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


75e8634c0bf63d96395584e7f4e4022f0840fe08
[apple/xnu.git] / bsd / netinet / tcp_output.c
1 /*
2 * Copyright (c) 2000-2018 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 <sys/kdebug.h>
116 #include <mach/sdt.h>
117
118 #if IPSEC
119 #include <netinet6/ipsec.h>
120 #endif /*IPSEC*/
121
122 #if CONFIG_MACF_NET
123 #include <security/mac_framework.h>
124 #endif /* MAC_SOCKET */
125
126 #include <netinet/lro_ext.h>
127 #if MPTCP
128 #include <netinet/mptcp_var.h>
129 #include <netinet/mptcp.h>
130 #include <netinet/mptcp_opt.h>
131 #endif
132
133 #include <corecrypto/ccaes.h>
134
135 #define DBG_LAYER_BEG NETDBG_CODE(DBG_NETTCP, 1)
136 #define DBG_LAYER_END NETDBG_CODE(DBG_NETTCP, 3)
137 #define DBG_FNC_TCP_OUTPUT NETDBG_CODE(DBG_NETTCP, (4 << 8) | 1)
138
139 SYSCTL_SKMEM_TCP_INT(OID_AUTO, path_mtu_discovery,
140 CTLFLAG_RW | CTLFLAG_LOCKED, int, path_mtu_discovery, 1,
141 "Enable Path MTU Discovery");
142
143 SYSCTL_SKMEM_TCP_INT(OID_AUTO, slowstart_flightsize,
144 CTLFLAG_RW | CTLFLAG_LOCKED, int, ss_fltsz, 1,
145 "Slow start flight size");
146
147 SYSCTL_SKMEM_TCP_INT(OID_AUTO, local_slowstart_flightsize,
148 CTLFLAG_RW | CTLFLAG_LOCKED, int, ss_fltsz_local, 8,
149 "Slow start flight size for local networks");
150
151 int tcp_do_tso = 1;
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW | CTLFLAG_LOCKED,
153 &tcp_do_tso, 0, "Enable TCP Segmentation Offload");
154
155 SYSCTL_SKMEM_TCP_INT(OID_AUTO, ecn_setup_percentage,
156 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_ecn_setup_percentage, 100,
157 "Max ECN setup percentage");
158
159 static int
160 sysctl_change_ecn_setting SYSCTL_HANDLER_ARGS
161 {
162 #pragma unused(oidp, arg1, arg2)
163 int i, err = 0, changed = 0;
164 struct ifnet *ifp;
165
166 err = sysctl_io_number(req, tcp_ecn_outbound, sizeof(int32_t),
167 &i, &changed);
168 if (err != 0 || req->newptr == USER_ADDR_NULL)
169 return(err);
170
171 if (changed) {
172 if ((tcp_ecn_outbound == 0 || tcp_ecn_outbound == 1) &&
173 (i == 0 || i == 1)) {
174 tcp_ecn_outbound = i;
175 SYSCTL_SKMEM_UPDATE_FIELD(tcp.ecn_initiate_out, tcp_ecn_outbound);
176 return(err);
177 }
178 if (tcp_ecn_outbound == 2 && (i == 0 || i == 1)) {
179 /*
180 * Reset ECN enable flags on non-cellular
181 * interfaces so that the system default will take
182 * over
183 */
184 ifnet_head_lock_shared();
185 TAILQ_FOREACH(ifp, &ifnet_head, if_link) {
186 if (!IFNET_IS_CELLULAR(ifp)) {
187 ifnet_lock_exclusive(ifp);
188 ifp->if_eflags &= ~IFEF_ECN_DISABLE;
189 ifp->if_eflags &= ~IFEF_ECN_ENABLE;
190 ifnet_lock_done(ifp);
191 }
192 }
193 ifnet_head_done();
194 } else {
195 /*
196 * Set ECN enable flags on non-cellular
197 * interfaces
198 */
199 ifnet_head_lock_shared();
200 TAILQ_FOREACH(ifp, &ifnet_head, if_link) {
201 if (!IFNET_IS_CELLULAR(ifp)) {
202 ifnet_lock_exclusive(ifp);
203 ifp->if_eflags |= IFEF_ECN_ENABLE;
204 ifp->if_eflags &= ~IFEF_ECN_DISABLE;
205 ifnet_lock_done(ifp);
206 }
207 }
208 ifnet_head_done();
209 }
210 tcp_ecn_outbound = i;
211 SYSCTL_SKMEM_UPDATE_FIELD(tcp.ecn_initiate_out, tcp_ecn_outbound);
212 }
213 /* Change the other one too as the work is done */
214 if (i == 2 || tcp_ecn_inbound == 2) {
215 tcp_ecn_inbound = i;
216 SYSCTL_SKMEM_UPDATE_FIELD(tcp.ecn_negotiate_in, tcp_ecn_inbound);
217 }
218 return (err);
219 }
220
221 int tcp_ecn_outbound = 2;
222 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, ecn_initiate_out,
223 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_outbound, 0,
224 sysctl_change_ecn_setting, "IU",
225 "Initiate ECN for outbound connections");
226
227 int tcp_ecn_inbound = 2;
228 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, ecn_negotiate_in,
229 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_inbound, 0,
230 sysctl_change_ecn_setting, "IU",
231 "Initiate ECN for inbound connections");
232
233 SYSCTL_SKMEM_TCP_INT(OID_AUTO, packetchain,
234 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_packet_chaining, 50,
235 "Enable TCP output packet chaining");
236
237 SYSCTL_SKMEM_TCP_INT(OID_AUTO, socket_unlocked_on_output,
238 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_output_unlocked, 1,
239 "Unlock TCP when sending packets down to IP");
240
241 SYSCTL_SKMEM_TCP_INT(OID_AUTO, rfc3390,
242 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_do_rfc3390, 1,
243 "Calculate intial slowstart cwnd depending on MSS");
244
245 SYSCTL_SKMEM_TCP_INT(OID_AUTO, min_iaj_win,
246 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_min_iaj_win, MIN_IAJ_WIN,
247 "Minimum recv win based on inter-packet arrival jitter");
248
249 SYSCTL_SKMEM_TCP_INT(OID_AUTO, acc_iaj_react_limit,
250 CTLFLAG_RW | CTLFLAG_LOCKED, int, tcp_acc_iaj_react_limit,
251 ACC_IAJ_REACT_LIMIT, "Accumulated IAJ when receiver starts to react");
252
253 SYSCTL_SKMEM_TCP_INT(OID_AUTO, doautosndbuf,
254 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_do_autosendbuf, 1,
255 "Enable send socket buffer auto-tuning");
256
257 SYSCTL_SKMEM_TCP_INT(OID_AUTO, autosndbufinc,
258 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_autosndbuf_inc,
259 8 * 1024, "Increment in send socket bufffer size");
260
261 SYSCTL_SKMEM_TCP_INT(OID_AUTO, autosndbufmax,
262 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_autosndbuf_max, 512 * 1024,
263 "Maximum send socket buffer size");
264
265 SYSCTL_SKMEM_TCP_INT(OID_AUTO, ack_prioritize,
266 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_prioritize_acks, 1,
267 "Prioritize pure acks");
268
269 SYSCTL_SKMEM_TCP_INT(OID_AUTO, rtt_recvbg,
270 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_use_rtt_recvbg, 1,
271 "Use RTT for bg recv algorithm");
272
273 SYSCTL_SKMEM_TCP_INT(OID_AUTO, recv_throttle_minwin,
274 CTLFLAG_RW | CTLFLAG_LOCKED, uint32_t, tcp_recv_throttle_minwin, 16 * 1024,
275 "Minimum recv win for throttling");
276
277 SYSCTL_SKMEM_TCP_INT(OID_AUTO, enable_tlp,
278 CTLFLAG_RW | CTLFLAG_LOCKED,
279 int32_t, tcp_enable_tlp, 1, "Enable Tail loss probe");
280
281 static int32_t packchain_newlist = 0;
282 static int32_t packchain_looped = 0;
283 static int32_t packchain_sent = 0;
284
285 /* temporary: for testing */
286 #if IPSEC
287 extern int ipsec_bypass;
288 #endif
289
290 extern int slowlink_wsize; /* window correction for slow links */
291 #if IPFIREWALL
292 extern int fw_enable; /* firewall check for packet chaining */
293 extern int fw_bypass; /* firewall check: disable packet chaining if there is rules */
294 #endif /* IPFIREWALL */
295
296 extern u_int32_t dlil_filter_disable_tso_count;
297 extern u_int32_t kipf_count;
298
299 static int tcp_ip_output(struct socket *, struct tcpcb *, struct mbuf *,
300 int, struct mbuf *, int, int, boolean_t);
301 static struct mbuf* tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th);
302 static int tcp_recv_throttle(struct tcpcb *tp);
303
304 static int32_t tcp_tfo_check(struct tcpcb *tp, int32_t len)
305 {
306 struct socket *so = tp->t_inpcb->inp_socket;
307 unsigned int optlen = 0;
308 unsigned int cookie_len;
309
310 if (tp->t_flags & TF_NOOPT)
311 goto fallback;
312
313 if ((so->so_flags1 & SOF1_DATA_AUTHENTICATED) &&
314 !(tp->t_flagsext & TF_FASTOPEN_HEUR))
315 return (len);
316
317 if (!tcp_heuristic_do_tfo(tp)) {
318 tp->t_tfo_stats |= TFO_S_HEURISTICS_DISABLE;
319 tcpstat.tcps_tfo_heuristics_disable++;
320 goto fallback;
321 }
322
323 if (so->so_flags1 & SOF1_DATA_AUTHENTICATED)
324 return (len);
325
326 optlen += TCPOLEN_MAXSEG;
327
328 if (tp->t_flags & TF_REQ_SCALE)
329 optlen += 4;
330
331 #if MPTCP
332 if ((so->so_flags & SOF_MP_SUBFLOW) && mptcp_enable &&
333 tp->t_rxtshift <= mptcp_mpcap_retries)
334 optlen += sizeof(struct mptcp_mpcapable_opt_common) + sizeof(mptcp_key_t);
335 #endif /* MPTCP */
336
337 if (tp->t_flags & TF_REQ_TSTMP)
338 optlen += TCPOLEN_TSTAMP_APPA;
339
340 if (SACK_ENABLED(tp))
341 optlen += TCPOLEN_SACK_PERMITTED;
342
343 /* Now, decide whether to use TFO or not */
344
345 /* Don't even bother trying if there is no space at all... */
346 if (MAX_TCPOPTLEN - optlen < TCPOLEN_FASTOPEN_REQ)
347 goto fallback;
348
349 cookie_len = tcp_cache_get_cookie_len(tp);
350 if (cookie_len == 0)
351 /* No cookie, so we request one */
352 return (0);
353
354 /* There is not enough space for the cookie, so we cannot do TFO */
355 if (MAX_TCPOPTLEN - optlen < cookie_len)
356 goto fallback;
357
358 /* Do not send SYN+data if there is more in the queue than MSS */
359 if (so->so_snd.sb_cc > (tp->t_maxopd - MAX_TCPOPTLEN))
360 goto fallback;
361
362 /* Ok, everything looks good. We can go on and do TFO */
363 return (len);
364
365 fallback:
366 tp->t_flagsext &= ~TF_FASTOPEN;
367 return (0);
368 }
369
370 /* Returns the number of bytes written to the TCP option-space */
371 static unsigned
372 tcp_tfo_write_cookie_rep(struct tcpcb *tp, unsigned optlen, u_char *opt)
373 {
374 u_char out[CCAES_BLOCK_SIZE];
375 unsigned ret = 0;
376 u_char *bp;
377
378 if ((MAX_TCPOPTLEN - optlen) <
379 (TCPOLEN_FASTOPEN_REQ + TFO_COOKIE_LEN_DEFAULT))
380 return (ret);
381
382 tcp_tfo_gen_cookie(tp->t_inpcb, out, sizeof(out));
383
384 bp = opt + optlen;
385
386 *bp++ = TCPOPT_FASTOPEN;
387 *bp++ = 2 + TFO_COOKIE_LEN_DEFAULT;
388 memcpy(bp, out, TFO_COOKIE_LEN_DEFAULT);
389 ret += 2 + TFO_COOKIE_LEN_DEFAULT;
390
391 tp->t_tfo_stats |= TFO_S_COOKIE_SENT;
392 tcpstat.tcps_tfo_cookie_sent++;
393
394 return (ret);
395 }
396
397 static unsigned
398 tcp_tfo_write_cookie(struct tcpcb *tp, unsigned optlen, int32_t len,
399 u_char *opt)
400 {
401 u_int8_t tfo_len = MAX_TCPOPTLEN - optlen - TCPOLEN_FASTOPEN_REQ;
402 struct socket *so = tp->t_inpcb->inp_socket;
403 unsigned ret = 0;
404 int res;
405 u_char *bp;
406
407 if (so->so_flags1 & SOF1_DATA_AUTHENTICATED) {
408 /* If there is some data, let's track it */
409 if (len > 0) {
410 tp->t_tfo_stats |= TFO_S_SYN_DATA_SENT;
411 tcpstat.tcps_tfo_syn_data_sent++;
412 }
413
414 return (0);
415 }
416
417 bp = opt + optlen;
418
419 /*
420 * The cookie will be copied in the appropriate place within the
421 * TCP-option space. That way we avoid the need for an intermediate
422 * variable.
