net/ipv4/bpf_tcp_ca.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2019 Facebook  */
   3
   4 #include <linux/types.h>
   5 #include <linux/bpf_verifier.h>
   6 #include <linux/bpf.h>
   7 #include <linux/btf.h>
   8 #include <linux/btf_ids.h>
   9 #include <linux/filter.h>
  10 #include <net/tcp.h>
  11 #include <net/bpf_sk_storage.h>
  12
  13 /* "extern" is to avoid sparse warning.  It is only used in bpf_struct_ops.c. */
  14 extern struct bpf_struct_ops bpf_tcp_congestion_ops;
  15
  16 static u32 optional_ops[] = {
  17         offsetof(struct tcp_congestion_ops, init),
  18         offsetof(struct tcp_congestion_ops, release),
  19         offsetof(struct tcp_congestion_ops, set_state),
  20         offsetof(struct tcp_congestion_ops, cwnd_event),
  21         offsetof(struct tcp_congestion_ops, in_ack_event),
  22         offsetof(struct tcp_congestion_ops, pkts_acked),
  23         offsetof(struct tcp_congestion_ops, min_tso_segs),
  24         offsetof(struct tcp_congestion_ops, sndbuf_expand),
  25         offsetof(struct tcp_congestion_ops, cong_control),
  26 };
  27
  28 static u32 unsupported_ops[] = {
  29         offsetof(struct tcp_congestion_ops, get_info),
  30 };
  31
  32 static const struct btf_type *tcp_sock_type;
  33 static u32 tcp_sock_id, sock_id;
  34
  35 static int bpf_tcp_ca_init(struct btf *btf)
  36 {
  37         s32 type_id;
  38
  39         type_id = btf_find_by_name_kind(btf, "sock", BTF_KIND_STRUCT);
  40         if (type_id < 0)
  41                 return -EINVAL;
  42         sock_id = type_id;
  43
  44         type_id = btf_find_by_name_kind(btf, "tcp_sock", BTF_KIND_STRUCT);
  45         if (type_id < 0)
  46                 return -EINVAL;
  47         tcp_sock_id = type_id;
  48         tcp_sock_type = btf_type_by_id(btf, tcp_sock_id);
  49
  50         return 0;
  51 }
  52
  53 static bool is_optional(u32 member_offset)
  54 {
  55         unsigned int i;
  56
  57         for (i = 0; i < ARRAY_SIZE(optional_ops); i++) {
  58                 if (member_offset == optional_ops[i])
  59                         return true;
  60         }
  61
  62         return false;
  63 }
  64
  65 static bool is_unsupported(u32 member_offset)
  66 {
  67         unsigned int i;
  68
  69         for (i = 0; i < ARRAY_SIZE(unsupported_ops); i++) {
  70                 if (member_offset == unsupported_ops[i])
  71                         return true;
  72         }
  73
  74         return false;
  75 }
  76
  77 extern struct btf *btf_vmlinux;
  78
  79 static bool bpf_tcp_ca_is_valid_access(int off, int size,
  80                                        enum bpf_access_type type,
  81                                        const struct bpf_prog *prog,
  82                                        struct bpf_insn_access_aux *info)
  83 {
  84         if (!bpf_tracing_btf_ctx_access(off, size, type, prog, info))
  85                 return false;
  86
  87         if (info->reg_type == PTR_TO_BTF_ID && info->btf_id == sock_id)
  88                 /* promote it to tcp_sock */
  89                 info->btf_id = tcp_sock_id;
  90
  91         return true;
  92 }
  93
  94 static int bpf_tcp_ca_btf_struct_access(struct bpf_verifier_log *log,
  95                                         const struct btf *btf,
  96                                         const struct btf_type *t, int off,
  97                                         int size, enum bpf_access_type atype,
  98                                         u32 *next_btf_id)
  99 {
 100         size_t end;
 101
 102         if (atype == BPF_READ)
 103                 return btf_struct_access(log, btf, t, off, size, atype, next_btf_id);
 104
 105         if (t != tcp_sock_type) {
 106                 bpf_log(log, "only read is supported\n");
 107                 return -EACCES;
 108         }
 109
 110         switch (off) {
 111         case bpf_ctx_range(struct inet_connection_sock, icsk_ca_priv):
