David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 1 | #include <linux/rcupdate.h> |
| 2 | #include <linux/spinlock.h> |
| 3 | #include <linux/jiffies.h> |
| 4 | #include <linux/bootmem.h> |
David S. Miller | ab92bb2 | 2012-07-09 16:19:30 -0700 | [diff] [blame] | 5 | #include <linux/module.h> |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 6 | #include <linux/cache.h> |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 7 | #include <linux/slab.h> |
| 8 | #include <linux/init.h> |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 9 | #include <linux/tcp.h> |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 10 | #include <linux/hash.h> |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 11 | |
| 12 | #include <net/inet_connection_sock.h> |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 13 | #include <net/net_namespace.h> |
David S. Miller | ab92bb2 | 2012-07-09 16:19:30 -0700 | [diff] [blame] | 14 | #include <net/request_sock.h> |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 15 | #include <net/inetpeer.h> |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 16 | #include <net/sock.h> |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 17 | #include <net/ipv6.h> |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 18 | #include <net/dst.h> |
| 19 | #include <net/tcp.h> |
| 20 | |
| 21 | int sysctl_tcp_nometrics_save __read_mostly; |
| 22 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 23 | enum tcp_metric_index { |
| 24 | TCP_METRIC_RTT, |
| 25 | TCP_METRIC_RTTVAR, |
| 26 | TCP_METRIC_SSTHRESH, |
| 27 | TCP_METRIC_CWND, |
| 28 | TCP_METRIC_REORDERING, |
| 29 | |
| 30 | /* Always last. */ |
| 31 | TCP_METRIC_MAX, |
| 32 | }; |
| 33 | |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 34 | struct tcp_fastopen_metrics { |
| 35 | u16 mss; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 36 | u16 syn_loss:10; /* Recurring Fast Open SYN losses */ |
| 37 | unsigned long last_syn_loss; /* Last Fast Open SYN loss */ |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 38 | struct tcp_fastopen_cookie cookie; |
| 39 | }; |
| 40 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 41 | struct tcp_metrics_block { |
| 42 | struct tcp_metrics_block __rcu *tcpm_next; |
| 43 | struct inetpeer_addr tcpm_addr; |
| 44 | unsigned long tcpm_stamp; |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 45 | u32 tcpm_ts; |
| 46 | u32 tcpm_ts_stamp; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 47 | u32 tcpm_lock; |
| 48 | u32 tcpm_vals[TCP_METRIC_MAX]; |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 49 | struct tcp_fastopen_metrics tcpm_fastopen; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 50 | }; |
| 51 | |
| 52 | static bool tcp_metric_locked(struct tcp_metrics_block *tm, |
| 53 | enum tcp_metric_index idx) |
| 54 | { |
| 55 | return tm->tcpm_lock & (1 << idx); |
| 56 | } |
| 57 | |
| 58 | static u32 tcp_metric_get(struct tcp_metrics_block *tm, |
| 59 | enum tcp_metric_index idx) |
| 60 | { |
| 61 | return tm->tcpm_vals[idx]; |
| 62 | } |
| 63 | |
| 64 | static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm, |
| 65 | enum tcp_metric_index idx) |
| 66 | { |
| 67 | return msecs_to_jiffies(tm->tcpm_vals[idx]); |
| 68 | } |
| 69 | |
| 70 | static void tcp_metric_set(struct tcp_metrics_block *tm, |
| 71 | enum tcp_metric_index idx, |
| 72 | u32 val) |
| 73 | { |
| 74 | tm->tcpm_vals[idx] = val; |
| 75 | } |
| 76 | |
| 77 | static void tcp_metric_set_msecs(struct tcp_metrics_block *tm, |
| 78 | enum tcp_metric_index idx, |
| 79 | u32 val) |
| 80 | { |
| 81 | tm->tcpm_vals[idx] = jiffies_to_msecs(val); |
| 82 | } |
| 83 | |
| 84 | static bool addr_same(const struct inetpeer_addr *a, |
| 85 | const struct inetpeer_addr *b) |
| 86 | { |
| 87 | const struct in6_addr *a6, *b6; |
| 88 | |
| 89 | if (a->family != b->family) |
