| 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
| 2 | /* |
| 3 | * ipv4 in net namespaces |
| 4 | */ |
| 5 | |
| 6 | #ifndef __NETNS_IPV4_H__ |
| 7 | #define __NETNS_IPV4_H__ |
| 8 | |
| 9 | #include <linux/uidgid.h> |
| 10 | #include <net/inet_frag.h> |
| 11 | #include <linux/rcupdate.h> |
| 12 | #include <linux/seqlock.h> |
| 13 | #include <linux/siphash.h> |
| 14 | |
| 15 | struct ; |
| 16 | struct ipv4_devconf; |
| 17 | struct fib_rules_ops; |
| 18 | struct hlist_head; |
| 19 | struct fib_table; |
| 20 | struct sock; |
| 21 | struct local_ports { |
| 22 | u32 range; /* high << 16 | low */ |
| 23 | bool warned; |
| 24 | }; |
| 25 | |
| 26 | struct ping_group_range { |
| 27 | seqlock_t lock; |
| 28 | kgid_t range[2]; |
| 29 | }; |
| 30 | |
| 31 | struct inet_hashinfo; |
| 32 | |
| 33 | struct inet_timewait_death_row { |
| 34 | refcount_t tw_refcount; |
| 35 | |
| 36 | /* Padding to avoid false sharing, tw_refcount can be often written */ |
| 37 | struct inet_hashinfo *hashinfo ____cacheline_aligned_in_smp; |
| 38 | int sysctl_max_tw_buckets; |
| 39 | }; |
| 40 | |
| 41 | struct tcp_fastopen_context; |
| 42 | |
| 43 | #ifdef CONFIG_IP_ROUTE_MULTIPATH |
| 44 | struct sysctl_fib_multipath_hash_seed { |
| 45 | u32 user_seed; |
| 46 | u32 mp_seed; |
| 47 | }; |
| 48 | #endif |
| 49 | |
| 50 | struct udp_tunnel_gro { |
| 51 | struct sock __rcu *sk; |
| 52 | struct hlist_head list; |
| 53 | }; |
| 54 | |
| 55 | struct netns_ipv4 { |
| 56 | /* Cacheline organization can be found documented in |
| 57 | * Documentation/networking/net_cachelines/netns_ipv4_sysctl.rst. |
| 58 | * Please update the document when adding new fields. |
| 59 | */ |
| 60 | |
| 61 | /* TX readonly hotpath cache lines */ |
| 62 | __cacheline_group_begin(netns_ipv4_read_tx); |
| 63 | u8 sysctl_tcp_early_retrans; |
| 64 | u8 sysctl_tcp_tso_win_divisor; |
| 65 | u8 sysctl_tcp_tso_rtt_log; |
| 66 | u8 sysctl_tcp_autocorking; |
| 67 | int sysctl_tcp_min_snd_mss; |
| 68 | unsigned int sysctl_tcp_notsent_lowat; |
| 69 | int sysctl_tcp_limit_output_bytes; |
| 70 | int sysctl_tcp_min_rtt_wlen; |
| 71 | int sysctl_tcp_wmem[3]; |
| 72 | u8 sysctl_ip_fwd_use_pmtu; |
| 73 | __cacheline_group_end(netns_ipv4_read_tx); |
| 74 | |
| 75 | /* TXRX readonly hotpath cache lines */ |
| 76 | __cacheline_group_begin(netns_ipv4_read_txrx); |
| 77 | __cacheline_group_end(netns_ipv4_read_txrx); |
| 78 | |
| 79 | /* RX readonly hotpath cache line */ |
| 80 | __cacheline_group_begin(netns_ipv4_read_rx); |
| 81 | u8 sysctl_tcp_moderate_rcvbuf; |
| 82 | u8 sysctl_ip_early_demux; |
| 83 | u8 sysctl_tcp_early_demux; |
