linux/kernel/pid.c
<<
>>
Prefs
   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_fs.h>
  40
  41#define pid_hashfn(nr, ns)      \
  42        hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
  43static struct hlist_head *pid_hash;
  44static unsigned int pidhash_shift = 4;
  45struct pid init_struct_pid = INIT_STRUCT_PID;
  46
  47int pid_max = PID_MAX_DEFAULT;
  48
  49#define RESERVED_PIDS           300
  50
  51int pid_max_min = RESERVED_PIDS + 1;
  52int pid_max_max = PID_MAX_LIMIT;
  53
  54#define BITS_PER_PAGE           (PAGE_SIZE*8)
  55#define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  56
  57static inline int mk_pid(struct pid_namespace *pid_ns,
  58                struct pidmap *map, int off)
  59{
  60        return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
  61}
  62
  63#define find_next_offset(map, off)                                      \
  64                find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
  65
  66/*
  67 * PID-map pages start out as NULL, they get allocated upon
  68 * first use and are never deallocated. This way a low pid_max
  69 * value does not cause lots of bitmaps to be allocated, but
  70 * the scheme scales to up to 4 million PIDs, runtime.
  71 */
  72struct pid_namespace init_pid_ns = {
  73        .kref = {
  74                .refcount       = ATOMIC_INIT(2),
  75        },
  76        .pidmap = {
  77                [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
  78        },
  79        .last_pid = 0,
  80        .level = 0,
  81        .child_reaper = &init_task,
  82        .user_ns = &init_user_ns,
  83        .proc_inum = PROC_PID_INIT_INO,
  84};
  85EXPORT_SYMBOL_GPL(init_pid_ns);
  86
  87/*
  88 * Note: disable interrupts while the pidmap_lock is held as an
  89 * interrupt might come in and do read_lock(&tasklist_lock).
  90 *
  91 * If we don't disable interrupts there is a nasty deadlock between
  92 * detach_pid()->free_pid() and another cpu that does
  93 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
  94 * read_lock(&tasklist_lock);
  95 *
  96 * After we clean up the tasklist_lock and know there are no
  97 * irq handlers that take it we can leave the interrupts enabled.
  98 * For now it is easier to be safe than to prove it can't happen.
  99 */
 100
 101static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
 102
 103static void free_pidmap(struct upid *upid)
 104{
 105        int nr = upid->nr;
 106        struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
 107        int offset = nr & BITS_PER_PAGE_MASK;
 108
 109        clear_bit(offset, map->page);
 110        atomic_inc(&map->nr_free);
 111}
 112
 113/*
 114 * If we started walking pids at 'base', is 'a' seen before 'b'?
 115 */
 116static int pid_before(int base, int a, int b)
 117{
 118        /*
 119         * This is the same as saying
 120         *
 121         * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
 122         * and that mapping orders 'a' and 'b' with respect to 'base'.
 123         */
 124        return (unsigned)(a - base) < (unsigned)(b - base);
 125}
 126
 127/*
 128 * We might be racing with someone else trying to set pid_ns->last_pid
 129 * at the pid allocation time (there's also a sysctl for this, but racing
 130 * with this one is OK, see comment in kernel/pid_namespace.c about it).
 131 * We want the winner to have the "later" value, because if the
 132 * "earlier" value prevails, then a pid may get reused immediately.
 133 *
 134 * Since pids rollover, it is not sufficient to just pick the bigger
 135 * value.  We have to consider where we started counting from.
 136 *
 137 * 'base' is the value of pid_ns->last_pid that we observed when
 138 * we started looking for a pid.
 139 *
 140 * 'pid' is the pid that we eventually found.
 141 */
 142static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
 143{
 144        int prev;
 145        int last_write = base;
 146        do {
 147                prev = last_write;
 148                last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
 149        } while ((prev != last_write) && (pid_before(base, last_write, pid)));
 150}
 151
 152static int alloc_pidmap(struct pid_namespace *pid_ns)
 153{
 154        int i, offset, max_scan, pid, last = pid_ns->last_pid;
 155        struct pidmap *map;
 156
 157        pid = last + 1;
 158        if (pid >= pid_max)
 159                pid = RESERVED_PIDS;
 160        offset = pid & BITS_PER_PAGE_MASK;
 161        map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
 162        /*
 163         * If last_pid points into the middle of the map->page we
 164         * want to scan this bitmap block twice, the second time
 165         * we start with offset == 0 (or RESERVED_PIDS).
