2 * Generic pidhash and scalable, time-bounded PID allocator
4 * (C) 2002-2003 Nadia Yvette Chambers, IBM
5 * (C) 2004 Nadia Yvette Chambers, Oracle
6 * (C) 2002-2004 Ingo Molnar, Red Hat
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.
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.
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).
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
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>
42 #define pid_hashfn(nr, ns) \
43 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
44 static struct hlist_head *pid_hash;
45 static unsigned int pidhash_shift = 4;
46 struct pid init_struct_pid = INIT_STRUCT_PID;
48 int pid_max = PID_MAX_DEFAULT;
50 #define RESERVED_PIDS 300
52 int pid_max_min = RESERVED_PIDS + 1;
53 int pid_max_max = PID_MAX_LIMIT;
55 static inline int mk_pid(struct pid_namespace *pid_ns,
56 struct pidmap *map, int off)
58 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
61 #define find_next_offset(map, off) \
62 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
65 * PID-map pages start out as NULL, they get allocated upon
66 * first use and are never deallocated. This way a low pid_max
67 * value does not cause lots of bitmaps to be allocated, but
68 * the scheme scales to up to 4 million PIDs, runtime.
70 struct pid_namespace init_pid_ns = {
72 .refcount = ATOMIC_INIT(2),
75 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
78 .nr_hashed = PIDNS_HASH_ADDING,
80 .child_reaper = &init_task,
81 .user_ns = &init_user_ns,
82 .proc_inum = PROC_PID_INIT_INO,
84 EXPORT_SYMBOL_GPL(init_pid_ns);
87 * Note: disable interrupts while the pidmap_lock is held as an
88 * interrupt might come in and do read_lock(&tasklist_lock).
90 * If we don't disable interrupts there is a nasty deadlock between
91 * detach_pid()->free_pid() and another cpu that does
92 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
93 * read_lock(&tasklist_lock);
95 * After we clean up the tasklist_lock and know there are no
96 * irq handlers that take it we can leave the interrupts enabled.
97 * For now it is easier to be safe than to prove it can't happen.
100 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
102 static void free_pidmap(struct upid *upid)
105 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
106 int offset = nr & BITS_PER_PAGE_MASK;
108 clear_bit(offset, map->page);
109 atomic_inc(&map->nr_free);
113 * If we started walking pids at 'base', is 'a' seen before 'b'?
115 static int pid_before(int base, int a, int b)
118 * This is the same as saying
120 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
121 * and that mapping orders 'a' and 'b' with respect to 'base'.
123 return (unsigned)(a - base) < (unsigned)(b - base);
127 * We might be racing with someone else trying to set pid_ns->last_pid
128 * at the pid allocation time (there's also a sysctl for this, but racing
129 * with this one is OK, see comment in kernel/pid_namespace.c about it).
130 * We want the winner to have the "later" value, because if the
131 * "earlier" value prevails, then a pid may get reused immediately.
133 * Since pids rollover, it is not sufficient to just pick the bigger
134 * value. We have to consider where we started counting from.
136 * 'base' is the value of pid_ns->last_pid that we observed when
137 * we started looking for a pid.
139 * 'pid' is the pid that we eventually found.
141 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
144 int last_write = base;
147 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
148 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
151 static int alloc_pidmap(struct pid_namespace *pid_ns)
153 int i, offset, max_scan, pid, last = pid_ns->last_pid;
159 offset = pid & BITS_PER_PAGE_MASK;
160 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
162 * If last_pid points into the middle of the map->page we
163 * want to scan this bitmap block twice, the second time
164 * we start with offset == 0 (or RESERVED_PIDS).
