bsd/netinet/tcp_var.h

   1 /*
   2  * Copyright (c) 2000-2015 Apple Inc. All rights reserved.
   3  *
   4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
   5  *
   6  * This file contains Original Code and/or Modifications of Original Code
   7  * as defined in and that are subject to the Apple Public Source License
   8  * Version 2.0 (the 'License'). You may not use this file except in
   9  * compliance with the License. The rights granted to you under the License
  10  * may not be used to create, or enable the creation or redistribution of,
  11  * unlawful or unlicensed copies of an Apple operating system, or to
  12  * circumvent, violate, or enable the circumvention or violation of, any
  13  * terms of an Apple operating system software license agreement.
  14  *
  15  * Please obtain a copy of the License at
  16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
  17  *
  18  * The Original Code and all software distributed under the License are
  19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  23  * Please see the License for the specific language governing rights and
  24  * limitations under the License.
  25  *
  26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
  27  */
  28 /*
  29  * Copyright (c) 1982, 1986, 1993, 1994, 1995
  30  *      The Regents of the University of California.  All rights reserved.
  31  *
  32  * Redistribution and use in source and binary forms, with or without
  33  * modification, are permitted provided that the following conditions
  34  * are met:
  35  * 1. Redistributions of source code must retain the above copyright
  36  *    notice, this list of conditions and the following disclaimer.
  37  * 2. Redistributions in binary form must reproduce the above copyright
  38  *    notice, this list of conditions and the following disclaimer in the
  39  *    documentation and/or other materials provided with the distribution.
  40  * 3. All advertising materials mentioning features or use of this software
  41  *    must display the following acknowledgement:
  42  *      This product includes software developed by the University of
  43  *      California, Berkeley and its contributors.
  44  * 4. Neither the name of the University nor the names of its contributors
  45  *    may be used to endorse or promote products derived from this software
  46  *    without specific prior written permission.
  47  *
  48  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  49  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  50  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  51  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  52  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  53  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  54  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  55  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  56  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  57  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  58  * SUCH DAMAGE.
  59  *
  60  *      @(#)tcp_var.h   8.4 (Berkeley) 5/24/95
  61  * $FreeBSD: src/sys/netinet/tcp_var.h,v 1.56.2.8 2001/08/22 00:59:13 silby Exp $
  62  */
  63
  64 #ifndef _NETINET_TCP_VAR_H_
  65 #define _NETINET_TCP_VAR_H_
  66 #include <sys/types.h>
  67 #include <sys/appleapiopts.h>
  68 #include <sys/queue.h>
  69 #include <netinet/in_pcb.h>
  70 #include <netinet/tcp.h>
  71 #include <netinet/tcp_timer.h>
  72
  73 #if defined(__LP64__)
  74 #define _TCPCB_PTR(x)                   u_int32_t
  75 #define _TCPCB_LIST_HEAD(name, type)    \
  76 struct name {                           \
  77         u_int32_t       lh_first;       \
  78 };
  79 #else
  80 #define _TCPCB_PTR(x)                   x
  81 #define _TCPCB_LIST_HEAD(name, type)    LIST_HEAD(name, type)
  82 #endif
  83
  84 #ifdef KERNEL_PRIVATE
  85
  86 #define TCP_RETRANSHZ   1000    /* granularity of TCP timestamps, 1ms */
  87 /* Minimum time quantum within which the timers are coalesced */
  88 #define TCP_TIMER_10MS_QUANTUM  (TCP_RETRANSHZ/100) /* every 10ms */
  89 #define TCP_TIMER_100MS_QUANTUM   (TCP_RETRANSHZ/10) /* every 100ms */
  90 #define TCP_TIMER_500MS_QUANTUM   (TCP_RETRANSHZ/2) /* every 500ms */
  91
  92 #define TCP_RETRANSHZ_TO_USEC 1000
  93
  94 #define N_TIME_WAIT_SLOTS   128         /* must be power of 2 */
  95
  96 /* Base RTT is stored for N_MIN_RTT_HISTORY slots. This is used to
  97  * estimate expected minimum RTT for delay based congestion control
  98  * algorithms.
  99  */
 100 #define N_RTT_BASE      5
 101
 102 /* Always allow at least 4 packets worth of recv window when adjusting
 103  * recv window using inter-packet arrival jitter.
 104  */
 105 #define MIN_IAJ_WIN 4
 106
 107 /* A variation in delay of this many milliseconds is tolerable. This limit has to
 108  * be low but greater than zero. We also use standard deviation on jitter to adjust
 109  * this limit for different link and connection types.
 110  */
 111 #define ALLOWED_IAJ 5
 112
 113 /* Ignore the first few packets on a connection until the ACK clock gets going
 114  */
 115 #define IAJ_IGNORE_PKTCNT 40
 116
 117 /* Let the accumulated IAJ value increase by this threshold at most. This limit
 118  * will control how many ALLOWED_IAJ measurements a receiver will have to see
 119  * before opening the receive window
 120  */
 121 #define ACC_IAJ_HIGH_THRESH 100
 122
 123 /* When accumulated IAJ reaches this value, the receiver starts to react by
 124  * closing the window
 125  */
 126 #define ACC_IAJ_REACT_LIMIT 200
 127
 128 /* If the number of small packets (smaller than IAJ packet size) seen on a
 129  * connection is more than this threshold, reset the size and learn it again.
 130  * This is needed because the sender might send smaller segments after PMTU
 131  * discovery and the receiver has to learn the new size.
 132  */
 133 #define RESET_IAJ_SIZE_THRESH 20
 134
 135 /*
 136  * Adaptive timeout is a read/write timeout specified by the application to
 137  * get a socket event when the transport layer detects a stall in data
 138  * transfer. The value specified is the number of probes that can be sent
 139  * to the peer before generating an event. Since it is not specified as
 140  * a time value, the timeout will adjust based on the RTT seen on the link.
 141  * The timeout will start only when there is an indication that the read/write
 142  * operation is not making progress.
 143  *
 144  * If a write operation stalls, the probe will be retransmission of data.
 145  * If a read operation stalls, the probe will be a keep-alive packet.
 146  *
 147  * The maximum value of adaptive timeout is set to 10 which will allow
 148  * transmission of enough number of probes to the peer.
 149  */
 150 #define TCP_ADAPTIVE_TIMEOUT_MAX        10
 151
 152 #define TCP_CONNECTIVITY_PROBES_MAX     5
 153
 154 /*
 155  * Kernel variables for tcp.
 156  */
 157
 158 /* TCP segment queue entry */
 159 struct tseg_qent {
 160         LIST_ENTRY(tseg_qent) tqe_q;
 161         int     tqe_len;                /* TCP segment data length */
 162         struct  tcphdr *tqe_th;         /* a pointer to tcp header */
 163         struct  mbuf    *tqe_m;         /* mbuf contains packet */
 164 };
 165 LIST_HEAD(tsegqe_head, tseg_qent);
 166
 167 struct sackblk {
 168         tcp_seq start;          /* start seq no. of sack block */
 169         tcp_seq end;            /* end seq no. */
 170 };
 171
 172 struct sackhole {
 173         tcp_seq start;          /* start seq no. of hole */
 174         tcp_seq end;            /* end seq no. */
 175         tcp_seq rxmit;          /* next seq. no in hole to be retransmitted */
 176         u_int32_t rxmit_start;  /* timestamp of first retransmission */
 177         TAILQ_ENTRY(sackhole) scblink;  /* scoreboard linkage */
 178 };
 179
 180 struct sackhint {
 181         struct sackhole *nexthole;
 182         int     sack_bytes_rexmit;
 183 };
 184
 185 struct tcp_rxt_seg {
 186         tcp_seq rx_start;
 187         tcp_seq rx_end;
 188         u_int16_t rx_count;
 189         u_int16_t rx_flags;
 190 #define TCP_RXT_SPURIOUS        0x1     /* received DSACK notification */
 191 #define TCP_RXT_DSACK_FOR_TLP   0x2
 192         SLIST_ENTRY(tcp_rxt_seg) rx_link;
 193 };
 194
 195 struct tcptemp {
 196         u_char  tt_ipgen[40]; /* the size must be of max ip header, now IPv6 */
 197         struct  tcphdr tt_t;
 198 };
 199
 200 struct bwmeas {
 201         tcp_seq bw_start;               /* start of bw measurement */
 202         uint32_t bw_ts;         /* timestamp when bw measurement started */
 203         uint32_t bw_size;               /* burst size in bytes for this bw measurement */
 204         uint32_t bw_minsizepkts;        /* Min burst size as segments */
 205         uint32_t bw_maxsizepkts;        /* Max burst size as segments */
 206         uint32_t bw_minsize;    /* Min size in bytes */
 207         uint32_t bw_maxsize;    /* Max size in bytes */
 208         uint32_t bw_sndbw;              /* Measured send bw */
 209 };
 210
 211 /* MPTCP Data sequence map entry */
 212 struct mpt_dsn_map {
 213         uint64_t                mpt_dsn;        /* data seq num recvd */
 214         uint32_t                mpt_sseq;       /* relative subflow # */
 215         uint16_t                mpt_len;        /* length of mapping */
 216         uint16_t                mpt_csum;       /* checksum value if on */
 217 };
 218 #define tcp6cb          tcpcb  /* for KAME src sync over BSD*'s */
 219
 220 struct tcp_ccstate {
 221         union {
 222                 struct tcp_cubic_state {
 223                         u_int32_t tc_last_max;  /* cwnd at last loss */
 224                         u_int32_t tc_epoch_start; /* TS of last loss */
 225                         u_int32_t tc_origin_point; /* window at the start of an epoch */
 226                         u_int32_t tc_tcp_win; /* computed tcp win */
 227                         u_int32_t tc_tcp_bytes_acked; /* bytes acked */
 228                         u_int32_t tc_target_win; /* cubic target win */
 229                         u_int32_t tc_avg_lastmax; /* Average of last max */
 230                         u_int32_t tc_mean_deviation; /* Mean absolute deviation */
 231                         float     tc_epoch_period; /* K parameter */
 232                 } _cubic_state_;
 233 #define cub_last_max __u__._cubic_state_.tc_last_max
 234 #define cub_epoch_start __u__._cubic_state_.tc_epoch_start
 235 #define cub_origin_point __u__._cubic_state_.tc_origin_point
 236 #define cub_tcp_win __u__._cubic_state_.tc_tcp_win
 237 #define cub_tcp_bytes_acked __u__._cubic_state_.tc_tcp_bytes_acked
 238 #define cub_epoch_period __u__._cubic_state_.tc_epoch_period
 239 #define cub_target_win __u__._cubic_state_.tc_target_win
 240 #define cub_avg_lastmax __u__._cubic_state_.tc_avg_lastmax
 241 #define cub_mean_dev __u__._cubic_state_.tc_mean_deviation
 242         } __u__;
 243 };
 244
 245 /*
 246  * Tcp control block, one per tcp; fields:
 247  * Organized for 16 byte cacheline efficiency.
