]> 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-6153.141.1.tar.gz
[apple/xnu.git] / bsd / netinet / tcp_output.c
1 /*
2 * Copyright (c) 2000-2019 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/in_tclass.h>
94 #include <netinet/ip.h>
95 #include <netinet/in_pcb.h>
96 #include <netinet/ip_var.h>
97 #include <mach/sdt.h>
98 #if INET6
99 #include <netinet6/in6_pcb.h>
100 #include <netinet/ip6.h>
101 #include <netinet6/ip6_var.h>
102 #endif
103 #include <netinet/tcp.h>
104 #define TCPOUTFLAGS
105 #include <netinet/tcp_cache.h>
106 #include <netinet/tcp_fsm.h>
107 #include <netinet/tcp_seq.h>
108 #include <netinet/tcp_timer.h>
109 #include <netinet/tcp_var.h>
110 #include <netinet/tcpip.h>
111 #include <netinet/tcp_cc.h>
112 #if TCPDEBUG
113 #include <netinet/tcp_debug.h>
114 #endif
115 #include <netinet/tcp_log.h>
116 #include <sys/kdebug.h>
117 #include <mach/sdt.h>
118
119 #if IPSEC
120 #include <netinet6/ipsec.h>
121 #endif /*IPSEC*/
122
123 #if CONFIG_MACF_NET
124 #include <security/mac_framework.h>
125 #endif /* MAC_SOCKET */
126
127 #include <netinet/lro_ext.h>
128 #if MPTCP
129 #include <netinet/mptcp_var.h>
130 #include <netinet/mptcp.h>
131 #include <netinet/mptcp_opt.h>
132 #endif
133
134 #include <corecrypto/ccaes.h>
135
136 #define DBG_LAYER_BEG NETDBG_CODE(DBG_NETTCP, 1)
137 #define DBG_LAYER_END NETDBG_CODE(DBG_NETTCP, 3)
138 #define DBG_FNC_TCP_OUTPUT NETDBG_CODE(DBG_NETTCP, (4 << 8) | 1)
139
140 SYSCTL_SKMEM_TCP_INT(OID_AUTO, path_mtu_discovery,
141 CTLFLAG_RW | CTLFLAG_LOCKED, int, path_mtu_discovery, 1,
142 "Enable Path MTU Discovery");
143
144 SYSCTL_SKMEM_TCP_INT(OID_AUTO, slowstart_flightsize,
145 CTLFLAG_RW | CTLFLAG_LOCKED, int, ss_fltsz, 1,
146 "Slow start flight size");
147
148 SYSCTL_SKMEM_TCP_INT(OID_AUTO, local_slowstart_flightsize,
149 CTLFLAG_RW | CTLFLAG_LOCKED, int, ss_fltsz_local, 8,
150 "Slow start flight size for local networks");
151
152 int tcp_do_tso = 1;
153 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW | CTLFLAG_LOCKED,
154 &tcp_do_tso, 0, "Enable TCP Segmentation Offload");
155
156 SYSCTL_SKMEM_TCP_INT(OID_AUTO, ecn_setup_percentage,
157 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_ecn_setup_percentage, 100,
158 "Max ECN setup percentage");
159
160 static int
161 sysctl_change_ecn_setting SYSCTL_HANDLER_ARGS
162 {
163 #pragma unused(oidp, arg1, arg2)
164 int i, err = 0, changed = 0;
165 struct ifnet *ifp;
166
167 err = sysctl_io_number(req, tcp_ecn_outbound, sizeof(int32_t),
168 &i, &changed);
169 if (err != 0 || req->newptr == USER_ADDR_NULL)
170 return err;
171
172 if (changed) {
173 if ((tcp_ecn_outbound == 0 || tcp_ecn_outbound == 1) &&
174 (i == 0 || i == 1)) {
175 tcp_ecn_outbound = i;
176 SYSCTL_SKMEM_UPDATE_FIELD(tcp.ecn_initiate_out, tcp_ecn_outbound);
177 return err;
178 }
179 if (tcp_ecn_outbound == 2 && (i == 0 || i == 1)) {
180 /*
181 * Reset ECN enable flags on non-cellular
182 * interfaces so that the system default will take
183 * over
184 */
185 ifnet_head_lock_shared();
186 TAILQ_FOREACH(ifp, &ifnet_head, if_link) {
187 if (!IFNET_IS_CELLULAR(ifp)) {
188 ifnet_lock_exclusive(ifp);
189 ifp->if_eflags &= ~IFEF_ECN_DISABLE;
190 ifp->if_eflags &= ~IFEF_ECN_ENABLE;
191 ifnet_lock_done(ifp);
192 }
193 }
194 ifnet_head_done();
195 } else {
196 /*
197 * Set ECN enable flags on non-cellular
198 * interfaces
199 */
200 ifnet_head_lock_shared();
201 TAILQ_FOREACH(ifp, &ifnet_head, if_link) {
202 if (!IFNET_IS_CELLULAR(ifp)) {
203 ifnet_lock_exclusive(ifp);
204 ifp->if_eflags |= IFEF_ECN_ENABLE;
205 ifp->if_eflags &= ~IFEF_ECN_DISABLE;
206 ifnet_lock_done(ifp);
207 }
208 }
209 ifnet_head_done();
210 }
211 tcp_ecn_outbound = i;
212 SYSCTL_SKMEM_UPDATE_FIELD(tcp.ecn_initiate_out, tcp_ecn_outbound);
213 }
214 /* Change the other one too as the work is done */
215 if (i == 2 || tcp_ecn_inbound == 2) {
216 tcp_ecn_inbound = i;
217 SYSCTL_SKMEM_UPDATE_FIELD(tcp.ecn_negotiate_in, tcp_ecn_inbound);
218 }
219 return err;
220 }
221
222 int tcp_ecn_outbound = 2;
223 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, ecn_initiate_out,
224 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_outbound, 0,
225 sysctl_change_ecn_setting, "IU",
226 "Initiate ECN for outbound connections");
227
228 int tcp_ecn_inbound = 2;
229 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, ecn_negotiate_in,
230 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_inbound, 0,
231 sysctl_change_ecn_setting, "IU",
232 "Initiate ECN for inbound connections");
233
234 SYSCTL_SKMEM_TCP_INT(OID_AUTO, packetchain,
235 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_packet_chaining, 50,
236 "Enable TCP output packet chaining");
237
238 SYSCTL_SKMEM_TCP_INT(OID_AUTO, socket_unlocked_on_output,
239 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_output_unlocked, 1,
240 "Unlock TCP when sending packets down to IP");
241
242 SYSCTL_SKMEM_TCP_INT(OID_AUTO, rfc3390,
243 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_do_rfc3390, 1,
244 "Calculate intial slowstart cwnd depending on MSS");
245
246 SYSCTL_SKMEM_TCP_INT(OID_AUTO, min_iaj_win,
247 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_min_iaj_win, MIN_IAJ_WIN,
248 "Minimum recv win based on inter-packet arrival jitter");
249
250 SYSCTL_SKMEM_TCP_INT(OID_AUTO, acc_iaj_react_limit,
251 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_acc_iaj_react_limit,
252 ACC_IAJ_REACT_LIMIT, "Accumulated IAJ when receiver starts to react");
253
254 SYSCTL_SKMEM_TCP_INT(OID_AUTO, doautosndbuf,
255 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_do_autosendbuf, 1,
256 "Enable send socket buffer auto-tuning");
257
258 SYSCTL_SKMEM_TCP_INT(OID_AUTO, autosndbufinc,
259 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_autosndbuf_inc,
260 8 * 1024, "Increment in send socket bufffer size");
261
262 SYSCTL_SKMEM_TCP_INT(OID_AUTO, autosndbufmax,
263 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_autosndbuf_max, 512 * 1024,
264 "Maximum send socket buffer size");
265
266 SYSCTL_SKMEM_TCP_INT(OID_AUTO, ack_prioritize,
267 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_prioritize_acks, 1,
268 "Prioritize pure acks");
269
270 SYSCTL_SKMEM_TCP_INT(OID_AUTO, rtt_recvbg,
271 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_use_rtt_recvbg, 1,
272 "Use RTT for bg recv algorithm");
273
274 SYSCTL_SKMEM_TCP_INT(OID_AUTO, recv_throttle_minwin,
275 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_recv_throttle_minwin, 16 * 1024,
276 "Minimum recv win for throttling");
277
278 SYSCTL_SKMEM_TCP_INT(OID_AUTO, enable_tlp,
279 CTLFLAG_RW | CTLFLAG_LOCKED,
280 int32_t, tcp_enable_tlp, 1, "Enable Tail loss probe");
281
282 static int32_t packchain_newlist = 0;
283 static int32_t packchain_looped = 0;
284 static int32_t packchain_sent = 0;
285
286 /* temporary: for testing */
287 #if IPSEC
288 extern int ipsec_bypass;
289 #endif
290
291 extern int slowlink_wsize; /* window correction for slow links */
292 #if IPFIREWALL
293 extern int fw_enable; /* firewall check for packet chaining */
294 extern int fw_bypass; /* firewall check: disable packet chaining if there is rules */
295 #endif /* IPFIREWALL */
296
297 extern u_int32_t dlil_filter_disable_tso_count;
298 extern u_int32_t kipf_count;
299
300 static int tcp_ip_output(struct socket *, struct tcpcb *, struct mbuf *,
301 int, struct mbuf *, int, int, boolean_t);
302 static struct mbuf* tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th);
303 static int tcp_recv_throttle(struct tcpcb *tp);
304
305 static int32_t tcp_tfo_check(struct tcpcb *tp, int32_t len)
306 {
307 struct socket *so = tp->t_inpcb->inp_socket;
308 unsigned int optlen = 0;
309 unsigned int cookie_len;
310
311 if (tp->t_flags & TF_NOOPT)
312 goto fallback;
313
314 if (!(tp->t_flagsext & TF_FASTOPEN_FORCE_ENABLE) &&
315 !tcp_heuristic_do_tfo(tp)) {
316 tp->t_tfo_stats |= TFO_S_HEURISTICS_DISABLE;
317 tcpstat.tcps_tfo_heuristics_disable++;
318 goto fallback;
319 }
320
321 if (so->so_flags1 & SOF1_DATA_AUTHENTICATED)
322 return len;
323
324 optlen += TCPOLEN_MAXSEG;
325
326 if (tp->t_flags & TF_REQ_SCALE)
327 optlen += 4;
328
329 #if MPTCP
330 if ((so->so_flags & SOF_MP_SUBFLOW) && mptcp_enable &&
331 (tp->t_rxtshift <= mptcp_mpcap_retries ||
332 (tptomptp(tp)->mpt_mpte->mpte_flags & MPTE_FORCE_ENABLE)))
333 optlen += sizeof(struct mptcp_mpcapable_opt_common) + sizeof(mptcp_key_t);
334 #endif /* MPTCP */
335
336 if (tp->t_flags & TF_REQ_TSTMP)
337 optlen += TCPOLEN_TSTAMP_APPA;
338
339 if (SACK_ENABLED(tp))
340 optlen += TCPOLEN_SACK_PERMITTED;
341
342 /* Now, decide whether to use TFO or not */
343
344 /* Don't even bother trying if there is no space at all... */
345 if (MAX_TCPOPTLEN - optlen < TCPOLEN_FASTOPEN_REQ)
346 goto fallback;
347
348 cookie_len = tcp_cache_get_cookie_len(tp);
349 if (cookie_len == 0)
350 /* No cookie, so we request one */
351 return 0;
352
353 /* There is not enough space for the cookie, so we cannot do TFO */
354 if (MAX_TCPOPTLEN - optlen < cookie_len)
355 goto fallback;
356
357 /* Do not send SYN+data if there is more in the queue than MSS */
358 if (so->so_snd.sb_cc > (tp->t_maxopd - MAX_TCPOPTLEN))
359 goto fallback;
360
361 /* Ok, everything looks good. We can go on and do TFO */
362 return len;
363
364 fallback:
365 tcp_disable_tfo(tp);
366 return 0;
367 }
368
369 /* Returns the number of bytes written to the TCP option-space */
370 static unsigned
371 tcp_tfo_write_cookie_rep(struct tcpcb *tp, unsigned optlen, u_char *opt)
372 {
373 u_char out[CCAES_BLOCK_SIZE];
374 unsigned ret = 0;
375 u_char *bp;
376
377 if ((MAX_TCPOPTLEN - optlen) <
378 (TCPOLEN_FASTOPEN_REQ + TFO_COOKIE_LEN_DEFAULT))
379 return ret;
380
381 tcp_tfo_gen_cookie(tp->t_inpcb, out, sizeof(out));
382
383 bp = opt + optlen;
384
385 *bp++ = TCPOPT_FASTOPEN;
386 *bp++ = 2 + TFO_COOKIE_LEN_DEFAULT;
387 memcpy(bp, out, TFO_COOKIE_LEN_DEFAULT);
388 ret += 2 + TFO_COOKIE_LEN_DEFAULT;
389
390 tp->t_tfo_stats |= TFO_S_COOKIE_SENT;
391 tcpstat.tcps_tfo_cookie_sent++;
392
393 return ret;
394 }
395
396 static unsigned
397 tcp_tfo_write_cookie(struct tcpcb *tp, unsigned optlen, int32_t len,
398 u_char *opt)
399 {
400 u_int8_t tfo_len = MAX_TCPOPTLEN - optlen - TCPOLEN_FASTOPEN_REQ;
401 struct socket *so = tp->t_inpcb->inp_socket;
402 unsigned ret = 0;
403 int res;
404 u_char *bp;
405
406 if (so->so_flags1 & SOF1_DATA_AUTHENTICATED) {
407 /* If there is some data, let's track it */
408 if (len > 0) {
409 tp->t_tfo_stats |= TFO_S_SYN_DATA_SENT;
410 tcpstat.tcps_tfo_syn_data_sent++;
411 }
412
413 return 0;
414 }
415
416 bp = opt + optlen;
417
418 /*
419 * The cookie will be copied in the appropriate place within the
420 * TCP-option space. That way we avoid the need for an intermediate
421 * variable.
