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