| 90 | return false; |
| 91 | if (a->family == AF_INET) |
| 92 | return a->addr.a4 == b->addr.a4; |
| 93 | |
| 94 | a6 = (const struct in6_addr *) &a->addr.a6[0]; |
| 95 | b6 = (const struct in6_addr *) &b->addr.a6[0]; |
| 96 | |
| 97 | return ipv6_addr_equal(a6, b6); |
| 98 | } |
| 99 | |
| 100 | struct tcpm_hash_bucket { |
| 101 | struct tcp_metrics_block __rcu *chain; |
| 102 | }; |
| 103 | |
| 104 | static DEFINE_SPINLOCK(tcp_metrics_lock); |
| 105 | |
| 106 | static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst) |
| 107 | { |
| 108 | u32 val; |
| 109 | |
| 110 | val = 0; |
| 111 | if (dst_metric_locked(dst, RTAX_RTT)) |
| 112 | val |= 1 << TCP_METRIC_RTT; |
| 113 | if (dst_metric_locked(dst, RTAX_RTTVAR)) |
| 114 | val |= 1 << TCP_METRIC_RTTVAR; |
| 115 | if (dst_metric_locked(dst, RTAX_SSTHRESH)) |
| 116 | val |= 1 << TCP_METRIC_SSTHRESH; |
| 117 | if (dst_metric_locked(dst, RTAX_CWND)) |
| 118 | val |= 1 << TCP_METRIC_CWND; |
| 119 | if (dst_metric_locked(dst, RTAX_REORDERING)) |
| 120 | val |= 1 << TCP_METRIC_REORDERING; |
| 121 | tm->tcpm_lock = val; |
| 122 | |
| 123 | tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT); |
| 124 | tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR); |
| 125 | tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH); |
| 126 | tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND); |
| 127 | tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING); |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 128 | tm->tcpm_ts = 0; |
| 129 | tm->tcpm_ts_stamp = 0; |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 130 | tm->tcpm_fastopen.mss = 0; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 131 | tm->tcpm_fastopen.syn_loss = 0; |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 132 | tm->tcpm_fastopen.cookie.len = 0; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 133 | } |
| 134 | |
| 135 | static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst, |
| 136 | struct inetpeer_addr *addr, |
| 137 | unsigned int hash, |
| 138 | bool reclaim) |
| 139 | { |
| 140 | struct tcp_metrics_block *tm; |
| 141 | struct net *net; |
| 142 | |
| 143 | spin_lock_bh(&tcp_metrics_lock); |
| 144 | net = dev_net(dst->dev); |
| 145 | if (unlikely(reclaim)) { |
| 146 | struct tcp_metrics_block *oldest; |
| 147 | |
| 148 | oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); |
| 149 | for (tm = rcu_dereference(oldest->tcpm_next); tm; |
| 150 | tm = rcu_dereference(tm->tcpm_next)) { |
| 151 | if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp)) |
| 152 | oldest = tm; |
| 153 | } |
| 154 | tm = oldest; |
| 155 | } else { |
| 156 | tm = kmalloc(sizeof(*tm), GFP_ATOMIC); |
| 157 | if (!tm) |
| 158 | goto out_unlock; |
| 159 | } |
| 160 | tm->tcpm_addr = *addr; |
| 161 | tm->tcpm_stamp = jiffies; |
| 162 | |
| 163 | tcpm_suck_dst(tm, dst); |
| 164 | |
| 165 | if (likely(!reclaim)) { |
| 166 | tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain; |
| 167 | rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm); |
| 168 | } |
| 169 | |
| 170 | out_unlock: |
| 171 | spin_unlock_bh(&tcp_metrics_lock); |
| 172 | return tm; |
| 173 | } |
| 174 | |
| 175 | #define TCP_METRICS_TIMEOUT (60 * 60 * HZ) |
| 176 | |
| 177 | static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst) |
| 178 | { |
| 179 | if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT))) |
| 180 | tcpm_suck_dst(tm, dst); |
| 181 | } |
| 182 | |
| 183 | #define TCP_METRICS_RECLAIM_DEPTH 5 |
| 184 | #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL |
| 185 | |
| 186 | static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth) |
| 187 | { |
| 188 | if (tm) |
| 189 | return tm; |
| 190 | if (depth > TCP_METRICS_RECLAIM_DEPTH) |
| 191 | return TCP_METRICS_RECLAIM_PTR; |
| 192 | return NULL; |
| 193 | } |
| 194 | |
