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>
39 #include <linux/proc_ns.h>
40 #include <linux/proc_fs.h>
41 #include <linux/anon_inodes.h>
43 #define pid_hashfn(nr, ns) \
44 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
45 static struct hlist_head
*pid_hash
;
46 static unsigned int pidhash_shift
= 4;
47 struct pid init_struct_pid
= INIT_STRUCT_PID
;
49 int pid_max
= PID_MAX_DEFAULT
;
51 #define RESERVED_PIDS 300
53 int pid_max_min
= RESERVED_PIDS
+ 1;
54 int pid_max_max
= PID_MAX_LIMIT
;
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
= {
74 [ 0 ... PIDMAP_ENTRIES
-1] = { ATOMIC_INIT(BITS_PER_PAGE
), NULL
}
77 .nr_hashed
= PIDNS_HASH_ADDING
,
79 .child_reaper
= &init_task
,
80 .user_ns
= &init_user_ns
,
81 .ns
.inum
= PROC_PID_INIT_INO
,
83 .ns
.ops
= &pidns_operations
,
86 EXPORT_SYMBOL_GPL(init_pid_ns
);
89 * Note: disable interrupts while the pidmap_lock is held as an
90 * interrupt might come in and do read_lock(&tasklist_lock).
92 * If we don't disable interrupts there is a nasty deadlock between
93 * detach_pid()->free_pid() and another cpu that does
94 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
95 * read_lock(&tasklist_lock);
97 * After we clean up the tasklist_lock and know there are no
98 * irq handlers that take it we can leave the interrupts enabled.
99 * For now it is easier to be safe than to prove it can't happen.
102 static __cacheline_aligned_in_smp
DEFINE_SPINLOCK(pidmap_lock
);
104 static void free_pidmap(struct upid
*upid
)
107 struct pidmap
*map
= upid
->ns
->pidmap
+ nr
/ BITS_PER_PAGE
;
108 int offset
= nr
& BITS_PER_PAGE_MASK
;
110 clear_bit(offset
, map
->page
);
111 atomic_inc(&map
->nr_free
);
115 * If we started walking pids at 'base', is 'a' seen before 'b'?
117 static int pid_before(int base
, int a
, int b
)
120 * This is the same as saying
122 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
123 * and that mapping orders 'a' and 'b' with respect to 'base'.
125 return (unsigned)(a
- base
) < (unsigned)(b
- base
);
129 * We might be racing with someone else trying to set pid_ns->last_pid
130 * at the pid allocation time (there's also a sysctl for this, but racing
131 * with this one is OK, see comment in kernel/pid_namespace.c about it).
132 * We want the winner to have the "later" value, because if the
133 * "earlier" value prevails, then a pid may get reused immediately.
135 * Since pids rollover, it is not sufficient to just pick the bigger
136 * value. We have to consider where we started counting from.
138 * 'base' is the value of pid_ns->last_pid that we observed when
139 * we started looking for a pid.
141 * 'pid' is the pid that we eventually found.
143 static void set_last_pid(struct pid_namespace
*pid_ns
, int base
, int pid
)
146 int last_write
= base
;
149 last_write
= cmpxchg(&pid_ns
->last_pid
, prev
, pid
);
150 } while ((prev
!= last_write
) && (pid_before(base
, last_write
, pid
)));
153 static int alloc_pidmap(struct pid_namespace
*pid_ns
)
155 int i
, offset
, max_scan
, pid
, last
= pid_ns
->last_pid
;
161 offset
= pid
& BITS_PER_PAGE_MASK
;
162 map
= &pid_ns
->pidmap
[pid
/BITS_PER_PAGE
];
164 * If last_pid points into the middle of the map->page we
165 * want to scan this bitmap block twice, the second time
166 * we start with offset == 0 (or RESERVED_PIDS).
