devtmpfs.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * devtmpfs - kernel-maintained tmpfs-based /dev
  4 *
  5 * Copyright (C) 2009, Kay Sievers <kay.sievers@vrfy.org>
  6 *
  7 * During bootup, before any driver core device is registered,
  8 * devtmpfs, a tmpfs-based filesystem is created. Every driver-core
  9 * device which requests a device node, will add a node in this
 10 * filesystem.
 11 * By default, all devices are named after the name of the device,
 12 * owned by root and have a default mode of 0600. Subsystems can
 13 * overwrite the default setting if needed.
 14 */
 15
 16#define pr_fmt(fmt) "devtmpfs: " fmt
 17
 18#include <linux/kernel.h>
 19#include <linux/syscalls.h>
 20#include <linux/mount.h>
 21#include <linux/device.h>
 22#include <linux/blkdev.h>
 23#include <linux/namei.h>
 24#include <linux/fs.h>
 25#include <linux/shmem_fs.h>
 26#include <linux/ramfs.h>
 27#include <linux/sched.h>
 28#include <linux/slab.h>
 29#include <linux/kthread.h>
 30#include <linux/init_syscalls.h>
 31#include <uapi/linux/mount.h>
 32#include "base.h"
 33
 34#ifdef CONFIG_DEVTMPFS_SAFE
 35#define DEVTMPFS_MFLAGS       (MS_SILENT | MS_NOEXEC | MS_NOSUID)
 36#else
 37#define DEVTMPFS_MFLAGS       (MS_SILENT)
 38#endif
 39
 40static struct task_struct *thread;
 41
 42static int __initdata mount_dev = IS_ENABLED(CONFIG_DEVTMPFS_MOUNT);
 43
 44static DEFINE_SPINLOCK(req_lock);
 45
 46static struct req {
 47	struct req *next;
 48	struct completion done;
 49	int err;
 50	const char *name;
 51	umode_t mode;	/* 0 => delete */
 52	kuid_t uid;
 53	kgid_t gid;
 54	struct device *dev;
 55} *requests;
 56
 57static int __init mount_param(char *str)
 58{
 59	mount_dev = simple_strtoul(str, NULL, 0);
 60	return 1;
 61}
 62__setup("devtmpfs.mount=", mount_param);
 63
 64static struct vfsmount *mnt;
 65
 66static struct file_system_type internal_fs_type = {
 67	.name = "devtmpfs",
 68#ifdef CONFIG_TMPFS
 69	.init_fs_context = shmem_init_fs_context,
 70#else
 71	.init_fs_context = ramfs_init_fs_context,
 72#endif
 73	.kill_sb = kill_anon_super,
 74};
 75
 76/* Simply take a ref on the existing mount */
 77static int devtmpfs_get_tree(struct fs_context *fc)
 78{
 79	struct super_block *sb = mnt->mnt_sb;
 80
 81	atomic_inc(&sb->s_active);
 82	down_write(&sb->s_umount);
 83	fc->root = dget(sb->s_root);
 84	return 0;
 85}
 86
 87/* Ops are filled in during init depending on underlying shmem or ramfs type */
 88struct fs_context_operations devtmpfs_context_ops = {};
 89
 90/* Call the underlying initialization and set to our ops */
 91static int devtmpfs_init_fs_context(struct fs_context *fc)
 92{
 93	int ret;
 94#ifdef CONFIG_TMPFS
 95	ret = shmem_init_fs_context(fc);
 96#else
 97	ret = ramfs_init_fs_context(fc);
 98#endif
 99	if (ret < 0)
100		return ret;
101
102	fc->ops = &devtmpfs_context_ops;
103
104	return 0;
105}
106
107static struct file_system_type dev_fs_type = {