 248  */
 249 struct tcpcb {
 250         struct  tsegqe_head t_segq;
 251         int     t_dupacks;              /* consecutive dup acks recd */
 252         int     t_state;                /* state of this connection */
 253         uint32_t t_timer[TCPT_NTIMERS]; /* tcp timers */
 254         struct tcptimerentry tentry;    /* entry in timer list */
 255
 256         struct  inpcb *t_inpcb;         /* back pointer to internet pcb */
 257         uint32_t        t_flags;
 258 #define TF_ACKNOW       0x00001         /* ack peer immediately */
 259 #define TF_DELACK       0x00002         /* ack, but try to delay it */
 260 #define TF_NODELAY      0x00004         /* don't delay packets to coalesce */
 261 #define TF_NOOPT        0x00008         /* don't use tcp options */
 262 #define TF_SENTFIN      0x00010         /* have sent FIN */
 263 #define TF_REQ_SCALE    0x00020         /* have/will request window scaling */
 264 #define TF_RCVD_SCALE   0x00040         /* other side has requested scaling */
 265 #define TF_REQ_TSTMP    0x00080         /* have/will request timestamps */
 266 #define TF_RCVD_TSTMP   0x00100         /* a timestamp was received in SYN */
 267 #define TF_SACK_PERMIT  0x00200         /* other side said I could SACK */
 268 #define TF_NEEDSYN      0x00400         /* send SYN (implicit state) */
 269 #define TF_NEEDFIN      0x00800         /* send FIN (implicit state) */
 270 #define TF_NOPUSH       0x01000         /* don't push */
 271 #define TF_REQ_CC       0x02000         /* have/will request CC */
 272 #define TF_RCVD_CC      0x04000         /* a CC was received in SYN */
 273 #define TF_SENDCCNEW    0x08000         /* send CCnew instead of CC in SYN */
 274 #define TF_MORETOCOME   0x10000         /* More data to be appended to sock */
 275 #define TF_LOCAL        0x20000         /* connection to a host on local link */
 276 #define TF_RXWIN0SENT   0x40000         /* sent a receiver win 0 in response */
 277 #define TF_SLOWLINK     0x80000         /* route is a on a modem speed link */
 278 #define TF_LASTIDLE     0x100000        /* connection was previously idle */
 279 #define TF_FASTRECOVERY 0x200000        /* in NewReno Fast Recovery */
 280 #define TF_WASFRECOVERY 0x400000        /* was in NewReno Fast Recovery */
 281 #define TF_SIGNATURE    0x800000        /* require MD5 digests (RFC2385) */
 282 #define TF_MAXSEGSNT    0x1000000       /* last segment sent was a full segment */
 283 #define TF_ENABLE_ECN   0x2000000       /* Enable ECN */
 284 #define TF_PMTUD        0x4000000       /* Perform Path MTU Discovery for this connection */
 285 #define TF_CLOSING      0x8000000       /* pending tcp close */
 286 #define TF_TSO          0x10000000      /* TCP Segment Offloading is enable on this connection */
 287 #define TF_BLACKHOLE    0x20000000      /* Path MTU Discovery Black Hole detection */
 288 #define TF_TIMER_ONLIST 0x40000000      /* pcb is on tcp_timer_list */
 289 #define TF_STRETCHACK   0x80000000      /* receiver is going to delay acks */
 290
 291         tcp_seq snd_una;                /* send unacknowledged */
 292         tcp_seq snd_max;                /* highest sequence number sent;
 293                                          * used to recognize retransmits
 294                                          */
 295         tcp_seq snd_nxt;                /* send next */
 296         tcp_seq snd_up;                 /* send urgent pointer */
 297
 298         tcp_seq snd_wl1;                /* window update seg seq number */
 299         tcp_seq snd_wl2;                /* window update seg ack number */
 300         tcp_seq iss;                    /* initial send sequence number */
 301         tcp_seq irs;                    /* initial receive sequence number */
 302
 303         tcp_seq rcv_nxt;                /* receive next */
 304         tcp_seq rcv_adv;                /* advertised window */
 305         u_int32_t       rcv_wnd;                /* receive window */
 306         tcp_seq rcv_up;                 /* receive urgent pointer */
 307
 308         u_int32_t       snd_wnd;        /* send window */
 309         u_int32_t       snd_cwnd;       /* congestion-controlled window */
 310         u_int32_t       snd_ssthresh;   /* snd_cwnd size threshold for
 311                                          * for slow start exponential to
 312                                          * linear switch
 313                                          */
 314         tcp_seq snd_recover;            /* for use in NewReno Fast Recovery */
 315
 316         u_int   t_maxopd;               /* mss plus options */
 317
 318         u_int32_t       t_rcvtime;      /* time at which a packet was received */
 319         u_int32_t       t_starttime;    /* time connection was established */
 320         int     t_rtttime;              /* tcp clock when rtt calculation was started */
 321         tcp_seq t_rtseq;                /* sequence number being timed */
 322
 323         u_int32_t rfbuf_ts;             /* recv buffer autoscaling timestamp */
 324         u_int32_t rfbuf_cnt;            /* recv buffer autoscaling byte count */
 325
 326         int     t_rxtcur;               /* current retransmit value (ticks) */
 327         u_int   t_maxseg;               /* maximum segment size */
 328         int     t_srtt;                 /* smoothed round-trip time */
 329         int     t_rttvar;               /* variance in round-trip time */
 330
 331         u_int64_t       t_accsleep_ms;  /* accumulated sleep time since last boot */
 332         u_int16_t       t_reassqlen;    /* length of reassembly queue */
 333         u_int16_t       t_rxtshift;     /* log(2) of rexmt exp. backoff */
 334         u_int32_t       t_rttmin;       /* minimum rtt allowed */
 335         u_int32_t       t_rttbest;      /* best rtt we've seen */
 336         u_int32_t       t_rttcur;       /* most recent value of rtt */
 337         u_int32_t       t_rttupdated;   /* number of times rtt sampled */
 338         u_int32_t       t_rxt_conndroptime;     /* retxmt conn gets dropped after this time, when set */
 339         u_int32_t       t_rxtstart;     /* time at which retransmission started */
 340         u_int32_t       max_sndwnd;     /* largest window peer has offered */
 341
 342         int     t_softerror;            /* possible error not yet reported */
 343 /* out-of-band data */
 344         char    t_oobflags;             /* have some */
 345         char    t_iobc;                 /* input character */
 346 #define TCPOOB_HAVEDATA 0x01
 347 #define TCPOOB_HADDATA  0x02
 348 /* RFC 1323 variables */
 349         u_int8_t        snd_scale;      /* window scaling for send window */
 350         u_int8_t        rcv_scale;      /* window scaling for recv window */
 351         u_int8_t        request_r_scale; /* pending window scaling */
 352         u_int8_t        requested_s_scale;
 353         u_int8_t        tcp_cc_index;   /* index of congestion control algorithm */
 354         u_int8_t        t_adaptive_rtimo;       /* Read timeout used as a multiple of RTT */
 355         u_int8_t        t_adaptive_wtimo;       /* Write timeout used as a multiple of RTT */
 356         u_int8_t        t_stretchack_delayed;   /* stretch ack delayed */
 357
 358 /* State for limiting early retransmits when SACK is not enabled */
 359         u_int16_t       t_early_rexmt_count; /* count of early rexmts */
 360         u_int32_t       t_early_rexmt_win; /* window for limiting early rexmts */
 361
 362         u_int32_t       ts_recent;      /* timestamp echo data */
 363
 364         u_int32_t       ts_recent_age;  /* when last updated */
 365         tcp_seq last_ack_sent;
 366 /* RFC 1644 variables */
 367         tcp_cc  cc_send;                /* send connection count */
 368         tcp_cc  cc_recv;                /* receive connection count */
 369 /* RFC 3465 variables */
 370         u_int32_t       t_bytes_acked;  /* ABC "bytes_acked" parameter */
 371
 372         int     t_lastchain;            /* amount of packets chained last time around */
 373         u_int16_t       t_unacksegs;    /* received but unacked segments for delaying acks */
 374         u_int8_t        t_rexmtthresh;  /* duplicate ack threshold for entering fast recovery */
 375         u_int8_t        t_rtimo_probes; /* number of adaptive rtimo probes sent */
 376         u_int32_t       t_persist_timeout; /* ZWP persistence limit as set by PERSIST_TIMEOUT */
 377         u_int32_t       t_persist_stop;         /* persistence limit deadline if triggered by ZWP */
 378         u_int32_t       t_notsent_lowat;        /* Low water for not sent data */
 379
 380 /* Receiver state for stretch-ack algorithm */
 381         u_int32_t       rcv_unackwin;   /* to measure win for stretching acks */
 382         u_int32_t       rcv_by_unackwin; /* bytes seen during the last ack-stretching win */
 383         u_int32_t       rcv_nostrack_ts; /* timestamp when stretch ack was disabled automatically */
 384         u_int16_t       rcv_waitforss;  /* wait for packets during slow-start */
 385         u_int16_t               ecn_flags;
 386 #define TE_SETUPSENT            0x01    /* Indicate we have sent ECN-SETUP SYN or SYN-ACK */
 387 #define TE_SETUPRECEIVED        0x02    /* Indicate we have received ECN-SETUP SYN or SYN-ACK */
 388 #define TE_SENDIPECT            0x04    /* Indicate we haven't sent or received non-ECN-setup SYN or SYN-ACK */
 389 #define TE_SENDCWR              0x08    /* Indicate that the next non-retransmit should have the TCP CWR flag set */
 390 #define TE_SENDECE              0x10    /* Indicate that the next packet should have the TCP ECE flag set */
 391 #define TE_INRECOVERY           0x20    /* connection entered recovery after receiving ECE */
 392 #define TE_RECV_ECN_CE          0x40    /* Received IPTOS_ECN_CE marking atleast once */
 393 #define TE_RECV_ECN_ECE 0x80    /* Received ECE marking atleast once */
 394 #define TE_ECN_ON               (TE_SETUPSENT | TE_SETUPRECEIVED) /* Indicate ECN was successfully negotiated on a connection) */
 395
 396 /* state for bad retransmit recovery */
 397         u_int32_t       snd_cwnd_prev;  /* cwnd prior to retransmit */
 398         u_int32_t       snd_ssthresh_prev; /* ssthresh prior to retransmit */
 399         tcp_seq snd_recover_prev;       /* snd_recover prior to retransmit */
 400         int     t_srtt_prev;            /* srtt prior to retransmit */
 401         int     t_rttvar_prev;          /* rttvar prior to retransmit */
 402         u_int32_t       t_badrexmt_time; /* bad rexmt detection time */
 403
 404 /* Packet reordering metric */
 405         u_int16_t       t_reorderwin; /* Reordering late time offset */
 406
 407 /* SACK related state */
 408         int16_t snd_numholes;           /* number of holes seen by sender */
 409         tcp_seq sack_newdata;           /* New data xmitted in this recovery
 410                                            episode starts at this seq number */
 411         TAILQ_HEAD(sackhole_head, sackhole) snd_holes;
 412                                                 /* SACK scoreboard (sorted) */
 413         tcp_seq snd_fack;               /* last seq number(+1) sack'd by rcv'r*/
 414         int     rcv_numsacks;           /* # distinct sack blks present */