423 */
424 res = tcp_cache_get_cookie(tp, bp + TCPOLEN_FASTOPEN_REQ, &tfo_len);
425 if (res == 0) {
426 *bp++ = TCPOPT_FASTOPEN;
427 *bp++ = TCPOLEN_FASTOPEN_REQ;
428 ret += TCPOLEN_FASTOPEN_REQ;
429
430 tp->t_tfo_flags |= TFO_F_COOKIE_REQ;
431
432 tp->t_tfo_stats |= TFO_S_COOKIE_REQ;
433 tcpstat.tcps_tfo_cookie_req++;
434 } else {
435 *bp++ = TCPOPT_FASTOPEN;
436 *bp++ = TCPOLEN_FASTOPEN_REQ + tfo_len;
437
438 ret += TCPOLEN_FASTOPEN_REQ + tfo_len;
439
440 tp->t_tfo_flags |= TFO_F_COOKIE_SENT;
441
442 /* If there is some data, let's track it */
443 if (len > 0) {
444 tp->t_tfo_stats |= TFO_S_SYN_DATA_SENT;
445 tcpstat.tcps_tfo_syn_data_sent++;
446 }
447 }
448
449 return (ret);
450 }
451
452 static inline bool
453 tcp_send_ecn_flags_on_syn(struct tcpcb *tp, struct socket *so)
454 {
455 return(!((tp->ecn_flags & TE_SETUPSENT) ||
456 (so->so_flags & SOF_MP_SUBFLOW) ||
457 (tp->t_flagsext & TF_FASTOPEN)));
458 }
459
460 void
461 tcp_set_ecn(struct tcpcb *tp, struct ifnet *ifp)
462 {
463 boolean_t inbound;
464
465 /*
466 * Socket option has precedence
467 */
468 if (tp->ecn_flags & TE_ECN_MODE_ENABLE) {
469 tp->ecn_flags |= TE_ENABLE_ECN;
470 goto check_heuristic;
471 }
472
473 if (tp->ecn_flags & TE_ECN_MODE_DISABLE) {
474 tp->ecn_flags &= ~TE_ENABLE_ECN;
475 return;
476 }
477 /*
478 * Per interface setting comes next
479 */
480 if (ifp != NULL) {
481 if (ifp->if_eflags & IFEF_ECN_ENABLE) {
482 tp->ecn_flags |= TE_ENABLE_ECN;
483 goto check_heuristic;
484 }
485
486 if (ifp->if_eflags & IFEF_ECN_DISABLE) {
487 tp->ecn_flags &= ~TE_ENABLE_ECN;
488 return;
489 }
490 }
491 /*
492 * System wide settings come last
493 */
494 inbound = (tp->t_inpcb->inp_socket->so_head != NULL);
495 if ((inbound && tcp_ecn_inbound == 1) ||
496 (!inbound && tcp_ecn_outbound == 1)) {
497 tp->ecn_flags |= TE_ENABLE_ECN;
498 goto check_heuristic;
499 } else {
500 tp->ecn_flags &= ~TE_ENABLE_ECN;
501 }
502
503 return;
504
505 check_heuristic:
506 if (!tcp_heuristic_do_ecn(tp))
507 tp->ecn_flags &= ~TE_ENABLE_ECN;
508
509 /*
510 * If the interface setting, system-level setting and heuristics
511 * allow to enable ECN, randomly select 5% of connections to
512 * enable it
513 */
514 if ((tp->ecn_flags & (TE_ECN_MODE_ENABLE | TE_ECN_MODE_DISABLE
515 | TE_ENABLE_ECN)) == TE_ENABLE_ECN) {
516 /*
517 * Use the random value in iss for randomizing
518 * this selection
519 */
520 if ((tp->iss % 100) >= tcp_ecn_setup_percentage)
521 tp->ecn_flags &= ~TE_ENABLE_ECN;
522 }
523 }
524
525 /*
526 * Tcp output routine: figure out what should be sent and send it.
527 *
528 * Returns: 0 Success
529 * EADDRNOTAVAIL
530 * ENOBUFS
531 * EMSGSIZE
532 * EHOSTUNREACH
533 * ENETDOWN
534 * ip_output_list:ENOMEM
535 * ip_output_list:EADDRNOTAVAIL
536 * ip_output_list:ENETUNREACH
537 * ip_output_list:EHOSTUNREACH
538 * ip_output_list:EACCES
539 * ip_output_list:EMSGSIZE
540 * ip_output_list:ENOBUFS
541 * ip_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL]
542 * ip6_output_list:EINVAL
543 * ip6_output_list:EOPNOTSUPP
544 * ip6_output_list:EHOSTUNREACH
545 * ip6_output_list:EADDRNOTAVAIL
546 * ip6_output_list:ENETUNREACH
547 * ip6_output_list:EMSGSIZE
548 * ip6_output_list:ENOBUFS
549 * ip6_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL]
550 */
551 int
552 tcp_output(struct tcpcb *tp)
553 {
554 struct inpcb *inp = tp->t_inpcb;
555 struct socket *so = inp->inp_socket;
556 int32_t len, recwin, sendwin, off;
557 int flags, error;
558 struct mbuf *m;
559 struct ip *ip = NULL;
560 struct ipovly *ipov = NULL;
561 #if INET6
562 struct ip6_hdr *ip6 = NULL;
563 #endif /* INET6 */
564 struct tcphdr *th;
565 u_char opt[TCP_MAXOLEN];
566 unsigned ipoptlen, optlen, hdrlen;
567 int idle, sendalot, lost = 0;
568 int i, sack_rxmit;
569 int tso = 0;
570 int sack_bytes_rxmt;
571 tcp_seq old_snd_nxt = 0;
572 struct sackhole *p;
573 #if IPSEC
574 unsigned ipsec_optlen = 0;
575 #endif /* IPSEC */
576 int idle_time = 0;
577 struct mbuf *packetlist = NULL;
578 struct mbuf *tp_inp_options = inp->inp_depend4.inp4_options;
579 #if INET6
580 int isipv6 = inp->inp_vflag & INP_IPV6 ;
581 #else
582 int isipv6 = 0;
583 #endif
584 short packchain_listadd = 0;
585 int so_options = so->so_options;
586 struct rtentry *rt;
587 u_int32_t svc_flags = 0, allocated_len;
588 u_int32_t lro_ackmore = (tp->t_lropktlen != 0) ? 1 : 0;
589 struct mbuf *mnext = NULL;
590 int sackoptlen = 0;
591 #if MPTCP
592 boolean_t mptcp_acknow;
593 #endif /* MPTCP */
594 boolean_t cell = FALSE;
595 boolean_t wifi = FALSE;
596 boolean_t wired = FALSE;
597 boolean_t sack_rescue_rxt = FALSE;
598 int sotc = so->so_traffic_class;
599
600 /*
601 * Determine length of data that should be transmitted,
602 * and flags that will be used.
603 * If there is some data or critical controls (SYN, RST)
604 * to send, then transmit; otherwise, investigate further.
605 */
606 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
607
608 /* Since idle_time is signed integer, the following integer subtraction
609 * will take care of wrap around of tcp_now
610 */
611 idle_time = tcp_now - tp->t_rcvtime;
612 if (idle && idle_time >= TCP_IDLETIMEOUT(tp)) {
613 if (CC_ALGO(tp)->after_idle != NULL &&
614 (tp->tcp_cc_index != TCP_CC_ALGO_CUBIC_INDEX ||
615 idle_time >= TCP_CC_CWND_NONVALIDATED_PERIOD)) {
616 CC_ALGO(tp)->after_idle(tp);
617 tcp_ccdbg_trace(tp, NULL, TCP_CC_IDLE_TIMEOUT);
618 }
619
620 /*
621 * Do some other tasks that need to be done after
622 * idle time
623 */
624 if (!SLIST_EMPTY(&tp->t_rxt_segments))
625 tcp_rxtseg_clean(tp);
626
627 /* If stretch ack was auto-disabled, re-evaluate it */
628 tcp_cc_after_idle_stretchack(tp);
629 }
630 tp->t_flags &= ~TF_LASTIDLE;
631 if (idle) {
632 if (tp->t_flags & TF_MORETOCOME) {
633 tp->t_flags |= TF_LASTIDLE;
634 idle = 0;
635 }
636 }
637 #if MPTCP
638 if (tp->t_mpflags & TMPF_RESET) {
639 tcp_check_timer_state(tp);
640 /*
641 * Once a RST has been sent for an MPTCP subflow,
642 * the subflow socket stays around until deleted.
643 * No packets such as FINs must be sent after RST.
644 */
645 return (0);
646 }
647 #endif /* MPTCP */
648
649 again:
650 #if MPTCP
651 mptcp_acknow = FALSE;
652 #endif
653
654 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_START, 0,0,0,0,0);
655
656 #if INET6
657 if (isipv6) {
658 KERNEL_DEBUG(DBG_LAYER_BEG,
659 ((inp->inp_fport << 16) | inp->inp_lport),
660 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
661 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
662 sendalot,0,0);
663 } else
664 #endif
665
666 {
667 KERNEL_DEBUG(DBG_LAYER_BEG,
668 ((inp->inp_fport << 16) | inp->inp_lport),
669 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
670 (inp->inp_faddr.s_addr & 0xffff)),
671 sendalot,0,0);
672 }
673 /*
674 * If the route generation id changed, we need to check that our
675 * local (source) IP address is still valid. If it isn't either
676 * return error or silently do nothing (assuming the address will
677 * come back before the TCP connection times out).
678 */
679 rt = inp->inp_route.ro_rt;
680 if (rt != NULL && ROUTE_UNUSABLE(&tp->t_inpcb->inp_route)) {
681 struct ifnet *ifp;
682 struct in_ifaddr *ia = NULL;
683 struct in6_ifaddr *ia6 = NULL;
684 int found_srcaddr = 0;
685
686 /* disable multipages at the socket */
687 somultipages(so, FALSE);
688
689 /* Disable TSO for the socket until we know more */
690 tp->t_flags &= ~TF_TSO;
691
692 soif2kcl(so, FALSE);
693
694 if (isipv6) {
695 ia6 = ifa_foraddr6(&inp->in6p_laddr);
696 if (ia6 != NULL)
697 found_srcaddr = 1;
698 } else {
699 ia = ifa_foraddr(inp->inp_laddr.s_addr);
700 if (ia != NULL)
701 found_srcaddr = 1;
702 }
703
704 /* check that the source address is still valid */
705 if (found_srcaddr == 0) {
706 soevent(so,
707 (SO_FILT_HINT_LOCKED | SO_FILT_HINT_NOSRCADDR));
708
709 if (tp->t_state >= TCPS_CLOSE_WAIT) {
710 tcp_drop(tp, EADDRNOTAVAIL);
711 return(EADDRNOTAVAIL);
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
719 if (!tp->t_timer[TCPT_REXMT]) {
720 tp->t_timer[TCPT_REXMT] =
721 OFFSET_FROM_START(tp, tp->t_rxtcur);
722 if (tp->t_timer[TCPT_PERSIST]) {
723 tp->t_timer[TCPT_PERSIST] = 0;
724 tp->t_persist_stop = 0;
725 TCP_RESET_REXMT_STATE(tp);
726 }
727 }
728
729 if (tp->t_pktlist_head != NULL)
730 m_freem_list(tp->t_pktlist_head);
731 TCP_PKTLIST_CLEAR(tp);
732
733 /* drop connection if source address isn't available */
734 if (so->so_flags & SOF_NOADDRAVAIL) {
735 tcp_drop(tp, EADDRNOTAVAIL);
736 return(EADDRNOTAVAIL);
737 } else {
738 tcp_check_timer_state(tp);
739 return(0); /* silently ignore, keep data in socket: address may be back */
740 }
741 }
742 if (ia != NULL)
743 IFA_REMREF(&ia->ia_ifa);
744
745 if (ia6 != NULL)
746 IFA_REMREF(&ia6->ia_ifa);
747
748 /*
749 * Address is still valid; check for multipages capability
750 * again in case the outgoing interface has changed.
751 */
752 RT_LOCK(rt);
753 if ((ifp = rt->rt_ifp) != NULL) {
754 somultipages(so, (ifp->if_hwassist & IFNET_MULTIPAGES));
755 tcp_set_tso(tp, ifp);
756 soif2kcl(so, (ifp->if_eflags & IFEF_2KCL));
757 tcp_set_ecn(tp, ifp);
758 }
759 if (rt->rt_flags & RTF_UP)
760 RT_GENID_SYNC(rt);
761 /*
762 * See if we should do MTU discovery. Don't do it if:
763 * 1) it is disabled via the sysctl
764 * 2) the route isn't up
765 * 3) the MTU is locked (if it is, then discovery
766 * has been disabled)
767 */
768
769 if (!path_mtu_discovery || ((rt != NULL) &&
770 (!(rt->rt_flags & RTF_UP) ||
771 (rt->rt_rmx.rmx_locks & RTV_MTU))))
772 tp->t_flags &= ~TF_PMTUD;
773 else
774 tp->t_flags |= TF_PMTUD;
775
776 RT_UNLOCK(rt);
777 }
778
779 if (rt != NULL) {
780 cell = IFNET_IS_CELLULAR(rt->rt_ifp);
781 wifi = (!cell && IFNET_IS_WIFI(rt->rt_ifp));
782 wired = (!wifi && IFNET_IS_WIRED(rt->rt_ifp));
783 }
784
785 /*
786 * If we've recently taken a timeout, snd_max will be greater than
787 * snd_nxt. There may be SACK information that allows us to avoid
788 * resending already delivered data. Adjust snd_nxt accordingly.
789 */
790 if (SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
791 tcp_sack_adjust(tp);
792 sendalot = 0;
793 off = tp->snd_nxt - tp->snd_una;
794 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
795
796 if (tp->t_flags & TF_SLOWLINK && slowlink_wsize > 0)
797 sendwin = min(sendwin, slowlink_wsize);
798
799 flags = tcp_outflags[tp->t_state];
800 /*
801 * Send any SACK-generated retransmissions. If we're explicitly
802 * trying to send out new data (when sendalot is 1), bypass this
803 * function. If we retransmit in fast recovery mode, decrement
804 * snd_cwnd, since we're replacing a (future) new transmission
805 * with a retransmission now, and we previously incremented
806 * snd_cwnd in tcp_input().
807 */
808 /*
809 * Still in sack recovery , reset rxmit flag to zero.
810 */
811 sack_rxmit = 0;
812 sack_bytes_rxmt = 0;
813 len = 0;
814 p = NULL;
815 if (SACK_ENABLED(tp) && IN_FASTRECOVERY(tp) &&
816 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
817 int32_t cwin;
818
819 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
820 if (cwin < 0)
821 cwin = 0;
822 /* Do not retransmit SACK segments beyond snd_recover */
823 if (SEQ_GT(p->end, tp->snd_recover)) {
824 /*
825 * (At least) part of sack hole extends beyond
826 * snd_recover. Check to see if we can rexmit data
827 * for this hole.
828 */
829 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
830 /*
831 * Can't rexmit any more data for this hole.
832 * That data will be rexmitted in the next
833 * sack recovery episode, when snd_recover
834 * moves past p->rxmit.
835 */
836 p = NULL;
837 goto after_sack_rexmit;
838 } else
839 /* Can rexmit part of the current hole */
840 len = ((int32_t)min(cwin,
841 tp->snd_recover - p->rxmit));
842 } else {
843 len = ((int32_t)min(cwin, p->end - p->rxmit));
844 }
845 if (len > 0) {
846 off = p->rxmit - tp->snd_una;
847 sack_rxmit = 1;
848 sendalot = 1;
849 tcpstat.tcps_sack_rexmits++;
850 tcpstat.tcps_sack_rexmit_bytes +=
851 min(len, tp->t_maxseg);
852 } else {
853 len = 0;
854 }
855 }
856 after_sack_rexmit:
857 /*
858 * Get standard flags, and add SYN or FIN if requested by 'hidden'
859 * state flags.
860 */
861 if (tp->t_flags & TF_NEEDFIN)
862 flags |= TH_FIN;
863 if (tp->t_flags & TF_NEEDSYN)
864 flags |= TH_SYN;
865
866 /*
867 * If in persist timeout with window of 0, send 1 byte.
868 * Otherwise, if window is small but nonzero
869 * and timer expired, we will send what we can
870 * and go to transmit state.
871 */
872 if (tp->t_flagsext & TF_FORCE) {
873 if (sendwin == 0) {
874 /*
875 * If we still have some data to send, then
876 * clear the FIN bit. Usually this would
877 * happen below when it realizes that we
878 * aren't sending all the data. However,
879 * if we have exactly 1 byte of unsent data,
880 * then it won't clear the FIN bit below,
881 * and if we are in persist state, we wind
882 * up sending the packet without recording
883 * that we sent the FIN bit.
