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).
24 #include <linux/module.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
27 #include <linux/bootmem.h>
28 #include <linux/hash.h>
29 #include <linux/pid_namespace.h>
30 #include <linux/init_task.h>
32 #define pid_hashfn(nr, ns) \
33 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
34 static struct hlist_head *pid_hash;
35 static int pidhash_shift;
36 struct pid init_struct_pid = INIT_STRUCT_PID;
38 int pid_max = PID_MAX_DEFAULT;
40 #define RESERVED_PIDS 300
42 int pid_max_min = RESERVED_PIDS + 1;
43 int pid_max_max = PID_MAX_LIMIT;
45 #define BITS_PER_PAGE (PAGE_SIZE*8)
46 #define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
48 static inline int mk_pid(struct pid_namespace *pid_ns,
49 struct pidmap *map, int off)
51 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
54 #define find_next_offset(map, off) \
55 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
58 * PID-map pages start out as NULL, they get allocated upon
59 * first use and are never deallocated. This way a low pid_max
60 * value does not cause lots of bitmaps to be allocated, but
61 * the scheme scales to up to 4 million PIDs, runtime.
63 struct pid_namespace init_pid_ns = {
65 .refcount = ATOMIC_INIT(2),
68 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
72 .child_reaper = &init_task,
75 int is_global_init(struct task_struct *tsk)
77 return tsk == init_pid_ns.child_reaper;
81 * Note: disable interrupts while the pidmap_lock is held as an
82 * interrupt might come in and do read_lock(&tasklist_lock).
84 * If we don't disable interrupts there is a nasty deadlock between
85 * detach_pid()->free_pid() and another cpu that does
86 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
87 * read_lock(&tasklist_lock);
89 * After we clean up the tasklist_lock and know there are no
90 * irq handlers that take it we can leave the interrupts enabled.
91 * For now it is easier to be safe than to prove it can't happen.
94 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
96 static fastcall void free_pidmap(struct pid_namespace *pid_ns, int pid)
98 struct pidmap *map = pid_ns->pidmap + pid / BITS_PER_PAGE;
99 int offset = pid & BITS_PER_PAGE_MASK;
101 clear_bit(offset, map->page);
102 atomic_inc(&map->nr_free);
105 static int alloc_pidmap(struct pid_namespace *pid_ns)
107 int i, offset, max_scan, pid, last = pid_ns->last_pid;
113 offset = pid & BITS_PER_PAGE_MASK;
114 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
115 max_scan = (pid_max + BITS_PER_PAGE - 1)/BITS_PER_PAGE - !offset;
116 for (i = 0; i <= max_scan; ++i) {
117 if (unlikely(!map->page)) {
118 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
120 * Free the page if someone raced with us
123 spin_lock_irq(&pidmap_lock);
128 spin_unlock_irq(&pidmap_lock);
129 if (unlikely(!map->page))
132 if (likely(atomic_read(&map->nr_free))) {
134 if (!test_and_set_bit(offset, map->page)) {
135 atomic_dec(&map->nr_free);
136 pid_ns->last_pid = pid;
139 offset = find_next_offset(map, offset);
140 pid = mk_pid(pid_ns, map, offset);
142 * find_next_offset() found a bit, the pid from it
143 * is in-bounds, and if we fell back to the last
144 * bitmap block and the final block was the same
145 * as the starting point, pid is before last_pid.
147 } while (offset < BITS_PER_PAGE && pid < pid_max &&
148 (i != max_scan || pid < last ||
149 !((last+1) & BITS_PER_PAGE_MASK)));
151 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
155 map = &pid_ns->pidmap[0];
156 offset = RESERVED_PIDS;
157 if (unlikely(last == offset))
160 pid = mk_pid(pid_ns, map, offset);
165 static int next_pidmap(struct pid_namespace *pid_ns, int last)
168 struct pidmap *map, *end;
170 offset = (last + 1) & BITS_PER_PAGE_MASK;
171 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
172 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
173 for (; map < end; map++, offset = 0) {
174 if (unlikely(!map->page))
176 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
177 if (offset < BITS_PER_PAGE)
178 return mk_pid(pid_ns, map, offset);
183 fastcall void put_pid(struct pid *pid)
185 struct pid_namespace *ns;
190 ns = pid->numbers[pid->level].ns;
191 if ((atomic_read(&pid->count) == 1) ||
192 atomic_dec_and_test(&pid->count)) {
193 kmem_cache_free(ns->pid_cachep, pid);
194 if (ns != &init_pid_ns)
198 EXPORT_SYMBOL_GPL(put_pid);
200 static void delayed_put_pid(struct rcu_head *rhp)
202 struct pid *pid = container_of(rhp, struct pid, rcu);
206 fastcall void free_pid(struct pid *pid)
208 /* We can be called with write_lock_irq(&tasklist_lock) held */
212 spin_lock_irqsave(&pidmap_lock, flags);
213 hlist_del_rcu(&pid->pid_chain);
214 spin_unlock_irqrestore(&pidmap_lock, flags);
216 for (i = 0; i <= pid->level; i++)
217 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
219 call_rcu(&pid->rcu, delayed_put_pid);
222 struct pid *alloc_pid(struct pid_namespace *ns)
227 struct pid_namespace *tmp;
229 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
234 for (i = ns->level; i >= 0; i--) {
235 nr = alloc_pidmap(tmp);
239 pid->numbers[i].nr = nr;
240 pid->numbers[i].ns = tmp;
244 if (ns != &init_pid_ns)
247 pid->level = ns->level;
248 pid->nr = pid->numbers[0].nr;
249 atomic_set(&pid->count, 1);
250 for (type = 0; type < PIDTYPE_MAX; ++type)
251 INIT_HLIST_HEAD(&pid->tasks[type]);
253 spin_lock_irq(&pidmap_lock);
254 hlist_add_head_rcu(&pid->pid_chain, &pid_hash[pid_hashfn(pid->nr, ns)]);
255 spin_unlock_irq(&pidmap_lock);
261 for (i++; i <= ns->level; i++)
262 free_pidmap(pid->numbers[i].ns, pid->numbers[i].nr);
264 kmem_cache_free(ns->pid_cachep, pid);
269 struct pid * fastcall find_pid(int nr)
271 struct hlist_node *elem;
274 hlist_for_each_entry_rcu(pid, elem,
275 &pid_hash[pid_hashfn(nr, &init_pid_ns)], pid_chain) {
281 EXPORT_SYMBOL_GPL(find_pid);
284 * attach_pid() must be called with the tasklist_lock write-held.
