kernel/user.c

   1 /*
   2  * The "user cache".
   3  *
   4  * (C) Copyright 1991-2000 Linus Torvalds
   5  *
   6  * We have a per-user structure to keep track of how many
   7  * processes, files etc the user has claimed, in order to be
   8  * able to have per-user limits for system resources.
   9  */
  10
  11 #include <linux/init.h>
  12 #include <linux/sched.h>
  13 #include <linux/slab.h>
  14 #include <linux/bitops.h>
  15 #include <linux/key.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/module.h>
  18 #include <linux/user_namespace.h>
  19
  20 struct user_namespace init_user_ns = {
  21         .kref = {
  22                 .refcount       = ATOMIC_INIT(2),
  23         },
  24         .root_user = &root_user,
  25 };
  26 EXPORT_SYMBOL_GPL(init_user_ns);
  27
  28 /*
  29  * UID task count cache, to get fast user lookup in "alloc_uid"
  30  * when changing user ID's (ie setuid() and friends).
  31  */
  32
  33 #define UIDHASH_MASK            (UIDHASH_SZ - 1)
  34 #define __uidhashfn(uid)        (((uid >> UIDHASH_BITS) + uid) & UIDHASH_MASK)
  35 #define uidhashentry(ns, uid)   ((ns)->uidhash_table + __uidhashfn((uid)))
  36
  37 static struct kmem_cache *uid_cachep;
  38
  39 /*
  40  * The uidhash_lock is mostly taken from process context, but it is
  41  * occasionally also taken from softirq/tasklet context, when
  42  * task-structs get RCU-freed. Hence all locking must be softirq-safe.
  43  * But free_uid() is also called with local interrupts disabled, and running
  44  * local_bh_enable() with local interrupts disabled is an error - we'll run
  45  * softirq callbacks, and they can unconditionally enable interrupts, and
  46  * the caller of free_uid() didn't expect that..
  47  */
  48 static DEFINE_SPINLOCK(uidhash_lock);
  49
  50 struct user_struct root_user = {
  51         .__count        = ATOMIC_INIT(1),
  52         .processes      = ATOMIC_INIT(1),
  53         .files          = ATOMIC_INIT(0),
  54         .sigpending     = ATOMIC_INIT(0),
  55         .locked_shm     = 0,
  56 #ifdef CONFIG_KEYS
  57         .uid_keyring    = &root_user_keyring,
  58         .session_keyring = &root_session_keyring,
  59 #endif
  60 #ifdef CONFIG_USER_SCHED
  61         .tg             = &init_task_group,
  62 #endif
  63 };
  64
  65 /*
  66  * These routines must be called with the uidhash spinlock held!
  67  */
  68 static void uid_hash_insert(struct user_struct *up, struct hlist_head *hashent)
  69 {
  70         hlist_add_head(&up->uidhash_node, hashent);
  71 }
  72
  73 static void uid_hash_remove(struct user_struct *up)
  74 {
  75         hlist_del_init(&up->uidhash_node);
  76 }
  77
  78 static struct user_struct *uid_hash_find(uid_t uid, struct hlist_head *hashent)
  79 {
  80         struct user_struct *user;
  81         struct hlist_node *h;
  82
  83         hlist_for_each_entry(user, h, hashent, uidhash_node) {
  84                 if (user->uid == uid) {
  85                         atomic_inc(&user->__count);
  86                         return user;
  87                 }
  88         }
  89
  90         return NULL;
  91 }
  92
  93 #ifdef CONFIG_USER_SCHED
  94
  95 static void sched_destroy_user(struct user_struct *up)
  96 {
  97         sched_destroy_group(up->tg);
  98 }
  99
 100 static int sched_create_user(struct user_struct *up)
 101 {
 102         int rc = 0;
 103
 104         up->tg = sched_create_group();
 105         if (IS_ERR(up->tg))
 106                 rc = -ENOMEM;
 107
 108         return rc;
 109 }
 110
 111 static void sched_switch_user(struct task_struct *p)
 112 {
 113         sched_move_task(p);
 114 }
 115
 116 #else   /* CONFIG_USER_SCHED */
 117
 118 static void sched_destroy_user(struct user_struct *up) { }
 119 static int sched_create_user(struct user_struct *up) { return 0; }
 120 static void sched_switch_user(struct task_struct *p) { }
 121
 122 #endif  /* CONFIG_USER_SCHED */
 123
 124 #if defined(CONFIG_USER_SCHED) && defined(CONFIG_SYSFS)
 125
 126 static struct kset *uids_kset; /* represents the /sys/kernel/uids/ directory */
 127 static DEFINE_MUTEX(uids_mutex);
 128
 129 static inline void uids_mutex_lock(void)
