]> git.saurik.com Git - apple/xnu.git/blob - bsd/netinet/tcp_output.c
projects / apple / xnu.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
xnu-3248.20.55.tar.gz
[apple/xnu.git] / bsd / netinet / tcp_output.c
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, 1988, 1990, 1993, 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_output.c 8.4 (Berkeley) 5/24/95
61 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.10 2001/07/07 04:30:38 silby Exp $
62 */
63 /*
64 * NOTICE: This file was modified by SPARTA, Inc. in 2005 to introduce
65 * support for mandatory and extensible security protections. This notice
66 * is included in support of clause 2.2 (b) of the Apple Public License,
67 * Version 2.0.
68 */
69
70 #define _IP_VHL
71
72
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mbuf.h>
78 #include <sys/domain.h>
79 #include <sys/protosw.h>
80 #include <sys/socket.h>
81 #include <sys/socketvar.h>
82
83 #include <net/route.h>
84 #include <net/ntstat.h>
85 #include <net/if_var.h>
86 #include <net/if.h>
87 #include <net/if_types.h>
88 #include <net/dlil.h>
89
90 #include <netinet/in.h>
91 #include <netinet/in_systm.h>
92 #include <netinet/in_var.h>
93 #include <netinet/ip.h>
94 #include <netinet/in_pcb.h>
95 #include <netinet/ip_var.h>
96 #include <mach/sdt.h>
97 #if INET6
98 #include <netinet6/in6_pcb.h>
99 #include <netinet/ip6.h>
100 #include <netinet6/ip6_var.h>
101 #endif
102 #include <netinet/tcp.h>
103 #define TCPOUTFLAGS
104 #include <netinet/tcp_cache.h>
105 #include <netinet/tcp_fsm.h>
106 #include <netinet/tcp_seq.h>
107 #include <netinet/tcp_timer.h>
108 #include <netinet/tcp_var.h>
109 #include <netinet/tcpip.h>
110 #include <netinet/tcp_cc.h>
111 #if TCPDEBUG
112 #include <netinet/tcp_debug.h>
113 #endif
114 #include <sys/kdebug.h>
115 #include <mach/sdt.h>
116
117 #if IPSEC
118 #include <netinet6/ipsec.h>
119 #endif /*IPSEC*/
120
121 #if CONFIG_MACF_NET
122 #include <security/mac_framework.h>
123 #endif /* MAC_SOCKET */
124
125 #include <netinet/lro_ext.h>
126 #if MPTCP
127 #include <netinet/mptcp_var.h>
128 #include <netinet/mptcp.h>
129 #include <netinet/mptcp_opt.h>
130 #endif
131
132 #include <corecrypto/ccaes.h>
133
134 #define DBG_LAYER_BEG NETDBG_CODE(DBG_NETTCP, 1)
135 #define DBG_LAYER_END NETDBG_CODE(DBG_NETTCP, 3)
136 #define DBG_FNC_TCP_OUTPUT NETDBG_CODE(DBG_NETTCP, (4 << 8) | 1)
137
138 int path_mtu_discovery = 1;
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery,
140 CTLFLAG_RW | CTLFLAG_LOCKED, &path_mtu_discovery, 1,
141 "Enable Path MTU Discovery");
142
143 int ss_fltsz = 1;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize,
145 CTLFLAG_RW | CTLFLAG_LOCKED,&ss_fltsz, 1,
146 "Slow start flight size");
147
148 int ss_fltsz_local = 8; /* starts with eight segments max */
149 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
150 CTLFLAG_RW | CTLFLAG_LOCKED, &ss_fltsz_local, 1,
151 "Slow start flight size for local networks");
152
153 int tcp_do_tso = 1;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW | CTLFLAG_LOCKED,
155 &tcp_do_tso, 0, "Enable TCP Segmentation Offload");
156
157 static int
158 sysctl_change_ecn_setting SYSCTL_HANDLER_ARGS
159 {
160 #pragma unused(oidp, arg1, arg2)
161 int i, err = 0, changed = 0;
162 struct ifnet *ifp;
163
164 err = sysctl_io_number(req, tcp_ecn_outbound, sizeof(int32_t),
165 &i, &changed);
166 if (err != 0 || req->newptr == USER_ADDR_NULL)
167 return(err);
168
169 if (changed) {
170 if ((tcp_ecn_outbound == 0 || tcp_ecn_outbound == 1) &&
171 (i == 0 || i == 1)) {
172 tcp_ecn_outbound = i;
173 return(err);
174 }
175 if (tcp_ecn_outbound == 2 && (i == 0 || i == 1)) {
176 /*
177 * Reset ECN enable flags on non-cellular
178 * interfaces so that the system default will take
179 * over
180 */
181 ifnet_head_lock_shared();
182 TAILQ_FOREACH(ifp, &ifnet_head, if_link) {
183 if (!IFNET_IS_CELLULAR(ifp)) {
184 ifnet_lock_exclusive(ifp);
185 ifp->if_eflags &= ~IFEF_ECN_DISABLE;
186 ifp->if_eflags &= ~IFEF_ECN_ENABLE;
187 ifnet_lock_done(ifp);
188 }
189 }
190 ifnet_head_done();
191 } else {
192 /*
193 * Set ECN enable flags on non-cellular
194 * interfaces
195 */
196 ifnet_head_lock_shared();
197 TAILQ_FOREACH(ifp, &ifnet_head, if_link) {
198 if (!IFNET_IS_CELLULAR(ifp)) {
199 ifnet_lock_exclusive(ifp);
200 ifp->if_eflags |= IFEF_ECN_ENABLE;
201 ifp->if_eflags &= ~IFEF_ECN_DISABLE;
202 ifnet_lock_done(ifp);
203 }
204 }
205 ifnet_head_done();
206 }
207 tcp_ecn_outbound = i;
208 }
209 /* Change the other one too as the work is done */
210 if (i == 2 || tcp_ecn_inbound == 2)
211 tcp_ecn_inbound = i;
212 return (err);
213 }
214
215 int tcp_ecn_outbound = 0;
216 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, ecn_initiate_out,
217 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_outbound, 0,
218 sysctl_change_ecn_setting, "IU",
219 "Initiate ECN for outbound connections");
220
221 int tcp_ecn_inbound = 0;
222 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, ecn_negotiate_in,
223 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_inbound, 0,
224 sysctl_change_ecn_setting, "IU",
225 "Initiate ECN for inbound connections");
226
227 int tcp_packet_chaining = 50;
228 SYSCTL_INT(_net_inet_tcp, OID_AUTO, packetchain,
229 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_packet_chaining, 0,
230 "Enable TCP output packet chaining");
231
232 int tcp_output_unlocked = 1;
233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, socket_unlocked_on_output,
234 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_output_unlocked, 0,
235 "Unlock TCP when sending packets down to IP");
236
237 int tcp_do_rfc3390 = 1;
238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390,
239 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_do_rfc3390, 1,
240 "Calculate intial slowstart cwnd depending on MSS");
241
242 int tcp_min_iaj_win = MIN_IAJ_WIN;
243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, min_iaj_win,
244 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_min_iaj_win, 1,
245 "Minimum recv win based on inter-packet arrival jitter");
246
247 int tcp_acc_iaj_react_limit = ACC_IAJ_REACT_LIMIT;
248 SYSCTL_INT(_net_inet_tcp, OID_AUTO, acc_iaj_react_limit,
249 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_acc_iaj_react_limit, 1,
250 "Accumulated IAJ when receiver starts to react");
251
252 uint32_t tcp_do_autosendbuf = 1;
253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, doautosndbuf,
254 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_do_autosendbuf, 1,
255 "Enable send socket buffer auto-tuning");
256
257 uint32_t tcp_autosndbuf_inc = 8 * 1024;
258 SYSCTL_INT(_net_inet_tcp, OID_AUTO, autosndbufinc,
259 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_autosndbuf_inc, 1,
260 "Increment in send socket bufffer size");
261
262 uint32_t tcp_autosndbuf_max = 512 * 1024;
263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, autosndbufmax,
264 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_autosndbuf_max, 1,
265 "Maximum send socket buffer size");
266
267 uint32_t tcp_prioritize_acks = 1;
268 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ack_prioritize,
269 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_prioritize_acks, 1,
270 "Prioritize pure acks");
271
272 uint32_t tcp_use_rtt_recvbg = 1;
273 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rtt_recvbg,
274 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_use_rtt_recvbg, 1,
275 "Use RTT for bg recv algorithm");
276
277 uint32_t tcp_recv_throttle_minwin = 16 * 1024;
278 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recv_throttle_minwin,
279 CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_recv_throttle_minwin, 1,
280 "Minimum recv win for throttling");
281
282 int32_t tcp_enable_tlp = 1;
283 SYSCTL_INT(_net_inet_tcp, OID_AUTO, enable_tlp,
284 CTLFLAG_RW | CTLFLAG_LOCKED,
285 &tcp_enable_tlp, 1, "Enable Tail loss probe");
286
287 static int32_t packchain_newlist = 0;
288 static int32_t packchain_looped = 0;
289 static int32_t packchain_sent = 0;
290
291 /* temporary: for testing */
292 #if IPSEC
293 extern int ipsec_bypass;
294 #endif
295
296 extern int slowlink_wsize; /* window correction for slow links */
297 #if IPFIREWALL
298 extern int fw_enable; /* firewall check for packet chaining */
299 extern int fw_bypass; /* firewall check: disable packet chaining if there is rules */
300 #endif /* IPFIREWALL */
301
302 extern u_int32_t dlil_filter_disable_tso_count;
303 extern u_int32_t kipf_count;
304 extern int tcp_recv_bg;
305
306 static int tcp_ip_output(struct socket *, struct tcpcb *, struct mbuf *, int,
307 struct mbuf *, int, int, int32_t, boolean_t);
308 static struct mbuf* tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th);
309 static int tcp_recv_throttle(struct tcpcb *tp);
310
311 static int32_t tcp_tfo_check(struct tcpcb *tp, int32_t len)
312 {
313 struct socket *so = tp->t_inpcb->inp_socket;
314 unsigned int optlen = 0;
315 unsigned int cookie_len;
316
317 if (tp->t_flags & TF_NOOPT)
318 goto fallback;
319
320 if (!tcp_heuristic_do_tfo(tp))
321 goto fallback;
322
323 optlen += TCPOLEN_MAXSEG;
324
325 if (tp->t_flags & TF_REQ_SCALE)
326 optlen += 4;
327
328 #if MPTCP
329 if ((so->so_flags & SOF_MP_SUBFLOW) && mptcp_enable &&
330 tp->t_rxtshift <= mptcp_mpcap_retries)
331 optlen += sizeof(struct mptcp_mpcapable_opt_common) + sizeof(mptcp_key_t);
332 #endif /* MPTCP */
333
334 if (tp->t_flags & TF_REQ_TSTMP)
335 optlen += TCPOLEN_TSTAMP_APPA;
336
337 if (SACK_ENABLED(tp))
338 optlen += TCPOLEN_SACK_PERMITTED;
339
340 /* Now, decide whether to use TFO or not */
341
342 /* Don't even bother trying if there is no space at all... */
343 if (MAX_TCPOPTLEN - optlen < TCPOLEN_FASTOPEN_REQ)
344 goto fallback;
345
346 cookie_len = tcp_cache_get_cookie_len(tp);
347 if (cookie_len == 0)
348 /* No cookie, so we request one */
349 return (0);
350
351 /* Do not send SYN+data if there is more in the queue than MSS */
352 if (so->so_snd.sb_cc > (tp->t_maxopd - MAX_TCPOPTLEN))
353 goto fallback;
354
355 /* Ok, everything looks good. We can go on and do TFO */
356 return (len);
357
358 fallback:
359 tp->t_flagsext &= ~TF_FASTOPEN;
360 return (0);
361 }
362
363 /* Returns the number of bytes written to the TCP option-space */
364 static unsigned
365 tcp_tfo_write_cookie_rep(struct tcpcb *tp, unsigned optlen, u_char *opt)
366 {
367 u_char out[CCAES_BLOCK_SIZE];
368 unsigned ret = 0;
369 u_char *bp;
370
371 if ((MAX_TCPOPTLEN - optlen) <
372 (TCPOLEN_FASTOPEN_REQ + TFO_COOKIE_LEN_DEFAULT))
373 return (ret);
374
375 tcp_tfo_gen_cookie(tp->t_inpcb, out, sizeof(out));
376
377 bp = opt + optlen;
378
379 *bp++ = TCPOPT_FASTOPEN;
380 *bp++ = 2 + TFO_COOKIE_LEN_DEFAULT;
381 memcpy(bp, out, TFO_COOKIE_LEN_DEFAULT);
382 ret += 2 + TFO_COOKIE_LEN_DEFAULT;
383
384 tp->t_tfo_stats |= TFO_S_COOKIE_SENT;
385 tcpstat.tcps_tfo_cookie_sent++;
386
387 return (ret);
388 }
389
390 static unsigned
391 tcp_tfo_write_cookie(struct tcpcb *tp, unsigned optlen, int32_t *len,
392 u_char *opt)
393 {
394 u_int8_t tfo_len = MAX_TCPOPTLEN - optlen - TCPOLEN_FASTOPEN_REQ;
395 unsigned ret = 0;
396 int res;
397 u_char *bp;
398
399 bp = opt + optlen;
400
401 /*
402 * The cookie will be copied in the appropriate place within the
403 * TCP-option space. That way we avoid the need for an intermediate
404 * variable.
