pids: fix a race in pid generation that causes pids to be reused immediately
[linux-block.git] / kernel / pid.c
CommitLineData
1da177e4
LT
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).
30e49c26
PE
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 *
1da177e4
LT
27 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/slab.h>
32#include <linux/init.h>
82524746 33#include <linux/rculist.h>
1da177e4
LT
34#include <linux/bootmem.h>
35#include <linux/hash.h>
61a58c6c 36#include <linux/pid_namespace.h>
820e45db 37#include <linux/init_task.h>
3eb07c8c 38#include <linux/syscalls.h>
1da177e4 39
8ef047aa
PE
40#define pid_hashfn(nr, ns) \
41 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
92476d7f 42static struct hlist_head *pid_hash;
2c85f51d 43static unsigned int pidhash_shift = 4;
820e45db 44struct pid init_struct_pid = INIT_STRUCT_PID;
1da177e4
LT
45
46int pid_max = PID_MAX_DEFAULT;
1da177e4
LT
47
48#define RESERVED_PIDS 300
49
50int pid_max_min = RESERVED_PIDS + 1;
51int pid_max_max = PID_MAX_LIMIT;
52
1da177e4
LT
53#define BITS_PER_PAGE (PAGE_SIZE*8)
54#define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
3fbc9648 55
61a58c6c
SB
56static inline int mk_pid(struct pid_namespace *pid_ns,
57 struct pidmap *map, int off)
3fbc9648 58{
61a58c6c 59 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
3fbc9648
SB
60}
61
1da177e4
LT
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 */
61a58c6c 71struct pid_namespace init_pid_ns = {
9a575a92
CLG
72 .kref = {
73 .refcount = ATOMIC_INIT(2),
74 },
3fbc9648
SB
75 .pidmap = {
76 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
77 },
84d73786 78 .last_pid = 0,
faacbfd3
PE
79 .level = 0,
80 .child_reaper = &init_task,
3fbc9648 81};
198fe21b 82EXPORT_SYMBOL_GPL(init_pid_ns);
1da177e4 83
b461cc03 84int is_container_init(struct task_struct *tsk)
b460cbc5 85{
b461cc03
PE
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;
b460cbc5 96}
b461cc03 97EXPORT_SYMBOL(is_container_init);
b460cbc5 98
92476d7f
EB
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 */
3fbc9648 112
1da177e4
LT
113static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
114
b7127aa4 115static void free_pidmap(struct upid *upid)
1da177e4 116{
b7127aa4
ON
117 int nr = upid->nr;
118 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
119 int offset = nr & BITS_PER_PAGE_MASK;
1da177e4
LT
120
121 clear_bit(offset, map->page);
122 atomic_inc(&map->nr_free);
123}
124
5fdee8c4
S
125/*
126 * If we started walking pids at 'base', is 'a' seen before 'b'?
127 */
128static 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 */
152static 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
61a58c6c 162static int alloc_pidmap(struct pid_namespace *pid_ns)
1da177e4 163{
61a58c6c 164 int i, offset, max_scan, pid, last = pid_ns->last_pid;
6a1f3b84 165 struct pidmap *map;
1da177e4
LT
166
167 pid = last + 1;
168 if (pid >= pid_max)
169 pid = RESERVED_PIDS;
170 offset = pid & BITS_PER_PAGE_MASK;
61a58c6c 171 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
1da177e4
LT
172 max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
173 for (i = 0; i <= max_scan; ++i) {
174 if (unlikely(!map->page)) {
3fbc9648 175 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
1da177e4
LT
176 /*
177 * Free the page if someone raced with us
178 * installing it:
179 */
92476d7f 180 spin_lock_irq(&pidmap_lock);
7be6d991 181 if (!map->page) {
3fbc9648 182 map->page = page;
7be6d991
AGR
183 page = NULL;
184 }
92476d7f 185 spin_unlock_irq(&pidmap_lock);
7be6d991 186 kfree(page);
1da177e4
LT
187 if (unlikely(!map->page))
188 break;
189 }
190 if (likely(atomic_read(&map->nr_free))) {
191 do {
192 if (!test_and_set_bit(offset, map->page)) {
193 atomic_dec(&map->nr_free);
5fdee8c4 194 set_last_pid(pid_ns, last, pid);
1da177e4
LT
195 return pid;
196 }
197 offset = find_next_offset(map, offset);
61a58c6c 198 pid = mk_pid(pid_ns, map, offset);
1da177e4
LT
199 /*
200 * find_next_offset() found a bit, the pid from it
201 * is in-bounds, and if we fell back to the last
202 * bitmap block and the final block was the same
203 * as the starting point, pid is before last_pid.