 415         struct sackblk sackblks[MAX_SACK_BLKS]; /* seq nos. of sack blocks */
 416         struct sackhint sackhint;       /* SACK scoreboard hint */
 417
 418         struct mbuf     *t_pktlist_head; /* First packet in transmit chain */
 419         struct mbuf     *t_pktlist_tail; /* Last packet in transmit chain */
 420         u_int32_t       t_pktlist_sentlen; /* total bytes in transmit chain */
 421
 422         u_int32_t       t_keepidle;     /* keepalive idle timer (override global if > 0) */
 423         u_int32_t       t_keepinit;     /* connection timeout, i.e. idle time
 424                                         in SYN_SENT or SYN_RECV state */
 425         u_int32_t       t_keepintvl;    /* interval between keepalives */
 426         u_int32_t       t_keepcnt;      /* number of keepalives before close */
 427
 428         u_int32_t       tso_max_segment_size;   /* TSO maximum segment unit for NIC */
 429         u_int16_t       t_pmtud_lastseg_size;   /* size of the last sent segment */
 430         u_int16_t       t_pmtud_saved_maxopd;   /* MSS saved before performing PMTU-D BlackHole detection */
 431         u_int32_t       t_pmtud_start_ts;       /* Time of PMTUD blackhole detection */
 432
 433         struct
 434         {
 435                 u_int32_t       rxduplicatebytes;
 436                 u_int32_t       rxoutoforderbytes;
 437                 u_int32_t       txretransmitbytes;
 438                 u_int8_t        synrxtshift;
 439                 u_int8_t        unused;
 440                 u_int16_t       unused_pad_to_8;
 441         } t_stat;
 442
 443         /* Background congestion related state */
 444         uint32_t        rtt_hist[N_RTT_BASE];   /* history of minimum RTT */
 445         uint32_t        rtt_count;              /* Number of RTT samples in recent base history */
 446         uint32_t        bg_ssthresh;            /* Slow start threshold until delay increases */
 447         uint32_t        t_flagsext;             /* Another field to accommodate more flags */
 448 #define TF_RXTFINDROP   0x1                     /* Drop conn after retransmitting FIN 3 times */
 449 #define TF_RCVUNACK_WAITSS      0x2             /* set when the receiver should not stretch acks */
 450 #define TF_BWMEAS_INPROGRESS    0x4             /* Indicate BW meas is happening */
 451 #define TF_MEASURESNDBW         0x8             /* Measure send bw on this connection */
 452 #define TF_LRO_OFFLOADED        0x10            /* Connection LRO offloaded */
 453 #define TF_SACK_ENABLE          0x20            /* SACK is enabled */
 454 #define TF_RECOMPUTE_RTT        0x40            /* recompute RTT after spurious retransmit */
 455 #define TF_DETECT_READSTALL     0x80            /* Used to detect a stall during read operation */
 456 #define TF_RECV_THROTTLE        0x100           /* Input throttling active */
 457 #define TF_NOSTRETCHACK         0x200           /* ack every other packet */
 458 #define TF_STREAMEOW            0x400           /* Last packet was small indicating end of write */
 459 #define TF_NOTIMEWAIT           0x800           /* Avoid going into time-wait */
 460 #define TF_SENT_TLPROBE         0x1000          /* Sent data in PTO */
 461 #define TF_PKTS_REORDERED       0x2000          /* Detected reordering */
 462 #define TF_DELAY_RECOVERY       0x4000          /* delay fast recovery */
 463 #define TF_FORCE                0x8000          /* force 1 byte out */
 464 #define TF_DISABLE_STRETCHACK   0x10000         /* auto-disable stretch ack */
 465 #define TF_NOBLACKHOLE_DETECTION 0x20000        /* Disable PMTU blackhole detection */
 466 #define TF_DISABLE_DSACK        0x40000         /* Ignore DSACK due to n/w duplication */
 467 #define TF_RESCUE_RXT           0x80000         /* SACK rescue retransmit */
 468 #define TF_CWND_NONVALIDATED    0x100000        /* cwnd non validated */
 469 #define TF_PROBING              0x200000        /* Trigger probe timeout */
 470 #define TF_FASTOPEN             0x400000        /* TCP Fastopen is enabled */
 471
 472 #if TRAFFIC_MGT
 473         /* Inter-arrival jitter related state */
 474         uint32_t        iaj_rcv_ts;             /* tcp clock when the first packet was received */
 475         uint16_t        iaj_size;               /* Size of packet for iaj measurement */
 476         uint8_t         iaj_small_pkt;          /* Count of packets smaller than iaj_size */
 477         uint8_t         t_pipeack_ind;          /* index for next pipeack sample */
 478         uint16_t        iaj_pktcnt;             /* packet count, to avoid throttling initially */
 479         uint16_t        acc_iaj;                /* Accumulated iaj */
 480         tcp_seq         iaj_rwintop;            /* recent max advertised window */
 481         uint32_t        avg_iaj;                /* Mean */
 482         uint32_t        std_dev_iaj;            /* Standard deviation */
 483 #endif /* TRAFFIC_MGT */
 484         struct bwmeas   *t_bwmeas;              /* State for bandwidth measurement */
 485         uint32_t        t_lropktlen;            /* Bytes in a LRO frame */
 486         tcp_seq         t_idleat;               /* rcv_nxt at idle time */
 487         TAILQ_ENTRY(tcpcb) t_twentry;           /* link for time wait queue */
 488         struct tcp_ccstate      *t_ccstate;     /* congestion control related state */
 489 /* Tail loss probe related state */
 490         tcp_seq         t_tlphighrxt;           /* snd_nxt after PTO */
 491         u_int32_t       t_tlpstart;             /* timestamp at PTO */
 492 /* DSACK data receiver state */
 493         tcp_seq         t_dsack_lseq;           /* DSACK left sequence */
 494         tcp_seq         t_dsack_rseq;           /* DSACK right sequence */
 495 /* DSACK data sender state */
 496         SLIST_HEAD(tcp_rxt_seghead, tcp_rxt_seg) t_rxt_segments;
 497         tcp_seq         t_dsack_lastuna;        /* snd_una when last recovery episode started */
 498 /* state for congestion window validation (draft-ietf-tcpm-newcwv-07) */
 499 #define TCP_PIPEACK_SAMPLE_COUNT        3
 500         u_int32_t       t_pipeack_sample[TCP_PIPEACK_SAMPLE_COUNT];     /* pipeack, bytes acked within RTT */
 501         tcp_seq         t_pipeack_lastuna; /* una when pipeack measurement started */
 502         u_int32_t       t_pipeack;
 503         u_int32_t       t_lossflightsize;
 504
 505 #if MPTCP
 506         u_int32_t       t_mpflags;              /* flags for multipath TCP */
 507
 508 #define TMPF_PREESTABLISHED     0x00000001 /* conn in pre-established state */
 509 #define TMPF_SENT_KEYS          0x00000002 /* indicates that keys were sent */
 510 #define TMPF_MPTCP_TRUE         0x00000004 /* negotiated MPTCP successfully */
 511 #define TMPF_MPTCP_RCVD_KEY     0x00000008 /* state for 3-way handshake */
 512 #define TMPF_SND_MPPRIO         0x00000010 /* send priority of subflow */
 513 #define TMPF_SND_REM_ADDR       0x00000020 /* initiate address removal */
 514 #define TMPF_UNUSED             0x00000040 /* address addition acked by peer */
 515 #define TMPF_JOINED_FLOW        0x00000080 /* Indicates additional flow */
 516 #define TMPF_BACKUP_PATH        0x00000100 /* Indicates backup path */
 517 #define TMPF_MPTCP_ACKNOW       0x00000200 /* Send Data ACK */
 518 #define TMPF_SEND_DSN           0x00000400 /* Send DSN mapping */
 519 #define TMPF_SEND_DFIN          0x00000800 /* Send Data FIN */
 520 #define TMPF_RECV_DFIN          0x00001000 /* Recv Data FIN */
 521 #define TMPF_SENT_JOIN          0x00002000 /* Sent Join */
 522 #define TMPF_RECVD_JOIN         0x00004000 /* Received Join */
 523 #define TMPF_RESET              0x00008000 /* Send RST */
 524 #define TMPF_TCP_FALLBACK       0x00010000 /* Fallback to TCP */
 525 #define TMPF_FASTCLOSE          0x00020000 /* Send Fastclose option */
 526 #define TMPF_EMBED_DSN          0x00040000 /* tp has DSN mapping */
 527 #define TMPF_MPTCP_READY        0x00080000 /* Can send DSS options on data */
 528 #define TMPF_INFIN_SENT         0x00100000 /* Sent infinite mapping */
 529 #define TMPF_SND_MPFAIL         0x00200000 /* Received mapping csum failure */
 530 #define TMPF_FASTJOIN_SEND      0x00400000 /* Fast join early data send */
 531 #define TMPF_FASTJOINBY2_SEND   0x00800000 /* Fast join send after 3 WHS */
 532 #define TMPF_MPCAP_RETRANSMIT   0x01000000 /* Retransmission of 3rd ACK */
 533
 534         tcp_seq                 t_mpuna;        /* unacknowledged sequence */
 535         void                    *t_mptcb;       /* pointer to MPTCP TCB */
 536         struct mpt_dsn_map      t_rcv_map;      /* Receive mapping list */
 537         u_int8_t                t_local_aid;    /* Addr Id for authentication */
 538         u_int8_t                t_rem_aid;      /* Addr ID of another subflow */
 539         u_int8_t                t_mprxtshift;   /* join retransmission */
 540 #endif /* MPTCP */
 541
 542 #define TFO_F_OFFER_COOKIE      0x01 /* We will offer a cookie */
 543 #define TFO_F_COOKIE_VALID      0x02 /* The received cookie is valid */
 544 #define TFO_F_COOKIE_REQ        0x04 /* Client requested a new cookie */
 545 #define TFO_F_COOKIE_SENT       0x08 /* Client did send a cookie in the SYN */
 546 #define TFO_F_SYN_LOSS          0x10 /* A SYN-loss triggered a fallback to regular TCP on the client-side */
 547 #define TFO_F_NO_RCVPROBING     0x20 /* This network is guaranteed to support TFO in the downstream direction */
 548 #define TFO_F_NO_SNDPROBING     0x40 /* This network is guaranteed to support TFO in the upstream direction */
 549         u_int8_t                t_tfo_flags;
 550 #define TFO_S_SYNDATA_RCV       0x01 /* SYN+data has been received */
 551 #define TFO_S_COOKIEREQ_RECV    0x02 /* TFO-cookie request received */
 552 #define TFO_S_COOKIE_SENT       0x04 /* TFO-cookie announced in SYN/ACK */
 553 #define TFO_S_COOKIE_INVALID    0x08 /* Received TFO-cookie is invalid */
 554 #define TFO_S_COOKIE_REQ        0x10 /* TFO-cookie requested within the SYN */
 555 #define TFO_S_COOKIE_RCV        0x20 /* TFO-cookie received in SYN/ACK */
 556 #define TFO_S_SYN_DATA_SENT     0x40 /* SYN+data sent */
 557 #define TFO_S_SYN_DATA_ACKED    0x80 /* SYN+data has been acknowledged in SYN/ACK */
 558 #define TFO_S_SYN_LOSS          0x0100 /* SYN+TFO has been lost - fallback to regular TCP */
 559         u_int16_t               t_tfo_stats;
 560
 561         u_int8_t                t_tfo_probes; /* TFO-probes we did send */
 562 /*
 563  * This here is the TFO-probing state-machine. Transitions are as follows:
 564  *
 565  * Current state: PROBE_NONE
 566  *                Event: SYN+DATA acknowledged
 567  *                       Action: Transition to PROBE_PROBING and set keepalive-timer
 568  *
 569  * Current state: PROBE_PROBING (initial state)
 570  *                Event: Receive data
 571  *                       Action: Transition to PROBE_NONE and cancel keepalive-timer
 572  *                Event: Receive ACK that does not indicate a hole
 573  *                       Action: Transition to PROBE_NONE and cancel keepalive-timer
 574  *                Event: Receive ACK that indicates a hole
 575  *                       Action: Transition to PROBE_WAIT_DATA and set a short timer
 576  *                               to wait for the final segment.