168 max_scan
= DIV_ROUND_UP(pid_max
, BITS_PER_PAGE
) - !offset
;
169 for (i
= 0; i
<= max_scan
; ++i
) {
170 if (unlikely(!map
->page
)) {
171 void *page
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
173 * Free the page if someone raced with us
176 spin_lock_irq(&pidmap_lock
);
181 spin_unlock_irq(&pidmap_lock
);
183 if (unlikely(!map
->page
))
186 if (likely(atomic_read(&map
->nr_free
))) {
188 if (!test_and_set_bit(offset
, map
->page
)) {
189 atomic_dec(&map
->nr_free
);
190 set_last_pid(pid_ns
, last
, pid
);
193 offset
= find_next_offset(map
, offset
);
194 if (offset
>= BITS_PER_PAGE
)
196 pid
= mk_pid(pid_ns
, map
, offset
);
201 if (map
< &pid_ns
->pidmap
[(pid_max
-1)/BITS_PER_PAGE
]) {
205 map
= &pid_ns
->pidmap
[0];
206 offset
= RESERVED_PIDS
;
207 if (unlikely(last
== offset
))
210 pid
= mk_pid(pid_ns
, map
, offset
);
215 int next_pidmap(struct pid_namespace
*pid_ns
, unsigned int last
)
218 struct pidmap
*map
, *end
;
220 if (last
>= PID_MAX_LIMIT
)
223 offset
= (last
+ 1) & BITS_PER_PAGE_MASK
;
224 map
= &pid_ns
->pidmap
[(last
+ 1)/BITS_PER_PAGE
];
225 end
= &pid_ns
->pidmap
[PIDMAP_ENTRIES
];
226 for (; map
< end
; map
++, offset
= 0) {
227 if (unlikely(!map
->page
))
229 offset
= find_next_bit((map
)->page
, BITS_PER_PAGE
, offset
);
230 if (offset
< BITS_PER_PAGE
)
231 return mk_pid(pid_ns
, map
, offset
);
236 void put_pid(struct pid
*pid
)
238 struct pid_namespace
*ns
;
243 ns
= pid
->numbers
[pid
->level
].ns
;
244 if ((atomic_read(&pid
->count
) == 1) ||
245 atomic_dec_and_test(&pid
->count
)) {
246 kmem_cache_free(ns
->pid_cachep
, pid
);
250 EXPORT_SYMBOL_GPL(put_pid
);
252 static void delayed_put_pid(struct rcu_head
*rhp
)
254 struct pid
*pid
= container_of(rhp
, struct pid
, rcu
);
258 void free_pid(struct pid
*pid
)
260 /* We can be called with write_lock_irq(&tasklist_lock) held */
264 spin_lock_irqsave(&pidmap_lock
, flags
);
265 for (i
= 0; i
<= pid
->level
; i
++) {
266 struct upid
*upid
= pid
->numbers
+ i
;
267 struct pid_namespace
*ns
= upid
->ns
;
268 hlist_del_rcu(&upid
->pid_chain
);
269 switch(--ns
->nr_hashed
) {
272 /* When all that is left in the pid namespace
273 * is the reaper wake up the reaper. The reaper
274 * may be sleeping in zap_pid_ns_processes().
276 wake_up_process(ns
->child_reaper
);
278 case PIDNS_HASH_ADDING
:
279 /* Handle a fork failure of the first process */
280 WARN_ON(ns
->child_reaper
);
284 schedule_work(&ns
->proc_work
);
288 spin_unlock_irqrestore(&pidmap_lock
, flags
);
290 for (i
= 0; i
<= pid
->level
; i
++)
291 free_pidmap(pid
->numbers
+ i
);
293 call_rcu(&pid
->rcu
, delayed_put_pid
);
296 struct pid
*alloc_pid(struct pid_namespace
*ns
)
301 struct pid_namespace
*tmp
;
303 int retval
= -ENOMEM
;
305 pid
= kmem_cache_alloc(ns
->pid_cachep
, GFP_KERNEL
);
307 return ERR_PTR(retval
);
310 pid
->level
= ns
->level
;
311 for (i
= ns
->level
; i
>= 0; i
--) {
312 nr
= alloc_pidmap(tmp
);
318 pid
->numbers
[i
].nr
= nr
;
319 pid
->numbers
[i
].ns
= tmp
;
323 if (unlikely(is_child_reaper(pid
))) {
324 if (pid_ns_prepare_proc(ns
)) {
325 disable_pid_allocation(ns
);
331 atomic_set(&pid
->count
, 1);
332 for (type
= 0; type
< PIDTYPE_MAX
; ++type
)
333 INIT_HLIST_HEAD(&pid
->tasks
[type
]);
335 init_waitqueue_head(&pid
->wait_pidfd
);
337 upid
= pid
->numbers
+ ns
->level
;
338 spin_lock_irq(&pidmap_lock
);
339 if (!(ns
->nr_hashed
& PIDNS_HASH_ADDING
))
341 for ( ; upid
>= pid
->numbers
; --upid
) {
342 hlist_add_head_rcu(&upid
->pid_chain
,
343 &pid_hash
[pid_hashfn(upid
->nr
, upid
->ns
)]);
344 upid
->ns
->nr_hashed
++;
346 spin_unlock_irq(&pidmap_lock
);
351 spin_unlock_irq(&pidmap_lock
);
355 while (++i
<= ns
->level
)
356 free_pidmap(pid
->numbers
+ i
);
358 kmem_cache_free(ns
->pid_cachep
, pid
);
359 return ERR_PTR(retval
);
362 void disable_pid_allocation(struct pid_namespace
*ns
)
364 spin_lock_irq(&pidmap_lock
);
365 ns
->nr_hashed
&= ~PIDNS_HASH_ADDING
;
366 spin_unlock_irq(&pidmap_lock
);
369 struct pid
*find_pid_ns(int nr
, struct pid_namespace
*ns
)
373 hlist_for_each_entry_rcu(pnr
,
374 &pid_hash
[pid_hashfn(nr
, ns
)], pid_chain
)
375 if (pnr
->nr
== nr
&& pnr
->ns
== ns
)
376 return container_of(pnr
, struct pid
,
381 EXPORT_SYMBOL_GPL(find_pid_ns
);
383 struct pid
*find_vpid(int nr
)
385 return find_pid_ns(nr
, task_active_pid_ns(current
));
387 EXPORT_SYMBOL_GPL(find_vpid
);
390 * attach_pid() must be called with the tasklist_lock write-held.
