1 /*-
2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 *
29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
30 */
31
32 #include <errno.h>
33 #include <string.h>
34
35 #include "../tcplp.h"
36 #include "tcp.h"
37 #include "tcp_fastopen.h"
38 #include "tcp_fsm.h"
39 #include "tcp_var.h"
40 #include "tcp_seq.h"
41 #include "tcp_timer.h"
42 #include "ip.h"
43 #include "../lib/cbuf.h"
44
45 #include "tcp_const.h"
46
47 #include <openthread/ip6.h>
48 #include <openthread/message.h>
49 #include <openthread/tcp.h>
50
51 static inline void
cc_after_idle(struct tcpcb * tp)52 cc_after_idle(struct tcpcb *tp)
53 {
54 /* samkumar: Removed synchronization. */
55 if (CC_ALGO(tp)->after_idle != NULL)
56 CC_ALGO(tp)->after_idle(tp->ccv);
57 }
58
min(long a,long b)59 long min(long a, long b) {
60 if (a < b) {
61 return a;
62 } else {
63 return b;
64 }
65 }
66
ulmin(unsigned long a,unsigned long b)67 unsigned long ulmin(unsigned long a, unsigned long b) {
68 if (a < b) {
69 return a;
70 } else {
71 return b;
72 }
73 }
74
75 #define lmin(a, b) min(a, b)
76
77 void
tcp_setpersist(struct tcpcb * tp)78 tcp_setpersist(struct tcpcb *tp)
79 {
80 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
81 int tt;
82
83 tp->t_flags &= ~TF_PREVVALID;
84 if (tcp_timer_active(tp, TT_REXMT))
85 tcplp_sys_panic("PANIC: tcp_setpersist: retransmit pending");
86 /*
87 * Start/restart persistance timer.
88 */
89 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
90 TCPTV_PERSMIN, TCPTV_PERSMAX);
91 tcp_timer_activate(tp, TT_PERSIST, tt);
92 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
93 tp->t_rxtshift++;
94 }
95
96 /*
97 * Tcp output routine: figure out what should be sent and send it.
98 */
99 int
tcp_output(struct tcpcb * tp)100 tcp_output(struct tcpcb *tp)
101 {
102 /*
103 * samkumar: The biggest change in this function is in how outgoing
104 * segments are built and sent out. That code has been updated to account
105 * for TCPlp's buffering, and using otMessages rather than mbufs to
106 * construct the outgoing segments.
107 *
108 * And, of course, all code corresponding to locks, stats, and debugging
109 * has been removed, and all code specific to IPv4 or to decide between
110 * IPv6 and IPv4 handling has been removed.
111 */
112
113 struct tcphdr* th = NULL;
114 int idle;
115 long len, recwin, sendwin;
116 int off, flags, error = 0; /* Keep compiler happy */
117 int sendalot, mtu;
118 int sack_rxmit, sack_bytes_rxmt;
119 struct sackhole* p;
120 unsigned ipoptlen, optlen, hdrlen;
121 struct tcpopt to;
122 unsigned int wanted_cookie = 0;
123 unsigned int dont_sendalot = 0;
124 uint8_t opt[TCP_MAXOLEN];
125 uint32_t ticks = tcplp_sys_get_ticks();
126
127 /* samkumar: Code for TCP offload has been removed. */
128
129 /*
130 * For TFO connections in SYN_SENT or SYN_RECEIVED,
131 * only allow the initial SYN or SYN|ACK and those sent
132 * by the retransmit timer.
133 */
134 if (IS_FASTOPEN(tp->t_flags) &&
135 ((tp->t_state == TCPS_SYN_SENT) ||
136 (tp->t_state == TCPS_SYN_RECEIVED)) &&
137 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
138 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
139 return (0);
140
141 /*
142 * Determine length of data that should be transmitted,
143 * and flags that will be used.
144 * If there is some data or critical controls (SYN, RST)
145 * to send, then transmit; otherwise, investigate further.
146 */
147 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
148 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
149 cc_after_idle(tp);
150
151 tp->t_flags &= ~TF_LASTIDLE;
152 if (idle) {
153 if (tp->t_flags & TF_MORETOCOME) {
154 tp->t_flags |= TF_LASTIDLE;
155 idle = 0;
156 }
157 }
158 /* samkumar: This would be printed once per _window_ that is transmitted. */
159 #ifdef INSTRUMENT_TCP
160 tcplp_sys_log("TCP output %u %d %d", (unsigned int) tcplp_sys_get_millis(), (int) tp->snd_wnd, (int) tp->snd_cwnd);
161 #endif
162
163 again:
164 /*
165 * If we've recently taken a timeout, snd_max will be greater than
166 * snd_nxt. There may be SACK information that allows us to avoid
167 * resending already delivered data. Adjust snd_nxt accordingly.
168 */
169 if ((tp->t_flags & TF_SACK_PERMIT) &&
170 SEQ_LT(tp->snd_nxt, tp->snd_max))
171 tcp_sack_adjust(tp);
172 sendalot = 0;
173 /* samkumar: Removed code for supporting TSO. */
174 mtu = 0;
175 off = tp->snd_nxt - tp->snd_una;
176 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
177
178 flags = tcp_outflags[tp->t_state];
179 /*
180 * Send any SACK-generated retransmissions. If we're explicitly trying
181 * to send out new data (when sendalot is 1), bypass this function.
182 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
183 * we're replacing a (future) new transmission with a retransmission
184 * now, and we previously incremented snd_cwnd in tcp_input().
185 */
186 /*
187 * Still in sack recovery , reset rxmit flag to zero.
188 */
189 sack_rxmit = 0;
190 sack_bytes_rxmt = 0;
191 len = 0;
192 p = NULL;
193 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
194 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
195 long cwin;
196
197 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
198 if (cwin < 0)
199 cwin = 0;
200 /* Do not retransmit SACK segments beyond snd_recover */
201 if (SEQ_GT(p->end, tp->snd_recover)) {
202 /*
203 * (At least) part of sack hole extends beyond
204 * snd_recover. Check to see if we can rexmit data
205 * for this hole.
206 */
207 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
208 /*
209 * Can't rexmit any more data for this hole.
210 * That data will be rexmitted in the next
211 * sack recovery episode, when snd_recover
212 * moves past p->rxmit.
213 */
214 p = NULL;
215 goto after_sack_rexmit;
216 } else
217 /* Can rexmit part of the current hole */
218 len = ((long)ulmin(cwin,
219 tp->snd_recover - p->rxmit));
220 } else
221 len = ((long)ulmin(cwin, p->end - p->rxmit));
222 off = p->rxmit - tp->snd_una;
223 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
224 __func__, off));
225 if (len > 0) {
226 sack_rxmit = 1;
227 sendalot = 1;
228 }
229 }
230 after_sack_rexmit:
231 /*
232 * Get standard flags, and add SYN or FIN if requested by 'hidden'
233 * state flags.