422 */
423 res = tcp_cache_get_cookie(tp, bp + TCPOLEN_FASTOPEN_REQ, &tfo_len);
424 if (res == 0) {
425 *bp++ = TCPOPT_FASTOPEN;
426 *bp++ = TCPOLEN_FASTOPEN_REQ;
427 ret += TCPOLEN_FASTOPEN_REQ;
428
429 tp->t_tfo_flags |= TFO_F_COOKIE_REQ;
430
431 tp->t_tfo_stats |= TFO_S_COOKIE_REQ;
432 tcpstat.tcps_tfo_cookie_req++;
433 } else {
434 *bp++ = TCPOPT_FASTOPEN;
435 *bp++ = TCPOLEN_FASTOPEN_REQ + tfo_len;
436
437 ret += TCPOLEN_FASTOPEN_REQ + tfo_len;
438
439 tp->t_tfo_flags |= TFO_F_COOKIE_SENT;
440
441 /* If there is some data, let's track it */
442 if (len > 0) {
443 tp->t_tfo_stats |= TFO_S_SYN_DATA_SENT;
444 tcpstat.tcps_tfo_syn_data_sent++;
445 }
446 }
447
448 return ret;
449 }
450
451 static inline bool
452 tcp_send_ecn_flags_on_syn(struct tcpcb *tp, struct socket *so)
453 {
454 return !((tp->ecn_flags & TE_SETUPSENT ||
455 (so->so_flags & SOF_MP_SUBFLOW) ||
456 (tfo_enabled(tp))));
457 }
458
459 void
460 tcp_set_ecn(struct tcpcb *tp, struct ifnet *ifp)
461 {
462 boolean_t inbound;
463
464 /*
465 * Socket option has precedence
466 */
467 if (tp->ecn_flags & TE_ECN_MODE_ENABLE) {
468 tp->ecn_flags |= TE_ENABLE_ECN;
469 goto check_heuristic;
470 }
471
472 if (tp->ecn_flags & TE_ECN_MODE_DISABLE) {
473 tp->ecn_flags &= ~TE_ENABLE_ECN;
474 return;
475 }
476 /*
477 * Per interface setting comes next
478 */
479 if (ifp != NULL) {
480 if (ifp->if_eflags & IFEF_ECN_ENABLE) {
481 tp->ecn_flags |= TE_ENABLE_ECN;
482 goto check_heuristic;
483 }
484
485 if (ifp->if_eflags & IFEF_ECN_DISABLE) {
486 tp->ecn_flags &= ~TE_ENABLE_ECN;
487 return;
488 }
489 }
490 /*
491 * System wide settings come last
492 */
493 inbound = (tp->t_inpcb->inp_socket->so_head != NULL);
494 if ((inbound && tcp_ecn_inbound == 1) ||
495 (!inbound && tcp_ecn_outbound == 1)) {
496 tp->ecn_flags |= TE_ENABLE_ECN;
497 goto check_heuristic;
498 } else {
499 tp->ecn_flags &= ~TE_ENABLE_ECN;
500 }
501
502 return;
503
504 check_heuristic:
505 if (!tcp_heuristic_do_ecn(tp))
506 tp->ecn_flags &= ~TE_ENABLE_ECN;
507
508 /*
509 * If the interface setting, system-level setting and heuristics
510 * allow to enable ECN, randomly select 5% of connections to
511 * enable it
512 */
513 if ((tp->ecn_flags & (TE_ECN_MODE_ENABLE | TE_ECN_MODE_DISABLE
514 | TE_ENABLE_ECN)) == TE_ENABLE_ECN) {
515 /*
516 * Use the random value in iss for randomizing
517 * this selection
518 */
519 if ((tp->iss % 100) >= tcp_ecn_setup_percentage)
520 tp->ecn_flags &= ~TE_ENABLE_ECN;
521 }
522 }
523
524 /*
525 * Tcp output routine: figure out what should be sent and send it.
526 *
527 * Returns: 0 Success
528 * EADDRNOTAVAIL
529 * ENOBUFS
530 * EMSGSIZE
531 * EHOSTUNREACH
532 * ENETDOWN
533 * ip_output_list:ENOMEM
534 * ip_output_list:EADDRNOTAVAIL
535 * ip_output_list:ENETUNREACH
536 * ip_output_list:EHOSTUNREACH
537 * ip_output_list:EACCES
538 * ip_output_list:EMSGSIZE
539 * ip_output_list:ENOBUFS
540 * ip_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL]
541 * ip6_output_list:EINVAL
542 * ip6_output_list:EOPNOTSUPP
543 * ip6_output_list:EHOSTUNREACH
544 * ip6_output_list:EADDRNOTAVAIL
545 * ip6_output_list:ENETUNREACH
546 * ip6_output_list:EMSGSIZE
547 * ip6_output_list:ENOBUFS
548 * ip6_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL]
549 */
550 int
551 tcp_output(struct tcpcb *tp)
552 {
553 struct inpcb *inp = tp->t_inpcb;
554 struct socket *so = inp->inp_socket;
555 int32_t len, recwin, sendwin, off;
556 int flags, error;
557 struct mbuf *m;
558 struct ip *ip = NULL;
559 struct ipovly *ipov = NULL;
560 #if INET6
561 struct ip6_hdr *ip6 = NULL;
562 #endif /* INET6 */
563 struct tcphdr *th;
564 u_char opt[TCP_MAXOLEN];
565 unsigned ipoptlen, optlen, hdrlen;
566 int idle, sendalot, lost = 0;
567 int i, sack_rxmit;
568 int tso = 0;
569 int sack_bytes_rxmt;
570 tcp_seq old_snd_nxt = 0;
571 struct sackhole *p;
572 #if IPSEC
573 unsigned ipsec_optlen = 0;
574 #endif /* IPSEC */
575 int idle_time = 0;
576 struct mbuf *packetlist = NULL;
577 struct mbuf *tp_inp_options = inp->inp_depend4.inp4_options;
578 #if INET6
579 int isipv6 = inp->inp_vflag & INP_IPV6 ;
580 #else
581 int isipv6 = 0;
582 #endif
583 short packchain_listadd = 0;
584 int so_options = so->so_options;
585 struct rtentry *rt;
586 u_int32_t svc_flags = 0, allocated_len;
587 u_int32_t lro_ackmore = (tp->t_lropktlen != 0) ? 1 : 0;
588 struct mbuf *mnext = NULL;
589 int sackoptlen = 0;
590 #if MPTCP
591 boolean_t mptcp_acknow;
592 #endif /* MPTCP */
593 boolean_t cell = FALSE;
594 boolean_t wifi = FALSE;
595 boolean_t wired = FALSE;
596 boolean_t sack_rescue_rxt = FALSE;
597 int sotc = so->so_traffic_class;
598
599 /*
600 * Determine length of data that should be transmitted,
601 * and flags that will be used.
602 * If there is some data or critical controls (SYN, RST)
603 * to send, then transmit; otherwise, investigate further.
604 */
605 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
606
607 /* Since idle_time is signed integer, the following integer subtraction
608 * will take care of wrap around of tcp_now
609 */
610 idle_time = tcp_now - tp->t_rcvtime;
611 if (idle && idle_time >= TCP_IDLETIMEOUT(tp)) {
612 if (CC_ALGO(tp)->after_idle != NULL &&
613 (tp->tcp_cc_index != TCP_CC_ALGO_CUBIC_INDEX ||
614 idle_time >= TCP_CC_CWND_NONVALIDATED_PERIOD)) {
615 CC_ALGO(tp)->after_idle(tp);
616 tcp_ccdbg_trace(tp, NULL, TCP_CC_IDLE_TIMEOUT);
617 }
618
619 /*
620 * Do some other tasks that need to be done after
621 * idle time
622 */
623 if (!SLIST_EMPTY(&tp->t_rxt_segments))
624 tcp_rxtseg_clean(tp);
625
626 /* If stretch ack was auto-disabled, re-evaluate it */
627 tcp_cc_after_idle_stretchack(tp);
628 }
629 tp->t_flags &= ~TF_LASTIDLE;
630 if (idle) {
631 if (tp->t_flags & TF_MORETOCOME) {
632 tp->t_flags |= TF_LASTIDLE;
633 idle = 0;
634 }
635 }
636 #if MPTCP
637 if (tp->t_mpflags & TMPF_RESET) {
638 tcp_check_timer_state(tp);
639 /*
640 * Once a RST has been sent for an MPTCP subflow,
641 * the subflow socket stays around until deleted.
642 * No packets such as FINs must be sent after RST.
643 */
644 return 0;
645 }
646 #endif /* MPTCP */
647
648 again:
649 #if MPTCP
650 mptcp_acknow = FALSE;
651 #endif
652
653 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_START, 0,0,0,0,0);
654
655 #if INET6
656 if (isipv6) {
657 KERNEL_DEBUG(DBG_LAYER_BEG,
658 ((inp->inp_fport << 16) | inp->inp_lport),
659 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
660 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
661 sendalot,0,0);
662 } else
663 #endif
664
665 {
666 KERNEL_DEBUG(DBG_LAYER_BEG,
667 ((inp->inp_fport << 16) | inp->inp_lport),
668 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
669 (inp->inp_faddr.s_addr & 0xffff)),
670 sendalot,0,0);
671 }
672 /*
673 * If the route generation id changed, we need to check that our
674 * local (source) IP address is still valid. If it isn't either
675 * return error or silently do nothing (assuming the address will
676 * come back before the TCP connection times out).
677 */
678 rt = inp->inp_route.ro_rt;
679 if (rt != NULL && ROUTE_UNUSABLE(&tp->t_inpcb->inp_route)) {
680 struct ifnet *ifp;
681 struct in_ifaddr *ia = NULL;
682 struct in6_ifaddr *ia6 = NULL;
683 int found_srcaddr = 0;
684
685 /* disable multipages at the socket */
686 somultipages(so, FALSE);
687
688 /* Disable TSO for the socket until we know more */
689 tp->t_flags &= ~TF_TSO;
690
691 soif2kcl(so, FALSE);
692
693 if (isipv6) {
694 ia6 = ifa_foraddr6(&inp->in6p_laddr);
695 if (ia6 != NULL)
696 found_srcaddr = 1;
697 } else {
698 ia = ifa_foraddr(inp->inp_laddr.s_addr);
699 if (ia != NULL)
700 found_srcaddr = 1;
701 }
702
703 /* check that the source address is still valid */
704 if (found_srcaddr == 0) {
705 soevent(so,
706 (SO_FILT_HINT_LOCKED | SO_FILT_HINT_NOSRCADDR));
707
708 if (tp->t_state >= TCPS_CLOSE_WAIT) {
709 tcp_drop(tp, EADDRNOTAVAIL);
710 return EADDRNOTAVAIL;
711 }
712
713 /*
714 * Set retransmit timer if it wasn't set,
715 * reset Persist timer and shift register as the
716 * advertised peer window may not be valid anymore
717 */
718 if (tp->t_timer[TCPT_REXMT] == 0) {
719 tp->t_timer[TCPT_REXMT] =
720 OFFSET_FROM_START(tp, tp->t_rxtcur);
721 if (tp->t_timer[TCPT_PERSIST] != 0) {
722 tp->t_timer[TCPT_PERSIST] = 0;
723 tp->t_persist_stop = 0;
724 TCP_RESET_REXMT_STATE(tp);
725 }
726 }
727
728 if (tp->t_pktlist_head != NULL)
729 m_freem_list(tp->t_pktlist_head);
730 TCP_PKTLIST_CLEAR(tp);
731
732 /* drop connection if source address isn't available */
733 if (so->so_flags & SOF_NOADDRAVAIL) {
734 tcp_drop(tp, EADDRNOTAVAIL);
735 return EADDRNOTAVAIL;
736 } else {
737 tcp_check_timer_state(tp);
738 return 0; /* silently ignore, keep data in socket: address may be back */
739 }
740 }
741 if (ia != NULL)
742 IFA_REMREF(&ia->ia_ifa);
743
744 if (ia6 != NULL)
745 IFA_REMREF(&ia6->ia_ifa);
746
747 /*
748 * Address is still valid; check for multipages capability
749 * again in case the outgoing interface has changed.
750 */
751 RT_LOCK(rt);
752 if ((ifp = rt->rt_ifp) != NULL) {
753 somultipages(so, (ifp->if_hwassist & IFNET_MULTIPAGES));
754 tcp_set_tso(tp, ifp);
755 soif2kcl(so, (ifp->if_eflags & IFEF_2KCL));
756 tcp_set_ecn(tp, ifp);
757 }
758 if (rt->rt_flags & RTF_UP)
759 RT_GENID_SYNC(rt);
760 /*
761 * See if we should do MTU discovery. Don't do it if:
762 * 1) it is disabled via the sysctl
763 * 2) the route isn't up
764 * 3) the MTU is locked (if it is, then discovery
765 * has been disabled)
766 */
767
768 if (!path_mtu_discovery || ((rt != NULL) &&
769 (!(rt->rt_flags & RTF_UP) ||
770 (rt->rt_rmx.rmx_locks & RTV_MTU))))
771 tp->t_flags &= ~TF_PMTUD;
772 else
773 tp->t_flags |= TF_PMTUD;
774
775 RT_UNLOCK(rt);
776 }
777
778 if (rt != NULL) {
779 cell = IFNET_IS_CELLULAR(rt->rt_ifp);
780 wifi = (!cell && IFNET_IS_WIFI(rt->rt_ifp));
781 wired = (!wifi && IFNET_IS_WIRED(rt->rt_ifp));
782 }
783
784 /*
785 * If we've recently taken a timeout, snd_max will be greater than
786 * snd_nxt. There may be SACK information that allows us to avoid
787 * resending already delivered data. Adjust snd_nxt accordingly.
788 */
789 if (SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
790 tcp_sack_adjust(tp);
791 sendalot = 0;
792 off = tp->snd_nxt - tp->snd_una;
793 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
794
795 if (tp->t_flags & TF_SLOWLINK && slowlink_wsize > 0)
796 sendwin = min(sendwin, slowlink_wsize);
797
798 flags = tcp_outflags[tp->t_state];
799 /*
800 * Send any SACK-generated retransmissions. If we're explicitly
801 * trying to send out new data (when sendalot is 1), bypass this
802 * function. If we retransmit in fast recovery mode, decrement
803 * snd_cwnd, since we're replacing a (future) new transmission
804 * with a retransmission now, and we previously incremented
805 * snd_cwnd in tcp_input().
806 */
807 /*
808 * Still in sack recovery , reset rxmit flag to zero.
809 */
810 sack_rxmit = 0;
811 sack_bytes_rxmt = 0;
812 len = 0;
813 p = NULL;
814 if (SACK_ENABLED(tp) && IN_FASTRECOVERY(tp) &&
815 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
816 int32_t cwin;
817
818 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
819 if (cwin < 0)
820 cwin = 0;
821 /* Do not retransmit SACK segments beyond snd_recover */
822 if (SEQ_GT(p->end, tp->snd_recover)) {
823 /*
824 * (At least) part of sack hole extends beyond
825 * snd_recover. Check to see if we can rexmit data
826 * for this hole.
827 */
828 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
829 /*
830 * Can't rexmit any more data for this hole.
831 * That data will be rexmitted in the next
832 * sack recovery episode, when snd_recover
833 * moves past p->rxmit.
834 */
835 p = NULL;
836 goto after_sack_rexmit;
837 } else
838 /* Can rexmit part of the current hole */
839 len = ((int32_t)min(cwin,
840 tp->snd_recover - p->rxmit));
841 } else {
842 len = ((int32_t)min(cwin, p->end - p->rxmit));
843 }
844 if (len > 0) {
845 off = p->rxmit - tp->snd_una;
846 sack_rxmit = 1;
847 sendalot = 1;
848 tcpstat.tcps_sack_rexmits++;
849 tcpstat.tcps_sack_rexmit_bytes +=
850 min(len, tp->t_maxseg);
851 } else {
852 len = 0;
853 }
854 }
855 after_sack_rexmit:
856 /*
857 * Get standard flags, and add SYN or FIN if requested by 'hidden'
858 * state flags.
859 */
860 if (tp->t_flags & TF_NEEDFIN)
861 flags |= TH_FIN;
862 if (tp->t_flags & TF_NEEDSYN)
863 flags |= TH_SYN;
864
865 /*
866 * If in persist timeout with window of 0, send 1 byte.
867 * Otherwise, if window is small but nonzero
868 * and timer expired, we will send what we can
869 * and go to transmit state.
