1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 *
31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
32 */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD: releng/12.0/sys/netinet/tcp_output.c 336934 2018-07-30 20:35:50Z tuexen $");
36
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
41
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
45 #ifdef TCP_HHOOK
46 #include <sys/hhook.h>
47 #endif
48 #include <sys/kernel.h>
49 #include <sys/lock.h>
50 #include <sys/mbuf.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
53 #include <sys/sdt.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
57
58 #include <net/if.h>
59 #include <net/route.h>
60 #include <net/vnet.h>
61
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
69 #ifdef INET6
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
73 #endif
74 #include <netinet/tcp.h>
75 #define TCPOUTFLAGS
76 #include <netinet/tcp_fsm.h>
77 #include <netinet/tcp_log_buf.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet/tcpip.h>
82 #include <netinet/cc/cc.h>
83 #include <netinet/tcp_fastopen.h>
84 #ifdef TCPPCAP
85 #include <netinet/tcp_pcap.h>
86 #endif
87 #ifdef TCPDEBUG
88 #include <netinet/tcp_debug.h>
89 #endif
90 #ifdef TCP_OFFLOAD
91 #include <netinet/tcp_offload.h>
92 #endif
93
94 #include <netipsec/ipsec_support.h>
95
96 #include <machine/in_cksum.h>
97
98 #include <security/mac/mac_framework.h>
99
100 VNET_DEFINE(int, path_mtu_discovery) = 1;
101 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
102 &VNET_NAME(path_mtu_discovery), 1,
103 "Enable Path MTU Discovery");
104
105 VNET_DEFINE(int, tcp_do_tso) = 1;
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
107 &VNET_NAME(tcp_do_tso), 0,
108 "Enable TCP Segmentation Offload");
109
110 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
111 #define V_tcp_sendspace VNET(tcp_sendspace)
112 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
113 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
114
115 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
117 &VNET_NAME(tcp_do_autosndbuf), 0,
118 "Enable automatic send buffer sizing");
119
120 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
122 &VNET_NAME(tcp_autosndbuf_inc), 0,
123 "Incrementor step size of automatic send buffer");
124
125 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
126 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
127 &VNET_NAME(tcp_autosndbuf_max), 0,
128 "Max size of automatic send buffer");
129
130 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
131 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
133 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
134 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
135
136 /*
137 * Make sure that either retransmit or persist timer is set for SYN, FIN and
138 * non-ACK.
139 */
140 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
141 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
142 tcp_timer_active((tp), TT_REXMT) || \
143 tcp_timer_active((tp), TT_PERSIST), \
144 ("neither rexmt nor persist timer is set"))
145
146 static void inline cc_after_idle(struct tcpcb *tp);
147
148 #ifdef TCP_HHOOK
149 /*
150 * Wrapper for the TCP established output helper hook.
151 */
152 void
153 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
154 struct tcpopt *to, uint32_t len, int tso)
155 {
156 struct tcp_hhook_data hhook_data;
157
158 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
159 hhook_data.tp = tp;
160 hhook_data.th = th;
161 hhook_data.to = to;
162 hhook_data.len = len;
163 hhook_data.tso = tso;
164
165 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
166 tp->osd);
167 }
168 }
169 #endif
170
171 /*
172 * CC wrapper hook functions
173 */
174 static void inline
175 cc_after_idle(struct tcpcb *tp)
176 {
177 INP_WLOCK_ASSERT(tp->t_inpcb);
178
179 if (CC_ALGO(tp)->after_idle != NULL)
180 CC_ALGO(tp)->after_idle(tp->ccv);
181 }
182
183 /*
184 * Tcp output routine: figure out what should be sent and send it.
185 */
186 int
187 tcp_output(struct tcpcb *tp)
188 {
189 struct socket *so = tp->t_inpcb->inp_socket;
190 int32_t len;
191 uint32_t recwin, sendwin;
192 int off, flags, error = 0; /* Keep compiler happy */
193 u_int if_hw_tsomaxsegcount = 0;
194 u_int if_hw_tsomaxsegsize;
195 struct mbuf *m;
196 struct ip *ip = NULL;
197 #ifdef TCPDEBUG
198 struct ipovly *ipov = NULL;
199 #endif
200 struct tcphdr *th;
201 u_char opt[TCP_MAXOLEN];
202 unsigned ipoptlen, optlen, hdrlen;
203 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
204 unsigned ipsec_optlen = 0;
205 #endif
206 int idle, sendalot, curticks;
207 int sack_rxmit, sack_bytes_rxmt;
208 struct sackhole *p;
209 int tso, mtu;
210 struct tcpopt to;
211 unsigned int wanted_cookie = 0;
212 unsigned int dont_sendalot = 0;
213 #if 0
214 int maxburst = TCP_MAXBURST;
215 #endif
216 #ifdef INET6
217 struct ip6_hdr *ip6 = NULL;
218 int isipv6;
219
220 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
221 #endif
222
223 INP_WLOCK_ASSERT(tp->t_inpcb);
224
225 #ifdef TCP_OFFLOAD
226 if (tp->t_flags & TF_TOE)
227 return (tcp_offload_output(tp));
228 #endif
229
230 /*
231 * For TFO connections in SYN_SENT or SYN_RECEIVED,
232 * only allow the initial SYN or SYN|ACK and those sent
233 * by the retransmit timer.
234 */
235 if (IS_FASTOPEN(tp->t_flags) &&
236 ((tp->t_state == TCPS_SYN_SENT) ||
237 (tp->t_state == TCPS_SYN_RECEIVED)) &&
238 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
239 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
240 return (0);
241
242 /*
243 * Determine length of data that should be transmitted,
244 * and flags that will be used.
245 * If there is some data or critical controls (SYN, RST)
246 * to send, then transmit; otherwise, investigate further.
247 */
248 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
249 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
250 cc_after_idle(tp);
251 tp->t_flags &= ~TF_LASTIDLE;
252 if (idle) {
253 if (tp->t_flags & TF_MORETOCOME) {
254 tp->t_flags |= TF_LASTIDLE;
255 idle = 0;
256 }
257 }
258 again:
259 /*
260 * If we've recently taken a timeout, snd_max will be greater than
261 * snd_nxt. There may be SACK information that allows us to avoid
262 * resending already delivered data. Adjust snd_nxt accordingly.
263 */
264 if ((tp->t_flags & TF_SACK_PERMIT) &&
265 SEQ_LT(tp->snd_nxt, tp->snd_max))
266 tcp_sack_adjust(tp);
267 sendalot = 0;
268 tso = 0;
269 mtu = 0;
270 off = tp->snd_nxt - tp->snd_una;
271 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
272
273 flags = tcp_outflags[tp->t_state];
274 /*
275 * Send any SACK-generated retransmissions. If we're explicitly trying
276 * to send out new data (when sendalot is 1), bypass this function.
277 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
278 * we're replacing a (future) new transmission with a retransmission
279 * now, and we previously incremented snd_cwnd in tcp_input().
280 */
281 /*
282 * Still in sack recovery , reset rxmit flag to zero.
283 */
284 sack_rxmit = 0;
285 sack_bytes_rxmt = 0;
286 len = 0;
287 p = NULL;
288 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
289 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
290 uint32_t cwin;
291
292 cwin =
293 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
294 /* Do not retransmit SACK segments beyond snd_recover */
295 if (SEQ_GT(p->end, tp->snd_recover)) {
296 /*
297 * (At least) part of sack hole extends beyond
298 * snd_recover. Check to see if we can rexmit data
299 * for this hole.
300 */
301 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
302 /*
303 * Can't rexmit any more data for this hole.
304 * That data will be rexmitted in the next
305 * sack recovery episode, when snd_recover
306 * moves past p->rxmit.