| 84 | u8 sysctl_tcp_l3mdev_accept; |
| 85 | /* 3 bytes hole, try to pack */ |
| 86 | int sysctl_tcp_reordering; |
| 87 | int sysctl_tcp_rmem[3]; |
| 88 | int sysctl_tcp_rcvbuf_low_rtt; |
| 89 | __cacheline_group_end(netns_ipv4_read_rx); |
| 90 | |
| 91 | struct inet_timewait_death_row tcp_death_row; |
| 92 | struct udp_table *udp_table; |
| 93 | |
| 94 | #if IS_ENABLED(CONFIG_NET_UDP_TUNNEL) |
| 95 | /* Not in a pernet subsys because need to be available at GRO stage */ |
| 96 | struct udp_tunnel_gro udp_tunnel_gro[2]; |
| 97 | #endif |
| 98 | |
| 99 | #ifdef CONFIG_SYSCTL |
| 100 | struct ctl_table_header *forw_hdr; |
| 101 | struct ctl_table_header *frags_hdr; |
| 102 | struct ctl_table_header *ipv4_hdr; |
| 103 | struct ctl_table_header *route_hdr; |
| 104 | struct ctl_table_header *xfrm4_hdr; |
| 105 | #endif |
| 106 | struct ipv4_devconf *devconf_all; |
| 107 | struct ipv4_devconf *devconf_dflt; |
| 108 | struct ip_ra_chain __rcu *ra_chain; |
| 109 | struct mutex ra_mutex; |
| 110 | #ifdef CONFIG_IP_MULTIPLE_TABLES |
| 111 | struct fib_rules_ops *rules_ops; |
| 112 | struct fib_table __rcu *fib_main; |
| 113 | struct fib_table __rcu *fib_default; |
| 114 | unsigned int fib_rules_require_fldissect; |
| 115 | bool fib_has_custom_rules; |
| 116 | #endif |
| 117 | bool fib_has_custom_local_routes; |
| 118 | bool fib_offload_disabled; |
| 119 | u8 sysctl_tcp_shrink_window; |
| 120 | #ifdef CONFIG_IP_ROUTE_CLASSID |
| 121 | atomic_t fib_num_tclassid_users; |
| 122 | #endif |
| 123 | struct hlist_head *fib_table_hash; |
| 124 | struct sock *fibnl; |
| 125 | struct hlist_head *fib_info_hash; |
| 126 | unsigned int fib_info_hash_bits; |
| 127 | unsigned int fib_info_cnt; |
| 128 | |
| 129 | struct sock *mc_autojoin_sk; |
| 130 | |
| 131 | struct inet_peer_base *peers; |
| 132 | struct fqdir *fqdir; |
| 133 | |
| 134 | u8 sysctl_icmp_echo_ignore_all; |
| 135 | u8 sysctl_icmp_echo_enable_probe; |
| 136 | u8 sysctl_icmp_echo_ignore_broadcasts; |
| 137 | u8 sysctl_icmp_ignore_bogus_error_responses; |
| 138 | u8 sysctl_icmp_errors_use_inbound_ifaddr; |
| 139 | u8 sysctl_icmp_errors_extension_mask; |
| 140 | int sysctl_icmp_ratelimit; |
| 141 | int sysctl_icmp_ratemask; |
| 142 | int sysctl_icmp_msgs_per_sec; |
| 143 | int sysctl_icmp_msgs_burst; |
| 144 | atomic_t icmp_global_credit; |
| 145 | u32 icmp_global_stamp; |
| 146 | u32 ip_rt_min_pmtu; |
| 147 | int ip_rt_mtu_expires; |
| 148 | int ip_rt_min_advmss; |
| 149 | |
| 150 | struct local_ports ip_local_ports; |
| 151 | |
| 152 | u8 sysctl_tcp_ecn; |
| 153 | u8 sysctl_tcp_ecn_option; |
| 154 | u8 sysctl_tcp_ecn_option_beacon; |
