kernel/pid.c

   1 /*
   2  * Generic pidhash and scalable, time-bounded PID allocator
   3  *
   4  * (C) 2002-2003 William Irwin, IBM
   5  * (C) 2004 William Irwin, Oracle
   6  * (C) 2002-2004 Ingo Molnar, Red Hat
   7  *
   8  * pid-structures are backing objects for tasks sharing a given ID to chain
   9  * against. There is very little to them aside from hashing them and
  10  * parking tasks using given ID's on a list.
  11  *
  12  * The hash is always changed with the tasklist_lock write-acquired,
  13  * and the hash is only accessed with the tasklist_lock at least
  14  * read-acquired, so there's no additional SMP locking needed here.
  15  *
  16  * We have a list of bitmap pages, which bitmaps represent the PID space.
  17  * Allocating and freeing PIDs is completely lockless. The worst-case
  18  * allocation scenario when all but one out of 1 million PIDs possible are
  19  * allocated already: the scanning of 32 list entries and at most PAGE_SIZE
  20  * bytes. The typical fastpath is a single successful setbit. Freeing is O(1).
  21  *
  22  * Pid namespaces:
  23  *    (C) 2007 Pavel Emelyanov <xemul@openvz.org>, OpenVZ, SWsoft Inc.
  24  *    (C) 2007 Sukadev Bhattiprolu <sukadev@us.ibm.com>, IBM
  25  *     Many thanks to Oleg Nesterov for comments and help
  26  *
  27  */
  28
  29 #include <linux/mm.h>
  30 #include <linux/module.h>
  31 #include <linux/slab.h>
  32 #include <linux/init.h>
  33 #include <linux/rculist.h>
  34 #include <linux/bootmem.h>
  35 #include <linux/hash.h>
  36 #include <linux/pid_namespace.h>
  37 #include <linux/init_task.h>
  38 #include <linux/syscalls.h>
  39
  40 #define pid_hashfn(nr, ns)      \
  41         hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
  42 static struct hlist_head *pid_hash;
  43 static unsigned int pidhash_shift = 4;
  44 struct pid init_struct_pid = INIT_STRUCT_PID;
  45
  46 int pid_max = PID_MAX_DEFAULT;
  47
  48 #define RESERVED_PIDS           300
  49
  50 int pid_max_min = RESERVED_PIDS + 1;
  51 int pid_max_max = PID_MAX_LIMIT;
  52
  53 #define BITS_PER_PAGE           (PAGE_SIZE*8)
  54 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  55
  56 static inline int mk_pid(struct pid_namespace *pid_ns,
  57                 struct pidmap *map, int off)
  58 {
  59         return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
  60 }
  61
  62 #define find_next_offset(map, off)                                      \
  63                 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
  64
  65 /*
  66  * PID-map pages start out as NULL, they get allocated upon
  67  * first use and are never deallocated. This way a low pid_max
  68  * value does not cause lots of bitmaps to be allocated, but
  69  * the scheme scales to up to 4 million PIDs, runtime.
  70  */
  71 struct pid_namespace init_pid_ns = {
  72         .kref = {
  73                 .refcount       = ATOMIC_INIT(2),
  74         },
  75         .pidmap = {
  76                 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
  77         },
  78         .last_pid = 0,
  79         .level = 0,
  80         .child_reaper = &init_task,
  81 };
  82 EXPORT_SYMBOL_GPL(init_pid_ns);
  83
  84 int is_container_init(struct task_struct *tsk)
  85 {
  86         int ret = 0;
  87         struct pid *pid;
  88
  89         rcu_read_lock();
  90         pid = task_pid(tsk);
  91         if (pid != NULL && pid->numbers[pid->level].nr == 1)
  92                 ret = 1;
  93         rcu_read_unlock();
  94
  95         return ret;
  96 }
  97 EXPORT_SYMBOL(is_container_init);
  98
  99 /*
 100  * Note: disable interrupts while the pidmap_lock is held as an
 101  * interrupt might come in and do read_lock(&tasklist_lock).
 102  *
 103  * If we don't disable interrupts there is a nasty deadlock between
 104  * detach_pid()->free_pid() and another cpu that does
 105  * spin_lock(&pidmap_lock) followed by an interrupt routine that does
 106  * read_lock(&tasklist_lock);
 107  *
 108  * After we clean up the tasklist_lock and know there are no
 109  * irq handlers that take it we can leave the interrupts enabled.
 110  * For now it is easier to be safe than to prove it can't happen.
