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 simple_inode_init_ts(inode);
122
123 inode_init_owner(&nop_mnt_idmap, inode, dir, mode);
124
125 return inode;
126}
127
128static int bpf_inode_type(const struct inode *inode, enum bpf_type *type)
129{
130 *type = BPF_TYPE_UNSPEC;
131 if (inode->i_op == &bpf_prog_iops)
132 *type = BPF_TYPE_PROG;
133 else if (inode->i_op == &bpf_map_iops)
134 *type = BPF_TYPE_MAP;
135 else if (inode->i_op == &bpf_link_iops)
136 *type = BPF_TYPE_LINK;
137 else
138 return -EACCES;
139
140 return 0;
141}
142
143static void bpf_dentry_finalize(struct dentry *dentry, struct inode *inode,
144 struct inode *dir)
145{
146 d_instantiate(dentry, inode);
147 dget(dentry);
148
149 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
150}
151
152static int bpf_mkdir(struct mnt_idmap *idmap, struct inode *dir,
153 struct dentry *dentry, umode_t mode)
154{
155 struct inode *inode;
156
157 inode = bpf_get_inode(dir->i_sb, dir, mode | S_IFDIR);
158 if (IS_ERR(inode))
159 return PTR_ERR(inode);
160
161 inode->i_op = &bpf_dir_iops;
162 inode->i_fop = &simple_dir_operations;
163
164 inc_nlink(inode);
165 inc_nlink(dir);
166
167 bpf_dentry_finalize(dentry, inode, dir);
168 return 0;
169}
170
171struct map_iter {
172 void *key;
173 bool done;
174};
175
176static struct map_iter *map_iter(struct seq_file *m)
177{
178 return m->private;
179}
180
181static struct bpf_map *seq_file_to_map(struct seq_file *m)
182{
183 return file_inode(m->file)->i_private;
184}
185
186static void map_iter_free(struct map_iter *iter)
187{
188 if (iter) {
189 kfree(iter->key);
190 kfree(iter);
191 }
192}
193
194static struct map_iter *map_iter_alloc(struct bpf_map *map)
195{
196 struct map_iter *iter;
197
198 iter = kzalloc(sizeof(*iter), GFP_KERNEL | __GFP_NOWARN);
199 if (!iter)
200 goto error;
201
202 iter->key = kzalloc(map->key_size, GFP_KERNEL | __GFP_NOWARN);
203 if (!iter->key)
204 goto error;
205
206 return iter;
207
208error:
209 map_iter_free(iter);
210 return NULL;
211}
212
213static void *map_seq_next(struct seq_file *m, void *v, loff_t *pos)
214{
215 struct bpf_map *map = seq_file_to_map(m);
216 void *key = map_iter(m)->key;
217 void *prev_key;
218
219 (*pos)++;
220 if (map_iter(m)->done)
221 return NULL;
222
223 if (unlikely(v == SEQ_START_TOKEN))
224 prev_key = NULL;
225 else
226 prev_key = key;
227
228 rcu_read_lock();
229 if (map->ops->map_get_next_key(map, prev_key, key)) {
230 map_iter(m)->done = true;
231 key = NULL;
232 }
233 rcu_read_unlock();
234 return key;
235}
236
237static void *map_seq_start(struct seq_file *m, loff_t *pos)
238{
239 if (map_iter(m)->done)
240 return NULL;
241
242 return *pos ? map_iter(m)->key : SEQ_START_TOKEN;
243}
244
245static void map_seq_stop(struct seq_file *m, void *v)
246{
247}
248
249static int map_seq_show(struct seq_file *m, void *v)
250{
251 struct bpf_map *map = seq_file_to_map(m);
252 void *key = map_iter(m)->key;
253
254 if (unlikely(v == SEQ_START_TOKEN)) {
255 seq_puts(m, "# WARNING!! The output is for debug purpose only\n");
256 seq_puts(m, "# WARNING!! The output format will change\n");
257 } else {
258 map->ops->map_seq_show_elem(map, key, m);
259 }
260
261 return 0;
262}
263
264static const struct seq_operations bpffs_map_seq_ops = {
265 .start = map_seq_start,
266 .next = map_seq_next,
267 .show = map_seq_show,
268 .stop = map_seq_stop,
269};
270
271static int bpffs_map_open(struct inode *inode, struct file *file)
272{
273 struct bpf_map *map = inode->i_private;
274 struct map_iter *iter;
275 struct seq_file *m;
276 int err;
277
278 iter = map_iter_alloc(map);
279 if (!iter)
280 return -ENOMEM;
281
282 err = seq_open(file, &bpffs_map_seq_ops);
283 if (err) {
284 map_iter_free(iter);
285 return err;
286 }
287
288 m = file->private_data;
289 m->private = iter;
290
291 return 0;
292}
293
294static int bpffs_map_release(struct inode *inode, struct file *file)
295{
296 struct seq_file *m = file->private_data;
297
298 map_iter_free(map_iter(m));
299
300 return seq_release(inode, file);
301}
302
303/* bpffs_map_fops should only implement the basic
304 * read operation for a BPF map. The purpose is to
305 * provide a simple user intuitive way to do
306 * "cat bpffs/pathto/a-pinned-map".