166 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
167 for (i = 0; i <= max_scan; ++i) {
168 if (unlikely(!map->page)) {
169 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
171 * Free the page if someone raced with us
174 spin_lock_irq(&pidmap_lock);
179 spin_unlock_irq(&pidmap_lock);
181 if (unlikely(!map->page))
184 if (likely(atomic_read(&map->nr_free))) {
186 if (!test_and_set_bit(offset, map->page)) {
187 atomic_dec(&map->nr_free);
188 set_last_pid(pid_ns, last, pid);
191 offset = find_next_offset(map, offset);
192 if (offset >= BITS_PER_PAGE)
194 pid = mk_pid(pid_ns, map, offset);
199 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
203 map = &pid_ns->pidmap[0];
204 offset = RESERVED_PIDS;
205 if (unlikely(last == offset))
208 pid = mk_pid(pid_ns, map, offset);
213 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
216 struct pidmap *map, *end;
218 if (last >= PID_MAX_LIMIT)
221 offset = (last + 1) & BITS_PER_PAGE_MASK;
222 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
223 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
224 for (; map < end; map++, offset = 0) {
225 if (unlikely(!map->page))
227 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
228 if (offset < BITS_PER_PAGE)
229 return mk_pid(pid_ns, map, offset);
234 void put_pid(struct pid *pid)
236 struct pid_namespace *ns;
241 ns = pid->numbers[pid->level].ns;
242 if ((atomic_read(&pid->count) == 1) ||
243 atomic_dec_and_test(&pid->count)) {
244 kmem_cache_free(ns->pid_cachep, pid);
248 EXPORT_SYMBOL_GPL(put_pid);
250 static void delayed_put_pid(struct rcu_head *rhp)
252 struct pid *pid = container_of(rhp, struct pid, rcu);
256 void free_pid(struct pid *pid)
258 /* We can be called with write_lock_irq(&tasklist_lock) held */
262 spin_lock_irqsave(&pidmap_lock, flags);
263 for (i = 0; i <= pid->level; i++) {
264 struct upid *upid = pid->numbers + i;
265 struct pid_namespace *ns = upid->ns;
266 hlist_del_rcu(&upid->pid_chain);
267 switch(--ns->nr_hashed) {
270 /* When all that is left in the pid namespace
271 * is the reaper wake up the reaper. The reaper
272 * may be sleeping in zap_pid_ns_processes().
274 wake_up_process(ns->child_reaper);
276 case PIDNS_HASH_ADDING:
277 /* Handle a fork failure of the first process */
278 WARN_ON(ns->child_reaper);
282 schedule_work(&ns->proc_work);
286 spin_unlock_irqrestore(&pidmap_lock, flags);
288 for (i = 0; i <= pid->level; i++)
289 free_pidmap(pid->numbers + i);
291 call_rcu(&pid->rcu, delayed_put_pid);
294 struct pid *alloc_pid(struct pid_namespace *ns)
299 struct pid_namespace *tmp;
302 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
307 pid->level = ns->level;
308 for (i = ns->level; i >= 0; i--) {
309 nr = alloc_pidmap(tmp);
313 pid->numbers[i].nr = nr;
314 pid->numbers[i].ns = tmp;
318 if (unlikely(is_child_reaper(pid))) {
319 if (pid_ns_prepare_proc(ns))
324 atomic_set(&pid->count, 1);
325 for (type = 0; type < PIDTYPE_MAX; ++type)
326 INIT_HLIST_HEAD(&pid->tasks[type]);
328 upid = pid->numbers + ns->level;
329 spin_lock_irq(&pidmap_lock);
330 if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
332 for ( ; upid >= pid->numbers; --upid) {
333 hlist_add_head_rcu(&upid->pid_chain,
334 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
335 upid->ns->nr_hashed++;
337 spin_unlock_irq(&pidmap_lock);
343 spin_unlock_irq(&pidmap_lock);
345 while (++i <= ns->level)
346 free_pidmap(pid->numbers + i);
348 kmem_cache_free(ns->pid_cachep, pid);
353 void disable_pid_allocation(struct pid_namespace *ns)
355 spin_lock_irq(&pidmap_lock);
356 ns->nr_hashed &= ~PIDNS_HASH_ADDING;
357 spin_unlock_irq(&pidmap_lock);
360 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
364 hlist_for_each_entry_rcu(pnr,
365 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
366 if (pnr->nr == nr && pnr->ns == ns)
367 return container_of(pnr, struct pid,
372 EXPORT_SYMBOL_GPL(find_pid_ns);
374 struct pid *find_vpid(int nr)
376 return find_pid_ns(nr, task_active_pid_ns(current));
378 EXPORT_SYMBOL_GPL(find_vpid);
381 * attach_pid() must be called with the tasklist_lock write-held.
