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_fs.h>
41 #define pid_hashfn(nr, ns) \
42 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
43 static struct hlist_head *pid_hash;
44 static unsigned int pidhash_shift = 4;
45 struct pid init_struct_pid = INIT_STRUCT_PID;
47 int pid_max = PID_MAX_DEFAULT;
49 #define RESERVED_PIDS 300
51 int pid_max_min = RESERVED_PIDS + 1;
52 int pid_max_max = PID_MAX_LIMIT;
54 #define BITS_PER_PAGE (PAGE_SIZE*8)
55 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
57 static inline int mk_pid(struct pid_namespace *pid_ns,
58 struct pidmap *map, int off)
60 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
63 #define find_next_offset(map, off) \
64 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
67 * PID-map pages start out as NULL, they get allocated upon
68 * first use and are never deallocated. This way a low pid_max
69 * value does not cause lots of bitmaps to be allocated, but
70 * the scheme scales to up to 4 million PIDs, runtime.
72 struct pid_namespace init_pid_ns = {
74 .refcount = ATOMIC_INIT(2),
77 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
81 .child_reaper = &init_task,
82 .user_ns = &init_user_ns,
83 .proc_inum = PROC_PID_INIT_INO,
85 EXPORT_SYMBOL_GPL(init_pid_ns);
87 int is_container_init(struct task_struct *tsk)
94 if (pid != NULL && pid->numbers[pid->level].nr == 1)
100 EXPORT_SYMBOL(is_container_init);
103 * Note: disable interrupts while the pidmap_lock is held as an
104 * interrupt might come in and do read_lock(&tasklist_lock).
106 * If we don't disable interrupts there is a nasty deadlock between
107 * detach_pid()->free_pid() and another cpu that does
108 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
109 * read_lock(&tasklist_lock);
111 * After we clean up the tasklist_lock and know there are no
112 * irq handlers that take it we can leave the interrupts enabled.
113 * For now it is easier to be safe than to prove it can't happen.
116 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
118 static void free_pidmap(struct upid *upid)
121 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
122 int offset = nr & BITS_PER_PAGE_MASK;
124 clear_bit(offset, map->page);
125 atomic_inc(&map->nr_free);
129 * If we started walking pids at 'base', is 'a' seen before 'b'?
131 static int pid_before(int base, int a, int b)
134 * This is the same as saying
136 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
137 * and that mapping orders 'a' and 'b' with respect to 'base'.
139 return (unsigned)(a - base) < (unsigned)(b - base);
143 * We might be racing with someone else trying to set pid_ns->last_pid
144 * at the pid allocation time (there's also a sysctl for this, but racing
145 * with this one is OK, see comment in kernel/pid_namespace.c about it).
146 * We want the winner to have the "later" value, because if the
147 * "earlier" value prevails, then a pid may get reused immediately.
149 * Since pids rollover, it is not sufficient to just pick the bigger
150 * value. We have to consider where we started counting from.
152 * 'base' is the value of pid_ns->last_pid that we observed when
153 * we started looking for a pid.
155 * 'pid' is the pid that we eventually found.
157 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
160 int last_write = base;
163 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
164 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
167 static int alloc_pidmap(struct pid_namespace *pid_ns)
169 int i, offset, max_scan, pid, last = pid_ns->last_pid;
175 offset = pid & BITS_PER_PAGE_MASK;
176 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
178 * If last_pid points into the middle of the map->page we
179 * want to scan this bitmap block twice, the second time
180 * we start with offset == 0 (or RESERVED_PIDS).
