xref: /linux/fs/pidfs.c (revision 49fba3725910c54878212ca08b968b9e1285866c)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/anon_inodes.h>
3 #include <linux/exportfs.h>
4 #include <linux/file.h>
5 #include <linux/fs.h>
6 #include <linux/cgroup.h>
7 #include <linux/magic.h>
8 #include <linux/mount.h>
9 #include <linux/pid.h>
10 #include <linux/pidfs.h>
11 #include <linux/pid_namespace.h>
12 #include <linux/poll.h>
13 #include <linux/proc_fs.h>
14 #include <linux/proc_ns.h>
15 #include <linux/pseudo_fs.h>
16 #include <linux/ptrace.h>
17 #include <linux/seq_file.h>
18 #include <uapi/linux/pidfd.h>
19 #include <linux/ipc_namespace.h>
20 #include <linux/time_namespace.h>
21 #include <linux/utsname.h>
22 #include <net/net_namespace.h>
23 #include <linux/coredump.h>
24 #include <linux/xattr.h>
25 
26 #include "internal.h"
27 #include "mount.h"
28 
29 #define PIDFS_PID_DEAD ERR_PTR(-ESRCH)
30 
31 static struct kmem_cache *pidfs_attr_cachep __ro_after_init;
32 static struct kmem_cache *pidfs_xattr_cachep __ro_after_init;
33 
34 /*
35  * Stashes information that userspace needs to access even after the
36  * process has been reaped.
37  */
38 struct pidfs_exit_info {
39 	__u64 cgroupid;
40 	__s32 exit_code;
41 	__u32 coredump_mask;
42 };
43 
44 struct pidfs_attr {
45 	struct simple_xattrs *xattrs;
46 	struct pidfs_exit_info __pei;
47 	struct pidfs_exit_info *exit_info;
48 };
49 
50 static struct rb_root pidfs_ino_tree = RB_ROOT;
51 
52 #if BITS_PER_LONG == 32
53 static inline unsigned long pidfs_ino(u64 ino)
54 {
55 	return lower_32_bits(ino);
56 }
57 
58 /* On 32 bit the generation number are the upper 32 bits. */
59 static inline u32 pidfs_gen(u64 ino)
60 {
61 	return upper_32_bits(ino);
62 }
63 
64 #else
65 
66 /* On 64 bit simply return ino. */
67 static inline unsigned long pidfs_ino(u64 ino)
68 {
69 	return ino;
70 }
71 
72 /* On 64 bit the generation number is 0. */
73 static inline u32 pidfs_gen(u64 ino)
74 {
75 	return 0;
76 }
77 #endif
78 
79 static int pidfs_ino_cmp(struct rb_node *a, const struct rb_node *b)
80 {
81 	struct pid *pid_a = rb_entry(a, struct pid, pidfs_node);
82 	struct pid *pid_b = rb_entry(b, struct pid, pidfs_node);
83 	u64 pid_ino_a = pid_a->ino;
84 	u64 pid_ino_b = pid_b->ino;
85 
86 	if (pid_ino_a < pid_ino_b)
87 		return -1;
88 	if (pid_ino_a > pid_ino_b)
89 		return 1;
90 	return 0;
91 }
92 
93 void pidfs_add_pid(struct pid *pid)
94 {
95 	static u64 pidfs_ino_nr = 2;
96 
97 	/*
98 	 * On 64 bit nothing special happens. The 64bit number assigned
99 	 * to struct pid is the inode number.
100 	 *
101 	 * On 32 bit the 64 bit number assigned to struct pid is split
102 	 * into two 32 bit numbers. The lower 32 bits are used as the
103 	 * inode number and the upper 32 bits are used as the inode
104 	 * generation number.
105 	 *
106 	 * On 32 bit pidfs_ino() will return the lower 32 bit. When
107 	 * pidfs_ino() returns zero a wrap around happened. When a
108 	 * wraparound happens the 64 bit number will be incremented by 2
109 	 * so inode numbering starts at 2 again.
110 	 *
111 	 * On 64 bit comparing two pidfds is as simple as comparing
112 	 * inode numbers.
113 	 *
114 	 * When a wraparound happens on 32 bit multiple pidfds with the
115 	 * same inode number are likely to exist (This isn't a problem
116 	 * since before pidfs pidfds used the anonymous inode meaning
117 	 * all pidfds had the same inode number.). Userspace can
118 	 * reconstruct the 64 bit identifier by retrieving both the
119 	 * inode number and the inode generation number to compare or
120 	 * use file handles.
121 	 */
122 	if (pidfs_ino(pidfs_ino_nr) == 0)
123 		pidfs_ino_nr += 2;
124 
125 	pid->ino = pidfs_ino_nr;
126 	pid->stashed = NULL;
127 	pid->attr = NULL;
128 	pidfs_ino_nr++;
129 
130 	write_seqcount_begin(&pidmap_lock_seq);
131 	rb_find_add_rcu(&pid->pidfs_node, &pidfs_ino_tree, pidfs_ino_cmp);
132 	write_seqcount_end(&pidmap_lock_seq);
133 }
134 
135 void pidfs_remove_pid(struct pid *pid)
136 {
137 	write_seqcount_begin(&pidmap_lock_seq);
138 	rb_erase(&pid->pidfs_node, &pidfs_ino_tree);
139 	write_seqcount_end(&pidmap_lock_seq);
140 }
141 
142 void pidfs_free_pid(struct pid *pid)
143 {
144 	struct pidfs_attr *attr __free(kfree) = no_free_ptr(pid->attr);
145 	struct simple_xattrs *xattrs __free(kfree) = NULL;
146 
147 	/*
148 	 * Any dentry must've been wiped from the pid by now.