870 */
871 if (tp->t_flagsext & TF_FORCE) {
872 if (sendwin == 0) {
873 /*
874 * If we still have some data to send, then
875 * clear the FIN bit. Usually this would
876 * happen below when it realizes that we
877 * aren't sending all the data. However,
878 * if we have exactly 1 byte of unsent data,
879 * then it won't clear the FIN bit below,
880 * and if we are in persist state, we wind
881 * up sending the packet without recording
882 * that we sent the FIN bit.
883 *
884 * We can't just blindly clear the FIN bit,
885 * because if we don't have any more data
886 * to send then the probe will be the FIN
887 * itself.
888 */
889 if (off < so->so_snd.sb_cc)
890 flags &= ~TH_FIN;
891 sendwin = 1;
892 } else {
893 tp->t_timer[TCPT_PERSIST] = 0;
894 tp->t_persist_stop = 0;
895 TCP_RESET_REXMT_STATE(tp);
896 }
897 }
898
899 /*
900 * If snd_nxt == snd_max and we have transmitted a FIN, the
901 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
902 * a negative length. This can also occur when TCP opens up
903 * its congestion window while receiving additional duplicate
904 * acks after fast-retransmit because TCP will reset snd_nxt
905 * to snd_max after the fast-retransmit.
906 *
907 * In the normal retransmit-FIN-only case, however, snd_nxt will
908 * be set to snd_una, the offset will be 0, and the length may
909 * wind up 0.
910 *
911 * If sack_rxmit is true we are retransmitting from the scoreboard
912 * in which case len is already set.
913 */
914 if (sack_rxmit == 0) {
915 if (sack_bytes_rxmt == 0) {
916 len = min(so->so_snd.sb_cc, sendwin) - off;
917 } else {
918 int32_t cwin;
919
920 cwin = tp->snd_cwnd -
921 (tp->snd_nxt - tp->sack_newdata) -
922 sack_bytes_rxmt;
923 if (cwin < 0)
924 cwin = 0;
925 /*
926 * We are inside of a SACK recovery episode and are
927 * sending new data, having retransmitted all the
928 * data possible in the scoreboard.
929 */
930 len = min(so->so_snd.sb_cc, tp->snd_wnd)
931 - off;
932 /*
933 * Don't remove this (len > 0) check !
934 * We explicitly check for len > 0 here (although it
935 * isn't really necessary), to work around a gcc
936 * optimization issue - to force gcc to compute
937 * len above. Without this check, the computation
938 * of len is bungled by the optimizer.
939 */
940 if (len > 0) {
941 len = imin(len, cwin);
942 } else {
943 len = 0;
944 }
945 /*
946 * At this point SACK recovery can not send any
947 * data from scoreboard or any new data. Check
948 * if we can do a rescue retransmit towards the
949 * tail end of recovery window.
950 */
951 if (len == 0 && cwin > 0 &&
952 SEQ_LT(tp->snd_fack, tp->snd_recover) &&
953 !(tp->t_flagsext & TF_RESCUE_RXT)) {
954 len = min((tp->snd_recover - tp->snd_fack),
955 tp->t_maxseg);
956 len = imin(len, cwin);
957 old_snd_nxt = tp->snd_nxt;
958 sack_rescue_rxt = TRUE;
959 tp->snd_nxt = tp->snd_recover - len;
960 /*
961 * If FIN has been sent, snd_max
962 * must have been advanced to cover it.
963 */
964 if ((tp->t_flags & TF_SENTFIN) &&
965 tp->snd_max == tp->snd_recover)
966 tp->snd_nxt--;
967
968 off = tp->snd_nxt - tp->snd_una;
969 sendalot = 0;
970 tp->t_flagsext |= TF_RESCUE_RXT;
971 }
972 }
973 }
974
975 /*
976 * Lop off SYN bit if it has already been sent. However, if this
977 * is SYN-SENT state and if segment contains data and if we don't
978 * know that foreign host supports TAO, suppress sending segment.
979 */
980 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
981 if (tp->t_state == TCPS_SYN_RECEIVED && tfo_enabled(tp) && tp->snd_nxt == tp->snd_una + 1) {
982 /* We are sending the SYN again! */
983 off--;
984 len++;
985 } else {
986 if (tp->t_state != TCPS_SYN_RECEIVED || tfo_enabled(tp)) {
987 flags &= ~TH_SYN;
988 }
989
990 off--;
991 len++;
992 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
993 while (inp->inp_sndinprog_cnt == 0 &&
994 tp->t_pktlist_head != NULL) {
995 packetlist = tp->t_pktlist_head;
996 packchain_listadd = tp->t_lastchain;
997 packchain_sent++;
998 TCP_PKTLIST_CLEAR(tp);
999
1000 error = tcp_ip_output(so, tp, packetlist,
1001 packchain_listadd, tp_inp_options,
1002 (so_options & SO_DONTROUTE),
1003 (sack_rxmit || (sack_bytes_rxmt != 0)),
1004 isipv6);
1005 }
1006
1007 /*
1008 * tcp was closed while we were in ip,
1009 * resume close
1010 */
1011 if (inp->inp_sndinprog_cnt == 0 &&
1012 (tp->t_flags & TF_CLOSING)) {
1013 tp->t_flags &= ~TF_CLOSING;
1014 (void) tcp_close(tp);
1015 } else {
1016 tcp_check_timer_state(tp);
1017 }
1018 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,
1019 0,0,0,0,0);
1020 return 0;
1021 }
1022 }
1023 }
1024
1025 /*
1026 * Be careful not to send data and/or FIN on SYN segments.
1027 * This measure is needed to prevent interoperability problems
1028 * with not fully conformant TCP implementations.
1029 *
1030 * In case of TFO, we handle the setting of the len in
1031 * tcp_tfo_check. In case TFO is not enabled, never ever send
1032 * SYN+data.
1033 */
1034 if ((flags & TH_SYN) && !tfo_enabled(tp)) {
1035 len = 0;
1036 flags &= ~TH_FIN;
1037 }
1038
1039 /*
1040 * Don't send a RST with data.
1041 */
1042 if (flags & TH_RST)
1043 len = 0;
1044
1045 if ((flags & TH_SYN) && tp->t_state <= TCPS_SYN_SENT && tfo_enabled(tp))
1046 len = tcp_tfo_check(tp, len);
1047
1048 /*
1049 * The check here used to be (len < 0). Some times len is zero
1050 * when the congestion window is closed and we need to check
1051 * if persist timer has to be set in that case. But don't set
1052 * persist until connection is established.
1053 */
1054 if (len <= 0 && !(flags & TH_SYN)) {
1055 /*
1056 * If FIN has been sent but not acked,
1057 * but we haven't been called to retransmit,
1058 * len will be < 0. Otherwise, window shrank
1059 * after we sent into it. If window shrank to 0,
1060 * cancel pending retransmit, pull snd_nxt back
1061 * to (closed) window, and set the persist timer
1062 * if it isn't already going. If the window didn't
1063 * close completely, just wait for an ACK.
1064 */
1065 len = 0;
1066 if (sendwin == 0) {
1067 tp->t_timer[TCPT_REXMT] = 0;
1068 tp->t_timer[TCPT_PTO] = 0;
1069 TCP_RESET_REXMT_STATE(tp);
1070 tp->snd_nxt = tp->snd_una;
1071 off = 0;
1072 if (tp->t_timer[TCPT_PERSIST] == 0)
1073 tcp_setpersist(tp);
1074 }
1075 }
1076
1077 /*
1078 * Automatic sizing of send socket buffer. Increase the send
1079 * socket buffer size if all of the following criteria are met
1080 * 1. the receiver has enough buffer space for this data
1081 * 2. send buffer is filled to 7/8th with data (so we actually
1082 * have data to make use of it);
1083 * 3. our send window (slow start and congestion controlled) is
1084 * larger than sent but unacknowledged data in send buffer.
1085 */
1086 if (tcp_do_autosendbuf == 1 &&
1087 !INP_WAIT_FOR_IF_FEEDBACK(inp) && !IN_FASTRECOVERY(tp) &&
1088 (so->so_snd.sb_flags & (SB_AUTOSIZE | SB_TRIM)) == SB_AUTOSIZE &&
1089 tcp_cansbgrow(&so->so_snd)) {
1090 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
1091 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
1092 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
1093 if (sbreserve(&so->so_snd,
1094 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
1095 tcp_autosndbuf_max)) == 1) {
1096 so->so_snd.sb_idealsize = so->so_snd.sb_hiwat;
1097 }
1098 }
1099 }
1100
1101 /*
1102 * Truncate to the maximum segment length or enable TCP Segmentation
1103 * Offloading (if supported by hardware) and ensure that FIN is removed
1104 * if the length no longer contains the last data byte.
1105 *
1106 * TSO may only be used if we are in a pure bulk sending state.
1107 * The presence of TCP-MD5, SACK retransmits, SACK advertizements,
1108 * ipfw rules and IP options, as well as disabling hardware checksum
1109 * offload prevent using TSO. With TSO the TCP header is the same
1110 * (except for the sequence number) for all generated packets. This
1111 * makes it impossible to transmit any options which vary per generated
1112 * segment or packet.
1113 *
1114 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
1115 * removal of FIN (if not already catched here) are handled later after
1116 * the exact length of the TCP options are known.
1117 */
1118 #if IPSEC
1119 /*
1120 * Pre-calculate here as we save another lookup into the darknesses
1121 * of IPsec that way and can actually decide if TSO is ok.
1122 */
1123 if (ipsec_bypass == 0)
1124 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
1125 #endif
1126 if (len > tp->t_maxseg) {
1127 if ((tp->t_flags & TF_TSO) && tcp_do_tso && hwcksum_tx &&
1128 ip_use_randomid && kipf_count == 0 &&
1129 dlil_filter_disable_tso_count == 0 &&
1130 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
1131 sack_bytes_rxmt == 0 &&
1132 inp->inp_options == NULL &&
1133 inp->in6p_options == NULL
1134 #if IPSEC
1135 && ipsec_optlen == 0
1136 #endif
1137 #if IPFIREWALL
1138 && (fw_enable == 0 || fw_bypass)
1139 #endif
1140 ) {
1141 tso = 1;
1142 sendalot = 0;
1143 } else {
1144 len = tp->t_maxseg;
1145 sendalot = 1;
1146 tso = 0;
1147 }
1148 }
1149
1150 /* Send one segment or less as a tail loss probe */
1151 if (tp->t_flagsext & TF_SENT_TLPROBE) {
1152 len = min(len, tp->t_maxseg);
1153 sendalot = 0;
1154 tso = 0;
1155 }
1156
1157 #if MPTCP
1158 if (so->so_flags & SOF_MP_SUBFLOW && off < 0) {
1159 os_log_error(mptcp_log_handle, "%s - %lx: offset is negative! len %d off %d\n",
1160 __func__, (unsigned long)VM_KERNEL_ADDRPERM(tp->t_mpsub->mpts_mpte),
1161 len, off);
1162 }
1163
1164 if ((so->so_flags & SOF_MP_SUBFLOW) &&
1165 !(tp->t_mpflags & TMPF_TCP_FALLBACK)) {
1166 int newlen = len;
1167 if (tp->t_state >= TCPS_ESTABLISHED &&
1168 (tp->t_mpflags & TMPF_SND_MPPRIO ||
1169 tp->t_mpflags & TMPF_SND_REM_ADDR ||
1170 tp->t_mpflags & TMPF_SND_MPFAIL ||
1171 tp->t_mpflags & TMPF_SND_KEYS ||
1172 tp->t_mpflags & TMPF_SND_JACK)) {
1173 if (len > 0) {
1174 len = 0;
1175 }
1176 /*
1177 * On a new subflow, don't try to send again, because
1178 * we are still waiting for the fourth ack.
1179 */
1180 if (!(tp->t_mpflags & TMPF_PREESTABLISHED))
1181 sendalot = 1;
1182 mptcp_acknow = TRUE;
1183 } else {
1184 mptcp_acknow = FALSE;
1185 }
1186 /*
1187 * The contiguous bytes in the subflow socket buffer can be
1188 * discontiguous at the MPTCP level. Since only one DSS
1189 * option can be sent in one packet, reduce length to match
1190 * the contiguous MPTCP level. Set sendalot to send remainder.
1191 */
1192 if (len > 0 && off >= 0) {
1193 newlen = mptcp_adj_sendlen(so, off);
1194 }
1195
1196 if (newlen < len) {
1197 len = newlen;
1198 }
1199 }
1200 #endif /* MPTCP */
1201
1202 /*
1203 * If the socket is capable of doing unordered send,
1204 * pull the amount of data that can be sent from the
1205 * unordered priority queues to the serial queue in
1206 * the socket buffer. If bytes are not yet available
1207 * in the highest priority message, we may not be able
1208 * to send any new data.
1209 */
1210 if (so->so_flags & SOF_ENABLE_MSGS) {
1211 if ((off + len) >
1212 so->so_msg_state->msg_serial_bytes) {
1213 sbpull_unordered_data(so, off, len);
1214
1215 /* check if len needs to be modified */
1216 if ((off + len) >
1217 so->so_msg_state->msg_serial_bytes) {
1218 len = so->so_msg_state->msg_serial_bytes - off;
1219 if (len <= 0) {
1220 len = 0;
1221 tcpstat.tcps_msg_sndwaithipri++;
1222 }
1223 }
1224 }
1225 }
1226
1227 if (sack_rxmit) {
1228 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
1229 flags &= ~TH_FIN;
1230 } else {
1231 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
1232 flags &= ~TH_FIN;
1233 }
1234 /*
1235 * Compare available window to amount of window
1236 * known to peer (as advertised window less
1237 * next expected input). If the difference is at least two
1238 * max size segments, or at least 25% of the maximum possible
1239 * window, then want to send a window update to peer.
1240 */
1241 recwin = tcp_sbspace(tp);
1242
1243 if (!(so->so_flags & SOF_MP_SUBFLOW)) {
1244 if (recwin < (int32_t)(so->so_rcv.sb_hiwat / 4) &&
1245 recwin < (int)tp->t_maxseg) {
1246 recwin = 0;
1247 }
1248 } else {
1249 struct mptcb *mp_tp = tptomptp(tp);
1250 struct socket *mp_so = mptetoso(mp_tp->mpt_mpte);
1251
1252 if (recwin < (int32_t)(mp_so->so_rcv.sb_hiwat / 4) &&
1253 recwin < (int)tp->t_maxseg) {
1254 recwin = 0;
1255 }
1256 }
1257
1258 #if TRAFFIC_MGT
1259 if (tcp_recv_bg == 1 || IS_TCP_RECV_BG(so)) {
1260 if (recwin > 0 && tcp_recv_throttle(tp)) {
1261 uint32_t min_iaj_win = tcp_min_iaj_win * tp->t_maxseg;
1262 uint32_t bg_rwintop = tp->rcv_adv;
1263 if (SEQ_LT(bg_rwintop, tp->rcv_nxt + min_iaj_win))
1264 bg_rwintop = tp->rcv_nxt + min_iaj_win;
1265 recwin = imin((int32_t)(bg_rwintop - tp->rcv_nxt),
1266 recwin);
1267 if (recwin < 0)
1268 recwin = 0;
1269 }
1270 }
1271 #endif /* TRAFFIC_MGT */
1272
1273 if (recwin > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
1274 recwin = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
1275
1276 if (!(so->so_flags & SOF_MP_SUBFLOW)) {
1277 if (recwin < (int32_t)(tp->rcv_adv - tp->rcv_nxt)) {
1278 recwin = (int32_t)(tp->rcv_adv - tp->rcv_nxt);
1279 }
1280 } else {
1281 struct mptcb *mp_tp = tptomptp(tp);
1282
1283 /* Don't remove what we announced at the MPTCP-layer */
1284 if (recwin < (int32_t)(mp_tp->mpt_rcvadv - (uint32_t)mp_tp->mpt_rcvnxt)) {
1285 recwin = (int32_t)(mp_tp->mpt_rcvadv - (uint32_t)mp_tp->mpt_rcvnxt);
1286 }
1287 }
1288
1289 /*
1290 * Sender silly window avoidance. We transmit under the following
1291 * conditions when len is non-zero:
1292 *
1293 * - we've timed out (e.g. persist timer)
1294 * - we need to retransmit
1295 * - We have a full segment (or more with TSO)
1296 * - This is the last buffer in a write()/send() and we are
1297 * either idle or running NODELAY
1298 * - we have more then 1/2 the maximum send window's worth of
1299 * data (receiver may be limited the window size)
1300 */
1301 if (len) {
1302 if (tp->t_flagsext & TF_FORCE)
1303 goto send;
1304 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1305 goto send;
1306 if (sack_rxmit)
1307 goto send;
1308
1309 /*
1310 * If this here is the first segment after SYN/ACK and TFO
1311 * is being used, then we always send it, regardless of Nagle,...
