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/filter.h>
   9 #include <net/tcp.h>
  10
  11 static u32 optional_ops[] = {
  12         offsetof(struct tcp_congestion_ops, init),
  13         offsetof(struct tcp_congestion_ops, release),
  14         offsetof(struct tcp_congestion_ops, set_state),
  15         offsetof(struct tcp_congestion_ops, cwnd_event),
  16         offsetof(struct tcp_congestion_ops, in_ack_event),
  17         offsetof(struct tcp_congestion_ops, pkts_acked),
  18         offsetof(struct tcp_congestion_ops, min_tso_segs),
  19         offsetof(struct tcp_congestion_ops, sndbuf_expand),
  20         offsetof(struct tcp_congestion_ops, cong_control),
  21 };
  22
  23 static u32 unsupported_ops[] = {
  24         offsetof(struct tcp_congestion_ops, get_info),
  25 };
  26
  27 static const struct btf_type *tcp_sock_type;
  28 static u32 tcp_sock_id, sock_id;
  29
  30 static int bpf_tcp_ca_init(struct btf *btf)
  31 {
  32         s32 type_id;
  33
  34         type_id = btf_find_by_name_kind(btf, "sock", BTF_KIND_STRUCT);
  35         if (type_id < 0)
  36                 return -EINVAL;
  37         sock_id = type_id;
  38
  39         type_id = btf_find_by_name_kind(btf, "tcp_sock", BTF_KIND_STRUCT);
  40         if (type_id < 0)
  41                 return -EINVAL;
  42         tcp_sock_id = type_id;
  43         tcp_sock_type = btf_type_by_id(btf, tcp_sock_id);
  44
  45         return 0;
  46 }
  47
  48 static bool is_optional(u32 member_offset)
  49 {
  50         unsigned int i;
  51
  52         for (i = 0; i < ARRAY_SIZE(optional_ops); i++) {
  53                 if (member_offset == optional_ops[i])
  54                         return true;
  55         }
  56
  57         return false;
  58 }
  59
  60 static bool is_unsupported(u32 member_offset)
  61 {
  62         unsigned int i;
  63
  64         for (i = 0; i < ARRAY_SIZE(unsupported_ops); i++) {
  65                 if (member_offset == unsupported_ops[i])
  66                         return true;
  67         }
  68
  69         return false;
  70 }
  71
  72 extern struct btf *btf_vmlinux;
  73
  74 static bool bpf_tcp_ca_is_valid_access(int off, int size,
  75                                        enum bpf_access_type type,
  76                                        const struct bpf_prog *prog,
  77                                        struct bpf_insn_access_aux *info)
  78 {
  79         if (off < 0 || off >= sizeof(__u64) * MAX_BPF_FUNC_ARGS)
  80                 return false;
  81         if (type != BPF_READ)
  82                 return false;
  83         if (off % size != 0)
  84                 return false;
  85
  86         if (!btf_ctx_access(off, size, type, prog, info))
  87                 return false;
  88
  89         if (info->reg_type == PTR_TO_BTF_ID && info->btf_id == sock_id)
  90                 /* promote it to tcp_sock */
  91                 info->btf_id = tcp_sock_id;
  92
  93         return true;
  94 }
  95
  96 static int bpf_tcp_ca_btf_struct_access(struct bpf_verifier_log *log,
  97                                         const struct btf_type *t, int off,
  98                                         int size, enum bpf_access_type atype,
  99                                         u32 *next_btf_id)
 100 {
 101         size_t end;
 102
 103         if (atype == BPF_READ)
 104                 return btf_struct_access(log, t, off, size, atype, next_btf_id);
 105
 106         if (t != tcp_sock_type) {
 107                 bpf_log(log, "only read is supported\n");
 108                 return -EACCES;
 109         }
 110
 111         switch (off) {
 112         case bpf_ctx_range(struct inet_connection_sock, icsk_ca_priv):
 113                 end = offsetofend(struct inet_connection_sock, icsk_ca_priv);
 114                 break;
 115         case offsetof(struct inet_connection_sock, icsk_ack.pending):
 116                 end = offsetofend(struct inet_connection_sock,
 117                                   icsk_ack.pending);
 118                 break;
 119         case offsetof(struct tcp_sock, snd_cwnd):
 120                 end = offsetofend(struct tcp_sock, snd_cwnd);
 121                 break;
 122         case offsetof(struct tcp_sock, snd_cwnd_cnt):
 123                 end = offsetofend(struct tcp_sock, snd_cwnd_cnt);
 124                 break;
 125         case offsetof(struct tcp_sock, snd_ssthresh):
 126                 end = offsetofend(struct tcp_sock, snd_ssthresh);
 127                 break;
 128         case offsetof(struct tcp_sock, ecn_flags):
 129                 end = offsetofend(struct tcp_sock, ecn_flags);
 130                 break;
