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>
38 #define pid_hashfn(nr, ns) \
39 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
40 static struct hlist_head *pid_hash;
41 static int pidhash_shift;
42 struct pid init_struct_pid = INIT_STRUCT_PID;
43 static struct kmem_cache *pid_ns_cachep;
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 fastcall void free_pidmap(struct pid_namespace *pid_ns, int pid)
116 struct pidmap *map = pid_ns->pidmap + pid / BITS_PER_PAGE;
117 int offset = pid & BITS_PER_PAGE_MASK;
119 clear_bit(offset, map->page);
120 atomic_inc(&map->nr_free);
123 static int alloc_pidmap(struct pid_namespace *pid_ns)
125 int i, offset, max_scan, pid, last = pid_ns->last_pid;
131 offset = pid & BITS_PER_PAGE_MASK;
132 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
133 max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
134 for (i = 0; i <= max_scan; ++i) {
135 if (unlikely(!map->page)) {
136 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
138 * Free the page if someone raced with us
141 spin_lock_irq(&pidmap_lock);
146 spin_unlock_irq(&pidmap_lock);
147 if (unlikely(!map->page))
150 if (likely(atomic_read(&map->nr_free))) {
152 if (!test_and_set_bit(offset, map->page)) {
153 atomic_dec(&map->nr_free);
154 pid_ns->last_pid = pid;
157 offset = find_next_offset(map, offset);
158 pid = mk_pid(pid_ns, map, offset);
160 * find_next_offset() found a bit, the pid from it
161 * is in-bounds, and if we fell back to the last
162 * bitmap block and the final block was the same
163 * as the starting point, pid is before last_pid.
165 } while (offset < BITS_PER_PAGE && pid < pid_max &&
166 (i != max_scan || pid < last ||
167 !((last+1) & BITS_PER_PAGE_MASK)));
169 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
173 map = &pid_ns->pidmap[0];
174 offset = RESERVED_PIDS;
175 if (unlikely(last == offset))
178 pid = mk_pid(pid_ns, map, offset);
183 static int next_pidmap(struct pid_namespace *pid_ns, int last)
186 struct pidmap *map, *end;
188 offset = (last + 1) & BITS_PER_PAGE_MASK;
189 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
190 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
191 for (; map < end; map++, offset = 0) {
192 if (unlikely(!map->page))
194 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
195 if (offset < BITS_PER_PAGE)
196 return mk_pid(pid_ns, map, offset);
201 fastcall void put_pid(struct pid *pid)
203 struct pid_namespace *ns;
208 ns = pid->numbers[pid->level].ns;
209 if ((atomic_read(&pid->count) == 1) ||
210 atomic_dec_and_test(&pid->count)) {
211 kmem_cache_free(ns->pid_cachep, pid);
215 EXPORT_SYMBOL_GPL(put_pid);
217 static void delayed_put_pid(struct rcu_head *rhp)
219 struct pid *pid = container_of(rhp, struct pid, rcu);
223 fastcall void free_pid(struct pid *pid)
225 /* We can be called with write_lock_irq(&tasklist_lock) held */
229 spin_lock_irqsave(&pidmap_lock, flags);
230 for (i = 0; i <= pid->level; i++)
231 hlist_del_rcu(&pid->numbers[i].pid_chain);
232 spin_unlock_irqrestore(&pidmap_lock, flags);
234 for (i = 0; i <= pid->level; i++)
235 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
237 call_rcu(&pid->rcu, delayed_put_pid);
240 struct pid *alloc_pid(struct pid_namespace *ns)
245 struct pid_namespace *tmp;
248 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
253 for (i = ns->level; i >= 0; i--) {
254 nr = alloc_pidmap(tmp);
258 pid->numbers[i].nr = nr;
259 pid->numbers[i].ns = tmp;
264 pid->level = ns->level;
265 pid->nr = pid->numbers[0].nr;
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 for (i++; i <= ns->level; i++)
283 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
285 kmem_cache_free(ns->pid_cachep, pid);
290 struct pid * fastcall 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);
306 * attach_pid() must be called with the tasklist_lock write-held.
308 int fastcall attach_pid(struct task_struct *task, enum pid_type type,
311 struct pid_link *link;
313 link = &task->pids[type];
315 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
320 void fastcall detach_pid(struct task_struct *task, enum pid_type type)
322 struct pid_link *link;
326 link = &task->pids[type];
329 hlist_del_rcu(&link->node);
332 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
333 if (!hlist_empty(&pid->tasks[tmp]))
339 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
340 void fastcall transfer_pid(struct task_struct *old, struct task_struct *new,
343 new->pids[type].pid = old->pids[type].pid;
344 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
345 old->pids[type].pid = NULL;
348 struct task_struct * fastcall pid_task(struct pid *pid, enum pid_type type)
350 struct task_struct *result = NULL;
352 struct hlist_node *first;
353 first = rcu_dereference(pid->tasks[type].first);
355 result = hlist_entry(first, struct task_struct, pids[(type)].node);
361 * Must be called under rcu_read_lock() or with tasklist_lock read-held.
