2 * Generic pidhash and scalable, time-bounded PID allocator
4 * (C) 2002-2003 William Irwin, IBM
5 * (C) 2004 William Irwin, 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/module.h>
31 #include <linux/slab.h>
32 #include <linux/init.h>
33 #include <linux/bootmem.h>
34 #include <linux/hash.h>
35 #include <linux/pid_namespace.h>
36 #include <linux/init_task.h>
37 #include <linux/syscalls.h>
39 #define pid_hashfn(nr, ns) \
40 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
41 static struct hlist_head *pid_hash;
42 static int pidhash_shift;
43 struct pid init_struct_pid = INIT_STRUCT_PID;
44 static struct kmem_cache *pid_ns_cachep;
46 int pid_max = PID_MAX_DEFAULT;
48 #define RESERVED_PIDS 300
50 int pid_max_min = RESERVED_PIDS + 1;
51 int pid_max_max = PID_MAX_LIMIT;
53 #define BITS_PER_PAGE (PAGE_SIZE*8)
54 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
56 static inline int mk_pid(struct pid_namespace *pid_ns,
57 struct pidmap *map, int off)
59 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
62 #define find_next_offset(map, off) \
63 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
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.
71 struct pid_namespace init_pid_ns = {
73 .refcount = ATOMIC_INIT(2),
76 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
80 .child_reaper = &init_task,
82 EXPORT_SYMBOL_GPL(init_pid_ns);
84 int is_container_init(struct task_struct *tsk)
91 if (pid != NULL && pid->numbers[pid->level].nr == 1)
97 EXPORT_SYMBOL(is_container_init);
100 * Note: disable interrupts while the pidmap_lock is held as an
101 * interrupt might come in and do read_lock(&tasklist_lock).
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);
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.
113 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
115 static fastcall void free_pidmap(struct pid_namespace *pid_ns, int pid)
117 struct pidmap *map = pid_ns->pidmap + pid / BITS_PER_PAGE;
118 int offset = pid & BITS_PER_PAGE_MASK;
120 clear_bit(offset, map->page);
121 atomic_inc(&map->nr_free);
124 static int alloc_pidmap(struct pid_namespace *pid_ns)
126 int i, offset, max_scan, pid, last = pid_ns->last_pid;
132 offset = pid & BITS_PER_PAGE_MASK;
133 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
134 max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
135 for (i = 0; i <= max_scan; ++i) {
136 if (unlikely(!map->page)) {
137 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
139 * Free the page if someone raced with us
142 spin_lock_irq(&pidmap_lock);
147 spin_unlock_irq(&pidmap_lock);
148 if (unlikely(!map->page))
151 if (likely(atomic_read(&map->nr_free))) {
153 if (!test_and_set_bit(offset, map->page)) {
154 atomic_dec(&map->nr_free);
155 pid_ns->last_pid = pid;
158 offset = find_next_offset(map, offset);
159 pid = mk_pid(pid_ns, map, offset);
161 * find_next_offset() found a bit, the pid from it
162 * is in-bounds, and if we fell back to the last
163 * bitmap block and the final block was the same
164 * as the starting point, pid is before last_pid.
166 } while (offset < BITS_PER_PAGE && pid < pid_max &&
167 (i != max_scan || pid < last ||
168 !((last+1) & BITS_PER_PAGE_MASK)));
170 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
174 map = &pid_ns->pidmap[0];
175 offset = RESERVED_PIDS;
176 if (unlikely(last == offset))
179 pid = mk_pid(pid_ns, map, offset);
184 static int next_pidmap(struct pid_namespace *pid_ns, int last)
187 struct pidmap *map, *end;
189 offset = (last + 1) & BITS_PER_PAGE_MASK;
190 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
191 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
192 for (; map < end; map++, offset = 0) {
193 if (unlikely(!map->page))
195 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
196 if (offset < BITS_PER_PAGE)
197 return mk_pid(pid_ns, map, offset);
202 fastcall void put_pid(struct pid *pid)
204 struct pid_namespace *ns;
209 ns = pid->numbers[pid->level].ns;
210 if ((atomic_read(&pid->count) == 1) ||
211 atomic_dec_and_test(&pid->count)) {
212 kmem_cache_free(ns->pid_cachep, pid);
216 EXPORT_SYMBOL_GPL(put_pid);
218 static void delayed_put_pid(struct rcu_head *rhp)
220 struct pid *pid = container_of(rhp, struct pid, rcu);
224 fastcall void free_pid(struct pid *pid)