149 	 * Otherwise there's a reference count bug.
150 	 */
151 	VFS_WARN_ON_ONCE(pid->stashed);
152 
153 	if (IS_ERR(attr))
154 		return;
155 
156 	/*
157 	 * Any dentry must've been wiped from the pid by now. Otherwise
158 	 * there's a reference count bug.
159 	 */
160 	VFS_WARN_ON_ONCE(pid->stashed);
161 
162 	xattrs = attr->xattrs;
163 	if (xattrs)
164 		simple_xattrs_free(attr->xattrs, NULL);
165 }
166 
167 #ifdef CONFIG_PROC_FS
168 /**
169  * pidfd_show_fdinfo - print information about a pidfd
170  * @m: proc fdinfo file
171  * @f: file referencing a pidfd
172  *
173  * Pid:
174  * This function will print the pid that a given pidfd refers to in the
175  * pid namespace of the procfs instance.
176  * If the pid namespace of the process is not a descendant of the pid
177  * namespace of the procfs instance 0 will be shown as its pid. This is
178  * similar to calling getppid() on a process whose parent is outside of
179  * its pid namespace.
180  *
181  * NSpid:
182  * If pid namespaces are supported then this function will also print
183  * the pid of a given pidfd refers to for all descendant pid namespaces
184  * starting from the current pid namespace of the instance, i.e. the
185  * Pid field and the first entry in the NSpid field will be identical.
186  * If the pid namespace of the process is not a descendant of the pid
187  * namespace of the procfs instance 0 will be shown as its first NSpid
188  * entry and no others will be shown.
189  * Note that this differs from the Pid and NSpid fields in
190  * /proc/<pid>/status where Pid and NSpid are always shown relative to
191  * the  pid namespace of the procfs instance. The difference becomes
192  * obvious when sending around a pidfd between pid namespaces from a
193  * different branch of the tree, i.e. where no ancestral relation is
194  * present between the pid namespaces:
195  * - create two new pid namespaces ns1 and ns2 in the initial pid
196  *   namespace (also take care to create new mount namespaces in the
197  *   new pid namespace and mount procfs)
198  * - create a process with a pidfd in ns1
199  * - send pidfd from ns1 to ns2
200  * - read /proc/self/fdinfo/<pidfd> and observe that both Pid and NSpid
201  *   have exactly one entry, which is 0
202  */
203 static void pidfd_show_fdinfo(struct seq_file *m, struct file *f)
204 {
205 	struct pid *pid = pidfd_pid(f);
206 	struct pid_namespace *ns;
207 	pid_t nr = -1;
208 
209 	if (likely(pid_has_task(pid, PIDTYPE_PID))) {
210 		ns = proc_pid_ns(file_inode(m->file)->i_sb);
211 		nr = pid_nr_ns(pid, ns);
212 	}
213 
214 	seq_put_decimal_ll(m, "Pid:\t", nr);
215 
216 #ifdef CONFIG_PID_NS
217 	seq_put_decimal_ll(m, "\nNSpid:\t", nr);
218 	if (nr > 0) {
219 		int i;
220 
221 		/* If nr is non-zero it means that 'pid' is valid and that
222 		 * ns, i.e. the pid namespace associated with the procfs
223 		 * instance, is in the pid namespace hierarchy of pid.
224 		 * Start at one below the already printed level.
225 		 */
226 		for (i = ns->level + 1; i <= pid->level; i++)
227 			seq_put_decimal_ll(m, "\t", pid->numbers[i].nr);
228 	}
229 #endif
230 	seq_putc(m, '\n');
231 }
232 #endif
233 
234 /*
235  * Poll support for process exit notification.
236  */
237 static __poll_t pidfd_poll(struct file *file, struct poll_table_struct *pts)
238 {
239 	struct pid *pid = pidfd_pid(file);
240 	struct task_struct *task;
241 	__poll_t poll_flags = 0;
242 
243 	poll_wait(file, &pid->wait_pidfd, pts);
244 	/*
245 	 * Don't wake waiters if the thread-group leader exited
246 	 * prematurely. They either get notified when the last subthread
247 	 * exits or not at all if one of the remaining subthreads execs
248 	 * and assumes the struct pid of the old thread-group leader.
