net/ipv4/bpf_tcp_ca.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2019 Facebook  */
   3
   4 #include <linux/init.h>
   5 #include <linux/types.h>
   6 #include <linux/bpf_verifier.h>
   7 #include <linux/bpf.h>
   8 #include <linux/btf.h>
   9 #include <linux/btf_ids.h>
  10 #include <linux/filter.h>
  11 #include <net/tcp.h>
  12 #include <net/bpf_sk_storage.h>
  13
  14 /* "extern" is to avoid sparse warning.  It is only used in bpf_struct_ops.c. */
  15 static struct bpf_struct_ops bpf_tcp_congestion_ops;
  16
  17 static const struct btf_type *tcp_sock_type;
  18 static u32 tcp_sock_id, sock_id;
  19 static const struct btf_type *tcp_congestion_ops_type;
  20
  21 static int bpf_tcp_ca_init(struct btf *btf)
  22 {
  23         s32 type_id;
  24
  25         type_id = btf_find_by_name_kind(btf, "sock", BTF_KIND_STRUCT);
  26         if (type_id < 0)
  27                 return -EINVAL;
  28         sock_id = type_id;
  29
  30         type_id = btf_find_by_name_kind(btf, "tcp_sock", BTF_KIND_STRUCT);
  31         if (type_id < 0)
  32                 return -EINVAL;
  33         tcp_sock_id = type_id;
  34         tcp_sock_type = btf_type_by_id(btf, tcp_sock_id);
  35
  36         type_id = btf_find_by_name_kind(btf, "tcp_congestion_ops", BTF_KIND_STRUCT);
  37         if (type_id < 0)
  38                 return -EINVAL;
  39         tcp_congestion_ops_type = btf_type_by_id(btf, type_id);
  40
  41         return 0;
  42 }
  43
  44 static bool bpf_tcp_ca_is_valid_access(int off, int size,
  45                                        enum bpf_access_type type,
  46                                        const struct bpf_prog *prog,
  47                                        struct bpf_insn_access_aux *info)
  48 {
  49         if (!bpf_tracing_btf_ctx_access(off, size, type, prog, info))
  50                 return false;
  51
  52         if (base_type(info->reg_type) == PTR_TO_BTF_ID &&
  53             !bpf_type_has_unsafe_modifiers(info->reg_type) &&
  54             info->btf_id == sock_id)
  55                 /* promote it to tcp_sock */
  56                 info->btf_id = tcp_sock_id;
  57
  58         return true;
  59 }
  60
  61 static int bpf_tcp_ca_btf_struct_access(struct bpf_verifier_log *log,
  62                                         const struct bpf_reg_state *reg,
  63                                         int off, int size)
  64 {
  65         const struct btf_type *t;
  66         size_t end;
  67
  68         t = btf_type_by_id(reg->btf, reg->btf_id);
  69         if (t != tcp_sock_type) {
  70                 bpf_log(log, "only read is supported\n");
  71                 return -EACCES;
  72         }
  73
  74         switch (off) {
  75         case offsetof(struct sock, sk_pacing_rate):
  76                 end = offsetofend(struct sock, sk_pacing_rate);
  77                 break;
  78         case offsetof(struct sock, sk_pacing_status):
  79                 end = offsetofend(struct sock, sk_pacing_status);
  80                 break;
  81         case bpf_ctx_range(struct inet_connection_sock, icsk_ca_priv):
  82                 end = offsetofend(struct inet_connection_sock, icsk_ca_priv);
  83                 break;
  84         case offsetof(struct inet_connection_sock, icsk_ack.pending):
  85                 end = offsetofend(struct inet_connection_sock,
  86                                   icsk_ack.pending);
  87                 break;
  88         case offsetof(struct tcp_sock, snd_cwnd):
  89                 end = offsetofend(struct tcp_sock, snd_cwnd);
  90                 break;
  91         case offsetof(struct tcp_sock, snd_cwnd_cnt):
  92                 end = offsetofend(struct tcp_sock, snd_cwnd_cnt);
  93                 break;
  94         case offsetof(struct tcp_sock, snd_cwnd_stamp):
  95                 end = offsetofend(struct tcp_sock, snd_cwnd_stamp);
  96                 break;
  97         case offsetof(struct tcp_sock, snd_ssthresh):
  98                 end = offsetofend(struct tcp_sock, snd_ssthresh);
  99                 break;
 100         case offsetof(struct tcp_sock, ecn_flags):
 101                 end = offsetofend(struct tcp_sock, ecn_flags);
 102                 break;
 103         case offsetof(struct tcp_sock, app_limited):
 104                 end = offsetofend(struct tcp_sock, app_limited);