226 /* We can be called with write_lock_irq(&tasklist_lock) held */
230 spin_lock_irqsave(&pidmap_lock, flags);
231 for (i = 0; i <= pid->level; i++)
232 hlist_del_rcu(&pid->numbers[i].pid_chain);
233 spin_unlock_irqrestore(&pidmap_lock, flags);
235 for (i = 0; i <= pid->level; i++)
236 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
238 call_rcu(&pid->rcu, delayed_put_pid);
241 struct pid *alloc_pid(struct pid_namespace *ns)
246 struct pid_namespace *tmp;
249 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
254 for (i = ns->level; i >= 0; i--) {
255 nr = alloc_pidmap(tmp);
259 pid->numbers[i].nr = nr;
260 pid->numbers[i].ns = tmp;
265 pid->level = ns->level;
266 pid->nr = pid->numbers[0].nr;
267 atomic_set(&pid->count, 1);
268 for (type = 0; type < PIDTYPE_MAX; ++type)
269 INIT_HLIST_HEAD(&pid->tasks[type]);
271 spin_lock_irq(&pidmap_lock);
272 for (i = ns->level; i >= 0; i--) {
273 upid = &pid->numbers[i];
274 hlist_add_head_rcu(&upid->pid_chain,
275 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
277 spin_unlock_irq(&pidmap_lock);
283 for (i++; i <= ns->level; i++)
284 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
286 kmem_cache_free(ns->pid_cachep, pid);
291 struct pid * fastcall find_pid_ns(int nr, struct pid_namespace *ns)
293 struct hlist_node *elem;
296 hlist_for_each_entry_rcu(pnr, elem,
297 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
298 if (pnr->nr == nr && pnr->ns == ns)
299 return container_of(pnr, struct pid,
304 EXPORT_SYMBOL_GPL(find_pid_ns);
307 * attach_pid() must be called with the tasklist_lock write-held.
309 int fastcall attach_pid(struct task_struct *task, enum pid_type type,
312 struct pid_link *link;
314 link = &task->pids[type];
316 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
321 void fastcall detach_pid(struct task_struct *task, enum pid_type type)
323 struct pid_link *link;
327 link = &task->pids[type];
330 hlist_del_rcu(&link->node);
333 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
334 if (!hlist_empty(&pid->tasks[tmp]))
340 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
341 void fastcall transfer_pid(struct task_struct *old, struct task_struct *new,
344 new->pids[type].pid = old->pids[type].pid;
345 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
346 old->pids[type].pid = NULL;
349 struct task_struct * fastcall pid_task(struct pid *pid, enum pid_type type)
351 struct task_struct *result = NULL;
353 struct hlist_node *first;
354 first = rcu_dereference(pid->tasks[type].first);
356 result = hlist_entry(first, struct task_struct, pids[(type)].node);
362 * Must be called under rcu_read_lock() or with tasklist_lock read-held.
364 struct task_struct *find_task_by_pid_type_ns(int type, int nr,
365 struct pid_namespace *ns)
367 return pid_task(find_pid_ns(nr, ns), type);
370 EXPORT_SYMBOL(find_task_by_pid_type_ns);
372 struct task_struct *find_task_by_pid(pid_t nr)
374 return find_task_by_pid_type_ns(PIDTYPE_PID, nr, &init_pid_ns);
376 EXPORT_SYMBOL(find_task_by_pid);
378 struct task_struct *find_task_by_vpid(pid_t vnr)
380 return find_task_by_pid_type_ns(PIDTYPE_PID, vnr,
381 current->nsproxy->pid_ns);
383 EXPORT_SYMBOL(find_task_by_vpid);
385 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
387 return find_task_by_pid_type_ns(PIDTYPE_PID, nr, ns);
389 EXPORT_SYMBOL(find_task_by_pid_ns);
391 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
395 pid = get_pid(task->pids[type].pid);
400 struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
402 struct task_struct *result;
404 result = pid_task(pid, type);
406 get_task_struct(result);
411 struct pid *find_get_pid(pid_t nr)
416 pid = get_pid(find_vpid(nr));
422 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
427 if (pid && ns->level <= pid->level) {
428 upid = &pid->numbers[ns->level];
436 * Used by proc to find the first pid that is greater then or equal to nr.
438 * If there is a pid at nr this function is exactly the same as find_pid.
440 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
445 pid = find_pid_ns(nr, ns);
448 nr = next_pidmap(ns, nr);
453 EXPORT_SYMBOL_GPL(find_get_pid);
458 struct kmem_cache *cachep;
459 struct list_head list;
462 static LIST_HEAD(pid_caches_lh);
463 static DEFINE_MUTEX(pid_caches_mutex);
466 * creates the kmem cache to allocate pids from.
