Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Minimal file system backend for holding eBPF maps and programs,
4 * used by bpf(2) object pinning.
5 *
6 * Authors:
7 *
8 * Daniel Borkmann <daniel@iogearbox.net>
9 */
10
11#include <linux/init.h>
12#include <linux/magic.h>
13#include <linux/major.h>
14#include <linux/mount.h>
15#include <linux/namei.h>
16#include <linux/fs.h>
17#include <linux/fs_context.h>
18#include <linux/fs_parser.h>
19#include <linux/kdev_t.h>
20#include <linux/filter.h>
21#include <linux/bpf.h>
22#include <linux/bpf_trace.h>
23#include "preload/bpf_preload.h"
24
25enum bpf_type {
26 BPF_TYPE_UNSPEC = 0,
27 BPF_TYPE_PROG,
28 BPF_TYPE_MAP,
29 BPF_TYPE_LINK,
30};
31
32static void *bpf_any_get(void *raw, enum bpf_type type)
33{
34 switch (type) {
35 case BPF_TYPE_PROG:
36 bpf_prog_inc(raw);
37 break;
38 case BPF_TYPE_MAP:
39 bpf_map_inc_with_uref(raw);
40 break;
41 case BPF_TYPE_LINK:
42 bpf_link_inc(raw);
43 break;
44 default:
45 WARN_ON_ONCE(1);
46 break;
47 }
48
49 return raw;
50}
51
52static void bpf_any_put(void *raw, enum bpf_type type)
53{
54 switch (type) {
55 case BPF_TYPE_PROG:
56 bpf_prog_put(raw);
57 break;
58 case BPF_TYPE_MAP:
59 bpf_map_put_with_uref(raw);
60 break;
61 case BPF_TYPE_LINK:
62 bpf_link_put(raw);
63 break;
64 default:
65 WARN_ON_ONCE(1);
66 break;
67 }
68}
69
70static void *bpf_fd_probe_obj(u32 ufd, enum bpf_type *type)
71{
72 void *raw;
73
74 raw = bpf_map_get_with_uref(ufd);
75 if (!IS_ERR(raw)) {
76 *type = BPF_TYPE_MAP;
77 return raw;
78 }
79
80 raw = bpf_prog_get(ufd);
81 if (!IS_ERR(raw)) {
82 *type = BPF_TYPE_PROG;
83 return raw;
84 }
85
86 raw = bpf_link_get_from_fd(ufd);
87 if (!IS_ERR(raw)) {
88 *type = BPF_TYPE_LINK;
89 return raw;
90 }
91
92 return ERR_PTR(-EINVAL);
93}
94
95static const struct inode_operations bpf_dir_iops;
96
97static const struct inode_operations bpf_prog_iops = { };
98static const struct inode_operations bpf_map_iops = { };
99static const struct inode_operations bpf_link_iops = { };
100
101static struct inode *bpf_get_inode(struct super_block *sb,
102 const struct inode *dir,
103 umode_t mode)
104{
105 struct inode *inode;
106
107 switch (mode & S_IFMT) {
108 case S_IFDIR:
109 case S_IFREG:
110 case S_IFLNK:
111 break;
112 default:
113 return ERR_PTR(-EINVAL);
114 }
115
116 inode = new_inode(sb);
117 if (!inode)
118 return ERR_PTR(-ENOSPC);
119
120 inode->i_ino = get_next_ino();
121 inode->i_atime = current_time(inode);
122 inode->i_mtime = inode->i_atime;
123 inode->i_ctime = inode->i_atime;
124
125 inode_init_owner(&init_user_ns, inode, dir, mode);
126
127 return inode;
128}
129
130static int bpf_inode_type(const struct inode *inode, enum bpf_type *type)
131{
132 *type = BPF_TYPE_UNSPEC;
133 if (inode->i_op == &bpf_prog_iops)
134 *type = BPF_TYPE_PROG;
135 else if (inode->i_op == &bpf_map_iops)
136 *type = BPF_TYPE_MAP;
137 else if (inode->i_op == &bpf_link_iops)
138 *type = BPF_TYPE_LINK;
139 else
140 return -EACCES;
141
142 return 0;
143}
144
145static void bpf_dentry_finalize(struct dentry *dentry, struct inode *inode,
146 struct inode *dir)
147{
148 d_instantiate(dentry, inode);
149 dget(dentry);
150
151 dir->i_mtime = current_time(dir);
152 dir->i_ctime = dir->i_mtime;
153}
154
155static int bpf_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
156 struct dentry *dentry, umode_t mode)
157{
158 struct inode *inode;
159
160 inode = bpf_get_inode(dir->i_sb, dir, mode | S_IFDIR);
161 if (IS_ERR(inode))
162 return PTR_ERR(inode);
163
164 inode->i_op = &bpf_dir_iops;
165 inode->i_fop = &simple_dir_operations;
166
167 inc_nlink(inode);
168 inc_nlink(dir);
169
170 bpf_dentry_finalize(dentry, inode, dir);
171 return 0;
172}
173
174struct map_iter {
175 void *key;
176 bool done;
177};
178
179static struct map_iter *map_iter(struct seq_file *m)
180{
181 return m->private;
182}
183
184static struct bpf_map *seq_file_to_map(struct seq_file *m)
185{
186 return file_inode(m->file)->i_private;
187}
188
189static void map_iter_free(struct map_iter *iter)
190{
191 if (iter) {
192 kfree(iter->key);
193 kfree(iter);
194 }
195}
196
197static struct map_iter *map_iter_alloc(struct bpf_map *map)
198{
199 struct map_iter *iter;
200
201 iter = kzalloc(sizeof(*iter), GFP_KERNEL | __GFP_NOWARN);
202 if (!iter)
203 goto error;
204
205 iter->key = kzalloc(map->key_size, GFP_KERNEL | __GFP_NOWARN);
206 if (!iter->key)
