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
  23 #include <linux/mm.h>
  24 #include <linux/module.h>
  25 #include <linux/slab.h>
  26 #include <linux/init.h>
  27 #include <linux/bootmem.h>
  28 #include <linux/hash.h>
  29 #include <linux/pid_namespace.h>
  30 #include <linux/init_task.h>
  31
  32 #define pid_hashfn(nr, ns)      \
  33         hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
  34 static struct hlist_head *pid_hash;
  35 static int pidhash_shift;
  36 struct pid init_struct_pid = INIT_STRUCT_PID;
  37
  38 int pid_max = PID_MAX_DEFAULT;
  39
  40 #define RESERVED_PIDS           300
  41
  42 int pid_max_min = RESERVED_PIDS + 1;
  43 int pid_max_max = PID_MAX_LIMIT;
  44
  45 #define BITS_PER_PAGE           (PAGE_SIZE*8)
  46 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
  47
  48 static inline int mk_pid(struct pid_namespace *pid_ns,
  49                 struct pidmap *map, int off)
  50 {
  51         return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
  52 }
  53
  54 #define find_next_offset(map, off)                                      \
  55                 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
  56
  57 /*
  58  * PID-map pages start out as NULL, they get allocated upon
  59  * first use and are never deallocated. This way a low pid_max
  60  * value does not cause lots of bitmaps to be allocated, but
  61  * the scheme scales to up to 4 million PIDs, runtime.
  62  */
  63 struct pid_namespace init_pid_ns = {
  64         .kref = {
  65                 .refcount       = ATOMIC_INIT(2),
  66         },
  67         .pidmap = {
  68                 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
  69         },
  70         .last_pid = 0,
  71         .level = 0,
  72         .child_reaper = &init_task,
  73 };
  74 EXPORT_SYMBOL_GPL(init_pid_ns);
  75
  76 int is_container_init(struct task_struct *tsk)
  77 {
  78         int ret = 0;
  79         struct pid *pid;
  80
  81         rcu_read_lock();
  82         pid = task_pid(tsk);
  83         if (pid != NULL && pid->numbers[pid->level].nr == 1)
  84                 ret = 1;
  85         rcu_read_unlock();
  86
  87         return ret;
  88 }
  89 EXPORT_SYMBOL(is_container_init);
  90
  91 /*
  92  * Note: disable interrupts while the pidmap_lock is held as an
  93  * interrupt might come in and do read_lock(&tasklist_lock).
  94  *
  95  * If we don't disable interrupts there is a nasty deadlock between
  96  * detach_pid()->free_pid() and another cpu that does
  97  * spin_lock(&pidmap_lock) followed by an interrupt routine that does
  98  * read_lock(&tasklist_lock);
  99  *
 100  * After we clean up the tasklist_lock and know there are no
 101  * irq handlers that take it we can leave the interrupts enabled.
 102  * For now it is easier to be safe than to prove it can't happen.
 103  */
 104
 105 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
 106
 107 static fastcall void free_pidmap(struct pid_namespace *pid_ns, int pid)
 108 {
 109         struct pidmap *map = pid_ns->pidmap + pid / BITS_PER_PAGE;
 110         int offset = pid & BITS_PER_PAGE_MASK;
 111
 112         clear_bit(offset, map->page);
 113         atomic_inc(&map->nr_free);
 114 }
 115
 116 static int alloc_pidmap(struct pid_namespace *pid_ns)
 117 {
 118         int i, offset, max_scan, pid, last = pid_ns->last_pid;
 119         struct pidmap *map;
 120
 121         pid = last + 1;
 122         if (pid >= pid_max)
 123                 pid = RESERVED_PIDS;
 124         offset = pid & BITS_PER_PAGE_MASK;
 125         map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
 126         max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
 127         for (i = 0; i <= max_scan; ++i) {
 128                 if (unlikely(!map->page)) {
 129                         void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 130                         /*
 131                          * Free the page if someone raced with us
 132                          * installing it:
 133                          */
 134                         spin_lock_irq(&pidmap_lock);
 135                         if (map->page)
 136                                 kfree(page);
 137                         else
 138                                 map->page = page;
 139                         spin_unlock_irq(&pidmap_lock);
 140                         if (unlikely(!map->page))
 141                                 break;
 142                 }
 143                 if (likely(atomic_read(&map->nr_free))) {
 144                         do {
 145                                 if (!test_and_set_bit(offset, map->page)) {
 146                                         atomic_dec(&map->nr_free);
 147                                         pid_ns->last_pid = pid;
 148                                         return pid;
 149                                 }
 150                                 offset = find_next_offset(map, offset);
 151                                 pid = mk_pid(pid_ns, map, offset);
 152                         /*
 153                          * find_next_offset() found a bit, the pid from it
 154                          * is in-bounds, and if we fell back to the last
 155                          * bitmap block and the final block was the same
 156                          * as the starting point, pid is before last_pid.