 105                 break;
 106         default:
 107                 bpf_log(log, "no write support to tcp_sock at off %d\n", off);
 108                 return -EACCES;
 109         }
 110
 111         if (off + size > end) {
 112                 bpf_log(log,
 113                         "write access at off %d with size %d beyond the member of tcp_sock ended at %zu\n",
 114                         off, size, end);
 115                 return -EACCES;
 116         }
 117
 118         return 0;
 119 }
 120
 121 BPF_CALL_2(bpf_tcp_send_ack, struct tcp_sock *, tp, u32, rcv_nxt)
 122 {
 123         /* bpf_tcp_ca prog cannot have NULL tp */
 124         __tcp_send_ack((struct sock *)tp, rcv_nxt);
 125         return 0;
 126 }
 127
 128 static const struct bpf_func_proto bpf_tcp_send_ack_proto = {
 129         .func           = bpf_tcp_send_ack,
 130         .gpl_only       = false,
 131         /* In case we want to report error later */
 132         .ret_type       = RET_INTEGER,
 133         .arg1_type      = ARG_PTR_TO_BTF_ID,
 134         .arg1_btf_id    = &tcp_sock_id,
 135         .arg2_type      = ARG_ANYTHING,
 136 };
 137
 138 static u32 prog_ops_moff(const struct bpf_prog *prog)
 139 {
 140         const struct btf_member *m;
 141         const struct btf_type *t;
 142         u32 midx;
 143
 144         midx = prog->expected_attach_type;
 145         t = tcp_congestion_ops_type;
 146         m = &btf_type_member(t)[midx];
 147
 148         return __btf_member_bit_offset(t, m) / 8;
 149 }
 150
 151 static const struct bpf_func_proto *
 152 bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id,
 153                           const struct bpf_prog *prog)
 154 {
 155         switch (func_id) {
 156         case BPF_FUNC_tcp_send_ack:
 157                 return &bpf_tcp_send_ack_proto;
 158         case BPF_FUNC_sk_storage_get:
 159                 return &bpf_sk_storage_get_proto;
 160         case BPF_FUNC_sk_storage_delete:
 161                 return &bpf_sk_storage_delete_proto;
 162         case BPF_FUNC_setsockopt:
 163                 /* Does not allow release() to call setsockopt.
 164                  * release() is called when the current bpf-tcp-cc
 165                  * is retiring.  It is not allowed to call
 166                  * setsockopt() to make further changes which
 167                  * may potentially allocate new resources.
 168                  */
 169                 if (prog_ops_moff(prog) !=
 170                     offsetof(struct tcp_congestion_ops, release))
 171                         return &bpf_sk_setsockopt_proto;
 172                 return NULL;
 173         case BPF_FUNC_getsockopt:
 174                 /* Since get/setsockopt is usually expected to
 175                  * be available together, disable getsockopt for
 176                  * release also to avoid usage surprise.
 177                  * The bpf-tcp-cc already has a more powerful way
 178                  * to read tcp_sock from the PTR_TO_BTF_ID.
 179                  */
 180                 if (prog_ops_moff(prog) !=
 181                     offsetof(struct tcp_congestion_ops, release))
 182                         return &bpf_sk_getsockopt_proto;
 183                 return NULL;
 184         case BPF_FUNC_ktime_get_coarse_ns:
 185                 return &bpf_ktime_get_coarse_ns_proto;
 186         default:
 187                 return bpf_base_func_proto(func_id, prog);
 188         }
 189 }
 190
 191 BTF_KFUNCS_START(bpf_tcp_ca_check_kfunc_ids)
 192 BTF_ID_FLAGS(func, tcp_reno_ssthresh)
 193 BTF_ID_FLAGS(func, tcp_reno_cong_avoid)
 194 BTF_ID_FLAGS(func, tcp_reno_undo_cwnd)
 195 BTF_ID_FLAGS(func, tcp_slow_start)
 196 BTF_ID_FLAGS(func, tcp_cong_avoid_ai)
 197 BTF_KFUNCS_END(bpf_tcp_ca_check_kfunc_ids)
 198
 199 static const struct btf_kfunc_id_set bpf_tcp_ca_kfunc_set = {
 200         .owner = THIS_MODULE,
 201         .set   = &bpf_tcp_ca_check_kfunc_ids,
 202 };
 203
 204 static const struct bpf_verifier_ops bpf_tcp_ca_verifier_ops = {
 205         .get_func_proto         = bpf_tcp_ca_get_func_proto,
 206         .is_valid_access        = bpf_tcp_ca_is_valid_access,
 207         .btf_struct_access      = bpf_tcp_ca_btf_struct_access,
 208 };
 209
 210 static int bpf_tcp_ca_init_member(const struct btf_type *t,
 211                                   const struct btf_member *member,