 130 {
 131         mutex_lock(&uids_mutex);
 132 }
 133
 134 static inline void uids_mutex_unlock(void)
 135 {
 136         mutex_unlock(&uids_mutex);
 137 }
 138
 139 /* uid directory attributes */
 140 #ifdef CONFIG_FAIR_GROUP_SCHED
 141 static ssize_t cpu_shares_show(struct kobject *kobj,
 142                                struct kobj_attribute *attr,
 143                                char *buf)
 144 {
 145         struct user_struct *up = container_of(kobj, struct user_struct, kobj);
 146
 147         return sprintf(buf, "%lu\n", sched_group_shares(up->tg));
 148 }
 149
 150 static ssize_t cpu_shares_store(struct kobject *kobj,
 151                                 struct kobj_attribute *attr,
 152                                 const char *buf, size_t size)
 153 {
 154         struct user_struct *up = container_of(kobj, struct user_struct, kobj);
 155         unsigned long shares;
 156         int rc;
 157
 158         sscanf(buf, "%lu", &shares);
 159
 160         rc = sched_group_set_shares(up->tg, shares);
 161
 162         return (rc ? rc : size);
 163 }
 164
 165 static struct kobj_attribute cpu_share_attr =
 166         __ATTR(cpu_share, 0644, cpu_shares_show, cpu_shares_store);
 167 #endif
 168
 169 #ifdef CONFIG_RT_GROUP_SCHED
 170 static ssize_t cpu_rt_runtime_show(struct kobject *kobj,
 171                                    struct kobj_attribute *attr,
 172                                    char *buf)
 173 {
 174         struct user_struct *up = container_of(kobj, struct user_struct, kobj);
 175
 176         return sprintf(buf, "%lu\n", sched_group_rt_runtime(up->tg));
 177 }
 178
 179 static ssize_t cpu_rt_runtime_store(struct kobject *kobj,
 180                                     struct kobj_attribute *attr,
 181                                     const char *buf, size_t size)
 182 {
 183         struct user_struct *up = container_of(kobj, struct user_struct, kobj);
 184         unsigned long rt_runtime;
 185         int rc;
 186
 187         sscanf(buf, "%lu", &rt_runtime);
 188
 189         rc = sched_group_set_rt_runtime(up->tg, rt_runtime);
 190
 191         return (rc ? rc : size);
 192 }
 193
 194 static struct kobj_attribute cpu_rt_runtime_attr =
 195         __ATTR(cpu_rt_runtime, 0644, cpu_rt_runtime_show, cpu_rt_runtime_store);
 196 #endif
 197
 198 /* default attributes per uid directory */
 199 static struct attribute *uids_attributes[] = {
 200 #ifdef CONFIG_FAIR_GROUP_SCHED
 201         &cpu_share_attr.attr,
 202 #endif
 203 #ifdef CONFIG_RT_GROUP_SCHED
 204         &cpu_rt_runtime_attr.attr,
 205 #endif
 206         NULL
 207 };
 208
 209 /* the lifetime of user_struct is not managed by the core (now) */
 210 static void uids_release(struct kobject *kobj)
 211 {
 212         return;
 213 }
 214
 215 static struct kobj_type uids_ktype = {
 216         .sysfs_ops = &kobj_sysfs_ops,
 217         .default_attrs = uids_attributes,
 218         .release = uids_release,
 219 };
 220
 221 /* create /sys/kernel/uids/<uid>/cpu_share file for this user */
 222 static int uids_user_create(struct user_struct *up)
 223 {
 224         struct kobject *kobj = &up->kobj;
 225         int error;
 226
 227         memset(kobj, 0, sizeof(struct kobject));
 228         kobj->kset = uids_kset;
 229         error = kobject_init_and_add(kobj, &uids_ktype, NULL, "%d", up->uid);
 230         if (error) {
 231                 kobject_put(kobj);
 232                 goto done;
 233         }
 234
 235         kobject_uevent(kobj, KOBJ_ADD);
 236 done:
 237         return error;
 238 }
 239
 240 /* create these entries in sysfs:
 241  *      "/sys/kernel/uids" directory
 242  *      "/sys/kernel/uids/0" directory (for root user)
 243  *      "/sys/kernel/uids/0/cpu_share" file (for root user)
 244  */
 245 int __init uids_sysfs_init(void)
 246 {
 247         uids_kset = kset_create_and_add("uids", NULL, kernel_kobj);
 248         if (!uids_kset)
 249                 return -ENOMEM;
 250
 251         return uids_user_create(&root_user);
 252 }
 253
 254 /* work function to remove sysfs directory for a user and free up
 255  * corresponding structures.