234 */
235 if (tp->t_flags & TF_NEEDFIN)
236 flags |= TH_FIN;
237 if (tp->t_flags & TF_NEEDSYN)
238 flags |= TH_SYN;
239
240 /*
241 * If in persist timeout with window of 0, send 1 byte.
242 * Otherwise, if window is small but nonzero
243 * and timer expired, we will send what we can
244 * and go to transmit state.
245 */
246 if (tp->t_flags & TF_FORCEDATA) {
247 if (sendwin == 0) {
248 /*
249 * If we still have some data to send, then
250 * clear the FIN bit. Usually this would
251 * happen below when it realizes that we
252 * aren't sending all the data. However,
253 * if we have exactly 1 byte of unsent data,
254 * then it won't clear the FIN bit below,
255 * and if we are in persist state, we wind
256 * up sending the packet without recording
257 * that we sent the FIN bit.
258 *
259 * We can't just blindly clear the FIN bit,
260 * because if we don't have any more data
261 * to send then the probe will be the FIN
262 * itself.
263 */
264 /*
265 * samkumar: Replaced call to sbused(&so->so_snd) with the call to
266 * lbuf_used_space below.
267 */
268 if (off < lbuf_used_space(&tp->sendbuf))
269 flags &= ~TH_FIN;
270 sendwin = 1;
271 } else {
272 tcp_timer_activate(tp, TT_PERSIST, 0);
273 tp->t_rxtshift = 0;
274 }
275 }
276
277 /*
278 * If snd_nxt == snd_max and we have transmitted a FIN, the
279 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
280 * a negative length. This can also occur when TCP opens up
281 * its congestion window while receiving additional duplicate
282 * acks after fast-retransmit because TCP will reset snd_nxt
283 * to snd_max after the fast-retransmit.
284 *
285 * In the normal retransmit-FIN-only case, however, snd_nxt will
286 * be set to snd_una, the offset will be 0, and the length may
287 * wind up 0.
288 *
289 * If sack_rxmit is true we are retransmitting from the scoreboard
290 * in which case len is already set.
291 */
292 if (sack_rxmit == 0) {
293 if (sack_bytes_rxmt == 0)
294 /*
295 * samkumar: Replaced sbavail(&so->so_snd) with this call to
296 * lbuf_used_space.
297 */
298 len = ((long)ulmin(lbuf_used_space(&tp->sendbuf), sendwin) -
299 off);
300 else {
301 long cwin;
302
303 /*
304 * We are inside of a SACK recovery episode and are
305 * sending new data, having retransmitted all the
306 * data possible in the scoreboard.
307 */
308 /*
309 * samkumar: Replaced sbavail(&so->so_snd) with this call to
310 * lbuf_used_space.
311 */
312 len = ((long)ulmin(lbuf_used_space(&tp->sendbuf), tp->snd_wnd) -
313 off);
314 /*
315 * Don't remove this (len > 0) check !
316 * We explicitly check for len > 0 here (although it
317 * isn't really necessary), to work around a gcc
318 * optimization issue - to force gcc to compute
319 * len above. Without this check, the computation
320 * of len is bungled by the optimizer.
321 */
322 if (len > 0) {
323 cwin = tp->snd_cwnd -
324 (tp->snd_nxt - tp->sack_newdata) -
325 sack_bytes_rxmt;
326 if (cwin < 0)
327 cwin = 0;
328 len = lmin(len, cwin);
329 }
330 }
331 }
332
333 /*
334 * Lop off SYN bit if it has already been sent. However, if this
335 * is SYN-SENT state and if segment contains data and if we don't
336 * know that foreign host supports TAO, suppress sending segment.
337 */
338 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
339 if (tp->t_state != TCPS_SYN_RECEIVED)
340 flags &= ~TH_SYN;
341 /*
342 * When sending additional segments following a TFO SYN|ACK,
343 * do not include the SYN bit.
344 */
345 if (IS_FASTOPEN(tp->t_flags) &&
346 (tp->t_state == TCPS_SYN_RECEIVED))
347 flags &= ~TH_SYN;
348 off--, len++;
349 }
350
351 /*
352 * Be careful not to send data and/or FIN on SYN segments.
353 * This measure is needed to prevent interoperability problems
354 * with not fully conformant TCP implementations.
355 */
356 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
357 len = 0;
358 flags &= ~TH_FIN;
359 }
360
361 /*
362 * On TFO sockets, ensure no data is sent in the following cases:
363 *
364 * - When retransmitting SYN|ACK on a passively-created socket
365 *
366 * - When retransmitting SYN on an actively created socket
367 *
368 * - When sending a zero-length cookie (cookie request) on an
369 * actively created socket
370 *
371 * - When the socket is in the CLOSED state (RST is being sent)
372 */
373 /*
374 * samkumar: I commented out the check to ensure no data is sent
375 * on a TFO cookie request. As far as I am aware, this is still
376 * compliant with the RFC.
377 */
378 if (IS_FASTOPEN(tp->t_flags) &&
379 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
380 /*((tp->t_state == TCPS_SYN_SENT) &&
381 (tp->t_tfo_client_cookie_len == 0)) ||*/
382 (flags & TH_RST)))
383 len = 0;
384 if (len <= 0) {
385 /*
386 * If FIN has been sent but not acked,
387 * but we haven't been called to retransmit,
388 * len will be < 0. Otherwise, window shrank
389 * after we sent into it. If window shrank to 0,
390 * cancel pending retransmit, pull snd_nxt back
391 * to (closed) window, and set the persist timer
392 * if it isn't already going. If the window didn't
393 * close completely, just wait for an ACK.
394 *
395 * We also do a general check here to ensure that
396 * we will set the persist timer when we have data
397 * to send, but a 0-byte window. This makes sure
398 * the persist timer is set even if the packet
399 * hits one of the "goto send" lines below.
400 */
401 len = 0;
402 /*
403 * samkumar: Replaced sbavail(&so->so_snd) with this call to
404 * lbuf_used_space.
405 */
406 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
407 (off < (int) lbuf_used_space(&tp->sendbuf))) {
408 tcp_timer_activate(tp, TT_REXMT, 0);
409 tp->t_rxtshift = 0;
410 tp->snd_nxt = tp->snd_una;
411 if (!tcp_timer_active(tp, TT_PERSIST)) {
412 tcp_setpersist(tp);
413 }
414 }
415 }
416
417
418 /* len will be >= 0 after this point. */
419 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
420
421 /*
422 * Automatic sizing of send socket buffer. Often the send buffer
423 * size is not optimally adjusted to the actual network conditions
424 * at hand (delay bandwidth product). Setting the buffer size too
425 * small limits throughput on links with high bandwidth and high
426 * delay (eg. trans-continental/oceanic links). Setting the
427 * buffer size too big consumes too much real kernel memory,
428 * especially with many connections on busy servers.
429 *
430 * The criteria to step up the send buffer one notch are:
431 * 1. receive window of remote host is larger than send buffer
432 * (with a fudge factor of 5/4th);
433 * 2. send buffer is filled to 7/8th with data (so we actually
434 * have data to make use of it);
435 * 3. send buffer fill has not hit maximal automatic size;
436 * 4. our send window (slow start and cogestion controlled) is
437 * larger than sent but unacknowledged data in send buffer.