 577  *                Event: Keepalive-timeout (did not receive any segment)
 578  *                       Action: Signal ETIMEDOUT as with regular keepalive-timers
 579  *
 580  * Current state: PROBE_WAIT_DATA
 581  *                Event: Receive data
 582  *                       Action: Transition to PROBE_NONE and cancel keepalive-timer
 583  *                Event: Data-timeout (did not receive the expected data)
 584  *                       Action: Signal ENODATA up to the app and close everything.
 585  */
 586 #define TFO_PROBE_NONE          0 /* Not probing now */
 587 #define TFO_PROBE_PROBING       1 /* Sending out TCP-keepalives waiting for reply */
 588 #define TFO_PROBE_WAIT_DATA     2 /* Received reply, waiting for data */
 589         u_int8_t                t_tfo_probe_state;
 590 };
 591
 592 #define IN_FASTRECOVERY(tp)     (tp->t_flags & TF_FASTRECOVERY)
 593 #define SACK_ENABLED(tp)        (tp->t_flagsext & TF_SACK_ENABLE)
 594
 595 /*
 596  * If the connection is in a throttled state due to advisory feedback from
 597  * the interface output queue, reset that state. We do this in favor
 598  * of entering recovery because the data transfer during recovery
 599  * should be just a trickle and it will help to improve performance.
 600  * We also do not want to back off twice in the same RTT.
 601  */
 602 #define ENTER_FASTRECOVERY(_tp_) do {                           \
 603         (_tp_)->t_flags |= TF_FASTRECOVERY;                     \
 604         if (INP_IS_FLOW_CONTROLLED((_tp_)->t_inpcb))            \
 605                 inp_reset_fc_state((_tp_)->t_inpcb);            \
 606         if (!SLIST_EMPTY(&tp->t_rxt_segments))                  \
 607                 tcp_rxtseg_clean(tp);                           \
 608 } while(0)
 609
 610 #define EXIT_FASTRECOVERY(_tp_) do {            \
 611         (_tp_)->t_flags &= ~TF_FASTRECOVERY;    \
 612         (_tp_)->t_dupacks = 0;                  \
 613         (_tp_)->t_rexmtthresh = tcprexmtthresh; \
 614         (_tp_)->t_bytes_acked = 0;              \
 615         (_tp_)->ecn_flags &= ~TE_INRECOVERY;    \
 616         (_tp_)->t_timer[TCPT_PTO] = 0;          \
 617         (_tp_)->t_flagsext &= ~TF_RESCUE_RXT;   \
 618         (_tp_)->t_lossflightsize = 0;           \
 619 } while(0)
 620
 621 /*
 622  * When the number of duplicate acks received is less than
 623  * the retransmit threshold, use Limited Transmit algorithm
 624  */
 625 extern int tcprexmtthresh;
 626 #define ALLOW_LIMITED_TRANSMIT(_tp_) \
 627         ((_tp_)->t_dupacks > 0 && \
 628         (_tp_)->t_dupacks < (_tp_)->t_rexmtthresh && \
 629         ((_tp_)->t_flagsext & (TF_PKTS_REORDERED|TF_DELAY_RECOVERY)) \
 630             != (TF_PKTS_REORDERED|TF_DELAY_RECOVERY))
 631
 632 /*
 633  * This condition is true is timestamp option is supported
 634  * on a connection.
 635  */
 636 #define TSTMP_SUPPORTED(_tp_) \
 637         (((_tp_)->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP)) == \
 638                 (TF_REQ_TSTMP|TF_RCVD_TSTMP))
 639
 640 /*
 641  * This condition is true if window scale option is supported
 642  * on a connection
 643  */
 644 #define TCP_WINDOW_SCALE_ENABLED(_tp_) \
 645         (((_tp_)->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == \
 646                 (TF_RCVD_SCALE|TF_REQ_SCALE))
 647
 648 /* Is ECN enabled end-to-end */
 649 #define TCP_ECN_ENABLED(_tp_) \
 650         (((_tp_)->ecn_flags & (TE_ECN_ON)) == (TE_ECN_ON))
 651
 652 /*
 653  * Gives number of bytes acked by this ack
 654  */
 655 #define BYTES_ACKED(_th_, _tp_) \
 656         ((_th_)->th_ack - (_tp_)->snd_una)
 657
 658 /* Returns true if a DSACK option should be sent */
 659 #define TCP_SEND_DSACK_OPT(_tp_) \
 660         ((_tp_)->t_dsack_lseq > 0 && (_tp_)->t_dsack_rseq > 0)
 661
 662 /* Check if DSACK option should be processed */
 663 #define TCP_DSACK_ENABLED(tp)   (tcp_dsack_enable == 1 && \
 664     !(tp->t_flagsext & TF_DISABLE_DSACK))
 665
 666 /*
 667  * Returns true if a DSACK sequence is within the max send window that will
 668  * be accepted. In order to set a window to validate sequence numbers, the
 669  * max send window within which a DSACK option is processed is limited.
 670  *
 671  * We need to choose a maximum window to check if the sequence number is
 672  * within the window. One arbitrary choice is 256 * MSS because if the
 673  * window is as large as 256 segments it might be big enough to ignore the
 674  * DSACK option. Choosing a much larger limit means that the memory for
 675  * retransmit segments can be held for a longer time.
 676  */
 677 #define TCP_DSACK_MAX_SEND_WINDOW(_tp_) ((_tp_)->t_maxseg << 8)
 678 #define TCP_DSACK_SEQ_IN_WINDOW(_tp_, _seq_, _una_) \
 679     (SEQ_LEQ((_seq_), (_tp_)->snd_max) && \
 680     SEQ_GEQ((_seq_), ((_una_) - TCP_DSACK_MAX_SEND_WINDOW(_tp_))))
 681
 682
 683 enum tcp_cc_event {
 684         TCP_CC_CWND_INIT,       /* 0 */
 685         TCP_CC_INSEQ_ACK_RCVD,  /* 1 */
 686         TCP_CC_ACK_RCVD,        /* 2 */
 687         TCP_CC_ENTER_FASTRECOVERY, /* 3 */
 688         TCP_CC_IN_FASTRECOVERY, /* 4 */
 689         TCP_CC_EXIT_FASTRECOVERY,  /* 5 */
 690         TCP_CC_PARTIAL_ACK,     /* 6 */
 691         TCP_CC_IDLE_TIMEOUT,    /* 7 */
 692         TCP_CC_REXMT_TIMEOUT,   /* 8 */
 693         TCP_CC_ECN_RCVD,        /* 9 */
 694         TCP_CC_BAD_REXMT_RECOVERY, /* 10 */
 695         TCP_CC_OUTPUT_ERROR,    /* 11 */
 696         TCP_CC_CHANGE_ALGO,     /* 12 */
 697         TCP_CC_FLOW_CONTROL,    /* 13 */
 698         TCP_CC_SUSPEND,         /* 14 */
 699         TCP_CC_LIMITED_TRANSMIT, /* 15 */
 700         TCP_CC_EARLY_RETRANSMIT, /* 16 */
 701         TCP_CC_TLP_RECOVERY,    /* 17 */
 702         TCP_CC_TLP_RECOVER_LASTPACKET, /* 18 */
 703         TCP_CC_DELAY_FASTRECOVERY, /* 19 */
 704         TCP_CC_TLP_IN_FASTRECOVERY, /* 20 */
 705         TCP_CC_DSACK_BAD_REXMT  /* 21 */
 706 };
 707
 708 /*
 709  * Structure to hold TCP options that are only used during segment
 710  * processing (in tcp_input), but not held in the tcpcb.
 711  * It's basically used to reduce the number of parameters
 712  * to tcp_dooptions.
 713  */
 714 struct tcpopt {
 715         u_int32_t       to_flags;               /* which options are present */
 716 #define TOF_TS          0x0001          /* timestamp */
 717 #define TOF_MSS         0x0010
 718 #define TOF_SCALE       0x0020
 719 #define TOF_SIGNATURE   0x0040  /* signature option present */
 720 #define TOF_SIGLEN      0x0080  /* signature length valid (RFC2385) */
 721 #define TOF_SACK        0x0100          /* Peer sent SACK option */
 722 #define TOF_MPTCP       0x0200  /* MPTCP options to be dropped */
 723 #define TOF_TFO         0x0400  /* TFO cookie option present */
 724 #define TOF_TFOREQ      0x0800  /* TFO cookie request present */
 725         u_int32_t               to_tsval;
 726         u_int32_t               to_tsecr;
 727         u_int16_t       to_mss;
 728         u_int8_t        to_requested_s_scale;
 729         u_int8_t        to_nsacks;      /* number of SACK blocks */
 730         u_char          *to_sacks;      /* pointer to the first SACK blocks */
 731         u_char          *to_tfo;        /* pointer to the TFO cookie */
 732 };
 733
 734 /*
 735  * The TAO cache entry which is stored in the protocol family specific
 736  * portion of the route metrics.
 737  */
 738 struct rmxp_tao {
 739         tcp_cc  tao_cc;                 /* latest CC in valid SYN */
 740         tcp_cc  tao_ccsent;             /* latest CC sent to peer */
 741         u_short tao_mssopt;             /* peer's cached MSS */
 742 #ifdef notyet
 743         u_short tao_flags;              /* cache status flags */
 744 #define TAOF_DONT       0x0001          /* peer doesn't understand rfc1644 */
 745 #define TAOF_OK         0x0002          /* peer does understand rfc1644 */
 746 #define TAOF_UNDEF      0               /* we don't know yet */
 747 #endif /* notyet */
 748 };
 749 #define rmx_taop(r)     ((struct rmxp_tao *)(r).rmx_filler)
 750
 751 #define intotcpcb(ip)   ((struct tcpcb *)(ip)->inp_ppcb)
 752 #define sototcpcb(so)   (intotcpcb(sotoinpcb(so)))
 753
 754 /*
 755  * The rtt measured is in milliseconds as the timestamp granularity is
 756  * a millisecond. The smoothed round-trip time and estimated variance
 757  * are stored as fixed point numbers scaled by the values below.
 758  * For convenience, these scales are also used in smoothing the average
 759  * (smoothed = (1/scale)sample + ((scale-1)/scale)smoothed).
 760  * With these scales, srtt has 5 bits to the right of the binary point,
 761  * and thus an "ALPHA" of 0.875.  rttvar has 4 bits to the right of the
 762  * binary point, and is smoothed with an ALPHA of 0.75.
 763  */
 764 #define TCP_RTT_SCALE           32      /* multiplier for srtt; 3 bits frac. */
 765 #define TCP_RTT_SHIFT           5       /* shift for srtt; 5 bits frac. */
 766 #define TCP_RTTVAR_SCALE        16      /* multiplier for rttvar; 4 bits */
 767 #define TCP_RTTVAR_SHIFT        4       /* shift for rttvar; 4 bits */
 768 #define TCP_DELTA_SHIFT         2       /* see tcp_input.c */
 769
 770 /* TFO-specific defines */
 771 #define TFO_COOKIE_LEN_MIN      4
 772 #define TFO_COOKIE_LEN_DEFAULT  8
 773 #define TFO_COOKIE_LEN_MAX      16
 774
 775 /*
 776  * The initial retransmission should happen at rtt + 4 * rttvar.
