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