1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2014, 2017 by Delphix. All rights reserved.
25 * Copyright 2019 Joyent, Inc.
26 */
27
28 /* This file contains all TCP output processing functions. */
29
30 #include <sys/types.h>
31 #include <sys/stream.h>
32 #include <sys/strsun.h>
33 #include <sys/strsubr.h>
34 #include <sys/stropts.h>
35 #include <sys/strlog.h>
36 #define _SUN_TPI_VERSION 2
37 #include <sys/tihdr.h>
38 #include <sys/suntpi.h>
39 #include <sys/xti_inet.h>
40 #include <sys/timod.h>
41 #include <sys/pattr.h>
42 #include <sys/squeue_impl.h>
43 #include <sys/squeue.h>
44 #include <sys/sockio.h>
45 #include <sys/tsol/tnet.h>
46
47 #include <inet/common.h>
48 #include <inet/ip.h>
49 #include <inet/tcp.h>
50 #include <inet/tcp_impl.h>
51 #include <inet/snmpcom.h>
52 #include <inet/proto_set.h>
53 #include <inet/ipsec_impl.h>
54 #include <inet/ip_ndp.h>
55
56 static mblk_t *tcp_get_seg_mp(tcp_t *, uint32_t, int32_t *);
57 static void tcp_wput_cmdblk(queue_t *, mblk_t *);
58 static void tcp_wput_flush(tcp_t *, mblk_t *);
59 static void tcp_wput_iocdata(tcp_t *tcp, mblk_t *mp);
60 static int tcp_xmit_end(tcp_t *);
61 static int tcp_send(tcp_t *, const int, const int, const int,
62 const int, int *, uint32_t *, int *, mblk_t **, mblk_t *);
63 static void tcp_xmit_early_reset(char *, mblk_t *, uint32_t, uint32_t,
64 int, ip_recv_attr_t *, ip_stack_t *, conn_t *);
65 static boolean_t tcp_send_rst_chk(tcp_stack_t *);
66 static void tcp_process_shrunk_swnd(tcp_t *, uint32_t);
67 static void tcp_fill_header(tcp_t *, uchar_t *, int);
68
69 /*
70 * Functions called directly via squeue having a prototype of edesc_t.
71 */
72 static void tcp_wput_nondata(void *, mblk_t *, void *, ip_recv_attr_t *);
73 static void tcp_wput_ioctl(void *, mblk_t *, void *, ip_recv_attr_t *);
74 static void tcp_wput_proto(void *, mblk_t *, void *, ip_recv_attr_t *);
75
76 /*
77 * This controls how tiny a write must be before we try to copy it
78 * into the mblk on the tail of the transmit queue. Not much
79 * speedup is observed for values larger than sixteen. Zero will
80 * disable the optimisation.
81 */
82 static int tcp_tx_pull_len = 16;
83
84 static void
85 cc_after_idle(tcp_t *tcp)
86 {
87 uint32_t old_cwnd = tcp->tcp_cwnd;
88
89 if (CC_ALGO(tcp)->after_idle != NULL)
90 CC_ALGO(tcp)->after_idle(&tcp->tcp_ccv);
91
92 DTRACE_PROBE3(cwnd__cc__after__idle, tcp_t *, tcp, uint32_t, old_cwnd,
93 uint32_t, tcp->tcp_cwnd);
94 }
95
96 int
97 tcp_wput(queue_t *q, mblk_t *mp)
98 {
99 conn_t *connp = Q_TO_CONN(q);
100 tcp_t *tcp;
101 void (*output_proc)();
102 t_scalar_t type;
103 uchar_t *rptr;
104 struct iocblk *iocp;
105 size_t size;
106
107 ASSERT(connp->conn_ref >= 2);
108
109 switch (DB_TYPE(mp)) {
110 case M_DATA:
111 tcp = connp->conn_tcp;
112 ASSERT(tcp != NULL);
113
114 size = msgdsize(mp);
115
116 mutex_enter(&tcp->tcp_non_sq_lock);
117 tcp->tcp_squeue_bytes += size;
118 if (TCP_UNSENT_BYTES(tcp) > connp->conn_sndbuf) {
119 tcp_setqfull(tcp);
120 }
121 mutex_exit(&tcp->tcp_non_sq_lock);
122
123 CONN_INC_REF(connp);
124 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_output, connp,
125 NULL, tcp_squeue_flag, SQTAG_TCP_OUTPUT);
126 return (0);
127
128 case M_CMD:
129 tcp_wput_cmdblk(q, mp);
130 return (0);
131
132 case M_PROTO:
133 case M_PCPROTO:
134 /*
135 * if it is a snmp message, don't get behind the squeue
136 */
137 tcp = connp->conn_tcp;
138 rptr = mp->b_rptr;
139 if ((mp->b_wptr - rptr) >= sizeof (t_scalar_t)) {
140 type = ((union T_primitives *)rptr)->type;
141 } else {
142 if (connp->conn_debug) {
143 (void) strlog(TCP_MOD_ID, 0, 1,
144 SL_ERROR|SL_TRACE,
145 "tcp_wput_proto, dropping one...");
146 }
147 freemsg(mp);
148 return (0);
149 }
150 if (type == T_SVR4_OPTMGMT_REQ) {
151 /*
152 * All Solaris components should pass a db_credp
153 * for this TPI message, hence we ASSERT.
154 * But in case there is some other M_PROTO that looks
155 * like a TPI message sent by some other kernel
156 * component, we check and return an error.
157 */
158 cred_t *cr = msg_getcred(mp, NULL);
159
160 ASSERT(cr != NULL);
161 if (cr == NULL) {
162 tcp_err_ack(tcp, mp, TSYSERR, EINVAL);
163 return (0);
164 }
165 if (snmpcom_req(q, mp, tcp_snmp_set, ip_snmp_get,
166 cr)) {
167 /*
168 * This was a SNMP request
169 */
170 return (0);
171 } else {
172 output_proc = tcp_wput_proto;
173 }
174 } else {
175 output_proc = tcp_wput_proto;
176 }
177 break;
178 case M_IOCTL:
179 /*
180 * Most ioctls can be processed right away without going via
181 * squeues - process them right here. Those that do require
182 * squeue (currently _SIOCSOCKFALLBACK)
183 * are processed by tcp_wput_ioctl().
184 */
185 iocp = (struct iocblk *)mp->b_rptr;
186 tcp = connp->conn_tcp;
187
188 switch (iocp->ioc_cmd) {
189 case TCP_IOC_ABORT_CONN:
190 tcp_ioctl_abort_conn(q, mp);
191 return (0);
192 case TI_GETPEERNAME:
193 case TI_GETMYNAME:
194 mi_copyin(q, mp, NULL,
195 SIZEOF_STRUCT(strbuf, iocp->ioc_flag));
196 return (0);
197
198 default:
199 output_proc = tcp_wput_ioctl;
200 break;
201 }
202 break;
203 default:
204 output_proc = tcp_wput_nondata;
205 break;
206 }
207
208 CONN_INC_REF(connp);
209 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, output_proc, connp,
210 NULL, tcp_squeue_flag, SQTAG_TCP_WPUT_OTHER);
211 return (0);
212 }
213
214 /*
215 * The TCP normal data output path.
216 * NOTE: the logic of the fast path is duplicated from this function.
217 */
218 void
219 tcp_wput_data(tcp_t *tcp, mblk_t *mp, boolean_t urgent)
220 {
221 int len;
222 mblk_t *local_time;
223 mblk_t *mp1;
224 uint32_t snxt;
225 int tail_unsent;
226 int tcpstate;
227 int usable = 0;
228 mblk_t *xmit_tail;
229 int32_t mss;
230 int32_t num_sack_blk = 0;
231 int32_t total_hdr_len;
232 int32_t tcp_hdr_len;
233 int rc;
234 conn_t *connp = tcp->tcp_connp;
235 clock_t now = LBOLT_FASTPATH;
236
237 tcpstate = tcp->tcp_state;
238 if (mp == NULL) {
239 /*
240 * tcp_wput_data() with NULL mp should only be called when
241 * there is unsent data.
242 */
243 ASSERT(tcp->tcp_unsent > 0);
244 /* Really tacky... but we need this for detached closes. */
245 len = tcp->tcp_unsent;
246 goto data_null;
247 }
248
249 ASSERT(mp->b_datap->db_type == M_DATA);
250 /*
251 * Don't allow data after T_ORDREL_REQ or T_DISCON_REQ,
252 * or before a connection attempt has begun.
253 */
254 if (tcpstate < TCPS_SYN_SENT || tcpstate > TCPS_CLOSE_WAIT ||
255 (tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
256 if ((tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
257 #ifdef DEBUG
258 cmn_err(CE_WARN,
259 "tcp_wput_data: data after ordrel, %s",
260 tcp_display(tcp, NULL,
261 DISP_ADDR_AND_PORT));
262 #else
263 if (connp->conn_debug) {
264 (void) strlog(TCP_MOD_ID, 0, 1,
265 SL_TRACE|SL_ERROR,
266 "tcp_wput_data: data after ordrel, %s\n",
267 tcp_display(tcp, NULL,
268 DISP_ADDR_AND_PORT));
269 }
270 #endif /* DEBUG */
271 }
272 if (tcp->tcp_snd_zcopy_aware &&
273 (mp->b_datap->db_struioflag & STRUIO_ZCNOTIFY))
274 tcp_zcopy_notify(tcp);
275 freemsg(mp);
276 mutex_enter(&tcp->tcp_non_sq_lock);
277 if (tcp->tcp_flow_stopped &&
278 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
279 tcp_clrqfull(tcp);
280 }
281 mutex_exit(&tcp->tcp_non_sq_lock);
282 return;
283 }
284
285 /* Strip empties */
286 for (;;) {
287 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
288 (uintptr_t)INT_MAX);
289 len = (int)(mp->b_wptr - mp->b_rptr);
290 if (len > 0)
291 break;
292 mp1 = mp;
293 mp = mp->b_cont;
294 freeb(mp1);
295 if (mp == NULL) {
296 return;
297 }
298 }
299
300 /* If we are the first on the list ... */
301 if (tcp->tcp_xmit_head == NULL) {
302 tcp->tcp_xmit_head = mp;
303 tcp->tcp_xmit_tail = mp;
304 tcp->tcp_xmit_tail_unsent = len;
305 } else {
306 /* If tiny tx and room in txq tail, pullup to save mblks. */
307 struct datab *dp;
308
309 mp1 = tcp->tcp_xmit_last;
310 if (len < tcp_tx_pull_len &&
311 (dp = mp1->b_datap)->db_ref == 1 &&
312 dp->db_lim - mp1->b_wptr >= len) {
313 ASSERT(len > 0);
314 ASSERT(!mp1->b_cont);
315 if (len == 1) {
316 *mp1->b_wptr++ = *mp->b_rptr;
317 } else {
318 bcopy(mp->b_rptr, mp1->b_wptr, len);
319 mp1->b_wptr += len;
320 }
321 if (mp1 == tcp->tcp_xmit_tail)
322 tcp->tcp_xmit_tail_unsent += len;
323 mp1->b_cont = mp->b_cont;
324 if (tcp->tcp_snd_zcopy_aware &&
325 (mp->b_datap->db_struioflag & STRUIO_ZCNOTIFY))
326 mp1->b_datap->db_struioflag |= STRUIO_ZCNOTIFY;
327 freeb(mp);
328 mp = mp1;
329 } else {
330 tcp->tcp_xmit_last->b_cont = mp;
331 }
332 len += tcp->tcp_unsent;
333 }
334
335 /* Tack on however many more positive length mblks we have */
336 if ((mp1 = mp->b_cont) != NULL) {
337 do {
338 int tlen;
339 ASSERT((uintptr_t)(mp1->b_wptr - mp1->b_rptr) <=
340 (uintptr_t)INT_MAX);
341 tlen = (int)(mp1->b_wptr - mp1->b_rptr);
342 if (tlen <= 0) {
343 mp->b_cont = mp1->b_cont;
344 freeb(mp1);
345 } else {
346 len += tlen;
347 mp = mp1;
348 }
349 } while ((mp1 = mp->b_cont) != NULL);
350 }
351 tcp->tcp_xmit_last = mp;
352 tcp->tcp_unsent = len;
353
354 if (urgent)
355 usable = 1;
356
357 data_null:
358 snxt = tcp->tcp_snxt;
359 xmit_tail = tcp->tcp_xmit_tail;
360 tail_unsent = tcp->tcp_xmit_tail_unsent;
361
362 /*
363 * Note that tcp_mss has been adjusted to take into account the
364 * timestamp option if applicable. Because SACK options do not
365 * appear in every TCP segments and they are of variable lengths,
366 * they cannot be included in tcp_mss. Thus we need to calculate
367 * the actual segment length when we need to send a segment which
368 * includes SACK options.
369 */
370 if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
371 int32_t opt_len;
372
373 num_sack_blk = MIN(tcp->tcp_max_sack_blk,
374 tcp->tcp_num_sack_blk);
375 opt_len = num_sack_blk * sizeof (sack_blk_t) + TCPOPT_NOP_LEN *
376 2 + TCPOPT_HEADER_LEN;
377 mss = tcp->tcp_mss - opt_len;
378 total_hdr_len = connp->conn_ht_iphc_len + opt_len;
379 tcp_hdr_len = connp->conn_ht_ulp_len + opt_len;
380 } else {
381 mss = tcp->tcp_mss;
382 total_hdr_len = connp->conn_ht_iphc_len;
383 tcp_hdr_len = connp->conn_ht_ulp_len;
384 }
385
386 if ((tcp->tcp_suna == snxt) && !tcp->tcp_localnet &&
387 (TICK_TO_MSEC(now - tcp->tcp_last_recv_time) >= tcp->tcp_rto)) {
388 cc_after_idle(tcp);
389 }
390 if (tcpstate == TCPS_SYN_RCVD) {
391 /*
392 * The three-way connection establishment handshake is not
393 * complete yet. We want to queue the data for transmission
394 * after entering ESTABLISHED state (RFC793). A jump to
395 * "done" label effectively leaves data on the queue.
396 */
397 goto done;
398 } else {
399 int usable_r;
400
401 /*
402 * In the special case when cwnd is zero, which can only
403 * happen if the connection is ECN capable, return now.
404 * New segments is sent using tcp_timer(). The timer
405 * is set in tcp_input_data().
406 */
407 if (tcp->tcp_cwnd == 0) {
408 /*
409 * Note that tcp_cwnd is 0 before 3-way handshake is
410 * finished.
411 */
412 ASSERT(tcp->tcp_ecn_ok ||
413 tcp->tcp_state < TCPS_ESTABLISHED);
414 return;
415 }
416
417 /* NOTE: trouble if xmitting while SYN not acked? */
418 usable_r = snxt - tcp->tcp_suna;
419 usable_r = tcp->tcp_swnd - usable_r;
420
421 /*
422 * Check if the receiver has shrunk the window. If
423 * tcp_wput_data() with NULL mp is called, tcp_fin_sent
424 * cannot be set as there is unsent data, so FIN cannot
425 * be sent out. Otherwise, we need to take into account
426 * of FIN as it consumes an "invisible" sequence number.
427 */
428 ASSERT(tcp->tcp_fin_sent == 0);
429 if (usable_r < 0) {
430 /*
431 * The receiver has shrunk the window and we have sent
432 * -usable_r date beyond the window, re-adjust.
433 *
434 * If TCP window scaling is enabled, there can be
435 * round down error as the advertised receive window
436 * is actually right shifted n bits. This means that
437 * the lower n bits info is wiped out. It will look
438 * like the window is shrunk. Do a check here to
439 * see if the shrunk amount is actually within the
440 * error in window calculation. If it is, just
441 * return. Note that this check is inside the
442 * shrunk window check. This makes sure that even
443 * though tcp_process_shrunk_swnd() is not called,
444 * we will stop further processing.
