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_flags {
  26         CGROUP_FREEZER_ONLINE   = (1 << 0), /* freezer is fully online */
  27         CGROUP_FREEZING_SELF    = (1 << 1), /* this freezer is freezing */
  28         CGROUP_FREEZING_PARENT  = (1 << 2), /* the parent freezer is freezing */
  29         CGROUP_FROZEN           = (1 << 3), /* this and its descendants frozen */
  30
  31         /* mask for all FREEZING flags */
  32         CGROUP_FREEZING         = CGROUP_FREEZING_SELF | CGROUP_FREEZING_PARENT,
  33 };
  34
  35 struct freezer {
  36         struct cgroup_subsys_state      css;
  37         unsigned int                    state;
  38         spinlock_t                      lock;
  39 };
  40
  41 static inline struct freezer *cgroup_freezer(struct cgroup *cgroup)
  42 {
  43         return container_of(cgroup_subsys_state(cgroup, freezer_subsys_id),
  44                             struct freezer, css);
  45 }
  46
  47 static inline struct freezer *task_freezer(struct task_struct *task)
  48 {
  49         return container_of(task_subsys_state(task, freezer_subsys_id),
  50                             struct freezer, css);
  51 }
  52
  53 bool cgroup_freezing(struct task_struct *task)
  54 {
  55         bool ret;
  56
  57         rcu_read_lock();
  58         ret = task_freezer(task)->state & CGROUP_FREEZING;
  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(unsigned int state)
  69 {
  70         if (state & CGROUP_FROZEN)
  71                 return "FROZEN";
  72         if (state & CGROUP_FREEZING)
  73                 return "FREEZING";
  74         return "THAWED";
  75 };
  76
  77 /*
  78  * State diagram
  79  * Transitions are caused by userspace writes to the freezer.state file.
  80  * The values in parenthesis are state labels. The rest are edge labels.
  81  *
  82  * (THAWED) --FROZEN--> (FREEZING) --FROZEN--> (FROZEN)
  83  *    ^ ^                    |                     |
  84  *    | \_______THAWED_______/                     |
  85  *    \__________________________THAWED____________/
  86  */
  87
  88 struct cgroup_subsys freezer_subsys;
  89
  90 static struct cgroup_subsys_state *freezer_create(struct cgroup *cgroup)
  91 {
  92         struct freezer *freezer;
  93
  94         freezer = kzalloc(sizeof(struct freezer), GFP_KERNEL);
  95         if (!freezer)
  96                 return ERR_PTR(-ENOMEM);
  97
  98         spin_lock_init(&freezer->lock);
  99         return &freezer->css;
 100 }
 101
 102 /**
 103  * freezer_post_create - commit creation of a freezer cgroup
 104  * @cgroup: cgroup being created
 105  *
 106  * We're committing to creation of @cgroup.  Mark it online.
 107  */
 108 static void freezer_post_create(struct cgroup *cgroup)
 109 {
 110         struct freezer *freezer = cgroup_freezer(cgroup);
 111
 112         spin_lock_irq(&freezer->lock);
 113         freezer->state |= CGROUP_FREEZER_ONLINE;
 114         spin_unlock_irq(&freezer->lock);
 115 }
 116
 117 /**
 118  * freezer_pre_destroy - initiate destruction of @cgroup
 119  * @cgroup: cgroup being destroyed
 120  *
 121  * @cgroup is going away.  Mark it dead and decrement system_freezing_count
 122  * if it was holding one.
 123  */
 124 static void freezer_pre_destroy(struct cgroup *cgroup)
 125 {
 126         struct freezer *freezer = cgroup_freezer(cgroup);
 127
 128         spin_lock_irq(&freezer->lock);
 129
 130         if (freezer->state & CGROUP_FREEZING)
 131                 atomic_dec(&system_freezing_cnt);
 132
 133         freezer->state = 0;
 134
 135         spin_unlock_irq(&freezer->lock);
 136 }
 137
 138 static void freezer_destroy(struct cgroup *cgroup)
 139 {
 140         kfree(cgroup_freezer(cgroup));
 141 }
 142
 143 /*
 144  * Tasks can be migrated into a different freezer anytime regardless of its
 145  * current state.  freezer_attach() is responsible for making new tasks
 146  * conform to the current state.
 147  *
 148  * Freezer state changes and task migration are synchronized via
 149  * @freezer->lock.  freezer_attach() makes the new tasks conform to the
 150  * current state and all following state changes can see the new tasks.
