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  * cgroup_mutex (AKA cgroup_lock)
  93  * freezer->lock
  94  * css_set_lock
  95  * task->alloc_lock (AKA task_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  * freezer_can_attach():
 104  * cgroup_mutex (held by caller of can_attach)
 105  *
 106  * cgroup_freezing_or_frozen():
 107  * task->alloc_lock (to get task's cgroup)
 108  *
 109  * freezer_fork() (preserving fork() performance means can't take cgroup_mutex):
 110  * freezer->lock
 111  *  sighand->siglock (if the cgroup is freezing)
 112  *
 113  * freezer_read():
 114  * cgroup_mutex
 115  *  freezer->lock
 116  *   write_lock css_set_lock (cgroup iterator start)
 117  *    task->alloc_lock
 118  *   read_lock css_set_lock (cgroup iterator start)
 119  *
 120  * freezer_write() (freeze):
 121  * cgroup_mutex
 122  *  freezer->lock
 123  *   write_lock css_set_lock (cgroup iterator start)
 124  *    task->alloc_lock
 125  *   read_lock css_set_lock (cgroup iterator start)
 126  *    sighand->siglock (fake signal delivery inside freeze_task())
 127  *
 128  * freezer_write() (unfreeze):
 129  * cgroup_mutex
 130  *  freezer->lock
 131  *   write_lock css_set_lock (cgroup iterator start)
 132  *    task->alloc_lock
 133  *   read_lock css_set_lock (cgroup iterator start)
 134  *    task->alloc_lock (inside thaw_process(), prevents race with refrigerator())
 135  *     sighand->siglock
 136  */
 137 static struct cgroup_subsys_state *freezer_create(struct cgroup_subsys *ss,
 138                                                   struct cgroup *cgroup)
 139 {
 140         struct freezer *freezer;
 141
 142         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
 143         if (!freezer)
 144                 return ERR_PTR(-ENOMEM);
 145
 146         spin_lock_init(&freezer->lock);
 147         freezer->state = CGROUP_THAWED;
 148         return &freezer->css;
 149 }
 150
 151 static void freezer_destroy(struct cgroup_subsys *ss,
 152                             struct cgroup *cgroup)
 153 {
 154         kfree(cgroup_freezer(cgroup));
 155 }
 156
 157 /* Task is frozen or will freeze immediately when next it gets woken */
 158 static bool is_task_frozen_enough(struct task_struct *task)
 159 {
 160         return frozen(task) ||
 161                 (task_is_stopped_or_traced(task) && freezing(task));
 162 }
 163
 164 /*
 165  * The call to cgroup_lock() in the freezer.state write method prevents
 166  * a write to that file racing against an attach, and hence the
 167  * can_attach() result will remain valid until the attach completes.
 168  */
 169 static int freezer_can_attach(struct cgroup_subsys *ss,
 170                               struct cgroup *new_cgroup,
 171                               struct task_struct *task, bool threadgroup)
 172 {
 173         struct freezer *freezer;
 174
 175         if ((current != task) && (!capable(CAP_SYS_ADMIN))) {
 176                 const struct cred *cred = current_cred(), *tcred;
 177
 178                 tcred = __task_cred(task);
 179                 if (cred->euid != tcred->uid && cred->euid != tcred->suid)
 180                         return -EPERM;
 181         }
 182
 183         /*
 184          * Anything frozen can't move or be moved to/from.
 185          *
 186          * Since orig_freezer->state == FROZEN means that @task has been
 187          * frozen, so it's sufficient to check the latter condition.
 188          */
 189
 190         if (is_task_frozen_enough(task))
 191                 return -EBUSY;
 192
 193         freezer = cgroup_freezer(new_cgroup);
 194         if (freezer->state == CGROUP_FROZEN)
 195                 return -EBUSY;
 196
 197         if (threadgroup) {
 198                 struct task_struct *c;
 199
 200                 rcu_read_lock();
 201                 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
 202                         if (is_task_frozen_enough(c)) {
 203                                 rcu_read_unlock();
 204                                 return -EBUSY;
 205                         }
 206                 }
 207                 rcu_read_unlock();
 208         }
 209
 210         return 0;
 211 }
 212
 213 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
 214 {
 215         struct freezer *freezer;
 216
 217         /*
 218          * No lock is needed, since the task isn't on tasklist yet,
 219          * so it can't be moved to another cgroup, which means the
 220          * freezer won't be removed and will be valid during this
 221          * function call.  Nevertheless, apply RCU read-side critical
 222          * section to suppress RCU lockdep false positives.
 223          */
 224         rcu_read_lock();
 225         freezer = task_freezer(task);
 226         rcu_read_unlock();
 227
 228         /*
 229          * The root cgroup is non-freezable, so we can skip the
 230          * following check.
