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_subsys *ss,
 132                                                   struct cgroup *cgroup)
 133 {
 134         struct freezer *freezer;
 135
 136         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
 137         if (!freezer)
 138                 return ERR_PTR(-ENOMEM);
 139
 140         spin_lock_init(&freezer->lock);
 141         freezer->state = CGROUP_THAWED;
 142         return &freezer->css;
 143 }
 144
 145 static void freezer_destroy(struct cgroup_subsys *ss,
 146                             struct cgroup *cgroup)
 147 {
 148         struct freezer *freezer = cgroup_freezer(cgroup);
 149
 150         if (freezer->state != CGROUP_THAWED)
 151                 atomic_dec(&system_freezing_cnt);
 152         kfree(freezer);
 153 }
 154
 155 /*
 156  * The call to cgroup_lock() in the freezer.state write method prevents
 157  * a write to that file racing against an attach, and hence the
 158  * can_attach() result will remain valid until the attach completes.
 159  */
 160 static int freezer_can_attach(struct cgroup_subsys *ss,
 161                               struct cgroup *new_cgroup,
 162                               struct task_struct *task)
 163 {
 164         struct freezer *freezer;
 165
 166         /*
 167          * Anything frozen can't move or be moved to/from.
 168          */
 169
 170         freezer = cgroup_freezer(new_cgroup);
 171         if (freezer->state != CGROUP_THAWED)
 172                 return -EBUSY;
 173
 174         return 0;
 175 }
 176
 177 static int freezer_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 178 {
 179         return cgroup_freezing(tsk) ? -EBUSY : 0;
 180 }
 181
 182 static void freezer_fork(struct cgroup_subsys *ss, struct task_struct *task)
 183 {
 184         struct freezer *freezer;
 185
 186         /*
 187          * No lock is needed, since the task isn't on tasklist yet,
 188          * so it can't be moved to another cgroup, which means the
 189          * freezer won't be removed and will be valid during this
 190          * function call.  Nevertheless, apply RCU read-side critical
 191          * section to suppress RCU lockdep false positives.
 192          */
 193         rcu_read_lock();
 194         freezer = task_freezer(task);
 195         rcu_read_unlock();
 196
 197         /*
 198          * The root cgroup is non-freezable, so we can skip the
 199          * following check.
 200          */
 201         if (!freezer->css.cgroup->parent)
 202                 return;
 203
 204         spin_lock_irq(&freezer->lock);
 205         BUG_ON(freezer->state == CGROUP_FROZEN);
 206
 207         /* Locking avoids race with FREEZING -> THAWED transitions. */
 208         if (freezer->state == CGROUP_FREEZING)
 209                 freeze_task(task);
 210         spin_unlock_irq(&freezer->lock);
 211 }
 212
 213 /*
 214  * caller must hold freezer->lock
 215  */
 216 static void update_if_frozen(struct cgroup *cgroup,
 217                                  struct freezer *freezer)
 218 {
 219         struct cgroup_iter it;
 220         struct task_struct *task;
 221         unsigned int nfrozen = 0, ntotal = 0;
 222         enum freezer_state old_state = freezer->state;
 223
 224         cgroup_iter_start(cgroup, &it);
 225         while ((task = cgroup_iter_next(cgroup, &it))) {
 226                 ntotal++;
 227                 if (freezing(task) && frozen(task))
 228                         nfrozen++;
 229         }
 230
 231         if (old_state == CGROUP_THAWED) {
 232                 BUG_ON(nfrozen > 0);
 233         } else if (old_state == CGROUP_FREEZING) {
 234                 if (nfrozen == ntotal)
 235                         freezer->state = CGROUP_FROZEN;
 236         } else { /* old_state == CGROUP_FROZEN */
 237                 BUG_ON(nfrozen != ntotal);
 238         }
 239
 240         cgroup_iter_end(cgroup, &it);
 241 }
 242
 243 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 244                         struct seq_file *m)
 245 {
 246         struct freezer *freezer;
 247         enum freezer_state state;
 248
 249         if (!cgroup_lock_live_group(cgroup))
 250                 return -ENODEV;
 251
 252         freezer = cgroup_freezer(cgroup);
 253         spin_lock_irq(&freezer->lock);
 254         state = freezer->state;
 255         if (state == CGROUP_FREEZING) {
 256                 /* We change from FREEZING to FROZEN lazily if the cgroup was