438 *
439 * The remote host receive window scaling factor may limit the
440 * growing of the send buffer before it reaches its allowed
441 * maximum.
442 *
443 * It scales directly with slow start or congestion window
444 * and does at most one step per received ACK. This fast
445 * scaling has the drawback of growing the send buffer beyond
446 * what is strictly necessary to make full use of a given
447 * delay*bandwith product. However testing has shown this not
448 * to be much of an problem. At worst we are trading wasting
449 * of available bandwith (the non-use of it) for wasting some
450 * socket buffer memory.
451 *
452 * TODO: Shrink send buffer during idle periods together
453 * with congestion window. Requires another timer. Has to
454 * wait for upcoming tcp timer rewrite.
455 *
456 * XXXGL: should there be used sbused() or sbavail()?
457 */
458 /*
459 * samkumar: There used to be code here to dynamically size the
460 * send buffer (by calling sbreserve_locked). In TCPlp, we don't support
461 * this, as the send buffer doesn't have a well-defined size (and even if
462 * we were to use a circular buffer, it would be a fixed-size buffer
463 * allocated by the application). Therefore, I removed the code that does
464 * this.
465 */
466
467 /*
468 * samkumar: There used to be code here to handle TCP Segmentation
469 * Offloading (TSO); I removed it becuase we don't support that in TCPlp.
470 */
471
472 if (sack_rxmit) {
473 /*
474 * samkumar: Replaced sbused(&so->so_snd) with this call to
475 * lbuf_used_space.
476 */
477 if (SEQ_LT(p->rxmit + len, tp->snd_una + lbuf_used_space(&tp->sendbuf)))
478 flags &= ~TH_FIN;
479 } else {
480 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
481 /*
482 * samkumar: Replaced sbused(&so->so_snd) with this call to
483 * lbuf_used_space.
484 */
485 lbuf_used_space(&tp->sendbuf)))
486 flags &= ~TH_FIN;
487 }
488
489 /*
490 * samkumar: Replaced sbspace(&so->so_rcv) with this call to
491 * cbuf_free_space.
492 */
493 recwin = cbuf_free_space(&tp->recvbuf);
494
495 /*
496 * Sender silly window avoidance. We transmit under the following
497 * conditions when len is non-zero:
498 *
499 * - We have a full segment (or more with TSO)
500 * - This is the last buffer in a write()/send() and we are
501 * either idle or running NODELAY
502 * - we've timed out (e.g. persist timer)
503 * - we have more then 1/2 the maximum send window's worth of
504 * data (receiver may be limited the window size)
505 * - we need to retransmit
506 */
507 if (len) {
508 if (len >= tp->t_maxseg)
509 goto send;
510 /*
511 * NOTE! on localhost connections an 'ack' from the remote
512 * end may occur synchronously with the output and cause
513 * us to flush a buffer queued with moretocome. XXX
514 *
515 * note: the len + off check is almost certainly unnecessary.
516 */
517 /*
518 * samkumar: Replaced sbavail(&so->so_snd) with this call to
519 * lbuf_used_space.
520 */
521 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
522 (idle || (tp->t_flags & TF_NODELAY)) &&
523 len + off >= lbuf_used_space(&tp->sendbuf) &&
524 (tp->t_flags & TF_NOPUSH) == 0) {
525 goto send;
526 }
527 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
528 goto send;
529 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
530 goto send;
531 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
532 goto send;
533 if (sack_rxmit)
534 goto send;
535 }
536
537 /*
538 * Sending of standalone window updates.
539 *
540 * Window updates are important when we close our window due to a
541 * full socket buffer and are opening it again after the application
542 * reads data from it. Once the window has opened again and the
543 * remote end starts to send again the ACK clock takes over and
544 * provides the most current window information.
545 *
546 * We must avoid the silly window syndrome whereas every read
547 * from the receive buffer, no matter how small, causes a window
548 * update to be sent. We also should avoid sending a flurry of
549 * window updates when the socket buffer had queued a lot of data
550 * and the application is doing small reads.
551 *
552 * Prevent a flurry of pointless window updates by only sending
553 * an update when we can increase the advertized window by more
554 * than 1/4th of the socket buffer capacity. When the buffer is
555 * getting full or is very small be more aggressive and send an
556 * update whenever we can increase by two mss sized segments.
557 * In all other situations the ACK's to new incoming data will
558 * carry further window increases.
559 *
560 * Don't send an independent window update if a delayed
561 * ACK is pending (it will get piggy-backed on it) or the
562 * remote side already has done a half-close and won't send
563 * more data. Skip this if the connection is in T/TCP
564 * half-open state.
565 */
566 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
567 !(tp->t_flags & TF_DELACK) &&
568 !TCPS_HAVERCVDFIN(tp->t_state)) {
569 /*
570 * "adv" is the amount we could increase the window,
571 * taking into account that we are limited by
572 * TCP_MAXWIN << tp->rcv_scale.
573 */
574 long adv;
575 int oldwin;
576
577 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
578 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
579 oldwin = (tp->rcv_adv - tp->rcv_nxt);
580 adv -= oldwin;
581 } else
582 oldwin = 0;
583
584 /*
585 * If the new window size ends up being the same as the old
586 * size when it is scaled, then don't force a window update.
587 */
588 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
589 goto dontupdate;
590
591 /*
592 * samkumar: Here, FreeBSD has some heuristics to decide whether or
593 * not to send a window update. The code for the original heuristics
594 * is commented out, using #if 0. These heuristics compare "adv,"
595 * the size of the window update, with the size of the local receive
596 * buffer. The FreeBSD heuristics aren't applicable because they are
597 * orders of magnitude off from what we see in TCPlp. For example,
598 * FreeBSD only sends a window update if it is at least two segments
599 * big. Note that, in the experiments I did, the second case did not
600 * filter window updates further because, in the experiments, the
601 * receive buffer was smaller than 8 segments.
602 *
603 * I replaced these heuristics with a simpler version, which you can
604 * see below. For the experiments I did, the first condition
605 * (checking if adv >= (long)(2 * tp->t_maxseg)) wasn't included; this
606 * did not matter because the receive buffer was smaller than 8
607 * segments, so any condition that would have triggered the first
608 * condition would have triggered the second one anyway. I've included
609 * the first condition in this version in an effort to be more robust,
610 * in case someone does try to run TCPlp with a large receive buffer.
611 *
612 * It may be worth studying this more and revisiting the heuristic to
613 * use here. In case we try to resurrect the old FreeBSD heuristics,
614 * note that so->so_rcv.sb_hiwat in FreeBSD corresponds roughly to
615 * cbuf_size(&tp->recvbuf) in TCPlp.
616 */
617 #if 0
618 if (adv >= (long)(2 * tp->t_maxseg) &&
619 (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
620 recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
621 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
622 goto send;
623 #endif
624 if (adv >= (long)(2 * tp->t_maxseg) ||
625 adv >= (long)cbuf_size(&tp->recvbuf) / 4)
626 goto send;
627 }
628 dontupdate:
629
630 /*
631 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
632 * is also a catch-all for the retransmit timer timeout case.