467 * @nr_ids: the number of numerical ids this pid will have to carry
470 static struct kmem_cache *create_pid_cachep(int nr_ids)
472 struct pid_cache *pcache;
473 struct kmem_cache *cachep;
475 mutex_lock(&pid_caches_mutex);
476 list_for_each_entry (pcache, &pid_caches_lh, list)
477 if (pcache->nr_ids == nr_ids)
480 pcache = kmalloc(sizeof(struct pid_cache), GFP_KERNEL);
484 snprintf(pcache->name, sizeof(pcache->name), "pid_%d", nr_ids);
485 cachep = kmem_cache_create(pcache->name,
486 sizeof(struct pid) + (nr_ids - 1) * sizeof(struct upid),
487 0, SLAB_HWCACHE_ALIGN, NULL);
491 pcache->nr_ids = nr_ids;
492 pcache->cachep = cachep;
493 list_add(&pcache->list, &pid_caches_lh);
495 mutex_unlock(&pid_caches_mutex);
496 return pcache->cachep;
501 mutex_unlock(&pid_caches_mutex);
505 static struct pid_namespace *create_pid_namespace(int level)
507 struct pid_namespace *ns;
510 ns = kmem_cache_alloc(pid_ns_cachep, GFP_KERNEL);
514 ns->pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
515 if (!ns->pidmap[0].page)
518 ns->pid_cachep = create_pid_cachep(level + 1);
519 if (ns->pid_cachep == NULL)
522 kref_init(&ns->kref);
524 ns->child_reaper = NULL;
527 set_bit(0, ns->pidmap[0].page);
528 atomic_set(&ns->pidmap[0].nr_free, BITS_PER_PAGE - 1);
530 for (i = 1; i < PIDMAP_ENTRIES; i++) {
531 ns->pidmap[i].page = 0;
532 atomic_set(&ns->pidmap[i].nr_free, BITS_PER_PAGE);
538 kfree(ns->pidmap[0].page);
540 kmem_cache_free(pid_ns_cachep, ns);
542 return ERR_PTR(-ENOMEM);
545 static void destroy_pid_namespace(struct pid_namespace *ns)
549 for (i = 0; i < PIDMAP_ENTRIES; i++)
550 kfree(ns->pidmap[i].page);
551 kmem_cache_free(pid_ns_cachep, ns);
554 struct pid_namespace *copy_pid_ns(unsigned long flags, struct pid_namespace *old_ns)
556 struct pid_namespace *new_ns;
559 new_ns = get_pid_ns(old_ns);
560 if (!(flags & CLONE_NEWPID))
563 new_ns = ERR_PTR(-EINVAL);
564 if (flags & CLONE_THREAD)
567 new_ns = create_pid_namespace(old_ns->level + 1);
569 new_ns->parent = get_pid_ns(old_ns);
577 void free_pid_ns(struct kref *kref)
579 struct pid_namespace *ns, *parent;
581 ns = container_of(kref, struct pid_namespace, kref);
584 destroy_pid_namespace(ns);
590 void zap_pid_ns_processes(struct pid_namespace *pid_ns)
596 * The last thread in the cgroup-init thread group is terminating.
597 * Find remaining pid_ts in the namespace, signal and wait for them
600 * Note: This signals each threads in the namespace - even those that
601 * belong to the same thread group, To avoid this, we would have
602 * to walk the entire tasklist looking a processes in this
603 * namespace, but that could be unnecessarily expensive if the
604 * pid namespace has just a few processes. Or we need to
605 * maintain a tasklist for each pid namespace.
608 read_lock(&tasklist_lock);
609 nr = next_pidmap(pid_ns, 1);
611 kill_proc_info(SIGKILL, SEND_SIG_PRIV, nr);
612 nr = next_pidmap(pid_ns, nr);
614 read_unlock(&tasklist_lock);
617 clear_thread_flag(TIF_SIGPENDING);
618 rc = sys_wait4(-1, NULL, __WALL, NULL);
619 } while (rc != -ECHILD);
622 /* Child reaper for the pid namespace is going away */
623 pid_ns->child_reaper = NULL;
628 * The pid hash table is scaled according to the amount of memory in the
629 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
632 void __init pidhash_init(void)
635 unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
637 pidhash_shift = max(4, fls(megabytes * 4));
638 pidhash_shift = min(12, pidhash_shift);
639 pidhash_size = 1 << pidhash_shift;
641 printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
642 pidhash_size, pidhash_shift,
643 pidhash_size * sizeof(struct hlist_head));
645 pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash)));
647 panic("Could not alloc pidhash!\n");
648 for (i = 0; i < pidhash_size; i++)
649 INIT_HLIST_HEAD(&pid_hash[i]);
652 void __init pidmap_init(void)
654 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
655 /* Reserve PID 0. We never call free_pidmap(0) */
656 set_bit(0, init_pid_ns.pidmap[0].page);
657 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
659 init_pid_ns.pid_cachep = create_pid_cachep(1);
660 if (init_pid_ns.pid_cachep == NULL)
661 panic("Can't create pid_1 cachep\n");
663 pid_ns_cachep = KMEM_CACHE(pid_namespace, SLAB_PANIC);