307 */
308 p = NULL;
309 goto after_sack_rexmit;
310 } else
311 /* Can rexmit part of the current hole */
312 len = ((int32_t)ulmin(cwin,
313 tp->snd_recover - p->rxmit));
314 } else
315 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
316 off = p->rxmit - tp->snd_una;
317 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
318 __func__, off));
319 if (len > 0) {
320 sack_rxmit = 1;
321 sendalot = 1;
322 TCPSTAT_INC(tcps_sack_rexmits);
323 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
324 min(len, tp->t_maxseg));
325 }
326 }
327 after_sack_rexmit:
328 /*
329 * Get standard flags, and add SYN or FIN if requested by 'hidden'
330 * state flags.
331 */
332 if (tp->t_flags & TF_NEEDFIN)
333 flags |= TH_FIN;
334 if (tp->t_flags & TF_NEEDSYN)
335 flags |= TH_SYN;
336
337 SOCKBUF_LOCK(&so->so_snd);
338 /*
339 * If in persist timeout with window of 0, send 1 byte.
340 * Otherwise, if window is small but nonzero
341 * and timer expired, we will send what we can
342 * and go to transmit state.
343 */
344 if (tp->t_flags & TF_FORCEDATA) {
345 if (sendwin == 0) {
346 /*
347 * If we still have some data to send, then
348 * clear the FIN bit. Usually this would
349 * happen below when it realizes that we
350 * aren't sending all the data. However,
351 * if we have exactly 1 byte of unsent data,
352 * then it won't clear the FIN bit below,
353 * and if we are in persist state, we wind
354 * up sending the packet without recording
355 * that we sent the FIN bit.
356 *
357 * We can't just blindly clear the FIN bit,
358 * because if we don't have any more data
359 * to send then the probe will be the FIN
360 * itself.
361 */
362 if (off < sbused(&so->so_snd))
363 flags &= ~TH_FIN;
364 sendwin = 1;
365 } else {
366 tcp_timer_activate(tp, TT_PERSIST, 0);
367 tp->t_rxtshift = 0;
368 }
369 }
370
371 /*
372 * If snd_nxt == snd_max and we have transmitted a FIN, the
373 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
374 * a negative length. This can also occur when TCP opens up
375 * its congestion window while receiving additional duplicate
376 * acks after fast-retransmit because TCP will reset snd_nxt
377 * to snd_max after the fast-retransmit.
378 *
379 * In the normal retransmit-FIN-only case, however, snd_nxt will
380 * be set to snd_una, the offset will be 0, and the length may
381 * wind up 0.
382 *
383 * If sack_rxmit is true we are retransmitting from the scoreboard
384 * in which case len is already set.
385 */
386 if (sack_rxmit == 0) {
387 if (sack_bytes_rxmt == 0)
388 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
389 off);
390 else {
391 int32_t cwin;
392
393 /*
394 * We are inside of a SACK recovery episode and are
395 * sending new data, having retransmitted all the
396 * data possible in the scoreboard.
397 */
398 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
399 off);
400 /*
401 * Don't remove this (len > 0) check !
402 * We explicitly check for len > 0 here (although it
403 * isn't really necessary), to work around a gcc
404 * optimization issue - to force gcc to compute
405 * len above. Without this check, the computation
406 * of len is bungled by the optimizer.
407 */
408 if (len > 0) {
409 cwin = tp->snd_cwnd -
410 (tp->snd_nxt - tp->sack_newdata) -
411 sack_bytes_rxmt;
412 if (cwin < 0)
413 cwin = 0;
414 len = imin(len, cwin);
415 }
416 }
417 }
418
419 /*
420 * Lop off SYN bit if it has already been sent. However, if this
421 * is SYN-SENT state and if segment contains data and if we don't
422 * know that foreign host supports TAO, suppress sending segment.
423 */
424 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
425 if (tp->t_state != TCPS_SYN_RECEIVED)
426 flags &= ~TH_SYN;
427 /*
428 * When sending additional segments following a TFO SYN|ACK,
429 * do not include the SYN bit.
430 */
431 if (IS_FASTOPEN(tp->t_flags) &&
432 (tp->t_state == TCPS_SYN_RECEIVED))
433 flags &= ~TH_SYN;
434 off--, len++;
435 }
436
437 /*
438 * Be careful not to send data and/or FIN on SYN segments.
439 * This measure is needed to prevent interoperability problems
440 * with not fully conformant TCP implementations.
441 */
442 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
443 len = 0;
444 flags &= ~TH_FIN;
445 }
446
447 /*
448 * On TFO sockets, ensure no data is sent in the following cases:
449 *
450 * - When retransmitting SYN|ACK on a passively-created socket
451 *
452 * - When retransmitting SYN on an actively created socket
453 *
454 * - When sending a zero-length cookie (cookie request) on an
455 * actively created socket
456 *
457 * - When the socket is in the CLOSED state (RST is being sent)
458 */
459 if (IS_FASTOPEN(tp->t_flags) &&
460 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
461 ((tp->t_state == TCPS_SYN_SENT) &&
462 (tp->t_tfo_client_cookie_len == 0)) ||
463 (flags & TH_RST)))
464 len = 0;
465 if (len <= 0) {
466 /*
467 * If FIN has been sent but not acked,
468 * but we haven't been called to retransmit,
469 * len will be < 0. Otherwise, window shrank
470 * after we sent into it. If window shrank to 0,
471 * cancel pending retransmit, pull snd_nxt back
472 * to (closed) window, and set the persist timer
473 * if it isn't already going. If the window didn't
474 * close completely, just wait for an ACK.
475 *
476 * We also do a general check here to ensure that
477 * we will set the persist timer when we have data
478 * to send, but a 0-byte window. This makes sure
479 * the persist timer is set even if the packet
480 * hits one of the "goto send" lines below.
481 */
482 len = 0;
483 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
484 (off < (int) sbavail(&so->so_snd))) {
485 tcp_timer_activate(tp, TT_REXMT, 0);
486 tp->t_rxtshift = 0;
487 tp->snd_nxt = tp->snd_una;
488 if (!tcp_timer_active(tp, TT_PERSIST))
489 tcp_setpersist(tp);
490 }
491 }
492
493 /* len will be >= 0 after this point. */
494 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
495
496 tcp_sndbuf_autoscale(tp, so, sendwin);
497
498 /*
499 * Decide if we can use TCP Segmentation Offloading (if supported by
500 * hardware).
501 *
502 * TSO may only be used if we are in a pure bulk sending state. The
503 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
504 * IP options prevent using TSO. With TSO the TCP header is the same
505 * (except for the sequence number) for all generated packets. This
506 * makes it impossible to transmit any options which vary per generated
507 * segment or packet.
508 *
509 * IPv4 handling has a clear separation of ip options and ip header
510 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
511 * the right thing below to provide length of just ip options and thus
512 * checking for ipoptlen is enough to decide if ip options are present.
513 */
514 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
515 /*
516 * Pre-calculate here as we save another lookup into the darknesses
517 * of IPsec that way and can actually decide if TSO is ok.