633 */
634 if (tp->t_flags & TF_ACKNOW) {
635 goto send;
636 }
637 if ((flags & TH_RST) ||
638 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
639 goto send;
640 if (SEQ_GT(tp->snd_up, tp->snd_una))
641 goto send;
642 /*
643 * If our state indicates that FIN should be sent
644 * and we have not yet done so, then we need to send.
645 */
646 if (flags & TH_FIN &&
647 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
648 goto send;
649 /*
650 * In SACK, it is possible for tcp_output to fail to send a segment
651 * after the retransmission timer has been turned off. Make sure
652 * that the retransmission timer is set.
653 */
654 if ((tp->t_flags & TF_SACK_PERMIT) &&
655 SEQ_GT(tp->snd_max, tp->snd_una) &&
656 !tcp_timer_active(tp, TT_REXMT) &&
657 !tcp_timer_active(tp, TT_PERSIST)) {
658 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
659 goto just_return;
660 }
661
662 /*
663 * TCP window updates are not reliable, rather a polling protocol
664 * using ``persist'' packets is used to insure receipt of window
665 * updates. The three ``states'' for the output side are:
666 * idle not doing retransmits or persists
667 * persisting to move a small or zero window
668 * (re)transmitting and thereby not persisting
669 *
670 * tcp_timer_active(tp, TT_PERSIST)
671 * is true when we are in persist state.
672 * (tp->t_flags & TF_FORCEDATA)
673 * is set when we are called to send a persist packet.
674 * tcp_timer_active(tp, TT_REXMT)
675 * is set when we are retransmitting
676 * The output side is idle when both timers are zero.
677 *
678 * If send window is too small, there is data to transmit, and no
679 * retransmit or persist is pending, then go to persist state.
680 * If nothing happens soon, send when timer expires:
681 * if window is nonzero, transmit what we can,
682 * otherwise force out a byte.
683 */
684 /*
685 * samkumar: Replaced sbavail(&so->so_snd) with this call to
686 * lbuf_used_space.
687 */
688 if (lbuf_used_space(&tp->sendbuf) && !tcp_timer_active(tp, TT_REXMT) &&
689 !tcp_timer_active(tp, TT_PERSIST)) {
690 tp->t_rxtshift = 0;
691 tcp_setpersist(tp);
692 }
693
694 /*
695 * No reason to send a segment, just return.
696 */
697 just_return:
698 return (0);
699
700 send:
701 if (len > 0) {
702 if (len >= tp->t_maxseg)
703 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
704 else
705 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
706 }
707 /*
708 * Before ESTABLISHED, force sending of initial options
709 * unless TCP set not to do any options.
710 * NOTE: we assume that the IP/TCP header plus TCP options
711 * always fit in a single mbuf, leaving room for a maximum
712 * link header, i.e.
713 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
714 */
715 optlen = 0;
716 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
717
718 /*
719 * Compute options for segment.
720 * We only have to care about SYN and established connection
721 * segments. Options for SYN-ACK segments are handled in TCP
722 * syncache.
723 */
724 /*
725 * samkumar: I've done away with the syncache. However, it
726 * seems that the existing logic works fine for SYN-ACK as
727 * well.
728 */
729 to.to_flags = 0;
730 if ((tp->t_flags & TF_NOOPT) == 0) {
731 /* Maximum segment size. */
732 if (flags & TH_SYN) {
733 tp->snd_nxt = tp->iss;
734 to.to_mss = tcp_mssopt(tp);
735 to.to_flags |= TOF_MSS;
736
737 /*
738 * On SYN or SYN|ACK transmits on TFO connections,
739 * only include the TFO option if it is not a
740 * retransmit, as the presence of the TFO option may
741 * have caused the original SYN or SYN|ACK to have
742 * been dropped by a middlebox.
743 */
744 if (IS_FASTOPEN(tp->t_flags) &&
745 (tp->t_rxtshift == 0)) {
746 if (tp->t_state == TCPS_SYN_RECEIVED) {
747 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
748 to.to_tfo_cookie =
749 (u_int8_t *)&tp->t_tfo_cookie.server;
750 to.to_flags |= TOF_FASTOPEN;
751 wanted_cookie = 1;
752 } else if (tp->t_state == TCPS_SYN_SENT) {
753 to.to_tfo_len =
754 tp->t_tfo_client_cookie_len;
755 to.to_tfo_cookie =
756 tp->t_tfo_cookie.client;
757 to.to_flags |= TOF_FASTOPEN;
758 wanted_cookie = 1;
759 /*
760 * If we wind up having more data to
761 * send with the SYN than can fit in
762 * one segment, don't send any more
763 * until the SYN|ACK comes back from
764 * the other end.
765 */
766 dont_sendalot = 1;
767 }
768 }
769 }
770 /* Window scaling. */
771 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
772 to.to_wscale = tp->request_r_scale;
773 to.to_flags |= TOF_SCALE;
774 }
775 /* Timestamps. */
776 if ((tp->t_flags & TF_RCVD_TSTMP) ||
777 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
778 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
779 to.to_tsecr = tp->ts_recent;
780 to.to_flags |= TOF_TS;
781 /*
782 * samkumar: I removed the code to set the timestamp tp->rfbuf_ts
783 * for receive buffer autosizing, since we don't do autosizing on
784 * the receive buffer in TCPlp.
785 */
786 }
787
788 /* Selective ACK's. */
789 if (tp->t_flags & TF_SACK_PERMIT) {
790 if (flags & TH_SYN)
791 to.to_flags |= TOF_SACKPERM;
792 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
793 (tp->t_flags & TF_SACK_PERMIT) &&
794 tp->rcv_numsacks > 0) {
795 to.to_flags |= TOF_SACK;
796 to.to_nsacks = tp->rcv_numsacks;
797 to.to_sacks = (uint8_t *)tp->sackblks;
798 }
799 }
800
801 /*
802 * samkumar: Remove logic to set TOF_SIGNATURE flag in to.to_flags,
803 * since TCPlp does not support TCP signatures.
804 */
805
806 /* Processing the options. */
807 hdrlen += optlen = tcp_addoptions(&to, opt);
808 /*
809 * If we wanted a TFO option to be added, but it was unable
810 * to fit, ensure no data is sent.
811 */
812 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
813 !(to.to_flags & TOF_FASTOPEN))
814 len = 0;
815 }
816 /*
817 * samkumar: This used to be set to ip6_optlen(tp->t_inpcb), instead of 0,
818 * along with some additional code to handle IPSEC. In TCPlp we don't set
819 * IPv6 options here; we expect those to be set by the host network stack.
820 * Of course, code that supports IPv4 has been removed as well.
821 */
822 ipoptlen = 0;
823
824 /*
825 * Adjust data length if insertion of options will
826 * bump the packet length beyond the t_maxopd length.