1312 */
1313 if (tp->t_state == TCPS_SYN_RECEIVED &&
1314 tfo_enabled(tp) &&
1315 (tp->t_tfo_flags & TFO_F_COOKIE_VALID) &&
1316 tp->snd_nxt == tp->iss + 1)
1317 goto send;
1318
1319 /*
1320 * Send new data on the connection only if it is
1321 * not flow controlled
1322 */
1323 if (!INP_WAIT_FOR_IF_FEEDBACK(inp) ||
1324 tp->t_state != TCPS_ESTABLISHED) {
1325 if (len >= tp->t_maxseg)
1326 goto send;
1327
1328 if (!(tp->t_flags & TF_MORETOCOME) &&
1329 (idle || tp->t_flags & TF_NODELAY ||
1330 (tp->t_flags & TF_MAXSEGSNT) ||
1331 ALLOW_LIMITED_TRANSMIT(tp)) &&
1332 (tp->t_flags & TF_NOPUSH) == 0 &&
1333 (len + off >= so->so_snd.sb_cc ||
1334 /*
1335 * MPTCP needs to respect the DSS-mappings. So, it
1336 * may be sending data that *could* have been
1337 * coalesced, but cannot because of
1338 * mptcp_adj_sendlen().
1339 */
1340 so->so_flags & SOF_MP_SUBFLOW))
1341 goto send;
1342 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
1343 goto send;
1344 } else {
1345 tcpstat.tcps_fcholdpacket++;
1346 }
1347 }
1348
1349 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
1350 /*
1351 * "adv" is the amount we can increase the window,
1352 * taking into account that we are limited by
1353 * TCP_MAXWIN << tp->rcv_scale.
1354 */
1355 int32_t adv, oldwin = 0;
1356 adv = imin(recwin, (int)TCP_MAXWIN << tp->rcv_scale) -
1357 (tp->rcv_adv - tp->rcv_nxt);
1358
1359 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
1360 oldwin = tp->rcv_adv - tp->rcv_nxt;
1361
1362 if (adv >= (int32_t) (2 * tp->t_maxseg)) {
1363 /*
1364 * Update only if the resulting scaled value of
1365 * the window changed, or if there is a change in
1366 * the sequence since the last ack. This avoids
1367 * what appears as dupe ACKS (see rdar://5640997)
1368 *
1369 * If streaming is detected avoid sending too many
1370 * window updates. We will depend on the delack
1371 * timer to send a window update when needed.
1372 */
1373 if (!(tp->t_flags & TF_STRETCHACK) &&
1374 (tp->last_ack_sent != tp->rcv_nxt ||
1375 ((oldwin + adv) >> tp->rcv_scale) >
1376 (oldwin >> tp->rcv_scale))) {
1377 goto send;
1378 }
1379
1380 }
1381 if (4 * adv >= (int32_t) so->so_rcv.sb_hiwat)
1382 goto send;
1383
1384 /*
1385 * Make sure that the delayed ack timer is set if
1386 * we delayed sending a window update because of
1387 * streaming detection.
1388 */
1389 if ((tp->t_flags & TF_STRETCHACK) &&
1390 !(tp->t_flags & TF_DELACK)) {
1391 tp->t_flags |= TF_DELACK;
1392 tp->t_timer[TCPT_DELACK] =
1393 OFFSET_FROM_START(tp, tcp_delack);
1394 }
1395 }
1396
1397 /*
1398 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
1399 * is also a catch-all for the retransmit timer timeout case.
1400 */
1401 if (tp->t_flags & TF_ACKNOW)
1402 goto send;
1403 if ((flags & TH_RST) ||
1404 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
1405 goto send;
1406 if (SEQ_GT(tp->snd_up, tp->snd_una))
1407 goto send;
1408 #if MPTCP
1409 if (mptcp_acknow)
1410 goto send;
1411 #endif /* MPTCP */
1412 /*
1413 * If our state indicates that FIN should be sent
1414 * and we have not yet done so, then we need to send.
1415 */
1416 if ((flags & TH_FIN) &&
1417 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
1418 goto send;
1419 /*
1420 * In SACK, it is possible for tcp_output to fail to send a segment
1421 * after the retransmission timer has been turned off. Make sure
1422 * that the retransmission timer is set.
1423 */
1424 if (SACK_ENABLED(tp) && (tp->t_state >= TCPS_ESTABLISHED) &&
1425 SEQ_GT(tp->snd_max, tp->snd_una) &&
1426 tp->t_timer[TCPT_REXMT] == 0 &&
1427 tp->t_timer[TCPT_PERSIST] == 0) {
1428 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp,
1429 tp->t_rxtcur);
1430 goto just_return;
1431 }
1432 /*
1433 * TCP window updates are not reliable, rather a polling protocol
1434 * using ``persist'' packets is used to insure receipt of window
1435 * updates. The three ``states'' for the output side are:
1436 * idle not doing retransmits or persists
1437 * persisting to move a small or zero window
1438 * (re)transmitting and thereby not persisting
1439 *
1440 * tp->t_timer[TCPT_PERSIST]
1441 * is set when we are in persist state.
1442 * tp->t_force
1443 * is set when we are called to send a persist packet.
1444 * tp->t_timer[TCPT_REXMT]
1445 * is set when we are retransmitting
1446 * The output side is idle when both timers are zero.
1447 *
1448 * If send window is too small, there is data to transmit, and no
1449 * retransmit or persist is pending, then go to persist state.
1450 * If nothing happens soon, send when timer expires:
1451 * if window is nonzero, transmit what we can,
1452 * otherwise force out a byte.
1453 */
1454 if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
1455 tp->t_timer[TCPT_PERSIST] == 0) {
1456 TCP_RESET_REXMT_STATE(tp);
1457 tcp_setpersist(tp);
1458 }
1459 just_return:
1460 /*
1461 * If there is no reason to send a segment, just return.
1462 * but if there is some packets left in the packet list, send them now.
1463 */
1464 while (inp->inp_sndinprog_cnt == 0 &&
1465 tp->t_pktlist_head != NULL) {
1466 packetlist = tp->t_pktlist_head;
1467 packchain_listadd = tp->t_lastchain;
1468 packchain_sent++;
1469 TCP_PKTLIST_CLEAR(tp);
1470
1471 error = tcp_ip_output(so, tp, packetlist,
1472 packchain_listadd,
1473 tp_inp_options, (so_options & SO_DONTROUTE),
1474 (sack_rxmit || (sack_bytes_rxmt != 0)), isipv6);
1475 }
1476 /* tcp was closed while we were in ip; resume close */
1477 if (inp->inp_sndinprog_cnt == 0 &&
1478 (tp->t_flags & TF_CLOSING)) {
1479 tp->t_flags &= ~TF_CLOSING;
1480 (void) tcp_close(tp);
1481 } else {
1482 tcp_check_timer_state(tp);
1483 }
1484 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
1485 return 0;
1486
1487 send:
1488 /*
1489 * Set TF_MAXSEGSNT flag if the segment size is greater than
1490 * the max segment size.
1491 */
1492 if (len > 0) {
1493 if (len >= tp->t_maxseg)
1494 tp->t_flags |= TF_MAXSEGSNT;
1495 else
1496 tp->t_flags &= ~TF_MAXSEGSNT;
1497 }
1498 /*
1499 * Before ESTABLISHED, force sending of initial options
1500 * unless TCP set not to do any options.
1501 * NOTE: we assume that the IP/TCP header plus TCP options
1502 * always fit in a single mbuf, leaving room for a maximum
1503 * link header, i.e.
1504 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1505 */
1506 optlen = 0;
1507 #if INET6
1508 if (isipv6)
1509 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
1510 else
1511 #endif
1512 hdrlen = sizeof (struct tcpiphdr);
1513 if (flags & TH_SYN) {
1514 tp->snd_nxt = tp->iss;
1515 if ((tp->t_flags & TF_NOOPT) == 0) {
1516 u_short mss;
1517
1518 opt[0] = TCPOPT_MAXSEG;
1519 opt[1] = TCPOLEN_MAXSEG;
1520 mss = htons((u_short) tcp_mssopt(tp));
1521 (void)memcpy(opt + 2, &mss, sizeof(mss));
1522 optlen = TCPOLEN_MAXSEG;
1523
1524 if ((tp->t_flags & TF_REQ_SCALE) &&
1525 ((flags & TH_ACK) == 0 ||
1526 (tp->t_flags & TF_RCVD_SCALE))) {
1527 *((u_int32_t *)(void *)(opt + optlen)) = htonl(
1528 TCPOPT_NOP << 24 |
1529 TCPOPT_WINDOW << 16 |
1530 TCPOLEN_WINDOW << 8 |
1531 tp->request_r_scale);
1532 optlen += 4;
1533 }
1534 #if MPTCP
1535 if (mptcp_enable && (so->so_flags & SOF_MP_SUBFLOW)) {
1536 optlen = mptcp_setup_syn_opts(so, opt, optlen);
1537 }
1538 #endif /* MPTCP */
1539 }
1540 }
1541
1542 /*
1543 * Send a timestamp and echo-reply if this is a SYN and our side
1544 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1545 * and our peer have sent timestamps in our SYN's.
1546 */
1547 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1548 (flags & TH_RST) == 0 &&
1549 ((flags & TH_ACK) == 0 ||
1550 (tp->t_flags & TF_RCVD_TSTMP))) {
1551 u_int32_t *lp = (u_int32_t *)(void *)(opt + optlen);
1552
1553 /* Form timestamp option as shown in appendix A of RFC 1323. */
1554 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1555 *lp++ = htonl(tcp_now);
1556 *lp = htonl(tp->ts_recent);
1557 optlen += TCPOLEN_TSTAMP_APPA;
1558 }
1559
1560 /* Note the timestamp for receive buffer autosizing */
1561 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1562 tp->rfbuf_ts = tcp_now;
1563
1564 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1565 /*
1566 * Tack on the SACK permitted option *last*.
1567 * And do padding of options after tacking this on.
1568 * This is because of MSS, TS, WinScale and Signatures are
1569 * all present, we have just 2 bytes left for the SACK
1570 * permitted option, which is just enough.
1571 */
1572 /*
1573 * If this is the first SYN of connection (not a SYN
1574 * ACK), include SACK permitted option. If this is a
1575 * SYN ACK, include SACK permitted option if peer has
1576 * already done so. This is only for active connect,
1577 * since the syncache takes care of the passive connect.
1578 */
1579 if ((flags & TH_SYN) &&
1580 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) {
1581 u_char *bp;
1582 bp = (u_char *)opt + optlen;
1583
1584 *bp++ = TCPOPT_SACK_PERMITTED;
1585 *bp++ = TCPOLEN_SACK_PERMITTED;
1586 optlen += TCPOLEN_SACK_PERMITTED;
1587 }
1588 }
1589 #if MPTCP
1590 if (so->so_flags & SOF_MP_SUBFLOW) {
1591 /*
1592 * Its important to piggyback acks with data as ack only packets
1593 * may get lost and data packets that don't send Data ACKs
1594 * still advance the subflow level ACK and therefore make it
1595 * hard for the remote end to recover in low cwnd situations.
1596 */
1597 if (len != 0) {
1598 tp->t_mpflags |= (TMPF_SEND_DSN |
1599 TMPF_MPTCP_ACKNOW);
1600 } else {
1601 tp->t_mpflags |= TMPF_MPTCP_ACKNOW;
1602 }
1603 optlen = mptcp_setup_opts(tp, off, &opt[0], optlen, flags,
1604 len, &mptcp_acknow);
1605 tp->t_mpflags &= ~TMPF_SEND_DSN;
1606 }
1607 #endif /* MPTCP */
1608
1609 if (tfo_enabled(tp) && !(tp->t_flags & TF_NOOPT) &&
1610 (flags & (TH_SYN | TH_ACK)) == TH_SYN)
1611 optlen += tcp_tfo_write_cookie(tp, optlen, len, opt);
1612
1613 if (tfo_enabled(tp) &&
1614 (flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1615 (tp->t_tfo_flags & TFO_F_OFFER_COOKIE))
1616 optlen += tcp_tfo_write_cookie_rep(tp, optlen, opt);
1617
1618 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1619 /*
1620 * Send SACKs if necessary. This should be the last
1621 * option processed. Only as many SACKs are sent as
1622 * are permitted by the maximum options size.
1623 *
1624 * In general, SACK blocks consume 8*n+2 bytes.
1625 * So a full size SACK blocks option is 34 bytes
1626 * (to generate 4 SACK blocks). At a minimum,
1627 * we need 10 bytes (to generate 1 SACK block).
1628 * If TCP Timestamps (12 bytes) and TCP Signatures
1629 * (18 bytes) are both present, we'll just have
1630 * 10 bytes for SACK options 40 - (12 + 18).
1631 */
1632 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1633 (tp->t_flags & TF_SACK_PERMIT) &&
1634 (tp->rcv_numsacks > 0 || TCP_SEND_DSACK_OPT(tp)) &&
1635 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) {
1636 int nsack, padlen;
1637 u_char *bp = (u_char *)opt + optlen;
1638 u_int32_t *lp;
1639
1640 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK;
1641 nsack = min(nsack, (tp->rcv_numsacks +
1642 (TCP_SEND_DSACK_OPT(tp) ? 1 : 0)));
1643 sackoptlen = (2 + nsack * TCPOLEN_SACK);
1644
1645 /*
1646 * First we need to pad options so that the
1647 * SACK blocks can start at a 4-byte boundary
1648 * (sack option and length are at a 2 byte offset).
1649 */
1650 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4;
1651 optlen += padlen;
1652 while (padlen-- > 0)
1653 *bp++ = TCPOPT_NOP;
1654
1655 tcpstat.tcps_sack_send_blocks++;
1656 *bp++ = TCPOPT_SACK;
1657 *bp++ = sackoptlen;
1658 lp = (u_int32_t *)(void *)bp;
1659
1660 /*
1661 * First block of SACK option should represent
1662 * DSACK. Prefer to send SACK information if there
1663 * is space for only one SACK block. This will
1664 * allow for faster recovery.