249 	 */
250 	guard(rcu)();
251 	task = pid_task(pid, PIDTYPE_PID);
252 	if (!task)
253 		poll_flags = EPOLLIN | EPOLLRDNORM | EPOLLHUP;
254 	else if (task->exit_state && !delay_group_leader(task))
255 		poll_flags = EPOLLIN | EPOLLRDNORM;
256 
257 	return poll_flags;
258 }
259 
260 static inline bool pid_in_current_pidns(const struct pid *pid)
261 {
262 	const struct pid_namespace *ns = task_active_pid_ns(current);
263 
264 	if (ns->level <= pid->level)
265 		return pid->numbers[ns->level].ns == ns;
266 
267 	return false;
268 }
269 
270 static __u32 pidfs_coredump_mask(unsigned long mm_flags)
271 {
272 	switch (__get_dumpable(mm_flags)) {
273 	case SUID_DUMP_USER:
274 		return PIDFD_COREDUMP_USER;
275 	case SUID_DUMP_ROOT:
276 		return PIDFD_COREDUMP_ROOT;
277 	case SUID_DUMP_DISABLE:
278 		return PIDFD_COREDUMP_SKIP;
279 	default:
280 		WARN_ON_ONCE(true);
281 	}
282 
283 	return 0;
284 }
285 
286 static long pidfd_info(struct file *file, unsigned int cmd, unsigned long arg)
287 {
288 	struct pidfd_info __user *uinfo = (struct pidfd_info __user *)arg;
289 	struct pid *pid = pidfd_pid(file);
290 	size_t usize = _IOC_SIZE(cmd);
291 	struct pidfd_info kinfo = {};
292 	struct pidfs_exit_info *exit_info;
293 	struct user_namespace *user_ns;
294 	struct task_struct *task;
295 	struct pidfs_attr *attr;
296 	const struct cred *c;
297 	__u64 mask;
298 
299 	if (!uinfo)
300 		return -EINVAL;
301 	if (usize < PIDFD_INFO_SIZE_VER0)
302 		return -EINVAL; /* First version, no smaller struct possible */
303 
304 	if (copy_from_user(&mask, &uinfo->mask, sizeof(mask)))
305 		return -EFAULT;
306 
307 	/*
308 	 * Restrict information retrieval to tasks within the caller's pid
309 	 * namespace hierarchy.
310 	 */
311 	if (!pid_in_current_pidns(pid))
312 		return -ESRCH;
313 
314 	attr = READ_ONCE(pid->attr);
315 	if (mask & PIDFD_INFO_EXIT) {
316 		exit_info = READ_ONCE(attr->exit_info);
317 		if (exit_info) {
318 			kinfo.mask |= PIDFD_INFO_EXIT;
319 #ifdef CONFIG_CGROUPS
320 			kinfo.cgroupid = exit_info->cgroupid;
321 			kinfo.mask |= PIDFD_INFO_CGROUPID;
322 #endif
323 			kinfo.exit_code = exit_info->exit_code;
324 		}
325 	}
326 
327 	if (mask & PIDFD_INFO_COREDUMP) {
328 		kinfo.mask |= PIDFD_INFO_COREDUMP;
329 		kinfo.coredump_mask = READ_ONCE(attr->__pei.coredump_mask);
330 	}
331 
332 	task = get_pid_task(pid, PIDTYPE_PID);
333 	if (!task) {
334 		/*
335 		 * If the task has already been reaped, only exit
336 		 * information is available
337 		 */
338 		if (!(mask & PIDFD_INFO_EXIT))
339 			return -ESRCH;
340 
341 		goto copy_out;
342 	}
343 
344 	c = get_task_cred(task);
345 	if (!c)
346 		return -ESRCH;
347 
348 	if (!(kinfo.mask & PIDFD_INFO_COREDUMP)) {
349 		task_lock(task);
350 		if (task->mm)
351 			kinfo.coredump_mask = pidfs_coredump_mask(task->mm->flags);
352 		task_unlock(task);
353 	}
354 
355 	/* Unconditionally return identifiers and credentials, the rest only on request */
356 
357 	user_ns = current_user_ns();
358 	kinfo.ruid = from_kuid_munged(user_ns, c->uid);
359 	kinfo.rgid = from_kgid_munged(user_ns, c->gid);
360 	kinfo.euid = from_kuid_munged(user_ns, c->euid);
361 	kinfo.egid = from_kgid_munged(user_ns, c->egid);
362 	kinfo.suid = from_kuid_munged(user_ns, c->suid);
363 	kinfo.sgid = from_kgid_munged(user_ns, c->sgid);
364 	kinfo.fsuid = from_kuid_munged(user_ns, c->fsuid);
365 	kinfo.fsgid = from_kgid_munged(user_ns, c->fsgid);
366 	kinfo.mask |= PIDFD_INFO_CREDS;
367 	put_cred(c);
368 
369 #ifdef CONFIG_CGROUPS
370 	if (!kinfo.cgroupid) {
371 		struct cgroup *cgrp;
372 
373 		rcu_read_lock();
374 		cgrp = task_dfl_cgroup(task);
375 		kinfo.cgroupid = cgroup_id(cgrp);
376 		kinfo.mask |= PIDFD_INFO_CGROUPID;
377 		rcu_read_unlock();
378 	}
379 #endif
380 
381 	/*
382 	 * Copy pid/tgid last, to reduce the chances the information might be
383 	 * stale. Note that it is not possible to ensure it will be valid as the
384 	 * task might return as soon as the copy_to_user finishes, but that's ok
385 	 * and userspace expects that might happen and can act accordingly, so
386 	 * this is just best-effort. What we can do however is checking that all
387 	 * the fields are set correctly, or return ESRCH to avoid providing
388 	 * incomplete information. */
389 
390 	kinfo.ppid = task_ppid_nr_ns(task, NULL);
391 	kinfo.tgid = task_tgid_vnr(task);
392 	kinfo.pid = task_pid_vnr(task);
393 	kinfo.mask |= PIDFD_INFO_PID;
394 
395 	if (kinfo.pid == 0 || kinfo.tgid == 0 || (kinfo.ppid == 0 && kinfo.pid != 1))
396 		return -ESRCH;
397 
398 copy_out:
399 	/*
400 	 * If userspace and the kernel have the same struct size it can just
401 	 * be copied. If userspace provides an older struct, only the bits that
402 	 * userspace knows about will be copied. If userspace provides a new
403 	 * struct, only the bits that the kernel knows about will be copied.
