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).
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 = {
73 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
76 .nr_hashed = PIDNS_HASH_ADDING,
78 .child_reaper = &init_task,
79 .user_ns = &init_user_ns,
80 .ns.inum = PROC_PID_INIT_INO,
82 .ns.ops = &pidns_operations,
85 EXPORT_SYMBOL_GPL(init_pid_ns);
88 * Note: disable interrupts while the pidmap_lock is held as an
89 * interrupt might come in and do read_lock(&tasklist_lock).
91 * If we don't disable interrupts there is a nasty deadlock between
92 * detach_pid()->free_pid() and another cpu that does
93 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
94 * read_lock(&tasklist_lock);
96 * After we clean up the tasklist_lock and know there are no
97 * irq handlers that take it we can leave the interrupts enabled.
98 * For now it is easier to be safe than to prove it can't happen.
101 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
103 static void free_pidmap(struct upid *upid)
106 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
107 int offset = nr & BITS_PER_PAGE_MASK;
109 clear_bit(offset, map->page);
110 atomic_inc(&map->nr_free);
114 * If we started walking pids at 'base', is 'a' seen before 'b'?
116 static int pid_before(int base, int a, int b)
119 * This is the same as saying
121 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
122 * and that mapping orders 'a' and 'b' with respect to 'base'.
124 return (unsigned)(a - base) < (unsigned)(b - base);
128 * We might be racing with someone else trying to set pid_ns->last_pid
129 * at the pid allocation time (there's also a sysctl for this, but racing
130 * with this one is OK, see comment in kernel/pid_namespace.c about it).
131 * We want the winner to have the "later" value, because if the
132 * "earlier" value prevails, then a pid may get reused immediately.
134 * Since pids rollover, it is not sufficient to just pick the bigger
135 * value. We have to consider where we started counting from.
137 * 'base' is the value of pid_ns->last_pid that we observed when
138 * we started looking for a pid.
140 * 'pid' is the pid that we eventually found.
142 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
145 int last_write = base;
148 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
149 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
152 static int alloc_pidmap(struct pid_namespace *pid_ns)
154 int i, offset, max_scan, pid, last = pid_ns->last_pid;
160 offset = pid & BITS_PER_PAGE_MASK;
161 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
163 * If last_pid points into the middle of the map->page we
164 * want to scan this bitmap block twice, the second time
165 * we start with offset == 0 (or RESERVED_PIDS).
167 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
168 for (i = 0; i <= max_scan; ++i) {
169 if (unlikely(!map->page)) {
170 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
172 * Free the page if someone raced with us
175 spin_lock_irq(&pidmap_lock);
180 spin_unlock_irq(&pidmap_lock);
182 if (unlikely(!map->page))
185 if (likely(atomic_read(&map->nr_free))) {
187 if (!test_and_set_bit(offset, map->page)) {
188 atomic_dec(&map->nr_free);
189 set_last_pid(pid_ns, last, pid);
192 offset = find_next_offset(map, offset);
193 if (offset >= BITS_PER_PAGE)
195 pid = mk_pid(pid_ns, map, offset);
200 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
204 map = &pid_ns->pidmap[0];
205 offset = RESERVED_PIDS;
206 if (unlikely(last == offset))
209 pid = mk_pid(pid_ns, map, offset);
214 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
217 struct pidmap *map, *end;
219 if (last >= PID_MAX_LIMIT)
222 offset = (last + 1) & BITS_PER_PAGE_MASK;
223 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
224 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
225 for (; map < end; map++, offset = 0) {
226 if (unlikely(!map->page))
228 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
229 if (offset < BITS_PER_PAGE)
230 return mk_pid(pid_ns, map, offset);
235 void put_pid(struct pid *pid)
237 struct pid_namespace *ns;
242 ns = pid->numbers[pid->level].ns;
243 if ((atomic_read(&pid->count) == 1) ||
244 atomic_dec_and_test(&pid->count)) {
245 kmem_cache_free(ns->pid_cachep, pid);
249 EXPORT_SYMBOL_GPL(put_pid);
251 static void delayed_put_pid(struct rcu_head *rhp)
253 struct pid *pid = container_of(rhp, struct pid, rcu);
257 void free_pid(struct pid *pid)
259 /* We can be called with write_lock_irq(&tasklist_lock) held */
263 spin_lock_irqsave(&pidmap_lock, flags);
264 for (i = 0; i <= pid->level; i++) {
265 struct upid *upid = pid->numbers + i;
266 struct pid_namespace *ns = upid->ns;
267 hlist_del_rcu(&upid->pid_chain);
268 switch(--ns->nr_hashed) {
271 /* When all that is left in the pid namespace
272 * is the reaper wake up the reaper. The reaper
273 * may be sleeping in zap_pid_ns_processes().