| 155 | u8 sysctl_tcp_ecn_fallback; |
| 156 | |
| 157 | u8 sysctl_ip_default_ttl; |
| 158 | u8 sysctl_ip_no_pmtu_disc; |
| 159 | u8 sysctl_ip_fwd_update_priority; |
| 160 | u8 sysctl_ip_nonlocal_bind; |
| 161 | u8 sysctl_ip_autobind_reuse; |
| 162 | /* Shall we try to damage output packets if routing dev changes? */ |
| 163 | u8 sysctl_ip_dynaddr; |
| 164 | #ifdef CONFIG_NET_L3_MASTER_DEV |
| 165 | u8 sysctl_raw_l3mdev_accept; |
| 166 | #endif |
| 167 | u8 sysctl_udp_early_demux; |
| 168 | |
| 169 | u8 sysctl_nexthop_compat_mode; |
| 170 | |
| 171 | u8 sysctl_fwmark_reflect; |
| 172 | u8 sysctl_tcp_fwmark_accept; |
| 173 | u8 sysctl_tcp_mtu_probing; |
| 174 | int sysctl_tcp_mtu_probe_floor; |
| 175 | int sysctl_tcp_base_mss; |
| 176 | int sysctl_tcp_probe_threshold; |
| 177 | u32 sysctl_tcp_probe_interval; |
| 178 | |
| 179 | int sysctl_tcp_keepalive_time; |
| 180 | int sysctl_tcp_keepalive_intvl; |
| 181 | u8 sysctl_tcp_keepalive_probes; |
| 182 | |
| 183 | u8 sysctl_tcp_syn_retries; |
| 184 | u8 sysctl_tcp_synack_retries; |
| 185 | u8 sysctl_tcp_syncookies; |
| 186 | u8 sysctl_tcp_migrate_req; |
| 187 | u8 sysctl_tcp_comp_sack_nr; |
| 188 | u8 sysctl_tcp_backlog_ack_defer; |
| 189 | u8 sysctl_tcp_pingpong_thresh; |
| 190 | |
| 191 | u8 sysctl_tcp_retries1; |
| 192 | u8 sysctl_tcp_retries2; |
| 193 | u8 sysctl_tcp_orphan_retries; |
| 194 | u8 sysctl_tcp_tw_reuse; |
| 195 | unsigned int sysctl_tcp_tw_reuse_delay; |
| 196 | int sysctl_tcp_fin_timeout; |
| 197 | u8 sysctl_tcp_sack; |
| 198 | u8 sysctl_tcp_window_scaling; |
| 199 | u8 sysctl_tcp_timestamps; |
| 200 | int sysctl_tcp_rto_min_us; |
| 201 | int sysctl_tcp_rto_max_ms; |
| 202 | u8 sysctl_tcp_recovery; |
| 203 | u8 sysctl_tcp_thin_linear_timeouts; |
| 204 | u8 sysctl_tcp_slow_start_after_idle; |
| 205 | u8 sysctl_tcp_retrans_collapse; |
| 206 | u8 sysctl_tcp_stdurg; |
| 207 | u8 sysctl_tcp_rfc1337; |
| 208 | u8 sysctl_tcp_abort_on_overflow; |
| 209 | u8 sysctl_tcp_fack; /* obsolete */ |
| 210 | int sysctl_tcp_max_reordering; |
| 211 | int sysctl_tcp_adv_win_scale; /* obsolete */ |
| 212 | u8 sysctl_tcp_dsack; |
| 213 | u8 sysctl_tcp_app_win; |
| 214 | u8 sysctl_tcp_frto; |
| 215 | u8 sysctl_tcp_nometrics_save; |
| 216 | u8 sysctl_tcp_no_ssthresh_metrics_save; |
| 217 | u8 sysctl_tcp_workaround_signed_windows; |
| 218 | int sysctl_tcp_challenge_ack_limit; |
| 219 | u8 sysctl_tcp_min_tso_segs; |
| 220 | u8 sysctl_tcp_reflect_tos; |
| 221 | int sysctl_tcp_invalid_ratelimit; |
| 222 | int sysctl_tcp_pacing_ss_ratio; |
| 223 | int sysctl_tcp_pacing_ca_ratio; |
| 224 | unsigned int sysctl_tcp_child_ehash_entries; |