884 *
885 * We can't just blindly clear the FIN bit,
886 * because if we don't have any more data
887 * to send then the probe will be the FIN
888 * itself.
889 */
890 if (off < so->so_snd.sb_cc)
891 flags &= ~TH_FIN;
892 sendwin = 1;
893 } else {
894 tp->t_timer[TCPT_PERSIST] = 0;
895 tp->t_persist_stop = 0;
896 TCP_RESET_REXMT_STATE(tp);
897 }
898 }
899
900 /*
901 * If snd_nxt == snd_max and we have transmitted a FIN, the
902 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
903 * a negative length. This can also occur when TCP opens up
904 * its congestion window while receiving additional duplicate
905 * acks after fast-retransmit because TCP will reset snd_nxt
906 * to snd_max after the fast-retransmit.
907 *
908 * In the normal retransmit-FIN-only case, however, snd_nxt will
909 * be set to snd_una, the offset will be 0, and the length may
910 * wind up 0.
911 *
912 * If sack_rxmit is true we are retransmitting from the scoreboard
913 * in which case len is already set.
914 */
915 if (sack_rxmit == 0) {
916 if (sack_bytes_rxmt == 0) {
917 len = min(so->so_snd.sb_cc, sendwin) - off;
918 } else {
919 int32_t cwin;
920
921 cwin = tp->snd_cwnd -
922 (tp->snd_nxt - tp->sack_newdata) -
923 sack_bytes_rxmt;
924 if (cwin < 0)
925 cwin = 0;
926 /*
927 * We are inside of a SACK recovery episode and are
928 * sending new data, having retransmitted all the
929 * data possible in the scoreboard.
930 */
931 len = min(so->so_snd.sb_cc, tp->snd_wnd)
932 - off;
933 /*
934 * Don't remove this (len > 0) check !
935 * We explicitly check for len > 0 here (although it
936 * isn't really necessary), to work around a gcc
937 * optimization issue - to force gcc to compute
938 * len above. Without this check, the computation
939 * of len is bungled by the optimizer.
940 */
941 if (len > 0) {
942 len = imin(len, cwin);
943 } else {
944 len = 0;
945 }
946 /*
947 * At this point SACK recovery can not send any
948 * data from scoreboard or any new data. Check
949 * if we can do a rescue retransmit towards the
950 * tail end of recovery window.
951 */
952 if (len == 0 && cwin > 0 &&
953 SEQ_LT(tp->snd_fack, tp->snd_recover) &&
954 !(tp->t_flagsext & TF_RESCUE_RXT)) {
955 len = min((tp->snd_recover - tp->snd_fack),
956 tp->t_maxseg);
957 len = imin(len, cwin);
958 old_snd_nxt = tp->snd_nxt;
959 sack_rescue_rxt = TRUE;
960 tp->snd_nxt = tp->snd_recover - len;
961 /*
962 * If FIN has been sent, snd_max
963 * must have been advanced to cover it.
964 */
965 if ((tp->t_flags & TF_SENTFIN) &&
966 tp->snd_max == tp->snd_recover)
967 tp->snd_nxt--;
968
969 off = tp->snd_nxt - tp->snd_una;
970 sendalot = 0;
971 tp->t_flagsext |= TF_RESCUE_RXT;
972 }
973 }
974 }
975
976 /*
977 * Lop off SYN bit if it has already been sent. However, if this
978 * is SYN-SENT state and if segment contains data and if we don't
979 * know that foreign host supports TAO, suppress sending segment.
980 */
981 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
982 if (tp->t_state != TCPS_SYN_RECEIVED || tfo_enabled(tp))
983 flags &= ~TH_SYN;
984 off--;
985 len++;
986 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
987 while (inp->inp_sndinprog_cnt == 0 &&
988 tp->t_pktlist_head != NULL) {
989 packetlist = tp->t_pktlist_head;
990 packchain_listadd = tp->t_lastchain;
991 packchain_sent++;
992 TCP_PKTLIST_CLEAR(tp);
993
994 error = tcp_ip_output(so, tp, packetlist,
995 packchain_listadd, tp_inp_options,
996 (so_options & SO_DONTROUTE),
997 (sack_rxmit || (sack_bytes_rxmt != 0)),
998 isipv6);
999 }
1000
1001 /*
1002 * tcp was closed while we were in ip,
1003 * resume close
1004 */
1005 if (inp->inp_sndinprog_cnt == 0 &&
1006 (tp->t_flags & TF_CLOSING)) {
1007 tp->t_flags &= ~TF_CLOSING;
1008 (void) tcp_close(tp);
1009 } else {
1010 tcp_check_timer_state(tp);
1011 }
1012 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,
1013 0,0,0,0,0);
1014 return(0);
1015 }
1016 }
1017
1018 /*
1019 * Be careful not to send data and/or FIN on SYN segments.
1020 * This measure is needed to prevent interoperability problems
1021 * with not fully conformant TCP implementations.
1022 *
1023 * In case of TFO, we handle the setting of the len in
1024 * tcp_tfo_check. In case TFO is not enabled, never ever send
1025 * SYN+data.
1026 */
1027 if ((flags & TH_SYN) && !tfo_enabled(tp)) {
1028 len = 0;
1029 flags &= ~TH_FIN;
1030 }
1031
1032 if ((flags & TH_SYN) && tp->t_state <= TCPS_SYN_SENT && tfo_enabled(tp))
1033 len = tcp_tfo_check(tp, len);
1034
1035 /*
1036 * The check here used to be (len < 0). Some times len is zero
1037 * when the congestion window is closed and we need to check
1038 * if persist timer has to be set in that case. But don't set
1039 * persist until connection is established.
1040 */
1041 if (len <= 0 && !(flags & TH_SYN)) {
1042 /*
1043 * If FIN has been sent but not acked,
1044 * but we haven't been called to retransmit,
1045 * len will be < 0. Otherwise, window shrank
1046 * after we sent into it. If window shrank to 0,
1047 * cancel pending retransmit, pull snd_nxt back
1048 * to (closed) window, and set the persist timer
1049 * if it isn't already going. If the window didn't
1050 * close completely, just wait for an ACK.
1051 */
1052 len = 0;
1053 if (sendwin == 0) {
1054 tp->t_timer[TCPT_REXMT] = 0;
1055 tp->t_timer[TCPT_PTO] = 0;
1056 TCP_RESET_REXMT_STATE(tp);
1057 tp->snd_nxt = tp->snd_una;
1058 off = 0;
1059 if (tp->t_timer[TCPT_PERSIST] == 0)
1060 tcp_setpersist(tp);
1061 }
1062 }
1063
1064 /*
1065 * Automatic sizing of send socket buffer. Increase the send
1066 * socket buffer size if all of the following criteria are met
1067 * 1. the receiver has enough buffer space for this data
1068 * 2. send buffer is filled to 7/8th with data (so we actually
1069 * have data to make use of it);
1070 * 3. our send window (slow start and congestion controlled) is
1071 * larger than sent but unacknowledged data in send buffer.
1072 */
1073 if (tcp_do_autosendbuf == 1 &&
1074 !INP_WAIT_FOR_IF_FEEDBACK(inp) && !IN_FASTRECOVERY(tp) &&
1075 (so->so_snd.sb_flags & (SB_AUTOSIZE | SB_TRIM)) == SB_AUTOSIZE &&
1076 tcp_cansbgrow(&so->so_snd)) {
1077 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
1078 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
1079 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
1080 if (sbreserve(&so->so_snd,
1081 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
1082 tcp_autosndbuf_max)) == 1) {
1083 so->so_snd.sb_idealsize = so->so_snd.sb_hiwat;
1084 }
1085 }
1086 }
1087
1088 /*
1089 * Truncate to the maximum segment length or enable TCP Segmentation
1090 * Offloading (if supported by hardware) and ensure that FIN is removed
1091 * if the length no longer contains the last data byte.
1092 *
1093 * TSO may only be used if we are in a pure bulk sending state.
1094 * The presence of TCP-MD5, SACK retransmits, SACK advertizements,
1095 * ipfw rules and IP options, as well as disabling hardware checksum
1096 * offload prevent using TSO. With TSO the TCP header is the same
1097 * (except for the sequence number) for all generated packets. This
1098 * makes it impossible to transmit any options which vary per generated
1099 * segment or packet.
1100 *
1101 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
1102 * removal of FIN (if not already catched here) are handled later after
1103 * the exact length of the TCP options are known.
1104 */
1105 #if IPSEC
1106 /*
1107 * Pre-calculate here as we save another lookup into the darknesses
1108 * of IPsec that way and can actually decide if TSO is ok.
1109 */
1110 if (ipsec_bypass == 0)
1111 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
1112 #endif
1113 if (len > tp->t_maxseg) {
1114 if ((tp->t_flags & TF_TSO) && tcp_do_tso && hwcksum_tx &&
1115 ip_use_randomid && kipf_count == 0 &&
1116 dlil_filter_disable_tso_count == 0 &&
1117 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
1118 sack_bytes_rxmt == 0 &&
1119 inp->inp_options == NULL &&
1120 inp->in6p_options == NULL
1121 #if IPSEC
1122 && ipsec_optlen == 0
1123 #endif
1124 #if IPFIREWALL
1125 && (fw_enable == 0 || fw_bypass)
1126 #endif
1127 ) {
1128 tso = 1;
1129 sendalot = 0;
1130 } else {
1131 len = tp->t_maxseg;
1132 sendalot = 1;
1133 tso = 0;
1134 }
1135 }
1136
1137 /* Send one segment or less as a tail loss probe */
1138 if (tp->t_flagsext & TF_SENT_TLPROBE) {
1139 len = min(len, tp->t_maxseg);
1140 sendalot = 0;
1141 tso = 0;
1142 }
1143
1144 #if MPTCP
1145 if ((so->so_flags & SOF_MP_SUBFLOW) &&
1146 !(tp->t_mpflags & TMPF_TCP_FALLBACK)) {
1147 int newlen = len;
1148 if (tp->t_state >= TCPS_ESTABLISHED &&
1149 (tp->t_mpflags & TMPF_SND_MPPRIO ||
1150 tp->t_mpflags & TMPF_SND_REM_ADDR ||
1151 tp->t_mpflags & TMPF_SND_MPFAIL ||
1152 tp->t_mpflags & TMPF_SND_KEYS ||
1153 tp->t_mpflags & TMPF_SND_JACK)) {
1154 if (len > 0) {
1155 len = 0;
1156 }
1157 /*
1158 * On a new subflow, don't try to send again, because
1159 * we are still waiting for the fourth ack.
1160 */
1161 if (!(tp->t_mpflags & TMPF_PREESTABLISHED))
1162 sendalot = 1;
1163 mptcp_acknow = TRUE;
1164 } else {
1165 mptcp_acknow = FALSE;
1166 }
1167 /*
1168 * The contiguous bytes in the subflow socket buffer can be
1169 * discontiguous at the MPTCP level. Since only one DSS
1170 * option can be sent in one packet, reduce length to match
1171 * the contiguous MPTCP level. Set sendalot to send remainder.
1172 */
1173 if (len > 0)
1174 newlen = mptcp_adj_sendlen(so, off);
1175 if (newlen < len) {
1176 len = newlen;
1177 sendalot = 1;
1178 }
1179 }
1180 #endif /* MPTCP */
1181
1182 /*
1183 * If the socket is capable of doing unordered send,
1184 * pull the amount of data that can be sent from the
1185 * unordered priority queues to the serial queue in
1186 * the socket buffer. If bytes are not yet available
1187 * in the highest priority message, we may not be able
1188 * to send any new data.
1189 */
1190 if (so->so_flags & SOF_ENABLE_MSGS) {
1191 if ((off + len) >
1192 so->so_msg_state->msg_serial_bytes) {
1193 sbpull_unordered_data(so, off, len);
1194
1195 /* check if len needs to be modified */
1196 if ((off + len) >
1197 so->so_msg_state->msg_serial_bytes) {
1198 len = so->so_msg_state->msg_serial_bytes - off;
1199 if (len <= 0) {
1200 len = 0;
1201 tcpstat.tcps_msg_sndwaithipri++;
1202 }
1203 }
1204 }
1205 }
1206
1207 if (sack_rxmit) {
1208 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
1209 flags &= ~TH_FIN;
1210 } else {
1211 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
1212 flags &= ~TH_FIN;
1213 }
1214 /*
1215 * Compare available window to amount of window
1216 * known to peer (as advertised window less
1217 * next expected input). If the difference is at least two
1218 * max size segments, or at least 25% of the maximum possible
1219 * window, then want to send a window update to peer.
1220 * Skip this if the connection is in T/TCP half-open state.
1221 */
1222 recwin = tcp_sbspace(tp);
1223 #if MPTCP
1224 if (so->so_flags & SOF_MP_SUBFLOW) {
1225 struct mptcb *mp_tp = tptomptp(tp);
1226
1227 if (mp_tp != NULL) {
1228 mpte_lock_assert_held(mp_tp->mpt_mpte);
1229 recwin = imin(recwin, mptcp_sbspace(mp_tp));
1230 }
1231 }
1232 #endif
1233
1234 if (recwin < (int32_t)(so->so_rcv.sb_hiwat / 4) &&
1235 recwin < (int)tp->t_maxseg)
1236 recwin = 0;
1237
1238 #if TRAFFIC_MGT
1239 if (tcp_recv_bg == 1 || IS_TCP_RECV_BG(so)) {
1240 if (recwin > 0 && tcp_recv_throttle(tp)) {
1241 uint32_t min_iaj_win = tcp_min_iaj_win * tp->t_maxseg;
1242 uint32_t bg_rwintop = tp->rcv_adv;
1243 if (SEQ_LT(bg_rwintop, tp->rcv_nxt + min_iaj_win))
1244 bg_rwintop = tp->rcv_nxt + min_iaj_win;
1245 recwin = imin((int32_t)(bg_rwintop - tp->rcv_nxt),
1246 recwin);
1247 if (recwin < 0)
1248 recwin = 0;
1249 }
1250 }
1251 #endif /* TRAFFIC_MGT */
1252
1253 if (recwin > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
1254 recwin = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
1255
1256 /*
1257 * MPTCP needs to be able to announce a smaller window than previously,
1258 * because the other subflow may have filled up the available window-
1259 * space. So we have to be able to go backwards and announce a smaller
1260 * window.
