1// SPDX-License-Identifier: GPL-2.0
  2/* XSKMAP used for AF_XDP sockets
  3 * Copyright(c) 2018 Intel Corporation.
  4 */
  5
  6#include <linux/bpf.h>
  7#include <linux/capability.h>
  8#include <net/xdp_sock.h>
  9#include <linux/slab.h>
 10#include <linux/sched.h>
 11
 12#include "xsk.h"
 13
 14static struct xsk_map_node *xsk_map_node_alloc(struct xsk_map *map,
 15					       struct xdp_sock __rcu **map_entry)
 16{
 17	struct xsk_map_node *node;
 18
 19	node = bpf_map_kzalloc(&map->map, sizeof(*node),
 20			       GFP_ATOMIC | __GFP_NOWARN);
 21	if (!node)
 22		return ERR_PTR(-ENOMEM);
 23
 24	bpf_map_inc(&map->map);
 25
 26	node->map = map;
 27	node->map_entry = map_entry;
 28	return node;
 29}
 30
 31static void xsk_map_node_free(struct xsk_map_node *node)
 32{
 33	bpf_map_put(&node->map->map);
 34	kfree(node);
 35}
 36
 37static void xsk_map_sock_add(struct xdp_sock *xs, struct xsk_map_node *node)
 38{
 39	spin_lock_bh(&xs->map_list_lock);
 40	list_add_tail(&node->node, &xs->map_list);
 41	spin_unlock_bh(&xs->map_list_lock);
 42}
 43
 44static void xsk_map_sock_delete(struct xdp_sock *xs,
 45				struct xdp_sock __rcu **map_entry)
 46{
 47	struct xsk_map_node *n, *tmp;
 48
 49	spin_lock_bh(&xs->map_list_lock);
 50	list_for_each_entry_safe(n, tmp, &xs->map_list, node) {
 51		if (map_entry == n->map_entry) {
 52			list_del(&n->node);
 53			xsk_map_node_free(n);
 54		}
 55	}
 56	spin_unlock_bh(&xs->map_list_lock);
 57}
 58
 59static struct bpf_map *xsk_map_alloc(union bpf_attr *attr)
 60{
 61	struct xsk_map *m;
 62	int numa_node;
 63	u64 size;
 64
 65	if (!capable(CAP_NET_ADMIN))
 66		return ERR_PTR(-EPERM);
 67
 68	if (attr->max_entries == 0 || attr->key_size != 4 ||
 69	    attr->value_size != 4 ||
 70	    attr->map_flags & ~(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY))
 71		return ERR_PTR(-EINVAL);
 72
 73	numa_node = bpf_map_attr_numa_node(attr);
 74	size = struct_size(m, xsk_map, attr->max_entries);
 75
 76	m = bpf_map_area_alloc(size, numa_node);
 77	if (!m)
 78		return ERR_PTR(-ENOMEM);
 79
 80	bpf_map_init_from_attr(&m->map, attr);
 81	spin_lock_init(&m->lock);
 82
 83	return &m->map;
 84}
 85
 86static void xsk_map_free(struct bpf_map *map)
 87{
 88	struct xsk_map *m = container_of(map, struct xsk_map, map);
 89
 90	synchronize_net();
 91	bpf_map_area_free(m);
 92}
 93
 94static int xsk_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 95{
 96	struct xsk_map *m = container_of(map, struct xsk_map, map);
 97	u32 index = key ? *(u32 *)key : U32_MAX;
 98	u32 *next = next_key;
 99
100	if (index >= m->map.max_entries) {
101		*next = 0;
102		return 0;
103	}
104
105	if (index == m->map.max_entries - 1)
106		return -ENOENT;
107	*next = index + 1;
108	return 0;
109}
110
111static int xsk_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
112{
113	const int ret = BPF_REG_0, mp = BPF_REG_1, index = BPF_REG_2;
114	struct bpf_insn *insn = insn_buf;
115
116	*insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
117	*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
118	*insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(sizeof(struct xsk_sock *)));
119	*insn++ = BPF_ALU64_IMM(BPF_ADD, mp, offsetof(struct xsk_map, xsk_map));
120	*insn++ = BPF_ALU64_REG(BPF_ADD, ret, mp);
121	*insn++ = BPF_LDX_MEM(BPF_SIZEOF(struct xsk_sock *), ret, ret, 0);
122	*insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
123	*insn++ = BPF_MOV64_IMM(ret, 0);
124	return insn - insn_buf;
125}
126
127/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
128 * by local_bh_disable() (from XDP calls inside NAPI). The
129 * rcu_read_lock_bh_held() below makes lockdep accept both.
