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         if ((current != task) && (!capable(CAP_SYS_ADMIN))) {
 168                 const struct cred *cred = current_cred(), *tcred;
 169
 170                 tcred = __task_cred(task);
 171                 if (cred->euid != tcred->uid && cred->euid != tcred->suid)
 172                         return -EPERM;
 173         }
 174
 175         /*
 176          * Anything frozen can't move or be moved to/from.
 177          */
 178
 179         freezer = cgroup_freezer(new_cgroup);
 180         if (freezer->state != CGROUP_THAWED)
 181                 return -EBUSY;
 182
 183         rcu_read_lock();
 184         if (__cgroup_freezing_or_frozen(task)) {
 185                 rcu_read_unlock();
 186                 return -EBUSY;
 187         }
 188         rcu_read_unlock();
 189
 190         if (threadgroup) {
 191                 struct task_struct *c;
 192
 193                 rcu_read_lock();
 194                 list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
 195                         if (__cgroup_freezing_or_frozen(c)) {
 196                                 rcu_read_unlock();
 197                                 return -EBUSY;
 198                         }
 199                 }
 200                 rcu_read_unlock();
 201         }
 202
 203         return 0;
 204 }
 205
 206 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
 207 {
 208         struct freezer *freezer;
 209
 210         /*
 211          * No lock is needed, since the task isn't on tasklist yet,
 212          * so it can't be moved to another cgroup, which means the
 213          * freezer won't be removed and will be valid during this
 214          * function call.  Nevertheless, apply RCU read-side critical
 215          * section to suppress RCU lockdep false positives.
 216          */
 217         rcu_read_lock();
 218         freezer = task_freezer(task);
 219         rcu_read_unlock();
 220
 221         /*
 222          * The root cgroup is non-freezable, so we can skip the
 223          * following check.
 224          */
 225         if (!freezer->css.cgroup->parent)
 226                 return;
 227
 228         spin_lock_irq(&freezer->lock);
 229         BUG_ON(freezer->state == CGROUP_FROZEN);
 230
 231         /* Locking avoids race with FREEZING -> THAWED transitions. */
 232         if (freezer->state == CGROUP_FREEZING)
 233                 freeze_task(task, true);
 234         spin_unlock_irq(&freezer->lock);
 235 }
 236
 237 /*
 238  * caller must hold freezer->lock
 239  */
 240 static void update_if_frozen(struct cgroup *cgroup,
 241                                  struct freezer *freezer)
 242 {
 243         struct cgroup_iter it;
 244         struct task_struct *task;
 245         unsigned int nfrozen = 0, ntotal = 0;
 246         enum freezer_state old_state = freezer->state;
 247
 248         cgroup_iter_start(cgroup, &it);
 249         while ((task = cgroup_iter_next(cgroup, &it))) {
 250                 ntotal++;
 251                 if (frozen(task))
 252                         nfrozen++;
 253         }
 254
 255         if (old_state == CGROUP_THAWED) {
 256                 BUG_ON(nfrozen > 0);
 257         } else if (old_state == CGROUP_FREEZING) {
 258                 if (nfrozen == ntotal)
 259                         freezer->state = CGROUP_FROZEN;
 260         } else { /* old_state == CGROUP_FROZEN */
 261                 BUG_ON(nfrozen != ntotal);
 262         }
 263
 264         cgroup_iter_end(cgroup, &it);
 265 }
 266
 267 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 268                         struct seq_file *m)
 269 {
 270         struct freezer *freezer;
 271         enum freezer_state state;
 272
 273         if (!cgroup_lock_live_group(cgroup))
 274                 return -ENODEV;
 275
 276         freezer = cgroup_freezer(cgroup);
 277         spin_lock_irq(&freezer->lock);
 278         state = freezer->state;
 279         if (state == CGROUP_FREEZING) {
 280                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 281                  * only partially frozen when we exitted write. */
 282                 update_if_frozen(cgroup, freezer);
 283                 state = freezer->state;
 284         }
 285         spin_unlock_irq(&freezer->lock);
 286         cgroup_unlock();
 287
 288         seq_puts(m, freezer_state_strs[state]);
 289         seq_putc(m, '\n');
 290         return 0;
 291 }
 292
 293 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 294 {
 295         struct cgroup_iter it;
 296         struct task_struct *task;
 297         unsigned int num_cant_freeze_now = 0;
 298
 299         freezer->state = CGROUP_FREEZING;
 300         cgroup_iter_start(cgroup, &it);
 301         while ((task = cgroup_iter_next(cgroup, &it))) {
 302                 if (!freeze_task(task, true))
 303                         continue;
 304                 if (frozen(task))
 305                         continue;
 306                 if (!freezing(task) && !freezer_should_skip(task))
 307                         num_cant_freeze_now++;
 308         }
 309         cgroup_iter_end(cgroup, &it);
 310
 311         return num_cant_freeze_now ? -EBUSY : 0;
 312 }
 313
 314 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 315 {
 316         struct cgroup_iter it;
 317         struct task_struct *task;
 318
 319         cgroup_iter_start(cgroup, &it);
 320         while ((task = cgroup_iter_next(cgroup, &it))) {
 321                 thaw_process(task);
 322         }
 323         cgroup_iter_end(cgroup, &it);
 324
 325         freezer->state = CGROUP_THAWED;
 326 }
 327
 328 static int freezer_change_state(struct cgroup *cgroup,
 329                                 enum freezer_state goal_state)
 330 {
 331         struct freezer *freezer;
 332         int retval = 0;
 333
 334         freezer = cgroup_freezer(cgroup);
 335
 336         spin_lock_irq(&freezer->lock);
 337
 338         update_if_frozen(cgroup, freezer);
 339         if (goal_state == freezer->state)
 340                 goto out;
 341
 342         switch (goal_state) {
 343         case CGROUP_THAWED:
 344                 unfreeze_cgroup(cgroup, freezer);
 345                 break;
 346         case CGROUP_FROZEN:
 347                 retval = try_to_freeze_cgroup(cgroup, freezer);
 348                 break;
 349         default:
 350                 BUG();
 351         }
 352 out:
 353         spin_unlock_irq(&freezer->lock);
 354
 355         return retval;
 356 }
 357
 358 static int freezer_write(struct cgroup *cgroup,
 359                          struct cftype *cft,
 360                          const char *buffer)
 361 {
 362         int retval;
 363         enum freezer_state goal_state;
 364
 365         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 366                 goal_state = CGROUP_THAWED;
 367         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 368                 goal_state = CGROUP_FROZEN;
 369         else
 370                 return -EINVAL;
 371
 372         if (!cgroup_lock_live_group(cgroup))
 373                 return -ENODEV;
 374         retval = freezer_change_state(cgroup, goal_state);
 375         cgroup_unlock();
 376         return retval;
 377 }
 378
 379 static struct cftype files[] = {
 380         {
 381                 .name = "state",
 382                 .read_seq_string = freezer_read,
 383                 .write_string = freezer_write,
 384         },
 385 };
 386
 387 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
 388 {
 389         if (!cgroup->parent)
 390                 return 0;
 391         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
 392 }
 393
 394 struct cgroup_subsys freezer_subsys = {
 395         .name           = "freezer",
 396         .create         = freezer_create,
 397         .destroy        = freezer_destroy,
 398         .populate       = freezer_populate,
 399         .subsys_id      = freezer_subsys_id,
 400         .can_attach     = freezer_can_attach,
 401         .attach         = NULL,
 402         .fork           = freezer_fork,
 403         .exit           = NULL,
 404 };