Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Generic pidhash and scalable, time-bounded PID allocator
4 *
5 * (C) 2002-2003 Nadia Yvette Chambers, IBM
6 * (C) 2004 Nadia Yvette Chambers, Oracle
7 * (C) 2002-2004 Ingo Molnar, Red Hat
8 *
9 * pid-structures are backing objects for tasks sharing a given ID to chain
10 * against. There is very little to them aside from hashing them and
11 * parking tasks using given ID's on a list.
12 *
13 * The hash is always changed with the tasklist_lock write-acquired,
14 * and the hash is only accessed with the tasklist_lock at least
15 * read-acquired, so there's no additional SMP locking needed here.
16 *
17 * We have a list of bitmap pages, which bitmaps represent the PID space.
18 * Allocating and freeing PIDs is completely lockless. The worst-case
19 * allocation scenario when all but one out of 1 million PIDs possible are
20 * allocated already: the scanning of 32 list entries and at most PAGE_SIZE
21 * bytes. The typical fastpath is a single successful setbit. Freeing is O(1).
22 *
23 * Pid namespaces:
24 * (C) 2007 Pavel Emelyanov <xemul@openvz.org>, OpenVZ, SWsoft Inc.
25 * (C) 2007 Sukadev Bhattiprolu <sukadev@us.ibm.com>, IBM
26 * Many thanks to Oleg Nesterov for comments and help
27 *
28 */
29
30#include <linux/mm.h>
31#include <linux/export.h>
32#include <linux/slab.h>
33#include <linux/init.h>
34#include <linux/rculist.h>
35#include <linux/memblock.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/refcount.h>
41#include <linux/anon_inodes.h>
42#include <linux/sched/signal.h>
43#include <linux/sched/task.h>
44#include <linux/idr.h>
45#include <net/sock.h>
46#include <uapi/linux/pidfd.h>
47
48struct pid init_struct_pid = {
49 .count = REFCOUNT_INIT(1),
50 .tasks = {
51 { .first = NULL },
52 { .first = NULL },
53 { .first = NULL },
54 },
55 .level = 0,
56 .numbers = { {
57 .nr = 0,
58 .ns = &init_pid_ns,
59 }, }
60};
61
62int pid_max = PID_MAX_DEFAULT;
63
64#define RESERVED_PIDS 300
65
66int pid_max_min = RESERVED_PIDS + 1;
67int pid_max_max = PID_MAX_LIMIT;
68
69/*
70 * PID-map pages start out as NULL, they get allocated upon
71 * first use and are never deallocated. This way a low pid_max
72 * value does not cause lots of bitmaps to be allocated, but
73 * the scheme scales to up to 4 million PIDs, runtime.
74 */
75struct pid_namespace init_pid_ns = {
76 .ns.count = REFCOUNT_INIT(2),
77 .idr = IDR_INIT(init_pid_ns.idr),
78 .pid_allocated = PIDNS_ADDING,
79 .level = 0,
80 .child_reaper = &init_task,
81 .user_ns = &init_user_ns,
82 .ns.inum = PROC_PID_INIT_INO,
83#ifdef CONFIG_PID_NS
84 .ns.ops = &pidns_operations,
85#endif
86};
87EXPORT_SYMBOL_GPL(init_pid_ns);
88
89/*
90 * Note: disable interrupts while the pidmap_lock is held as an
91 * interrupt might come in and do read_lock(&tasklist_lock).
92 *
93 * If we don't disable interrupts there is a nasty deadlock between
94 * detach_pid()->free_pid() and another cpu that does
95 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
96 * read_lock(&tasklist_lock);
97 *
98 * After we clean up the tasklist_lock and know there are no
99 * irq handlers that take it we can leave the interrupts enabled.
100 * For now it is easier to be safe than to prove it can't happen.
101 */
102
103static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
104
105void put_pid(struct pid *pid)
106{
107 struct pid_namespace *ns;
108
109 if (!pid)
110 return;
111
112 ns = pid->numbers[pid->level].ns;
113 if (refcount_dec_and_test(&pid->count)) {
114 kmem_cache_free(ns->pid_cachep, pid);
115 put_pid_ns(ns);
116 }
117}
118EXPORT_SYMBOL_GPL(put_pid);
119
120static void delayed_put_pid(struct rcu_head *rhp)
121{
122 struct pid *pid = container_of(rhp, struct pid, rcu);
123 put_pid(pid);
124}
125
126void free_pid(struct pid *pid)
127{
128 /* We can be called with write_lock_irq(&tasklist_lock) held */
129 int i;
130 unsigned long flags;
131
132 spin_lock_irqsave(&pidmap_lock, flags);
133 for (i = 0; i <= pid->level; i++) {
134 struct upid *upid = pid->numbers + i;
135 struct pid_namespace *ns = upid->ns;
136 switch (--ns->pid_allocated) {
137 case 2:
138 case 1:
139 /* When all that is left in the pid namespace
140 * is the reaper wake up the reaper. The reaper
141 * may be sleeping in zap_pid_ns_processes().