1261 */
1262 if (!(so->so_flags & SOF_MP_SUBFLOW) &&
1263 recwin < (int32_t)(tp->rcv_adv - tp->rcv_nxt))
1264 recwin = (int32_t)(tp->rcv_adv - tp->rcv_nxt);
1265
1266 /*
1267 * Sender silly window avoidance. We transmit under the following
1268 * conditions when len is non-zero:
1269 *
1270 * - we've timed out (e.g. persist timer)
1271 * - we need to retransmit
1272 * - We have a full segment (or more with TSO)
1273 * - This is the last buffer in a write()/send() and we are
1274 * either idle or running NODELAY
1275 * - we have more then 1/2 the maximum send window's worth of
1276 * data (receiver may be limited the window size)
1277 */
1278 if (len) {
1279 if (tp->t_flagsext & TF_FORCE)
1280 goto send;
1281 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1282 goto send;
1283 if (sack_rxmit)
1284 goto send;
1285
1286 /*
1287 * Send new data on the connection only if it is
1288 * not flow controlled
1289 */
1290 if (!INP_WAIT_FOR_IF_FEEDBACK(inp) ||
1291 tp->t_state != TCPS_ESTABLISHED) {
1292 if (len >= tp->t_maxseg)
1293 goto send;
1294
1295 if (!(tp->t_flags & TF_MORETOCOME) &&
1296 (idle || tp->t_flags & TF_NODELAY ||
1297 (tp->t_flags & TF_MAXSEGSNT) ||
1298 ALLOW_LIMITED_TRANSMIT(tp)) &&
1299 (tp->t_flags & TF_NOPUSH) == 0 &&
1300 (len + off >= so->so_snd.sb_cc ||
1301 /*
1302 * MPTCP needs to respect the DSS-mappings. So, it
1303 * may be sending data that *could* have been
1304 * coalesced, but cannot because of
1305 * mptcp_adj_sendlen().
1306 */
1307 so->so_flags & SOF_MP_SUBFLOW))
1308 goto send;
1309 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
1310 goto send;
1311 } else {
1312 tcpstat.tcps_fcholdpacket++;
1313 }
1314 }
1315
1316 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN)) {
1317 /*
1318 * "adv" is the amount we can increase the window,
1319 * taking into account that we are limited by
1320 * TCP_MAXWIN << tp->rcv_scale.
1321 */
1322 int32_t adv, oldwin = 0;
1323 adv = imin(recwin, (int)TCP_MAXWIN << tp->rcv_scale) -
1324 (tp->rcv_adv - tp->rcv_nxt);
1325
1326 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
1327 oldwin = tp->rcv_adv - tp->rcv_nxt;
1328
1329 if (adv >= (int32_t) (2 * tp->t_maxseg)) {
1330 /*
1331 * Update only if the resulting scaled value of
1332 * the window changed, or if there is a change in
1333 * the sequence since the last ack. This avoids
1334 * what appears as dupe ACKS (see rdar://5640997)
1335 *
1336 * If streaming is detected avoid sending too many
1337 * window updates. We will depend on the delack
1338 * timer to send a window update when needed.
1339 */
1340 if (!(tp->t_flags & TF_STRETCHACK) &&
1341 (tp->last_ack_sent != tp->rcv_nxt ||
1342 ((oldwin + adv) >> tp->rcv_scale) >
1343 (oldwin >> tp->rcv_scale))) {
1344 goto send;
1345 }
1346
1347 }
1348 if (4 * adv >= (int32_t) so->so_rcv.sb_hiwat)
1349 goto send;
1350
1351 /*
1352 * Make sure that the delayed ack timer is set if
1353 * we delayed sending a window update because of
1354 * streaming detection.
1355 */
1356 if ((tp->t_flags & TF_STRETCHACK) &&
1357 !(tp->t_flags & TF_DELACK)) {
1358 tp->t_flags |= TF_DELACK;
1359 tp->t_timer[TCPT_DELACK] =
1360 OFFSET_FROM_START(tp, tcp_delack);
1361 }
1362 }
1363
1364 /*
1365 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
1366 * is also a catch-all for the retransmit timer timeout case.
1367 */
1368 if (tp->t_flags & TF_ACKNOW)
1369 goto send;
1370 if ((flags & TH_RST) ||
1371 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
1372 goto send;
1373 if (SEQ_GT(tp->snd_up, tp->snd_una))
1374 goto send;
1375 #if MPTCP
1376 if (mptcp_acknow)
1377 goto send;
1378 #endif /* MPTCP */
1379 /*
1380 * If our state indicates that FIN should be sent
1381 * and we have not yet done so, then we need to send.
1382 */
1383 if ((flags & TH_FIN) &&
1384 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
1385 goto send;
1386 /*
1387 * In SACK, it is possible for tcp_output to fail to send a segment
1388 * after the retransmission timer has been turned off. Make sure
1389 * that the retransmission timer is set.
1390 */
1391 if (SACK_ENABLED(tp) && (tp->t_state >= TCPS_ESTABLISHED) &&
1392 SEQ_GT(tp->snd_max, tp->snd_una) &&
1393 tp->t_timer[TCPT_REXMT] == 0 &&
1394 tp->t_timer[TCPT_PERSIST] == 0) {
1395 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp,
1396 tp->t_rxtcur);
1397 goto just_return;
1398 }
1399 /*
1400 * TCP window updates are not reliable, rather a polling protocol
1401 * using ``persist'' packets is used to insure receipt of window
1402 * updates. The three ``states'' for the output side are:
1403 * idle not doing retransmits or persists
1404 * persisting to move a small or zero window
1405 * (re)transmitting and thereby not persisting
1406 *
1407 * tp->t_timer[TCPT_PERSIST]
1408 * is set when we are in persist state.
1409 * tp->t_force
1410 * is set when we are called to send a persist packet.
1411 * tp->t_timer[TCPT_REXMT]
1412 * is set when we are retransmitting
1413 * The output side is idle when both timers are zero.
1414 *
1415 * If send window is too small, there is data to transmit, and no
1416 * retransmit or persist is pending, then go to persist state.
1417 * If nothing happens soon, send when timer expires:
1418 * if window is nonzero, transmit what we can,
1419 * otherwise force out a byte.
1420 */
1421 if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 &&
1422 tp->t_timer[TCPT_PERSIST] == 0) {
1423 TCP_RESET_REXMT_STATE(tp);
1424 tcp_setpersist(tp);
1425 }
1426 just_return:
1427 /*
1428 * If there is no reason to send a segment, just return.
1429 * but if there is some packets left in the packet list, send them now.
1430 */
1431 while (inp->inp_sndinprog_cnt == 0 &&
1432 tp->t_pktlist_head != NULL) {
1433 packetlist = tp->t_pktlist_head;
1434 packchain_listadd = tp->t_lastchain;
1435 packchain_sent++;
1436 TCP_PKTLIST_CLEAR(tp);
1437
1438 error = tcp_ip_output(so, tp, packetlist,
1439 packchain_listadd,
1440 tp_inp_options, (so_options & SO_DONTROUTE),
1441 (sack_rxmit || (sack_bytes_rxmt != 0)), isipv6);
1442 }
1443 /* tcp was closed while we were in ip; resume close */
1444 if (inp->inp_sndinprog_cnt == 0 &&
1445 (tp->t_flags & TF_CLOSING)) {
1446 tp->t_flags &= ~TF_CLOSING;
1447 (void) tcp_close(tp);
1448 } else {
1449 tcp_check_timer_state(tp);
1450 }
1451 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
1452 return (0);
1453
1454 send:
1455 /*
1456 * Set TF_MAXSEGSNT flag if the segment size is greater than
1457 * the max segment size.
1458 */
1459 if (len > 0) {
1460 if (len >= tp->t_maxseg)
1461 tp->t_flags |= TF_MAXSEGSNT;
1462 else
1463 tp->t_flags &= ~TF_MAXSEGSNT;
1464 }
1465 /*
1466 * Before ESTABLISHED, force sending of initial options
1467 * unless TCP set not to do any options.
1468 * NOTE: we assume that the IP/TCP header plus TCP options
1469 * always fit in a single mbuf, leaving room for a maximum
1470 * link header, i.e.
1471 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1472 */
1473 optlen = 0;
1474 #if INET6
1475 if (isipv6)
1476 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
1477 else
1478 #endif
1479 hdrlen = sizeof (struct tcpiphdr);
1480 if (flags & TH_SYN) {
1481 tp->snd_nxt = tp->iss;
1482 if ((tp->t_flags & TF_NOOPT) == 0) {
1483 u_short mss;
1484
1485 opt[0] = TCPOPT_MAXSEG;
1486 opt[1] = TCPOLEN_MAXSEG;
1487 mss = htons((u_short) tcp_mssopt(tp));
1488 (void)memcpy(opt + 2, &mss, sizeof(mss));
1489 optlen = TCPOLEN_MAXSEG;
1490
1491 if ((tp->t_flags & TF_REQ_SCALE) &&
1492 ((flags & TH_ACK) == 0 ||
1493 (tp->t_flags & TF_RCVD_SCALE))) {
1494 *((u_int32_t *)(void *)(opt + optlen)) = htonl(
1495 TCPOPT_NOP << 24 |
1496 TCPOPT_WINDOW << 16 |
1497 TCPOLEN_WINDOW << 8 |
1498 tp->request_r_scale);
1499 optlen += 4;
1500 }
1501 #if MPTCP
1502 if (mptcp_enable && (so->so_flags & SOF_MP_SUBFLOW)) {
1503 optlen = mptcp_setup_syn_opts(so, opt, optlen);
1504 }
1505 #endif /* MPTCP */
1506 }
1507 }
1508
1509 /*
1510 * Send a timestamp and echo-reply if this is a SYN and our side
1511 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1512 * and our peer have sent timestamps in our SYN's.
1513 */
1514 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1515 (flags & TH_RST) == 0 &&
1516 ((flags & TH_ACK) == 0 ||
1517 (tp->t_flags & TF_RCVD_TSTMP))) {
1518 u_int32_t *lp = (u_int32_t *)(void *)(opt + optlen);
1519
1520 /* Form timestamp option as shown in appendix A of RFC 1323. */
1521 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1522 *lp++ = htonl(tcp_now);
1523 *lp = htonl(tp->ts_recent);
1524 optlen += TCPOLEN_TSTAMP_APPA;
1525 }
1526
1527 /* Note the timestamp for receive buffer autosizing */
1528 if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE))
1529 tp->rfbuf_ts = tcp_now;
1530
1531 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1532 /*
1533 * Tack on the SACK permitted option *last*.
1534 * And do padding of options after tacking this on.
1535 * This is because of MSS, TS, WinScale and Signatures are
1536 * all present, we have just 2 bytes left for the SACK
1537 * permitted option, which is just enough.
1538 */
1539 /*
1540 * If this is the first SYN of connection (not a SYN
1541 * ACK), include SACK permitted option. If this is a
1542 * SYN ACK, include SACK permitted option if peer has
1543 * already done so. This is only for active connect,
1544 * since the syncache takes care of the passive connect.
1545 */
1546 if ((flags & TH_SYN) &&
1547 (!(flags & TH_ACK) || (tp->t_flags & TF_SACK_PERMIT))) {
1548 u_char *bp;
1549 bp = (u_char *)opt + optlen;
1550
1551 *bp++ = TCPOPT_SACK_PERMITTED;
1552 *bp++ = TCPOLEN_SACK_PERMITTED;
1553 optlen += TCPOLEN_SACK_PERMITTED;
1554 }
1555 }
1556 #if MPTCP
1557 if (so->so_flags & SOF_MP_SUBFLOW) {
1558 /*
1559 * Its important to piggyback acks with data as ack only packets
1560 * may get lost and data packets that don't send Data ACKs
1561 * still advance the subflow level ACK and therefore make it
1562 * hard for the remote end to recover in low cwnd situations.
1563 */
1564 if (len != 0) {
1565 tp->t_mpflags |= (TMPF_SEND_DSN |
1566 TMPF_MPTCP_ACKNOW);
1567 } else {
1568 tp->t_mpflags |= TMPF_MPTCP_ACKNOW;
1569 }
1570 optlen = mptcp_setup_opts(tp, off, &opt[0], optlen, flags,
1571 len, &mptcp_acknow);
1572 tp->t_mpflags &= ~TMPF_SEND_DSN;
1573 }
1574 #endif /* MPTCP */
1575
1576 if (tfo_enabled(tp) && !(tp->t_flags & TF_NOOPT) &&
1577 (flags & (TH_SYN | TH_ACK)) == TH_SYN)
1578 optlen += tcp_tfo_write_cookie(tp, optlen, len, opt);
1579
1580 if (tfo_enabled(tp) &&
1581 (flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1582 (tp->t_tfo_flags & TFO_F_OFFER_COOKIE))
1583 optlen += tcp_tfo_write_cookie_rep(tp, optlen, opt);
1584
1585 if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) {
1586 /*
1587 * Send SACKs if necessary. This should be the last
1588 * option processed. Only as many SACKs are sent as
1589 * are permitted by the maximum options size.
1590 *
1591 * In general, SACK blocks consume 8*n+2 bytes.
1592 * So a full size SACK blocks option is 34 bytes
1593 * (to generate 4 SACK blocks). At a minimum,
1594 * we need 10 bytes (to generate 1 SACK block).
1595 * If TCP Timestamps (12 bytes) and TCP Signatures
1596 * (18 bytes) are both present, we'll just have
1597 * 10 bytes for SACK options 40 - (12 + 18).
1598 */
1599 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1600 (tp->t_flags & TF_SACK_PERMIT) &&
1601 (tp->rcv_numsacks > 0 || TCP_SEND_DSACK_OPT(tp)) &&
1602 MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) {
1603 int nsack, padlen;
1604 u_char *bp = (u_char *)opt + optlen;
1605 u_int32_t *lp;
1606
1607 nsack = (MAX_TCPOPTLEN - optlen - 2) / TCPOLEN_SACK;
1608 nsack = min(nsack, (tp->rcv_numsacks +
1609 (TCP_SEND_DSACK_OPT(tp) ? 1 : 0)));
1610 sackoptlen = (2 + nsack * TCPOLEN_SACK);
1611
1612 /*
1613 * First we need to pad options so that the
1614 * SACK blocks can start at a 4-byte boundary
1615 * (sack option and length are at a 2 byte offset).
1616 */
1617 padlen = (MAX_TCPOPTLEN - optlen - sackoptlen) % 4;
1618 optlen += padlen;
1619 while (padlen-- > 0)
1620 *bp++ = TCPOPT_NOP;
1621
1622 tcpstat.tcps_sack_send_blocks++;
1623 *bp++ = TCPOPT_SACK;
1624 *bp++ = sackoptlen;
1625 lp = (u_int32_t *)(void *)bp;
1626
1627 /*
1628 * First block of SACK option should represent
1629 * DSACK. Prefer to send SACK information if there
1630 * is space for only one SACK block. This will
1631 * allow for faster recovery.