207 goto error;
208
209 return iter;
210
211error:
212 map_iter_free(iter);
213 return NULL;
214}
215
216static void *map_seq_next(struct seq_file *m, void *v, loff_t *pos)
217{
218 struct bpf_map *map = seq_file_to_map(m);
219 void *key = map_iter(m)->key;
220 void *prev_key;
221
222 (*pos)++;
223 if (map_iter(m)->done)
224 return NULL;
225
226 if (unlikely(v == SEQ_START_TOKEN))
227 prev_key = NULL;
228 else
229 prev_key = key;
230
231 rcu_read_lock();
232 if (map->ops->map_get_next_key(map, prev_key, key)) {
233 map_iter(m)->done = true;
234 key = NULL;
235 }
236 rcu_read_unlock();
237 return key;
238}
239
240static void *map_seq_start(struct seq_file *m, loff_t *pos)
241{
242 if (map_iter(m)->done)
243 return NULL;
244
245 return *pos ? map_iter(m)->key : SEQ_START_TOKEN;
246}
247
248static void map_seq_stop(struct seq_file *m, void *v)
249{
250}
251
252static int map_seq_show(struct seq_file *m, void *v)
253{
254 struct bpf_map *map = seq_file_to_map(m);
255 void *key = map_iter(m)->key;
256
257 if (unlikely(v == SEQ_START_TOKEN)) {
258 seq_puts(m, "# WARNING!! The output is for debug purpose only\n");
259 seq_puts(m, "# WARNING!! The output format will change\n");
260 } else {
261 map->ops->map_seq_show_elem(map, key, m);
262 }
263
264 return 0;
265}
266
267static const struct seq_operations bpffs_map_seq_ops = {
268 .start = map_seq_start,
269 .next = map_seq_next,
270 .show = map_seq_show,
271 .stop = map_seq_stop,
272};
273
274static int bpffs_map_open(struct inode *inode, struct file *file)
275{
276 struct bpf_map *map = inode->i_private;
277 struct map_iter *iter;
278 struct seq_file *m;
279 int err;
280
281 iter = map_iter_alloc(map);
282 if (!iter)
283 return -ENOMEM;
284
285 err = seq_open(file, &bpffs_map_seq_ops);
286 if (err) {
287 map_iter_free(iter);
288 return err;
289 }
290
291 m = file->private_data;
292 m->private = iter;
293
294 return 0;
295}
296
297static int bpffs_map_release(struct inode *inode, struct file *file)
298{
299 struct seq_file *m = file->private_data;
300
301 map_iter_free(map_iter(m));
302
303 return seq_release(inode, file);
304}
305
306/* bpffs_map_fops should only implement the basic
307 * read operation for a BPF map. The purpose is to
308 * provide a simple user intuitive way to do
309 * "cat bpffs/pathto/a-pinned-map".
310 *
311 * Other operations (e.g. write, lookup...) should be realized by
312 * the userspace tools (e.g. bpftool) through the
313 * BPF_OBJ_GET_INFO_BY_FD and the map's lookup/update
314 * interface.
315 */
316static const struct file_operations bpffs_map_fops = {
317 .open = bpffs_map_open,
318 .read = seq_read,
319 .release = bpffs_map_release,
320};
321
322static int bpffs_obj_open(struct inode *inode, struct file *file)
323{
324 return -EIO;
325}
326
327static const struct file_operations bpffs_obj_fops = {
328 .open = bpffs_obj_open,
329};
330
331static int bpf_mkobj_ops(struct dentry *dentry, umode_t mode, void *raw,
332 const struct inode_operations *iops,
333 const struct file_operations *fops)
334{
335 struct inode *dir = dentry->d_parent->d_inode;
336 struct inode *inode = bpf_get_inode(dir->i_sb, dir, mode);
337 if (IS_ERR(inode))
338 return PTR_ERR(inode);
339
340 inode->i_op = iops;
341 inode->i_fop = fops;
342 inode->i_private = raw;
343
344 bpf_dentry_finalize(dentry, inode, dir);
345 return 0;
346}
347
348static int bpf_mkprog(struct dentry *dentry, umode_t mode, void *arg)
349{
350 return bpf_mkobj_ops(dentry, mode, arg, &bpf_prog_iops,
351 &bpffs_obj_fops);
352}
353
354static int bpf_mkmap(struct dentry *dentry, umode_t mode, void *arg)
355{
356 struct bpf_map *map = arg;
357
358 return bpf_mkobj_ops(dentry, mode, arg, &bpf_map_iops,
359 bpf_map_support_seq_show(map) ?
360 &bpffs_map_fops : &bpffs_obj_fops);
361}
362
363static int bpf_mklink(struct dentry *dentry, umode_t mode, void *arg)
364{
365 struct bpf_link *link = arg;
366
367 return bpf_mkobj_ops(dentry, mode, arg, &bpf_link_iops,
368 bpf_link_is_iter(link) ?
369 &bpf_iter_fops : &bpffs_obj_fops);
370}
371
372static struct dentry *
373bpf_lookup(struct inode *dir, struct dentry *dentry, unsigned flags)
374{
375 /* Dots in names (e.g. "/sys/fs/bpf/foo.bar") are reserved for future
376 * extensions. That allows popoulate_bpffs() create special files.
377 */
378 if ((dir->i_mode & S_IALLUGO) &&
379 strchr(dentry->d_name.name, '.'))