142 */
143 wake_up_process(ns->child_reaper);
144 break;
145 case PIDNS_ADDING:
146 /* Handle a fork failure of the first process */
147 WARN_ON(ns->child_reaper);
148 ns->pid_allocated = 0;
149 break;
150 }
151
152 idr_remove(&ns->idr, upid->nr);
153 }
154 spin_unlock_irqrestore(&pidmap_lock, flags);
155
156 call_rcu(&pid->rcu, delayed_put_pid);
157}
158
159struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid,
160 size_t set_tid_size)
161{
162 struct pid *pid;
163 enum pid_type type;
164 int i, nr;
165 struct pid_namespace *tmp;
166 struct upid *upid;
167 int retval = -ENOMEM;
168
169 /*
170 * set_tid_size contains the size of the set_tid array. Starting at
171 * the most nested currently active PID namespace it tells alloc_pid()
172 * which PID to set for a process in that most nested PID namespace
173 * up to set_tid_size PID namespaces. It does not have to set the PID
174 * for a process in all nested PID namespaces but set_tid_size must
175 * never be greater than the current ns->level + 1.
176 */
177 if (set_tid_size > ns->level + 1)
178 return ERR_PTR(-EINVAL);
179
180 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
181 if (!pid)
182 return ERR_PTR(retval);
183
184 tmp = ns;
185 pid->level = ns->level;
186
187 for (i = ns->level; i >= 0; i--) {
188 int tid = 0;
189
190 if (set_tid_size) {
191 tid = set_tid[ns->level - i];
192
193 retval = -EINVAL;
194 if (tid < 1 || tid >= pid_max)
195 goto out_free;
196 /*
197 * Also fail if a PID != 1 is requested and
198 * no PID 1 exists.
199 */
200 if (tid != 1 && !tmp->child_reaper)
201 goto out_free;
202 retval = -EPERM;
203 if (!checkpoint_restore_ns_capable(tmp->user_ns))
204 goto out_free;
205 set_tid_size--;
206 }
207
208 idr_preload(GFP_KERNEL);
209 spin_lock_irq(&pidmap_lock);
210
211 if (tid) {
212 nr = idr_alloc(&tmp->idr, NULL, tid,
213 tid + 1, GFP_ATOMIC);
214 /*
215 * If ENOSPC is returned it means that the PID is
216 * alreay in use. Return EEXIST in that case.
217 */
218 if (nr == -ENOSPC)
219 nr = -EEXIST;
220 } else {
221 int pid_min = 1;
222 /*
223 * init really needs pid 1, but after reaching the
224 * maximum wrap back to RESERVED_PIDS
225 */
226 if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
227 pid_min = RESERVED_PIDS;
228
229 /*
230 * Store a null pointer so find_pid_ns does not find
231 * a partially initialized PID (see below).
232 */
233 nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
234 pid_max, GFP_ATOMIC);
235 }
236 spin_unlock_irq(&pidmap_lock);
237 idr_preload_end();
238
239 if (nr < 0) {
240 retval = (nr == -ENOSPC) ? -EAGAIN : nr;
241 goto out_free;
242 }
243
244 pid->numbers[i].nr = nr;
245 pid->numbers[i].ns = tmp;
246 tmp = tmp->parent;
247 }
248
249 /*
250 * ENOMEM is not the most obvious choice especially for the case
251 * where the child subreaper has already exited and the pid
252 * namespace denies the creation of any new processes. But ENOMEM
253 * is what we have exposed to userspace for a long time and it is
254 * documented behavior for pid namespaces. So we can't easily
255 * change it even if there were an error code better suited.
256 */
257 retval = -ENOMEM;
258
259 get_pid_ns(ns);
260 refcount_set(&pid->count, 1);
261 spin_lock_init(&pid->lock);
262 for (type = 0; type < PIDTYPE_MAX; ++type)
263 INIT_HLIST_HEAD(&pid->tasks[type]);
264
265 init_waitqueue_head(&pid->wait_pidfd);
266 INIT_HLIST_HEAD(&pid->inodes);
267
268 upid = pid->numbers + ns->level;
269 spin_lock_irq(&pidmap_lock);
270 if (!(ns->pid_allocated & PIDNS_ADDING))
271 goto out_unlock;
272 for ( ; upid >= pid->numbers; --upid) {
273 /* Make the PID visible to find_pid_ns. */
274 idr_replace(&upid->ns->idr, pid, upid->nr);
275 upid->ns->pid_allocated++;
276 }
277 spin_unlock_irq(&pidmap_lock);
278
279 return pid;
280
281out_unlock:
282 spin_unlock_irq(&pidmap_lock);
283 put_pid_ns(ns);
284
285out_free:
286 spin_lock_irq(&pidmap_lock);
287 while (++i <= ns->level) {
288 upid = pid->numbers + i;
289 idr_remove(&upid->ns->idr, upid->nr);
290 }
291
292 /* On failure to allocate the first pid, reset the state */
293 if (ns->pid_allocated == PIDNS_ADDING)
294 idr_set_cursor(&ns->idr, 0);
295
296 spin_unlock_irq(&pidmap_lock);
297
298 kmem_cache_free(ns->pid_cachep, pid);
299 return ERR_PTR(retval);
300}
301
302void disable_pid_allocation(struct pid_namespace *ns)
303{
304 spin_lock_irq(&pidmap_lock);
305 ns->pid_allocated &= ~PIDNS_ADDING;
306 spin_unlock_irq(&pidmap_lock);
307}
308
309struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
310{
311 return idr_find(&ns->idr, nr);
312}
313EXPORT_SYMBOL_GPL(find_pid_ns);
314
315struct pid *find_vpid(int nr)
316{
317 return find_pid_ns(nr, task_active_pid_ns(current));
318}
319EXPORT_SYMBOL_GPL(find_vpid);
320
321static struct pid **task_pid_ptr(struct task_struct *task, enum pid_type type)
322{
323 return (type == PIDTYPE_PID) ?