307 *
308 * Other operations (e.g. write, lookup...) should be realized by
309 * the userspace tools (e.g. bpftool) through the
310 * BPF_OBJ_GET_INFO_BY_FD and the map's lookup/update
311 * interface.
312 */
313static const struct file_operations bpffs_map_fops = {
314 .open = bpffs_map_open,
315 .read = seq_read,
316 .release = bpffs_map_release,
317};
318
319static int bpffs_obj_open(struct inode *inode, struct file *file)
320{
321 return -EIO;
322}
323
324static const struct file_operations bpffs_obj_fops = {
325 .open = bpffs_obj_open,
326};
327
328static int bpf_mkobj_ops(struct dentry *dentry, umode_t mode, void *raw,
329 const struct inode_operations *iops,
330 const struct file_operations *fops)
331{
332 struct inode *dir = dentry->d_parent->d_inode;
333 struct inode *inode = bpf_get_inode(dir->i_sb, dir, mode);
334 if (IS_ERR(inode))
335 return PTR_ERR(inode);
336
337 inode->i_op = iops;
338 inode->i_fop = fops;
339 inode->i_private = raw;
340
341 bpf_dentry_finalize(dentry, inode, dir);
342 return 0;
343}
344
345static int bpf_mkprog(struct dentry *dentry, umode_t mode, void *arg)
346{
347 return bpf_mkobj_ops(dentry, mode, arg, &bpf_prog_iops,
348 &bpffs_obj_fops);
349}
350
351static int bpf_mkmap(struct dentry *dentry, umode_t mode, void *arg)
352{
353 struct bpf_map *map = arg;
354
355 return bpf_mkobj_ops(dentry, mode, arg, &bpf_map_iops,
356 bpf_map_support_seq_show(map) ?
357 &bpffs_map_fops : &bpffs_obj_fops);
358}
359
360static int bpf_mklink(struct dentry *dentry, umode_t mode, void *arg)
361{
362 struct bpf_link *link = arg;
363
364 return bpf_mkobj_ops(dentry, mode, arg, &bpf_link_iops,
365 bpf_link_is_iter(link) ?
366 &bpf_iter_fops : &bpffs_obj_fops);
367}
368
369static struct dentry *
370bpf_lookup(struct inode *dir, struct dentry *dentry, unsigned flags)
371{
372 /* Dots in names (e.g. "/sys/fs/bpf/foo.bar") are reserved for future
373 * extensions. That allows popoulate_bpffs() create special files.