130 */
131static void *__xsk_map_lookup_elem(struct bpf_map *map, u32 key)
132{
133	struct xsk_map *m = container_of(map, struct xsk_map, map);
134
135	if (key >= map->max_entries)
136		return NULL;
137
138	return rcu_dereference_check(m->xsk_map[key], rcu_read_lock_bh_held());
139}
140
141static void *xsk_map_lookup_elem(struct bpf_map *map, void *key)
142{
143	return __xsk_map_lookup_elem(map, *(u32 *)key);
144}
145
146static void *xsk_map_lookup_elem_sys_only(struct bpf_map *map, void *key)
147{
148	return ERR_PTR(-EOPNOTSUPP);
149}
150
151static int xsk_map_update_elem(struct bpf_map *map, void *key, void *value,
152			       u64 map_flags)
153{
154	struct xsk_map *m = container_of(map, struct xsk_map, map);
155	struct xdp_sock __rcu **map_entry;
156	struct xdp_sock *xs, *old_xs;
157	u32 i = *(u32 *)key, fd = *(u32 *)value;
158	struct xsk_map_node *node;
159	struct socket *sock;
160	int err;
161
162	if (unlikely(map_flags > BPF_EXIST))
163		return -EINVAL;
164	if (unlikely(i >= m->map.max_entries))
165		return -E2BIG;
166
167	sock = sockfd_lookup(fd, &err);
168	if (!sock)
169		return err;
170
171	if (sock->sk->sk_family != PF_XDP) {
172		sockfd_put(sock);
173		return -EOPNOTSUPP;
174	}
175
176	xs = (struct xdp_sock *)sock->sk;
177
178	map_entry = &m->xsk_map[i];
179	node = xsk_map_node_alloc(m, map_entry);
180	if (IS_ERR(node)) {
181		sockfd_put(sock);
182		return PTR_ERR(node);
183	}
184
185	spin_lock_bh(&m->lock);
186	old_xs = rcu_dereference_protected(*map_entry, lockdep_is_held(&m->lock));
187	if (old_xs == xs) {
188		err = 0;
189		goto out;
190	} else if (old_xs && map_flags == BPF_NOEXIST) {
191		err = -EEXIST;
192		goto out;
193	} else if (!old_xs && map_flags == BPF_EXIST) {
194		err = -ENOENT;
195		goto out;
196	}
197	xsk_map_sock_add(xs, node);
198	rcu_assign_pointer(*map_entry, xs);
199	if (old_xs)
200		xsk_map_sock_delete(old_xs, map_entry);
201	spin_unlock_bh(&m->lock);
202	sockfd_put(sock);
203	return 0;
204
205out:
206	spin_unlock_bh(&m->lock);
207	sockfd_put(sock);
208	xsk_map_node_free(node);
209	return err;
210}
211
212static int xsk_map_delete_elem(struct bpf_map *map, void *key)
213{
214	struct xsk_map *m = container_of(map, struct xsk_map, map);
215	struct xdp_sock __rcu **map_entry;
216	struct xdp_sock *old_xs;
217	int k = *(u32 *)key;
218
219	if (k >= map->max_entries)
220		return -EINVAL;
221
222	spin_lock_bh(&m->lock);
223	map_entry = &m->xsk_map[k];
224	old_xs = unrcu_pointer(xchg(map_entry, NULL));
225	if (old_xs)
226		xsk_map_sock_delete(old_xs, map_entry);
227	spin_unlock_bh(&m->lock);
228
229	return 0;
230}
231
232static int xsk_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
233{
234	return __bpf_xdp_redirect_map(map, ifindex, flags, 0,
235				      __xsk_map_lookup_elem);
236}
237
238void xsk_map_try_sock_delete(struct xsk_map *map, struct xdp_sock *xs,
239			     struct xdp_sock __rcu **map_entry)
240{
241	spin_lock_bh(&map->lock);
242	if (rcu_access_pointer(*map_entry) == xs) {
243		rcu_assign_pointer(*map_entry, NULL);
244		xsk_map_sock_delete(xs, map_entry);
245	}
246	spin_unlock_bh(&map->lock);
247}
248
249static bool xsk_map_meta_equal(const struct bpf_map *meta0,
250			       const struct bpf_map *meta1)
251{
252	return meta0->max_entries == meta1->max_entries &&
253		bpf_map_meta_equal(meta0, meta1);
254}
255
256static int xsk_map_btf_id;
257const struct bpf_map_ops xsk_map_ops = {
258	.map_meta_equal = xsk_map_meta_equal,
259	.map_alloc = xsk_map_alloc,
260	.map_free = xsk_map_free,
261	.map_get_next_key = xsk_map_get_next_key,
262	.map_lookup_elem = xsk_map_lookup_elem,
263	.map_gen_lookup = xsk_map_gen_lookup,
264	.map_lookup_elem_sys_only = xsk_map_lookup_elem_sys_only,
265	.map_update_elem = xsk_map_update_elem,
266	.map_delete_elem = xsk_map_delete_elem,
267	.map_check_btf = map_check_no_btf,
268	.map_btf_name = "xsk_map",
269	.map_btf_id = &xsk_map_btf_id,
270	.map_redirect = xsk_map_redirect,
271};