Merge branch 'for-3.3' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup
[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_subsys *ss,
132                                                   struct cgroup *cgroup)
133 {
134         struct freezer *freezer;
135
136         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
137         if (!freezer)
138                 return ERR_PTR(-ENOMEM);
139
140         spin_lock_init(&freezer->lock);
141         freezer->state = CGROUP_THAWED;
142         return &freezer->css;
143 }
144
145 static void freezer_destroy(struct cgroup_subsys *ss,
146                             struct cgroup *cgroup)
147 {
148         struct freezer *freezer = cgroup_freezer(cgroup);
149
150         if (freezer->state != CGROUP_THAWED)
151                 atomic_dec(&system_freezing_cnt);
152         kfree(freezer);
153 }
154
155 /* task is frozen or will freeze immediately when next it gets woken */
156 static bool is_task_frozen_enough(struct task_struct *task)
157 {
158         return frozen(task) ||
159                 (task_is_stopped_or_traced(task) && freezing(task));
160 }
161
162 /*
163  * The call to cgroup_lock() in the freezer.state write method prevents
164  * a write to that file racing against an attach, and hence the
165  * can_attach() result will remain valid until the attach completes.
166  */
167 static int freezer_can_attach(struct cgroup_subsys *ss,
168                               struct cgroup *new_cgroup,
169                               struct cgroup_taskset *tset)
170 {
171         struct freezer *freezer;
172         struct task_struct *task;
173
174         /*
175          * Anything frozen can't move or be moved to/from.
176          */
177         cgroup_taskset_for_each(task, new_cgroup, tset)
178                 if (cgroup_freezing(task))
179                         return -EBUSY;
180
181         freezer = cgroup_freezer(new_cgroup);
182         if (freezer->state != CGROUP_THAWED)
183                 return -EBUSY;
184
185         return 0;
186 }
187
188 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
189 {
190         struct freezer *freezer;
191
192         /*
193          * No lock is needed, since the task isn't on tasklist yet,
194          * so it can't be moved to another cgroup, which means the
195          * freezer won't be removed and will be valid during this
196          * function call.  Nevertheless, apply RCU read-side critical
197          * section to suppress RCU lockdep false positives.
198          */
199         rcu_read_lock();
200         freezer = task_freezer(task);
201         rcu_read_unlock();
202
203         /*
204          * The root cgroup is non-freezable, so we can skip the
205          * following check.
206          */
207         if (!freezer->css.cgroup->parent)
208                 return;
209
210         spin_lock_irq(&freezer->lock);
211         BUG_ON(freezer->state == CGROUP_FROZEN);
212
213         /* Locking avoids race with FREEZING -> THAWED transitions. */
214         if (freezer->state == CGROUP_FREEZING)
215                 freeze_task(task);
216         spin_unlock_irq(&freezer->lock);
217 }
218
219 /*
220  * caller must hold freezer->lock
221  */
222 static void update_if_frozen(struct cgroup *cgroup,
223                                  struct freezer *freezer)
224 {
225         struct cgroup_iter it;
226         struct task_struct *task;
227         unsigned int nfrozen = 0, ntotal = 0;
228         enum freezer_state old_state = freezer->state;
229
230         cgroup_iter_start(cgroup, &it);
231         while ((task = cgroup_iter_next(cgroup, &it))) {
232                 ntotal++;
233                 if (freezing(task) && is_task_frozen_enough(task))
234                         nfrozen++;
235         }
236
237         if (old_state == CGROUP_THAWED) {
238                 BUG_ON(nfrozen > 0);
239         } else if (old_state == CGROUP_FREEZING) {
240                 if (nfrozen == ntotal)
241                         freezer->state = CGROUP_FROZEN;
242         } else { /* old_state == CGROUP_FROZEN */
243                 BUG_ON(nfrozen != ntotal);
244         }
245
246         cgroup_iter_end(cgroup, &it);
247 }
248
249 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
250                         struct seq_file *m)
251 {
252         struct freezer *freezer;
253         enum freezer_state state;
254