1665 */
1666 if (TCP_SEND_DSACK_OPT(tp) && nsack > 0 &&
1667 (tp->rcv_numsacks == 0 || nsack > 1)) {
1668 *lp++ = htonl(tp->t_dsack_lseq);
1669 *lp++ = htonl(tp->t_dsack_rseq);
1670 tcpstat.tcps_dsack_sent++;
1671 tp->t_dsack_sent++;
1672 nsack--;
1673 }
1674 VERIFY(nsack == 0 || tp->rcv_numsacks >= nsack);
1675 for (i = 0; i < nsack; i++) {
1676 struct sackblk sack = tp->sackblks[i];
1677 *lp++ = htonl(sack.start);
1678 *lp++ = htonl(sack.end);
1679 }
1680 optlen += sackoptlen;
1681 }
1682 }
1683
1684 /* Pad TCP options to a 4 byte boundary */
1685 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) {
1686 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t));
1687 u_char *bp = (u_char *)opt + optlen;
1688
1689 optlen += pad;
1690 while (pad) {
1691 *bp++ = TCPOPT_EOL;
1692 pad--;
1693 }
1694 }
1695
1696 /*
1697 * RFC 3168 states that:
1698 * - If you ever sent an ECN-setup SYN/SYN-ACK you must be prepared
1699 * to handle the TCP ECE flag, even if you also later send a
1700 * non-ECN-setup SYN/SYN-ACK.
1701 * - If you ever send a non-ECN-setup SYN/SYN-ACK, you must not set
1702 * the ip ECT flag.
1703 *
1704 * It is not clear how the ECE flag would ever be set if you never
1705 * set the IP ECT flag on outbound packets. All the same, we use
1706 * the TE_SETUPSENT to indicate that we have committed to handling
1707 * the TCP ECE flag correctly. We use the TE_SENDIPECT to indicate
1708 * whether or not we should set the IP ECT flag on outbound packet
1709 *
1710 * For a SYN-ACK, send an ECN setup SYN-ACK
1711 */
1712 if ((flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1713 (tp->ecn_flags & TE_ENABLE_ECN)) {
1714 if (tp->ecn_flags & TE_SETUPRECEIVED) {
1715 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1716 /*
1717 * Setting TH_ECE makes this an ECN-setup
1718 * SYN-ACK
1719 */
1720 flags |= TH_ECE;
1721
1722 /*
1723 * Record that we sent the ECN-setup and
1724 * default to setting IP ECT.
1725 */
1726 tp->ecn_flags |= (TE_SETUPSENT|TE_SENDIPECT);
1727 tcpstat.tcps_ecn_server_setup++;
1728 tcpstat.tcps_ecn_server_success++;
1729 } else {
1730 /*
1731 * We sent an ECN-setup SYN-ACK but it was
1732 * dropped. Fallback to non-ECN-setup
1733 * SYN-ACK and clear flag to indicate that
1734 * we should not send data with IP ECT set
1735 *
1736 * Pretend we didn't receive an
1737 * ECN-setup SYN.
1738 *
1739 * We already incremented the counter
1740 * assuming that the ECN setup will
1741 * succeed. Decrementing here
1742 * tcps_ecn_server_success to correct it.
1743 */
1744 if (tp->ecn_flags & TE_SETUPSENT) {
1745 tcpstat.tcps_ecn_lost_synack++;
1746 tcpstat.tcps_ecn_server_success--;
1747 tp->ecn_flags |= TE_LOST_SYNACK;
1748 }
1749
1750 tp->ecn_flags &=
1751 ~(TE_SETUPRECEIVED | TE_SENDIPECT |
1752 TE_SENDCWR);
1753 }
1754 }
1755 } else if ((flags & (TH_SYN | TH_ACK)) == TH_SYN &&
1756 (tp->ecn_flags & TE_ENABLE_ECN)) {
1757 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1758 /*
1759 * Setting TH_ECE and TH_CWR makes this an
1760 * ECN-setup SYN
1761 */
1762 flags |= (TH_ECE | TH_CWR);
1763 tcpstat.tcps_ecn_client_setup++;
1764 tp->ecn_flags |= TE_CLIENT_SETUP;
1765
1766 /*
1767 * Record that we sent the ECN-setup and default to
1768 * setting IP ECT.
1769 */
1770 tp->ecn_flags |= (TE_SETUPSENT | TE_SENDIPECT);
1771 } else {
1772 /*
1773 * We sent an ECN-setup SYN but it was dropped.
1774 * Fall back to non-ECN and clear flag indicating
1775 * we should send data with IP ECT set.
1776 */
1777 if (tp->ecn_flags & TE_SETUPSENT) {
1778 tcpstat.tcps_ecn_lost_syn++;
1779 tp->ecn_flags |= TE_LOST_SYN;
1780 }
1781 tp->ecn_flags &= ~TE_SENDIPECT;
1782 }
1783 }
1784
1785 /*
1786 * Check if we should set the TCP CWR flag.
1787 * CWR flag is sent when we reduced the congestion window because
1788 * we received a TCP ECE or we performed a fast retransmit. We
1789 * never set the CWR flag on retransmitted packets. We only set
1790 * the CWR flag on data packets. Pure acks don't have this set.
1791 */
1792 if ((tp->ecn_flags & TE_SENDCWR) != 0 && len != 0 &&
1793 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
1794 flags |= TH_CWR;
1795 tp->ecn_flags &= ~TE_SENDCWR;
1796 }
1797
1798 /*
1799 * Check if we should set the TCP ECE flag.
1800 */
1801 if ((tp->ecn_flags & TE_SENDECE) != 0 && len == 0) {
1802 flags |= TH_ECE;
1803 tcpstat.tcps_ecn_sent_ece++;
1804 }
1805
1806
1807 hdrlen += optlen;
1808
1809 /* Reset DSACK sequence numbers */
1810 tp->t_dsack_lseq = 0;
1811 tp->t_dsack_rseq = 0;
1812
1813 #if INET6
1814 if (isipv6)
1815 ipoptlen = ip6_optlen(inp);
1816 else
1817 #endif
1818 {
1819 if (tp_inp_options) {
1820 ipoptlen = tp_inp_options->m_len -
1821 offsetof(struct ipoption, ipopt_list);
1822 } else {
1823 ipoptlen = 0;
1824 }
1825 }
1826 #if IPSEC
1827 ipoptlen += ipsec_optlen;
1828 #endif
1829
1830 /*
1831 * Adjust data length if insertion of options will
1832 * bump the packet length beyond the t_maxopd length.
1833 * Clear the FIN bit because we cut off the tail of
1834 * the segment.
1835 *
1836 * When doing TSO limit a burst to TCP_MAXWIN minus the
1837 * IP, TCP and Options length to keep ip->ip_len from
1838 * overflowing. Prevent the last segment from being
1839 * fractional thus making them all equal sized and set
1840 * the flag to continue sending. TSO is disabled when
1841 * IP options or IPSEC are present.
1842 */
1843 if (len + optlen + ipoptlen > tp->t_maxopd) {
1844 /*
1845 * If there is still more to send,
1846 * don't close the connection.
1847 */
1848 flags &= ~TH_FIN;
1849 if (tso) {
1850 int32_t tso_maxlen;
1851
1852 tso_maxlen = tp->tso_max_segment_size ?
1853 tp->tso_max_segment_size : TCP_MAXWIN;
1854
1855 if (len > tso_maxlen - hdrlen - optlen) {
1856 len = tso_maxlen - hdrlen - optlen;
1857 len = len - (len % (tp->t_maxopd - optlen));
1858 sendalot = 1;
1859 } else if (tp->t_flags & TF_NEEDFIN) {
1860 sendalot = 1;
1861 }
1862 } else {
1863 len = tp->t_maxopd - optlen - ipoptlen;
1864 sendalot = 1;
1865 }
1866 }
1867
1868 if (max_linkhdr + hdrlen > MCLBYTES)
1869 panic("tcphdr too big");
1870
1871 /* Check if there is enough data in the send socket
1872 * buffer to start measuring bandwidth
1873 */
1874 if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 &&
1875 (tp->t_bwmeas != NULL) &&
1876 (tp->t_flagsext & TF_BWMEAS_INPROGRESS) == 0) {
1877 tp->t_bwmeas->bw_size = min(min(
1878 (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)),
1879 tp->snd_cwnd), tp->snd_wnd);
1880 if (tp->t_bwmeas->bw_minsize > 0 &&
1881 tp->t_bwmeas->bw_size < tp->t_bwmeas->bw_minsize)
1882 tp->t_bwmeas->bw_size = 0;
1883 if (tp->t_bwmeas->bw_maxsize > 0)
1884 tp->t_bwmeas->bw_size = min(tp->t_bwmeas->bw_size,
1885 tp->t_bwmeas->bw_maxsize);
1886 if (tp->t_bwmeas->bw_size > 0) {
1887 tp->t_flagsext |= TF_BWMEAS_INPROGRESS;
1888 tp->t_bwmeas->bw_start = tp->snd_max;
1889 tp->t_bwmeas->bw_ts = tcp_now;
1890 }
1891 }
1892
1893 VERIFY(inp->inp_flowhash != 0);
1894 /*
1895 * Grab a header mbuf, attaching a copy of data to
1896 * be transmitted, and initialize the header from
1897 * the template for sends on this connection.
1898 */
1899 if (len) {
1900 tp->t_pmtud_lastseg_size = len + optlen + ipoptlen;
1901 if ((tp->t_flagsext & TF_FORCE) && len == 1)
1902 tcpstat.tcps_sndprobe++;
1903 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1904 tcpstat.tcps_sndrexmitpack++;
1905 tcpstat.tcps_sndrexmitbyte += len;
1906 if (nstat_collect) {
1907 nstat_route_tx(inp->inp_route.ro_rt, 1,
1908 len, NSTAT_TX_FLAG_RETRANSMIT);
1909 INP_ADD_STAT(inp, cell, wifi, wired,
1910 txpackets, 1);
1911 INP_ADD_STAT(inp, cell, wifi, wired,
1912 txbytes, len);
1913 tp->t_stat.txretransmitbytes += len;
1914 tp->t_stat.rxmitpkts++;
1915 }
1916 } else {
1917 tcpstat.tcps_sndpack++;
1918 tcpstat.tcps_sndbyte += len;
1919
1920 if (nstat_collect) {
1921 INP_ADD_STAT(inp, cell, wifi, wired,
1922 txpackets, 1);
1923 INP_ADD_STAT(inp, cell, wifi, wired,
1924 txbytes, len);
1925 }
1926 inp_decr_sndbytes_unsent(so, len);
1927 }
1928 inp_set_activity_bitmap(inp);
1929 #if MPTCP
1930 if (tp->t_mpflags & TMPF_MPTCP_TRUE) {
1931 tcpstat.tcps_mp_sndpacks++;
1932 tcpstat.tcps_mp_sndbytes += len;
1933 }
1934 #endif /* MPTCP */
1935 /*
1936 * try to use the new interface that allocates all
1937 * the necessary mbuf hdrs under 1 mbuf lock and
1938 * avoids rescanning the socket mbuf list if
1939 * certain conditions are met. This routine can't
1940 * be used in the following cases...
1941 * 1) the protocol headers exceed the capacity of
1942 * of a single mbuf header's data area (no cluster attached)
1943 * 2) the length of the data being transmitted plus
1944 * the protocol headers fits into a single mbuf header's
1945 * data area (no cluster attached)
1946 */
1947 m = NULL;
1948
1949 /* minimum length we are going to allocate */
1950 allocated_len = MHLEN;
1951 if (MHLEN < hdrlen + max_linkhdr) {
1952 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1953 if (m == NULL) {
1954 error = ENOBUFS;
1955 goto out;
1956 }
1957 MCLGET(m, M_DONTWAIT);
1958 if ((m->m_flags & M_EXT) == 0) {
1959 m_freem(m);
1960 error = ENOBUFS;
1961 goto out;
1962 }
1963 m->m_data += max_linkhdr;
1964 m->m_len = hdrlen;
1965 allocated_len = MCLBYTES;
1966 }
1967 if (len <= allocated_len - hdrlen - max_linkhdr) {
1968 if (m == NULL) {
1969 VERIFY(allocated_len <= MHLEN);
1970 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1971 if (m == NULL) {
1972 error = ENOBUFS;
1973 goto out;
1974 }
1975 m->m_data += max_linkhdr;
1976 m->m_len = hdrlen;
1977 }
1978 /* makes sure we still have data left to be sent at this point */
1979 if (so->so_snd.sb_mb == NULL || off < 0) {
1980 if (m != NULL) m_freem(m);
1981 error = 0; /* should we return an error? */
1982 goto out;
1983 }
1984 m_copydata(so->so_snd.sb_mb, off, (int) len,
1985 mtod(m, caddr_t) + hdrlen);
1986 m->m_len += len;
1987 } else {
1988 uint32_t copymode;
1989 /*
1990 * Retain packet header metadata at the socket
1991 * buffer if this is is an MPTCP subflow,
1992 * otherwise move it.
1993 */
1994 copymode = M_COPYM_MOVE_HDR;
1995 #if MPTCP
1996 if (so->so_flags & SOF_MP_SUBFLOW) {
1997 copymode = M_COPYM_NOOP_HDR;
1998 }
1999 #endif /* MPTCP */
2000 if (m != NULL) {
2001 m->m_next = m_copym_mode(so->so_snd.sb_mb,
2002 off, (int)len, M_DONTWAIT, copymode);
2003 if (m->m_next == NULL) {
2004 (void) m_free(m);
2005 error = ENOBUFS;
2006 goto out;
2007 }
2008 } else {
2009 /*
2010 * make sure we still have data left
2011 * to be sent at this point
2012 */
2013 if (so->so_snd.sb_mb == NULL) {
2014 error = 0; /* should we return an error? */
2015 goto out;
2016 }
2017
2018 /*
2019 * m_copym_with_hdrs will always return the
2020 * last mbuf pointer and the offset into it that
2021 * it acted on to fullfill the current request,
2022 * whether a valid 'hint' was passed in or not.
2023 */
2024 if ((m = m_copym_with_hdrs(so->so_snd.sb_mb,
2025 off, len, M_DONTWAIT, NULL, NULL,
2026 copymode)) == NULL) {
2027 error = ENOBUFS;
2028 goto out;
2029 }
2030 m->m_data += max_linkhdr;
2031 m->m_len = hdrlen;
2032 }
2033 }
2034 /*
2035 * If we're sending everything we've got, set PUSH.
2036 * (This will keep happy those implementations which only
2037 * give data to the user when a buffer fills or
2038 * a PUSH comes in.)
2039 *
2040 * On SYN-segments we should not add the PUSH-flag.