374 */
375 if ((dir->i_mode & S_IALLUGO) &&
376 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 mnt_idmap *idmap, struct inode *dir,
383 struct dentry *dentry, 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
414/* pin iterator link into bpffs */
415static int bpf_iter_link_pin_kernel(struct dentry *parent,
416 const char *name, struct bpf_link *link)
417{
418 umode_t mode = S_IFREG | S_IRUSR;
419 struct dentry *dentry;
420 int ret;
421
422 inode_lock(parent->d_inode);
423 dentry = lookup_one_len(name, parent, strlen(name));
424 if (IS_ERR(dentry)) {
425 inode_unlock(parent->d_inode);
426 return PTR_ERR(dentry);
427 }
428 ret = bpf_mkobj_ops(dentry, mode, link, &bpf_link_iops,
429 &bpf_iter_fops);
430 dput(dentry);
431 inode_unlock(parent->d_inode);
432 return ret;
433}
434
435static int bpf_obj_do_pin(int path_fd, const char __user *pathname, void *raw,
436 enum bpf_type type)
437{
438 struct dentry *dentry;
439 struct inode *dir;
440 struct path path;
441 umode_t mode;
442 int ret;
443
444 dentry = user_path_create(path_fd, pathname, &path, 0);
445 if (IS_ERR(dentry))
446 return PTR_ERR(dentry);
447
448 dir = d_inode(path.dentry);
449 if (dir->i_op != &bpf_dir_iops) {
450 ret = -EPERM;
451 goto out;
452 }
453
454 mode = S_IFREG | ((S_IRUSR | S_IWUSR) & ~current_umask());
455 ret = security_path_mknod(&path, dentry, mode, 0);
456 if (ret)
457 goto out;
458
459 switch (type) {
460 case BPF_TYPE_PROG:
461 ret = vfs_mkobj(dentry, mode, bpf_mkprog, raw);
462 break;
463 case BPF_TYPE_MAP:
464 ret = vfs_mkobj(dentry, mode, bpf_mkmap, raw);
465 break;
466 case BPF_TYPE_LINK:
467 ret = vfs_mkobj(dentry, mode, bpf_mklink, raw);
468 break;
469 default:
470 ret = -EPERM;
471 }
472out:
473 done_path_create(&path, dentry);
474 return ret;
475}
476
477int bpf_obj_pin_user(u32 ufd, int path_fd, const char __user *pathname)
478{
479 enum bpf_type type;
480 void *raw;
481 int ret;
482
483 raw = bpf_fd_probe_obj(ufd, &type);
484 if (IS_ERR(raw))
485 return PTR_ERR(raw);
486
487 ret = bpf_obj_do_pin(path_fd, pathname, raw, type);
488 if (ret != 0)
489 bpf_any_put(raw, type);
490
491 return ret;
492}
493
494static void *bpf_obj_do_get(int path_fd, const char __user *pathname,
495 enum bpf_type *type, int flags)
496{
497 struct inode *inode;
498 struct path path;
499 void *raw;
500 int ret;
501
502 ret = user_path_at(path_fd, pathname, LOOKUP_FOLLOW, &path);
503 if (ret)
504 return ERR_PTR(ret);
505
506 inode = d_backing_inode(path.dentry);
507 ret = path_permission(&path, ACC_MODE(flags));
508 if (ret)
509 goto out;
510
511 ret = bpf_inode_type(inode, type);
512 if (ret)
513 goto out;
514
515 raw = bpf_any_get(inode->i_private, *type);
516 if (!IS_ERR(raw))
517 touch_atime(&path);
518
519 path_put(&path);
520 return raw;
521out:
522 path_put(&path);
523 return ERR_PTR(ret);
524}
525
526int bpf_obj_get_user(int path_fd, const char __user *pathname, int flags)
527{
528 enum bpf_type type = BPF_TYPE_UNSPEC;
529 int f_flags;
530 void *raw;
531 int ret;
532
533 f_flags = bpf_get_file_flag(flags);
534 if (f_flags < 0)
535 return f_flags;
536
537 raw = bpf_obj_do_get(path_fd, pathname, &type, f_flags);
538 if (IS_ERR(raw))
539 return PTR_ERR(raw);
540
541 if (type == BPF_TYPE_PROG)
542 ret = bpf_prog_new_fd(raw);
543 else if (type == BPF_TYPE_MAP)
544 ret = bpf_map_new_fd(raw, f_flags);
545 else if (type == BPF_TYPE_LINK)
546 ret = (f_flags != O_RDWR) ? -EINVAL : bpf_link_new_fd(raw);
547 else
548 return -ENOENT;
549
550 if (ret < 0)
551 bpf_any_put(raw, type);
552 return ret;
553}
554
555static struct bpf_prog *__get_prog_inode(struct inode *inode, enum bpf_prog_type type)
556{
557 struct bpf_prog *prog;
558 int ret = inode_permission(&nop_mnt_idmap, inode, MAY_READ);
559 if (ret)
560 return ERR_PTR(ret);
561
562 if (inode->i_op == &bpf_map_iops)
563 return ERR_PTR(-EINVAL);
564 if (inode->i_op == &bpf_link_iops)
565 return ERR_PTR(-EINVAL);
566 if (inode->i_op != &bpf_prog_iops)
567 return ERR_PTR(-EACCES);
568
569 prog = inode->i_private;
570
571 ret = security_bpf_prog(prog);
572 if (ret < 0)
573 return ERR_PTR(ret);
574
575 if (!bpf_prog_get_ok(prog, &type, false))
576 return ERR_PTR(-EINVAL);
577
578 bpf_prog_inc(prog);
579 return prog;
580}
581
582struct bpf_prog *bpf_prog_get_type_path(const char *name, enum bpf_prog_type type)
583{
584 struct bpf_prog *prog;
585 struct path path;
586 int ret = kern_path(name, LOOKUP_FOLLOW, &path);
587 if (ret)
588 return ERR_PTR(ret);
589 prog = __get_prog_inode(d_backing_inode(path.dentry), type);
590 if (!IS_ERR(prog))
591 touch_atime(&path);
592 path_put(&path);
593 return prog;
594}
595EXPORT_SYMBOL(bpf_prog_get_type_path);
596
597/*
598 * Display the mount options in /proc/mounts.
