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 .ns.inum = PROC_PID_INIT_INO,
84 .ns.ops = &pidns_operations,
87 EXPORT_SYMBOL_GPL(init_pid_ns);
90 * Note: disable interrupts while the pidmap_lock is held as an
91 * interrupt might come in and do read_lock(&tasklist_lock).
93 * If we don't disable interrupts there is a nasty deadlock between
94 * detach_pid()->free_pid() and another cpu that does
95 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
96 * read_lock(&tasklist_lock);
98 * After we clean up the tasklist_lock and know there are no
99 * irq handlers that take it we can leave the interrupts enabled.
100 * For now it is easier to be safe than to prove it can't happen.
103 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
105 static void free_pidmap(struct upid *upid)
108 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
109 int offset = nr & BITS_PER_PAGE_MASK;
111 clear_bit(offset, map->page);
112 atomic_inc(&map->nr_free);
116 * If we started walking pids at 'base', is 'a' seen before 'b'?
118 static int pid_before(int base, int a, int b)
121 * This is the same as saying
123 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
124 * and that mapping orders 'a' and 'b' with respect to 'base'.
126 return (unsigned)(a - base) < (unsigned)(b - base);
130 * We might be racing with someone else trying to set pid_ns->last_pid
131 * at the pid allocation time (there's also a sysctl for this, but racing
132 * with this one is OK, see comment in kernel/pid_namespace.c about it).
133 * We want the winner to have the "later" value, because if the
134 * "earlier" value prevails, then a pid may get reused immediately.
136 * Since pids rollover, it is not sufficient to just pick the bigger
137 * value. We have to consider where we started counting from.
139 * 'base' is the value of pid_ns->last_pid that we observed when
140 * we started looking for a pid.
142 * 'pid' is the pid that we eventually found.
144 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
147 int last_write = base;
150 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
151 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
154 static int alloc_pidmap(struct pid_namespace *pid_ns)
156 int i, offset, max_scan, pid, last = pid_ns->last_pid;
162 offset = pid & BITS_PER_PAGE_MASK;
163 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
165 * If last_pid points into the middle of the map->page we
166 * want to scan this bitmap block twice, the second time
167 * we start with offset == 0 (or RESERVED_PIDS).
169 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
170 for (i = 0; i <= max_scan; ++i) {
171 if (unlikely(!map->page)) {
172 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
174 * Free the page if someone raced with us
177 spin_lock_irq(&pidmap_lock);
182 spin_unlock_irq(&pidmap_lock);
184 if (unlikely(!map->page))
187 if (likely(atomic_read(&map->nr_free))) {
189 if (!test_and_set_bit(offset, map->page)) {
190 atomic_dec(&map->nr_free);
191 set_last_pid(pid_ns, last, pid);
194 offset = find_next_offset(map, offset);
195 if (offset >= BITS_PER_PAGE)
197 pid = mk_pid(pid_ns, map, offset);
202 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
206 map = &pid_ns->pidmap[0];
207 offset = RESERVED_PIDS;
208 if (unlikely(last == offset))
211 pid = mk_pid(pid_ns, map, offset);
216 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
219 struct pidmap *map, *end;
221 if (last >= PID_MAX_LIMIT)
224 offset = (last + 1) & BITS_PER_PAGE_MASK;
225 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
226 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
227 for (; map < end; map++, offset = 0) {
228 if (unlikely(!map->page))
230 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
231 if (offset < BITS_PER_PAGE)
232 return mk_pid(pid_ns, map, offset);
237 void put_pid(struct pid *pid)
239 struct pid_namespace *ns;
244 ns = pid->numbers[pid->level].ns;
245 if ((atomic_read(&pid->count) == 1) ||
246 atomic_dec_and_test(&pid->count)) {
247 kmem_cache_free(ns->pid_cachep, pid);
251 EXPORT_SYMBOL_GPL(put_pid);
253 static void delayed_put_pid(struct rcu_head *rhp)
255 struct pid *pid = container_of(rhp, struct pid, rcu);
259 void free_pid(struct pid *pid)
261 /* We can be called with write_lock_irq(&tasklist_lock) held */
265 spin_lock_irqsave(&pidmap_lock, flags);
266 for (i = 0; i <= pid->level; i++) {
267 struct upid *upid = pid->numbers + i;
268 struct pid_namespace *ns = upid->ns;
269 hlist_del_rcu(&upid->pid_chain);
270 switch(--ns->nr_hashed) {
273 /* When all that is left in the pid namespace
274 * is the reaper wake up the reaper. The reaper
275 * may be sleeping in zap_pid_ns_processes().