 166         */
 167        max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
 168        for (i = 0; i <= max_scan; ++i) {
 169                if (unlikely(!map->page)) {
 170                        void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 171                        /*
 172                         * Free the page if someone raced with us
 173                         * installing it:
 174                         */
 175                        spin_lock_irq(&pidmap_lock);
 176                        if (!map->page) {
 177                                map->page = page;
 178                                page = NULL;
 179                        }
 180                        spin_unlock_irq(&pidmap_lock);
 181                        kfree(page);
 182                        if (unlikely(!map->page))
 183                                break;
 184                }
 185                if (likely(atomic_read(&map->nr_free))) {
 186                        do {
 187                                if (!test_and_set_bit(offset, map->page)) {
 188                                        atomic_dec(&map->nr_free);
 189                                        set_last_pid(pid_ns, last, pid);
 190                                        return pid;
 191                                }
 192                                offset = find_next_offset(map, offset);
 193                                pid = mk_pid(pid_ns, map, offset);
 194                        } while (offset < BITS_PER_PAGE && pid < pid_max);
 195                }
 196                if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
 197                        ++map;
 198                        offset = 0;
 199                } else {
 200                        map = &pid_ns->pidmap[0];
 201                        offset = RESERVED_PIDS;
 202                        if (unlikely(last == offset))
 203                                break;
 204                }
 205                pid = mk_pid(pid_ns, map, offset);
 206        }
 207        return -1;
 208}
 209
 210int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
 211{
 212        int offset;
 213        struct pidmap *map, *end;
 214
 215        if (last >= PID_MAX_LIMIT)
 216                return -1;
 217
 218        offset = (last + 1) & BITS_PER_PAGE_MASK;
 219        map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
 220        end = &pid_ns->pidmap[PIDMAP_ENTRIES];
 221        for (; map < end; map++, offset = 0) {
 222                if (unlikely(!map->page))
 223                        continue;
 224                offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
 225                if (offset < BITS_PER_PAGE)
 226                        return mk_pid(pid_ns, map, offset);
 227        }
 228        return -1;
 229}
 230
 231void put_pid(struct pid *pid)
 232{
 233        struct pid_namespace *ns;
 234
 235        if (!pid)
 236                return;
 237
 238        ns = pid->numbers[pid->level].ns;
 239        if ((atomic_read(&pid->count) == 1) ||
 240             atomic_dec_and_test(&pid->count)) {
 241                kmem_cache_free(ns->pid_cachep, pid);
 242                put_pid_ns(ns);
 243        }
 244}
 245EXPORT_SYMBOL_GPL(put_pid);
 246
 247static void delayed_put_pid(struct rcu_head *rhp)
 248{
 249        struct pid *pid = container_of(rhp, struct pid, rcu);
 250        put_pid(pid);
 251}
 252
 253void free_pid(struct pid *pid)
 254{
 255        /* We can be called with write_lock_irq(&tasklist_lock) held */
 256        int i;
 257        unsigned long flags;
 258
 259        spin_lock_irqsave(&pidmap_lock, flags);
 260        for (i = 0; i <= pid->level; i++) {
 261                struct upid *upid = pid->numbers + i;
 262                struct pid_namespace *ns = upid->ns;
 263                hlist_del_rcu(&upid->pid_chain);
 264                switch(--ns->nr_hashed) {
 265                case 1:
 266                        /* When all that is left in the pid namespace
 267                         * is the reaper wake up the reaper.  The reaper
 268                         * may be sleeping in zap_pid_ns_processes().