 157                          */
 158                         } while (offset < BITS_PER_PAGE && pid < pid_max &&
 159                                         (i != max_scan || pid < last ||
 160                                             !((last+1) & BITS_PER_PAGE_MASK)));
 161                 }
 162                 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
 163                         ++map;
 164                         offset = 0;
 165                 } else {
 166                         map = &pid_ns->pidmap[0];
 167                         offset = RESERVED_PIDS;
 168                         if (unlikely(last == offset))
 169                                 break;
 170                 }
 171                 pid = mk_pid(pid_ns, map, offset);
 172         }
 173         return -1;
 174 }
 175
 176 static int next_pidmap(struct pid_namespace *pid_ns, int last)
 177 {
 178         int offset;
 179         struct pidmap *map, *end;
 180
 181         offset = (last + 1) & BITS_PER_PAGE_MASK;
 182         map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
 183         end = &pid_ns->pidmap[PIDMAP_ENTRIES];
 184         for (; map < end; map++, offset = 0) {
 185                 if (unlikely(!map->page))
 186                         continue;
 187                 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
 188                 if (offset < BITS_PER_PAGE)
 189                         return mk_pid(pid_ns, map, offset);
 190         }
 191         return -1;
 192 }
 193
 194 fastcall void put_pid(struct pid *pid)
 195 {
 196         struct pid_namespace *ns;
 197
 198         if (!pid)
 199                 return;
 200
 201         ns = pid->numbers[pid->level].ns;
 202         if ((atomic_read(&pid->count) == 1) ||
 203              atomic_dec_and_test(&pid->count)) {
 204                 kmem_cache_free(ns->pid_cachep, pid);
 205                 put_pid_ns(ns);
 206         }
 207 }
 208 EXPORT_SYMBOL_GPL(put_pid);
 209
 210 static void delayed_put_pid(struct rcu_head *rhp)
 211 {
 212         struct pid *pid = container_of(rhp, struct pid, rcu);
 213         put_pid(pid);
 214 }
 215
 216 fastcall void free_pid(struct pid *pid)
 217 {
 218         /* We can be called with write_lock_irq(&tasklist_lock) held */
 219         int i;
 220         unsigned long flags;
 221
 222         spin_lock_irqsave(&pidmap_lock, flags);
 223         for (i = 0; i <= pid->level; i++)
 224                 hlist_del_rcu(&pid->numbers[i].pid_chain);
 225         spin_unlock_irqrestore(&pidmap_lock, flags);
 226
 227         for (i = 0; i <= pid->level; i++)
 228                 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
 229
 230         call_rcu(&pid->rcu, delayed_put_pid);
 231 }
 232
 233 struct pid *alloc_pid(struct pid_namespace *ns)
 234 {
 235         struct pid *pid;
 236         enum pid_type type;
 237         int i, nr;
 238         struct pid_namespace *tmp;
 239         struct upid *upid;
 240
 241         pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
 242         if (!pid)
 243                 goto out;
 244
 245         tmp = ns;
 246         for (i = ns->level; i >= 0; i--) {
 247                 nr = alloc_pidmap(tmp);
 248                 if (nr < 0)
 249                         goto out_free;
 250
 251                 pid->numbers[i].nr = nr;
 252                 pid->numbers[i].ns = tmp;
 253                 tmp = tmp->parent;
 254         }
 255
 256         get_pid_ns(ns);
 257         pid->level = ns->level;
 258         pid->nr = pid->numbers[0].nr;
 259         atomic_set(&pid->count, 1);
 260         for (type = 0; type < PIDTYPE_MAX; ++type)
 261                 INIT_HLIST_HEAD(&pid->tasks[type]);
 262
 263         spin_lock_irq(&pidmap_lock);
 264         for (i = ns->level; i >= 0; i--) {
 265                 upid = &pid->numbers[i];
 266                 hlist_add_head_rcu(&upid->pid_chain,
 267                                 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
 268         }
 269         spin_unlock_irq(&pidmap_lock);
 270
 271 out:
 272         return pid;
 273
 274 out_free:
 275         for (i++; i <= ns->level; i++)
 276                 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
 277
 278         kmem_cache_free(ns->pid_cachep, pid);
 279         pid = NULL;
 280         goto out;
 281 }
 282
 283 struct pid * fastcall find_pid_ns(int nr, struct pid_namespace *ns)
 284 {
 285         struct hlist_node *elem;
 286         struct upid *pnr;
 287
 288         hlist_for_each_entry_rcu(pnr, elem,
 289                         &pid_hash[pid_hashfn(nr, ns)], pid_chain)
 290                 if (pnr->nr == nr && pnr->ns == ns)
 291                         return container_of(pnr, struct pid,
 292                                         numbers[ns->level]);
 293
 294         return NULL;
 295 }
 296 EXPORT_SYMBOL_GPL(find_pid_ns);
 297
 298 /*
 299  * attach_pid() must be called with the tasklist_lock write-held.