405 */
406 res = tcp_cache_get_cookie(tp, bp + TCPOLEN_FASTOPEN_REQ, &tfo_len);
407 if (res == 0) {
408 *bp++ = TCPOPT_FASTOPEN;
409 *bp++ = TCPOLEN_FASTOPEN_REQ;
410 ret += TCPOLEN_FASTOPEN_REQ;
411
412 tp->t_tfo_flags |= TFO_F_COOKIE_REQ;
413
414 tp->t_tfo_stats |= TFO_S_COOKIE_REQ;
415 tcpstat.tcps_tfo_cookie_req++;
416 } else {
417 *bp++ = TCPOPT_FASTOPEN;
418 *bp++ = TCPOLEN_FASTOPEN_REQ + tfo_len;
419
420 ret += TCPOLEN_FASTOPEN_REQ + tfo_len;
421
422 tp->t_tfo_flags |= TFO_F_COOKIE_SENT;
423
424 /* If there is some data, let's track it */
425 if (*len) {
426 tp->t_tfo_stats |= TFO_S_SYN_DATA_SENT;
427 tcpstat.tcps_tfo_syn_data_sent++;
428 }
429 }
430
431 return (ret);
432 }
433
434 static inline bool
435 tcp_send_ecn_flags_on_syn(struct tcpcb *tp, struct socket *so)
436 {
437 return(!((tp->ecn_flags & TE_SETUPSENT) ||
438 (so->so_flags & SOF_MP_SUBFLOW) ||
439 (tp->t_flagsext & TF_FASTOPEN)));
440 }
441
442 void
443 tcp_set_ecn(struct tcpcb *tp, struct ifnet *ifp)
444 {
445 boolean_t inbound;
446
447 /*
448 * Socket option has precedence
449 */
450 if (tp->ecn_flags & TE_ECN_MODE_ENABLE) {
451 tp->ecn_flags |= TE_ENABLE_ECN;
452 goto check_heuristic;
453 }
454
455 if (tp->ecn_flags & TE_ECN_MODE_DISABLE) {
456 tp->ecn_flags &= ~TE_ENABLE_ECN;
457 return;
458 }
459 /*
460 * Per interface setting comes next
461 */
462 if (ifp != NULL) {
463 if (ifp->if_eflags & IFEF_ECN_ENABLE) {
464 tp->ecn_flags |= TE_ENABLE_ECN;
465 goto check_heuristic;
466 }
467
468 if (ifp->if_eflags & IFEF_ECN_DISABLE) {
469 tp->ecn_flags &= ~TE_ENABLE_ECN;
470 return;
471 }
472 }
473 /*
474 * System wide settings come last
475 */
476 inbound = (tp->t_inpcb->inp_socket->so_head != NULL);
477 if ((inbound && tcp_ecn_inbound == 1) ||
478 (!inbound && tcp_ecn_outbound == 1)) {
479 tp->ecn_flags |= TE_ENABLE_ECN;
480 goto check_heuristic;
481 } else {
482 tp->ecn_flags &= ~TE_ENABLE_ECN;
483 }
484
485 return;
486
487 check_heuristic:
488 if (!tcp_heuristic_do_ecn(tp))
489 tp->ecn_flags &= ~TE_ENABLE_ECN;
490 }
491
492 /*
493 * Tcp output routine: figure out what should be sent and send it.
494 *
495 * Returns: 0 Success
496 * EADDRNOTAVAIL
497 * ENOBUFS
498 * EMSGSIZE
499 * EHOSTUNREACH
500 * ENETDOWN
501 * ip_output_list:ENOMEM
502 * ip_output_list:EADDRNOTAVAIL
503 * ip_output_list:ENETUNREACH
504 * ip_output_list:EHOSTUNREACH
505 * ip_output_list:EACCES
506 * ip_output_list:EMSGSIZE
507 * ip_output_list:ENOBUFS
508 * ip_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL]
509 * ip6_output_list:EINVAL
510 * ip6_output_list:EOPNOTSUPP
511 * ip6_output_list:EHOSTUNREACH
512 * ip6_output_list:EADDRNOTAVAIL
513 * ip6_output_list:ENETUNREACH
514 * ip6_output_list:EMSGSIZE
515 * ip6_output_list:ENOBUFS
516 * ip6_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL]
517 */
518 int
519 tcp_output(struct tcpcb *tp)
520 {
521 struct inpcb *inp = tp->t_inpcb;
522 struct socket *so = inp->inp_socket;
523 int32_t len, recwin, sendwin, off;
524 int flags, error;
525 struct mbuf *m;
526 struct ip *ip = NULL;
527 struct ipovly *ipov = NULL;
528 #if INET6
529 struct ip6_hdr *ip6 = NULL;
530 #endif /* INET6 */
531 struct tcphdr *th;
532 u_char opt[TCP_MAXOLEN];
533 unsigned ipoptlen, optlen, hdrlen;
534 int idle, sendalot, lost = 0;
535 int i, sack_rxmit;
536 int tso = 0;
537 int sack_bytes_rxmt;
538 tcp_seq old_snd_nxt = 0;
539 struct sackhole *p;
540 #if IPSEC
541 unsigned ipsec_optlen = 0;
542 #endif /* IPSEC */
543 int idle_time = 0;
544 struct mbuf *packetlist = NULL;
545 struct mbuf *tp_inp_options = inp->inp_depend4.inp4_options;
546 #if INET6
547 int isipv6 = inp->inp_vflag & INP_IPV6 ;
548 #endif
549 short packchain_listadd = 0;
550 int so_options = so->so_options;
551 struct rtentry *rt;
552 u_int32_t basertt, svc_flags = 0, allocated_len;
553 u_int32_t lro_ackmore = (tp->t_lropktlen != 0) ? 1 : 0;
554 struct mbuf *mnext = NULL;
555 int sackoptlen = 0;
556 #if MPTCP
557 unsigned int *dlenp = NULL;
558 u_int8_t *finp = NULL;
559 u_int32_t *sseqp = NULL;
560 u_int64_t dss_val = 0;
561 boolean_t mptcp_acknow = FALSE;
562 boolean_t early_data_sent = FALSE;
563 #endif /* MPTCP */
564 boolean_t cell = FALSE;
565 boolean_t wifi = FALSE;
566 boolean_t wired = FALSE;
567 boolean_t sack_rescue_rxt = FALSE;
568
569 /*
570 * Determine length of data that should be transmitted,
571 * and flags that will be used.
572 * If there is some data or critical controls (SYN, RST)
573 * to send, then transmit; otherwise, investigate further.
574 */
575 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
576
577 /* Since idle_time is signed integer, the following integer subtraction
578 * will take care of wrap around of tcp_now
579 */
580 idle_time = tcp_now - tp->t_rcvtime;
581 if (idle && idle_time >= TCP_IDLETIMEOUT(tp)) {
582 if (CC_ALGO(tp)->after_idle != NULL &&
583 (tp->tcp_cc_index != TCP_CC_ALGO_CUBIC_INDEX ||
584 idle_time >= TCP_CC_CWND_NONVALIDATED_PERIOD)) {
585 CC_ALGO(tp)->after_idle(tp);
586 tcp_ccdbg_trace(tp, NULL, TCP_CC_IDLE_TIMEOUT);
587 }
588
589 /*
590 * Do some other tasks that need to be done after
591 * idle time
592 */
593 if (!SLIST_EMPTY(&tp->t_rxt_segments))
594 tcp_rxtseg_clean(tp);
595
596 /* If stretch ack was auto-disabled, re-evaluate it */
597 tcp_cc_after_idle_stretchack(tp);
598 }
599 tp->t_flags &= ~TF_LASTIDLE;
600 if (idle) {
601 if (tp->t_flags & TF_MORETOCOME) {
602 tp->t_flags |= TF_LASTIDLE;
603 idle = 0;
604 }
605 }
606 #if MPTCP
607 if (tp->t_mpflags & TMPF_RESET) {
608 tcp_check_timer_state(tp);
609 /*
610 * Once a RST has been sent for an MPTCP subflow,
611 * the subflow socket stays around until deleted.
612 * No packets such as FINs must be sent after RST.
613 */
614 return (0);
615 }
616 #endif /* MPTCP */
617
618 again:
619 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_START, 0,0,0,0,0);
620
621 #if INET6
622 if (isipv6) {
623 KERNEL_DEBUG(DBG_LAYER_BEG,
624 ((inp->inp_fport << 16) | inp->inp_lport),
625 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
626 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
627 sendalot,0,0);
628 } else
629 #endif
630
631 {
632 KERNEL_DEBUG(DBG_LAYER_BEG,
633 ((inp->inp_fport << 16) | inp->inp_lport),
634 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
635 (inp->inp_faddr.s_addr & 0xffff)),
636 sendalot,0,0);
637 }
638 /*
639 * If the route generation id changed, we need to check that our
640 * local (source) IP address is still valid. If it isn't either
641 * return error or silently do nothing (assuming the address will
642 * come back before the TCP connection times out).
643 */
644 rt = inp->inp_route.ro_rt;
645 if (rt != NULL && ROUTE_UNUSABLE(&tp->t_inpcb->inp_route)) {
646 struct ifnet *ifp;
647 struct in_ifaddr *ia = NULL;
648 struct in6_ifaddr *ia6 = NULL;
649 int found_srcaddr = 0;
650
651 /* disable multipages at the socket */
652 somultipages(so, FALSE);
653
654 /* Disable TSO for the socket until we know more */
655 tp->t_flags &= ~TF_TSO;
656
657 soif2kcl(so, FALSE);
658
659 if (isipv6) {
660 ia6 = ifa_foraddr6(&inp->in6p_laddr);
661 if (ia6 != NULL)
662 found_srcaddr = 1;
663 } else {
664 ia = ifa_foraddr(inp->inp_laddr.s_addr);
665 if (ia != NULL)
666 found_srcaddr = 1;
667 }
668
669 /* check that the source address is still valid */
670 if (found_srcaddr == 0) {
671 soevent(so,
672 (SO_FILT_HINT_LOCKED | SO_FILT_HINT_NOSRCADDR));
673
674 if (tp->t_state >= TCPS_CLOSE_WAIT) {
675 tcp_drop(tp, EADDRNOTAVAIL);
676 return(EADDRNOTAVAIL);
677 }
678
679 /* Set retransmit timer if it wasn't set,
680 * reset Persist timer and shift register as the
681 * advertised peer window may not be valid anymore
682 */
683
684 if (!tp->t_timer[TCPT_REXMT]) {
685 tp->t_timer[TCPT_REXMT] =
686 OFFSET_FROM_START(tp, tp->t_rxtcur);
687 if (tp->t_timer[TCPT_PERSIST]) {
688 tp->t_timer[TCPT_PERSIST] = 0;
689 tp->t_rxtshift = 0;
690 tp->t_persist_stop = 0;
691 tp->t_rxtstart = 0;
692 }
693 }
694
695 if (tp->t_pktlist_head != NULL)
696 m_freem_list(tp->t_pktlist_head);
697 TCP_PKTLIST_CLEAR(tp);
698
699 /* drop connection if source address isn't available */
700 if (so->so_flags & SOF_NOADDRAVAIL) {
701 tcp_drop(tp, EADDRNOTAVAIL);
702 return(EADDRNOTAVAIL);
703 } else {
704 tcp_check_timer_state(tp);
705 return(0); /* silently ignore, keep data in socket: address may be back */
706 }
707 }
708 if (ia != NULL)
709 IFA_REMREF(&ia->ia_ifa);
710
711 if (ia6 != NULL)
712 IFA_REMREF(&ia6->ia_ifa);
713
714 /*
715 * Address is still valid; check for multipages capability
716 * again in case the outgoing interface has changed.
717 */
718 RT_LOCK(rt);
719 if ((ifp = rt->rt_ifp) != NULL) {
720 somultipages(so, (ifp->if_hwassist & IFNET_MULTIPAGES));
721 tcp_set_tso(tp, ifp);
722 soif2kcl(so, (ifp->if_eflags & IFEF_2KCL));
723 tcp_set_ecn(tp, ifp);
724 }
725 if (rt->rt_flags & RTF_UP)
726 RT_GENID_SYNC(rt);
727 /*
728 * See if we should do MTU discovery. Don't do it if:
729 * 1) it is disabled via the sysctl
730 * 2) the route isn't up
731 * 3) the MTU is locked (if it is, then discovery
732 * has been disabled)
733 */
734
735 if (!path_mtu_discovery || ((rt != NULL) &&
736 (!(rt->rt_flags & RTF_UP) ||
737 (rt->rt_rmx.rmx_locks & RTV_MTU))))
738 tp->t_flags &= ~TF_PMTUD;
739 else
740 tp->t_flags |= TF_PMTUD;
741
742 RT_UNLOCK(rt);
743 }
744
745 if (rt != NULL) {
746 cell = IFNET_IS_CELLULAR(rt->rt_ifp);
747 wifi = (!cell && IFNET_IS_WIFI(rt->rt_ifp));
748 wired = (!wifi && IFNET_IS_WIRED(rt->rt_ifp));
749 }
750
751 /*
752 * If we've recently taken a timeout, snd_max will be greater than
753 * snd_nxt. There may be SACK information that allows us to avoid
754 * resending already delivered data. Adjust snd_nxt accordingly.
755 */
756 if (SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
757 tcp_sack_adjust(tp);
758 sendalot = 0;
759 off = tp->snd_nxt - tp->snd_una;
760 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
761
762 if (tp->t_flags & TF_SLOWLINK && slowlink_wsize > 0)
763 sendwin = min(sendwin, slowlink_wsize);
764
765 flags = tcp_outflags[tp->t_state];
766 /*
767 * Send any SACK-generated retransmissions. If we're explicitly
768 * trying to send out new data (when sendalot is 1), bypass this
769 * function. If we retransmit in fast recovery mode, decrement
770 * snd_cwnd, since we're replacing a (future) new transmission
771 * with a retransmission now, and we previously incremented
772 * snd_cwnd in tcp_input().
773 */
774 /*
775 * Still in sack recovery , reset rxmit flag to zero.
776 */
777 sack_rxmit = 0;
778 sack_bytes_rxmt = 0;
779 len = 0;
780 p = NULL;
781 if (SACK_ENABLED(tp) && IN_FASTRECOVERY(tp) &&
782 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
783 int32_t cwin;
784
785 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
786 if (cwin < 0)
787 cwin = 0;
788 /* Do not retransmit SACK segments beyond snd_recover */
789 if (SEQ_GT(p->end, tp->snd_recover)) {
790 /*
791 * (At least) part of sack hole extends beyond
792 * snd_recover. Check to see if we can rexmit data
793 * for this hole.
794 */
795 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
796 /*
797 * Can't rexmit any more data for this hole.
798 * That data will be rexmitted in the next
799 * sack recovery episode, when snd_recover
800 * moves past p->rxmit.
801 */
802 p = NULL;
803 goto after_sack_rexmit;
804 } else
805 /* Can rexmit part of the current hole */
806 len = ((int32_t)min(cwin,
807 tp->snd_recover - p->rxmit));
808 } else {
809 len = ((int32_t)min(cwin, p->end - p->rxmit));
810 }
811 if (len > 0) {
812 off = p->rxmit - tp->snd_una;
813 sack_rxmit = 1;
814 sendalot = 1;
815 tcpstat.tcps_sack_rexmits++;
816 tcpstat.tcps_sack_rexmit_bytes +=
817 min(len, tp->t_maxseg);
818 if (nstat_collect) {
819 nstat_route_tx(inp->inp_route.ro_rt, 1,
820 min(len, tp->t_maxseg),
821 NSTAT_TX_FLAG_RETRANSMIT);
822 INP_ADD_STAT(inp, cell, wifi, wired,
823 txpackets, 1);
824 INP_ADD_STAT(inp, cell, wifi, wired,
825 txbytes, min(len, tp->t_maxseg));
826 tp->t_stat.txretransmitbytes += min(len, tp->t_maxseg);
827 }
828 } else {
829 len = 0;
830 }
831 }
832 after_sack_rexmit:
833 /*
834 * Get standard flags, and add SYN or FIN if requested by 'hidden'
835 * state flags.