 269                         */
 270                        wake_up_process(ns->child_reaper);
 271                        break;
 272                case 0:
 273                        schedule_work(&ns->proc_work);
 274                        break;
 275                }
 276        }
 277        spin_unlock_irqrestore(&pidmap_lock, flags);
 278
 279        for (i = 0; i <= pid->level; i++)
 280                free_pidmap(pid->numbers + i);
 281
 282        call_rcu(&pid->rcu, delayed_put_pid);
 283}
 284
 285struct pid *alloc_pid(struct pid_namespace *ns)
 286{
 287        struct pid *pid;
 288        enum pid_type type;
 289        int i, nr;
 290        struct pid_namespace *tmp;
 291        struct upid *upid;
 292
 293        pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
 294        if (!pid)
 295                goto out;
 296
 297        tmp = ns;
 298        pid->level = ns->level;
 299        for (i = ns->level; i >= 0; i--) {
 300                nr = alloc_pidmap(tmp);
 301                if (nr < 0)
 302                        goto out_free;
 303
 304                pid->numbers[i].nr = nr;
 305                pid->numbers[i].ns = tmp;
 306                tmp = tmp->parent;
 307        }
 308
 309        if (unlikely(is_child_reaper(pid))) {
 310                if (pid_ns_prepare_proc(ns))
 311                        goto out_free;
 312        }
 313
 314        get_pid_ns(ns);
 315        atomic_set(&pid->count, 1);
 316        for (type = 0; type < PIDTYPE_MAX; ++type)
 317                INIT_HLIST_HEAD(&pid->tasks[type]);
 318
 319        upid = pid->numbers + ns->level;
 320        spin_lock_irq(&pidmap_lock);
 321        if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
 322                goto out_unlock;
 323        for ( ; upid >= pid->numbers; --upid) {
 324                hlist_add_head_rcu(&upid->pid_chain,
 325                                &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
 326                upid->ns->nr_hashed++;
 327        }
 328        spin_unlock_irq(&pidmap_lock);
 329
 330out:
 331        return pid;
 332
 333out_unlock:
 334        spin_unlock_irq(&pidmap_lock);
 335out_free:
 336        while (++i <= ns->level)
 337                free_pidmap(pid->numbers + i);
 338
 339        kmem_cache_free(ns->pid_cachep, pid);
 340        pid = NULL;
 341        goto out;
 342}
 343
 344void disable_pid_allocation(struct pid_namespace *ns)
 345{
 346        spin_lock_irq(&pidmap_lock);
 347        ns->nr_hashed &= ~PIDNS_HASH_ADDING;
 348        spin_unlock_irq(&pidmap_lock);
 349}
 350
 351struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
 352{
 353        struct upid *pnr;
 354
 355        hlist_for_each_entry_rcu(pnr,
 356                        &pid_hash[pid_hashfn(nr, ns)], pid_chain)
 357                if (pnr->nr == nr && pnr->ns == ns)
 358                        return container_of(pnr, struct pid,
 359                                        numbers[ns->level]);
 360
 361        return NULL;
 362}
 363EXPORT_SYMBOL_GPL(find_pid_ns);
 364
 365struct pid *find_vpid(int nr)
 366{
 367        return find_pid_ns(nr, task_active_pid_ns(current));
 368}
 369EXPORT_SYMBOL_GPL(find_vpid);
 370
 371/*
 372 * attach_pid() must be called with the tasklist_lock write-held.
 373 */
 374void attach_pid(struct task_struct *task, enum pid_type type,
 375                struct pid *pid)
 376{
 377        struct pid_link *link;
 378
 379        link = &task->pids[type];
 380        link->pid = pid;
 381        hlist_add_head_rcu(&link->node, &pid->tasks[type]);
 382}
 383
 384static void __change_pid(struct task_struct *task, enum pid_type type,
 385                        struct pid *new)
 386{
 387        struct pid_link *link;
 388        struct pid *pid;
 389        int tmp;
 390
 391        link = &task->pids[type];
 392        pid = link->pid;
 393
 394        hlist_del_rcu(&link->node);
 395        link->pid = new;
 396
 397        for (tmp = PIDTYPE_MAX; --tmp >= 0; )
 398                if (!hlist_empty(&pid->tasks[tmp]))
 399                        return;
 400
 401        free_pid(pid);
 402}
 403
 404void detach_pid(struct task_struct *task, enum pid_type type)
 405{
 406        __change_pid(task, type, NULL);
 407}
 408
 409void change_pid(struct task_struct *task, enum pid_type type,
 410                struct pid *pid)
 411{
 412        __change_pid(task, type, pid);
 413        attach_pid(task, type, pid);
 414}
 415
 416/* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
 417void transfer_pid(struct task_struct *old, struct task_struct *new,
 418                           enum pid_type type)
 419{
 420        new->pids[type].pid = old->pids[type].pid;
 421        hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
 422}
 423
 424struct task_struct *pid_task(struct pid *pid, enum pid_type type)
 425{
 426        struct task_struct *result = NULL;
 427        if (pid) {
 428                struct hlist_node *first;
 429                first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
 430                                              lockdep_tasklist_lock_is_held());
 431                if (first)
 432                        result = hlist_entry(first, struct task_struct, pids[(type)].node);
 433        }
 434        return result;
 435}
 436EXPORT_SYMBOL(pid_task);
 437
 438/*
 439 * Must be called under rcu_read_lock().
