SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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_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         state = freezer->state;
  59         task_unlock(task);
  60
  61         return state == CGROUP_FROZEN;
  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  * css_set_lock
  90  * cgroup_mutex (AKA cgroup_lock)
  91  * task->alloc_lock (AKA task_lock)
  92  * freezer->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  * can_attach():
 101  * cgroup_mutex
 102  *
 103  * cgroup_frozen():
 104  * task->alloc_lock (to get task's cgroup)
 105  *
 106  * freezer_fork() (preserving fork() performance means can't take cgroup_mutex):
 107  * task->alloc_lock (to get task's cgroup)
 108  * freezer->lock
 109  *  sighand->siglock (if the cgroup is freezing)
 110  *
 111  * freezer_read():
 112  * cgroup_mutex
 113  *  freezer->lock
 114  *   read_lock css_set_lock (cgroup iterator start)
 115  *
 116  * freezer_write() (freeze):
 117  * cgroup_mutex
 118  *  freezer->lock
 119  *   read_lock css_set_lock (cgroup iterator start)
 120  *    sighand->siglock
 121  *
 122  * freezer_write() (unfreeze):
 123  * cgroup_mutex
 124  *  freezer->lock
 125  *   read_lock css_set_lock (cgroup iterator start)
 126  *    task->alloc_lock (to prevent races with freeze_task())
 127  *     sighand->siglock
 128  */
 129 static struct cgroup_subsys_state *freezer_create(struct cgroup_subsys *ss,
 130                                                   struct cgroup *cgroup)
 131 {
 132         struct freezer *freezer;
 133
 134         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
 135         if (!freezer)
 136                 return ERR_PTR(-ENOMEM);
 137
 138         spin_lock_init(&freezer->lock);
 139         freezer->state = CGROUP_THAWED;
 140         return &freezer->css;
 141 }
 142
 143 static void freezer_destroy(struct cgroup_subsys *ss,
 144                             struct cgroup *cgroup)
 145 {
 146         kfree(cgroup_freezer(cgroup));
 147 }
 148
 149 /* Task is frozen or will freeze immediately when next it gets woken */
 150 static bool is_task_frozen_enough(struct task_struct *task)
 151 {
 152         return frozen(task) ||
 153                 (task_is_stopped_or_traced(task) && freezing(task));
 154 }
 155
 156 /*
 157  * The call to cgroup_lock() in the freezer.state write method prevents
 158  * a write to that file racing against an attach, and hence the
 159  * can_attach() result will remain valid until the attach completes.
 160  */
 161 static int freezer_can_attach(struct cgroup_subsys *ss,
 162                               struct cgroup *new_cgroup,
 163                               struct task_struct *task, bool threadgroup)
 164 {
 165         struct freezer *freezer;
 166
 167         /*
 168          * Anything frozen can't move or be moved to/from.
 169          *
 170          * Since orig_freezer->state == FROZEN means that @task has been
 171          * frozen, so it's sufficient to check the latter condition.
 172          */
 173
 174         if (is_task_frozen_enough(task))
 175                 return -EBUSY;
 176
 177         freezer = cgroup_freezer(new_cgroup);
 178         if (freezer->state == CGROUP_FROZEN)
 179                 return -EBUSY;
 180
 181         if (threadgroup) {
 182                 struct task_struct *c;
 183
 184                 rcu_read_lock();
 185                 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
 186                         if (is_task_frozen_enough(c)) {
 187                                 rcu_read_unlock();
 188                                 return -EBUSY;
 189                         }
 190                 }
 191                 rcu_read_unlock();
 192         }
 193
 194         return 0;
 195 }
 196
 197 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
 198 {
 199         struct freezer *freezer;
 200
 201         /*
 202          * No lock is needed, since the task isn't on tasklist yet,
 203          * so it can't be moved to another cgroup, which means the
 204          * freezer won't be removed and will be valid during this
 205          * function call.
 206          */
 207         freezer = task_freezer(task);
 208
 209         /*
 210          * The root cgroup is non-freezable, so we can skip the
 211          * following check.
 212          */
 213         if (!freezer->css.cgroup->parent)
 214                 return;
 215
 216         spin_lock_irq(&freezer->lock);
 217         BUG_ON(freezer->state == CGROUP_FROZEN);
 218
 219         /* Locking avoids race with FREEZING -> THAWED transitions. */
 220         if (freezer->state == CGROUP_FREEZING)
 221                 freeze_task(task, true);
 222         spin_unlock_irq(&freezer->lock);
 223 }
 224
 225 /*
 226  * caller must hold freezer->lock
 227  */
 228 static void update_freezer_state(struct cgroup *cgroup,
 229                                  struct freezer *freezer)
 230 {
 231         struct cgroup_iter it;
 232         struct task_struct *task;
 233         unsigned int nfrozen = 0, ntotal = 0;
 234
 235         cgroup_iter_start(cgroup, &it);
 236         while ((task = cgroup_iter_next(cgroup, &it))) {
 237                 ntotal++;
 238                 if (is_task_frozen_enough(task))
 239                         nfrozen++;
 240         }
 241
 242         /*
 243          * Transition to FROZEN when no new tasks can be added ensures
 244          * that we never exist in the FROZEN state while there are unfrozen
 245          * tasks.
