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// 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
47struct pid init_struct_pid = {
48 .count = REFCOUNT_INIT(1),
49 .tasks = {
50 { .first = NULL },
51 { .first = NULL },
52 { .first = NULL },
53 },
54 .level = 0,
55 .numbers = { {
56 .nr = 0,
57 .ns = &init_pid_ns,
58 }, }
59};
60
61int pid_max = PID_MAX_DEFAULT;
62
63#define RESERVED_PIDS 300
64
65int pid_max_min = RESERVED_PIDS + 1;
66int pid_max_max = PID_MAX_LIMIT;
67
68/*
69 * PID-map pages start out as NULL, they get allocated upon
70 * first use and are never deallocated. This way a low pid_max
71 * value does not cause lots of bitmaps to be allocated, but
72 * the scheme scales to up to 4 million PIDs, runtime.
73 */
74struct pid_namespace init_pid_ns = {
75 .kref = KREF_INIT(2),
76 .idr = IDR_INIT(init_pid_ns.idr),
77 .pid_allocated = PIDNS_ADDING,
78 .level = 0,
79 .child_reaper = &init_task,
80 .user_ns = &init_user_ns,
81 .ns.inum = PROC_PID_INIT_INO,
82#ifdef CONFIG_PID_NS
83 .ns.ops = &pidns_operations,
84#endif
85};
86EXPORT_SYMBOL_GPL(init_pid_ns);
87
88/*
89 * Note: disable interrupts while the pidmap_lock is held as an
90 * interrupt might come in and do read_lock(&tasklist_lock).
91 *
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);
96 *
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.
100 */
101
102static __cacheline_aligned_in_smp DEFINE_SPINLOCK(pidmap_lock);
103
104void put_pid(struct pid *pid)
105{
106 struct pid_namespace *ns;
107
108 if (!pid)
109 return;
110
111 ns = pid->numbers[pid->level].ns;
112 if (refcount_dec_and_test(&pid->count)) {
113 kmem_cache_free(ns->pid_cachep, pid);
114 put_pid_ns(ns);
115 }
116}
117EXPORT_SYMBOL_GPL(put_pid);
118
119static void delayed_put_pid(struct rcu_head *rhp)
120{
121 struct pid *pid = container_of(rhp, struct pid, rcu);
122 put_pid(pid);
123}
124
125void free_pid(struct pid *pid)
126{
127 /* We can be called with write_lock_irq(&tasklist_lock) held */
128 int i;
129 unsigned long flags;
130
131 spin_lock_irqsave(&pidmap_lock, flags);
132 for (i = 0; i <= pid->level; i++) {
133 struct upid *upid = pid->numbers + i;
134 struct pid_namespace *ns = upid->ns;
135 switch (--ns->pid_allocated) {
136 case 2:
137 case 1:
138 /* When all that is left in the pid namespace
139 * is the reaper wake up the reaper. The reaper
140 * may be sleeping in zap_pid_ns_processes().
141 */
142 wake_up_process(ns->child_reaper);
143 break;
144 case PIDNS_ADDING:
145 /* Handle a fork failure of the first process */
146 WARN_ON(ns->child_reaper);
147 ns->pid_allocated = 0;
148 break;
149 }
150
151 idr_remove(&ns->idr, upid->nr);
152 }
153 spin_unlock_irqrestore(&pidmap_lock, flags);
154
155 call_rcu(&pid->rcu, delayed_put_pid);
156}
157
158struct pid *alloc_pid(struct pid_namespace *ns, pid_t *set_tid,
159 size_t set_tid_size)
160{
161 struct pid *pid;
162 enum pid_type type;
163 int i, nr;
164 struct pid_namespace *tmp;
165 struct upid *upid;
166 int retval = -ENOMEM;
167
168 /*
169 * set_tid_size contains the size of the set_tid array. Starting at
170 * the most nested currently active PID namespace it tells alloc_pid()
171 * which PID to set for a process in that most nested PID namespace
172 * up to set_tid_size PID namespaces. It does not have to set the PID
173 * for a process in all nested PID namespaces but set_tid_size must
174 * never be greater than the current ns->level + 1.