 777  * Because of the way we do the smoothing, srtt and rttvar
 778  * will each average +1/2 tick of bias.  When we compute
 779  * the retransmit timer, we want 1/2 tick of rounding and
 780  * 1 extra tick because of +-1/2 tick uncertainty in the
 781  * firing of the timer.  The bias will give us exactly the
 782  * 1.5 tick we need.  But, because the bias is
 783  * statistical, we have to test that we don't drop below
 784  * the minimum feasible timer (which is 2 ticks).
 785  * This version of the macro adapted from a paper by Lawrence
 786  * Brakmo and Larry Peterson which outlines a problem caused
 787  * by insufficient precision in the original implementation,
 788  * which results in inappropriately large RTO values for very
 789  * fast networks.
 790  */
 791 #define TCP_REXMTVAL(tp) \
 792         max((tp)->t_rttmin, (((tp)->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT))  \
 793           + (tp)->t_rttvar) >> TCP_DELTA_SHIFT)
 794
 795 /*
 796  * Jaguar compatible TCP control block, for xtcpcb
 797  * Does not have the old fields
 798  */
 799 struct otcpcb {
 800 #else
 801 struct tseg_qent;
 802 _TCPCB_LIST_HEAD(tsegqe_head, tseg_qent);
 803
 804 struct tcpcb {
 805 #endif /* KERNEL_PRIVATE */
 806 #if defined(KERNEL_PRIVATE)
 807         u_int32_t t_segq;
 808 #else
 809         struct  tsegqe_head t_segq;
 810 #endif /* KERNEL_PRIVATE */
 811         int     t_dupacks;              /* consecutive dup acks recd */
 812         u_int32_t unused;               /* unused now: was t_template */
 813
 814         int     t_timer[TCPT_NTIMERS_EXT];      /* tcp timers */
 815
 816         _TCPCB_PTR(struct inpcb *) t_inpcb;     /* back pointer to internet pcb */
 817         int     t_state;                /* state of this connection */
 818         u_int   t_flags;
 819 #define TF_ACKNOW       0x00001         /* ack peer immediately */
 820 #define TF_DELACK       0x00002         /* ack, but try to delay it */
 821 #define TF_NODELAY      0x00004         /* don't delay packets to coalesce */
 822 #define TF_NOOPT        0x00008         /* don't use tcp options */
 823 #define TF_SENTFIN      0x00010         /* have sent FIN */
 824 #define TF_REQ_SCALE    0x00020         /* have/will request window scaling */
 825 #define TF_RCVD_SCALE   0x00040         /* other side has requested scaling */
 826 #define TF_REQ_TSTMP    0x00080         /* have/will request timestamps */
 827 #define TF_RCVD_TSTMP   0x00100         /* a timestamp was received in SYN */
 828 #define TF_SACK_PERMIT  0x00200         /* other side said I could SACK */
 829 #define TF_NEEDSYN      0x00400         /* send SYN (implicit state) */
 830 #define TF_NEEDFIN      0x00800         /* send FIN (implicit state) */
 831 #define TF_NOPUSH       0x01000         /* don't push */
 832 #define TF_REQ_CC       0x02000         /* have/will request CC */
 833 #define TF_RCVD_CC      0x04000         /* a CC was received in SYN */
 834 #define TF_SENDCCNEW    0x08000         /* send CCnew instead of CC in SYN */
 835 #define TF_MORETOCOME   0x10000         /* More data to be appended to sock */
 836 #define TF_LQ_OVERFLOW  0x20000         /* listen queue overflow */
 837 #define TF_RXWIN0SENT   0x40000         /* sent a receiver win 0 in response */
 838 #define TF_SLOWLINK     0x80000         /* route is a on a modem speed link */
 839
 840         int     t_force;                /* 1 if forcing out a byte */
 841
 842         tcp_seq snd_una;                /* send unacknowledged */
 843         tcp_seq snd_max;                /* highest sequence number sent;
 844                                          * used to recognize retransmits
 845                                          */
 846         tcp_seq snd_nxt;                /* send next */
 847         tcp_seq snd_up;                 /* send urgent pointer */
 848
 849         tcp_seq snd_wl1;                /* window update seg seq number */
 850         tcp_seq snd_wl2;                /* window update seg ack number */
 851         tcp_seq iss;                    /* initial send sequence number */
 852         tcp_seq irs;                    /* initial receive sequence number */
 853
 854         tcp_seq rcv_nxt;                /* receive next */
 855         tcp_seq rcv_adv;                /* advertised window */
 856         u_int32_t rcv_wnd;              /* receive window */
 857         tcp_seq rcv_up;                 /* receive urgent pointer */
 858
 859         u_int32_t snd_wnd;              /* send window */
 860         u_int32_t snd_cwnd;             /* congestion-controlled window */
 861         u_int32_t snd_ssthresh;         /* snd_cwnd size threshold for
 862                                          * for slow start exponential to
 863                                          * linear switch
 864                                          */
 865         u_int   t_maxopd;               /* mss plus options */
 866
 867         u_int32_t t_rcvtime;            /* time at which a packet was received */
 868         u_int32_t t_starttime;          /* time connection was established */
 869         int     t_rtttime;              /* round trip time */
 870         tcp_seq t_rtseq;                /* sequence number being timed */
 871
 872         int     t_rxtcur;               /* current retransmit value (ticks) */
 873         u_int   t_maxseg;               /* maximum segment size */
 874         int     t_srtt;                 /* smoothed round-trip time */
 875         int     t_rttvar;               /* variance in round-trip time */
 876
 877         int     t_rxtshift;             /* log(2) of rexmt exp. backoff */
 878         u_int   t_rttmin;               /* minimum rtt allowed */
 879         u_int32_t t_rttupdated;         /* number of times rtt sampled */
 880         u_int32_t max_sndwnd;           /* largest window peer has offered */
 881
 882         int     t_softerror;            /* possible error not yet reported */
 883 /* out-of-band data */
 884         char    t_oobflags;             /* have some */
 885         char    t_iobc;                 /* input character */
 886 #define TCPOOB_HAVEDATA 0x01
 887 #define TCPOOB_HADDATA  0x02
 888 /* RFC 1323 variables */
 889         u_char  snd_scale;              /* window scaling for send window */
 890         u_char  rcv_scale;              /* window scaling for recv window */
 891         u_char  request_r_scale;        /* pending window scaling */
 892         u_char  requested_s_scale;
 893         u_int32_t ts_recent;            /* timestamp echo data */
 894
 895         u_int32_t ts_recent_age;        /* when last updated */
 896         tcp_seq last_ack_sent;
 897 /* RFC 1644 variables */
 898         tcp_cc  cc_send;                /* send connection count */
 899         tcp_cc  cc_recv;                /* receive connection count */
 900         tcp_seq snd_recover;            /* for use in fast recovery */
 901 /* experimental */
 902         u_int32_t snd_cwnd_prev;        /* cwnd prior to retransmit */
 903         u_int32_t snd_ssthresh_prev;    /* ssthresh prior to retransmit */
 904         u_int32_t t_badrxtwin;          /* window for retransmit recovery */
 905 };
 906
 907 #define        tcps_ecn_setup  tcps_ecn_client_success
 908 #define        tcps_sent_cwr   tcps_ecn_recv_ece
 909 #define        tcps_sent_ece   tcps_ecn_sent_ece
 910
 911 /*
 912  * TCP statistics.
 913  * Many of these should be kept per connection,
 914  * but that's inconvenient at the moment.
 915  */
 916 struct  tcpstat {
 917         u_int32_t       tcps_connattempt;       /* connections initiated */
 918         u_int32_t       tcps_accepts;           /* connections accepted */
 919         u_int32_t       tcps_connects;          /* connections established */
 920         u_int32_t       tcps_drops;             /* connections dropped */
 921         u_int32_t       tcps_conndrops;         /* embryonic connections dropped */
 922         u_int32_t       tcps_closed;            /* conn. closed (includes drops) */
 923         u_int32_t       tcps_segstimed;         /* segs where we tried to get rtt */
 924         u_int32_t       tcps_rttupdated;        /* times we succeeded */
 925         u_int32_t       tcps_delack;            /* delayed acks sent */
 926         u_int32_t       tcps_timeoutdrop;       /* conn. dropped in rxmt timeout */
 927         u_int32_t       tcps_rexmttimeo;        /* retransmit timeouts */
 928         u_int32_t       tcps_persisttimeo;      /* persist timeouts */
 929         u_int32_t       tcps_keeptimeo;         /* keepalive timeouts */
 930         u_int32_t       tcps_keepprobe;         /* keepalive probes sent */
 931         u_int32_t       tcps_keepdrops;         /* connections dropped in keepalive */
 932
 933         u_int32_t       tcps_sndtotal;          /* total packets sent */
 934         u_int32_t       tcps_sndpack;           /* data packets sent */
 935         u_int32_t       tcps_sndbyte;           /* data bytes sent */
 936         u_int32_t       tcps_sndrexmitpack;     /* data packets retransmitted */
 937         u_int32_t       tcps_sndrexmitbyte;     /* data bytes retransmitted */
 938         u_int32_t       tcps_sndacks;           /* ack-only packets sent */
 939         u_int32_t       tcps_sndprobe;          /* window probes sent */
 940         u_int32_t       tcps_sndurg;            /* packets sent with URG only */
 941         u_int32_t       tcps_sndwinup;          /* window update-only packets sent */
 942         u_int32_t       tcps_sndctrl;           /* control (SYN|FIN|RST) packets sent */
 943
 944         u_int32_t       tcps_rcvtotal;          /* total packets received */
 945         u_int32_t       tcps_rcvpack;           /* packets received in sequence */
 946         u_int32_t       tcps_rcvbyte;           /* bytes received in sequence */
 947         u_int32_t       tcps_rcvbadsum;         /* packets received with ccksum errs */
 948         u_int32_t       tcps_rcvbadoff;         /* packets received with bad offset */
 949         u_int32_t       tcps_rcvmemdrop;        /* packets dropped for lack of memory */
 950         u_int32_t       tcps_rcvshort;          /* packets received too short */
 951         u_int32_t       tcps_rcvduppack;        /* duplicate-only packets received */
 952         u_int32_t       tcps_rcvdupbyte;        /* duplicate-only bytes received */
 953         u_int32_t       tcps_rcvpartduppack;    /* packets with some duplicate data */
 954         u_int32_t       tcps_rcvpartdupbyte;    /* dup. bytes in part-dup. packets */
 955         u_int32_t       tcps_rcvoopack;         /* out-of-order packets received */
 956         u_int32_t       tcps_rcvoobyte;         /* out-of-order bytes received */
 957         u_int32_t       tcps_rcvpackafterwin;   /* packets with data after window */
 958         u_int32_t       tcps_rcvbyteafterwin;   /* bytes rcvd after window */
 959         u_int32_t       tcps_rcvafterclose;     /* packets rcvd after "close" */
 960         u_int32_t       tcps_rcvwinprobe;       /* rcvd window probe packets */
 961         u_int32_t       tcps_rcvdupack;         /* rcvd duplicate acks */
 962         u_int32_t       tcps_rcvacktoomuch;     /* rcvd acks for unsent data */
 963         u_int32_t       tcps_rcvackpack;        /* rcvd ack packets */
 964         u_int32_t       tcps_rcvackbyte;        /* bytes acked by rcvd acks */
 965         u_int32_t       tcps_rcvwinupd;         /* rcvd window update packets */
 966         u_int32_t       tcps_pawsdrop;          /* segments dropped due to PAWS */
 967         u_int32_t       tcps_predack;           /* times hdr predict ok for acks */
 968         u_int32_t       tcps_preddat;           /* times hdr predict ok for data pkts */
 969         u_int32_t       tcps_pcbcachemiss;
 970         u_int32_t       tcps_cachedrtt;         /* times cached RTT in route updated */
 971         u_int32_t       tcps_cachedrttvar;      /* times cached rttvar updated */
 972         u_int32_t       tcps_cachedssthresh;    /* times cached ssthresh updated */
 973         u_int32_t       tcps_usedrtt;           /* times RTT initialized from route */
