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/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 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 /* 126 * If we started walking pids at 'base', is 'a' seen before 'b'? 127 */ 128 static int pid_before(int base, int a, int b) 129 { 130 /* 131 * This is the same as saying 132 * 133 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT 134 * and that mapping orders 'a' and 'b' with respect to 'base'. 135 */ 136 return (unsigned)(a - base) < (unsigned)(b - base); 137 } 138 139 /* 140 * We might be racing with someone else trying to set pid_ns->last_pid. 141 * We want the winner to have the "later" value, because if the 142 * "earlier" value prevails, then a pid may get reused immediately. 143 * 144 * Since pids rollover, it is not sufficient to just pick the bigger 145 * value. We have to consider where we started counting from. 146 * 147 * 'base' is the value of pid_ns->last_pid that we observed when 148 * we started looking for a pid. 149 * 150 * 'pid' is the pid that we eventually found. 151 */ 152 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid) 153 { 154 int prev; 155 int last_write = base; 156 do { 157 prev = last_write; 158 last_write = cmpxchg(&pid_ns->last_pid, prev, pid); 159 } while ((prev != last_write) && (pid_before(base, last_write, pid))); 160 } 161 162 static int alloc_pidmap(struct pid_namespace *pid_ns) 163 { 164 int i, offset, max_scan, pid, last = pid_ns->last_pid; 165 struct pidmap *map; 166 167 pid = last + 1; 168 if (pid >= pid_max) 169 pid = RESERVED_PIDS; 170 offset = pid & BITS_PER_PAGE_MASK; 171 map = &pid_ns->pidmap[pid/BITS_PER_PAGE]; 172 /* 173 * If last_pid points into the middle of the map->page we 174 * want to scan this bitmap block twice, the second time 175 * we start with offset == 0 (or RESERVED_PIDS). 176 */ 177 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset; 178 for (i = 0; i <= max_scan; ++i) { 179 if (unlikely(!map->page)) { 180 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL); 181 /* 182 * Free the page if someone raced with us 183 * installing it: 184 */ 185 spin_lock_irq(&pidmap_lock); 186 if (!map->page) { 187 map->page = page; 188 page = NULL; 189 } 190 spin_unlock_irq(&pidmap_lock); 191 kfree(page); 192 if (unlikely(!map->page)) 193 break; 194 } 195 if (likely(atomic_read(&map->nr_free))) { 196 do { 197 if (!test_and_set_bit(offset, map->page)) { 198 atomic_dec(&map->nr_free); 199 set_last_pid(pid_ns, last, pid); 200 return pid; 201 } 202 offset = find_next_offset(map, offset); 203 pid = mk_pid(pid_ns, map, offset); 204 } while (offset < BITS_PER_PAGE && pid < pid_max); 205 } 206 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) { 207 ++map; 208 offset = 0; 209 } else { 210 map = &pid_ns->pidmap[0]; 211 offset = RESERVED_PIDS; 212 if (unlikely(last == offset)) 213 break; 214 } 215 pid = mk_pid(pid_ns, map, offset); 216 } 217 return -1; 218 } 219 220 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last) 221 { 222 int offset; 223 struct pidmap *map, *end; 224 225 if (last >= PID_MAX_LIMIT) 226 return -1; 227 228 offset = (last + 1) & BITS_PER_PAGE_MASK; 229 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE]; 230 end = &pid_ns->pidmap[PIDMAP_ENTRIES]; 231 for (; map < end; map++, offset = 0) { 232 if (unlikely(!map->page)) 233 continue; 234 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset); 235 if (offset < BITS_PER_PAGE) 236 return mk_pid(pid_ns, map, offset); 237 } 238 return -1; 239 } 240 241 void put_pid(struct pid *pid) 242 { 243 struct pid_namespace *ns; 244 245 if (!pid) 246 return; 247 248 ns = pid->numbers[pid->level].ns; 249 if ((atomic_read(&pid->count) == 1) || 250 atomic_dec_and_test(&pid->count)) { 251 kmem_cache_free(ns->pid_cachep, pid); 252 put_pid_ns(ns); 253 } 254 } 255 EXPORT_SYMBOL_GPL(put_pid); 256 257 static void delayed_put_pid(struct rcu_head *rhp) 258 { 259 