 300  */
 301 int fastcall attach_pid(struct task_struct *task, enum pid_type type,
 302                 struct pid *pid)
 303 {
 304         struct pid_link *link;
 305
 306         link = &task->pids[type];
 307         link->pid = pid;
 308         hlist_add_head_rcu(&link->node, &pid->tasks[type]);
 309
 310         return 0;
 311 }
 312
 313 void fastcall detach_pid(struct task_struct *task, enum pid_type type)
 314 {
 315         struct pid_link *link;
 316         struct pid *pid;
 317         int tmp;
 318
 319         link = &task->pids[type];
 320         pid = link->pid;
 321
 322         hlist_del_rcu(&link->node);
 323         link->pid = NULL;
 324
 325         for (tmp = PIDTYPE_MAX; --tmp >= 0; )
 326                 if (!hlist_empty(&pid->tasks[tmp]))
 327                         return;
 328
 329         free_pid(pid);
 330 }
 331
 332 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
 333 void fastcall transfer_pid(struct task_struct *old, struct task_struct *new,
 334                            enum pid_type type)
 335 {
 336         new->pids[type].pid = old->pids[type].pid;
 337         hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
 338         old->pids[type].pid = NULL;
 339 }
 340
 341 struct task_struct * fastcall pid_task(struct pid *pid, enum pid_type type)
 342 {
 343         struct task_struct *result = NULL;
 344         if (pid) {
 345                 struct hlist_node *first;
 346                 first = rcu_dereference(pid->tasks[type].first);
 347                 if (first)
 348                         result = hlist_entry(first, struct task_struct, pids[(type)].node);
 349         }
 350         return result;
 351 }
 352
 353 /*
 354  * Must be called under rcu_read_lock() or with tasklist_lock read-held.
 355  */
 356 struct task_struct *find_task_by_pid_type_ns(int type, int nr,
 357                 struct pid_namespace *ns)
 358 {
 359         return pid_task(find_pid_ns(nr, ns), type);
 360 }
 361
 362 EXPORT_SYMBOL(find_task_by_pid_type_ns);
 363
 364 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
 365 {
 366         struct pid *pid;
 367         rcu_read_lock();
 368         pid = get_pid(task->pids[type].pid);
 369         rcu_read_unlock();
 370         return pid;
 371 }
 372
 373 struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
 374 {
 375         struct task_struct *result;
 376         rcu_read_lock();
 377         result = pid_task(pid, type);
 378         if (result)
 379                 get_task_struct(result);
 380         rcu_read_unlock();
 381         return result;
 382 }
 383
 384 struct pid *find_get_pid(pid_t nr)
 385 {
 386         struct pid *pid;
 387
 388         rcu_read_lock();
 389         pid = get_pid(find_vpid(nr));
 390         rcu_read_unlock();
 391
 392         return pid;
 393 }
 394
 395 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
 396 {
 397         struct upid *upid;
 398         pid_t nr = 0;
 399
 400         if (pid && ns->level <= pid->level) {
 401                 upid = &pid->numbers[ns->level];
 402                 if (upid->ns == ns)
 403                         nr = upid->nr;
 404         }
 405         return nr;
 406 }
 407
 408 /*
 409  * Used by proc to find the first pid that is greater then or equal to nr.
