kernel/pid.c

   1 /*
   2  * Generic pidhash and scalable, time-bounded PID allocator
   3  *
   4  * (C) 2002-2003 Nadia Yvette Chambers, IBM
   5  * (C) 2004 Nadia Yvette Chambers, 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/export.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 #include <linux/proc_ns.h>
  40 #include <linux/proc_fs.h>
  41
  42 #define pid_hashfn(nr, ns)      \
  43         hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
  44 static struct hlist_head *pid_hash;
  45 static unsigned int pidhash_shift = 4;
  46 struct pid init_struct_pid = INIT_STRUCT_PID;
  47
  48 int pid_max = PID_MAX_DEFAULT;
  49
  50 #define RESERVED_PIDS           300
  51
  52 int pid_max_min = RESERVED_PIDS + 1;
  53 int pid_max_max = PID_MAX_LIMIT;
  54
  55 static inline int mk_pid(struct pid_namespace *pid_ns,
  56                 struct pidmap *map, int off)
  57 {
  58         return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
  59 }
  60
  61 #define find_next_offset(map, off)                                      \
  62                 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
  63
  64 /*
  65  * PID-map pages start out as NULL, they get allocated upon
  66  * first use and are never deallocated. This way a low pid_max
  67  * value does not cause lots of bitmaps to be allocated, but
  68  * the scheme scales to up to 4 million PIDs, runtime.
  69  */
  70 struct pid_namespace init_pid_ns = {
  71         .kref = {
  72                 .refcount       = ATOMIC_INIT(2),
  73         },
  74         .pidmap = {
  75                 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
  76         },
  77         .last_pid = 0,
  78         .nr_hashed = PIDNS_HASH_ADDING,
  79         .level = 0,
  80         .child_reaper = &init_task,
  81         .user_ns = &init_user_ns,
  82         .proc_inum = PROC_PID_INIT_INO,
  83 };
  84 EXPORT_SYMBOL_GPL(init_pid_ns);
  85
  86 /*
  87  * Note: disable interrupts while the pidmap_lock is held as an
  88  * interrupt might come in and do read_lock(&tasklist_lock).
  89  *
  90  * If we don't disable interrupts there is a nasty deadlock between
  91  * detach_pid()->free_pid() and another cpu that does
  92  * spin_lock(&pidmap_lock) followed by an interrupt routine that does
  93  * read_lock(&tasklist_lock);
  94  *
  95  * After we clean up the tasklist_lock and know there are no
  96  * irq handlers that take it we can leave the interrupts enabled.
  97  * For now it is easier to be safe than to prove it can't happen.
  98  */
  99
 100 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
 101
 102 static void free_pidmap(struct upid *upid)
 103 {
 104         int nr = upid->nr;
 105         struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
 106         int offset = nr & BITS_PER_PAGE_MASK;
 107
 108         clear_bit(offset, map->page);
 109         atomic_inc(&map->nr_free);
 110 }
 111
 112 /*
 113  * If we started walking pids at 'base', is 'a' seen before 'b'?
 114  */
 115 static int pid_before(int base, int a, int b)
 116 {
 117         /*
 118          * This is the same as saying
 119          *
 120          * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
 121          * and that mapping orders 'a' and 'b' with respect to 'base'.
 122          */
 123         return (unsigned)(a - base) < (unsigned)(b - base);
 124 }
 125
 126 /*
 127  * We might be racing with someone else trying to set pid_ns->last_pid
 128  * at the pid allocation time (there's also a sysctl for this, but racing
 129  * with this one is OK, see comment in kernel/pid_namespace.c about it).
 130  * We want the winner to have the "later" value, because if the
 131  * "earlier" value prevails, then a pid may get reused immediately.
 132  *
 133  * Since pids rollover, it is not sufficient to just pick the bigger
 134  * value.  We have to consider where we started counting from.