 151  */
 152 static void freezer_attach(struct cgroup *new_cgrp, struct cgroup_taskset *tset)
 153 {
 154         struct freezer *freezer = cgroup_freezer(new_cgrp);
 155         struct task_struct *task;
 156
 157         spin_lock_irq(&freezer->lock);
 158
 159         /*
 160          * Make the new tasks conform to the current state of @new_cgrp.
 161          * For simplicity, when migrating any task to a FROZEN cgroup, we
 162          * revert it to FREEZING and let update_if_frozen() determine the
 163          * correct state later.
 164          *
 165          * Tasks in @tset are on @new_cgrp but may not conform to its
 166          * current state before executing the following - !frozen tasks may
 167          * be visible in a FROZEN cgroup and frozen tasks in a THAWED one.
 168          */
 169         cgroup_taskset_for_each(task, new_cgrp, tset) {
 170                 if (!(freezer->state & CGROUP_FREEZING)) {
 171                         __thaw_task(task);
 172                 } else {
 173                         freeze_task(task);
 174                         freezer->state &= ~CGROUP_FROZEN;
 175                 }
 176         }
 177
 178         spin_unlock_irq(&freezer->lock);
 179 }
 180
 181 static void freezer_fork(struct task_struct *task)
 182 {
 183         struct freezer *freezer;
 184
 185         rcu_read_lock();
 186         freezer = task_freezer(task);
 187
 188         /*
 189          * The root cgroup is non-freezable, so we can skip the
 190          * following check.
 191          */
 192         if (!freezer->css.cgroup->parent)
 193                 goto out;
 194
 195         spin_lock_irq(&freezer->lock);
 196         if (freezer->state & CGROUP_FREEZING)
 197                 freeze_task(task);
 198         spin_unlock_irq(&freezer->lock);
 199 out:
 200         rcu_read_unlock();
 201 }
 202
 203 /*
 204  * We change from FREEZING to FROZEN lazily if the cgroup was only
 205  * partially frozen when we exitted write.  Caller must hold freezer->lock.
 206  *
 207  * Task states and freezer state might disagree while tasks are being
 208  * migrated into or out of @cgroup, so we can't verify task states against
 209  * @freezer state here.  See freezer_attach() for details.
 210  */
 211 static void update_if_frozen(struct freezer *freezer)
 212 {
 213         struct cgroup *cgroup = freezer->css.cgroup;
 214         struct cgroup_iter it;
 215         struct task_struct *task;
 216
 217         if (!(freezer->state & CGROUP_FREEZING) ||
 218             (freezer->state & CGROUP_FROZEN))
 219                 return;
 220
 221         cgroup_iter_start(cgroup, &it);
 222
 223         while ((task = cgroup_iter_next(cgroup, &it))) {
 224                 if (freezing(task)) {
 225                         /*
 226                          * freezer_should_skip() indicates that the task
 227                          * should be skipped when determining freezing
 228                          * completion.  Consider it frozen in addition to
 229                          * the usual frozen condition.
 230                          */
 231                         if (!frozen(task) && !freezer_should_skip(task))
 232                                 goto notyet;
 233                 }
 234         }
 235
 236         freezer->state |= CGROUP_FROZEN;
 237 notyet:
 238         cgroup_iter_end(cgroup, &it);
 239 }
 240
 241 static int freezer_read(struct cgroup *cgroup, struct cftype *cft,
 242                         struct seq_file *m)
 243 {
 244         struct freezer *freezer = cgroup_freezer(cgroup);
 245         unsigned int state;
 246
 247         spin_lock_irq(&freezer->lock);
 248         update_if_frozen(freezer);
 249         state = freezer->state;
 250         spin_unlock_irq(&freezer->lock);
 251
 252         seq_puts(m, freezer_state_strs(state));
 253         seq_putc(m, '\n');
 254         return 0;
 255 }
 256
 257 static void freeze_cgroup(struct freezer *freezer)
 258 {
 259         struct cgroup *cgroup = freezer->css.cgroup;
 260         struct cgroup_iter it;
 261         struct task_struct *task;
 262
 263         cgroup_iter_start(cgroup, &it);
 264         while ((task = cgroup_iter_next(cgroup, &it)))
 265                 freeze_task(task);
 266         cgroup_iter_end(cgroup, &it);
 267 }
 268
 269 static void unfreeze_cgroup(struct freezer *freezer)
 270 {
 271         struct cgroup *cgroup = freezer->css.cgroup;
 272         struct cgroup_iter it;
 273         struct task_struct *task;
 274
 275         cgroup_iter_start(cgroup, &it);
 276         while ((task = cgroup_iter_next(cgroup, &it)))
 277                 __thaw_task(task);
 278         cgroup_iter_end(cgroup, &it);
 279 }
 280
 281 /**
 282  * freezer_apply_state - apply state change to a single cgroup_freezer
 283  * @freezer: freezer to apply state change to
 284  * @freeze: whether to freeze or unfreeze
 285  * @state: CGROUP_FREEZING_* flag to set or clear
 286  *
 287  * Set or clear @state on @cgroup according to @freeze, and perform
 288  * freezing or thawing as necessary.