518 */
519 #ifdef INET6
520 if (isipv6 && IPSEC_ENABLED(ipv6))
521 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
522 #ifdef INET
523 else
524 #endif
525 #endif /* INET6 */
526 #ifdef INET
527 if (IPSEC_ENABLED(ipv4))
528 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
529 #endif /* INET */
530 #endif /* IPSEC */
531 #ifdef INET6
532 if (isipv6)
533 ipoptlen = ip6_optlen(tp->t_inpcb);
534 else
535 #endif
536 if (tp->t_inpcb->inp_options)
537 ipoptlen = tp->t_inpcb->inp_options->m_len -
538 offsetof(struct ipoption, ipopt_list);
539 else
540 ipoptlen = 0;
541 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
542 ipoptlen += ipsec_optlen;
543 #endif
544
545 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
546 ((tp->t_flags & TF_SIGNATURE) == 0) &&
547 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
548 ipoptlen == 0 && !(flags & TH_SYN))
549 tso = 1;
550
551 if (sack_rxmit) {
552 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
553 flags &= ~TH_FIN;
554 } else {
555 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
556 sbused(&so->so_snd)))
557 flags &= ~TH_FIN;
558 }
559
560 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
561 (long)TCP_MAXWIN << tp->rcv_scale);
562
563 /*
564 * Sender silly window avoidance. We transmit under the following
565 * conditions when len is non-zero:
566 *
567 * - We have a full segment (or more with TSO)
568 * - This is the last buffer in a write()/send() and we are
569 * either idle or running NODELAY
570 * - we've timed out (e.g. persist timer)
571 * - we have more then 1/2 the maximum send window's worth of
572 * data (receiver may be limited the window size)
573 * - we need to retransmit
574 */
575 if (len) {
576 if (len >= tp->t_maxseg)
577 goto send;
578 /*
579 * NOTE! on localhost connections an 'ack' from the remote
580 * end may occur synchronously with the output and cause
581 * us to flush a buffer queued with moretocome. XXX
582 *
583 * note: the len + off check is almost certainly unnecessary.
584 */
585 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
586 (idle || (tp->t_flags & TF_NODELAY)) &&
587 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
588 (tp->t_flags & TF_NOPUSH) == 0) {
589 goto send;
590 }
591 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
592 goto send;
593 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
594 goto send;
595 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
596 goto send;
597 if (sack_rxmit)
598 goto send;
599 }
600
601 /*
602 * Sending of standalone window updates.
603 *
604 * Window updates are important when we close our window due to a
605 * full socket buffer and are opening it again after the application
606 * reads data from it. Once the window has opened again and the
607 * remote end starts to send again the ACK clock takes over and
608 * provides the most current window information.
609 *
610 * We must avoid the silly window syndrome whereas every read
611 * from the receive buffer, no matter how small, causes a window
612 * update to be sent. We also should avoid sending a flurry of
613 * window updates when the socket buffer had queued a lot of data
614 * and the application is doing small reads.
615 *
616 * Prevent a flurry of pointless window updates by only sending
617 * an update when we can increase the advertized window by more
618 * than 1/4th of the socket buffer capacity. When the buffer is
619 * getting full or is very small be more aggressive and send an
620 * update whenever we can increase by two mss sized segments.
621 * In all other situations the ACK's to new incoming data will
622 * carry further window increases.
623 *
624 * Don't send an independent window update if a delayed
625 * ACK is pending (it will get piggy-backed on it) or the
626 * remote side already has done a half-close and won't send
627 * more data. Skip this if the connection is in T/TCP
628 * half-open state.
629 */
630 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
631 !(tp->t_flags & TF_DELACK) &&
632 !TCPS_HAVERCVDFIN(tp->t_state)) {
633 /*
634 * "adv" is the amount we could increase the window,
635 * taking into account that we are limited by
636 * TCP_MAXWIN << tp->rcv_scale.
637 */
638 int32_t adv;
639 int oldwin;
640
641 adv = recwin;
642 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
643 oldwin = (tp->rcv_adv - tp->rcv_nxt);
644 adv -= oldwin;
645 } else
646 oldwin = 0;
647
648 /*
649 * If the new window size ends up being the same as or less
650 * than the old size when it is scaled, then don't force
651 * a window update.
652 */
653 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
654 goto dontupdate;
655
656 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
657 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
658 recwin <= (so->so_rcv.sb_hiwat / 8) ||
659 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
660 goto send;
661 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
662 goto send;
663 }
664 dontupdate:
665
666 /*
667 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
668 * is also a catch-all for the retransmit timer timeout case.
669 */
670 if (tp->t_flags & TF_ACKNOW)
671 goto send;
672 if ((flags & TH_RST) ||
673 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
674 goto send;
675 if (SEQ_GT(tp->snd_up, tp->snd_una))
676 goto send;
677 /*
678 * If our state indicates that FIN should be sent
679 * and we have not yet done so, then we need to send.
680 */
681 if (flags & TH_FIN &&
682 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
683 goto send;
684 /*
685 * In SACK, it is possible for tcp_output to fail to send a segment
686 * after the retransmission timer has been turned off. Make sure
687 * that the retransmission timer is set.
688 */
689 if ((tp->t_flags & TF_SACK_PERMIT) &&
690 SEQ_GT(tp->snd_max, tp->snd_una) &&
691 !tcp_timer_active(tp, TT_REXMT) &&
692 !tcp_timer_active(tp, TT_PERSIST)) {
693 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
694 goto just_return;
695 }
696 /*
697 * TCP window updates are not reliable, rather a polling protocol
698 * using ``persist'' packets is used to insure receipt of window
699 * updates. The three ``states'' for the output side are:
700 * idle not doing retransmits or persists
701 * persisting to move a small or zero window
702 * (re)transmitting and thereby not persisting
703 *
704 * tcp_timer_active(tp, TT_PERSIST)
705 * is true when we are in persist state.
706 * (tp->t_flags & TF_FORCEDATA)
707 * is set when we are called to send a persist packet.
708 * tcp_timer_active(tp, TT_REXMT)
709 * is set when we are retransmitting
710 * The output side is idle when both timers are zero.
711 *
712 * If send window is too small, there is data to transmit, and no
713 * retransmit or persist is pending, then go to persist state.
714 * If nothing happens soon, send when timer expires:
715 * if window is nonzero, transmit what we can,
716 * otherwise force out a byte.
717 */
718 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
719 !tcp_timer_active(tp, TT_PERSIST)) {
720 tp->t_rxtshift = 0;
721 tcp_setpersist(tp);
722 }
723
724 /*
725 * No reason to send a segment, just return.
726 */
727 just_return:
728 SOCKBUF_UNLOCK(&so->so_snd);
729 return (0);
730
731 send:
732 SOCKBUF_LOCK_ASSERT(&so->so_snd);
733 if (len > 0) {
734 if (len >= tp->t_maxseg)
735 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
736 else
737 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
738 }
739 /*
740 * Before ESTABLISHED, force sending of initial options
741 * unless TCP set not to do any options.
742 * NOTE: we assume that the IP/TCP header plus TCP options
743 * always fit in a single mbuf, leaving room for a maximum
744 * link header, i.e.
745 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
746 */
747 optlen = 0;
748 #ifdef INET6
749 if (isipv6)
750 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
751 else
752 #endif
753 hdrlen = sizeof (struct tcpiphdr);
754
755 /*
756 * Compute options for segment.
757 * We only have to care about SYN and established connection
758 * segments. Options for SYN-ACK segments are handled in TCP
759 * syncache.
760 */
761 to.to_flags = 0;
762 if ((tp->t_flags & TF_NOOPT) == 0) {
763 /* Maximum segment size. */
764 if (flags & TH_SYN) {
765 tp->snd_nxt = tp->iss;
766 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
767 to.to_flags |= TOF_MSS;
768
769 /*
770 * On SYN or SYN|ACK transmits on TFO connections,
771 * only include the TFO option if it is not a
772 * retransmit, as the presence of the TFO option may
773 * have caused the original SYN or SYN|ACK to have
774 * been dropped by a middlebox.
775 */
776 if (IS_FASTOPEN(tp->t_flags) &&
777 (tp->t_rxtshift == 0)) {
778 if (tp->t_state == TCPS_SYN_RECEIVED) {
779 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
780 to.to_tfo_cookie =
781 (u_int8_t *)&tp->t_tfo_cookie.server;
782 to.to_flags |= TOF_FASTOPEN;
783 wanted_cookie = 1;
784 } else if (tp->t_state == TCPS_SYN_SENT) {
785 to.to_tfo_len =
786 tp->t_tfo_client_cookie_len;
787 to.to_tfo_cookie =
788 tp->t_tfo_cookie.client;
789 to.to_flags |= TOF_FASTOPEN;
790 wanted_cookie = 1;
791 /*
792 * If we wind up having more data to
793 * send with the SYN than can fit in
794 * one segment, don't send any more
795 * until the SYN|ACK comes back from
796 * the other end.