175 */
176 if (set_tid_size > ns->level + 1)
177 return ERR_PTR(-EINVAL);
178
179 pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
180 if (!pid)
181 return ERR_PTR(retval);
182
183 tmp = ns;
184 pid->level = ns->level;
185
186 for (i = ns->level; i >= 0; i--) {
187 int tid = 0;
188
189 if (set_tid_size) {
190 tid = set_tid[ns->level - i];
191
192 retval = -EINVAL;
193 if (tid < 1 || tid >= pid_max)
194 goto out_free;
195 /*
196 * Also fail if a PID != 1 is requested and
197 * no PID 1 exists.
198 */
199 if (tid != 1 && !tmp->child_reaper)
200 goto out_free;
201 retval = -EPERM;
202 if (!checkpoint_restore_ns_capable(tmp->user_ns))
203 goto out_free;
204 set_tid_size--;
205 }
206
207 idr_preload(GFP_KERNEL);
208 spin_lock_irq(&pidmap_lock);
209
210 if (tid) {
211 nr = idr_alloc(&tmp->idr, NULL, tid,
212 tid + 1, GFP_ATOMIC);
213 /*
214 * If ENOSPC is returned it means that the PID is
215 * alreay in use. Return EEXIST in that case.
216 */
217 if (nr == -ENOSPC)
218 nr = -EEXIST;
219 } else {
220 int pid_min = 1;
221 /*
222 * init really needs pid 1, but after reaching the
223 * maximum wrap back to RESERVED_PIDS
224 */
225 if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
226 pid_min = RESERVED_PIDS;
227
228 /*
229 * Store a null pointer so find_pid_ns does not find
230 * a partially initialized PID (see below).
231 */
232 nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
233 pid_max, GFP_ATOMIC);
234 }
235 spin_unlock_irq(&pidmap_lock);
236 idr_preload_end();
237
238 if (nr < 0) {
239 retval = (nr == -ENOSPC) ? -EAGAIN : nr;
240 goto out_free;
241 }
242
243 pid->numbers[i].nr = nr;
244 pid->numbers[i].ns = tmp;
245 tmp = tmp->parent;
246 }
247
248 /*
249 * ENOMEM is not the most obvious choice especially for the case
250 * where the child subreaper has already exited and the pid
251 * namespace denies the creation of any new processes. But ENOMEM
252 * is what we have exposed to userspace for a long time and it is
253 * documented behavior for pid namespaces. So we can't easily
254 * change it even if there were an error code better suited.
255 */
256 retval = -ENOMEM;
257
258 get_pid_ns(ns);
259 refcount_set(&pid->count, 1);
260 spin_lock_init(&pid->lock);
261 for (type = 0; type < PIDTYPE_MAX; ++type)
262 INIT_HLIST_HEAD(&pid->tasks[type]);
263
264 init_waitqueue_head(&pid->wait_pidfd);
265 INIT_HLIST_HEAD(&pid->inodes);
266
267 upid = pid->numbers + ns->level;
268 spin_lock_irq(&pidmap_lock);
269 if (!(ns->pid_allocated & PIDNS_ADDING))
270 goto out_unlock;
271 for ( ; upid >= pid->numbers; --upid) {
272 /* Make the PID visible to find_pid_ns. */
273 idr_replace(&upid->ns->idr, pid, upid->nr);
274 upid->ns->pid_allocated++;
275 }
276 spin_unlock_irq(&pidmap_lock);
277
278 return pid;
279
280out_unlock:
281 spin_unlock_irq(&pidmap_lock);
282 put_pid_ns(ns);
283
284out_free:
285 spin_lock_irq(&pidmap_lock);
286 while (++i <= ns->level) {
287 upid = pid->numbers + i;
288 idr_remove(&upid->ns->idr, upid->nr);
289 }
290
291 /* On failure to allocate the first pid, reset the state */
292 if (ns->pid_allocated == PIDNS_ADDING)
293 idr_set_cursor(&ns->idr, 0);
294
295 spin_unlock_irq(&pidmap_lock);
296
297 kmem_cache_free(ns->pid_cachep, pid);
298 return ERR_PTR(retval);
299}
300
301void disable_pid_allocation(struct pid_namespace *ns)
302{
303 spin_lock_irq(&pidmap_lock);
304 ns->pid_allocated &= ~PIDNS_ADDING;
305 spin_unlock_irq(&pidmap_lock);
306}
307
308struct pid *find_pid_ns(int nr, struct pid_namespace *ns)
309{
310 return idr_find(&ns->idr, nr);
311}
312EXPORT_SYMBOL_GPL(find_pid_ns);
313
314struct pid *find_vpid(int nr)
315{
316 return find_pid_ns(nr, task_active_pid_ns(current));
317}
318EXPORT_SYMBOL_GPL(find_vpid);
319
320static struct pid **task_pid_ptr(struct task_struct *task, enum pid_type type)
321{
322 return (type == PIDTYPE_PID) ?