836 */
837 if (tp->t_flags & TF_NEEDFIN)
838 flags |= TH_FIN;
839 if (tp->t_flags & TF_NEEDSYN)
840 flags |= TH_SYN;
841
842 /*
843 * If in persist timeout with window of 0, send 1 byte.
844 * Otherwise, if window is small but nonzero
845 * and timer expired, we will send what we can
846 * and go to transmit state.
847 */
848 if (tp->t_flagsext & TF_FORCE) {
849 if (sendwin == 0) {
850 /*
851 * If we still have some data to send, then
852 * clear the FIN bit. Usually this would
853 * happen below when it realizes that we
854 * aren't sending all the data. However,
855 * if we have exactly 1 byte of unsent data,
856 * then it won't clear the FIN bit below,
857 * and if we are in persist state, we wind
858 * up sending the packet without recording
859 * that we sent the FIN bit.
860 *
861 * We can't just blindly clear the FIN bit,
862 * because if we don't have any more data
863 * to send then the probe will be the FIN
864 * itself.
865 */
866 if (off < so->so_snd.sb_cc)
867 flags &= ~TH_FIN;
868 sendwin = 1;
869 } else {
870 tp->t_timer[TCPT_PERSIST] = 0;
871 tp->t_rxtshift = 0;
872 tp->t_rxtstart = 0;
873 tp->t_persist_stop = 0;
874 }
875 }
876
877 /*
878 * If snd_nxt == snd_max and we have transmitted a FIN, the
879 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
880 * a negative length. This can also occur when TCP opens up
881 * its congestion window while receiving additional duplicate
882 * acks after fast-retransmit because TCP will reset snd_nxt
883 * to snd_max after the fast-retransmit.
884 *
885 * In the normal retransmit-FIN-only case, however, snd_nxt will
886 * be set to snd_una, the offset will be 0, and the length may
887 * wind up 0.
888 *
889 * If sack_rxmit is true we are retransmitting from the scoreboard
890 * in which case len is already set.
891 */
892 if (sack_rxmit == 0) {
893 if (sack_bytes_rxmt == 0) {
894 len = min(so->so_snd.sb_cc, sendwin) - off;
895 } else {
896 int32_t cwin;
897
898 cwin = tp->snd_cwnd -
899 (tp->snd_nxt - tp->sack_newdata) -
900 sack_bytes_rxmt;
901 if (cwin < 0)
902 cwin = 0;
903 /*
904 * We are inside of a SACK recovery episode and are
905 * sending new data, having retransmitted all the
906 * data possible in the scoreboard.
907 */
908 len = min(so->so_snd.sb_cc, tp->snd_wnd)
909 - off;
910 /*
911 * Don't remove this (len > 0) check !
912 * We explicitly check for len > 0 here (although it
913 * isn't really necessary), to work around a gcc
914 * optimization issue - to force gcc to compute
915 * len above. Without this check, the computation
916 * of len is bungled by the optimizer.
917 */
918 if (len > 0) {
919 len = imin(len, cwin);
920 } else {
921 len = 0;
922 }
923 /*
924 * At this point SACK recovery can not send any
925 * data from scoreboard or any new data. Check
926 * if we can do a rescue retransmit towards the
927 * tail end of recovery window.
928 */
929 if (len == 0 && cwin > 0 &&
930 SEQ_LT(tp->snd_fack, tp->snd_recover) &&
931 !(tp->t_flagsext & TF_RESCUE_RXT)) {
932 len = min((tp->snd_recover - tp->snd_fack),
933 tp->t_maxseg);
934 len = imin(len, cwin);
935 old_snd_nxt = tp->snd_nxt;
936 sack_rescue_rxt = TRUE;
937 tp->snd_nxt = tp->snd_recover - len;
938 /*
939 * If FIN has been sent, snd_max
940 * must have been advanced to cover it.
941 */
942 if ((tp->t_flags & TF_SENTFIN) &&
943 tp->snd_max == tp->snd_recover)
944 tp->snd_nxt--;
945
946 off = tp->snd_nxt - tp->snd_una;
947 sendalot = 0;
948 tp->t_flagsext |= TF_RESCUE_RXT;
949 }
950 }
951 }
952
953 #if MPTCP
954 if ((tp->t_mpflags & TMPF_FASTJOIN_SEND) &&
955 (tp->t_state == TCPS_SYN_SENT) &&
956 (!(tp->t_flags & TF_CLOSING)) &&
957 (so->so_snd.sb_cc != 0) &&
958 (tp->t_rxtshift == 0)) {
959 flags &= ~TH_SYN;
960 flags |= TH_ACK;
961 off = 0;
962 len = min(so->so_snd.sb_cc, tp->t_maxseg);
963 early_data_sent = TRUE;
964 } else if (early_data_sent) {
965 /* for now, we allow only one data segment to be sent */
966 return (0);
967 }
968 #endif /* MPTCP */
969 /*
970 * Lop off SYN bit if it has already been sent. However, if this
971 * is SYN-SENT state and if segment contains data and if we don't
972 * know that foreign host supports TAO, suppress sending segment.
973 */
974 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
975 if (tp->t_state != TCPS_SYN_RECEIVED || tfo_enabled(tp))
976 flags &= ~TH_SYN;
977 off--, len++;
978 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
979 while (inp->inp_sndinprog_cnt == 0 &&
980 tp->t_pktlist_head != NULL) {
981 packetlist = tp->t_pktlist_head;
982 packchain_listadd = tp->t_lastchain;
983 packchain_sent++;
984 TCP_PKTLIST_CLEAR(tp);
985
986 error = tcp_ip_output(so, tp, packetlist,
987 packchain_listadd, tp_inp_options,
988 (so_options & SO_DONTROUTE),
989 (sack_rxmit | (sack_bytes_rxmt != 0)), 0,
990 #if INET6
991 isipv6);
992 #else /* INET6 */
993 0);
994 #endif /* !INET6 */
995
996
997 }
998
999 /*
1000 * tcp was closed while we were in ip,
1001 * resume close
1002 */
1003 if (inp->inp_sndinprog_cnt == 0 &&
1004 (tp->t_flags & TF_CLOSING)) {
1005 tp->t_flags &= ~TF_CLOSING;
1006 (void) tcp_close(tp);
1007 } else {
1008 tcp_check_timer_state(tp);
1009 }
1010 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,
1011 0,0,0,0,0);
1012 return(0);
1013 }
1014 }
1015
1016 /*
1017 * Be careful not to send data and/or FIN on SYN segments.
1018 * This measure is needed to prevent interoperability problems
1019 * with not fully conformant TCP implementations.
1020 *
1021 * In case of TFO, we handle the setting of the len in
1022 * tcp_tfo_check. In case TFO is not enabled, never ever send
1023 * SYN+data.
1024 */
1025 if ((flags & TH_SYN) && !tfo_enabled(tp)) {
1026 len = 0;
1027 flags &= ~TH_FIN;
1028 }
1029
1030 if ((flags & TH_SYN) && tp->t_state <= TCPS_SYN_SENT && tfo_enabled(tp))
1031 len = tcp_tfo_check(tp, len);
1032
1033 /*
1034 * The check here used to be (len < 0). Some times len is zero
1035 * when the congestion window is closed and we need to check
1036 * if persist timer has to be set in that case. But don't set
1037 * persist until connection is established.
1038 */
1039 if (len <= 0 && !(flags & TH_SYN)) {
1040 /*
1041 * If FIN has been sent but not acked,
1042 * but we haven't been called to retransmit,
1043 * len will be < 0. Otherwise, window shrank
1044 * after we sent into it. If window shrank to 0,
1045 * cancel pending retransmit, pull snd_nxt back
1046 * to (closed) window, and set the persist timer
1047 * if it isn't already going. If the window didn't
1048 * close completely, just wait for an ACK.
1049 */
1050 len = 0;
1051 if (sendwin == 0) {
1052 tp->t_timer[TCPT_REXMT] = 0;
1053 tp->t_timer[TCPT_PTO] = 0;
1054 tp->t_rxtshift = 0;
1055 tp->t_rxtstart = 0;
1056 tp->snd_nxt = tp->snd_una;
1057 off = 0;
1058 if (tp->t_timer[TCPT_PERSIST] == 0)
1059 tcp_setpersist(tp);
1060 }
1061 }
1062
1063 /*
1064 * Automatic sizing of send socket buffer. Increase the send
1065 * socket buffer size if all of the following criteria are met
1066 * 1. the receiver has enough buffer space for this data
1067 * 2. send buffer is filled to 7/8th with data (so we actually
1068 * have data to make use of it);
1069 * 3. our send window (slow start and congestion controlled) is
1070 * larger than sent but unacknowledged data in send buffer.
1071 */
1072 basertt = get_base_rtt(tp);
1073 if (tcp_do_autosendbuf == 1 &&
1074 !INP_WAIT_FOR_IF_FEEDBACK(inp) && !IN_FASTRECOVERY(tp) &&
1075 (so->so_snd.sb_flags & (SB_AUTOSIZE | SB_TRIM)) == SB_AUTOSIZE &&
1076 tcp_cansbgrow(&so->so_snd)) {
1077 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
1078 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
1079 sendwin >= (so->so_snd.sb_cc -
1080 (tp->snd_nxt - tp->snd_una))) {
1081 /* Also increase the send buffer only if the
1082 * round-trip time is not increasing because we do
1083 * not want to contribute to latency by filling
1084 * buffers.
1085 * We also do not want to hold onto application's
1086 * old data for too long. Interactive applications
1087 * would rather discard old data.
1088 */
1089 if (tp->t_rttcur <= (basertt + 25)) {
1090 if (sbreserve(&so->so_snd,
1091 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
1092 tcp_autosndbuf_max)) == 1) {
1093 so->so_snd.sb_idealsize = so->so_snd.sb_hiwat;
1094 }
1095 } else {
1096 so->so_snd.sb_idealsize =
1097 max(tcp_sendspace, so->so_snd.sb_hiwat -
1098 (2 * tcp_autosndbuf_inc));
1099 so->so_snd.sb_flags |= SB_TRIM;
1100 }
1101 }
1102 }
1103
1104 /*
1105 * Truncate to the maximum segment length or enable TCP Segmentation
1106 * Offloading (if supported by hardware) and ensure that FIN is removed
1107 * if the length no longer contains the last data byte.
1108 *
1109 * TSO may only be used if we are in a pure bulk sending state.
1110 * The presence of TCP-MD5, SACK retransmits, SACK advertizements,
1111 * ipfw rules and IP options, as well as disabling hardware checksum
1112 * offload prevent using TSO. With TSO the TCP header is the same
1113 * (except for the sequence number) for all generated packets. This
1114 * makes it impossible to transmit any options which vary per generated
1115 * segment or packet.
1116 *
1117 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
1118 * removal of FIN (if not already catched here) are handled later after
1119 * the exact length of the TCP options are known.
1120 */
1121 #if IPSEC
1122 /*
1123 * Pre-calculate here as we save another lookup into the darknesses
1124 * of IPsec that way and can actually decide if TSO is ok.
1125 */
1126 if (ipsec_bypass == 0)
1127 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
1128 #endif
1129 if (len > tp->t_maxseg) {
1130 if ((tp->t_flags & TF_TSO) && tcp_do_tso && hwcksum_tx &&
1131 ip_use_randomid && kipf_count == 0 &&
1132 dlil_filter_disable_tso_count == 0 &&
1133 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
1134 sack_bytes_rxmt == 0 &&
1135 inp->inp_options == NULL &&
1136 inp->in6p_options == NULL
1137 #if IPSEC
1138 && ipsec_optlen == 0
1139 #endif
1140 #if IPFIREWALL
1141 && (fw_enable == 0 || fw_bypass)
1142 #endif
1143 ) {
1144 tso = 1;
1145 sendalot = 0;
1146 } else {
1147 len = tp->t_maxseg;
1148 sendalot = 1;
1149 tso = 0;
1150 }
1151 }
1152
1153 /* Send one segment or less as a tail loss probe */
1154 if (tp->t_flagsext & TF_SENT_TLPROBE) {
1155 len = min(len, tp->t_maxseg);
1156 sendalot = 0;
1157 tso = 0;
1158 }
1159
1160 #if MPTCP
1161 if ((so->so_flags & SOF_MP_SUBFLOW) &&
1162 !(tp->t_mpflags & TMPF_TCP_FALLBACK)) {
1163 int newlen = len;
1164 if (!(tp->t_mpflags & TMPF_PREESTABLISHED) &&
1165 (tp->t_state > TCPS_CLOSED) &&
1166 ((tp->t_mpflags & TMPF_SND_MPPRIO) ||
1167 (tp->t_mpflags & TMPF_SND_REM_ADDR) ||
1168 (tp->t_mpflags & TMPF_SND_MPFAIL) ||
1169 (tp->t_mpflags & TMPF_MPCAP_RETRANSMIT))) {
1170 if (len > 0) {
1171 len = 0;
1172 }
1173 sendalot = 1;
1174 mptcp_acknow = TRUE;
1175 } else {
1176 mptcp_acknow = FALSE;
1177 }
1178 /*
1179 * The contiguous bytes in the subflow socket buffer can be
1180 * discontiguous at the MPTCP level. Since only one DSS
1181 * option can be sent in one packet, reduce length to match
1182 * the contiguous MPTCP level. Set sendalot to send remainder.
1183 */
1184 if (len > 0)
1185 newlen = mptcp_adj_sendlen(so, off, len);
1186 if (newlen < len) {
1187 len = newlen;
1188 sendalot = 1;
1189 }
1190 }
1191 #endif /* MPTCP */
1192
1193 /*
1194 * If the socket is capable of doing unordered send,
1195 * pull the amount of data that can be sent from the
1196 * unordered priority queues to the serial queue in
1197 * the socket buffer. If bytes are not yet available
1198 * in the highest priority message, we may not be able
1199 * to send any new data.