380 return ERR_PTR(-EPERM);
381
382 return simple_lookup(dir, dentry, flags);
383}
384
385static int bpf_symlink(struct user_namespace *mnt_userns, struct inode *dir,
386 struct dentry *dentry, const char *target)
387{
388 char *link = kstrdup(target, GFP_USER | __GFP_NOWARN);
389 struct inode *inode;
390
391 if (!link)
392 return -ENOMEM;
393
394 inode = bpf_get_inode(dir->i_sb, dir, S_IRWXUGO | S_IFLNK);
395 if (IS_ERR(inode)) {
396 kfree(link);
397 return PTR_ERR(inode);
398 }
399
400 inode->i_op = &simple_symlink_inode_operations;
401 inode->i_link = link;
402
403 bpf_dentry_finalize(dentry, inode, dir);
404 return 0;
405}
406
407static const struct inode_operations bpf_dir_iops = {
408 .lookup = bpf_lookup,
409 .mkdir = bpf_mkdir,
410 .symlink = bpf_symlink,
411 .rmdir = simple_rmdir,
412 .rename = simple_rename,
413 .link = simple_link,
414 .unlink = simple_unlink,
415};
416
417/* pin iterator link into bpffs */
418static int bpf_iter_link_pin_kernel(struct dentry *parent,
419 const char *name, struct bpf_link *link)
420{
421 umode_t mode = S_IFREG | S_IRUSR;
422 struct dentry *dentry;
423 int ret;
424
425 inode_lock(parent->d_inode);
426 dentry = lookup_one_len(name, parent, strlen(name));
427 if (IS_ERR(dentry)) {
428 inode_unlock(parent->d_inode);
429 return PTR_ERR(dentry);
430 }
431 ret = bpf_mkobj_ops(dentry, mode, link, &bpf_link_iops,
432 &bpf_iter_fops);
433 dput(dentry);
434 inode_unlock(parent->d_inode);
435 return ret;
436}
437
438static int bpf_obj_do_pin(const char __user *pathname, void *raw,
439 enum bpf_type type)
440{
441 struct dentry *dentry;
442 struct inode *dir;
443 struct path path;
444 umode_t mode;
445 int ret;
446
447 dentry = user_path_create(AT_FDCWD, pathname, &path, 0);
448 if (IS_ERR(dentry))
449 return PTR_ERR(dentry);
450
451 mode = S_IFREG | ((S_IRUSR | S_IWUSR) & ~current_umask());
452
453 ret = security_path_mknod(&path, dentry, mode, 0);
454 if (ret)
455 goto out;
456
457 dir = d_inode(path.dentry);
458 if (dir->i_op != &bpf_dir_iops) {
459 ret = -EPERM;
460 goto out;
461 }
462
463 switch (type) {
464 case BPF_TYPE_PROG:
465 ret = vfs_mkobj(dentry, mode, bpf_mkprog, raw);
466 break;
467 case BPF_TYPE_MAP:
468 ret = vfs_mkobj(dentry, mode, bpf_mkmap, raw);
469 break;
470 case BPF_TYPE_LINK:
471 ret = vfs_mkobj(dentry, mode, bpf_mklink, raw);
472 break;
473 default:
474 ret = -EPERM;
475 }
476out:
477 done_path_create(&path, dentry);
478 return ret;
479}
480
481int bpf_obj_pin_user(u32 ufd, const char __user *pathname)
482{
483 enum bpf_type type;
484 void *raw;
485 int ret;
486
487 raw = bpf_fd_probe_obj(ufd, &type);
488 if (IS_ERR(raw))
489 return PTR_ERR(raw);
490
491 ret = bpf_obj_do_pin(pathname, raw, type);
492 if (ret != 0)
493 bpf_any_put(raw, type);
494
495 return ret;
496}
497
498static void *bpf_obj_do_get(const char __user *pathname,
499 enum bpf_type *type, int flags)
500{
501 struct inode *inode;
502 struct path path;
503 void *raw;
504 int ret;
505
506 ret = user_path_at(AT_FDCWD, pathname, LOOKUP_FOLLOW, &path);
507 if (ret)
508 return ERR_PTR(ret);
509
510 inode = d_backing_inode(path.dentry);
511 ret = path_permission(&path, ACC_MODE(flags));
512 if (ret)
513 goto out;
514
515 ret = bpf_inode_type(inode, type);
516 if (ret)
517 goto out;
518
519 raw = bpf_any_get(inode->i_private, *type);
520 if (!IS_ERR(raw))
521 touch_atime(&path);
522
523 path_put(&path);
524 return raw;
525out:
526 path_put(&path);
527 return ERR_PTR(ret);
528}
529
530int bpf_obj_get_user(const char __user *pathname, int flags)
531{
532 enum bpf_type type = BPF_TYPE_UNSPEC;
533 int f_flags;
534 void *raw;
535 int ret;
536
537 f_flags = bpf_get_file_flag(flags);
538 if (f_flags < 0)
539 return f_flags;
540
541 raw = bpf_obj_do_get(pathname, &type, f_flags);
542 if (IS_ERR(raw))
543 return PTR_ERR(raw);
544
545 if (type == BPF_TYPE_PROG)
546 ret = bpf_prog_new_fd(raw);
547 else if (type == BPF_TYPE_MAP)
548 ret = bpf_map_new_fd(raw, f_flags);
549 else if (type == BPF_TYPE_LINK)
550 ret = (f_flags != O_RDWR) ? -EINVAL : bpf_link_new_fd(raw);
551 else
552 return -ENOENT;
553
554 if (ret < 0)
555 bpf_any_put(raw, type);
556 return ret;
557}
558
559static struct bpf_prog *__get_prog_inode(struct inode *inode, enum bpf_prog_type type)
560{
561 struct bpf_prog *prog;
562 int ret = inode_permission(&init_user_ns, inode, MAY_READ);
563 if (ret)
564 return ERR_PTR(ret);
565
566 if (inode->i_op == &bpf_map_iops)
567 return ERR_PTR(-EINVAL);
568 if (inode->i_op == &bpf_link_iops)
569 return ERR_PTR(-EINVAL);
570 if (inode->i_op != &bpf_prog_iops)
571 return ERR_PTR(-EACCES);
572
573 prog = inode->i_private;
574
575 ret = security_bpf_prog(prog);
576 if (ret < 0)
577 return ERR_PTR(ret);
578
579 if (!bpf_prog_get_ok(prog, &type, false))
580 return ERR_PTR(-EINVAL);
581
582 bpf_prog_inc(prog);
583 return prog;
584}
585
586struct bpf_prog *bpf_prog_get_type_path(const char *name, enum bpf_prog_type type)
587{
588 struct bpf_prog *prog;
589 struct path path;
590 int ret = kern_path(name, LOOKUP_FOLLOW, &path);
591 if (ret)
592 return ERR_PTR(ret);
593 prog = __get_prog_inode(d_backing_inode(path.dentry), type);
594 if (!IS_ERR(prog))
595 touch_atime(&path);
596 path_put(&path);
597 return prog;
598}
599EXPORT_SYMBOL(bpf_prog_get_type_path);
600
601/*
602 * Display the mount options in /proc/mounts.
