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 }
79 .child_reaper = &init_task,
80 .user_ns = &init_user_ns,
81 .proc_inum = PROC_PID_INIT_INO,
83 EXPORT_SYMBOL_GPL(init_pid_ns);
86 * Note: disable interrupts while the pidmap_lock is held as an
87 * interrupt might come in and do read_lock(&tasklist_lock).
89 * If we don't disable interrupts there is a nasty deadlock between
90 * detach_pid()->free_pid() and another cpu that does
91 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
92 * read_lock(&tasklist_lock);
94 * After we clean up the tasklist_lock and know there are no
95 * irq handlers that take it we can leave the interrupts enabled.
96 * For now it is easier to be safe than to prove it can't happen.
99 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
101 static void free_pidmap(struct upid *upid)
104 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
105 int offset = nr & BITS_PER_PAGE_MASK;
107 clear_bit(offset, map->page);
108 atomic_inc(&map->nr_free);
112 * If we started walking pids at 'base', is 'a' seen before 'b'?
114 static int pid_before(int base, int a, int b)
117 * This is the same as saying
119 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
120 * and that mapping orders 'a' and 'b' with respect to 'base'.
122 return (unsigned)(a - base) < (unsigned)(b - base);
126 * We might be racing with someone else trying to set pid_ns->last_pid
127 * at the pid allocation time (there's also a sysctl for this, but racing
128 * with this one is OK, see comment in kernel/pid_namespace.c about it).
129 * We want the winner to have the "later" value, because if the
130 * "earlier" value prevails, then a pid may get reused immediately.
132 * Since pids rollover, it is not sufficient to just pick the bigger
133 * value. We have to consider where we started counting from.
135 * 'base' is the value of pid_ns->last_pid that we observed when
136 * we started looking for a pid.
138 * 'pid' is the pid that we eventually found.
140 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
143 int last_write = base;
146 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
147 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
150 static int alloc_pidmap(struct pid_namespace *pid_ns)
152 int i, offset, max_scan, pid, last = pid_ns->last_pid;
158 offset = pid & BITS_PER_PAGE_MASK;
159 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
161 * If last_pid points into the middle of the map->page we
162 * want to scan this bitmap block twice, the second time
163 * we start with offset == 0 (or RESERVED_PIDS).
165 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
166 for (i = 0; i <= max_scan; ++i) {
167 if (unlikely(!map->page)) {
168 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
170 * Free the page if someone raced with us
173 spin_lock_irq(&pidmap_lock);
178 spin_unlock_irq(&pidmap_lock);
180 if (unlikely(!map->page))
183 if (likely(atomic_read(&map->nr_free))) {
185 if (!test_and_set_bit(offset, map->page)) {
186 atomic_dec(&map->nr_free);
187 set_last_pid(pid_ns, last, pid);
190 offset = find_next_offset(map, offset);
191 if (offset >= BITS_PER_PAGE)
193 pid = mk_pid(pid_ns, map, offset);
198 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
202 map = &pid_ns->pidmap[0];
203 offset = RESERVED_PIDS;
204 if (unlikely(last == offset))
207 pid = mk_pid(pid_ns, map, offset);
212 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
215 struct pidmap *map, *end;
217 if (last >= PID_MAX_LIMIT)
220 offset = (last + 1) & BITS_PER_PAGE_MASK;
221 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
222 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
223 for (; map < end; map++, offset = 0) {
224 if (unlikely(!map->page))
226 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
227 if (offset < BITS_PER_PAGE)
228 return mk_pid(pid_ns, map, offset);
233 void put_pid(struct pid *pid)
235 struct pid_namespace *ns;
240 ns = pid->numbers[pid->level].ns;
241 if ((atomic_read(&pid->count) == 1) ||
242 atomic_dec_and_test(&pid->count)) {
243 kmem_cache_free(ns->pid_cachep, pid);
247 EXPORT_SYMBOL_GPL(put_pid);
249 static void delayed_put_pid(struct rcu_head *rhp)
251 struct pid *pid = container_of(rhp, struct pid, rcu);
255 void free_pid(struct pid *pid)
257 /* We can be called with write_lock_irq(&tasklist_lock) held */
261 spin_lock_irqsave(&pidmap_lock, flags);
262 for (i = 0; i <= pid->level; i++) {
263 struct upid *upid = pid->numbers + i;
264 struct pid_namespace *ns = upid->ns;
265 hlist_del_rcu(&upid->pid_chain);
266 switch(--ns->nr_hashed) {
268 /* When all that is left in the pid namespace
269 * is the reaper wake up the reaper. The reaper
270 * may be sleeping in zap_pid_ns_processes().