| 195 | static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr, |
| 196 | struct net *net, unsigned int hash) |
| 197 | { |
| 198 | struct tcp_metrics_block *tm; |
| 199 | int depth = 0; |
| 200 | |
| 201 | for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; |
| 202 | tm = rcu_dereference(tm->tcpm_next)) { |
| 203 | if (addr_same(&tm->tcpm_addr, addr)) |
| 204 | break; |
| 205 | depth++; |
| 206 | } |
| 207 | return tcp_get_encode(tm, depth); |
| 208 | } |
| 209 | |
| 210 | static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req, |
| 211 | struct dst_entry *dst) |
| 212 | { |
| 213 | struct tcp_metrics_block *tm; |
| 214 | struct inetpeer_addr addr; |
| 215 | unsigned int hash; |
| 216 | struct net *net; |
| 217 | |
| 218 | addr.family = req->rsk_ops->family; |
| 219 | switch (addr.family) { |
| 220 | case AF_INET: |
| 221 | addr.addr.a4 = inet_rsk(req)->rmt_addr; |
| 222 | hash = (__force unsigned int) addr.addr.a4; |
| 223 | break; |
| 224 | case AF_INET6: |
| 225 | *(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr; |
Eric Dumazet | ddbe503 | 2012-07-18 08:11:12 +0000 | [diff] [blame] | 226 | hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 227 | break; |
| 228 | default: |
| 229 | return NULL; |
| 230 | } |
| 231 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 232 | net = dev_net(dst->dev); |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 233 | hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 234 | |
| 235 | for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; |
| 236 | tm = rcu_dereference(tm->tcpm_next)) { |
| 237 | if (addr_same(&tm->tcpm_addr, &addr)) |
| 238 | break; |
| 239 | } |
| 240 | tcpm_check_stamp(tm, dst); |
| 241 | return tm; |
| 242 | } |
| 243 | |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 244 | static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw) |
| 245 | { |
| 246 | struct inet6_timewait_sock *tw6; |
| 247 | struct tcp_metrics_block *tm; |
| 248 | struct inetpeer_addr addr; |
| 249 | unsigned int hash; |
| 250 | struct net *net; |
| 251 | |
| 252 | addr.family = tw->tw_family; |
| 253 | switch (addr.family) { |
| 254 | case AF_INET: |
| 255 | addr.addr.a4 = tw->tw_daddr; |
| 256 | hash = (__force unsigned int) addr.addr.a4; |
| 257 | break; |
| 258 | case AF_INET6: |
| 259 | tw6 = inet6_twsk((struct sock *)tw); |
| 260 | *(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr; |
Eric Dumazet | ddbe503 | 2012-07-18 08:11:12 +0000 | [diff] [blame] | 261 | hash = ipv6_addr_hash(&tw6->tw_v6_daddr); |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 262 | break; |
| 263 | default: |
| 264 | return NULL; |
| 265 | } |
| 266 | |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 267 | net = twsk_net(tw); |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 268 | hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 269 | |
| 270 | for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; |
| 271 | tm = rcu_dereference(tm->tcpm_next)) { |
| 272 | if (addr_same(&tm->tcpm_addr, &addr)) |
| 273 | break; |
| 274 | } |
| 275 | return tm; |
| 276 | } |
| 277 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 278 | static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk, |
| 279 | struct dst_entry *dst, |
| 280 | bool create) |
| 281 | { |
| 282 | struct tcp_metrics_block *tm; |
| 283 | struct inetpeer_addr addr; |
| 284 | unsigned int hash; |
| 285 | struct net *net; |
| 286 | bool reclaim; |
| 287 | |
| 288 | addr.family = sk->sk_family; |
| 289 | switch (addr.family) { |
| 290 | case AF_INET: |
| 291 | addr.addr.a4 = inet_sk(sk)->inet_daddr; |
| 292 | hash = (__force unsigned int) addr.addr.a4; |
| 293 | break; |
| 294 | case AF_INET6: |
| 295 | *(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr; |
Eric Dumazet | ddbe503 | 2012-07-18 08:11:12 +0000 | [diff] [blame] | 296 | hash = ipv6_addr_hash(&inet6_sk(sk)->daddr); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 297 | break; |