1200 */
1201 if (so->so_flags & SOF_ENABLE_MSGS) {
1202 if ((off + len) >
1203 so->so_msg_state->msg_serial_bytes) {
1204 sbpull_unordered_data(so, off, len);
1205
1206 /* check if len needs to be modified */
1207 if ((off + len) >
1208 so->so_msg_state->msg_serial_bytes) {
1209 len = so->so_msg_state->msg_serial_bytes - off;
1210 if (len <= 0) {
1211 len = 0;
1212 tcpstat.tcps_msg_sndwaithipri++;
1213 }
1214 }
1215 }
1216 }
1217
1218 if (sack_rxmit) {
1219 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
1220 flags &= ~TH_FIN;
1221 } else {
1222 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
1223 flags &= ~TH_FIN;
1224 }
1225
1226 recwin = tcp_sbspace(tp);
1227
1228 /*
1229 * Sender silly window avoidance. We transmit under the following
1230 * conditions when len is non-zero:
1231 *
1232 * - we've timed out (e.g. persist timer)
1233 * - we need to retransmit
1234 * - We have a full segment (or more with TSO)
1235 * - This is the last buffer in a write()/send() and we are
1236 * either idle or running NODELAY
1237 * - we have more then 1/2 the maximum send window's worth of
1238 * data (receiver may be limited the window size)
1239 */
1240 if (len) {
1241 if (tp->t_flagsext & TF_FORCE)
1242 goto send;
1243 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1244 goto send;
1245 if (sack_rxmit)
1246 goto send;
1247
1248 /*
1249 * Send new data on the connection only if it is
1250 * not flow controlled
1251 */
1252 if (!INP_WAIT_FOR_IF_FEEDBACK(inp) ||
1253 tp->t_state != TCPS_ESTABLISHED) {
1254 if (len >= tp->t_maxseg)
1255 goto send;
1256 if (!(tp->t_flags & TF_MORETOCOME) &&
1257 (idle || tp->t_flags & TF_NODELAY ||
1258 tp->t_flags & TF_MAXSEGSNT ||
1259 ALLOW_LIMITED_TRANSMIT(tp)) &&
1260 (tp->t_flags & TF_NOPUSH) == 0 &&
1261 len + off >= so->so_snd.sb_cc)
1262 goto send;
1263 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
1264 goto send;
1265 } else {
1266 tcpstat.tcps_fcholdpacket++;
1267 }
1268 }
1269
1270 /*
1271 * Compare available window to amount of window
1272 * known to peer (as advertised window less
1273 * next expected input). If the difference is at least two
1274 * max size segments, or at least 25% of the maximum possible
1275 * window, then want to send a window update to peer.
1276 * Skip this if the connection is in T/TCP half-open state.
1277 */
1278 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
1279 /*
1280 * "adv" is the amount we can increase the window,
1281 * taking into account that we are limited by
1282 * TCP_MAXWIN << tp->rcv_scale.
1283 */
1284 int32_t adv, oldwin = 0;
1285 adv = imin(recwin, (int)TCP_MAXWIN << tp->rcv_scale) -
1286 (tp->rcv_adv - tp->rcv_nxt);
1287
1288 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
1289 oldwin = tp->rcv_adv - tp->rcv_nxt;
1290
1291 if (adv >= (int32_t) (2 * tp->t_maxseg)) {
1292 /*
1293 * Update only if the resulting scaled value of
1294 * the window changed, or if there is a change in
1295 * the sequence since the last ack. This avoids
1296 * what appears as dupe ACKS (see rdar://5640997)
1297 *
1298 * If streaming is detected avoid sending too many
1299 * window updates. We will depend on the delack
1300 * timer to send a window update when needed.
1301 */
1302 if (!(tp->t_flags & TF_STRETCHACK) &&
1303 (tp->last_ack_sent != tp->rcv_nxt ||
1304 ((oldwin + adv) >> tp->rcv_scale) >
1305 (oldwin >> tp->rcv_scale))) {
1306 goto send;
1307 }
1308
1309 /*
1310 * Make sure that the delayed ack timer is set if
1311 * we delayed sending a window update because of
1312 * streaming detection.
1313 */
1314 if ((tp->t_flags & TF_STRETCHACK) &&
1315 !(tp->t_flags & TF_DELACK)) {
1316 tp->t_flags |= TF_DELACK;
1317 tp->t_timer[TCPT_DELACK] =
1318 OFFSET_FROM_START(tp, tcp_delack);
1319 }
1320 }
1321 if (4 * adv >= (int32_t) so->so_rcv.sb_hiwat)
1322 goto send;
1323 }
1324
1325 /*
1326 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
1327 * is also a catch-all for the retransmit timer timeout case.
1328 */
1329 if (tp->t_flags & TF_ACKNOW)
1330 goto send;
1331 if ((flags & TH_RST) ||
1332 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
1333 goto send;
1334 if (SEQ_GT(tp->snd_up, tp->snd_una))
1335 goto send;
1336 #if MPTCP
1337 if (mptcp_acknow)
1338 goto send;
1339 #endif /* MPTCP */
1340 /*
1341 * If our state indicates that FIN should be sent
1342 * and we have not yet done so, then we need to send.
1343 */
1344 if ((flags & TH_FIN) &&
1345 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
1346 goto send;
1347 /*
1348 * In SACK, it is possible for tcp_output to fail to send a segment
1349 * after the retransmission timer has been turned off. Make sure
1350 * that the retransmission timer is set.
1351 */
1352 if (SACK_ENABLED(tp) && (tp->t_state >= TCPS_ESTABLISHED) &&
1353 SEQ_GT(tp->snd_max, tp->snd_una) &&
1354 tp->t_timer[TCPT_REXMT] == 0 &&
1355 tp->t_timer[TCPT_PERSIST] == 0) {
1356 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp,
1357 tp->t_rxtcur);
1358 goto just_return;
1359 }
1360 /*
1361 * TCP window updates are not reliable, rather a polling protocol
1362 * using ``persist'' packets is used to insure receipt of window
1363 * updates. The three ``states'' for the output side are:
1364 * idle not doing retransmits or persists
1365 * persisting to move a small or zero window
1366 * (re)transmitting and thereby not persisting
1367 *
1368 * tp->t_timer[TCPT_PERSIST]
1369 * is set when we are in persist state.
1370 * tp->t_force
1371 * is set when we are called to send a persist packet.
1372 * tp->t_timer[TCPT_REXMT]
1373 * is set when we are retransmitting
1374 * The output side is idle when both timers are zero.
1375 *
1376 * If send window is too small, there is data to transmit, and no
1377 * retransmit or persist is pending, then go to persist state.
1378 * If nothing happens soon, send when timer expires:
1379 * if window is nonzero, transmit what we can,
1380 * otherwise force out a byte.
1381 */
1382 if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
1383 tp->t_timer[TCPT_PERSIST] == 0) {
1384 tp->t_rxtshift = 0;
1385 tp->t_rxtstart = 0;
1386 tcp_setpersist(tp);
1387 }
1388 just_return:
1389 /*
1390 * If there is no reason to send a segment, just return.
1391 * but if there is some packets left in the packet list, send them now.
1392 */
1393 while (inp->inp_sndinprog_cnt == 0 &&
1394 tp->t_pktlist_head != NULL) {
1395 packetlist = tp->t_pktlist_head;
1396 packchain_listadd = tp->t_lastchain;
1397 packchain_sent++;
1398 TCP_PKTLIST_CLEAR(tp);
1399
1400 error = tcp_ip_output(so, tp, packetlist,
1401 packchain_listadd,
1402 tp_inp_options, (so_options & SO_DONTROUTE),
1403 (sack_rxmit | (sack_bytes_rxmt != 0)), recwin,
1404 #if INET6
1405 isipv6);
1406 #else /* INET6 */
1407 0);
1408 #endif /* !INET6 */
1409 }
1410 /* tcp was closed while we were in ip; resume close */
1411 if (inp->inp_sndinprog_cnt == 0 &&
1412 (tp->t_flags & TF_CLOSING)) {
1413 tp->t_flags &= ~TF_CLOSING;
1414 (void) tcp_close(tp);
1415 } else {
1416 tcp_check_timer_state(tp);
1417 }
1418 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
1419 return (0);
1420
1421 send:
1422 /*
1423 * Set TF_MAXSEGSNT flag if the segment size is greater than
1424 * the max segment size.
1425 */
1426 if (len > 0) {
1427 if (len >= tp->t_maxseg)
1428 tp->t_flags |= TF_MAXSEGSNT;
1429 else
1430 tp->t_flags &= ~TF_MAXSEGSNT;
1431 }
1432 /*
1433 * Before ESTABLISHED, force sending of initial options
1434 * unless TCP set not to do any options.
1435 * NOTE: we assume that the IP/TCP header plus TCP options
1436 * always fit in a single mbuf, leaving room for a maximum
1437 * link header, i.e.
1438 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1439 */
1440 optlen = 0;
1441 #if INET6
1442 if (isipv6)
1443 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
1444 else
1445 #endif
1446 hdrlen = sizeof (struct tcpiphdr);
1447 if (flags & TH_SYN) {
1448 tp->snd_nxt = tp->iss;
1449 if ((tp->t_flags & TF_NOOPT) == 0) {
1450 u_short mss;
1451
1452 opt[0] = TCPOPT_MAXSEG;
1453 opt[1] = TCPOLEN_MAXSEG;
1454 mss = htons((u_short) tcp_mssopt(tp));
1455 (void)memcpy(opt + 2, &mss, sizeof(mss));
1456 optlen = TCPOLEN_MAXSEG;
1457
1458 if ((tp->t_flags & TF_REQ_SCALE) &&
1459 ((flags & TH_ACK) == 0 ||
1460 (tp->t_flags & TF_RCVD_SCALE))) {
1461 *((u_int32_t *)(void *)(opt + optlen)) = htonl(
1462 TCPOPT_NOP << 24 |
1463 TCPOPT_WINDOW << 16 |
1464 TCPOLEN_WINDOW << 8 |
1465 tp->request_r_scale);
1466 optlen += 4;
1467 }
1468 #if MPTCP
1469 if (mptcp_enable) {
1470 optlen = mptcp_setup_syn_opts(so, flags, opt,
1471 optlen);
1472 }
1473 #endif /* MPTCP */
1474 }
1475 }
1476
1477 /*
1478 * Send a timestamp and echo-reply if this is a SYN and our side
1479 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1480 * and our peer have sent timestamps in our SYN's.
1481 */
1482 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1483 (flags & TH_RST) == 0 &&
1484 ((flags & TH_ACK) == 0 ||
1485 (tp->t_flags & TF_RCVD_TSTMP))) {
1486 u_int32_t *lp = (u_int32_t *)(void *)(opt + optlen);
1487
1488 /* Form timestamp option as shown in appendix A of RFC 1323. */
1489 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1490 *lp++ = htonl(tcp_now);
1491 *lp = htonl(tp->ts_recent);
1492 optlen += TCPOLEN_TSTAMP_APPA;
1493 }
1494
1495 /* Note the timestamp for receive buffer autosizing */
1496 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1497 tp->rfbuf_ts = tcp_now;
1498
1499 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1500 /*
1501 * Tack on the SACK permitted option *last*.
1502 * And do padding of options after tacking this on.
1503 * This is because of MSS, TS, WinScale and Signatures are
1504 * all present, we have just 2 bytes left for the SACK
1505 * permitted option, which is just enough.
1506 */
1507 /*
1508 * If this is the first SYN of connection (not a SYN
1509 * ACK), include SACK permitted option. If this is a
1510 * SYN ACK, include SACK permitted option if peer has
1511 * already done so. This is only for active connect,
1512 * since the syncache takes care of the passive connect.
1513 */
1514 if ((flags & TH_SYN) &&
1515 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) {
1516 u_char *bp;
1517 bp = (u_char *)opt + optlen;
1518
1519 *bp++ = TCPOPT_SACK_PERMITTED;
1520 *bp++ = TCPOLEN_SACK_PERMITTED;
1521 optlen += TCPOLEN_SACK_PERMITTED;
1522 }
1523 }
1524 #if MPTCP
1525 if (so->so_flags & SOF_MP_SUBFLOW) {
1526 /*
1527 * Its important to piggyback acks with data as ack only packets
1528 * may get lost and data packets that don't send Data ACKs
1529 * still advance the subflow level ACK and therefore make it
1530 * hard for the remote end to recover in low cwnd situations.
1531 */
1532 if (len != 0) {
1533 tp->t_mpflags |= (TMPF_SEND_DSN |
1534 TMPF_MPTCP_ACKNOW);
1535 } else {
1536 tp->t_mpflags |= TMPF_MPTCP_ACKNOW;
1537 }
1538 optlen = mptcp_setup_opts(tp, off, &opt[0], optlen, flags,
1539 len, &dlenp, &finp, &dss_val, &sseqp, &mptcp_acknow);
1540 tp->t_mpflags &= ~TMPF_SEND_DSN;
1541 }
1542 #endif /* MPTCP */
1543
1544 if (tfo_enabled(tp) && !(tp->t_flags & TF_NOOPT) &&
1545 (flags & (TH_SYN | TH_ACK)) == TH_SYN)
1546 optlen += tcp_tfo_write_cookie(tp, optlen, &len, opt);
1547
1548 if (tfo_enabled(tp) &&
1549 (flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1550 (tp->t_tfo_flags & TFO_F_OFFER_COOKIE))
1551 optlen += tcp_tfo_write_cookie_rep(tp, optlen, opt);
1552
1553 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1554 /*
1555 * Send SACKs if necessary. This should be the last
1556 * option processed. Only as many SACKs are sent as
1557 * are permitted by the maximum options size.
1558 *
1559 * In general, SACK blocks consume 8*n+2 bytes.
1560 * So a full size SACK blocks option is 34 bytes
1561 * (to generate 4 SACK blocks). At a minimum,
1562 * we need 10 bytes (to generate 1 SACK block).
1563 * If TCP Timestamps (12 bytes) and TCP Signatures
1564 * (18 bytes) are both present, we'll just have
1565 * 10 bytes for SACK options 40 - (12 + 18).
1566 */
1567 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1568 (tp->t_flags & TF_SACK_PERMIT) &&
1569 (tp->rcv_numsacks > 0 || TCP_SEND_DSACK_OPT(tp)) &&
1570 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) {
1571 int nsack, padlen;
1572 u_char *bp = (u_char *)opt + optlen;
1573 u_int32_t *lp;
1574
1575 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK;
1576 nsack = min(nsack, (tp->rcv_numsacks +
1577 (TCP_SEND_DSACK_OPT(tp) ? 1 : 0)));
1578 sackoptlen = (2 + nsack * TCPOLEN_SACK);
1579
1580 /*
1581 * First we need to pad options so that the
1582 * SACK blocks can start at a 4-byte boundary
1583 * (sack option and length are at a 2 byte offset).