 257                  * only partially frozen when we exitted write. */
 258                 update_if_frozen(cgroup, freezer);
 259                 state = freezer->state;
 260         }
 261         spin_unlock_irq(&freezer->lock);
 262         cgroup_unlock();
 263
 264         seq_puts(m, freezer_state_strs[state]);
 265         seq_putc(m, '\n');
 266         return 0;
 267 }
 268
 269 static int try_to_freeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 270 {
 271         struct cgroup_iter it;
 272         struct task_struct *task;
 273         unsigned int num_cant_freeze_now = 0;
 274
 275         cgroup_iter_start(cgroup, &it);
 276         while ((task = cgroup_iter_next(cgroup, &it))) {
 277                 if (!freeze_task(task))
 278                         continue;
 279                 if (frozen(task))
 280                         continue;
 281                 if (!freezing(task) && !freezer_should_skip(task))
 282                         num_cant_freeze_now++;
 283         }
 284         cgroup_iter_end(cgroup, &it);
 285
 286         return num_cant_freeze_now ? -EBUSY : 0;
 287 }
 288
 289 static void unfreeze_cgroup(struct cgroup *cgroup, struct freezer *freezer)
 290 {
 291         struct cgroup_iter it;
 292         struct task_struct *task;
 293
 294         cgroup_iter_start(cgroup, &it);
 295         while ((task = cgroup_iter_next(cgroup, &it)))
 296                 __thaw_task(task);
 297         cgroup_iter_end(cgroup, &it);
 298 }
 299
 300 static int freezer_change_state(struct cgroup *cgroup,
 301                                 enum freezer_state goal_state)
 302 {
 303         struct freezer *freezer;
 304         int retval = 0;
 305
 306         freezer = cgroup_freezer(cgroup);
 307
 308         spin_lock_irq(&freezer->lock);
 309
 310         update_if_frozen(cgroup, freezer);
 311
 312         switch (goal_state) {
 313         case CGROUP_THAWED:
 314                 if (freezer->state != CGROUP_THAWED)
 315                         atomic_dec(&system_freezing_cnt);
 316                 freezer->state = CGROUP_THAWED;
 317                 unfreeze_cgroup(cgroup, freezer);
 318                 break;
 319         case CGROUP_FROZEN:
 320                 if (freezer->state == CGROUP_THAWED)
 321                         atomic_inc(&system_freezing_cnt);
 322                 freezer->state = CGROUP_FREEZING;
 323                 retval = try_to_freeze_cgroup(cgroup, freezer);
 324                 break;
 325         default:
 326                 BUG();
 327         }
 328
 329         spin_unlock_irq(&freezer->lock);
 330
 331         return retval;
 332 }
 333
 334 static int freezer_write(struct cgroup *cgroup,
 335                          struct cftype *cft,
 336                          const char *buffer)
 337 {
 338         int retval;
 339         enum freezer_state goal_state;
 340
 341         if (strcmp(buffer, freezer_state_strs[CGROUP_THAWED]) == 0)
 342                 goal_state = CGROUP_THAWED;
 343         else if (strcmp(buffer, freezer_state_strs[CGROUP_FROZEN]) == 0)
 344                 goal_state = CGROUP_FROZEN;
 345         else
 346                 return -EINVAL;
 347
 348         if (!cgroup_lock_live_group(cgroup))
 349                 return -ENODEV;
 350         retval = freezer_change_state(cgroup, goal_state);
 351         cgroup_unlock();
 352         return retval;
 353 }
 354
 355 static struct cftype files[] = {
 356         {
 357                 .name = "state",
 358                 .read_seq_string = freezer_read,
 359                 .write_string = freezer_write,
 360         },
 361 };
 362
 363 static int freezer_populate(struct cgroup_subsys *ss, struct cgroup *cgroup)
 364 {
 365         if (!cgroup->parent)
 366                 return 0;
 367         return cgroup_add_files(cgroup, ss, files, ARRAY_SIZE(files));
 368 }
 369
 370 struct cgroup_subsys freezer_subsys = {
 371         .name           = "freezer",
 372         .create         = freezer_create,
 373         .destroy        = freezer_destroy,
 374         .populate       = freezer_populate,
 375         .subsys_id      = freezer_subsys_id,
 376         .can_attach     = freezer_can_attach,
 377         .can_attach_task = freezer_can_attach_task,
 378         .pre_attach     = NULL,
 379         .attach_task    = NULL,
 380         .attach         = NULL,
 381         .fork           = freezer_fork,
 382         .exit           = NULL,
 383 };