| 298 | default: |
| 299 | return NULL; |
| 300 | } |
| 301 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 302 | net = dev_net(dst->dev); |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 303 | hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 304 | |
| 305 | tm = __tcp_get_metrics(&addr, net, hash); |
| 306 | reclaim = false; |
| 307 | if (tm == TCP_METRICS_RECLAIM_PTR) { |
| 308 | reclaim = true; |
| 309 | tm = NULL; |
| 310 | } |
| 311 | if (!tm && create) |
| 312 | tm = tcpm_new(dst, &addr, hash, reclaim); |
| 313 | else |
| 314 | tcpm_check_stamp(tm, dst); |
| 315 | |
| 316 | return tm; |
| 317 | } |
| 318 | |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 319 | /* Save metrics learned by this TCP session. This function is called |
| 320 | * only, when TCP finishes successfully i.e. when it enters TIME-WAIT |
| 321 | * or goes from LAST-ACK to CLOSE. |
| 322 | */ |
| 323 | void tcp_update_metrics(struct sock *sk) |
| 324 | { |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 325 | const struct inet_connection_sock *icsk = inet_csk(sk); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 326 | struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 327 | struct tcp_sock *tp = tcp_sk(sk); |
| 328 | struct tcp_metrics_block *tm; |
| 329 | unsigned long rtt; |
| 330 | u32 val; |
| 331 | int m; |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 332 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 333 | if (sysctl_tcp_nometrics_save || !dst) |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 334 | return; |
| 335 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 336 | if (dst->flags & DST_HOST) |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 337 | dst_confirm(dst); |
| 338 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 339 | rcu_read_lock(); |
| 340 | if (icsk->icsk_backoff || !tp->srtt) { |
| 341 | /* This session failed to estimate rtt. Why? |
| 342 | * Probably, no packets returned in time. Reset our |
| 343 | * results. |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 344 | */ |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 345 | tm = tcp_get_metrics(sk, dst, false); |
| 346 | if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT)) |
| 347 | tcp_metric_set(tm, TCP_METRIC_RTT, 0); |
| 348 | goto out_unlock; |
| 349 | } else |
| 350 | tm = tcp_get_metrics(sk, dst, true); |
| 351 | |
| 352 | if (!tm) |
| 353 | goto out_unlock; |
| 354 | |
| 355 | rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT); |
| 356 | m = rtt - tp->srtt; |
| 357 | |
| 358 | /* If newly calculated rtt larger than stored one, store new |
| 359 | * one. Otherwise, use EWMA. Remember, rtt overestimation is |
| 360 | * always better than underestimation. |
| 361 | */ |
| 362 | if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) { |
| 363 | if (m <= 0) |
| 364 | rtt = tp->srtt; |
| 365 | else |
| 366 | rtt -= (m >> 3); |
| 367 | tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt); |
| 368 | } |
| 369 | |
| 370 | if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) { |
| 371 | unsigned long var; |
| 372 | |
| 373 | if (m < 0) |
| 374 | m = -m; |
| 375 | |
| 376 | /* Scale deviation to rttvar fixed point */ |
| 377 | m >>= 1; |
| 378 | if (m < tp->mdev) |
| 379 | m = tp->mdev; |
| 380 | |
| 381 | var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR); |
| 382 | if (m >= var) |
| 383 | var = m; |
| 384 | else |
| 385 | var -= (var - m) >> 2; |
| 386 | |
| 387 | tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var); |
| 388 | } |
| 389 | |
| 390 | if (tcp_in_initial_slowstart(tp)) { |
| 391 | /* Slow start still did not finish. */ |
| 392 | if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { |
| 393 | val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); |
| 394 | if (val && (tp->snd_cwnd >> 1) > val) |
| 395 | tcp_metric_set(tm, TCP_METRIC_SSTHRESH, |
| 396 | tp->snd_cwnd >> 1); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 397 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 398 | if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { |
| 399 | val = tcp_metric_get(tm, TCP_METRIC_CWND); |
| 400 | if (tp->snd_cwnd > val) |
| 401 | tcp_metric_set(tm, TCP_METRIC_CWND, |
| 402 | tp->snd_cwnd); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 403 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 404 | } else if (tp->snd_cwnd > tp->snd_ssthresh && |
| 405 | icsk->icsk_ca_state == TCP_CA_Open) { |
| 406 | /* Cong. avoidance phase, cwnd is reliable. */ |
| 407 | if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) |
| 408 | tcp_metric_set(tm, TCP_METRIC_SSTHRESH, |
| 409 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); |
| 410 | if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { |
| 411 | val = tcp_metric_get(tm, TCP_METRIC_CWND); |
Alexander Duyck | 2100844 | 2012-07-11 17:18:04 -0700 | [diff] [blame] | 412 | tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 413 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 414 | } else { |
| 415 | /* Else slow start did not finish, cwnd is non-sense, |
| 416 | * ssthresh may be also invalid. |
| 417 | */ |
| 418 | if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { |
| 419 | val = tcp_metric_get(tm, TCP_METRIC_CWND); |
| 420 | tcp_metric_set(tm, TCP_METRIC_CWND, |
| 421 | (val + tp->snd_ssthresh) >> 1); |
| 422 | } |
| 423 | if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { |
| 424 | val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); |
| 425 | if (val && tp->snd_ssthresh > val) |
| 426 | tcp_metric_set(tm, TCP_METRIC_SSTHRESH, |
| 427 | tp->snd_ssthresh); |
| 428 | } |
| 429 | if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) { |
| 430 | val = tcp_metric_get(tm, TCP_METRIC_REORDERING); |
| 431 | if (val < tp->reordering && |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 432 | tp->reordering != sysctl_tcp_reordering) |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 433 | tcp_metric_set(tm, TCP_METRIC_REORDERING, |
| 434 | tp->reordering); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 435 | } |
| 436 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 437 | tm->tcpm_stamp = jiffies; |
| 438 | out_unlock: |
| 439 | rcu_read_unlock(); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 440 | } |
| 441 | |
| 442 | /* Initialize metrics on socket. */ |
| 443 | |
| 444 | void tcp_init_metrics(struct sock *sk) |
| 445 | { |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 446 | struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 447 | struct tcp_sock *tp = tcp_sk(sk); |
| 448 | struct tcp_metrics_block *tm; |
| 449 | u32 val; |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 450 | |
| 451 | if (dst == NULL) |
| 452 | goto reset; |
| 453 | |
| 454 | dst_confirm(dst); |
| 455 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 456 | rcu_read_lock(); |
| 457 | tm = tcp_get_metrics(sk, dst, true); |
| 458 | if (!tm) { |
| 459 | rcu_read_unlock(); |
| 460 | goto reset; |
| 461 | } |
| 462 | |
| 463 | if (tcp_metric_locked(tm, TCP_METRIC_CWND)) |
| 464 | tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND); |
| 465 | |
| 466 | val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); |
| 467 | if (val) { |
| 468 | tp->snd_ssthresh = val; |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 469 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) |
| 470 | tp->snd_ssthresh = tp->snd_cwnd_clamp; |
| 471 | } else { |
| 472 | /* ssthresh may have been reduced unnecessarily during. |
| 473 | * 3WHS. Restore it back to its initial default. |
| 474 | */ |