827 * Clear the FIN bit because we cut off the tail of
828 * the segment.
829 */
830 if (len + optlen + ipoptlen > tp->t_maxopd) {
831 flags &= ~TH_FIN;
832 /*
833 * samkumar: Remove code for TCP segmentation offloading.
834 */
835 len = tp->t_maxopd - optlen - ipoptlen;
836 sendalot = 1;
837 if (dont_sendalot)
838 sendalot = 0;
839 }
840 /*
841 * samkumar: The else case of the above "if" statement would set tso to 0.
842 * Removing this since we no longer need a tso variable.
843 */
844 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
845 ("%s: len > IP_MAXPACKET", __func__));
846
847 /*
848 * This KASSERT is here to catch edge cases at a well defined place.
849 * Before, those had triggered (random) panic conditions further down.
850 */
851 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
852
853 /*
854 * Grab a header mbuf, attaching a copy of data to
855 * be transmitted, and initialize the header from
856 * the template for sends on this connection.
857 */
858
859 /*
860 * samkumar: The code to allocate, build, and send outgoing segments has
861 * been rewritten. I've left the original code to build the output mbuf
862 * here in a comment, for reference. The new code is below.
863 */
864 #if 0
865 if (len) {
866 struct mbuf *mb;
867 uint32_t moff;
868
869 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
870 TCPSTAT_INC(tcps_sndprobe);
871 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
872 tp->t_sndrexmitpack++;
873 TCPSTAT_INC(tcps_sndrexmitpack);
874 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
875 } else {
876 TCPSTAT_INC(tcps_sndpack);
877 TCPSTAT_ADD(tcps_sndbyte, len);
878 }
879 #ifdef INET6
880 if (MHLEN < hdrlen + max_linkhdr)
881 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
882 else
883 #endif
884 m = m_gethdr(M_NOWAIT, MT_DATA);
885
886 if (m == NULL) {
887 SOCKBUF_UNLOCK(&so->so_snd);
888 error = ENOBUFS;
889 sack_rxmit = 0;
890 goto out;
891 }
892
893 m->m_data += max_linkhdr;
894 m->m_len = hdrlen;
895
896 /*
897 * Start the m_copy functions from the closest mbuf
898 * to the offset in the socket buffer chain.
899 */
900 mb = sbsndptr(&so->so_snd, off, len, &moff);
901
902 if (len <= MHLEN - hdrlen - max_linkhdr) {
903 m_copydata(mb, moff, (int)len,
904 mtod(m, caddr_t) + hdrlen);
905 m->m_len += len;
906 } else {
907 m->m_next = m_copy(mb, moff, (int)len);
908 if (m->m_next == NULL) {
909 SOCKBUF_UNLOCK(&so->so_snd);
910 (void) m_free(m);
911 error = ENOBUFS;
912 sack_rxmit = 0;
913 goto out;
914 }
915 }
916
917 /*
918 * If we're sending everything we've got, set PUSH.
919 * (This will keep happy those implementations which only
920 * give data to the user when a buffer fills or
921 * a PUSH comes in.)
922 */
923 if (off + len == sbused(&so->so_snd))
924 flags |= TH_PUSH;
925 SOCKBUF_UNLOCK(&so->so_snd);
926 } else {
927 SOCKBUF_UNLOCK(&so->so_snd);
928 if (tp->t_flags & TF_ACKNOW)
929 TCPSTAT_INC(tcps_sndacks);
930 else if (flags & (TH_SYN|TH_FIN|TH_RST))
931 TCPSTAT_INC(tcps_sndctrl);
932 else if (SEQ_GT(tp->snd_up, tp->snd_una))
933 TCPSTAT_INC(tcps_sndurg);
934 else
935 TCPSTAT_INC(tcps_sndwinup);
936
937 m = m_gethdr(M_NOWAIT, MT_DATA);
938 if (m == NULL) {
939 error = ENOBUFS;
940 sack_rxmit = 0;
941 goto out;
942 }
943 #ifdef INET6
944 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
945 MHLEN >= hdrlen) {
946 M_ALIGN(m, hdrlen);
947 } else
948 #endif
949 m->m_data += max_linkhdr;
950 m->m_len = hdrlen;
951 }
952 #endif
953
954 KASSERT(ipoptlen == 0, ("No IP options supported")); // samkumar
955
956 otMessage* message = tcplp_sys_new_message(tp->instance);
957 if (message == NULL) {
958 error = ENOBUFS;
959 sack_rxmit = 0;
960 goto out;
961 }
962 if (otMessageSetLength(message, sizeof(struct tcphdr) + optlen + len) != OT_ERROR_NONE) {
963 tcplp_sys_free_message(tp->instance, message);
964 error = ENOBUFS;
965 sack_rxmit = 0;
966 goto out;
967 }
968 if (len) {
969 uint32_t used_space = lbuf_used_space(&tp->sendbuf);
970
971 /*
972 * The TinyOS version has a way to avoid the copying we have to do here.
973 * Because it is possible to send iovecs directly in the BLIP stack, and
974 * an lbuf is made of iovecs, we could just "save" the starting and ending
975 * iovecs, modify them to get exactly the slice we want, call "send" on
976 * the resulting chain, and then restore the starting and ending iovecs
977 * once "send" returns.
978 *
979 * In RIOT, pktsnips have additional behavior regarding memory management
980 * that precludes this optimization. But, now that we have moved to
981 * cbufs, this is not relevant anymore.
982 */
983 {
984 otLinkedBuffer* start;
985 size_t start_offset;
986 otLinkedBuffer* end;
987 size_t end_offset;
988 otLinkedBuffer* curr;
989 int rv = lbuf_getrange(&tp->sendbuf, off, len, &start, &start_offset, &end, &end_offset);
990 size_t message_offset = otMessageGetOffset(message) + sizeof(struct tcphdr) + optlen;
991 KASSERT(rv == 0, ("Reading send buffer out of range!"));
992 for (curr = start; curr != end->mNext; curr = curr->mNext) {
993 const uint8_t* data_to_copy = curr->mData;
994 size_t length_to_copy = curr->mLength;
995 if (curr == start) {
996 data_to_copy += start_offset;
997 length_to_copy -= start_offset;
998 }
999 if (curr == end) {
1000 length_to_copy -= end_offset;
1001 }
1002 otMessageWrite(message, message_offset, data_to_copy, length_to_copy);
1003 message_offset += length_to_copy;
1004 }
1005 }
1006
1007 /*
1008 * If we're sending everything we've got, set PUSH.
1009 * (This will keep happy those implementations which only
1010 * give data to the user when a buffer fills or
1011 * a PUSH comes in.)
1012 */
1013 /* samkumar: Replaced call to sbused(&so->so_snd) with used_space. */
1014 if (off + len == used_space)
1015 flags |= TH_PUSH;
1016 }
1017
1018 char outbuf[sizeof(struct tcphdr) + TCP_MAXOLEN];
1019 th = (struct tcphdr*) (&outbuf[0]);
1020
1021 /*
1022 * samkumar: I replaced the original call to tcpip_fillheaders with the
1023 * one below.