324 &task->thread_pid :
325 &task->signal->pids[type];
326}
327
328/*
329 * attach_pid() must be called with the tasklist_lock write-held.
330 */
331void attach_pid(struct task_struct *task, enum pid_type type)
332{
333 struct pid *pid = *task_pid_ptr(task, type);
334 hlist_add_head_rcu(&task->pid_links[type], &pid->tasks[type]);
335}
336
337static void __change_pid(struct task_struct *task, enum pid_type type,
338 struct pid *new)
339{
340 struct pid **pid_ptr = task_pid_ptr(task, type);
341 struct pid *pid;
342 int tmp;
343
344 pid = *pid_ptr;
345
346 hlist_del_rcu(&task->pid_links[type]);
347 *pid_ptr = new;
348
349 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
350 if (pid_has_task(pid, tmp))
351 return;
352
353 free_pid(pid);
354}
355
356void detach_pid(struct task_struct *task, enum pid_type type)
357{
358 __change_pid(task, type, NULL);
359}
360
361void change_pid(struct task_struct *task, enum pid_type type,
362 struct pid *pid)
363{
364 __change_pid(task, type, pid);
365 attach_pid(task, type);
366}
367
368void exchange_tids(struct task_struct *left, struct task_struct *right)
369{
370 struct pid *pid1 = left->thread_pid;
371 struct pid *pid2 = right->thread_pid;
372 struct hlist_head *head1 = &pid1->tasks[PIDTYPE_PID];
373 struct hlist_head *head2 = &pid2->tasks[PIDTYPE_PID];
374
375 /* Swap the single entry tid lists */
376 hlists_swap_heads_rcu(head1, head2);
377
378 /* Swap the per task_struct pid */
379 rcu_assign_pointer(left->thread_pid, pid2);
380 rcu_assign_pointer(right->thread_pid, pid1);
381
382 /* Swap the cached value */
383 WRITE_ONCE(left->pid, pid_nr(pid2));
384 WRITE_ONCE(right->pid, pid_nr(pid1));
385}
386
387/* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
388void transfer_pid(struct task_struct *old, struct task_struct *new,
389 enum pid_type type)
390{
391 if (type == PIDTYPE_PID)
392 new->thread_pid = old->thread_pid;
393 hlist_replace_rcu(&old->pid_links[type], &new->pid_links[type]);
394}
395
396struct task_struct *pid_task(struct pid *pid, enum pid_type type)
397{
398 struct task_struct *result = NULL;
399 if (pid) {
400 struct hlist_node *first;
401 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
402 lockdep_tasklist_lock_is_held());
403 if (first)
404 result = hlist_entry(first, struct task_struct, pid_links[(type)]);
405 }
406 return result;
407}
408EXPORT_SYMBOL(pid_task);
409
410/*
411 * Must be called under rcu_read_lock().
412 */
413struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
414{
415 RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
416 "find_task_by_pid_ns() needs rcu_read_lock() protection");
417 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
418}
419
420struct task_struct *find_task_by_vpid(pid_t vnr)
421{
422 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
423}
424
425struct task_struct *find_get_task_by_vpid(pid_t nr)
426{
427 struct task_struct *task;
428
429 rcu_read_lock();
430 task = find_task_by_vpid(nr);
431 if (task)
432 get_task_struct(task);
433 rcu_read_unlock();
434
435 return task;
436}
437
438struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
439{
440 struct pid *pid;
441 rcu_read_lock();
442 pid = get_pid(rcu_dereference(*task_pid_ptr(task, type)));
443 rcu_read_unlock();
444 return pid;
445}
446EXPORT_SYMBOL_GPL(get_task_pid);
447
448struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
449{
450 struct task_struct *result;
451 rcu_read_lock();
452 result = pid_task(pid, type);
453 if (result)
454 get_task_struct(result);
455 rcu_read_unlock();
456 return result;
457}
458EXPORT_SYMBOL_GPL(get_pid_task);
459
460struct pid *find_get_pid(pid_t nr)
461{
462 struct pid *pid;
463
464 rcu_read_lock();
465 pid = get_pid(find_vpid(nr));
466 rcu_read_unlock();
467
468 return pid;
469}
470EXPORT_SYMBOL_GPL(find_get_pid);
471
472pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
473{
474 struct upid *upid;
475 pid_t nr = 0;
476
477 if (pid && ns->level <= pid->level) {
478 upid = &pid->numbers[ns->level];
479 if (upid->ns == ns)
480 nr = upid->nr;
481 }
482 return nr;
483}
484EXPORT_SYMBOL_GPL(pid_nr_ns);
485
486pid_t pid_vnr(struct pid *pid)
487{
488 return pid_nr_ns(pid, task_active_pid_ns(current));
489}
490EXPORT_SYMBOL_GPL(pid_vnr);
491
492pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
493 struct pid_namespace *ns)
494{
495 pid_t nr = 0;
496
497 rcu_read_lock();
498 if (!ns)
499 ns = task_active_pid_ns(current);
500 nr = pid_nr_ns(rcu_dereference(*task_pid_ptr(task, type)), ns);
501 rcu_read_unlock();
502
503 return nr;
504}
505EXPORT_SYMBOL(__task_pid_nr_ns);
506
507struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
508{
509 return ns_of_pid(task_pid(tsk));
510}
511EXPORT_SYMBOL_GPL(task_active_pid_ns);
512
513/*
514 * Used by proc to find the first pid that is greater than or equal to nr.
