]> git.saurik.com Git - apple/xnu.git/blob - bsd/netinet/tcp_output.c
projects / apple / xnu.git / blob
? search:


6f63e40f7b226d39429bed8aca5512f5be845e36
[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))
982 flags &= ~TH_SYN;
983 off--;
984 len++;
985 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
986 while (inp->inp_sndinprog_cnt == 0 &&
987 tp->t_pktlist_head != NULL) {
988 packetlist = tp->t_pktlist_head;
989 packchain_listadd = tp->t_lastchain;
990 packchain_sent++;
991 TCP_PKTLIST_CLEAR(tp);
992
993 error = tcp_ip_output(so, tp, packetlist,
994 packchain_listadd, tp_inp_options,
995 (so_options & SO_DONTROUTE),
996 (sack_rxmit || (sack_bytes_rxmt != 0)),
997 isipv6);
998 }
999
1000 /*
1001 * tcp was closed while we were in ip,
1002 * resume close
1003 */
1004 if (inp->inp_sndinprog_cnt == 0 &&
1005 (tp->t_flags & TF_CLOSING)) {
1006 tp->t_flags &= ~TF_CLOSING;
1007 (void) tcp_close(tp);
1008 } else {
1009 tcp_check_timer_state(tp);
1010 }
1011 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,
1012 0,0,0,0,0);
1013 return 0;
1014 }
1015 }
1016
1017 /*
1018 * Be careful not to send data and/or FIN on SYN segments.
1019 * This measure is needed to prevent interoperability problems
1020 * with not fully conformant TCP implementations.
1021 *
1022 * In case of TFO, we handle the setting of the len in
1023 * tcp_tfo_check. In case TFO is not enabled, never ever send
1024 * SYN+data.
1025 */
1026 if ((flags & TH_SYN) && !tfo_enabled(tp)) {
1027 len = 0;
1028 flags &= ~TH_FIN;
1029 }
1030
1031 /*
1032 * Don't send a RST with data.
1033 */
1034 if (flags & TH_RST)
1035 len = 0;
1036
1037 if ((flags & TH_SYN) && tp->t_state <= TCPS_SYN_SENT && tfo_enabled(tp))
1038 len = tcp_tfo_check(tp, len);
1039
1040 /*
1041 * The check here used to be (len < 0). Some times len is zero
1042 * when the congestion window is closed and we need to check
1043 * if persist timer has to be set in that case. But don't set
1044 * persist until connection is established.
1045 */
1046 if (len <= 0 && !(flags & TH_SYN)) {
1047 /*
1048 * If FIN has been sent but not acked,
1049 * but we haven't been called to retransmit,
1050 * len will be < 0. Otherwise, window shrank
1051 * after we sent into it. If window shrank to 0,
1052 * cancel pending retransmit, pull snd_nxt back
1053 * to (closed) window, and set the persist timer
1054 * if it isn't already going. If the window didn't
1055 * close completely, just wait for an ACK.
1056 */
1057 len = 0;
1058 if (sendwin == 0) {
1059 tp->t_timer[TCPT_REXMT] = 0;
1060 tp->t_timer[TCPT_PTO] = 0;
1061 TCP_RESET_REXMT_STATE(tp);
1062 tp->snd_nxt = tp->snd_una;
1063 off = 0;
1064 if (tp->t_timer[TCPT_PERSIST] == 0)
1065 tcp_setpersist(tp);
1066 }
1067 }
1068
1069 /*
1070 * Automatic sizing of send socket buffer. Increase the send
1071 * socket buffer size if all of the following criteria are met
1072 * 1. the receiver has enough buffer space for this data
1073 * 2. send buffer is filled to 7/8th with data (so we actually
1074 * have data to make use of it);
1075 * 3. our send window (slow start and congestion controlled) is
1076 * larger than sent but unacknowledged data in send buffer.
1077 */
1078 if (tcp_do_autosendbuf == 1 &&
1079 !INP_WAIT_FOR_IF_FEEDBACK(inp) && !IN_FASTRECOVERY(tp) &&
1080 (so->so_snd.sb_flags & (SB_AUTOSIZE | SB_TRIM)) == SB_AUTOSIZE &&
1081 tcp_cansbgrow(&so->so_snd)) {
1082 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
1083 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
1084 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
1085 if (sbreserve(&so->so_snd,
1086 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
1087 tcp_autosndbuf_max)) == 1) {
1088 so->so_snd.sb_idealsize = so->so_snd.sb_hiwat;
1089 }
1090 }
1091 }
1092
1093 /*
1094 * Truncate to the maximum segment length or enable TCP Segmentation
1095 * Offloading (if supported by hardware) and ensure that FIN is removed
1096 * if the length no longer contains the last data byte.
1097 *
1098 * TSO may only be used if we are in a pure bulk sending state.
1099 * The presence of TCP-MD5, SACK retransmits, SACK advertizements,
1100 * ipfw rules and IP options, as well as disabling hardware checksum
1101 * offload prevent using TSO. With TSO the TCP header is the same
1102 * (except for the sequence number) for all generated packets. This
1103 * makes it impossible to transmit any options which vary per generated
1104 * segment or packet.
1105 *
1106 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
1107 * removal of FIN (if not already catched here) are handled later after
1108 * the exact length of the TCP options are known.
1109 */
1110 #if IPSEC
1111 /*
1112 * Pre-calculate here as we save another lookup into the darknesses
1113 * of IPsec that way and can actually decide if TSO is ok.
1114 */
1115 if (ipsec_bypass == 0)
1116 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
1117 #endif
1118 if (len > tp->t_maxseg) {
1119 if ((tp->t_flags & TF_TSO) && tcp_do_tso && hwcksum_tx &&
1120 ip_use_randomid && kipf_count == 0 &&
1121 dlil_filter_disable_tso_count == 0 &&
1122 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
1123 sack_bytes_rxmt == 0 &&
1124 inp->inp_options == NULL &&
1125 inp->in6p_options == NULL
1126 #if IPSEC
1127 && ipsec_optlen == 0
1128 #endif
1129 #if IPFIREWALL
1130 && (fw_enable == 0 || fw_bypass)
1131 #endif
1132 ) {
1133 tso = 1;
1134 sendalot = 0;
1135 } else {
1136 len = tp->t_maxseg;
1137 sendalot = 1;
1138 tso = 0;
1139 }
1140 }
1141
1142 /* Send one segment or less as a tail loss probe */
1143 if (tp->t_flagsext & TF_SENT_TLPROBE) {
1144 len = min(len, tp->t_maxseg);
1145 sendalot = 0;
1146 tso = 0;
1147 }
1148
1149 #if MPTCP
1150 if (so->so_flags & SOF_MP_SUBFLOW && off < 0) {
1151 os_log_error(mptcp_log_handle, "%s - %lx: offset is negative! len %d off %d\n",
1152 __func__, (unsigned long)VM_KERNEL_ADDRPERM(tp->t_mpsub->mpts_mpte),
1153 len, off);
1154 }
1155
1156 if ((so->so_flags & SOF_MP_SUBFLOW) &&
1157 !(tp->t_mpflags & TMPF_TCP_FALLBACK)) {
1158 int newlen = len;
1159 if (tp->t_state >= TCPS_ESTABLISHED &&
1160 (tp->t_mpflags & TMPF_SND_MPPRIO ||
1161 tp->t_mpflags & TMPF_SND_REM_ADDR ||
1162 tp->t_mpflags & TMPF_SND_MPFAIL ||
1163 tp->t_mpflags & TMPF_SND_KEYS ||
1164 tp->t_mpflags & TMPF_SND_JACK)) {
1165 if (len > 0) {
1166 len = 0;
1167 }
1168 /*
1169 * On a new subflow, don't try to send again, because
1170 * we are still waiting for the fourth ack.
1171 */
1172 if (!(tp->t_mpflags & TMPF_PREESTABLISHED))
1173 sendalot = 1;
1174 mptcp_acknow = TRUE;
1175 } else {
1176 mptcp_acknow = FALSE;
1177 }
1178 /*
1179 * The contiguous bytes in the subflow socket buffer can be
1180 * discontiguous at the MPTCP level. Since only one DSS
1181 * option can be sent in one packet, reduce length to match
1182 * the contiguous MPTCP level. Set sendalot to send remainder.
1183 */
1184 if (len > 0 && off >= 0) {
1185 newlen = mptcp_adj_sendlen(so, off);
1186 }
1187
1188 if (newlen < len) {
1189 len = newlen;
1190 }
1191 }
1192 #endif /* MPTCP */
1193
1194 /*
1195 * If the socket is capable of doing unordered send,
1196 * pull the amount of data that can be sent from the
1197 * unordered priority queues to the serial queue in
1198 * the socket buffer. If bytes are not yet available
1199 * in the highest priority message, we may not be able
1200 * to send any new data.
1201 */
1202 if (so->so_flags & SOF_ENABLE_MSGS) {
1203 if ((off + len) >
1204 so->so_msg_state->msg_serial_bytes) {
1205 sbpull_unordered_data(so, off, len);
1206
1207 /* check if len needs to be modified */
1208 if ((off + len) >
1209 so->so_msg_state->msg_serial_bytes) {
1210 len = so->so_msg_state->msg_serial_bytes - off;
1211 if (len <= 0) {
1212 len = 0;
1213 tcpstat.tcps_msg_sndwaithipri++;
1214 }
1215 }
1216 }
1217 }
1218
1219 if (sack_rxmit) {
1220 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
1221 flags &= ~TH_FIN;
1222 } else {
1223 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
1224 flags &= ~TH_FIN;
1225 }
1226 /*
1227 * Compare available window to amount of window
1228 * known to peer (as advertised window less
1229 * next expected input). If the difference is at least two
1230 * max size segments, or at least 25% of the maximum possible
1231 * window, then want to send a window update to peer.