1632 */
1633 if (TCP_SEND_DSACK_OPT(tp) && nsack > 0 &&
1634 (tp->rcv_numsacks == 0 || nsack > 1)) {
1635 *lp++ = htonl(tp->t_dsack_lseq);
1636 *lp++ = htonl(tp->t_dsack_rseq);
1637 tcpstat.tcps_dsack_sent++;
1638 tp->t_dsack_sent++;
1639 nsack--;
1640 }
1641 VERIFY(nsack == 0 || tp->rcv_numsacks >= nsack);
1642 for (i = 0; i < nsack; i++) {
1643 struct sackblk sack = tp->sackblks[i];
1644 *lp++ = htonl(sack.start);
1645 *lp++ = htonl(sack.end);
1646 }
1647 optlen += sackoptlen;
1648 }
1649 }
1650
1651 /* Pad TCP options to a 4 byte boundary */
1652 if (optlen < MAX_TCPOPTLEN && (optlen % sizeof(u_int32_t))) {
1653 int pad = sizeof(u_int32_t) - (optlen % sizeof(u_int32_t));
1654 u_char *bp = (u_char *)opt + optlen;
1655
1656 optlen += pad;
1657 while (pad) {
1658 *bp++ = TCPOPT_EOL;
1659 pad--;
1660 }
1661 }
1662
1663 /*
1664 * RFC 3168 states that:
1665 * - If you ever sent an ECN-setup SYN/SYN-ACK you must be prepared
1666 * to handle the TCP ECE flag, even if you also later send a
1667 * non-ECN-setup SYN/SYN-ACK.
1668 * - If you ever send a non-ECN-setup SYN/SYN-ACK, you must not set
1669 * the ip ECT flag.
1670 *
1671 * It is not clear how the ECE flag would ever be set if you never
1672 * set the IP ECT flag on outbound packets. All the same, we use
1673 * the TE_SETUPSENT to indicate that we have committed to handling
1674 * the TCP ECE flag correctly. We use the TE_SENDIPECT to indicate
1675 * whether or not we should set the IP ECT flag on outbound packet
1676 *
1677 * For a SYN-ACK, send an ECN setup SYN-ACK
1678 */
1679 if ((flags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK) &&
1680 (tp->ecn_flags & TE_ENABLE_ECN)) {
1681 if (tp->ecn_flags & TE_SETUPRECEIVED) {
1682 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1683 /*
1684 * Setting TH_ECE makes this an ECN-setup
1685 * SYN-ACK
1686 */
1687 flags |= TH_ECE;
1688
1689 /*
1690 * Record that we sent the ECN-setup and
1691 * default to setting IP ECT.
1692 */
1693 tp->ecn_flags |= (TE_SETUPSENT|TE_SENDIPECT);
1694 tcpstat.tcps_ecn_server_setup++;
1695 tcpstat.tcps_ecn_server_success++;
1696 } else {
1697 /*
1698 * We sent an ECN-setup SYN-ACK but it was
1699 * dropped. Fallback to non-ECN-setup
1700 * SYN-ACK and clear flag to indicate that
1701 * we should not send data with IP ECT set
1702 *
1703 * Pretend we didn't receive an
1704 * ECN-setup SYN.
1705 *
1706 * We already incremented the counter
1707 * assuming that the ECN setup will
1708 * succeed. Decrementing here
1709 * tcps_ecn_server_success to correct it.
1710 */
1711 if (tp->ecn_flags & TE_SETUPSENT) {
1712 tcpstat.tcps_ecn_lost_synack++;
1713 tcpstat.tcps_ecn_server_success--;
1714 tp->ecn_flags |= TE_LOST_SYNACK;
1715 }
1716
1717 tp->ecn_flags &=
1718 ~(TE_SETUPRECEIVED | TE_SENDIPECT |
1719 TE_SENDCWR);
1720 }
1721 }
1722 } else if ((flags & (TH_SYN | TH_ACK)) == TH_SYN &&
1723 (tp->ecn_flags & TE_ENABLE_ECN)) {
1724 if (tcp_send_ecn_flags_on_syn(tp, so)) {
1725 /*
1726 * Setting TH_ECE and TH_CWR makes this an
1727 * ECN-setup SYN
1728 */
1729 flags |= (TH_ECE | TH_CWR);
1730 tcpstat.tcps_ecn_client_setup++;
1731 tp->ecn_flags |= TE_CLIENT_SETUP;
1732
1733 /*
1734 * Record that we sent the ECN-setup and default to
1735 * setting IP ECT.
1736 */
1737 tp->ecn_flags |= (TE_SETUPSENT | TE_SENDIPECT);
1738 } else {
1739 /*
1740 * We sent an ECN-setup SYN but it was dropped.
1741 * Fall back to non-ECN and clear flag indicating
1742 * we should send data with IP ECT set.
1743 */
1744 if (tp->ecn_flags & TE_SETUPSENT) {
1745 tcpstat.tcps_ecn_lost_syn++;
1746 tp->ecn_flags |= TE_LOST_SYN;
1747 }
1748 tp->ecn_flags &= ~TE_SENDIPECT;
1749 }
1750 }
1751
1752 /*
1753 * Check if we should set the TCP CWR flag.
1754 * CWR flag is sent when we reduced the congestion window because
1755 * we received a TCP ECE or we performed a fast retransmit. We
1756 * never set the CWR flag on retransmitted packets. We only set
1757 * the CWR flag on data packets. Pure acks don't have this set.
1758 */
1759 if ((tp->ecn_flags & TE_SENDCWR) != 0 && len != 0 &&
1760 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
1761 flags |= TH_CWR;
1762 tp->ecn_flags &= ~TE_SENDCWR;
1763 }
1764
1765 /*
1766 * Check if we should set the TCP ECE flag.
1767 */
1768 if ((tp->ecn_flags & TE_SENDECE) != 0 && len == 0) {
1769 flags |= TH_ECE;
1770 tcpstat.tcps_ecn_sent_ece++;
1771 }
1772
1773
1774 hdrlen += optlen;
1775
1776 /* Reset DSACK sequence numbers */
1777 tp->t_dsack_lseq = 0;
1778 tp->t_dsack_rseq = 0;
1779
1780 #if INET6
1781 if (isipv6)
1782 ipoptlen = ip6_optlen(inp);
1783 else
1784 #endif
1785 {
1786 if (tp_inp_options) {
1787 ipoptlen = tp_inp_options->m_len -
1788 offsetof(struct ipoption, ipopt_list);
1789 } else {
1790 ipoptlen = 0;
1791 }
1792 }
1793 #if IPSEC
1794 ipoptlen += ipsec_optlen;
1795 #endif
1796
1797 /*
1798 * Adjust data length if insertion of options will
1799 * bump the packet length beyond the t_maxopd length.
1800 * Clear the FIN bit because we cut off the tail of
1801 * the segment.
1802 *
1803 * When doing TSO limit a burst to TCP_MAXWIN minus the
1804 * IP, TCP and Options length to keep ip->ip_len from
1805 * overflowing. Prevent the last segment from being
1806 * fractional thus making them all equal sized and set
1807 * the flag to continue sending. TSO is disabled when
1808 * IP options or IPSEC are present.
1809 */
1810 if (len + optlen + ipoptlen > tp->t_maxopd) {
1811 /*
1812 * If there is still more to send,
1813 * don't close the connection.
1814 */
1815 flags &= ~TH_FIN;
1816 if (tso) {
1817 int32_t tso_maxlen;
1818
1819 tso_maxlen = tp->tso_max_segment_size ?
1820 tp->tso_max_segment_size : TCP_MAXWIN;
1821
1822 if (len > tso_maxlen - hdrlen - optlen) {
1823 len = tso_maxlen - hdrlen - optlen;
1824 len = len - (len % (tp->t_maxopd - optlen));
1825 sendalot = 1;
1826 } else if (tp->t_flags & TF_NEEDFIN) {
1827 sendalot = 1;
1828 }
1829 } else {
1830 len = tp->t_maxopd - optlen - ipoptlen;
1831 sendalot = 1;
1832 }
1833 }
1834
1835 if (max_linkhdr + hdrlen > MCLBYTES)
1836 panic("tcphdr too big");
1837
1838 /* Check if there is enough data in the send socket
1839 * buffer to start measuring bandwidth
1840 */
1841 if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 &&
1842 (tp->t_bwmeas != NULL) &&
1843 (tp->t_flagsext & TF_BWMEAS_INPROGRESS) == 0) {
1844 tp->t_bwmeas->bw_size = min(min(
1845 (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)),
1846 tp->snd_cwnd), tp->snd_wnd);
1847 if (tp->t_bwmeas->bw_minsize > 0 &&
1848 tp->t_bwmeas->bw_size < tp->t_bwmeas->bw_minsize)
1849 tp->t_bwmeas->bw_size = 0;
1850 if (tp->t_bwmeas->bw_maxsize > 0)
1851 tp->t_bwmeas->bw_size = min(tp->t_bwmeas->bw_size,
1852 tp->t_bwmeas->bw_maxsize);
1853 if (tp->t_bwmeas->bw_size > 0) {
1854 tp->t_flagsext |= TF_BWMEAS_INPROGRESS;
1855 tp->t_bwmeas->bw_start = tp->snd_max;
1856 tp->t_bwmeas->bw_ts = tcp_now;
1857 }
1858 }
1859
1860 VERIFY(inp->inp_flowhash != 0);
1861 /*
1862 * Grab a header mbuf, attaching a copy of data to
1863 * be transmitted, and initialize the header from
1864 * the template for sends on this connection.
1865 */
1866 if (len) {
1867 tp->t_pmtud_lastseg_size = len + optlen + ipoptlen;
1868 if ((tp->t_flagsext & TF_FORCE) && len == 1)
1869 tcpstat.tcps_sndprobe++;
1870 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1871 tcpstat.tcps_sndrexmitpack++;
1872 tcpstat.tcps_sndrexmitbyte += len;
1873 if (nstat_collect) {
1874 nstat_route_tx(inp->inp_route.ro_rt, 1,
1875 len, NSTAT_TX_FLAG_RETRANSMIT);
1876 INP_ADD_STAT(inp, cell, wifi, wired,
1877 txpackets, 1);
1878 INP_ADD_STAT(inp, cell, wifi, wired,
1879 txbytes, len);
1880 tp->t_stat.txretransmitbytes += len;
1881 tp->t_stat.rxmitpkts++;
1882 }
1883 } else {
1884 tcpstat.tcps_sndpack++;
1885 tcpstat.tcps_sndbyte += len;
1886
1887 if (nstat_collect) {
1888 INP_ADD_STAT(inp, cell, wifi, wired,
1889 txpackets, 1);
1890 INP_ADD_STAT(inp, cell, wifi, wired,
1891 txbytes, len);
1892 }
1893 inp_decr_sndbytes_unsent(so, len);
1894 }
1895 inp_set_activity_bitmap(inp);
1896 #if MPTCP
1897 if (tp->t_mpflags & TMPF_MPTCP_TRUE) {
1898 tcpstat.tcps_mp_sndpacks++;
1899 tcpstat.tcps_mp_sndbytes += len;
1900 }
1901 #endif /* MPTCP */
1902 /*
1903 * try to use the new interface that allocates all
1904 * the necessary mbuf hdrs under 1 mbuf lock and
1905 * avoids rescanning the socket mbuf list if
1906 * certain conditions are met. This routine can't
1907 * be used in the following cases...
1908 * 1) the protocol headers exceed the capacity of
1909 * of a single mbuf header's data area (no cluster attached)
1910 * 2) the length of the data being transmitted plus
1911 * the protocol headers fits into a single mbuf header's
1912 * data area (no cluster attached)
1913 */
1914 m = NULL;
1915
1916 /* minimum length we are going to allocate */
1917 allocated_len = MHLEN;
1918 if (MHLEN < hdrlen + max_linkhdr) {
1919 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1920 if (m == NULL) {
1921 error = ENOBUFS;
1922 goto out;
1923 }
1924 MCLGET(m, M_DONTWAIT);
1925 if ((m->m_flags & M_EXT) == 0) {
1926 m_freem(m);
1927 error = ENOBUFS;
1928 goto out;
1929 }
1930 m->m_data += max_linkhdr;
1931 m->m_len = hdrlen;
1932 allocated_len = MCLBYTES;
1933 }
1934 if (len <= allocated_len - hdrlen - max_linkhdr) {
1935 if (m == NULL) {
1936 VERIFY(allocated_len <= MHLEN);
1937 MGETHDR(m, M_DONTWAIT, MT_HEADER);
1938 if (m == NULL) {
1939 error = ENOBUFS;
1940 goto out;
1941 }
1942 m->m_data += max_linkhdr;
1943 m->m_len = hdrlen;
1944 }
1945 /* makes sure we still have data left to be sent at this point */
1946 if (so->so_snd.sb_mb == NULL || off < 0) {
1947 if (m != NULL) m_freem(m);
1948 error = 0; /* should we return an error? */
1949 goto out;
1950 }
1951 m_copydata(so->so_snd.sb_mb, off, (int) len,
1952 mtod(m, caddr_t) + hdrlen);
1953 m->m_len += len;
1954 } else {
1955 uint32_t copymode;
1956 /*
1957 * Retain packet header metadata at the socket
1958 * buffer if this is is an MPTCP subflow,
1959 * otherwise move it.
1960 */
1961 copymode = M_COPYM_MOVE_HDR;
1962 #if MPTCP
1963 if (so->so_flags & SOF_MP_SUBFLOW) {
1964 copymode = M_COPYM_NOOP_HDR;
1965 }
1966 #endif /* MPTCP */
1967 if (m != NULL) {
1968 m->m_next = m_copym_mode(so->so_snd.sb_mb,
1969 off, (int)len, M_DONTWAIT, copymode);
1970 if (m->m_next == NULL) {
1971 (void) m_free(m);
1972 error = ENOBUFS;
1973 goto out;
1974 }
1975 } else {
1976 /*
1977 * make sure we still have data left
1978 * to be sent at this point
1979 */
1980 if (so->so_snd.sb_mb == NULL) {
1981 error = 0; /* should we return an error? */
1982 goto out;
1983 }
1984
1985 /*
1986 * m_copym_with_hdrs will always return the
1987 * last mbuf pointer and the offset into it that
1988 * it acted on to fullfill the current request,
1989 * whether a valid 'hint' was passed in or not.
1990 */
1991 if ((m = m_copym_with_hdrs(so->so_snd.sb_mb,
1992 off, len, M_DONTWAIT, NULL, NULL,
1993 copymode)) == NULL) {
1994 error = ENOBUFS;
1995 goto out;
1996 }
1997 m->m_data += max_linkhdr;
1998 m->m_len = hdrlen;
1999 }
2000 }
2001 /*
2002 * If we're sending everything we've got, set PUSH.
2003 * (This will keep happy those implementations which only
2004 * give data to the user when a buffer fills or
2005 * a PUSH comes in.)
2006 *
2007 * On SYN-segments we should not add the PUSH-flag.