404 	 */
405 	return copy_struct_to_user(uinfo, usize, &kinfo, sizeof(kinfo), NULL);
406 }
407 
408 static bool pidfs_ioctl_valid(unsigned int cmd)
409 {
410 	switch (cmd) {
411 	case FS_IOC_GETVERSION:
412 	case PIDFD_GET_CGROUP_NAMESPACE:
413 	case PIDFD_GET_IPC_NAMESPACE:
414 	case PIDFD_GET_MNT_NAMESPACE:
415 	case PIDFD_GET_NET_NAMESPACE:
416 	case PIDFD_GET_PID_FOR_CHILDREN_NAMESPACE:
417 	case PIDFD_GET_TIME_NAMESPACE:
418 	case PIDFD_GET_TIME_FOR_CHILDREN_NAMESPACE:
419 	case PIDFD_GET_UTS_NAMESPACE:
420 	case PIDFD_GET_USER_NAMESPACE:
421 	case PIDFD_GET_PID_NAMESPACE:
422 		return true;
423 	}
424 
425 	/* Extensible ioctls require some more careful checks. */
426 	switch (_IOC_NR(cmd)) {
427 	case _IOC_NR(PIDFD_GET_INFO):
428 		/*
429 		 * Try to prevent performing a pidfd ioctl when someone
430 		 * erronously mistook the file descriptor for a pidfd.
431 		 * This is not perfect but will catch most cases.
432 		 */
433 		return (_IOC_TYPE(cmd) == _IOC_TYPE(PIDFD_GET_INFO));
434 	}
435 
436 	return false;
437 }
438 
439 static long pidfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
440 {
441 	struct task_struct *task __free(put_task) = NULL;
442 	struct nsproxy *nsp __free(put_nsproxy) = NULL;
443 	struct ns_common *ns_common = NULL;
444 	struct pid_namespace *pid_ns;
445 
446 	if (!pidfs_ioctl_valid(cmd))
447 		return -ENOIOCTLCMD;
448 
449 	if (cmd == FS_IOC_GETVERSION) {
450 		if (!arg)
451 			return -EINVAL;
452 
453 		__u32 __user *argp = (__u32 __user *)arg;
454 		return put_user(file_inode(file)->i_generation, argp);
455 	}
456 
457 	/* Extensible IOCTL that does not open namespace FDs, take a shortcut */
458 	if (_IOC_NR(cmd) == _IOC_NR(PIDFD_GET_INFO))
459 		return pidfd_info(file, cmd, arg);
460 
461 	task = get_pid_task(pidfd_pid(file), PIDTYPE_PID);
462 	if (!task)
463 		return -ESRCH;
464 
465 	if (arg)
466 		return -EINVAL;
467 
468 	scoped_guard(task_lock, task) {
469 		nsp = task->nsproxy;
470 		if (nsp)
471 			get_nsproxy(nsp);
472 	}
473 	if (!nsp)
474 		return -ESRCH; /* just pretend it didn't exist */
475 
476 	/*
477 	 * We're trying to open a file descriptor to the namespace so perform a
478 	 * filesystem cred ptrace check. Also, we mirror nsfs behavior.
479 	 */
480 	if (!ptrace_may_access(task, PTRACE_MODE_READ_FSCREDS))
481 		return -EACCES;
482 
483 	switch (cmd) {
484 	/* Namespaces that hang of nsproxy. */
485 	case PIDFD_GET_CGROUP_NAMESPACE:
486 		if (IS_ENABLED(CONFIG_CGROUPS)) {
487 			get_cgroup_ns(nsp->cgroup_ns);
488 			ns_common = to_ns_common(nsp->cgroup_ns);
489 		}
490 		break;
491 	case PIDFD_GET_IPC_NAMESPACE:
492 		if (IS_ENABLED(CONFIG_IPC_NS)) {
493 			get_ipc_ns(nsp->ipc_ns);
494 			ns_common = to_ns_common(nsp->ipc_ns);
495 		}
496 		break;
497 	case PIDFD_GET_MNT_NAMESPACE:
498 		get_mnt_ns(nsp->mnt_ns);
499 		ns_common = to_ns_common(nsp->mnt_ns);
500 		break;
501 	case PIDFD_GET_NET_NAMESPACE:
502 		if (IS_ENABLED(CONFIG_NET_NS)) {
503 			ns_common = to_ns_common(nsp->net_ns);
504 			get_net_ns(ns_common);
505 		}
506 		break;
507 	case PIDFD_GET_PID_FOR_CHILDREN_NAMESPACE:
508 		if (IS_ENABLED(CONFIG_PID_NS)) {
509 			get_pid_ns(nsp->pid_ns_for_children);
510 			ns_common = to_ns_common(nsp->pid_ns_for_children);
511 		}
512 		break;
513 	case PIDFD_GET_TIME_NAMESPACE:
514 		if (IS_ENABLED(CONFIG_TIME_NS)) {
515 			get_time_ns(nsp->time_ns);
516 			ns_common = to_ns_common(nsp->time_ns);
517 		}
518 		break;
519 	case PIDFD_GET_TIME_FOR_CHILDREN_NAMESPACE:
520 		if (IS_ENABLED(CONFIG_TIME_NS)) {
521 			get_time_ns(nsp->time_ns_for_children);
522 			ns_common = to_ns_common(nsp->time_ns_for_children);
523 		}
524 		break;
525 	case PIDFD_GET_UTS_NAMESPACE:
526 		if (IS_ENABLED(CONFIG_UTS_NS)) {
527 			get_uts_ns(nsp->uts_ns);
528 			ns_common = to_ns_common(nsp->uts_ns);
529 		}
530 		break;
531 	/* Namespaces that don't hang of nsproxy. */
532 	case PIDFD_GET_USER_NAMESPACE:
533 		if (IS_ENABLED(CONFIG_USER_NS)) {
534 			rcu_read_lock();
535 			ns_common = to_ns_common(get_user_ns(task_cred_xxx(task, user_ns)));
536 			rcu_read_unlock();
537 		}
538 		break;
539 	case PIDFD_GET_PID_NAMESPACE:
