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  * caller must hold freezer->lock
 217  */
 218 static void update_if_frozen(struct cgroup *cgroup,
 219                                  struct freezer *freezer)
 220 {
 221         struct cgroup_iter it;
 222         struct task_struct *task;
 223         unsigned int nfrozen = 0, ntotal = 0;
 224         enum freezer_state old_state = freezer->state;
 225
 226         cgroup_iter_start(cgroup, &it);
 227         while ((task = cgroup_iter_next(cgroup, &it))) {
 228                 if (freezing(task)) {
 229                         ntotal++;
 230                         /*
 231                          * freezer_should_skip() indicates that the task
 232                          * should be skipped when determining freezing
 233                          * completion.  Consider it frozen in addition to
 234                          * the usual frozen condition.
 235                          */
 236                         if (frozen(task) || task_is_stopped_or_traced(task) ||
 237                             freezer_should_skip(task))
 238                                 nfrozen++;
 239                 }
 240         }
 241
 242         if (old_state == CGROUP_THAWED) {
 243                 BUG_ON(nfrozen > 0);
 244         } else if (old_state == CGROUP_FREEZING) {
 245                 if (nfrozen == ntotal)
 246                         freezer->state = CGROUP_FROZEN;
 247         } else { /* old_state == CGROUP_FROZEN */
 248                 BUG_ON(nfrozen != ntotal);
 249         }
 250
 251         cgroup_iter_end(cgroup, &it);
 252 }
 253
 254 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 255                         struct seq_file *m)
 256 {
 257         struct freezer *freezer;
 258         enum freezer_state state;
 259
 260         if (!cgroup_lock_live_group(cgroup))
 261                 return -ENODEV;
 262
 263         freezer = cgroup_freezer(cgroup);
 264         spin_lock_irq(&freezer->lock);
 265         state = freezer->state;
 266         if (state == CGROUP_FREEZING) {
 267                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 268                  * only partially frozen when we exitted write. */
 269                 update_if_frozen(cgroup, freezer);
 270                 state = freezer->state;
 271         }
 272         spin_unlock_irq(&freezer->lock);
 273         cgroup_unlock();
 274
 275         seq_puts(m, freezer_state_strs[state]);
 276         seq_putc(m, '\n');
 277         return 0;
 278 }
 279
 280 static void freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 281 {
 282         struct cgroup_iter it;
 283         struct task_struct *task;
 284
 285         cgroup_iter_start(cgroup, &it);
 286         while ((task = cgroup_iter_next(cgroup, &it)))
 287                 freeze_task(task);
 288         cgroup_iter_end(cgroup, &it);
 289 }
 290
 291 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 292 {
 293         struct cgroup_iter it;
 294         struct task_struct *task;
 295
 296         cgroup_iter_start(cgroup, &it);
 297         while ((task = cgroup_iter_next(cgroup, &it)))
 298                 __thaw_task(task);
 299         cgroup_iter_end(cgroup, &it);
 300 }
 301
 302 static void freezer_change_state(struct cgroup *cgroup,
 303                                  enum freezer_state goal_state)
 304 {
 305         struct freezer *freezer = cgroup_freezer(cgroup);
 306
 307         spin_lock_irq(&freezer->lock);
 308
 309         update_if_frozen(cgroup, freezer);
 310
 311         switch (goal_state) {
 312         case CGROUP_THAWED:
 313                 if (freezer->state != CGROUP_THAWED)
 314                         atomic_dec(&system_freezing_cnt);
 315                 freezer->state = CGROUP_THAWED;
 316                 unfreeze_cgroup(cgroup, freezer);
 317                 break;
 318         case CGROUP_FROZEN:
 319                 if (freezer->state == CGROUP_THAWED)
 320                         atomic_inc(&system_freezing_cnt);
 321                 freezer->state = CGROUP_FREEZING;
 322                 freeze_cgroup(cgroup, freezer);
 323                 break;
 324         default:
 325                 BUG();
 326         }
 327
 328         spin_unlock_irq(&freezer->lock);
 329 }
 330
 331 static int freezer_write(struct cgroup *cgroup,
 332                          struct cftype *cft,
 333                          const char *buffer)
 334 {
 335         enum freezer_state goal_state;
 336
 337         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 338                 goal_state = CGROUP_THAWED;
 339         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 340                 goal_state = CGROUP_FROZEN;
 341         else
 342                 return -EINVAL;
 343
 344         if (!cgroup_lock_live_group(cgroup))
 345                 return -ENODEV;
 346         freezer_change_state(cgroup, goal_state);
 347         cgroup_unlock();
 348         return 0;
 349 }
 350
 351 static struct cftype files[] = {
 352         {
 353                 .name = "state",
 354                 .flags = CFTYPE_NOT_ON_ROOT,
 355                 .read_seq_string = freezer_read,
 356                 .write_string = freezer_write,
 357         },
 358         { }     /* terminate */
 359 };
 360
 361 struct cgroup_subsys freezer_subsys = {
 362         .name           = "freezer",
 363         .create         = freezer_create,
 364         .destroy        = freezer_destroy,
 365         .subsys_id      = freezer_subsys_id,
 366         .attach         = freezer_attach,
 367         .fork           = freezer_fork,
 368         .base_cftypes   = files,
 369
 370         /*
 371          * freezer subsys doesn't handle hierarchy at all.  Frozen state
 372          * should be inherited through the hierarchy - if a parent is
 373          * frozen, all its children should be frozen.  Fix it and remove
 374          * the following.
 375          */
 376         .broken_hierarchy = true,
 377 };