204 */
205 } while (offset < BITS_PER_PAGE && pid < pid_max &&
206 (i != max_scan || pid < last ||
207 !((last+1) & BITS_PER_PAGE_MASK)));
208 }
61a58c6c 209 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
1da177e4
LT
210 ++map;
211 offset = 0;
212 } else {
61a58c6c 213 map = &pid_ns->pidmap[0];
1da177e4
LT
214 offset = RESERVED_PIDS;
215 if (unlikely(last == offset))
216 break;
217 }
61a58c6c 218 pid = mk_pid(pid_ns, map, offset);
1da177e4
LT
219 }
220 return -1;
221}
222
74bd59bb 223int next_pidmap(struct pid_namespace *pid_ns, int last)
0804ef4b
EB
224{
225 int offset;
f40f50d3 226 struct pidmap *map, *end;
0804ef4b
EB
227
228 offset = (last + 1) & BITS_PER_PAGE_MASK;
61a58c6c
SB
229 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
230 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
f40f50d3 231 for (; map < end; map++, offset = 0) {
0804ef4b
EB
232 if (unlikely(!map->page))
233 continue;
234 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
235 if (offset < BITS_PER_PAGE)
61a58c6c 236 return mk_pid(pid_ns, map, offset);
0804ef4b
EB
237 }
238 return -1;
239}
240
7ad5b3a5 241void put_pid(struct pid *pid)
92476d7f 242{
baf8f0f8
PE
243 struct pid_namespace *ns;
244
92476d7f
EB
245 if (!pid)
246 return;
baf8f0f8 247
8ef047aa 248 ns = pid->numbers[pid->level].ns;
92476d7f 249 if ((atomic_read(&pid->count) == 1) ||
8ef047aa 250 atomic_dec_and_test(&pid->count)) {
baf8f0f8 251 kmem_cache_free(ns->pid_cachep, pid);
b461cc03 252 put_pid_ns(ns);
8ef047aa 253 }
92476d7f 254}
bbf73147 255EXPORT_SYMBOL_GPL(put_pid);
92476d7f
EB
256
257static 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
7ad5b3a5 263void free_pid(struct pid *pid)
92476d7f
EB
264{
265 /* We can be called with write_lock_irq(&tasklist_lock) held */
8ef047aa 266 int i;
92476d7f
EB
267 unsigned long flags;
268
269 spin_lock_irqsave(&pidmap_lock, flags);
198fe21b
PE
270 for (i = 0; i <= pid->level; i++)
271 hlist_del_rcu(&pid->numbers[i].pid_chain);
92476d7f
EB
272 spin_unlock_irqrestore(&pidmap_lock, flags);
273
8ef047aa 274 for (i = 0; i <= pid->level; i++)
b7127aa4 275 free_pidmap(pid->numbers + i);
8ef047aa 276
92476d7f
EB
277 call_rcu(&pid->rcu, delayed_put_pid);
278}
279
8ef047aa 280struct pid *alloc_pid(struct pid_namespace *ns)
92476d7f
EB
281{
282 struct pid *pid;
283 enum pid_type type;
8ef047aa
PE
284 int i, nr;
285 struct pid_namespace *tmp;
198fe21b 286 struct upid *upid;
92476d7f 287
baf8f0f8 288 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
92476d7f
EB
289 if (!pid)
290 goto out;
291
8ef047aa
PE
292 tmp = ns;
293 for (i = ns->level; i >= 0; i--) {
294 nr = alloc_pidmap(tmp);
295 if (nr < 0)
296 goto out_free;
92476d7f 297
8ef047aa
PE
298 pid->numbers[i].nr = nr;
299 pid->numbers[i].ns = tmp;
300 tmp = tmp->parent;
301 }
302
b461cc03 303 get_pid_ns(ns);
8ef047aa 304 pid->level = ns->level;
92476d7f 305 atomic_set(&pid->count, 1);
92476d7f
EB
306 for (type = 0; type < PIDTYPE_MAX; ++type)