603 */
604static int bpf_show_options(struct seq_file *m, struct dentry *root)
605{
606 umode_t mode = d_inode(root)->i_mode & S_IALLUGO & ~S_ISVTX;
607
608 if (mode != S_IRWXUGO)
609 seq_printf(m, ",mode=%o", mode);
610 return 0;
611}
612
613static void bpf_free_inode(struct inode *inode)
614{
615 enum bpf_type type;
616
617 if (S_ISLNK(inode->i_mode))
618 kfree(inode->i_link);
619 if (!bpf_inode_type(inode, &type))
620 bpf_any_put(inode->i_private, type);
621 free_inode_nonrcu(inode);
622}
623
624static const struct super_operations bpf_super_ops = {
625 .statfs = simple_statfs,
626 .drop_inode = generic_delete_inode,
627 .show_options = bpf_show_options,
628 .free_inode = bpf_free_inode,
629};
630
631enum {
632 OPT_MODE,
633};
634
635static const struct fs_parameter_spec bpf_fs_parameters[] = {
636 fsparam_u32oct ("mode", OPT_MODE),
637 {}
638};
639
640struct bpf_mount_opts {
641 umode_t mode;
642};
643
644static int bpf_parse_param(struct fs_context *fc, struct fs_parameter *param)
645{
646 struct bpf_mount_opts *opts = fc->fs_private;
647 struct fs_parse_result result;
648 int opt;
649
650 opt = fs_parse(fc, bpf_fs_parameters, param, &result);
651 if (opt < 0) {
652 /* We might like to report bad mount options here, but
653 * traditionally we've ignored all mount options, so we'd
654 * better continue to ignore non-existing options for bpf.
655 */
656 if (opt == -ENOPARAM) {
657 opt = vfs_parse_fs_param_source(fc, param);
658 if (opt != -ENOPARAM)
659 return opt;
660
661 return 0;
662 }
663
664 if (opt < 0)
665 return opt;
666 }
667
668 switch (opt) {
669 case OPT_MODE:
670 opts->mode = result.uint_32 & S_IALLUGO;
671 break;
672 }
673
674 return 0;
675}
676
677struct bpf_preload_ops *bpf_preload_ops;
678EXPORT_SYMBOL_GPL(bpf_preload_ops);
679
680static bool bpf_preload_mod_get(void)
681{
682 /* If bpf_preload.ko wasn't loaded earlier then load it now.
683 * When bpf_preload is built into vmlinux the module's __init
684 * function will populate it.
685 */
686 if (!bpf_preload_ops) {
687 request_module("bpf_preload");
688 if (!bpf_preload_ops)
689 return false;
690 }
691 /* And grab the reference, so the module doesn't disappear while the
692 * kernel is interacting with the kernel module and its UMD.
693 */
694 if (!try_module_get(bpf_preload_ops->owner)) {
695 pr_err("bpf_preload module get failed.\n");
696 return false;
697 }
698 return true;
699}
700
701static void bpf_preload_mod_put(void)
702{
703 if (bpf_preload_ops)
704 /* now user can "rmmod bpf_preload" if necessary */
705 module_put(bpf_preload_ops->owner);
706}
707
708static DEFINE_MUTEX(bpf_preload_lock);
709
710static int populate_bpffs(struct dentry *parent)
711{
712 struct bpf_preload_info objs[BPF_PRELOAD_LINKS] = {};
713 int err = 0, i;
714
715 /* grab the mutex to make sure the kernel interactions with bpf_preload
716 * are serialized
717 */
718 mutex_lock(&bpf_preload_lock);
719
720 /* if bpf_preload.ko wasn't built into vmlinux then load it */
721 if (!bpf_preload_mod_get())
722 goto out;
723
724 err = bpf_preload_ops->preload(objs);
725 if (err)
726 goto out_put;
727 for (i = 0; i < BPF_PRELOAD_LINKS; i++) {
728 bpf_link_inc(objs[i].link);
729 err = bpf_iter_link_pin_kernel(parent,
730 objs[i].link_name, objs[i].link);
731 if (err) {
732 bpf_link_put(objs[i].link);
733 goto out_put;
734 }
735 }
736out_put:
737 bpf_preload_mod_put();
738out:
739 mutex_unlock(&bpf_preload_lock);
740 return err;
741}
742
743static int bpf_fill_super(struct super_block *sb, struct fs_context *fc)
744{
745 static const struct tree_descr bpf_rfiles[] = { { "" } };
746 struct bpf_mount_opts *opts = fc->fs_private;
747 struct inode *inode;
748 int ret;
749
750 ret = simple_fill_super(sb, BPF_FS_MAGIC, bpf_rfiles);
751 if (ret)
752 return ret;
753
754 sb->s_op = &bpf_super_ops;
755
756 inode = sb->s_root->d_inode;
757 inode->i_op = &bpf_dir_iops;
758 inode->i_mode &= ~S_IALLUGO;
759 populate_bpffs(sb->s_root);
760 inode->i_mode |= S_ISVTX | opts->mode;
761 return 0;
762}
763
764static int bpf_get_tree(struct fs_context *fc)
765{
766 return get_tree_nodev(fc, bpf_fill_super);
767}
768
769static void bpf_free_fc(struct fs_context *fc)
770{
771 kfree(fc->fs_private);
772}
773
774static const struct fs_context_operations bpf_context_ops = {
775 .free = bpf_free_fc,
776 .parse_param = bpf_parse_param,
777 .get_tree = bpf_get_tree,
778};
779
780/*
781 * Set up the filesystem mount context.
