cgroup: cgroup_subsys->fork() should be called after the task is added to css_set
[linux-3.10.git] / kernel / cgroup_freezer.c
1 /*
2  * cgroup_freezer.c -  control group freezer subsystem
3  *
4  * Copyright IBM Corporation, 2007
5  *
6  * Author : Cedric Le Goater <clg@fr.ibm.com>
7  *
8  * This program is free software; you can redistribute it and/or modify it
9  * under the terms of version 2.1 of the GNU Lesser General Public License
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it would be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
15  */
16
17 #include <linux/export.h>
18 #include <linux/slab.h>
19 #include <linux/cgroup.h>
20 #include <linux/fs.h>
21 #include <linux/uaccess.h>
22 #include <linux/freezer.h>
23 #include <linux/seq_file.h>
24
25 enum freezer_state {
26         CGROUP_THAWED = 0,
27         CGROUP_FREEZING,
28         CGROUP_FROZEN,
29 };
30
31 struct freezer {
32         struct cgroup_subsys_state css;
33         enum freezer_state state;
34         spinlock_t lock; /* protects _writes_ to state */
35 };
36
37 static inline struct freezer *cgroup_freezer(
38                 struct cgroup *cgroup)
39 {
40         return container_of(
41                 cgroup_subsys_state(cgroup, freezer_subsys_id),
42                 struct freezer, css);
43 }
44
45 static inline struct freezer *task_freezer(struct task_struct *task)
46 {
47         return container_of(task_subsys_state(task, freezer_subsys_id),
48                             struct freezer, css);
49 }
50
51 bool cgroup_freezing(struct task_struct *task)
52 {
53         enum freezer_state state;
54         bool ret;
55
56         rcu_read_lock();
57         state = task_freezer(task)->state;
58         ret = state == CGROUP_FREEZING || state == CGROUP_FROZEN;
59         rcu_read_unlock();
60
61         return ret;
62 }
63
64 /*
65  * cgroups_write_string() limits the size of freezer state strings to
66  * CGROUP_LOCAL_BUFFER_SIZE
67  */
68 static const char *freezer_state_strs[] = {
69         "THAWED",
70         "FREEZING",
71         "FROZEN",
72 };
73
74 /*
75  * State diagram
76  * Transitions are caused by userspace writes to the freezer.state file.
77  * The values in parenthesis are state labels. The rest are edge labels.
78  *
79  * (THAWED) --FROZEN--> (FREEZING) --FROZEN--> (FROZEN)
80  *    ^ ^                    |                     |
81  *    | \_______THAWED_______/                     |
82  *    \__________________________THAWED____________/
83  */
84
85 struct cgroup_subsys freezer_subsys;
86
87 /* Locks taken and their ordering
88  * ------------------------------
89  * cgroup_mutex (AKA cgroup_lock)
90  * freezer->lock
91  * css_set_lock
92  * task->alloc_lock (AKA task_lock)
93  * task->sighand->siglock
94  *
95  * cgroup code forces css_set_lock to be taken before task->alloc_lock
96  *
97  * freezer_create(), freezer_destroy():
98  * cgroup_mutex [ by cgroup core ]
99  *
100  * freezer_can_attach():
101  * cgroup_mutex (held by caller of can_attach)
102  *
103  * freezer_fork() (preserving fork() performance means can't take cgroup_mutex):
104  * freezer->lock
105  *  sighand->siglock (if the cgroup is freezing)
106  *
107  * freezer_read():
108  * cgroup_mutex
109  *  freezer->lock
110  *   write_lock css_set_lock (cgroup iterator start)
111  *    task->alloc_lock
112  *   read_lock css_set_lock (cgroup iterator start)
113  *
114  * freezer_write() (freeze):
115  * cgroup_mutex
116  *  freezer->lock
117  *   write_lock css_set_lock (cgroup iterator start)
118  *    task->alloc_lock
119  *   read_lock css_set_lock (cgroup iterator start)
120  *    sighand->siglock (fake signal delivery inside freeze_task())
121  *
122  * freezer_write() (unfreeze):
123  * cgroup_mutex
124  *  freezer->lock
125  *   write_lock css_set_lock (cgroup iterator start)
126  *    task->alloc_lock
127  *   read_lock css_set_lock (cgroup iterator start)
128  *    task->alloc_lock (inside __thaw_task(), prevents race with refrigerator())
129  *     sighand->siglock
130  */
131 static struct cgroup_subsys_state *freezer_create(struct cgroup *cgroup)
132 {
133         struct freezer *freezer;