 135  *
 136  * 'base' is the value of pid_ns->last_pid that we observed when
 137  * we started looking for a pid.
 138  *
 139  * 'pid' is the pid that we eventually found.
 140  */
 141 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
 142 {
 143         int prev;
 144         int last_write = base;
 145         do {
 146                 prev = last_write;
 147                 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
 148         } while ((prev != last_write) && (pid_before(base, last_write, pid)));
 149 }
 150
 151 static int alloc_pidmap(struct pid_namespace *pid_ns)
 152 {
 153         int i, offset, max_scan, pid, last = pid_ns->last_pid;
 154         struct pidmap *map;
 155
 156         pid = last + 1;
 157         if (pid >= pid_max)
 158                 pid = RESERVED_PIDS;
 159         offset = pid & BITS_PER_PAGE_MASK;
 160         map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
 161         /*
 162          * If last_pid points into the middle of the map->page we
 163          * want to scan this bitmap block twice, the second time
 164          * we start with offset == 0 (or RESERVED_PIDS).
 165          */
 166         max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
 167         for (i = 0; i <= max_scan; ++i) {
 168                 if (unlikely(!map->page)) {
 169                         void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 170                         /*
 171                          * Free the page if someone raced with us
 172                          * installing it:
 173                          */
 174                         spin_lock_irq(&pidmap_lock);
 175                         if (!map->page) {
 176                                 map->page = page;
 177                                 page = NULL;
 178                         }
 179                         spin_unlock_irq(&pidmap_lock);
 180                         kfree(page);
 181                         if (unlikely(!map->page))
 182                                 break;
 183                 }
 184                 if (likely(atomic_read(&map->nr_free))) {
 185                         for ( ; ; ) {
 186                                 if (!test_and_set_bit(offset, map->page)) {
 187                                         atomic_dec(&map->nr_free);
 188                                         set_last_pid(pid_ns, last, pid);
 189                                         return pid;
 190                                 }
 191                                 offset = find_next_offset(map, offset);
 192                                 if (offset >= BITS_PER_PAGE)
 193                                         break;
 194                                 pid = mk_pid(pid_ns, map, offset);
 195                                 if (pid >= pid_max)
 196                                         break;
 197                         }
 198                 }
 199                 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
 200                         ++map;
 201                         offset = 0;
 202                 } else {
 203                         map = &pid_ns->pidmap[0];
 204                         offset = RESERVED_PIDS;
 205                         if (unlikely(last == offset))
 206                                 break;
 207                 }
 208                 pid = mk_pid(pid_ns, map, offset);
 209         }
 210         return -1;
 211 }
 212
 213 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
 214 {
 215         int offset;
 216         struct pidmap *map, *end;
 217
 218         if (last >= PID_MAX_LIMIT)
 219                 return -1;
 220
 221         offset = (last + 1) & BITS_PER_PAGE_MASK;
 222         map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
 223         end = &pid_ns->pidmap[PIDMAP_ENTRIES];
 224         for (; map < end; map++, offset = 0) {
 225                 if (unlikely(!map->page))
 226                         continue;
 227                 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
 228                 if (offset < BITS_PER_PAGE)
 229                         return mk_pid(pid_ns, map, offset);
 230         }
 231         return -1;
 232 }
 233
 234 void put_pid(struct pid *pid)
 235 {
 236         struct pid_namespace *ns;
 237
 238         if (!pid)
 239                 return;
 240
 241         ns = pid->numbers[pid->level].ns;
 242         if ((atomic_read(&pid->count) == 1) ||
 243              atomic_dec_and_test(&pid->count)) {
 244                 kmem_cache_free(ns->pid_cachep, pid);
 245                 put_pid_ns(ns);
 246         }
 247 }
 248 EXPORT_SYMBOL_GPL(put_pid);
 249
 250 static void delayed_put_pid(struct rcu_head *rhp)
 251 {
 252         struct pid *pid = container_of(rhp, struct pid, rcu);
 253         put_pid(pid);
 254 }
 255
 256 void free_pid(struct pid *pid)
 257 {
 258         /* We can be called with write_lock_irq(&tasklist_lock) held */
 259         int i;
 260         unsigned long flags;
 261
 262         spin_lock_irqsave(&pidmap_lock, flags);
 263         for (i = 0; i <= pid->level; i++) {
 264                 struct upid *upid = pid->numbers + i;
 265                 struct pid_namespace *ns = upid->ns;
 266                 hlist_del_rcu(&upid->pid_chain);
 267                 switch(--ns->nr_hashed) {
 268                 case 2:
 269                 case 1:
 270                         /* When all that is left in the pid namespace
 271                          * is the reaper wake up the reaper.  The reaper
 272                          * may be sleeping in zap_pid_ns_processes().