782 */
783static int bpf_init_fs_context(struct fs_context *fc)
784{
785 struct bpf_mount_opts *opts;
786
787 opts = kzalloc(sizeof(struct bpf_mount_opts), GFP_KERNEL);
788 if (!opts)
789 return -ENOMEM;
790
791 opts->mode = S_IRWXUGO;
792
793 fc->fs_private = opts;
794 fc->ops = &bpf_context_ops;
795 return 0;
796}
797
798static struct file_system_type bpf_fs_type = {
799 .owner = THIS_MODULE,
800 .name = "bpf",
801 .init_fs_context = bpf_init_fs_context,
802 .parameters = bpf_fs_parameters,
803 .kill_sb = kill_litter_super,
804};
805
806static int __init bpf_init(void)
807{
808 int ret;
809
810 ret = sysfs_create_mount_point(fs_kobj, "bpf");
811 if (ret)
812 return ret;
813
814 ret = register_filesystem(&bpf_fs_type);
815 if (ret)
816 sysfs_remove_mount_point(fs_kobj, "bpf");
817
818 return ret;
819}
820fs_initcall(bpf_init);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Minimal file system backend for holding eBPF maps and programs,
4 * used by bpf(2) object pinning.
5 *
6 * Authors:
7 *
8 * Daniel Borkmann <daniel@iogearbox.net>
9 */
10
11#include <linux/init.h>
12#include <linux/magic.h>
13#include <linux/major.h>
14#include <linux/mount.h>
15#include <linux/namei.h>
16#include <linux/fs.h>
17#include <linux/fs_context.h>
18#include <linux/fs_parser.h>
19#include <linux/kdev_t.h>
20#include <linux/filter.h>
21#include <linux/bpf.h>
22#include <linux/bpf_trace.h>
23
24enum bpf_type {
25 BPF_TYPE_UNSPEC = 0,
26 BPF_TYPE_PROG,
27 BPF_TYPE_MAP,
28 BPF_TYPE_LINK,
29};
30
31static void *bpf_any_get(void *raw, enum bpf_type type)
32{
33 switch (type) {
34 case BPF_TYPE_PROG:
35 bpf_prog_inc(raw);
36 break;
37 case BPF_TYPE_MAP:
38 bpf_map_inc_with_uref(raw);
39 break;
40 case BPF_TYPE_LINK:
41 bpf_link_inc(raw);
42 break;
43 default:
44 WARN_ON_ONCE(1);
45 break;
46 }
47
48 return raw;
49}
50
51static void bpf_any_put(void *raw, enum bpf_type type)
52{
53 switch (type) {
54 case BPF_TYPE_PROG:
55 bpf_prog_put(raw);
56 break;
57 case BPF_TYPE_MAP:
58 bpf_map_put_with_uref(raw);
59 break;
60 case BPF_TYPE_LINK:
61 bpf_link_put(raw);
62 break;
63 default:
64 WARN_ON_ONCE(1);
65 break;
66 }
67}
68
69static void *bpf_fd_probe_obj(u32 ufd, enum bpf_type *type)
70{
71 void *raw;
72
73 raw = bpf_map_get_with_uref(ufd);
74 if (!IS_ERR(raw)) {
75 *type = BPF_TYPE_MAP;
76 return raw;
77 }
78
79 raw = bpf_prog_get(ufd);
80 if (!IS_ERR(raw)) {
81 *type = BPF_TYPE_PROG;
82 return raw;
83 }
84
85 raw = bpf_link_get_from_fd(ufd);
86 if (!IS_ERR(raw)) {
87 *type = BPF_TYPE_LINK;
88 return raw;
89 }
90
91 return ERR_PTR(-EINVAL);
92}
93
94static const struct inode_operations bpf_dir_iops;
95
96static const struct inode_operations bpf_prog_iops = { };
97static const struct inode_operations bpf_map_iops = { };
98static const struct inode_operations bpf_link_iops = { };
99
100static struct inode *bpf_get_inode(struct super_block *sb,
101 const struct inode *dir,
102 umode_t mode)
103{
104 struct inode *inode;
105
106 switch (mode & S_IFMT) {
107 case S_IFDIR:
108 case S_IFREG:
109 case S_IFLNK:
110 break;
111 default:
112 return ERR_PTR(-EINVAL);
113 }
114
115 inode = new_inode(sb);
116 if (!inode)
117 return ERR_PTR(-ENOSPC);
118
119 inode->i_ino = get_next_ino();
120 inode->i_atime = current_time(inode);
121 inode->i_mtime = inode->i_atime;
122 inode->i_ctime = inode->i_atime;
123
124 inode_init_owner(inode, dir, mode);
125
126 return inode;
127}
128
129static int bpf_inode_type(const struct inode *inode, enum bpf_type *type)
130{
131 *type = BPF_TYPE_UNSPEC;
132 if (inode->i_op == &bpf_prog_iops)
133 *type = BPF_TYPE_PROG;
134 else if (inode->i_op == &bpf_map_iops)
135 *type = BPF_TYPE_MAP;
136 else if (inode->i_op == &bpf_link_iops)
137 *type = BPF_TYPE_LINK;
138 else
139 return -EACCES;
140
141 return 0;
142}
143
144static void bpf_dentry_finalize(struct dentry *dentry, struct inode *inode,
145 struct inode *dir)
146{
147 d_instantiate(dentry, inode);
148 dget(dentry);
149
150 dir->i_mtime = current_time(dir);
151 dir->i_ctime = dir->i_mtime;
152}
153
154static int bpf_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
155{
156 struct inode *inode;
157
158 inode = bpf_get_inode(dir->i_sb, dir, mode | S_IFDIR);
159 if (IS_ERR(inode))
160 return PTR_ERR(inode);
161
162 inode->i_op = &bpf_dir_iops;
163 inode->i_fop = &simple_dir_operations;
164
165 inc_nlink(inode);
166 inc_nlink(dir);
167
168 bpf_dentry_finalize(dentry, inode, dir);
169 return 0;