540 		if (IS_ENABLED(CONFIG_PID_NS)) {
541 			rcu_read_lock();
542 			pid_ns = task_active_pid_ns(task);
543 			if (pid_ns)
544 				ns_common = to_ns_common(get_pid_ns(pid_ns));
545 			rcu_read_unlock();
546 		}
547 		break;
548 	default:
549 		return -ENOIOCTLCMD;
550 	}
551 
552 	if (!ns_common)
553 		return -EOPNOTSUPP;
554 
555 	/* open_namespace() unconditionally consumes the reference */
556 	return open_namespace(ns_common);
557 }
558 
559 static const struct file_operations pidfs_file_operations = {
560 	.poll		= pidfd_poll,
561 #ifdef CONFIG_PROC_FS
562 	.show_fdinfo	= pidfd_show_fdinfo,
563 #endif
564 	.unlocked_ioctl	= pidfd_ioctl,
565 	.compat_ioctl   = compat_ptr_ioctl,
566 };
567 
568 struct pid *pidfd_pid(const struct file *file)
569 {
570 	if (file->f_op != &pidfs_file_operations)
571 		return ERR_PTR(-EBADF);
572 	return file_inode(file)->i_private;
573 }
574 
575 /*
576  * We're called from release_task(). We know there's at least one
577  * reference to struct pid being held that won't be released until the
578  * task has been reaped which cannot happen until we're out of
579  * release_task().
580  *
581  * If this struct pid has at least once been referred to by a pidfd then
582  * pid->attr will be allocated. If not we mark the struct pid as dead so
583  * anyone who is trying to register it with pidfs will fail to do so.
584  * Otherwise we would hand out pidfs for reaped tasks without having
585  * exit information available.
586  *
587  * Worst case is that we've filled in the info and the pid gets freed
588  * right away in free_pid() when no one holds a pidfd anymore. Since
589  * pidfs_exit() currently is placed after exit_task_work() we know that
590  * it cannot be us aka the exiting task holding a pidfd to itself.
591  */
592 void pidfs_exit(struct task_struct *tsk)
593 {
594 	struct pid *pid = task_pid(tsk);
595 	struct pidfs_attr *attr;
596 	struct pidfs_exit_info *exit_info;
597 #ifdef CONFIG_CGROUPS
598 	struct cgroup *cgrp;
599 #endif
600 
601 	might_sleep();
602 
603 	guard(spinlock_irq)(&pid->wait_pidfd.lock);
604 	attr = pid->attr;
605 	if (!attr) {
606 		/*
607 		 * No one ever held a pidfd for this struct pid.
608 		 * Mark it as dead so no one can add a pidfs
609 		 * entry anymore. We're about to be reaped and
610 		 * so no exit information would be available.
611 		 */
612 		pid->attr = PIDFS_PID_DEAD;
613 		return;
614 	}
615 
616 	/*
617 	 * If @pid->attr is set someone might still legitimately hold a
618 	 * pidfd to @pid or someone might concurrently still be getting
619 	 * a reference to an already stashed dentry from @pid->stashed.
620 	 * So defer cleaning @pid->attr until the last reference to @pid
621 	 * is put
622 	 */
623 
624 	exit_info = &attr->__pei;
625 
626 #ifdef CONFIG_CGROUPS
627 	rcu_read_lock();
628 	cgrp = task_dfl_cgroup(tsk);
629 	exit_info->cgroupid = cgroup_id(cgrp);
630 	rcu_read_unlock();
631 #endif
632 	exit_info->exit_code = tsk->exit_code;
633 
634 	/* Ensure that PIDFD_GET_INFO sees either all or nothing. */
635 	smp_store_release(&attr->exit_info, &attr->__pei);
636 }
637 
638 #ifdef CONFIG_COREDUMP
639 void pidfs_coredump(const struct coredump_params *cprm)
640 {
641 	struct pid *pid = cprm->pid;
642 	struct pidfs_exit_info *exit_info;
643 	struct pidfs_attr *attr;
644 	__u32 coredump_mask = 0;
645 
646 	attr = READ_ONCE(pid->attr);
647 
648 	VFS_WARN_ON_ONCE(!attr);
649 	VFS_WARN_ON_ONCE(attr == PIDFS_PID_DEAD);
650 
651 	exit_info = &attr->__pei;
652 	/* Note how we were coredumped. */
653 	coredump_mask = pidfs_coredump_mask(cprm->mm_flags);
654 	/* Note that we actually did coredump. */
655 	coredump_mask |= PIDFD_COREDUMPED;
656 	/* If coredumping is set to skip we should never end up here. */
657 	VFS_WARN_ON_ONCE(coredump_mask & PIDFD_COREDUMP_SKIP);
658 	smp_store_release(&exit_info->coredump_mask, coredump_mask);
659 }
660 #endif
661 
662 static struct vfsmount *pidfs_mnt __ro_after_init;
663 
664 /*
665  * The vfs falls back to simple_setattr() if i_op->setattr() isn't
666  * implemented. Let's reject it completely until we have a clean
667  * permission concept for pidfds.