323 &task->thread_pid :
324 &task->signal->pids[type];
325}
326
327/*
328 * attach_pid() must be called with the tasklist_lock write-held.
329 */
330void attach_pid(struct task_struct *task, enum pid_type type)
331{
332 struct pid *pid = *task_pid_ptr(task, type);
333 hlist_add_head_rcu(&task->pid_links[type], &pid->tasks[type]);
334}
335
336static void __change_pid(struct task_struct *task, enum pid_type type,
337 struct pid *new)
338{
339 struct pid **pid_ptr = task_pid_ptr(task, type);
340 struct pid *pid;
341 int tmp;
342
343 pid = *pid_ptr;
344
345 hlist_del_rcu(&task->pid_links[type]);
346 *pid_ptr = new;
347
348 for (tmp = PIDTYPE_MAX; --tmp >= 0; )
349 if (pid_has_task(pid, tmp))
350 return;
351
352 free_pid(pid);
353}
354
355void detach_pid(struct task_struct *task, enum pid_type type)
356{
357 __change_pid(task, type, NULL);
358}
359
360void change_pid(struct task_struct *task, enum pid_type type,
361 struct pid *pid)
362{
363 __change_pid(task, type, pid);
364 attach_pid(task, type);
365}
366
367void exchange_tids(struct task_struct *left, struct task_struct *right)
368{
369 struct pid *pid1 = left->thread_pid;
370 struct pid *pid2 = right->thread_pid;
371 struct hlist_head *head1 = &pid1->tasks[PIDTYPE_PID];
372 struct hlist_head *head2 = &pid2->tasks[PIDTYPE_PID];
373
374 /* Swap the single entry tid lists */
375 hlists_swap_heads_rcu(head1, head2);
376
377 /* Swap the per task_struct pid */
378 rcu_assign_pointer(left->thread_pid, pid2);
379 rcu_assign_pointer(right->thread_pid, pid1);
380
381 /* Swap the cached value */
382 WRITE_ONCE(left->pid, pid_nr(pid2));
383 WRITE_ONCE(right->pid, pid_nr(pid1));
384}
385
386/* transfer_pid is an optimization of attach_pid(new), detach_pid(old) */
387void transfer_pid(struct task_struct *old, struct task_struct *new,
388 enum pid_type type)
389{
390 if (type == PIDTYPE_PID)
391 new->thread_pid = old->thread_pid;
392 hlist_replace_rcu(&old->pid_links[type], &new->pid_links[type]);
393}
394
395struct task_struct *pid_task(struct pid *pid, enum pid_type type)
396{
397 struct task_struct *result = NULL;
398 if (pid) {
399 struct hlist_node *first;
400 first = rcu_dereference_check(hlist_first_rcu(&pid->tasks[type]),
401 lockdep_tasklist_lock_is_held());
402 if (first)
403 result = hlist_entry(first, struct task_struct, pid_links[(type)]);
404 }
405 return result;
406}
407EXPORT_SYMBOL(pid_task);
408
409/*
410 * Must be called under rcu_read_lock().