363 struct task_struct *find_task_by_pid_type_ns(int type, int nr,
364 struct pid_namespace *ns)
366 return pid_task(find_pid_ns(nr, ns), type);
369 EXPORT_SYMBOL(find_task_by_pid_type_ns);
371 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
375 pid = get_pid(task->pids[type].pid);
380 struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
382 struct task_struct *result;
384 result = pid_task(pid, type);
386 get_task_struct(result);
391 struct pid *find_get_pid(pid_t nr)
396 pid = get_pid(find_vpid(nr));
402 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
407 if (pid && ns->level <= pid->level) {
408 upid = &pid->numbers[ns->level];
416 * Used by proc to find the first pid that is greater then or equal to nr.
418 * If there is a pid at nr this function is exactly the same as find_pid.
420 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
425 pid = find_pid_ns(nr, ns);
428 nr = next_pidmap(ns, nr);
433 EXPORT_SYMBOL_GPL(find_get_pid);
438 struct kmem_cache *cachep;
439 struct list_head list;
442 static LIST_HEAD(pid_caches_lh);
443 static DEFINE_MUTEX(pid_caches_mutex);
446 * creates the kmem cache to allocate pids from.
447 * @nr_ids: the number of numerical ids this pid will have to carry
450 static struct kmem_cache *create_pid_cachep(int nr_ids)
452 struct pid_cache *pcache;
453 struct kmem_cache *cachep;
455 mutex_lock(&pid_caches_mutex);
456 list_for_each_entry (pcache, &pid_caches_lh, list)
457 if (pcache->nr_ids == nr_ids)
460 pcache = kmalloc(sizeof(struct pid_cache), GFP_KERNEL);
464 snprintf(pcache->name, sizeof(pcache->name), "pid_%d", nr_ids);
465 cachep = kmem_cache_create(pcache->name,
466 sizeof(struct pid) + (nr_ids - 1) * sizeof(struct upid),
467 0, SLAB_HWCACHE_ALIGN, NULL);
471 pcache->nr_ids = nr_ids;
472 pcache->cachep = cachep;
473 list_add(&pcache->list, &pid_caches_lh);
475 mutex_unlock(&pid_caches_mutex);
476 return pcache->cachep;
481 mutex_unlock(&pid_caches_mutex);
485 static struct pid_namespace *create_pid_namespace(int level)
487 struct pid_namespace *ns;
490 ns = kmem_cache_alloc(pid_ns_cachep, GFP_KERNEL);
494 ns->pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
495 if (!ns->pidmap[0].page)
498 ns->pid_cachep = create_pid_cachep(level + 1);
499 if (ns->pid_cachep == NULL)
502 kref_init(&ns->kref);
504 ns->child_reaper = NULL;
507 set_bit(0, ns->pidmap[0].page);
508 atomic_set(&ns->pidmap[0].nr_free, BITS_PER_PAGE - 1);
510 for (i = 1; i < PIDMAP_ENTRIES; i++) {
511 ns->pidmap[i].page = 0;
512 atomic_set(&ns->pidmap[i].nr_free, BITS_PER_PAGE);
518 kfree(ns->pidmap[0].page);
520 kmem_cache_free(pid_ns_cachep, ns);
522 return ERR_PTR(-ENOMEM);
525 static void destroy_pid_namespace(struct pid_namespace *ns)
529 for (i = 0; i < PIDMAP_ENTRIES; i++)
530 kfree(ns->pidmap[i].page);
531 kmem_cache_free(pid_ns_cachep, ns);
534 struct pid_namespace *copy_pid_ns(unsigned long flags, struct pid_namespace *old_ns)
536 struct pid_namespace *new_ns;
539 new_ns = get_pid_ns(old_ns);
540 if (!(flags & CLONE_NEWPID))
543 new_ns = ERR_PTR(-EINVAL);
544 if (flags & CLONE_THREAD)
547 new_ns = create_pid_namespace(old_ns->level + 1);
549 new_ns->parent = get_pid_ns(old_ns);
557 void free_pid_ns(struct kref *kref)
559 struct pid_namespace *ns, *parent;
561 ns = container_of(kref, struct pid_namespace, kref);
564 destroy_pid_namespace(ns);
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)
578 unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
580 pidhash_shift = max(4, fls(megabytes * 4));
581 pidhash_shift = min(12, pidhash_shift);
582 pidhash_size = 1 << pidhash_shift;
584 printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
585 pidhash_size, pidhash_shift,
586 pidhash_size * sizeof(struct hlist_head));
588 pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash)));
590 panic("Could not alloc pidhash!\n");
591 for (i = 0; i < pidhash_size; i++)
592 INIT_HLIST_HEAD(&pid_hash[i]);
595 void __init pidmap_init(void)
597 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
598 /* Reserve PID 0. We never call free_pidmap(0) */
599 set_bit(0, init_pid_ns.pidmap[0].page);
600 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
602 init_pid_ns.pid_cachep = create_pid_cachep(1);
603 if (init_pid_ns.pid_cachep == NULL)
604 panic("Can't create pid_1 cachep\n");
606 pid_ns_cachep = KMEM_CACHE(pid_namespace, SLAB_PANIC);