182 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
183 for (i = 0; i <= max_scan; ++i) {
184 if (unlikely(!map->page)) {
185 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
187 * Free the page if someone raced with us
190 spin_lock_irq(&pidmap_lock);
195 spin_unlock_irq(&pidmap_lock);
197 if (unlikely(!map->page))
200 if (likely(atomic_read(&map->nr_free))) {
202 if (!test_and_set_bit(offset, map->page)) {
203 atomic_dec(&map->nr_free);
204 set_last_pid(pid_ns, last, pid);
207 offset = find_next_offset(map, offset);
208 pid = mk_pid(pid_ns, map, offset);
209 } while (offset < BITS_PER_PAGE && pid < pid_max);
211 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
215 map = &pid_ns->pidmap[0];
216 offset = RESERVED_PIDS;
217 if (unlikely(last == offset))
220 pid = mk_pid(pid_ns, map, offset);
225 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
228 struct pidmap *map, *end;
230 if (last >= PID_MAX_LIMIT)
233 offset = (last + 1) & BITS_PER_PAGE_MASK;
234 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
235 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
236 for (; map < end; map++, offset = 0) {
237 if (unlikely(!map->page))
239 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
240 if (offset < BITS_PER_PAGE)
241 return mk_pid(pid_ns, map, offset);
246 void put_pid(struct pid *pid)
248 struct pid_namespace *ns;
253 ns = pid->numbers[pid->level].ns;
254 if ((atomic_read(&pid->count) == 1) ||
255 atomic_dec_and_test(&pid->count)) {
256 kmem_cache_free(ns->pid_cachep, pid);
260 EXPORT_SYMBOL_GPL(put_pid);
262 static void delayed_put_pid(struct rcu_head *rhp)
264 struct pid *pid = container_of(rhp, struct pid, rcu);
268 void free_pid(struct pid *pid)
270 /* We can be called with write_lock_irq(&tasklist_lock) held */
274 spin_lock_irqsave(&pidmap_lock, flags);
275 for (i = 0; i <= pid->level; i++) {
276 struct upid *upid = pid->numbers + i;
277 struct pid_namespace *ns = upid->ns;
278 hlist_del_rcu(&upid->pid_chain);
279 switch(--ns->nr_hashed) {
281 /* When all that is left in the pid namespace
282 * is the reaper wake up the reaper. The reaper
283 * may be sleeping in zap_pid_ns_processes().
285 wake_up_process(ns->child_reaper);
289 schedule_work(&ns->proc_work);
293 spin_unlock_irqrestore(&pidmap_lock, flags);
295 for (i = 0; i <= pid->level; i++)
296 free_pidmap(pid->numbers + i);
298 call_rcu(&pid->rcu, delayed_put_pid);
301 struct pid *alloc_pid(struct pid_namespace *ns)
306 struct pid_namespace *tmp;
309 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
314 pid->level = ns->level;
315 for (i = ns->level; i >= 0; i--) {
316 nr = alloc_pidmap(tmp);
320 pid->numbers[i].nr = nr;
321 pid->numbers[i].ns = tmp;
325 if (unlikely(is_child_reaper(pid))) {
326 if (pid_ns_prepare_proc(ns))
331 atomic_set(&pid->count, 1);
332 for (type = 0; type < PIDTYPE_MAX; ++type)
333 INIT_HLIST_HEAD(&pid->tasks[type]);
335 upid = pid->numbers + ns->level;
336 spin_lock_irq(&pidmap_lock);
337 if (ns->nr_hashed < 0)
339 for ( ; upid >= pid->numbers; --upid) {
340 hlist_add_head_rcu(&upid->pid_chain,
341 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
342 upid->ns->nr_hashed++;
344 spin_unlock_irq(&pidmap_lock);
350 spin_unlock(&pidmap_lock);
352 while (++i <= ns->level)
353 free_pidmap(pid->numbers + i);
355 kmem_cache_free(ns->pid_cachep, pid);
360 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
362 struct hlist_node *elem;
365 hlist_for_each_entry_rcu(pnr, elem,
366 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
367 if (pnr->nr == nr && pnr->ns == ns)
368 return container_of(pnr, struct pid,
373 EXPORT_SYMBOL_GPL(find_pid_ns);
375 struct pid *find_vpid(int nr)
377 return find_pid_ns(nr, task_active_pid_ns(current));
379 EXPORT_SYMBOL_GPL(find_vpid);
382 * attach_pid() must be called with the tasklist_lock write-held.
