2 * Functions to manage eBPF programs attached to cgroups
4 * Copyright (c) 2016 Daniel Mack
6 * This file is subject to the terms and conditions of version 2 of the GNU
7 * General Public License. See the file COPYING in the main directory of the
8 * Linux distribution for more details.
11 #include <linux/kernel.h>
12 #include <linux/atomic.h>
13 #include <linux/cgroup.h>
14 #include <linux/slab.h>
15 #include <linux/bpf.h>
16 #include <linux/bpf-cgroup.h>
19 DEFINE_STATIC_KEY_FALSE(cgroup_bpf_enabled_key);
20 EXPORT_SYMBOL(cgroup_bpf_enabled_key);
23 * cgroup_bpf_put() - put references of all bpf programs
24 * @cgrp: the cgroup to modify
26 void cgroup_bpf_put(struct cgroup *cgrp)
30 for (type = 0; type < ARRAY_SIZE(cgrp->bpf.progs); type++) {
31 struct list_head *progs = &cgrp->bpf.progs[type];
32 struct bpf_prog_list *pl, *tmp;
34 list_for_each_entry_safe(pl, tmp, progs, node) {
36 bpf_prog_put(pl->prog);
38 static_branch_dec(&cgroup_bpf_enabled_key);
40 bpf_prog_array_free(cgrp->bpf.effective[type]);
44 /* count number of elements in the list.
45 * it's slow but the list cannot be long
47 static u32 prog_list_length(struct list_head *head)
49 struct bpf_prog_list *pl;
52 list_for_each_entry(pl, head, node) {
60 /* if parent has non-overridable prog attached,
61 * disallow attaching new programs to the descendent cgroup.
62 * if parent has overridable or multi-prog, allow attaching
64 static bool hierarchy_allows_attach(struct cgroup *cgrp,
65 enum bpf_attach_type type,
70 p = cgroup_parent(cgrp);
74 u32 flags = p->bpf.flags[type];
77 if (flags & BPF_F_ALLOW_MULTI)
79 cnt = prog_list_length(&p->bpf.progs[type]);
80 WARN_ON_ONCE(cnt > 1);
82 return !!(flags & BPF_F_ALLOW_OVERRIDE);
88 /* compute a chain of effective programs for a given cgroup:
89 * start from the list of programs in this cgroup and add
90 * all parent programs.
91 * Note that parent's F_ALLOW_OVERRIDE-type program is yielding
92 * to programs in this cgroup
94 static int compute_effective_progs(struct cgroup *cgrp,
95 enum bpf_attach_type type,
96 struct bpf_prog_array __rcu **array)
98 struct bpf_prog_array __rcu *progs;
99 struct bpf_prog_list *pl;
100 struct cgroup *p = cgrp;
103 /* count number of effective programs by walking parents */
105 if (cnt == 0 || (p->bpf.flags[type] & BPF_F_ALLOW_MULTI))
106 cnt += prog_list_length(&p->bpf.progs[type]);
107 p = cgroup_parent(p);
110 progs = bpf_prog_array_alloc(cnt, GFP_KERNEL);
114 /* populate the array with effective progs */
118 if (cnt == 0 || (p->bpf.flags[type] & BPF_F_ALLOW_MULTI))
119 list_for_each_entry(pl,
120 &p->bpf.progs[type], node) {
123 rcu_dereference_protected(progs, 1)->
124 progs[cnt++] = pl->prog;
126 p = cgroup_parent(p);
133 static void activate_effective_progs(struct cgroup *cgrp,
134 enum bpf_attach_type type,
135 struct bpf_prog_array __rcu *array)
137 struct bpf_prog_array __rcu *old_array;
139 old_array = xchg(&cgrp->bpf.effective[type], array);
140 /* free prog array after grace period, since __cgroup_bpf_run_*()
141 * might be still walking the array
143 bpf_prog_array_free(old_array);
147 * cgroup_bpf_inherit() - inherit effective programs from parent
148 * @cgrp: the cgroup to modify
150 int cgroup_bpf_inherit(struct cgroup *cgrp)
152 /* has to use marco instead of const int, since compiler thinks
153 * that array below is variable length
155 #define NR ARRAY_SIZE(cgrp->bpf.effective)
156 struct bpf_prog_array __rcu *arrays[NR] = {};
159 for (i = 0; i < NR; i++)
160 INIT_LIST_HEAD(&cgrp->bpf.progs[i]);
162 for (i = 0; i < NR; i++)
163 if (compute_effective_progs(cgrp, i, &arrays[i]))
166 for (i = 0; i < NR; i++)
167 activate_effective_progs(cgrp, i, arrays[i]);
171 for (i = 0; i < NR; i++)
172 bpf_prog_array_free(arrays[i]);
176 #define BPF_CGROUP_MAX_PROGS 64
179 * __cgroup_bpf_attach() - Attach the program to a cgroup, and
180 * propagate the change to descendants
181 * @cgrp: The cgroup which descendants to traverse
182 * @prog: A program to attach
183 * @type: Type of attach operation
185 * Must be called with cgroup_mutex held.