797 */
798 dont_sendalot = 1;
799 }
800 }
801 }
802 /* Window scaling. */
803 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
804 to.to_wscale = tp->request_r_scale;
805 to.to_flags |= TOF_SCALE;
806 }
807 /* Timestamps. */
808 if ((tp->t_flags & TF_RCVD_TSTMP) ||
809 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
810 curticks = tcp_ts_getticks();
811 to.to_tsval = curticks + tp->ts_offset;
812 to.to_tsecr = tp->ts_recent;
813 to.to_flags |= TOF_TS;
814 if (tp->t_rxtshift == 1)
815 tp->t_badrxtwin = curticks;
816 }
817
818 /* Set receive buffer autosizing timestamp. */
819 if (tp->rfbuf_ts == 0 &&
820 (so->so_rcv.sb_flags & SB_AUTOSIZE))
821 tp->rfbuf_ts = tcp_ts_getticks();
822
823 /* Selective ACK's. */
824 if (tp->t_flags & TF_SACK_PERMIT) {
825 if (flags & TH_SYN)
826 to.to_flags |= TOF_SACKPERM;
827 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
828 (tp->t_flags & TF_SACK_PERMIT) &&
829 tp->rcv_numsacks > 0) {
830 to.to_flags |= TOF_SACK;
831 to.to_nsacks = tp->rcv_numsacks;
832 to.to_sacks = (u_char *)tp->sackblks;
833 }
834 }
835 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
836 /* TCP-MD5 (RFC2385). */
837 /*
838 * Check that TCP_MD5SIG is enabled in tcpcb to
839 * account the size needed to set this TCP option.
840 */
841 if (tp->t_flags & TF_SIGNATURE)
842 to.to_flags |= TOF_SIGNATURE;
843 #endif /* TCP_SIGNATURE */
844
845 /* Processing the options. */
846 hdrlen += optlen = tcp_addoptions(&to, opt);
847 /*
848 * If we wanted a TFO option to be added, but it was unable
849 * to fit, ensure no data is sent.
850 */
851 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
852 !(to.to_flags & TOF_FASTOPEN))
853 len = 0;
854 }
855
856 /*
857 * Adjust data length if insertion of options will
858 * bump the packet length beyond the t_maxseg length.
859 * Clear the FIN bit because we cut off the tail of
860 * the segment.
861 */
862 if (len + optlen + ipoptlen > tp->t_maxseg) {
863 flags &= ~TH_FIN;
864
865 if (tso) {
866 u_int if_hw_tsomax;
867 u_int moff;
868 int max_len;
869
870 /* extract TSO information */
871 if_hw_tsomax = tp->t_tsomax;
872 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
873 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
874
875 /*
876 * Limit a TSO burst to prevent it from
877 * overflowing or exceeding the maximum length
878 * allowed by the network interface:
879 */
880 KASSERT(ipoptlen == 0,
881 ("%s: TSO can't do IP options", __func__));
882
883 /*
884 * Check if we should limit by maximum payload
885 * length:
886 */
887 if (if_hw_tsomax != 0) {
888 /* compute maximum TSO length */
889 max_len = (if_hw_tsomax - hdrlen -
890 max_linkhdr);
891 if (max_len <= 0) {
892 len = 0;
893 } else if (len > max_len) {
894 sendalot = 1;
895 len = max_len;
896 }
897 }
898 /*
899 * Prevent the last segment from being
900 * fractional unless the send sockbuf can be
901 * emptied:
902 */
903 max_len = (tp->t_maxseg - optlen);
904 if (((uint32_t)off + (uint32_t)len) <
905 sbavail(&so->so_snd)) {
906 moff = len % max_len;
907 if (moff != 0) {
908 len -= moff;
909 sendalot = 1;
910 }
911 }
912
913 /*
914 * In case there are too many small fragments
915 * don't use TSO:
916 */
917 if (len <= max_len) {
918 len = max_len;
919 sendalot = 1;
920 tso = 0;
921 }
922
923 /*
924 * Send the FIN in a separate segment
925 * after the bulk sending is done.
926 * We don't trust the TSO implementations
927 * to clear the FIN flag on all but the
928 * last segment.
929 */
930 if (tp->t_flags & TF_NEEDFIN)
931 sendalot = 1;
932 } else {
933 len = tp->t_maxseg - optlen - ipoptlen;
934 sendalot = 1;
935 if (dont_sendalot)
936 sendalot = 0;
937 }
938 } else
939 tso = 0;
940
941 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
942 ("%s: len > IP_MAXPACKET", __func__));
943
944 /*#ifdef DIAGNOSTIC*/
945 #ifdef INET6
946 if (max_linkhdr + hdrlen > MCLBYTES)
947 #else
948 if (max_linkhdr + hdrlen > MHLEN)
949 #endif
950 panic("tcphdr too big");
951 /*#endif*/
952
953 /*
954 * This KASSERT is here to catch edge cases at a well defined place.
955 * Before, those had triggered (random) panic conditions further down.
956 */
957 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
958
959 /*
960 * Grab a header mbuf, attaching a copy of data to
961 * be transmitted, and initialize the header from
962 * the template for sends on this connection.
963 */
964 if (len) {
965 struct mbuf *mb;
966 struct sockbuf *msb;
967 u_int moff;
968
969 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
970 TCPSTAT_INC(tcps_sndprobe);
971 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
972 tp->t_sndrexmitpack++;
973 TCPSTAT_INC(tcps_sndrexmitpack);
974 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
975 } else {
976 TCPSTAT_INC(tcps_sndpack);
977 TCPSTAT_ADD(tcps_sndbyte, len);
978 }
979 #ifdef INET6
980 if (MHLEN < hdrlen + max_linkhdr)
981 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
982 else
983 #endif
984 m = m_gethdr(M_NOWAIT, MT_DATA);
985
986 if (m == NULL) {
987 SOCKBUF_UNLOCK(&so->so_snd);
988 error = ENOBUFS;
989 sack_rxmit = 0;
990 goto out;
991 }
992
993 m->m_data += max_linkhdr;
994 m->m_len = hdrlen;
995
996 /*
997 * Start the m_copy functions from the closest mbuf
998 * to the offset in the socket buffer chain.
999 */
1000 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1001 if (len <= MHLEN - hdrlen - max_linkhdr) {
1002 m_copydata(mb, moff, len,
1003 mtod(m, caddr_t) + hdrlen);
1004 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1005 sbsndptr_adv(&so->so_snd, mb, len);
1006 m->m_len += len;
1007 } else {
1008 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1009 msb = NULL;
1010 else
1011 msb = &so->so_snd;
1012 m->m_next = tcp_m_copym(mb, moff,
1013 &len, if_hw_tsomaxsegcount,
1014 if_hw_tsomaxsegsize, msb);
1015 if (len <= (tp->t_maxseg - optlen)) {
1016 /*
1017 * Must have ran out of mbufs for the copy
1018 * shorten it to no longer need tso. Lets
1019 * not put on sendalot since we are low on
1020 * mbufs.
1021 */
1022 tso = 0;
1023 }
1024 if (m->m_next == NULL) {
1025 SOCKBUF_UNLOCK(&so->so_snd);
1026 (void) m_free(m);
1027 error = ENOBUFS;
1028 sack_rxmit = 0;
1029 goto out;
1030 }
1031 }
1032
1033 /*
1034 * If we're sending everything we've got, set PUSH.
1035 * (This will keep happy those implementations which only
1036 * give data to the user when a buffer fills or
1037 * a PUSH comes in.)