384 void attach_pid(struct task_struct *task, enum pid_type type,
387 struct pid_link *link;
389 link = &task->pids[type];
391 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
394 static void __change_pid(struct task_struct *task, enum pid_type type,
397 struct pid_link *link;
401 link = &task->pids[type];
404 hlist_del_rcu(&link->node);
407 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
408 if (!hlist_empty(&pid->tasks[tmp]))
414 void detach_pid(struct task_struct *task, enum pid_type type)
416 __change_pid(task, type, NULL);
419 void change_pid(struct task_struct *task, enum pid_type type,
422 __change_pid(task, type, pid);
423 attach_pid(task, type, pid);
426 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
427 void transfer_pid(struct task_struct *old, struct task_struct *new,
430 new->pids[type].pid = old->pids[type].pid;
431 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
434 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
436 struct task_struct *result = NULL;
438 struct hlist_node *first;
439 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
440 lockdep_tasklist_lock_is_held());
442 result = hlist_entry(first, struct task_struct, pids[(type)].node);
446 EXPORT_SYMBOL(pid_task);
449 * Must be called under rcu_read_lock().
451 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
453 rcu_lockdep_assert(rcu_read_lock_held(),
454 "find_task_by_pid_ns() needs rcu_read_lock()"
456 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
459 struct task_struct *find_task_by_vpid(pid_t vnr)
461 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
464 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
468 if (type != PIDTYPE_PID)
469 task = task->group_leader;
470 pid = get_pid(task->pids[type].pid);
474 EXPORT_SYMBOL_GPL(get_task_pid);
476 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
478 struct task_struct *result;
480 result = pid_task(pid, type);
482 get_task_struct(result);
486 EXPORT_SYMBOL_GPL(get_pid_task);
488 struct pid *find_get_pid(pid_t nr)
493 pid = get_pid(find_vpid(nr));
498 EXPORT_SYMBOL_GPL(find_get_pid);
500 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
505 if (pid && ns->level <= pid->level) {
506 upid = &pid->numbers[ns->level];
512 EXPORT_SYMBOL_GPL(pid_nr_ns);
514 pid_t pid_vnr(struct pid *pid)
516 return pid_nr_ns(pid, task_active_pid_ns(current));
518 EXPORT_SYMBOL_GPL(pid_vnr);
520 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
521 struct pid_namespace *ns)
527 ns = task_active_pid_ns(current);
528 if (likely(pid_alive(task))) {
529 if (type != PIDTYPE_PID)
530 task = task->group_leader;
531 nr = pid_nr_ns(task->pids[type].pid, ns);
537 EXPORT_SYMBOL(__task_pid_nr_ns);
539 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
541 return pid_nr_ns(task_tgid(tsk), ns);
543 EXPORT_SYMBOL(task_tgid_nr_ns);
545 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
547 return ns_of_pid(task_pid(tsk));
549 EXPORT_SYMBOL_GPL(task_active_pid_ns);
552 * Used by proc to find the first pid that is greater than or equal to nr.
554 * If there is a pid at nr this function is exactly the same as find_pid_ns.
556 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
561 pid = find_pid_ns(nr, ns);
564 nr = next_pidmap(ns, nr);
571 * The pid hash table is scaled according to the amount of memory in the
572 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
575 void __init pidhash_init(void)
577 unsigned int i, pidhash_size;
579 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
580 HASH_EARLY | HASH_SMALL,
581 &pidhash_shift, NULL,
583 pidhash_size = 1U << pidhash_shift;
585 for (i = 0; i < pidhash_size; i++)
586 INIT_HLIST_HEAD(&pid_hash[i]);
589 void __init pidmap_init(void)
591 /* bump default and minimum pid_max based on number of cpus */
592 pid_max = min(pid_max_max, max_t(int, pid_max,
593 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
594 pid_max_min = max_t(int, pid_max_min,
595 PIDS_PER_CPU_MIN * num_possible_cpus());
596 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
598 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
599 /* Reserve PID 0. We never call free_pidmap(0) */
600 set_bit(0, init_pid_ns.pidmap[0].page);
601 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
602 init_pid_ns.nr_hashed = 1;
604 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
605 SLAB_HWCACHE_ALIGN | SLAB_PANIC);