| 475 | tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; |
| 476 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 477 | val = tcp_metric_get(tm, TCP_METRIC_REORDERING); |
| 478 | if (val && tp->reordering != val) { |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 479 | tcp_disable_fack(tp); |
| 480 | tcp_disable_early_retrans(tp); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 481 | tp->reordering = val; |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 482 | } |
| 483 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 484 | val = tcp_metric_get(tm, TCP_METRIC_RTT); |
| 485 | if (val == 0 || tp->srtt == 0) { |
| 486 | rcu_read_unlock(); |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 487 | goto reset; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 488 | } |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 489 | /* Initial rtt is determined from SYN,SYN-ACK. |
| 490 | * The segment is small and rtt may appear much |
| 491 | * less than real one. Use per-dst memory |
| 492 | * to make it more realistic. |
| 493 | * |
| 494 | * A bit of theory. RTT is time passed after "normal" sized packet |
| 495 | * is sent until it is ACKed. In normal circumstances sending small |
| 496 | * packets force peer to delay ACKs and calculation is correct too. |
| 497 | * The algorithm is adaptive and, provided we follow specs, it |
| 498 | * NEVER underestimate RTT. BUT! If peer tries to make some clever |
| 499 | * tricks sort of "quick acks" for time long enough to decrease RTT |
| 500 | * to low value, and then abruptly stops to do it and starts to delay |
| 501 | * ACKs, wait for troubles. |
| 502 | */ |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 503 | val = msecs_to_jiffies(val); |
| 504 | if (val > tp->srtt) { |
| 505 | tp->srtt = val; |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 506 | tp->rtt_seq = tp->snd_nxt; |
| 507 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 508 | val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR); |
| 509 | if (val > tp->mdev) { |
| 510 | tp->mdev = val; |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 511 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
| 512 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 513 | rcu_read_unlock(); |
| 514 | |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 515 | tcp_set_rto(sk); |
| 516 | reset: |
| 517 | if (tp->srtt == 0) { |
| 518 | /* RFC6298: 5.7 We've failed to get a valid RTT sample from |
| 519 | * 3WHS. This is most likely due to retransmission, |
| 520 | * including spurious one. Reset the RTO back to 3secs |
| 521 | * from the more aggressive 1sec to avoid more spurious |
| 522 | * retransmission. |
| 523 | */ |
| 524 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; |
| 525 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; |
| 526 | } |
| 527 | /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been |
| 528 | * retransmitted. In light of RFC6298 more aggressive 1sec |
| 529 | * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK |
| 530 | * retransmission has occurred. |
| 531 | */ |
| 532 | if (tp->total_retrans > 1) |
| 533 | tp->snd_cwnd = 1; |
| 534 | else |
| 535 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); |
| 536 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 537 | } |
David S. Miller | ab92bb2 | 2012-07-09 16:19:30 -0700 | [diff] [blame] | 538 | |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 539 | bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check) |
David S. Miller | ab92bb2 | 2012-07-09 16:19:30 -0700 | [diff] [blame] | 540 | { |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 541 | struct tcp_metrics_block *tm; |
| 542 | bool ret; |
| 543 | |
David S. Miller | ab92bb2 | 2012-07-09 16:19:30 -0700 | [diff] [blame] | 544 | if (!dst) |