445 */
446 if ((-usable_r >> tcp->tcp_snd_ws) > 0) {
447 tcp_process_shrunk_swnd(tcp, -usable_r);
448 }
449 return;
450 }
451
452 /* usable = MIN(swnd, cwnd) - unacked_bytes */
453 if (tcp->tcp_swnd > tcp->tcp_cwnd)
454 usable_r -= tcp->tcp_swnd - tcp->tcp_cwnd;
455
456 /* usable = MIN(usable, unsent) */
457 if (usable_r > len)
458 usable_r = len;
459
460 /* usable = MAX(usable, {1 for urgent, 0 for data}) */
461 if (usable_r > 0) {
462 usable = usable_r;
463 } else {
464 /* Bypass all other unnecessary processing. */
465 goto done;
466 }
467 }
468
469 local_time = (mblk_t *)(intptr_t)gethrtime();
470
471 /*
472 * "Our" Nagle Algorithm. This is not the same as in the old
473 * BSD. This is more in line with the true intent of Nagle.
474 *
475 * The conditions are:
476 * 1. The amount of unsent data (or amount of data which can be
477 * sent, whichever is smaller) is less than Nagle limit.
478 * 2. The last sent size is also less than Nagle limit.
479 * 3. There is unack'ed data.
480 * 4. Urgent pointer is not set. Send urgent data ignoring the
481 * Nagle algorithm. This reduces the probability that urgent
482 * bytes get "merged" together.
483 * 5. The app has not closed the connection. This eliminates the
484 * wait time of the receiving side waiting for the last piece of
485 * (small) data.
486 *
487 * If all are satisified, exit without sending anything. Note
488 * that Nagle limit can be smaller than 1 MSS. Nagle limit is
489 * the smaller of 1 MSS and global tcp_naglim_def (default to be
490 * 4095).
491 */
492 if (usable < (int)tcp->tcp_naglim &&
493 tcp->tcp_naglim > tcp->tcp_last_sent_len &&
494 snxt != tcp->tcp_suna &&
495 !(tcp->tcp_valid_bits & TCP_URG_VALID) &&
496 !(tcp->tcp_valid_bits & TCP_FSS_VALID)) {
497 goto done;
498 }
499
500 /*
501 * If tcp_zero_win_probe is not set and the tcp->tcp_cork option
502 * is set, then we have to force TCP not to send partial segment
503 * (smaller than MSS bytes). We are calculating the usable now
504 * based on full mss and will save the rest of remaining data for
505 * later. When tcp_zero_win_probe is set, TCP needs to send out
506 * something to do zero window probe.
507 */
508 if (tcp->tcp_cork && !tcp->tcp_zero_win_probe) {
509 if (usable < mss)
510 goto done;
511 usable = (usable / mss) * mss;
512 }
513
514 /* Update the latest receive window size in TCP header. */
515 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
516
517 /* Send the packet. */
518 rc = tcp_send(tcp, mss, total_hdr_len, tcp_hdr_len,
519 num_sack_blk, &usable, &snxt, &tail_unsent, &xmit_tail,
520 local_time);
521
522 /* Pretend that all we were trying to send really got sent */
523 if (rc < 0 && tail_unsent < 0) {
524 do {
525 xmit_tail = xmit_tail->b_cont;
526 xmit_tail->b_prev = local_time;
527 ASSERT((uintptr_t)(xmit_tail->b_wptr -
528 xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
529 tail_unsent += (int)(xmit_tail->b_wptr -
530 xmit_tail->b_rptr);
531 } while (tail_unsent < 0);
532 }
533 done:;
534 tcp->tcp_xmit_tail = xmit_tail;
535 tcp->tcp_xmit_tail_unsent = tail_unsent;
536 len = tcp->tcp_snxt - snxt;
537 if (len) {
538 /*
539 * If new data was sent, need to update the notsack
540 * list, which is, afterall, data blocks that have
541 * not been sack'ed by the receiver. New data is
542 * not sack'ed.
543 */
544 if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
545 /* len is a negative value. */
546 tcp->tcp_pipe -= len;
547 tcp_notsack_update(&(tcp->tcp_notsack_list),
548 tcp->tcp_snxt, snxt,
549 &(tcp->tcp_num_notsack_blk),
550 &(tcp->tcp_cnt_notsack_list));
551 }
552 tcp->tcp_snxt = snxt + tcp->tcp_fin_sent;
553 tcp->tcp_rack = tcp->tcp_rnxt;
554 tcp->tcp_rack_cnt = 0;
555 if ((snxt + len) == tcp->tcp_suna) {
556 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
557 }
558 } else if (snxt == tcp->tcp_suna && tcp->tcp_swnd == 0) {
559 /*
560 * Didn't send anything. Make sure the timer is running
561 * so that we will probe a zero window.
562 */
563 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
564 }
565 /* Note that len is the amount we just sent but with a negative sign */
566 tcp->tcp_unsent += len;
567 mutex_enter(&tcp->tcp_non_sq_lock);
568 if (tcp->tcp_flow_stopped) {
569 if (TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
570 tcp_clrqfull(tcp);
571 }
572 } else if (TCP_UNSENT_BYTES(tcp) >= connp->conn_sndbuf) {
573 if (!(tcp->tcp_detached))
574 tcp_setqfull(tcp);
575 }
576 mutex_exit(&tcp->tcp_non_sq_lock);
577 }
578
579 /*
580 * Initial STREAMS write side put() procedure for sockets. It tries to
581 * handle the T_CAPABILITY_REQ which sockfs sends down while setting
582 * up the socket without using the squeue. Non T_CAPABILITY_REQ messages
583 * are handled by tcp_wput() as usual.
584 *
585 * All further messages will also be handled by tcp_wput() because we cannot
586 * be sure that the above short cut is safe later.
587 */
588 int
589 tcp_wput_sock(queue_t *wq, mblk_t *mp)
590 {
591 conn_t *connp = Q_TO_CONN(wq);
592 tcp_t *tcp = connp->conn_tcp;
593 struct T_capability_req *car = (struct T_capability_req *)mp->b_rptr;
594
595 ASSERT(wq->q_qinfo == &tcp_sock_winit);
596 wq->q_qinfo = &tcp_winit;
597
598 ASSERT(IPCL_IS_TCP(connp));
599 ASSERT(TCP_IS_SOCKET(tcp));
600
601 if (DB_TYPE(mp) == M_PCPROTO &&
602 MBLKL(mp) == sizeof (struct T_capability_req) &&
603 car->PRIM_type == T_CAPABILITY_REQ) {
604 tcp_capability_req(tcp, mp);
605 return (0);
606 }
607
608 tcp_wput(wq, mp);
609 return (0);
610 }
611
612 /* ARGSUSED */
613 int
614 tcp_wput_fallback(queue_t *wq, mblk_t *mp)
615 {
616 #ifdef DEBUG
617 cmn_err(CE_CONT, "tcp_wput_fallback: Message during fallback \n");
618 #endif
619 freemsg(mp);
620 return (0);
621 }
622
623 /*
624 * Call by tcp_wput() to handle misc non M_DATA messages.
625 */
626 /* ARGSUSED */
627 static void
628 tcp_wput_nondata(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
629 {
630 conn_t *connp = (conn_t *)arg;
631 tcp_t *tcp = connp->conn_tcp;
632
633 ASSERT(DB_TYPE(mp) != M_IOCTL);
634 /*
635 * TCP is D_MP and qprocsoff() is done towards the end of the tcp_close.
636 * Once the close starts, streamhead and sockfs will not let any data
637 * packets come down (close ensures that there are no threads using the
638 * queue and no new threads will come down) but since qprocsoff()
639 * hasn't happened yet, a M_FLUSH or some non data message might
640 * get reflected back (in response to our own FLUSHRW) and get
641 * processed after tcp_close() is done. The conn would still be valid
642 * because a ref would have added but we need to check the state
643 * before actually processing the packet.
644 */
645 if (TCP_IS_DETACHED(tcp) || (tcp->tcp_state == TCPS_CLOSED)) {
646 freemsg(mp);
647 return;
648 }
649
650 switch (DB_TYPE(mp)) {
651 case M_IOCDATA:
652 tcp_wput_iocdata(tcp, mp);
653 break;
654 case M_FLUSH:
655 tcp_wput_flush(tcp, mp);
656 break;
657 default:
658 ip_wput_nondata(connp->conn_wq, mp);
659 break;
660 }
661 }
662
663 /* tcp_wput_flush is called by tcp_wput_nondata to handle M_FLUSH messages. */
664 static void
665 tcp_wput_flush(tcp_t *tcp, mblk_t *mp)
666 {
667 uchar_t fval = *mp->b_rptr;
668 mblk_t *tail;
669 conn_t *connp = tcp->tcp_connp;
670 queue_t *q = connp->conn_wq;
671
672 /* TODO: How should flush interact with urgent data? */
673 if ((fval & FLUSHW) && tcp->tcp_xmit_head != NULL &&
674 !(tcp->tcp_valid_bits & TCP_URG_VALID)) {
675 /*
676 * Flush only data that has not yet been put on the wire. If
677 * we flush data that we have already transmitted, life, as we
678 * know it, may come to an end.
679 */
680 tail = tcp->tcp_xmit_tail;
681 tail->b_wptr -= tcp->tcp_xmit_tail_unsent;
682 tcp->tcp_xmit_tail_unsent = 0;
683 tcp->tcp_unsent = 0;
684 if (tail->b_wptr != tail->b_rptr)
685 tail = tail->b_cont;
686 if (tail) {
687 mblk_t **excess = &tcp->tcp_xmit_head;
688 for (;;) {
689 mblk_t *mp1 = *excess;
690 if (mp1 == tail)
691 break;
692 tcp->tcp_xmit_tail = mp1;
693 tcp->tcp_xmit_last = mp1;
694 excess = &mp1->b_cont;
695 }
696 *excess = NULL;
697 tcp_close_mpp(&tail);
698 if (tcp->tcp_snd_zcopy_aware)
699 tcp_zcopy_notify(tcp);
700 }
701 /*
702 * We have no unsent data, so unsent must be less than
703 * conn_sndlowat, so re-enable flow.
704 */
705 mutex_enter(&tcp->tcp_non_sq_lock);
706 if (tcp->tcp_flow_stopped) {
707 tcp_clrqfull(tcp);
708 }
709 mutex_exit(&tcp->tcp_non_sq_lock);
710 }
711 /*
712 * TODO: you can't just flush these, you have to increase rwnd for one
713 * thing. For another, how should urgent data interact?
714 */
715 if (fval & FLUSHR) {
716 *mp->b_rptr = fval & ~FLUSHW;
717 /* XXX */
718 qreply(q, mp);
719 return;
720 }
721 freemsg(mp);
722 }
723
724 /*
725 * tcp_wput_iocdata is called by tcp_wput_nondata to handle all M_IOCDATA
726 * messages.
727 */
728 static void
729 tcp_wput_iocdata(tcp_t *tcp, mblk_t *mp)
730 {
731 mblk_t *mp1;
732 struct iocblk *iocp = (struct iocblk *)mp->b_rptr;
733 STRUCT_HANDLE(strbuf, sb);
734 uint_t addrlen;
735 conn_t *connp = tcp->tcp_connp;
736 queue_t *q = connp->conn_wq;
737
738 /* Make sure it is one of ours. */
739 switch (iocp->ioc_cmd) {
740 case TI_GETMYNAME:
741 case TI_GETPEERNAME:
742 break;
743 default:
744 /*
745 * If the conn is closing, then error the ioctl here. Otherwise
746 * use the CONN_IOCTLREF_* macros to hold off tcp_close until
747 * we're done here.
748 */
749 mutex_enter(&connp->conn_lock);
750 if (connp->conn_state_flags & CONN_CLOSING) {
751 mutex_exit(&connp->conn_lock);
752 iocp->ioc_error = EINVAL;
753 mp->b_datap->db_type = M_IOCNAK;
754 iocp->ioc_count = 0;
755 qreply(q, mp);
756 return;
757 }
758
759 CONN_INC_IOCTLREF_LOCKED(connp);
760 ip_wput_nondata(q, mp);
761 CONN_DEC_IOCTLREF(connp);
762 return;
763 }
764 switch (mi_copy_state(q, mp, &mp1)) {
765 case -1:
766 return;
767 case MI_COPY_CASE(MI_COPY_IN, 1):
768 break;
769 case MI_COPY_CASE(MI_COPY_OUT, 1):
770 /* Copy out the strbuf. */
771 mi_copyout(q, mp);
772 return;
773 case MI_COPY_CASE(MI_COPY_OUT, 2):
774 /* All done. */
775 mi_copy_done(q, mp, 0);
776 return;
777 default:
778 mi_copy_done(q, mp, EPROTO);
779 return;
780 }
781 /* Check alignment of the strbuf */
782 if (!OK_32PTR(mp1->b_rptr)) {
783 mi_copy_done(q, mp, EINVAL);
784 return;
785 }
786
787 STRUCT_SET_HANDLE(sb, iocp->ioc_flag, (void *)mp1->b_rptr);
788
789 if (connp->conn_family == AF_INET)
790 addrlen = sizeof (sin_t);
791 else
792 addrlen = sizeof (sin6_t);
793
794 if (STRUCT_FGET(sb, maxlen) < addrlen) {
795 mi_copy_done(q, mp, EINVAL);
796 return;
797 }
798
799 switch (iocp->ioc_cmd) {
800 case TI_GETMYNAME:
801 break;
802 case TI_GETPEERNAME:
803 if (tcp->tcp_state < TCPS_SYN_RCVD) {
804 mi_copy_done(q, mp, ENOTCONN);
805 return;
806 }
807 break;
808 }
809 mp1 = mi_copyout_alloc(q, mp, STRUCT_FGETP(sb, buf), addrlen, B_TRUE);
810 if (!mp1)
811 return;
812
813 STRUCT_FSET(sb, len, addrlen);
814 switch (((struct iocblk *)mp->b_rptr)->ioc_cmd) {
815 case TI_GETMYNAME:
816 (void) conn_getsockname(connp, (struct sockaddr *)mp1->b_wptr,
817 &addrlen);
818 break;
819 case TI_GETPEERNAME:
820 (void) conn_getpeername(connp, (struct sockaddr *)mp1->b_wptr,
821 &addrlen);
822 break;
823 }
824 mp1->b_wptr += addrlen;
825 /* Copy out the address */
826 mi_copyout(q, mp);
827 }
828
829 /*
830 * tcp_wput_ioctl is called by tcp_wput_nondata() to handle all M_IOCTL
831 * messages.
832 */
833 /* ARGSUSED */
834 static void
835 tcp_wput_ioctl(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
836 {
837 conn_t *connp = (conn_t *)arg;
838 tcp_t *tcp = connp->conn_tcp;
839 queue_t *q = connp->conn_wq;
840 struct iocblk *iocp;
841
842 ASSERT(DB_TYPE(mp) == M_IOCTL);
843 /*
844 * Try and ASSERT the minimum possible references on the
845 * conn early enough. Since we are executing on write side,
846 * the connection is obviously not detached and that means
847 * there is a ref each for TCP and IP. Since we are behind
848 * the squeue, the minimum references needed are 3. If the
849 * conn is in classifier hash list, there should be an
850 * extra ref for that (we check both the possibilities).
851 */
852 ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
853 (connp->conn_fanout == NULL && connp->conn_ref >= 3));
854
855 iocp = (struct iocblk *)mp->b_rptr;
856 switch (iocp->ioc_cmd) {
857 case _SIOCSOCKFALLBACK:
858 /*
859 * Either sockmod is about to be popped and the socket
860 * would now be treated as a plain stream, or a module
861 * is about to be pushed so we could no longer use read-
862 * side synchronous streams for fused loopback tcp.
863 * Drain any queued data and disable direct sockfs
864 * interface from now on.
865 */
866 if (!tcp->tcp_issocket) {
867 DB_TYPE(mp) = M_IOCNAK;
868 iocp->ioc_error = EINVAL;
869 } else {
870 tcp_use_pure_tpi(tcp);
871 DB_TYPE(mp) = M_IOCACK;
872 iocp->ioc_error = 0;
873 }
874 iocp->ioc_count = 0;
875 iocp->ioc_rval = 0;
876 qreply(q, mp);
877 return;
878 }
879
880 /*
881 * If the conn is closing, then error the ioctl here. Otherwise bump the
882 * conn_ioctlref to hold off tcp_close until we're done here.
