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 /*
 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
 171         freezer = cgroup_freezer(new_cgroup);
 172         if (freezer->state != CGROUP_THAWED)
 173                 return -EBUSY;
 174
 175         rcu_read_lock();
 176         if (__cgroup_freezing_or_frozen(task)) {
 177                 rcu_read_unlock();
 178                 return -EBUSY;
 179         }
 180         rcu_read_unlock();
 181
 182         if (threadgroup) {
 183                 struct task_struct *c;
 184
 185                 rcu_read_lock();
 186                 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
 187                         if (__cgroup_freezing_or_frozen(c)) {
 188                                 rcu_read_unlock();
 189                                 return -EBUSY;
 190                         }
 191                 }
 192                 rcu_read_unlock();
 193         }
 194
 195         return 0;
 196 }
 197
 198 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
 199 {
 200         struct freezer *freezer;
 201
 202         /*
 203          * No lock is needed, since the task isn't on tasklist yet,
 204          * so it can't be moved to another cgroup, which means the
 205          * freezer won't be removed and will be valid during this
 206          * function call.  Nevertheless, apply RCU read-side critical
 207          * section to suppress RCU lockdep false positives.
 208          */
 209         rcu_read_lock();
 210         freezer = task_freezer(task);
 211         rcu_read_unlock();
 212
 213         /*
 214          * The root cgroup is non-freezable, so we can skip the
 215          * following check.
 216          */
 217         if (!freezer->css.cgroup->parent)
 218                 return;
 219
 220         spin_lock_irq(&freezer->lock);
 221         BUG_ON(freezer->state == CGROUP_FROZEN);
 222
 223         /* Locking avoids race with FREEZING -> THAWED transitions. */
 224         if (freezer->state == CGROUP_FREEZING)
 225                 freeze_task(task, true);
 226         spin_unlock_irq(&freezer->lock);
 227 }
 228
 229 /*
 230  * caller must hold freezer->lock
 231  */
 232 static void update_if_frozen(struct cgroup *cgroup,
 233                                  struct freezer *freezer)
 234 {
 235         struct cgroup_iter it;
 236         struct task_struct *task;
 237         unsigned int nfrozen = 0, ntotal = 0;
 238         enum freezer_state old_state = freezer->state;
 239
 240         cgroup_iter_start(cgroup, &it);
 241         while ((task = cgroup_iter_next(cgroup, &it))) {
 242                 ntotal++;
 243                 if (frozen(task))
 244                         nfrozen++;
 245         }
 246
 247         if (old_state == CGROUP_THAWED) {
 248                 BUG_ON(nfrozen > 0);
 249         } else if (old_state == CGROUP_FREEZING) {
 250                 if (nfrozen == ntotal)
 251                         freezer->state = CGROUP_FROZEN;
 252         } else { /* old_state == CGROUP_FROZEN */
 253                 BUG_ON(nfrozen != ntotal);
 254         }
 255
 256         cgroup_iter_end(cgroup, &it);
 257 }
 258
 259 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 260                         struct seq_file *m)
 261 {
 262         struct freezer *freezer;
 263         enum freezer_state state;
 264
 265         if (!cgroup_lock_live_group(cgroup))
 266                 return -ENODEV;
 267
 268         freezer = cgroup_freezer(cgroup);
 269         spin_lock_irq(&freezer->lock);
 270         state = freezer->state;
 271         if (state == CGROUP_FREEZING) {
 272                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 273                  * only partially frozen when we exitted write. */
 274                 update_if_frozen(cgroup, freezer);
 275                 state = freezer->state;
 276         }
 277         spin_unlock_irq(&freezer->lock);
 278         cgroup_unlock();
 279
 280         seq_puts(m, freezer_state_strs[state]);
 281         seq_putc(m, '\n');
 282         return 0;
 283 }
 284
 285 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 286 {
 287         struct cgroup_iter it;
 288         struct task_struct *task;
 289         unsigned int num_cant_freeze_now = 0;
 290
 291         freezer->state = CGROUP_FREEZING;
 292         cgroup_iter_start(cgroup, &it);
 293         while ((task = cgroup_iter_next(cgroup, &it))) {
 294                 if (!freeze_task(task, true))
 295                         continue;
 296                 if (frozen(task))
 297                         continue;
 298                 if (!freezing(task) && !freezer_should_skip(task))
 299                         num_cant_freeze_now++;
 300         }
 301         cgroup_iter_end(cgroup, &it);
 302
 303         return num_cant_freeze_now ? -EBUSY : 0;
 304 }
 305
 306 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 307 {
 308         struct cgroup_iter it;
 309         struct task_struct *task;
 310
 311         cgroup_iter_start(cgroup, &it);
 312         while ((task = cgroup_iter_next(cgroup, &it))) {
 313                 thaw_process(task);
 314         }
 315         cgroup_iter_end(cgroup, &it);
 316
 317         freezer->state = CGROUP_THAWED;
 318 }
 319
 320 static int freezer_change_state(struct cgroup *cgroup,
 321                                 enum freezer_state goal_state)
 322 {
 323         struct freezer *freezer;
 324         int retval = 0;
 325
 326         freezer = cgroup_freezer(cgroup);
 327
 328         spin_lock_irq(&freezer->lock);
 329
 330         update_if_frozen(cgroup, freezer);
 331         if (goal_state == freezer->state)
 332                 goto out;
 333
 334         switch (goal_state) {
 335         case CGROUP_THAWED:
 336                 unfreeze_cgroup(cgroup, freezer);
 337                 break;
 338         case CGROUP_FROZEN:
 339                 retval = try_to_freeze_cgroup(cgroup, freezer);
 340                 break;
 341         default:
 342                 BUG();
 343         }
 344 out:
 345         spin_unlock_irq(&freezer->lock);
 346
 347         return retval;
 348 }
 349
 350 static int freezer_write(struct cgroup *cgroup,
 351                          struct cftype *cft,
 352                          const char *buffer)
 353 {
 354         int retval;
 355         enum freezer_state goal_state;
 356
 357         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 358                 goal_state = CGROUP_THAWED;
 359         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 360                 goal_state = CGROUP_FROZEN;
 361         else
 362                 return -EINVAL;
 363
 364         if (!cgroup_lock_live_group(cgroup))
 365                 return -ENODEV;
 366         retval = freezer_change_state(cgroup, goal_state);
 367         cgroup_unlock();
 368         return retval;
 369 }
 370
 371 static struct cftype files[] = {
 372         {
 373                 .name = "state",
 374                 .read_seq_string = freezer_read,
 375                 .write_string = freezer_write,
 376         },
 377 };
 378
 379 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
 380 {
 381         if (!cgroup->parent)
 382                 return 0;
 383         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
 384 }
 385
 386 struct cgroup_subsys freezer_subsys = {
 387         .name           = "freezer",
 388         .create         = freezer_create,
 389         .destroy        = freezer_destroy,
 390         .populate       = freezer_populate,
 391         .subsys_id      = freezer_subsys_id,
 392         .can_attach     = freezer_can_attach,
 393         .attach         = NULL,
 394         .fork           = freezer_fork,
 395         .exit           = NULL,
 396 };