255         if (!cgroup_lock_live_group(cgroup))
256                 return -ENODEV;
257
258         freezer = cgroup_freezer(cgroup);
259         spin_lock_irq(&freezer->lock);
260         state = freezer->state;
261         if (state == CGROUP_FREEZING) {
262                 /* We change from FREEZING to FROZEN lazily if the cgroup was
263                  * only partially frozen when we exitted write. */
264                 update_if_frozen(cgroup, freezer);
265                 state = freezer->state;
266         }
267         spin_unlock_irq(&freezer->lock);
268         cgroup_unlock();
269
270         seq_puts(m, freezer_state_strs[state]);
271         seq_putc(m, '\n');
272         return 0;
273 }
274
275 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
276 {
277         struct cgroup_iter it;
278         struct task_struct *task;
279         unsigned int num_cant_freeze_now = 0;
280
281         cgroup_iter_start(cgroup, &it);
282         while ((task = cgroup_iter_next(cgroup, &it))) {
283                 if (!freeze_task(task))
284                         continue;
285                 if (is_task_frozen_enough(task))
286                         continue;
287                 if (!freezing(task) && !freezer_should_skip(task))
288                         num_cant_freeze_now++;
289         }
290         cgroup_iter_end(cgroup, &it);
291
292         return num_cant_freeze_now ? -EBUSY : 0;
293 }
294
295 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
296 {
297         struct cgroup_iter it;
298         struct task_struct *task;
299
300         cgroup_iter_start(cgroup, &it);
301         while ((task = cgroup_iter_next(cgroup, &it)))
302                 __thaw_task(task);
303         cgroup_iter_end(cgroup, &it);
304 }
305
306 static int freezer_change_state(struct cgroup *cgroup,
307                                 enum freezer_state goal_state)
308 {
309         struct freezer *freezer;
310         int retval = 0;
311
312         freezer = cgroup_freezer(cgroup);
313
314         spin_lock_irq(&freezer->lock);
315
316         update_if_frozen(cgroup, freezer);
317
318         switch (goal_state) {
319         case CGROUP_THAWED:
320                 if (freezer->state != CGROUP_THAWED)
321                         atomic_dec(&system_freezing_cnt);
322                 freezer->state = CGROUP_THAWED;
323                 unfreeze_cgroup(cgroup, freezer);
324                 break;
325         case CGROUP_FROZEN:
326                 if (freezer->state == CGROUP_THAWED)
327                         atomic_inc(&system_freezing_cnt);
328                 freezer->state = CGROUP_FREEZING;
329                 retval = try_to_freeze_cgroup(cgroup, freezer);
330                 break;
331         default:
332                 BUG();
333         }
334
335         spin_unlock_irq(&freezer->lock);
336
337         return retval;
338 }
339
340 static int freezer_write(struct cgroup *cgroup,
341                          struct cftype *cft,
342                          const char *buffer)
343 {
344         int retval;
345         enum freezer_state goal_state;
346
347         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
348                 goal_state = CGROUP_THAWED;
349         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
350                 goal_state = CGROUP_FROZEN;
351         else
352                 return -EINVAL;
353
354         if (!cgroup_lock_live_group(cgroup))
355                 return -ENODEV;
356         retval = freezer_change_state(cgroup, goal_state);
357         cgroup_unlock();
358         return retval;
359 }
360
361 static struct cftype files[] = {
362         {
363                 .name = "state",
364                 .read_seq_string = freezer_read,
365                 .write_string = freezer_write,
366         },
367 };
368
369 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
370 {
371         if (!cgroup->parent)
372                 return 0;
373         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
374 }
375
376 struct cgroup_subsys freezer_subsys = {
377         .name           = "freezer",
378         .create         = freezer_create,
379         .destroy        = freezer_destroy,
380         .populate       = freezer_populate,
381         .subsys_id      = freezer_subsys_id,
382         .can_attach     = freezer_can_attach,
383         .fork           = freezer_fork,
384 };