1584 */
1585 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4;
1586 optlen += padlen;
1587 while (padlen-- > 0)
1588 *bp++ = TCPOPT_NOP;
1589
1590 tcpstat.tcps_sack_send_blocks++;
1591 *bp++ = TCPOPT_SACK;
1592 *bp++ = sackoptlen;
1593 lp = (u_int32_t *)(void *)bp;
1594
1595 /*
1596 * First block of SACK option should represent
1597 * DSACK. Prefer to send SACK information if there
1598 * is space for only one SACK block. This will
1599 * allow for faster recovery.
1600 */
1601 if (TCP_SEND_DSACK_OPT(tp) && nsack > 0 &&
1602 (tp->rcv_numsacks == 0 || nsack > 1)) {
1603 *lp++ = htonl(tp->t_dsack_lseq);
1604 *lp++ = htonl(tp->t_dsack_rseq);
1605 tcpstat.tcps_dsack_sent++;
1606 tp->t_dsack_sent++;
1607 nsack--;
1608 }
1609 VERIFY(nsack == 0 || tp->rcv_numsacks >= nsack);
1610 for (i = 0; i < nsack; i++) {
1611 struct sackblk sack = tp->sackblks[i];
1612 *lp++ = htonl(sack.start);
1613 *lp++ = htonl(sack.end);
1614 }
1615 optlen += sackoptlen;
1616 }
1617 }
1618
1619 /* Pad TCP options to a 4 byte boundary */
1620 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) {
1621 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t));
1622 u_char *bp = (u_char *)opt + optlen;
1623
1624 optlen += pad;
1625 while (pad) {
1626 *bp++ = TCPOPT_EOL;
1627 pad--;
1628 }
1629 }
1630
1631 /*
1632 * RFC 3168 states that:
1633 * - If you ever sent an ECN-setup SYN/SYN-ACK you must be prepared
1634 * to handle the TCP ECE flag, even if you also later send a
1635 * non-ECN-setup SYN/SYN-ACK.
1636 * - If you ever send a non-ECN-setup SYN/SYN-ACK, you must not set
1637 * the ip ECT flag.
1638 *
1639 * It is not clear how the ECE flag would ever be set if you never
1640 * set the IP ECT flag on outbound packets. All the same, we use
1641 * the TE_SETUPSENT to indicate that we have committed to handling
1642 * the TCP ECE flag correctly. We use the TE_SENDIPECT to indicate
1643 * whether or not we should set the IP ECT flag on outbound packet
1644 *
1645 * For a SYN-ACK, send an ECN setup SYN-ACK
1646 */
1647 if ((flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1648 (tp->ecn_flags & TE_ENABLE_ECN)) {
1649 if (tp->ecn_flags & TE_SETUPRECEIVED) {
1650 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1651 /*
1652 * Setting TH_ECE makes this an ECN-setup
1653 * SYN-ACK
1654 */
1655 flags |= TH_ECE;
1656
1657 /*
1658 * Record that we sent the ECN-setup and
1659 * default to setting IP ECT.
1660 */
1661 tp->ecn_flags |= (TE_SETUPSENT|TE_SENDIPECT);
1662 tcpstat.tcps_ecn_server_setup++;
1663 tcpstat.tcps_ecn_server_success++;
1664 } else {
1665 /*
1666 * We sent an ECN-setup SYN-ACK but it was
1667 * dropped. Fallback to non-ECN-setup
1668 * SYN-ACK and clear flag to indicate that
1669 * we should not send data with IP ECT set
1670 *
1671 * Pretend we didn't receive an
1672 * ECN-setup SYN.
1673 *
1674 * We already incremented the counter
1675 * assuming that the ECN setup will
1676 * succeed. Decrementing here
1677 * tcps_ecn_server_success to correct it.
1678 */
1679 if (tp->ecn_flags & TE_SETUPSENT) {
1680 tcpstat.tcps_ecn_lost_synack++;
1681 tcpstat.tcps_ecn_server_success--;
1682 tp->ecn_flags |= TE_LOST_SYNACK;
1683 }
1684
1685 tp->ecn_flags &=
1686 ~(TE_SETUPRECEIVED | TE_SENDIPECT |
1687 TE_SENDCWR);
1688 }
1689 }
1690 } else if ((flags & (TH_SYN | TH_ACK)) == TH_SYN &&
1691 (tp->ecn_flags & TE_ENABLE_ECN)) {
1692 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1693 /*
1694 * Setting TH_ECE and TH_CWR makes this an
1695 * ECN-setup SYN
1696 */
1697 flags |= (TH_ECE | TH_CWR);
1698 tcpstat.tcps_ecn_client_setup++;
1699 tp->ecn_flags |= TE_CLIENT_SETUP;
1700
1701 /*
1702 * Record that we sent the ECN-setup and default to
1703 * setting IP ECT.
1704 */
1705 tp->ecn_flags |= (TE_SETUPSENT | TE_SENDIPECT);
1706 } else {
1707 /*
1708 * We sent an ECN-setup SYN but it was dropped.
1709 * Fall back to non-ECN and clear flag indicating
1710 * we should send data with IP ECT set.
1711 */
1712 if (tp->ecn_flags & TE_SETUPSENT) {
1713 tcpstat.tcps_ecn_lost_syn++;
1714 tp->ecn_flags |= TE_LOST_SYN;
1715 }
1716 tp->ecn_flags &= ~TE_SENDIPECT;
1717 }
1718 }
1719
1720 /*
1721 * Check if we should set the TCP CWR flag.
1722 * CWR flag is sent when we reduced the congestion window because
1723 * we received a TCP ECE or we performed a fast retransmit. We
1724 * never set the CWR flag on retransmitted packets. We only set
1725 * the CWR flag on data packets. Pure acks don't have this set.
1726 */
1727 if ((tp->ecn_flags & TE_SENDCWR) != 0 && len != 0 &&
1728 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
1729 flags |= TH_CWR;
1730 tp->ecn_flags &= ~TE_SENDCWR;
1731 }
1732
1733 /*
1734 * Check if we should set the TCP ECE flag.
1735 */
1736 if ((tp->ecn_flags & TE_SENDECE) != 0 && len == 0) {
1737 flags |= TH_ECE;
1738 tcpstat.tcps_ecn_sent_ece++;
1739 }
1740
1741
1742 hdrlen += optlen;
1743
1744 /* Reset DSACK sequence numbers */
1745 tp->t_dsack_lseq = 0;
1746 tp->t_dsack_rseq = 0;
1747
1748 #if INET6
1749 if (isipv6)
1750 ipoptlen = ip6_optlen(inp);
1751 else
1752 #endif
1753 {
1754 if (tp_inp_options) {
1755 ipoptlen = tp_inp_options->m_len -
1756 offsetof(struct ipoption, ipopt_list);
1757 } else {
1758 ipoptlen = 0;
1759 }
1760 }
1761 #if IPSEC
1762 ipoptlen += ipsec_optlen;
1763 #endif
1764
1765 /*
1766 * Adjust data length if insertion of options will
1767 * bump the packet length beyond the t_maxopd length.
1768 * Clear the FIN bit because we cut off the tail of
1769 * the segment.
1770 *
1771 * When doing TSO limit a burst to TCP_MAXWIN minus the
1772 * IP, TCP and Options length to keep ip->ip_len from
1773 * overflowing. Prevent the last segment from being
1774 * fractional thus making them all equal sized and set
1775 * the flag to continue sending. TSO is disabled when
1776 * IP options or IPSEC are present.
1777 */
1778 if (len + optlen + ipoptlen > tp->t_maxopd) {
1779 /*
1780 * If there is still more to send,
1781 * don't close the connection.
1782 */
1783 flags &= ~TH_FIN;
1784 if (tso) {
1785 int32_t tso_maxlen;
1786
1787 tso_maxlen = tp->tso_max_segment_size ?
1788 tp->tso_max_segment_size : TCP_MAXWIN;
1789
1790 if (len > tso_maxlen - hdrlen - optlen) {
1791 len = tso_maxlen - hdrlen - optlen;
1792 len = len - (len % (tp->t_maxopd - optlen));
1793 sendalot = 1;
1794 } else if (tp->t_flags & TF_NEEDFIN) {
1795 sendalot = 1;
1796 }
1797 } else {
1798 len = tp->t_maxopd - optlen - ipoptlen;
1799 sendalot = 1;
1800 }
1801 }
1802 #if MPTCP
1803 /* Adjust the length in the DSS option, if it is lesser than len */
1804 if (dlenp) {
1805 /*
1806 * To test this path without SACK, artificially
1807 * decrement len with something like
1808 * if (len > 10)
1809 len -= 10;
1810 */
1811 if (ntohs(*dlenp) > len) {
1812 *dlenp = htons(len);
1813 /* Unset the FIN flag, if len was adjusted */
1814 if (finp) {
1815 *finp &= ~MDSS_F;
1816 }
1817 sendalot = 1;
1818 }
1819 }
1820 #endif /* MPTCP */
1821
1822 if (max_linkhdr + hdrlen > MCLBYTES)
1823 panic("tcphdr too big");
1824
1825 /* Check if there is enough data in the send socket
1826 * buffer to start measuring bw
1827 */
1828 if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 &&
1829 (tp->t_bwmeas != NULL) &&
1830 (tp->t_flagsext & TF_BWMEAS_INPROGRESS) == 0 &&
1831 (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)) >=
1832 tp->t_bwmeas->bw_minsize) {
1833 tp->t_bwmeas->bw_size = min(
1834 (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)),
1835 tp->t_bwmeas->bw_maxsize);
1836 tp->t_flagsext |= TF_BWMEAS_INPROGRESS;
1837 tp->t_bwmeas->bw_start = tp->snd_max;
1838 tp->t_bwmeas->bw_ts = tcp_now;
1839 }
1840
1841 VERIFY(inp->inp_flowhash != 0);
1842 /*
1843 * Grab a header mbuf, attaching a copy of data to
1844 * be transmitted, and initialize the header from
1845 * the template for sends on this connection.
1846 */
1847 if (len) {
1848 tp->t_pmtud_lastseg_size = len + optlen + ipoptlen;
1849 if ((tp->t_flagsext & TF_FORCE) && len == 1)
1850 tcpstat.tcps_sndprobe++;
1851 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1852 tcpstat.tcps_sndrexmitpack++;
1853 tcpstat.tcps_sndrexmitbyte += len;
1854 if (nstat_collect) {
1855 nstat_route_tx(inp->inp_route.ro_rt, 1,
1856 len, NSTAT_TX_FLAG_RETRANSMIT);
1857 INP_ADD_STAT(inp, cell, wifi, wired,
1858 txpackets, 1);
1859 INP_ADD_STAT(inp, cell, wifi, wired,
1860 txbytes, len);
1861 tp->t_stat.txretransmitbytes += len;
1862 }
1863 } else {
1864 tcpstat.tcps_sndpack++;
1865 tcpstat.tcps_sndbyte += len;
1866
1867 if (nstat_collect) {
1868 INP_ADD_STAT(inp, cell, wifi, wired,
1869 txpackets, 1);
1870 INP_ADD_STAT(inp, cell, wifi, wired,
1871 txbytes, len);
1872 }
1873 }
1874 #if MPTCP
1875 if (tp->t_mpflags & TMPF_MPTCP_TRUE) {
1876 tcpstat.tcps_mp_sndpacks++;
1877 tcpstat.tcps_mp_sndbytes += len;
1878 }
1879 #endif /* MPTCP */
1880 /*
1881 * try to use the new interface that allocates all
1882 * the necessary mbuf hdrs under 1 mbuf lock and
1883 * avoids rescanning the socket mbuf list if
1884 * certain conditions are met. This routine can't
1885 * be used in the following cases...
1886 * 1) the protocol headers exceed the capacity of
1887 * of a single mbuf header's data area (no cluster attached)
1888 * 2) the length of the data being transmitted plus
1889 * the protocol headers fits into a single mbuf header's
1890 * data area (no cluster attached)
1891 */
1892 m = NULL;
1893
1894 /* minimum length we are going to allocate */
1895 allocated_len = MHLEN;
1896 if (MHLEN < hdrlen + max_linkhdr) {
1897 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1898 if (m == NULL) {
1899 error = ENOBUFS;
1900 goto out;
1901 }
1902 MCLGET(m, M_DONTWAIT);
1903 if ((m->m_flags & M_EXT) == 0) {
1904 m_freem(m);
1905 error = ENOBUFS;
1906 goto out;
1907 }
1908 m->m_data += max_linkhdr;
1909 m->m_len = hdrlen;
1910 allocated_len = MCLBYTES;
1911 }
1912 if (len <= allocated_len - hdrlen - max_linkhdr) {
1913 if (m == NULL) {
1914 VERIFY(allocated_len <= MHLEN);
1915 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1916 if (m == NULL) {
1917 error = ENOBUFS;
1918 goto out;
1919 }
1920 m->m_data += max_linkhdr;
1921 m->m_len = hdrlen;
1922 }
1923 /* makes sure we still have data left to be sent at this point */
1924 if (so->so_snd.sb_mb == NULL || off < 0) {
1925 if (m != NULL) m_freem(m);
1926 error = 0; /* should we return an error? */
1927 goto out;
1928 }
1929 m_copydata(so->so_snd.sb_mb, off, (int) len,
1930 mtod(m, caddr_t) + hdrlen);
1931 m->m_len += len;
1932 } else {
1933 uint32_t copymode;
1934 /*
1935 * Retain packet header metadata at the socket
1936 * buffer if this is is an MPTCP subflow,
1937 * otherwise move it.
1938 */
1939 copymode = M_COPYM_MOVE_HDR;
1940 #if MPTCP
1941 if (so->so_flags & SOF_MP_SUBFLOW) {
1942 copymode = M_COPYM_NOOP_HDR;
1943 }
1944 #endif /* MPTCP */
1945 if (m != NULL) {
1946 m->m_next = m_copym_mode(so->so_snd.sb_mb,
1947 off, (int)len, M_DONTWAIT, copymode);
1948 if (m->m_next == NULL) {
1949 (void) m_free(m);
1950 error = ENOBUFS;
1951 goto out;
1952 }
1953 } else {
1954 /*
1955 * make sure we still have data left
1956 * to be sent at this point
1957 */
1958 if (so->so_snd.sb_mb == NULL) {
1959 error = 0; /* should we return an error? */
1960 goto out;
1961 }
1962
1963 /*
1964 * m_copym_with_hdrs will always return the
1965 * last mbuf pointer and the offset into it that
1966 * it acted on to fullfill the current request,
1967 * whether a valid 'hint' was passed in or not.
1968 */
1969 if ((m = m_copym_with_hdrs(so->so_snd.sb_mb,
1970 off, len, M_DONTWAIT, NULL, NULL,
1971 copymode)) == NULL) {
1972 error = ENOBUFS;
1973 goto out;
1974 }
1975 m->m_data += max_linkhdr;
1976 m->m_len = hdrlen;
1977 }
1978 }
1979 /*
1980 * If we're sending everything we've got, set PUSH.