187 int __cgroup_bpf_attach(struct cgroup *cgrp, struct bpf_prog *prog,
188 enum bpf_attach_type type, u32 flags)
190 struct list_head *progs = &cgrp->bpf.progs[type];
191 struct bpf_prog *old_prog = NULL;
192 struct cgroup_subsys_state *css;
193 struct bpf_prog_list *pl;
194 bool pl_was_allocated;
197 if ((flags & BPF_F_ALLOW_OVERRIDE) && (flags & BPF_F_ALLOW_MULTI))
198 /* invalid combination */
201 if (!hierarchy_allows_attach(cgrp, type, flags))
204 if (!list_empty(progs) && cgrp->bpf.flags[type] != flags)
205 /* Disallow attaching non-overridable on top
206 * of existing overridable in this cgroup.
207 * Disallow attaching multi-prog if overridable or none
211 if (prog_list_length(progs) >= BPF_CGROUP_MAX_PROGS)
214 if (flags & BPF_F_ALLOW_MULTI) {
215 list_for_each_entry(pl, progs, node)
216 if (pl->prog == prog)
217 /* disallow attaching the same prog twice */
220 pl = kmalloc(sizeof(*pl), GFP_KERNEL);
223 pl_was_allocated = true;
225 list_add_tail(&pl->node, progs);
227 if (list_empty(progs)) {
228 pl = kmalloc(sizeof(*pl), GFP_KERNEL);
231 pl_was_allocated = true;
232 list_add_tail(&pl->node, progs);
234 pl = list_first_entry(progs, typeof(*pl), node);
236 pl_was_allocated = false;
241 cgrp->bpf.flags[type] = flags;
243 /* allocate and recompute effective prog arrays */
244 css_for_each_descendant_pre(css, &cgrp->self) {
245 struct cgroup *desc = container_of(css, struct cgroup, self);
247 err = compute_effective_progs(desc, type, &desc->bpf.inactive);
252 /* all allocations were successful. Activate all prog arrays */
253 css_for_each_descendant_pre(css, &cgrp->self) {
254 struct cgroup *desc = container_of(css, struct cgroup, self);
256 activate_effective_progs(desc, type, desc->bpf.inactive);
257 desc->bpf.inactive = NULL;
260 static_branch_inc(&cgroup_bpf_enabled_key);
262 bpf_prog_put(old_prog);
263 static_branch_dec(&cgroup_bpf_enabled_key);
268 /* oom while computing effective. Free all computed effective arrays
269 * since they were not activated
271 css_for_each_descendant_pre(css, &cgrp->self) {
272 struct cgroup *desc = container_of(css, struct cgroup, self);
274 bpf_prog_array_free(desc->bpf.inactive);
275 desc->bpf.inactive = NULL;
278 /* and cleanup the prog list */
280 if (pl_was_allocated) {
288 * __cgroup_bpf_detach() - Detach the program from a cgroup, and
289 * propagate the change to descendants
290 * @cgrp: The cgroup which descendants to traverse
291 * @prog: A program to detach or NULL
292 * @type: Type of detach operation
294 * Must be called with cgroup_mutex held.