108	.name = "devtmpfs",
109	.init_fs_context = devtmpfs_init_fs_context,
110};
111
112static int devtmpfs_submit_req(struct req *req, const char *tmp)
113{
114	init_completion(&req->done);
115
116	spin_lock(&req_lock);
117	req->next = requests;
118	requests = req;
119	spin_unlock(&req_lock);
120
121	wake_up_process(thread);
122	wait_for_completion(&req->done);
123
124	kfree(tmp);
125
126	return req->err;
127}
128
129int devtmpfs_create_node(struct device *dev)
130{
131	const char *tmp = NULL;
132	struct req req;
133
134	if (!thread)
135		return 0;
136
137	req.mode = 0;
138	req.uid = GLOBAL_ROOT_UID;
139	req.gid = GLOBAL_ROOT_GID;
140	req.name = device_get_devnode(dev, &req.mode, &req.uid, &req.gid, &tmp);
141	if (!req.name)
142		return -ENOMEM;
143
144	if (req.mode == 0)
145		req.mode = 0600;
146	if (is_blockdev(dev))
147		req.mode |= S_IFBLK;
148	else
149		req.mode |= S_IFCHR;
150
151	req.dev = dev;
152
153	return devtmpfs_submit_req(&req, tmp);
154}
155
156int devtmpfs_delete_node(struct device *dev)
157{
158	const char *tmp = NULL;
159	struct req req;
160
161	if (!thread)
162		return 0;
163
164	req.name = device_get_devnode(dev, NULL, NULL, NULL, &tmp);
165	if (!req.name)
166		return -ENOMEM;
167
168	req.mode = 0;
169	req.dev = dev;
170
171	return devtmpfs_submit_req(&req, tmp);
172}
173
174static int dev_mkdir(const char *name, umode_t mode)
175{
176	struct dentry *dentry;
177	struct path path;
178
179	dentry = start_creating_path(AT_FDCWD, name, &path, LOOKUP_DIRECTORY);
180	if (IS_ERR(dentry))
181		return PTR_ERR(dentry);
182
183	dentry = vfs_mkdir(&nop_mnt_idmap, d_inode(path.dentry), dentry, mode, NULL);
184	if (!IS_ERR(dentry))
185		/* mark as kernel-created inode */
186		d_inode(dentry)->i_private = &thread;
187	end_creating_path(&path, dentry);
188	return PTR_ERR_OR_ZERO(dentry);
189}
190
191static int create_path(const char *nodepath)
192{
193	char *path;
194	char *s;
195	int err = 0;
196
197	/* parent directories do not exist, create them */
198	path = kstrdup(nodepath, GFP_KERNEL);
199	if (!path)
200		return -ENOMEM;
201
202	s = path;
203	for (;;) {
204		s = strchr(s, '/');
205		if (!s)
206			break;
207		s[0] = '\0';
208		err = dev_mkdir(path, 0755);
209		if (err && err != -EEXIST)
210			break;
211		s[0] = '/';
212		s++;
213	}
214	kfree(path);
215	return err;
216}
217
218static int handle_create(const char *nodename, umode_t mode, kuid_t uid,
219			 kgid_t gid, struct device *dev)
220{
221	struct dentry *dentry;
222	struct path path;
223	int err;
224
225	dentry = start_creating_path(AT_FDCWD, nodename, &path, 0);
226	if (dentry == ERR_PTR(-ENOENT)) {
227		create_path(nodename);
228		dentry = start_creating_path(AT_FDCWD, nodename, &path, 0);
229	}
230	if (IS_ERR(dentry))
231		return PTR_ERR(dentry);
232
233	err = vfs_mknod(&nop_mnt_idmap, d_inode(path.dentry), dentry, mode,
234			dev->devt, NULL);
235	if (!err) {
236		struct iattr newattrs;
237
238		newattrs.ia_mode = mode;