1981 * (This will keep happy those implementations which only
1982 * give data to the user when a buffer fills or
1983 * a PUSH comes in.)
1984 *
1985 * On SYN-segments we should not add the PUSH-flag.
1986 */
1987 if (off + len == so->so_snd.sb_cc && !(flags & TH_SYN))
1988 flags |= TH_PUSH;
1989 } else {
1990 if (tp->t_flags & TF_ACKNOW)
1991 tcpstat.tcps_sndacks++;
1992 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1993 tcpstat.tcps_sndctrl++;
1994 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1995 tcpstat.tcps_sndurg++;
1996 else
1997 tcpstat.tcps_sndwinup++;
1998
1999 MGETHDR(m, M_DONTWAIT, MT_HEADER); /* MAC-OK */
2000 if (m == NULL) {
2001 error = ENOBUFS;
2002 goto out;
2003 }
2004 if (MHLEN < (hdrlen + max_linkhdr)) {
2005 MCLGET(m, M_DONTWAIT);
2006 if ((m->m_flags & M_EXT) == 0) {
2007 m_freem(m);
2008 error = ENOBUFS;
2009 goto out;
2010 }
2011 }
2012 m->m_data += max_linkhdr;
2013 m->m_len = hdrlen;
2014 }
2015 m->m_pkthdr.rcvif = 0;
2016 #if MPTCP
2017 /* Before opt is copied to the mbuf, set the csum field */
2018 mptcp_output_csum(tp, m, len, hdrlen, dss_val, sseqp);
2019 #endif /* MPTCP */
2020 #if CONFIG_MACF_NET
2021 mac_mbuf_label_associate_inpcb(inp, m);
2022 #endif
2023 #if INET6
2024 if (isipv6) {
2025 ip6 = mtod(m, struct ip6_hdr *);
2026 th = (struct tcphdr *)(void *)(ip6 + 1);
2027 tcp_fillheaders(tp, ip6, th);
2028 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2029 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
2030 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
2031 }
2032 svc_flags |= PKT_SCF_IPV6;
2033 #if PF_ECN
2034 m->m_pkthdr.pf_mtag.pftag_hdr = (void *)ip6;
2035 m->m_pkthdr.pf_mtag.pftag_flags |= PF_TAG_HDR_INET6;
2036 #endif /* PF_ECN */
2037 } else
2038 #endif /* INET6 */
2039 {
2040 ip = mtod(m, struct ip *);
2041 ipov = (struct ipovly *)ip;
2042 th = (struct tcphdr *)(void *)(ip + 1);
2043 /* this picks up the pseudo header (w/o the length) */
2044 tcp_fillheaders(tp, ip, th);
2045 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2046 !SEQ_LT(tp->snd_nxt, tp->snd_max) &&
2047 !sack_rxmit && !(flags & TH_SYN)) {
2048 ip->ip_tos |= IPTOS_ECN_ECT0;
2049 }
2050 #if PF_ECN
2051 m->m_pkthdr.pf_mtag.pftag_hdr = (void *)ip;
2052 m->m_pkthdr.pf_mtag.pftag_flags |= PF_TAG_HDR_INET;
2053 #endif /* PF_ECN */
2054 }
2055
2056 /*
2057 * Fill in fields, remembering maximum advertised
2058 * window for use in delaying messages about window sizes.
2059 * If resending a FIN, be sure not to use a new sequence number.
2060 */
2061 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
2062 tp->snd_nxt == tp->snd_max)
2063 tp->snd_nxt--;
2064 /*
2065 * If we are doing retransmissions, then snd_nxt will
2066 * not reflect the first unsent octet. For ACK only
2067 * packets, we do not want the sequence number of the
2068 * retransmitted packet, we want the sequence number
2069 * of the next unsent octet. So, if there is no data
2070 * (and no SYN or FIN), use snd_max instead of snd_nxt
2071 * when filling in ti_seq. But if we are in persist
2072 * state, snd_max might reflect one byte beyond the
2073 * right edge of the window, so use snd_nxt in that
2074 * case, since we know we aren't doing a retransmission.
2075 * (retransmit and persist are mutually exclusive...)
2076 *
2077 * Note the state of this retransmit segment to detect spurious
2078 * retransmissions.
2079 */
2080 if (sack_rxmit == 0) {
2081 if (len || (flags & (TH_SYN|TH_FIN)) ||
2082 tp->t_timer[TCPT_PERSIST]) {
2083 th->th_seq = htonl(tp->snd_nxt);
2084 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
2085 if (SACK_ENABLED(tp) && len > 1) {
2086 tcp_rxtseg_insert(tp, tp->snd_nxt,
2087 (tp->snd_nxt + len - 1));
2088 }
2089 if (len > 0)
2090 m->m_pkthdr.pkt_flags |=
2091 PKTF_TCP_REXMT;
2092 }
2093 } else {
2094 th->th_seq = htonl(tp->snd_max);
2095 }
2096 } else {
2097 th->th_seq = htonl(p->rxmit);
2098 tcp_rxtseg_insert(tp, p->rxmit, (p->rxmit + len - 1));
2099 p->rxmit += len;
2100 tp->sackhint.sack_bytes_rexmit += len;
2101 if (len > 0)
2102 m->m_pkthdr.pkt_flags |= PKTF_TCP_REXMT;
2103 }
2104 th->th_ack = htonl(tp->rcv_nxt);
2105 tp->last_ack_sent = tp->rcv_nxt;
2106 #if MPTCP
2107 /* Initialize the ACK field to a value as 0 ack fields are dropped */
2108 if (early_data_sent) {
2109 th->th_ack = th->th_seq + 1;
2110 }
2111 #endif /* MPTCP */
2112 if (optlen) {
2113 bcopy(opt, th + 1, optlen);
2114 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
2115 }
2116 th->th_flags = flags;
2117 /*
2118 * Calculate receive window. Don't shrink window,
2119 * but avoid silly window syndrome.
2120 */
2121 if (recwin < (int32_t)(so->so_rcv.sb_hiwat / 4) && recwin < (int)tp->t_maxseg)
2122 recwin = 0;
2123 if (recwin < (int32_t)(tp->rcv_adv - tp->rcv_nxt))
2124 recwin = (int32_t)(tp->rcv_adv - tp->rcv_nxt);
2125 if (tp->t_flags & TF_SLOWLINK && slowlink_wsize > 0) {
2126 if (recwin > (int32_t)slowlink_wsize)
2127 recwin = slowlink_wsize;
2128 }
2129
2130 #if TRAFFIC_MGT
2131 if (tcp_recv_bg == 1 || IS_TCP_RECV_BG(so)) {
2132 if (tcp_recv_throttle(tp)) {
2133 uint32_t min_iaj_win =
2134 tcp_min_iaj_win * tp->t_maxseg;
2135 if (tp->iaj_rwintop == 0 ||
2136 SEQ_LT(tp->iaj_rwintop, tp->rcv_adv))
2137 tp->iaj_rwintop = tp->rcv_adv;
2138 if (SEQ_LT(tp->iaj_rwintop,
2139 tp->rcv_nxt + min_iaj_win))
2140 tp->iaj_rwintop = tp->rcv_nxt + min_iaj_win;
2141 recwin = min(tp->iaj_rwintop - tp->rcv_nxt, recwin);
2142 }
2143 }
2144 #endif /* TRAFFIC_MGT */
2145
2146 if (recwin > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
2147 recwin = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
2148 th->th_win = htons((u_short) (recwin>>tp->rcv_scale));
2149
2150 /*
2151 * Adjust the RXWIN0SENT flag - indicate that we have advertised
2152 * a 0 window. This may cause the remote transmitter to stall. This
2153 * flag tells soreceive() to disable delayed acknowledgements when
2154 * draining the buffer. This can occur if the receiver is attempting
2155 * to read more data then can be buffered prior to transmitting on
2156 * the connection.
2157 */
2158 if (th->th_win == 0)
2159 tp->t_flags |= TF_RXWIN0SENT;
2160 else
2161 tp->t_flags &= ~TF_RXWIN0SENT;
2162 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
2163 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
2164 th->th_flags |= TH_URG;
2165 } else {
2166 /*
2167 * If no urgent pointer to send, then we pull
2168 * the urgent pointer to the left edge of the send window
2169 * so that it doesn't drift into the send window on sequence
2170 * number wraparound.
2171 */
2172 tp->snd_up = tp->snd_una; /* drag it along */
2173 }
2174
2175 /*
2176 * Put TCP length in extended header, and then
2177 * checksum extended header and data.
2178 */
2179 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
2180 #if INET6
2181 if (isipv6) {
2182 /*
2183 * ip6_plen is not need to be filled now, and will be filled
2184 * in ip6_output.
2185 */
2186 m->m_pkthdr.csum_flags = CSUM_TCPIPV6;
2187 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2188 if (len + optlen)
2189 th->th_sum = in_addword(th->th_sum,
2190 htons((u_short)(optlen + len)));
2191 }
2192 else
2193 #endif /* INET6 */
2194 {
2195 m->m_pkthdr.csum_flags = CSUM_TCP;
2196 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2197 if (len + optlen)
2198 th->th_sum = in_addword(th->th_sum,
2199 htons((u_short)(optlen + len)));
2200 }
2201
2202 /*
2203 * Enable TSO and specify the size of the segments.
2204 * The TCP pseudo header checksum is always provided.
2205 */
2206 if (tso) {
2207 #if INET6
2208 if (isipv6)
2209 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV6;
2210 else
2211 #endif /* INET6 */
2212 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV4;
2213
2214 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
2215 } else {
2216 m->m_pkthdr.tso_segsz = 0;
2217 }
2218
2219 /*
2220 * In transmit state, time the transmission and arrange for
2221 * the retransmit. In persist state, just set snd_max.
2222 */
2223 if (!(tp->t_flagsext & TF_FORCE)
2224 || tp->t_timer[TCPT_PERSIST] == 0) {
2225 tcp_seq startseq = tp->snd_nxt;
2226
2227 /*
2228 * Advance snd_nxt over sequence space of this segment.
2229 */
2230 if (flags & (TH_SYN|TH_FIN)) {
2231 if (flags & TH_SYN)
2232 tp->snd_nxt++;
2233 if ((flags & TH_FIN) &&
2234 !(tp->t_flags & TF_SENTFIN)) {
2235 tp->snd_nxt++;
2236 tp->t_flags |= TF_SENTFIN;
2237 }
2238 }
2239 if (sack_rxmit)
2240 goto timer;
2241 if (sack_rescue_rxt == TRUE) {
2242 tp->snd_nxt = old_snd_nxt;
2243 sack_rescue_rxt = FALSE;
2244 tcpstat.tcps_pto_in_recovery++;
2245 } else {
2246 tp->snd_nxt += len;
2247 }
2248 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
2249 tp->snd_max = tp->snd_nxt;
2250 /*
2251 * Time this transmission if not a retransmission and
2252 * not currently timing anything.
2253 */
2254 if (tp->t_rtttime == 0) {
2255 tp->t_rtttime = tcp_now;
2256 tp->t_rtseq = startseq;
2257 tcpstat.tcps_segstimed++;
2258
2259 /* update variables related to pipe ack */
2260 tp->t_pipeack_lastuna = tp->snd_una;
2261 }
2262 }
2263
2264 /*
2265 * Set retransmit timer if not currently set,
2266 * and not doing an ack or a keep-alive probe.
2267 */
2268 timer:
2269 if (tp->t_timer[TCPT_REXMT] == 0 &&
2270 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
2271 tp->snd_nxt != tp->snd_una || (flags & TH_FIN))) {
2272 if (tp->t_timer[TCPT_PERSIST]) {
2273 tp->t_timer[TCPT_PERSIST] = 0;
2274 tp->t_rxtshift = 0;
2275 tp->t_rxtstart = 0;
2276 tp->t_persist_stop = 0;
2277 }
2278 tp->t_timer[TCPT_REXMT] =
2279 OFFSET_FROM_START(tp, tp->t_rxtcur);
2280 }
2281
2282 /*
2283 * Set tail loss probe timeout if new data is being
2284 * transmitted. This will be supported only when
2285 * SACK option is enabled on a connection.
2286 *
2287 * Every time new data is sent PTO will get reset.
2288 */
2289 if (tcp_enable_tlp && tp->t_state == TCPS_ESTABLISHED &&
2290 SACK_ENABLED(tp) && !IN_FASTRECOVERY(tp)
2291 && tp->snd_nxt == tp->snd_max
2292 && SEQ_GT(tp->snd_nxt, tp->snd_una)
2293 && tp->t_rxtshift == 0
2294 && (tp->t_flagsext & (TF_SENT_TLPROBE|TF_PKTS_REORDERED)) == 0) {
2295 u_int32_t pto, srtt, new_rto = 0;
2296
2297 /*
2298 * Using SRTT alone to set PTO can cause spurious
2299 * retransmissions on wireless networks where there
2300 * is a lot of variance in RTT. Taking variance
2301 * into account will avoid this.
2302 */
2303 srtt = tp->t_srtt >> TCP_RTT_SHIFT;
2304 pto = ((TCP_REXMTVAL(tp)) * 3) >> 1;
2305 pto = max (2 * srtt, pto);
2306 if ((tp->snd_max - tp->snd_una) == tp->t_maxseg)
2307 pto = max(pto,
2308 (((3 * pto) >> 2) + tcp_delack * 2));
2309 else
2310 pto = max(10, pto);
2311
2312 /* if RTO is less than PTO, choose RTO instead */
2313 if (tp->t_rxtcur < pto) {
2314 /*
2315 * Schedule PTO instead of RTO in favor of
2316 * fast recovery.
2317 */
2318 pto = tp->t_rxtcur;
2319
2320 /* Reset the next RTO to be after PTO. */
2321 TCPT_RANGESET(new_rto,
2322 (pto + TCP_REXMTVAL(tp)),
2323 max(tp->t_rttmin, tp->t_rttcur + 2),
2324 TCPTV_REXMTMAX, 0);
2325 tp->t_timer[TCPT_REXMT] =
2326 OFFSET_FROM_START(tp, new_rto);
2327 }
2328 tp->t_timer[TCPT_PTO] = OFFSET_FROM_START(tp, pto);
2329 }
2330 } else {
2331 /*
2332 * Persist case, update snd_max but since we are in
2333 * persist mode (no window) we do not update snd_nxt.