668  */
669 static int pidfs_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
670 			 struct iattr *attr)
671 {
672 	return anon_inode_setattr(idmap, dentry, attr);
673 }
674 
675 static int pidfs_getattr(struct mnt_idmap *idmap, const struct path *path,
676 			 struct kstat *stat, u32 request_mask,
677 			 unsigned int query_flags)
678 {
679 	return anon_inode_getattr(idmap, path, stat, request_mask, query_flags);
680 }
681 
682 static ssize_t pidfs_listxattr(struct dentry *dentry, char *buf, size_t size)
683 {
684 	struct inode *inode = d_inode(dentry);
685 	struct pid *pid = inode->i_private;
686 	struct pidfs_attr *attr = pid->attr;
687 	struct simple_xattrs *xattrs;
688 
689 	xattrs = READ_ONCE(attr->xattrs);
690 	if (!xattrs)
691 		return 0;
692 
693 	return simple_xattr_list(inode, xattrs, buf, size);
694 }
695 
696 static const struct inode_operations pidfs_inode_operations = {
697 	.getattr	= pidfs_getattr,
698 	.setattr	= pidfs_setattr,
699 	.listxattr	= pidfs_listxattr,
700 };
701 
702 static void pidfs_evict_inode(struct inode *inode)
703 {
704 	struct pid *pid = inode->i_private;
705 
706 	clear_inode(inode);
707 	put_pid(pid);
708 }
709 
710 static const struct super_operations pidfs_sops = {
711 	.drop_inode	= generic_delete_inode,
712 	.evict_inode	= pidfs_evict_inode,
713 	.statfs		= simple_statfs,
714 };
715 
716 /*
717  * 'lsof' has knowledge of out historical anon_inode use, and expects
718  * the pidfs dentry name to start with 'anon_inode'.
719  */
720 static char *pidfs_dname(struct dentry *dentry, char *buffer, int buflen)
721 {
722 	return dynamic_dname(buffer, buflen, "anon_inode:[pidfd]");
723 }
724 
725 const struct dentry_operations pidfs_dentry_operations = {
726 	.d_dname	= pidfs_dname,
727 	.d_prune	= stashed_dentry_prune,
728 };
729 
730 static int pidfs_encode_fh(struct inode *inode, u32 *fh, int *max_len,
731 			   struct inode *parent)
732 {
733 	const struct pid *pid = inode->i_private;
734 
735 	if (*max_len < 2) {
736 		*max_len = 2;
737 		return FILEID_INVALID;
738 	}
739 
740 	*max_len = 2;
741 	*(u64 *)fh = pid->ino;
742 	return FILEID_KERNFS;
743 }
744 
745 static int pidfs_ino_find(const void *key, const struct rb_node *node)
746 {
747 	const u64 pid_ino = *(u64 *)key;
748 	const struct pid *pid = rb_entry(node, struct pid, pidfs_node);
749 
750 	if (pid_ino < pid->ino)
751 		return -1;
752 	if (pid_ino > pid->ino)
753 		return 1;
754 	return 0;
755 }
756 
757 /* Find a struct pid based on the inode number. */
758 static struct pid *pidfs_ino_get_pid(u64 ino)
759 {
760 	struct pid *pid;
761 	struct rb_node *node;
762 	unsigned int seq;
763 
764 	guard(rcu)();
765 	do {
766 		seq = read_seqcount_begin(&pidmap_lock_seq);
767 		node = rb_find_rcu(&ino, &pidfs_ino_tree, pidfs_ino_find);
768 		if (node)
769 			break;
770 	} while (read_seqcount_retry(&pidmap_lock_seq, seq));
771 
772 	if (!node)
773 		return NULL;
774 
775 	pid = rb_entry(node, struct pid, pidfs_node);
776 
777 	/* Within our pid namespace hierarchy? */
778 	if (pid_vnr(pid) == 0)
779 		return NULL;
780 
781 	return get_pid(pid);
782 }
783 
784 static struct dentry *pidfs_fh_to_dentry(struct super_block *sb,
785 					 struct fid *fid, int fh_len,
786 					 int fh_type)
787 {
788 	int ret;
789 	u64 pid_ino;
790 	struct path path;
791 	struct pid *pid;
792 
793 	if (fh_len < 2)
794 		return NULL;
795 
796 	switch (fh_type) {
797 	case FILEID_KERNFS:
798 		pid_ino = *(u64 *)fid;
799 		break;
800 	default:
801 		return NULL;
802 	}
803 
804 	pid = pidfs_ino_get_pid(pid_ino);
805 	if (!pid)
806 		return NULL;
807 
808 	ret = path_from_stashed(&pid->stashed, pidfs_mnt, pid, &path);
809 	if (ret < 0)
810 		return ERR_PTR(ret);
811 
812 	mntput(path.mnt);
813 	return path.dentry;
814 }
815 
816 /*
817  * Make sure that we reject any nonsensical flags that users pass via
818  * open_by_handle_at(). Note that PIDFD_THREAD is defined as O_EXCL, and
819  * PIDFD_NONBLOCK as O_NONBLOCK.
