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/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>
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;
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,
81 .user_ns = &init_user_ns,
83 EXPORT_SYMBOL_GPL(init_pid_ns);
85 int is_container_init(struct task_struct *tsk)
92 if (pid != NULL && pid->numbers[pid->level].nr == 1)
98 EXPORT_SYMBOL(is_container_init);
101 * Note: disable interrupts while the pidmap_lock is held as an
102 * interrupt might come in and do read_lock(&tasklist_lock).
104 * If we don't disable interrupts there is a nasty deadlock between
105 * detach_pid()->free_pid() and another cpu that does
106 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
107 * read_lock(&tasklist_lock);
109 * After we clean up the tasklist_lock and know there are no
110 * irq handlers that take it we can leave the interrupts enabled.
111 * For now it is easier to be safe than to prove it can't happen.
114 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
116 static void free_pidmap(struct upid *upid)
119 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
120 int offset = nr & BITS_PER_PAGE_MASK;
122 clear_bit(offset, map->page);
123 atomic_inc(&map->nr_free);
127 * If we started walking pids at 'base', is 'a' seen before 'b'?
129 static int pid_before(int base, int a, int b)
132 * This is the same as saying
134 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
135 * and that mapping orders 'a' and 'b' with respect to 'base'.
137 return (unsigned)(a - base) < (unsigned)(b - base);
141 * We might be racing with someone else trying to set pid_ns->last_pid
142 * at the pid allocation time (there's also a sysctl for this, but racing
143 * with this one is OK, see comment in kernel/pid_namespace.c about it).
144 * We want the winner to have the "later" value, because if the
145 * "earlier" value prevails, then a pid may get reused immediately.
147 * Since pids rollover, it is not sufficient to just pick the bigger
148 * value. We have to consider where we started counting from.
150 * 'base' is the value of pid_ns->last_pid that we observed when
151 * we started looking for a pid.
153 * 'pid' is the pid that we eventually found.
155 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
158 int last_write = base;
161 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
162 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
165 static int alloc_pidmap(struct pid_namespace *pid_ns)
167 int i, offset, max_scan, pid, last = pid_ns->last_pid;
173 offset = pid & BITS_PER_PAGE_MASK;
174 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
176 * If last_pid points into the middle of the map->page we
177 * want to scan this bitmap block twice, the second time
178 * we start with offset == 0 (or RESERVED_PIDS).
180 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
181 for (i = 0; i <= max_scan; ++i) {
182 if (unlikely(!map->page)) {
183 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
185 * Free the page if someone raced with us
188 spin_lock_irq(&pidmap_lock);
193 spin_unlock_irq(&pidmap_lock);
195 if (unlikely(!map->page))
198 if (likely(atomic_read(&map->nr_free))) {
200 if (!test_and_set_bit(offset, map->page)) {
201 atomic_dec(&map->nr_free);
202 set_last_pid(pid_ns, last, pid);
205 offset = find_next_offset(map, offset);
206 pid = mk_pid(pid_ns, map, offset);
207 } while (offset < BITS_PER_PAGE && pid < pid_max);
209 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
213 map = &pid_ns->pidmap[0];
214 offset = RESERVED_PIDS;
215 if (unlikely(last == offset))
218 pid = mk_pid(pid_ns, map, offset);
223 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
226 struct pidmap *map, *end;
228 if (last >= PID_MAX_LIMIT)
231 offset = (last + 1) & BITS_PER_PAGE_MASK;
232 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
233 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
234 for (; map < end; map++, offset = 0) {
235 if (unlikely(!map->page))
237 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
238 if (offset < BITS_PER_PAGE)
239 return mk_pid(pid_ns, map, offset);
244 void put_pid(struct pid *pid)
246 struct pid_namespace *ns;
251 ns = pid->numbers[pid->level].ns;
252 if ((atomic_read(&pid->count) == 1) ||
253 atomic_dec_and_test(&pid->count)) {
254 kmem_cache_free(ns->pid_cachep, pid);
258 EXPORT_SYMBOL_GPL(put_pid);
260 static void delayed_put_pid(struct rcu_head *rhp)
262 struct pid *pid = container_of(rhp, struct pid, rcu);
266 void free_pid(struct pid *pid)
268 /* We can be called with write_lock_irq(&tasklist_lock) held */
272 spin_lock_irqsave(&pidmap_lock, flags);
273 for (i = 0; i <= pid->level; i++)
274 hlist_del_rcu(&pid->numbers[i].pid_chain);
275 spin_unlock_irqrestore(&pidmap_lock, flags);
277 for (i = 0; i <= pid->level; i++)
278 free_pidmap(pid->numbers + i);
280 call_rcu(&pid->rcu, delayed_put_pid);
283 struct pid *alloc_pid(struct pid_namespace *ns)
288 struct pid_namespace *tmp;
291 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
296 for (i = ns->level; i >= 0; i--) {
297 nr = alloc_pidmap(tmp);
301 pid->numbers[i].nr = nr;
302 pid->numbers[i].ns = tmp;
307 pid->level = ns->level;
308 atomic_set(&pid->count, 1);
309 for (type = 0; type < PIDTYPE_MAX; ++type)
310 INIT_HLIST_HEAD(&pid->tasks[type]);
312 upid = pid->numbers + ns->level;
313 spin_lock_irq(&pidmap_lock);
314 for ( ; upid >= pid->numbers; --upid)
315 hlist_add_head_rcu(&upid->pid_chain,
316 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
317 spin_unlock_irq(&pidmap_lock);
323 while (++i <= ns->level)
324 free_pidmap(pid->numbers + i);
326 kmem_cache_free(ns->pid_cachep, pid);
331 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
333 struct hlist_node *elem;
336 hlist_for_each_entry_rcu(pnr, elem,
337 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
338 if (pnr->nr == nr && pnr->ns == ns)
339 return container_of(pnr, struct pid,
344 EXPORT_SYMBOL_GPL(find_pid_ns);
346 struct pid *find_vpid(int nr)
348 return find_pid_ns(nr, task_active_pid_ns(current));
350 EXPORT_SYMBOL_GPL(find_vpid);
353 * attach_pid() must be called with the tasklist_lock write-held.