134
135         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
136         if (!freezer)
137                 return ERR_PTR(-ENOMEM);
138
139         spin_lock_init(&freezer->lock);
140         freezer->state = CGROUP_THAWED;
141         return &freezer->css;
142 }
143
144 static void freezer_destroy(struct cgroup *cgroup)
145 {
146         struct freezer *freezer = cgroup_freezer(cgroup);
147
148         if (freezer->state != CGROUP_THAWED)
149                 atomic_dec(&system_freezing_cnt);
150         kfree(freezer);
151 }
152
153 /* task is frozen or will freeze immediately when next it gets woken */
154 static bool is_task_frozen_enough(struct task_struct *task)
155 {
156         return frozen(task) ||
157                 (task_is_stopped_or_traced(task) && freezing(task));
158 }
159
160 /*
161  * The call to cgroup_lock() in the freezer.state write method prevents
162  * a write to that file racing against an attach, and hence the
163  * can_attach() result will remain valid until the attach completes.
164  */
165 static int freezer_can_attach(struct cgroup *new_cgroup,
166                               struct cgroup_taskset *tset)
167 {
168         struct freezer *freezer;
169         struct task_struct *task;
170
171         /*
172          * Anything frozen can't move or be moved to/from.
173          */
174         cgroup_taskset_for_each(task, new_cgroup, tset)
175                 if (cgroup_freezing(task))
176                         return -EBUSY;
177
178         freezer = cgroup_freezer(new_cgroup);
179         if (freezer->state != CGROUP_THAWED)
180                 return -EBUSY;
181
182         return 0;
183 }
184
185 static void freezer_fork(struct task_struct *task)
186 {
187         struct freezer *freezer;
188
189         rcu_read_lock();
190         freezer = task_freezer(task);
191
192         /*
193          * The root cgroup is non-freezable, so we can skip the
194          * following check.
195          */
196         if (!freezer->css.cgroup->parent)
197                 goto out;
198
199         spin_lock_irq(&freezer->lock);
200         BUG_ON(freezer->state == CGROUP_FROZEN);
201
202         /* Locking avoids race with FREEZING -> THAWED transitions. */
203         if (freezer->state == CGROUP_FREEZING)
204                 freeze_task(task);
205
206         spin_unlock_irq(&freezer->lock);
207 out:
208         rcu_read_unlock();
209 }
210
211 /*
212  * caller must hold freezer->lock
213  */
214 static void update_if_frozen(struct cgroup *cgroup,
215                                  struct freezer *freezer)
216 {
217         struct cgroup_iter it;
218         struct task_struct *task;
219         unsigned int nfrozen = 0, ntotal = 0;
220         enum freezer_state old_state = freezer->state;
221
222         cgroup_iter_start(cgroup, &it);
223         while ((task = cgroup_iter_next(cgroup, &it))) {
224                 ntotal++;
225                 if (freezing(task) && is_task_frozen_enough(task))
226                         nfrozen++;
227         }
228
229         if (old_state == CGROUP_THAWED) {
230                 BUG_ON(nfrozen > 0);
231         } else if (old_state == CGROUP_FREEZING) {
232                 if (nfrozen == ntotal)
233                         freezer->state = CGROUP_FROZEN;
234         } else { /* old_state == CGROUP_FROZEN */
235                 BUG_ON(nfrozen != ntotal);
236         }
237
238         cgroup_iter_end(cgroup, &it);
239 }
240
241 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
242                         struct seq_file *m)
243 {
244         struct freezer *freezer;
245         enum freezer_state state;
246
247         if (!cgroup_lock_live_group(cgroup))
248                 return -ENODEV;
249
250         freezer = cgroup_freezer(cgroup);
251         spin_lock_irq(&freezer->lock);
252         state = freezer->state;
253         if (state == CGROUP_FREEZING) {
254                 /* We change from FREEZING to FROZEN lazily if the cgroup was
255                  * only partially frozen when we exitted write. */
256                 update_if_frozen(cgroup, freezer);