 410  *
 411  * If there is a pid at nr this function is exactly the same as find_pid.
 412  */
 413 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
 414 {
 415         struct pid *pid;
 416
 417         do {
 418                 pid = find_pid_ns(nr, ns);
 419                 if (pid)
 420                         break;
 421                 nr = next_pidmap(ns, nr);
 422         } while (nr > 0);
 423
 424         return pid;
 425 }
 426 EXPORT_SYMBOL_GPL(find_get_pid);
 427
 428 struct pid_cache {
 429         int nr_ids;
 430         char name[16];
 431         struct kmem_cache *cachep;
 432         struct list_head list;
 433 };
 434
 435 static LIST_HEAD(pid_caches_lh);
 436 static DEFINE_MUTEX(pid_caches_mutex);
 437
 438 /*
 439  * creates the kmem cache to allocate pids from.
 440  * @nr_ids: the number of numerical ids this pid will have to carry
 441  */
 442
 443 static struct kmem_cache *create_pid_cachep(int nr_ids)
 444 {
 445         struct pid_cache *pcache;
 446         struct kmem_cache *cachep;
 447
 448         mutex_lock(&pid_caches_mutex);
 449         list_for_each_entry (pcache, &pid_caches_lh, list)
 450                 if (pcache->nr_ids == nr_ids)
 451                         goto out;
 452
 453         pcache = kmalloc(sizeof(struct pid_cache), GFP_KERNEL);
 454         if (pcache == NULL)
 455                 goto err_alloc;
 456
 457         snprintf(pcache->name, sizeof(pcache->name), "pid_%d", nr_ids);
 458         cachep = kmem_cache_create(pcache->name,
 459                         /* FIXME add numerical ids here */
 460                         sizeof(struct pid), 0, SLAB_HWCACHE_ALIGN, NULL);
 461         if (cachep == NULL)
 462                 goto err_cachep;
 463
 464         pcache->nr_ids = nr_ids;
 465         pcache->cachep = cachep;
 466         list_add(&pcache->list, &pid_caches_lh);
 467 out:
 468         mutex_unlock(&pid_caches_mutex);
 469         return pcache->cachep;
 470
 471 err_cachep:
 472         kfree(pcache);
 473 err_alloc:
 474         mutex_unlock(&pid_caches_mutex);
 475         return NULL;
 476 }
 477
 478 struct pid_namespace *copy_pid_ns(unsigned long flags, struct pid_namespace *old_ns)
 479 {
 480         BUG_ON(!old_ns);
 481         get_pid_ns(old_ns);
 482         return old_ns;
 483 }
 484
 485 void free_pid_ns(struct kref *kref)
 486 {
 487         struct pid_namespace *ns;
 488
 489         ns = container_of(kref, struct pid_namespace, kref);
 490         kfree(ns);
 491 }
 492
 493 /*
 494  * The pid hash table is scaled according to the amount of memory in the
 495  * machine.  From a minimum of 16 slots up to 4096 slots at one gigabyte or
 496  * more.
 497  */
 498 void __init pidhash_init(void)
 499 {
 500         int i, pidhash_size;
 501         unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
 502
 503         pidhash_shift = max(4, fls(megabytes * 4));
 504         pidhash_shift = min(12, pidhash_shift);
 505         pidhash_size = 1 << pidhash_shift;
 506
 507         printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
 508                 pidhash_size, pidhash_shift,
 509                 pidhash_size * sizeof(struct hlist_head));
 510
 511         pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash)));
 512         if (!pid_hash)
 513                 panic("Could not alloc pidhash!\n");
 514         for (i = 0; i < pidhash_size; i++)
 515                 INIT_HLIST_HEAD(&pid_hash[i]);
 516 }
 517
 518 void __init pidmap_init(void)
 519 {
 520         init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
 521         /* Reserve PID 0. We never call free_pidmap(0) */
 522         set_bit(0, init_pid_ns.pidmap[0].page);
 523         atomic_dec(&init_pid_ns.pidmap[0].nr_free);
 524
 525         init_pid_ns.pid_cachep = create_pid_cachep(1);
 526         if (init_pid_ns.pid_cachep == NULL)
 527                 panic("Can't create pid_1 cachep\n");
 528 }