struct pid *pid = container_of(rhp, struct pid, rcu); 260 put_pid(pid); 261 } 262 263 void free_pid(struct pid *pid) 264 { 265 /* We can be called with write_lock_irq(&tasklist_lock) held */ 266 int i; 267 unsigned long flags; 268 269 spin_lock_irqsave(&pidmap_lock, flags); 270 for (i = 0; i <= pid->level; i++) 271 hlist_del_rcu(&pid->numbers[i].pid_chain); 272 spin_unlock_irqrestore(&pidmap_lock, flags); 273 274 for (i = 0; i <= pid->level; i++) 275 free_pidmap(pid->numbers + i); 276 277 call_rcu(&pid->rcu, delayed_put_pid); 278 } 279 280 struct pid *alloc_pid(struct pid_namespace *ns) 281 { 282 struct pid *pid; 283 enum pid_type type; 284 int i, nr; 285 struct pid_namespace *tmp; 286 struct upid *upid; 287 288 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL); 289 if (!pid) 290 goto out; 291 292 tmp = ns; 293 for (i = ns->level; i >= 0; i--) { 294 nr = alloc_pidmap(tmp); 295 if (nr < 0) 296 goto out_free; 297 298 pid->numbers[i].nr = nr; 299 pid->numbers[i].ns = tmp; 300 tmp = tmp->parent; 301 } 302 303 get_pid_ns(ns); 304 pid->level = ns->level; 305 atomic_set(&pid->count, 1); 306 for (type = 0; type < PIDTYPE_MAX; ++type) 307 INIT_HLIST_HEAD(&pid->tasks[type]); 308 309 upid = pid->numbers + ns->level; 310 spin_lock_irq(&pidmap_lock); 311 for ( ; upid >= pid->numbers; --upid) 312 hlist_add_head_rcu(&upid->pid_chain, 313 &pid_hash[pid_hashfn(upid->nr, upid->ns)]); 314 spin_unlock_irq(&pidmap_lock); 315 316 out: 317 return pid; 318 319 out_free: 320 while (++i <= ns->level) 321 free_pidmap(pid->numbers + i); 322 323 kmem_cache_free(ns->pid_cachep, pid); 324 pid = NULL; 325 goto out; 326 } 327 328 struct pid *find_pid_ns(int nr, struct pid_namespace *ns) 329 { 330 struct hlist_node *elem; 331 struct upid *pnr; 332 333 hlist_for_each_entry_rcu(pnr, elem, 334 &pid_hash[pid_hashfn(nr, ns)], pid_chain) 335 if (pnr->nr == nr && pnr->ns == ns) 336 return container_of(pnr, struct pid, 337 numbers[ns->level]); 338 339 return NULL; 340 } 341 EXPORT_SYMBOL_GPL(find_pid_ns); 342 343 struct pid *find_vpid(int nr) 344 { 345 return find_pid_ns(nr, current->nsproxy->pid_ns); 346 } 347 EXPORT_SYMBOL_GPL(find_vpid); 348 349 /* 350 * attach_pid() must be called with the tasklist_lock write-held. 351 */ 352 void attach_pid(struct task_struct *task, enum pid_type type, 353 struct pid *pid) 354 { 355 struct pid_link *link; 356 357 link = &task->pids[type]; 358 link->pid = pid; 359 hlist_add_head_rcu(&link->node, &pid->tasks[type]); 360 } 361 362 static void __change_pid(struct task_struct *task, enum pid_type type, 363 struct pid *new) 364 { 365 struct pid_link *link; 366 struct pid *pid; 367 int tmp; 368 369 link = &task->pids[type]; 370 pid = link->pid; 371 372 hlist_del_rcu(&link->node); 373 link->pid = new; 374 375 for (tmp = PIDTYPE_MAX; --tmp >= 0; ) 376 if (!hlist_empty(&pid->tasks[tmp])) 377 return; 378 379 free_pid(pid); 380 } 381 382 void detach_pid(struct task_struct *task, enum pid_type type) 383 { 384 __change_pid(task, type, NULL); 385 } 386 387 void change_pid(struct task_struct *task, enum pid_type type, 388 struct pid *pid) 389 { 390 __change_pid(task, type, pid); 391 attach_pid(task, type, pid); 392 } 393 394 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */ 395 void transfer_pid(struct task_struct *old, struct task_struct *new, 396 enum pid_type type) 397 { 398 new->pids[type].pid = old->pids[type].pid; 399 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node); 400 } 401 402 struct task_struct *pid_task(struct pid *pid, enum pid_type type) 403 { 404 struct task_struct *result = NULL; 405 if (pid) { 406 struct hlist_node *first; 407 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]), 408 lockdep_tasklist_lock_is_held()); 409 if (first) 410 result = hlist_entry(first, struct task_struct, pids[(type)].node); 411 } 412 return result; 413 } 414 EXPORT_SYMBOL(pid_task); 415 416 /* 417 * Must be called under rcu_read_lock(). 