| 545 | return false; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 546 | |
| 547 | rcu_read_lock(); |
| 548 | tm = __tcp_get_metrics_req(req, dst); |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 549 | if (paws_check) { |
| 550 | if (tm && |
| 551 | (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL && |
| 552 | (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW) |
| 553 | ret = false; |
| 554 | else |
| 555 | ret = true; |
| 556 | } else { |
| 557 | if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp) |
| 558 | ret = true; |
| 559 | else |
| 560 | ret = false; |
| 561 | } |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 562 | rcu_read_unlock(); |
| 563 | |
| 564 | return ret; |
David S. Miller | ab92bb2 | 2012-07-09 16:19:30 -0700 | [diff] [blame] | 565 | } |
| 566 | EXPORT_SYMBOL_GPL(tcp_peer_is_proven); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 567 | |
David S. Miller | 81166dd | 2012-07-10 03:14:24 -0700 | [diff] [blame] | 568 | void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst) |
| 569 | { |
| 570 | struct tcp_metrics_block *tm; |
| 571 | |
| 572 | rcu_read_lock(); |
| 573 | tm = tcp_get_metrics(sk, dst, true); |
| 574 | if (tm) { |
| 575 | struct tcp_sock *tp = tcp_sk(sk); |
| 576 | |
| 577 | if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) { |
| 578 | tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp; |
| 579 | tp->rx_opt.ts_recent = tm->tcpm_ts; |
| 580 | } |
| 581 | } |
| 582 | rcu_read_unlock(); |
| 583 | } |
| 584 | EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp); |
| 585 | |
| 586 | /* VJ's idea. Save last timestamp seen from this destination and hold |
| 587 | * it at least for normal timewait interval to use for duplicate |
| 588 | * segment detection in subsequent connections, before they enter |
| 589 | * synchronized state. |
| 590 | */ |
| 591 | bool tcp_remember_stamp(struct sock *sk) |
| 592 | { |
| 593 | struct dst_entry *dst = __sk_dst_get(sk); |
| 594 | bool ret = false; |
| 595 | |
| 596 | if (dst) { |
| 597 | struct tcp_metrics_block *tm; |
| 598 | |
| 599 | rcu_read_lock(); |
| 600 | tm = tcp_get_metrics(sk, dst, true); |
| 601 | if (tm) { |
| 602 | struct tcp_sock *tp = tcp_sk(sk); |
| 603 | |
| 604 | if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 || |
| 605 | ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && |
| 606 | tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { |
| 607 | tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; |
| 608 | tm->tcpm_ts = tp->rx_opt.ts_recent; |
| 609 | } |
| 610 | ret = true; |
| 611 | } |
| 612 | rcu_read_unlock(); |
| 613 | } |
| 614 | return ret; |
| 615 | } |
| 616 | |
| 617 | bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) |
| 618 | { |
| 619 | struct tcp_metrics_block *tm; |
| 620 | bool ret = false; |
| 621 | |
| 622 | rcu_read_lock(); |
| 623 | tm = __tcp_get_metrics_tw(tw); |
| 624 | if (tw) { |
| 625 | const struct tcp_timewait_sock *tcptw; |
| 626 | struct sock *sk = (struct sock *) tw; |
| 627 | |
| 628 | tcptw = tcp_twsk(sk); |
| 629 | if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 || |
| 630 | ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && |
| 631 | tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { |
| 632 | tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; |
| 633 | tm->tcpm_ts = tcptw->tw_ts_recent; |
| 634 | } |
| 635 | ret = true; |
| 636 | } |
| 637 | rcu_read_unlock(); |
| 638 | |
| 639 | return ret; |
| 640 | } |
| 641 | |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 642 | static DEFINE_SEQLOCK(fastopen_seqlock); |
| 643 | |
| 644 | void tcp_fastopen_cache_get(struct sock *sk, u16 *mss, |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 645 | struct tcp_fastopen_cookie *cookie, |
| 646 | int *syn_loss, unsigned long *last_syn_loss) |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 647 | { |