883 */
884 mutex_enter(&(connp)->conn_lock);
885 if ((connp)->conn_state_flags & CONN_CLOSING) {
886 mutex_exit(&(connp)->conn_lock);
887 iocp->ioc_error = EINVAL;
888 mp->b_datap->db_type = M_IOCNAK;
889 iocp->ioc_count = 0;
890 qreply(q, mp);
891 return;
892 }
893
894 CONN_INC_IOCTLREF_LOCKED(connp);
895 ip_wput_nondata(q, mp);
896 CONN_DEC_IOCTLREF(connp);
897 }
898
899 /*
900 * This routine is called by tcp_wput() to handle all TPI requests.
901 */
902 /* ARGSUSED */
903 static void
904 tcp_wput_proto(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
905 {
906 conn_t *connp = (conn_t *)arg;
907 tcp_t *tcp = connp->conn_tcp;
908 union T_primitives *tprim = (union T_primitives *)mp->b_rptr;
909 uchar_t *rptr;
910 t_scalar_t type;
911 cred_t *cr;
912
913 /*
914 * Try and ASSERT the minimum possible references on the
915 * conn early enough. Since we are executing on write side,
916 * the connection is obviously not detached and that means
917 * there is a ref each for TCP and IP. Since we are behind
918 * the squeue, the minimum references needed are 3. If the
919 * conn is in classifier hash list, there should be an
920 * extra ref for that (we check both the possibilities).
921 */
922 ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
923 (connp->conn_fanout == NULL && connp->conn_ref >= 3));
924
925 rptr = mp->b_rptr;
926 ASSERT((uintptr_t)(mp->b_wptr - rptr) <= (uintptr_t)INT_MAX);
927 if ((mp->b_wptr - rptr) >= sizeof (t_scalar_t)) {
928 type = ((union T_primitives *)rptr)->type;
929 if (type == T_EXDATA_REQ) {
930 tcp_output_urgent(connp, mp, arg2, NULL);
931 } else if (type != T_DATA_REQ) {
932 goto non_urgent_data;
933 } else {
934 /* TODO: options, flags, ... from user */
935 /* Set length to zero for reclamation below */
936 tcp_wput_data(tcp, mp->b_cont, B_TRUE);
937 freeb(mp);
938 }
939 return;
940 } else {
941 if (connp->conn_debug) {
942 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
943 "tcp_wput_proto, dropping one...");
944 }
945 freemsg(mp);
946 return;
947 }
948
949 non_urgent_data:
950
951 switch ((int)tprim->type) {
952 case O_T_BIND_REQ: /* bind request */
953 case T_BIND_REQ: /* new semantics bind request */
954 tcp_tpi_bind(tcp, mp);
955 break;
956 case T_UNBIND_REQ: /* unbind request */
957 tcp_tpi_unbind(tcp, mp);
958 break;
959 case O_T_CONN_RES: /* old connection response XXX */
960 case T_CONN_RES: /* connection response */
961 tcp_tli_accept(tcp, mp);
962 break;
963 case T_CONN_REQ: /* connection request */
964 tcp_tpi_connect(tcp, mp);
965 break;
966 case T_DISCON_REQ: /* disconnect request */
967 tcp_disconnect(tcp, mp);
968 break;
969 case T_CAPABILITY_REQ:
970 tcp_capability_req(tcp, mp); /* capability request */
971 break;
972 case T_INFO_REQ: /* information request */
973 tcp_info_req(tcp, mp);
974 break;
975 case T_SVR4_OPTMGMT_REQ: /* manage options req */
976 case T_OPTMGMT_REQ:
977 /*
978 * Note: no support for snmpcom_req() through new
979 * T_OPTMGMT_REQ. See comments in ip.c
980 */
981
982 /*
983 * All Solaris components should pass a db_credp
984 * for this TPI message, hence we ASSERT.
985 * But in case there is some other M_PROTO that looks
986 * like a TPI message sent by some other kernel
987 * component, we check and return an error.
988 */
989 cr = msg_getcred(mp, NULL);
990 ASSERT(cr != NULL);
991 if (cr == NULL) {
992 tcp_err_ack(tcp, mp, TSYSERR, EINVAL);
993 return;
994 }
995 /*
996 * If EINPROGRESS is returned, the request has been queued
997 * for subsequent processing by ip_restart_optmgmt(), which
998 * will do the CONN_DEC_REF().
999 */
1000 if ((int)tprim->type == T_SVR4_OPTMGMT_REQ) {
1001 svr4_optcom_req(connp->conn_wq, mp, cr, &tcp_opt_obj);
1002 } else {
1003 tpi_optcom_req(connp->conn_wq, mp, cr, &tcp_opt_obj);
1004 }
1005 break;
1006
1007 case T_UNITDATA_REQ: /* unitdata request */
1008 tcp_err_ack(tcp, mp, TNOTSUPPORT, 0);
1009 break;
1010 case T_ORDREL_REQ: /* orderly release req */
1011 freemsg(mp);
1012
1013 if (tcp->tcp_fused)
1014 tcp_unfuse(tcp);
1015
1016 if (tcp_xmit_end(tcp) != 0) {
1017 /*
1018 * We were crossing FINs and got a reset from
1019 * the other side. Just ignore it.
1020 */
1021 if (connp->conn_debug) {
1022 (void) strlog(TCP_MOD_ID, 0, 1,
1023 SL_ERROR|SL_TRACE,
1024 "tcp_wput_proto, T_ORDREL_REQ out of "
1025 "state %s",
1026 tcp_display(tcp, NULL,
1027 DISP_ADDR_AND_PORT));
1028 }
1029 }
1030 break;
1031 case T_ADDR_REQ:
1032 tcp_addr_req(tcp, mp);
1033 break;
1034 default:
1035 if (connp->conn_debug) {
1036 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1037 "tcp_wput_proto, bogus TPI msg, type %d",
1038 tprim->type);
1039 }
1040 /*
1041 * We used to M_ERROR. Sending TNOTSUPPORT gives the user
1042 * to recover.
1043 */
1044 tcp_err_ack(tcp, mp, TNOTSUPPORT, 0);
1045 break;
1046 }
1047 }
1048
1049 /*
1050 * Handle special out-of-band ioctl requests (see PSARC/2008/265).
1051 */
1052 static void
1053 tcp_wput_cmdblk(queue_t *q, mblk_t *mp)
1054 {
1055 void *data;
1056 mblk_t *datamp = mp->b_cont;
1057 conn_t *connp = Q_TO_CONN(q);
1058 tcp_t *tcp = connp->conn_tcp;
1059 cmdblk_t *cmdp = (cmdblk_t *)mp->b_rptr;
1060
1061 if (datamp == NULL || MBLKL(datamp) < cmdp->cb_len) {
1062 cmdp->cb_error = EPROTO;
1063 qreply(q, mp);
1064 return;
1065 }
1066
1067 data = datamp->b_rptr;
1068
1069 switch (cmdp->cb_cmd) {
1070 case TI_GETPEERNAME:
1071 if (tcp->tcp_state < TCPS_SYN_RCVD)
1072 cmdp->cb_error = ENOTCONN;
1073 else
1074 cmdp->cb_error = conn_getpeername(connp, data,
1075 &cmdp->cb_len);
1076 break;
1077 case TI_GETMYNAME:
1078 cmdp->cb_error = conn_getsockname(connp, data, &cmdp->cb_len);
1079 break;
1080 default:
1081 cmdp->cb_error = EINVAL;
1082 break;
1083 }
1084
1085 qreply(q, mp);
1086 }
1087
1088 /*
1089 * The TCP fast path write put procedure.
1090 * NOTE: the logic of the fast path is duplicated from tcp_wput_data()
1091 */
1092 /* ARGSUSED */
1093 void
1094 tcp_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1095 {
1096 int len;
1097 int hdrlen;
1098 int plen;
1099 mblk_t *mp1;
1100 uchar_t *rptr;
1101 uint32_t snxt;
1102 tcpha_t *tcpha;
1103 struct datab *db;
1104 uint32_t suna;
1105 uint32_t mss;
1106 ipaddr_t *dst;
1107 ipaddr_t *src;
1108 uint32_t sum;
1109 int usable;
1110 conn_t *connp = (conn_t *)arg;
1111 tcp_t *tcp = connp->conn_tcp;
1112 uint32_t msize;
1113 tcp_stack_t *tcps = tcp->tcp_tcps;
1114 ip_xmit_attr_t *ixa;
1115 clock_t now;
1116
1117 /*
1118 * Try and ASSERT the minimum possible references on the
1119 * conn early enough. Since we are executing on write side,
1120 * the connection is obviously not detached and that means
1121 * there is a ref each for TCP and IP. Since we are behind
1122 * the squeue, the minimum references needed are 3. If the
1123 * conn is in classifier hash list, there should be an
1124 * extra ref for that (we check both the possibilities).
1125 */
1126 ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
1127 (connp->conn_fanout == NULL && connp->conn_ref >= 3));
1128
1129 ASSERT(DB_TYPE(mp) == M_DATA);
1130 msize = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1131
1132 mutex_enter(&tcp->tcp_non_sq_lock);
1133 tcp->tcp_squeue_bytes -= msize;
1134 mutex_exit(&tcp->tcp_non_sq_lock);
1135
1136 /* Bypass tcp protocol for fused tcp loopback */
1137 if (tcp->tcp_fused && tcp_fuse_output(tcp, mp, msize))
1138 return;
1139
1140 mss = tcp->tcp_mss;
1141 /*
1142 * If ZEROCOPY has turned off, try not to send any zero-copy message
1143 * down. Do backoff, now.
1144 */
1145 if (tcp->tcp_snd_zcopy_aware && !tcp->tcp_snd_zcopy_on)
1146 mp = tcp_zcopy_backoff(tcp, mp, B_FALSE);
1147
1148
1149 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1150 len = (int)(mp->b_wptr - mp->b_rptr);
1151
1152 /*
1153 * Criteria for fast path:
1154 *
1155 * 1. no unsent data
1156 * 2. single mblk in request
1157 * 3. connection established
1158 * 4. data in mblk
1159 * 5. len <= mss
1160 * 6. no tcp_valid bits
1161 */
1162 if ((tcp->tcp_unsent != 0) ||
1163 (tcp->tcp_cork) ||
1164 (mp->b_cont != NULL) ||
1165 (tcp->tcp_state != TCPS_ESTABLISHED) ||
1166 (len == 0) ||
1167 (len > mss) ||
1168 (tcp->tcp_valid_bits != 0)) {
1169 tcp_wput_data(tcp, mp, B_FALSE);
1170 return;
1171 }
1172
1173 ASSERT(tcp->tcp_xmit_tail_unsent == 0);
1174 ASSERT(tcp->tcp_fin_sent == 0);
1175
1176 /* queue new packet onto retransmission queue */
1177 if (tcp->tcp_xmit_head == NULL) {
1178 tcp->tcp_xmit_head = mp;
1179 } else {
1180 tcp->tcp_xmit_last->b_cont = mp;
1181 }
1182 tcp->tcp_xmit_last = mp;
1183 tcp->tcp_xmit_tail = mp;
1184
1185 /* find out how much we can send */
1186 /* BEGIN CSTYLED */
1187 /*
1188 * un-acked usable
1189 * |--------------|-----------------|
1190 * tcp_suna tcp_snxt tcp_suna+tcp_swnd
1191 */
1192 /* END CSTYLED */
1193
1194 /* start sending from tcp_snxt */
1195 snxt = tcp->tcp_snxt;
1196
1197 /*
1198 * Check to see if this connection has been idle for some time and no
1199 * ACK is expected. If so, then the congestion window size is no longer
1200 * meaningfully tied to current network conditions.
1201 *
1202 * We reinitialize tcp_cwnd, and slow start again to get back the
1203 * connection's "self-clock" as described in Van Jacobson's 1988 paper
1204 * "Congestion avoidance and control".
1205 */
1206 now = LBOLT_FASTPATH;
1207 if ((tcp->tcp_suna == snxt) && !tcp->tcp_localnet &&
1208 (TICK_TO_MSEC(now - tcp->tcp_last_recv_time) >= tcp->tcp_rto)) {
1209 cc_after_idle(tcp);
1210 }
1211
1212 usable = tcp->tcp_swnd; /* tcp window size */
1213 if (usable > tcp->tcp_cwnd)
1214 usable = tcp->tcp_cwnd; /* congestion window smaller */
1215 usable -= snxt; /* subtract stuff already sent */
1216 suna = tcp->tcp_suna;
1217 usable += suna;
1218 /* usable can be < 0 if the congestion window is smaller */
1219 if (len > usable) {
1220 /* Can't send complete M_DATA in one shot */
1221 goto slow;
1222 }
1223
1224 mutex_enter(&tcp->tcp_non_sq_lock);
1225 if (tcp->tcp_flow_stopped &&
1226 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1227 tcp_clrqfull(tcp);
1228 }
1229 mutex_exit(&tcp->tcp_non_sq_lock);
1230
1231 /*
1232 * determine if anything to send (Nagle).
1233 *
1234 * 1. len < tcp_mss (i.e. small)
1235 * 2. unacknowledged data present
1236 * 3. len < nagle limit
1237 * 4. last packet sent < nagle limit (previous packet sent)
1238 */
1239 if ((len < mss) && (snxt != suna) &&
1240 (len < (int)tcp->tcp_naglim) &&
1241 (tcp->tcp_last_sent_len < tcp->tcp_naglim)) {
1242 /*
1243 * This was the first unsent packet and normally
1244 * mss < xmit_hiwater so there is no need to worry
1245 * about flow control. The next packet will go
1246 * through the flow control check in tcp_wput_data().
1247 */
1248 /* leftover work from above */
1249 tcp->tcp_unsent = len;
1250 tcp->tcp_xmit_tail_unsent = len;
1251
1252 return;
1253 }
1254
1255 /*
1256 * len <= tcp->tcp_mss && len == unsent so no sender silly window. Can
1257 * send now.
1258 */
1259
1260 if (snxt == suna) {
1261 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1262 }
1263
1264 /* we have always sent something */
1265 tcp->tcp_rack_cnt = 0;
1266
1267 tcp->tcp_snxt = snxt + len;
1268 tcp->tcp_rack = tcp->tcp_rnxt;
1269
1270 if ((mp1 = dupb(mp)) == 0)
1271 goto no_memory;
1272 mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
1273 mp->b_next = (mblk_t *)(uintptr_t)snxt;
1274
1275 /* adjust tcp header information */
1276 tcpha = tcp->tcp_tcpha;
1277 tcpha->tha_flags = (TH_ACK|TH_PUSH);
1278
1279 sum = len + connp->conn_ht_ulp_len + connp->conn_sum;
1280 sum = (sum >> 16) + (sum & 0xFFFF);
1281 tcpha->tha_sum = htons(sum);
1282
1283 tcpha->tha_seq = htonl(snxt);
1284
1285 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1286 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1287 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
1288 tcp->tcp_cs.tcp_out_data_segs++;
1289 tcp->tcp_cs.tcp_out_data_bytes += len;
1290
1291 /* Update the latest receive window size in TCP header. */
1292 tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
1293
1294 tcp->tcp_last_sent_len = (ushort_t)len;
1295
1296 plen = len + connp->conn_ht_iphc_len;
1297
1298 ixa = connp->conn_ixa;
1299 ixa->ixa_pktlen = plen;
1300
1301 if (ixa->ixa_flags & IXAF_IS_IPV4) {
1302 tcp->tcp_ipha->ipha_length = htons(plen);
1303 } else {
1304 tcp->tcp_ip6h->ip6_plen = htons(plen - IPV6_HDR_LEN);
1305 }
1306
1307 /* see if we need to allocate a mblk for the headers */
1308 hdrlen = connp->conn_ht_iphc_len;
1309 rptr = mp1->b_rptr - hdrlen;
1310 db = mp1->b_datap;
1311 if ((db->db_ref != 2) || rptr < db->db_base ||
1312 (!OK_32PTR(rptr))) {
1313 /* NOTE: we assume allocb returns an OK_32PTR */
1314 mp = allocb(hdrlen + tcps->tcps_wroff_xtra, BPRI_MED);
1315 if (!mp) {
1316 freemsg(mp1);
1317 goto no_memory;
1318 }
1319 mp->b_cont = mp1;
1320 mp1 = mp;
1321 /* Leave room for Link Level header */
1322 rptr = &mp1->b_rptr[tcps->tcps_wroff_xtra];
1323 mp1->b_wptr = &rptr[hdrlen];
1324 }
1325 mp1->b_rptr = rptr;
1326
1327 /* Fill in the timestamp option. */
1328 if (tcp->tcp_snd_ts_ok) {
1329 U32_TO_BE32(now,
1330 (char *)tcpha + TCP_MIN_HEADER_LENGTH + 4);
1331 U32_TO_BE32(tcp->tcp_ts_recent,
1332 (char *)tcpha + TCP_MIN_HEADER_LENGTH + 8);
1333 } else {
1334 ASSERT(connp->conn_ht_ulp_len == TCP_MIN_HEADER_LENGTH);
1335 }
1336
1337 /* copy header into outgoing packet */
1338 dst = (ipaddr_t *)rptr;
1339 src = (ipaddr_t *)connp->conn_ht_iphc;
1340 dst[0] = src[0];
1341 dst[1] = src[1];
1342 dst[2] = src[2];
1343 dst[3] = src[3];
1344 dst[4] = src[4];
1345 dst[5] = src[5];
1346 dst[6] = src[6];
1347 dst[7] = src[7];
1348 dst[8] = src[8];
1349 dst[9] = src[9];
1350 if (hdrlen -= 40) {
1351 hdrlen >>= 2;
1352 dst += 10;
1353 src += 10;
1354 do {
1355 *dst++ = *src++;
1356 } while (--hdrlen);
1357 }
1358
1359 /*
1360 * Set the ECN info in the TCP header. Note that this
1361 * is not the template header.