1232 */
1233 recwin = tcp_sbspace(tp);
1234
1235 if (!(so->so_flags & SOF_MP_SUBFLOW)) {
1236 if (recwin < (int32_t)(so->so_rcv.sb_hiwat / 4) &&
1237 recwin < (int)tp->t_maxseg) {
1238 recwin = 0;
1239 }
1240 } else {
1241 struct mptcb *mp_tp = tptomptp(tp);
1242 struct socket *mp_so = mptetoso(mp_tp->mpt_mpte);
1243
1244 if (recwin < (int32_t)(mp_so->so_rcv.sb_hiwat / 4) &&
1245 recwin < (int)tp->t_maxseg) {
1246 recwin = 0;
1247 }
1248 }
1249
1250 #if TRAFFIC_MGT
1251 if (tcp_recv_bg == 1 || IS_TCP_RECV_BG(so)) {
1252 if (recwin > 0 && tcp_recv_throttle(tp)) {
1253 uint32_t min_iaj_win = tcp_min_iaj_win * tp->t_maxseg;
1254 uint32_t bg_rwintop = tp->rcv_adv;
1255 if (SEQ_LT(bg_rwintop, tp->rcv_nxt + min_iaj_win))
1256 bg_rwintop = tp->rcv_nxt + min_iaj_win;
1257 recwin = imin((int32_t)(bg_rwintop - tp->rcv_nxt),
1258 recwin);
1259 if (recwin < 0)
1260 recwin = 0;
1261 }
1262 }
1263 #endif /* TRAFFIC_MGT */
1264
1265 if (recwin > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
1266 recwin = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
1267
1268 if (!(so->so_flags & SOF_MP_SUBFLOW)) {
1269 if (recwin < (int32_t)(tp->rcv_adv - tp->rcv_nxt)) {
1270 recwin = (int32_t)(tp->rcv_adv - tp->rcv_nxt);
1271 }
1272 } else {
1273 struct mptcb *mp_tp = tptomptp(tp);
1274
1275 /* Don't remove what we announced at the MPTCP-layer */
1276 if (recwin < (int32_t)(mp_tp->mpt_rcvadv - (uint32_t)mp_tp->mpt_rcvnxt)) {
1277 recwin = (int32_t)(mp_tp->mpt_rcvadv - (uint32_t)mp_tp->mpt_rcvnxt);
1278 }
1279 }
1280
1281 /*
1282 * Sender silly window avoidance. We transmit under the following
1283 * conditions when len is non-zero:
1284 *
1285 * - we've timed out (e.g. persist timer)
1286 * - we need to retransmit
1287 * - We have a full segment (or more with TSO)
1288 * - This is the last buffer in a write()/send() and we are
1289 * either idle or running NODELAY
1290 * - we have more then 1/2 the maximum send window's worth of
1291 * data (receiver may be limited the window size)
1292 */
1293 if (len) {
1294 if (tp->t_flagsext & TF_FORCE)
1295 goto send;
1296 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1297 goto send;
1298 if (sack_rxmit)
1299 goto send;
1300
1301 /*
1302 * If this here is the first segment after SYN/ACK and TFO
1303 * is being used, then we always send it, regardless of Nagle,...
1304 */
1305 if (tp->t_state == TCPS_SYN_RECEIVED &&
1306 tfo_enabled(tp) &&
1307 (tp->t_tfo_flags & TFO_F_COOKIE_VALID) &&
1308 tp->snd_nxt == tp->iss + 1)
1309 goto send;
1310
1311 /*
1312 * Send new data on the connection only if it is
1313 * not flow controlled
1314 */
1315 if (!INP_WAIT_FOR_IF_FEEDBACK(inp) ||
1316 tp->t_state != TCPS_ESTABLISHED) {
1317 if (len >= tp->t_maxseg)
1318 goto send;
1319
1320 if (!(tp->t_flags & TF_MORETOCOME) &&
1321 (idle || tp->t_flags & TF_NODELAY ||
1322 (tp->t_flags & TF_MAXSEGSNT) ||
1323 ALLOW_LIMITED_TRANSMIT(tp)) &&
1324 (tp->t_flags & TF_NOPUSH) == 0 &&
1325 (len + off >= so->so_snd.sb_cc ||
1326 /*
1327 * MPTCP needs to respect the DSS-mappings. So, it
1328 * may be sending data that *could* have been
1329 * coalesced, but cannot because of
1330 * mptcp_adj_sendlen().
1331 */
1332 so->so_flags & SOF_MP_SUBFLOW))
1333 goto send;
1334 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
1335 goto send;
1336 } else {
1337 tcpstat.tcps_fcholdpacket++;
1338 }
1339 }
1340
1341 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
1342 /*
1343 * "adv" is the amount we can increase the window,
1344 * taking into account that we are limited by
1345 * TCP_MAXWIN << tp->rcv_scale.
1346 */
1347 int32_t adv, oldwin = 0;
1348 adv = imin(recwin, (int)TCP_MAXWIN << tp->rcv_scale) -
1349 (tp->rcv_adv - tp->rcv_nxt);
1350
1351 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
1352 oldwin = tp->rcv_adv - tp->rcv_nxt;
1353
1354 if (adv >= (int32_t) (2 * tp->t_maxseg)) {
1355 /*
1356 * Update only if the resulting scaled value of
1357 * the window changed, or if there is a change in
1358 * the sequence since the last ack. This avoids
1359 * what appears as dupe ACKS (see rdar://5640997)
1360 *
1361 * If streaming is detected avoid sending too many
1362 * window updates. We will depend on the delack
1363 * timer to send a window update when needed.
1364 */
1365 if (!(tp->t_flags & TF_STRETCHACK) &&
1366 (tp->last_ack_sent != tp->rcv_nxt ||
1367 ((oldwin + adv) >> tp->rcv_scale) >
1368 (oldwin >> tp->rcv_scale))) {
1369 goto send;
1370 }
1371
1372 }
1373 if (4 * adv >= (int32_t) so->so_rcv.sb_hiwat)
1374 goto send;
1375
1376 /*
1377 * Make sure that the delayed ack timer is set if
1378 * we delayed sending a window update because of
1379 * streaming detection.
1380 */
1381 if ((tp->t_flags & TF_STRETCHACK) &&
1382 !(tp->t_flags & TF_DELACK)) {
1383 tp->t_flags |= TF_DELACK;
1384 tp->t_timer[TCPT_DELACK] =
1385 OFFSET_FROM_START(tp, tcp_delack);
1386 }
1387 }
1388
1389 /*
1390 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
1391 * is also a catch-all for the retransmit timer timeout case.
1392 */
1393 if (tp->t_flags & TF_ACKNOW)
1394 goto send;
1395 if ((flags & TH_RST) ||
1396 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
1397 goto send;
1398 if (SEQ_GT(tp->snd_up, tp->snd_una))
1399 goto send;
1400 #if MPTCP
1401 if (mptcp_acknow)
1402 goto send;
1403 #endif /* MPTCP */
1404 /*
1405 * If our state indicates that FIN should be sent
1406 * and we have not yet done so, then we need to send.
1407 */
1408 if ((flags & TH_FIN) &&
1409 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
1410 goto send;
1411 /*
1412 * In SACK, it is possible for tcp_output to fail to send a segment
1413 * after the retransmission timer has been turned off. Make sure
1414 * that the retransmission timer is set.
1415 */
1416 if (SACK_ENABLED(tp) && (tp->t_state >= TCPS_ESTABLISHED) &&
1417 SEQ_GT(tp->snd_max, tp->snd_una) &&
1418 tp->t_timer[TCPT_REXMT] == 0 &&
1419 tp->t_timer[TCPT_PERSIST] == 0) {
1420 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp,
1421 tp->t_rxtcur);
1422 goto just_return;
1423 }
1424 /*
1425 * TCP window updates are not reliable, rather a polling protocol
1426 * using ``persist'' packets is used to insure receipt of window
1427 * updates. The three ``states'' for the output side are:
1428 * idle not doing retransmits or persists
1429 * persisting to move a small or zero window
1430 * (re)transmitting and thereby not persisting
1431 *
1432 * tp->t_timer[TCPT_PERSIST]
1433 * is set when we are in persist state.
1434 * tp->t_force
1435 * is set when we are called to send a persist packet.
1436 * tp->t_timer[TCPT_REXMT]
1437 * is set when we are retransmitting
1438 * The output side is idle when both timers are zero.
1439 *
1440 * If send window is too small, there is data to transmit, and no
1441 * retransmit or persist is pending, then go to persist state.
1442 * If nothing happens soon, send when timer expires:
1443 * if window is nonzero, transmit what we can,
1444 * otherwise force out a byte.
1445 */
1446 if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
1447 tp->t_timer[TCPT_PERSIST] == 0) {
1448 TCP_RESET_REXMT_STATE(tp);
1449 tcp_setpersist(tp);
1450 }
1451 just_return:
1452 /*
1453 * If there is no reason to send a segment, just return.
1454 * but if there is some packets left in the packet list, send them now.
1455 */
1456 while (inp->inp_sndinprog_cnt == 0 &&
1457 tp->t_pktlist_head != NULL) {
1458 packetlist = tp->t_pktlist_head;
1459 packchain_listadd = tp->t_lastchain;
1460 packchain_sent++;
1461 TCP_PKTLIST_CLEAR(tp);
1462
1463 error = tcp_ip_output(so, tp, packetlist,
1464 packchain_listadd,
1465 tp_inp_options, (so_options & SO_DONTROUTE),
1466 (sack_rxmit || (sack_bytes_rxmt != 0)), isipv6);
1467 }
1468 /* tcp was closed while we were in ip; resume close */
1469 if (inp->inp_sndinprog_cnt == 0 &&
1470 (tp->t_flags & TF_CLOSING)) {
1471 tp->t_flags &= ~TF_CLOSING;
1472 (void) tcp_close(tp);
1473 } else {
1474 tcp_check_timer_state(tp);
1475 }
1476 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
1477 return 0;
1478
1479 send:
1480 /*
1481 * Set TF_MAXSEGSNT flag if the segment size is greater than
1482 * the max segment size.
1483 */
1484 if (len > 0) {
1485 if (len >= tp->t_maxseg)
1486 tp->t_flags |= TF_MAXSEGSNT;
1487 else
1488 tp->t_flags &= ~TF_MAXSEGSNT;
1489 }
1490 /*
1491 * Before ESTABLISHED, force sending of initial options
1492 * unless TCP set not to do any options.
1493 * NOTE: we assume that the IP/TCP header plus TCP options
1494 * always fit in a single mbuf, leaving room for a maximum
1495 * link header, i.e.
1496 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1497 */
1498 optlen = 0;
1499 #if INET6
1500 if (isipv6)
1501 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
1502 else
1503 #endif
1504 hdrlen = sizeof (struct tcpiphdr);
1505 if (flags & TH_SYN) {
1506 tp->snd_nxt = tp->iss;
1507 if ((tp->t_flags & TF_NOOPT) == 0) {
1508 u_short mss;
1509
1510 opt[0] = TCPOPT_MAXSEG;
1511 opt[1] = TCPOLEN_MAXSEG;
1512 mss = htons((u_short) tcp_mssopt(tp));
1513 (void)memcpy(opt + 2, &mss, sizeof(mss));
1514 optlen = TCPOLEN_MAXSEG;
1515
1516 if ((tp->t_flags & TF_REQ_SCALE) &&
1517 ((flags & TH_ACK) == 0 ||
1518 (tp->t_flags & TF_RCVD_SCALE))) {
1519 *((u_int32_t *)(void *)(opt + optlen)) = htonl(
1520 TCPOPT_NOP << 24 |
1521 TCPOPT_WINDOW << 16 |
1522 TCPOLEN_WINDOW << 8 |
1523 tp->request_r_scale);
1524 optlen += 4;
1525 }
1526 #if MPTCP
1527 if (mptcp_enable && (so->so_flags & SOF_MP_SUBFLOW)) {
1528 optlen = mptcp_setup_syn_opts(so, opt, optlen);
1529 }
1530 #endif /* MPTCP */
1531 }
1532 }
1533
1534 /*
1535 * Send a timestamp and echo-reply if this is a SYN and our side
1536 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1537 * and our peer have sent timestamps in our SYN's.