275 wake_up_process(ns->child_reaper);
277 case PIDNS_HASH_ADDING:
278 /* Handle a fork failure of the first process */
279 WARN_ON(ns->child_reaper);
283 schedule_work(&ns->proc_work);
287 spin_unlock_irqrestore(&pidmap_lock, flags);
289 for (i = 0; i <= pid->level; i++)
290 free_pidmap(pid->numbers + i);
292 call_rcu(&pid->rcu, delayed_put_pid);
295 struct pid *alloc_pid(struct pid_namespace *ns)
300 struct pid_namespace *tmp;
302 int retval = -ENOMEM;
304 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
306 return ERR_PTR(retval);
309 pid->level = ns->level;
310 for (i = ns->level; i >= 0; i--) {
311 nr = alloc_pidmap(tmp);
317 pid->numbers[i].nr = nr;
318 pid->numbers[i].ns = tmp;
322 if (unlikely(is_child_reaper(pid))) {
323 if (pid_ns_prepare_proc(ns))
328 atomic_set(&pid->count, 1);
329 for (type = 0; type < PIDTYPE_MAX; ++type)
330 INIT_HLIST_HEAD(&pid->tasks[type]);
332 upid = pid->numbers + ns->level;
333 spin_lock_irq(&pidmap_lock);
334 if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
336 for ( ; upid >= pid->numbers; --upid) {
337 hlist_add_head_rcu(&upid->pid_chain,
338 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
339 upid->ns->nr_hashed++;
341 spin_unlock_irq(&pidmap_lock);
346 spin_unlock_irq(&pidmap_lock);
350 while (++i <= ns->level)
351 free_pidmap(pid->numbers + i);
353 kmem_cache_free(ns->pid_cachep, pid);
354 return ERR_PTR(retval);
357 void disable_pid_allocation(struct pid_namespace *ns)
359 spin_lock_irq(&pidmap_lock);
360 ns->nr_hashed &= ~PIDNS_HASH_ADDING;
361 spin_unlock_irq(&pidmap_lock);
364 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
368 hlist_for_each_entry_rcu(pnr,
369 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
370 if (pnr->nr == nr && pnr->ns == ns)
371 return container_of(pnr, struct pid,
376 EXPORT_SYMBOL_GPL(find_pid_ns);
378 struct pid *find_vpid(int nr)
380 return find_pid_ns(nr, task_active_pid_ns(current));
382 EXPORT_SYMBOL_GPL(find_vpid);
385 * attach_pid() must be called with the tasklist_lock write-held.
387 void attach_pid(struct task_struct *task, enum pid_type type)
389 struct pid_link *link = &task->pids[type];
390 hlist_add_head_rcu(&link->node, &link->pid->tasks[type]);
393 static void __change_pid(struct task_struct *task, enum pid_type type,
396 struct pid_link *link;
400 link = &task->pids[type];
403 hlist_del_rcu(&link->node);
406 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
407 if (!hlist_empty(&pid->tasks[tmp]))
413 void detach_pid(struct task_struct *task, enum pid_type type)
415 __change_pid(task, type, NULL);
418 void change_pid(struct task_struct *task, enum pid_type type,
421 __change_pid(task, type, pid);
422 attach_pid(task, type);
425 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
426 void transfer_pid(struct task_struct *old, struct task_struct *new,
429 new->pids[type].pid = old->pids[type].pid;
430 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
433 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
435 struct task_struct *result = NULL;
437 struct hlist_node *first;
438 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
439 lockdep_tasklist_lock_is_held());
441 result = hlist_entry(first, struct task_struct, pids[(type)].node);
445 EXPORT_SYMBOL(pid_task);
448 * Must be called under rcu_read_lock().
450 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
452 RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
453 "find_task_by_pid_ns() needs rcu_read_lock() protection");
454 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
457 struct task_struct *find_task_by_vpid(pid_t vnr)
459 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
462 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
466 if (type != PIDTYPE_PID)
467 task = task->group_leader;
468 pid = get_pid(rcu_dereference(task->pids[type].pid));
472 EXPORT_SYMBOL_GPL(get_task_pid);
474 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
476 struct task_struct *result;
478 result = pid_task(pid, type);
480 get_task_struct(result);
484 EXPORT_SYMBOL_GPL(get_pid_task);
486 struct pid *find_get_pid(pid_t nr)
491 pid = get_pid(find_vpid(nr));
496 EXPORT_SYMBOL_GPL(find_get_pid);
498 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
503 if (pid && ns->level <= pid->level) {
504 upid = &pid->numbers[ns->level];
510 EXPORT_SYMBOL_GPL(pid_nr_ns);
512 pid_t pid_vnr(struct pid *pid)
514 return pid_nr_ns(pid, task_active_pid_ns(current));
516 EXPORT_SYMBOL_GPL(pid_vnr);
518 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
519 struct pid_namespace *ns)
525 ns = task_active_pid_ns(current);
526 if (likely(pid_alive(task))) {
527 if (type != PIDTYPE_PID)
528 task = task->group_leader;
529 nr = pid_nr_ns(rcu_dereference(task->pids[type].pid), ns);
535 EXPORT_SYMBOL(__task_pid_nr_ns);
537 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
539 return pid_nr_ns(task_tgid(tsk), ns);
541 EXPORT_SYMBOL(task_tgid_nr_ns);
543 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
545 return ns_of_pid(task_pid(tsk));
547 EXPORT_SYMBOL_GPL(task_active_pid_ns);
550 * Used by proc to find the first pid that is greater than or equal to nr.
552 * If there is a pid at nr this function is exactly the same as find_pid_ns.
554 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
559 pid = find_pid_ns(nr, ns);
562 nr = next_pidmap(ns, nr);
569 * The pid hash table is scaled according to the amount of memory in the
570 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
573 void __init pidhash_init(void)
575 unsigned int i, pidhash_size;
577 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
578 HASH_EARLY | HASH_SMALL,
579 &pidhash_shift, NULL,
581 pidhash_size = 1U << pidhash_shift;
583 for (i = 0; i < pidhash_size; i++)
584 INIT_HLIST_HEAD(&pid_hash[i]);
587 void __init pidmap_init(void)
589 /* Verify no one has done anything silly: */
590 BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_HASH_ADDING);
592 /* bump default and minimum pid_max based on number of cpus */
593 pid_max = min(pid_max_max, max_t(int, pid_max,
594 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
595 pid_max_min = max_t(int, pid_max_min,
596 PIDS_PER_CPU_MIN * num_possible_cpus());
597 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
599 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
600 /* Reserve PID 0. We never call free_pidmap(0) */
601 set_bit(0, init_pid_ns.pidmap[0].page);
602 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
604 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
605 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);