1038 */
1039 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1040 !(flags & TH_SYN))
1041 flags |= TH_PUSH;
1042 SOCKBUF_UNLOCK(&so->so_snd);
1043 } else {
1044 SOCKBUF_UNLOCK(&so->so_snd);
1045 if (tp->t_flags & TF_ACKNOW)
1046 TCPSTAT_INC(tcps_sndacks);
1047 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1048 TCPSTAT_INC(tcps_sndctrl);
1049 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1050 TCPSTAT_INC(tcps_sndurg);
1051 else
1052 TCPSTAT_INC(tcps_sndwinup);
1053
1054 m = m_gethdr(M_NOWAIT, MT_DATA);
1055 if (m == NULL) {
1056 error = ENOBUFS;
1057 sack_rxmit = 0;
1058 goto out;
1059 }
1060 #ifdef INET6
1061 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1062 MHLEN >= hdrlen) {
1063 M_ALIGN(m, hdrlen);
1064 } else
1065 #endif
1066 m->m_data += max_linkhdr;
1067 m->m_len = hdrlen;
1068 }
1069 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1070 m->m_pkthdr.rcvif = (struct ifnet *)0;
1071 #ifdef MAC
1072 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1073 #endif
1074 #ifdef INET6
1075 if (isipv6) {
1076 ip6 = mtod(m, struct ip6_hdr *);
1077 th = (struct tcphdr *)(ip6 + 1);
1078 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1079 } else
1080 #endif /* INET6 */
1081 {
1082 ip = mtod(m, struct ip *);
1083 #ifdef TCPDEBUG
1084 ipov = (struct ipovly *)ip;
1085 #endif
1086 th = (struct tcphdr *)(ip + 1);
1087 tcpip_fillheaders(tp->t_inpcb, ip, th);
1088 }
1089
1090 /*
1091 * Fill in fields, remembering maximum advertised
1092 * window for use in delaying messages about window sizes.
1093 * If resending a FIN, be sure not to use a new sequence number.
1094 */
1095 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1096 tp->snd_nxt == tp->snd_max)
1097 tp->snd_nxt--;
1098 /*
1099 * If we are starting a connection, send ECN setup
1100 * SYN packet. If we are on a retransmit, we may
1101 * resend those bits a number of times as per
1102 * RFC 3168.
1103 */
1104 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1105 if (tp->t_rxtshift >= 1) {
1106 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1107 flags |= TH_ECE|TH_CWR;
1108 } else
1109 flags |= TH_ECE|TH_CWR;
1110 }
1111
1112 if (tp->t_state == TCPS_ESTABLISHED &&
1113 (tp->t_flags & TF_ECN_PERMIT)) {
1114 /*
1115 * If the peer has ECN, mark data packets with
1116 * ECN capable transmission (ECT).
1117 * Ignore pure ack packets, retransmissions and window probes.
1118 */
1119 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1120 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1121 #ifdef INET6
1122 if (isipv6)
1123 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1124 else
1125 #endif
1126 ip->ip_tos |= IPTOS_ECN_ECT0;
1127 TCPSTAT_INC(tcps_ecn_ect0);
1128 }
1129
1130 /*
1131 * Reply with proper ECN notifications.
1132 */
1133 if (tp->t_flags & TF_ECN_SND_CWR) {
1134 flags |= TH_CWR;
1135 tp->t_flags &= ~TF_ECN_SND_CWR;
1136 }
1137 if (tp->t_flags & TF_ECN_SND_ECE)
1138 flags |= TH_ECE;
1139 }
1140
1141 /*
1142 * If we are doing retransmissions, then snd_nxt will
1143 * not reflect the first unsent octet. For ACK only
1144 * packets, we do not want the sequence number of the
1145 * retransmitted packet, we want the sequence number
1146 * of the next unsent octet. So, if there is no data
1147 * (and no SYN or FIN), use snd_max instead of snd_nxt
1148 * when filling in ti_seq. But if we are in persist
1149 * state, snd_max might reflect one byte beyond the
1150 * right edge of the window, so use snd_nxt in that
1151 * case, since we know we aren't doing a retransmission.
1152 * (retransmit and persist are mutually exclusive...)
1153 */
1154 if (sack_rxmit == 0) {
1155 if (len || (flags & (TH_SYN|TH_FIN)) ||
1156 tcp_timer_active(tp, TT_PERSIST))
1157 th->th_seq = htonl(tp->snd_nxt);
1158 else
1159 th->th_seq = htonl(tp->snd_max);
1160 } else {
1161 th->th_seq = htonl(p->rxmit);
1162 p->rxmit += len;
1163 tp->sackhint.sack_bytes_rexmit += len;
1164 }
1165 th->th_ack = htonl(tp->rcv_nxt);
1166 if (optlen) {
1167 bcopy(opt, th + 1, optlen);
1168 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1169 }
1170 th->th_flags = flags;
1171 /*
1172 * Calculate receive window. Don't shrink window,
1173 * but avoid silly window syndrome.
1174 */
1175 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1176 recwin < tp->t_maxseg)
1177 recwin = 0;
1178 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1179 recwin < (tp->rcv_adv - tp->rcv_nxt))
1180 recwin = (tp->rcv_adv - tp->rcv_nxt);
1181
1182 /*
1183 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1184 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1185 * case is handled in syncache.
1186 */
1187 if (flags & TH_SYN)
1188 th->th_win = htons((u_short)
1189 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1190 else
1191 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1192
1193 /*
1194 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1195 * a 0 window. This may cause the remote transmitter to stall. This
1196 * flag tells soreceive() to disable delayed acknowledgements when
1197 * draining the buffer. This can occur if the receiver is attempting
1198 * to read more data than can be buffered prior to transmitting on
1199 * the connection.
1200 */
1201 if (th->th_win == 0) {
1202 tp->t_sndzerowin++;
1203 tp->t_flags |= TF_RXWIN0SENT;
1204 } else
1205 tp->t_flags &= ~TF_RXWIN0SENT;
1206 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1207 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1208 th->th_flags |= TH_URG;
1209 } else
1210 /*
1211 * If no urgent pointer to send, then we pull
1212 * the urgent pointer to the left edge of the send window
1213 * so that it doesn't drift into the send window on sequence
1214 * number wraparound.
1215 */
1216 tp->snd_up = tp->snd_una; /* drag it along */
1217
1218 /*
1219 * Put TCP length in extended header, and then
1220 * checksum extended header and data.
1221 */
1222 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1223 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1224
1225 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1226 if (to.to_flags & TOF_SIGNATURE) {
1227 /*
1228 * Calculate MD5 signature and put it into the place
1229 * determined before.
1230 * NOTE: since TCP options buffer doesn't point into
1231 * mbuf's data, calculate offset and use it.
1232 */
1233 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1234 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1235 /*
1236 * Do not send segment if the calculation of MD5
1237 * digest has failed.
1238 */
1239 m_freem(m);
1240 goto out;
1241 }
1242 }
1243 #endif
1244 #ifdef INET6
1245 if (isipv6) {
1246 /*
1247 * There is no need to fill in ip6_plen right now.
1248 * It will be filled later by ip6_output.
1249 */
1250 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1251 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1252 optlen + len, IPPROTO_TCP, 0);
1253 }
1254 #endif
1255 #if defined(INET6) && defined(INET)
1256 else
1257 #endif
1258 #ifdef INET
1259 {
1260 m->m_pkthdr.csum_flags = CSUM_TCP;
1261 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1262 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1263
1264 /* IP version must be set here for ipv4/ipv6 checking later */
1265 KASSERT(ip->ip_v == IPVERSION,
1266 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1267 }
1268 #endif
1269
1270 /*
1271 * Enable TSO and specify the size of the segments.
1272 * The TCP pseudo header checksum is always provided.