 112                 end = offsetofend(struct inet_connection_sock, icsk_ca_priv);
 113                 break;
 114         case offsetof(struct inet_connection_sock, icsk_ack.pending):
 115                 end = offsetofend(struct inet_connection_sock,
 116                                   icsk_ack.pending);
 117                 break;
 118         case offsetof(struct tcp_sock, snd_cwnd):
 119                 end = offsetofend(struct tcp_sock, snd_cwnd);
 120                 break;
 121         case offsetof(struct tcp_sock, snd_cwnd_cnt):
 122                 end = offsetofend(struct tcp_sock, snd_cwnd_cnt);
 123                 break;
 124         case offsetof(struct tcp_sock, snd_ssthresh):
 125                 end = offsetofend(struct tcp_sock, snd_ssthresh);
 126                 break;
 127         case offsetof(struct tcp_sock, ecn_flags):
 128                 end = offsetofend(struct tcp_sock, ecn_flags);
 129                 break;
 130         default:
 131                 bpf_log(log, "no write support to tcp_sock at off %d\n", off);
 132                 return -EACCES;
 133         }
 134
 135         if (off + size > end) {
 136                 bpf_log(log,
 137                         "write access at off %d with size %d beyond the member of tcp_sock ended at %zu\n",
 138                         off, size, end);
 139                 return -EACCES;
 140         }
 141
 142         return NOT_INIT;
 143 }
 144
 145 BPF_CALL_2(bpf_tcp_send_ack, struct tcp_sock *, tp, u32, rcv_nxt)
 146 {
 147         /* bpf_tcp_ca prog cannot have NULL tp */
 148         __tcp_send_ack((struct sock *)tp, rcv_nxt);
 149         return 0;
 150 }
 151
 152 static const struct bpf_func_proto bpf_tcp_send_ack_proto = {
 153         .func           = bpf_tcp_send_ack,
 154         .gpl_only       = false,
 155         /* In case we want to report error later */
 156         .ret_type       = RET_INTEGER,
 157         .arg1_type      = ARG_PTR_TO_BTF_ID,
 158         .arg1_btf_id    = &tcp_sock_id,
 159         .arg2_type      = ARG_ANYTHING,
 160 };
 161
 162 static u32 prog_ops_moff(const struct bpf_prog *prog)
 163 {
 164         const struct btf_member *m;
 165         const struct btf_type *t;
 166         u32 midx;
 167
 168         midx = prog->expected_attach_type;
 169         t = bpf_tcp_congestion_ops.type;
 170         m = &btf_type_member(t)[midx];
 171
 172         return __btf_member_bit_offset(t, m) / 8;
 173 }
 174
 175 static const struct bpf_func_proto *
 176 bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id,
 177                           const struct bpf_prog *prog)
 178 {
 179         switch (func_id) {
 180         case BPF_FUNC_tcp_send_ack:
 181                 return &bpf_tcp_send_ack_proto;
 182         case BPF_FUNC_sk_storage_get:
 183                 return &bpf_sk_storage_get_proto;
 184         case BPF_FUNC_sk_storage_delete:
 185                 return &bpf_sk_storage_delete_proto;
 186         case BPF_FUNC_setsockopt:
 187                 /* Does not allow release() to call setsockopt.
 188                  * release() is called when the current bpf-tcp-cc
 189                  * is retiring.  It is not allowed to call
 190                  * setsockopt() to make further changes which
 191                  * may potentially allocate new resources.
 192                  */
 193                 if (prog_ops_moff(prog) !=
 194                     offsetof(struct tcp_congestion_ops, release))
 195                         return &bpf_sk_setsockopt_proto;
 196                 return NULL;
 197         case BPF_FUNC_getsockopt:
 198                 /* Since get/setsockopt is usually expected to
 199                  * be available together, disable getsockopt for
 200                  * release also to avoid usage surprise.
 201                  * The bpf-tcp-cc already has a more powerful way
 202                  * to read tcp_sock from the PTR_TO_BTF_ID.
