net/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 <linux/capability.h>
   9 #include <net/xdp_sock.h>
  10 #include <linux/slab.h>
  11 #include <linux/sched.h>
  12 #include <linux/btf_ids.h>
  13
  14 #include "xsk.h"
  15
  16 static struct xsk_map_node *xsk_map_node_alloc(struct xsk_map *map,
  17                                                struct xdp_sock __rcu **map_entry)
  18 {
  19         struct xsk_map_node *node;
  20
  21         node = bpf_map_kzalloc(&map->map, sizeof(*node),
  22                                GFP_ATOMIC | __GFP_NOWARN);
  23         if (!node)
  24                 return ERR_PTR(-ENOMEM);
  25
  26         bpf_map_inc(&map->map);
  27         atomic_inc(&map->count);
  28
  29         node->map = map;
  30         node->map_entry = map_entry;
  31         return node;
  32 }
  33
  34 static void xsk_map_node_free(struct xsk_map_node *node)
  35 {
  36         struct xsk_map *map = node->map;
  37
  38         bpf_map_put(&node->map->map);
  39         kfree(node);
  40         atomic_dec(&map->count);
  41 }
  42
  43 static void xsk_map_sock_add(struct xdp_sock *xs, struct xsk_map_node *node)
  44 {
  45         spin_lock_bh(&xs->map_list_lock);
  46         list_add_tail(&node->node, &xs->map_list);
  47         spin_unlock_bh(&xs->map_list_lock);
  48 }
  49
  50 static void xsk_map_sock_delete(struct xdp_sock *xs,
  51                                 struct xdp_sock __rcu **map_entry)
  52 {
  53         struct xsk_map_node *n, *tmp;
  54
  55         spin_lock_bh(&xs->map_list_lock);
  56         list_for_each_entry_safe(n, tmp, &xs->map_list, node) {
  57                 if (map_entry == n->map_entry) {
  58                         list_del(&n->node);
  59                         xsk_map_node_free(n);
  60                 }
  61         }
  62         spin_unlock_bh(&xs->map_list_lock);
  63 }
  64
  65 static struct bpf_map *xsk_map_alloc(union bpf_attr *attr)
  66 {
  67         struct xsk_map *m;
  68         int numa_node;
  69         u64 size;
  70
  71         if (!capable(CAP_NET_ADMIN))
  72                 return ERR_PTR(-EPERM);
  73
  74         if (attr->max_entries == 0 || attr->key_size != 4 ||
  75             attr->value_size != 4 ||
  76             attr->map_flags & ~(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY))
  77                 return ERR_PTR(-EINVAL);
  78
  79         numa_node = bpf_map_attr_numa_node(attr);
  80         size = struct_size(m, xsk_map, attr->max_entries);
  81
  82         m = bpf_map_area_alloc(size, numa_node);
  83         if (!m)
  84                 return ERR_PTR(-ENOMEM);
  85
  86         bpf_map_init_from_attr(&m->map, attr);
  87         spin_lock_init(&m->lock);
  88
  89         return &m->map;
  90 }
  91
  92 static u64 xsk_map_mem_usage(const struct bpf_map *map)
  93 {
  94         struct xsk_map *m = container_of(map, struct xsk_map, map);
  95
  96         return struct_size(m, xsk_map, map->max_entries) +
  97                    (u64)atomic_read(&m->count) * sizeof(struct xsk_map_node);
  98 }
  99
 100 static void xsk_map_free(struct bpf_map *map)
 101 {
 102         struct xsk_map *m = container_of(map, struct xsk_map, map);
 103
 104         synchronize_net();
 105         bpf_map_area_free(m);
 106 }
 107
 108 static int xsk_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 109 {
 110         struct xsk_map *m = container_of(map, struct xsk_map, map);
 111         u32 index = key ? *(u32 *)key : U32_MAX;
 112         u32 *next = next_key;
 113
 114         if (index >= m->map.max_entries) {
 115                 *next = 0;
 116                 return 0;
 117         }
 118
 119         if (index == m->map.max_entries - 1)
 120                 return -ENOENT;
 121         *next = index + 1;
 122         return 0;
 123 }
 124
 125 static int xsk_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 126 {
 127         const int ret = BPF_REG_0, mp = BPF_REG_1, index = BPF_REG_2;
 128         struct bpf_insn *insn = insn_buf;
 129
 130         *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
 131         *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
 132         *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(sizeof(struct xsk_sock *)));
 133         *insn++ = BPF_ALU64_IMM(BPF_ADD, mp, offsetof(struct xsk_map, xsk_map));
 134         *insn++ = BPF_ALU64_REG(BPF_ADD, ret, mp);
 135         *insn++ = BPF_LDX_MEM(BPF_SIZEOF(struct xsk_sock *), ret, ret, 0);
 136         *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
 137         *insn++ = BPF_MOV64_IMM(ret, 0);
 138         return insn - insn_buf;
 139 }
 140
 141 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 142  * by local_bh_disable() (from XDP calls inside NAPI). The
 143  * rcu_read_lock_bh_held() below makes lockdep accept both.
