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>
40 #define pid_hashfn(nr, ns) \
41 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
42 static struct hlist_head *pid_hash;
43 static unsigned int pidhash_shift = 4;
44 struct pid init_struct_pid = INIT_STRUCT_PID;
46 int pid_max = PID_MAX_DEFAULT;
48 #define RESERVED_PIDS 300
50 int pid_max_min = RESERVED_PIDS + 1;
51 int pid_max_max = PID_MAX_LIMIT;
53 #define BITS_PER_PAGE (PAGE_SIZE*8)
54 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
56 static inline int mk_pid(struct pid_namespace *pid_ns,
57 struct pidmap *map, int off)
59 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
62 #define find_next_offset(map, off) \
63 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
66 * PID-map pages start out as NULL, they get allocated upon
67 * first use and are never deallocated. This way a low pid_max
68 * value does not cause lots of bitmaps to be allocated, but
69 * the scheme scales to up to 4 million PIDs, runtime.
71 struct pid_namespace init_pid_ns = {
73 .refcount = ATOMIC_INIT(2),
76 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
80 .child_reaper = &init_task,
82 EXPORT_SYMBOL_GPL(init_pid_ns);
85 * Note: disable interrupts while the pidmap_lock is held as an
86 * interrupt might come in and do read_lock(&tasklist_lock).
88 * If we don't disable interrupts there is a nasty deadlock between
89 * detach_pid()->free_pid() and another cpu that does
90 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
91 * read_lock(&tasklist_lock);
93 * After we clean up the tasklist_lock and know there are no
94 * irq handlers that take it we can leave the interrupts enabled.
95 * For now it is easier to be safe than to prove it can't happen.
98 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
100 static void free_pidmap(struct upid *upid)
103 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
104 int offset = nr & BITS_PER_PAGE_MASK;
106 clear_bit(offset, map->page);
107 atomic_inc(&map->nr_free);
111 * If we started walking pids at 'base', is 'a' seen before 'b'?
113 static int pid_before(int base, int a, int b)
116 * This is the same as saying
118 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
119 * and that mapping orders 'a' and 'b' with respect to 'base'.
121 return (unsigned)(a - base) < (unsigned)(b - base);
125 * We might be racing with someone else trying to set pid_ns->last_pid
126 * at the pid allocation time (there's also a sysctl for this, but racing
127 * with this one is OK, see comment in kernel/pid_namespace.c about it).
128 * We want the winner to have the "later" value, because if the
129 * "earlier" value prevails, then a pid may get reused immediately.
131 * Since pids rollover, it is not sufficient to just pick the bigger
132 * value. We have to consider where we started counting from.
134 * 'base' is the value of pid_ns->last_pid that we observed when
135 * we started looking for a pid.
137 * 'pid' is the pid that we eventually found.
139 static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
142 int last_write = base;
145 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
146 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
149 static int alloc_pidmap(struct pid_namespace *pid_ns)
151 int i, offset, max_scan, pid, last = pid_ns->last_pid;
157 offset = pid & BITS_PER_PAGE_MASK;
158 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
160 * If last_pid points into the middle of the map->page we
161 * want to scan this bitmap block twice, the second time
162 * we start with offset == 0 (or RESERVED_PIDS).
164 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
165 for (i = 0; i <= max_scan; ++i) {
166 if (unlikely(!map->page)) {
167 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
169 * Free the page if someone raced with us
172 spin_lock_irq(&pidmap_lock);
177 spin_unlock_irq(&pidmap_lock);
179 if (unlikely(!map->page))
182 if (likely(atomic_read(&map->nr_free))) {
184 if (!test_and_set_bit(offset, map->page)) {
185 atomic_dec(&map->nr_free);
186 set_last_pid(pid_ns, last, pid);
189 offset = find_next_offset(map, offset);
190 pid = mk_pid(pid_ns, map, offset);
191 } while (offset < BITS_PER_PAGE && pid < pid_max);
193 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
197 map = &pid_ns->pidmap[0];
198 offset = RESERVED_PIDS;
199 if (unlikely(last == offset))
202 pid = mk_pid(pid_ns, map, offset);
207 int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
210 struct pidmap *map, *end;
212 if (last >= PID_MAX_LIMIT)
215 offset = (last + 1) & BITS_PER_PAGE_MASK;
216 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
217 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
218 for (; map < end; map++, offset = 0) {
219 if (unlikely(!map->page))
221 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
222 if (offset < BITS_PER_PAGE)
223 return mk_pid(pid_ns, map, offset);
228 void put_pid(struct pid *pid)
230 struct pid_namespace *ns;
235 ns = pid->numbers[pid->level].ns;
236 if ((atomic_read(&pid->count) == 1) ||
237 atomic_dec_and_test(&pid->count)) {
238 kmem_cache_free(ns->pid_cachep, pid);
242 EXPORT_SYMBOL_GPL(put_pid);
244 static void delayed_put_pid(struct rcu_head *rhp)
246 struct pid *pid = container_of(rhp, struct pid, rcu);
250 void free_pid(struct pid *pid)
252 /* We can be called with write_lock_irq(&tasklist_lock) held */
256 spin_lock_irqsave(&pidmap_lock, flags);
257 for (i = 0; i <= pid->level; i++)
258 hlist_del_rcu(&pid->numbers[i].pid_chain);
259 spin_unlock_irqrestore(&pidmap_lock, flags);
261 for (i = 0; i <= pid->level; i++)
262 free_pidmap(pid->numbers + i);
264 call_rcu(&pid->rcu, delayed_put_pid);
267 struct pid *alloc_pid(struct pid_namespace *ns)
272 struct pid_namespace *tmp;
275 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
280 for (i = ns->level; i >= 0; i--) {
281 nr = alloc_pidmap(tmp);
285 pid->numbers[i].nr = nr;
286 pid->numbers[i].ns = tmp;
291 pid->level = ns->level;
292 atomic_set(&pid->count, 1);
293 for (type = 0; type < PIDTYPE_MAX; ++type)
294 INIT_HLIST_HEAD(&pid->tasks[type]);
296 upid = pid->numbers + ns->level;
297 spin_lock_irq(&pidmap_lock);
298 for ( ; upid >= pid->numbers; --upid)
299 hlist_add_head_rcu(&upid->pid_chain,
300 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
301 spin_unlock_irq(&pidmap_lock);
307 while (++i <= ns->level)
308 free_pidmap(pid->numbers + i);
310 kmem_cache_free(ns->pid_cachep, pid);
315 struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
317 struct hlist_node *elem;
320 hlist_for_each_entry_rcu(pnr, elem,
321 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
322 if (pnr->nr == nr && pnr->ns == ns)
323 return container_of(pnr, struct pid,
328 EXPORT_SYMBOL_GPL(find_pid_ns);
330 struct pid *find_vpid(int nr)
332 return find_pid_ns(nr, current->nsproxy->pid_ns);
334 EXPORT_SYMBOL_GPL(find_vpid);
337 * attach_pid() must be called with the tasklist_lock write-held.