286 int fastcall attach_pid(struct task_struct *task, enum pid_type type,
289 struct pid_link *link;
291 link = &task->pids[type];
293 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
298 void fastcall detach_pid(struct task_struct *task, enum pid_type type)
300 struct pid_link *link;
304 link = &task->pids[type];
307 hlist_del_rcu(&link->node);
310 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
311 if (!hlist_empty(&pid->tasks[tmp]))
317 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
318 void fastcall transfer_pid(struct task_struct *old, struct task_struct *new,
321 new->pids[type].pid = old->pids[type].pid;
322 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
323 old->pids[type].pid = NULL;
326 struct task_struct * fastcall pid_task(struct pid *pid, enum pid_type type)
328 struct task_struct *result = NULL;
330 struct hlist_node *first;
331 first = rcu_dereference(pid->tasks[type].first);
333 result = hlist_entry(first, struct task_struct, pids[(type)].node);
339 * Must be called under rcu_read_lock() or with tasklist_lock read-held.
341 struct task_struct *find_task_by_pid_type(int type, int nr)
343 return pid_task(find_pid(nr), type);
346 EXPORT_SYMBOL(find_task_by_pid_type);
348 struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
352 pid = get_pid(task->pids[type].pid);
357 struct task_struct *fastcall get_pid_task(struct pid *pid, enum pid_type type)
359 struct task_struct *result;
361 result = pid_task(pid, type);
363 get_task_struct(result);
368 struct pid *find_get_pid(pid_t nr)
373 pid = get_pid(find_pid(nr));
380 * Used by proc to find the first pid that is greater then or equal to nr.
382 * If there is a pid at nr this function is exactly the same as find_pid.
384 struct pid *find_ge_pid(int nr)
392 nr = next_pidmap(task_active_pid_ns(current), nr);
397 EXPORT_SYMBOL_GPL(find_get_pid);
402 struct kmem_cache *cachep;
403 struct list_head list;
406 static LIST_HEAD(pid_caches_lh);
407 static DEFINE_MUTEX(pid_caches_mutex);
410 * creates the kmem cache to allocate pids from.
411 * @nr_ids: the number of numerical ids this pid will have to carry
414 static struct kmem_cache *create_pid_cachep(int nr_ids)
416 struct pid_cache *pcache;
417 struct kmem_cache *cachep;
419 mutex_lock(&pid_caches_mutex);
420 list_for_each_entry (pcache, &pid_caches_lh, list)
421 if (pcache->nr_ids == nr_ids)
424 pcache = kmalloc(sizeof(struct pid_cache), GFP_KERNEL);
428 snprintf(pcache->name, sizeof(pcache->name), "pid_%d", nr_ids);
429 cachep = kmem_cache_create(pcache->name,
430 /* FIXME add numerical ids here */
431 sizeof(struct pid), 0, SLAB_HWCACHE_ALIGN, NULL);
435 pcache->nr_ids = nr_ids;
436 pcache->cachep = cachep;
437 list_add(&pcache->list, &pid_caches_lh);
439 mutex_unlock(&pid_caches_mutex);
440 return pcache->cachep;
445 mutex_unlock(&pid_caches_mutex);
449 struct pid_namespace *copy_pid_ns(unsigned long flags, struct pid_namespace *old_ns)
456 void free_pid_ns(struct kref *kref)
458 struct pid_namespace *ns;
460 ns = container_of(kref, struct pid_namespace, kref);
465 * The pid hash table is scaled according to the amount of memory in the
466 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
469 void __init pidhash_init(void)
472 unsigned long megabytes = nr_kernel_pages >> (20 - PAGE_SHIFT);
474 pidhash_shift = max(4, fls(megabytes * 4));
475 pidhash_shift = min(12, pidhash_shift);
476 pidhash_size = 1 << pidhash_shift;
478 printk("PID hash table entries: %d (order: %d, %Zd bytes)\n",
479 pidhash_size, pidhash_shift,
480 pidhash_size * sizeof(struct hlist_head));
482 pid_hash = alloc_bootmem(pidhash_size * sizeof(*(pid_hash)));
484 panic("Could not alloc pidhash!\n");
485 for (i = 0; i < pidhash_size; i++)
486 INIT_HLIST_HEAD(&pid_hash[i]);
489 void __init pidmap_init(void)
491 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
492 /* Reserve PID 0. We never call free_pidmap(0) */
493 set_bit(0, init_pid_ns.pidmap[0].page);
494 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
496 init_pid_ns.pid_cachep = create_pid_cachep(1);
497 if (init_pid_ns.pid_cachep == NULL)
498 panic("Can't create pid_1 cachep\n");