2041 */
2042 if (off + len == so->so_snd.sb_cc && !(flags & TH_SYN))
2043 flags |= TH_PUSH;
2044 } else {
2045 if (tp->t_flags & TF_ACKNOW)
2046 tcpstat.tcps_sndacks++;
2047 else if (flags & (TH_SYN|TH_FIN|TH_RST))
2048 tcpstat.tcps_sndctrl++;
2049 else if (SEQ_GT(tp->snd_up, tp->snd_una))
2050 tcpstat.tcps_sndurg++;
2051 else
2052 tcpstat.tcps_sndwinup++;
2053
2054 MGETHDR(m, M_DONTWAIT, MT_HEADER); /* MAC-OK */
2055 if (m == NULL) {
2056 error = ENOBUFS;
2057 goto out;
2058 }
2059 if (MHLEN < (hdrlen + max_linkhdr)) {
2060 MCLGET(m, M_DONTWAIT);
2061 if ((m->m_flags & M_EXT) == 0) {
2062 m_freem(m);
2063 error = ENOBUFS;
2064 goto out;
2065 }
2066 }
2067 m->m_data += max_linkhdr;
2068 m->m_len = hdrlen;
2069 }
2070 m->m_pkthdr.rcvif = 0;
2071 #if CONFIG_MACF_NET
2072 mac_mbuf_label_associate_inpcb(inp, m);
2073 #endif
2074 #if INET6
2075 if (isipv6) {
2076 ip6 = mtod(m, struct ip6_hdr *);
2077 th = (struct tcphdr *)(void *)(ip6 + 1);
2078 tcp_fillheaders(tp, ip6, th);
2079 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2080 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
2081 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
2082 }
2083 svc_flags |= PKT_SCF_IPV6;
2084 #if PF_ECN
2085 m_pftag(m)->pftag_hdr = (void *)ip6;
2086 m_pftag(m)->pftag_flags |= PF_TAG_HDR_INET6;
2087 #endif /* PF_ECN */
2088 } else
2089 #endif /* INET6 */
2090 {
2091 ip = mtod(m, struct ip *);
2092 ipov = (struct ipovly *)ip;
2093 th = (struct tcphdr *)(void *)(ip + 1);
2094 /* this picks up the pseudo header (w/o the length) */
2095 tcp_fillheaders(tp, ip, th);
2096 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2097 !SEQ_LT(tp->snd_nxt, tp->snd_max) &&
2098 !sack_rxmit && !(flags & TH_SYN)) {
2099 ip->ip_tos |= IPTOS_ECN_ECT0;
2100 }
2101 #if PF_ECN
2102 m_pftag(m)->pftag_hdr = (void *)ip;
2103 m_pftag(m)->pftag_flags |= PF_TAG_HDR_INET;
2104 #endif /* PF_ECN */
2105 }
2106
2107 /*
2108 * Fill in fields, remembering maximum advertised
2109 * window for use in delaying messages about window sizes.
2110 * If resending a FIN, be sure not to use a new sequence number.
2111 */
2112 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
2113 tp->snd_nxt == tp->snd_max)
2114 tp->snd_nxt--;
2115 /*
2116 * If we are doing retransmissions, then snd_nxt will
2117 * not reflect the first unsent octet. For ACK only
2118 * packets, we do not want the sequence number of the
2119 * retransmitted packet, we want the sequence number
2120 * of the next unsent octet. So, if there is no data
2121 * (and no SYN or FIN), use snd_max instead of snd_nxt
2122 * when filling in ti_seq. But if we are in persist
2123 * state, snd_max might reflect one byte beyond the
2124 * right edge of the window, so use snd_nxt in that
2125 * case, since we know we aren't doing a retransmission.
2126 * (retransmit and persist are mutually exclusive...)
2127 *
2128 * Note the state of this retransmit segment to detect spurious
2129 * retransmissions.
2130 */
2131 if (sack_rxmit == 0) {
2132 if (len || (flags & (TH_SYN|TH_FIN)) ||
2133 tp->t_timer[TCPT_PERSIST]) {
2134 th->th_seq = htonl(tp->snd_nxt);
2135 if (len > 0) {
2136 m->m_pkthdr.tx_start_seq = tp->snd_nxt;
2137 m->m_pkthdr.pkt_flags |= PKTF_START_SEQ;
2138 }
2139 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
2140 if (SACK_ENABLED(tp) && len > 1) {
2141 tcp_rxtseg_insert(tp, tp->snd_nxt,
2142 (tp->snd_nxt + len - 1));
2143 }
2144 if (len > 0)
2145 m->m_pkthdr.pkt_flags |=
2146 PKTF_TCP_REXMT;
2147 }
2148 } else {
2149 th->th_seq = htonl(tp->snd_max);
2150 }
2151 } else {
2152 th->th_seq = htonl(p->rxmit);
2153 if (len > 0) {
2154 m->m_pkthdr.pkt_flags |=
2155 (PKTF_TCP_REXMT | PKTF_START_SEQ);
2156 m->m_pkthdr.tx_start_seq = p->rxmit;
2157 }
2158 tcp_rxtseg_insert(tp, p->rxmit, (p->rxmit + len - 1));
2159 p->rxmit += len;
2160 tp->sackhint.sack_bytes_rexmit += len;
2161 }
2162 th->th_ack = htonl(tp->rcv_nxt);
2163 tp->last_ack_sent = tp->rcv_nxt;
2164 if (optlen) {
2165 bcopy(opt, th + 1, optlen);
2166 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
2167 }
2168 th->th_flags = flags;
2169 th->th_win = htons((u_short) (recwin>>tp->rcv_scale));
2170 if (!(so->so_flags & SOF_MP_SUBFLOW)) {
2171 if (recwin > 0 && SEQ_LT(tp->rcv_adv, tp->rcv_nxt + recwin)) {
2172 tp->rcv_adv = tp->rcv_nxt + recwin;
2173 }
2174 } else {
2175 struct mptcb *mp_tp = tptomptp(tp);
2176 if (recwin > 0) {
2177 tp->rcv_adv = tp->rcv_nxt + recwin;
2178 }
2179
2180 if (recwin > 0 && SEQ_LT(mp_tp->mpt_rcvadv, (uint32_t)mp_tp->mpt_rcvnxt + recwin)) {
2181 mp_tp->mpt_rcvadv = (uint32_t)mp_tp->mpt_rcvnxt + recwin;
2182 }
2183 }
2184
2185 /*
2186 * Adjust the RXWIN0SENT flag - indicate that we have advertised
2187 * a 0 window. This may cause the remote transmitter to stall. This
2188 * flag tells soreceive() to disable delayed acknowledgements when
2189 * draining the buffer. This can occur if the receiver is attempting
2190 * to read more data then can be buffered prior to transmitting on
2191 * the connection.
2192 */
2193 if (th->th_win == 0)
2194 tp->t_flags |= TF_RXWIN0SENT;
2195 else
2196 tp->t_flags &= ~TF_RXWIN0SENT;
2197
2198 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
2199 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
2200 th->th_flags |= TH_URG;
2201 } else {
2202 /*
2203 * If no urgent pointer to send, then we pull
2204 * the urgent pointer to the left edge of the send window
2205 * so that it doesn't drift into the send window on sequence
2206 * number wraparound.
2207 */
2208 tp->snd_up = tp->snd_una; /* drag it along */
2209 }
2210
2211 /*
2212 * Put TCP length in extended header, and then
2213 * checksum extended header and data.
2214 */
2215 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
2216
2217 /*
2218 * If this is potentially the last packet on the stream, then mark
2219 * it in order to enable some optimizations in the underlying
2220 * layers
2221 */
2222 if (tp->t_state != TCPS_ESTABLISHED &&
2223 (tp->t_state == TCPS_CLOSING || tp->t_state == TCPS_TIME_WAIT
2224 || tp->t_state == TCPS_LAST_ACK || (th->th_flags & TH_RST)))
2225 m->m_pkthdr.pkt_flags |= PKTF_LAST_PKT;
2226
2227 #if INET6
2228 if (isipv6) {
2229 /*
2230 * ip6_plen is not need to be filled now, and will be filled
2231 * in ip6_output.
2232 */
2233 m->m_pkthdr.csum_flags = CSUM_TCPIPV6;
2234 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2235 if (len + optlen)
2236 th->th_sum = in_addword(th->th_sum,
2237 htons((u_short)(optlen + len)));
2238 }
2239 else
2240 #endif /* INET6 */
2241 {
2242 m->m_pkthdr.csum_flags = CSUM_TCP;
2243 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2244 if (len + optlen)
2245 th->th_sum = in_addword(th->th_sum,
2246 htons((u_short)(optlen + len)));
2247 }
2248
2249 /*
2250 * Enable TSO and specify the size of the segments.
2251 * The TCP pseudo header checksum is always provided.
2252 */
2253 if (tso) {
2254 #if INET6
2255 if (isipv6)
2256 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV6;
2257 else
2258 #endif /* INET6 */
2259 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV4;
2260
2261 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
2262 } else {
2263 m->m_pkthdr.tso_segsz = 0;
2264 }
2265
2266 /*
2267 * In transmit state, time the transmission and arrange for
2268 * the retransmit. In persist state, just set snd_max.
2269 */
2270 if (!(tp->t_flagsext & TF_FORCE)
2271 || tp->t_timer[TCPT_PERSIST] == 0) {
2272 tcp_seq startseq = tp->snd_nxt;
2273
2274 /*
2275 * Advance snd_nxt over sequence space of this segment.
2276 */
2277 if (flags & (TH_SYN|TH_FIN)) {
2278 if (flags & TH_SYN)
2279 tp->snd_nxt++;
2280 if ((flags & TH_FIN) &&
2281 !(tp->t_flags & TF_SENTFIN)) {
2282 tp->snd_nxt++;
2283 tp->t_flags |= TF_SENTFIN;
2284 }
2285 }
2286 if (sack_rxmit)
2287 goto timer;
2288 if (sack_rescue_rxt == TRUE) {
2289 tp->snd_nxt = old_snd_nxt;
2290 sack_rescue_rxt = FALSE;
2291 tcpstat.tcps_pto_in_recovery++;
2292 } else {
2293 tp->snd_nxt += len;
2294 }
2295 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
2296 tp->snd_max = tp->snd_nxt;
2297 tp->t_sndtime = tcp_now;
2298 /*
2299 * Time this transmission if not a retransmission and
2300 * not currently timing anything.
2301 */
2302 if (tp->t_rtttime == 0) {
2303 tp->t_rtttime = tcp_now;
2304 tp->t_rtseq = startseq;
2305 tcpstat.tcps_segstimed++;
2306
2307 /* update variables related to pipe ack */
2308 tp->t_pipeack_lastuna = tp->snd_una;
2309 }
2310 }
2311
2312 /*
2313 * Set retransmit timer if not currently set,
2314 * and not doing an ack or a keep-alive probe.
2315 */
2316 timer:
2317 if (tp->t_timer[TCPT_REXMT] == 0 &&
2318 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
2319 tp->snd_nxt != tp->snd_una || (flags & TH_FIN))) {
2320 if (tp->t_timer[TCPT_PERSIST]) {
2321 tp->t_timer[TCPT_PERSIST] = 0;
2322 tp->t_persist_stop = 0;
2323 TCP_RESET_REXMT_STATE(tp);
2324 }
2325 tp->t_timer[TCPT_REXMT] =
2326 OFFSET_FROM_START(tp, tp->t_rxtcur);
2327 }
2328
2329 /*
2330 * Set tail loss probe timeout if new data is being
2331 * transmitted. This will be supported only when
2332 * SACK option is enabled on a connection.
2333 *
2334 * Every time new data is sent PTO will get reset.
2335 */
2336 if (tcp_enable_tlp && len != 0 && tp->t_state == TCPS_ESTABLISHED &&
2337 SACK_ENABLED(tp) && !IN_FASTRECOVERY(tp) &&
2338 tp->snd_nxt == tp->snd_max &&
2339 SEQ_GT(tp->snd_nxt, tp->snd_una) &&
2340 tp->t_rxtshift == 0 &&
2341 (tp->t_flagsext & (TF_SENT_TLPROBE|TF_PKTS_REORDERED)) == 0) {
2342 u_int32_t pto, srtt;
2343
2344 /*
2345 * Using SRTT alone to set PTO can cause spurious
2346 * retransmissions on wireless networks where there
2347 * is a lot of variance in RTT. Taking variance
2348 * into account will avoid this.
2349 */
2350 srtt = tp->t_srtt >> TCP_RTT_SHIFT;
2351 pto = ((TCP_REXMTVAL(tp)) * 3) >> 1;
2352 pto = max (2 * srtt, pto);
2353 if ((tp->snd_max - tp->snd_una) == tp->t_maxseg)
2354 pto = max(pto,
2355 (((3 * pto) >> 2) + tcp_delack * 2));
2356 else
2357 pto = max(10, pto);
2358
2359 /* if RTO is less than PTO, choose RTO instead */
2360 if (tp->t_rxtcur < pto)
2361 pto = tp->t_rxtcur;
2362
2363 tp->t_timer[TCPT_PTO] = OFFSET_FROM_START(tp, pto);
2364 }
2365 } else {
2366 /*
2367 * Persist case, update snd_max but since we are in
2368 * persist mode (no window) we do not update snd_nxt.
2369 */
2370 int xlen = len;
2371 if (flags & TH_SYN)
2372 ++xlen;
2373 if ((flags & TH_FIN) &&
2374 !(tp->t_flags & TF_SENTFIN)) {
2375 ++xlen;
2376 tp->t_flags |= TF_SENTFIN;
2377 }
2378 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) {
2379 tp->snd_max = tp->snd_nxt + len;
2380 tp->t_sndtime = tcp_now;
2381 }
2382 }
2383
2384 #if TCPDEBUG
2385 /*
2386 * Trace.
2387 */
2388 if (so_options & SO_DEBUG)
2389 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
2390 #endif
2391
2392 /*
2393 * Fill in IP length and desired time to live and
2394 * send to IP level. There should be a better way
2395 * to handle ttl and tos; we could keep them in
2396 * the template, but need a way to checksum without them.
2397 */
2398 #if INET6
2399 /*
2400 * m->m_pkthdr.len should have been set before cksum calcuration,
2401 * because in6_cksum() need it.
2402 */
2403 if (isipv6) {
2404 /*
2405 * we separately set hoplimit for every segment, since the
2406 * user might want to change the value via setsockopt.
2407 * Also, desired default hop limit might be changed via
2408 * Neighbor Discovery.
2409 */
2410 ip6->ip6_hlim = in6_selecthlim(inp, inp->in6p_route.ro_rt ?
2411 inp->in6p_route.ro_rt->rt_ifp : NULL);
2412
2413 /* TODO: IPv6 IP6TOS_ECT bit on */
2414 KERNEL_DEBUG(DBG_LAYER_BEG,
2415 ((inp->inp_fport << 16) | inp->inp_lport),
2416 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
2417 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
2418 sendalot,0,0);
2419 } else
2420 #endif /* INET6 */
2421 {
2422 ip->ip_len = m->m_pkthdr.len;
2423 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
2424 ip->ip_tos |= (inp->inp_ip_tos & ~IPTOS_ECN_MASK);/* XXX */
2425 KERNEL_DEBUG(DBG_LAYER_BEG,
2426 ((inp->inp_fport << 16) | inp->inp_lport),
2427 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
2428 (inp->inp_faddr.s_addr & 0xffff)), 0,0,0);
2429 }
2430
2431 /*
2432 * See if we should do MTU discovery.