2008 */
2009 if (off + len == so->so_snd.sb_cc && !(flags & TH_SYN))
2010 flags |= TH_PUSH;
2011 } else {
2012 if (tp->t_flags & TF_ACKNOW)
2013 tcpstat.tcps_sndacks++;
2014 else if (flags & (TH_SYN|TH_FIN|TH_RST))
2015 tcpstat.tcps_sndctrl++;
2016 else if (SEQ_GT(tp->snd_up, tp->snd_una))
2017 tcpstat.tcps_sndurg++;
2018 else
2019 tcpstat.tcps_sndwinup++;
2020
2021 MGETHDR(m, M_DONTWAIT, MT_HEADER); /* MAC-OK */
2022 if (m == NULL) {
2023 error = ENOBUFS;
2024 goto out;
2025 }
2026 if (MHLEN < (hdrlen + max_linkhdr)) {
2027 MCLGET(m, M_DONTWAIT);
2028 if ((m->m_flags & M_EXT) == 0) {
2029 m_freem(m);
2030 error = ENOBUFS;
2031 goto out;
2032 }
2033 }
2034 m->m_data += max_linkhdr;
2035 m->m_len = hdrlen;
2036 }
2037 m->m_pkthdr.rcvif = 0;
2038 #if CONFIG_MACF_NET
2039 mac_mbuf_label_associate_inpcb(inp, m);
2040 #endif
2041 #if INET6
2042 if (isipv6) {
2043 ip6 = mtod(m, struct ip6_hdr *);
2044 th = (struct tcphdr *)(void *)(ip6 + 1);
2045 tcp_fillheaders(tp, ip6, th);
2046 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2047 !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) {
2048 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
2049 }
2050 svc_flags |= PKT_SCF_IPV6;
2051 #if PF_ECN
2052 m_pftag(m)->pftag_hdr = (void *)ip6;
2053 m_pftag(m)->pftag_flags |= PF_TAG_HDR_INET6;
2054 #endif /* PF_ECN */
2055 } else
2056 #endif /* INET6 */
2057 {
2058 ip = mtod(m, struct ip *);
2059 ipov = (struct ipovly *)ip;
2060 th = (struct tcphdr *)(void *)(ip + 1);
2061 /* this picks up the pseudo header (w/o the length) */
2062 tcp_fillheaders(tp, ip, th);
2063 if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len &&
2064 !SEQ_LT(tp->snd_nxt, tp->snd_max) &&
2065 !sack_rxmit && !(flags & TH_SYN)) {
2066 ip->ip_tos |= IPTOS_ECN_ECT0;
2067 }
2068 #if PF_ECN
2069 m_pftag(m)->pftag_hdr = (void *)ip;
2070 m_pftag(m)->pftag_flags |= PF_TAG_HDR_INET;
2071 #endif /* PF_ECN */
2072 }
2073
2074 /*
2075 * Fill in fields, remembering maximum advertised
2076 * window for use in delaying messages about window sizes.
2077 * If resending a FIN, be sure not to use a new sequence number.
2078 */
2079 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
2080 tp->snd_nxt == tp->snd_max)
2081 tp->snd_nxt--;
2082 /*
2083 * If we are doing retransmissions, then snd_nxt will
2084 * not reflect the first unsent octet. For ACK only
2085 * packets, we do not want the sequence number of the
2086 * retransmitted packet, we want the sequence number
2087 * of the next unsent octet. So, if there is no data
2088 * (and no SYN or FIN), use snd_max instead of snd_nxt
2089 * when filling in ti_seq. But if we are in persist
2090 * state, snd_max might reflect one byte beyond the
2091 * right edge of the window, so use snd_nxt in that
2092 * case, since we know we aren't doing a retransmission.
2093 * (retransmit and persist are mutually exclusive...)
2094 *
2095 * Note the state of this retransmit segment to detect spurious
2096 * retransmissions.
2097 */
2098 if (sack_rxmit == 0) {
2099 if (len || (flags & (TH_SYN|TH_FIN)) ||
2100 tp->t_timer[TCPT_PERSIST]) {
2101 th->th_seq = htonl(tp->snd_nxt);
2102 if (len > 0) {
2103 m->m_pkthdr.tx_start_seq = tp->snd_nxt;
2104 m->m_pkthdr.pkt_flags |= PKTF_START_SEQ;
2105 }
2106 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
2107 if (SACK_ENABLED(tp) && len > 1) {
2108 tcp_rxtseg_insert(tp, tp->snd_nxt,
2109 (tp->snd_nxt + len - 1));
2110 }
2111 if (len > 0)
2112 m->m_pkthdr.pkt_flags |=
2113 PKTF_TCP_REXMT;
2114 }
2115 } else {
2116 th->th_seq = htonl(tp->snd_max);
2117 }
2118 } else {
2119 th->th_seq = htonl(p->rxmit);
2120 if (len > 0) {
2121 m->m_pkthdr.pkt_flags |=
2122 (PKTF_TCP_REXMT | PKTF_START_SEQ);
2123 m->m_pkthdr.tx_start_seq = p->rxmit;
2124 }
2125 tcp_rxtseg_insert(tp, p->rxmit, (p->rxmit + len - 1));
2126 p->rxmit += len;
2127 tp->sackhint.sack_bytes_rexmit += len;
2128 }
2129 th->th_ack = htonl(tp->rcv_nxt);
2130 tp->last_ack_sent = tp->rcv_nxt;
2131 if (optlen) {
2132 bcopy(opt, th + 1, optlen);
2133 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
2134 }
2135 th->th_flags = flags;
2136 th->th_win = htons((u_short) (recwin>>tp->rcv_scale));
2137 if (recwin > 0 && SEQ_LT(tp->rcv_adv, tp->rcv_nxt + recwin))
2138 tp->rcv_adv = tp->rcv_nxt + recwin;
2139
2140 /*
2141 * Adjust the RXWIN0SENT flag - indicate that we have advertised
2142 * a 0 window. This may cause the remote transmitter to stall. This
2143 * flag tells soreceive() to disable delayed acknowledgements when
2144 * draining the buffer. This can occur if the receiver is attempting
2145 * to read more data then can be buffered prior to transmitting on
2146 * the connection.
2147 */
2148 if (th->th_win == 0)
2149 tp->t_flags |= TF_RXWIN0SENT;
2150 else
2151 tp->t_flags &= ~TF_RXWIN0SENT;
2152 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
2153 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
2154 th->th_flags |= TH_URG;
2155 } else {
2156 /*
2157 * If no urgent pointer to send, then we pull
2158 * the urgent pointer to the left edge of the send window
2159 * so that it doesn't drift into the send window on sequence
2160 * number wraparound.
2161 */
2162 tp->snd_up = tp->snd_una; /* drag it along */
2163 }
2164
2165 /*
2166 * Put TCP length in extended header, and then
2167 * checksum extended header and data.
2168 */
2169 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
2170
2171 /*
2172 * If this is potentially the last packet on the stream, then mark
2173 * it in order to enable some optimizations in the underlying
2174 * layers
2175 */
2176 if (tp->t_state != TCPS_ESTABLISHED &&
2177 (tp->t_state == TCPS_CLOSING || tp->t_state == TCPS_TIME_WAIT
2178 || tp->t_state == TCPS_LAST_ACK || (th->th_flags & TH_RST)))
2179 m->m_pkthdr.pkt_flags |= PKTF_LAST_PKT;
2180
2181 #if INET6
2182 if (isipv6) {
2183 /*
2184 * ip6_plen is not need to be filled now, and will be filled
2185 * in ip6_output.
2186 */
2187 m->m_pkthdr.csum_flags = CSUM_TCPIPV6;
2188 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2189 if (len + optlen)
2190 th->th_sum = in_addword(th->th_sum,
2191 htons((u_short)(optlen + len)));
2192 }
2193 else
2194 #endif /* INET6 */
2195 {
2196 m->m_pkthdr.csum_flags = CSUM_TCP;
2197 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
2198 if (len + optlen)
2199 th->th_sum = in_addword(th->th_sum,
2200 htons((u_short)(optlen + len)));
2201 }
2202
2203 /*
2204 * Enable TSO and specify the size of the segments.
2205 * The TCP pseudo header checksum is always provided.
2206 */
2207 if (tso) {
2208 #if INET6
2209 if (isipv6)
2210 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV6;
2211 else
2212 #endif /* INET6 */
2213 m->m_pkthdr.csum_flags |= CSUM_TSO_IPV4;
2214
2215 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
2216 } else {
2217 m->m_pkthdr.tso_segsz = 0;
2218 }
2219
2220 /*
2221 * In transmit state, time the transmission and arrange for
2222 * the retransmit. In persist state, just set snd_max.
2223 */
2224 if (!(tp->t_flagsext & TF_FORCE)
2225 || tp->t_timer[TCPT_PERSIST] == 0) {
2226 tcp_seq startseq = tp->snd_nxt;
2227
2228 /*
2229 * Advance snd_nxt over sequence space of this segment.
2230 */
2231 if (flags & (TH_SYN|TH_FIN)) {
2232 if (flags & TH_SYN)
2233 tp->snd_nxt++;
2234 if ((flags & TH_FIN) &&
2235 !(tp->t_flags & TF_SENTFIN)) {
2236 tp->snd_nxt++;
2237 tp->t_flags |= TF_SENTFIN;
2238 }
2239 }
2240 if (sack_rxmit)
2241 goto timer;
2242 if (sack_rescue_rxt == TRUE) {
2243 tp->snd_nxt = old_snd_nxt;
2244 sack_rescue_rxt = FALSE;
2245 tcpstat.tcps_pto_in_recovery++;
2246 } else {
2247 tp->snd_nxt += len;
2248 }
2249 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
2250 tp->snd_max = tp->snd_nxt;
2251 tp->t_sndtime = tcp_now;
2252 /*
2253 * Time this transmission if not a retransmission and
2254 * not currently timing anything.
2255 */
2256 if (tp->t_rtttime == 0) {
2257 tp->t_rtttime = tcp_now;
2258 tp->t_rtseq = startseq;
2259 tcpstat.tcps_segstimed++;
2260
2261 /* update variables related to pipe ack */
2262 tp->t_pipeack_lastuna = tp->snd_una;
2263 }
2264 }
2265
2266 /*
2267 * Set retransmit timer if not currently set,
2268 * and not doing an ack or a keep-alive probe.
2269 */
2270 timer:
2271 if (tp->t_timer[TCPT_REXMT] == 0 &&
2272 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
2273 tp->snd_nxt != tp->snd_una || (flags & TH_FIN))) {
2274 if (tp->t_timer[TCPT_PERSIST]) {
2275 tp->t_timer[TCPT_PERSIST] = 0;
2276 tp->t_persist_stop = 0;
2277 TCP_RESET_REXMT_STATE(tp);
2278 }
2279 tp->t_timer[TCPT_REXMT] =
2280 OFFSET_FROM_START(tp, tp->t_rxtcur);
2281 }
2282
2283 /*
2284 * Set tail loss probe timeout if new data is being
2285 * transmitted. This will be supported only when
2286 * SACK option is enabled on a connection.
2287 *
2288 * Every time new data is sent PTO will get reset.
2289 */
2290 if (tcp_enable_tlp && len != 0 && tp->t_state == TCPS_ESTABLISHED &&
2291 SACK_ENABLED(tp) && !IN_FASTRECOVERY(tp) &&
2292 tp->snd_nxt == tp->snd_max &&
2293 SEQ_GT(tp->snd_nxt, tp->snd_una) &&
2294 tp->t_rxtshift == 0 &&
2295 (tp->t_flagsext & (TF_SENT_TLPROBE|TF_PKTS_REORDERED)) == 0) {
2296 u_int32_t pto, srtt;
2297
2298 /*
2299 * Using SRTT alone to set PTO can cause spurious
2300 * retransmissions on wireless networks where there
2301 * is a lot of variance in RTT. Taking variance
2302 * into account will avoid this.
2303 */
2304 srtt = tp->t_srtt >> TCP_RTT_SHIFT;
2305 pto = ((TCP_REXMTVAL(tp)) * 3) >> 1;
2306 pto = max (2 * srtt, pto);
2307 if ((tp->snd_max - tp->snd_una) == tp->t_maxseg)
2308 pto = max(pto,
2309 (((3 * pto) >> 2) + tcp_delack * 2));
2310 else
2311 pto = max(10, pto);
2312
2313 /* if RTO is less than PTO, choose RTO instead */
2314 if (tp->t_rxtcur < pto)
2315 pto = tp->t_rxtcur;
2316
2317 tp->t_timer[TCPT_PTO] = OFFSET_FROM_START(tp, pto);
2318 }
2319 } else {
2320 /*
2321 * Persist case, update snd_max but since we are in
2322 * persist mode (no window) we do not update snd_nxt.
2323 */
2324 int xlen = len;
2325 if (flags & TH_SYN)
2326 ++xlen;
2327 if ((flags & TH_FIN) &&
2328 !(tp->t_flags & TF_SENTFIN)) {
2329 ++xlen;
2330 tp->t_flags |= TF_SENTFIN;
2331 }
2332 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) {
2333 tp->snd_max = tp->snd_nxt + len;
2334 tp->t_sndtime = tcp_now;
2335 }
2336 }
2337
2338 #if TCPDEBUG
2339 /*
2340 * Trace.
2341 */
2342 if (so_options & SO_DEBUG)
2343 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
2344 #endif
2345
2346 /*
2347 * Fill in IP length and desired time to live and
2348 * send to IP level. There should be a better way
2349 * to handle ttl and tos; we could keep them in
2350 * the template, but need a way to checksum without them.
2351 */
2352 #if INET6
2353 /*
2354 * m->m_pkthdr.len should have been set before cksum calcuration,
2355 * because in6_cksum() need it.
2356 */
2357 if (isipv6) {
2358 /*
2359 * we separately set hoplimit for every segment, since the
2360 * user might want to change the value via setsockopt.
2361 * Also, desired default hop limit might be changed via
2362 * Neighbor Discovery.
2363 */
2364 ip6->ip6_hlim = in6_selecthlim(inp, inp->in6p_route.ro_rt ?
2365 inp->in6p_route.ro_rt->rt_ifp : NULL);
2366
2367 /* TODO: IPv6 IP6TOS_ECT bit on */
2368 KERNEL_DEBUG(DBG_LAYER_BEG,
2369 ((inp->inp_fport << 16) | inp->inp_lport),
2370 (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) |
2371 (inp->in6p_faddr.s6_addr16[0] & 0xffff)),
2372 sendalot,0,0);
2373 } else
2374 #endif /* INET6 */
2375 {
2376 ip->ip_len = m->m_pkthdr.len;
2377 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
2378 ip->ip_tos |= (inp->inp_ip_tos & ~IPTOS_ECN_MASK);/* XXX */
2379 KERNEL_DEBUG(DBG_LAYER_BEG,
2380 ((inp->inp_fport << 16) | inp->inp_lport),
2381 (((inp->inp_laddr.s_addr & 0xffff) << 16) |
2382 (inp->inp_faddr.s_addr & 0xffff)), 0,0,0);
2383 }
2384
2385 /*
2386 * See if we should do MTU discovery.