599 */
600static int bpf_show_options(struct seq_file *m, struct dentry *root)
601{
602 struct inode *inode = d_inode(root);
603 umode_t mode = inode->i_mode & S_IALLUGO & ~S_ISVTX;
604
605 if (!uid_eq(inode->i_uid, GLOBAL_ROOT_UID))
606 seq_printf(m, ",uid=%u",
607 from_kuid_munged(&init_user_ns, inode->i_uid));
608 if (!gid_eq(inode->i_gid, GLOBAL_ROOT_GID))
609 seq_printf(m, ",gid=%u",
610 from_kgid_munged(&init_user_ns, inode->i_gid));
611 if (mode != S_IRWXUGO)
612 seq_printf(m, ",mode=%o", mode);
613 return 0;
614}
615
616static void bpf_free_inode(struct inode *inode)
617{
618 enum bpf_type type;
619
620 if (S_ISLNK(inode->i_mode))
621 kfree(inode->i_link);
622 if (!bpf_inode_type(inode, &type))
623 bpf_any_put(inode->i_private, type);
624 free_inode_nonrcu(inode);
625}
626
627static const struct super_operations bpf_super_ops = {
628 .statfs = simple_statfs,
629 .drop_inode = generic_delete_inode,
630 .show_options = bpf_show_options,
631 .free_inode = bpf_free_inode,
632};
633
634enum {
635 OPT_UID,
636 OPT_GID,
637 OPT_MODE,
638};
639
640static const struct fs_parameter_spec bpf_fs_parameters[] = {
641 fsparam_u32 ("uid", OPT_UID),
642 fsparam_u32 ("gid", OPT_GID),
643 fsparam_u32oct ("mode", OPT_MODE),
644 {}
645};
646
647struct bpf_mount_opts {
648 kuid_t uid;
649 kgid_t gid;
650 umode_t mode;
651};
652
653static int bpf_parse_param(struct fs_context *fc, struct fs_parameter *param)
654{
655 struct bpf_mount_opts *opts = fc->fs_private;
656 struct fs_parse_result result;
657 kuid_t uid;
658 kgid_t gid;
659 int opt;
660
661 opt = fs_parse(fc, bpf_fs_parameters, param, &result);
662 if (opt < 0) {
663 /* We might like to report bad mount options here, but
664 * traditionally we've ignored all mount options, so we'd
665 * better continue to ignore non-existing options for bpf.
666 */
667 if (opt == -ENOPARAM) {
668 opt = vfs_parse_fs_param_source(fc, param);
669 if (opt != -ENOPARAM)
670 return opt;
671
672 return 0;
673 }
674
675 if (opt < 0)
676 return opt;
677 }
678
679 switch (opt) {
680 case OPT_UID:
681 uid = make_kuid(current_user_ns(), result.uint_32);
682 if (!uid_valid(uid))
683 goto bad_value;
684
685 /*
686 * The requested uid must be representable in the
687 * filesystem's idmapping.
688 */
689 if (!kuid_has_mapping(fc->user_ns, uid))
690 goto bad_value;
691
692 opts->uid = uid;
693 break;
694 case OPT_GID:
695 gid = make_kgid(current_user_ns(), result.uint_32);
696 if (!gid_valid(gid))
697 goto bad_value;
698
699 /*
700 * The requested gid must be representable in the
701 * filesystem's idmapping.
702 */
703 if (!kgid_has_mapping(fc->user_ns, gid))
704 goto bad_value;
705
706 opts->gid = gid;
707 break;
708 case OPT_MODE:
709 opts->mode = result.uint_32 & S_IALLUGO;
710 break;
711 }
712
713 return 0;
714bad_value:
715 return invalfc(fc, "Bad value for '%s'", param->key);
716}
717
718struct bpf_preload_ops *bpf_preload_ops;
719EXPORT_SYMBOL_GPL(bpf_preload_ops);
720
721static bool bpf_preload_mod_get(void)
722{
723 /* If bpf_preload.ko wasn't loaded earlier then load it now.