2433 * Look at the flag updated on the following criterias:
2434 * 1) Path MTU discovery is authorized by the sysctl
2435 * 2) The route isn't set yet (unlikely but could happen)
2436 * 3) The route is up
2437 * 4) the MTU is not locked (if it is, then discovery has been
2438 * disabled for that route)
2439 */
2440 #if INET6
2441 if (!isipv6)
2442 #endif /* INET6 */
2443 if (path_mtu_discovery && (tp->t_flags & TF_PMTUD))
2444 ip->ip_off |= IP_DF;
2445
2446 #if NECP
2447 {
2448 necp_kernel_policy_id policy_id;
2449 necp_kernel_policy_id skip_policy_id;
2450 u_int32_t route_rule_id;
2451 if (!necp_socket_is_allowed_to_send_recv(inp, NULL, &policy_id, &route_rule_id, &skip_policy_id)) {
2452 TCP_LOG_DROP_NECP(isipv6 ? (void *)ip6 : (void *)ip, th, tp, true);
2453 m_freem(m);
2454 error = EHOSTUNREACH;
2455 goto out;
2456 }
2457 necp_mark_packet_from_socket(m, inp, policy_id, route_rule_id, skip_policy_id);
2458
2459 if (net_qos_policy_restricted != 0) {
2460 necp_socket_update_qos_marking(inp, inp->inp_route.ro_rt,
2461 NULL, route_rule_id);
2462 }
2463 }
2464 #endif /* NECP */
2465
2466 #if IPSEC
2467 if (inp->inp_sp != NULL)
2468 ipsec_setsocket(m, so);
2469 #endif /*IPSEC*/
2470
2471 /*
2472 * The socket is kept locked while sending out packets in ip_output, even if packet chaining is not active.
2473 */
2474 lost = 0;
2475
2476 /*
2477 * Embed the flow hash in pkt hdr and mark the packet as
2478 * capable of flow controlling
2479 */
2480 m->m_pkthdr.pkt_flowsrc = FLOWSRC_INPCB;
2481 m->m_pkthdr.pkt_flowid = inp->inp_flowhash;
2482 m->m_pkthdr.pkt_flags |= (PKTF_FLOW_ID | PKTF_FLOW_LOCALSRC | PKTF_FLOW_ADV);
2483 m->m_pkthdr.pkt_proto = IPPROTO_TCP;
2484 m->m_pkthdr.tx_tcp_pid = so->last_pid;
2485 if (so->so_flags & SOF_DELEGATED)
2486 m->m_pkthdr.tx_tcp_e_pid = so->e_pid;
2487 else
2488 m->m_pkthdr.tx_tcp_e_pid = 0;
2489
2490 m->m_nextpkt = NULL;
2491
2492 if (inp->inp_last_outifp != NULL &&
2493 !(inp->inp_last_outifp->if_flags & IFF_LOOPBACK)) {
2494 /* Hint to prioritize this packet if
2495 * 1. if the packet has no data
2496 * 2. the interface supports transmit-start model and did
2497 * not disable ACK prioritization.
2498 * 3. Only ACK flag is set.
2499 * 4. there is no outstanding data on this connection.
2500 */
2501 if (tcp_prioritize_acks != 0 && len == 0 &&
2502 (inp->inp_last_outifp->if_eflags &
2503 (IFEF_TXSTART | IFEF_NOACKPRI)) == IFEF_TXSTART) {
2504 if (th->th_flags == TH_ACK &&
2505 tp->snd_una == tp->snd_max &&
2506 tp->t_timer[TCPT_REXMT] == 0)
2507 svc_flags |= PKT_SCF_TCP_ACK;
2508 if (th->th_flags & TH_SYN)
2509 svc_flags |= PKT_SCF_TCP_SYN;
2510 }
2511 set_packet_service_class(m, so, sotc, svc_flags);
2512 } else {
2513 /*
2514 * Optimization for loopback just set the mbuf
2515 * service class
2516 */
2517 (void) m_set_service_class(m, so_tc2msc(sotc));
2518 }
2519
2520 TCP_LOG_TH_FLAGS(isipv6 ? (void *)ip6 : (void *)ip, th, tp, true,
2521 inp->inp_last_outifp != NULL ? inp->inp_last_outifp :
2522 inp->inp_boundifp);
2523
2524 tp->t_pktlist_sentlen += len;
2525 tp->t_lastchain++;
2526
2527 #if INET6
2528 if (isipv6) {
2529 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2530 struct ip6 *, ip6, struct tcpcb *, tp, struct tcphdr *,
2531 th);
2532 } else
2533 #endif /* INET6 */
2534 {
2535 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2536 struct ip *, ip, struct tcpcb *, tp, struct tcphdr *, th);
2537 }
2538
2539 if (tp->t_pktlist_head != NULL) {
2540 tp->t_pktlist_tail->m_nextpkt = m;
2541 tp->t_pktlist_tail = m;
2542 } else {
2543 packchain_newlist++;
2544 tp->t_pktlist_head = tp->t_pktlist_tail = m;
2545 }
2546
2547 if (lro_ackmore && !sackoptlen && tp->t_timer[TCPT_PERSIST] == 0 &&
2548 (th->th_flags & TH_ACK) == TH_ACK && len == 0 &&
2549 tp->t_state == TCPS_ESTABLISHED) {
2550 /* For a pure ACK, see if you need to send more of them */
2551 mnext = tcp_send_lroacks(tp, m, th);
2552 if (mnext) {
2553 tp->t_pktlist_tail->m_nextpkt = mnext;
2554 if (mnext->m_nextpkt == NULL) {
2555 tp->t_pktlist_tail = mnext;
2556 tp->t_lastchain++;
2557 } else {
2558 struct mbuf *tail, *next;
2559 next = mnext->m_nextpkt;
2560 tail = next->m_nextpkt;
2561 while (tail) {
2562 next = tail;
2563 tail = tail->m_nextpkt;
2564 tp->t_lastchain++;
2565 }
2566 tp->t_pktlist_tail = next;
2567 }
2568 }
2569 }
2570
2571 if (sendalot == 0 || (tp->t_state != TCPS_ESTABLISHED) ||
2572 (tp->snd_cwnd <= (tp->snd_wnd / 8)) ||
2573 (tp->t_flags & TF_ACKNOW) ||
2574 (tp->t_flagsext & TF_FORCE) ||
2575 tp->t_lastchain >= tcp_packet_chaining) {
2576 error = 0;
2577 while (inp->inp_sndinprog_cnt == 0 &&
2578 tp->t_pktlist_head != NULL) {
2579 packetlist = tp->t_pktlist_head;
2580 packchain_listadd = tp->t_lastchain;
2581 packchain_sent++;
2582 lost = tp->t_pktlist_sentlen;
2583 TCP_PKTLIST_CLEAR(tp);
2584
2585 error = tcp_ip_output(so, tp, packetlist,
2586 packchain_listadd, tp_inp_options,
2587 (so_options & SO_DONTROUTE),
2588 (sack_rxmit || (sack_bytes_rxmt != 0)), isipv6);
2589 if (error) {
2590 /*
2591 * Take into account the rest of unsent
2592 * packets in the packet list for this tcp
2593 * into "lost", since we're about to free
2594 * the whole list below.
2595 */
2596 lost += tp->t_pktlist_sentlen;
2597 break;
2598 } else {
2599 lost = 0;
2600 }
2601 }
2602 /* tcp was closed while we were in ip; resume close */
2603 if (inp->inp_sndinprog_cnt == 0 &&
2604 (tp->t_flags & TF_CLOSING)) {
2605 tp->t_flags &= ~TF_CLOSING;
2606 (void) tcp_close(tp);
2607 return 0;
2608 }
2609 } else {
2610 error = 0;
2611 packchain_looped++;
2612 tcpstat.tcps_sndtotal++;
2613
2614 goto again;
2615 }
2616 if (error) {
2617 /*
2618 * Assume that the packets were lost, so back out the
2619 * sequence number advance, if any. Note that the "lost"
2620 * variable represents the amount of user data sent during
2621 * the recent call to ip_output_list() plus the amount of
2622 * user data in the packet list for this tcp at the moment.
2623 */
2624 if (!(tp->t_flagsext & TF_FORCE)
2625 || tp->t_timer[TCPT_PERSIST] == 0) {
2626 /*
2627 * No need to check for TH_FIN here because
2628 * the TF_SENTFIN flag handles that case.
2629 */
2630 if ((flags & TH_SYN) == 0) {
2631 if (sack_rxmit) {
2632 if (SEQ_GT((p->rxmit - lost),
2633 tp->snd_una)) {
2634 p->rxmit -= lost;
2635 } else {
2636 lost = p->rxmit - tp->snd_una;
2637 p->rxmit = tp->snd_una;
2638 }
2639 tp->sackhint.sack_bytes_rexmit -= lost;
2640 } else {
2641 if (SEQ_GT((tp->snd_nxt - lost),
2642 tp->snd_una))
2643 tp->snd_nxt -= lost;
2644 else
2645 tp->snd_nxt = tp->snd_una;
2646 }
2647 }
2648 }
2649 out:
2650 if (tp->t_pktlist_head != NULL)
2651 m_freem_list(tp->t_pktlist_head);
2652 TCP_PKTLIST_CLEAR(tp);
2653
2654 if (error == ENOBUFS) {
2655 /*
2656 * Set retransmit timer if not currently set
2657 * when we failed to send a segment that can be
2658 * retransmitted (i.e. not pure ack or rst)
2659 */
2660 if (tp->t_timer[TCPT_REXMT] == 0 &&
2661 tp->t_timer[TCPT_PERSIST] == 0 &&
2662 (len != 0 || (flags & (TH_SYN | TH_FIN)) != 0 ||
2663 so->so_snd.sb_cc > 0))
2664 tp->t_timer[TCPT_REXMT] =
2665 OFFSET_FROM_START(tp, tp->t_rxtcur);
2666 tp->snd_cwnd = tp->t_maxseg;
2667 tp->t_bytes_acked = 0;
2668 tcp_check_timer_state(tp);
2669 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2670
2671 tcp_ccdbg_trace(tp, NULL, TCP_CC_OUTPUT_ERROR);
2672 return 0;
2673 }
2674 if (error == EMSGSIZE) {
2675 /*
2676 * ip_output() will have already fixed the route
2677 * for us. tcp_mtudisc() will, as its last action,
2678 * initiate retransmission, so it is important to
2679 * not do so here.
2680 *
2681 * If TSO was active we either got an interface
2682 * without TSO capabilits or TSO was turned off.
2683 * Disable it for this connection as too and
2684 * immediatly retry with MSS sized segments generated
2685 * by this function.
2686 */
2687 if (tso)
2688 tp->t_flags &= ~TF_TSO;
2689
2690 tcp_mtudisc(inp, 0);
2691 tcp_check_timer_state(tp);
2692
2693 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2694 return 0;
2695 }
2696 /*
2697 * Unless this is due to interface restriction policy,
2698 * treat EHOSTUNREACH/ENETDOWN as a soft error.
2699 */
2700 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
2701 TCPS_HAVERCVDSYN(tp->t_state) &&
2702 !inp_restricted_send(inp, inp->inp_last_outifp)) {
2703 tp->t_softerror = error;
2704 error = 0;
2705 }
2706 tcp_check_timer_state(tp);
2707 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2708 return error;
2709 }
2710
2711 tcpstat.tcps_sndtotal++;
2712
2713 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,0,0,0,0,0);
2714 if (sendalot)
2715 goto again;
2716
2717 tcp_check_timer_state(tp);
2718
2719 return 0;
2720 }
2721
2722 static int
2723 tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt,
2724 int cnt, struct mbuf *opt, int flags, int sack_in_progress, boolean_t isipv6)
2725 {
2726 int error = 0;
2727 boolean_t chain;
2728 boolean_t unlocked = FALSE;
2729 boolean_t ifdenied = FALSE;
2730 struct inpcb *inp = tp->t_inpcb;
2731 struct ip_out_args ipoa;
2732 struct route ro;
2733 struct ifnet *outif = NULL;
2734
2735 bzero(&ipoa, sizeof(ipoa));
2736 ipoa.ipoa_boundif = IFSCOPE_NONE;
2737 ipoa.ipoa_flags = IPOAF_SELECT_SRCIF | IPOAF_BOUND_SRCADDR;
2738 ipoa.ipoa_sotc = SO_TC_UNSPEC;
2739 ipoa.ipoa_netsvctype = _NET_SERVICE_TYPE_UNSPEC;
2740 #if INET6
2741 struct ip6_out_args ip6oa;
2742 struct route_in6 ro6;
2743
2744 bzero(&ip6oa, sizeof(ip6oa));
2745 ip6oa.ip6oa_boundif = IFSCOPE_NONE;
2746 ip6oa.ip6oa_flags = IP6OAF_SELECT_SRCIF | IP6OAF_BOUND_SRCADDR;
2747 ip6oa.ip6oa_sotc = SO_TC_UNSPEC;
2748 ip6oa.ip6oa_netsvctype = _NET_SERVICE_TYPE_UNSPEC;
2749
2750 struct flowadv *adv =
2751 (isipv6 ? &ip6oa.ip6oa_flowadv : &ipoa.ipoa_flowadv);
2752 #else /* INET6 */
2753 struct flowadv *adv = &ipoa.ipoa_flowadv;
2754 #endif /* !INET6 */
2755
2756 /* If socket was bound to an ifindex, tell ip_output about it */
2757 if (inp->inp_flags & INP_BOUND_IF) {
2758 #if INET6
2759 if (isipv6) {
2760 ip6oa.ip6oa_boundif = inp->inp_boundifp->if_index;
2761 ip6oa.ip6oa_flags |= IP6OAF_BOUND_IF;
2762 } else
2763 #endif /* INET6 */
2764 {
2765 ipoa.ipoa_boundif = inp->inp_boundifp->if_index;
2766 ipoa.ipoa_flags |= IPOAF_BOUND_IF;
2767 }
2768 }
2769
2770 if (INP_NO_CELLULAR(inp)) {
2771 #if INET6
2772 if (isipv6)
2773 ip6oa.ip6oa_flags |= IP6OAF_NO_CELLULAR;
2774 else
2775 #endif /* INET6 */
2776 ipoa.ipoa_flags |= IPOAF_NO_CELLULAR;
2777 }
2778 if (INP_NO_EXPENSIVE(inp)) {
2779 #if INET6
2780 if (isipv6)
2781 ip6oa.ip6oa_flags |= IP6OAF_NO_EXPENSIVE;
2782 else
2783 #endif /* INET6 */
2784 ipoa.ipoa_flags |= IPOAF_NO_EXPENSIVE;
2785
2786 }
2787 if (INP_NO_CONSTRAINED(inp)) {
2788 #if INET6
2789 if (isipv6)
2790 ip6oa.ip6oa_flags |= IP6OAF_NO_CONSTRAINED;
2791 else
2792 #endif /* INET6 */
2793 ipoa.ipoa_flags |= IPOAF_NO_CONSTRAINED;
2794 }
2795 if (INP_AWDL_UNRESTRICTED(inp)) {
2796 #if INET6
2797 if (isipv6)
2798 ip6oa.ip6oa_flags |= IP6OAF_AWDL_UNRESTRICTED;
2799 else
2800 #endif /* INET6 */
2801 ipoa.ipoa_flags |= IPOAF_AWDL_UNRESTRICTED;
2802
2803 }
2804 #if INET6
2805 if (INP_INTCOPROC_ALLOWED(inp) && isipv6) {
2806 ip6oa.ip6oa_flags |= IP6OAF_INTCOPROC_ALLOWED;
2807 }
2808 if (isipv6) {
2809 ip6oa.ip6oa_sotc = so->so_traffic_class;
2810 ip6oa.ip6oa_netsvctype = so->so_netsvctype;
2811 } else
2812 #endif /* INET6 */
2813 {
2814 ipoa.ipoa_sotc = so->so_traffic_class;
2815 ipoa.ipoa_netsvctype = so->so_netsvctype;
2816 }
2817 if ((so->so_flags1 & SOF1_QOSMARKING_ALLOWED)) {
2818 #if INET6
2819 if (isipv6)
2820 ip6oa.ip6oa_flags |= IP6OAF_QOSMARKING_ALLOWED;
2821 else
2822 #endif /* INET6 */
2823 ipoa.ipoa_flags |= IPOAF_QOSMARKING_ALLOWED;
2824 }
2825 #if INET6
2826 if (isipv6)
2827 flags |= IPV6_OUTARGS;
2828 else
2829 #endif /* INET6 */
2830 flags |= IP_OUTARGS;
2831
2832 /* Copy the cached route and take an extra reference */
2833 #if INET6
2834 if (isipv6)
2835 in6p_route_copyout(inp, &ro6);
2836 else
2837 #endif /* INET6 */
2838 inp_route_copyout(inp, &ro);
2839
2840 /*
2841 * Make sure ACK/DELACK conditions are cleared before
2842 * we unlock the socket.