296 int __cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
297 enum bpf_attach_type type, u32 unused_flags)
299 struct list_head *progs = &cgrp->bpf.progs[type];
300 u32 flags = cgrp->bpf.flags[type];
301 struct bpf_prog *old_prog = NULL;
302 struct cgroup_subsys_state *css;
303 struct bpf_prog_list *pl;
306 if (flags & BPF_F_ALLOW_MULTI) {
308 /* to detach MULTI prog the user has to specify valid FD
309 * of the program to be detached
313 if (list_empty(progs))
314 /* report error when trying to detach and nothing is attached */
318 if (flags & BPF_F_ALLOW_MULTI) {
319 /* find the prog and detach it */
320 list_for_each_entry(pl, progs, node) {
321 if (pl->prog != prog)
324 /* mark it deleted, so it's ignored while
325 * recomputing effective
333 /* to maintain backward compatibility NONE and OVERRIDE cgroups
334 * allow detaching with invalid FD (prog==NULL)
336 pl = list_first_entry(progs, typeof(*pl), node);
341 /* allocate and recompute effective prog arrays */
342 css_for_each_descendant_pre(css, &cgrp->self) {
343 struct cgroup *desc = container_of(css, struct cgroup, self);
345 err = compute_effective_progs(desc, type, &desc->bpf.inactive);
350 /* all allocations were successful. Activate all prog arrays */
351 css_for_each_descendant_pre(css, &cgrp->self) {
352 struct cgroup *desc = container_of(css, struct cgroup, self);
354 activate_effective_progs(desc, type, desc->bpf.inactive);
355 desc->bpf.inactive = NULL;
358 /* now can actually delete it from this cgroup list */
361 if (list_empty(progs))
362 /* last program was detached, reset flags to zero */
363 cgrp->bpf.flags[type] = 0;
365 bpf_prog_put(old_prog);
366 static_branch_dec(&cgroup_bpf_enabled_key);
370 /* oom while computing effective. Free all computed effective arrays
371 * since they were not activated
373 css_for_each_descendant_pre(css, &cgrp->self) {
374 struct cgroup *desc = container_of(css, struct cgroup, self);
376 bpf_prog_array_free(desc->bpf.inactive);
377 desc->bpf.inactive = NULL;
380 /* and restore back old_prog */
385 /* Must be called with cgroup_mutex held to avoid races. */
386 int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
387 union bpf_attr __user *uattr)
389 __u32 __user *prog_ids = u64_to_user_ptr(attr->query.prog_ids);
390 enum bpf_attach_type type = attr->query.attach_type;
391 struct list_head *progs = &cgrp->bpf.progs[type];
392 u32 flags = cgrp->bpf.flags[type];
395 if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE)
396 cnt = bpf_prog_array_length(cgrp->bpf.effective[type]);
398 cnt = prog_list_length(progs);
400 if (copy_to_user(&uattr->query.attach_flags, &flags, sizeof(flags)))
402 if (copy_to_user(&uattr->query.prog_cnt, &cnt, sizeof(cnt)))
404 if (attr->query.prog_cnt == 0 || !prog_ids || !cnt)
405 /* return early if user requested only program count + flags */
407 if (attr->query.prog_cnt < cnt) {
408 cnt = attr->query.prog_cnt;
412 if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE) {
413 return bpf_prog_array_copy_to_user(cgrp->bpf.effective[type],
416 struct bpf_prog_list *pl;
420 list_for_each_entry(pl, progs, node) {
421 id = pl->prog->aux->id;
422 if (copy_to_user(prog_ids + i, &id, sizeof(id)))
432 * __cgroup_bpf_run_filter_skb() - Run a program for packet filtering
433 * @sk: The socket sending or receiving traffic
434 * @skb: The skb that is being sent or received
435 * @type: The type of program to be exectuted
437 * If no socket is passed, or the socket is not of type INET or INET6,
438 * this function does nothing and returns 0.
440 * The program type passed in via @type must be suitable for network
441 * filtering. No further check is performed to assert that.
443 * This function will return %-EPERM if any if an attached program was found
444 * and if it returned != 1 during execution. In all other cases, 0 is returned.
446 int __cgroup_bpf_run_filter_skb(struct sock *sk,
448 enum bpf_attach_type type)
450 unsigned int offset = skb->data - skb_network_header(skb);
451 struct sock *save_sk;
455 if (!sk || !sk_fullsock(sk))
458 if (sk->sk_family != AF_INET && sk->sk_family != AF_INET6)
461 cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
464 __skb_push(skb, offset);
465 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], skb,
466 bpf_prog_run_save_cb);
467 __skb_pull(skb, offset);
469 return ret == 1 ? 0 : -EPERM;
471 EXPORT_SYMBOL(__cgroup_bpf_run_filter_skb);
474 * __cgroup_bpf_run_filter_sk() - Run a program on a sock
475 * @sk: sock structure to manipulate
476 * @type: The type of program to be exectuted
478 * socket is passed is expected to be of type INET or INET6.
480 * The program type passed in via @type must be suitable for sock
481 * filtering. No further check is performed to assert that.
483 * This function will return %-EPERM if any if an attached program was found
484 * and if it returned != 1 during execution. In all other cases, 0 is returned.