1024 */
1025 otMessageInfo ip6info;
1026 tcpip_fillheaders(tp, &ip6info, th);
1027
1028 /*
1029 * Fill in fields, remembering maximum advertised
1030 * window for use in delaying messages about window sizes.
1031 * If resending a FIN, be sure not to use a new sequence number.
1032 */
1033 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1034 tp->snd_nxt == tp->snd_max)
1035 tp->snd_nxt--;
1036 /*
1037 * If we are starting a connection, send ECN setup
1038 * SYN packet. If we are on a retransmit, we may
1039 * resend those bits a number of times as per
1040 * RFC 3168.
1041 */
1042 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1043 if (tp->t_rxtshift >= 1) {
1044 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1045 flags |= TH_ECE|TH_CWR;
1046 } else
1047 flags |= TH_ECE|TH_CWR;
1048 }
1049
1050 /*
1051 * samkumar: Make tcp_output reply with ECE flag in the SYN-ACK for
1052 * ECN-enabled connections. The existing code in FreeBSD didn't have to do
1053 * this, because it didn't use tcp_output to send the SYN-ACK; it
1054 * constructed the SYN-ACK segment manually. Yet another consequnce of
1055 * removing the SYN cache...
1056 */
1057 if (tp->t_state == TCPS_SYN_RECEIVED && tp->t_flags & TF_ECN_PERMIT &&
1058 V_tcp_do_ecn) {
1059 flags |= TH_ECE;
1060 }
1061
1062 if (tp->t_state == TCPS_ESTABLISHED &&
1063 (tp->t_flags & TF_ECN_PERMIT)) {
1064 /*
1065 * If the peer has ECN, mark data packets with
1066 * ECN capable transmission (ECT).
1067 * Ignore pure ack packets, retransmissions and window probes.
1068 */
1069 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1070 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1071 /*
1072 * samkumar: Replaced ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1073 * with the following code, which will cause OpenThread to set the
1074 * ECT0 bit in the header.
1075 */
1076 ip6info.mEcn = OT_ECN_CAPABLE_0;
1077 }
1078
1079 /*
1080 * Reply with proper ECN notifications.
1081 */
1082 if (tp->t_flags & TF_ECN_SND_CWR) {
1083 flags |= TH_CWR;
1084 tp->t_flags &= ~TF_ECN_SND_CWR;
1085 }
1086 if (tp->t_flags & TF_ECN_SND_ECE)
1087 flags |= TH_ECE;
1088 }
1089
1090 /*
1091 * If we are doing retransmissions, then snd_nxt will
1092 * not reflect the first unsent octet. For ACK only
1093 * packets, we do not want the sequence number of the
1094 * retransmitted packet, we want the sequence number
1095 * of the next unsent octet. So, if there is no data
1096 * (and no SYN or FIN), use snd_max instead of snd_nxt
1097 * when filling in ti_seq. But if we are in persist
1098 * state, snd_max might reflect one byte beyond the
1099 * right edge of the window, so use snd_nxt in that
1100 * case, since we know we aren't doing a retransmission.
1101 * (retransmit and persist are mutually exclusive...)
1102 */
1103 if (sack_rxmit == 0) {
1104 if (len || (flags & (TH_SYN|TH_FIN)) ||
1105 tcp_timer_active(tp, TT_PERSIST))
1106 th->th_seq = htonl(tp->snd_nxt);
1107 else
1108 th->th_seq = htonl(tp->snd_max);
1109 } else {
1110 th->th_seq = htonl(p->rxmit);
1111 p->rxmit += len;
1112 tp->sackhint.sack_bytes_rexmit += len;
1113 }
1114
1115 /*
1116 * samkumar: Check if this is a retransmission (added as part of TCPlp).
1117 * This kind of stats collection is useful but not necessary for TCP, so
1118 * I've left it as a comment in case we want to bring this back to measure
1119 * performance.
1120 */
1121 #if 0
1122 if (len > 0 && !tcp_timer_active(tp, TT_PERSIST) && SEQ_LT(ntohl(th->th_seq), tp->snd_max)) {
1123 tcplp_totalRexmitCnt++;
1124 }
1125 #endif
1126
1127 th->th_ack = htonl(tp->rcv_nxt);
1128 if (optlen) {
1129 bcopy(opt, th + 1, optlen);
1130 th->th_off_x2 = ((sizeof (struct tcphdr) + optlen) >> 2) << TH_OFF_SHIFT;
1131 }
1132 th->th_flags = flags;
1133 /*
1134 * Calculate receive window. Don't shrink window,
1135 * but avoid silly window syndrome.
1136 */
1137 /* samkumar: Replaced so->so_rcv.sb_hiwat with this call to cbuf_size. */
1138 if (recwin < (long)(cbuf_size(&tp->recvbuf) / 4) &&
1139 recwin < (long)tp->t_maxseg)
1140 recwin = 0;
1141 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1142 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1143 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1144 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1145 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1146
1147 /*
1148 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1149 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1150 * case is handled in syncache.
1151 */
1152 if (flags & TH_SYN)
1153 th->th_win = htons((uint16_t)
1154 (min(cbuf_size(&tp->recvbuf), TCP_MAXWIN)));
1155 else
1156 th->th_win = htons((uint16_t)(recwin >> tp->rcv_scale));
1157
1158 /*
1159 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1160 * a 0 window. This may cause the remote transmitter to stall. This
1161 * flag tells soreceive() to disable delayed acknowledgements when
1162 * draining the buffer. This can occur if the receiver is attempting
1163 * to read more data than can be buffered prior to transmitting on
1164 * the connection.
1165 */
1166 if (th->th_win == 0) {
1167 tp->t_flags |= TF_RXWIN0SENT;
1168 } else
1169 tp->t_flags &= ~TF_RXWIN0SENT;
1170 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1171 th->th_urp = htons((uint16_t)(tp->snd_up - tp->snd_nxt));
1172 th->th_flags |= TH_URG;
1173 } else
1174 /*
1175 * If no urgent pointer to send, then we pull
1176 * the urgent pointer to the left edge of the send window
1177 * so that it doesn't drift into the send window on sequence
1178 * number wraparound.
1179 */
1180 tp->snd_up = tp->snd_una; /* drag it along */
1181
1182 /*
1183 * samkumar: Removed code for TCP signatures.
1184 */
1185 /*
1186 * Put TCP length in extended header, and then
1187 * checksum extended header and data.
1188 */
1189 /*
1190 * samkumar: The code to implement the above comment isn't relevant to us.
1191 * Checksum computation is not handled using FreeBSD code, so we don't need
1192 * to build an extended header.
1193 */
1194 /*
1195 * samkumar: Removed code for TCP Segmentation Offloading.
1196 */
1197 /* samkumar: Removed mbuf-specific assertions an debug code. */
1198 /*
1199 * Fill in IP length and desired time to live and
1200 * send to IP level. There should be a better way
1201 * to handle ttl and tos; we could keep them in
1202 * the template, but need a way to checksum without them.