2334 */
2335 int xlen = len;
2336 if (flags & TH_SYN)
2337 ++xlen;
2338 if ((flags & TH_FIN) &&
2339 !(tp->t_flags & TF_SENTFIN)) {
2340 ++xlen;
2341 tp->t_flags |= TF_SENTFIN;
2342 }
2343 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
2344 tp->snd_max = tp->snd_nxt + len;
2345 }
2346
2347 #if TCPDEBUG
2348 /*
2349 * Trace.
2350 */
2351 if (so_options & SO_DEBUG)
2352 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
2353 #endif
2354
2355 /*
2356 * Fill in IP length and desired time to live and
2357 * send to IP level. There should be a better way
2358 * to handle ttl and tos; we could keep them in
2359 * the template, but need a way to checksum without them.
2360 */
2361 #if INET6
2362 /*
2363 * m->m_pkthdr.len should have been set before cksum calcuration,
2364 * because in6_cksum() need it.
2365 */
2366 if (isipv6) {
2367 /*
2368 * we separately set hoplimit for every segment, since the
2369 * user might want to change the value via setsockopt.
2370 * Also, desired default hop limit might be changed via
2371 * Neighbor Discovery.
2372 */
2373 ip6->ip6_hlim = in6_selecthlim(inp, inp->in6p_route.ro_rt ?
2374 inp->in6p_route.ro_rt->rt_ifp : NULL);
2375
2376 /* TODO: IPv6 IP6TOS_ECT bit on */
2377 KERNEL_DEBUG(DBG_LAYER_BEG,
2378 ((inp->inp_fport << 16) | inp->inp_lport),
2379 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
2380 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
2381 sendalot,0,0);
2382 } else
2383 #endif /* INET6 */
2384 {
2385 ip->ip_len = m->m_pkthdr.len;
2386 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
2387 ip->ip_tos |= (inp->inp_ip_tos & ~IPTOS_ECN_MASK);/* XXX */
2388 KERNEL_DEBUG(DBG_LAYER_BEG,
2389 ((inp->inp_fport << 16) | inp->inp_lport),
2390 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
2391 (inp->inp_faddr.s_addr & 0xffff)), 0,0,0);
2392 }
2393
2394 /*
2395 * See if we should do MTU discovery.
2396 * Look at the flag updated on the following criterias:
2397 * 1) Path MTU discovery is authorized by the sysctl
2398 * 2) The route isn't set yet (unlikely but could happen)
2399 * 3) The route is up
2400 * 4) the MTU is not locked (if it is, then discovery has been
2401 * disabled for that route)
2402 */
2403 #if INET6
2404 if (!isipv6)
2405 #endif /* INET6 */
2406 if (path_mtu_discovery && (tp->t_flags & TF_PMTUD))
2407 ip->ip_off |= IP_DF;
2408
2409 #if NECP
2410 {
2411 necp_kernel_policy_id policy_id;
2412 u_int32_t route_rule_id;
2413 if (!necp_socket_is_allowed_to_send_recv(inp, &policy_id, &route_rule_id)) {
2414 m_freem(m);
2415 error = EHOSTUNREACH;
2416 goto out;
2417 }
2418
2419 necp_mark_packet_from_socket(m, inp, policy_id, route_rule_id);
2420 }
2421 #endif /* NECP */
2422
2423 #if IPSEC
2424 if (inp->inp_sp != NULL)
2425 ipsec_setsocket(m, so);
2426 #endif /*IPSEC*/
2427
2428 /*
2429 * The socket is kept locked while sending out packets in ip_output, even if packet chaining is not active.
2430 */
2431 lost = 0;
2432
2433 /*
2434 * Embed the flow hash in pkt hdr and mark the packet as
2435 * capable of flow controlling
2436 */
2437 m->m_pkthdr.pkt_flowsrc = FLOWSRC_INPCB;
2438 m->m_pkthdr.pkt_flowid = inp->inp_flowhash;
2439 m->m_pkthdr.pkt_flags |= PKTF_FLOW_ID | PKTF_FLOW_LOCALSRC;
2440 #if MPTCP
2441 /* Disable flow advisory when using MPTCP. */
2442 if (!(tp->t_mpflags & TMPF_MPTCP_TRUE))
2443 #endif /* MPTCP */
2444 m->m_pkthdr.pkt_flags |= PKTF_FLOW_ADV;
2445 m->m_pkthdr.pkt_proto = IPPROTO_TCP;
2446
2447 m->m_nextpkt = NULL;
2448
2449 if (inp->inp_last_outifp != NULL &&
2450 !(inp->inp_last_outifp->if_flags & IFF_LOOPBACK)) {
2451 /* Hint to prioritize this packet if
2452 * 1. if the packet has no data
2453 * 2. the interface supports transmit-start model and did
2454 * not disable ACK prioritization.
2455 * 3. Only ACK flag is set.
2456 * 4. there is no outstanding data on this connection.
2457 */
2458 if (tcp_prioritize_acks != 0 && len == 0 &&
2459 (inp->inp_last_outifp->if_eflags &
2460 (IFEF_TXSTART | IFEF_NOACKPRI)) == IFEF_TXSTART &&
2461 th->th_flags == TH_ACK && tp->snd_una == tp->snd_max &&
2462 tp->t_timer[TCPT_REXMT] == 0) {
2463 svc_flags |= PKT_SCF_TCP_ACK;
2464 }
2465 set_packet_service_class(m, so, MBUF_SC_UNSPEC, svc_flags);
2466 }
2467
2468 tp->t_pktlist_sentlen += len;
2469 tp->t_lastchain++;
2470
2471 #if INET6
2472 if (isipv6) {
2473 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2474 struct ip6 *, ip6, struct tcpcb *, tp, struct tcphdr *,
2475 th);
2476 } else
2477 #endif /* INET6 */
2478 {
2479 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2480 struct ip *, ip, struct tcpcb *, tp, struct tcphdr *, th);
2481 }
2482
2483 if (tp->t_pktlist_head != NULL) {
2484 tp->t_pktlist_tail->m_nextpkt = m;
2485 tp->t_pktlist_tail = m;
2486 } else {
2487 packchain_newlist++;
2488 tp->t_pktlist_head = tp->t_pktlist_tail = m;
2489 }
2490
2491 if ((lro_ackmore) && (!sackoptlen) && (!tp->t_timer[TCPT_PERSIST]) &&
2492 ((th->th_flags & TH_ACK) == TH_ACK) && (!len) &&
2493 (tp->t_state == TCPS_ESTABLISHED)) {
2494 /* For a pure ACK, see if you need to send more of them */
2495 mnext = tcp_send_lroacks(tp, m, th);
2496 if (mnext) {
2497 tp->t_pktlist_tail->m_nextpkt = mnext;
2498 if (mnext->m_nextpkt == NULL) {
2499 tp->t_pktlist_tail = mnext;
2500 tp->t_lastchain++;
2501 } else {
2502 struct mbuf *tail, *next;
2503 next = mnext->m_nextpkt;
2504 tail = next->m_nextpkt;
2505 while (tail) {
2506 next = tail;
2507 tail = tail->m_nextpkt;
2508 tp->t_lastchain++;
2509 }
2510 tp->t_pktlist_tail = next;
2511 }
2512 }
2513 }
2514
2515 if (sendalot == 0 || (tp->t_state != TCPS_ESTABLISHED) ||
2516 (tp->snd_cwnd <= (tp->snd_wnd / 8)) ||
2517 (tp->t_flags & (TH_PUSH | TF_ACKNOW)) ||
2518 (tp->t_flagsext & TF_FORCE) ||
2519 tp->t_lastchain >= tcp_packet_chaining) {
2520 error = 0;
2521 while (inp->inp_sndinprog_cnt == 0 &&
2522 tp->t_pktlist_head != NULL) {
2523 packetlist = tp->t_pktlist_head;
2524 packchain_listadd = tp->t_lastchain;
2525 packchain_sent++;
2526 lost = tp->t_pktlist_sentlen;
2527 TCP_PKTLIST_CLEAR(tp);
2528
2529 error = tcp_ip_output(so, tp, packetlist,
2530 packchain_listadd, tp_inp_options,
2531 (so_options & SO_DONTROUTE),
2532 (sack_rxmit | (sack_bytes_rxmt != 0)), recwin,
2533 #if INET6
2534 isipv6);
2535 #else /* INET6 */
2536 0);
2537 #endif /* !INET6 */
2538 if (error) {
2539 /*
2540 * Take into account the rest of unsent
2541 * packets in the packet list for this tcp
2542 * into "lost", since we're about to free
2543 * the whole list below.
2544 */
2545 lost += tp->t_pktlist_sentlen;
2546 break;
2547 } else {
2548 lost = 0;
2549 }
2550 }
2551 /* tcp was closed while we were in ip; resume close */
2552 if (inp->inp_sndinprog_cnt == 0 &&
2553 (tp->t_flags & TF_CLOSING)) {
2554 tp->t_flags &= ~TF_CLOSING;
2555 (void) tcp_close(tp);
2556 return (0);
2557 }
2558 } else {
2559 error = 0;
2560 packchain_looped++;
2561 tcpstat.tcps_sndtotal++;
2562
2563 goto again;
2564 }
2565 if (error) {
2566 /*
2567 * Assume that the packets were lost, so back out the
2568 * sequence number advance, if any. Note that the "lost"
2569 * variable represents the amount of user data sent during
2570 * the recent call to ip_output_list() plus the amount of
2571 * user data in the packet list for this tcp at the moment.
2572 */
2573 if (!(tp->t_flagsext & TF_FORCE)
2574 || tp->t_timer[TCPT_PERSIST] == 0) {
2575 /*
2576 * No need to check for TH_FIN here because
2577 * the TF_SENTFIN flag handles that case.
2578 */
2579 if ((flags & TH_SYN) == 0) {
2580 if (sack_rxmit) {
2581 if (SEQ_GT((p->rxmit - lost),
2582 tp->snd_una)) {
2583 p->rxmit -= lost;
2584 } else {
2585 lost = p->rxmit - tp->snd_una;
2586 p->rxmit = tp->snd_una;
2587 }
2588 tp->sackhint.sack_bytes_rexmit -= lost;
2589 } else {
2590 if (SEQ_GT((tp->snd_nxt - lost),
2591 tp->snd_una))
2592 tp->snd_nxt -= lost;
2593 else
2594 tp->snd_nxt = tp->snd_una;
2595 }
2596 }
2597 }
2598 out:
2599 if (tp->t_pktlist_head != NULL)
2600 m_freem_list(tp->t_pktlist_head);
2601 TCP_PKTLIST_CLEAR(tp);
2602
2603 if (error == ENOBUFS) {
2604 if (!tp->t_timer[TCPT_REXMT] &&
2605 !tp->t_timer[TCPT_PERSIST])
2606 tp->t_timer[TCPT_REXMT] =
2607 OFFSET_FROM_START(tp, tp->t_rxtcur);
2608 tp->snd_cwnd = tp->t_maxseg;
2609 tp->t_bytes_acked = 0;
2610 tcp_check_timer_state(tp);
2611 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2612
2613 tcp_ccdbg_trace(tp, NULL, TCP_CC_OUTPUT_ERROR);
2614 return (0);
2615 }
2616 if (error == EMSGSIZE) {
2617 /*
2618 * ip_output() will have already fixed the route
2619 * for us. tcp_mtudisc() will, as its last action,
2620 * initiate retransmission, so it is important to
2621 * not do so here.
2622 *
2623 * If TSO was active we either got an interface
2624 * without TSO capabilits or TSO was turned off.
2625 * Disable it for this connection as too and
2626 * immediatly retry with MSS sized segments generated
2627 * by this function.
2628 */
2629 if (tso)
2630 tp->t_flags &= ~TF_TSO;
2631
2632 tcp_mtudisc(inp, 0);
2633 tcp_check_timer_state(tp);
2634
2635 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2636 return 0;
2637 }
2638 /*
2639 * Unless this is due to interface restriction policy,
2640 * treat EHOSTUNREACH/ENETDOWN as a soft error.
2641 */
2642 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
2643 TCPS_HAVERCVDSYN(tp->t_state) &&
2644 !inp_restricted_send(inp, inp->inp_last_outifp)) {
2645 tp->t_softerror = error;
2646 error = 0;
2647 }
2648 tcp_check_timer_state(tp);
2649 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2650 return (error);
2651 }
2652
2653 tcpstat.tcps_sndtotal++;
2654
2655 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,0,0,0,0,0);
2656 if (sendalot)
2657 goto again;
2658
2659 tcp_check_timer_state(tp);
2660 return (0);
2661 }
2662
2663 static int
2664 tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt,
2665 int cnt, struct mbuf *opt, int flags, int sack_in_progress, int recwin,
2666 boolean_t isipv6)
2667 {
2668 int error = 0;
2669 boolean_t chain;
2670 boolean_t unlocked = FALSE;
2671 boolean_t ifdenied = FALSE;
2672 struct inpcb *inp = tp->t_inpcb;
2673 struct ip_out_args ipoa =
2674 { IFSCOPE_NONE, { 0 }, IPOAF_SELECT_SRCIF|IPOAF_BOUND_SRCADDR, 0 };
2675 struct route ro;
2676 struct ifnet *outif = NULL;
2677 #if INET6
2678 struct ip6_out_args ip6oa =
2679 { IFSCOPE_NONE, { 0 }, IP6OAF_SELECT_SRCIF|IP6OAF_BOUND_SRCADDR, 0 };
2680 struct route_in6 ro6;
2681 struct flowadv *adv =
2682 (isipv6 ? &ip6oa.ip6oa_flowadv : &ipoa.ipoa_flowadv);
2683 #else /* INET6 */
2684 struct flowadv *adv = &ipoa.ipoa_flowadv;
2685 #endif /* !INET6 */
2686
2687 /* If socket was bound to an ifindex, tell ip_output about it */
2688 if (inp->inp_flags & INP_BOUND_IF) {
2689 #if INET6
2690 if (isipv6) {
2691 ip6oa.ip6oa_boundif = inp->inp_boundifp->if_index;
2692 ip6oa.ip6oa_flags |= IP6OAF_BOUND_IF;
2693 } else
2694 #endif /* INET6 */
2695 {
2696 ipoa.ipoa_boundif = inp->inp_boundifp->if_index;
2697 ipoa.ipoa_flags |= IPOAF_BOUND_IF;
2698 }
2699 }
2700
2701 if (INP_NO_CELLULAR(inp)) {
2702 #if INET6
2703 if (isipv6)
2704 ip6oa.ip6oa_flags |= IP6OAF_NO_CELLULAR;
2705 else
2706 #endif /* INET6 */
2707 ipoa.ipoa_flags |= IPOAF_NO_CELLULAR;
2708 }
2709 if (INP_NO_EXPENSIVE(inp)) {
2710 #if INET6
2711 if (isipv6)
2712 ip6oa.ip6oa_flags |= IP6OAF_NO_EXPENSIVE;
2713 else
2714 #endif /* INET6 */
2715 ipoa.ipoa_flags |= IPOAF_NO_EXPENSIVE;
2716
2717 }
2718 if (INP_AWDL_UNRESTRICTED(inp)) {
2719 #if INET6
2720 if (isipv6)
2721 ip6oa.ip6oa_flags |= IP6OAF_AWDL_UNRESTRICTED;
2722 else
2723 #endif /* INET6 */
2724 ipoa.ipoa_flags |= IPOAF_AWDL_UNRESTRICTED;
2725
2726 }
2727 #if INET6
2728 if (isipv6)
2729 flags |= IPV6_OUTARGS;
2730 else
2731 #endif /* INET6 */
2732 flags |= IP_OUTARGS;
2733
2734 /* Copy the cached route and take an extra reference */
2735 #if INET6
2736 if (isipv6)
2737 in6p_route_copyout(inp, &ro6);
2738 else
2739 #endif /* INET6 */
2740 inp_route_copyout(inp, &ro);
2741
2742 /*
2743 * Data sent (as far as we can tell).