486 int __cgroup_bpf_run_filter_sk(struct sock *sk,
487 enum bpf_attach_type type)
489 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
492 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], sk, BPF_PROG_RUN);
493 return ret == 1 ? 0 : -EPERM;
495 EXPORT_SYMBOL(__cgroup_bpf_run_filter_sk);
498 * __cgroup_bpf_run_filter_sock_addr() - Run a program on a sock and
499 * provided by user sockaddr
500 * @sk: sock struct that will use sockaddr
501 * @uaddr: sockaddr struct provided by user
502 * @type: The type of program to be exectuted
504 * socket is expected to be of type INET or INET6.
506 * This function will return %-EPERM if an attached program is found and
507 * returned value != 1 during execution. In all other cases, 0 is returned.
509 int __cgroup_bpf_run_filter_sock_addr(struct sock *sk,
510 struct sockaddr *uaddr,
511 enum bpf_attach_type type)
513 struct bpf_sock_addr_kern ctx = {
520 /* Check socket family since not all sockets represent network
521 * endpoint (e.g. AF_UNIX).
523 if (sk->sk_family != AF_INET && sk->sk_family != AF_INET6)
526 cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
527 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], &ctx, BPF_PROG_RUN);
529 return ret == 1 ? 0 : -EPERM;
531 EXPORT_SYMBOL(__cgroup_bpf_run_filter_sock_addr);
534 * __cgroup_bpf_run_filter_sock_ops() - Run a program on a sock
535 * @sk: socket to get cgroup from
536 * @sock_ops: bpf_sock_ops_kern struct to pass to program. Contains
537 * sk with connection information (IP addresses, etc.) May not contain
538 * cgroup info if it is a req sock.
539 * @type: The type of program to be exectuted
541 * socket passed is expected to be of type INET or INET6.
543 * The program type passed in via @type must be suitable for sock_ops
544 * filtering. No further check is performed to assert that.
546 * This function will return %-EPERM if any if an attached program was found
547 * and if it returned != 1 during execution. In all other cases, 0 is returned.
549 int __cgroup_bpf_run_filter_sock_ops(struct sock *sk,
550 struct bpf_sock_ops_kern *sock_ops,
551 enum bpf_attach_type type)
553 struct cgroup *cgrp = sock_cgroup_ptr(&sk->sk_cgrp_data);
556 ret = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], sock_ops,
558 return ret == 1 ? 0 : -EPERM;
560 EXPORT_SYMBOL(__cgroup_bpf_run_filter_sock_ops);
562 int __cgroup_bpf_check_dev_permission(short dev_type, u32 major, u32 minor,
563 short access, enum bpf_attach_type type)
566 struct bpf_cgroup_dev_ctx ctx = {
567 .access_type = (access << 16) | dev_type,
574 cgrp = task_dfl_cgroup(current);
575 allow = BPF_PROG_RUN_ARRAY(cgrp->bpf.effective[type], &ctx,
581 EXPORT_SYMBOL(__cgroup_bpf_check_dev_permission);
583 static const struct bpf_func_proto *
584 cgroup_dev_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
587 case BPF_FUNC_map_lookup_elem:
588 return &bpf_map_lookup_elem_proto;
589 case BPF_FUNC_map_update_elem:
590 return &bpf_map_update_elem_proto;
591 case BPF_FUNC_map_delete_elem:
592 return &bpf_map_delete_elem_proto;
593 case BPF_FUNC_get_current_uid_gid:
594 return &bpf_get_current_uid_gid_proto;
595 case BPF_FUNC_trace_printk:
596 if (capable(CAP_SYS_ADMIN))
597 return bpf_get_trace_printk_proto();
603 static bool cgroup_dev_is_valid_access(int off, int size,
604 enum bpf_access_type type,
605 const struct bpf_prog *prog,
606 struct bpf_insn_access_aux *info)
608 const int size_default = sizeof(__u32);
610 if (type == BPF_WRITE)
613 if (off < 0 || off + size > sizeof(struct bpf_cgroup_dev_ctx))
615 /* The verifier guarantees that size > 0. */
620 case bpf_ctx_range(struct bpf_cgroup_dev_ctx, access_type):
621 bpf_ctx_record_field_size(info, size_default);
622 if (!bpf_ctx_narrow_access_ok(off, size, size_default))
626 if (size != size_default)
633 const struct bpf_prog_ops cg_dev_prog_ops = {
636 const struct bpf_verifier_ops cg_dev_verifier_ops = {
637 .get_func_proto = cgroup_dev_func_proto,
638 .is_valid_access = cgroup_dev_is_valid_access,