]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - kernel/cgroup_freezer.c
3d45503a21a2696aa9bba958cd1b5171aea13f4d
[~shefty/rdma-dev.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 /*
154  * The call to cgroup_lock() in the freezer.state write method prevents
155  * a write to that file racing against an attach, and hence we don't need
156  * to worry about racing against migration.
157  */
158 static void freezer_attach(struct cgroup *new_cgrp, struct cgroup_taskset *tset)
159 {
160         struct freezer *freezer = cgroup_freezer(new_cgrp);
161         struct task_struct *task;
162
163         spin_lock_irq(&freezer->lock);
164
165         /*
166          * Make the new tasks conform to the current state of @new_cgrp.
167          * For simplicity, when migrating any task to a FROZEN cgroup, we
168          * revert it to FREEZING and let update_if_frozen() determine the
169          * correct state later.
170          *
171          * Tasks in @tset are on @new_cgrp but may not conform to its
172          * current state before executing the following - !frozen tasks may
173          * be visible in a FROZEN cgroup and frozen tasks in a THAWED one.
174          * This means that, to determine whether to freeze, one should test
175          * whether the state equals THAWED.
176          */
177         cgroup_taskset_for_each(task, new_cgrp, tset) {
178                 if (freezer->state == CGROUP_THAWED) {
179                         __thaw_task(task);
180                 } else {
181                         freeze_task(task);
182                         freezer->state = CGROUP_FREEZING;
183                 }
184         }
185
186         spin_unlock_irq(&freezer->lock);
187 }
188
189 static void freezer_fork(struct task_struct *task)
190 {
191         struct freezer *freezer;
192
193         rcu_read_lock();
194         freezer = task_freezer(task);
195
196         /*
197          * The root cgroup is non-freezable, so we can skip the
198          * following check.
199          */
200         if (!freezer->css.cgroup->parent)
201                 goto out;
202
203         spin_lock_irq(&freezer->lock);
204         /*
205          * @task might have been just migrated into a FROZEN cgroup.  Test
206          * equality with THAWED.  Read the comment in freezer_attach().
207          */
208         if (freezer->state != CGROUP_THAWED)
209                 freeze_task(task);
210         spin_unlock_irq(&freezer->lock);
211 out:
212         rcu_read_unlock();
213 }
214
215 /*
216  * We change from FREEZING to FROZEN lazily if the cgroup was only
217  * partially frozen when we exitted write.  Caller must hold freezer->lock.
218  *
219  * Task states and freezer state might disagree while tasks are being
220  * migrated into @cgroup, so we can't verify task states against @freezer
221  * state here.  See freezer_attach() for details.
222  */
223 static void update_if_frozen(struct cgroup *cgroup, struct freezer *freezer)
224 {
225         struct cgroup_iter it;
226         struct task_struct *task;
227
228         if (freezer->state != CGROUP_FREEZING)
229                 return;
230
231         cgroup_iter_start(cgroup, &it);
232
233         while ((task = cgroup_iter_next(cgroup, &it))) {
234                 if (freezing(task)) {
235                         /*
236                          * freezer_should_skip() indicates that the task
237                          * should be skipped when determining freezing
238                          * completion.  Consider it frozen in addition to
239                          * the usual frozen condition.
240                          */
241                         if (!frozen(task) && !task_is_stopped_or_traced(task) &&
242                             !freezer_should_skip(task))
243                                 goto notyet;
244                 }
245         }
246
247         freezer->state = CGROUP_FROZEN;
248 notyet:
249         cgroup_iter_end(cgroup, &it);
250 }
251
252 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
253                         struct seq_file *m)
254 {
255         struct freezer *freezer;
256         enum freezer_state state;
257
258         if (!cgroup_lock_live_group(cgroup))
259                 return -ENODEV;
260
261         freezer = cgroup_freezer(cgroup);
262         spin_lock_irq(&freezer->lock);
263         update_if_frozen(cgroup, freezer);
264         state = freezer->state;
265         spin_unlock_irq(&freezer->lock);
266         cgroup_unlock();
267
268         seq_puts(m, freezer_state_strs[state]);
269         seq_putc(m, '\n');
270         return 0;
271 }
272
273 static void freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
274 {
275         struct cgroup_iter it;
276         struct task_struct *task;
277
278         cgroup_iter_start(cgroup, &it);
279         while ((task = cgroup_iter_next(cgroup, &it)))
280                 freeze_task(task);
281         cgroup_iter_end(cgroup, &it);
282 }
283
284 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
285 {
286         struct cgroup_iter it;
287         struct task_struct *task;
288
289         cgroup_iter_start(cgroup, &it);
290         while ((task = cgroup_iter_next(cgroup, &it)))
291                 __thaw_task(task);
292         cgroup_iter_end(cgroup, &it);
293 }
294
295 static void freezer_change_state(struct cgroup *cgroup,
296                                  enum freezer_state goal_state)
297 {
298         struct freezer *freezer = cgroup_freezer(cgroup);
299
300         spin_lock_irq(&freezer->lock);
301
302         switch (goal_state) {
303         case CGROUP_THAWED:
304                 if (freezer->state != CGROUP_THAWED)
305                         atomic_dec(&system_freezing_cnt);
306                 freezer->state = CGROUP_THAWED;
307                 unfreeze_cgroup(cgroup, freezer);
308                 break;
309         case CGROUP_FROZEN:
310                 if (freezer->state == CGROUP_THAWED)
311                         atomic_inc(&system_freezing_cnt);
312                 freezer->state = CGROUP_FREEZING;
313                 freeze_cgroup(cgroup, freezer);
314                 break;
315         default:
316                 BUG();
317         }
318
319         spin_unlock_irq(&freezer->lock);
320 }
321
322 static int freezer_write(struct cgroup *cgroup,
323                          struct cftype *cft,
324                          const char *buffer)
325 {
326         enum freezer_state goal_state;
327
328         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
329                 goal_state = CGROUP_THAWED;
330         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
331                 goal_state = CGROUP_FROZEN;
332         else
333                 return -EINVAL;
334
335         if (!cgroup_lock_live_group(cgroup))
336                 return -ENODEV;
337         freezer_change_state(cgroup, goal_state);
338         cgroup_unlock();
339         return 0;
340 }
341
342 static struct cftype files[] = {
343         {
344                 .name = "state",
345                 .flags = CFTYPE_NOT_ON_ROOT,
346                 .read_seq_string = freezer_read,
347                 .write_string = freezer_write,
348         },
349         { }     /* terminate */
350 };
351
352 struct cgroup_subsys freezer_subsys = {
353         .name           = "freezer",
354         .create         = freezer_create,
355         .destroy        = freezer_destroy,
356         .subsys_id      = freezer_subsys_id,
357         .attach         = freezer_attach,
358         .fork           = freezer_fork,
359         .base_cftypes   = files,
360
361         /*
362          * freezer subsys doesn't handle hierarchy at all.  Frozen state
363          * should be inherited through the hierarchy - if a parent is
364          * frozen, all its children should be frozen.  Fix it and remove
365          * the following.
366          */
367         .broken_hierarchy = true,
368 };