Merge branch 'master' into export-slabh
[kernel.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/module.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 int cgroup_freezing_or_frozen(struct task_struct *task)
52 {
53         struct freezer *freezer;
54         enum freezer_state state;
55
56         task_lock(task);
57         freezer = task_freezer(task);
58         if (!freezer->css.cgroup->parent)
59                 state = CGROUP_THAWED; /* root cgroup can't be frozen */
60         else
61                 state = freezer->state;
62         task_unlock(task);
63
64         return (state == CGROUP_FREEZING) || (state == CGROUP_FROZEN);
65 }
66
67 /*
68  * cgroups_write_string() limits the size of freezer state strings to
69  * CGROUP_LOCAL_BUFFER_SIZE
70  */
71 static const char *freezer_state_strs[] = {
72         "THAWED",
73         "FREEZING",
74         "FROZEN",
75 };
76
77 /*
78  * State diagram
79  * Transitions are caused by userspace writes to the freezer.state file.
80  * The values in parenthesis are state labels. The rest are edge labels.
81  *
82  * (THAWED) --FROZEN--> (FREEZING) --FROZEN--> (FROZEN)
83  *    ^ ^                    |                     |
84  *    | \_______THAWED_______/                     |
85  *    \__________________________THAWED____________/
86  */
87
88 struct cgroup_subsys freezer_subsys;
89
90 /* Locks taken and their ordering
91  * ------------------------------
92  * css_set_lock
93  * cgroup_mutex (AKA cgroup_lock)
94  * task->alloc_lock (AKA task_lock)
95  * freezer->lock
96  * task->sighand->siglock
97  *
98  * cgroup code forces css_set_lock to be taken before task->alloc_lock
99  *
100  * freezer_create(), freezer_destroy():
101  * cgroup_mutex [ by cgroup core ]
102  *
103  * can_attach():
104  * cgroup_mutex
105  *
106  * cgroup_frozen():
107  * task->alloc_lock (to get task's cgroup)
108  *
109  * freezer_fork() (preserving fork() performance means can't take cgroup_mutex):
110  * task->alloc_lock (to get task's cgroup)
111  * freezer->lock
112  *  sighand->siglock (if the cgroup is freezing)
113  *
114  * freezer_read():
115  * cgroup_mutex
116  *  freezer->lock
117  *   read_lock css_set_lock (cgroup iterator start)
118  *
119  * freezer_write() (freeze):
120  * cgroup_mutex
121  *  freezer->lock
122  *   read_lock css_set_lock (cgroup iterator start)
123  *    sighand->siglock
124  *
125  * freezer_write() (unfreeze):
126  * cgroup_mutex
127  *  freezer->lock
128  *   read_lock css_set_lock (cgroup iterator start)
129  *    task->alloc_lock (to prevent races with freeze_task())
130  *     sighand->siglock
131  */
132 static struct cgroup_subsys_state *freezer_create(struct cgroup_subsys *ss,
133                                                   struct cgroup *cgroup)
134 {
135         struct freezer *freezer;
136
137         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
138         if (!freezer)
139                 return ERR_PTR(-ENOMEM);
140
141         spin_lock_init(&freezer->lock);
142         freezer->state = CGROUP_THAWED;
143         return &freezer->css;
144 }
145
146 static void freezer_destroy(struct cgroup_subsys *ss,
147                             struct cgroup *cgroup)
148 {
149         kfree(cgroup_freezer(cgroup));
150 }
151
152 /* Task is frozen or will freeze immediately when next it gets woken */
153 static bool is_task_frozen_enough(struct task_struct *task)
154 {
155         return frozen(task) ||
156                 (task_is_stopped_or_traced(task) && freezing(task));
157 }
158
159 /*
160  * The call to cgroup_lock() in the freezer.state write method prevents
161  * a write to that file racing against an attach, and hence the
162  * can_attach() result will remain valid until the attach completes.
163  */
164 static int freezer_can_attach(struct cgroup_subsys *ss,
165                               struct cgroup *new_cgroup,
166                               struct task_struct *task, bool threadgroup)
167 {
168         struct freezer *freezer;
169
170         /*
171          * Anything frozen can't move or be moved to/from.
172          *
173          * Since orig_freezer->state == FROZEN means that @task has been
174          * frozen, so it's sufficient to check the latter condition.
175          */
176
177         if (is_task_frozen_enough(task))
178                 return -EBUSY;
179
180         freezer = cgroup_freezer(new_cgroup);
181         if (freezer->state == CGROUP_FROZEN)
182                 return -EBUSY;
183
184         if (threadgroup) {
185                 struct task_struct *c;
186
187                 rcu_read_lock();
188                 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
189                         if (is_task_frozen_enough(c)) {
190                                 rcu_read_unlock();
191                                 return -EBUSY;
192                         }
193                 }
194                 rcu_read_unlock();
195         }
196
197         return 0;
198 }
199
200 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
201 {
202         struct freezer *freezer;
203
204         /*
205          * No lock is needed, since the task isn't on tasklist yet,
206          * so it can't be moved to another cgroup, which means the
207          * freezer won't be removed and will be valid during this
208          * function call.
209          */
210         freezer = task_freezer(task);
211
212         /*
213          * The root cgroup is non-freezable, so we can skip the
214          * following check.