1273 */
1274 if (tso) {
1275 KASSERT(len > tp->t_maxseg - optlen,
1276 ("%s: len <= tso_segsz", __func__));
1277 m->m_pkthdr.csum_flags |= CSUM_TSO;
1278 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1279 }
1280
1281 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1282 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1283 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1284 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1285 #else
1286 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1287 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1288 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1289 #endif
1290
1291 #ifdef TCP_HHOOK
1292 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1293 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1294 #endif
1295
1296 #ifdef TCPDEBUG
1297 /*
1298 * Trace.
1299 */
1300 if (so->so_options & SO_DEBUG) {
1301 u_short save = 0;
1302 #ifdef INET6
1303 if (!isipv6)
1304 #endif
1305 {
1306 save = ipov->ih_len;
1307 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1308 }
1309 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1310 #ifdef INET6
1311 if (!isipv6)
1312 #endif
1313 ipov->ih_len = save;
1314 }
1315 #endif /* TCPDEBUG */
1316 TCP_PROBE3(debug__output, tp, th, m);
1317
1318 /* We're getting ready to send; log now. */
1319 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1320 len, NULL, false);
1321
1322 /*
1323 * Fill in IP length and desired time to live and
1324 * send to IP level. There should be a better way
1325 * to handle ttl and tos; we could keep them in
1326 * the template, but need a way to checksum without them.
1327 */
1328 /*
1329 * m->m_pkthdr.len should have been set before checksum calculation,
1330 * because in6_cksum() need it.
1331 */
1332 #ifdef INET6
1333 if (isipv6) {
1334 /*
1335 * we separately set hoplimit for every segment, since the
1336 * user might want to change the value via setsockopt.
1337 * Also, desired default hop limit might be changed via
1338 * Neighbor Discovery.
1339 */
1340 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1341
1342 /*
1343 * Set the packet size here for the benefit of DTrace probes.
1344 * ip6_output() will set it properly; it's supposed to include
1345 * the option header lengths as well.
1346 */
1347 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1348
1349 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1350 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1351 else
1352 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1353
1354 if (tp->t_state == TCPS_SYN_SENT)
1355 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1356
1357 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1358
1359 #ifdef TCPPCAP
1360 /* Save packet, if requested. */
1361 tcp_pcap_add(th, m, &(tp->t_outpkts));
1362 #endif
1363
1364 /* TODO: IPv6 IP6TOS_ECT bit on */
1365 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1366 &tp->t_inpcb->inp_route6,
1367 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1368 NULL, NULL, tp->t_inpcb);
1369
1370 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1371 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1372 }
1373 #endif /* INET6 */
1374 #if defined(INET) && defined(INET6)
1375 else
1376 #endif
1377 #ifdef INET
1378 {
1379 ip->ip_len = htons(m->m_pkthdr.len);
1380 #ifdef INET6
1381 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1382 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1383 #endif /* INET6 */
1384 /*
1385 * If we do path MTU discovery, then we set DF on every packet.
1386 * This might not be the best thing to do according to RFC3390
1387 * Section 2. However the tcp hostcache migitates the problem
1388 * so it affects only the first tcp connection with a host.
1389 *
1390 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1391 */
1392 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1393 ip->ip_off |= htons(IP_DF);
1394 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1395 } else {
1396 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1397 }
1398
1399 if (tp->t_state == TCPS_SYN_SENT)
1400 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1401
1402 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1403
1404 #ifdef TCPPCAP
1405 /* Save packet, if requested. */
1406 tcp_pcap_add(th, m, &(tp->t_outpkts));
1407 #endif
1408
1409 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1410 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1411 tp->t_inpcb);
1412
1413 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1414 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1415 }
1416 #endif /* INET */
1417
1418 out:
1419 /*
1420 * In transmit state, time the transmission and arrange for
1421 * the retransmit. In persist state, just set snd_max.
1422 */
1423 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1424 !tcp_timer_active(tp, TT_PERSIST)) {
1425 tcp_seq startseq = tp->snd_nxt;
1426
1427 /*
1428 * Advance snd_nxt over sequence space of this segment.
1429 */
1430 if (flags & (TH_SYN|TH_FIN)) {
1431 if (flags & TH_SYN)
1432 tp->snd_nxt++;
1433 if (flags & TH_FIN) {
1434 tp->snd_nxt++;
1435 tp->t_flags |= TF_SENTFIN;
1436 }
1437 }
1438 if (sack_rxmit)
1439 goto timer;
1440 tp->snd_nxt += len;
1441 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1442 tp->snd_max = tp->snd_nxt;
1443 /*
1444 * Time this transmission if not a retransmission and
1445 * not currently timing anything.
1446 */
1447 if (tp->t_rtttime == 0) {
1448 tp->t_rtttime = ticks;
1449 tp->t_rtseq = startseq;
1450 TCPSTAT_INC(tcps_segstimed);
1451 }
1452 }
1453
1454 /*
1455 * Set retransmit timer if not currently set,
1456 * and not doing a pure ack or a keep-alive probe.
1457 * Initial value for retransmit timer is smoothed
1458 * round-trip time + 2 * round-trip time variance.
1459 * Initialize shift counter which is used for backoff
1460 * of retransmit time.
1461 */
1462 timer:
1463 if (!tcp_timer_active(tp, TT_REXMT) &&
1464 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1465 (tp->snd_nxt != tp->snd_una))) {
1466 if (tcp_timer_active(tp, TT_PERSIST)) {
1467 tcp_timer_activate(tp, TT_PERSIST, 0);
1468 tp->t_rxtshift = 0;
1469 }
1470 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1471 } else if (len == 0 && sbavail(&so->so_snd) &&
1472 !tcp_timer_active(tp, TT_REXMT) &&
1473 !tcp_timer_active(tp, TT_PERSIST)) {
1474 /*
1475 * Avoid a situation where we do not set persist timer
1476 * after a zero window condition. For example:
1477 * 1) A -> B: packet with enough data to fill the window
1478 * 2) B -> A: ACK for #1 + new data (0 window
1479 * advertisement)
1480 * 3) A -> B: ACK for #2, 0 len packet
1481 *
1482 * In this case, A will not activate the persist timer,
1483 * because it chose to send a packet. Unless tcp_output
1484 * is called for some other reason (delayed ack timer,
1485 * another input packet from B, socket syscall), A will
1486 * not send zero window probes.
1487 *
1488 * So, if you send a 0-length packet, but there is data
1489 * in the socket buffer, and neither the rexmt or
1490 * persist timer is already set, then activate the
1491 * persist timer.
1492 */
1493 tp->t_rxtshift = 0;
1494 tcp_setpersist(tp);
1495 }
1496 } else {
1497 /*
1498 * Persist case, update snd_max but since we are in
1499 * persist mode (no window) we do not update snd_nxt.
1500 */
1501 int xlen = len;
1502 if (flags & TH_SYN)
1503 ++xlen;
1504 if (flags & TH_FIN) {
1505 ++xlen;
1506 tp->t_flags |= TF_SENTFIN;
1507 }
1508 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1509 tp->snd_max = tp->snd_nxt + xlen;
1510 }
1511
1512 if (error) {
1513 /* Record the error. */
1514 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1515 error, 0, NULL, false);
1516
1517 /*
1518 * We know that the packet was lost, so back out the
1519 * sequence number advance, if any.
1520 *
1521 * If the error is EPERM the packet got blocked by the
1522 * local firewall. Normally we should terminate the
1523 * connection but the blocking may have been spurious
1524 * due to a firewall reconfiguration cycle. So we treat
1525 * it like a packet loss and let the retransmit timer and
1526 * timeouts do their work over time.
1527 * XXX: It is a POLA question whether calling tcp_drop right
1528 * away would be the really correct behavior instead.