170}
171
172struct map_iter {
173 void *key;
174 bool done;
175};
176
177static struct map_iter *map_iter(struct seq_file *m)
178{
179 return m->private;
180}
181
182static struct bpf_map *seq_file_to_map(struct seq_file *m)
183{
184 return file_inode(m->file)->i_private;
185}
186
187static void map_iter_free(struct map_iter *iter)
188{
189 if (iter) {
190 kfree(iter->key);
191 kfree(iter);
192 }
193}
194
195static struct map_iter *map_iter_alloc(struct bpf_map *map)
196{
197 struct map_iter *iter;
198
199 iter = kzalloc(sizeof(*iter), GFP_KERNEL | __GFP_NOWARN);
200 if (!iter)
201 goto error;
202
203 iter->key = kzalloc(map->key_size, GFP_KERNEL | __GFP_NOWARN);
204 if (!iter->key)
205 goto error;
206
207 return iter;
208
209error:
210 map_iter_free(iter);
211 return NULL;
212}
213
214static void *map_seq_next(struct seq_file *m, void *v, loff_t *pos)
215{
216 struct bpf_map *map = seq_file_to_map(m);
217 void *key = map_iter(m)->key;
218 void *prev_key;
219
220 (*pos)++;
221 if (map_iter(m)->done)
222 return NULL;
223
224 if (unlikely(v == SEQ_START_TOKEN))
225 prev_key = NULL;
226 else
227 prev_key = key;
228
229 rcu_read_lock();
230 if (map->ops->map_get_next_key(map, prev_key, key)) {
231 map_iter(m)->done = true;
232 key = NULL;
233 }
234 rcu_read_unlock();
235 return key;
236}
237
238static void *map_seq_start(struct seq_file *m, loff_t *pos)
239{
240 if (map_iter(m)->done)
241 return NULL;
242
243 return *pos ? map_iter(m)->key : SEQ_START_TOKEN;
244}
245
246static void map_seq_stop(struct seq_file *m, void *v)
247{
248}
249
250static int map_seq_show(struct seq_file *m, void *v)
251{
252 struct bpf_map *map = seq_file_to_map(m);
253 void *key = map_iter(m)->key;
254
255 if (unlikely(v == SEQ_START_TOKEN)) {
256 seq_puts(m, "# WARNING!! The output is for debug purpose only\n");
257 seq_puts(m, "# WARNING!! The output format will change\n");
258 } else {
259 map->ops->map_seq_show_elem(map, key, m);
260 }
261
262 return 0;
263}
264
265static const struct seq_operations bpffs_map_seq_ops = {
266 .start = map_seq_start,
267 .next = map_seq_next,
268 .show = map_seq_show,
269 .stop = map_seq_stop,
270};
271
272static int bpffs_map_open(struct inode *inode, struct file *file)
273{
274 struct bpf_map *map = inode->i_private;
275 struct map_iter *iter;
276 struct seq_file *m;
277 int err;
278
279 iter = map_iter_alloc(map);
280 if (!iter)
281 return -ENOMEM;
282
283 err = seq_open(file, &bpffs_map_seq_ops);
284 if (err) {
285 map_iter_free(iter);
286 return err;
287 }
288
289 m = file->private_data;
290 m->private = iter;
291
292 return 0;
293}
294
295static int bpffs_map_release(struct inode *inode, struct file *file)
296{
297 struct seq_file *m = file->private_data;
298
299 map_iter_free(map_iter(m));
300
301 return seq_release(inode, file);
302}
303
304/* bpffs_map_fops should only implement the basic
305 * read operation for a BPF map. The purpose is to
306 * provide a simple user intuitive way to do
307 * "cat bpffs/pathto/a-pinned-map".
308 *
309 * Other operations (e.g. write, lookup...) should be realized by
310 * the userspace tools (e.g. bpftool) through the
311 * BPF_OBJ_GET_INFO_BY_FD and the map's lookup/update
312 * interface.
313 */
314static const struct file_operations bpffs_map_fops = {
315 .open = bpffs_map_open,
316 .read = seq_read,
317 .release = bpffs_map_release,
318};
319
320static int bpffs_obj_open(struct inode *inode, struct file *file)
321{
322 return -EIO;
323}
324
325static const struct file_operations bpffs_obj_fops = {
326 .open = bpffs_obj_open,
327};
328
329static int bpf_mkobj_ops(struct dentry *dentry, umode_t mode, void *raw,
330 const struct inode_operations *iops,
331 const struct file_operations *fops)
332{
333 struct inode *dir = dentry->d_parent->d_inode;
334 struct inode *inode = bpf_get_inode(dir->i_sb, dir, mode);
335 if (IS_ERR(inode))
336 return PTR_ERR(inode);
337
338 inode->i_op = iops;
339 inode->i_fop = fops;
340 inode->i_private = raw;
341
342 bpf_dentry_finalize(dentry, inode, dir);
343 return 0;
344}
345
346static int bpf_mkprog(struct dentry *dentry, umode_t mode, void *arg)
347{
348 return bpf_mkobj_ops(dentry, mode, arg, &bpf_prog_iops,
349 &bpffs_obj_fops);
350}
351
352static int bpf_mkmap(struct dentry *dentry, umode_t mode, void *arg)
353{
354 struct bpf_map *map = arg;
355
356 return bpf_mkobj_ops(dentry, mode, arg, &bpf_map_iops,
357 bpf_map_support_seq_show(map) ?