1538 */
1539 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1540 (flags & TH_RST) == 0 &&
1541 ((flags & TH_ACK) == 0 ||
1542 (tp->t_flags & TF_RCVD_TSTMP))) {
1543 u_int32_t *lp = (u_int32_t *)(void *)(opt + optlen);
1544
1545 /* Form timestamp option as shown in appendix A of RFC 1323. */
1546 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1547 *lp++ = htonl(tcp_now);
1548 *lp = htonl(tp->ts_recent);
1549 optlen += TCPOLEN_TSTAMP_APPA;
1550 }
1551
1552 /* Note the timestamp for receive buffer autosizing */
1553 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1554 tp->rfbuf_ts = tcp_now;
1555
1556 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1557 /*
1558 * Tack on the SACK permitted option *last*.
1559 * And do padding of options after tacking this on.
1560 * This is because of MSS, TS, WinScale and Signatures are
1561 * all present, we have just 2 bytes left for the SACK
1562 * permitted option, which is just enough.
1563 */
1564 /*
1565 * If this is the first SYN of connection (not a SYN
1566 * ACK), include SACK permitted option. If this is a
1567 * SYN ACK, include SACK permitted option if peer has
1568 * already done so. This is only for active connect,
1569 * since the syncache takes care of the passive connect.
1570 */
1571 if ((flags & TH_SYN) &&
1572 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) {
1573 u_char *bp;
1574 bp = (u_char *)opt + optlen;
1575
1576 *bp++ = TCPOPT_SACK_PERMITTED;
1577 *bp++ = TCPOLEN_SACK_PERMITTED;
1578 optlen += TCPOLEN_SACK_PERMITTED;
1579 }
1580 }
1581 #if MPTCP
1582 if (so->so_flags & SOF_MP_SUBFLOW) {
1583 /*
1584 * Its important to piggyback acks with data as ack only packets
1585 * may get lost and data packets that don't send Data ACKs
1586 * still advance the subflow level ACK and therefore make it
1587 * hard for the remote end to recover in low cwnd situations.
1588 */
1589 if (len != 0) {
1590 tp->t_mpflags |= (TMPF_SEND_DSN |
1591 TMPF_MPTCP_ACKNOW);
1592 } else {
1593 tp->t_mpflags |= TMPF_MPTCP_ACKNOW;
1594 }
1595 optlen = mptcp_setup_opts(tp, off, &opt[0], optlen, flags,
1596 len, &mptcp_acknow);
1597 tp->t_mpflags &= ~TMPF_SEND_DSN;
1598 }
1599 #endif /* MPTCP */
1600
1601 if (tfo_enabled(tp) && !(tp->t_flags & TF_NOOPT) &&
1602 (flags & (TH_SYN | TH_ACK)) == TH_SYN)
1603 optlen += tcp_tfo_write_cookie(tp, optlen, len, opt);
1604
1605 if (tfo_enabled(tp) &&
1606 (flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1607 (tp->t_tfo_flags & TFO_F_OFFER_COOKIE))
1608 optlen += tcp_tfo_write_cookie_rep(tp, optlen, opt);
1609
1610 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1611 /*
1612 * Send SACKs if necessary. This should be the last
1613 * option processed. Only as many SACKs are sent as
1614 * are permitted by the maximum options size.
1615 *
1616 * In general, SACK blocks consume 8*n+2 bytes.
1617 * So a full size SACK blocks option is 34 bytes
1618 * (to generate 4 SACK blocks). At a minimum,
1619 * we need 10 bytes (to generate 1 SACK block).
1620 * If TCP Timestamps (12 bytes) and TCP Signatures
1621 * (18 bytes) are both present, we'll just have
1622 * 10 bytes for SACK options 40 - (12 + 18).
1623 */
1624 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1625 (tp->t_flags & TF_SACK_PERMIT) &&
1626 (tp->rcv_numsacks > 0 || TCP_SEND_DSACK_OPT(tp)) &&
1627 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) {
1628 int nsack, padlen;
1629 u_char *bp = (u_char *)opt + optlen;
1630 u_int32_t *lp;
1631
1632 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK;
1633 nsack = min(nsack, (tp->rcv_numsacks +
1634 (TCP_SEND_DSACK_OPT(tp) ? 1 : 0)));
1635 sackoptlen = (2 + nsack * TCPOLEN_SACK);
1636
1637 /*
1638 * First we need to pad options so that the
1639 * SACK blocks can start at a 4-byte boundary
1640 * (sack option and length are at a 2 byte offset).
1641 */
1642 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4;
1643 optlen += padlen;
1644 while (padlen-- > 0)
1645 *bp++ = TCPOPT_NOP;
1646
1647 tcpstat.tcps_sack_send_blocks++;
1648 *bp++ = TCPOPT_SACK;
1649 *bp++ = sackoptlen;
1650 lp = (u_int32_t *)(void *)bp;
1651
1652 /*
1653 * First block of SACK option should represent
1654 * DSACK. Prefer to send SACK information if there
1655 * is space for only one SACK block. This will
1656 * allow for faster recovery.
1657 */
1658 if (TCP_SEND_DSACK_OPT(tp) && nsack > 0 &&
1659 (tp->rcv_numsacks == 0 || nsack > 1)) {
1660 *lp++ = htonl(tp->t_dsack_lseq);
1661 *lp++ = htonl(tp->t_dsack_rseq);
1662 tcpstat.tcps_dsack_sent++;
1663 tp->t_dsack_sent++;
1664 nsack--;
1665 }
1666 VERIFY(nsack == 0 || tp->rcv_numsacks >= nsack);
1667 for (i = 0; i < nsack; i++) {
1668 struct sackblk sack = tp->sackblks[i];
1669 *lp++ = htonl(sack.start);
1670 *lp++ = htonl(sack.end);
1671 }
1672 optlen += sackoptlen;
1673 }
1674 }
1675
1676 /* Pad TCP options to a 4 byte boundary */
1677 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) {
1678 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t));
1679 u_char *bp = (u_char *)opt + optlen;
1680
1681 optlen += pad;
1682 while (pad) {
1683 *bp++ = TCPOPT_EOL;
1684 pad--;
1685 }
1686 }
1687
1688 /*
1689 * RFC 3168 states that:
1690 * - If you ever sent an ECN-setup SYN/SYN-ACK you must be prepared
1691 * to handle the TCP ECE flag, even if you also later send a
1692 * non-ECN-setup SYN/SYN-ACK.
1693 * - If you ever send a non-ECN-setup SYN/SYN-ACK, you must not set
1694 * the ip ECT flag.
1695 *
1696 * It is not clear how the ECE flag would ever be set if you never
1697 * set the IP ECT flag on outbound packets. All the same, we use
1698 * the TE_SETUPSENT to indicate that we have committed to handling
1699 * the TCP ECE flag correctly. We use the TE_SENDIPECT to indicate
1700 * whether or not we should set the IP ECT flag on outbound packet
1701 *
1702 * For a SYN-ACK, send an ECN setup SYN-ACK
1703 */
1704 if ((flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1705 (tp->ecn_flags & TE_ENABLE_ECN)) {
1706 if (tp->ecn_flags & TE_SETUPRECEIVED) {
1707 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1708 /*
1709 * Setting TH_ECE makes this an ECN-setup
1710 * SYN-ACK
1711 */
1712 flags |= TH_ECE;
1713
1714 /*
1715 * Record that we sent the ECN-setup and
1716 * default to setting IP ECT.
1717 */
1718 tp->ecn_flags |= (TE_SETUPSENT|TE_SENDIPECT);
1719 tcpstat.tcps_ecn_server_setup++;
1720 tcpstat.tcps_ecn_server_success++;
1721 } else {
1722 /*
1723 * We sent an ECN-setup SYN-ACK but it was
1724 * dropped. Fallback to non-ECN-setup
1725 * SYN-ACK and clear flag to indicate that
1726 * we should not send data with IP ECT set
1727 *
1728 * Pretend we didn't receive an
1729 * ECN-setup SYN.
1730 *
1731 * We already incremented the counter
1732 * assuming that the ECN setup will
1733 * succeed. Decrementing here
1734 * tcps_ecn_server_success to correct it.
1735 */
1736 if (tp->ecn_flags & TE_SETUPSENT) {
1737 tcpstat.tcps_ecn_lost_synack++;
1738 tcpstat.tcps_ecn_server_success--;
1739 tp->ecn_flags |= TE_LOST_SYNACK;
1740 }
1741
1742 tp->ecn_flags &=
1743 ~(TE_SETUPRECEIVED | TE_SENDIPECT |
1744 TE_SENDCWR);
1745 }
1746 }
1747 } else if ((flags & (TH_SYN | TH_ACK)) == TH_SYN &&
1748 (tp->ecn_flags & TE_ENABLE_ECN)) {
1749 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1750 /*
1751 * Setting TH_ECE and TH_CWR makes this an
1752 * ECN-setup SYN
1753 */
1754 flags |= (TH_ECE | TH_CWR);
1755 tcpstat.tcps_ecn_client_setup++;
1756 tp->ecn_flags |= TE_CLIENT_SETUP;
1757
1758 /*
1759 * Record that we sent the ECN-setup and default to
1760 * setting IP ECT.
1761 */
1762 tp->ecn_flags |= (TE_SETUPSENT | TE_SENDIPECT);
1763 } else {
1764 /*
1765 * We sent an ECN-setup SYN but it was dropped.
1766 * Fall back to non-ECN and clear flag indicating
1767 * we should send data with IP ECT set.
1768 */
1769 if (tp->ecn_flags & TE_SETUPSENT) {
1770 tcpstat.tcps_ecn_lost_syn++;
1771 tp->ecn_flags |= TE_LOST_SYN;
1772 }
1773 tp->ecn_flags &= ~TE_SENDIPECT;
1774 }
1775 }
1776
1777 /*
1778 * Check if we should set the TCP CWR flag.
1779 * CWR flag is sent when we reduced the congestion window because
1780 * we received a TCP ECE or we performed a fast retransmit. We
1781 * never set the CWR flag on retransmitted packets. We only set
1782 * the CWR flag on data packets. Pure acks don't have this set.
1783 */
1784 if ((tp->ecn_flags & TE_SENDCWR) != 0 && len != 0 &&
1785 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
1786 flags |= TH_CWR;
1787 tp->ecn_flags &= ~TE_SENDCWR;
1788 }
1789
1790 /*
1791 * Check if we should set the TCP ECE flag.
1792 */
1793 if ((tp->ecn_flags & TE_SENDECE) != 0 && len == 0) {
1794 flags |= TH_ECE;
1795 tcpstat.tcps_ecn_sent_ece++;
1796 }
1797
1798
1799 hdrlen += optlen;
1800
1801 /* Reset DSACK sequence numbers */
1802 tp->t_dsack_lseq = 0;
1803 tp->t_dsack_rseq = 0;
1804
1805 #if INET6
1806 if (isipv6)
1807 ipoptlen = ip6_optlen(inp);
1808 else
1809 #endif
1810 {
1811 if (tp_inp_options) {
1812 ipoptlen = tp_inp_options->m_len -
1813 offsetof(struct ipoption, ipopt_list);
1814 } else {
1815 ipoptlen = 0;
1816 }
1817 }
1818 #if IPSEC
1819 ipoptlen += ipsec_optlen;
1820 #endif
1821
1822 /*
1823 * Adjust data length if insertion of options will
1824 * bump the packet length beyond the t_maxopd length.