| 648 | struct tcp_metrics_block *tm; |
| 649 | |
| 650 | rcu_read_lock(); |
| 651 | tm = tcp_get_metrics(sk, __sk_dst_get(sk), false); |
| 652 | if (tm) { |
| 653 | struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; |
| 654 | unsigned int seq; |
| 655 | |
| 656 | do { |
| 657 | seq = read_seqbegin(&fastopen_seqlock); |
| 658 | if (tfom->mss) |
| 659 | *mss = tfom->mss; |
| 660 | *cookie = tfom->cookie; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 661 | *syn_loss = tfom->syn_loss; |
| 662 | *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0; |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 663 | } while (read_seqretry(&fastopen_seqlock, seq)); |
| 664 | } |
| 665 | rcu_read_unlock(); |
| 666 | } |
| 667 | |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 668 | void tcp_fastopen_cache_set(struct sock *sk, u16 mss, |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 669 | struct tcp_fastopen_cookie *cookie, bool syn_lost) |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 670 | { |
| 671 | struct tcp_metrics_block *tm; |
| 672 | |
| 673 | rcu_read_lock(); |
| 674 | tm = tcp_get_metrics(sk, __sk_dst_get(sk), true); |
| 675 | if (tm) { |
| 676 | struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; |
| 677 | |
| 678 | write_seqlock_bh(&fastopen_seqlock); |
| 679 | tfom->mss = mss; |
| 680 | if (cookie->len > 0) |
| 681 | tfom->cookie = *cookie; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 682 | if (syn_lost) { |
| 683 | ++tfom->syn_loss; |
| 684 | tfom->last_syn_loss = jiffies; |
| 685 | } else |
| 686 | tfom->syn_loss = 0; |
Yuchung Cheng | 1fe4c48 | 2012-07-19 06:43:06 +0000 | [diff] [blame] | 687 | write_sequnlock_bh(&fastopen_seqlock); |
| 688 | } |
| 689 | rcu_read_unlock(); |
| 690 | } |
| 691 | |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 692 | static unsigned int tcpmhash_entries; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 693 | static int __init set_tcpmhash_entries(char *str) |
| 694 | { |
| 695 | ssize_t ret; |
| 696 | |
| 697 | if (!str) |
| 698 | return 0; |
| 699 | |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 700 | ret = kstrtouint(str, 0, &tcpmhash_entries); |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 701 | if (ret) |
| 702 | return 0; |
| 703 | |
| 704 | return 1; |
| 705 | } |
| 706 | __setup("tcpmhash_entries=", set_tcpmhash_entries); |
| 707 | |
| 708 | static int __net_init tcp_net_metrics_init(struct net *net) |
| 709 | { |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 710 | size_t size; |
| 711 | unsigned int slots; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 712 | |
| 713 | slots = tcpmhash_entries; |
| 714 | if (!slots) { |
| 715 | if (totalram_pages >= 128 * 1024) |
| 716 | slots = 16 * 1024; |
| 717 | else |
| 718 | slots = 8 * 1024; |
| 719 | } |
| 720 | |
Eric Dumazet | 5815d5e | 2012-07-19 23:02:34 +0000 | [diff] [blame] | 721 | net->ipv4.tcp_metrics_hash_log = order_base_2(slots); |
| 722 | size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log; |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 723 | |
| 724 | net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL); |
| 725 | if (!net->ipv4.tcp_metrics_hash) |
| 726 | return -ENOMEM; |
| 727 | |
David S. Miller | 51c5d0c | 2012-07-10 00:49:14 -0700 | [diff] [blame] | 728 | return 0; |
| 729 | } |
| 730 | |
| 731 | static void __net_exit tcp_net_metrics_exit(struct net *net) |
| 732 | { |
| 733 | kfree(net->ipv4.tcp_metrics_hash); |
| 734 | } |
| 735 | |
| 736 | static __net_initdata struct pernet_operations tcp_net_metrics_ops = { |
| 737 | .init = tcp_net_metrics_init, |
| 738 | .exit = tcp_net_metrics_exit, |
| 739 | }; |
| 740 | |
| 741 | void __init tcp_metrics_init(void) |
| 742 | { |
| 743 | register_pernet_subsys(&tcp_net_metrics_ops); |
| 744 | } |