724 * When bpf_preload is built into vmlinux the module's __init
725 * function will populate it.
726 */
727 if (!bpf_preload_ops) {
728 request_module("bpf_preload");
729 if (!bpf_preload_ops)
730 return false;
731 }
732 /* And grab the reference, so the module doesn't disappear while the
733 * kernel is interacting with the kernel module and its UMD.
734 */
735 if (!try_module_get(bpf_preload_ops->owner)) {
736 pr_err("bpf_preload module get failed.\n");
737 return false;
738 }
739 return true;
740}
741
742static void bpf_preload_mod_put(void)
743{
744 if (bpf_preload_ops)
745 /* now user can "rmmod bpf_preload" if necessary */
746 module_put(bpf_preload_ops->owner);
747}
748
749static DEFINE_MUTEX(bpf_preload_lock);
750
751static int populate_bpffs(struct dentry *parent)
752{
753 struct bpf_preload_info objs[BPF_PRELOAD_LINKS] = {};
754 int err = 0, i;
755
756 /* grab the mutex to make sure the kernel interactions with bpf_preload
757 * are serialized
758 */
759 mutex_lock(&bpf_preload_lock);
760
761 /* if bpf_preload.ko wasn't built into vmlinux then load it */
762 if (!bpf_preload_mod_get())
763 goto out;
764
765 err = bpf_preload_ops->preload(objs);
766 if (err)
767 goto out_put;
768 for (i = 0; i < BPF_PRELOAD_LINKS; i++) {
769 bpf_link_inc(objs[i].link);
770 err = bpf_iter_link_pin_kernel(parent,
771 objs[i].link_name, objs[i].link);
772 if (err) {
773 bpf_link_put(objs[i].link);
774 goto out_put;
775 }
776 }
777out_put:
778 bpf_preload_mod_put();
779out:
780 mutex_unlock(&bpf_preload_lock);
781 return err;
782}
783
784static int bpf_fill_super(struct super_block *sb, struct fs_context *fc)
785{
786 static const struct tree_descr bpf_rfiles[] = { { "" } };
787 struct bpf_mount_opts *opts = fc->fs_private;
788 struct inode *inode;
789 int ret;
790
791 ret = simple_fill_super(sb, BPF_FS_MAGIC, bpf_rfiles);
792 if (ret)
793 return ret;
794
795 sb->s_op = &bpf_super_ops;
796
797 inode = sb->s_root->d_inode;
798 inode->i_uid = opts->uid;
799 inode->i_gid = opts->gid;
800 inode->i_op = &bpf_dir_iops;
801 inode->i_mode &= ~S_IALLUGO;
802 populate_bpffs(sb->s_root);
803 inode->i_mode |= S_ISVTX | opts->mode;
804 return 0;
805}
806
807static int bpf_get_tree(struct fs_context *fc)
808{
809 return get_tree_nodev(fc, bpf_fill_super);
810}
811
812static void bpf_free_fc(struct fs_context *fc)
813{
814 kfree(fc->fs_private);
815}
816
817static const struct fs_context_operations bpf_context_ops = {
818 .free = bpf_free_fc,
819 .parse_param = bpf_parse_param,
820 .get_tree = bpf_get_tree,
821};
822
823/*
824 * Set up the filesystem mount context.
825 */
826static int bpf_init_fs_context(struct fs_context *fc)
827{
828 struct bpf_mount_opts *opts;
829
830 opts = kzalloc(sizeof(struct bpf_mount_opts), GFP_KERNEL);
831 if (!opts)
832 return -ENOMEM;
833
834 opts->mode = S_IRWXUGO;
835 opts->uid = current_fsuid();
836 opts->gid = current_fsgid();
837
838 fc->fs_private = opts;
839 fc->ops = &bpf_context_ops;
840 return 0;
841}
842
843static struct file_system_type bpf_fs_type = {
844 .owner = THIS_MODULE,
845 .name = "bpf",
846 .init_fs_context = bpf_init_fs_context,
847 .parameters = bpf_fs_parameters,
848 .kill_sb = kill_litter_super,
849};
850
851static int __init bpf_init(void)
852{
853 int ret;
854
855 ret = sysfs_create_mount_point(fs_kobj, "bpf");
856 if (ret)
857 return ret;
858
859 ret = register_filesystem(&bpf_fs_type);
860 if (ret)
861 sysfs_remove_mount_point(fs_kobj, "bpf");
862
863 return ret;
864}
865fs_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);