
202
203                                 /* If we haven't killed this process
204                                  * yet, kill it */
205
206                                 if (kill(pid, sig) < 0 && errno != ESRCH) {
207                                         r = -errno;
208                                         break;
209                                 }
210
211                                 done = false;
212
213                                 if ((r = set_put(s, INT_TO_PTR(pid))) < 0)
214                                     break;
215                         }
216
217                         if ((r = cgroup_get_task_next(&iterator, &pid)) != 0) {
218
219                                 if (r == ECGEOF)
220                                         r = 0;
221                                 else
222                                         r = translate_error(r, errno);
223
224                                 break;
225                         }
226                 }
227
228         kill_done:
229                 assert_se(cgroup_get_task_end(&iterator) == 0);
230
231                 /* To avoid racing against processes which fork
232                  * quicker than we can kill them we repeat this until
233                  * no new pids need to be killed. */
234
235         } while (!done && r >= 0);
236
237         set_free(s);
238         return r;
239 }
240
241 int cgroup_bonding_kill_list(CGroupBonding *first, int sig) {
242         CGroupBonding *b;
243
244         LIST_FOREACH(by_unit, b, first) {
245                 int r;
246
247                 if ((r = cgroup_bonding_kill(b, sig)) < 0)
248                         return r;
249         }
250
251         return 0;
252 }
253
254 /* Returns 1 if the group is empty, 0 if it is not, -EAGAIN if we
255  * cannot know */
256 int cgroup_bonding_is_empty(CGroupBonding *b) {
257         void *iterator;
258         pid_t pid;
259         int r;
260
261         assert(b);
262
263         r = cgroup_get_task_begin(b->path, b->controller, &iterator, &pid);
264
265         if (r == 0 || r == ECGEOF)
266                 cgroup_get_task_end(&iterator);
267
268         /* Hmm, no PID in this group? Then it is definitely empty */
269         if (r == ECGEOF)
270                 return 1;
271
272         /* Some error? Let's return it */
273         if (r != 0)
274                 return translate_error(r, errno);
275
276         /* It's not empty, and we are the only user, then it is
277          * definitely not empty */
278         if (b->only_us)
279                 return 0;
280
281         /* There are PIDs in the group but we aren't the only users,
282          * hence we cannot say */
283         return -EAGAIN;
284 }
285
286 int cgroup_bonding_is_empty_list(CGroupBonding *first) {
287         CGroupBonding *b;
288
289         LIST_FOREACH(by_unit, b, first) {
290                 int r;
291
292                 if ((r = cgroup_bonding_is_empty(b)) < 0) {
293                         /* If this returned -EAGAIN, then we don't know if the
294                          * group is empty, so let's see if another group can
295                          * tell us */
296
297                         if (r != -EAGAIN)
298                                 return r;
299                 } else
300                         return r;
301         }
302
303         return -EAGAIN;
304 }
305
306 static int install_release_agent(Manager *m, const char *mount_point) {
307         char *p, *c, *sc;
308         int r;
309
310         assert(m);
311         assert(mount_point);
312
313         if (asprintf(&p, "%s/release_agent", mount_point) < 0)
314                 return -ENOMEM;
315
316         if ((r = read_one_line_file(p, &c)) < 0) {
317                 free(p);
318                 return r;
319         }
320
321         sc = strstrip(c);
322
323         if (sc[0] == 0) {
324                 if ((r = write_one_line_file(p, CGROUP_AGENT_PATH "\n" )) < 0) {
325                         free(p);
326                         free(c);
327                         return r;
328                 }
329         } else if (!streq(sc, CGROUP_AGENT_PATH)) {
330                 free(p);
331                 free(c);
332                 return -EEXIST;
333         }
334
335         free(c);
336         free(p);
337
338         if (asprintf(&p, "%s/notify_on_release", mount_point) < 0)
339                 return -ENOMEM;
340
341         if ((r = read_one_line_file(p, &c)) < 0) {
342                 free(p);
343                 return r;
344         }
345
346         sc = strstrip(c);
347
348         if (streq(sc, "0")) {
349                 if ((r = write_one_line_file(p, "1\n")) < 0) {
350                         free(p);
351                         free(c);
352                         return r;
353                 }
354         } else if (!streq(sc, "1")) {
355                 free(p);
356                 free(c);
357                 return -EIO;
358         }
359
360         free(p);
361         free(c);
362
363         return 0;
364 }
365
366 static int create_hierarchy_cgroup(Manager *m) {