392 void attach_pid(struct task_struct
*task
, enum pid_type type
)
394 struct pid_link
*link
= &task
->pids
[type
];
395 hlist_add_head_rcu(&link
->node
, &link
->pid
->tasks
[type
]);
398 static void __change_pid(struct task_struct
*task
, enum pid_type type
,
401 struct pid_link
*link
;
405 link
= &task
->pids
[type
];
408 hlist_del_rcu(&link
->node
);
411 for (tmp
= PIDTYPE_MAX
; --tmp
>= 0; )
412 if (!hlist_empty(&pid
->tasks
[tmp
]))
418 void detach_pid(struct task_struct
*task
, enum pid_type type
)
420 __change_pid(task
, type
, NULL
);
423 void change_pid(struct task_struct
*task
, enum pid_type type
,
426 __change_pid(task
, type
, pid
);
427 attach_pid(task
, type
);
430 /* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
431 void transfer_pid(struct task_struct
*old
, struct task_struct
*new,
434 new->pids
[type
].pid
= old
->pids
[type
].pid
;
435 hlist_replace_rcu(&old
->pids
[type
].node
, &new->pids
[type
].node
);
438 struct task_struct
*pid_task(struct pid
*pid
, enum pid_type type
)
440 struct task_struct
*result
= NULL
;
442 struct hlist_node
*first
;
443 first
= rcu_dereference_check(hlist_first_rcu(&pid
->tasks
[type
]),
444 lockdep_tasklist_lock_is_held());
446 result
= hlist_entry(first
, struct task_struct
, pids
[(type
)].node
);
450 EXPORT_SYMBOL(pid_task
);
453 * Must be called under rcu_read_lock().
455 struct task_struct
*find_task_by_pid_ns(pid_t nr
, struct pid_namespace
*ns
)
457 RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
458 "find_task_by_pid_ns() needs rcu_read_lock() protection");
459 return pid_task(find_pid_ns(nr
, ns
), PIDTYPE_PID
);
462 struct task_struct
*find_task_by_vpid(pid_t vnr
)
464 return find_task_by_pid_ns(vnr
, task_active_pid_ns(current
));
467 struct pid
*get_task_pid(struct task_struct
*task
, enum pid_type type
)
471 if (type
!= PIDTYPE_PID
)
472 task
= task
->group_leader
;
473 pid
= get_pid(rcu_dereference(task
->pids
[type
].pid
));
477 EXPORT_SYMBOL_GPL(get_task_pid
);
479 struct task_struct
*get_pid_task(struct pid
*pid
, enum pid_type type
)
481 struct task_struct
*result
;
483 result
= pid_task(pid
, type
);
485 get_task_struct(result
);
489 EXPORT_SYMBOL_GPL(get_pid_task
);
491 struct pid
*find_get_pid(pid_t nr
)
496 pid
= get_pid(find_vpid(nr
));
501 EXPORT_SYMBOL_GPL(find_get_pid
);
503 pid_t
pid_nr_ns(struct pid
*pid
, struct pid_namespace
*ns
)
508 if (pid
&& ns
->level
<= pid
->level
) {
509 upid
= &pid
->numbers
[ns
->level
];
515 EXPORT_SYMBOL_GPL(pid_nr_ns
);
517 pid_t
pid_vnr(struct pid
*pid
)
519 return pid_nr_ns(pid
, task_active_pid_ns(current
));
521 EXPORT_SYMBOL_GPL(pid_vnr
);
523 pid_t
__task_pid_nr_ns(struct task_struct
*task
, enum pid_type type
,
524 struct pid_namespace
*ns
)
530 ns
= task_active_pid_ns(current
);
531 if (likely(pid_alive(task
))) {
532 if (type
!= PIDTYPE_PID
) {
533 if (type
== __PIDTYPE_TGID
)
535 task
= task
->group_leader
;
537 nr
= pid_nr_ns(rcu_dereference(task
->pids
[type
].pid
), ns
);
543 EXPORT_SYMBOL(__task_pid_nr_ns
);
545 struct pid_namespace
*task_active_pid_ns(struct task_struct
*tsk
)
547 return ns_of_pid(task_pid(tsk
));
549 EXPORT_SYMBOL_GPL(task_active_pid_ns
);
552 * Used by proc to find the first pid that is greater than or equal to nr.
