mm: correctly synchronize rss-counters at exit/exec
[linux-2.6.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         /*
190          * No lock is needed, since the task isn't on tasklist yet,
191          * so it can't be moved to another cgroup, which means the
192          * freezer won't be removed and will be valid during this
193          * function call.  Nevertheless, apply RCU read-side critical
194          * section to suppress RCU lockdep false positives.
195          */
196         rcu_read_lock();
197         freezer = task_freezer(task);
198         rcu_read_unlock();
199
200         /*
201          * The root cgroup is non-freezable, so we can skip the
202          * following check.
203          */
204         if (!freezer->css.cgroup->parent)
205                 return;
206
207         spin_lock_irq(&freezer->lock);
208         BUG_ON(freezer->state == CGROUP_FROZEN);
209
210         /* Locking avoids race with FREEZING -> THAWED transitions. */
211         if (freezer->state == CGROUP_FREEZING)
212                 freeze_task(task);
213         spin_unlock_irq(&freezer->lock);
214 }
215
216 /*
217  * caller must hold freezer->lock
218  */
219 static void update_if_frozen(struct cgroup *cgroup,
220                                  struct freezer *freezer)
221 {
222         struct cgroup_iter it;
223         struct task_struct *task;
224         unsigned int nfrozen = 0, ntotal = 0;
225         enum freezer_state old_state = freezer->state;
226
227         cgroup_iter_start(cgroup, &it);
228         while ((task = cgroup_iter_next(cgroup, &it))) {
229                 ntotal++;
230                 if (freezing(task) && is_task_frozen_enough(task))
231                         nfrozen++;
232         }
233
234         if (old_state == CGROUP_THAWED) {
235                 BUG_ON(nfrozen > 0);
236         } else if (old_state == CGROUP_FREEZING) {
237                 if (nfrozen == ntotal)
238                         freezer->state = CGROUP_FROZEN;
239         } else { /* old_state == CGROUP_FROZEN */
240                 BUG_ON(nfrozen != ntotal);
241         }
242
243         cgroup_iter_end(cgroup, &it);
244 }
245
246 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
247                         struct seq_file *m)
248 {
249         struct freezer *freezer;
250         enum freezer_state state;
251
252         if (!cgroup_lock_live_group(cgroup))
253                 return -ENODEV;
254
255         freezer = cgroup_freezer(cgroup);
256         spin_lock_irq(&freezer->lock);
257         state = freezer->state;
258         if (state == CGROUP_FREEZING) {
259                 /* We change from FREEZING to FROZEN lazily if the cgroup was
260                  * only partially frozen when we exitted write. */
261                 update_if_frozen(cgroup, freezer);
262                 state = freezer->state;
263         }
264         spin_unlock_irq(&freezer->lock);
265         cgroup_unlock();
266
267         seq_puts(m, freezer_state_strs[state]);
268         seq_putc(m, '\n');
269         return 0;
270 }
271
272 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
273 {
274         struct cgroup_iter it;
275         struct task_struct *task;
276         unsigned int num_cant_freeze_now = 0;
277
278         cgroup_iter_start(cgroup, &it);
279         while ((task = cgroup_iter_next(cgroup, &it))) {
280                 if (!freeze_task(task))
281                         continue;
282                 if (is_task_frozen_enough(task))
283                         continue;
284                 if (!freezing(task) && !freezer_should_skip(task))
285                         num_cant_freeze_now++;
286         }
287         cgroup_iter_end(cgroup, &it);
288
289         return num_cant_freeze_now ? -EBUSY : 0;
290 }
291
292 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
293 {
294         struct cgroup_iter it;
295         struct task_struct *task;
296
297         cgroup_iter_start(cgroup, &it);
298         while ((task = cgroup_iter_next(cgroup, &it)))
299                 __thaw_task(task);
300         cgroup_iter_end(cgroup, &it);
301 }
302
303 static int freezer_change_state(struct cgroup *cgroup,
304                                 enum freezer_state goal_state)
305 {
306         struct freezer *freezer;
307         int retval = 0;
308
309         freezer = cgroup_freezer(cgroup);
310
311         spin_lock_irq(&freezer->lock);
312
313         update_if_frozen(cgroup, freezer);
314
315         switch (goal_state) {
316         case CGROUP_THAWED:
317                 if (freezer->state != CGROUP_THAWED)
318                         atomic_dec(&system_freezing_cnt);
319                 freezer->state = CGROUP_THAWED;
320                 unfreeze_cgroup(cgroup, freezer);
321                 break;
322         case CGROUP_FROZEN:
323                 if (freezer->state == CGROUP_THAWED)
324                         atomic_inc(&system_freezing_cnt);
325                 freezer->state = CGROUP_FREEZING;
326                 retval = try_to_freeze_cgroup(cgroup, freezer);
327                 break;
328         default:
329                 BUG();
330         }
331
332         spin_unlock_irq(&freezer->lock);
333
334         return retval;
335 }
336
337 static int freezer_write(struct cgroup *cgroup,
338                          struct cftype *cft,
339                          const char *buffer)
340 {
341         int retval;
342         enum freezer_state goal_state;
343
344         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
345                 goal_state = CGROUP_THAWED;
346         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
347                 goal_state = CGROUP_FROZEN;
348         else
349                 return -EINVAL;
350
351         if (!cgroup_lock_live_group(cgroup))
352                 return -ENODEV;
353         retval = freezer_change_state(cgroup, goal_state);
354         cgroup_unlock();
355         return retval;
356 }
357
358 static struct cftype files[] = {
359         {
360                 .name = "state",
361                 .read_seq_string = freezer_read,
362                 .write_string = freezer_write,
363         },
364 };
365
366 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
367 {
368         if (!cgroup->parent)
369                 return 0;
370         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
371 }
372
373 struct cgroup_subsys freezer_subsys = {
374         .name           = "freezer",
375         .create         = freezer_create,
376         .destroy        = freezer_destroy,
377         .populate       = freezer_populate,
378         .subsys_id      = freezer_subsys_id,
379         .can_attach     = freezer_can_attach,
380         .fork           = freezer_fork,
381 };