 111  */
 112
 113 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
 114
 115 static void free_pidmap(struct upid *upid)
 116 {
 117         int nr = upid->nr;
 118         struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
 119         int offset = nr & BITS_PER_PAGE_MASK;
 120
 121         clear_bit(offset, map->page);
 122         atomic_inc(&map->nr_free);
 123 }
 124
 125 static int alloc_pidmap(struct pid_namespace *pid_ns)
 126 {
 127         int i, offset, max_scan, pid, last = pid_ns->last_pid;
 128         struct pidmap *map;
 129
 130         pid = last + 1;
 131         if (pid >= pid_max)
 132                 pid = RESERVED_PIDS;
 133         offset = pid & BITS_PER_PAGE_MASK;
 134         map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
 135         max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
 136         for (i = 0; i <= max_scan; ++i) {
 137                 if (unlikely(!map->page)) {
 138                         void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 139                         /*
 140                          * Free the page if someone raced with us
 141                          * installing it:
 142                          */
 143                         spin_lock_irq(&pidmap_lock);
 144                         if (!map->page) {
 145                                 map->page = page;
 146                                 page = NULL;
 147                         }
 148                         spin_unlock_irq(&pidmap_lock);
 149                         kfree(page);
 150                         if (unlikely(!map->page))
 151                                 break;
 152                 }
 153                 if (likely(atomic_read(&map->nr_free))) {
 154                         do {
 155                                 if (!test_and_set_bit(offset, map->page)) {
 156                                         atomic_dec(&map->nr_free);
 157                                         pid_ns->last_pid = pid;
 158                                         return pid;
 159                                 }
 160                                 offset = find_next_offset(map, offset);
 161                                 pid = mk_pid(pid_ns, map, offset);
 162                         /*
 163                          * find_next_offset() found a bit, the pid from it
 164                          * is in-bounds, and if we fell back to the last
 165                          * bitmap block and the final block was the same
 166                          * as the starting point, pid is before last_pid.
 167                          */
 168                         } while (offset < BITS_PER_PAGE && pid < pid_max &&
 169                                         (i != max_scan || pid < last ||
 170                                             !((last+1) & BITS_PER_PAGE_MASK)));
 171                 }
 172                 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
 173                         ++map;
 174                         offset = 0;
 175                 } else {
 176                         map = &pid_ns->pidmap[0];
 177                         offset = RESERVED_PIDS;
 178                         if (unlikely(last == offset))
 179                                 break;
 180                 }
 181                 pid = mk_pid(pid_ns, map, offset);
 182         }
 183         return -1;
 184 }
 185
 186 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
 187 {
 188         int offset;
 189         struct pidmap *map, *end;
 190
 191         if (last >= PID_MAX_LIMIT)
 192                 return -1;
 193
 194         offset = (last + 1) & BITS_PER_PAGE_MASK;
 195         map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
 196         end = &pid_ns->pidmap[PIDMAP_ENTRIES];
 197         for (; map < end; map++, offset = 0) {
 198                 if (unlikely(!map->page))
 199                         continue;
 200                 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
 201                 if (offset < BITS_PER_PAGE)
 202                         return mk_pid(pid_ns, map, offset);
 203         }
 204         return -1;
 205 }
 206
 207 void put_pid(struct pid *pid)
 208 {
 209         struct pid_namespace *ns;
 210
 211         if (!pid)
 212                 return;
 213
 214         ns = pid->numbers[pid->level].ns;
 215         if ((atomic_read(&pid->count) == 1) ||
 216              atomic_dec_and_test(&pid->count)) {
 217                 kmem_cache_free(ns->pid_cachep, pid);
 218                 put_pid_ns(ns);
 219         }
 220 }
 221 EXPORT_SYMBOL_GPL(put_pid);
 222
 223 static void delayed_put_pid(struct rcu_head *rhp)
 224 {
 225         struct pid *pid = container_of(rhp, struct pid, rcu);
 226         put_pid(pid);
 227 }
 228
 229 void free_pid(struct pid *pid)
 230 {
 231         /* We can be called with write_lock_irq(&tasklist_lock) held */
 232         int i;