272 wake_up_process(ns->child_reaper);
275 schedule_work(&ns->proc_work);
279 spin_unlock_irqrestore(&pidmap_lock, flags);
281 for (i = 0; i <= pid->level; i++)
282 free_pidmap(pid->numbers + i);
284 call_rcu(&pid->rcu, delayed_put_pid);
287 struct pid *alloc_pid(struct pid_namespace *ns)
292 struct pid_namespace *tmp;
295 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
300 pid->level = ns->level;
301 for (i = ns->level; i >= 0; i--) {
302 nr = alloc_pidmap(tmp);
306 pid->numbers[i].nr = nr;
307 pid->numbers[i].ns = tmp;
311 if (unlikely(is_child_reaper(pid))) {
312 if (pid_ns_prepare_proc(ns))
317 atomic_set(&pid->count, 1);
318 for (type = 0; type < PIDTYPE_MAX; ++type)
319 INIT_HLIST_HEAD(&pid->tasks[type]);
321 upid = pid->numbers + ns->level;
322 spin_lock_irq(&pidmap_lock);
323 if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
325 for ( ; upid >= pid->numbers; --upid) {
326 hlist_add_head_rcu(&upid->pid_chain,
327 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
328 upid->ns->nr_hashed++;
330 spin_unlock_irq(&pidmap_lock);
336 spin_unlock_irq(&pidmap_lock);
338 while (++i <= ns->level)
339 free_pidmap(pid->numbers + i);
341 kmem_cache_free(ns->pid_cachep, pid);
346 void disable_pid_allocation(struct pid_namespace *ns)
348 spin_lock_irq(&pidmap_lock);
349 ns->nr_hashed &= ~PIDNS_HASH_ADDING;
350 spin_unlock_irq(&pidmap_lock);
353 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
357 hlist_for_each_entry_rcu(pnr,
358 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
359 if (pnr->nr == nr && pnr->ns == ns)
360 return container_of(pnr, struct pid,
365 EXPORT_SYMBOL_GPL(find_pid_ns);
367 struct pid *find_vpid(int nr)
369 return find_pid_ns(nr, task_active_pid_ns(current));
371 EXPORT_SYMBOL_GPL(find_vpid);
374 * attach_pid() must be called with the tasklist_lock write-held.
376 void attach_pid(struct task_struct *task, enum pid_type type,
379 struct pid_link *link;
381 link = &task->pids[type];
383 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
386 static void __change_pid(struct task_struct *task, enum pid_type type,
389 struct pid_link *link;
393 link = &task->pids[type];
396 hlist_del_rcu(&link->node);
399 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
400 if (!hlist_empty(&pid->tasks[tmp]))
406 void detach_pid(struct task_struct *task, enum pid_type type)
408 __change_pid(task, type, NULL);
411 void change_pid(struct task_struct *task, enum pid_type type,
414 __change_pid(task, type, pid);
415 attach_pid(task, type, pid);
418 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
419 void transfer_pid(struct task_struct *old, struct task_struct *new,
422 new->pids[type].pid = old->pids[type].pid;
423 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
426 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
428 struct task_struct *result = NULL;
430 struct hlist_node *first;
431 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
432 lockdep_tasklist_lock_is_held());
434 result = hlist_entry(first, struct task_struct, pids[(type)].node);
438 EXPORT_SYMBOL(pid_task);
441 * Must be called under rcu_read_lock().
443 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
445 rcu_lockdep_assert(rcu_read_lock_held(),
446 "find_task_by_pid_ns() needs rcu_read_lock()"
448 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
451 struct task_struct *find_task_by_vpid(pid_t vnr)
453 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
456 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
460 if (type != PIDTYPE_PID)
461 task = task->group_leader;
462 pid = get_pid(task->pids[type].pid);
466 EXPORT_SYMBOL_GPL(get_task_pid);
468 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
470 struct task_struct *result;
472 result = pid_task(pid, type);
474 get_task_struct(result);
478 EXPORT_SYMBOL_GPL(get_pid_task);
480 struct pid *find_get_pid(pid_t nr)
485 pid = get_pid(find_vpid(nr));
490 EXPORT_SYMBOL_GPL(find_get_pid);
492 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
497 if (pid && ns->level <= pid->level) {
498 upid = &pid->numbers[ns->level];
504 EXPORT_SYMBOL_GPL(pid_nr_ns);
506 pid_t pid_vnr(struct pid *pid)
508 return pid_nr_ns(pid, task_active_pid_ns(current));
510 EXPORT_SYMBOL_GPL(pid_vnr);
512 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
513 struct pid_namespace *ns)
519 ns = task_active_pid_ns(current);
520 if (likely(pid_alive(task))) {
521 if (type != PIDTYPE_PID)
522 task = task->group_leader;
523 nr = pid_nr_ns(task->pids[type].pid, ns);
529 EXPORT_SYMBOL(__task_pid_nr_ns);
531 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
533 return pid_nr_ns(task_tgid(tsk), ns);
535 EXPORT_SYMBOL(task_tgid_nr_ns);
537 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
539 return ns_of_pid(task_pid(tsk));
541 EXPORT_SYMBOL_GPL(task_active_pid_ns);
544 * Used by proc to find the first pid that is greater than or equal to nr.
546 * If there is a pid at nr this function is exactly the same as find_pid_ns.
548 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
553 pid = find_pid_ns(nr, ns);
556 nr = next_pidmap(ns, nr);
563 * The pid hash table is scaled according to the amount of memory in the
564 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
567 void __init pidhash_init(void)
569 unsigned int i, pidhash_size;
571 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
572 HASH_EARLY | HASH_SMALL,
573 &pidhash_shift, NULL,
575 pidhash_size = 1U << pidhash_shift;
577 for (i = 0; i < pidhash_size; i++)
578 INIT_HLIST_HEAD(&pid_hash[i]);
581 void __init pidmap_init(void)
583 /* Veryify no one has done anything silly */
584 BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_HASH_ADDING);
586 /* bump default and minimum pid_max based on number of cpus */
587 pid_max = min(pid_max_max, max_t(int, pid_max,
588 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
589 pid_max_min = max_t(int, pid_max_min,
590 PIDS_PER_CPU_MIN * num_possible_cpus());
591 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
593 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
594 /* Reserve PID 0. We never call free_pidmap(0) */
595 set_bit(0, init_pid_ns.pidmap[0].page);
596 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
597 init_pid_ns.nr_hashed = PIDNS_HASH_ADDING;
599 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
600 SLAB_HWCACHE_ALIGN | SLAB_PANIC);