 974         u_int32_t       tcps_usedrttvar;        /* times RTTVAR initialized from rt */
 975         u_int32_t       tcps_usedssthresh;      /* times ssthresh initialized from rt*/
 976         u_int32_t       tcps_persistdrop;       /* timeout in persist state */
 977         u_int32_t       tcps_badsyn;            /* bogus SYN, e.g. premature ACK */
 978         u_int32_t       tcps_mturesent;         /* resends due to MTU discovery */
 979         u_int32_t       tcps_listendrop;        /* listen queue overflows */
 980
 981         /* new stats from FreeBSD 5.4 sync up */
 982         u_int32_t       tcps_minmssdrops;       /* average minmss too low drops */
 983         u_int32_t       tcps_sndrexmitbad;      /* unnecessary packet retransmissions */
 984         u_int32_t       tcps_badrst;            /* ignored RSTs in the window */
 985
 986         u_int32_t       tcps_sc_added;          /* entry added to syncache */
 987         u_int32_t       tcps_sc_retransmitted;  /* syncache entry was retransmitted */
 988         u_int32_t       tcps_sc_dupsyn;         /* duplicate SYN packet */
 989         u_int32_t       tcps_sc_dropped;        /* could not reply to packet */
 990         u_int32_t       tcps_sc_completed;      /* successful extraction of entry */
 991         u_int32_t       tcps_sc_bucketoverflow; /* syncache per-bucket limit hit */
 992         u_int32_t       tcps_sc_cacheoverflow;  /* syncache cache limit hit */
 993         u_int32_t       tcps_sc_reset;          /* RST removed entry from syncache */
 994         u_int32_t       tcps_sc_stale;          /* timed out or listen socket gone */
 995         u_int32_t       tcps_sc_aborted;        /* syncache entry aborted */
 996         u_int32_t       tcps_sc_badack;         /* removed due to bad ACK */
 997         u_int32_t       tcps_sc_unreach;        /* ICMP unreachable received */
 998         u_int32_t       tcps_sc_zonefail;       /* zalloc() failed */
 999         u_int32_t       tcps_sc_sendcookie;     /* SYN cookie sent */
1000         u_int32_t       tcps_sc_recvcookie;     /* SYN cookie received */
1001
1002         u_int32_t       tcps_hc_added;          /* entry added to hostcache */
1003         u_int32_t       tcps_hc_bucketoverflow; /* hostcache per bucket limit hit */
1004
1005         /* SACK related stats */
1006         u_int32_t       tcps_sack_recovery_episode; /* SACK recovery episodes */
1007         u_int32_t       tcps_sack_rexmits;          /* SACK rexmit segments   */
1008         u_int32_t       tcps_sack_rexmit_bytes;     /* SACK rexmit bytes      */
1009         u_int32_t       tcps_sack_rcv_blocks;       /* SACK blocks (options) received */
1010         u_int32_t       tcps_sack_send_blocks;      /* SACK blocks (options) sent     */
1011         u_int32_t       tcps_sack_sboverflow;       /* SACK sendblock overflow   */
1012
1013         u_int32_t       tcps_bg_rcvtotal;       /* total background packets received */
1014         u_int32_t       tcps_rxtfindrop;        /* drop conn after retransmitting FIN */
1015         u_int32_t       tcps_fcholdpacket;      /* packets withheld because of flow control */
1016
1017         /* LRO related stats */
1018         u_int32_t       tcps_coalesced_pack;    /* number of coalesced packets */
1019         u_int32_t       tcps_flowtbl_full;      /* times flow table was full */
1020         u_int32_t       tcps_flowtbl_collision; /* collisions in flow tbl */
1021         u_int32_t       tcps_lro_twopack;       /* 2 packets coalesced */
1022         u_int32_t       tcps_lro_multpack;      /* 3 or 4 pkts coalesced */
1023         u_int32_t       tcps_lro_largepack;     /* 5 or more pkts coalesced */
1024
1025         u_int32_t       tcps_limited_txt;       /* Limited transmit used */
1026         u_int32_t       tcps_early_rexmt;       /* Early retransmit used */
1027         u_int32_t       tcps_sack_ackadv;       /* Cumulative ack advanced along with sack */
1028
1029         /* Checksum related stats */
1030         u_int32_t       tcps_rcv_swcsum;        /* tcp swcksum (inbound), packets */
1031         u_int32_t       tcps_rcv_swcsum_bytes;  /* tcp swcksum (inbound), bytes */
1032         u_int32_t       tcps_rcv6_swcsum;       /* tcp6 swcksum (inbound), packets */
1033         u_int32_t       tcps_rcv6_swcsum_bytes; /* tcp6 swcksum (inbound), bytes */
1034         u_int32_t       tcps_snd_swcsum;        /* tcp swcksum (outbound), packets */
1035         u_int32_t       tcps_snd_swcsum_bytes;  /* tcp swcksum (outbound), bytes */
1036         u_int32_t       tcps_snd6_swcsum;       /* tcp6 swcksum (outbound), packets */
1037         u_int32_t       tcps_snd6_swcsum_bytes; /* tcp6 swcksum (outbound), bytes */
1038         u_int32_t       tcps_msg_unopkts;       /* unordered packet on TCP msg stream */
1039         u_int32_t       tcps_msg_unoappendfail; /* failed to append unordered pkt */
1040         u_int32_t       tcps_msg_sndwaithipri; /* send is waiting for high priority data */
1041
1042         /* MPTCP Related stats */
1043         u_int32_t       tcps_invalid_mpcap;     /* Invalid MPTCP capable opts */
1044         u_int32_t       tcps_invalid_joins;     /* Invalid MPTCP joins */
1045         u_int32_t       tcps_mpcap_fallback;    /* TCP fallback in primary */
1046         u_int32_t       tcps_join_fallback;     /* No MPTCP in secondary */
1047         u_int32_t       tcps_estab_fallback;    /* DSS option dropped */
1048         u_int32_t       tcps_invalid_opt;       /* Catchall error stat */
1049         u_int32_t       tcps_mp_outofwin;       /* Packet lies outside the
1050                                                    shared recv window */
1051         u_int32_t       tcps_mp_reducedwin;     /* Reduced subflow window */
1052         u_int32_t       tcps_mp_badcsum;        /* Bad DSS csum */
1053         u_int32_t       tcps_mp_oodata;         /* Out of order data */
1054         u_int32_t       tcps_mp_switches;       /* number of subflow switch */
1055         u_int32_t       tcps_mp_rcvtotal;       /* number of rcvd packets */
1056         u_int32_t       tcps_mp_rcvbytes;       /* number of bytes received */
1057         u_int32_t       tcps_mp_sndpacks;       /* number of data packs sent */
1058         u_int32_t       tcps_mp_sndbytes;       /* number of bytes sent */
1059         u_int32_t       tcps_join_rxmts;        /* join ack retransmits */
1060         u_int32_t       tcps_tailloss_rto;      /* RTO due to tail loss */
1061         u_int32_t       tcps_reordered_pkts;    /* packets reorderd */
1062         u_int32_t       tcps_recovered_pkts;    /* recovered after loss */
1063         u_int32_t       tcps_pto;               /* probe timeout */
1064         u_int32_t       tcps_rto_after_pto;     /* RTO after a probe */
1065         u_int32_t       tcps_tlp_recovery;      /* TLP induced fast recovery */
1066         u_int32_t       tcps_tlp_recoverlastpkt; /* TLP recoverd last pkt */
1067         u_int32_t       tcps_ecn_client_success; /* client-side connection negotiated ECN */
1068         u_int32_t       tcps_ecn_recv_ece;      /* ECE received, sent CWR */
1069         u_int32_t       tcps_ecn_sent_ece;      /* Sent ECE notification */
1070         u_int32_t       tcps_detect_reordering; /* Detect pkt reordering */
1071         u_int32_t       tcps_delay_recovery;    /* Delay fast recovery */
1072         u_int32_t       tcps_avoid_rxmt;        /* Retransmission was avoided */
1073         u_int32_t       tcps_unnecessary_rxmt;  /* Retransmission was not needed */
1074         u_int32_t       tcps_nostretchack;      /* disabled stretch ack algorithm on a connection */
1075         u_int32_t       tcps_rescue_rxmt;       /* SACK rescue retransmit */
1076         u_int32_t       tcps_pto_in_recovery;   /* rescue retransmit in fast recovery */
1077         u_int32_t       tcps_pmtudbh_reverted;  /* PMTU Blackhole detection, segment size reverted */
1078
1079         /* DSACK related statistics */
1080         u_int32_t       tcps_dsack_disable;     /* DSACK disabled due to n/w duplication */
1081         u_int32_t       tcps_dsack_ackloss;     /* ignore DSACK due to ack loss */
1082         u_int32_t       tcps_dsack_badrexmt;    /* DSACK based bad rexmt recovery */
1083         u_int32_t       tcps_dsack_sent;        /* Sent DSACK notification */
1084         u_int32_t       tcps_dsack_recvd;       /* Received a valid DSACK option */
1085         u_int32_t       tcps_dsack_recvd_old;   /* Received an out of window DSACK option */
1086
1087         /* MPTCP Subflow selection stats */
1088         u_int32_t       tcps_mp_sel_symtomsd;   /* By symptomsd */
1089         u_int32_t       tcps_mp_sel_rtt;        /* By RTT comparison */
1090         u_int32_t       tcps_mp_sel_rto;        /* By RTO comparison */
1091         u_int32_t       tcps_mp_sel_peer;       /* By peer's output pattern */
1092         u_int32_t       tcps_mp_num_probes;     /* Number of probes sent */
1093         u_int32_t       tcps_mp_verdowngrade;   /* MPTCP version downgrade */
1094         u_int32_t       tcps_drop_after_sleep;  /* drop after long AP sleep */
1095         u_int32_t       tcps_probe_if;          /* probe packets after interface availability */
1096         u_int32_t       tcps_probe_if_conflict; /* Can't send probe packets for interface */
1097
1098         u_int32_t       tcps_ecn_client_setup;  /* Attempted ECN setup from client side */
1099         u_int32_t       tcps_ecn_server_setup;  /* Attempted ECN setup from server side */
1100         u_int32_t       tcps_ecn_server_success; /* server-side connection negotiated ECN */
1101         u_int32_t       tcps_ecn_lost_synack;   /* Lost SYN-ACK with ECN setup */
1102         u_int32_t       tcps_ecn_lost_syn;      /* Lost SYN with ECN setup */
1103         u_int32_t       tcps_ecn_not_supported; /* Server did not support ECN setup */
1104         u_int32_t       tcps_ecn_recv_ce;       /* Received CE from the network */
1105         u_int32_t       tcps_ecn_conn_recv_ce;  /* Number of connections received CE atleast once */
1106         u_int32_t       tcps_ecn_conn_recv_ece; /* Number of connections received ECE atleast once */
1107         u_int32_t       tcps_ecn_conn_plnoce;   /* Number of connections that received no CE and sufferred packet loss */
1108         u_int32_t       tcps_ecn_conn_pl_ce;    /* Number of connections that received CE and sufferred packet loss */
1109         u_int32_t       tcps_ecn_conn_nopl_ce;  /* Number of connections that received CE and sufferred no packet loss */
1110
1111         /* TFO-related statistics */
1112         u_int32_t       tcps_tfo_syn_data_rcv;  /* Received a SYN+data with valid cookie */
1113         u_int32_t       tcps_tfo_cookie_req_rcv;/* Received a TFO cookie-request */
1114         u_int32_t       tcps_tfo_cookie_sent;   /* Offered a TFO-cookie to the client */
1115         u_int32_t       tcps_tfo_cookie_invalid;/* Received an invalid TFO-cookie */
1116         u_int32_t       tcps_tfo_cookie_req;    /* Cookie requested with the SYN */
1117         u_int32_t       tcps_tfo_cookie_rcv;    /* Cookie received in a SYN/ACK */
1118         u_int32_t       tcps_tfo_syn_data_sent; /* SYN+data+cookie sent */
1119         u_int32_t       tcps_tfo_syn_data_acked;/* SYN+data has been acknowledged */
1120         u_int32_t       tcps_tfo_syn_loss;      /* SYN+TFO has been lost and we fallback */
1121         u_int32_t       tcps_tfo_blackhole;     /* TFO got blackholed by a middlebox. */
1122 };
1123
1124 struct tcpstat_local {
1125         u_int64_t badformat;
1126         u_int64_t unspecv6;
1127         u_int64_t synfin;
1128         u_int64_t badformatipsec;
1129         u_int64_t noconnnolist;
1130         u_int64_t noconnlist;
1131         u_int64_t listbadsyn;
1132         u_int64_t icmp6unreach;
1133         u_int64_t deprecate6;
1134         u_int64_t ooopacket;
1135         u_int64_t rstinsynrcv;
1136         u_int64_t dospacket;
1137         u_int64_t cleanup;
1138         u_int64_t synwindow;
1139 };
1140
1141 #pragma pack(4)
1142
1143 /*
1144  * TCB structure exported to user-land via sysctl(3).