 203                  */
 204                 if (prog_ops_moff(prog) !=
 205                     offsetof(struct tcp_congestion_ops, release))
 206                         return &bpf_sk_getsockopt_proto;
 207                 return NULL;
 208         case BPF_FUNC_ktime_get_coarse_ns:
 209                 return &bpf_ktime_get_coarse_ns_proto;
 210         default:
 211                 return bpf_base_func_proto(func_id);
 212         }
 213 }
 214
 215 BTF_SET_START(bpf_tcp_ca_kfunc_ids)
 216 BTF_ID(func, tcp_reno_ssthresh)
 217 BTF_ID(func, tcp_reno_cong_avoid)
 218 BTF_ID(func, tcp_reno_undo_cwnd)
 219 BTF_ID(func, tcp_slow_start)
 220 BTF_ID(func, tcp_cong_avoid_ai)
 221 BTF_SET_END(bpf_tcp_ca_kfunc_ids)
 222
 223 static bool bpf_tcp_ca_check_kfunc_call(u32 kfunc_btf_id, struct module *owner)
 224 {
 225         if (btf_id_set_contains(&bpf_tcp_ca_kfunc_ids, kfunc_btf_id))
 226                 return true;
 227         return bpf_check_mod_kfunc_call(&bpf_tcp_ca_kfunc_list, kfunc_btf_id, owner);
 228 }
 229
 230 static const struct bpf_verifier_ops bpf_tcp_ca_verifier_ops = {
 231         .get_func_proto         = bpf_tcp_ca_get_func_proto,
 232         .is_valid_access        = bpf_tcp_ca_is_valid_access,
 233         .btf_struct_access      = bpf_tcp_ca_btf_struct_access,
 234         .check_kfunc_call       = bpf_tcp_ca_check_kfunc_call,
 235 };
 236
 237 static int bpf_tcp_ca_init_member(const struct btf_type *t,
 238                                   const struct btf_member *member,
 239                                   void *kdata, const void *udata)
 240 {
 241         const struct tcp_congestion_ops *utcp_ca;
 242         struct tcp_congestion_ops *tcp_ca;
 243         int prog_fd;
 244         u32 moff;
 245
 246         utcp_ca = (const struct tcp_congestion_ops *)udata;
 247         tcp_ca = (struct tcp_congestion_ops *)kdata;
 248
 249         moff = __btf_member_bit_offset(t, member) / 8;
 250         switch (moff) {
 251         case offsetof(struct tcp_congestion_ops, flags):
 252                 if (utcp_ca->flags & ~TCP_CONG_MASK)
 253                         return -EINVAL;
 254                 tcp_ca->flags = utcp_ca->flags;
 255                 return 1;
 256         case offsetof(struct tcp_congestion_ops, name):
 257                 if (bpf_obj_name_cpy(tcp_ca->name, utcp_ca->name,
 258                                      sizeof(tcp_ca->name)) <= 0)
 259                         return -EINVAL;
 260                 if (tcp_ca_find(utcp_ca->name))
 261                         return -EEXIST;
 262                 return 1;
 263         }
 264
 265         if (!btf_type_resolve_func_ptr(btf_vmlinux, member->type, NULL))
 266                 return 0;
 267
 268         /* Ensure bpf_prog is provided for compulsory func ptr */
 269         prog_fd = (int)(*(unsigned long *)(udata + moff));
 270         if (!prog_fd && !is_optional(moff) && !is_unsupported(moff))
 271                 return -EINVAL;
 272
 273         return 0;
 274 }
 275
 276 static int bpf_tcp_ca_check_member(const struct btf_type *t,
 277                                    const struct btf_member *member)
 278 {
 279         if (is_unsupported(__btf_member_bit_offset(t, member) / 8))
 280                 return -ENOTSUPP;
 281         return 0;
 282 }
 283
 284 static int bpf_tcp_ca_reg(void *kdata)
 285 {
 286         return tcp_register_congestion_control(kdata);
 287 }
 288
 289 static void bpf_tcp_ca_unreg(void *kdata)
 290 {
 291         tcp_unregister_congestion_control(kdata);
 292 }
 293
 294 struct bpf_struct_ops bpf_tcp_congestion_ops = {
 295         .verifier_ops = &bpf_tcp_ca_verifier_ops,
 296         .reg = bpf_tcp_ca_reg,
 297         .unreg = bpf_tcp_ca_unreg,
 298         .check_member = bpf_tcp_ca_check_member,
 299         .init_member = bpf_tcp_ca_init_member,
 300         .init = bpf_tcp_ca_init,
 301         .name = "tcp_congestion_ops",
 302 };