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