1145  * Evil hack: declare only if in_pcb.h and sys/socketvar.h have been
1146  * included.  Not all of our clients do.
1147  */
1148
1149 struct  xtcpcb {
1150         u_int32_t       xt_len;
1151 #ifdef KERNEL_PRIVATE
1152         struct  inpcb_compat    xt_inp;
1153 #else
1154         struct  inpcb   xt_inp;
1155 #endif
1156 #ifdef KERNEL_PRIVATE
1157         struct  otcpcb  xt_tp;
1158 #else
1159         struct  tcpcb   xt_tp;
1160 #endif
1161         struct  xsocket xt_socket;
1162         u_quad_t        xt_alignment_hack;
1163 };
1164
1165
1166 struct  xtcpcb64 {
1167         u_int32_t               xt_len;
1168         struct xinpcb64         xt_inpcb;
1169
1170         u_int64_t t_segq;
1171         int     t_dupacks;              /* consecutive dup acks recd */
1172
1173         int t_timer[TCPT_NTIMERS_EXT];  /* tcp timers */
1174
1175         int     t_state;                /* state of this connection */
1176         u_int   t_flags;
1177
1178         int     t_force;                /* 1 if forcing out a byte */
1179
1180         tcp_seq snd_una;                /* send unacknowledged */
1181         tcp_seq snd_max;                /* highest sequence number sent;
1182                                          * used to recognize retransmits
1183                                          */
1184         tcp_seq snd_nxt;                /* send next */
1185         tcp_seq snd_up;                 /* send urgent pointer */
1186
1187         tcp_seq snd_wl1;                /* window update seg seq number */
1188         tcp_seq snd_wl2;                /* window update seg ack number */
1189         tcp_seq iss;                    /* initial send sequence number */
1190         tcp_seq irs;                    /* initial receive sequence number */
1191
1192         tcp_seq rcv_nxt;                /* receive next */
1193         tcp_seq rcv_adv;                /* advertised window */
1194         u_int32_t rcv_wnd;              /* receive window */
1195         tcp_seq rcv_up;                 /* receive urgent pointer */
1196
1197         u_int32_t snd_wnd;              /* send window */
1198         u_int32_t snd_cwnd;             /* congestion-controlled window */
1199         u_int32_t snd_ssthresh;         /* snd_cwnd size threshold for
1200                                          * for slow start exponential to
1201                                          * linear switch
1202                                          */
1203         u_int   t_maxopd;               /* mss plus options */
1204
1205         u_int32_t t_rcvtime;            /* time at which a packet was received */
1206         u_int32_t t_starttime;          /* time connection was established */
1207         int     t_rtttime;              /* round trip time */
1208         tcp_seq t_rtseq;                /* sequence number being timed */
1209
1210         int     t_rxtcur;               /* current retransmit value (ticks) */
1211         u_int   t_maxseg;               /* maximum segment size */
1212         int     t_srtt;                 /* smoothed round-trip time */
1213         int     t_rttvar;               /* variance in round-trip time */
1214
1215         int     t_rxtshift;             /* log(2) of rexmt exp. backoff */
1216         u_int   t_rttmin;               /* minimum rtt allowed */
1217         u_int32_t t_rttupdated;         /* number of times rtt sampled */
1218         u_int32_t max_sndwnd;           /* largest window peer has offered */
1219
1220         int     t_softerror;            /* possible error not yet reported */
1221 /* out-of-band data */
1222         char    t_oobflags;             /* have some */
1223         char    t_iobc;                 /* input character */
1224 /* RFC 1323 variables */
1225         u_char  snd_scale;              /* window scaling for send window */
1226         u_char  rcv_scale;              /* window scaling for recv window */
1227         u_char  request_r_scale;        /* pending window scaling */
1228         u_char  requested_s_scale;
1229         u_int32_t ts_recent;            /* timestamp echo data */
1230
1231         u_int32_t ts_recent_age;        /* when last updated */
1232         tcp_seq last_ack_sent;
1233 /* RFC 1644 variables */
1234         tcp_cc  cc_send;                /* send connection count */
1235         tcp_cc  cc_recv;                /* receive connection count */
1236         tcp_seq snd_recover;            /* for use in fast recovery */
1237 /* experimental */
1238         u_int32_t snd_cwnd_prev;        /* cwnd prior to retransmit */
1239         u_int32_t snd_ssthresh_prev;    /* ssthresh prior to retransmit */
1240         u_int32_t t_badrxtwin;          /* window for retransmit recovery */
1241
1242         u_quad_t                xt_alignment_hack;
1243 };
1244
1245
1246 #ifdef PRIVATE
1247
1248 struct  xtcpcb_n {
1249         u_int32_t               xt_len;
1250         u_int32_t                       xt_kind;                /* XSO_TCPCB */
1251
1252         u_int64_t t_segq;
1253         int     t_dupacks;              /* consecutive dup acks recd */
1254
1255         int t_timer[TCPT_NTIMERS_EXT];  /* tcp timers */
1256
1257         int     t_state;                /* state of this connection */
1258         u_int   t_flags;
1259
1260         int     t_force;                /* 1 if forcing out a byte */
1261
1262         tcp_seq snd_una;                /* send unacknowledged */
1263         tcp_seq snd_max;                /* highest sequence number sent;
1264                                                                          * used to recognize retransmits
1265                                                                          */
1266         tcp_seq snd_nxt;                /* send next */
1267         tcp_seq snd_up;                 /* send urgent pointer */
1268
1269         tcp_seq snd_wl1;                /* window update seg seq number */
1270         tcp_seq snd_wl2;                /* window update seg ack number */
1271         tcp_seq iss;                    /* initial send sequence number */
1272         tcp_seq irs;                    /* initial receive sequence number */
1273
1274         tcp_seq rcv_nxt;                /* receive next */
1275         tcp_seq rcv_adv;                /* advertised window */
1276         u_int32_t rcv_wnd;              /* receive window */
1277         tcp_seq rcv_up;                 /* receive urgent pointer */
1278
1279         u_int32_t snd_wnd;              /* send window */
1280         u_int32_t snd_cwnd;             /* congestion-controlled window */
1281         u_int32_t snd_ssthresh;         /* snd_cwnd size threshold for
1282                                                                          * for slow start exponential to
1283                                                                          * linear switch
1284                                                                          */
1285         u_int   t_maxopd;               /* mss plus options */
1286
1287         u_int32_t t_rcvtime;            /* time at which a packet was received */
1288         u_int32_t t_starttime;          /* time connection was established */
1289         int     t_rtttime;              /* round trip time */
1290         tcp_seq t_rtseq;                /* sequence number being timed */
1291
1292         int     t_rxtcur;               /* current retransmit value (ticks) */
1293         u_int   t_maxseg;               /* maximum segment size */
1294         int     t_srtt;                 /* smoothed round-trip time */
1295         int     t_rttvar;               /* variance in round-trip time */
1296
1297         int     t_rxtshift;             /* log(2) of rexmt exp. backoff */
1298         u_int   t_rttmin;               /* minimum rtt allowed */
1299         u_int32_t t_rttupdated;         /* number of times rtt sampled */
1300         u_int32_t max_sndwnd;           /* largest window peer has offered */
1301
1302         int     t_softerror;            /* possible error not yet reported */
1303         /* out-of-band data */
1304         char    t_oobflags;             /* have some */
1305         char    t_iobc;                 /* input character */
1306         /* RFC 1323 variables */
1307         u_char  snd_scale;              /* window scaling for send window */
1308         u_char  rcv_scale;              /* window scaling for recv window */
1309         u_char  request_r_scale;        /* pending window scaling */
1310         u_char  requested_s_scale;
1311         u_int32_t ts_recent;            /* timestamp echo data */
1312
1313         u_int32_t ts_recent_age;        /* when last updated */
1314         tcp_seq last_ack_sent;
1315         /* RFC 1644 variables */
1316         tcp_cc  cc_send;                /* send connection count */
1317         tcp_cc  cc_recv;                /* receive connection count */
1318         tcp_seq snd_recover;            /* for use in fast recovery */
1319         /* experimental */
1320         u_int32_t snd_cwnd_prev;        /* cwnd prior to retransmit */
1321         u_int32_t snd_ssthresh_prev;    /* ssthresh prior to retransmit */
1322 };
1323
1324 #endif /* PRIVATE */
1325
1326 #pragma pack()
1327
1328 /*
1329  * Names for TCP sysctl objects
1330  */
1331 #define TCPCTL_DO_RFC1323       1       /* use RFC-1323 extensions */