1362 */
1363 if (tcp->tcp_ecn_ok) {
1364 TCP_SET_ECT(tcp, rptr);
1365
1366 tcpha = (tcpha_t *)(rptr + ixa->ixa_ip_hdr_length);
1367 if (tcp->tcp_ecn_echo_on)
1368 tcpha->tha_flags |= TH_ECE;
1369 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
1370 tcpha->tha_flags |= TH_CWR;
1371 tcp->tcp_ecn_cwr_sent = B_TRUE;
1372 }
1373 }
1374
1375 if (tcp->tcp_ip_forward_progress) {
1376 tcp->tcp_ip_forward_progress = B_FALSE;
1377 connp->conn_ixa->ixa_flags |= IXAF_REACH_CONF;
1378 } else {
1379 connp->conn_ixa->ixa_flags &= ~IXAF_REACH_CONF;
1380 }
1381 tcp_send_data(tcp, mp1);
1382 return;
1383
1384 /*
1385 * If we ran out of memory, we pretend to have sent the packet
1386 * and that it was lost on the wire.
1387 */
1388 no_memory:
1389 return;
1390
1391 slow:
1392 /* leftover work from above */
1393 tcp->tcp_unsent = len;
1394 tcp->tcp_xmit_tail_unsent = len;
1395 tcp_wput_data(tcp, NULL, B_FALSE);
1396 }
1397
1398 /* ARGSUSED2 */
1399 void
1400 tcp_output_urgent(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1401 {
1402 int len;
1403 uint32_t msize;
1404 conn_t *connp = (conn_t *)arg;
1405 tcp_t *tcp = connp->conn_tcp;
1406
1407 msize = msgdsize(mp);
1408
1409 len = msize - 1;
1410 if (len < 0) {
1411 freemsg(mp);
1412 return;
1413 }
1414
1415 /*
1416 * Try to force urgent data out on the wire. Even if we have unsent
1417 * data this will at least send the urgent flag.
1418 * XXX does not handle more flag correctly.
1419 */
1420 len += tcp->tcp_unsent;
1421 len += tcp->tcp_snxt;
1422 tcp->tcp_urg = len;
1423 tcp->tcp_valid_bits |= TCP_URG_VALID;
1424
1425 /* Bypass tcp protocol for fused tcp loopback */
1426 if (tcp->tcp_fused && tcp_fuse_output(tcp, mp, msize))
1427 return;
1428
1429 /* Strip off the T_EXDATA_REQ if the data is from TPI */
1430 if (DB_TYPE(mp) != M_DATA) {
1431 mblk_t *mp1 = mp;
1432 ASSERT(!IPCL_IS_NONSTR(connp));
1433 mp = mp->b_cont;
1434 freeb(mp1);
1435 }
1436 tcp_wput_data(tcp, mp, B_TRUE);
1437 }
1438
1439 /*
1440 * Called by streams close routine via squeues when our client blows off its
1441 * descriptor, we take this to mean: "close the stream state NOW, close the tcp
1442 * connection politely" When SO_LINGER is set (with a non-zero linger time and
1443 * it is not a nonblocking socket) then this routine sleeps until the FIN is
1444 * acked.
1445 *
1446 * NOTE: tcp_close potentially returns error when lingering.
1447 * However, the stream head currently does not pass these errors
1448 * to the application. 4.4BSD only returns EINTR and EWOULDBLOCK
1449 * errors to the application (from tsleep()) and not errors
1450 * like ECONNRESET caused by receiving a reset packet.
1451 */
1452
1453 /* ARGSUSED */
1454 void
1455 tcp_close_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1456 {
1457 char *msg;
1458 conn_t *connp = (conn_t *)arg;
1459 tcp_t *tcp = connp->conn_tcp;
1460 clock_t delta = 0;
1461 tcp_stack_t *tcps = tcp->tcp_tcps;
1462
1463 /*
1464 * When a non-STREAMS socket is being closed, it does not always
1465 * stick around waiting for tcp_close_output to run and can therefore
1466 * have dropped a reference already. So adjust the asserts accordingly.
1467 */
1468 ASSERT((connp->conn_fanout != NULL &&
1469 connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 3 : 4)) ||
1470 (connp->conn_fanout == NULL &&
1471 connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 2 : 3)));
1472
1473 mutex_enter(&tcp->tcp_eager_lock);
1474 if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
1475 /*
1476 * Cleanup for listener. For non-STREAM sockets sockfs will
1477 * close all the eagers on 'q', so in that case only deal
1478 * with 'q0'.
1479 */
1480 tcp_eager_cleanup(tcp, IPCL_IS_NONSTR(connp) ? 1 : 0);
1481 tcp->tcp_wait_for_eagers = 1;
1482 }
1483 mutex_exit(&tcp->tcp_eager_lock);
1484
1485 tcp->tcp_lso = B_FALSE;
1486
1487 msg = NULL;
1488 switch (tcp->tcp_state) {
1489 case TCPS_CLOSED:
1490 case TCPS_IDLE:
1491 break;
1492 case TCPS_BOUND:
1493 if (tcp->tcp_listener != NULL) {
1494 ASSERT(IPCL_IS_NONSTR(connp));
1495 /*
1496 * Unlink from the listener and drop the reference
1497 * put on it by the eager. tcp_closei_local will not
1498 * do it because tcp_tconnind_started is TRUE.
1499 */
1500 mutex_enter(&tcp->tcp_saved_listener->tcp_eager_lock);
1501 tcp_eager_unlink(tcp);
1502 mutex_exit(&tcp->tcp_saved_listener->tcp_eager_lock);
1503 CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
1504 }
1505 break;
1506 case TCPS_LISTEN:
1507 break;
1508 case TCPS_SYN_SENT:
1509 msg = "tcp_close, during connect";
1510 break;
1511 case TCPS_SYN_RCVD:
1512 /*
1513 * Close during the connect 3-way handshake
1514 * but here there may or may not be pending data
1515 * already on queue. Process almost same as in
1516 * the ESTABLISHED state.
1517 */
1518 /* FALLTHRU */
1519 default:
1520 if (tcp->tcp_fused)
1521 tcp_unfuse(tcp);
1522
1523 /*
1524 * If SO_LINGER has set a zero linger time, abort the
1525 * connection with a reset.
1526 */
1527 if (connp->conn_linger && connp->conn_lingertime == 0) {
1528 msg = "tcp_close, zero lingertime";
1529 break;
1530 }
1531
1532 /*
1533 * Abort connection if there is unread data queued.
1534 */
1535 if (tcp->tcp_rcv_list || tcp->tcp_reass_head) {
1536 msg = "tcp_close, unread data";
1537 break;
1538 }
1539
1540 /*
1541 * Abort connection if it is being closed without first
1542 * being accepted. This can happen if a listening non-STREAM
1543 * socket wants to get rid of the socket, for example, if the
1544 * listener is closing.
1545 */
1546 if (tcp->tcp_listener != NULL) {
1547 ASSERT(IPCL_IS_NONSTR(connp));
1548 msg = "tcp_close, close before accept";
1549
1550 /*
1551 * Unlink from the listener and drop the reference
1552 * put on it by the eager. tcp_closei_local will not
1553 * do it because tcp_tconnind_started is TRUE.
1554 */
1555 mutex_enter(&tcp->tcp_saved_listener->tcp_eager_lock);
1556 tcp_eager_unlink(tcp);
1557 mutex_exit(&tcp->tcp_saved_listener->tcp_eager_lock);
1558 CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
1559 break;
1560 }
1561
1562 /*
1563 * Transmit the FIN before detaching the tcp_t.
1564 * After tcp_detach returns this queue/perimeter
1565 * no longer owns the tcp_t thus others can modify it.
1566 */
1567 (void) tcp_xmit_end(tcp);
1568
1569 /*
1570 * If lingering on close then wait until the fin is acked,
1571 * the SO_LINGER time passes, or a reset is sent/received.
1572 */
1573 if (connp->conn_linger && connp->conn_lingertime > 0 &&
1574 !(tcp->tcp_fin_acked) &&
1575 tcp->tcp_state >= TCPS_ESTABLISHED) {
1576 if (tcp->tcp_closeflags & (FNDELAY|FNONBLOCK)) {
1577 tcp->tcp_client_errno = EWOULDBLOCK;
1578 } else if (tcp->tcp_client_errno == 0) {
1579
1580 ASSERT(tcp->tcp_linger_tid == 0);
1581
1582 /* conn_lingertime is in sec. */
1583 tcp->tcp_linger_tid = TCP_TIMER(tcp,
1584 tcp_close_linger_timeout,
1585 connp->conn_lingertime * MILLISEC);
1586
1587 /* tcp_close_linger_timeout will finish close */
1588 if (tcp->tcp_linger_tid == 0)
1589 tcp->tcp_client_errno = ENOSR;
1590 else
1591 return;
1592 }
1593
1594 /*
1595 * Check if we need to detach or just close
1596 * the instance.
1597 */
1598 if (tcp->tcp_state <= TCPS_LISTEN)
1599 break;
1600 }
1601
1602 /*
1603 * Make sure that no other thread will access the conn_rq of
1604 * this instance (through lookups etc.) as conn_rq will go
1605 * away shortly.
1606 */
1607 tcp_acceptor_hash_remove(tcp);
1608
1609 mutex_enter(&tcp->tcp_non_sq_lock);
1610 if (tcp->tcp_flow_stopped) {
1611 tcp_clrqfull(tcp);
1612 }
1613 mutex_exit(&tcp->tcp_non_sq_lock);
1614
1615 if (tcp->tcp_timer_tid != 0) {
1616 delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
1617 tcp->tcp_timer_tid = 0;
1618 }
1619 /*
1620 * Need to cancel those timers which will not be used when
1621 * TCP is detached. This has to be done before the conn_wq
1622 * is set to NULL.
1623 */
1624 tcp_timers_stop(tcp);
1625
1626 tcp->tcp_detached = B_TRUE;
1627 if (tcp->tcp_state == TCPS_TIME_WAIT) {
1628 tcp_time_wait_append(tcp);
1629 TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1630 ASSERT(connp->conn_ref >=
1631 (IPCL_IS_NONSTR(connp) ? 2 : 3));
1632 goto finish;
1633 }
1634
1635 /*
1636 * If delta is zero the timer event wasn't executed and was
1637 * successfully canceled. In this case we need to restart it
1638 * with the minimal delta possible.
1639 */
1640 if (delta >= 0)
1641 tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1642 delta ? delta : 1);
1643
1644 ASSERT(connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 2 : 3));
1645 goto finish;
1646 }
1647
1648 /* Detach did not complete. Still need to remove q from stream. */
1649 if (msg) {
1650 if (tcp->tcp_state == TCPS_ESTABLISHED ||
1651 tcp->tcp_state == TCPS_CLOSE_WAIT)
1652 TCPS_BUMP_MIB(tcps, tcpEstabResets);
1653 if (tcp->tcp_state == TCPS_SYN_SENT ||
1654 tcp->tcp_state == TCPS_SYN_RCVD)
1655 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1656 tcp_xmit_ctl(msg, tcp, tcp->tcp_snxt, 0, TH_RST);
1657 }
1658
1659 tcp_closei_local(tcp);
1660 CONN_DEC_REF(connp);
1661 ASSERT(connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 1 : 2));
1662
1663 finish:
1664 /*
1665 * Don't change the queues in the case of a listener that has
1666 * eagers in its q or q0. It could surprise the eagers.
1667 * Instead wait for the eagers outside the squeue.
1668 *
1669 * For non-STREAMS sockets tcp_wait_for_eagers implies that
1670 * we should delay the su_closed upcall until all eagers have
1671 * dropped their references.
1672 */
1673 if (!tcp->tcp_wait_for_eagers) {
1674 tcp->tcp_detached = B_TRUE;
1675 connp->conn_rq = NULL;
1676 connp->conn_wq = NULL;
1677
1678 /* non-STREAM socket, release the upper handle */
1679 if (IPCL_IS_NONSTR(connp)) {
1680 ASSERT(connp->conn_upper_handle != NULL);
1681 (*connp->conn_upcalls->su_closed)
1682 (connp->conn_upper_handle);
1683 connp->conn_upper_handle = NULL;
1684 connp->conn_upcalls = NULL;
1685 }
1686 }
1687
1688 /* Signal tcp_close() to finish closing. */
1689 mutex_enter(&tcp->tcp_closelock);
1690 tcp->tcp_closed = 1;
1691 cv_signal(&tcp->tcp_closecv);
1692 mutex_exit(&tcp->tcp_closelock);
1693 }
1694
1695 /* ARGSUSED */
1696 void
1697 tcp_shutdown_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1698 {
1699 conn_t *connp = (conn_t *)arg;
1700 tcp_t *tcp = connp->conn_tcp;
1701
1702 freemsg(mp);
1703
1704 if (tcp->tcp_fused)
1705 tcp_unfuse(tcp);
1706
1707 if (tcp_xmit_end(tcp) != 0) {
1708 /*
1709 * We were crossing FINs and got a reset from
1710 * the other side. Just ignore it.
1711 */
1712 if (connp->conn_debug) {
1713 (void) strlog(TCP_MOD_ID, 0, 1,
1714 SL_ERROR|SL_TRACE,
1715 "tcp_shutdown_output() out of state %s",
1716 tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
1717 }
1718 }
1719 }
1720
1721 #pragma inline(tcp_send_data)
1722
1723 void
1724 tcp_send_data(tcp_t *tcp, mblk_t *mp)
1725 {
1726 conn_t *connp = tcp->tcp_connp;
1727
1728 /*
1729 * Check here to avoid sending zero-copy message down to IP when
1730 * ZEROCOPY capability has turned off. We only need to deal with
1731 * the race condition between sockfs and the notification here.
1732 * Since we have tried to backoff the tcp_xmit_head when turning
1733 * zero-copy off and new messages in tcp_output(), we simply drop
1734 * the dup'ed packet here and let tcp retransmit, if tcp_xmit_zc_clean
1735 * is not true.
