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 static inline int __cgroup_freezing_or_frozen(struct task_struct *task)
  52 {
  53         enum freezer_state state = task_freezer(task)->state;
  54         return (state == CGROUP_FREEZING) || (state == CGROUP_FROZEN);
  55 }
  56
  57 int cgroup_freezing_or_frozen(struct task_struct *task)
  58 {
  59         int result;
  60         task_lock(task);
  61         result = __cgroup_freezing_or_frozen(task);
  62         task_unlock(task);
  63         return result;
  64 }
  65
  66 /*
  67  * cgroups_write_string() limits the size of freezer state strings to
  68  * CGROUP_LOCAL_BUFFER_SIZE
  69  */
  70 static const char *freezer_state_strs[] = {
  71         "THAWED",
  72         "FREEZING",
  73         "FROZEN",
  74 };
  75
  76 /*
  77  * State diagram
  78  * Transitions are caused by userspace writes to the freezer.state file.
  79  * The values in parenthesis are state labels. The rest are edge labels.
  80  *
  81  * (THAWED) --FROZEN--> (FREEZING) --FROZEN--> (FROZEN)
  82  *    ^ ^                    |                     |
  83  *    | \_______THAWED_______/                     |
  84  *    \__________________________THAWED____________/
  85  */
  86
  87 struct cgroup_subsys freezer_subsys;
  88
  89 /* Locks taken and their ordering
  90  * ------------------------------
  91  * cgroup_mutex (AKA cgroup_lock)
  92  * freezer->lock
  93  * css_set_lock
  94  * task->alloc_lock (AKA task_lock)
  95  * task->sighand->siglock
  96  *
  97  * cgroup code forces css_set_lock to be taken before task->alloc_lock
  98  *
  99  * freezer_create(), freezer_destroy():
 100  * cgroup_mutex [ by cgroup core ]
 101  *
 102  * freezer_can_attach():
 103  * cgroup_mutex (held by caller of can_attach)
 104  *
 105  * cgroup_freezing_or_frozen():
 106  * task->alloc_lock (to get task's cgroup)
 107  *
 108  * freezer_fork() (preserving fork() performance means can't take cgroup_mutex):
 109  * freezer->lock
 110  *  sighand->siglock (if the cgroup is freezing)
 111  *
 112  * freezer_read():
 113  * cgroup_mutex
 114  *  freezer->lock
 115  *   write_lock css_set_lock (cgroup iterator start)
 116  *    task->alloc_lock
 117  *   read_lock css_set_lock (cgroup iterator start)
 118  *
 119  * freezer_write() (freeze):
 120  * cgroup_mutex
 121  *  freezer->lock
 122  *   write_lock css_set_lock (cgroup iterator start)
 123  *    task->alloc_lock
 124  *   read_lock css_set_lock (cgroup iterator start)
 125  *    sighand->siglock (fake signal delivery inside freeze_task())
 126  *
 127  * freezer_write() (unfreeze):
 128  * cgroup_mutex
 129  *  freezer->lock
 130  *   write_lock css_set_lock (cgroup iterator start)
 131  *    task->alloc_lock
 132  *   read_lock css_set_lock (cgroup iterator start)
 133  *    task->alloc_lock (inside thaw_process(), prevents race with refrigerator())
 134  *     sighand->siglock
 135  */
 136 static struct cgroup_subsys_state *freezer_create(struct cgroup_subsys *ss,
 137                                                   struct cgroup *cgroup)
 138 {
 139         struct freezer *freezer;
 140
 141         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
 142         if (!freezer)
 143                 return ERR_PTR(-ENOMEM);
 144
 145         spin_lock_init(&freezer->lock);
 146         freezer->state = CGROUP_THAWED;
 147         return &freezer->css;
 148 }
 149
 150 static void freezer_destroy(struct cgroup_subsys *ss,
 151                             struct cgroup *cgroup)
 152 {
 153         kfree(cgroup_freezer(cgroup));
 154 }
 155
 156 /* Task is frozen or will freeze immediately when next it gets woken */
 157 static bool is_task_frozen_enough(struct task_struct *task)
 158 {
 159         return frozen(task) ||
 160                 (task_is_stopped_or_traced(task) && freezing(task));
 161 }
 162
 163 /*
 164  * The call to cgroup_lock() in the freezer.state write method prevents
 165  * a write to that file racing against an attach, and hence the
 166  * can_attach() result will remain valid until the attach completes.