 273                          */
 274                         wake_up_process(ns->child_reaper);
 275                         break;
 276                 case PIDNS_HASH_ADDING:
 277                         /* Handle a fork failure of the first process */
 278                         WARN_ON(ns->child_reaper);
 279                         ns->nr_hashed = 0;
 280                         /* fall through */
 281                 case 0:
 282                         schedule_work(&ns->proc_work);
 283                         break;
 284                 }
 285         }
 286         spin_unlock_irqrestore(&pidmap_lock, flags);
 287
 288         for (i = 0; i <= pid->level; i++)
 289                 free_pidmap(pid->numbers + i);
 290
 291         call_rcu(&pid->rcu, delayed_put_pid);
 292 }
 293
 294 struct pid *alloc_pid(struct pid_namespace *ns)
 295 {
 296         struct pid *pid;
 297         enum pid_type type;
 298         int i, nr;
 299         struct pid_namespace *tmp;
 300         struct upid *upid;
 301
 302         pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
 303         if (!pid)
 304                 goto out;
 305
 306         tmp = ns;
 307         pid->level = ns->level;
 308         for (i = ns->level; i >= 0; i--) {
 309                 nr = alloc_pidmap(tmp);
 310                 if (nr < 0)
 311                         goto out_free;
 312
 313                 pid->numbers[i].nr = nr;
 314                 pid->numbers[i].ns = tmp;
 315                 tmp = tmp->parent;
 316         }
 317
 318         if (unlikely(is_child_reaper(pid))) {
 319                 if (pid_ns_prepare_proc(ns))
 320                         goto out_free;
 321         }
 322
 323         get_pid_ns(ns);
 324         atomic_set(&pid->count, 1);
 325         for (type = 0; type < PIDTYPE_MAX; ++type)
 326                 INIT_HLIST_HEAD(&pid->tasks[type]);
 327
 328         upid = pid->numbers + ns->level;
 329         spin_lock_irq(&pidmap_lock);
 330         if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
 331                 goto out_unlock;
 332         for ( ; upid >= pid->numbers; --upid) {
 333                 hlist_add_head_rcu(&upid->pid_chain,
 334                                 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
 335                 upid->ns->nr_hashed++;
 336         }
 337         spin_unlock_irq(&pidmap_lock);
 338
 339 out:
 340         return pid;
 341
 342 out_unlock:
 343         spin_unlock_irq(&pidmap_lock);
 344         put_pid_ns(ns);
 345
 346 out_free:
 347         while (++i <= ns->level)
 348                 free_pidmap(pid->numbers + i);
 349
 350         kmem_cache_free(ns->pid_cachep, pid);
 351         pid = NULL;
 352         goto out;
 353 }
 354
 355 void disable_pid_allocation(struct pid_namespace *ns)
 356 {
 357         spin_lock_irq(&pidmap_lock);
 358         ns->nr_hashed &= ~PIDNS_HASH_ADDING;
 359         spin_unlock_irq(&pidmap_lock);
 360 }
 361
 362 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
 363 {
 364         struct upid *pnr;
 365
 366         hlist_for_each_entry_rcu(pnr,
 367                         &pid_hash[pid_hashfn(nr, ns)], pid_chain)
 368                 if (pnr->nr == nr && pnr->ns == ns)
 369                         return container_of(pnr, struct pid,
 370                                         numbers[ns->level]);
 371
 372         return NULL;
 373 }
 374 EXPORT_SYMBOL_GPL(find_pid_ns);
 375
 376 struct pid *find_vpid(int nr)
 377 {
 378         return find_pid_ns(nr, task_active_pid_ns(current));
 379 }
 380 EXPORT_SYMBOL_GPL(find_vpid);
 381
 382 /*
 383  * attach_pid() must be called with the tasklist_lock write-held.