554 * If there is a pid at nr this function is exactly the same as find_pid_ns.
556 struct pid
*find_ge_pid(int nr
, struct pid_namespace
*ns
)
561 pid
= find_pid_ns(nr
, ns
);
564 nr
= next_pidmap(ns
, nr
);
571 * pidfd_create() - Create a new pid file descriptor.
573 * @pid: struct pid that the pidfd will reference
575 * This creates a new pid file descriptor with the O_CLOEXEC flag set.
577 * Note, that this function can only be called after the fd table has
578 * been unshared to avoid leaking the pidfd to the new process.
580 * Return: On success, a cloexec pidfd is returned.
581 * On error, a negative errno number will be returned.
583 static int pidfd_create(struct pid
*pid
)
587 fd
= anon_inode_getfd("[pidfd]", &pidfd_fops
, get_pid(pid
),
596 * pidfd_open() - Open new pid file descriptor.
598 * @pid: pid for which to retrieve a pidfd
599 * @flags: flags to pass
601 * This creates a new pid file descriptor with the O_CLOEXEC flag set for
602 * the process identified by @pid. Currently, the process identified by
603 * @pid must be a thread-group leader. This restriction currently exists
604 * for all aspects of pidfds including pidfd creation (CLONE_PIDFD cannot
605 * be used with CLONE_THREAD) and pidfd polling (only supports thread group
608 * Return: On success, a cloexec pidfd is returned.
609 * On error, a negative errno number will be returned.
611 SYSCALL_DEFINE2(pidfd_open
, pid_t
, pid
, unsigned int, flags
)
615 struct task_struct
*tsk
;
623 p
= find_get_pid(pid
);
629 tsk
= pid_task(p
, PIDTYPE_PID
);
630 /* Check that pid belongs to a group leader task */
631 if (!tsk
|| !thread_group_leader(tsk
))
635 fd
= ret
?: pidfd_create(p
);
641 * The pid hash table is scaled according to the amount of memory in the
642 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
645 void __init
pidhash_init(void)
647 unsigned int i
, pidhash_size
;
649 pid_hash
= alloc_large_system_hash("PID", sizeof(*pid_hash
), 0, 18,
650 HASH_EARLY
| HASH_SMALL
,
651 &pidhash_shift
, NULL
,
653 pidhash_size
= 1U << pidhash_shift
;
655 for (i
= 0; i
< pidhash_size
; i
++)
656 INIT_HLIST_HEAD(&pid_hash
[i
]);
659 void __init
pidmap_init(void)
661 /* Verify no one has done anything silly: */
662 BUILD_BUG_ON(PID_MAX_LIMIT
>= PIDNS_HASH_ADDING
);
664 /* bump default and minimum pid_max based on number of cpus */
665 pid_max
= min(pid_max_max
, max_t(int, pid_max
,
666 PIDS_PER_CPU_DEFAULT
* num_possible_cpus()));
667 pid_max_min
= max_t(int, pid_max_min
,
668 PIDS_PER_CPU_MIN
* num_possible_cpus());
669 pr_info("pid_max: default: %u minimum: %u\n", pid_max
, pid_max_min
);
671 init_pid_ns
.pidmap
[0].page
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
672 /* Reserve PID 0. We never call free_pidmap(0) */
673 set_bit(0, init_pid_ns
.pidmap
[0].page
);
674 atomic_dec(&init_pid_ns
.pidmap
[0].nr_free
);
676 init_pid_ns
.pid_cachep
= KMEM_CACHE(pid
,
677 SLAB_HWCACHE_ALIGN
| SLAB_PANIC
| SLAB_ACCOUNT
);