515 *
516 * If there is a pid at nr this function is exactly the same as find_pid_ns.
517 */
518struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
519{
520 return idr_get_next(&ns->idr, &nr);
521}
522EXPORT_SYMBOL_GPL(find_ge_pid);
523
524struct pid *pidfd_get_pid(unsigned int fd, unsigned int *flags)
525{
526 struct fd f;
527 struct pid *pid;
528
529 f = fdget(fd);
530 if (!f.file)
531 return ERR_PTR(-EBADF);
532
533 pid = pidfd_pid(f.file);
534 if (!IS_ERR(pid)) {
535 get_pid(pid);
536 *flags = f.file->f_flags;
537 }
538
539 fdput(f);
540 return pid;
541}
542
543/**
544 * pidfd_get_task() - Get the task associated with a pidfd
545 *
546 * @pidfd: pidfd for which to get the task
547 * @flags: flags associated with this pidfd
548 *
549 * Return the task associated with @pidfd. The function takes a reference on
550 * the returned task. The caller is responsible for releasing that reference.
551 *
552 * Currently, the process identified by @pidfd is always a thread-group leader.
553 * This restriction currently exists for all aspects of pidfds including pidfd
554 * creation (CLONE_PIDFD cannot be used with CLONE_THREAD) and pidfd polling
555 * (only supports thread group leaders).
556 *
557 * Return: On success, the task_struct associated with the pidfd.
558 * On error, a negative errno number will be returned.
559 */
560struct task_struct *pidfd_get_task(int pidfd, unsigned int *flags)
561{
562 unsigned int f_flags;
563 struct pid *pid;
564 struct task_struct *task;
565
566 pid = pidfd_get_pid(pidfd, &f_flags);
567 if (IS_ERR(pid))
568 return ERR_CAST(pid);
569
570 task = get_pid_task(pid, PIDTYPE_TGID);
571 put_pid(pid);
572 if (!task)
573 return ERR_PTR(-ESRCH);
574
575 *flags = f_flags;
576 return task;
577}
578
579/**
580 * pidfd_create() - Create a new pid file descriptor.
581 *
582 * @pid: struct pid that the pidfd will reference
583 * @flags: flags to pass
584 *
585 * This creates a new pid file descriptor with the O_CLOEXEC flag set.
586 *
587 * Note, that this function can only be called after the fd table has
588 * been unshared to avoid leaking the pidfd to the new process.
589 *
590 * This symbol should not be explicitly exported to loadable modules.
591 *
592 * Return: On success, a cloexec pidfd is returned.
593 * On error, a negative errno number will be returned.
594 */
595int pidfd_create(struct pid *pid, unsigned int flags)
596{
597 int fd;
598
599 if (!pid || !pid_has_task(pid, PIDTYPE_TGID))
600 return -EINVAL;
601
602 if (flags & ~(O_NONBLOCK | O_RDWR | O_CLOEXEC))
603 return -EINVAL;
604
605 fd = anon_inode_getfd("[pidfd]", &pidfd_fops, get_pid(pid),
606 flags | O_RDWR | O_CLOEXEC);
607 if (fd < 0)
608 put_pid(pid);
609
610 return fd;
611}
612
613/**
614 * pidfd_open() - Open new pid file descriptor.
615 *
616 * @pid: pid for which to retrieve a pidfd
617 * @flags: flags to pass
618 *
619 * This creates a new pid file descriptor with the O_CLOEXEC flag set for
620 * the process identified by @pid. Currently, the process identified by
621 * @pid must be a thread-group leader. This restriction currently exists
622 * for all aspects of pidfds including pidfd creation (CLONE_PIDFD cannot
623 * be used with CLONE_THREAD) and pidfd polling (only supports thread group
624 * leaders).
625 *
626 * Return: On success, a cloexec pidfd is returned.
627 * On error, a negative errno number will be returned.