 384  */
 385 void attach_pid(struct task_struct *task, enum pid_type type)
 386 {
 387         struct pid_link *link = &task->pids[type];
 388         hlist_add_head_rcu(&link->node, &link->pid->tasks[type]);
 389 }
 390
 391 static void __change_pid(struct task_struct *task, enum pid_type type,
 392                         struct pid *new)
 393 {
 394         struct pid_link *link;
 395         struct pid *pid;
 396         int tmp;
 397
 398         link = &task->pids[type];
 399         pid = link->pid;
 400
 401         hlist_del_rcu(&link->node);
 402         link->pid = new;
 403
 404         for (tmp = PIDTYPE_MAX; --tmp >= 0; )
 405                 if (!hlist_empty(&pid->tasks[tmp]))
 406                         return;
 407
 408         free_pid(pid);
 409 }
 410
 411 void detach_pid(struct task_struct *task, enum pid_type type)
 412 {
 413         __change_pid(task, type, NULL);
 414 }
 415
 416 void change_pid(struct task_struct *task, enum pid_type type,
 417                 struct pid *pid)
 418 {
 419         __change_pid(task, type, pid);
 420         attach_pid(task, type);
 421 }
 422
 423 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
 424 void transfer_pid(struct task_struct *old, struct task_struct *new,
 425                            enum pid_type type)
 426 {
 427         new->pids[type].pid = old->pids[type].pid;
 428         hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
 429 }
 430
 431 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
 432 {
 433         struct task_struct *result = NULL;
 434         if (pid) {
 435                 struct hlist_node *first;
 436                 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
 437                                               lockdep_tasklist_lock_is_held());
 438                 if (first)
 439                         result = hlist_entry(first, struct task_struct, pids[(type)].node);
 440         }
 441         return result;
 442 }
 443 EXPORT_SYMBOL(pid_task);
 444
 445 /*
 446  * Must be called under rcu_read_lock().
 447  */
 448 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
 449 {
 450         rcu_lockdep_assert(rcu_read_lock_held(),
 451                            "find_task_by_pid_ns() needs rcu_read_lock()"
 452                            " protection");
 453         return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
 454 }
 455
 456 struct task_struct *find_task_by_vpid(pid_t vnr)
 457 {
 458         return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
 459 }
 460
 461 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
 462 {
 463         struct pid *pid;
 464         rcu_read_lock();
 465         if (type != PIDTYPE_PID)
 466                 task = task->group_leader;
 467         pid = get_pid(task->pids[type].pid);
 468         rcu_read_unlock();
 469         return pid;
 470 }
 471 EXPORT_SYMBOL_GPL(get_task_pid);
 472
 473 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
 474 {
 475         struct task_struct *result;
 476         rcu_read_lock();
 477         result = pid_task(pid, type);
 478         if (result)
 479                 get_task_struct(result);
 480         rcu_read_unlock();
 481         return result;
 482 }
 483 EXPORT_SYMBOL_GPL(get_pid_task);
 484
 485 struct pid *find_get_pid(pid_t nr)
 486 {
 487         struct pid *pid;
 488
 489         rcu_read_lock();
 490         pid = get_pid(find_vpid(nr));
 491         rcu_read_unlock();
 492
 493         return pid;
 494 }
 495 EXPORT_SYMBOL_GPL(find_get_pid);
 496
 497 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
 498 {
 499         struct upid *upid;