307 INIT_HLIST_HEAD(&pid->tasks[type]);
308
417e3152 309 upid = pid->numbers + ns->level;
92476d7f 310 spin_lock_irq(&pidmap_lock);
417e3152 311 for ( ; upid >= pid->numbers; --upid)
198fe21b
PE
312 hlist_add_head_rcu(&upid->pid_chain,
313 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
92476d7f
EB
314 spin_unlock_irq(&pidmap_lock);
315
316out:
317 return pid;
318
319out_free:
b7127aa4
ON
320 while (++i <= ns->level)
321 free_pidmap(pid->numbers + i);
8ef047aa 322
baf8f0f8 323 kmem_cache_free(ns->pid_cachep, pid);
92476d7f
EB
324 pid = NULL;
325 goto out;
326}
327
7ad5b3a5 328struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
1da177e4
LT
329{
330 struct hlist_node *elem;
198fe21b
PE
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]);
1da177e4 338
1da177e4
LT
339 return NULL;
340}
198fe21b 341EXPORT_SYMBOL_GPL(find_pid_ns);
1da177e4 342
8990571e
PE
343struct pid *find_vpid(int nr)
344{
345 return find_pid_ns(nr, current->nsproxy->pid_ns);
346}
347EXPORT_SYMBOL_GPL(find_vpid);
348
e713d0da
SB
349/*
350 * attach_pid() must be called with the tasklist_lock write-held.
351 */
24336eae 352void attach_pid(struct task_struct *task, enum pid_type type,
e713d0da 353 struct pid *pid)
1da177e4 354{
92476d7f 355 struct pid_link *link;
92476d7f 356
92476d7f 357 link = &task->pids[type];
e713d0da 358 link->pid = pid;
92476d7f 359 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
1da177e4
LT
360}
361
24336eae
ON
362static void __change_pid(struct task_struct *task, enum pid_type type,
363 struct pid *new)
1da177e4 364{
92476d7f
EB
365 struct pid_link *link;
366 struct pid *pid;
367 int tmp;
1da177e4 368
92476d7f
EB
369 link = &task->pids[type];
370 pid = link->pid;
1da177e4 371
92476d7f 372 hlist_del_rcu(&link->node);
24336eae 373 link->pid = new;
1da177e4 374
92476d7f
EB
375 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
376 if (!hlist_empty(&pid->tasks[tmp]))
377 return;
1da177e4 378
92476d7f 379 free_pid(pid);
1da177e4
LT
380}
381
24336eae
ON
382void detach_pid(struct task_struct *task, enum pid_type type)
383{
384 __change_pid(task, type, NULL);
385}
386
387void 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
c18258c6 394/* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
7ad5b3a5 395void transfer_pid(struct task_struct *old, struct task_struct *new,
c18258c6
EB
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);
c18258c6
EB
400}
401
7ad5b3a5 402struct task_struct *pid_task(struct pid *pid, enum pid_type type)
1da177e4 403{
92476d7f
EB
404 struct task_struct *result = NULL;
405 if (pid) {
406 struct hlist_node *first;
db1466b3
PM
407 first = rcu_dereference_check(pid->tasks[type].first,
408 rcu_read_lock_held() ||
409 lockdep_tasklist_lock_is_held());
92476d7f
EB
410 if (first)
411 result = hlist_entry(first, struct task_struct, pids[(type)].node);
412 }
413 return result;
414}
eccba068 415EXPORT_SYMBOL(pid_task);
1da177e4 416
92476d7f 417/*
9728e5d6 418 * Must be called under rcu_read_lock().
