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