xdp/xskmap.c

   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/filter.h>
   8 #include <net/xdp_sock.h>
   9 #include <linux/slab.h>
  10 #include <linux/sched.h>
  11 #include <linux/btf_ids.h>
  12
  13 #include "xsk.h"
  14
  15 static struct xsk_map_node *xsk_map_node_alloc(struct xsk_map *map,
  16                                                struct xdp_sock __rcu **map_entry)
  17 {
  18         struct xsk_map_node *node;
  19
  20         node = bpf_map_kzalloc(&map->map, sizeof(*node),
  21                                GFP_ATOMIC | __GFP_NOWARN);
  22         if (!node)
  23                 return ERR_PTR(-ENOMEM);
  24
  25         bpf_map_inc(&map->map);
  26         atomic_inc(&map->count);
  27
  28         node->map = map;
  29         node->map_entry = map_entry;
  30         return node;
  31 }
  32
  33 static void xsk_map_node_free(struct xsk_map_node *node)
  34 {
  35         struct xsk_map *map = node->map;
  36
  37         bpf_map_put(&node->map->map);
  38         kfree(node);
  39         atomic_dec(&map->count);
  40 }
  41
  42 static void xsk_map_sock_add(struct xdp_sock *xs, struct xsk_map_node *node)
  43 {
  44         spin_lock_bh(&xs->map_list_lock);
  45         list_add_tail(&node->node, &xs->map_list);
  46         spin_unlock_bh(&xs->map_list_lock);
  47 }
  48
  49 static void xsk_map_sock_delete(struct xdp_sock *xs,
  50                                 struct xdp_sock __rcu **map_entry)
  51 {
  52         struct xsk_map_node *n, *tmp;
  53
  54         spin_lock_bh(&xs->map_list_lock);
  55         list_for_each_entry_safe(n, tmp, &xs->map_list, node) {
  56                 if (map_entry == n->map_entry) {
  57                         list_del(&n->node);
  58                         xsk_map_node_free(n);
  59                 }
  60         }
  61         spin_unlock_bh(&xs->map_list_lock);
  62 }
  63
  64 static struct bpf_map *xsk_map_alloc(union bpf_attr *attr)
  65 {
  66         struct xsk_map *m;
  67         int numa_node;
  68         u64 size;
  69
  70         if (attr->max_entries == 0 || attr->key_size != 4 ||
  71             attr->value_size != 4 ||
  72             attr->map_flags & ~(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY))
  73                 return ERR_PTR(-EINVAL);
  74
  75         numa_node = bpf_map_attr_numa_node(attr);
  76         size = struct_size(m, xsk_map, attr->max_entries);
  77
  78         m = bpf_map_area_alloc(size, numa_node);
  79         if (!m)
  80                 return ERR_PTR(-ENOMEM);
  81
  82         bpf_map_init_from_attr(&m->map, attr);
  83         spin_lock_init(&m->lock);
  84
  85         return &m->map;
  86 }
  87
  88 static u64 xsk_map_mem_usage(const struct bpf_map *map)
  89 {
  90         struct xsk_map *m = container_of(map, struct xsk_map, map);
  91
  92         return struct_size(m, xsk_map, map->max_entries) +
  93                    (u64)atomic_read(&m->count) * sizeof(struct xsk_map_node);
  94 }
  95
  96 static void xsk_map_free(struct bpf_map *map)
  97 {
  98         struct xsk_map *m = container_of(map, struct xsk_map, map);
  99
 100         synchronize_net();
 101         bpf_map_area_free(m);
 102 }
 103
 104 static int xsk_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 105 {
 106         struct xsk_map *m = container_of(map, struct xsk_map, map);
 107         u32 index = key ? *(u32 *)key : U32_MAX;
 108         u32 *next = next_key;
 109
 110         if (index >= m->map.max_entries) {
 111                 *next = 0;
 112                 return 0;
 113         }
 114
 115         if (index == m->map.max_entries - 1)
 116                 return -ENOENT;
 117         *next = index + 1;
 118         return 0;
 119 }
 120
 121 static int xsk_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 122 {
 123         const int ret = BPF_REG_0, mp = BPF_REG_1, index = BPF_REG_2;
 124         struct bpf_insn *insn = insn_buf;
 125
 126         *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
 127         *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
 128         *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(sizeof(struct xsk_sock *)));
 129         *insn++ = BPF_ALU64_IMM(BPF_ADD, mp, offsetof(struct xsk_map, xsk_map));
 130         *insn++ = BPF_ALU64_REG(BPF_ADD, ret, mp);
 131         *insn++ = BPF_LDX_MEM(BPF_SIZEOF(struct xsk_sock *), ret, ret, 0);
 132         *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
 133         *insn++ = BPF_MOV64_IMM(ret, 0);
 134         return insn - insn_buf;
 135 }
 136
 137 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 138  * by local_bh_disable() (from XDP calls inside NAPI). The
 139  * rcu_read_lock_bh_held() below makes lockdep accept both.