2387 * Look at the flag updated on the following criterias:
2388 * 1) Path MTU discovery is authorized by the sysctl
2389 * 2) The route isn't set yet (unlikely but could happen)
2390 * 3) The route is up
2391 * 4) the MTU is not locked (if it is, then discovery has been
2392 * disabled for that route)
2393 */
2394 #if INET6
2395 if (!isipv6)
2396 #endif /* INET6 */
2397 if (path_mtu_discovery && (tp->t_flags & TF_PMTUD))
2398 ip->ip_off |= IP_DF;
2399
2400 #if NECP
2401 {
2402 necp_kernel_policy_id policy_id;
2403 necp_kernel_policy_id skip_policy_id;
2404 u_int32_t route_rule_id;
2405 if (!necp_socket_is_allowed_to_send_recv(inp, &policy_id, &route_rule_id, &skip_policy_id)) {
2406 m_freem(m);
2407 error = EHOSTUNREACH;
2408 goto out;
2409 }
2410 necp_mark_packet_from_socket(m, inp, policy_id, route_rule_id, skip_policy_id);
2411
2412 if (net_qos_policy_restricted != 0) {
2413 necp_socket_update_qos_marking(inp, inp->inp_route.ro_rt,
2414 NULL, route_rule_id);
2415 }
2416 }
2417 #endif /* NECP */
2418
2419 #if IPSEC
2420 if (inp->inp_sp != NULL)
2421 ipsec_setsocket(m, so);
2422 #endif /*IPSEC*/
2423
2424 /*
2425 * The socket is kept locked while sending out packets in ip_output, even if packet chaining is not active.
2426 */
2427 lost = 0;
2428
2429 /*
2430 * Embed the flow hash in pkt hdr and mark the packet as
2431 * capable of flow controlling
2432 */
2433 m->m_pkthdr.pkt_flowsrc = FLOWSRC_INPCB;
2434 m->m_pkthdr.pkt_flowid = inp->inp_flowhash;
2435 m->m_pkthdr.pkt_flags |= (PKTF_FLOW_ID | PKTF_FLOW_LOCALSRC | PKTF_FLOW_ADV);
2436 m->m_pkthdr.pkt_proto = IPPROTO_TCP;
2437 m->m_pkthdr.tx_tcp_pid = so->last_pid;
2438 if (so->so_flags & SOF_DELEGATED)
2439 m->m_pkthdr.tx_tcp_e_pid = so->e_pid;
2440 else
2441 m->m_pkthdr.tx_tcp_e_pid = 0;
2442
2443 m->m_nextpkt = NULL;
2444
2445 if (inp->inp_last_outifp != NULL &&
2446 !(inp->inp_last_outifp->if_flags & IFF_LOOPBACK)) {
2447 /* Hint to prioritize this packet if
2448 * 1. if the packet has no data
2449 * 2. the interface supports transmit-start model and did
2450 * not disable ACK prioritization.
2451 * 3. Only ACK flag is set.
2452 * 4. there is no outstanding data on this connection.
2453 */
2454 if (tcp_prioritize_acks != 0 && len == 0 &&
2455 (inp->inp_last_outifp->if_eflags &
2456 (IFEF_TXSTART | IFEF_NOACKPRI)) == IFEF_TXSTART) {
2457 if (th->th_flags == TH_ACK &&
2458 tp->snd_una == tp->snd_max &&
2459 tp->t_timer[TCPT_REXMT] == 0)
2460 svc_flags |= PKT_SCF_TCP_ACK;
2461 if (th->th_flags & TH_SYN)
2462 svc_flags |= PKT_SCF_TCP_SYN;
2463 }
2464 set_packet_service_class(m, so, sotc, svc_flags);
2465 } else {
2466 /*
2467 * Optimization for loopback just set the mbuf
2468 * service class
2469 */
2470 (void) m_set_service_class(m, so_tc2msc(sotc));
2471 }
2472
2473 tp->t_pktlist_sentlen += len;
2474 tp->t_lastchain++;
2475
2476 #if INET6
2477 if (isipv6) {
2478 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2479 struct ip6 *, ip6, struct tcpcb *, tp, struct tcphdr *,
2480 th);
2481 } else
2482 #endif /* INET6 */
2483 {
2484 DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp,
2485 struct ip *, ip, struct tcpcb *, tp, struct tcphdr *, th);
2486 }
2487
2488 if (tp->t_pktlist_head != NULL) {
2489 tp->t_pktlist_tail->m_nextpkt = m;
2490 tp->t_pktlist_tail = m;
2491 } else {
2492 packchain_newlist++;
2493 tp->t_pktlist_head = tp->t_pktlist_tail = m;
2494 }
2495
2496 if ((lro_ackmore) && (!sackoptlen) && (!tp->t_timer[TCPT_PERSIST]) &&
2497 ((th->th_flags & TH_ACK) == TH_ACK) && (!len) &&
2498 (tp->t_state == TCPS_ESTABLISHED)) {
2499 /* For a pure ACK, see if you need to send more of them */
2500 mnext = tcp_send_lroacks(tp, m, th);
2501 if (mnext) {
2502 tp->t_pktlist_tail->m_nextpkt = mnext;
2503 if (mnext->m_nextpkt == NULL) {
2504 tp->t_pktlist_tail = mnext;
2505 tp->t_lastchain++;
2506 } else {
2507 struct mbuf *tail, *next;
2508 next = mnext->m_nextpkt;
2509 tail = next->m_nextpkt;
2510 while (tail) {
2511 next = tail;
2512 tail = tail->m_nextpkt;
2513 tp->t_lastchain++;
2514 }
2515 tp->t_pktlist_tail = next;
2516 }
2517 }
2518 }
2519
2520 if (sendalot == 0 || (tp->t_state != TCPS_ESTABLISHED) ||
2521 (tp->snd_cwnd <= (tp->snd_wnd / 8)) ||
2522 (tp->t_flags & TF_ACKNOW) ||
2523 (tp->t_flagsext & TF_FORCE) ||
2524 tp->t_lastchain >= tcp_packet_chaining) {
2525 error = 0;
2526 while (inp->inp_sndinprog_cnt == 0 &&
2527 tp->t_pktlist_head != NULL) {
2528 packetlist = tp->t_pktlist_head;
2529 packchain_listadd = tp->t_lastchain;
2530 packchain_sent++;
2531 lost = tp->t_pktlist_sentlen;
2532 TCP_PKTLIST_CLEAR(tp);
2533
2534 error = tcp_ip_output(so, tp, packetlist,
2535 packchain_listadd, tp_inp_options,
2536 (so_options & SO_DONTROUTE),
2537 (sack_rxmit || (sack_bytes_rxmt != 0)), isipv6);
2538 if (error) {
2539 /*
2540 * Take into account the rest of unsent
2541 * packets in the packet list for this tcp
2542 * into "lost", since we're about to free
2543 * the whole list below.
2544 */
2545 lost += tp->t_pktlist_sentlen;
2546 break;
2547 } else {
2548 lost = 0;
2549 }
2550 }
2551 /* tcp was closed while we were in ip; resume close */
2552 if (inp->inp_sndinprog_cnt == 0 &&
2553 (tp->t_flags & TF_CLOSING)) {
2554 tp->t_flags &= ~TF_CLOSING;
2555 (void) tcp_close(tp);
2556 return (0);
2557 }
2558 } else {
2559 error = 0;
2560 packchain_looped++;
2561 tcpstat.tcps_sndtotal++;
2562
2563 goto again;
2564 }
2565 if (error) {
2566 /*
2567 * Assume that the packets were lost, so back out the
2568 * sequence number advance, if any. Note that the "lost"
2569 * variable represents the amount of user data sent during
2570 * the recent call to ip_output_list() plus the amount of
2571 * user data in the packet list for this tcp at the moment.
2572 */
2573 if (!(tp->t_flagsext & TF_FORCE)
2574 || tp->t_timer[TCPT_PERSIST] == 0) {
2575 /*
2576 * No need to check for TH_FIN here because
2577 * the TF_SENTFIN flag handles that case.
2578 */
2579 if ((flags & TH_SYN) == 0) {
2580 if (sack_rxmit) {
2581 if (SEQ_GT((p->rxmit - lost),
2582 tp->snd_una)) {
2583 p->rxmit -= lost;
2584 } else {
2585 lost = p->rxmit - tp->snd_una;
2586 p->rxmit = tp->snd_una;
2587 }
2588 tp->sackhint.sack_bytes_rexmit -= lost;
2589 } else {
2590 if (SEQ_GT((tp->snd_nxt - lost),
2591 tp->snd_una))
2592 tp->snd_nxt -= lost;
2593 else
2594 tp->snd_nxt = tp->snd_una;
2595 }
2596 }
2597 }
2598 out:
2599 if (tp->t_pktlist_head != NULL)
2600 m_freem_list(tp->t_pktlist_head);
2601 TCP_PKTLIST_CLEAR(tp);
2602
2603 if (error == ENOBUFS) {
2604 /*
2605 * Set retransmit timer if not currently set
2606 * when we failed to send a segment that can be
2607 * retransmitted (i.e. not pure ack or rst)
2608 */
2609 if (!tp->t_timer[TCPT_REXMT] &&
2610 !tp->t_timer[TCPT_PERSIST] &&
2611 (len != 0 || (flags & (TH_SYN | TH_FIN)) != 0 ||
2612 so->so_snd.sb_cc > 0))
2613 tp->t_timer[TCPT_REXMT] =
2614 OFFSET_FROM_START(tp, tp->t_rxtcur);
2615 tp->snd_cwnd = tp->t_maxseg;
2616 tp->t_bytes_acked = 0;
2617 tcp_check_timer_state(tp);
2618 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2619
2620 tcp_ccdbg_trace(tp, NULL, TCP_CC_OUTPUT_ERROR);
2621 return (0);
2622 }
2623 if (error == EMSGSIZE) {
2624 /*
2625 * ip_output() will have already fixed the route
2626 * for us. tcp_mtudisc() will, as its last action,
2627 * initiate retransmission, so it is important to
2628 * not do so here.
2629 *
2630 * If TSO was active we either got an interface
2631 * without TSO capabilits or TSO was turned off.
2632 * Disable it for this connection as too and
2633 * immediatly retry with MSS sized segments generated
2634 * by this function.
2635 */
2636 if (tso)
2637 tp->t_flags &= ~TF_TSO;
2638
2639 tcp_mtudisc(inp, 0);
2640 tcp_check_timer_state(tp);
2641
2642 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2643 return 0;
2644 }
2645 /*
2646 * Unless this is due to interface restriction policy,
2647 * treat EHOSTUNREACH/ENETDOWN as a soft error.
2648 */
2649 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
2650 TCPS_HAVERCVDSYN(tp->t_state) &&
2651 !inp_restricted_send(inp, inp->inp_last_outifp)) {
2652 tp->t_softerror = error;
2653 error = 0;
2654 }
2655 tcp_check_timer_state(tp);
2656 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0);
2657 return (error);
2658 }
2659
2660 tcpstat.tcps_sndtotal++;
2661
2662 KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,0,0,0,0,0);
2663 if (sendalot)
2664 goto again;
2665
2666 tcp_check_timer_state(tp);
2667 return (0);
2668 }
2669
2670 static int
2671 tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt,
2672 int cnt, struct mbuf *opt, int flags, int sack_in_progress, boolean_t isipv6)
2673 {
2674 int error = 0;
2675 boolean_t chain;
2676 boolean_t unlocked = FALSE;
2677 boolean_t ifdenied = FALSE;
2678 struct inpcb *inp = tp->t_inpcb;
2679 struct ip_out_args ipoa;
2680 struct route ro;
2681 struct ifnet *outif = NULL;
2682
2683 bzero(&ipoa, sizeof(ipoa));
2684 ipoa.ipoa_boundif = IFSCOPE_NONE;
2685 ipoa.ipoa_flags = IPOAF_SELECT_SRCIF | IPOAF_BOUND_SRCADDR;
2686 ipoa.ipoa_sotc = SO_TC_UNSPEC;
2687 ipoa.ipoa_netsvctype = _NET_SERVICE_TYPE_UNSPEC;
2688 #if INET6
2689 struct ip6_out_args ip6oa;
2690 struct route_in6 ro6;
2691
2692 bzero(&ip6oa, sizeof(ip6oa));
2693 ip6oa.ip6oa_boundif = IFSCOPE_NONE;
2694 ip6oa.ip6oa_flags = IP6OAF_SELECT_SRCIF | IP6OAF_BOUND_SRCADDR;
2695 ip6oa.ip6oa_sotc = SO_TC_UNSPEC;
2696 ip6oa.ip6oa_netsvctype = _NET_SERVICE_TYPE_UNSPEC;
2697
2698 struct flowadv *adv =
2699 (isipv6 ? &ip6oa.ip6oa_flowadv : &ipoa.ipoa_flowadv);
2700 #else /* INET6 */
2701 struct flowadv *adv = &ipoa.ipoa_flowadv;
2702 #endif /* !INET6 */
2703
2704 /* If socket was bound to an ifindex, tell ip_output about it */
2705 if (inp->inp_flags & INP_BOUND_IF) {
2706 #if INET6
2707 if (isipv6) {
2708 ip6oa.ip6oa_boundif = inp->inp_boundifp->if_index;
2709 ip6oa.ip6oa_flags |= IP6OAF_BOUND_IF;
2710 } else
2711 #endif /* INET6 */
2712 {
2713 ipoa.ipoa_boundif = inp->inp_boundifp->if_index;
2714 ipoa.ipoa_flags |= IPOAF_BOUND_IF;
2715 }
2716 }
2717
2718 if (INP_NO_CELLULAR(inp)) {
2719 #if INET6
2720 if (isipv6)
2721 ip6oa.ip6oa_flags |= IP6OAF_NO_CELLULAR;
2722 else
2723 #endif /* INET6 */
2724 ipoa.ipoa_flags |= IPOAF_NO_CELLULAR;
2725 }
2726 if (INP_NO_EXPENSIVE(inp)) {
2727 #if INET6
2728 if (isipv6)
2729 ip6oa.ip6oa_flags |= IP6OAF_NO_EXPENSIVE;
2730 else
2731 #endif /* INET6 */
2732 ipoa.ipoa_flags |= IPOAF_NO_EXPENSIVE;
2733
2734 }
2735 if (INP_AWDL_UNRESTRICTED(inp)) {
2736 #if INET6
2737 if (isipv6)
2738 ip6oa.ip6oa_flags |= IP6OAF_AWDL_UNRESTRICTED;
2739 else
2740 #endif /* INET6 */
2741 ipoa.ipoa_flags |= IPOAF_AWDL_UNRESTRICTED;
2742
2743 }
2744 #if INET6
2745 if (INP_INTCOPROC_ALLOWED(inp) && isipv6) {
2746 ip6oa.ip6oa_flags |= IP6OAF_INTCOPROC_ALLOWED;
2747 }
2748 if (isipv6) {
2749 ip6oa.ip6oa_sotc = so->so_traffic_class;
2750 ip6oa.ip6oa_netsvctype = so->so_netsvctype;
2751 } else
2752 #endif /* INET6 */
2753 {
2754 ipoa.ipoa_sotc = so->so_traffic_class;
2755 ipoa.ipoa_netsvctype = so->so_netsvctype;
2756 }
2757 if ((so->so_flags1 & SOF1_QOSMARKING_ALLOWED)) {
2758 #if INET6
2759 if (isipv6)
2760 ip6oa.ip6oa_flags |= IP6OAF_QOSMARKING_ALLOWED;
2761 else
2762 #endif /* INET6 */
2763 ipoa.ipoa_flags |= IPOAF_QOSMARKING_ALLOWED;
2764 }
2765 #if INET6
2766 if (isipv6)
2767 flags |= IPV6_OUTARGS;
2768 else
2769 #endif /* INET6 */
2770 flags |= IP_OUTARGS;
2771
2772 /* Copy the cached route and take an extra reference */
2773 #if INET6
2774 if (isipv6)
2775 in6p_route_copyout(inp, &ro6);
2776 else
2777 #endif /* INET6 */
2778 inp_route_copyout(inp, &ro);
2779
2780 /*
2781 * Make sure ACK/DELACK conditions are cleared before
2782 * we unlock the socket.