 233         unsigned long flags;
 234
 235         spin_lock_irqsave(&pidmap_lock, flags);
 236         for (i = 0; i <= pid->level; i++)
 237                 hlist_del_rcu(&pid->numbers[i].pid_chain);
 238         spin_unlock_irqrestore(&pidmap_lock, flags);
 239
 240         for (i = 0; i <= pid->level; i++)
 241                 free_pidmap(pid->numbers + i);
 242
 243         call_rcu(&pid->rcu, delayed_put_pid);
 244 }
 245
 246 struct pid *alloc_pid(struct pid_namespace *ns)
 247 {
 248         struct pid *pid;
 249         enum pid_type type;
 250         int i, nr;
 251         struct pid_namespace *tmp;
 252         struct upid *upid;
 253
 254         pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
 255         if (!pid)
 256                 goto out;
 257
 258         tmp = ns;
 259         for (i = ns->level; i >= 0; i--) {
 260                 nr = alloc_pidmap(tmp);
 261                 if (nr < 0)
 262                         goto out_free;
 263
 264                 pid->numbers[i].nr = nr;
 265                 pid->numbers[i].ns = tmp;
 266                 tmp = tmp->parent;
 267         }
 268
 269         get_pid_ns(ns);
 270         pid->level = ns->level;
 271         atomic_set(&pid->count, 1);
 272         for (type = 0; type < PIDTYPE_MAX; ++type)
 273                 INIT_HLIST_HEAD(&pid->tasks[type]);
 274
 275         upid = pid->numbers + ns->level;
 276         spin_lock_irq(&pidmap_lock);
 277         for ( ; upid >= pid->numbers; --upid)
 278                 hlist_add_head_rcu(&upid->pid_chain,
 279                                 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
 280         spin_unlock_irq(&pidmap_lock);
 281
 282 out:
 283         return pid;
 284
 285 out_free:
 286         while (++i <= ns->level)
 287                 free_pidmap(pid->numbers + i);
 288
 289         kmem_cache_free(ns->pid_cachep, pid);
 290         pid = NULL;
 291         goto out;
 292 }
 293
 294 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
 295 {
 296         struct hlist_node *elem;
 297         struct upid *pnr;
 298
 299         hlist_for_each_entry_rcu(pnr, elem,
 300                         &pid_hash[pid_hashfn(nr, ns)], pid_chain)
 301                 if (pnr->nr == nr && pnr->ns == ns)
 302                         return container_of(pnr, struct pid,
 303                                         numbers[ns->level]);
 304
 305         return NULL;
 306 }
 307 EXPORT_SYMBOL_GPL(find_pid_ns);
 308
 309 struct pid *find_vpid(int nr)
 310 {
 311         return find_pid_ns(nr, current->nsproxy->pid_ns);
 312 }
 313 EXPORT_SYMBOL_GPL(find_vpid);
 314
 315 /*
 316  * attach_pid() must be called with the tasklist_lock write-held.
 317  */
 318 void attach_pid(struct task_struct *task, enum pid_type type,
 319                 struct pid *pid)
 320 {
 321         struct pid_link *link;
 322
 323         link = &task->pids[type];
 324         link->pid = pid;
 325         hlist_add_head_rcu(&link->node, &pid->tasks[type]);
 326 }
 327
 328 static void __change_pid(struct task_struct *task, enum pid_type type,
 329                         struct pid *new)
 330 {
 331         struct pid_link *link;
 332         struct pid *pid;
 333         int tmp;
 334
 335         link = &task->pids[type];
 336         pid = link->pid;
 337
 338         hlist_del_rcu(&link->node);
 339         link->pid = new;
 340
 341         for (tmp = PIDTYPE_MAX; --tmp >= 0; )
 342                 if (!hlist_empty(&pid->tasks[tmp]))
 343                         return;
 344
 345         free_pid(pid);
 346 }
 347
 348 void detach_pid(struct task_struct *task, enum pid_type type)
 349 {
 350         __change_pid(task, type, NULL);
 351 }
 352
 353 void change_pid(struct task_struct *task, enum pid_type type,
 354                 struct pid *pid)
 355 {
 356         __change_pid(task, type, pid);
 357         attach_pid(task, type, pid);
 358 }
 359
 360 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
 361 void transfer_pid(struct task_struct *old, struct task_struct *new,
 362                            enum pid_type type)
 363 {
 364         new->pids[type].pid = old->pids[type].pid;
 365         hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
 366 }
 367
 368 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
 369 {
 370         struct task_struct *result = NULL;
 371         if (pid) {
 372                 struct hlist_node *first;
 373                 first = rcu_dereference_check(pid->tasks[type].first,
 374                                               rcu_read_lock_held() ||
 375                                               lockdep_tasklist_lock_is_held());
 376                 if (first)
 377                         result = hlist_entry(first, struct task_struct, pids[(type)].node);
 378         }
 379         return result;
 380 }
 381 EXPORT_SYMBOL(pid_task);
 382
 383 /*
 384  * Must be called under rcu_read_lock().