1736 */
1737 if (tcp->tcp_snd_zcopy_aware && !tcp->tcp_snd_zcopy_on &&
1738 !tcp->tcp_xmit_zc_clean) {
1739 ip_drop_output("TCP ZC was disabled but not clean", mp, NULL);
1740 freemsg(mp);
1741 return;
1742 }
1743
1744 DTRACE_TCP5(send, mblk_t *, NULL, ip_xmit_attr_t *, connp->conn_ixa,
1745 __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, tcp,
1746 __dtrace_tcp_tcph_t *,
1747 &mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
1748
1749 ASSERT(connp->conn_ixa->ixa_notify_cookie == connp->conn_tcp);
1750 (void) conn_ip_output(mp, connp->conn_ixa);
1751 }
1752
1753 /* ARGSUSED2 */
1754 void
1755 tcp_send_synack(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1756 {
1757 conn_t *econnp = (conn_t *)arg;
1758 tcp_t *tcp = econnp->conn_tcp;
1759 ip_xmit_attr_t *ixa = econnp->conn_ixa;
1760
1761 /* Guard against a RST having blown it away while on the squeue */
1762 if (tcp->tcp_state == TCPS_CLOSED) {
1763 freemsg(mp);
1764 return;
1765 }
1766
1767 /*
1768 * In the off-chance that the eager received and responded to
1769 * some other packet while the SYN|ACK was queued, we recalculate
1770 * the ixa_pktlen. It would be better to fix the SYN/accept
1771 * multithreading scheme to avoid this complexity.
1772 */
1773 ixa->ixa_pktlen = msgdsize(mp);
1774 (void) conn_ip_output(mp, ixa);
1775 }
1776
1777 /*
1778 * tcp_send() is called by tcp_wput_data() and returns one of the following:
1779 *
1780 * -1 = failed allocation.
1781 * 0 = We've either successfully sent data, or our usable send window is too
1782 * small and we'd rather wait until later before sending again.
1783 */
1784 static int
1785 tcp_send(tcp_t *tcp, const int mss, const int total_hdr_len,
1786 const int tcp_hdr_len, const int num_sack_blk, int *usable,
1787 uint32_t *snxt, int *tail_unsent, mblk_t **xmit_tail, mblk_t *local_time)
1788 {
1789 int num_lso_seg = 1;
1790 uint_t lso_usable;
1791 boolean_t do_lso_send = B_FALSE;
1792 tcp_stack_t *tcps = tcp->tcp_tcps;
1793 conn_t *connp = tcp->tcp_connp;
1794 ip_xmit_attr_t *ixa = connp->conn_ixa;
1795
1796 /*
1797 * Check LSO possibility. The value of tcp->tcp_lso indicates whether
1798 * the underlying connection is LSO capable. Will check whether having
1799 * enough available data to initiate LSO transmission in the for(){}
1800 * loops.
1801 */
1802 if (tcp->tcp_lso && (tcp->tcp_valid_bits & ~TCP_FSS_VALID) == 0)
1803 do_lso_send = B_TRUE;
1804
1805 for (;;) {
1806 struct datab *db;
1807 tcpha_t *tcpha;
1808 uint32_t sum;
1809 mblk_t *mp, *mp1;
1810 uchar_t *rptr;
1811 int len;
1812
1813 /*
1814 * Calculate the maximum payload length we can send at one
1815 * time.
1816 */
1817 if (do_lso_send) {
1818 /*
1819 * Determine whether or not it's possible to do LSO,
1820 * and if so, how much data we can send.
1821 */
1822 if ((*usable - 1) / mss >= 1) {
1823 lso_usable = MIN(tcp->tcp_lso_max, *usable);
1824 num_lso_seg = lso_usable / mss;
1825 if (lso_usable % mss) {
1826 num_lso_seg++;
1827 tcp->tcp_last_sent_len = (ushort_t)
1828 (lso_usable % mss);
1829 } else {
1830 tcp->tcp_last_sent_len = (ushort_t)mss;
1831 }
1832 } else {
1833 do_lso_send = B_FALSE;
1834 num_lso_seg = 1;
1835 lso_usable = mss;
1836 }
1837 }
1838
1839 ASSERT(num_lso_seg <= IP_MAXPACKET / mss + 1);
1840
1841 len = mss;
1842 if (len > *usable) {
1843 ASSERT(do_lso_send == B_FALSE);
1844
1845 len = *usable;
1846 if (len <= 0) {
1847 /* Terminate the loop */
1848 break; /* success; too small */
1849 }
1850 /*
1851 * Sender silly-window avoidance.
1852 * Ignore this if we are going to send a
1853 * zero window probe out.
1854 *
1855 * TODO: force data into microscopic window?
1856 * ==> (!pushed || (unsent > usable))
1857 */
1858 if (len < (tcp->tcp_max_swnd >> 1) &&
1859 (tcp->tcp_unsent - (*snxt - tcp->tcp_snxt)) > len &&
1860 !((tcp->tcp_valid_bits & TCP_URG_VALID) &&
1861 len == 1) && (! tcp->tcp_zero_win_probe)) {
1862 /*
1863 * If the retransmit timer is not running
1864 * we start it so that we will retransmit
1865 * in the case when the receiver has
1866 * decremented the window.
1867 */
1868 if (*snxt == tcp->tcp_snxt &&
1869 *snxt == tcp->tcp_suna) {
1870 /*
1871 * We are not supposed to send
1872 * anything. So let's wait a little
1873 * bit longer before breaking SWS
1874 * avoidance.
1875 *
1876 * What should the value be?
1877 * Suggestion: MAX(init rexmit time,
1878 * tcp->tcp_rto)
1879 */
1880 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1881 }
1882 break; /* success; too small */
1883 }
1884 }
1885
1886 tcpha = tcp->tcp_tcpha;
1887
1888 /*
1889 * The reason to adjust len here is that we need to set flags
1890 * and calculate checksum.
1891 */
1892 if (do_lso_send)
1893 len = lso_usable;
1894
1895 *usable -= len; /* Approximate - can be adjusted later */
1896 if (*usable > 0)
1897 tcpha->tha_flags = TH_ACK;
1898 else
1899 tcpha->tha_flags = (TH_ACK | TH_PUSH);
1900
1901 /*
1902 * Prime pump for IP's checksumming on our behalf.
1903 * Include the adjustment for a source route if any.
1904 * In case of LSO, the partial pseudo-header checksum should
1905 * exclusive TCP length, so zero tha_sum before IP calculate
1906 * pseudo-header checksum for partial checksum offload.
1907 */
1908 if (do_lso_send) {
1909 sum = 0;
1910 } else {
1911 sum = len + tcp_hdr_len + connp->conn_sum;
1912 sum = (sum >> 16) + (sum & 0xFFFF);
1913 }
1914 tcpha->tha_sum = htons(sum);
1915 tcpha->tha_seq = htonl(*snxt);
1916
1917 /*
1918 * Branch off to tcp_xmit_mp() if any of the VALID bits is
1919 * set. For the case when TCP_FSS_VALID is the only valid
1920 * bit (normal active close), branch off only when we think
1921 * that the FIN flag needs to be set. Note for this case,
1922 * that (snxt + len) may not reflect the actual seg_len,
1923 * as len may be further reduced in tcp_xmit_mp(). If len
1924 * gets modified, we will end up here again.
1925 */
1926 if (tcp->tcp_valid_bits != 0 &&
1927 (tcp->tcp_valid_bits != TCP_FSS_VALID ||
1928 ((*snxt + len) == tcp->tcp_fss))) {
1929 uchar_t *prev_rptr;
1930 uint32_t prev_snxt = tcp->tcp_snxt;
1931
1932 if (*tail_unsent == 0) {
1933 ASSERT((*xmit_tail)->b_cont != NULL);
1934 *xmit_tail = (*xmit_tail)->b_cont;
1935 prev_rptr = (*xmit_tail)->b_rptr;
1936 *tail_unsent = (int)((*xmit_tail)->b_wptr -
1937 (*xmit_tail)->b_rptr);
1938 } else {
1939 prev_rptr = (*xmit_tail)->b_rptr;
1940 (*xmit_tail)->b_rptr = (*xmit_tail)->b_wptr -
1941 *tail_unsent;
1942 }
1943 mp = tcp_xmit_mp(tcp, *xmit_tail, len, NULL, NULL,
1944 *snxt, B_FALSE, (uint32_t *)&len, B_FALSE);
1945 /* Restore tcp_snxt so we get amount sent right. */
1946 tcp->tcp_snxt = prev_snxt;
1947 if (prev_rptr == (*xmit_tail)->b_rptr) {
1948 /*
1949 * If the previous timestamp is still in use,
1950 * don't stomp on it.
1951 */
1952 if ((*xmit_tail)->b_next == NULL) {
1953 (*xmit_tail)->b_prev = local_time;
1954 (*xmit_tail)->b_next =
1955 (mblk_t *)(uintptr_t)(*snxt);
1956 }
1957 } else
1958 (*xmit_tail)->b_rptr = prev_rptr;
1959
1960 if (mp == NULL) {
1961 return (-1);
1962 }
1963 mp1 = mp->b_cont;
1964
1965 if (len <= mss) /* LSO is unusable (!do_lso_send) */
1966 tcp->tcp_last_sent_len = (ushort_t)len;
1967 while (mp1->b_cont) {
1968 *xmit_tail = (*xmit_tail)->b_cont;
1969 (*xmit_tail)->b_prev = local_time;
1970 (*xmit_tail)->b_next =
1971 (mblk_t *)(uintptr_t)(*snxt);
1972 mp1 = mp1->b_cont;
1973 }
1974 *snxt += len;
1975 *tail_unsent = (*xmit_tail)->b_wptr - mp1->b_wptr;
1976 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
1977 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1978 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1979 tcp->tcp_cs.tcp_out_data_segs++;
1980 tcp->tcp_cs.tcp_out_data_bytes += len;
1981 tcp_send_data(tcp, mp);
1982 continue;
1983 }
1984
1985 *snxt += len; /* Adjust later if we don't send all of len */
1986 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
1987 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1988 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1989 tcp->tcp_cs.tcp_out_data_segs++;
1990 tcp->tcp_cs.tcp_out_data_bytes += len;
1991
1992 if (*tail_unsent) {
1993 /* Are the bytes above us in flight? */
1994 rptr = (*xmit_tail)->b_wptr - *tail_unsent;
1995 if (rptr != (*xmit_tail)->b_rptr) {
1996 *tail_unsent -= len;
1997 if (len <= mss) /* LSO is unusable */
1998 tcp->tcp_last_sent_len = (ushort_t)len;
1999 len += total_hdr_len;
2000 ixa->ixa_pktlen = len;
2001
2002 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2003 tcp->tcp_ipha->ipha_length = htons(len);
2004 } else {
2005 tcp->tcp_ip6h->ip6_plen =
2006 htons(len - IPV6_HDR_LEN);
2007 }
2008
2009 mp = dupb(*xmit_tail);
2010 if (mp == NULL) {
2011 return (-1); /* out_of_mem */
2012 }
2013 mp->b_rptr = rptr;
2014 /*
2015 * If the old timestamp is no longer in use,
2016 * sample a new timestamp now.
2017 */
2018 if ((*xmit_tail)->b_next == NULL) {
2019 (*xmit_tail)->b_prev = local_time;
2020 (*xmit_tail)->b_next =
2021 (mblk_t *)(uintptr_t)(*snxt-len);
2022 }
2023 goto must_alloc;
2024 }
2025 } else {
2026 *xmit_tail = (*xmit_tail)->b_cont;
2027 ASSERT((uintptr_t)((*xmit_tail)->b_wptr -
2028 (*xmit_tail)->b_rptr) <= (uintptr_t)INT_MAX);
2029 *tail_unsent = (int)((*xmit_tail)->b_wptr -
2030 (*xmit_tail)->b_rptr);
2031 }
2032
2033 (*xmit_tail)->b_prev = local_time;
2034 (*xmit_tail)->b_next = (mblk_t *)(uintptr_t)(*snxt - len);
2035
2036 *tail_unsent -= len;
2037 if (len <= mss) /* LSO is unusable (!do_lso_send) */
2038 tcp->tcp_last_sent_len = (ushort_t)len;
2039
2040 len += total_hdr_len;
2041 ixa->ixa_pktlen = len;
2042
2043 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2044 tcp->tcp_ipha->ipha_length = htons(len);
2045 } else {
2046 tcp->tcp_ip6h->ip6_plen = htons(len - IPV6_HDR_LEN);
2047 }
2048
2049 mp = dupb(*xmit_tail);
2050 if (mp == NULL) {
2051 return (-1); /* out_of_mem */
2052 }
2053
2054 len = total_hdr_len;
2055 /*
2056 * There are four reasons to allocate a new hdr mblk:
2057 * 1) The bytes above us are in use by another packet
2058 * 2) We don't have good alignment
2059 * 3) The mblk is being shared
2060 * 4) We don't have enough room for a header
2061 */
2062 rptr = mp->b_rptr - len;
2063 if (!OK_32PTR(rptr) ||
2064 ((db = mp->b_datap), db->db_ref != 2) ||
2065 rptr < db->db_base) {
2066 /* NOTE: we assume allocb returns an OK_32PTR */
2067
2068 must_alloc:;
2069 mp1 = allocb(connp->conn_ht_iphc_allocated +
2070 tcps->tcps_wroff_xtra, BPRI_MED);
2071 if (mp1 == NULL) {
2072 freemsg(mp);
2073 return (-1); /* out_of_mem */
2074 }
2075 mp1->b_cont = mp;
2076 mp = mp1;
2077 /* Leave room for Link Level header */
2078 len = total_hdr_len;
2079 rptr = &mp->b_rptr[tcps->tcps_wroff_xtra];
2080 mp->b_wptr = &rptr[len];
2081 }
2082
2083 /*
2084 * Fill in the header using the template header, and add
2085 * options such as time-stamp, ECN and/or SACK, as needed.
2086 */
2087 tcp_fill_header(tcp, rptr, num_sack_blk);
2088
2089 mp->b_rptr = rptr;
2090
2091 if (*tail_unsent) {
2092 int spill = *tail_unsent;
2093
2094 mp1 = mp->b_cont;
2095 if (mp1 == NULL)
2096 mp1 = mp;
2097
2098 /*
2099 * If we're a little short, tack on more mblks until
2100 * there is no more spillover.
2101 */
2102 while (spill < 0) {
2103 mblk_t *nmp;
2104 int nmpsz;
2105
2106 nmp = (*xmit_tail)->b_cont;
2107 nmpsz = MBLKL(nmp);
2108
2109 /*
2110 * Excess data in mblk; can we split it?
2111 * If LSO is enabled for the connection,
2112 * keep on splitting as this is a transient
2113 * send path.
2114 */
2115 if (!do_lso_send && (spill + nmpsz > 0)) {
2116 /*
2117 * Don't split if stream head was
2118 * told to break up larger writes
2119 * into smaller ones.
2120 */
2121 if (tcp->tcp_maxpsz_multiplier > 0)
2122 break;
2123
2124 /*
2125 * Next mblk is less than SMSS/2
2126 * rounded up to nearest 64-byte;
2127 * let it get sent as part of the
2128 * next segment.
2129 */
2130 if (tcp->tcp_localnet &&
2131 !tcp->tcp_cork &&
2132 (nmpsz < roundup((mss >> 1), 64)))
2133 break;
2134 }
2135
2136 *xmit_tail = nmp;
2137 ASSERT((uintptr_t)nmpsz <= (uintptr_t)INT_MAX);
2138 /* Stash for rtt use later */
2139 (*xmit_tail)->b_prev = local_time;
2140 (*xmit_tail)->b_next =
2141 (mblk_t *)(uintptr_t)(*snxt - len);
2142 mp1->b_cont = dupb(*xmit_tail);
2143 mp1 = mp1->b_cont;
2144
2145 spill += nmpsz;
2146 if (mp1 == NULL) {
2147 *tail_unsent = spill;
2148 freemsg(mp);
2149 return (-1); /* out_of_mem */
2150 }
2151 }
2152
2153 /* Trim back any surplus on the last mblk */
2154 if (spill >= 0) {
2155 mp1->b_wptr -= spill;
2156 *tail_unsent = spill;
2157 } else {
2158 /*
2159 * We did not send everything we could in
2160 * order to remain within the b_cont limit.