 500         pid_t nr = 0;
 501
 502         if (pid && ns->level <= pid->level) {
 503                 upid = &pid->numbers[ns->level];
 504                 if (upid->ns == ns)
 505                         nr = upid->nr;
 506         }
 507         return nr;
 508 }
 509 EXPORT_SYMBOL_GPL(pid_nr_ns);
 510
 511 pid_t pid_vnr(struct pid *pid)
 512 {
 513         return pid_nr_ns(pid, task_active_pid_ns(current));
 514 }
 515 EXPORT_SYMBOL_GPL(pid_vnr);
 516
 517 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
 518                         struct pid_namespace *ns)
 519 {
 520         pid_t nr = 0;
 521
 522         rcu_read_lock();
 523         if (!ns)
 524                 ns = task_active_pid_ns(current);
 525         if (likely(pid_alive(task))) {
 526                 if (type != PIDTYPE_PID)
 527                         task = task->group_leader;
 528                 nr = pid_nr_ns(task->pids[type].pid, ns);
 529         }
 530         rcu_read_unlock();
 531
 532         return nr;
 533 }
 534 EXPORT_SYMBOL(__task_pid_nr_ns);
 535
 536 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
 537 {
 538         return pid_nr_ns(task_tgid(tsk), ns);
 539 }
 540 EXPORT_SYMBOL(task_tgid_nr_ns);
 541
 542 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
 543 {
 544         return ns_of_pid(task_pid(tsk));
 545 }
 546 EXPORT_SYMBOL_GPL(task_active_pid_ns);
 547
 548 /*
 549  * Used by proc to find the first pid that is greater than or equal to nr.
 550  *
 551  * If there is a pid at nr this function is exactly the same as find_pid_ns.
 552  */
 553 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
 554 {
 555         struct pid *pid;
 556
 557         do {
 558                 pid = find_pid_ns(nr, ns);
 559                 if (pid)
 560                         break;
 561                 nr = next_pidmap(ns, nr);
 562         } while (nr > 0);
 563
 564         return pid;
 565 }
 566
 567 /*
 568  * The pid hash table is scaled according to the amount of memory in the
 569  * machine.  From a minimum of 16 slots up to 4096 slots at one gigabyte or
 570  * more.
 571  */
 572 void __init pidhash_init(void)
 573 {
 574         unsigned int i, pidhash_size;
 575
 576         pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
 577                                            HASH_EARLY | HASH_SMALL,
 578                                            &pidhash_shift, NULL,
 579                                            0, 4096);
 580         pidhash_size = 1U << pidhash_shift;
 581
 582         for (i = 0; i < pidhash_size; i++)
 583                 INIT_HLIST_HEAD(&pid_hash[i]);
 584 }
 585
 586 void __init pidmap_init(void)
 587 {
 588         /* Veryify no one has done anything silly */
 589         BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_HASH_ADDING);
 590
 591         /* bump default and minimum pid_max based on number of cpus */
 592         pid_max = min(pid_max_max, max_t(int, pid_max,
 593                                 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
 594         pid_max_min = max_t(int, pid_max_min,
 595                                 PIDS_PER_CPU_MIN * num_possible_cpus());
 596         pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
 597
 598         init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 599         /* Reserve PID 0. We never call free_pidmap(0) */
 600         set_bit(0, init_pid_ns.pidmap[0].page);
 601         atomic_dec(&init_pid_ns.pidmap[0].nr_free);
 602
 603         init_pid_ns.pid_cachep = KMEM_CACHE(pid,
 604                         SLAB_HWCACHE_ALIGN | SLAB_PANIC);
 605 }