 440 */
 441struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
 442{
 443        rcu_lockdep_assert(rcu_read_lock_held(),
 444                           "find_task_by_pid_ns() needs rcu_read_lock()"
 445                           " protection");
 446        return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
 447}
 448
 449struct task_struct *find_task_by_vpid(pid_t vnr)
 450{
 451        return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
 452}
 453
 454struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
 455{
 456        struct pid *pid;
 457        rcu_read_lock();
 458        if (type != PIDTYPE_PID)
 459                task = task->group_leader;
 460        pid = get_pid(task->pids[type].pid);
 461        rcu_read_unlock();
 462        return pid;
 463}
 464EXPORT_SYMBOL_GPL(get_task_pid);
 465
 466struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
 467{
 468        struct task_struct *result;
 469        rcu_read_lock();
 470        result = pid_task(pid, type);
 471        if (result)
 472                get_task_struct(result);
 473        rcu_read_unlock();
 474        return result;
 475}
 476EXPORT_SYMBOL_GPL(get_pid_task);
 477
 478struct pid *find_get_pid(pid_t nr)
 479{
 480        struct pid *pid;
 481
 482        rcu_read_lock();
 483        pid = get_pid(find_vpid(nr));
 484        rcu_read_unlock();
 485
 486        return pid;
 487}
 488EXPORT_SYMBOL_GPL(find_get_pid);
 489
 490pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
 491{
 492        struct upid *upid;
 493        pid_t nr = 0;
 494
 495        if (pid && ns->level <= pid->level) {
 496                upid = &pid->numbers[ns->level];
 497                if (upid->ns == ns)
 498                        nr = upid->nr;
 499        }
 500        return nr;
 501}
 502EXPORT_SYMBOL_GPL(pid_nr_ns);
 503
 504pid_t pid_vnr(struct pid *pid)
 505{
 506        return pid_nr_ns(pid, task_active_pid_ns(current));
 507}
 508EXPORT_SYMBOL_GPL(pid_vnr);
 509
 510pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
 511                        struct pid_namespace *ns)
 512{
 513        pid_t nr = 0;
 514
 515        rcu_read_lock();
 516        if (!ns)
 517                ns = task_active_pid_ns(current);
 518        if (likely(pid_alive(task))) {
 519                if (type != PIDTYPE_PID)
 520                        task = task->group_leader;
 521                nr = pid_nr_ns(task->pids[type].pid, ns);
 522        }
 523        rcu_read_unlock();
 524
 525        return nr;
 526}
 527EXPORT_SYMBOL(__task_pid_nr_ns);
 528
 529pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
 530{
 531        return pid_nr_ns(task_tgid(tsk), ns);
 532}
 533EXPORT_SYMBOL(task_tgid_nr_ns);
 534
 535struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
 536{
 537        return ns_of_pid(task_pid(tsk));
 538}
 539EXPORT_SYMBOL_GPL(task_active_pid_ns);
 540
 541/*
 542 * Used by proc to find the first pid that is greater than or equal to nr.
 543 *
 544 * If there is a pid at nr this function is exactly the same as find_pid_ns.
 545 */
 546struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
 547{
 548        struct pid *pid;
 549
 550        do {
 551                pid = find_pid_ns(nr, ns);
 552                if (pid)
 553                        break;
 554                nr = next_pidmap(ns, nr);
 555        } while (nr > 0);
 556
 557        return pid;
 558}
 559
 560/*
 561 * The pid hash table is scaled according to the amount of memory in the
 562 * machine.  From a minimum of 16 slots up to 4096 slots at one gigabyte or
 563 * more.
 564 */
 565void __init pidhash_init(void)
 566{
 567        unsigned int i, pidhash_size;
 568
 569        pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
 570                                           HASH_EARLY | HASH_SMALL,
 571                                           &pidhash_shift, NULL,
 572                                           0, 4096);
 573        pidhash_size = 1U << pidhash_shift;
 574
 575        for (i = 0; i < pidhash_size; i++)
 576                INIT_HLIST_HEAD(&pid_hash[i]);
 577}
 578
 579void __init pidmap_init(void)
 580{
 581        /* Veryify no one has done anything silly */
 582        BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_HASH_ADDING);
 583
 584        /* bump default and minimum pid_max based on number of cpus */
 585        pid_max = min(pid_max_max, max_t(int, pid_max,
 586                                PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
 587        pid_max_min = max_t(int, pid_max_min,
 588                                PIDS_PER_CPU_MIN * num_possible_cpus());
 589        pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
 590
 591        init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 592        /* Reserve PID 0. We never call free_pidmap(0) */
 593        set_bit(0, init_pid_ns.pidmap[0].page);
 594        atomic_dec(&init_pid_ns.pidmap[0].nr_free);
 595        init_pid_ns.nr_hashed = PIDNS_HASH_ADDING;
 596
 597        init_pid_ns.pid_cachep = KMEM_CACHE(pid,
 598                        SLAB_HWCACHE_ALIGN | SLAB_PANIC);
 599}
 600
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.