| 225 | int sysctl_tcp_comp_sack_rtt_percent; |
| 226 | unsigned long sysctl_tcp_comp_sack_delay_ns; |
| 227 | unsigned long sysctl_tcp_comp_sack_slack_ns; |
| 228 | int sysctl_max_syn_backlog; |
| 229 | int sysctl_tcp_fastopen; |
| 230 | const struct tcp_congestion_ops __rcu *tcp_congestion_control; |
| 231 | struct tcp_fastopen_context __rcu *tcp_fastopen_ctx; |
| 232 | unsigned int sysctl_tcp_fastopen_blackhole_timeout; |
| 233 | atomic_t tfo_active_disable_times; |
| 234 | unsigned long tfo_active_disable_stamp; |
| 235 | u32 tcp_challenge_timestamp; |
| 236 | u32 tcp_challenge_count; |
| 237 | u8 sysctl_tcp_plb_enabled; |
| 238 | u8 sysctl_tcp_plb_idle_rehash_rounds; |
| 239 | u8 sysctl_tcp_plb_rehash_rounds; |
| 240 | u8 sysctl_tcp_plb_suspend_rto_sec; |
| 241 | int sysctl_tcp_plb_cong_thresh; |
| 242 | |
| 243 | int sysctl_udp_wmem_min; |
| 244 | int sysctl_udp_rmem_min; |
| 245 | |
| 246 | u8 sysctl_fib_notify_on_flag_change; |
| 247 | u8 sysctl_tcp_syn_linear_timeouts; |
| 248 | |
| 249 | #ifdef CONFIG_NET_L3_MASTER_DEV |
| 250 | u8 sysctl_udp_l3mdev_accept; |
| 251 | #endif |
| 252 | |
| 253 | u8 sysctl_igmp_llm_reports; |
| 254 | int sysctl_igmp_max_memberships; |
| 255 | int sysctl_igmp_max_msf; |
| 256 | int sysctl_igmp_qrv; |
| 257 | |
| 258 | struct ping_group_range ping_group_range; |
| 259 | u16 ping_port_rover; |
| 260 | |
| 261 | atomic_t dev_addr_genid; |
| 262 | |
| 263 | unsigned int sysctl_udp_child_hash_entries; |
| 264 | |
| 265 | #ifdef CONFIG_SYSCTL |
| 266 | unsigned long *sysctl_local_reserved_ports; |
| 267 | int sysctl_ip_prot_sock; |
| 268 | #endif |
| 269 | |
| 270 | #ifdef CONFIG_IP_MROUTE |
| 271 | #ifndef CONFIG_IP_MROUTE_MULTIPLE_TABLES |
| 272 | struct mr_table *mrt; |
| 273 | #else |
| 274 | struct list_head mr_tables; |
| 275 | struct fib_rules_ops *mr_rules_ops; |
| 276 | #endif |
| 277 | #endif |
| 278 | #ifdef CONFIG_IP_ROUTE_MULTIPATH |
| 279 | struct sysctl_fib_multipath_hash_seed sysctl_fib_multipath_hash_seed; |
| 280 | u32 sysctl_fib_multipath_hash_fields; |
| 281 | u8 sysctl_fib_multipath_use_neigh; |
| 282 | u8 sysctl_fib_multipath_hash_policy; |
| 283 | #endif |
| 284 | |
| 285 | struct fib_notifier_ops *notifier_ops; |
| 286 | unsigned int fib_seq; /* writes protected by rtnl_mutex */ |
| 287 | |
| 288 | struct fib_notifier_ops *ipmr_notifier_ops; |
| 289 | unsigned int ipmr_seq; /* protected by rtnl_mutex */ |
| 290 | |
| 291 | atomic_t rt_genid; |
| 292 | siphash_key_t ip_id_key; |
| 293 | struct hlist_head *inet_addr_lst; |
| 294 | struct delayed_work addr_chk_work; |
| 295 | }; |
| 296 | |
| 297 | #endif |
| 298 | |