2161 */
2162 *usable -= spill;
2163 *snxt += spill;
2164 tcp->tcp_last_sent_len += spill;
2165 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, spill);
2166 tcp->tcp_cs.tcp_out_data_bytes += spill;
2167 /*
2168 * Adjust the checksum
2169 */
2170 tcpha = (tcpha_t *)(rptr +
2171 ixa->ixa_ip_hdr_length);
2172 sum += spill;
2173 sum = (sum >> 16) + (sum & 0xFFFF);
2174 tcpha->tha_sum = htons(sum);
2175 if (connp->conn_ipversion == IPV4_VERSION) {
2176 sum = ntohs(
2177 ((ipha_t *)rptr)->ipha_length) +
2178 spill;
2179 ((ipha_t *)rptr)->ipha_length =
2180 htons(sum);
2181 } else {
2182 sum = ntohs(
2183 ((ip6_t *)rptr)->ip6_plen) +
2184 spill;
2185 ((ip6_t *)rptr)->ip6_plen =
2186 htons(sum);
2187 }
2188 ixa->ixa_pktlen += spill;
2189 *tail_unsent = 0;
2190 }
2191 }
2192 if (tcp->tcp_ip_forward_progress) {
2193 tcp->tcp_ip_forward_progress = B_FALSE;
2194 ixa->ixa_flags |= IXAF_REACH_CONF;
2195 } else {
2196 ixa->ixa_flags &= ~IXAF_REACH_CONF;
2197 }
2198
2199 if (do_lso_send) {
2200 /* Append LSO information to the mp. */
2201 lso_info_set(mp, mss, HW_LSO);
2202 ixa->ixa_fragsize = IP_MAXPACKET;
2203 ixa->ixa_extra_ident = num_lso_seg - 1;
2204
2205 DTRACE_PROBE2(tcp_send_lso, int, num_lso_seg,
2206 boolean_t, B_TRUE);
2207
2208 tcp_send_data(tcp, mp);
2209
2210 /*
2211 * Restore values of ixa_fragsize and ixa_extra_ident.
2212 */
2213 ixa->ixa_fragsize = ixa->ixa_pmtu;
2214 ixa->ixa_extra_ident = 0;
2215 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
2216 TCP_STAT(tcps, tcp_lso_times);
2217 TCP_STAT_UPDATE(tcps, tcp_lso_pkt_out, num_lso_seg);
2218 } else {
2219 /*
2220 * Make sure to clean up LSO information. Wherever a
2221 * new mp uses the prepended header room after dupb(),
2222 * lso_info_cleanup() should be called.
2223 */
2224 lso_info_cleanup(mp);
2225 tcp_send_data(tcp, mp);
2226 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
2227 }
2228 }
2229
2230 return (0);
2231 }
2232
2233 /*
2234 * Initiate closedown sequence on an active connection. (May be called as
2235 * writer.) Return value zero for OK return, non-zero for error return.
2236 */
2237 static int
2238 tcp_xmit_end(tcp_t *tcp)
2239 {
2240 mblk_t *mp;
2241 tcp_stack_t *tcps = tcp->tcp_tcps;
2242 iulp_t uinfo;
2243 ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
2244 conn_t *connp = tcp->tcp_connp;
2245
2246 if (tcp->tcp_state < TCPS_SYN_RCVD ||
2247 tcp->tcp_state > TCPS_CLOSE_WAIT) {
2248 /*
2249 * Invalid state, only states TCPS_SYN_RCVD,
2250 * TCPS_ESTABLISHED and TCPS_CLOSE_WAIT are valid
2251 */
2252 return (-1);
2253 }
2254
2255 tcp->tcp_fss = tcp->tcp_snxt + tcp->tcp_unsent;
2256 tcp->tcp_valid_bits |= TCP_FSS_VALID;
2257 /*
2258 * If there is nothing more unsent, send the FIN now.
2259 * Otherwise, it will go out with the last segment.
2260 */
2261 if (tcp->tcp_unsent == 0) {
2262 mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
2263 tcp->tcp_fss, B_FALSE, NULL, B_FALSE);
2264
2265 if (mp) {
2266 tcp_send_data(tcp, mp);
2267 } else {
2268 /*
2269 * Couldn't allocate msg. Pretend we got it out.
2270 * Wait for rexmit timeout.
2271 */
2272 tcp->tcp_snxt = tcp->tcp_fss + 1;
2273 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
2274 }
2275
2276 /*
2277 * If needed, update tcp_rexmit_snxt as tcp_snxt is
2278 * changed.
2279 */
2280 if (tcp->tcp_rexmit && tcp->tcp_rexmit_nxt == tcp->tcp_fss) {
2281 tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
2282 }
2283 } else {
2284 /*
2285 * If tcp->tcp_cork is set, then the data will not get sent,
2286 * so we have to check that and unset it first.
2287 */
2288 if (tcp->tcp_cork)
2289 tcp->tcp_cork = B_FALSE;
2290 tcp_wput_data(tcp, NULL, B_FALSE);
2291 }
2292
2293 /*
2294 * If TCP does not get enough samples of RTT or tcp_rtt_updates
2295 * is 0, don't update the cache.
2296 */
2297 if (tcps->tcps_rtt_updates == 0 ||
2298 tcp->tcp_rtt_update < tcps->tcps_rtt_updates)
2299 return (0);
2300
2301 /*
2302 * We do not have a good algorithm to update ssthresh at this time.
2303 * So don't do any update.
2304 */
2305 bzero(&uinfo, sizeof (uinfo));
2306 uinfo.iulp_rtt = NSEC2MSEC(tcp->tcp_rtt_sa);
2307 uinfo.iulp_rtt_sd = NSEC2MSEC(tcp->tcp_rtt_sd);
2308
2309 /*
2310 * Note that uinfo is kept for conn_faddr in the DCE. Could update even
2311 * if source routed but we don't.
2312 */
2313 if (connp->conn_ipversion == IPV4_VERSION) {
2314 if (connp->conn_faddr_v4 != tcp->tcp_ipha->ipha_dst) {
2315 return (0);
2316 }
2317 (void) dce_update_uinfo_v4(connp->conn_faddr_v4, &uinfo, ipst);
2318 } else {
2319 uint_t ifindex;
2320
2321 if (!(IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6,
2322 &tcp->tcp_ip6h->ip6_dst))) {
2323 return (0);
2324 }
2325 ifindex = 0;
2326 if (IN6_IS_ADDR_LINKSCOPE(&connp->conn_faddr_v6)) {
2327 ip_xmit_attr_t *ixa = connp->conn_ixa;
2328
2329 /*
2330 * If we are going to create a DCE we'd better have
2331 * an ifindex
2332 */
2333 if (ixa->ixa_nce != NULL) {
2334 ifindex = ixa->ixa_nce->nce_common->ncec_ill->
2335 ill_phyint->phyint_ifindex;
2336 } else {
2337 return (0);
2338 }
2339 }
2340
2341 (void) dce_update_uinfo(&connp->conn_faddr_v6, ifindex, &uinfo,
2342 ipst);
2343 }
2344 return (0);
2345 }
2346
2347 /*
2348 * Send out a control packet on the tcp connection specified. This routine
2349 * is typically called where we need a simple ACK or RST generated.
2350 */
2351 void
2352 tcp_xmit_ctl(char *str, tcp_t *tcp, uint32_t seq, uint32_t ack, int ctl)
2353 {
2354 uchar_t *rptr;
2355 tcpha_t *tcpha;
2356 ipha_t *ipha = NULL;
2357 ip6_t *ip6h = NULL;
2358 uint32_t sum;
2359 int total_hdr_len;
2360 int ip_hdr_len;
2361 mblk_t *mp;
2362 tcp_stack_t *tcps = tcp->tcp_tcps;
2363 conn_t *connp = tcp->tcp_connp;
2364 ip_xmit_attr_t *ixa = connp->conn_ixa;
2365
2366 /*
2367 * Save sum for use in source route later.
2368 */
2369 sum = connp->conn_ht_ulp_len + connp->conn_sum;
2370 total_hdr_len = connp->conn_ht_iphc_len;
2371 ip_hdr_len = ixa->ixa_ip_hdr_length;
2372
2373 /* If a text string is passed in with the request, pass it to strlog. */
2374 if (str != NULL && connp->conn_debug) {
2375 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE,
2376 "tcp_xmit_ctl: '%s', seq 0x%x, ack 0x%x, ctl 0x%x",
2377 str, seq, ack, ctl);
2378 }
2379 mp = allocb(connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra,
2380 BPRI_MED);
2381 if (mp == NULL) {
2382 return;
2383 }
2384 rptr = &mp->b_rptr[tcps->tcps_wroff_xtra];
2385 mp->b_rptr = rptr;
2386 mp->b_wptr = &rptr[total_hdr_len];
2387 bcopy(connp->conn_ht_iphc, rptr, total_hdr_len);
2388
2389 ixa->ixa_pktlen = total_hdr_len;
2390
2391 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2392 ipha = (ipha_t *)rptr;
2393 ipha->ipha_length = htons(total_hdr_len);
2394 } else {
2395 ip6h = (ip6_t *)rptr;
2396 ip6h->ip6_plen = htons(total_hdr_len - IPV6_HDR_LEN);
2397 }
2398 tcpha = (tcpha_t *)&rptr[ip_hdr_len];
2399 tcpha->tha_flags = (uint8_t)ctl;
2400 if (ctl & TH_RST) {
2401 TCPS_BUMP_MIB(tcps, tcpOutRsts);
2402 TCPS_BUMP_MIB(tcps, tcpOutControl);
2403 /*
2404 * Don't send TSopt w/ TH_RST packets per RFC 1323.
2405 */
2406 if (tcp->tcp_snd_ts_ok &&
2407 tcp->tcp_state > TCPS_SYN_SENT) {
2408 mp->b_wptr = &rptr[total_hdr_len - TCPOPT_REAL_TS_LEN];
2409 *(mp->b_wptr) = TCPOPT_EOL;
2410
2411 ixa->ixa_pktlen = total_hdr_len - TCPOPT_REAL_TS_LEN;
2412
2413 if (connp->conn_ipversion == IPV4_VERSION) {
2414 ipha->ipha_length = htons(total_hdr_len -
2415 TCPOPT_REAL_TS_LEN);
2416 } else {
2417 ip6h->ip6_plen = htons(total_hdr_len -
2418 IPV6_HDR_LEN - TCPOPT_REAL_TS_LEN);
2419 }
2420 tcpha->tha_offset_and_reserved -= (3 << 4);
2421 sum -= TCPOPT_REAL_TS_LEN;
2422 }
2423 }
2424 if (ctl & TH_ACK) {
2425 if (tcp->tcp_snd_ts_ok) {
2426 uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
2427
2428 U32_TO_BE32(llbolt,
2429 (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
2430 U32_TO_BE32(tcp->tcp_ts_recent,
2431 (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
2432 }
2433
2434 /* Update the latest receive window size in TCP header. */
2435 tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2436 /* Track what we sent to the peer */
2437 tcp->tcp_tcpha->tha_win = tcpha->tha_win;
2438 tcp->tcp_rack = ack;
2439 tcp->tcp_rack_cnt = 0;
2440 TCPS_BUMP_MIB(tcps, tcpOutAck);
2441 }
2442 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
2443 tcpha->tha_seq = htonl(seq);
2444 tcpha->tha_ack = htonl(ack);
2445 /*
2446 * Include the adjustment for a source route if any.
2447 */
2448 sum = (sum >> 16) + (sum & 0xFFFF);
2449 tcpha->tha_sum = htons(sum);
2450 tcp_send_data(tcp, mp);
2451 }
2452
2453 /*
2454 * Generate a reset based on an inbound packet, connp is set by caller
2455 * when RST is in response to an unexpected inbound packet for which
2456 * there is active tcp state in the system.
2457 *
2458 * IPSEC NOTE : Try to send the reply with the same protection as it came
2459 * in. We have the ip_recv_attr_t which is reversed to form the ip_xmit_attr_t.
2460 * That way the packet will go out at the same level of protection as it
2461 * came in with.
2462 */
2463 static void
2464 tcp_xmit_early_reset(char *str, mblk_t *mp, uint32_t seq, uint32_t ack, int ctl,
2465 ip_recv_attr_t *ira, ip_stack_t *ipst, conn_t *connp)
2466 {
2467 ipha_t *ipha = NULL;
2468 ip6_t *ip6h = NULL;
2469 ushort_t len;
2470 tcpha_t *tcpha;
2471 int i;
2472 ipaddr_t v4addr;
2473 in6_addr_t v6addr;
2474 netstack_t *ns = ipst->ips_netstack;
2475 tcp_stack_t *tcps = ns->netstack_tcp;
2476 ip_xmit_attr_t ixas, *ixa;
2477 uint_t ip_hdr_len = ira->ira_ip_hdr_length;
2478 boolean_t need_refrele = B_FALSE; /* ixa_refrele(ixa) */
2479 ushort_t port;
2480
2481 if (!tcp_send_rst_chk(tcps)) {
2482 TCP_STAT(tcps, tcp_rst_unsent);
2483 freemsg(mp);
2484 return;
2485 }
2486
2487 /*
2488 * If connp != NULL we use conn_ixa to keep IP_NEXTHOP and other
2489 * options from the listener. In that case the caller must ensure that
2490 * we are running on the listener = connp squeue.
2491 *
2492 * We get a safe copy of conn_ixa so we don't need to restore anything
2493 * we or ip_output_simple might change in the ixa.
2494 */
2495 if (connp != NULL) {
2496 ASSERT(connp->conn_on_sqp);
2497
2498 ixa = conn_get_ixa_exclusive(connp);
2499 if (ixa == NULL) {
2500 TCP_STAT(tcps, tcp_rst_unsent);
2501 freemsg(mp);
2502 return;
2503 }
2504 need_refrele = B_TRUE;
2505 } else {
2506 bzero(&ixas, sizeof (ixas));
2507 ixa = &ixas;
2508 /*
2509 * IXAF_VERIFY_SOURCE is overkill since we know the
2510 * packet was for us.
2511 */
2512 ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE;
2513 ixa->ixa_protocol = IPPROTO_TCP;
2514 ixa->ixa_zoneid = ira->ira_zoneid;
2515 ixa->ixa_ifindex = 0;
2516 ixa->ixa_ipst = ipst;
2517 ixa->ixa_cred = kcred;
2518 ixa->ixa_cpid = NOPID;
2519 }
2520
2521 if (str && tcps->tcps_dbg) {
2522 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE,
2523 "tcp_xmit_early_reset: '%s', seq 0x%x, ack 0x%x, "
2524 "flags 0x%x",
2525 str, seq, ack, ctl);
2526 }
2527 if (mp->b_datap->db_ref != 1) {
2528 mblk_t *mp1 = copyb(mp);
2529 freemsg(mp);
2530 mp = mp1;
2531 if (mp == NULL)
2532 goto done;
2533 } else if (mp->b_cont) {
2534 freemsg(mp->b_cont);
2535 mp->b_cont = NULL;
2536 DB_CKSUMFLAGS(mp) = 0;
2537 }
2538 /*
2539 * We skip reversing source route here.
2540 * (for now we replace all IP options with EOL)
2541 */
2542 if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2543 ipha = (ipha_t *)mp->b_rptr;
2544 for (i = IP_SIMPLE_HDR_LENGTH; i < (int)ip_hdr_len; i++)
2545 mp->b_rptr[i] = IPOPT_EOL;
2546 /*
2547 * Make sure that src address isn't flagrantly invalid.
2548 * Not all broadcast address checking for the src address
2549 * is possible, since we don't know the netmask of the src
2550 * addr. No check for destination address is done, since
2551 * IP will not pass up a packet with a broadcast dest
2552 * address to TCP. Similar checks are done below for IPv6.