 212                                   void *kdata, const void *udata)
 213 {
 214         const struct tcp_congestion_ops *utcp_ca;
 215         struct tcp_congestion_ops *tcp_ca;
 216         u32 moff;
 217
 218         utcp_ca = (const struct tcp_congestion_ops *)udata;
 219         tcp_ca = (struct tcp_congestion_ops *)kdata;
 220
 221         moff = __btf_member_bit_offset(t, member) / 8;
 222         switch (moff) {
 223         case offsetof(struct tcp_congestion_ops, flags):
 224                 if (utcp_ca->flags & ~TCP_CONG_MASK)
 225                         return -EINVAL;
 226                 tcp_ca->flags = utcp_ca->flags;
 227                 return 1;
 228         case offsetof(struct tcp_congestion_ops, name):
 229                 if (bpf_obj_name_cpy(tcp_ca->name, utcp_ca->name,
 230                                      sizeof(tcp_ca->name)) <= 0)
 231                         return -EINVAL;
 232                 return 1;
 233         }
 234
 235         return 0;
 236 }
 237
 238 static int bpf_tcp_ca_reg(void *kdata, struct bpf_link *link)
 239 {
 240         return tcp_register_congestion_control(kdata);
 241 }
 242
 243 static void bpf_tcp_ca_unreg(void *kdata, struct bpf_link *link)
 244 {
 245         tcp_unregister_congestion_control(kdata);
 246 }
 247
 248 static int bpf_tcp_ca_update(void *kdata, void *old_kdata, struct bpf_link *link)
 249 {
 250         return tcp_update_congestion_control(kdata, old_kdata);
 251 }
 252
 253 static int bpf_tcp_ca_validate(void *kdata)
 254 {
 255         return tcp_validate_congestion_control(kdata);
 256 }
 257
 258 static u32 bpf_tcp_ca_ssthresh(struct sock *sk)
 259 {
 260         return 0;
 261 }
 262
 263 static void bpf_tcp_ca_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 264 {
 265 }
 266
 267 static void bpf_tcp_ca_set_state(struct sock *sk, u8 new_state)
 268 {
 269 }
 270
 271 static void bpf_tcp_ca_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
 272 {
 273 }
 274
 275 static void bpf_tcp_ca_in_ack_event(struct sock *sk, u32 flags)
 276 {
 277 }
 278
 279 static void bpf_tcp_ca_pkts_acked(struct sock *sk, const struct ack_sample *sample)
 280 {
 281 }
 282
 283 static u32 bpf_tcp_ca_min_tso_segs(struct sock *sk)
 284 {
 285         return 0;
 286 }
 287
 288 static void bpf_tcp_ca_cong_control(struct sock *sk, u32 ack, int flag,
 289                                     const struct rate_sample *rs)
 290 {
 291 }
 292
 293 static u32 bpf_tcp_ca_undo_cwnd(struct sock *sk)
 294 {
 295         return 0;
 296 }
 297
 298 static u32 bpf_tcp_ca_sndbuf_expand(struct sock *sk)
 299 {
 300         return 0;
 301 }
 302
 303 static void __bpf_tcp_ca_init(struct sock *sk)
 304 {
 305 }
 306
 307 static void __bpf_tcp_ca_release(struct sock *sk)
 308 {
 309 }
 310
 311 static struct tcp_congestion_ops __bpf_ops_tcp_congestion_ops = {
 312         .ssthresh = bpf_tcp_ca_ssthresh,
 313         .cong_avoid = bpf_tcp_ca_cong_avoid,
 314         .set_state = bpf_tcp_ca_set_state,
 315         .cwnd_event = bpf_tcp_ca_cwnd_event,
 316         .in_ack_event = bpf_tcp_ca_in_ack_event,
 317         .pkts_acked = bpf_tcp_ca_pkts_acked,
 318         .min_tso_segs = bpf_tcp_ca_min_tso_segs,
 319         .cong_control = bpf_tcp_ca_cong_control,
 320         .undo_cwnd = bpf_tcp_ca_undo_cwnd,
 321         .sndbuf_expand = bpf_tcp_ca_sndbuf_expand,
 322
 323         .init = __bpf_tcp_ca_init,
 324         .release = __bpf_tcp_ca_release,
 325 };
 326
 327 static struct bpf_struct_ops bpf_tcp_congestion_ops = {
 328         .verifier_ops = &bpf_tcp_ca_verifier_ops,
 329         .reg = bpf_tcp_ca_reg,
 330         .unreg = bpf_tcp_ca_unreg,
 331         .update = bpf_tcp_ca_update,
 332         .init_member = bpf_tcp_ca_init_member,
 333         .init = bpf_tcp_ca_init,
 334         .validate = bpf_tcp_ca_validate,
 335         .name = "tcp_congestion_ops",
 336         .cfi_stubs = &__bpf_ops_tcp_congestion_ops,
 337         .owner = THIS_MODULE,
 338 };
 339
 340 static int __init bpf_tcp_ca_kfunc_init(void)
 341 {
 342         int ret;
 343
 344         ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &bpf_tcp_ca_kfunc_set);
 345         ret = ret ?: register_bpf_struct_ops(&bpf_tcp_congestion_ops, tcp_congestion_ops);
 346
 347         return ret;
 348 }
 349 late_initcall(bpf_tcp_ca_kfunc_init);