1825 * Clear the FIN bit because we cut off the tail of
1826 * the segment.
1827 *
1828 * When doing TSO limit a burst to TCP_MAXWIN minus the
1829 * IP, TCP and Options length to keep ip->ip_len from
1830 * overflowing. Prevent the last segment from being
1831 * fractional thus making them all equal sized and set
1832 * the flag to continue sending. TSO is disabled when
1833 * IP options or IPSEC are present.
1834 */
1835 if (len + optlen + ipoptlen > tp->t_maxopd) {
1836 /*
1837 * If there is still more to send,
1838 * don't close the connection.
1839 */
1840 flags &= ~TH_FIN;
1841 if (tso) {
1842 int32_t tso_maxlen;
1843
1844 tso_maxlen = tp->tso_max_segment_size ?
1845 tp->tso_max_segment_size : TCP_MAXWIN;
1846
1847 if (len > tso_maxlen - hdrlen - optlen) {
1848 len = tso_maxlen - hdrlen - optlen;
1849 len = len - (len % (tp->t_maxopd - optlen));
1850 sendalot = 1;
1851 } else if (tp->t_flags & TF_NEEDFIN) {
1852 sendalot = 1;
1853 }
1854 } else {
1855 len = tp->t_maxopd - optlen - ipoptlen;
1856 sendalot = 1;
1857 }
1858 }
1859
1860 if (max_linkhdr + hdrlen > MCLBYTES)
1861 panic("tcphdr too big");
1862
1863 /* Check if there is enough data in the send socket
1864 * buffer to start measuring bandwidth
1865 */
1866 if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 &&
1867 (tp->t_bwmeas != NULL) &&
1868 (tp->t_flagsext & TF_BWMEAS_INPROGRESS) == 0) {
1869 tp->t_bwmeas->bw_size = min(min(
1870 (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)),
1871 tp->snd_cwnd), tp->snd_wnd);
1872 if (tp->t_bwmeas->bw_minsize > 0 &&
1873 tp->t_bwmeas->bw_size < tp->t_bwmeas->bw_minsize)
1874 tp->t_bwmeas->bw_size = 0;
1875 if (tp->t_bwmeas->bw_maxsize > 0)
1876 tp->t_bwmeas->bw_size = min(tp->t_bwmeas->bw_size,
1877 tp->t_bwmeas->bw_maxsize);
1878 if (tp->t_bwmeas->bw_size > 0) {
1879 tp->t_flagsext |= TF_BWMEAS_INPROGRESS;
1880 tp->t_bwmeas->bw_start = tp->snd_max;
1881 tp->t_bwmeas->bw_ts = tcp_now;
1882 }
1883 }
1884
1885 VERIFY(inp->inp_flowhash != 0);
1886 /*
1887 * Grab a header mbuf, attaching a copy of data to
1888 * be transmitted, and initialize the header from
1889 * the template for sends on this connection.
1890 */
1891 if (len) {
1892 tp->t_pmtud_lastseg_size = len + optlen + ipoptlen;
1893 if ((tp->t_flagsext & TF_FORCE) && len == 1)
1894 tcpstat.tcps_sndprobe++;
1895 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1896 tcpstat.tcps_sndrexmitpack++;
1897 tcpstat.tcps_sndrexmitbyte += len;
1898 if (nstat_collect) {
1899 nstat_route_tx(inp->inp_route.ro_rt, 1,
1900 len, NSTAT_TX_FLAG_RETRANSMIT);
1901 INP_ADD_STAT(inp, cell, wifi, wired,
1902 txpackets, 1);
1903 INP_ADD_STAT(inp, cell, wifi, wired,
1904 txbytes, len);
1905 tp->t_stat.txretransmitbytes += len;
1906 tp->t_stat.rxmitpkts++;
1907 }
1908 } else {
1909 tcpstat.tcps_sndpack++;
1910 tcpstat.tcps_sndbyte += len;
1911
1912 if (nstat_collect) {
1913 INP_ADD_STAT(inp, cell, wifi, wired,
1914 txpackets, 1);
1915 INP_ADD_STAT(inp, cell, wifi, wired,
1916 txbytes, len);
1917 }
1918 inp_decr_sndbytes_unsent(so, len);
1919 }
1920 inp_set_activity_bitmap(inp);
1921 #if MPTCP
1922 if (tp->t_mpflags & TMPF_MPTCP_TRUE) {
1923 tcpstat.tcps_mp_sndpacks++;
1924 tcpstat.tcps_mp_sndbytes += len;
1925 }
1926 #endif /* MPTCP */
1927 /*
1928 * try to use the new interface that allocates all
1929 * the necessary mbuf hdrs under 1 mbuf lock and
1930 * avoids rescanning the socket mbuf list if
1931 * certain conditions are met. This routine can't
1932 * be used in the following cases...
1933 * 1) the protocol headers exceed the capacity of
1934 * of a single mbuf header's data area (no cluster attached)
1935 * 2) the length of the data being transmitted plus
1936 * the protocol headers fits into a single mbuf header's
1937 * data area (no cluster attached)
1938 */
1939 m = NULL;
1940
1941 /* minimum length we are going to allocate */
1942 allocated_len = MHLEN;
1943 if (MHLEN < hdrlen + max_linkhdr) {
1944 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1945 if (m == NULL) {
1946 error = ENOBUFS;
1947 goto out;
1948 }
1949 MCLGET(m, M_DONTWAIT);
1950 if ((m->m_flags & M_EXT) == 0) {
1951 m_freem(m);
1952 error = ENOBUFS;
1953 goto out;
1954 }
1955 m->m_data += max_linkhdr;
1956 m->m_len = hdrlen;
1957 allocated_len = MCLBYTES;
1958 }
1959 if (len <= allocated_len - hdrlen - max_linkhdr) {
1960 if (m == NULL) {
1961 VERIFY(allocated_len <= MHLEN);
1962 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1963 if (m == NULL) {
1964 error = ENOBUFS;
1965 goto out;
1966 }
1967 m->m_data += max_linkhdr;
1968 m->m_len = hdrlen;
1969 }
1970 /* makes sure we still have data left to be sent at this point */
1971 if (so->so_snd.sb_mb == NULL || off < 0) {
1972 if (m != NULL) m_freem(m);
1973 error = 0; /* should we return an error? */
1974 goto out;
1975 }
1976 m_copydata(so->so_snd.sb_mb, off, (int) len,
1977 mtod(m, caddr_t) + hdrlen);
1978 m->m_len += len;
1979 } else {
1980 uint32_t copymode;
1981 /*
1982 * Retain packet header metadata at the socket
1983 * buffer if this is is an MPTCP subflow,
1984 * otherwise move it.
1985 */
1986 copymode = M_COPYM_MOVE_HDR;
1987 #if MPTCP
1988 if (so->so_flags & SOF_MP_SUBFLOW) {
1989 copymode = M_COPYM_NOOP_HDR;
1990 }
1991 #endif /* MPTCP */
1992 if (m != NULL) {
1993 m->m_next = m_copym_mode(so->so_snd.sb_mb,
1994 off, (int)len, M_DONTWAIT, copymode);
1995 if (m->m_next == NULL) {
1996 (void) m_free(m);
1997 error = ENOBUFS;
1998 goto out;
1999 }
2000 } else {
2001 /*
2002 * make sure we still have data left
2003 * to be sent at this point
2004 */
2005 if (so->so_snd.sb_mb == NULL) {
2006 error = 0; /* should we return an error? */
2007 goto out;
2008 }
2009
2010 /*
2011 * m_copym_with_hdrs will always return the
2012 * last mbuf pointer and the offset into it that
2013 * it acted on to fullfill the current request,
2014 * whether a valid 'hint' was passed in or not.
2015 */
2016 if ((m = m_copym_with_hdrs(so->so_snd.sb_mb,
2017 off, len, M_DONTWAIT, NULL, NULL,
2018 copymode)) == NULL) {
2019 error = ENOBUFS;
2020 goto out;
2021 }
2022 m->m_data += max_linkhdr;
2023 m->m_len = hdrlen;
2024 }
2025 }
2026 /*
2027 * If we're sending everything we've got, set PUSH.
2028 * (This will keep happy those implementations which only
2029 * give data to the user when a buffer fills or
2030 * a PUSH comes in.)
2031 *
2032 * On SYN-segments we should not add the PUSH-flag.
2033 */
2034 if (off + len == so->so_snd.sb_cc && !(flags & TH_SYN))
2035 flags |= TH_PUSH;
2036 } else {
2037 if (tp->t_flags & TF_ACKNOW)
2038 tcpstat.tcps_sndacks++;
2039 else if (flags & (TH_SYN|TH_FIN|TH_RST))
2040 tcpstat.tcps_sndctrl++;
2041 else if (SEQ_GT(tp->snd_up, tp->snd_una))
2042 tcpstat.tcps_sndurg++;
2043 else
2044 tcpstat.tcps_sndwinup++;
2045
2046 MGETHDR(m, M_DONTWAIT, MT_HEADER); /* MAC-OK */
2047 if (m == NULL) {
2048 error = ENOBUFS;
2049 goto out;
2050 }
2051 if (MHLEN < (hdrlen + max_linkhdr)) {
2052 MCLGET(m, M_DONTWAIT);
2053 if ((m->m_flags & M_EXT) == 0) {
2054 m_freem(m);
2055 error = ENOBUFS;
2056 goto out;
2057 }
2058 }
2059 m->m_data += max_linkhdr;
2060 m->m_len = hdrlen;
2061 }
2062 m->m_pkthdr.rcvif = 0;
2063 #if CONFIG_MACF_NET
2064 mac_mbuf_label_associate_inpcb(inp, m);
2065 #endif
2066 #if INET6
2067 if (isipv6) {
2068 ip6 = mtod(m, struct ip6_hdr *);
2069 th = (struct tcphdr *)(void *)(ip6 + 1);
2070 tcp_fillheaders(tp, ip6, th);
2071 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2072 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
2073 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
2074 }
2075 svc_flags |= PKT_SCF_IPV6;
2076 #if PF_ECN
2077 m_pftag(m)->pftag_hdr = (void *)ip6;
2078 m_pftag(m)->pftag_flags |= PF_TAG_HDR_INET6;
2079 #endif /* PF_ECN */
2080 } else
2081 #endif /* INET6 */
2082 {
2083 ip = mtod(m, struct ip *);
2084 ipov = (struct ipovly *)ip;
2085 th = (struct tcphdr *)(void *)(ip + 1);
2086 /* this picks up the pseudo header (w/o the length) */
2087 tcp_fillheaders(tp, ip, th);
2088 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2089 !SEQ_LT(tp->snd_nxt, tp->snd_max) &&
2090 !sack_rxmit && !(flags & TH_SYN)) {
2091 ip->ip_tos |= IPTOS_ECN_ECT0;
2092 }
2093 #if PF_ECN
2094 m_pftag(m)->pftag_hdr = (void *)ip;
2095 m_pftag(m)->pftag_flags |= PF_TAG_HDR_INET;
2096 #endif /* PF_ECN */
2097 }
2098
2099 /*
2100 * Fill in fields, remembering maximum advertised
2101 * window for use in delaying messages about window sizes.