1529 */
1530 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1531 !tcp_timer_active(tp, TT_PERSIST)) &&
1532 ((flags & TH_SYN) == 0) &&
1533 (error != EPERM)) {
1534 if (sack_rxmit) {
1535 p->rxmit -= len;
1536 tp->sackhint.sack_bytes_rexmit -= len;
1537 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1538 ("sackhint bytes rtx >= 0"));
1539 } else
1540 tp->snd_nxt -= len;
1541 }
1542 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1543 switch (error) {
1544 case EACCES:
1545 case EPERM:
1546 tp->t_softerror = error;
1547 return (error);
1548 case ENOBUFS:
1549 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1550 tp->snd_cwnd = tp->t_maxseg;
1551 return (0);
1552 case EMSGSIZE:
1553 /*
1554 * For some reason the interface we used initially
1555 * to send segments changed to another or lowered
1556 * its MTU.
1557 * If TSO was active we either got an interface
1558 * without TSO capabilits or TSO was turned off.
1559 * If we obtained mtu from ip_output() then update
1560 * it and try again.
1561 */
1562 if (tso)
1563 tp->t_flags &= ~TF_TSO;
1564 if (mtu != 0) {
1565 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1566 goto again;
1567 }
1568 return (error);
1569 case EHOSTDOWN:
1570 case EHOSTUNREACH:
1571 case ENETDOWN:
1572 case ENETUNREACH:
1573 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1574 tp->t_softerror = error;
1575 return (0);
1576 }
1577 /* FALLTHROUGH */
1578 default:
1579 return (error);
1580 }
1581 }
1582 TCPSTAT_INC(tcps_sndtotal);
1583
1584 /*
1585 * Data sent (as far as we can tell).
1586 * If this advertises a larger window than any other segment,
1587 * then remember the size of the advertised window.
1588 * Any pending ACK has now been sent.
1589 */
1590 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1591 tp->rcv_adv = tp->rcv_nxt + recwin;
1592 tp->last_ack_sent = tp->rcv_nxt;
1593 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1594 if (tcp_timer_active(tp, TT_DELACK))
1595 tcp_timer_activate(tp, TT_DELACK, 0);
1596 #if 0
1597 /*
1598 * This completely breaks TCP if newreno is turned on. What happens
1599 * is that if delayed-acks are turned on on the receiver, this code
1600 * on the transmitter effectively destroys the TCP window, forcing
1601 * it to four packets (1.5Kx4 = 6K window).
1602 */
1603 if (sendalot && --maxburst)
1604 goto again;
1605 #endif
1606 if (sendalot)
1607 goto again;
1608 return (0);
1609 }
1610
1611 void
1612 tcp_setpersist(struct tcpcb *tp)
1613 {
1614 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1615 int tt;
1616
1617 tp->t_flags &= ~TF_PREVVALID;
1618 if (tcp_timer_active(tp, TT_REXMT))
1619 panic("tcp_setpersist: retransmit pending");
1620 /*
1621 * Start/restart persistence timer.
1622 */
1623 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1624 tcp_persmin, tcp_persmax);
1625 tcp_timer_activate(tp, TT_PERSIST, tt);
1626 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1627 tp->t_rxtshift++;
1628 }
1629
1630 /*
1631 * Insert TCP options according to the supplied parameters to the place
1632 * optp in a consistent way. Can handle unaligned destinations.
1633 *
1634 * The order of the option processing is crucial for optimal packing and
1635 * alignment for the scarce option space.
1636 *
1637 * The optimal order for a SYN/SYN-ACK segment is:
1638 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1639 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1640 *
1641 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1642 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1643 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1644 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1645 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1646 */
1647 int
1648 tcp_addoptions(struct tcpopt *to, u_char *optp)
1649 {
1650 u_int32_t mask, optlen = 0;
1651
1652 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1653 if ((to->to_flags & mask) != mask)
1654 continue;
1655 if (optlen == TCP_MAXOLEN)
1656 break;
1657 switch (to->to_flags & mask) {
1658 case TOF_MSS:
1659 while (optlen % 4) {
1660 optlen += TCPOLEN_NOP;
1661 *optp++ = TCPOPT_NOP;
1662 }
1663 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1664 continue;
1665 optlen += TCPOLEN_MAXSEG;
1666 *optp++ = TCPOPT_MAXSEG;
1667 *optp++ = TCPOLEN_MAXSEG;
1668 to->to_mss = htons(to->to_mss);
1669 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1670 optp += sizeof(to->to_mss);
1671 break;
1672 case TOF_SCALE:
1673 while (!optlen || optlen % 2 != 1) {
1674 optlen += TCPOLEN_NOP;
1675 *optp++ = TCPOPT_NOP;
1676 }
1677 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1678 continue;
1679 optlen += TCPOLEN_WINDOW;
1680 *optp++ = TCPOPT_WINDOW;
1681 *optp++ = TCPOLEN_WINDOW;
1682 *optp++ = to->to_wscale;
1683 break;
1684 case TOF_SACKPERM:
1685 while (optlen % 2) {
1686 optlen += TCPOLEN_NOP;
1687 *optp++ = TCPOPT_NOP;
1688 }
1689 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1690 continue;
1691 optlen += TCPOLEN_SACK_PERMITTED;
1692 *optp++ = TCPOPT_SACK_PERMITTED;
1693 *optp++ = TCPOLEN_SACK_PERMITTED;
1694 break;
1695 case TOF_TS:
1696 while (!optlen || optlen % 4 != 2) {
1697 optlen += TCPOLEN_NOP;
1698 *optp++ = TCPOPT_NOP;
1699 }
1700 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1701 continue;
1702 optlen += TCPOLEN_TIMESTAMP;
1703 *optp++ = TCPOPT_TIMESTAMP;
1704 *optp++ = TCPOLEN_TIMESTAMP;
1705 to->to_tsval = htonl(to->to_tsval);
1706 to->to_tsecr = htonl(to->to_tsecr);
1707 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1708 optp += sizeof(to->to_tsval);
1709 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1710 optp += sizeof(to->to_tsecr);
1711 break;
1712 case TOF_SIGNATURE:
1713 {
1714 int siglen = TCPOLEN_SIGNATURE - 2;
1715
1716 while (!optlen || optlen % 4 != 2) {
1717 optlen += TCPOLEN_NOP;
1718 *optp++ = TCPOPT_NOP;
1719 }
1720 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1721 to->to_flags &= ~TOF_SIGNATURE;
1722 continue;
1723 }
1724 optlen += TCPOLEN_SIGNATURE;
1725 *optp++ = TCPOPT_SIGNATURE;
1726 *optp++ = TCPOLEN_SIGNATURE;
1727 to->to_signature = optp;
1728 while (siglen--)
1729 *optp++ = 0;
1730 break;
1731 }
1732 case TOF_SACK:
1733 {
1734 int sackblks = 0;
1735 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1736 tcp_seq sack_seq;
1737
1738 while (!optlen || optlen % 4 != 2) {
1739 optlen += TCPOLEN_NOP;
1740 *optp++ = TCPOPT_NOP;
1741 }
1742 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1743 continue;
1744 optlen += TCPOLEN_SACKHDR;
1745 *optp++ = TCPOPT_SACK;
1746 sackblks = min(to->to_nsacks,
1747 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1748 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1749 while (sackblks--) {
1750 sack_seq = htonl(sack->start);
1751 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1752 optp += sizeof(sack_seq);
1753 sack_seq = htonl(sack->end);
1754 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1755 optp += sizeof(sack_seq);
1756 optlen += TCPOLEN_SACK;
1757 sack++;
1758 }
1759 TCPSTAT_INC(tcps_sack_send_blocks);
1760 break;
1761 }
1762 case TOF_FASTOPEN:
1763 {
1764 int total_len;
1765
1766 /* XXX is there any point to aligning this option? */
1767 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1768 if (TCP_MAXOLEN - optlen < total_len) {
1769 to->to_flags &= ~TOF_FASTOPEN;
1770 continue;
1771 }
1772 *optp++ = TCPOPT_FAST_OPEN;
1773 *optp++ = total_len;
1774 if (to->to_tfo_len > 0) {
1775 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1776 optp += to->to_tfo_len;
1777 }
1778 optlen += total_len;
1779 break;
1780 }
1781 default:
1782 panic("%s: unknown TCP option type", __func__);
1783 break;
1784 }
1785 }
1786
1787 /* Terminate and pad TCP options to a 4 byte boundary. */
1788 if (optlen % 4) {
1789 optlen += TCPOLEN_EOL;
1790 *optp++ = TCPOPT_EOL;
1791 }
1792 /*
1793 * According to RFC 793 (STD0007):
1794 * "The content of the header beyond the End-of-Option option
1795 * must be header padding (i.e., zero)."