92476d7f 419 */
17f98dcf 420struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
92476d7f 421{
17f98dcf 422 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
92476d7f 423}
1da177e4 424
228ebcbe
PE
425struct task_struct *find_task_by_vpid(pid_t vnr)
426{
17f98dcf 427 return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns);
228ebcbe 428}
228ebcbe 429
1a657f78
ON
430struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
431{
432 struct pid *pid;
433 rcu_read_lock();
2ae448ef
ON
434 if (type != PIDTYPE_PID)
435 task = task->group_leader;
1a657f78
ON
436 pid = get_pid(task->pids[type].pid);
437 rcu_read_unlock();
438 return pid;
439}
440
7ad5b3a5 441struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
92476d7f
EB
442{
443 struct task_struct *result;
444 rcu_read_lock();
445 result = pid_task(pid, type);
446 if (result)
447 get_task_struct(result);
448 rcu_read_unlock();
449 return result;
1da177e4
LT
450}
451
92476d7f 452struct pid *find_get_pid(pid_t nr)
1da177e4
LT
453{
454 struct pid *pid;
455
92476d7f 456 rcu_read_lock();
198fe21b 457 pid = get_pid(find_vpid(nr));
92476d7f 458 rcu_read_unlock();
1da177e4 459
92476d7f 460 return pid;
1da177e4 461}
339caf2a 462EXPORT_SYMBOL_GPL(find_get_pid);
1da177e4 463
7af57294
PE
464pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
465{
466 struct upid *upid;
467 pid_t nr = 0;
468
469 if (pid && ns->level <= pid->level) {
470 upid = &pid->numbers[ns->level];
471 if (upid->ns == ns)
472 nr = upid->nr;
473 }
474 return nr;
475}
476
44c4e1b2
EB
477pid_t pid_vnr(struct pid *pid)
478{
479 return pid_nr_ns(pid, current->nsproxy->pid_ns);
480}
481EXPORT_SYMBOL_GPL(pid_vnr);
482
52ee2dfd
ON
483pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
484 struct pid_namespace *ns)
2f2a3a46 485{
52ee2dfd
ON
486 pid_t nr = 0;
487
488 rcu_read_lock();
489 if (!ns)
490 ns = current->nsproxy->pid_ns;
491 if (likely(pid_alive(task))) {
492 if (type != PIDTYPE_PID)
493 task = task->group_leader;
494 nr = pid_nr_ns(task->pids[type].pid, ns);
495 }
496 rcu_read_unlock();
497
498 return nr;
2f2a3a46 499}
52ee2dfd 500EXPORT_SYMBOL(__task_pid_nr_ns);
2f2a3a46
PE
501
502pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
503{
504 return pid_nr_ns(task_tgid(tsk), ns);
505}
506EXPORT_SYMBOL(task_tgid_nr_ns);
507
61bce0f1
EB
508struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
509{
510 return ns_of_pid(task_pid(tsk));
511}
512EXPORT_SYMBOL_GPL(task_active_pid_ns);
513
0804ef4b 514/*
025dfdaf 515 * Used by proc to find the first pid that is greater than or equal to nr.
0804ef4b 516 *
e49859e7 517 * If there is a pid at nr this function is exactly the same as find_pid_ns.
0804ef4b 518 */
198fe21b 519struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
0804ef4b
EB
520{
521 struct pid *pid;
522
523 do {
198fe21b 524 pid = find_pid_ns(nr, ns);
0804ef4b
EB
525 if (pid)
526 break;
198fe21b 527 nr = next_pidmap(ns, nr);
0804ef4b
EB
528 } while (nr > 0);
529
530 return pid;
531}
532
1da177e4
LT
533/*
534 * The pid hash table is scaled according to the amount of memory in the
535 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
536 * more.
537 */
538void __init pidhash_init(void)
539{
92476d7f 540 int i, pidhash_size;
1da177e4 541
2c85f51d
JB
542 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
543 HASH_EARLY | HASH_SMALL,
544 &pidhash_shift, NULL, 4096);
1da177e4
LT
545 pidhash_size = 1 << pidhash_shift;
546
92476d7f
EB
547 for (i = 0; i < pidhash_size; i++)
548 INIT_HLIST_HEAD(&pid_hash[i]);
1da177e4
LT
549}
550
551void __init pidmap_init(void)
552{
72680a19
HB
553 /* bump default and minimum pid_max based on number of cpus */
554 pid_max = min(pid_max_max, max_t(int, pid_max,
555 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
556 pid_max_min = max_t(int, pid_max_min,
557 PIDS_PER_CPU_MIN * num_possible_cpus());
558 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
559
61a58c6c 560 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
73b9ebfe 561 /* Reserve PID 0. We never call free_pidmap(0) */
61a58c6c
SB
562 set_bit(0, init_pid_ns.pidmap[0].page);
563 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
92476d7f 564
74bd59bb
PE
565 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
566 SLAB_HWCACHE_ALIGN | SLAB_PANIC);
1da177e4 567}