 231          */
 232         if (!freezer->css.cgroup->parent)
 233                 return;
 234
 235         spin_lock_irq(&freezer->lock);
 236         BUG_ON(freezer->state == CGROUP_FROZEN);
 237
 238         /* Locking avoids race with FREEZING -> THAWED transitions. */
 239         if (freezer->state == CGROUP_FREEZING)
 240                 freeze_task(task, true);
 241         spin_unlock_irq(&freezer->lock);
 242 }
 243
 244 /*
 245  * caller must hold freezer->lock
 246  */
 247 static void update_freezer_state(struct cgroup *cgroup,
 248                                  struct freezer *freezer)
 249 {
 250         struct cgroup_iter it;
 251         struct task_struct *task;
 252         unsigned int nfrozen = 0, ntotal = 0;
 253
 254         cgroup_iter_start(cgroup, &it);
 255         while ((task = cgroup_iter_next(cgroup, &it))) {
 256                 ntotal++;
 257                 if (is_task_frozen_enough(task))
 258                         nfrozen++;
 259         }
 260
 261         /*
 262          * Transition to FROZEN when no new tasks can be added ensures
 263          * that we never exist in the FROZEN state while there are unfrozen
 264          * tasks.
 265          */
 266         if (nfrozen == ntotal)
 267                 freezer->state = CGROUP_FROZEN;
 268         else if (nfrozen > 0)
 269                 freezer->state = CGROUP_FREEZING;
 270         else
 271                 freezer->state = CGROUP_THAWED;
 272         cgroup_iter_end(cgroup, &it);
 273 }
 274
 275 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 276                         struct seq_file *m)
 277 {
 278         struct freezer *freezer;
 279         enum freezer_state state;
 280
 281         if (!cgroup_lock_live_group(cgroup))
 282                 return -ENODEV;
 283
 284         freezer = cgroup_freezer(cgroup);
 285         spin_lock_irq(&freezer->lock);
 286         state = freezer->state;
 287         if (state == CGROUP_FREEZING) {
 288                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 289                  * only partially frozen when we exitted write. */
 290                 update_freezer_state(cgroup, freezer);
 291                 state = freezer->state;
 292         }
 293         spin_unlock_irq(&freezer->lock);
 294         cgroup_unlock();
 295
 296         seq_puts(m, freezer_state_strs[state]);
 297         seq_putc(m, '\n');
 298         return 0;
 299 }
 300
 301 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 302 {
 303         struct cgroup_iter it;
 304         struct task_struct *task;
 305         unsigned int num_cant_freeze_now = 0;
 306
 307         freezer->state = CGROUP_FREEZING;
 308         cgroup_iter_start(cgroup, &it);
 309         while ((task = cgroup_iter_next(cgroup, &it))) {
 310                 if (!freeze_task(task, true))
 311                         continue;
 312                 if (is_task_frozen_enough(task))
 313                         continue;
 314                 if (!freezing(task) && !freezer_should_skip(task))
 315                         num_cant_freeze_now++;
 316         }
 317         cgroup_iter_end(cgroup, &it);
 318
 319         return num_cant_freeze_now ? -EBUSY : 0;
 320 }
 321
 322 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 323 {
 324         struct cgroup_iter it;
 325         struct task_struct *task;
 326
 327         cgroup_iter_start(cgroup, &it);
 328         while ((task = cgroup_iter_next(cgroup, &it))) {
 329                 thaw_process(task);
 330         }
 331         cgroup_iter_end(cgroup, &it);
 332
 333         freezer->state = CGROUP_THAWED;
 334 }
 335
 336 static int freezer_change_state(struct cgroup *cgroup,
 337                                 enum freezer_state goal_state)
 338 {
 339         struct freezer *freezer;
 340         int retval = 0;
 341
 342         freezer = cgroup_freezer(cgroup);
 343
 344         spin_lock_irq(&freezer->lock);
 345
 346         update_freezer_state(cgroup, freezer);
 347         if (goal_state == freezer->state)
 348                 goto out;
 349
 350         switch (goal_state) {
 351         case CGROUP_THAWED:
 352                 unfreeze_cgroup(cgroup, freezer);
 353                 break;
 354         case CGROUP_FROZEN:
 355                 retval = try_to_freeze_cgroup(cgroup, freezer);
 356                 break;
 357         default:
 358                 BUG();
 359         }
 360 out:
 361         spin_unlock_irq(&freezer->lock);
 362
 363         return retval;
 364 }
 365
 366 static int freezer_write(struct cgroup *cgroup,
 367                          struct cftype *cft,
 368                          const char *buffer)
 369 {
 370         int retval;
 371         enum freezer_state goal_state;
 372
 373         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 374                 goal_state = CGROUP_THAWED;
 375         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 376                 goal_state = CGROUP_FROZEN;
 377         else
 378                 return -EINVAL;
 379
 380         if (!cgroup_lock_live_group(cgroup))
 381                 return -ENODEV;
 382         retval = freezer_change_state(cgroup, goal_state);
 383         cgroup_unlock();
 384         return retval;
 385 }
 386
 387 static struct cftype files[] = {
 388         {
 389                 .name = "state",
 390                 .read_seq_string = freezer_read,
 391                 .write_string = freezer_write,
 392         },
 393 };
 394
 395 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
 396 {
 397         if (!cgroup->parent)
 398                 return 0;
 399         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
 400 }
 401
 402 struct cgroup_subsys freezer_subsys = {
 403         .name           = "freezer",
 404         .create         = freezer_create,
 405         .destroy        = freezer_destroy,
 406         .populate       = freezer_populate,
 407         .subsys_id      = freezer_subsys_id,
 408         .can_attach     = freezer_can_attach,
 409         .attach         = NULL,
 410         .fork           = freezer_fork,
 411         .exit           = NULL,
 412 };