411 */
412struct task_struct *find_task_by_pid_ns(pid_t nr, struct pid_namespace *ns)
413{
414 RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
415 "find_task_by_pid_ns() needs rcu_read_lock() protection");
416 return pid_task(find_pid_ns(nr, ns), PIDTYPE_PID);
417}
418
419struct task_struct *find_task_by_vpid(pid_t vnr)
420{
421 return find_task_by_pid_ns(vnr, task_active_pid_ns(current));
422}
423
424struct task_struct *find_get_task_by_vpid(pid_t nr)
425{
426 struct task_struct *task;
427
428 rcu_read_lock();
429 task = find_task_by_vpid(nr);
430 if (task)
431 get_task_struct(task);
432 rcu_read_unlock();
433
434 return task;
435}
436
437struct pid *get_task_pid(struct task_struct *task, enum pid_type type)
438{
439 struct pid *pid;
440 rcu_read_lock();
441 pid = get_pid(rcu_dereference(*task_pid_ptr(task, type)));
442 rcu_read_unlock();
443 return pid;
444}
445EXPORT_SYMBOL_GPL(get_task_pid);
446
447struct task_struct *get_pid_task(struct pid *pid, enum pid_type type)
448{
449 struct task_struct *result;
450 rcu_read_lock();
451 result = pid_task(pid, type);
452 if (result)
453 get_task_struct(result);
454 rcu_read_unlock();
455 return result;
456}
457EXPORT_SYMBOL_GPL(get_pid_task);
458
459struct pid *find_get_pid(pid_t nr)
460{
461 struct pid *pid;
462
463 rcu_read_lock();
464 pid = get_pid(find_vpid(nr));
465 rcu_read_unlock();
466
467 return pid;
468}
469EXPORT_SYMBOL_GPL(find_get_pid);
470
471pid_t pid_nr_ns(struct pid *pid, struct pid_namespace *ns)
472{
473 struct upid *upid;
474 pid_t nr = 0;
475
476 if (pid && ns->level <= pid->level) {
477 upid = &pid->numbers[ns->level];
478 if (upid->ns == ns)
479 nr = upid->nr;
480 }
481 return nr;
482}
483EXPORT_SYMBOL_GPL(pid_nr_ns);
484
485pid_t pid_vnr(struct pid *pid)
486{
487 return pid_nr_ns(pid, task_active_pid_ns(current));
488}
489EXPORT_SYMBOL_GPL(pid_vnr);
490
491pid_t __task_pid_nr_ns(struct task_struct *task, enum pid_type type,
492 struct pid_namespace *ns)
493{
494 pid_t nr = 0;
495
496 rcu_read_lock();
497 if (!ns)
498 ns = task_active_pid_ns(current);
499 nr = pid_nr_ns(rcu_dereference(*task_pid_ptr(task, type)), ns);
500 rcu_read_unlock();
501
502 return nr;
503}
504EXPORT_SYMBOL(__task_pid_nr_ns);
505
506struct pid_namespace *task_active_pid_ns(struct task_struct *tsk)
507{
508 return ns_of_pid(task_pid(tsk));
509}
510EXPORT_SYMBOL_GPL(task_active_pid_ns);
511
512/*
513 * Used by proc to find the first pid that is greater than or equal to nr.
514 *
515 * If there is a pid at nr this function is exactly the same as find_pid_ns.
516 */
517struct pid *find_ge_pid(int nr, struct pid_namespace *ns)
518{
519 return idr_get_next(&ns->idr, &nr);
520}
521
522/**
523 * pidfd_create() - Create a new pid file descriptor.
524 *
525 * @pid: struct pid that the pidfd will reference
526 *
527 * This creates a new pid file descriptor with the O_CLOEXEC flag set.
528 *
529 * Note, that this function can only be called after the fd table has
530 * been unshared to avoid leaking the pidfd to the new process.
531 *
532 * Return: On success, a cloexec pidfd is returned.
533 * On error, a negative errno number will be returned.
534 */
535static int pidfd_create(struct pid *pid)
536{
537 int fd;
538
539 fd = anon_inode_getfd("[pidfd]", &pidfd_fops, get_pid(pid),
540 O_RDWR | O_CLOEXEC);
541 if (fd < 0)
542 put_pid(pid);
543
544 return fd;
545}
546
547/**
548 * pidfd_open() - Open new pid file descriptor.