 256  */
 257 static void remove_user_sysfs_dir(struct work_struct *w)
 258 {
 259         struct user_struct *up = container_of(w, struct user_struct, work);
 260         unsigned long flags;
 261         int remove_user = 0;
 262
 263         /* Make uid_hash_remove() + sysfs_remove_file() + kobject_del()
 264          * atomic.
 265          */
 266         uids_mutex_lock();
 267
 268         local_irq_save(flags);
 269
 270         if (atomic_dec_and_lock(&up->__count, &uidhash_lock)) {
 271                 uid_hash_remove(up);
 272                 remove_user = 1;
 273                 spin_unlock_irqrestore(&uidhash_lock, flags);
 274         } else {
 275                 local_irq_restore(flags);
 276         }
 277
 278         if (!remove_user)
 279                 goto done;
 280
 281         kobject_uevent(&up->kobj, KOBJ_REMOVE);
 282         kobject_del(&up->kobj);
 283         kobject_put(&up->kobj);
 284
 285         sched_destroy_user(up);
 286         key_put(up->uid_keyring);
 287         key_put(up->session_keyring);
 288         kmem_cache_free(uid_cachep, up);
 289
 290 done:
 291         uids_mutex_unlock();
 292 }
 293
 294 /* IRQs are disabled and uidhash_lock is held upon function entry.
 295  * IRQ state (as stored in flags) is restored and uidhash_lock released
 296  * upon function exit.
 297  */
 298 static inline void free_user(struct user_struct *up, unsigned long flags)
 299 {
 300         /* restore back the count */
 301         atomic_inc(&up->__count);
 302         spin_unlock_irqrestore(&uidhash_lock, flags);
 303
 304         INIT_WORK(&up->work, remove_user_sysfs_dir);
 305         schedule_work(&up->work);
 306 }
 307
 308 #else   /* CONFIG_USER_SCHED && CONFIG_SYSFS */
 309
 310 int uids_sysfs_init(void) { return 0; }
 311 static inline int uids_user_create(struct user_struct *up) { return 0; }
 312 static inline void uids_mutex_lock(void) { }
 313 static inline void uids_mutex_unlock(void) { }
 314
 315 /* IRQs are disabled and uidhash_lock is held upon function entry.
 316  * IRQ state (as stored in flags) is restored and uidhash_lock released
 317  * upon function exit.
 318  */
 319 static inline void free_user(struct user_struct *up, unsigned long flags)
 320 {
 321         uid_hash_remove(up);
 322         spin_unlock_irqrestore(&uidhash_lock, flags);
 323         sched_destroy_user(up);
 324         key_put(up->uid_keyring);
 325         key_put(up->session_keyring);
 326         kmem_cache_free(uid_cachep, up);
 327 }
 328
 329 #endif
 330
 331 /*
 332  * Locate the user_struct for the passed UID.  If found, take a ref on it.  The
 333  * caller must undo that ref with free_uid().
 334  *
 335  * If the user_struct could not be found, return NULL.
 336  */
 337 struct user_struct *find_user(uid_t uid)
 338 {
 339         struct user_struct *ret;
 340         unsigned long flags;
 341         struct user_namespace *ns = current->nsproxy->user_ns;
 342
 343         spin_lock_irqsave(&uidhash_lock, flags);
 344         ret = uid_hash_find(uid, uidhashentry(ns, uid));
 345         spin_unlock_irqrestore(&uidhash_lock, flags);
 346         return ret;
 347 }
 348
 349 void free_uid(struct user_struct *up)
 350 {
 351         unsigned long flags;
 352
 353         if (!up)
 354                 return;
 355
 356         local_irq_save(flags);
 357         if (atomic_dec_and_lock(&up->__count, &uidhash_lock))
 358                 free_user(up, flags);
 359         else
 360                 local_irq_restore(flags);
 361 }
 362
 363 struct user_struct * alloc_uid(struct user_namespace *ns, uid_t uid)
 364 {
 365         struct hlist_head *hashent = uidhashentry(ns, uid);
 366         struct user_struct *up, *new;
 367
 368         /* Make uid_hash_find() + uids_user_create() + uid_hash_insert()
 369          * atomic.