418 */ 419 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns) 420 { 421 rcu_lockdep_assert(rcu_read_lock_held(), 422 "find_task_by_pid_ns() needs rcu_read_lock()" 423 " protection"); 424 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID); 425 } 426 427 struct task_struct *find_task_by_vpid(pid_t vnr) 428 { 429 return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns); 430 } 431 432 struct pid *get_task_pid(struct task_struct *task, enum pid_type type) 433 { 434 struct pid *pid; 435 rcu_read_lock(); 436 if (type != PIDTYPE_PID) 437 task = task->group_leader; 438 pid = get_pid(task->pids[type].pid); 439 rcu_read_unlock(); 440 return pid; 441 } 442 EXPORT_SYMBOL_GPL(get_task_pid); 443 444 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type) 445 { 446 struct task_struct *result; 447 rcu_read_lock(); 448 result = pid_task(pid, type); 449 if (result) 450 get_task_struct(result); 451 rcu_read_unlock(); 452 return result; 453 } 454 EXPORT_SYMBOL_GPL(get_pid_task); 455 456 struct pid *find_get_pid(pid_t nr) 457 { 458 struct pid *pid; 459 460 rcu_read_lock(); 461 pid = get_pid(find_vpid(nr)); 462 rcu_read_unlock(); 463 464 return pid; 465 } 466 EXPORT_SYMBOL_GPL(find_get_pid); 467 468 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns) 469 { 470 struct upid *upid; 471 pid_t nr = 0; 472 473 if (pid && ns->level <= pid->level) { 474 upid = &pid->numbers[ns->level]; 475 if (upid->ns == ns) 476 nr = upid->nr; 477 } 478 return nr; 479 } 480 481 pid_t pid_vnr(struct pid *pid) 482 { 483 return pid_nr_ns(pid, current->nsproxy->pid_ns); 484 } 485 EXPORT_SYMBOL_GPL(pid_vnr); 486 487 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type, 488 struct pid_namespace *ns) 489 { 490 pid_t nr = 0; 491 492 rcu_read_lock(); 493 if (!ns) 494 ns = current->nsproxy->pid_ns; 495 if (likely(pid_alive(task))) { 496 if (type != PIDTYPE_PID) 497 task = task->group_leader; 498 nr = pid_nr_ns(task->pids[type].pid, ns); 499 } 500 rcu_read_unlock(); 501 502 return nr; 503 } 504 EXPORT_SYMBOL(__task_pid_nr_ns); 505 506 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns) 507 { 508 return pid_nr_ns(task_tgid(tsk), ns); 509 } 510 EXPORT_SYMBOL(task_tgid_nr_ns); 511 512 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk) 513 { 514 return ns_of_pid(task_pid(tsk)); 515 } 516 EXPORT_SYMBOL_GPL(task_active_pid_ns); 517 518 /* 519 * Used by proc to find the first pid that is greater than or equal to nr. 520 * 521 * If there is a pid at nr this function is exactly the same as find_pid_ns. 522 */ 523 struct pid *find_ge_pid(int nr, struct pid_namespace *ns) 524 { 525 struct pid *pid; 526 527 do { 528 pid = find_pid_ns(nr, ns); 529 if (pid) 530 break; 531 nr = next_pidmap(ns, nr); 532 } while (nr > 0); 533 534 return pid; 535 } 536 537 /* 538 * The pid hash table is scaled according to the amount of memory in the 539 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or 540 * more. 541 */ 542 void __init pidhash_init(void) 543 { 544 int i, pidhash_size; 545 546 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18, 547 HASH_EARLY | HASH_SMALL, 548 &pidhash_shift, NULL, 4096); 549 pidhash_size = 1 << pidhash_shift; 550 551 for (i = 0; i < pidhash_size; i++) 552 INIT_HLIST_HEAD(&pid_hash[i]); 553 } 554 555 void __init pidmap_init(void) 556 { 557 /* bump default and minimum pid_max based on number of cpus */ 558 pid_max = min(pid_max_max, max_t(int, pid_max, 559 PIDS_PER_CPU_DEFAULT * num_possible_cpus())); 560 pid_max_min = max_t(int, pid_max_min, 561 PIDS_PER_CPU_MIN * num_possible_cpus()); 562 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min); 563 564 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL); 565 /* Reserve PID 0. We never call free_pidmap(0) */ 566 set_bit(0, init_pid_ns.pidmap[0].page); 567 atomic_dec(&init_pid_ns.pidmap[0].nr_free); 568 569 init_pid_ns.pid_cachep = KMEM_CACHE(pid, 570 SLAB_HWCACHE_ALIGN | SLAB_PANIC); 571 } 572