 289  */
 290 static void freezer_apply_state(struct freezer *freezer, bool freeze,
 291                                 unsigned int state)
 292 {
 293         /* also synchronizes against task migration, see freezer_attach() */
 294         lockdep_assert_held(&freezer->lock);
 295
 296         if (!(freezer->state & CGROUP_FREEZER_ONLINE))
 297                 return;
 298
 299         if (freeze) {
 300                 if (!(freezer->state & CGROUP_FREEZING))
 301                         atomic_inc(&system_freezing_cnt);
 302                 freezer->state |= state;
 303                 freeze_cgroup(freezer);
 304         } else {
 305                 bool was_freezing = freezer->state & CGROUP_FREEZING;
 306
 307                 freezer->state &= ~state;
 308
 309                 if (!(freezer->state & CGROUP_FREEZING)) {
 310                         if (was_freezing)
 311                                 atomic_dec(&system_freezing_cnt);
 312                         freezer->state &= ~CGROUP_FROZEN;
 313                         unfreeze_cgroup(freezer);
 314                 }
 315         }
 316 }
 317
 318 /**
 319  * freezer_change_state - change the freezing state of a cgroup_freezer
 320  * @freezer: freezer of interest
 321  * @freeze: whether to freeze or thaw
 322  *
 323  * Freeze or thaw @cgroup according to @freeze.
 324  */
 325 static void freezer_change_state(struct freezer *freezer, bool freeze)
 326 {
 327         /* update @freezer */
 328         spin_lock_irq(&freezer->lock);
 329         freezer_apply_state(freezer, freeze, CGROUP_FREEZING_SELF);
 330         spin_unlock_irq(&freezer->lock);
 331 }
 332
 333 static int freezer_write(struct cgroup *cgroup, struct cftype *cft,
 334                          const char *buffer)
 335 {
 336         bool freeze;
 337
 338         if (strcmp(buffer, freezer_state_strs(0)) == 0)
 339                 freeze = false;
 340         else if (strcmp(buffer, freezer_state_strs(CGROUP_FROZEN)) == 0)
 341                 freeze = true;
 342         else
 343                 return -EINVAL;
 344
 345         freezer_change_state(cgroup_freezer(cgroup), freeze);
 346         return 0;
 347 }
 348
 349 static u64 freezer_self_freezing_read(struct cgroup *cgroup, struct cftype *cft)
 350 {
 351         struct freezer *freezer = cgroup_freezer(cgroup);
 352
 353         return (bool)(freezer->state & CGROUP_FREEZING_SELF);
 354 }
 355
 356 static u64 freezer_parent_freezing_read(struct cgroup *cgroup, struct cftype *cft)
 357 {
 358         struct freezer *freezer = cgroup_freezer(cgroup);
 359
 360         return (bool)(freezer->state & CGROUP_FREEZING_PARENT);
 361 }
 362
 363 static struct cftype files[] = {
 364         {
 365                 .name = "state",
 366                 .flags = CFTYPE_NOT_ON_ROOT,
 367                 .read_seq_string = freezer_read,
 368                 .write_string = freezer_write,
 369         },
 370         {
 371                 .name = "self_freezing",
 372                 .flags = CFTYPE_NOT_ON_ROOT,
 373                 .read_u64 = freezer_self_freezing_read,
 374         },
 375         {
 376                 .name = "parent_freezing",
 377                 .flags = CFTYPE_NOT_ON_ROOT,
 378                 .read_u64 = freezer_parent_freezing_read,
 379         },
 380         { }     /* terminate */
 381 };
 382
 383 struct cgroup_subsys freezer_subsys = {
 384         .name           = "freezer",
 385         .create         = freezer_create,
 386         .post_create    = freezer_post_create,
 387         .pre_destroy    = freezer_pre_destroy,
 388         .destroy        = freezer_destroy,
 389         .subsys_id      = freezer_subsys_id,
 390         .attach         = freezer_attach,
 391         .fork           = freezer_fork,
 392         .base_cftypes   = files,
 393
 394         /*
 395          * freezer subsys doesn't handle hierarchy at all.  Frozen state
 396          * should be inherited through the hierarchy - if a parent is
 397          * frozen, all its children should be frozen.  Fix it and remove
 398          * the following.
 399          */
 400         .broken_hierarchy = true,
 401 };