 385  */
 386 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
 387 {
 388         return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
 389 }
 390
 391 struct task_struct *find_task_by_vpid(pid_t vnr)
 392 {
 393         return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns);
 394 }
 395
 396 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
 397 {
 398         struct pid *pid;
 399         rcu_read_lock();
 400         if (type != PIDTYPE_PID)
 401                 task = task->group_leader;
 402         pid = get_pid(task->pids[type].pid);
 403         rcu_read_unlock();
 404         return pid;
 405 }
 406
 407 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
 408 {
 409         struct task_struct *result;
 410         rcu_read_lock();
 411         result = pid_task(pid, type);
 412         if (result)
 413                 get_task_struct(result);
 414         rcu_read_unlock();
 415         return result;
 416 }
 417
 418 struct pid *find_get_pid(pid_t nr)
 419 {
 420         struct pid *pid;
 421
 422         rcu_read_lock();
 423         pid = get_pid(find_vpid(nr));
 424         rcu_read_unlock();
 425
 426         return pid;
 427 }
 428 EXPORT_SYMBOL_GPL(find_get_pid);
 429
 430 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
 431 {
 432         struct upid *upid;
 433         pid_t nr = 0;
 434
 435         if (pid && ns->level <= pid->level) {
 436                 upid = &pid->numbers[ns->level];
 437                 if (upid->ns == ns)
 438                         nr = upid->nr;
 439         }
 440         return nr;
 441 }
 442
 443 pid_t pid_vnr(struct pid *pid)
 444 {
 445         return pid_nr_ns(pid, current->nsproxy->pid_ns);
 446 }
 447 EXPORT_SYMBOL_GPL(pid_vnr);
 448
 449 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
 450                         struct pid_namespace *ns)
 451 {
 452         pid_t nr = 0;
 453
 454         rcu_read_lock();
 455         if (!ns)
 456                 ns = current->nsproxy->pid_ns;
 457         if (likely(pid_alive(task))) {
 458                 if (type != PIDTYPE_PID)
 459                         task = task->group_leader;
 460                 nr = pid_nr_ns(task->pids[type].pid, ns);
 461         }
 462         rcu_read_unlock();
 463
 464         return nr;
 465 }
 466 EXPORT_SYMBOL(__task_pid_nr_ns);
 467
 468 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
 469 {
 470         return pid_nr_ns(task_tgid(tsk), ns);
 471 }
 472 EXPORT_SYMBOL(task_tgid_nr_ns);
 473
 474 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
 475 {
 476         return ns_of_pid(task_pid(tsk));
 477 }
 478 EXPORT_SYMBOL_GPL(task_active_pid_ns);
 479
 480 /*
 481  * Used by proc to find the first pid that is greater than or equal to nr.
 482  *
 483  * If there is a pid at nr this function is exactly the same as find_pid_ns.
 484  */
 485 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
 486 {
 487         struct pid *pid;
 488
 489         do {
 490                 pid = find_pid_ns(nr, ns);
 491                 if (pid)
 492                         break;
 493                 nr = next_pidmap(ns, nr);
 494         } while (nr > 0);
 495
 496         return pid;
 497 }
 498
 499 /*
 500  * The pid hash table is scaled according to the amount of memory in the
 501  * machine.  From a minimum of 16 slots up to 4096 slots at one gigabyte or
 502  * more.
 503  */
 504 void __init pidhash_init(void)
 505 {
 506         int i, pidhash_size;
 507
 508         pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
 509                                            HASH_EARLY | HASH_SMALL,
 510                                            &pidhash_shift, NULL, 4096);
 511         pidhash_size = 1 << pidhash_shift;
 512
 513         for (i = 0; i < pidhash_size; i++)
 514                 INIT_HLIST_HEAD(&pid_hash[i]);
 515 }
 516
 517 void __init pidmap_init(void)
 518 {
 519         /* bump default and minimum pid_max based on number of cpus */
 520         pid_max = min(pid_max_max, max_t(int, pid_max,
 521                                 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
 522         pid_max_min = max_t(int, pid_max_min,
 523                                 PIDS_PER_CPU_MIN * num_possible_cpus());
 524         pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
 525
 526         init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 527         /* Reserve PID 0. We never call free_pidmap(0) */
 528         set_bit(0, init_pid_ns.pidmap[0].page);
 529         atomic_dec(&init_pid_ns.pidmap[0].nr_free);
 530
 531         init_pid_ns.pid_cachep = KMEM_CACHE(pid,
 532                         SLAB_HWCACHE_ALIGN | SLAB_PANIC);
 533 }