 167  */
 168 static int freezer_can_attach(struct cgroup_subsys *ss,
 169                               struct cgroup *new_cgroup,
 170                               struct task_struct *task, bool threadgroup)
 171 {
 172         struct freezer *freezer;
 173
 174         /*
 175          * Anything frozen can't move or be moved to/from.
 176          *
 177          * Since orig_freezer->state == FROZEN means that @task has been
 178          * frozen, so it's sufficient to check the latter condition.
 179          */
 180
 181         if (is_task_frozen_enough(task))
 182                 return -EBUSY;
 183
 184         freezer = cgroup_freezer(new_cgroup);
 185         if (freezer->state == CGROUP_FROZEN)
 186                 return -EBUSY;
 187
 188         if (threadgroup) {
 189                 struct task_struct *c;
 190
 191                 rcu_read_lock();
 192                 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
 193                         if (is_task_frozen_enough(c)) {
 194                                 rcu_read_unlock();
 195                                 return -EBUSY;
 196                         }
 197                 }
 198                 rcu_read_unlock();
 199         }
 200
 201         return 0;
 202 }
 203
 204 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
 205 {
 206         struct freezer *freezer;
 207
 208         /*
 209          * No lock is needed, since the task isn't on tasklist yet,
 210          * so it can't be moved to another cgroup, which means the
 211          * freezer won't be removed and will be valid during this
 212          * function call.  Nevertheless, apply RCU read-side critical
 213          * section to suppress RCU lockdep false positives.
 214          */
 215         rcu_read_lock();
 216         freezer = task_freezer(task);
 217         rcu_read_unlock();
 218
 219         /*
 220          * The root cgroup is non-freezable, so we can skip the
 221          * following check.
 222          */
 223         if (!freezer->css.cgroup->parent)
 224                 return;
 225
 226         spin_lock_irq(&freezer->lock);
 227         BUG_ON(freezer->state == CGROUP_FROZEN);
 228
 229         /* Locking avoids race with FREEZING -> THAWED transitions. */
 230         if (freezer->state == CGROUP_FREEZING)
 231                 freeze_task(task, true);
 232         spin_unlock_irq(&freezer->lock);
 233 }
 234
 235 /*
 236  * caller must hold freezer->lock
 237  */
 238 static void update_freezer_state(struct cgroup *cgroup,
 239                                  struct freezer *freezer)
 240 {
 241         struct cgroup_iter it;
 242         struct task_struct *task;
 243         unsigned int nfrozen = 0, ntotal = 0;
 244
 245         cgroup_iter_start(cgroup, &it);
 246         while ((task = cgroup_iter_next(cgroup, &it))) {
 247                 ntotal++;
 248                 if (is_task_frozen_enough(task))
 249                         nfrozen++;
 250         }
 251
 252         /*
 253          * Transition to FROZEN when no new tasks can be added ensures
 254          * that we never exist in the FROZEN state while there are unfrozen
 255          * tasks.