257                 state = freezer->state;
258         }
259         spin_unlock_irq(&freezer->lock);
260         cgroup_unlock();
261
262         seq_puts(m, freezer_state_strs[state]);
263         seq_putc(m, '\n');
264         return 0;
265 }
266
267 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
268 {
269         struct cgroup_iter it;
270         struct task_struct *task;
271         unsigned int num_cant_freeze_now = 0;
272
273         cgroup_iter_start(cgroup, &it);
274         while ((task = cgroup_iter_next(cgroup, &it))) {
275                 if (!freeze_task(task))
276                         continue;
277                 if (is_task_frozen_enough(task))
278                         continue;
279                 if (!freezing(task) && !freezer_should_skip(task))
280                         num_cant_freeze_now++;
281         }
282         cgroup_iter_end(cgroup, &it);
283
284         return num_cant_freeze_now ? -EBUSY : 0;
285 }
286
287 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
288 {
289         struct cgroup_iter it;
290         struct task_struct *task;
291
292         cgroup_iter_start(cgroup, &it);
293         while ((task = cgroup_iter_next(cgroup, &it)))
294                 __thaw_task(task);
295         cgroup_iter_end(cgroup, &it);
296 }
297
298 static int freezer_change_state(struct cgroup *cgroup,
299                                 enum freezer_state goal_state)
300 {
301         struct freezer *freezer;
302         int retval = 0;
303
304         freezer = cgroup_freezer(cgroup);
305
306         spin_lock_irq(&freezer->lock);
307
308         update_if_frozen(cgroup, freezer);
309
310         switch (goal_state) {
311         case CGROUP_THAWED:
312                 if (freezer->state != CGROUP_THAWED)
313                         atomic_dec(&system_freezing_cnt);
314                 freezer->state = CGROUP_THAWED;
315                 unfreeze_cgroup(cgroup, freezer);
316                 break;
317         case CGROUP_FROZEN:
318                 if (freezer->state == CGROUP_THAWED)
319                         atomic_inc(&system_freezing_cnt);
320                 freezer->state = CGROUP_FREEZING;
321                 retval = try_to_freeze_cgroup(cgroup, freezer);
322                 break;
323         default:
324                 BUG();
325         }
326
327         spin_unlock_irq(&freezer->lock);
328
329         return retval;
330 }
331
332 static int freezer_write(struct cgroup *cgroup,
333                          struct cftype *cft,
334                          const char *buffer)
335 {
336         int retval;
337         enum freezer_state goal_state;
338
339         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
340                 goal_state = CGROUP_THAWED;
341         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
342                 goal_state = CGROUP_FROZEN;
343         else
344                 return -EINVAL;
345
346         if (!cgroup_lock_live_group(cgroup))
347                 return -ENODEV;
348         retval = freezer_change_state(cgroup, goal_state);
349         cgroup_unlock();
350         return retval;
351 }
352
353 static struct cftype files[] = {
354         {
355                 .name = "state",
356                 .flags = CFTYPE_NOT_ON_ROOT,
357                 .read_seq_string = freezer_read,
358                 .write_string = freezer_write,
359         },
360         { }     /* terminate */
361 };
362
363 struct cgroup_subsys freezer_subsys = {
364         .name           = "freezer",
365         .create         = freezer_create,
366         .destroy        = freezer_destroy,
367         .subsys_id      = freezer_subsys_id,
368         .can_attach     = freezer_can_attach,
369         .fork           = freezer_fork,
370         .base_cftypes   = files,
371
372         /*
373          * freezer subsys doesn't handle hierarchy at all.  Frozen state
374          * should be inherited through the hierarchy - if a parent is
375          * frozen, all its children should be frozen.  Fix it and remove
376          * the following.
377          */
378         .broken_hierarchy = true,
379 };