 246          */
 247         if (nfrozen == ntotal)
 248                 freezer->state = CGROUP_FROZEN;
 249         else if (nfrozen > 0)
 250                 freezer->state = CGROUP_FREEZING;
 251         else
 252                 freezer->state = CGROUP_THAWED;
 253         cgroup_iter_end(cgroup, &it);
 254 }
 255
 256 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 257                         struct seq_file *m)
 258 {
 259         struct freezer *freezer;
 260         enum freezer_state state;
 261
 262         if (!cgroup_lock_live_group(cgroup))
 263                 return -ENODEV;
 264
 265         freezer = cgroup_freezer(cgroup);
 266         spin_lock_irq(&freezer->lock);
 267         state = freezer->state;
 268         if (state == CGROUP_FREEZING) {
 269                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 270                  * only partially frozen when we exitted write. */
 271                 update_freezer_state(cgroup, freezer);
 272                 state = freezer->state;
 273         }
 274         spin_unlock_irq(&freezer->lock);
 275         cgroup_unlock();
 276
 277         seq_puts(m, freezer_state_strs[state]);
 278         seq_putc(m, '\n');
 279         return 0;
 280 }
 281
 282 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 283 {
 284         struct cgroup_iter it;
 285         struct task_struct *task;
 286         unsigned int num_cant_freeze_now = 0;
 287
 288         freezer->state = CGROUP_FREEZING;
 289         cgroup_iter_start(cgroup, &it);
 290         while ((task = cgroup_iter_next(cgroup, &it))) {
 291                 if (!freeze_task(task, true))
 292                         continue;
 293                 if (is_task_frozen_enough(task))
 294                         continue;
 295                 if (!freezing(task) && !freezer_should_skip(task))
 296                         num_cant_freeze_now++;
 297         }
 298         cgroup_iter_end(cgroup, &it);
 299
 300         return num_cant_freeze_now ? -EBUSY : 0;
 301 }
 302
 303 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 304 {
 305         struct cgroup_iter it;
 306         struct task_struct *task;
 307
 308         cgroup_iter_start(cgroup, &it);
 309         while ((task = cgroup_iter_next(cgroup, &it))) {
 310                 thaw_process(task);
 311         }
 312         cgroup_iter_end(cgroup, &it);
 313
 314         freezer->state = CGROUP_THAWED;
 315 }
 316
 317 static int freezer_change_state(struct cgroup *cgroup,
 318                                 enum freezer_state goal_state)
 319 {
 320         struct freezer *freezer;
 321         int retval = 0;
 322
 323         freezer = cgroup_freezer(cgroup);
 324
 325         spin_lock_irq(&freezer->lock);
 326
 327         update_freezer_state(cgroup, freezer);
 328         if (goal_state == freezer->state)
 329                 goto out;
 330
 331         switch (goal_state) {
 332         case CGROUP_THAWED:
 333                 unfreeze_cgroup(cgroup, freezer);
 334                 break;
 335         case CGROUP_FROZEN:
 336                 retval = try_to_freeze_cgroup(cgroup, freezer);
 337                 break;
 338         default:
 339                 BUG();
 340         }
 341 out:
 342         spin_unlock_irq(&freezer->lock);
 343
 344         return retval;
 345 }
 346
 347 static int freezer_write(struct cgroup *cgroup,
 348                          struct cftype *cft,
 349                          const char *buffer)
 350 {
 351         int retval;
 352         enum freezer_state goal_state;
 353
 354         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 355                 goal_state = CGROUP_THAWED;
 356         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 357                 goal_state = CGROUP_FROZEN;
 358         else
 359                 return -EINVAL;
 360
 361         if (!cgroup_lock_live_group(cgroup))
 362                 return -ENODEV;
 363         retval = freezer_change_state(cgroup, goal_state);
 364         cgroup_unlock();
 365         return retval;
 366 }
 367
 368 static struct cftype files[] = {
 369         {
 370                 .name = "state",
 371                 .read_seq_string = freezer_read,
 372                 .write_string = freezer_write,
 373         },
 374 };
 375
 376 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
 377 {
 378         if (!cgroup->parent)
 379                 return 0;
 380         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
 381 }
 382
 383 struct cgroup_subsys freezer_subsys = {
 384         .name           = "freezer",
 385         .create         = freezer_create,
 386         .destroy        = freezer_destroy,
 387         .populate       = freezer_populate,
 388         .subsys_id      = freezer_subsys_id,
 389         .can_attach     = freezer_can_attach,
 390         .attach         = NULL,
 391         .fork           = freezer_fork,
 392         .exit           = NULL,
 393 };