367         struct cgroup *cg;
368         int r;
369
370         assert(m);
371
372         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
373                 return -ENOMEM;
374
375         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
376                 r = -ENOMEM;
377                 goto finish;
378         }
379
380         if ((r = cgroup_create_cgroup(cg, true)) != 0) {
381                 log_error("Failed to create cgroup hierarchy group: %s", cgroup_strerror(r));
382                 r = translate_error(r, errno);
383                 goto finish;
384         }
385
386         if ((r = cgroup_attach_task(cg)) != 0) {
387                 log_error("Failed to add ourselves to hierarchy group: %s", cgroup_strerror(r));
388                 r = translate_error(r, errno);
389                 goto finish;
390         }
391
392         r = 0;
393
394 finish:
395         cgroup_free(&cg);
396         return r;
397 }
398
399 int manager_setup_cgroup(Manager *m) {
400         char *mp, *cp;
401         int r;
402         pid_t pid;
403
404         assert(m);
405
406         if ((r = cgroup_init()) != 0) {
407                 log_error("Failed to initialize libcg: %s", cgroup_strerror(r));
408                 return translate_error(r, errno);
409         }
410
411         free(m->cgroup_controller);
412         if (!(m->cgroup_controller = strdup("debug")))
413                 return -ENOMEM;
414
415         if ((r = cgroup_get_subsys_mount_point(m->cgroup_controller, &mp)))
416                 return translate_error(r, errno);
417
418         pid = getpid();
419
420         if ((r = cgroup_get_current_controller_path(pid, m->cgroup_controller, &cp))) {
421                 free(mp);
422                 return translate_error(r, errno);
423         }
424
425         free(m->cgroup_hierarchy);
426         m->cgroup_hierarchy = NULL;
427         if (asprintf(&m->cgroup_hierarchy, "%s/systemd-%llu", strcmp(cp, "/") == 0 ? "" : cp, (unsigned long long) pid) < 0) {
428                 free(cp);
429                 free(mp);
430                 return -ENOMEM;
431         }
432
433         log_info("Using cgroup controller <%s>, hierarchy mounted at <%s>, using root group <%s>.",
434                  m->cgroup_controller,
435                  mp,
436                  m->cgroup_hierarchy);
437
438         if ((r = install_release_agent(m, mp)) < 0)
439                 log_warning("Failed to install release agent, ignoring: %s", strerror(-r));
440         else
441                 log_info("Installed release agent, or already installed.");
442
443         free(mp);
444         free(cp);
445
446         if ((r = create_hierarchy_cgroup(m)) < 0)
447                 log_error("Failed to create root cgroup hierarchy: %s", strerror(-r));
448         else
449                 log_info("Created root group.");
450
451         return r;
452 }
453
454 int manager_shutdown_cgroup(Manager *m) {
455         struct cgroup *cg;
456         int r;
457
458         assert(m);
459
460         if (!m->cgroup_hierarchy)
461                 return 0;
462
463         if (!(cg = cgroup_new_cgroup(m->cgroup_hierarchy)))
464                 return -ENOMEM;
465
466         if (!(cgroup_add_controller(cg, m->cgroup_controller))) {
467                 r = -ENOMEM;
468                 goto finish;
469         }
470
471         if ((r = cgroup_delete_cgroup_ext(cg, CGFLAG_DELETE_IGNORE_MIGRATION|CGFLAG_DELETE_RECURSIVE)) != 0) {
472                 log_error("Failed to delete cgroup hierarchy group: %s", cgroup_strerror(r));
473                 r = translate_error(r, errno);
474                 goto finish;
475         }
476         r = 0;
477
478 finish:
479         cgroup_free(&cg);
480         return r;
481
482 }
483
484 int cgroup_notify_empty(Manager *m, const char *group) {
485         CGroupBonding *l, *b;
486
487         assert(m);
488         assert(group);
489
490         if (!(l = hashmap_get(m->cgroup_bondings, group)))
491                 return 0;
492
493         LIST_FOREACH(by_path, b, l) {
494                 int t;
495
496                 if (!b->unit)
497                         continue;
498
499                 if ((t = cgroup_bonding_is_empty_list(b)) < 0) {
500
501                         /* If we don't know, we don't know */
502                         if (t != -EAGAIN)
503                                 log_warning("Failed to check whether cgroup is empty: %s", strerror(errno));
504
505                         continue;
506                 }
507
508                 if (t > 0)
509                         if (UNIT_VTABLE(b->unit)->cgroup_notify_empty)
510                                 UNIT_VTABLE(b->unit)->cgroup_notify_empty(b->unit);
511         }
512
513         return 0;
514 }
515
516 CGroupBonding *cgroup_bonding_find_list(CGroupBonding *first, const char *controller) {
517         CGroupBonding *b;
518
519         assert(controller);
520
521         LIST_FOREACH(by_unit, b, first)
522                 if (streq(b->controller, controller))
523                         return b;
524
525         return NULL;
526 }