355 void attach_pid(struct task_struct *task, enum pid_type type,
358 struct pid_link *link;
360 link = &task->pids[type];
362 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
365 static void __change_pid(struct task_struct *task, enum pid_type type,
368 struct pid_link *link;
372 link = &task->pids[type];
375 hlist_del_rcu(&link->node);
378 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
379 if (!hlist_empty(&pid->tasks[tmp]))
385 void detach_pid(struct task_struct *task, enum pid_type type)
387 __change_pid(task, type, NULL);
390 void change_pid(struct task_struct *task, enum pid_type type,
393 __change_pid(task, type, pid);
394 attach_pid(task, type, pid);
397 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
398 void transfer_pid(struct task_struct *old, struct task_struct *new,
401 new->pids[type].pid = old->pids[type].pid;
402 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
405 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
407 struct task_struct *result = NULL;
409 struct hlist_node *first;
410 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
411 lockdep_tasklist_lock_is_held());
413 result = hlist_entry(first, struct task_struct, pids[(type)].node);
417 EXPORT_SYMBOL(pid_task);
420 * Must be called under rcu_read_lock().
422 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
424 rcu_lockdep_assert(rcu_read_lock_held(),
425 "find_task_by_pid_ns() needs rcu_read_lock()"
427 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
430 struct task_struct *find_task_by_vpid(pid_t vnr)
432 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
435 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
439 if (type != PIDTYPE_PID)
440 task = task->group_leader;
441 pid = get_pid(task->pids[type].pid);
445 EXPORT_SYMBOL_GPL(get_task_pid);
447 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
449 struct task_struct *result;
451 result = pid_task(pid, type);
453 get_task_struct(result);
457 EXPORT_SYMBOL_GPL(get_pid_task);
459 struct pid *find_get_pid(pid_t nr)
464 pid = get_pid(find_vpid(nr));
469 EXPORT_SYMBOL_GPL(find_get_pid);
471 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
476 if (pid && ns->level <= pid->level) {
477 upid = &pid->numbers[ns->level];
483 EXPORT_SYMBOL_GPL(pid_nr_ns);
485 pid_t pid_vnr(struct pid *pid)
487 return pid_nr_ns(pid, task_active_pid_ns(current));
489 EXPORT_SYMBOL_GPL(pid_vnr);
491 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
492 struct pid_namespace *ns)
498 ns = task_active_pid_ns(current);
499 if (likely(pid_alive(task))) {
500 if (type != PIDTYPE_PID)
501 task = task->group_leader;
502 nr = pid_nr_ns(task->pids[type].pid, ns);
508 EXPORT_SYMBOL(__task_pid_nr_ns);
510 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
512 return pid_nr_ns(task_tgid(tsk), ns);
514 EXPORT_SYMBOL(task_tgid_nr_ns);
516 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
518 return ns_of_pid(task_pid(tsk));
520 EXPORT_SYMBOL_GPL(task_active_pid_ns);
523 * Used by proc to find the first pid that is greater than or equal to nr.
525 * If there is a pid at nr this function is exactly the same as find_pid_ns.
527 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
532 pid = find_pid_ns(nr, ns);
535 nr = next_pidmap(ns, nr);
542 * The pid hash table is scaled according to the amount of memory in the
543 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
546 void __init pidhash_init(void)
548 unsigned int i, pidhash_size;
550 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
551 HASH_EARLY | HASH_SMALL,
552 &pidhash_shift, NULL,
554 pidhash_size = 1U << pidhash_shift;
556 for (i = 0; i < pidhash_size; i++)
557 INIT_HLIST_HEAD(&pid_hash[i]);
560 void __init pidmap_init(void)
562 /* bump default and minimum pid_max based on number of cpus */
563 pid_max = min(pid_max_max, max_t(int, pid_max,
564 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
565 pid_max_min = max_t(int, pid_max_min,
566 PIDS_PER_CPU_MIN * num_possible_cpus());
567 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
569 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
570 /* Reserve PID 0. We never call free_pidmap(0) */
571 set_bit(0, init_pid_ns.pidmap[0].page);
572 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
574 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
575 SLAB_HWCACHE_ALIGN | SLAB_PANIC);