2843 */
2844 tp->last_ack_sent = tp->rcv_nxt;
2845 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
2846 tp->t_timer[TCPT_DELACK] = 0;
2847 tp->t_unacksegs = 0;
2848
2849 /* Increment the count of outstanding send operations */
2850 inp->inp_sndinprog_cnt++;
2851
2852 /*
2853 * If allowed, unlock TCP socket while in IP
2854 * but only if the connection is established and
2855 * in a normal mode where reentrancy on the tcpcb won't be
2856 * an issue:
2857 * - there is no SACK episode
2858 * - we're not in Fast Recovery mode
2859 * - if we're not sending from an upcall.
2860 */
2861 if (tcp_output_unlocked && !so->so_upcallusecount &&
2862 (tp->t_state == TCPS_ESTABLISHED) && (sack_in_progress == 0) &&
2863 !IN_FASTRECOVERY(tp) && !(so->so_flags & SOF_MP_SUBFLOW)) {
2864
2865 unlocked = TRUE;
2866 socket_unlock(so, 0);
2867 }
2868
2869 /*
2870 * Don't send down a chain of packets when:
2871 * - TCP chaining is disabled
2872 * - there is an IPsec rule set
2873 * - there is a non default rule set for the firewall
2874 */
2875
2876 chain = tcp_packet_chaining > 1
2877 #if IPSEC
2878 && ipsec_bypass
2879 #endif
2880 #if IPFIREWALL
2881 && (fw_enable == 0 || fw_bypass)
2882 #endif
2883 ; // I'm important, not extraneous
2884
2885 while (pkt != NULL) {
2886 struct mbuf *npkt = pkt->m_nextpkt;
2887
2888 if (!chain) {
2889 pkt->m_nextpkt = NULL;
2890 /*
2891 * If we are not chaining, make sure to set the packet
2892 * list count to 0 so that IP takes the right path;
2893 * this is important for cases such as IPsec where a
2894 * single mbuf might result in multiple mbufs as part
2895 * of the encapsulation. If a non-zero count is passed
2896 * down to IP, the head of the chain might change and
2897 * we could end up skipping it (thus generating bogus
2898 * packets). Fixing it in IP would be desirable, but
2899 * for now this would do it.
2900 */
2901 cnt = 0;
2902 }
2903 #if INET6
2904 if (isipv6) {
2905 error = ip6_output_list(pkt, cnt,
2906 inp->in6p_outputopts, &ro6, flags, NULL, NULL,
2907 &ip6oa);
2908 ifdenied = (ip6oa.ip6oa_retflags & IP6OARF_IFDENIED);
2909 } else {
2910 #endif /* INET6 */
2911 error = ip_output_list(pkt, cnt, opt, &ro, flags, NULL,
2912 &ipoa);
2913 ifdenied = (ipoa.ipoa_retflags & IPOARF_IFDENIED);
2914 }
2915
2916 if (chain || error) {
2917 /*
2918 * If we sent down a chain then we are done since
2919 * the callee had taken care of everything; else
2920 * we need to free the rest of the chain ourselves.
2921 */
2922 if (!chain)
2923 m_freem_list(npkt);
2924 break;
2925 }
2926 pkt = npkt;
2927 }
2928
2929 if (unlocked)
2930 socket_lock(so, 0);
2931
2932 /*
2933 * Enter flow controlled state if the connection is established
2934 * and is not in recovery. Flow control is allowed only if there
2935 * is outstanding data.
2936 *
2937 * A connection will enter suspended state even if it is in
2938 * recovery.
2939 */
2940 if (((adv->code == FADV_FLOW_CONTROLLED && !IN_FASTRECOVERY(tp)) ||
2941 adv->code == FADV_SUSPENDED) &&
2942 !(tp->t_flags & TF_CLOSING) &&
2943 tp->t_state == TCPS_ESTABLISHED &&
2944 SEQ_GT(tp->snd_max, tp->snd_una)) {
2945 int rc;
2946 rc = inp_set_fc_state(inp, adv->code);
2947
2948 if (rc == 1)
2949 tcp_ccdbg_trace(tp, NULL,
2950 ((adv->code == FADV_FLOW_CONTROLLED) ?
2951 TCP_CC_FLOW_CONTROL : TCP_CC_SUSPEND));
2952 }
2953
2954 /*
2955 * When an interface queue gets suspended, some of the
2956 * packets are dropped. Return ENOBUFS, to update the
2957 * pcb state.
2958 */
2959 if (adv->code == FADV_SUSPENDED)
2960 error = ENOBUFS;
2961
2962 VERIFY(inp->inp_sndinprog_cnt > 0);
2963 if ( --inp->inp_sndinprog_cnt == 0) {
2964 inp->inp_flags &= ~(INP_FC_FEEDBACK);
2965 if (inp->inp_sndingprog_waiters > 0) {
2966 wakeup(&inp->inp_sndinprog_cnt);
2967 }
2968 }
2969
2970 #if INET6
2971 if (isipv6) {
2972 /*
2973 * When an NECP IP tunnel policy forces the outbound interface,
2974 * ip6_output_list() informs the transport layer what is the actual
2975 * outgoing interface
2976 */
2977 if (ip6oa.ip6oa_flags & IP6OAF_BOUND_IF) {
2978 outif = ifindex2ifnet[ip6oa.ip6oa_boundif];
2979 } else if (ro6.ro_rt != NULL) {
2980 outif = ro6.ro_rt->rt_ifp;
2981 }
2982 } else
2983 #endif /* INET6 */
2984 if (ro.ro_rt != NULL)
2985 outif = ro.ro_rt->rt_ifp;
2986
2987 if (outif != NULL && outif != inp->inp_last_outifp) {
2988 /* Update the send byte count */
2989 if (so->so_snd.sb_cc > 0 && so->so_snd.sb_flags & SB_SNDBYTE_CNT) {
2990 inp_decr_sndbytes_total(so, so->so_snd.sb_cc);
2991 inp_decr_sndbytes_allunsent(so, tp->snd_una);
2992 so->so_snd.sb_flags &= ~SB_SNDBYTE_CNT;
2993 }
2994 inp->inp_last_outifp = outif;
2995
2996 }
2997
2998 if (error != 0 && ifdenied &&
2999 (INP_NO_CELLULAR(inp) || INP_NO_EXPENSIVE(inp) || INP_NO_CONSTRAINED(inp)))
3000 soevent(so,
3001 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_IFDENIED));
3002
3003 /* Synchronize cached PCB route & options */
3004 #if INET6
3005 if (isipv6)
3006 in6p_route_copyin(inp, &ro6);
3007 else
3008 #endif /* INET6 */
3009 inp_route_copyin(inp, &ro);
3010
3011 if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift == 0 &&
3012 tp->t_inpcb->inp_route.ro_rt != NULL) {
3013 /* If we found the route and there is an rtt on it
3014 * reset the retransmit timer
3015 */
3016 tcp_getrt_rtt(tp, tp->t_inpcb->in6p_route.ro_rt);
3017 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur);
3018 }
3019 return error;
3020 }
3021
3022 int tcptv_persmin_val = TCPTV_PERSMIN;
3023
3024 void
3025 tcp_setpersist(struct tcpcb *tp)
3026 {
3027 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
3028
3029 /* If a PERSIST_TIMER option was set we will limit the
3030 * time the persist timer will be active for that connection
3031 * in order to avoid DOS by using zero window probes.
3032 * see rdar://5805356
3033 */
3034
3035 if (tp->t_persist_timeout != 0 &&
3036 tp->t_timer[TCPT_PERSIST] == 0 &&
3037 tp->t_persist_stop == 0) {
3038 tp->t_persist_stop = tcp_now + tp->t_persist_timeout;
3039 }
3040
3041 /*
3042 * Start/restart persistance timer.
3043 */
3044 TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
3045 t * tcp_backoff[tp->t_rxtshift],
3046 tcptv_persmin_val, TCPTV_PERSMAX, 0);
3047 tp->t_timer[TCPT_PERSIST] = OFFSET_FROM_START(tp, tp->t_timer[TCPT_PERSIST]);
3048
3049 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
3050 tp->t_rxtshift++;
3051 }
3052
3053 /*
3054 * Send as many acks as data coalesced. Every other packet when stretch
3055 * ACK is not enabled. Every 8 packets, if stretch ACK is enabled.
3056 */
3057 static struct mbuf*
3058 tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th)
3059 {
3060 struct mbuf *mnext = NULL, *ack_chain = NULL, *tail = NULL;
3061 int count = 0;
3062 tcp_seq org_ack = ntohl(th->th_ack);
3063 tcp_seq prev_ack = 0;
3064 int tack_offset = 28; /* IPv6 and IP options not supported */
3065 int twin_offset = 34; /* IPv6 and IP options not supported */
3066 int ack_size = (tp->t_flags & TF_STRETCHACK) ?
3067 (maxseg_unacked * tp->t_maxseg) : (tp->t_maxseg << 1);
3068 int segs_acked = (tp->t_flags & TF_STRETCHACK) ? maxseg_unacked : 2;
3069 struct mbuf *prev_ack_pkt = NULL;
3070 struct socket *so = tp->t_inpcb->inp_socket;
3071 unsigned short winsz = ntohs(th->th_win);
3072 unsigned int scaled_win = winsz<<tp->rcv_scale;
3073 tcp_seq win_rtedge = org_ack + scaled_win;
3074
3075 count = tp->t_lropktlen/tp->t_maxseg;
3076
3077 prev_ack = (org_ack - tp->t_lropktlen) + ack_size;
3078 if (prev_ack < org_ack) {
3079 ack_chain = m_dup(m, M_DONTWAIT);
3080 if (ack_chain) {
3081 th->th_ack = htonl(prev_ack);
3082 /* Keep adv window constant for duplicated ACK packets */
3083 scaled_win = win_rtedge - prev_ack;
3084 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3085 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3086 th->th_win = htons(scaled_win>>tp->rcv_scale);
3087 if (lrodebug == 5) {
3088 printf("%s: win = %d winsz = %d sc = %d"
3089 " lro_len %d %d\n",
3090 __func__, scaled_win>>tp->rcv_scale, winsz,
3091 tp->rcv_scale, tp->t_lropktlen, count);
3092 }
3093 tail = ack_chain;
3094 count -= segs_acked; /* accounts for prev_ack packet */
3095 count = (count <= segs_acked) ? 0 : count - segs_acked;
3096 tcpstat.tcps_sndacks++;
3097 so_tc_update_stats(m, so, m_get_service_class(m));
3098 } else {
3099 return NULL;
3100 }
3101 }
3102 else {
3103 tp->t_lropktlen = 0;
3104 return NULL;
3105 }
3106
3107 prev_ack_pkt = ack_chain;
3108
3109 while (count > 0) {
3110 if ((prev_ack + ack_size) < org_ack) {
3111 prev_ack += ack_size;
3112 } else {
3113 /*
3114 * The last ACK sent must have the ACK number that TCP
3115 * thinks is the last sent ACK number.
3116 */
3117 prev_ack = org_ack;
3118 }
3119 mnext = m_dup(prev_ack_pkt, M_DONTWAIT);
3120 if (mnext) {
3121 /* Keep adv window constant for duplicated ACK packets */
3122 scaled_win = win_rtedge - prev_ack;
3123 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3124 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3125 winsz = htons(scaled_win>>tp->rcv_scale);
3126 if (lrodebug == 5) {
3127 printf("%s: winsz = %d ack %x count %d\n",
3128 __func__, scaled_win>>tp->rcv_scale,
3129 prev_ack, count);
3130 }
3131 bcopy(&winsz, mtod(prev_ack_pkt, caddr_t) + twin_offset, 2);
3132 HTONL(prev_ack);
3133 bcopy(&prev_ack, mtod(prev_ack_pkt, caddr_t) + tack_offset, 4);
3134 NTOHL(prev_ack);
3135 tail->m_nextpkt = mnext;
3136 tail = mnext;
3137 count -= segs_acked;
3138 tcpstat.tcps_sndacks++;
3139 so_tc_update_stats(m, so, m_get_service_class(m));
3140 } else {
3141 if (lrodebug == 5) {
3142 printf("%s: failed to alloc mbuf.\n", __func__);
3143 }
3144 break;
3145 }
3146 prev_ack_pkt = mnext;
3147 }
3148 tp->t_lropktlen = 0;
3149 return ack_chain;
3150 }
3151
3152 static int
3153 tcp_recv_throttle (struct tcpcb *tp)
3154 {
3155 uint32_t base_rtt, newsize;
3156 struct sockbuf *sbrcv = &tp->t_inpcb->inp_socket->so_rcv;
3157
3158 if (tcp_use_rtt_recvbg == 1 &&
3159 TSTMP_SUPPORTED(tp)) {
3160 /*
3161 * Timestamps are supported on this connection. Use
3162 * RTT to look for an increase in latency.
3163 */
3164
3165 /*
3166 * If the connection is already being throttled, leave it
3167 * in that state until rtt comes closer to base rtt
3168 */
3169 if (tp->t_flagsext & TF_RECV_THROTTLE)
3170 return 1;
3171
3172 base_rtt = get_base_rtt(tp);
3173
3174 if (base_rtt != 0 && tp->t_rttcur != 0) {
3175 /*
3176 * if latency increased on a background flow,
3177 * return 1 to start throttling.
3178 */
3179 if (tp->t_rttcur > (base_rtt + target_qdelay)) {
3180 tp->t_flagsext |= TF_RECV_THROTTLE;
3181 if (tp->t_recv_throttle_ts == 0)
3182 tp->t_recv_throttle_ts = tcp_now;
3183 /*
3184 * Reduce the recv socket buffer size to
3185 * minimize latecy.
3186 */
3187 if (sbrcv->sb_idealsize >
3188 tcp_recv_throttle_minwin) {
3189 newsize = sbrcv->sb_idealsize >> 1;
3190 /* Set a minimum of 16 K */
3191 newsize =
3192 max(newsize,
3193 tcp_recv_throttle_minwin);
3194 sbrcv->sb_idealsize = newsize;
3195 }
3196 return 1;
3197 } else {
3198 return 0;
3199 }
3200 }
3201 }
3202
3203 /*
3204 * Timestamps are not supported or there is no good RTT
3205 * measurement. Use IPDV in this case.
3206 */
3207 if (tp->acc_iaj > tcp_acc_iaj_react_limit)
3208 return 1;
3209
3210 return 0;
3211 }
mirror of XNU