1796 * and later: "The padding is composed of zeros."
1797 */
1798 while (optlen % 4) {
1799 optlen += TCPOLEN_PAD;
1800 *optp++ = TCPOPT_PAD;
1801 }
1802
1803 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1804 return (optlen);
1805 }
1806
1807 /*
1808 * This is a copy of m_copym(), taking the TSO segment size/limit
1809 * constraints into account, and advancing the sndptr as it goes.
1810 */
1811 struct mbuf *
1812 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1813 int32_t seglimit, int32_t segsize, struct sockbuf *sb)
1814 {
1815 struct mbuf *n, **np;
1816 struct mbuf *top;
1817 int32_t off = off0;
1818 int32_t len = *plen;
1819 int32_t fragsize;
1820 int32_t len_cp = 0;
1821 int32_t *pkthdrlen;
1822 uint32_t mlen, frags;
1823 bool copyhdr;
1824
1825
1826 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1827 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1828 if (off == 0 && m->m_flags & M_PKTHDR)
1829 copyhdr = true;
1830 else
1831 copyhdr = false;
1832 while (off > 0) {
1833 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1834 if (off < m->m_len)
1835 break;
1836 off -= m->m_len;
1837 if ((sb) && (m == sb->sb_sndptr)) {
1838 sb->sb_sndptroff += m->m_len;
1839 sb->sb_sndptr = m->m_next;
1840 }
1841 m = m->m_next;
1842 }
1843 np = ⊤
1844 top = NULL;
1845 pkthdrlen = NULL;
1846 while (len > 0) {
1847 if (m == NULL) {
1848 KASSERT(len == M_COPYALL,
1849 ("tcp_m_copym, length > size of mbuf chain"));
1850 *plen = len_cp;
1851 if (pkthdrlen != NULL)
1852 *pkthdrlen = len_cp;
1853 break;
1854 }
1855 mlen = min(len, m->m_len - off);
1856 if (seglimit) {
1857 /*
1858 * For M_NOMAP mbufs, add 3 segments
1859 * + 1 in case we are crossing page boundaries
1860 * + 2 in case the TLS hdr/trailer are used
1861 * It is cheaper to just add the segments
1862 * than it is to take the cache miss to look
1863 * at the mbuf ext_pgs state in detail.
1864 */
1865 if (m->m_flags & M_NOMAP) {
1866 fragsize = min(segsize, PAGE_SIZE);
1867 frags = 3;
1868 } else {
1869 fragsize = segsize;
1870 frags = 0;
1871 }
1872
1873 /* Break if we really can't fit anymore. */
1874 if ((frags + 1) >= seglimit) {
1875 *plen = len_cp;
1876 if (pkthdrlen != NULL)
1877 *pkthdrlen = len_cp;
1878 break;
1879 }
1880
1881 /*
1882 * Reduce size if you can't copy the whole
1883 * mbuf. If we can't copy the whole mbuf, also
1884 * adjust len so the loop will end after this
1885 * mbuf.
1886 */
1887 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1888 mlen = (seglimit - frags - 1) * fragsize;
1889 len = mlen;
1890 *plen = len_cp + len;
1891 if (pkthdrlen != NULL)
1892 *pkthdrlen = *plen;
1893 }
1894 frags += howmany(mlen, fragsize);
1895 if (frags == 0)
1896 frags++;
1897 seglimit -= frags;
1898 KASSERT(seglimit > 0,
1899 ("%s: seglimit went too low", __func__));
1900 }
1901 if (copyhdr)
1902 n = m_gethdr(M_NOWAIT, m->m_type);
1903 else
1904 n = m_get(M_NOWAIT, m->m_type);
1905 *np = n;
1906 if (n == NULL)
1907 goto nospace;
1908 if (copyhdr) {
1909 if (!m_dup_pkthdr(n, m, M_NOWAIT))
1910 goto nospace;
1911 if (len == M_COPYALL)
1912 n->m_pkthdr.len -= off0;
1913 else
1914 n->m_pkthdr.len = len;
1915 pkthdrlen = &n->m_pkthdr.len;
1916 copyhdr = false;
1917 }
1918 n->m_len = mlen;
1919 len_cp += n->m_len;
1920 if (m->m_flags & M_EXT) {
1921 n->m_data = m->m_data + off;
1922 mb_dupcl(n, m);
1923 } else
1924 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1925 (u_int)n->m_len);
1926
1927 if (sb && (sb->sb_sndptr == m) &&
1928 ((n->m_len + off) >= m->m_len) && m->m_next) {
1929 sb->sb_sndptroff += m->m_len;
1930 sb->sb_sndptr = m->m_next;
1931 }
1932 off = 0;
1933 if (len != M_COPYALL) {
1934 len -= n->m_len;
1935 }
1936 m = m->m_next;
1937 np = &n->m_next;
1938 }
1939 return (top);
1940 nospace:
1941 m_freem(top);
1942 return (NULL);
1943 }
1944
1945 void
1946 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
1947 {
1948
1949 /*
1950 * Automatic sizing of send socket buffer. Often the send buffer
1951 * size is not optimally adjusted to the actual network conditions
1952 * at hand (delay bandwidth product). Setting the buffer size too
1953 * small limits throughput on links with high bandwidth and high
1954 * delay (eg. trans-continental/oceanic links). Setting the
1955 * buffer size too big consumes too much real kernel memory,
1956 * especially with many connections on busy servers.
1957 *
1958 * The criteria to step up the send buffer one notch are:
1959 * 1. receive window of remote host is larger than send buffer
1960 * (with a fudge factor of 5/4th);
1961 * 2. send buffer is filled to 7/8th with data (so we actually
1962 * have data to make use of it);
1963 * 3. send buffer fill has not hit maximal automatic size;
1964 * 4. our send window (slow start and cogestion controlled) is
1965 * larger than sent but unacknowledged data in send buffer.
1966 *
1967 * The remote host receive window scaling factor may limit the
1968 * growing of the send buffer before it reaches its allowed
1969 * maximum.
1970 *
1971 * It scales directly with slow start or congestion window
1972 * and does at most one step per received ACK. This fast
1973 * scaling has the drawback of growing the send buffer beyond
1974 * what is strictly necessary to make full use of a given
1975 * delay*bandwidth product. However testing has shown this not
1976 * to be much of an problem. At worst we are trading wasting
1977 * of available bandwidth (the non-use of it) for wasting some
1978 * socket buffer memory.
1979 *
1980 * TODO: Shrink send buffer during idle periods together
1981 * with congestion window. Requires another timer. Has to
1982 * wait for upcoming tcp timer rewrite.
1983 *
1984 * XXXGL: should there be used sbused() or sbavail()?
1985 */
1986 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
1987 int lowat;
1988
1989 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
1990 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
1991 sbused(&so->so_snd) >=
1992 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
1993 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
1994 sendwin >= (sbused(&so->so_snd) -
1995 (tp->snd_nxt - tp->snd_una))) {
1996 if (!sbreserve_locked(&so->so_snd,
1997 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
1998 V_tcp_autosndbuf_max), so, curthread))
1999 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2000 }
2001 }
2002 }
Cache object: d7edc070d204011cf5f9e9a2a6437bad
|