239		newattrs.ia_uid = uid;
240		newattrs.ia_gid = gid;
241		newattrs.ia_valid = ATTR_MODE|ATTR_UID|ATTR_GID;
242		inode_lock(d_inode(dentry));
243		notify_change(&nop_mnt_idmap, dentry, &newattrs, NULL);
244		inode_unlock(d_inode(dentry));
245
246		/* mark as kernel-created inode */
247		d_inode(dentry)->i_private = &thread;
248	}
249	end_creating_path(&path, dentry);
250	return err;
251}
252
253static int dev_rmdir(const char *name)
254{
255	struct path parent;
256	struct dentry *dentry;
257	int err;
258
259	dentry = start_removing_path(name, &parent);
260	if (IS_ERR(dentry))
261		return PTR_ERR(dentry);
262	if (d_inode(dentry)->i_private == &thread)
263		err = vfs_rmdir(&nop_mnt_idmap, d_inode(parent.dentry),
264				dentry, NULL);
265	else
266		err = -EPERM;
267
268	end_removing_path(&parent, dentry);
269	return err;
270}
271
272static int delete_path(const char *nodepath)
273{
274	char *path;
275	int err = 0;
276
277	path = kstrdup(nodepath, GFP_KERNEL);
278	if (!path)
279		return -ENOMEM;
280
281	for (;;) {
282		char *base;
283
284		base = strrchr(path, '/');
285		if (!base)
286			break;
287		base[0] = '\0';
288		err = dev_rmdir(path);
289		if (err)
290			break;
291	}
292
293	kfree(path);
294	return err;
295}
296
297static int dev_mynode(struct device *dev, struct inode *inode)
298{
299	/* did we create it */
300	if (inode->i_private != &thread)
301		return 0;
302
303	/* does the dev_t match */
304	if (is_blockdev(dev)) {
305		if (!S_ISBLK(inode->i_mode))
306			return 0;
307	} else {
308		if (!S_ISCHR(inode->i_mode))
309			return 0;
310	}
311	if (inode->i_rdev != dev->devt)
312		return 0;
313
314	/* ours */
315	return 1;
316}
317
318static int handle_remove(const char *nodename, struct device *dev)
319{
320	struct path parent;
321	struct dentry *dentry;
322	struct inode *inode;
323	int deleted = 0;
324	int err = 0;
325
326	dentry = start_removing_path(nodename, &parent);
327	if (IS_ERR(dentry))
328		return PTR_ERR(dentry);
329
330	inode = d_inode(dentry);
331	if (dev_mynode(dev, inode)) {
332		struct iattr newattrs;
333		/*
334		 * before unlinking this node, reset permissions
335		 * of possible references like hardlinks
336		 */
337		newattrs.ia_uid = GLOBAL_ROOT_UID;
338		newattrs.ia_gid = GLOBAL_ROOT_GID;
339		newattrs.ia_mode = inode->i_mode & ~0777;
340		newattrs.ia_valid =
341			ATTR_UID|ATTR_GID|ATTR_MODE;
342		inode_lock(d_inode(dentry));
343		notify_change(&nop_mnt_idmap, dentry, &newattrs, NULL);
344		inode_unlock(d_inode(dentry));
345		err = vfs_unlink(&nop_mnt_idmap, d_inode(parent.dentry),
346				 dentry, NULL);
347		if (!err || err == -ENOENT)
348			deleted = 1;
349	}
350	end_removing_path(&parent, dentry);
351
352	if (deleted && strchr(nodename, '/'))
353		delete_path(nodename);
354	return err;
355}
356
357/*
358 * If configured, or requested by the commandline, devtmpfs will be
359 * auto-mounted after the kernel mounted the root filesystem.