2553 */
2554 if (ipha->ipha_src == 0 || ipha->ipha_src == INADDR_BROADCAST ||
2555 CLASSD(ipha->ipha_src)) {
2556 BUMP_MIB(&ipst->ips_ip_mib, ipIfStatsInDiscards);
2557 ip_drop_input("ipIfStatsInDiscards", mp, NULL);
2558 freemsg(mp);
2559 goto done;
2560 }
2561 } else {
2562 ip6h = (ip6_t *)mp->b_rptr;
2563
2564 if (IN6_IS_ADDR_UNSPECIFIED(&ip6h->ip6_src) ||
2565 IN6_IS_ADDR_MULTICAST(&ip6h->ip6_src)) {
2566 BUMP_MIB(&ipst->ips_ip6_mib, ipIfStatsInDiscards);
2567 ip_drop_input("ipIfStatsInDiscards", mp, NULL);
2568 freemsg(mp);
2569 goto done;
2570 }
2571
2572 /* Remove any extension headers assuming partial overlay */
2573 if (ip_hdr_len > IPV6_HDR_LEN) {
2574 uint8_t *to;
2575
2576 to = mp->b_rptr + ip_hdr_len - IPV6_HDR_LEN;
2577 ovbcopy(ip6h, to, IPV6_HDR_LEN);
2578 mp->b_rptr += ip_hdr_len - IPV6_HDR_LEN;
2579 ip_hdr_len = IPV6_HDR_LEN;
2580 ip6h = (ip6_t *)mp->b_rptr;
2581 ip6h->ip6_nxt = IPPROTO_TCP;
2582 }
2583 }
2584 tcpha = (tcpha_t *)&mp->b_rptr[ip_hdr_len];
2585 if (tcpha->tha_flags & TH_RST) {
2586 freemsg(mp);
2587 goto done;
2588 }
2589 tcpha->tha_offset_and_reserved = (5 << 4);
2590 len = ip_hdr_len + sizeof (tcpha_t);
2591 mp->b_wptr = &mp->b_rptr[len];
2592 if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2593 ipha->ipha_length = htons(len);
2594 /* Swap addresses */
2595 v4addr = ipha->ipha_src;
2596 ipha->ipha_src = ipha->ipha_dst;
2597 ipha->ipha_dst = v4addr;
2598 ipha->ipha_ident = 0;
2599 ipha->ipha_ttl = (uchar_t)tcps->tcps_ipv4_ttl;
2600 ixa->ixa_flags |= IXAF_IS_IPV4;
2601 ixa->ixa_ip_hdr_length = ip_hdr_len;
2602 } else {
2603 ip6h->ip6_plen = htons(len - IPV6_HDR_LEN);
2604 /* Swap addresses */
2605 v6addr = ip6h->ip6_src;
2606 ip6h->ip6_src = ip6h->ip6_dst;
2607 ip6h->ip6_dst = v6addr;
2608 ip6h->ip6_hops = (uchar_t)tcps->tcps_ipv6_hoplimit;
2609 ixa->ixa_flags &= ~IXAF_IS_IPV4;
2610
2611 if (IN6_IS_ADDR_LINKSCOPE(&ip6h->ip6_dst)) {
2612 ixa->ixa_flags |= IXAF_SCOPEID_SET;
2613 ixa->ixa_scopeid = ira->ira_ruifindex;
2614 }
2615 ixa->ixa_ip_hdr_length = IPV6_HDR_LEN;
2616 }
2617 ixa->ixa_pktlen = len;
2618
2619 /* Swap the ports */
2620 port = tcpha->tha_fport;
2621 tcpha->tha_fport = tcpha->tha_lport;
2622 tcpha->tha_lport = port;
2623
2624 tcpha->tha_ack = htonl(ack);
2625 tcpha->tha_seq = htonl(seq);
2626 tcpha->tha_win = 0;
2627 tcpha->tha_sum = htons(sizeof (tcpha_t));
2628 tcpha->tha_flags = (uint8_t)ctl;
2629 if (ctl & TH_RST) {
2630 if (ctl & TH_ACK) {
2631 /*
2632 * Probe connection rejection here.
2633 * tcp_xmit_listeners_reset() drops non-SYN segments
2634 * that do not specify TH_ACK in their flags without
2635 * calling this function. As a consequence, if this
2636 * function is called with a TH_RST|TH_ACK ctl argument,
2637 * it is being called in response to a SYN segment
2638 * and thus the tcp:::accept-refused probe point
2639 * is valid here.
2640 */
2641 DTRACE_TCP5(accept__refused, mblk_t *, NULL,
2642 void, NULL, void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2643 tcph_t *, tcpha);
2644 }
2645 TCPS_BUMP_MIB(tcps, tcpOutRsts);
2646 TCPS_BUMP_MIB(tcps, tcpOutControl);
2647 }
2648
2649 /* Discard any old label */
2650 if (ixa->ixa_free_flags & IXA_FREE_TSL) {
2651 ASSERT(ixa->ixa_tsl != NULL);
2652 label_rele(ixa->ixa_tsl);
2653 ixa->ixa_free_flags &= ~IXA_FREE_TSL;
2654 }
2655 ixa->ixa_tsl = ira->ira_tsl; /* Behave as a multi-level responder */
2656
2657 if (ira->ira_flags & IRAF_IPSEC_SECURE) {
2658 /*
2659 * Apply IPsec based on how IPsec was applied to
2660 * the packet that caused the RST.
2661 */
2662 if (!ipsec_in_to_out(ira, ixa, mp, ipha, ip6h)) {
2663 BUMP_MIB(&ipst->ips_ip_mib, ipIfStatsOutDiscards);
2664 /* Note: mp already consumed and ip_drop_packet done */
2665 goto done;
2666 }
2667 } else {
2668 /*
2669 * This is in clear. The RST message we are building
2670 * here should go out in clear, independent of our policy.
2671 */
2672 ixa->ixa_flags |= IXAF_NO_IPSEC;
2673 }
2674
2675 DTRACE_TCP5(send, mblk_t *, NULL, ip_xmit_attr_t *, ixa,
2676 __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2677 __dtrace_tcp_tcph_t *, tcpha);
2678
2679 /*
2680 * NOTE: one might consider tracing a TCP packet here, but
2681 * this function has no active TCP state and no tcp structure
2682 * that has a trace buffer. If we traced here, we would have
2683 * to keep a local trace buffer in tcp_record_trace().
2684 */
2685
2686 (void) ip_output_simple(mp, ixa);
2687 done:
2688 ixa_cleanup(ixa);
2689 if (need_refrele) {
2690 ASSERT(ixa != &ixas);
2691 ixa_refrele(ixa);
2692 }
2693 }
2694
2695 /*
2696 * Generate a "no listener here" RST in response to an "unknown" segment.
2697 * connp is set by caller when RST is in response to an unexpected
2698 * inbound packet for which there is active tcp state in the system.
2699 * Note that we are reusing the incoming mp to construct the outgoing RST.
2700 */
2701 void
2702 tcp_xmit_listeners_reset(mblk_t *mp, ip_recv_attr_t *ira, ip_stack_t *ipst,
2703 conn_t *connp)
2704 {
2705 uchar_t *rptr;
2706 uint32_t seg_len;
2707 tcpha_t *tcpha;
2708 uint32_t seg_seq;
2709 uint32_t seg_ack;
2710 uint_t flags;
2711 ipha_t *ipha;
2712 ip6_t *ip6h;
2713 boolean_t policy_present;
2714 netstack_t *ns = ipst->ips_netstack;
2715 tcp_stack_t *tcps = ns->netstack_tcp;
2716 ipsec_stack_t *ipss = tcps->tcps_netstack->netstack_ipsec;
2717 uint_t ip_hdr_len = ira->ira_ip_hdr_length;
2718
2719 TCP_STAT(tcps, tcp_no_listener);
2720
2721 /*
2722 * DTrace this "unknown" segment as a tcp:::receive, as we did
2723 * just receive something that was TCP.
2724 */
2725 DTRACE_TCP5(receive, mblk_t *, NULL, ip_xmit_attr_t *, NULL,
2726 __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2727 __dtrace_tcp_tcph_t *, &mp->b_rptr[ip_hdr_len]);
2728
2729 if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2730 policy_present = ipss->ipsec_inbound_v4_policy_present;
2731 ipha = (ipha_t *)mp->b_rptr;
2732 ip6h = NULL;
2733 } else {
2734 policy_present = ipss->ipsec_inbound_v6_policy_present;
2735 ipha = NULL;
2736 ip6h = (ip6_t *)mp->b_rptr;
2737 }
2738
2739 if (policy_present) {
2740 /*
2741 * The conn_t parameter is NULL because we already know
2742 * nobody's home.
2743 */
2744 mp = ipsec_check_global_policy(mp, (conn_t *)NULL, ipha, ip6h,
2745 ira, ns);
2746 if (mp == NULL)
2747 return;
2748 }
2749 if (is_system_labeled() && !tsol_can_reply_error(mp, ira)) {
2750 DTRACE_PROBE2(
2751 tx__ip__log__error__nolistener__tcp,
2752 char *, "Could not reply with RST to mp(1)",
2753 mblk_t *, mp);
2754 ip2dbg(("tcp_xmit_listeners_reset: not permitted to reply\n"));
2755 freemsg(mp);
2756 return;
2757 }
2758
2759 rptr = mp->b_rptr;
2760
2761 tcpha = (tcpha_t *)&rptr[ip_hdr_len];
2762 seg_seq = ntohl(tcpha->tha_seq);
2763 seg_ack = ntohl(tcpha->tha_ack);
2764 flags = tcpha->tha_flags;
2765
2766 seg_len = msgdsize(mp) - (TCP_HDR_LENGTH(tcpha) + ip_hdr_len);
2767 if (flags & TH_RST) {
2768 freemsg(mp);
2769 } else if (flags & TH_ACK) {
2770 tcp_xmit_early_reset("no tcp, reset", mp, seg_ack, 0, TH_RST,
2771 ira, ipst, connp);
2772 } else {
2773 if (flags & TH_SYN) {
2774 seg_len++;
2775 } else {
2776 /*
2777 * Here we violate the RFC. Note that a normal
2778 * TCP will never send a segment without the ACK
2779 * flag, except for RST or SYN segment. This
2780 * segment is neither. Just drop it on the
2781 * floor.
2782 */
2783 freemsg(mp);
2784 TCP_STAT(tcps, tcp_rst_unsent);
2785 return;
2786 }
2787
2788 tcp_xmit_early_reset("no tcp, reset/ack", mp, 0,
2789 seg_seq + seg_len, TH_RST | TH_ACK, ira, ipst, connp);
2790 }
2791 }
2792
2793 /*
2794 * Helper function for tcp_xmit_mp() in handling connection set up flag
2795 * options setting.
2796 */
2797 static void
2798 tcp_xmit_mp_aux_iss(tcp_t *tcp, conn_t *connp, tcpha_t *tcpha, mblk_t *mp,
2799 uint_t *flags)
2800 {
2801 uint32_t u1;
2802 uint8_t *wptr = mp->b_wptr;
2803 tcp_stack_t *tcps = tcp->tcp_tcps;
2804 boolean_t add_sack = B_FALSE;
2805
2806 /*
2807 * If TCP_ISS_VALID and the seq number is tcp_iss,
2808 * TCP can only be in SYN-SENT, SYN-RCVD or
2809 * FIN-WAIT-1 state. It can be FIN-WAIT-1 if
2810 * our SYN is not ack'ed but the app closes this
2811 * TCP connection.
2812 */
2813 ASSERT(tcp->tcp_state == TCPS_SYN_SENT ||
2814 tcp->tcp_state == TCPS_SYN_RCVD ||
2815 tcp->tcp_state == TCPS_FIN_WAIT_1);
2816
2817 /*
2818 * Tack on the MSS option. It is always needed
2819 * for both active and passive open.
2820 *
2821 * MSS option value should be interface MTU - MIN
2822 * TCP/IP header according to RFC 793 as it means
2823 * the maximum segment size TCP can receive. But
2824 * to get around some broken middle boxes/end hosts
2825 * out there, we allow the option value to be the
2826 * same as the MSS option size on the peer side.
2827 * In this way, the other side will not send
2828 * anything larger than they can receive.
2829 *
2830 * Note that for SYN_SENT state, the ndd param
2831 * tcp_use_smss_as_mss_opt has no effect as we
2832 * don't know the peer's MSS option value. So
2833 * the only case we need to take care of is in
2834 * SYN_RCVD state, which is done later.
2835 */
2836 wptr[0] = TCPOPT_MAXSEG;
2837 wptr[1] = TCPOPT_MAXSEG_LEN;
2838 wptr += 2;
2839 u1 = tcp->tcp_initial_pmtu - (connp->conn_ipversion == IPV4_VERSION ?
2840 IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN) - TCP_MIN_HEADER_LENGTH;
2841 U16_TO_BE16(u1, wptr);
2842 wptr += 2;
2843
2844 /* Update the offset to cover the additional word */
2845 tcpha->tha_offset_and_reserved += (1 << 4);
2846
2847 switch (tcp->tcp_state) {
2848 case TCPS_SYN_SENT:
2849 *flags = TH_SYN;
2850
2851 if (tcp->tcp_snd_sack_ok)
2852 add_sack = B_TRUE;
2853
2854 if (tcp->tcp_snd_ts_ok) {
2855 uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
2856
2857 if (add_sack) {
2858 wptr[0] = TCPOPT_SACK_PERMITTED;
2859 wptr[1] = TCPOPT_SACK_OK_LEN;
2860 add_sack = B_FALSE;
2861 } else {
2862 wptr[0] = TCPOPT_NOP;
2863 wptr[1] = TCPOPT_NOP;
2864 }
2865 wptr[2] = TCPOPT_TSTAMP;
2866 wptr[3] = TCPOPT_TSTAMP_LEN;
2867 wptr += 4;
2868 U32_TO_BE32(llbolt, wptr);
2869 wptr += 4;
2870 ASSERT(tcp->tcp_ts_recent == 0);
2871 U32_TO_BE32(0L, wptr);
2872 wptr += 4;
2873 tcpha->tha_offset_and_reserved += (3 << 4);
2874 }
2875
2876 /*
2877 * Set up all the bits to tell other side
2878 * we are ECN capable.
2879 */
2880 if (tcp->tcp_ecn_ok)
2881 *flags |= (TH_ECE | TH_CWR);
2882
2883 break;
2884
2885 case TCPS_SYN_RCVD:
2886 *flags |= TH_SYN;
2887
2888 /*
2889 * Reset the MSS option value to be SMSS
2890 * We should probably add back the bytes
2891 * for timestamp option and IPsec. We
2892 * don't do that as this is a workaround
2893 * for broken middle boxes/end hosts, it
2894 * is better for us to be more cautious.
2895 * They may not take these things into
2896 * account in their SMSS calculation. Thus
2897 * the peer's calculated SMSS may be smaller
2898 * than what it can be. This should be OK.
2899 */
2900 if (tcps->tcps_use_smss_as_mss_opt) {
2901 u1 = tcp->tcp_mss;
2902 /*
2903 * Note that wptr points just past the MSS
2904 * option value.
2905 */
2906 U16_TO_BE16(u1, wptr - 2);
2907 }
2908
2909 /*
2910 * tcp_snd_ts_ok can only be set in TCPS_SYN_RCVD
2911 * when the peer also uses timestamps option. And
2912 * the TCP header template must have already been
2913 * updated to include the timestamps option.
2914 */
2915 if (tcp->tcp_snd_sack_ok) {
2916 if (tcp->tcp_snd_ts_ok) {
2917 uint8_t *tmp_wptr;
2918
2919 /*
2920 * Use the NOP in the header just
2921 * before timestamps opton.
2922 */
2923 tmp_wptr = (uint8_t *)tcpha +
2924 TCP_MIN_HEADER_LENGTH;
2925 ASSERT(tmp_wptr[0] == TCPOPT_NOP &&
2926 tmp_wptr[1] == TCPOPT_NOP);
2927 tmp_wptr[0] = TCPOPT_SACK_PERMITTED;
2928 tmp_wptr[1] = TCPOPT_SACK_OK_LEN;
2929 } else {
2930 add_sack = B_TRUE;
2931 }
2932 }
2933
2934
2935 /*
2936 * If the other side is ECN capable, reply
2937 * that we are also ECN capable.
2938 */
2939 if (tcp->tcp_ecn_ok)
2940 *flags |= TH_ECE;
2941 break;
2942
2943 default:
2944 /*
2945 * The above ASSERT() makes sure that this
2946 * must be FIN-WAIT-1 state. Our SYN has
2947 * not been ack'ed so retransmit it.