2102 * If resending a FIN, be sure not to use a new sequence number.
2103 */
2104 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
2105 tp->snd_nxt == tp->snd_max)
2106 tp->snd_nxt--;
2107 /*
2108 * If we are doing retransmissions, then snd_nxt will
2109 * not reflect the first unsent octet. For ACK only
2110 * packets, we do not want the sequence number of the
2111 * retransmitted packet, we want the sequence number
2112 * of the next unsent octet. So, if there is no data
2113 * (and no SYN or FIN), use snd_max instead of snd_nxt
2114 * when filling in ti_seq. But if we are in persist
2115 * state, snd_max might reflect one byte beyond the
2116 * right edge of the window, so use snd_nxt in that
2117 * case, since we know we aren't doing a retransmission.
2118 * (retransmit and persist are mutually exclusive...)
2119 *
2120 * Note the state of this retransmit segment to detect spurious
2121 * retransmissions.
2122 */
2123 if (sack_rxmit == 0) {
2124 if (len || (flags & (TH_SYN|TH_FIN)) ||
2125 tp->t_timer[TCPT_PERSIST]) {
2126 th->th_seq = htonl(tp->snd_nxt);
2127 if (len > 0) {
2128 m->m_pkthdr.tx_start_seq = tp->snd_nxt;
2129 m->m_pkthdr.pkt_flags |= PKTF_START_SEQ;
2130 }
2131 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
2132 if (SACK_ENABLED(tp) && len > 1) {
2133 tcp_rxtseg_insert(tp, tp->snd_nxt,
2134 (tp->snd_nxt + len - 1));
2135 }
2136 if (len > 0)
2137 m->m_pkthdr.pkt_flags |=
2138 PKTF_TCP_REXMT;
2139 }
2140 } else {
2141 th->th_seq = htonl(tp->snd_max);
2142 }
2143 } else {
2144 th->th_seq = htonl(p->rxmit);
2145 if (len > 0) {
2146 m->m_pkthdr.pkt_flags |=
2147 (PKTF_TCP_REXMT | PKTF_START_SEQ);
2148 m->m_pkthdr.tx_start_seq = p->rxmit;
2149 }
2150 tcp_rxtseg_insert(tp, p->rxmit, (p->rxmit + len - 1));
2151 p->rxmit += len;
2152 tp->sackhint.sack_bytes_rexmit += len;
2153 }
2154 th->th_ack = htonl(tp->rcv_nxt);
2155 tp->last_ack_sent = tp->rcv_nxt;
2156 if (optlen) {
2157 bcopy(opt, th + 1, optlen);
2158 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
2159 }
2160 th->th_flags = flags;
2161 th->th_win = htons((u_short) (recwin>>tp->rcv_scale));
2162 if (!(so->so_flags & SOF_MP_SUBFLOW)) {
2163 if (recwin > 0 && SEQ_LT(tp->rcv_adv, tp->rcv_nxt + recwin)) {
2164 tp->rcv_adv = tp->rcv_nxt + recwin;
2165 }
2166 } else {
2167 struct mptcb *mp_tp = tptomptp(tp);
2168 if (recwin > 0) {
2169 tp->rcv_adv = tp->rcv_nxt + recwin;
2170 }
2171
2172 if (recwin > 0 && SEQ_LT(mp_tp->mpt_rcvadv, (uint32_t)mp_tp->mpt_rcvnxt + recwin)) {
2173 mp_tp->mpt_rcvadv = (uint32_t)mp_tp->mpt_rcvnxt + recwin;
2174 }
2175 }
2176
2177 /*
2178 * Adjust the RXWIN0SENT flag - indicate that we have advertised
2179 * a 0 window. This may cause the remote transmitter to stall. This
2180 * flag tells soreceive() to disable delayed acknowledgements when
2181 * draining the buffer. This can occur if the receiver is attempting
2182 * to read more data then can be buffered prior to transmitting on
2183 * the connection.
2184 */
2185 if (th->th_win == 0)
2186 tp->t_flags |= TF_RXWIN0SENT;
2187 else
2188 tp->t_flags &= ~TF_RXWIN0SENT;
2189
2190 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
2191 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
2192 th->th_flags |= TH_URG;
2193 } else {
2194 /*
2195 * If no urgent pointer to send, then we pull
2196 * the urgent pointer to the left edge of the send window
2197 * so that it doesn't drift into the send window on sequence
2198 * number wraparound.
2199 */
2200 tp->snd_up = tp->snd_una; /* drag it along */
2201 }
2202
2203 /*
2204 * Put TCP length in extended header, and then
2205 * checksum extended header and data.
2206 */
2207 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
2208
2209 /*
2210 * If this is potentially the last packet on the stream, then mark
2211 * it in order to enable some optimizations in the underlying
2212 * layers
2213 */
2214 if (tp->t_state != TCPS_ESTABLISHED &&
2215 (tp->t_state == TCPS_CLOSING || tp->t_state == TCPS_TIME_WAIT
2216 || tp->t_state == TCPS_LAST_ACK || (th->th_flags & TH_RST)))
2217 m->m_pkthdr.pkt_flags |= PKTF_LAST_PKT;
2218
2219 #if INET6
2220 if (isipv6) {
2221 /*
2222 * ip6_plen is not need to be filled now, and will be filled
2223 * in ip6_output.
2224 */
2225 m->m_pkthdr.csum_flags = CSUM_TCPIPV6;
2226 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2227 if (len + optlen)
2228 th->th_sum = in_addword(th->th_sum,
2229 htons((u_short)(optlen + len)));
2230 }
2231 else
2232 #endif /* INET6 */
2233 {
2234 m->m_pkthdr.csum_flags = CSUM_TCP;
2235 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2236 if (len + optlen)
2237 th->th_sum = in_addword(th->th_sum,
2238 htons((u_short)(optlen + len)));
2239 }
2240
2241 /*
2242 * Enable TSO and specify the size of the segments.
2243 * The TCP pseudo header checksum is always provided.
2244 */
2245 if (tso) {
2246 #if INET6
2247 if (isipv6)
2248 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV6;
2249 else
2250 #endif /* INET6 */
2251 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV4;
2252
2253 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
2254 } else {
2255 m->m_pkthdr.tso_segsz = 0;
2256 }
2257
2258 /*
2259 * In transmit state, time the transmission and arrange for
2260 * the retransmit. In persist state, just set snd_max.
2261 */
2262 if (!(tp->t_flagsext & TF_FORCE)
2263 || tp->t_timer[TCPT_PERSIST] == 0) {
2264 tcp_seq startseq = tp->snd_nxt;
2265
2266 /*
2267 * Advance snd_nxt over sequence space of this segment.
2268 */
2269 if (flags & (TH_SYN|TH_FIN)) {
2270 if (flags & TH_SYN)
2271 tp->snd_nxt++;
2272 if ((flags & TH_FIN) &&
2273 !(tp->t_flags & TF_SENTFIN)) {
2274 tp->snd_nxt++;
2275 tp->t_flags |= TF_SENTFIN;
2276 }
2277 }
2278 if (sack_rxmit)
2279 goto timer;
2280 if (sack_rescue_rxt == TRUE) {
2281 tp->snd_nxt = old_snd_nxt;
2282 sack_rescue_rxt = FALSE;
2283 tcpstat.tcps_pto_in_recovery++;
2284 } else {
2285 tp->snd_nxt += len;
2286 }
2287 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
2288 tp->snd_max = tp->snd_nxt;
2289 tp->t_sndtime = tcp_now;
2290 /*
2291 * Time this transmission if not a retransmission and
2292 * not currently timing anything.
2293 */
2294 if (tp->t_rtttime == 0) {
2295 tp->t_rtttime = tcp_now;
2296 tp->t_rtseq = startseq;
2297 tcpstat.tcps_segstimed++;
2298
2299 /* update variables related to pipe ack */
2300 tp->t_pipeack_lastuna = tp->snd_una;
2301 }
2302 }
2303
2304 /*
2305 * Set retransmit timer if not currently set,
2306 * and not doing an ack or a keep-alive probe.
2307 */
2308 timer:
2309 if (tp->t_timer[TCPT_REXMT] == 0 &&
2310 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
2311 tp->snd_nxt != tp->snd_una || (flags & TH_FIN))) {
2312 if (tp->t_timer[TCPT_PERSIST]) {
2313 tp->t_timer[TCPT_PERSIST] = 0;
2314 tp->t_persist_stop = 0;
2315 TCP_RESET_REXMT_STATE(tp);
2316 }
2317 tp->t_timer[TCPT_REXMT] =
2318 OFFSET_FROM_START(tp, tp->t_rxtcur);
2319 }
2320
2321 /*
2322 * Set tail loss probe timeout if new data is being
2323 * transmitted. This will be supported only when
2324 * SACK option is enabled on a connection.
2325 *
2326 * Every time new data is sent PTO will get reset.
2327 */
2328 if (tcp_enable_tlp && len != 0 && tp->t_state == TCPS_ESTABLISHED &&
2329 SACK_ENABLED(tp) && !IN_FASTRECOVERY(tp) &&
2330 tp->snd_nxt == tp->snd_max &&
2331 SEQ_GT(tp->snd_nxt, tp->snd_una) &&
2332 tp->t_rxtshift == 0 &&
2333 (tp->t_flagsext & (TF_SENT_TLPROBE|TF_PKTS_REORDERED)) == 0) {
2334 u_int32_t pto, srtt;
2335
2336 /*
2337 * Using SRTT alone to set PTO can cause spurious
2338 * retransmissions on wireless networks where there
2339 * is a lot of variance in RTT. Taking variance
2340 * into account will avoid this.
2341 */
2342 srtt = tp->t_srtt >> TCP_RTT_SHIFT;
2343 pto = ((TCP_REXMTVAL(tp)) * 3) >> 1;
2344 pto = max (2 * srtt, pto);
2345 if ((tp->snd_max - tp->snd_una) == tp->t_maxseg)
2346 pto = max(pto,
2347 (((3 * pto) >> 2) + tcp_delack * 2));
2348 else
2349 pto = max(10, pto);
2350
2351 /* if RTO is less than PTO, choose RTO instead */
2352 if (tp->t_rxtcur < pto)
2353 pto = tp->t_rxtcur;
2354
2355 tp->t_timer[TCPT_PTO] = OFFSET_FROM_START(tp, pto);
2356 }
2357 } else {
2358 /*
2359 * Persist case, update snd_max but since we are in
2360 * persist mode (no window) we do not update snd_nxt.
2361 */
2362 int xlen = len;
2363 if (flags & TH_SYN)
2364 ++xlen;
2365 if ((flags & TH_FIN) &&
2366 !(tp->t_flags & TF_SENTFIN)) {
2367 ++xlen;
2368 tp->t_flags |= TF_SENTFIN;
2369 }
2370 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) {
2371 tp->snd_max = tp->snd_nxt + len;
2372 tp->t_sndtime = tcp_now;
2373 }
2374 }
2375
2376 #if TCPDEBUG
2377 /*
2378 * Trace.