820  */
821 #define VALID_FILE_HANDLE_OPEN_FLAGS \
822 	(O_RDONLY | O_WRONLY | O_RDWR | O_NONBLOCK | O_CLOEXEC | O_EXCL)
823 
824 static int pidfs_export_permission(struct handle_to_path_ctx *ctx,
825 				   unsigned int oflags)
826 {
827 	if (oflags & ~(VALID_FILE_HANDLE_OPEN_FLAGS | O_LARGEFILE))
828 		return -EINVAL;
829 
830 	/*
831 	 * pidfd_ino_get_pid() will verify that the struct pid is part
832 	 * of the caller's pid namespace hierarchy. No further
833 	 * permission checks are needed.
834 	 */
835 	return 0;
836 }
837 
838 static struct file *pidfs_export_open(struct path *path, unsigned int oflags)
839 {
840 	/*
841 	 * Clear O_LARGEFILE as open_by_handle_at() forces it and raise
842 	 * O_RDWR as pidfds always are.
843 	 */
844 	oflags &= ~O_LARGEFILE;
845 	return dentry_open(path, oflags | O_RDWR, current_cred());
846 }
847 
848 static const struct export_operations pidfs_export_operations = {
849 	.encode_fh	= pidfs_encode_fh,
850 	.fh_to_dentry	= pidfs_fh_to_dentry,
851 	.open		= pidfs_export_open,
852 	.permission	= pidfs_export_permission,
853 };
854 
855 static int pidfs_init_inode(struct inode *inode, void *data)
856 {
857 	const struct pid *pid = data;
858 
859 	inode->i_private = data;
860 	inode->i_flags |= S_PRIVATE | S_ANON_INODE;
861 	/* We allow to set xattrs. */
862 	inode->i_flags &= ~S_IMMUTABLE;
863 	inode->i_mode |= S_IRWXU;
864 	inode->i_op = &pidfs_inode_operations;
865 	inode->i_fop = &pidfs_file_operations;
866 	inode->i_ino = pidfs_ino(pid->ino);
867 	inode->i_generation = pidfs_gen(pid->ino);
868 	return 0;
869 }
870 
871 static void pidfs_put_data(void *data)
872 {
873 	struct pid *pid = data;
874 	put_pid(pid);
875 }
876 
877 /**
878  * pidfs_register_pid - register a struct pid in pidfs
879  * @pid: pid to pin
880  *
881  * Register a struct pid in pidfs.
882  *
883  * Return: On success zero, on error a negative error code is returned.
884  */
885 int pidfs_register_pid(struct pid *pid)
886 {
887 	struct pidfs_attr *new_attr __free(kfree) = NULL;
888 	struct pidfs_attr *attr;
889 
890 	might_sleep();
891 
892 	if (!pid)
893 		return 0;
894 
895 	attr = READ_ONCE(pid->attr);
896 	if (unlikely(attr == PIDFS_PID_DEAD))
897 		return PTR_ERR(PIDFS_PID_DEAD);
898 	if (attr)
899 		return 0;
900 
901 	new_attr = kmem_cache_zalloc(pidfs_attr_cachep, GFP_KERNEL);
902 	if (!new_attr)
903 		return -ENOMEM;
904 
905 	/* Synchronize with pidfs_exit(). */
906 	guard(spinlock_irq)(&pid->wait_pidfd.lock);
907 
908 	attr = pid->attr;
909 	if (unlikely(attr == PIDFS_PID_DEAD))
910 		return PTR_ERR(PIDFS_PID_DEAD);
911 	if (unlikely(attr))
912 		return 0;
913 
914 	pid->attr = no_free_ptr(new_attr);
915 	return 0;
916 }
917 
918 static struct dentry *pidfs_stash_dentry(struct dentry **stashed,
919 					 struct dentry *dentry)
920 {
921 	int ret;
922 	struct pid *pid = d_inode(dentry)->i_private;
923 
924 	VFS_WARN_ON_ONCE(stashed != &pid->stashed);
925 
926 	ret = pidfs_register_pid(pid);
927 	if (ret)
928 		return ERR_PTR(ret);
929 
930 	return stash_dentry(stashed, dentry);
931 }
932 
933 static const struct stashed_operations pidfs_stashed_ops = {
934 	.stash_dentry	= pidfs_stash_dentry,
935 	.init_inode	= pidfs_init_inode,
936 	.put_data	= pidfs_put_data,
937 };
938 
939 static int pidfs_xattr_get(const struct xattr_handler *handler,
940 			   struct dentry *unused, struct inode *inode,
941 			   const char *suffix, void *value, size_t size)
942 {
943 	struct pid *pid = inode->i_private;
944 	struct pidfs_attr *attr = pid->attr;
945 	const char *name;
946 	struct simple_xattrs *xattrs;
947 
948 	xattrs = READ_ONCE(attr->xattrs);
949 	if (!xattrs)