628 */
629SYSCALL_DEFINE2(pidfd_open, pid_t, pid, unsigned int, flags)
630{
631 int fd;
632 struct pid *p;
633
634 if (flags & ~PIDFD_NONBLOCK)
635 return -EINVAL;
636
637 if (pid <= 0)
638 return -EINVAL;
639
640 p = find_get_pid(pid);
641 if (!p)
642 return -ESRCH;
643
644 fd = pidfd_create(p, flags);
645
646 put_pid(p);
647 return fd;
648}
649
650void __init pid_idr_init(void)
651{
652 /* Verify no one has done anything silly: */
653 BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_ADDING);
654
655 /* bump default and minimum pid_max based on number of cpus */
656 pid_max = min(pid_max_max, max_t(int, pid_max,
657 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
658 pid_max_min = max_t(int, pid_max_min,
659 PIDS_PER_CPU_MIN * num_possible_cpus());
660 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
661
662 idr_init(&init_pid_ns.idr);
663
664 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
665 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
666}
667
668static struct file *__pidfd_fget(struct task_struct *task, int fd)
669{
670 struct file *file;
671 int ret;
672
673 ret = down_read_killable(&task->signal->exec_update_lock);
674 if (ret)
675 return ERR_PTR(ret);
676
677 if (ptrace_may_access(task, PTRACE_MODE_ATTACH_REALCREDS))
678 file = fget_task(task, fd);
679 else
680 file = ERR_PTR(-EPERM);
681
682 up_read(&task->signal->exec_update_lock);
683
684 return file ?: ERR_PTR(-EBADF);
685}
686
687static int pidfd_getfd(struct pid *pid, int fd)
688{
689 struct task_struct *task;
690 struct file *file;
691 int ret;
692
693 task = get_pid_task(pid, PIDTYPE_PID);
694 if (!task)
695 return -ESRCH;
696
697 file = __pidfd_fget(task, fd);
698 put_task_struct(task);
699 if (IS_ERR(file))
700 return PTR_ERR(file);
701
702 ret = receive_fd(file, O_CLOEXEC);
703 fput(file);
704
705 return ret;
706}
707
708/**
709 * sys_pidfd_getfd() - Get a file descriptor from another process
710 *
711 * @pidfd: the pidfd file descriptor of the process
712 * @fd: the file descriptor number to get
713 * @flags: flags on how to get the fd (reserved)
714 *
715 * This syscall gets a copy of a file descriptor from another process
716 * based on the pidfd, and file descriptor number. It requires that
717 * the calling process has the ability to ptrace the process represented
718 * by the pidfd. The process which is having its file descriptor copied
719 * is otherwise unaffected.
720 *
721 * Return: On success, a cloexec file descriptor is returned.
722 * On error, a negative errno number will be returned.
723 */
724SYSCALL_DEFINE3(pidfd_getfd, int, pidfd, int, fd,
725 unsigned int, flags)
726{
727 struct pid *pid;
728 struct fd f;
729 int ret;
730
731 /* flags is currently unused - make sure it's unset */
732 if (flags)
733 return -EINVAL;
734
735 f = fdget(pidfd);
736 if (!f.file)
737 return -EBADF;
738
739 pid = pidfd_pid(f.file);
740 if (IS_ERR(pid))
741 ret = PTR_ERR(pid);
742 else
743 ret = pidfd_getfd(pid, fd);
744
745 fdput(f);
746 return ret;
747}
1/*
2 * Generic pidhash and scalable, time-bounded PID allocator
3 *
4 * (C) 2002-2003 William Irwin, IBM
5 * (C) 2004 William Irwin, Oracle
6 * (C) 2002-2004 Ingo Molnar, Red Hat
7 *
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.
11 *
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.
15 *
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).
21 *
22 * Pid namespaces:
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
26 *
27 */
28
29#include <linux/mm.h>
30#include <linux/module.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
40#define pid_hashfn(nr, ns) \
41 hash_long((unsigned long)nr + (unsigned long)ns, pidhash_shift)
42static struct hlist_head *pid_hash;
43static unsigned int pidhash_shift = 4;
44struct pid init_struct_pid = INIT_STRUCT_PID;
45
46int pid_max = PID_MAX_DEFAULT;
47
48#define RESERVED_PIDS 300
49
50int pid_max_min = RESERVED_PIDS + 1;
51int pid_max_max = PID_MAX_LIMIT;
52
53#define BITS_PER_PAGE (PAGE_SIZE*8)
54#define BITS_PER_PAGE_MASK (BITS_PER_PAGE-1)
55
56static inline int mk_pid(struct pid_namespace *pid_ns,
57 struct pidmap *map, int off)
58{
59 return (map - pid_ns->pidmap)*BITS_PER_PAGE + off;
60}
61
62#define find_next_offset(map, off) \
63 find_next_zero_bit((map)->page, BITS_PER_PAGE, off)
64
65/*
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.