 140  */
 141 static void *__xsk_map_lookup_elem(struct bpf_map *map, u32 key)
 142 {
 143         struct xsk_map *m = container_of(map, struct xsk_map, map);
 144
 145         if (key >= map->max_entries)
 146                 return NULL;
 147
 148         return rcu_dereference_check(m->xsk_map[key], rcu_read_lock_bh_held());
 149 }
 150
 151 static void *xsk_map_lookup_elem(struct bpf_map *map, void *key)
 152 {
 153         return __xsk_map_lookup_elem(map, *(u32 *)key);
 154 }
 155
 156 static void *xsk_map_lookup_elem_sys_only(struct bpf_map *map, void *key)
 157 {
 158         return ERR_PTR(-EOPNOTSUPP);
 159 }
 160
 161 static long xsk_map_update_elem(struct bpf_map *map, void *key, void *value,
 162                                 u64 map_flags)
 163 {
 164         struct xsk_map *m = container_of(map, struct xsk_map, map);
 165         struct xdp_sock __rcu **map_entry;
 166         struct xdp_sock *xs, *old_xs;
 167         u32 i = *(u32 *)key, fd = *(u32 *)value;
 168         struct xsk_map_node *node;
 169         struct socket *sock;
 170         int err;
 171
 172         if (unlikely(map_flags > BPF_EXIST))
 173                 return -EINVAL;
 174         if (unlikely(i >= m->map.max_entries))
 175                 return -E2BIG;
 176
 177         sock = sockfd_lookup(fd, &err);
 178         if (!sock)
 179                 return err;
 180
 181         if (sock->sk->sk_family != PF_XDP) {
 182                 sockfd_put(sock);
 183                 return -EOPNOTSUPP;
 184         }
 185
 186         xs = (struct xdp_sock *)sock->sk;
 187
 188         map_entry = &m->xsk_map[i];
 189         node = xsk_map_node_alloc(m, map_entry);
 190         if (IS_ERR(node)) {
 191                 sockfd_put(sock);
 192                 return PTR_ERR(node);
 193         }
 194
 195         spin_lock_bh(&m->lock);
 196         old_xs = rcu_dereference_protected(*map_entry, lockdep_is_held(&m->lock));
 197         if (old_xs == xs) {
 198                 err = 0;
 199                 goto out;
 200         } else if (old_xs && map_flags == BPF_NOEXIST) {
 201                 err = -EEXIST;
 202                 goto out;
 203         } else if (!old_xs && map_flags == BPF_EXIST) {
 204                 err = -ENOENT;
 205                 goto out;
 206         }
 207         xsk_map_sock_add(xs, node);
 208         rcu_assign_pointer(*map_entry, xs);
 209         if (old_xs)
 210                 xsk_map_sock_delete(old_xs, map_entry);
 211         spin_unlock_bh(&m->lock);
 212         sockfd_put(sock);
 213         return 0;
 214
 215 out:
 216         spin_unlock_bh(&m->lock);
 217         sockfd_put(sock);
 218         xsk_map_node_free(node);
 219         return err;
 220 }
 221
 222 static long xsk_map_delete_elem(struct bpf_map *map, void *key)
 223 {
 224         struct xsk_map *m = container_of(map, struct xsk_map, map);
 225         struct xdp_sock __rcu **map_entry;
 226         struct xdp_sock *old_xs;
 227         int k = *(u32 *)key;
 228
 229         if (k >= map->max_entries)
 230                 return -EINVAL;
 231
 232         spin_lock_bh(&m->lock);
 233         map_entry = &m->xsk_map[k];
 234         old_xs = unrcu_pointer(xchg(map_entry, NULL));
 235         if (old_xs)
 236                 xsk_map_sock_delete(old_xs, map_entry);
 237         spin_unlock_bh(&m->lock);
 238
 239         return 0;
 240 }
 241
 242 static long xsk_map_redirect(struct bpf_map *map, u64 index, u64 flags)
 243 {
 244         return __bpf_xdp_redirect_map(map, index, flags, 0,
 245                                       __xsk_map_lookup_elem);
 246 }
 247
 248 void xsk_map_try_sock_delete(struct xsk_map *map, struct xdp_sock *xs,
 249                              struct xdp_sock __rcu **map_entry)
 250 {
 251         spin_lock_bh(&map->lock);
 252         if (rcu_access_pointer(*map_entry) == xs) {
 253                 rcu_assign_pointer(*map_entry, NULL);
 254                 xsk_map_sock_delete(xs, map_entry);
 255         }
 256         spin_unlock_bh(&map->lock);
 257 }
 258
 259 static bool xsk_map_meta_equal(const struct bpf_map *meta0,
 260                                const struct bpf_map *meta1)
 261 {
 262         return meta0->max_entries == meta1->max_entries &&
 263                 bpf_map_meta_equal(meta0, meta1);
 264 }
 265
 266 BTF_ID_LIST_SINGLE(xsk_map_btf_ids, struct, xsk_map)
 267 const struct bpf_map_ops xsk_map_ops = {
 268         .map_meta_equal = xsk_map_meta_equal,
 269         .map_alloc = xsk_map_alloc,
 270         .map_free = xsk_map_free,
 271         .map_get_next_key = xsk_map_get_next_key,
 272         .map_lookup_elem = xsk_map_lookup_elem,
 273         .map_gen_lookup = xsk_map_gen_lookup,
 274         .map_lookup_elem_sys_only = xsk_map_lookup_elem_sys_only,
 275         .map_update_elem = xsk_map_update_elem,
 276         .map_delete_elem = xsk_map_delete_elem,
 277         .map_check_btf = map_check_no_btf,
 278         .map_mem_usage = xsk_map_mem_usage,
 279         .map_btf_id = &xsk_map_btf_ids[0],
 280         .map_redirect = xsk_map_redirect,
 281 };