 144  */
 145 static void *__xsk_map_lookup_elem(struct bpf_map *map, u32 key)
 146 {
 147         struct xsk_map *m = container_of(map, struct xsk_map, map);
 148
 149         if (key >= map->max_entries)
 150                 return NULL;
 151
 152         return rcu_dereference_check(m->xsk_map[key], rcu_read_lock_bh_held());
 153 }
 154
 155 static void *xsk_map_lookup_elem(struct bpf_map *map, void *key)
 156 {
 157         return __xsk_map_lookup_elem(map, *(u32 *)key);
 158 }
 159
 160 static void *xsk_map_lookup_elem_sys_only(struct bpf_map *map, void *key)
 161 {
 162         return ERR_PTR(-EOPNOTSUPP);
 163 }
 164
 165 static long xsk_map_update_elem(struct bpf_map *map, void *key, void *value,
 166                                 u64 map_flags)
 167 {
 168         struct xsk_map *m = container_of(map, struct xsk_map, map);
 169         struct xdp_sock __rcu **map_entry;
 170         struct xdp_sock *xs, *old_xs;
 171         u32 i = *(u32 *)key, fd = *(u32 *)value;
 172         struct xsk_map_node *node;
 173         struct socket *sock;
 174         int err;
 175
 176         if (unlikely(map_flags > BPF_EXIST))
 177                 return -EINVAL;
 178         if (unlikely(i >= m->map.max_entries))
 179                 return -E2BIG;
 180
 181         sock = sockfd_lookup(fd, &err);
 182         if (!sock)
 183                 return err;
 184
 185         if (sock->sk->sk_family != PF_XDP) {
 186                 sockfd_put(sock);
 187                 return -EOPNOTSUPP;
 188         }
 189
 190         xs = (struct xdp_sock *)sock->sk;
 191
 192         map_entry = &m->xsk_map[i];
 193         node = xsk_map_node_alloc(m, map_entry);
 194         if (IS_ERR(node)) {
 195                 sockfd_put(sock);
 196                 return PTR_ERR(node);
 197         }
 198
 199         spin_lock_bh(&m->lock);
 200         old_xs = rcu_dereference_protected(*map_entry, lockdep_is_held(&m->lock));
 201         if (old_xs == xs) {
 202                 err = 0;
 203                 goto out;
 204         } else if (old_xs && map_flags == BPF_NOEXIST) {
 205                 err = -EEXIST;
 206                 goto out;
 207         } else if (!old_xs && map_flags == BPF_EXIST) {
 208                 err = -ENOENT;
 209                 goto out;
 210         }
 211         xsk_map_sock_add(xs, node);
 212         rcu_assign_pointer(*map_entry, xs);
 213         if (old_xs)
 214                 xsk_map_sock_delete(old_xs, map_entry);
 215         spin_unlock_bh(&m->lock);
 216         sockfd_put(sock);
 217         return 0;
 218
 219 out:
 220         spin_unlock_bh(&m->lock);
 221         sockfd_put(sock);
 222         xsk_map_node_free(node);
 223         return err;
 224 }
 225
 226 static long xsk_map_delete_elem(struct bpf_map *map, void *key)
 227 {
 228         struct xsk_map *m = container_of(map, struct xsk_map, map);
 229         struct xdp_sock __rcu **map_entry;
 230         struct xdp_sock *old_xs;
 231         int k = *(u32 *)key;
 232
 233         if (k >= map->max_entries)
 234                 return -EINVAL;
 235
 236         spin_lock_bh(&m->lock);
 237         map_entry = &m->xsk_map[k];
 238         old_xs = unrcu_pointer(xchg(map_entry, NULL));
 239         if (old_xs)
 240                 xsk_map_sock_delete(old_xs, map_entry);
 241         spin_unlock_bh(&m->lock);
 242
 243         return 0;
 244 }
 245
 246 static long xsk_map_redirect(struct bpf_map *map, u64 index, u64 flags)
 247 {
 248         return __bpf_xdp_redirect_map(map, index, flags, 0,
 249                                       __xsk_map_lookup_elem);
 250 }
 251
 252 void xsk_map_try_sock_delete(struct xsk_map *map, struct xdp_sock *xs,
 253                              struct xdp_sock __rcu **map_entry)
 254 {
 255         spin_lock_bh(&map->lock);
 256         if (rcu_access_pointer(*map_entry) == xs) {
 257                 rcu_assign_pointer(*map_entry, NULL);
 258                 xsk_map_sock_delete(xs, map_entry);
 259         }
 260         spin_unlock_bh(&map->lock);
 261 }
 262
 263 static bool xsk_map_meta_equal(const struct bpf_map *meta0,
 264                                const struct bpf_map *meta1)
 265 {
 266         return meta0->max_entries == meta1->max_entries &&
 267                 bpf_map_meta_equal(meta0, meta1);
 268 }
 269
 270 BTF_ID_LIST_SINGLE(xsk_map_btf_ids, struct, xsk_map)
 271 const struct bpf_map_ops xsk_map_ops = {
 272         .map_meta_equal = xsk_map_meta_equal,
 273         .map_alloc = xsk_map_alloc,
 274         .map_free = xsk_map_free,
 275         .map_get_next_key = xsk_map_get_next_key,
 276         .map_lookup_elem = xsk_map_lookup_elem,
 277         .map_gen_lookup = xsk_map_gen_lookup,
 278         .map_lookup_elem_sys_only = xsk_map_lookup_elem_sys_only,
 279         .map_update_elem = xsk_map_update_elem,
 280         .map_delete_elem = xsk_map_delete_elem,
 281         .map_check_btf = map_check_no_btf,
 282         .map_mem_usage = xsk_map_mem_usage,
 283         .map_btf_id = &xsk_map_btf_ids[0],
 284         .map_redirect = xsk_map_redirect,
 285 };