358 &bpffs_map_fops : &bpffs_obj_fops);
359}
360
361static int bpf_mklink(struct dentry *dentry, umode_t mode, void *arg)
362{
363 struct bpf_link *link = arg;
364
365 return bpf_mkobj_ops(dentry, mode, arg, &bpf_link_iops,
366 bpf_link_is_iter(link) ?
367 &bpf_iter_fops : &bpffs_obj_fops);
368}
369
370static struct dentry *
371bpf_lookup(struct inode *dir, struct dentry *dentry, unsigned flags)
372{
373 /* Dots in names (e.g. "/sys/fs/bpf/foo.bar") are reserved for future
374 * extensions.
375 */
376 if (strchr(dentry->d_name.name, '.'))
377 return ERR_PTR(-EPERM);
378
379 return simple_lookup(dir, dentry, flags);
380}
381
382static int bpf_symlink(struct inode *dir, struct dentry *dentry,
383 const char *target)
384{
385 char *link = kstrdup(target, GFP_USER | __GFP_NOWARN);
386 struct inode *inode;
387
388 if (!link)
389 return -ENOMEM;
390
391 inode = bpf_get_inode(dir->i_sb, dir, S_IRWXUGO | S_IFLNK);
392 if (IS_ERR(inode)) {
393 kfree(link);
394 return PTR_ERR(inode);
395 }
396
397 inode->i_op = &simple_symlink_inode_operations;
398 inode->i_link = link;
399
400 bpf_dentry_finalize(dentry, inode, dir);
401 return 0;
402}
403
404static const struct inode_operations bpf_dir_iops = {
405 .lookup = bpf_lookup,
406 .mkdir = bpf_mkdir,
407 .symlink = bpf_symlink,
408 .rmdir = simple_rmdir,
409 .rename = simple_rename,
410 .link = simple_link,
411 .unlink = simple_unlink,
412};
413
414static int bpf_obj_do_pin(const char __user *pathname, void *raw,
415 enum bpf_type type)
416{
417 struct dentry *dentry;
418 struct inode *dir;
419 struct path path;
420 umode_t mode;
421 int ret;
422
423 dentry = user_path_create(AT_FDCWD, pathname, &path, 0);
424 if (IS_ERR(dentry))
425 return PTR_ERR(dentry);
426
427 mode = S_IFREG | ((S_IRUSR | S_IWUSR) & ~current_umask());
428
429 ret = security_path_mknod(&path, dentry, mode, 0);
430 if (ret)
431 goto out;
432
433 dir = d_inode(path.dentry);
434 if (dir->i_op != &bpf_dir_iops) {
435 ret = -EPERM;
436 goto out;
437 }
438
439 switch (type) {
440 case BPF_TYPE_PROG:
441 ret = vfs_mkobj(dentry, mode, bpf_mkprog, raw);
442 break;
443 case BPF_TYPE_MAP:
444 ret = vfs_mkobj(dentry, mode, bpf_mkmap, raw);
445 break;
446 case BPF_TYPE_LINK:
447 ret = vfs_mkobj(dentry, mode, bpf_mklink, raw);
448 break;
449 default:
450 ret = -EPERM;
451 }
452out:
453 done_path_create(&path, dentry);
454 return ret;
455}
456
457int bpf_obj_pin_user(u32 ufd, const char __user *pathname)
458{
459 enum bpf_type type;
460 void *raw;
461 int ret;
462
463 raw = bpf_fd_probe_obj(ufd, &type);
464 if (IS_ERR(raw))
465 return PTR_ERR(raw);
466
467 ret = bpf_obj_do_pin(pathname, raw, type);
468 if (ret != 0)
469 bpf_any_put(raw, type);
470
471 return ret;
472}
473
474static void *bpf_obj_do_get(const char __user *pathname,
475 enum bpf_type *type, int flags)
476{
477 struct inode *inode;
478 struct path path;
479 void *raw;
480 int ret;
481
482 ret = user_path_at(AT_FDCWD, pathname, LOOKUP_FOLLOW, &path);
483 if (ret)
484 return ERR_PTR(ret);
485
486 inode = d_backing_inode(path.dentry);
487 ret = inode_permission(inode, ACC_MODE(flags));
488 if (ret)
489 goto out;
490
491 ret = bpf_inode_type(inode, type);
492 if (ret)
493 goto out;
494
495 raw = bpf_any_get(inode->i_private, *type);
496 if (!IS_ERR(raw))
497 touch_atime(&path);
498
499 path_put(&path);
500 return raw;
501out:
502 path_put(&path);
503 return ERR_PTR(ret);
504}
505
506int bpf_obj_get_user(const char __user *pathname, int flags)
507{
508 enum bpf_type type = BPF_TYPE_UNSPEC;
509 int f_flags;
510 void *raw;
511 int ret;
512
513 f_flags = bpf_get_file_flag(flags);
514 if (f_flags < 0)
515 return f_flags;
516
517 raw = bpf_obj_do_get(pathname, &type, f_flags);
518 if (IS_ERR(raw))
519 return PTR_ERR(raw);
520
521 if (type == BPF_TYPE_PROG)
522 ret = bpf_prog_new_fd(raw);
523 else if (type == BPF_TYPE_MAP)
524 ret = bpf_map_new_fd(raw, f_flags);
525 else if (type == BPF_TYPE_LINK)
526 ret = bpf_link_new_fd(raw);
527 else
528 return -ENOENT;
529
530 if (ret < 0)
531 bpf_any_put(raw, type);
532 return ret;
533}
534
535static struct bpf_prog *__get_prog_inode(struct inode *inode, enum bpf_prog_type type)
536{
537 struct bpf_prog *prog;
538 int ret = inode_permission(inode, MAY_READ);
539 if (ret)
540 return ERR_PTR(ret);
541
542 if (inode->i_op == &bpf_map_iops)
543 return ERR_PTR(-EINVAL);
544 if (inode->i_op == &bpf_link_iops)
545 return ERR_PTR(-EINVAL);
546 if (inode->i_op != &bpf_prog_iops)
547 return ERR_PTR(-EACCES);
548
549 prog = inode->i_private;
550
551 ret = security_bpf_prog(prog);
552 if (ret < 0)
553 return ERR_PTR(ret);
554
555 if (!bpf_prog_get_ok(prog, &type, false))
556 return ERR_PTR(-EINVAL);
557
558 bpf_prog_inc(prog);
559 return prog;
560}
561
562struct bpf_prog *bpf_prog_get_type_path(const char *name, enum bpf_prog_type type)
563{
564 struct bpf_prog *prog;
565 struct path path;
566 int ret = kern_path(name, LOOKUP_FOLLOW, &path);
567 if (ret)
568 return ERR_PTR(ret);
569 prog = __get_prog_inode(d_backing_inode(path.dentry), type);
570 if (!IS_ERR(prog))
571 touch_atime(&path);
572 path_put(&path);
573 return prog;
574}
575EXPORT_SYMBOL(bpf_prog_get_type_path);
576
577/*
578 * Display the mount options in /proc/mounts.
