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;
45 int pid_max = PID_MAX_DEFAULT;
47 #define RESERVED_PIDS 300
49 int pid_max_min = RESERVED_PIDS + 1;
50 int pid_max_max = PID_MAX_LIMIT;
52 #define BITS_PER_PAGE (PAGE_SIZE*8)
53 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
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 }
79 .child_reaper = &init_task,
81 EXPORT_SYMBOL_GPL(init_pid_ns);
83 int is_container_init(struct task_struct *tsk)
90 if (pid != NULL && pid->numbers[pid->level].nr == 1)
96 EXPORT_SYMBOL(is_container_init);
99 * Note: disable interrupts while the pidmap_lock is held as an
100 * interrupt might come in and do read_lock(&tasklist_lock).
102 * If we don't disable interrupts there is a nasty deadlock between
103 * detach_pid()->free_pid() and another cpu that does
104 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
105 * read_lock(&tasklist_lock);
107 * After we clean up the tasklist_lock and know there are no
108 * irq handlers that take it we can leave the interrupts enabled.
109 * For now it is easier to be safe than to prove it can't happen.
112 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
114 static void free_pidmap(struct upid *upid)
117 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
118 int offset = nr & 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 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 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 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);
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 atomic_set(&pid->count, 1);
267 for (type = 0; type < PIDTYPE_MAX; ++type)
268 INIT_HLIST_HEAD(&pid->tasks[type]);
270 spin_lock_irq(&pidmap_lock);
271 for (i = ns->level; i >= 0; i--) {
272 upid = &pid->numbers[i];
273 hlist_add_head_rcu(&upid->pid_chain,
274 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
276 spin_unlock_irq(&pidmap_lock);
282 while (++i <= ns->level)
283 free_pidmap(pid->numbers + i);
285 kmem_cache_free(ns->pid_cachep, pid);
290 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
292 struct hlist_node *elem;
295 hlist_for_each_entry_rcu(pnr, elem,
296 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
297 if (pnr->nr == nr && pnr->ns == ns)
298 return container_of(pnr, struct pid,
303 EXPORT_SYMBOL_GPL(find_pid_ns);
305 struct pid *find_vpid(int nr)
307 return find_pid_ns(nr, current->nsproxy->pid_ns);
309 EXPORT_SYMBOL_GPL(find_vpid);
311 struct pid *find_pid(int nr)
313 return find_pid_ns(nr, &init_pid_ns);
315 EXPORT_SYMBOL_GPL(find_pid);
318 * attach_pid() must be called with the tasklist_lock write-held.
320 void attach_pid(struct task_struct *task, enum pid_type type,
323 struct pid_link *link;
325 link = &task->pids[type];
327 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
330 static void __change_pid(struct task_struct *task, enum pid_type type,
333 struct pid_link *link;
337 link = &task->pids[type];
340 hlist_del_rcu(&link->node);
343 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
344 if (!hlist_empty(&pid->tasks[tmp]))
350 void detach_pid(struct task_struct *task, enum pid_type type)
352 __change_pid(task, type, NULL);
355 void change_pid(struct task_struct *task, enum pid_type type,
358 __change_pid(task, type, pid);
359 attach_pid(task, type, pid);
362 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
363 void transfer_pid(struct task_struct *old, struct task_struct *new,
366 new->pids[type].pid = old->pids[type].pid;
367 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
370 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
372 struct task_struct *result = NULL;
374 struct hlist_node *first;
375 first = rcu_dereference(pid->tasks[type].first);
377 result = hlist_entry(first, struct task_struct, pids[(type)].node);
381 EXPORT_SYMBOL(pid_task);
384 * Must be called under rcu_read_lock() or with tasklist_lock read-held.
386 struct task_struct *find_task_by_pid_type_ns(int type, int nr,
387 struct pid_namespace *ns)
389 return pid_task(find_pid_ns(nr, ns), type);
392 EXPORT_SYMBOL(find_task_by_pid_type_ns);
394 struct task_struct *find_task_by_vpid(pid_t vnr)
396 return find_task_by_pid_type_ns(PIDTYPE_PID, vnr,
397 current->nsproxy->pid_ns);
399 EXPORT_SYMBOL(find_task_by_vpid);
401 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
403 return find_task_by_pid_type_ns(PIDTYPE_PID, nr, ns);
405 EXPORT_SYMBOL(find_task_by_pid_ns);
407 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
411 pid = get_pid(task->pids[type].pid);
416 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
418 struct task_struct *result;
420 result = pid_task(pid, type);
422 get_task_struct(result);
427 struct pid *find_get_pid(pid_t nr)
432 pid = get_pid(find_vpid(nr));
438 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
443 if (pid && ns->level <= pid->level) {
444 upid = &pid->numbers[ns->level];
451 pid_t pid_vnr(struct pid *pid)
453 return pid_nr_ns(pid, current->nsproxy->pid_ns);
455 EXPORT_SYMBOL_GPL(pid_vnr);
457 pid_t task_pid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
459 return pid_nr_ns(task_pid(tsk), ns);
461 EXPORT_SYMBOL(task_pid_nr_ns);
463 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
465 return pid_nr_ns(task_tgid(tsk), ns);
467 EXPORT_SYMBOL(task_tgid_nr_ns);
469 pid_t task_pgrp_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
471 return pid_nr_ns(task_pgrp(tsk), ns);
473 EXPORT_SYMBOL(task_pgrp_nr_ns);
475 pid_t task_session_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
477 return pid_nr_ns(task_session(tsk), ns);
479 EXPORT_SYMBOL(task_session_nr_ns);
482 * Used by proc to find the first pid that is greater then or equal to nr.
484 * If there is a pid at nr this function is exactly the same as find_pid.
486 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
491 pid = find_pid_ns(nr, ns);
494 nr = next_pidmap(ns, nr);
499 EXPORT_SYMBOL_GPL(find_get_pid);
502 * The pid hash table is scaled according to the amount of memory in the
503 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
506 void __init pidhash_init(void)
509 unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
511 pidhash_shift = max(4, fls(megabytes * 4));
512 pidhash_shift = min(12, pidhash_shift);
513 pidhash_size = 1 << pidhash_shift;
515 printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
516 pidhash_size, pidhash_shift,
517 pidhash_size * sizeof(struct hlist_head));
519 pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash)));
521 panic("Could not alloc pidhash!\n");
522 for (i = 0; i < pidhash_size; i++)
523 INIT_HLIST_HEAD(&pid_hash[i]);
526 void __init pidmap_init(void)
528 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
529 /* Reserve PID 0. We never call free_pidmap(0) */
530 set_bit(0, init_pid_ns.pidmap[0].page);
531 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
533 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
534 SLAB_HWCACHE_ALIGN | SLAB_PANIC);