70 */
71struct pid_namespace init_pid_ns = {
72 .kref = {
73 .refcount = ATOMIC_INIT(2),
74 },
75 .pidmap = {
76 [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }
77 },
78 .last_pid = 0,
79 .level = 0,
80 .child_reaper = &init_task,
81};
82EXPORT_SYMBOL_GPL(init_pid_ns);
83
84int is_container_init(struct task_struct *tsk)
85{
86 int ret = 0;
87 struct pid *pid;
88
89 rcu_read_lock();
90 pid = task_pid(tsk);
91 if (pid != NULL && pid->numbers[pid->level].nr == 1)
92 ret = 1;
93 rcu_read_unlock();
94
95 return ret;
96}
97EXPORT_SYMBOL(is_container_init);
98
99/*
100 * Note: disable interrupts while the pidmap_lock is held as an
101 * interrupt might come in and do read_lock(&tasklist_lock).
102 *
103 * If we don't disable interrupts there is a nasty deadlock between
104 * detach_pid()->free_pid() and another cpu that does
105 * spin_lock(&pidmap_lock) followed by an interrupt routine that does
106 * read_lock(&tasklist_lock);
107 *
108 * After we clean up the tasklist_lock and know there are no
109 * irq handlers that take it we can leave the interrupts enabled.
110 * For now it is easier to be safe than to prove it can't happen.
111 */
112
113static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
114
115static void free_pidmap(struct upid *upid)
116{
117 int nr = upid->nr;
118 struct pidmap *map = upid->ns->pidmap + nr / BITS_PER_PAGE;
119 int offset = nr & BITS_PER_PAGE_MASK;
120
121 clear_bit(offset, map->page);
122 atomic_inc(&map->nr_free);
123}
124
125/*
126 * If we started walking pids at 'base', is 'a' seen before 'b'?
127 */
128static int pid_before(int base, int a, int b)
129{
130 /*
131 * This is the same as saying
132 *
133 * (a - base + MAXUINT) % MAXUINT < (b - base + MAXUINT) % MAXUINT
134 * and that mapping orders 'a' and 'b' with respect to 'base'.
135 */
136 return (unsigned)(a - base) < (unsigned)(b - base);
137}
138
139/*
140 * We might be racing with someone else trying to set pid_ns->last_pid.
141 * We want the winner to have the "later" value, because if the
142 * "earlier" value prevails, then a pid may get reused immediately.
143 *
144 * Since pids rollover, it is not sufficient to just pick the bigger
145 * value. We have to consider where we started counting from.
146 *
147 * 'base' is the value of pid_ns->last_pid that we observed when
148 * we started looking for a pid.
149 *
150 * 'pid' is the pid that we eventually found.
151 */
152static void set_last_pid(struct pid_namespace *pid_ns, int base, int pid)
153{
154 int prev;
155 int last_write = base;
156 do {
157 prev = last_write;
158 last_write = cmpxchg(&pid_ns->last_pid, prev, pid);
159 } while ((prev != last_write) && (pid_before(base, last_write, pid)));
160}
161
162static int alloc_pidmap(struct pid_namespace *pid_ns)
163{
164 int i, offset, max_scan, pid, last = pid_ns->last_pid;
165 struct pidmap *map;
166
167 pid = last + 1;
168 if (pid >= pid_max)
169 pid = RESERVED_PIDS;
170 offset = pid & BITS_PER_PAGE_MASK;
171 map = &pid_ns->pidmap[pid/BITS_PER_PAGE];
172 /*
173 * If last_pid points into the middle of the map->page we
174 * want to scan this bitmap block twice, the second time
175 * we start with offset == 0 (or RESERVED_PIDS).
176 */
177 max_scan = DIV_ROUND_UP(pid_max, BITS_PER_PAGE) - !offset;
178 for (i = 0; i <= max_scan; ++i) {
179 if (unlikely(!map->page)) {
180 void *page = kzalloc(PAGE_SIZE, GFP_KERNEL);
181 /*
182 * Free the page if someone raced with us
183 * installing it:
184 */
185 spin_lock_irq(&pidmap_lock);
186 if (!map->page) {
187 map->page = page;
188 page = NULL;
189 }
190 spin_unlock_irq(&pidmap_lock);
191 kfree(page);
192 if (unlikely(!map->page))
193 break;
194 }
195 if (likely(atomic_read(&map->nr_free))) {
196 do {
197 if (!test_and_set_bit(offset, map->page)) {
198 atomic_dec(&map->nr_free);
199 set_last_pid(pid_ns, last, pid);
200 return pid;
201 }
202 offset = find_next_offset(map, offset);
203 pid = mk_pid(pid_ns, map, offset);
204 } while (offset < BITS_PER_PAGE && pid < pid_max);
205 }
206 if (map < &pid_ns->pidmap[(pid_max-1)/BITS_PER_PAGE]) {
207 ++map;
208 offset = 0;
209 } else {
210 map = &pid_ns->pidmap[0];
211 offset = RESERVED_PIDS;
212 if (unlikely(last == offset))
213 break;
214 }
215 pid = mk_pid(pid_ns, map, offset);
216 }
217 return -1;
218}
219
220int next_pidmap(struct pid_namespace *pid_ns, unsigned int last)
221{
222 int offset;
223 struct pidmap *map, *end;
224
225 if (last >= PID_MAX_LIMIT)
226 return -1;
227
228 offset = (last + 1) & BITS_PER_PAGE_MASK;
229 map = &pid_ns->pidmap[(last + 1)/BITS_PER_PAGE];