339 void attach_pid(struct task_struct *task, enum pid_type type,
342 struct pid_link *link;
344 link = &task->pids[type];
346 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
349 static void __change_pid(struct task_struct *task, enum pid_type type,
352 struct pid_link *link;
356 link = &task->pids[type];
359 hlist_del_rcu(&link->node);
362 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
363 if (!hlist_empty(&pid->tasks[tmp]))
369 void detach_pid(struct task_struct *task, enum pid_type type)
371 __change_pid(task, type, NULL);
374 void change_pid(struct task_struct *task, enum pid_type type,
377 __change_pid(task, type, pid);
378 attach_pid(task, type, pid);
381 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
382 void transfer_pid(struct task_struct *old, struct task_struct *new,
385 new->pids[type].pid = old->pids[type].pid;
386 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
389 struct task_struct *pid_task(struct pid *pid, enum pid_type type)
391 struct task_struct *result = NULL;
393 struct hlist_node *first;
394 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
395 lockdep_tasklist_lock_is_held());
397 result = hlist_entry(first, struct task_struct, pids[(type)].node);
401 EXPORT_SYMBOL(pid_task);
404 * Must be called under rcu_read_lock().
406 struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
408 rcu_lockdep_assert(rcu_read_lock_held(),
409 "find_task_by_pid_ns() needs rcu_read_lock()"
411 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
414 struct task_struct *find_task_by_vpid(pid_t vnr)
416 return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns);
419 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
423 if (type != PIDTYPE_PID)
424 task = task->group_leader;
425 pid = get_pid(task->pids[type].pid);
429 EXPORT_SYMBOL_GPL(get_task_pid);
431 struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
433 struct task_struct *result;
435 result = pid_task(pid, type);
437 get_task_struct(result);
441 EXPORT_SYMBOL_GPL(get_pid_task);
443 struct pid *find_get_pid(pid_t nr)
448 pid = get_pid(find_vpid(nr));
453 EXPORT_SYMBOL_GPL(find_get_pid);
455 pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
460 if (pid && ns->level <= pid->level) {
461 upid = &pid->numbers[ns->level];
467 EXPORT_SYMBOL_GPL(pid_nr_ns);
469 pid_t pid_vnr(struct pid *pid)
471 return pid_nr_ns(pid, current->nsproxy->pid_ns);
473 EXPORT_SYMBOL_GPL(pid_vnr);
475 pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
476 struct pid_namespace *ns)
482 ns = current->nsproxy->pid_ns;
483 if (likely(pid_alive(task))) {
484 if (type != PIDTYPE_PID)
485 task = task->group_leader;
486 nr = pid_nr_ns(task->pids[type].pid, ns);
492 EXPORT_SYMBOL(__task_pid_nr_ns);
494 pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
496 return pid_nr_ns(task_tgid(tsk), ns);
498 EXPORT_SYMBOL(task_tgid_nr_ns);
500 struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
502 return ns_of_pid(task_pid(tsk));
504 EXPORT_SYMBOL_GPL(task_active_pid_ns);
507 * Used by proc to find the first pid that is greater than or equal to nr.
509 * If there is a pid at nr this function is exactly the same as find_pid_ns.
511 struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
516 pid = find_pid_ns(nr, ns);
519 nr = next_pidmap(ns, nr);
526 * The pid hash table is scaled according to the amount of memory in the
527 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
530 void __init pidhash_init(void)
532 unsigned int i, pidhash_size;
534 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
535 HASH_EARLY | HASH_SMALL,
536 &pidhash_shift, NULL,
538 pidhash_size = 1U << pidhash_shift;
540 for (i = 0; i < pidhash_size; i++)
541 INIT_HLIST_HEAD(&pid_hash[i]);
544 void __init pidmap_init(void)
546 /* bump default and minimum pid_max based on number of cpus */
547 pid_max = min(pid_max_max, max_t(int, pid_max,
548 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
549 pid_max_min = max_t(int, pid_max_min,
550 PIDS_PER_CPU_MIN * num_possible_cpus());
551 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
553 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
554 /* Reserve PID 0. We never call free_pidmap(0) */
555 set_bit(0, init_pid_ns.pidmap[0].page);
556 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
558 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
559 SLAB_HWCACHE_ALIGN | SLAB_PANIC);