1203 */
1204 /*
1205 * m->m_pkthdr.len should have been set before checksum calculation,
1206 * because in6_cksum() need it.
1207 */
1208 /*
1209 * samkumar: The IPv6 packet length and hop limit are handled by the host
1210 * network stack, not by TCPlp. I've also removed code for Path MTU
1211 * discovery. And of course, I've removed debug code as well.
1212 */
1213 /* samkumar: I've replaced the call to ip6_output with the following. */
1214 otMessageWrite(message, 0, outbuf, sizeof(struct tcphdr) + optlen);
1215 tcplp_sys_send_message(tp->instance, message, &ip6info);
1216
1217 out:
1218 /*
1219 * In transmit state, time the transmission and arrange for
1220 * the retransmit. In persist state, just set snd_max.
1221 */
1222 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1223 !tcp_timer_active(tp, TT_PERSIST)) {
1224 tcp_seq startseq = tp->snd_nxt;
1225
1226 /*
1227 * Advance snd_nxt over sequence space of this segment.
1228 */
1229 if (flags & (TH_SYN|TH_FIN)) {
1230 if (flags & TH_SYN)
1231 tp->snd_nxt++;
1232 if (flags & TH_FIN) {
1233 tp->snd_nxt++;
1234 tp->t_flags |= TF_SENTFIN;
1235 }
1236 }
1237 if (sack_rxmit)
1238 goto timer;
1239 tp->snd_nxt += len;
1240 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1241 tp->snd_max = tp->snd_nxt;
1242 /*
1243 * Time this transmission if not a retransmission and
1244 * not currently timing anything.
1245 */
1246 if (tp->t_rtttime == 0) {
1247 tp->t_rtttime = ticks;
1248 tp->t_rtseq = startseq;
1249 }
1250 }
1251
1252 /*
1253 * Set retransmit timer if not currently set,
1254 * and not doing a pure ack or a keep-alive probe.
1255 * Initial value for retransmit timer is smoothed
1256 * round-trip time + 2 * round-trip time variance.
1257 * Initialize shift counter which is used for backoff
1258 * of retransmit time.
1259 */
1260 timer:
1261 if (!tcp_timer_active(tp, TT_REXMT) &&
1262 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1263 (tp->snd_nxt != tp->snd_una))) {
1264 if (tcp_timer_active(tp, TT_PERSIST)) {
1265 tcp_timer_activate(tp, TT_PERSIST, 0);
1266 tp->t_rxtshift = 0;
1267 }
1268 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1269 /*
1270 * samkumar: Replaced sbavail(&so->so_snd) with this call to
1271 * lbuf_used_space.
1272 */
1273 } else if (len == 0 && lbuf_used_space(&tp->sendbuf) &&
1274 !tcp_timer_active(tp, TT_REXMT) &&
1275 !tcp_timer_active(tp, TT_PERSIST)) {
1276 /*
1277 * Avoid a situation where we do not set persist timer
1278 * after a zero window condition. For example:
1279 * 1) A -> B: packet with enough data to fill the window
1280 * 2) B -> A: ACK for #1 + new data (0 window
1281 * advertisement)
1282 * 3) A -> B: ACK for #2, 0 len packet
1283 *
1284 * In this case, A will not activate the persist timer,
1285 * because it chose to send a packet. Unless tcp_output
1286 * is called for some other reason (delayed ack timer,
1287 * another input packet from B, socket syscall), A will
1288 * not send zero window probes.
1289 *
1290 * So, if you send a 0-length packet, but there is data
1291 * in the socket buffer, and neither the rexmt or
1292 * persist timer is already set, then activate the
1293 * persist timer.
1294 */
1295 tp->t_rxtshift = 0;
1296 tcp_setpersist(tp);
1297 }
1298 } else {
1299 /*
1300 * Persist case, update snd_max but since we are in
1301 * persist mode (no window) we do not update snd_nxt.
1302 */
1303 int xlen = len;
1304 if (flags & TH_SYN)
1305 ++xlen;
1306 if (flags & TH_FIN) {
1307 ++xlen;
1308 tp->t_flags |= TF_SENTFIN;
1309 }
1310 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1311 tp->snd_max = tp->snd_nxt + len;
1312 }
1313
1314 if (error) {
1315
1316 /*
1317 * We know that the packet was lost, so back out the
1318 * sequence number advance, if any.
1319 *
1320 * If the error is EPERM the packet got blocked by the
1321 * local firewall. Normally we should terminate the
1322 * connection but the blocking may have been spurious
1323 * due to a firewall reconfiguration cycle. So we treat
1324 * it like a packet loss and let the retransmit timer and
1325 * timeouts do their work over time.
1326 * XXX: It is a POLA question whether calling tcp_drop right
1327 * away would be the really correct behavior instead.
1328 */
1329 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1330 !tcp_timer_active(tp, TT_PERSIST)) &&
1331 ((flags & TH_SYN) == 0) &&
1332 (error != EPERM)) {
1333 if (sack_rxmit) {
1334 p->rxmit -= len;
1335 tp->sackhint.sack_bytes_rexmit -= len;
1336 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1337 ("sackhint bytes rtx >= 0"));
1338 } else
1339 tp->snd_nxt -= len;
1340 }
1341 switch (error) {
1342 case EPERM:
1343 tp->t_softerror = error;
1344 return (error);
1345 case ENOBUFS:
1346 if (!tcp_timer_active(tp, TT_REXMT) &&
1347 !tcp_timer_active(tp, TT_PERSIST))
1348 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1349 tp->snd_cwnd = tp->t_maxseg;
1350 #ifdef INSTRUMENT_TCP
1351 tcplp_sys_log("TCP ALLOCFAIL %u %d", (unsigned int) tcplp_sys_get_millis(), (int) tp->snd_cwnd);
1352 #endif
1353 return (0);
1354 case EMSGSIZE:
1355 /*
1356 * For some reason the interface we used initially
1357 * to send segments changed to another or lowered
1358 * its MTU.
1359 * If TSO was active we either got an interface
1360 * without TSO capabilits or TSO was turned off.
1361 * If we obtained mtu from ip_output() then update
1362 * it and try again.
1363 */
1364 /* samkumar: Removed code for TCP Segmentation Offloading. */
1365 if (mtu != 0) {
1366 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1367 goto again;
1368 }
1369 return (error);
1370 case EHOSTDOWN:
1371 case EHOSTUNREACH:
1372 case ENETDOWN:
1373 case ENETUNREACH:
1374 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1375 tp->t_softerror = error;
1376 return (0);
1377 }
1378 /* FALLTHROUGH */
1379 default:
1380 return (error);
1381 }
1382 }
1383
1384 /*
1385 * Data sent (as far as we can tell).
1386 * If this advertises a larger window than any other segment,
1387 * then remember the size of the advertised window.
1388 * Any pending ACK has now been sent.