2379 */
2380 if (so_options & SO_DEBUG)
2381 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
2382 #endif
2383
2384 /*
2385 * Fill in IP length and desired time to live and
2386 * send to IP level. There should be a better way
2387 * to handle ttl and tos; we could keep them in
2388 * the template, but need a way to checksum without them.
2389 */
2390 #if INET6
2391 /*
2392 * m->m_pkthdr.len should have been set before cksum calcuration,
2393 * because in6_cksum() need it.
2394 */
2395 if (isipv6) {
2396 /*
2397 * we separately set hoplimit for every segment, since the
2398 * user might want to change the value via setsockopt.
2399 * Also, desired default hop limit might be changed via
2400 * Neighbor Discovery.
2401 */
2402 ip6->ip6_hlim = in6_selecthlim(inp, inp->in6p_route.ro_rt ?
2403 inp->in6p_route.ro_rt->rt_ifp : NULL);
2404
2405 /* TODO: IPv6 IP6TOS_ECT bit on */
2406 KERNEL_DEBUG(DBG_LAYER_BEG,
2407 ((inp->inp_fport << 16) | inp->inp_lport),
2408 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
2409 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
2410 sendalot,0,0);
2411 } else
2412 #endif /* INET6 */
2413 {
2414 ip->ip_len = m->m_pkthdr.len;
2415 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
2416 ip->ip_tos |= (inp->inp_ip_tos & ~IPTOS_ECN_MASK);/* XXX */
2417 KERNEL_DEBUG(DBG_LAYER_BEG,
2418 ((inp->inp_fport << 16) | inp->inp_lport),
2419 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
2420 (inp->inp_faddr.s_addr & 0xffff)), 0,0,0);
2421 }
2422
2423 /*
2424 * See if we should do MTU discovery.
2425 * Look at the flag updated on the following criterias:
2426 * 1) Path MTU discovery is authorized by the sysctl
2427 * 2) The route isn't set yet (unlikely but could happen)
2428 * 3) The route is up
2429 * 4) the MTU is not locked (if it is, then discovery has been
2430 * disabled for that route)
2431 */
2432 #if INET6
2433 if (!isipv6)
2434 #endif /* INET6 */
2435 if (path_mtu_discovery && (tp->t_flags & TF_PMTUD))
2436 ip->ip_off |= IP_DF;
2437
2438 #if NECP
2439 {
2440 necp_kernel_policy_id policy_id;
2441 necp_kernel_policy_id skip_policy_id;
2442 u_int32_t route_rule_id;
2443 if (!necp_socket_is_allowed_to_send_recv(inp, NULL, &policy_id, &route_rule_id, &skip_policy_id)) {
2444 TCP_LOG_DROP_NECP(isipv6 ? (void *)ip6 : (void *)ip, th, tp, true);
2445 m_freem(m);
2446 error = EHOSTUNREACH;
2447 goto out;
2448 }
2449 necp_mark_packet_from_socket(m, inp, policy_id, route_rule_id, skip_policy_id);
2450
2451 if (net_qos_policy_restricted != 0) {
2452 necp_socket_update_qos_marking(inp, inp->inp_route.ro_rt,
2453 NULL, route_rule_id);
2454 }
2455 }
2456 #endif /* NECP */
2457
2458 #if IPSEC
2459 if (inp->inp_sp != NULL)
2460 ipsec_setsocket(m, so);
2461 #endif /*IPSEC*/
2462
2463 /*
2464 * The socket is kept locked while sending out packets in ip_output, even if packet chaining is not active.
2465 */
2466 lost = 0;
2467
2468 /*
2469 * Embed the flow hash in pkt hdr and mark the packet as
2470 * capable of flow controlling
2471 */
2472 m->m_pkthdr.pkt_flowsrc = FLOWSRC_INPCB;
2473 m->m_pkthdr.pkt_flowid = inp->inp_flowhash;
2474 m->m_pkthdr.pkt_flags |= (PKTF_FLOW_ID | PKTF_FLOW_LOCALSRC | PKTF_FLOW_ADV);
2475 m->m_pkthdr.pkt_proto = IPPROTO_TCP;
2476 m->m_pkthdr.tx_tcp_pid = so->last_pid;
2477 if (so->so_flags & SOF_DELEGATED)
2478 m->m_pkthdr.tx_tcp_e_pid = so->e_pid;
2479 else
2480 m->m_pkthdr.tx_tcp_e_pid = 0;
2481
2482 m->m_nextpkt = NULL;
2483
2484 if (inp->inp_last_outifp != NULL &&
2485 !(inp->inp_last_outifp->if_flags & IFF_LOOPBACK)) {
2486 /* Hint to prioritize this packet if
2487 * 1. if the packet has no data
2488 * 2. the interface supports transmit-start model and did
2489 * not disable ACK prioritization.
2490 * 3. Only ACK flag is set.
2491 * 4. there is no outstanding data on this connection.
2492 */
2493 if (tcp_prioritize_acks != 0 && len == 0 &&
2494 (inp->inp_last_outifp->if_eflags &
2495 (IFEF_TXSTART | IFEF_NOACKPRI)) == IFEF_TXSTART) {
2496 if (th->th_flags == TH_ACK &&
2497 tp->snd_una == tp->snd_max &&
2498 tp->t_timer[TCPT_REXMT] == 0)
2499 svc_flags |= PKT_SCF_TCP_ACK;
2500 if (th->th_flags & TH_SYN)
2501 svc_flags |= PKT_SCF_TCP_SYN;
2502 }
2503 set_packet_service_class(m, so, sotc, svc_flags);
2504 } else {
2505 /*
2506 * Optimization for loopback just set the mbuf
2507 * service class
2508 */
2509 (void) m_set_service_class(m, so_tc2msc(sotc));
2510 }
2511
2512 TCP_LOG_TH_FLAGS(isipv6 ? (void *)ip6 : (void *)ip, th, tp, true,
2513 inp->inp_last_outifp != NULL ? inp->inp_last_outifp :
2514 inp->inp_boundifp);
2515
2516 tp->t_pktlist_sentlen += len;
2517 tp->t_lastchain++;
2518
2519 #if INET6
2520 if (isipv6) {
2521 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2522 struct ip6 *, ip6, struct tcpcb *, tp, struct tcphdr *,
2523 th);
2524 } else
2525 #endif /* INET6 */
2526 {
2527 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2528 struct ip *, ip, struct tcpcb *, tp, struct tcphdr *, th);
2529 }
2530
2531 if (tp->t_pktlist_head != NULL) {
2532 tp->t_pktlist_tail->m_nextpkt = m;
2533 tp->t_pktlist_tail = m;
2534 } else {
2535 packchain_newlist++;
2536 tp->t_pktlist_head = tp->t_pktlist_tail = m;
2537 }
2538
2539 if (lro_ackmore && !sackoptlen && tp->t_timer[TCPT_PERSIST] == 0 &&
2540 (th->th_flags & TH_ACK) == TH_ACK && len == 0 &&
2541 tp->t_state == TCPS_ESTABLISHED) {
2542 /* For a pure ACK, see if you need to send more of them */
2543 mnext = tcp_send_lroacks(tp, m, th);
2544 if (mnext) {
2545 tp->t_pktlist_tail->m_nextpkt = mnext;
2546 if (mnext->m_nextpkt == NULL) {
2547 tp->t_pktlist_tail = mnext;
2548 tp->t_lastchain++;
2549 } else {
2550 struct mbuf *tail, *next;
2551 next = mnext->m_nextpkt;
2552 tail = next->m_nextpkt;
2553 while (tail) {
2554 next = tail;
2555 tail = tail->m_nextpkt;
2556 tp->t_lastchain++;
2557 }
2558 tp->t_pktlist_tail = next;
2559 }
2560 }
2561 }
2562
2563 if (sendalot == 0 || (tp->t_state != TCPS_ESTABLISHED) ||
2564 (tp->snd_cwnd <= (tp->snd_wnd / 8)) ||
2565 (tp->t_flags & TF_ACKNOW) ||
2566 (tp->t_flagsext & TF_FORCE) ||
2567 tp->t_lastchain >= tcp_packet_chaining) {
2568 error = 0;
2569 while (inp->inp_sndinprog_cnt == 0 &&
2570 tp->t_pktlist_head != NULL) {
2571 packetlist = tp->t_pktlist_head;
2572 packchain_listadd = tp->t_lastchain;
2573 packchain_sent++;
2574 lost = tp->t_pktlist_sentlen;
2575 TCP_PKTLIST_CLEAR(tp);
2576
2577 error = tcp_ip_output(so, tp, packetlist,
2578 packchain_listadd, tp_inp_options,
2579 (so_options & SO_DONTROUTE),
2580 (sack_rxmit || (sack_bytes_rxmt != 0)), isipv6);
2581 if (error) {
2582 /*
2583 * Take into account the rest of unsent
2584 * packets in the packet list for this tcp
2585 * into "lost", since we're about to free
2586 * the whole list below.
2587 */
2588 lost += tp->t_pktlist_sentlen;
2589 break;
2590 } else {
2591 lost = 0;
2592 }
2593 }
2594 /* tcp was closed while we were in ip; resume close */
2595 if (inp->inp_sndinprog_cnt == 0 &&
2596 (tp->t_flags & TF_CLOSING)) {
2597 tp->t_flags &= ~TF_CLOSING;
2598 (void) tcp_close(tp);
2599 return 0;
2600 }
2601 } else {
2602 error = 0;
2603 packchain_looped++;
2604 tcpstat.tcps_sndtotal++;
2605
2606 goto again;
2607 }
2608 if (error) {
2609 /*
2610 * Assume that the packets were lost, so back out the
2611 * sequence number advance, if any. Note that the "lost"
2612 * variable represents the amount of user data sent during
2613 * the recent call to ip_output_list() plus the amount of
2614 * user data in the packet list for this tcp at the moment.
2615 */
2616 if (!(tp->t_flagsext & TF_FORCE)
2617 || tp->t_timer[TCPT_PERSIST] == 0) {
2618 /*
2619 * No need to check for TH_FIN here because
2620 * the TF_SENTFIN flag handles that case.