215          */
216         if (!freezer->css.cgroup->parent)
217                 return;
218
219         spin_lock_irq(&freezer->lock);
220         BUG_ON(freezer->state == CGROUP_FROZEN);
221
222         /* Locking avoids race with FREEZING -> THAWED transitions. */
223         if (freezer->state == CGROUP_FREEZING)
224                 freeze_task(task, true);
225         spin_unlock_irq(&freezer->lock);
226 }
227
228 /*
229  * caller must hold freezer->lock
230  */
231 static void update_freezer_state(struct cgroup *cgroup,
232                                  struct freezer *freezer)
233 {
234         struct cgroup_iter it;
235         struct task_struct *task;
236         unsigned int nfrozen = 0, ntotal = 0;
237
238         cgroup_iter_start(cgroup, &it);
239         while ((task = cgroup_iter_next(cgroup, &it))) {
240                 ntotal++;
241                 if (is_task_frozen_enough(task))
242                         nfrozen++;
243         }
244
245         /*
246          * Transition to FROZEN when no new tasks can be added ensures
247          * that we never exist in the FROZEN state while there are unfrozen
248          * tasks.
249          */
250         if (nfrozen == ntotal)
251                 freezer->state = CGROUP_FROZEN;
252         else if (nfrozen > 0)
253                 freezer->state = CGROUP_FREEZING;
254         else
255                 freezer->state = CGROUP_THAWED;
256         cgroup_iter_end(cgroup, &it);
257 }
258
259 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
260                         struct seq_file *m)
261 {
262         struct freezer *freezer;
263         enum freezer_state state;
264
265         if (!cgroup_lock_live_group(cgroup))
266                 return -ENODEV;
267
268         freezer = cgroup_freezer(cgroup);
269         spin_lock_irq(&freezer->lock);
270         state = freezer->state;
271         if (state == CGROUP_FREEZING) {
272                 /* We change from FREEZING to FROZEN lazily if the cgroup was
273                  * only partially frozen when we exitted write. */
274                 update_freezer_state(cgroup, freezer);
275                 state = freezer->state;
276         }
277         spin_unlock_irq(&freezer->lock);
278         cgroup_unlock();
279
280         seq_puts(m, freezer_state_strs[state]);
281         seq_putc(m, '\n');
282         return 0;
283 }
284
285 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
286 {
287         struct cgroup_iter it;
288         struct task_struct *task;
289         unsigned int num_cant_freeze_now = 0;
290
291         freezer->state = CGROUP_FREEZING;
292         cgroup_iter_start(cgroup, &it);
293         while ((task = cgroup_iter_next(cgroup, &it))) {
294                 if (!freeze_task(task, true))
295                         continue;
296                 if (is_task_frozen_enough(task))
297                         continue;
298                 if (!freezing(task) && !freezer_should_skip(task))
299                         num_cant_freeze_now++;
300         }
301         cgroup_iter_end(cgroup, &it);
302
303         return num_cant_freeze_now ? -EBUSY : 0;
304 }
305
306 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
307 {
308         struct cgroup_iter it;
309         struct task_struct *task;
310
311         cgroup_iter_start(cgroup, &it);
312         while ((task = cgroup_iter_next(cgroup, &it))) {
313                 thaw_process(task);
314         }
315         cgroup_iter_end(cgroup, &it);
316
317         freezer->state = CGROUP_THAWED;
318 }
319
320 static int freezer_change_state(struct cgroup *cgroup,
321                                 enum freezer_state goal_state)
322 {
323         struct freezer *freezer;
324         int retval = 0;
325
326         freezer = cgroup_freezer(cgroup);
327
328         spin_lock_irq(&freezer->lock);
329
330         update_freezer_state(cgroup, freezer);
331         if (goal_state == freezer->state)
332                 goto out;
333
334         switch (goal_state) {
335         case CGROUP_THAWED:
336                 unfreeze_cgroup(cgroup, freezer);
337                 break;
338         case CGROUP_FROZEN:
339                 retval = try_to_freeze_cgroup(cgroup, freezer);
340                 break;
341         default:
342                 BUG();
343         }
344 out:
345         spin_unlock_irq(&freezer->lock);
346
347         return retval;
348 }
349
350 static int freezer_write(struct cgroup *cgroup,
351                          struct cftype *cft,
352                          const char *buffer)
353 {
354         int retval;
355         enum freezer_state goal_state;
356
357         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
358                 goal_state = CGROUP_THAWED;
359         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
360                 goal_state = CGROUP_FROZEN;
361         else
362                 return -EINVAL;
363
364         if (!cgroup_lock_live_group(cgroup))
365                 return -ENODEV;
366         retval = freezer_change_state(cgroup, goal_state);
367         cgroup_unlock();
368         return retval;
369 }
370
371 static struct cftype files[] = {
372         {
373                 .name = "state",
374                 .read_seq_string = freezer_read,
375                 .write_string = freezer_write,
376         },
377 };
378
379 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
380 {
381         if (!cgroup->parent)
382                 return 0;
383         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
384 }
385
386 struct cgroup_subsys freezer_subsys = {
387         .name           = "freezer",
388         .create         = freezer_create,
389         .destroy        = freezer_destroy,
390         .populate       = freezer_populate,
391         .subsys_id      = freezer_subsys_id,
392         .can_attach     = freezer_can_attach,
393         .attach         = NULL,
394         .fork           = freezer_fork,
395         .exit           = NULL,
396 };