 256          */
 257         if (nfrozen == ntotal)
 258                 freezer->state = CGROUP_FROZEN;
 259         else if (nfrozen > 0)
 260                 freezer->state = CGROUP_FREEZING;
 261         else
 262                 freezer->state = CGROUP_THAWED;
 263         cgroup_iter_end(cgroup, &it);
 264 }
 265
 266 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 267                         struct seq_file *m)
 268 {
 269         struct freezer *freezer;
 270         enum freezer_state state;
 271
 272         if (!cgroup_lock_live_group(cgroup))
 273                 return -ENODEV;
 274
 275         freezer = cgroup_freezer(cgroup);
 276         spin_lock_irq(&freezer->lock);
 277         state = freezer->state;
 278         if (state == CGROUP_FREEZING) {
 279                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 280                  * only partially frozen when we exitted write. */
 281                 update_freezer_state(cgroup, freezer);
 282                 state = freezer->state;
 283         }
 284         spin_unlock_irq(&freezer->lock);
 285         cgroup_unlock();
 286
 287         seq_puts(m, freezer_state_strs[state]);
 288         seq_putc(m, '\n');
 289         return 0;
 290 }
 291
 292 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 293 {
 294         struct cgroup_iter it;
 295         struct task_struct *task;
 296         unsigned int num_cant_freeze_now = 0;
 297
 298         freezer->state = CGROUP_FREEZING;
 299         cgroup_iter_start(cgroup, &it);
 300         while ((task = cgroup_iter_next(cgroup, &it))) {
 301                 if (!freeze_task(task, true))
 302                         continue;
 303                 if (is_task_frozen_enough(task))
 304                         continue;
 305                 if (!freezing(task) && !freezer_should_skip(task))
 306                         num_cant_freeze_now++;
 307         }
 308         cgroup_iter_end(cgroup, &it);
 309
 310         return num_cant_freeze_now ? -EBUSY : 0;
 311 }
 312
 313 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 314 {
 315         struct cgroup_iter it;
 316         struct task_struct *task;
 317
 318         cgroup_iter_start(cgroup, &it);
 319         while ((task = cgroup_iter_next(cgroup, &it))) {
 320                 thaw_process(task);
 321         }
 322         cgroup_iter_end(cgroup, &it);
 323
 324         freezer->state = CGROUP_THAWED;
 325 }
 326
 327 static int freezer_change_state(struct cgroup *cgroup,
 328                                 enum freezer_state goal_state)
 329 {
 330         struct freezer *freezer;
 331         int retval = 0;
 332
 333         freezer = cgroup_freezer(cgroup);
 334
 335         spin_lock_irq(&freezer->lock);
 336
 337         update_freezer_state(cgroup, freezer);
 338         if (goal_state == freezer->state)
 339                 goto out;
 340
 341         switch (goal_state) {
 342         case CGROUP_THAWED:
 343                 unfreeze_cgroup(cgroup, freezer);
 344                 break;
 345         case CGROUP_FROZEN:
 346                 retval = try_to_freeze_cgroup(cgroup, freezer);
 347                 break;
 348         default:
 349                 BUG();
 350         }
 351 out:
 352         spin_unlock_irq(&freezer->lock);
 353
 354         return retval;
 355 }
 356
 357 static int freezer_write(struct cgroup *cgroup,
 358                          struct cftype *cft,
 359                          const char *buffer)
 360 {
 361         int retval;
 362         enum freezer_state goal_state;
 363
 364         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 365                 goal_state = CGROUP_THAWED;
 366         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 367                 goal_state = CGROUP_FROZEN;
 368         else
 369                 return -EINVAL;
 370
 371         if (!cgroup_lock_live_group(cgroup))
 372                 return -ENODEV;
 373         retval = freezer_change_state(cgroup, goal_state);
 374         cgroup_unlock();
 375         return retval;
 376 }
 377
 378 static struct cftype files[] = {
 379         {
 380                 .name = "state",
 381                 .read_seq_string = freezer_read,
 382                 .write_string = freezer_write,
 383         },
 384 };
 385
 386 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
 387 {
 388         if (!cgroup->parent)
 389                 return 0;
 390         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
 391 }
 392
 393 struct cgroup_subsys freezer_subsys = {
 394         .name           = "freezer",
 395         .create         = freezer_create,
 396         .destroy        = freezer_destroy,
 397         .populate       = freezer_populate,
 398         .subsys_id      = freezer_subsys_id,
 399         .can_attach     = freezer_can_attach,
 400         .attach         = NULL,
 401         .fork           = freezer_fork,
 402         .exit           = NULL,
 403 };