2621 */
2622 if ((flags & TH_SYN) == 0) {
2623 if (sack_rxmit) {
2624 if (SEQ_GT((p->rxmit - lost),
2625 tp->snd_una)) {
2626 p->rxmit -= lost;
2627 } else {
2628 lost = p->rxmit - tp->snd_una;
2629 p->rxmit = tp->snd_una;
2630 }
2631 tp->sackhint.sack_bytes_rexmit -= lost;
2632 } else {
2633 if (SEQ_GT((tp->snd_nxt - lost),
2634 tp->snd_una))
2635 tp->snd_nxt -= lost;
2636 else
2637 tp->snd_nxt = tp->snd_una;
2638 }
2639 }
2640 }
2641 out:
2642 if (tp->t_pktlist_head != NULL)
2643 m_freem_list(tp->t_pktlist_head);
2644 TCP_PKTLIST_CLEAR(tp);
2645
2646 if (error == ENOBUFS) {
2647 /*
2648 * Set retransmit timer if not currently set
2649 * when we failed to send a segment that can be
2650 * retransmitted (i.e. not pure ack or rst)
2651 */
2652 if (tp->t_timer[TCPT_REXMT] == 0 &&
2653 tp->t_timer[TCPT_PERSIST] == 0 &&
2654 (len != 0 || (flags & (TH_SYN | TH_FIN)) != 0 ||
2655 so->so_snd.sb_cc > 0))
2656 tp->t_timer[TCPT_REXMT] =
2657 OFFSET_FROM_START(tp, tp->t_rxtcur);
2658 tp->snd_cwnd = tp->t_maxseg;
2659 tp->t_bytes_acked = 0;
2660 tcp_check_timer_state(tp);
2661 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2662
2663 tcp_ccdbg_trace(tp, NULL, TCP_CC_OUTPUT_ERROR);
2664 return 0;
2665 }
2666 if (error == EMSGSIZE) {
2667 /*
2668 * ip_output() will have already fixed the route
2669 * for us. tcp_mtudisc() will, as its last action,
2670 * initiate retransmission, so it is important to
2671 * not do so here.
2672 *
2673 * If TSO was active we either got an interface
2674 * without TSO capabilits or TSO was turned off.
2675 * Disable it for this connection as too and
2676 * immediatly retry with MSS sized segments generated
2677 * by this function.
2678 */
2679 if (tso)
2680 tp->t_flags &= ~TF_TSO;
2681
2682 tcp_mtudisc(inp, 0);
2683 tcp_check_timer_state(tp);
2684
2685 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2686 return 0;
2687 }
2688 /*
2689 * Unless this is due to interface restriction policy,
2690 * treat EHOSTUNREACH/ENETDOWN as a soft error.
2691 */
2692 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
2693 TCPS_HAVERCVDSYN(tp->t_state) &&
2694 !inp_restricted_send(inp, inp->inp_last_outifp)) {
2695 tp->t_softerror = error;
2696 error = 0;
2697 }
2698 tcp_check_timer_state(tp);
2699 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2700 return error;
2701 }
2702
2703 tcpstat.tcps_sndtotal++;
2704
2705 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,0,0,0,0,0);
2706 if (sendalot)
2707 goto again;
2708
2709 tcp_check_timer_state(tp);
2710
2711 return 0;
2712 }
2713
2714 static int
2715 tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt,
2716 int cnt, struct mbuf *opt, int flags, int sack_in_progress, boolean_t isipv6)
2717 {
2718 int error = 0;
2719 boolean_t chain;
2720 boolean_t unlocked = FALSE;
2721 boolean_t ifdenied = FALSE;
2722 struct inpcb *inp = tp->t_inpcb;
2723 struct ip_out_args ipoa;
2724 struct route ro;
2725 struct ifnet *outif = NULL;
2726
2727 bzero(&ipoa, sizeof(ipoa));
2728 ipoa.ipoa_boundif = IFSCOPE_NONE;
2729 ipoa.ipoa_flags = IPOAF_SELECT_SRCIF | IPOAF_BOUND_SRCADDR;
2730 ipoa.ipoa_sotc = SO_TC_UNSPEC;
2731 ipoa.ipoa_netsvctype = _NET_SERVICE_TYPE_UNSPEC;
2732 #if INET6
2733 struct ip6_out_args ip6oa;
2734 struct route_in6 ro6;
2735
2736 bzero(&ip6oa, sizeof(ip6oa));
2737 ip6oa.ip6oa_boundif = IFSCOPE_NONE;
2738 ip6oa.ip6oa_flags = IP6OAF_SELECT_SRCIF | IP6OAF_BOUND_SRCADDR;
2739 ip6oa.ip6oa_sotc = SO_TC_UNSPEC;
2740 ip6oa.ip6oa_netsvctype = _NET_SERVICE_TYPE_UNSPEC;
2741
2742 struct flowadv *adv =
2743 (isipv6 ? &ip6oa.ip6oa_flowadv : &ipoa.ipoa_flowadv);
2744 #else /* INET6 */
2745 struct flowadv *adv = &ipoa.ipoa_flowadv;
2746 #endif /* !INET6 */
2747
2748 /* If socket was bound to an ifindex, tell ip_output about it */
2749 if (inp->inp_flags & INP_BOUND_IF) {
2750 #if INET6
2751 if (isipv6) {
2752 ip6oa.ip6oa_boundif = inp->inp_boundifp->if_index;
2753 ip6oa.ip6oa_flags |= IP6OAF_BOUND_IF;
2754 } else
2755 #endif /* INET6 */
2756 {
2757 ipoa.ipoa_boundif = inp->inp_boundifp->if_index;
2758 ipoa.ipoa_flags |= IPOAF_BOUND_IF;
2759 }
2760 }
2761
2762 if (INP_NO_CELLULAR(inp)) {
2763 #if INET6
2764 if (isipv6)
2765 ip6oa.ip6oa_flags |= IP6OAF_NO_CELLULAR;
2766 else
2767 #endif /* INET6 */
2768 ipoa.ipoa_flags |= IPOAF_NO_CELLULAR;
2769 }
2770 if (INP_NO_EXPENSIVE(inp)) {
2771 #if INET6
2772 if (isipv6)
2773 ip6oa.ip6oa_flags |= IP6OAF_NO_EXPENSIVE;
2774 else
2775 #endif /* INET6 */
2776 ipoa.ipoa_flags |= IPOAF_NO_EXPENSIVE;
2777
2778 }
2779 if (INP_NO_CONSTRAINED(inp)) {
2780 #if INET6
2781 if (isipv6)
2782 ip6oa.ip6oa_flags |= IP6OAF_NO_CONSTRAINED;
2783 else
2784 #endif /* INET6 */
2785 ipoa.ipoa_flags |= IPOAF_NO_CONSTRAINED;
2786 }
2787 if (INP_AWDL_UNRESTRICTED(inp)) {
2788 #if INET6
2789 if (isipv6)
2790 ip6oa.ip6oa_flags |= IP6OAF_AWDL_UNRESTRICTED;
2791 else
2792 #endif /* INET6 */
2793 ipoa.ipoa_flags |= IPOAF_AWDL_UNRESTRICTED;
2794
2795 }
2796 #if INET6
2797 if (INP_INTCOPROC_ALLOWED(inp) && isipv6) {
2798 ip6oa.ip6oa_flags |= IP6OAF_INTCOPROC_ALLOWED;
2799 }
2800 if (isipv6) {
2801 ip6oa.ip6oa_sotc = so->so_traffic_class;
2802 ip6oa.ip6oa_netsvctype = so->so_netsvctype;
2803 } else
2804 #endif /* INET6 */
2805 {
2806 ipoa.ipoa_sotc = so->so_traffic_class;
2807 ipoa.ipoa_netsvctype = so->so_netsvctype;
2808 }
2809 if ((so->so_flags1 & SOF1_QOSMARKING_ALLOWED)) {
2810 #if INET6
2811 if (isipv6)
2812 ip6oa.ip6oa_flags |= IP6OAF_QOSMARKING_ALLOWED;
2813 else
2814 #endif /* INET6 */
2815 ipoa.ipoa_flags |= IPOAF_QOSMARKING_ALLOWED;
2816 }
2817 #if INET6
2818 if (isipv6)
2819 flags |= IPV6_OUTARGS;
2820 else
2821 #endif /* INET6 */
2822 flags |= IP_OUTARGS;
2823
2824 /* Copy the cached route and take an extra reference */
2825 #if INET6
2826 if (isipv6)
2827 in6p_route_copyout(inp, &ro6);
2828 else
2829 #endif /* INET6 */
2830 inp_route_copyout(inp, &ro);
2831
2832 /*
2833 * Make sure ACK/DELACK conditions are cleared before
2834 * we unlock the socket.
2835 */
2836 tp->last_ack_sent = tp->rcv_nxt;
2837 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
2838 tp->t_timer[TCPT_DELACK] = 0;
2839 tp->t_unacksegs = 0;
2840
2841 /* Increment the count of outstanding send operations */
2842 inp->inp_sndinprog_cnt++;
2843
2844 /*
2845 * If allowed, unlock TCP socket while in IP
2846 * but only if the connection is established and
2847 * in a normal mode where reentrancy on the tcpcb won't be
2848 * an issue:
2849 * - there is no SACK episode
2850 * - we're not in Fast Recovery mode
2851 * - if we're not sending from an upcall.
2852 */
2853 if (tcp_output_unlocked && !so->so_upcallusecount &&
2854 (tp->t_state == TCPS_ESTABLISHED) && (sack_in_progress == 0) &&
2855 !IN_FASTRECOVERY(tp) && !(so->so_flags & SOF_MP_SUBFLOW)) {
2856
2857 unlocked = TRUE;
2858 socket_unlock(so, 0);
2859 }
2860
2861 /*
2862 * Don't send down a chain of packets when:
2863 * - TCP chaining is disabled
2864 * - there is an IPsec rule set
2865 * - there is a non default rule set for the firewall
2866 */
2867
2868 chain = tcp_packet_chaining > 1
2869 #if IPSEC
2870 && ipsec_bypass
2871 #endif
2872 #if IPFIREWALL
2873 && (fw_enable == 0 || fw_bypass)
2874 #endif
2875 ; // I'm important, not extraneous
2876
2877 while (pkt != NULL) {
2878 struct mbuf *npkt = pkt->m_nextpkt;
2879
2880 if (!chain) {
2881 pkt->m_nextpkt = NULL;
2882 /*
2883 * If we are not chaining, make sure to set the packet
2884 * list count to 0 so that IP takes the right path;
2885 * this is important for cases such as IPsec where a
2886 * single mbuf might result in multiple mbufs as part
2887 * of the encapsulation. If a non-zero count is passed
2888 * down to IP, the head of the chain might change and
2889 * we could end up skipping it (thus generating bogus
2890 * packets). Fixing it in IP would be desirable, but
2891 * for now this would do it.
2892 */
2893 cnt = 0;
2894 }
2895 #if INET6
2896 if (isipv6) {
2897 error = ip6_output_list(pkt, cnt,
2898 inp->in6p_outputopts, &ro6, flags, NULL, NULL,
2899 &ip6oa);
2900 ifdenied = (ip6oa.ip6oa_retflags & IP6OARF_IFDENIED);
2901 } else {
2902 #endif /* INET6 */
2903 error = ip_output_list(pkt, cnt, opt, &ro, flags, NULL,
2904 &ipoa);
2905 ifdenied = (ipoa.ipoa_retflags & IPOARF_IFDENIED);
2906 }
2907
2908 if (chain || error) {
2909 /*
2910 * If we sent down a chain then we are done since
2911 * the callee had taken care of everything; else
2912 * we need to free the rest of the chain ourselves.
2913 */
2914 if (!chain)
2915 m_freem_list(npkt);
2916 break;
2917 }
2918 pkt = npkt;
2919 }
2920
2921 if (unlocked)
2922 socket_lock(so, 0);
2923
2924 /*
2925 * Enter flow controlled state if the connection is established
2926 * and is not in recovery. Flow control is allowed only if there
2927 * is outstanding data.