549 *
550 * @pid: pid for which to retrieve a pidfd
551 * @flags: flags to pass
552 *
553 * This creates a new pid file descriptor with the O_CLOEXEC flag set for
554 * the process identified by @pid. Currently, the process identified by
555 * @pid must be a thread-group leader. This restriction currently exists
556 * for all aspects of pidfds including pidfd creation (CLONE_PIDFD cannot
557 * be used with CLONE_THREAD) and pidfd polling (only supports thread group
558 * leaders).
559 *
560 * Return: On success, a cloexec pidfd is returned.
561 * On error, a negative errno number will be returned.
562 */
563SYSCALL_DEFINE2(pidfd_open, pid_t, pid, unsigned int, flags)
564{
565 int fd;
566 struct pid *p;
567
568 if (flags)
569 return -EINVAL;
570
571 if (pid <= 0)
572 return -EINVAL;
573
574 p = find_get_pid(pid);
575 if (!p)
576 return -ESRCH;
577
578 if (pid_has_task(p, PIDTYPE_TGID))
579 fd = pidfd_create(p);
580 else
581 fd = -EINVAL;
582
583 put_pid(p);
584 return fd;
585}
586
587void __init pid_idr_init(void)
588{
589 /* Verify no one has done anything silly: */
590 BUILD_BUG_ON(PID_MAX_LIMIT >= PIDNS_ADDING);
591
592 /* bump default and minimum pid_max based on number of cpus */
593 pid_max = min(pid_max_max, max_t(int, pid_max,
594 PIDS_PER_CPU_DEFAULT * num_possible_cpus()));
595 pid_max_min = max_t(int, pid_max_min,
596 PIDS_PER_CPU_MIN * num_possible_cpus());
597 pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min);
598
599 idr_init(&init_pid_ns.idr);
600
601 init_pid_ns.pid_cachep = KMEM_CACHE(pid,
602 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
603}
604
605static struct file *__pidfd_fget(struct task_struct *task, int fd)
606{
607 struct file *file;
608 int ret;
609
610 ret = mutex_lock_killable(&task->signal->exec_update_mutex);
611 if (ret)
612 return ERR_PTR(ret);
613
614 if (ptrace_may_access(task, PTRACE_MODE_ATTACH_REALCREDS))
615 file = fget_task(task, fd);
616 else
617 file = ERR_PTR(-EPERM);
618
619 mutex_unlock(&task->signal->exec_update_mutex);
620
621 return file ?: ERR_PTR(-EBADF);
622}
623
624static int pidfd_getfd(struct pid *pid, int fd)
625{
626 struct task_struct *task;
627 struct file *file;
628 int ret;
629
630 task = get_pid_task(pid, PIDTYPE_PID);
631 if (!task)
632 return -ESRCH;
633
634 file = __pidfd_fget(task, fd);
635 put_task_struct(task);
636 if (IS_ERR(file))
637 return PTR_ERR(file);
638
639 ret = receive_fd(file, O_CLOEXEC);
640 fput(file);
641
642 return ret;
643}
644
645/**
646 * sys_pidfd_getfd() - Get a file descriptor from another process
647 *
648 * @pidfd: the pidfd file descriptor of the process
649 * @fd: the file descriptor number to get
650 * @flags: flags on how to get the fd (reserved)
651 *
652 * This syscall gets a copy of a file descriptor from another process
653 * based on the pidfd, and file descriptor number. It requires that
654 * the calling process has the ability to ptrace the process represented
655 * by the pidfd. The process which is having its file descriptor copied
656 * is otherwise unaffected.
657 *
658 * Return: On success, a cloexec file descriptor is returned.
659 * On error, a negative errno number will be returned.
660 */
661SYSCALL_DEFINE3(pidfd_getfd, int, pidfd, int, fd,
662 unsigned int, flags)
663{
664 struct pid *pid;
665 struct fd f;
666 int ret;
667
668 /* flags is currently unused - make sure it's unset */
669 if (flags)
670 return -EINVAL;
671
672 f = fdget(pidfd);
673 if (!f.file)
674 return -EBADF;
675
676 pid = pidfd_pid(f.file);
677 if (IS_ERR(pid))
678 ret = PTR_ERR(pid);
679 else
680 ret = pidfd_getfd(pid, fd);
681
682 fdput(f);
683 return ret;
684}