1389 */
1390 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1391 tp->rcv_adv = tp->rcv_nxt + recwin;
1392 tp->last_ack_sent = tp->rcv_nxt;
1393 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1394 if (tcp_timer_active(tp, TT_DELACK))
1395 tcp_timer_activate(tp, TT_DELACK, 0);
1396
1397 /*
1398 * samkumar: This was already commented out (using #if 0) in the original
1399 * FreeBSD code.
1400 */
1401 #if 0
1402 /*
1403 * This completely breaks TCP if newreno is turned on. What happens
1404 * is that if delayed-acks are turned on on the receiver, this code
1405 * on the transmitter effectively destroys the TCP window, forcing
1406 * it to four packets (1.5Kx4 = 6K window).
1407 */
1408 if (sendalot && --maxburst)
1409 goto again;
1410 #endif
1411 if (sendalot)
1412 goto again;
1413 return (0);
1414 }
1415
1416 /*
1417 * Insert TCP options according to the supplied parameters to the place
1418 * optp in a consistent way. Can handle unaligned destinations.
1419 *
1420 * The order of the option processing is crucial for optimal packing and
1421 * alignment for the scarce option space.
1422 *
1423 * The optimal order for a SYN/SYN-ACK segment is:
1424 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1425 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1426 *
1427 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1428 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1429 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1430 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1431 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1432 */
1433 int
tcp_addoptions(struct tcpopt * to,uint8_t * optp)1434 tcp_addoptions(struct tcpopt *to, uint8_t *optp)
1435 {
1436 uint32_t mask, optlen = 0;
1437
1438 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1439 if ((to->to_flags & mask) != mask)
1440 continue;
1441 if (optlen == TCP_MAXOLEN)
1442 break;
1443 switch (to->to_flags & mask) {
1444 case TOF_MSS:
1445 while (optlen % 4) {
1446 optlen += TCPOLEN_NOP;
1447 *optp++ = TCPOPT_NOP;
1448 }
1449 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1450 continue;
1451 optlen += TCPOLEN_MAXSEG;
1452 *optp++ = TCPOPT_MAXSEG;
1453 *optp++ = TCPOLEN_MAXSEG;
1454 to->to_mss = htons(to->to_mss);
1455 bcopy((uint8_t *)&to->to_mss, optp, sizeof(to->to_mss));
1456 optp += sizeof(to->to_mss);
1457 break;
1458 case TOF_SCALE:
1459 while (!optlen || optlen % 2 != 1) {
1460 optlen += TCPOLEN_NOP;
1461 *optp++ = TCPOPT_NOP;
1462 }
1463 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1464 continue;
1465 optlen += TCPOLEN_WINDOW;
1466 *optp++ = TCPOPT_WINDOW;
1467 *optp++ = TCPOLEN_WINDOW;
1468 *optp++ = to->to_wscale;
1469 break;
1470 case TOF_SACKPERM:
1471 while (optlen % 2) {
1472 optlen += TCPOLEN_NOP;
1473 *optp++ = TCPOPT_NOP;
1474 }
1475 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1476 continue;
1477 optlen += TCPOLEN_SACK_PERMITTED;
1478 *optp++ = TCPOPT_SACK_PERMITTED;
1479 *optp++ = TCPOLEN_SACK_PERMITTED;
1480 break;
1481 case TOF_TS:
1482 while (!optlen || optlen % 4 != 2) {
1483 optlen += TCPOLEN_NOP;
1484 *optp++ = TCPOPT_NOP;
1485 }
1486 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1487 continue;
1488 optlen += TCPOLEN_TIMESTAMP;
1489 *optp++ = TCPOPT_TIMESTAMP;
1490 *optp++ = TCPOLEN_TIMESTAMP;
1491 to->to_tsval = htonl(to->to_tsval);
1492 to->to_tsecr = htonl(to->to_tsecr);
1493 bcopy((uint8_t *)&to->to_tsval, optp, sizeof(to->to_tsval));
1494 optp += sizeof(to->to_tsval);
1495 bcopy((uint8_t *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1496 optp += sizeof(to->to_tsecr);
1497 break;
1498 case TOF_SIGNATURE:
1499 {
1500 int siglen = TCPOLEN_SIGNATURE - 2;
1501
1502 while (!optlen || optlen % 4 != 2) {
1503 optlen += TCPOLEN_NOP;
1504 *optp++ = TCPOPT_NOP;
1505 }
1506 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1507 continue;
1508 optlen += TCPOLEN_SIGNATURE;
1509 *optp++ = TCPOPT_SIGNATURE;
1510 *optp++ = TCPOLEN_SIGNATURE;
1511 to->to_signature = optp;
1512 while (siglen--)
1513 *optp++ = 0;
1514 break;
1515 }
1516 case TOF_SACK:
1517 {
1518 int sackblks = 0;
1519 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1520 tcp_seq sack_seq;
1521
1522 while (!optlen || optlen % 4 != 2) {
1523 optlen += TCPOLEN_NOP;
1524 *optp++ = TCPOPT_NOP;
1525 }
1526 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1527 continue;
1528 optlen += TCPOLEN_SACKHDR;
1529 *optp++ = TCPOPT_SACK;
1530 sackblks = min(to->to_nsacks,
1531 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1532 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1533 while (sackblks--) {
1534 sack_seq = htonl(sack->start);
1535 bcopy((uint8_t *)&sack_seq, optp, sizeof(sack_seq));
1536 optp += sizeof(sack_seq);
1537 sack_seq = htonl(sack->end);
1538 bcopy((uint8_t *)&sack_seq, optp, sizeof(sack_seq));
1539 optp += sizeof(sack_seq);
1540 optlen += TCPOLEN_SACK;
1541 sack++;
1542 }
1543 /* samkumar: Removed TCPSTAT_INC(tcps_sack_send_blocks); */
1544 break;
1545 }
1546 case TOF_FASTOPEN:
1547 {
1548 int total_len;
1549
1550 /* XXX is there any point to aligning this option? */
1551 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1552 if (TCP_MAXOLEN - optlen < total_len) {
1553 to->to_flags &= ~TOF_FASTOPEN;
1554 continue;
1555 }
1556 *optp++ = TCPOPT_FAST_OPEN;
1557 *optp++ = total_len;
1558 if (to->to_tfo_len > 0) {
1559 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1560 optp += to->to_tfo_len;
1561 }
1562 optlen += total_len;
1563 break;
1564 }
1565 default:
1566 tcplp_sys_panic("PANIC: %s: unknown TCP option type", __func__);
1567 break;
1568 }
1569 }
1570
1571 /* Terminate and pad TCP options to a 4 byte boundary. */
1572 if (optlen % 4) {
1573 optlen += TCPOLEN_EOL;
1574 *optp++ = TCPOPT_EOL;
1575 }
1576 /*
1577 * According to RFC 793 (STD0007):
1578 * "The content of the header beyond the End-of-Option option
1579 * must be header padding (i.e., zero)."
1580 * and later: "The padding is composed of zeros."
1581 */
1582 while (optlen % 4) {
1583 optlen += TCPOLEN_PAD;
1584 *optp++ = TCPOPT_PAD;
1585 }
1586
1587 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1588 return (optlen);
1589 }
1590