 370          */
 371         uids_mutex_lock();
 372
 373         spin_lock_irq(&uidhash_lock);
 374         up = uid_hash_find(uid, hashent);
 375         spin_unlock_irq(&uidhash_lock);
 376
 377         if (!up) {
 378                 new = kmem_cache_alloc(uid_cachep, GFP_KERNEL);
 379                 if (!new)
 380                         goto out_unlock;
 381
 382                 new->uid = uid;
 383                 atomic_set(&new->__count, 1);
 384                 atomic_set(&new->processes, 0);
 385                 atomic_set(&new->files, 0);
 386                 atomic_set(&new->sigpending, 0);
 387 #ifdef CONFIG_INOTIFY_USER
 388                 atomic_set(&new->inotify_watches, 0);
 389                 atomic_set(&new->inotify_devs, 0);
 390 #endif
 391 #ifdef CONFIG_POSIX_MQUEUE
 392                 new->mq_bytes = 0;
 393 #endif
 394                 new->locked_shm = 0;
 395
 396                 if (alloc_uid_keyring(new, current) < 0)
 397                         goto out_free_user;
 398
 399                 if (sched_create_user(new) < 0)
 400                         goto out_put_keys;
 401
 402                 if (uids_user_create(new))
 403                         goto out_destoy_sched;
 404
 405                 /*
 406                  * Before adding this, check whether we raced
 407                  * on adding the same user already..
 408                  */
 409                 spin_lock_irq(&uidhash_lock);
 410                 up = uid_hash_find(uid, hashent);
 411                 if (up) {
 412                         /* This case is not possible when CONFIG_USER_SCHED
 413                          * is defined, since we serialize alloc_uid() using
 414                          * uids_mutex. Hence no need to call
 415                          * sched_destroy_user() or remove_user_sysfs_dir().
 416                          */
 417                         key_put(new->uid_keyring);
 418                         key_put(new->session_keyring);
 419                         kmem_cache_free(uid_cachep, new);
 420                 } else {
 421                         uid_hash_insert(new, hashent);
 422                         up = new;
 423                 }
 424                 spin_unlock_irq(&uidhash_lock);
 425
 426         }
 427
 428         uids_mutex_unlock();
 429
 430         return up;
 431
 432 out_destoy_sched:
 433         sched_destroy_user(new);
 434 out_put_keys:
 435         key_put(new->uid_keyring);
 436         key_put(new->session_keyring);
 437 out_free_user:
 438         kmem_cache_free(uid_cachep, new);
 439 out_unlock:
 440         uids_mutex_unlock();
 441         return NULL;
 442 }
 443
 444 void switch_uid(struct user_struct *new_user)
 445 {
 446         struct user_struct *old_user;
 447
 448         /* What if a process setreuid()'s and this brings the
 449          * new uid over his NPROC rlimit?  We can check this now
 450          * cheaply with the new uid cache, so if it matters
 451          * we should be checking for it.  -DaveM
 452          */
 453         old_user = current->user;
 454         atomic_inc(&new_user->processes);
 455         atomic_dec(&old_user->processes);
 456         switch_uid_keyring(new_user);
 457         current->user = new_user;
 458         sched_switch_user(current);
 459
 460         /*
 461          * We need to synchronize with __sigqueue_alloc()
 462          * doing a get_uid(p->user).. If that saw the old
 463          * user value, we need to wait until it has exited
 464          * its critical region before we can free the old
 465          * structure.
 466          */
 467         smp_mb();
 468         spin_unlock_wait(&current->sighand->siglock);
 469
 470         free_uid(old_user);
 471         suid_keys(current);
 472 }
 473
 474 #ifdef CONFIG_USER_NS
 475 void release_uids(struct user_namespace *ns)
 476 {
 477         int i;
 478         unsigned long flags;
 479         struct hlist_head *head;
 480         struct hlist_node *nd;
 481
 482         spin_lock_irqsave(&uidhash_lock, flags);
 483         /*
 484          * collapse the chains so that the user_struct-s will
 485          * be still alive, but not in hashes. subsequent free_uid()
 486          * will free them.
 487          */
 488         for (i = 0; i < UIDHASH_SZ; i++) {
 489                 head = ns->uidhash_table + i;
 490                 while (!hlist_empty(head)) {
 491                         nd = head->first;
 492                         hlist_del_init(nd);
 493                 }
 494         }
 495         spin_unlock_irqrestore(&uidhash_lock, flags);
 496
 497         free_uid(ns->root_user);
 498 }
 499 #endif
 500
 501 static int __init uid_cache_init(void)
 502 {
 503         int n;
 504
 505         uid_cachep = kmem_cache_create("uid_cache", sizeof(struct user_struct),
 506                         0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
 507
 508         for(n = 0; n < UIDHASH_SZ; ++n)
 509                 INIT_HLIST_HEAD(init_user_ns.uidhash_table + n);
 510
 511         /* Insert the root user immediately (init already runs as root) */
 512         spin_lock_irq(&uidhash_lock);
 513         uid_hash_insert(&root_user, uidhashentry(&init_user_ns, 0));
 514         spin_unlock_irq(&uidhash_lock);
 515
 516         return 0;
 517 }
 518
 519 module_init(uid_cache_init);