277 wake_up_process(ns->child_reaper);
279 case PIDNS_HASH_ADDING:
280 /* Handle a fork failure of the first process */
281 WARN_ON(ns->child_reaper);
285 schedule_work(&ns->proc_work);
289 spin_unlock_irqrestore(&pidmap_lock, flags);
291 for (i = 0; i <= pid->level; i++)
292 free_pidmap(pid->numbers + i);
294 call_rcu(&pid->rcu, delayed_put_pid);
297 struct pid *alloc_pid(struct pid_namespace *ns)
302 struct pid_namespace *tmp;
304 int retval = -ENOMEM;
306 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
308 return ERR_PTR(retval);
311 pid->level = ns->level;
312 for (i = ns->level; i >= 0; i--) {
313 nr = alloc_pidmap(tmp);
314 if (IS_ERR_VALUE(nr)) {
319 pid->numbers[i].nr = nr;
320 pid->numbers[i].ns = tmp;
324 if (unlikely(is_child_reaper(pid))) {
325 if (pid_ns_prepare_proc(ns))
330 atomic_set(&pid->count, 1);
331 for (type = 0; type < PIDTYPE_MAX; ++type)
332 INIT_HLIST_HEAD(&pid->tasks[type]);
334 upid = pid->numbers + ns->level;
335 spin_lock_irq(&pidmap_lock);
336 if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
338 for ( ; upid >= pid->numbers; --upid) {
339 hlist_add_head_rcu(&upid->pid_chain,
340 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
341 upid->ns->nr_hashed++;
343 spin_unlock_irq(&pidmap_lock);
348 spin_unlock_irq(&pidmap_lock);
352 while (++i <= ns->level)
353 free_pidmap(pid->numbers + i);
355 kmem_cache_free(ns->pid_cachep, pid);
356 return ERR_PTR(retval);
359 void disable_pid_allocation(struct pid_namespace *ns)
361 spin_lock_irq(&pidmap_lock);
362 ns->nr_hashed &= ~PIDNS_HASH_ADDING;
363 spin_unlock_irq(&pidmap_lock);
366 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
370 hlist_for_each_entry_rcu(pnr,
371 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
372 if (pnr->nr == nr && pnr->ns == ns)
373 return container_of(pnr, struct pid,
378 EXPORT_SYMBOL_GPL(find_pid_ns);
380 struct pid *find_vpid(int nr)
382 return find_pid_ns(nr, task_active_pid_ns(current));
384 EXPORT_SYMBOL_GPL(find_vpid);
387 * attach_pid() must be called with the tasklist_lock write-held.
389 void attach_pid(struct task_struct *task, enum pid_type type)
391 struct pid_link *link = &task->pids[type];
392 hlist_add_head_rcu(&link->node, &link->pid->tasks[type]);
395 static void __change_pid(struct task_struct *task, enum pid_type type,
398 struct pid_link *link;
402 link = &task->pids[type];
405 hlist_del_rcu(&link->node);
408 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
409 if (!hlist_empty(&pid->tasks[tmp]))
415 void detach_pid(struct task_struct *task, enum pid_type type)
417 __change_pid(task, type, NULL);
420 void change_pid(struct task_struct *task, enum pid_type type,
423 __change_pid(task, type, pid);
424 attach_pid(task, type);
427 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
428 void transfer_pid(struct task_struct *old, struct task_struct *new,
431 new->pids[type].pid = old->pids[type].pid;
432 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
435 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
437 struct task_struct *result = NULL;
439 struct hlist_node *first;
440 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
441 lockdep_tasklist_lock_is_held());
443 result = hlist_entry(first, struct task_struct, pids[(type)].node);
447 EXPORT_SYMBOL(pid_task);
450 * Must be called under rcu_read_lock().
452 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
454 RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
455 "find_task_by_pid_ns() needs rcu_read_lock() protection");
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(rcu_dereference(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(rcu_dereference(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 /* Veryify no one has done anything silly */
592 BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_HASH_ADDING);
594 /* bump default and minimum pid_max based on number of cpus */
595 pid_max = min(pid_max_max, max_t(int, pid_max,
596 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
597 pid_max_min = max_t(int, pid_max_min,
598 PIDS_PER_CPU_MIN * num_possible_cpus());
599 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
601 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
602 /* Reserve PID 0. We never call free_pidmap(0) */
603 set_bit(0, init_pid_ns.pidmap[0].page);
604 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
606 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
607 SLAB_HWCACHE_ALIGN | SLAB_PANIC);