 131         default:
 132                 bpf_log(log, "no write support to tcp_sock at off %d\n", off);
 133                 return -EACCES;
 134         }
 135
 136         if (off + size > end) {
 137                 bpf_log(log,
 138                         "write access at off %d with size %d beyond the member of tcp_sock ended at %zu\n",
 139                         off, size, end);
 140                 return -EACCES;
 141         }
 142
 143         return NOT_INIT;
 144 }
 145
 146 BPF_CALL_2(bpf_tcp_send_ack, struct tcp_sock *, tp, u32, rcv_nxt)
 147 {
 148         /* bpf_tcp_ca prog cannot have NULL tp */
 149         __tcp_send_ack((struct sock *)tp, rcv_nxt);
 150         return 0;
 151 }
 152
 153 static const struct bpf_func_proto bpf_tcp_send_ack_proto = {
 154         .func           = bpf_tcp_send_ack,
 155         .gpl_only       = false,
 156         /* In case we want to report error later */
 157         .ret_type       = RET_INTEGER,
 158         .arg1_type      = ARG_PTR_TO_BTF_ID,
 159         .arg2_type      = ARG_ANYTHING,
 160         .btf_id         = &tcp_sock_id,
 161 };
 162
 163 static const struct bpf_func_proto *
 164 bpf_tcp_ca_get_func_proto(enum bpf_func_id func_id,
 165                           const struct bpf_prog *prog)
 166 {
 167         switch (func_id) {
 168         case BPF_FUNC_tcp_send_ack:
 169                 return &bpf_tcp_send_ack_proto;
 170         default:
 171                 return bpf_base_func_proto(func_id);
 172         }
 173 }
 174
 175 static const struct bpf_verifier_ops bpf_tcp_ca_verifier_ops = {
 176         .get_func_proto         = bpf_tcp_ca_get_func_proto,
 177         .is_valid_access        = bpf_tcp_ca_is_valid_access,
 178         .btf_struct_access      = bpf_tcp_ca_btf_struct_access,
 179 };
 180
 181 static int bpf_tcp_ca_init_member(const struct btf_type *t,
 182                                   const struct btf_member *member,
 183                                   void *kdata, const void *udata)
 184 {
 185         const struct tcp_congestion_ops *utcp_ca;
 186         struct tcp_congestion_ops *tcp_ca;
 187         size_t tcp_ca_name_len;
 188         int prog_fd;
 189         u32 moff;
 190
 191         utcp_ca = (const struct tcp_congestion_ops *)udata;
 192         tcp_ca = (struct tcp_congestion_ops *)kdata;
 193
 194         moff = btf_member_bit_offset(t, member) / 8;
 195         switch (moff) {
 196         case offsetof(struct tcp_congestion_ops, flags):
 197                 if (utcp_ca->flags & ~TCP_CONG_MASK)
 198                         return -EINVAL;
 199                 tcp_ca->flags = utcp_ca->flags;
 200                 return 1;
 201         case offsetof(struct tcp_congestion_ops, name):
 202                 tcp_ca_name_len = strnlen(utcp_ca->name, sizeof(utcp_ca->name));
 203                 if (!tcp_ca_name_len ||
 204                     tcp_ca_name_len == sizeof(utcp_ca->name))
 205                         return -EINVAL;
 206                 if (tcp_ca_find(utcp_ca->name))
 207                         return -EEXIST;
 208                 memcpy(tcp_ca->name, utcp_ca->name, sizeof(tcp_ca->name));
 209                 return 1;
 210         }
 211
 212         if (!btf_type_resolve_func_ptr(btf_vmlinux, member->type, NULL))
 213                 return 0;
 214
 215         /* Ensure bpf_prog is provided for compulsory func ptr */
 216         prog_fd = (int)(*(unsigned long *)(udata + moff));
 217         if (!prog_fd && !is_optional(moff) && !is_unsupported(moff))
 218                 return -EINVAL;
 219
 220         return 0;
 221 }
 222
 223 static int bpf_tcp_ca_check_member(const struct btf_type *t,
 224                                    const struct btf_member *member)
 225 {
 226         if (is_unsupported(btf_member_bit_offset(t, member) / 8))
 227                 return -ENOTSUPP;
 228         return 0;
 229 }
 230
 231 static int bpf_tcp_ca_reg(void *kdata)
 232 {
 233         return tcp_register_congestion_control(kdata);
 234 }
 235
 236 static void bpf_tcp_ca_unreg(void *kdata)
 237 {
 238         tcp_unregister_congestion_control(kdata);
 239 }
 240
 241 /* Avoid sparse warning.  It is only used in bpf_struct_ops.c. */
 242 extern struct bpf_struct_ops bpf_tcp_congestion_ops;
 243
 244 struct bpf_struct_ops bpf_tcp_congestion_ops = {
 245         .verifier_ops = &bpf_tcp_ca_verifier_ops,
 246         .reg = bpf_tcp_ca_reg,
 247         .unreg = bpf_tcp_ca_unreg,
 248         .check_member = bpf_tcp_ca_check_member,
 249         .init_member = bpf_tcp_ca_init_member,
 250         .init = bpf_tcp_ca_init,
 251         .name = "tcp_congestion_ops",
 252 };