579 */
580static int bpf_show_options(struct seq_file *m, struct dentry *root)
581{
582 umode_t mode = d_inode(root)->i_mode & S_IALLUGO & ~S_ISVTX;
583
584 if (mode != S_IRWXUGO)
585 seq_printf(m, ",mode=%o", mode);
586 return 0;
587}
588
589static void bpf_free_inode(struct inode *inode)
590{
591 enum bpf_type type;
592
593 if (S_ISLNK(inode->i_mode))
594 kfree(inode->i_link);
595 if (!bpf_inode_type(inode, &type))
596 bpf_any_put(inode->i_private, type);
597 free_inode_nonrcu(inode);
598}
599
600static const struct super_operations bpf_super_ops = {
601 .statfs = simple_statfs,
602 .drop_inode = generic_delete_inode,
603 .show_options = bpf_show_options,
604 .free_inode = bpf_free_inode,
605};
606
607enum {
608 OPT_MODE,
609};
610
611static const struct fs_parameter_spec bpf_fs_parameters[] = {
612 fsparam_u32oct ("mode", OPT_MODE),
613 {}
614};
615
616struct bpf_mount_opts {
617 umode_t mode;
618};
619
620static int bpf_parse_param(struct fs_context *fc, struct fs_parameter *param)
621{
622 struct bpf_mount_opts *opts = fc->fs_private;
623 struct fs_parse_result result;
624 int opt;
625
626 opt = fs_parse(fc, bpf_fs_parameters, param, &result);
627 if (opt < 0)
628 /* We might like to report bad mount options here, but
629 * traditionally we've ignored all mount options, so we'd
630 * better continue to ignore non-existing options for bpf.
631 */
632 return opt == -ENOPARAM ? 0 : opt;
633
634 switch (opt) {
635 case OPT_MODE:
636 opts->mode = result.uint_32 & S_IALLUGO;
637 break;
638 }
639
640 return 0;
641}
642
643static int bpf_fill_super(struct super_block *sb, struct fs_context *fc)
644{
645 static const struct tree_descr bpf_rfiles[] = { { "" } };
646 struct bpf_mount_opts *opts = fc->fs_private;
647 struct inode *inode;
648 int ret;
649
650 ret = simple_fill_super(sb, BPF_FS_MAGIC, bpf_rfiles);
651 if (ret)
652 return ret;
653
654 sb->s_op = &bpf_super_ops;
655
656 inode = sb->s_root->d_inode;
657 inode->i_op = &bpf_dir_iops;
658 inode->i_mode &= ~S_IALLUGO;
659 inode->i_mode |= S_ISVTX | opts->mode;
660
661 return 0;
662}
663
664static int bpf_get_tree(struct fs_context *fc)
665{
666 return get_tree_nodev(fc, bpf_fill_super);
667}
668
669static void bpf_free_fc(struct fs_context *fc)
670{
671 kfree(fc->fs_private);
672}
673
674static const struct fs_context_operations bpf_context_ops = {
675 .free = bpf_free_fc,
676 .parse_param = bpf_parse_param,
677 .get_tree = bpf_get_tree,
678};
679
680/*
681 * Set up the filesystem mount context.
682 */
683static int bpf_init_fs_context(struct fs_context *fc)
684{
685 struct bpf_mount_opts *opts;
686
687 opts = kzalloc(sizeof(struct bpf_mount_opts), GFP_KERNEL);
688 if (!opts)
689 return -ENOMEM;
690
691 opts->mode = S_IRWXUGO;
692
693 fc->fs_private = opts;
694 fc->ops = &bpf_context_ops;
695 return 0;
696}
697
698static struct file_system_type bpf_fs_type = {
699 .owner = THIS_MODULE,
700 .name = "bpf",
701 .init_fs_context = bpf_init_fs_context,
702 .parameters = bpf_fs_parameters,
703 .kill_sb = kill_litter_super,
704};
705
706static int __init bpf_init(void)
707{
708 int ret;
709
710 ret = sysfs_create_mount_point(fs_kobj, "bpf");
711 if (ret)
712 return ret;
713
714 ret = register_filesystem(&bpf_fs_type);
715 if (ret)
716 sysfs_remove_mount_point(fs_kobj, "bpf");
717
718 return ret;
719}
720fs_initcall(bpf_init);