2928 *
2929 * A connection will enter suspended state even if it is in
2930 * recovery.
2931 */
2932 if (((adv->code == FADV_FLOW_CONTROLLED && !IN_FASTRECOVERY(tp)) ||
2933 adv->code == FADV_SUSPENDED) &&
2934 !(tp->t_flags & TF_CLOSING) &&
2935 tp->t_state == TCPS_ESTABLISHED &&
2936 SEQ_GT(tp->snd_max, tp->snd_una)) {
2937 int rc;
2938 rc = inp_set_fc_state(inp, adv->code);
2939
2940 if (rc == 1)
2941 tcp_ccdbg_trace(tp, NULL,
2942 ((adv->code == FADV_FLOW_CONTROLLED) ?
2943 TCP_CC_FLOW_CONTROL : TCP_CC_SUSPEND));
2944 }
2945
2946 /*
2947 * When an interface queue gets suspended, some of the
2948 * packets are dropped. Return ENOBUFS, to update the
2949 * pcb state.
2950 */
2951 if (adv->code == FADV_SUSPENDED)
2952 error = ENOBUFS;
2953
2954 VERIFY(inp->inp_sndinprog_cnt > 0);
2955 if ( --inp->inp_sndinprog_cnt == 0) {
2956 inp->inp_flags &= ~(INP_FC_FEEDBACK);
2957 if (inp->inp_sndingprog_waiters > 0) {
2958 wakeup(&inp->inp_sndinprog_cnt);
2959 }
2960 }
2961
2962 #if INET6
2963 if (isipv6) {
2964 /*
2965 * When an NECP IP tunnel policy forces the outbound interface,
2966 * ip6_output_list() informs the transport layer what is the actual
2967 * outgoing interface
2968 */
2969 if (ip6oa.ip6oa_flags & IP6OAF_BOUND_IF) {
2970 outif = ifindex2ifnet[ip6oa.ip6oa_boundif];
2971 } else if (ro6.ro_rt != NULL) {
2972 outif = ro6.ro_rt->rt_ifp;
2973 }
2974 } else
2975 #endif /* INET6 */
2976 if (ro.ro_rt != NULL)
2977 outif = ro.ro_rt->rt_ifp;
2978
2979 if (outif != NULL && outif != inp->inp_last_outifp) {
2980 /* Update the send byte count */
2981 if (so->so_snd.sb_cc > 0 && so->so_snd.sb_flags & SB_SNDBYTE_CNT) {
2982 inp_decr_sndbytes_total(so, so->so_snd.sb_cc);
2983 inp_decr_sndbytes_allunsent(so, tp->snd_una);
2984 so->so_snd.sb_flags &= ~SB_SNDBYTE_CNT;
2985 }
2986 inp->inp_last_outifp = outif;
2987
2988 }
2989
2990 if (error != 0 && ifdenied &&
2991 (INP_NO_CELLULAR(inp) || INP_NO_EXPENSIVE(inp) || INP_NO_CONSTRAINED(inp)))
2992 soevent(so,
2993 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_IFDENIED));
2994
2995 /* Synchronize cached PCB route & options */
2996 #if INET6
2997 if (isipv6)
2998 in6p_route_copyin(inp, &ro6);
2999 else
3000 #endif /* INET6 */
3001 inp_route_copyin(inp, &ro);
3002
3003 if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift == 0 &&
3004 tp->t_inpcb->inp_route.ro_rt != NULL) {
3005 /* If we found the route and there is an rtt on it
3006 * reset the retransmit timer
3007 */
3008 tcp_getrt_rtt(tp, tp->t_inpcb->in6p_route.ro_rt);
3009 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur);
3010 }
3011 return error;
3012 }
3013
3014 int tcptv_persmin_val = TCPTV_PERSMIN;
3015
3016 void
3017 tcp_setpersist(struct tcpcb *tp)
3018 {
3019 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
3020
3021 /* If a PERSIST_TIMER option was set we will limit the
3022 * time the persist timer will be active for that connection
3023 * in order to avoid DOS by using zero window probes.
3024 * see rdar://5805356
3025 */
3026
3027 if (tp->t_persist_timeout != 0 &&
3028 tp->t_timer[TCPT_PERSIST] == 0 &&
3029 tp->t_persist_stop == 0) {
3030 tp->t_persist_stop = tcp_now + tp->t_persist_timeout;
3031 }
3032
3033 /*
3034 * Start/restart persistance timer.
3035 */
3036 TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
3037 t * tcp_backoff[tp->t_rxtshift],
3038 tcptv_persmin_val, TCPTV_PERSMAX, 0);
3039 tp->t_timer[TCPT_PERSIST] = OFFSET_FROM_START(tp, tp->t_timer[TCPT_PERSIST]);
3040
3041 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
3042 tp->t_rxtshift++;
3043 }
3044
3045 /*
3046 * Send as many acks as data coalesced. Every other packet when stretch
3047 * ACK is not enabled. Every 8 packets, if stretch ACK is enabled.
3048 */
3049 static struct mbuf*
3050 tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th)
3051 {
3052 struct mbuf *mnext = NULL, *ack_chain = NULL, *tail = NULL;
3053 int count = 0;
3054 tcp_seq org_ack = ntohl(th->th_ack);
3055 tcp_seq prev_ack = 0;
3056 int tack_offset = 28; /* IPv6 and IP options not supported */
3057 int twin_offset = 34; /* IPv6 and IP options not supported */
3058 int ack_size = (tp->t_flags & TF_STRETCHACK) ?
3059 (maxseg_unacked * tp->t_maxseg) : (tp->t_maxseg << 1);
3060 int segs_acked = (tp->t_flags & TF_STRETCHACK) ? maxseg_unacked : 2;
3061 struct mbuf *prev_ack_pkt = NULL;
3062 struct socket *so = tp->t_inpcb->inp_socket;
3063 unsigned short winsz = ntohs(th->th_win);
3064 unsigned int scaled_win = winsz<<tp->rcv_scale;
3065 tcp_seq win_rtedge = org_ack + scaled_win;
3066
3067 count = tp->t_lropktlen/tp->t_maxseg;
3068
3069 prev_ack = (org_ack - tp->t_lropktlen) + ack_size;
3070 if (prev_ack < org_ack) {
3071 ack_chain = m_dup(m, M_DONTWAIT);
3072 if (ack_chain) {
3073 th->th_ack = htonl(prev_ack);
3074 /* Keep adv window constant for duplicated ACK packets */
3075 scaled_win = win_rtedge - prev_ack;
3076 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3077 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3078 th->th_win = htons(scaled_win>>tp->rcv_scale);
3079 if (lrodebug == 5) {
3080 printf("%s: win = %d winsz = %d sc = %d"
3081 " lro_len %d %d\n",
3082 __func__, scaled_win>>tp->rcv_scale, winsz,
3083 tp->rcv_scale, tp->t_lropktlen, count);
3084 }
3085 tail = ack_chain;
3086 count -= segs_acked; /* accounts for prev_ack packet */
3087 count = (count <= segs_acked) ? 0 : count - segs_acked;
3088 tcpstat.tcps_sndacks++;
3089 so_tc_update_stats(m, so, m_get_service_class(m));
3090 } else {
3091 return NULL;
3092 }
3093 }
3094 else {
3095 tp->t_lropktlen = 0;
3096 return NULL;
3097 }
3098
3099 prev_ack_pkt = ack_chain;
3100
3101 while (count > 0) {
3102 if ((prev_ack + ack_size) < org_ack) {
3103 prev_ack += ack_size;
3104 } else {
3105 /*
3106 * The last ACK sent must have the ACK number that TCP
3107 * thinks is the last sent ACK number.
3108 */
3109 prev_ack = org_ack;
3110 }
3111 mnext = m_dup(prev_ack_pkt, M_DONTWAIT);
3112 if (mnext) {
3113 /* Keep adv window constant for duplicated ACK packets */
3114 scaled_win = win_rtedge - prev_ack;
3115 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3116 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3117 winsz = htons(scaled_win>>tp->rcv_scale);
3118 if (lrodebug == 5) {
3119 printf("%s: winsz = %d ack %x count %d\n",
3120 __func__, scaled_win>>tp->rcv_scale,
3121 prev_ack, count);
3122 }
3123 bcopy(&winsz, mtod(prev_ack_pkt, caddr_t) + twin_offset, 2);
3124 HTONL(prev_ack);
3125 bcopy(&prev_ack, mtod(prev_ack_pkt, caddr_t) + tack_offset, 4);
3126 NTOHL(prev_ack);
3127 tail->m_nextpkt = mnext;
3128 tail = mnext;
3129 count -= segs_acked;
3130 tcpstat.tcps_sndacks++;
3131 so_tc_update_stats(m, so, m_get_service_class(m));
3132 } else {
3133 if (lrodebug == 5) {
3134 printf("%s: failed to alloc mbuf.\n", __func__);
3135 }
3136 break;
3137 }
3138 prev_ack_pkt = mnext;
3139 }
3140 tp->t_lropktlen = 0;
3141 return ack_chain;
3142 }
3143
3144 static int
3145 tcp_recv_throttle (struct tcpcb *tp)
3146 {
3147 uint32_t base_rtt, newsize;
3148 struct sockbuf *sbrcv = &tp->t_inpcb->inp_socket->so_rcv;
3149
3150 if (tcp_use_rtt_recvbg == 1 &&
3151 TSTMP_SUPPORTED(tp)) {
3152 /*
3153 * Timestamps are supported on this connection. Use
3154 * RTT to look for an increase in latency.
3155 */
3156
3157 /*
3158 * If the connection is already being throttled, leave it
3159 * in that state until rtt comes closer to base rtt
3160 */
3161 if (tp->t_flagsext & TF_RECV_THROTTLE)
3162 return 1;
3163
3164 base_rtt = get_base_rtt(tp);
3165
3166 if (base_rtt != 0 && tp->t_rttcur != 0) {
3167 /*
3168 * if latency increased on a background flow,
3169 * return 1 to start throttling.
3170 */
3171 if (tp->t_rttcur > (base_rtt + target_qdelay)) {
3172 tp->t_flagsext |= TF_RECV_THROTTLE;
3173 if (tp->t_recv_throttle_ts == 0)
3174 tp->t_recv_throttle_ts = tcp_now;
3175 /*
3176 * Reduce the recv socket buffer size to
3177 * minimize latecy.
3178 */
3179 if (sbrcv->sb_idealsize >
3180 tcp_recv_throttle_minwin) {
3181 newsize = sbrcv->sb_idealsize >> 1;
3182 /* Set a minimum of 16 K */
3183 newsize =
3184 max(newsize,
3185 tcp_recv_throttle_minwin);
3186 sbrcv->sb_idealsize = newsize;
3187 }
3188 return 1;
3189 } else {
3190 return 0;
3191 }
3192 }
3193 }
3194
3195 /*
3196 * Timestamps are not supported or there is no good RTT
3197 * measurement. Use IPDV in this case.
3198 */
3199 if (tp->acc_iaj > tcp_acc_iaj_react_limit)
3200 return 1;
3201
3202 return 0;
3203 }
mirror of XNU