1332 #define TCPCTL_DO_RFC1644       2       /* use RFC-1644 extensions */
1333 #define TCPCTL_MSSDFLT          3       /* MSS default */
1334 #define TCPCTL_STATS            4       /* statistics (read-only) */
1335 #define TCPCTL_RTTDFLT          5       /* default RTT estimate */
1336 #define TCPCTL_KEEPIDLE         6       /* keepalive idle timer */
1337 #define TCPCTL_KEEPINTVL        7       /* interval to send keepalives */
1338 #define TCPCTL_SENDSPACE        8       /* send buffer space */
1339 #define TCPCTL_RECVSPACE        9       /* receive buffer space */
1340 #define TCPCTL_KEEPINIT         10      /* timeout for establishing syn */
1341 #define TCPCTL_PCBLIST          11      /* list of all outstanding PCBs */
1342 #define TCPCTL_DELACKTIME       12      /* time before sending delayed ACK */
1343 #define TCPCTL_V6MSSDFLT        13      /* MSS default for IPv6 */
1344 #define TCPCTL_MAXID            14
1345
1346 #ifdef BSD_KERNEL_PRIVATE
1347 #include <sys/bitstring.h>
1348
1349 #define TCP_PKTLIST_CLEAR(tp) {                                         \
1350         (tp)->t_pktlist_head = (tp)->t_pktlist_tail = NULL;             \
1351         (tp)->t_lastchain = (tp)->t_pktlist_sentlen = 0;                \
1352 }
1353
1354 #define TCPCTL_NAMES { \
1355         { 0, 0 }, \
1356         { "rfc1323", CTLTYPE_INT }, \
1357         { "rfc1644", CTLTYPE_INT }, \
1358         { "mssdflt", CTLTYPE_INT }, \
1359         { "stats", CTLTYPE_STRUCT }, \
1360         { "rttdflt", CTLTYPE_INT }, \
1361         { "keepidle", CTLTYPE_INT }, \
1362         { "keepintvl", CTLTYPE_INT }, \
1363         { "sendspace", CTLTYPE_INT }, \
1364         { "recvspace", CTLTYPE_INT }, \
1365         { "keepinit", CTLTYPE_INT }, \
1366         { "pcblist", CTLTYPE_STRUCT }, \
1367         { "delacktime", CTLTYPE_INT }, \
1368         { "v6mssdflt", CTLTYPE_INT }, \
1369 }
1370
1371 #ifdef SYSCTL_DECL
1372 SYSCTL_DECL(_net_inet_tcp);
1373 #endif /* SYSCTL_DECL */
1374
1375 extern  struct inpcbhead tcb;           /* head of queue of active tcpcb's */
1376 extern  struct inpcbinfo tcbinfo;
1377 extern  struct tcpstat tcpstat; /* tcp statistics */
1378 extern  int tcp_mssdflt;        /* XXX */
1379 extern  int tcp_minmss;
1380 #define TCP_FASTOPEN_SERVER 0x01
1381 #define TCP_FASTOPEN_CLIENT 0x02
1382
1383 extern  int tcp_tfo_halfcnt;
1384 extern  int tcp_tfo_backlog;
1385 extern  int tcp_fastopen;
1386 extern  int tcp_tfo_fallback_min;
1387 extern  int ss_fltsz;
1388 extern  int ss_fltsz_local;
1389 extern  int tcp_do_rfc3390;             /* Calculate ss_fltsz according to RFC 3390 */
1390 extern int tcp_do_rfc1323;
1391 extern int target_qdelay;
1392 extern  u_int32_t tcp_now;              /* for RFC 1323 timestamps */
1393 extern struct timeval tcp_uptime;
1394 extern lck_spin_t *tcp_uptime_lock;
1395 extern  int tcp_delack_enabled;
1396 extern  int tcp_do_sack;        /* SACK enabled/disabled */
1397 extern int tcp_do_rfc3465;
1398 extern int tcp_do_rfc3465_lim2;
1399 extern int maxseg_unacked;
1400 extern int tcp_use_newreno;
1401 extern struct zone *tcp_reass_zone;
1402 extern struct zone *tcp_rxt_seg_zone;
1403 extern int tcp_ecn_outbound;
1404 extern int tcp_ecn_inbound;
1405
1406
1407 #if CONFIG_IFEF_NOWINDOWSCALE
1408 extern int tcp_obey_ifef_nowindowscale;
1409 #endif
1410
1411 struct protosw;
1412 struct domain;
1413
1414 struct tcp_respond_args {
1415         unsigned int ifscope;
1416         unsigned int nocell:1,
1417                      noexpensive:1,
1418                      awdl_unrestricted:1;
1419 };
1420
1421 void     tcp_canceltimers(struct tcpcb *);
1422 struct tcpcb *
1423          tcp_close(struct tcpcb *);
1424 void     tcp_ctlinput(int, struct sockaddr *, void *);
1425 int      tcp_ctloutput(struct socket *, struct sockopt *);
1426 struct tcpcb *
1427          tcp_drop(struct tcpcb *, int);
1428 void     tcp_drain(void);
1429 void     tcp_getrt_rtt(struct tcpcb *tp, struct rtentry *rt);
1430 struct rmxp_tao *
1431          tcp_gettaocache(struct inpcb *);
1432 void     tcp_init(struct protosw *, struct domain *);
1433 void     tcp_input(struct mbuf *, int);
1434 void     tcp_mss(struct tcpcb *, int, unsigned int);
1435 int      tcp_mssopt(struct tcpcb *);
1436 void     tcp_drop_syn_sent(struct inpcb *, int);
1437 void     tcp_mtudisc(struct inpcb *, int);
1438 struct tcpcb *
1439          tcp_newtcpcb(struct inpcb *);
1440 int      tcp_output(struct tcpcb *);
1441 void     tcp_respond(struct tcpcb *, void *, struct tcphdr *, struct mbuf *,
1442     tcp_seq, tcp_seq, int, struct tcp_respond_args *);
1443 struct rtentry *
1444          tcp_rtlookup(struct inpcb *, unsigned int);
1445 void     tcp_setpersist(struct tcpcb *);
1446 void     tcp_gc(struct inpcbinfo *);
1447 void     tcp_itimer(struct inpcbinfo *ipi);
1448 void     tcp_check_timer_state(struct tcpcb *tp);
1449 void     tcp_run_timerlist(void *arg1, void *arg2);
1450
1451 struct tcptemp *tcp_maketemplate(struct tcpcb *);
1452 void     tcp_fillheaders(struct tcpcb *, void *, void *);
1453 struct tcpcb *tcp_timers(struct tcpcb *, int);
1454 void     tcp_trace(int, int, struct tcpcb *, void *, struct tcphdr *, int);
1455
1456 void tcp_sack_doack(struct tcpcb *, struct tcpopt *, struct tcphdr *,
1457     u_int32_t *);
1458 extern boolean_t tcp_sack_process_dsack(struct tcpcb *, struct tcpopt *,
1459     struct tcphdr *);
1460 int tcp_detect_bad_rexmt(struct tcpcb *, struct tcphdr *, struct tcpopt *,
1461     u_int32_t rxtime);
1462 void     tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_laststart, tcp_seq rcv_lastend);
1463 void     tcp_clean_sackreport(struct tcpcb *tp);
1464 void     tcp_sack_adjust(struct tcpcb *tp);
1465 struct sackhole *tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt);
1466 void     tcp_sack_partialack(struct tcpcb *, struct tcphdr *);
1467 void     tcp_free_sackholes(struct tcpcb *tp);
1468 int32_t  tcp_sbspace(struct tcpcb *tp);
1469 void     tcp_set_tso(struct tcpcb *tp, struct ifnet *ifp);
1470 void     tcp_reset_stretch_ack(struct tcpcb *tp);
1471 extern void tcp_get_ports_used(u_int32_t, int, u_int32_t, bitstr_t *);
1472 uint32_t tcp_count_opportunistic(unsigned int ifindex, u_int32_t flags);
1473 uint32_t tcp_find_anypcb_byaddr(struct ifaddr *ifa);
1474 void     tcp_set_max_rwinscale(struct tcpcb *tp, struct socket *so);
1475 struct bwmeas* tcp_bwmeas_alloc(struct tcpcb *tp);
1476 void tcp_bwmeas_free(struct tcpcb *tp);
1477 extern int32_t timer_diff(uint32_t t1, uint32_t toff1, uint32_t t2, uint32_t toff2);
1478
1479 extern void tcp_set_background_cc(struct socket *);
1480 extern void tcp_set_foreground_cc(struct socket *);
1481 extern void tcp_set_recv_bg(struct socket *);
1482 extern void tcp_clear_recv_bg(struct socket *);
1483 extern boolean_t tcp_sack_byte_islost(struct tcpcb *tp);
1484 #define IS_TCP_RECV_BG(_so)     \
1485         ((_so)->so_traffic_mgt_flags & TRAFFIC_MGT_TCP_RECVBG)
1486
1487 #if TRAFFIC_MGT
1488 #define CLEAR_IAJ_STATE(_tp_) (_tp_)->iaj_rcv_ts = 0
1489 void     reset_acc_iaj(struct tcpcb *tp);
1490 #endif /* TRAFFIC_MGT */
1491
1492 int      tcp_lock (struct socket *, int, void *);
1493 int      tcp_unlock (struct socket *, int, void *);
1494 void     calculate_tcp_clock(void);
1495
1496 extern void tcp_keepalive_reset(struct tcpcb *);
1497 extern uint32_t get_base_rtt(struct tcpcb *tp);
1498
1499 #ifdef _KERN_LOCKS_H_
1500 lck_mtx_t *      tcp_getlock (struct socket *, int);
1501 #else
1502 void *   tcp_getlock (struct socket *, int);
1503 #endif
1504
1505
1506 extern  struct pr_usrreqs tcp_usrreqs;
1507 extern  u_int32_t tcp_sendspace;
1508 extern  u_int32_t tcp_recvspace;
1509 tcp_seq tcp_new_isn(struct tcpcb *);
1510
1511 extern int tcp_input_checksum(int, struct mbuf *, struct tcphdr *, int, int);
1512 extern void tcp_getconninfo(struct socket *, struct conninfo_tcp *);
1513 extern void add_to_time_wait(struct tcpcb *, uint32_t delay);
1514 extern void tcp_pmtud_revert_segment_size(struct tcpcb *tp);
1515 extern void tcp_rxtseg_insert(struct tcpcb *, tcp_seq, tcp_seq);
1516 extern struct tcp_rxt_seg *tcp_rxtseg_find(struct tcpcb *, tcp_seq, tcp_seq);
1517 extern void tcp_rxtseg_clean(struct tcpcb *);
1518 extern boolean_t tcp_rxtseg_detect_bad_rexmt(struct tcpcb *, tcp_seq);
1519 extern boolean_t tcp_rxtseg_dsack_for_tlp(struct tcpcb *);
1520 extern u_int32_t tcp_rxtseg_total_size(struct tcpcb *tp);
1521 extern void tcp_rexmt_save_state(struct tcpcb *tp);
1522 extern void tcp_interface_send_probe(u_int16_t if_index_available);
1523 extern void tcp_probe_connectivity(struct ifnet *ifp, u_int32_t enable);
1524 extern void tcp_get_connectivity_status(struct tcpcb *,
1525     struct tcp_conn_status *);
1526
1527 extern boolean_t tfo_enabled(const struct tcpcb *tp);
1528 extern void tcp_disable_tfo(struct tcpcb *tp);
1529 extern void tcp_tfo_gen_cookie(struct inpcb *inp, u_char *out, size_t blk_size);
1530 #define TCP_FASTOPEN_KEYLEN 16
1531
1532 #if MPTCP
1533 extern int mptcp_input_preproc(struct tcpcb *, struct mbuf *, int);
1534 extern void mptcp_output_csum(struct tcpcb *, struct mbuf *, int32_t, unsigned,
1535     u_int64_t, u_int32_t *);
1536 extern int mptcp_adj_mss(struct tcpcb *, boolean_t);
1537 extern void mptcp_insert_rmap(struct tcpcb *, struct mbuf *);
1538 #endif
1539 #endif /* BSD_KERNEL_RPIVATE */
1540
1541 #endif /* _NETINET_TCP_VAR_H_ */