950 		return 0;
951 
952 	name = xattr_full_name(handler, suffix);
953 	return simple_xattr_get(xattrs, name, value, size);
954 }
955 
956 static int pidfs_xattr_set(const struct xattr_handler *handler,
957 			   struct mnt_idmap *idmap, struct dentry *unused,
958 			   struct inode *inode, const char *suffix,
959 			   const void *value, size_t size, int flags)
960 {
961 	struct pid *pid = inode->i_private;
962 	struct pidfs_attr *attr = pid->attr;
963 	const char *name;
964 	struct simple_xattrs *xattrs;
965 	struct simple_xattr *old_xattr;
966 
967 	/* Ensure we're the only one to set @attr->xattrs. */
968 	WARN_ON_ONCE(!inode_is_locked(inode));
969 
970 	xattrs = READ_ONCE(attr->xattrs);
971 	if (!xattrs) {
972 		xattrs = kmem_cache_zalloc(pidfs_xattr_cachep, GFP_KERNEL);
973 		if (!xattrs)
974 			return -ENOMEM;
975 
976 		simple_xattrs_init(xattrs);
977 		smp_store_release(&pid->attr->xattrs, xattrs);
978 	}
979 
980 	name = xattr_full_name(handler, suffix);
981 	old_xattr = simple_xattr_set(xattrs, name, value, size, flags);
982 	if (IS_ERR(old_xattr))
983 		return PTR_ERR(old_xattr);
984 
985 	simple_xattr_free(old_xattr);
986 	return 0;
987 }
988 
989 static const struct xattr_handler pidfs_trusted_xattr_handler = {
990 	.prefix = XATTR_TRUSTED_PREFIX,
991 	.get	= pidfs_xattr_get,
992 	.set	= pidfs_xattr_set,
993 };
994 
995 static const struct xattr_handler *const pidfs_xattr_handlers[] = {
996 	&pidfs_trusted_xattr_handler,
997 	NULL
998 };
999 
1000 static int pidfs_init_fs_context(struct fs_context *fc)
1001 {
1002 	struct pseudo_fs_context *ctx;
1003 
1004 	ctx = init_pseudo(fc, PID_FS_MAGIC);
1005 	if (!ctx)
1006 		return -ENOMEM;
1007 
1008 	fc->s_iflags |= SB_I_NOEXEC;
1009 	fc->s_iflags |= SB_I_NODEV;
1010 	ctx->ops = &pidfs_sops;
1011 	ctx->eops = &pidfs_export_operations;
1012 	ctx->dops = &pidfs_dentry_operations;
1013 	ctx->xattr = pidfs_xattr_handlers;
1014 	fc->s_fs_info = (void *)&pidfs_stashed_ops;
1015 	return 0;
1016 }
1017 
1018 static struct file_system_type pidfs_type = {
1019 	.name			= "pidfs",
1020 	.init_fs_context	= pidfs_init_fs_context,
1021 	.kill_sb		= kill_anon_super,
1022 };
1023 
1024 struct file *pidfs_alloc_file(struct pid *pid, unsigned int flags)
1025 {
1026 	struct file *pidfd_file;
1027 	struct path path __free(path_put) = {};
1028 	int ret;
1029 
1030 	/*
1031 	 * Ensure that PIDFD_STALE can be passed as a flag without
1032 	 * overloading other uapi pidfd flags.
1033 	 */
1034 	BUILD_BUG_ON(PIDFD_STALE == PIDFD_THREAD);
1035 	BUILD_BUG_ON(PIDFD_STALE == PIDFD_NONBLOCK);
1036 
1037 	ret = path_from_stashed(&pid->stashed, pidfs_mnt, get_pid(pid), &path);
1038 	if (ret < 0)
1039 		return ERR_PTR(ret);
1040 
1041 	flags &= ~PIDFD_STALE;
1042 	flags |= O_RDWR;
1043 	pidfd_file = dentry_open(&path, flags, current_cred());
1044 	/* Raise PIDFD_THREAD explicitly as do_dentry_open() strips it. */
1045 	if (!IS_ERR(pidfd_file))
1046 		pidfd_file->f_flags |= (flags & PIDFD_THREAD);
1047 
1048 	return pidfd_file;
1049 }
1050 
1051 void __init pidfs_init(void)
1052 {
1053 	pidfs_attr_cachep = kmem_cache_create("pidfs_attr_cache", sizeof(struct pidfs_attr), 0,
1054 					 (SLAB_HWCACHE_ALIGN | SLAB_RECLAIM_ACCOUNT |
1055 					  SLAB_ACCOUNT | SLAB_PANIC), NULL);
1056 
1057 	pidfs_xattr_cachep = kmem_cache_create("pidfs_xattr_cache",
1058 					       sizeof(struct simple_xattrs), 0,
1059 					       (SLAB_HWCACHE_ALIGN | SLAB_RECLAIM_ACCOUNT |
1060 						SLAB_ACCOUNT | SLAB_PANIC), NULL);
1061 
1062 	pidfs_mnt = kern_mount(&pidfs_type);
1063 	if (IS_ERR(pidfs_mnt))
1064 		panic("Failed to mount pidfs pseudo filesystem");
1065 }
1066