230 end = &pid_ns->pidmap[PIDMAP_ENTRIES];
231 for (; map < end; map++, offset = 0) {
232 if (unlikely(!map->page))
233 continue;
234 offset = find_next_bit((map)->page, BITS_PER_PAGE, offset);
235 if (offset < BITS_PER_PAGE)
236 return mk_pid(pid_ns, map, offset);
237 }
238 return -1;
239}
240
241void put_pid(struct pid *pid)
242{
243 struct pid_namespace *ns;
244
245 if (!pid)
246 return;
247
248 ns = pid->numbers[pid->level].ns;
249 if ((atomic_read(&pid->count) == 1) ||
250 atomic_dec_and_test(&pid->count)) {
251 kmem_cache_free(ns->pid_cachep, pid);
252 put_pid_ns(ns);
253 }
254}
255EXPORT_SYMBOL_GPL(put_pid);
256
257static void delayed_put_pid(struct rcu_head *rhp)
258{
259 struct pid *pid = container_of(rhp, struct pid, rcu);
260 put_pid(pid);
261}
262
263void free_pid(struct pid *pid)
264{
265 /* We can be called with write_lock_irq(&tasklist_lock) held */
266 int i;
267 unsigned long flags;
268
269 spin_lock_irqsave(&pidmap_lock, flags);
270 for (i = 0; i <= pid->level; i++)
271 hlist_del_rcu(&pid->numbers[i].pid_chain);
272 spin_unlock_irqrestore(&pidmap_lock, flags);
273
274 for (i = 0; i <= pid->level; i++)
275 free_pidmap(pid->numbers + i);
276
277 call_rcu(&pid->rcu, delayed_put_pid);
278}
279
280struct pid *alloc_pid(struct pid_namespace *ns)
281{
282 struct pid *pid;
283 enum pid_type type;
284 int i, nr;
285 struct pid_namespace *tmp;
286 struct upid *upid;
287
288 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
289 if (!pid)
290 goto out;
291
292 tmp = ns;
293 for (i = ns->level; i >= 0; i--) {
294 nr = alloc_pidmap(tmp);
295 if (nr < 0)
296 goto out_free;
297
298 pid->numbers[i].nr = nr;
299 pid->numbers[i].ns = tmp;
300 tmp = tmp->parent;
301 }
302
303 get_pid_ns(ns);
304 pid->level = ns->level;
305 atomic_set(&pid->count, 1);
306 for (type = 0; type < PIDTYPE_MAX; ++type)
307 INIT_HLIST_HEAD(&pid->tasks[type]);
308
309 upid = pid->numbers + ns->level;
310 spin_lock_irq(&pidmap_lock);
311 for ( ; upid >= pid->numbers; --upid)
312 hlist_add_head_rcu(&upid->pid_chain,
313 &pid_hash[pid_hashfn(upid->nr, upid->ns)]);
314 spin_unlock_irq(&pidmap_lock);
315
316out:
317 return pid;
318
319out_free:
320 while (++i <= ns->level)
321 free_pidmap(pid->numbers + i);
322
323 kmem_cache_free(ns->pid_cachep, pid);
324 pid = NULL;
325 goto out;
326}
327
328struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
329{
330 struct hlist_node *elem;
331 struct upid *pnr;
332
333 hlist_for_each_entry_rcu(pnr, elem,
334 &pid_hash[pid_hashfn(nr, ns)], pid_chain)
335 if (pnr->nr == nr && pnr->ns == ns)
336 return container_of(pnr, struct pid,
337 numbers[ns->level]);
338
339 return NULL;
340}
341EXPORT_SYMBOL_GPL(find_pid_ns);
342
343struct pid *find_vpid(int nr)
344{
345 return find_pid_ns(nr, current->nsproxy->pid_ns);
346}
347EXPORT_SYMBOL_GPL(find_vpid);
348
349/*
350 * attach_pid() must be called with the tasklist_lock write-held.
351 */
352void attach_pid(struct task_struct *task, enum pid_type type,
353 struct pid *pid)
354{
355 struct pid_link *link;
356
357 link = &task->pids[type];
358 link->pid = pid;
359 hlist_add_head_rcu(&link->node, &pid->tasks[type]);
360}
361
362static void __change_pid(struct task_struct *task, enum pid_type type,
363 struct pid *new)
364{
365 struct pid_link *link;
366 struct pid *pid;
367 int tmp;
368
369 link = &task->pids[type];
370 pid = link->pid;
371
372 hlist_del_rcu(&link->node);
373 link->pid = new;
374
375 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
376 if (!hlist_empty(&pid->tasks[tmp]))
377 return;
378
379 free_pid(pid);
380}
381
382void detach_pid(struct task_struct *task, enum pid_type type)
383{
384 __change_pid(task, type, NULL);
385}
386
387void change_pid(struct task_struct *task, enum pid_type type,
388 struct pid *pid)
389{
390 __change_pid(task, type, pid);
391 attach_pid(task, type, pid);
392}
393
394/* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
395void transfer_pid(struct task_struct *old, struct task_struct *new,
396 enum pid_type type)
397{
398 new->pids[type].pid = old->pids[type].pid;
399 hlist_replace_rcu(&old->pids[type].node, &new->pids[type].node);
400}
401
402struct task_struct *pid_task(struct pid *pid, enum pid_type type)
403{
404 struct task_struct *result = NULL;
405 if (pid) {
406 struct hlist_node *first;
407 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
408 lockdep_tasklist_lock_is_held());
409 if (first)
410 result = hlist_entry(first, struct task_struct, pids[(type)].node);
411 }
412 return result;
413}
414EXPORT_SYMBOL(pid_task);
415
416/*
417 * Must be called under rcu_read_lock().