383 void attach_pid(struct task_struct *task, enum pid_type type)
385 struct pid_link *link = &task->pids[type];
386 hlist_add_head_rcu(&link->node, &link->pid->tasks[type]);
389 static void __change_pid(struct task_struct *task, enum pid_type type,
392 struct pid_link *link;
396 link = &task->pids[type];
399 hlist_del_rcu(&link->node);
402 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
403 if (!hlist_empty(&pid->tasks[tmp]))
409 void detach_pid(struct task_struct *task, enum pid_type type)
411 __change_pid(task, type, NULL);
414 void change_pid(struct task_struct *task, enum pid_type type,
417 __change_pid(task, type, pid);
418 attach_pid(task, type);
421 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
422 void transfer_pid(struct task_struct *old, struct task_struct *new,
425 new->pids[type].pid = old->pids[type].pid;
426 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
429 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
431 struct task_struct *result = NULL;
433 struct hlist_node *first;
434 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
435 lockdep_tasklist_lock_is_held());
437 result = hlist_entry(first, struct task_struct, pids[(type)].node);
441 EXPORT_SYMBOL(pid_task);
444 * Must be called under rcu_read_lock().
446 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
448 rcu_lockdep_assert(rcu_read_lock_held(),
449 "find_task_by_pid_ns() needs rcu_read_lock()"
451 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
454 struct task_struct *find_task_by_vpid(pid_t vnr)
456 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
459 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
463 if (type != PIDTYPE_PID)
464 task = task->group_leader;
465 pid = get_pid(task->pids[type].pid);
469 EXPORT_SYMBOL_GPL(get_task_pid);
471 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
473 struct task_struct *result;
475 result = pid_task(pid, type);
477 get_task_struct(result);
481 EXPORT_SYMBOL_GPL(get_pid_task);
483 struct pid *find_get_pid(pid_t nr)
488 pid = get_pid(find_vpid(nr));
493 EXPORT_SYMBOL_GPL(find_get_pid);
495 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
500 if (pid && ns->level <= pid->level) {
501 upid = &pid->numbers[ns->level];
507 EXPORT_SYMBOL_GPL(pid_nr_ns);
509 pid_t pid_vnr(struct pid *pid)
511 return pid_nr_ns(pid, task_active_pid_ns(current));
513 EXPORT_SYMBOL_GPL(pid_vnr);
515 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
516 struct pid_namespace *ns)
522 ns = task_active_pid_ns(current);
523 if (likely(pid_alive(task))) {
524 if (type != PIDTYPE_PID)
525 task = task->group_leader;
526 nr = pid_nr_ns(task->pids[type].pid, ns);
532 EXPORT_SYMBOL(__task_pid_nr_ns);
534 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
536 return pid_nr_ns(task_tgid(tsk), ns);
538 EXPORT_SYMBOL(task_tgid_nr_ns);
540 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
542 return ns_of_pid(task_pid(tsk));
544 EXPORT_SYMBOL_GPL(task_active_pid_ns);
547 * Used by proc to find the first pid that is greater than or equal to nr.
549 * If there is a pid at nr this function is exactly the same as find_pid_ns.
551 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
556 pid = find_pid_ns(nr, ns);
559 nr = next_pidmap(ns, nr);
566 * The pid hash table is scaled according to the amount of memory in the
567 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
570 void __init pidhash_init(void)
572 unsigned int i, pidhash_size;
574 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
575 HASH_EARLY | HASH_SMALL,
576 &pidhash_shift, NULL,
578 pidhash_size = 1U << pidhash_shift;
580 for (i = 0; i < pidhash_size; i++)
581 INIT_HLIST_HEAD(&pid_hash[i]);
584 void __init pidmap_init(void)
586 /* Veryify no one has done anything silly */
587 BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_HASH_ADDING);
589 /* bump default and minimum pid_max based on number of cpus */
590 pid_max = min(pid_max_max, max_t(int, pid_max,
591 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
592 pid_max_min = max_t(int, pid_max_min,
593 PIDS_PER_CPU_MIN * num_possible_cpus());
594 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
596 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
597 /* Reserve PID 0. We never call free_pidmap(0) */
598 set_bit(0, init_pid_ns.pidmap[0].page);
599 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
601 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
602 SLAB_HWCACHE_ALIGN | SLAB_PANIC);