360 */
361int __init devtmpfs_mount(void)
362{
363	int err;
364
365	if (!mount_dev)
366		return 0;
367
368	if (!thread)
369		return 0;
370
371	err = init_mount("devtmpfs", "dev", "devtmpfs", DEVTMPFS_MFLAGS, NULL);
372	if (err)
373		pr_info("error mounting %d\n", err);
374	else
375		pr_info("mounted\n");
376	return err;
377}
378
379static __initdata DECLARE_COMPLETION(setup_done);
380
381static int handle(const char *name, umode_t mode, kuid_t uid, kgid_t gid,
382		  struct device *dev)
383{
384	if (mode)
385		return handle_create(name, mode, uid, gid, dev);
386	else
387		return handle_remove(name, dev);
388}
389
390static void __noreturn devtmpfs_work_loop(void)
391{
392	while (1) {
393		spin_lock(&req_lock);
394		while (requests) {
395			struct req *req = requests;
396			requests = NULL;
397			spin_unlock(&req_lock);
398			while (req) {
399				struct req *next = req->next;
400				req->err = handle(req->name, req->mode,
401						  req->uid, req->gid, req->dev);
402				complete(&req->done);
403				req = next;
404			}
405			spin_lock(&req_lock);
406		}
407		__set_current_state(TASK_INTERRUPTIBLE);
408		spin_unlock(&req_lock);
409		schedule();
410	}
411}
412
413static noinline int __init devtmpfs_setup(void *p)
414{
415	int err;
416
417	err = ksys_unshare(CLONE_NEWNS);
418	if (err)
419		goto out;
420	err = init_mount("devtmpfs", "/", "devtmpfs", DEVTMPFS_MFLAGS, NULL);
421	if (err)
422		goto out;
423	init_chdir("/.."); /* will traverse into overmounted root */
424	init_chroot(".");
425out:
426	*(int *)p = err;
427	return err;
428}
429
430/*
431 * The __ref is because devtmpfs_setup needs to be __init for the routines it
432 * calls.  That call is done while devtmpfs_init, which is marked __init,
433 * synchronously waits for it to complete.
434 */
435static int __ref devtmpfsd(void *p)
436{
437	int err = devtmpfs_setup(p);
438
439	complete(&setup_done);
440	if (err)
441		return err;
442	devtmpfs_work_loop();
443	return 0;
444}
445
446/*
447 * Get the underlying (shmem/ramfs) context ops to build ours
448 */
449static int devtmpfs_configure_context(void)
450{
451	struct fs_context *fc;
452
453	fc = fs_context_for_reconfigure(mnt->mnt_root, mnt->mnt_sb->s_flags,
454					MS_RMT_MASK);
455	if (IS_ERR(fc))
456		return PTR_ERR(fc);
457
458	/* Set up devtmpfs_context_ops based on underlying type */
459	devtmpfs_context_ops.free	      = fc->ops->free;
460	devtmpfs_context_ops.dup	      = fc->ops->dup;
461	devtmpfs_context_ops.parse_param      = fc->ops->parse_param;
462	devtmpfs_context_ops.parse_monolithic = fc->ops->parse_monolithic;
463	devtmpfs_context_ops.get_tree	      = &devtmpfs_get_tree;
464	devtmpfs_context_ops.reconfigure      = fc->ops->reconfigure;
465
466	put_fs_context(fc);
467
468	return 0;
469}
470
471/*
472 * Create devtmpfs instance, driver-core devices will add their device
473 * nodes here.
474 */
475int __init devtmpfs_init(void)
476{
477	char opts[] = "mode=0755";
478	int err;
479
480	mnt = vfs_kern_mount(&internal_fs_type, 0, "devtmpfs", opts);
481	if (IS_ERR(mnt)) {
482		pr_err("unable to create devtmpfs %ld\n", PTR_ERR(mnt));
483		return PTR_ERR(mnt);
484	}
485
486	err = devtmpfs_configure_context();
487	if (err) {
488		pr_err("unable to configure devtmpfs type %d\n", err);
489		return err;
490	}
491
492	err = register_filesystem(&dev_fs_type);
493	if (err) {
494		pr_err("unable to register devtmpfs type %d\n", err);
495		return err;
496	}
497
498	thread = kthread_run(devtmpfsd, &err, "kdevtmpfs");
499	if (!IS_ERR(thread)) {
500		wait_for_completion(&setup_done);
501	} else {
502		err = PTR_ERR(thread);
503		thread = NULL;
504	}
505
506	if (err) {
507		pr_err("unable to create devtmpfs %d\n", err);
508		unregister_filesystem(&dev_fs_type);
509		thread = NULL;
510		return err;
511	}
512
513	pr_info("initialized\n");
514	return 0;
515}