418 */
419struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
420{
421 rcu_lockdep_assert(rcu_read_lock_held());
422 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
423}
424
425struct task_struct *find_task_by_vpid(pid_t vnr)
426{
427 return find_task_by_pid_ns(vnr, current->nsproxy->pid_ns);
428}
429
430struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
431{
432 struct pid *pid;
433 rcu_read_lock();
434 if (type != PIDTYPE_PID)
435 task = task->group_leader;
436 pid = get_pid(task->pids[type].pid);
437 rcu_read_unlock();
438 return pid;
439}
440EXPORT_SYMBOL_GPL(get_task_pid);
441
442struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
443{
444 struct task_struct *result;
445 rcu_read_lock();
446 result = pid_task(pid, type);
447 if (result)
448 get_task_struct(result);
449 rcu_read_unlock();
450 return result;
451}
452EXPORT_SYMBOL_GPL(get_pid_task);
453
454struct pid *find_get_pid(pid_t nr)
455{
456 struct pid *pid;
457
458 rcu_read_lock();
459 pid = get_pid(find_vpid(nr));
460 rcu_read_unlock();
461
462 return pid;
463}
464EXPORT_SYMBOL_GPL(find_get_pid);
465
466pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
467{
468 struct upid *upid;
469 pid_t nr = 0;
470
471 if (pid && ns->level <= pid->level) {
472 upid = &pid->numbers[ns->level];
473 if (upid->ns == ns)
474 nr = upid->nr;
475 }
476 return nr;
477}
478
479pid_t pid_vnr(struct pid *pid)
480{
481 return pid_nr_ns(pid, current->nsproxy->pid_ns);
482}
483EXPORT_SYMBOL_GPL(pid_vnr);
484
485pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
486 struct pid_namespace *ns)
487{
488 pid_t nr = 0;
489
490 rcu_read_lock();
491 if (!ns)
492 ns = current->nsproxy->pid_ns;
493 if (likely(pid_alive(task))) {
494 if (type != PIDTYPE_PID)
495 task = task->group_leader;
496 nr = pid_nr_ns(task->pids[type].pid, ns);
497 }
498 rcu_read_unlock();
499
500 return nr;
501}
502EXPORT_SYMBOL(__task_pid_nr_ns);
503
504pid_t task_tgid_nr_ns(struct task_struct *tsk, struct pid_namespace *ns)
505{
506 return pid_nr_ns(task_tgid(tsk), ns);
507}
508EXPORT_SYMBOL(task_tgid_nr_ns);
509
510struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
511{
512 return ns_of_pid(task_pid(tsk));
513}
514EXPORT_SYMBOL_GPL(task_active_pid_ns);
515
516/*
517 * Used by proc to find the first pid that is greater than or equal to nr.
518 *
519 * If there is a pid at nr this function is exactly the same as find_pid_ns.
520 */
521struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
522{
523 struct pid *pid;
524
525 do {
526 pid = find_pid_ns(nr, ns);
527 if (pid)
528 break;
529 nr = next_pidmap(ns, nr);
530 } while (nr > 0);
531
532 return pid;
533}
534
535/*
536 * The pid hash table is scaled according to the amount of memory in the
537 * machine. From a minimum of 16 slots up to 4096 slots at one gigabyte or
538 * more.
539 */
540void __init pidhash_init(void)
541{
542 int i, pidhash_size;
543
544 pid_hash = alloc_large_system_hash("PID", sizeof(*pid_hash), 0, 18,
545 HASH_EARLY | HASH_SMALL,
546 &pidhash_shift, NULL, 4096);
547 pidhash_size = 1 << pidhash_shift;
548
549 for (i = 0; i < pidhash_size; i++)
550 INIT_HLIST_HEAD(&pid_hash[i]);
551}
552
553void __init pidmap_init(void)
554{
555 /* bump default and minimum pid_max based on number of cpus */
556 pid_max = min(pid_max_max, max_t(int, pid_max,
557 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
558 pid_max_min = max_t(int, pid_max_min,
559 PIDS_PER_CPU_MIN * num_possible_cpus());
560 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
561
562 init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL);
563 /* Reserve PID 0. We never call free_pidmap(0) */
564 set_bit(0, init_pid_ns.pidmap[0].page);
565 atomic_dec(&init_pid_ns.pidmap[0].nr_free);
566
567 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
568 SLAB_HWCACHE_ALIGN | SLAB_PANIC);
569}