2783 */
2784 tp->last_ack_sent = tp->rcv_nxt;
2785 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
2786 tp->t_timer[TCPT_DELACK] = 0;
2787 tp->t_unacksegs = 0;
2788
2789 /* Increment the count of outstanding send operations */
2790 inp->inp_sndinprog_cnt++;
2791
2792 /*
2793 * If allowed, unlock TCP socket while in IP
2794 * but only if the connection is established and
2795 * in a normal mode where reentrancy on the tcpcb won't be
2796 * an issue:
2797 * - there is no SACK episode
2798 * - we're not in Fast Recovery mode
2799 * - if we're not sending from an upcall.
2800 */
2801 if (tcp_output_unlocked && !so->so_upcallusecount &&
2802 (tp->t_state == TCPS_ESTABLISHED) && (sack_in_progress == 0) &&
2803 !IN_FASTRECOVERY(tp) && !(so->so_flags & SOF_MP_SUBFLOW)) {
2804
2805 unlocked = TRUE;
2806 socket_unlock(so, 0);
2807 }
2808
2809 /*
2810 * Don't send down a chain of packets when:
2811 * - TCP chaining is disabled
2812 * - there is an IPsec rule set
2813 * - there is a non default rule set for the firewall
2814 */
2815
2816 chain = tcp_packet_chaining > 1
2817 #if IPSEC
2818 && ipsec_bypass
2819 #endif
2820 #if IPFIREWALL
2821 && (fw_enable == 0 || fw_bypass)
2822 #endif
2823 ; // I'm important, not extraneous
2824
2825
2826 while (pkt != NULL) {
2827 struct mbuf *npkt = pkt->m_nextpkt;
2828
2829 if (!chain) {
2830 pkt->m_nextpkt = NULL;
2831 /*
2832 * If we are not chaining, make sure to set the packet
2833 * list count to 0 so that IP takes the right path;
2834 * this is important for cases such as IPSec where a
2835 * single mbuf might result in multiple mbufs as part
2836 * of the encapsulation. If a non-zero count is passed
2837 * down to IP, the head of the chain might change and
2838 * we could end up skipping it (thus generating bogus
2839 * packets). Fixing it in IP would be desirable, but
2840 * for now this would do it.
2841 */
2842 cnt = 0;
2843 }
2844 #if INET6
2845 if (isipv6) {
2846 error = ip6_output_list(pkt, cnt,
2847 inp->in6p_outputopts, &ro6, flags, NULL, NULL,
2848 &ip6oa);
2849 ifdenied = (ip6oa.ip6oa_retflags & IP6OARF_IFDENIED);
2850 } else {
2851 #endif /* INET6 */
2852 error = ip_output_list(pkt, cnt, opt, &ro, flags, NULL,
2853 &ipoa);
2854 ifdenied = (ipoa.ipoa_retflags & IPOARF_IFDENIED);
2855 }
2856
2857 if (chain || error) {
2858 /*
2859 * If we sent down a chain then we are done since
2860 * the callee had taken care of everything; else
2861 * we need to free the rest of the chain ourselves.
2862 */
2863 if (!chain)
2864 m_freem_list(npkt);
2865 break;
2866 }
2867 pkt = npkt;
2868 }
2869
2870 if (unlocked)
2871 socket_lock(so, 0);
2872
2873 /*
2874 * Enter flow controlled state if the connection is established
2875 * and is not in recovery. Flow control is allowed only if there
2876 * is outstanding data.
2877 *
2878 * A connection will enter suspended state even if it is in
2879 * recovery.
2880 */
2881 if (((adv->code == FADV_FLOW_CONTROLLED && !IN_FASTRECOVERY(tp)) ||
2882 adv->code == FADV_SUSPENDED) &&
2883 !(tp->t_flags & TF_CLOSING) &&
2884 tp->t_state == TCPS_ESTABLISHED &&
2885 SEQ_GT(tp->snd_max, tp->snd_una)) {
2886 int rc;
2887 rc = inp_set_fc_state(inp, adv->code);
2888
2889 if (rc == 1)
2890 tcp_ccdbg_trace(tp, NULL,
2891 ((adv->code == FADV_FLOW_CONTROLLED) ?
2892 TCP_CC_FLOW_CONTROL : TCP_CC_SUSPEND));
2893 }
2894
2895 /*
2896 * When an interface queue gets suspended, some of the
2897 * packets are dropped. Return ENOBUFS, to update the
2898 * pcb state.
2899 */
2900 if (adv->code == FADV_SUSPENDED)
2901 error = ENOBUFS;
2902
2903 VERIFY(inp->inp_sndinprog_cnt > 0);
2904 if ( --inp->inp_sndinprog_cnt == 0)
2905 inp->inp_flags &= ~(INP_FC_FEEDBACK);
2906
2907 #if INET6
2908 if (isipv6) {
2909 if (ro6.ro_rt != NULL)
2910 outif = ro6.ro_rt->rt_ifp;
2911 } else
2912 #endif /* INET6 */
2913 if (ro.ro_rt != NULL)
2914 outif = ro.ro_rt->rt_ifp;
2915
2916 if (outif != NULL && outif != inp->inp_last_outifp &&
2917 so->so_snd.sb_cc > 0) {
2918 /* Update the send byte count */
2919 if (so->so_snd.sb_flags & SB_SNDBYTE_CNT) {
2920 inp_decr_sndbytes_total(so, so->so_snd.sb_cc);
2921 inp_decr_sndbytes_allunsent(so, tp->snd_una);
2922 so->so_snd.sb_flags &= ~SB_SNDBYTE_CNT;
2923 }
2924 inp->inp_last_outifp = outif;
2925
2926 }
2927
2928 if (error != 0 && ifdenied &&
2929 (INP_NO_CELLULAR(inp) || INP_NO_EXPENSIVE(inp)))
2930 soevent(so,
2931 (SO_FILT_HINT_LOCKED|SO_FILT_HINT_IFDENIED));
2932
2933 /* Synchronize cached PCB route & options */
2934 #if INET6
2935 if (isipv6)
2936 in6p_route_copyin(inp, &ro6);
2937 else
2938 #endif /* INET6 */
2939 inp_route_copyin(inp, &ro);
2940
2941 if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift == 0 &&
2942 tp->t_inpcb->inp_route.ro_rt != NULL) {
2943 /* If we found the route and there is an rtt on it
2944 * reset the retransmit timer
2945 */
2946 tcp_getrt_rtt(tp, tp->t_inpcb->in6p_route.ro_rt);
2947 tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur);
2948 }
2949 return (error);
2950 }
2951
2952 int tcptv_persmin_val = TCPTV_PERSMIN;
2953
2954 void
2955 tcp_setpersist(struct tcpcb *tp)
2956 {
2957 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
2958
2959 /* If a PERSIST_TIMER option was set we will limit the
2960 * time the persist timer will be active for that connection
2961 * in order to avoid DOS by using zero window probes.
2962 * see rdar://5805356
2963 */
2964
2965 if ((tp->t_persist_timeout != 0) &&
2966 (tp->t_timer[TCPT_PERSIST] == 0) &&
2967 (tp->t_persist_stop == 0)) {
2968 tp->t_persist_stop = tcp_now + tp->t_persist_timeout;
2969 }
2970
2971 /*
2972 * Start/restart persistance timer.
2973 */
2974 TCPT_RANGESET(tp->t_timer[TCPT_PERSIST],
2975 t * tcp_backoff[tp->t_rxtshift],
2976 tcptv_persmin_val, TCPTV_PERSMAX, 0);
2977 tp->t_timer[TCPT_PERSIST] = OFFSET_FROM_START(tp, tp->t_timer[TCPT_PERSIST]);
2978
2979 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
2980 tp->t_rxtshift++;
2981 }
2982
2983 /*
2984 * Send as many acks as data coalesced. Every other packet when stretch
2985 * ACK is not enabled. Every 8 packets, if stretch ACK is enabled.
2986 */
2987 static struct mbuf*
2988 tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th)
2989 {
2990 struct mbuf *mnext = NULL, *ack_chain = NULL, *tail = NULL;
2991 int count = 0;
2992 tcp_seq org_ack = ntohl(th->th_ack);
2993 tcp_seq prev_ack = 0;
2994 int tack_offset = 28; /* IPv6 and IP options not supported */
2995 int twin_offset = 34; /* IPv6 and IP options not supported */
2996 int ack_size = (tp->t_flags & TF_STRETCHACK) ?
2997 (maxseg_unacked * tp->t_maxseg) : (tp->t_maxseg << 1);
2998 int segs_acked = (tp->t_flags & TF_STRETCHACK) ? maxseg_unacked : 2;
2999 struct mbuf *prev_ack_pkt = NULL;
3000 struct socket *so = tp->t_inpcb->inp_socket;
3001 unsigned short winsz = ntohs(th->th_win);
3002 unsigned int scaled_win = winsz<<tp->rcv_scale;
3003 tcp_seq win_rtedge = org_ack + scaled_win;
3004
3005 count = tp->t_lropktlen/tp->t_maxseg;
3006
3007 prev_ack = (org_ack - tp->t_lropktlen) + ack_size;
3008 if (prev_ack < org_ack) {
3009 ack_chain = m_dup(m, M_DONTWAIT);
3010 if (ack_chain) {
3011 th->th_ack = htonl(prev_ack);
3012 /* Keep adv window constant for duplicated ACK packets */
3013 scaled_win = win_rtedge - prev_ack;
3014 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3015 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3016 th->th_win = htons(scaled_win>>tp->rcv_scale);
3017 if (lrodebug == 5) {
3018 printf("%s: win = %d winsz = %d sc = %d"
3019 " lro_len %d %d\n",
3020 __func__, scaled_win>>tp->rcv_scale, winsz,
3021 tp->rcv_scale, tp->t_lropktlen, count);
3022 }
3023 tail = ack_chain;
3024 count -= segs_acked; /* accounts for prev_ack packet */
3025 count = (count <= segs_acked) ? 0 : count - segs_acked;
3026 tcpstat.tcps_sndacks++;
3027 so_tc_update_stats(m, so, m_get_service_class(m));
3028 } else {
3029 return NULL;
3030 }
3031 }
3032 else {
3033 tp->t_lropktlen = 0;
3034 return NULL;
3035 }
3036
3037 prev_ack_pkt = ack_chain;
3038
3039 while (count > 0) {
3040 if ((prev_ack + ack_size) < org_ack) {
3041 prev_ack += ack_size;
3042 } else {
3043 /*
3044 * The last ACK sent must have the ACK number that TCP
3045 * thinks is the last sent ACK number.
3046 */
3047 prev_ack = org_ack;
3048 }
3049 mnext = m_dup(prev_ack_pkt, M_DONTWAIT);
3050 if (mnext) {
3051 /* Keep adv window constant for duplicated ACK packets */
3052 scaled_win = win_rtedge - prev_ack;
3053 if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale))
3054 scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale);
3055 winsz = htons(scaled_win>>tp->rcv_scale);
3056 if (lrodebug == 5) {
3057 printf("%s: winsz = %d ack %x count %d\n",
3058 __func__, scaled_win>>tp->rcv_scale,
3059 prev_ack, count);
3060 }
3061 bcopy(&winsz, mtod(prev_ack_pkt, caddr_t) + twin_offset, 2);
3062 HTONL(prev_ack);
3063 bcopy(&prev_ack, mtod(prev_ack_pkt, caddr_t) + tack_offset, 4);
3064 NTOHL(prev_ack);
3065 tail->m_nextpkt = mnext;
3066 tail = mnext;
3067 count -= segs_acked;
3068 tcpstat.tcps_sndacks++;
3069 so_tc_update_stats(m, so, m_get_service_class(m));
3070 } else {
3071 if (lrodebug == 5) {
3072 printf("%s: failed to alloc mbuf.\n", __func__);
3073 }
3074 break;
3075 }
3076 prev_ack_pkt = mnext;
3077 }
3078 tp->t_lropktlen = 0;
3079 return ack_chain;
3080 }
3081
3082 static int
3083 tcp_recv_throttle (struct tcpcb *tp)
3084 {
3085 uint32_t base_rtt, newsize;
3086 struct sockbuf *sbrcv = &tp->t_inpcb->inp_socket->so_rcv;
3087
3088 if (tcp_use_rtt_recvbg == 1 &&
3089 TSTMP_SUPPORTED(tp)) {
3090 /*
3091 * Timestamps are supported on this connection. Use
3092 * RTT to look for an increase in latency.
3093 */
3094
3095 /*
3096 * If the connection is already being throttled, leave it
3097 * in that state until rtt comes closer to base rtt
3098 */
3099 if (tp->t_flagsext & TF_RECV_THROTTLE)
3100 return (1);
3101
3102 base_rtt = get_base_rtt(tp);
3103
3104 if (base_rtt != 0 && tp->t_rttcur != 0) {
3105 /*
3106 * if latency increased on a background flow,
3107 * return 1 to start throttling.
3108 */
3109 if (tp->t_rttcur > (base_rtt + target_qdelay)) {
3110 tp->t_flagsext |= TF_RECV_THROTTLE;
3111 if (tp->t_recv_throttle_ts == 0)
3112 tp->t_recv_throttle_ts = tcp_now;
3113 /*
3114 * Reduce the recv socket buffer size to
3115 * minimize latecy.
3116 */
3117 if (sbrcv->sb_idealsize >
3118 tcp_recv_throttle_minwin) {
3119 newsize = sbrcv->sb_idealsize >> 1;
3120 /* Set a minimum of 16 K */
3121 newsize =
3122 max(newsize,
3123 tcp_recv_throttle_minwin);
3124 sbrcv->sb_idealsize = newsize;
3125 }
3126 return (1);
3127 } else {
3128 return (0);
3129 }
3130 }
3131 }
3132
3133 /*
3134 * Timestamps are not supported or there is no good RTT
3135 * measurement. Use IPDV in this case.
3136 */
3137 if (tp->acc_iaj > tcp_acc_iaj_react_limit)
3138 return (1);
3139
3140 return (0);
3141 }
mirror of XNU