2744 * If this advertises a larger window than any other segment,
2745 * then remember the size of the advertised window.
2746 * Make sure ACK/DELACK conditions are cleared before
2747 * we unlock the socket.
2748 */
2749 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
2750 tp->rcv_adv = tp->rcv_nxt + recwin;
2751 tp->last_ack_sent = tp->rcv_nxt;
2752 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
2753 tp->t_timer[TCPT_DELACK] = 0;
2754 tp->t_unacksegs = 0;
2755
2756 /* Increment the count of outstanding send operations */
2757 inp->inp_sndinprog_cnt++;
2758
2759 /*
2760 * If allowed, unlock TCP socket while in IP
2761 * but only if the connection is established and
2762 * in a normal mode where reentrancy on the tcpcb won't be
2763 * an issue:
2764 * - there is no SACK episode
2765 * - we're not in Fast Recovery mode
2766 * - if we're not sending from an upcall.
2767 */
2768 if (tcp_output_unlocked && !so->so_upcallusecount &&
2769 (tp->t_state == TCPS_ESTABLISHED) && (sack_in_progress == 0) &&
2770 !IN_FASTRECOVERY(tp)) {
2771
2772 unlocked = TRUE;
2773 socket_unlock(so, 0);
2774 }
2775
2776 /*
2777 * Don't send down a chain of packets when:
2778 * - TCP chaining is disabled
2779 * - there is an IPsec rule set
2780 * - there is a non default rule set for the firewall
2781 */
2782
2783 chain = tcp_packet_chaining > 1
2784 #if IPSEC
2785 && ipsec_bypass
2786 #endif
2787 #if IPFIREWALL
2788 && (fw_enable == 0 || fw_bypass)
2789 #endif
2790 ; // I'm important, not extraneous
2791
2792
2793 while (pkt != NULL) {
2794 struct mbuf *npkt = pkt->m_nextpkt;
2795
2796 if (!chain) {
2797 pkt->m_nextpkt = NULL;
2798 /*
2799 * If we are not chaining, make sure to set the packet
2800 * list count to 0 so that IP takes the right path;
2801 * this is important for cases such as IPSec where a
2802 * single mbuf might result in multiple mbufs as part
2803 * of the encapsulation. If a non-zero count is passed
2804 * down to IP, the head of the chain might change and
2805 * we could end up skipping it (thus generating bogus
2806 * packets). Fixing it in IP would be desirable, but
2807 * for now this would do it.
2808 */
2809 cnt = 0;
2810 }
2811 #if INET6
2812 if (isipv6) {
2813 error = ip6_output_list(pkt, cnt,
2814 inp->in6p_outputopts, &ro6, flags, NULL, NULL,
2815 &ip6oa);
2816 ifdenied = (ip6oa.ip6oa_retflags & IP6OARF_IFDENIED);
2817 } else {
2818 #endif /* INET6 */
2819 error = ip_output_list(pkt, cnt, opt, &ro, flags, NULL,
2820 &ipoa);
2821 ifdenied = (ipoa.ipoa_retflags & IPOARF_IFDENIED);
2822 }
2823
2824 if (chain || error) {
2825 /*
2826 * If we sent down a chain then we are done since
2827 * the callee had taken care of everything; else
2828 * we need to free the rest of the chain ourselves.
2829 */
2830 if (!chain)
2831 m_freem_list(npkt);
2832 break;
2833 }
2834 pkt = npkt;
2835 }
2836
2837 if (unlocked)
2838 socket_lock(so, 0);
2839
2840 /*
2841 * Enter flow controlled state if the connection is established
2842 * and is not in recovery.
2843 *
2844 * A connection will enter suspended state even if it is in
2845 * recovery.
2846 */
2847 if (((adv->code == FADV_FLOW_CONTROLLED && !IN_FASTRECOVERY(tp)) ||
2848 adv->code == FADV_SUSPENDED) &&
2849 !(tp->t_flags & TF_CLOSING) &&
2850 tp->t_state == TCPS_ESTABLISHED) {
2851 int rc;
2852 rc = inp_set_fc_state(inp, adv->code);
2853
2854 if (rc == 1)
2855 tcp_ccdbg_trace(tp, NULL,
2856 ((adv->code == FADV_FLOW_CONTROLLED) ?
2857 TCP_CC_FLOW_CONTROL : TCP_CC_SUSPEND));
2858 }
2859
2860 /*
2861 * When an interface queue gets suspended, some of the
2862 * packets are dropped. Return ENOBUFS, to update the
2863 * pcb state.
2864 */
2865 if (adv->code == FADV_SUSPENDED)
2866 error = ENOBUFS;
2867
2868 VERIFY(inp->inp_sndinprog_cnt > 0);
2869 if ( --inp->inp_sndinprog_cnt == 0)
2870 inp->inp_flags &= ~(INP_FC_FEEDBACK);
2871
2872 #if INET6
2873 if (isipv6) {
2874 if (ro6.ro_rt != NULL && (outif = ro6.ro_rt->rt_ifp) !=
2875 inp->in6p_last_outifp)
2876 inp->in6p_last_outifp = outif;
2877 } else
2878 #endif /* INET6 */
2879 if (ro.ro_rt != NULL && (outif = ro.ro_rt->rt_ifp) !=
2880 inp->inp_last_outifp)
2881 inp->inp_last_outifp = outif;
2882
2883 if (error != 0 && ifdenied &&
2884 (INP_NO_CELLULAR(inp) || INP_NO_EXPENSIVE(inp)))
2885 soevent(inp->inp_socket,
2886 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_IFDENIED));
2887
2888 /* Synchronize cached PCB route & options */
2889 #if INET6
2890 if (isipv6)
2891 in6p_route_copyin(inp, &ro6);
2892 else
2893 #endif /* INET6 */
2894 inp_route_copyin(inp, &ro);
2895
2896 if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift == 0 &&
2897 tp->t_inpcb->inp_route.ro_rt != NULL) {
2898 /* If we found the route and there is an rtt on it
2899 * reset the retransmit timer
2900 */
2901 tcp_getrt_rtt(tp, tp->t_inpcb->in6p_route.ro_rt);
2902 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur);
2903 }
2904 return (error);
2905 }
2906
2907 void
2908 tcp_setpersist(tp)
2909 register struct tcpcb *tp;
2910 {
2911 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
2912
2913 /* If a PERSIST_TIMER option was set we will limit the
2914 * time the persist timer will be active for that connection
2915 * in order to avoid DOS by using zero window probes.
2916 * see rdar://5805356
2917 */
2918
2919 if ((tp->t_persist_timeout != 0) &&
2920 (tp->t_timer[TCPT_PERSIST] == 0) &&
2921 (tp->t_persist_stop == 0)) {
2922 tp->t_persist_stop = tcp_now + tp->t_persist_timeout;
2923 }
2924
2925 /*
2926 * Start/restart persistance timer.
2927 */
2928 TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
2929 t * tcp_backoff[tp->t_rxtshift],
2930 TCPTV_PERSMIN, TCPTV_PERSMAX, 0);
2931 tp->t_timer[TCPT_PERSIST] = OFFSET_FROM_START(tp, tp->t_timer[TCPT_PERSIST]);
2932
2933 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
2934 tp->t_rxtshift++;
2935 }
2936
2937 /*
2938 * Send as many acks as data coalesced. Every other packet when stretch
2939 * ACK is not enabled. Every 8 packets, if stretch ACK is enabled.
2940 */
2941 static struct mbuf*
2942 tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th)
2943 {
2944 struct mbuf *mnext = NULL, *ack_chain = NULL, *tail = NULL;
2945 int count = 0;
2946 tcp_seq org_ack = ntohl(th->th_ack);
2947 tcp_seq prev_ack = 0;
2948 int tack_offset = 28; /* XXX IPv6 and IP options not supported */
2949 int twin_offset = 34; /* XXX IPv6 and IP options not supported */
2950 int ack_size = (tp->t_flags & TF_STRETCHACK) ?
2951 (maxseg_unacked * tp->t_maxseg) : (tp->t_maxseg << 1);
2952 int segs_acked = (tp->t_flags & TF_STRETCHACK) ? maxseg_unacked : 2;
2953 struct mbuf *prev_ack_pkt = NULL;
2954 struct socket *so = tp->t_inpcb->inp_socket;
2955 unsigned short winsz = ntohs(th->th_win);
2956 unsigned int scaled_win = winsz<<tp->rcv_scale;
2957 tcp_seq win_rtedge = org_ack + scaled_win;
2958
2959 count = tp->t_lropktlen/tp->t_maxseg;
2960
2961 prev_ack = (org_ack - tp->t_lropktlen) + ack_size;
2962 if (prev_ack < org_ack) {
2963 ack_chain = m_dup(m, M_DONTWAIT);
2964 if (ack_chain) {
2965 th->th_ack = htonl(prev_ack);
2966 /* Keep adv window constant for duplicated ACK packets */
2967 scaled_win = win_rtedge - prev_ack;
2968 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
2969 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
2970 th->th_win = htons(scaled_win>>tp->rcv_scale);
2971 if (lrodebug == 5) {
2972 printf("%s: win = %d winsz = %d sc = %d"
2973 " lro_len %d %d\n",
2974 __func__, scaled_win>>tp->rcv_scale, winsz,
2975 tp->rcv_scale, tp->t_lropktlen, count);
2976 }
2977 tail = ack_chain;
2978 count -= segs_acked; /* accounts for prev_ack packet */
2979 count = (count <= segs_acked) ? 0 : count - segs_acked;
2980 tcpstat.tcps_sndacks++;
2981 so_tc_update_stats(m, so, m_get_service_class(m));
2982 } else {
2983 return NULL;
2984 }
2985 }
2986 else {
2987 tp->t_lropktlen = 0;
2988 return NULL;
2989 }
2990
2991 prev_ack_pkt = ack_chain;
2992
2993 while (count > 0) {
2994 if ((prev_ack + ack_size) < org_ack) {
2995 prev_ack += ack_size;
2996 } else {
2997 /*
2998 * The last ACK sent must have the ACK number that TCP
2999 * thinks is the last sent ACK number.
3000 */
3001 prev_ack = org_ack;
3002 }
3003 mnext = m_dup(prev_ack_pkt, M_DONTWAIT);
3004 if (mnext) {
3005 /* Keep adv window constant for duplicated ACK packets */
3006 scaled_win = win_rtedge - prev_ack;
3007 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3008 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3009 winsz = htons(scaled_win>>tp->rcv_scale);
3010 if (lrodebug == 5) {
3011 printf("%s: winsz = %d ack %x count %d\n",
3012 __func__, scaled_win>>tp->rcv_scale,
3013 prev_ack, count);
3014 }
3015 bcopy(&winsz, mtod(prev_ack_pkt, caddr_t) + twin_offset, 2);
3016 HTONL(prev_ack);
3017 bcopy(&prev_ack, mtod(prev_ack_pkt, caddr_t) + tack_offset, 4);
3018 NTOHL(prev_ack);
3019 tail->m_nextpkt = mnext;
3020 tail = mnext;
3021 count -= segs_acked;
3022 tcpstat.tcps_sndacks++;
3023 so_tc_update_stats(m, so, m_get_service_class(m));
3024 } else {
3025 if (lrodebug == 5) {
3026 printf("%s: failed to alloc mbuf.\n", __func__);
3027 }
3028 break;
3029 }
3030 prev_ack_pkt = mnext;
3031 }
3032 tp->t_lropktlen = 0;
3033 return ack_chain;
3034 }
3035
3036 static int
3037 tcp_recv_throttle (struct tcpcb *tp)
3038 {
3039 uint32_t base_rtt, newsize;
3040 int32_t qdelay;
3041 struct sockbuf *sbrcv = &tp->t_inpcb->inp_socket->so_rcv;
3042
3043 if (tcp_use_rtt_recvbg == 1 &&
3044 TSTMP_SUPPORTED(tp)) {
3045 /*
3046 * Timestamps are supported on this connection. Use
3047 * RTT to look for an increase in latency.
3048 */
3049
3050 /*
3051 * If the connection is already being throttled, leave it
3052 * in that state until rtt comes closer to base rtt
3053 */
3054 if (tp->t_flagsext & TF_RECV_THROTTLE)
3055 return (1);
3056
3057 base_rtt = get_base_rtt(tp);
3058
3059 if (base_rtt != 0 && tp->t_rttcur != 0) {
3060 qdelay = tp->t_rttcur - base_rtt;
3061 /*
3062 * if latency increased on a background flow,
3063 * return 1 to start throttling.
3064 */
3065 if (qdelay > target_qdelay) {
3066 tp->t_flagsext |= TF_RECV_THROTTLE;
3067
3068 /*
3069 * Reduce the recv socket buffer size to
3070 * minimize latecy.
3071 */
3072 if (sbrcv->sb_idealsize >
3073 tcp_recv_throttle_minwin) {
3074 newsize = sbrcv->sb_idealsize >> 1;
3075 /* Set a minimum of 16 K */
3076 newsize =
3077 max(newsize,
3078 tcp_recv_throttle_minwin);
3079 sbrcv->sb_idealsize = newsize;
3080 }
3081 return (1);
3082 } else {
3083 return (0);
3084 }
3085 }
3086 }
3087
3088 /*
3089 * Timestamps are not supported or there is no good RTT
3090 * measurement. Use IPDV in this case.
3091 */
3092 if (tp->acc_iaj > tcp_acc_iaj_react_limit)
3093 return (1);
3094
3095 return (0);
3096 }
mirror of XNU