2948 */
2949 *flags |= TH_SYN;
2950 break;
2951 }
2952
2953 if (add_sack) {
2954 wptr[0] = TCPOPT_NOP;
2955 wptr[1] = TCPOPT_NOP;
2956 wptr[2] = TCPOPT_SACK_PERMITTED;
2957 wptr[3] = TCPOPT_SACK_OK_LEN;
2958 wptr += TCPOPT_REAL_SACK_OK_LEN;
2959 tcpha->tha_offset_and_reserved += (1 << 4);
2960 }
2961
2962 if (tcp->tcp_snd_ws_ok) {
2963 wptr[0] = TCPOPT_NOP;
2964 wptr[1] = TCPOPT_WSCALE;
2965 wptr[2] = TCPOPT_WS_LEN;
2966 wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
2967 wptr += TCPOPT_REAL_WS_LEN;
2968 tcpha->tha_offset_and_reserved += (1 << 4);
2969 }
2970
2971 mp->b_wptr = wptr;
2972 u1 = (int)(mp->b_wptr - mp->b_rptr);
2973 /*
2974 * Get IP set to checksum on our behalf
2975 * Include the adjustment for a source route if any.
2976 */
2977 u1 += connp->conn_sum;
2978 u1 = (u1 >> 16) + (u1 & 0xFFFF);
2979 tcpha->tha_sum = htons(u1);
2980 TCPS_BUMP_MIB(tcps, tcpOutControl);
2981 }
2982
2983 /*
2984 * Helper function for tcp_xmit_mp() in handling connection tear down
2985 * flag setting and state changes.
2986 */
2987 static void
2988 tcp_xmit_mp_aux_fss(tcp_t *tcp, ip_xmit_attr_t *ixa, uint_t *flags)
2989 {
2990 if (!tcp->tcp_fin_acked) {
2991 *flags |= TH_FIN;
2992 TCPS_BUMP_MIB(tcp->tcp_tcps, tcpOutControl);
2993 }
2994 if (!tcp->tcp_fin_sent) {
2995 tcp->tcp_fin_sent = B_TRUE;
2996 switch (tcp->tcp_state) {
2997 case TCPS_SYN_RCVD:
2998 tcp->tcp_state = TCPS_FIN_WAIT_1;
2999 DTRACE_TCP6(state__change, void, NULL,
3000 ip_xmit_attr_t *, ixa, void, NULL,
3001 tcp_t *, tcp, void, NULL,
3002 int32_t, TCPS_SYN_RCVD);
3003 break;
3004 case TCPS_ESTABLISHED:
3005 tcp->tcp_state = TCPS_FIN_WAIT_1;
3006 DTRACE_TCP6(state__change, void, NULL,
3007 ip_xmit_attr_t *, ixa, void, NULL,
3008 tcp_t *, tcp, void, NULL,
3009 int32_t, TCPS_ESTABLISHED);
3010 break;
3011 case TCPS_CLOSE_WAIT:
3012 tcp->tcp_state = TCPS_LAST_ACK;
3013 DTRACE_TCP6(state__change, void, NULL,
3014 ip_xmit_attr_t *, ixa, void, NULL,
3015 tcp_t *, tcp, void, NULL,
3016 int32_t, TCPS_CLOSE_WAIT);
3017 break;
3018 }
3019 if (tcp->tcp_suna == tcp->tcp_snxt)
3020 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3021 tcp->tcp_snxt = tcp->tcp_fss + 1;
3022 }
3023 }
3024
3025 /*
3026 * tcp_xmit_mp is called to return a pointer to an mblk chain complete with
3027 * ip and tcp header ready to pass down to IP. If the mp passed in is
3028 * non-NULL, then up to max_to_send bytes of data will be dup'ed off that
3029 * mblk. (If sendall is not set the dup'ing will stop at an mblk boundary
3030 * otherwise it will dup partial mblks.)
3031 * Otherwise, an appropriate ACK packet will be generated. This
3032 * routine is not usually called to send new data for the first time. It
3033 * is mostly called out of the timer for retransmits, and to generate ACKs.
3034 *
3035 * If offset is not NULL, the returned mblk chain's first mblk's b_rptr will
3036 * be adjusted by *offset. And after dupb(), the offset and the ending mblk
3037 * of the original mblk chain will be returned in *offset and *end_mp.
3038 */
3039 mblk_t *
3040 tcp_xmit_mp(tcp_t *tcp, mblk_t *mp, int32_t max_to_send, int32_t *offset,
3041 mblk_t **end_mp, uint32_t seq, boolean_t sendall, uint32_t *seg_len,
3042 boolean_t rexmit)
3043 {
3044 int data_length;
3045 int32_t off = 0;
3046 uint_t flags;
3047 mblk_t *mp1;
3048 mblk_t *mp2;
3049 uchar_t *rptr;
3050 tcpha_t *tcpha;
3051 int32_t num_sack_blk = 0;
3052 int32_t sack_opt_len = 0;
3053 tcp_stack_t *tcps = tcp->tcp_tcps;
3054 conn_t *connp = tcp->tcp_connp;
3055 ip_xmit_attr_t *ixa = connp->conn_ixa;
3056
3057 /* Allocate for our maximum TCP header + link-level */
3058 mp1 = allocb(connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra,
3059 BPRI_MED);
3060 if (mp1 == NULL)
3061 return (NULL);
3062 data_length = 0;
3063
3064 /*
3065 * Note that tcp_mss has been adjusted to take into account the
3066 * timestamp option if applicable. Because SACK options do not
3067 * appear in every TCP segments and they are of variable lengths,
3068 * they cannot be included in tcp_mss. Thus we need to calculate
3069 * the actual segment length when we need to send a segment which
3070 * includes SACK options.
3071 */
3072 if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
3073 num_sack_blk = MIN(tcp->tcp_max_sack_blk,
3074 tcp->tcp_num_sack_blk);
3075 sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
3076 TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
3077 if (max_to_send + sack_opt_len > tcp->tcp_mss)
3078 max_to_send -= sack_opt_len;
3079 }
3080
3081 if (offset != NULL) {
3082 off = *offset;
3083 /* We use offset as an indicator that end_mp is not NULL. */
3084 *end_mp = NULL;
3085 }
3086 for (mp2 = mp1; mp && data_length != max_to_send; mp = mp->b_cont) {
3087 /* This could be faster with cooperation from downstream */
3088 if (mp2 != mp1 && !sendall &&
3089 data_length + (int)(mp->b_wptr - mp->b_rptr) >
3090 max_to_send)
3091 /*
3092 * Don't send the next mblk since the whole mblk
3093 * does not fit.
3094 */
3095 break;
3096 mp2->b_cont = dupb(mp);
3097 mp2 = mp2->b_cont;
3098 if (!mp2) {
3099 freemsg(mp1);
3100 return (NULL);
3101 }
3102 mp2->b_rptr += off;
3103 ASSERT((uintptr_t)(mp2->b_wptr - mp2->b_rptr) <=
3104 (uintptr_t)INT_MAX);
3105
3106 data_length += (int)(mp2->b_wptr - mp2->b_rptr);
3107 if (data_length > max_to_send) {
3108 mp2->b_wptr -= data_length - max_to_send;
3109 data_length = max_to_send;
3110 off = mp2->b_wptr - mp->b_rptr;
3111 break;
3112 } else {
3113 off = 0;
3114 }
3115 }
3116 if (offset != NULL) {
3117 *offset = off;
3118 *end_mp = mp;
3119 }
3120 if (seg_len != NULL) {
3121 *seg_len = data_length;
3122 }
3123
3124 /* Update the latest receive window size in TCP header. */
3125 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3126
3127 rptr = mp1->b_rptr + tcps->tcps_wroff_xtra;
3128 mp1->b_rptr = rptr;
3129 mp1->b_wptr = rptr + connp->conn_ht_iphc_len + sack_opt_len;
3130 bcopy(connp->conn_ht_iphc, rptr, connp->conn_ht_iphc_len);
3131 tcpha = (tcpha_t *)&rptr[ixa->ixa_ip_hdr_length];
3132 tcpha->tha_seq = htonl(seq);
3133
3134 /*
3135 * Use tcp_unsent to determine if the PUSH bit should be used assumes
3136 * that this function was called from tcp_wput_data. Thus, when called
3137 * to retransmit data the setting of the PUSH bit may appear some
3138 * what random in that it might get set when it should not. This
3139 * should not pose any performance issues.
3140 */
3141 if (data_length != 0 && (tcp->tcp_unsent == 0 ||
3142 tcp->tcp_unsent == data_length)) {
3143 flags = TH_ACK | TH_PUSH;
3144 } else {
3145 flags = TH_ACK;
3146 }
3147
3148 if (tcp->tcp_ecn_ok) {
3149 if (tcp->tcp_ecn_echo_on)
3150 flags |= TH_ECE;
3151
3152 /*
3153 * Only set ECT bit and ECN_CWR if a segment contains new data.
3154 * There is no TCP flow control for non-data segments, and
3155 * only data segment is transmitted reliably.
3156 */
3157 if (data_length > 0 && !rexmit) {
3158 TCP_SET_ECT(tcp, rptr);
3159 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
3160 flags |= TH_CWR;
3161 tcp->tcp_ecn_cwr_sent = B_TRUE;
3162 }
3163 }
3164 }
3165
3166 /* Check if there is any special processing needs to be done. */
3167 if (tcp->tcp_valid_bits) {
3168 uint32_t u1;
3169
3170 /* We don't allow having SYN and FIN in the same segment... */
3171 if ((tcp->tcp_valid_bits & TCP_ISS_VALID) &&
3172 seq == tcp->tcp_iss) {
3173 /* Need to do connection set up processing. */
3174 tcp_xmit_mp_aux_iss(tcp, connp, tcpha, mp1, &flags);
3175 } else if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
3176 (seq + data_length) == tcp->tcp_fss) {
3177 /* Need to do connection tear down processing. */
3178 tcp_xmit_mp_aux_fss(tcp, ixa, &flags);
3179 }
3180
3181 /*
3182 * Need to do urgent pointer processing.
3183 *
3184 * Note the trick here. u1 is unsigned. When tcp_urg
3185 * is smaller than seq, u1 will become a very huge value.
3186 * So the comparison will fail. Also note that tcp_urp
3187 * should be positive, see RFC 793 page 17.
3188 */
3189 u1 = tcp->tcp_urg - seq + TCP_OLD_URP_INTERPRETATION;
3190 if ((tcp->tcp_valid_bits & TCP_URG_VALID) && u1 != 0 &&
3191 u1 < (uint32_t)(64 * 1024)) {
3192 flags |= TH_URG;
3193 TCPS_BUMP_MIB(tcps, tcpOutUrg);
3194 tcpha->tha_urp = htons(u1);
3195 }
3196 }
3197 tcpha->tha_flags = (uchar_t)flags;
3198 tcp->tcp_rack = tcp->tcp_rnxt;
3199 tcp->tcp_rack_cnt = 0;
3200
3201 /* Fill in the current value of timestamps option. */
3202 if (tcp->tcp_snd_ts_ok) {
3203 if (tcp->tcp_state != TCPS_SYN_SENT) {
3204 uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
3205
3206 U32_TO_BE32(llbolt,
3207 (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
3208 U32_TO_BE32(tcp->tcp_ts_recent,
3209 (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
3210 }
3211 }
3212
3213 /* Fill in the SACK blocks. */
3214 if (num_sack_blk > 0) {
3215 uchar_t *wptr = (uchar_t *)tcpha + connp->conn_ht_ulp_len;
3216 sack_blk_t *tmp;
3217 int32_t i;
3218
3219 wptr[0] = TCPOPT_NOP;
3220 wptr[1] = TCPOPT_NOP;
3221 wptr[2] = TCPOPT_SACK;
3222 wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
3223 sizeof (sack_blk_t);
3224 wptr += TCPOPT_REAL_SACK_LEN;
3225
3226 tmp = tcp->tcp_sack_list;
3227 for (i = 0; i < num_sack_blk; i++) {
3228 U32_TO_BE32(tmp[i].begin, wptr);
3229 wptr += sizeof (tcp_seq);
3230 U32_TO_BE32(tmp[i].end, wptr);
3231 wptr += sizeof (tcp_seq);
3232 }
3233 tcpha->tha_offset_and_reserved += ((num_sack_blk * 2 + 1) << 4);
3234 }
3235 ASSERT((uintptr_t)(mp1->b_wptr - rptr) <= (uintptr_t)INT_MAX);
3236 data_length += (int)(mp1->b_wptr - rptr);
3237
3238 ixa->ixa_pktlen = data_length;
3239
3240 if (ixa->ixa_flags & IXAF_IS_IPV4) {
3241 ((ipha_t *)rptr)->ipha_length = htons(data_length);
3242 } else {
3243 ip6_t *ip6 = (ip6_t *)rptr;
3244
3245 ip6->ip6_plen = htons(data_length - IPV6_HDR_LEN);
3246 }
3247
3248 /*
3249 * Prime pump for IP
3250 * Include the adjustment for a source route if any.
3251 */
3252 data_length -= ixa->ixa_ip_hdr_length;
3253 data_length += connp->conn_sum;
3254 data_length = (data_length >> 16) + (data_length & 0xFFFF);
3255 tcpha->tha_sum = htons(data_length);
3256 if (tcp->tcp_ip_forward_progress) {
3257 tcp->tcp_ip_forward_progress = B_FALSE;
3258 connp->conn_ixa->ixa_flags |= IXAF_REACH_CONF;
3259 } else {
3260 connp->conn_ixa->ixa_flags &= ~IXAF_REACH_CONF;
3261 }
3262 return (mp1);
3263 }
3264
3265 /*
3266 * If this routine returns B_TRUE, TCP can generate a RST in response
3267 * to a segment. If it returns B_FALSE, TCP should not respond.
3268 */
3269 static boolean_t
3270 tcp_send_rst_chk(tcp_stack_t *tcps)
3271 {
3272 int64_t now;
3273
3274 /*
3275 * TCP needs to protect itself from generating too many RSTs.
3276 * This can be a DoS attack by sending us random segments
3277 * soliciting RSTs.
3278 *
3279 * What we do here is to have a limit of tcp_rst_sent_rate RSTs
3280 * in each 1 second interval. In this way, TCP still generate
3281 * RSTs in normal cases but when under attack, the impact is
3282 * limited.
3283 */
3284 if (tcps->tcps_rst_sent_rate_enabled != 0) {
3285 now = ddi_get_lbolt64();
3286 if (TICK_TO_MSEC(now - tcps->tcps_last_rst_intrvl) >
3287 1*SECONDS) {
3288 tcps->tcps_last_rst_intrvl = now;
3289 tcps->tcps_rst_cnt = 1;
3290 } else if (++tcps->tcps_rst_cnt > tcps->tcps_rst_sent_rate) {
3291 return (B_FALSE);
3292 }
3293 }
3294 return (B_TRUE);
3295 }
3296
3297 /*
3298 * This function handles all retransmissions if SACK is enabled for this
3299 * connection. First it calculates how many segments can be retransmitted
3300 * based on tcp_pipe. Then it goes thru the notsack list to find eligible
3301 * segments. A segment is eligible if sack_cnt for that segment is greater
3302 * than or equal tcp_dupack_fast_retransmit. After it has retransmitted
3303 * all eligible segments, it checks to see if TCP can send some new segments
3304 * (fast recovery). If it can, set the appropriate flag for tcp_input_data().
3305 *
3306 * Parameters:
3307 * tcp_t *tcp: the tcp structure of the connection.
3308 * uint_t *flags: in return, appropriate value will be set for
3309 * tcp_input_data().
3310 */
3311 void
3312 tcp_sack_rexmit(tcp_t *tcp, uint_t *flags)
3313 {
3314 notsack_blk_t *notsack_blk;
3315 int32_t usable_swnd;
3316 int32_t mss;
3317 uint32_t seg_len;
3318 mblk_t *xmit_mp;
3319 tcp_stack_t *tcps = tcp->tcp_tcps;
3320
3321 ASSERT(tcp->tcp_notsack_list != NULL);
3322 ASSERT(tcp->tcp_rexmit == B_FALSE);
3323
3324 /* Defensive coding in case there is a bug... */
3325 if (tcp->tcp_notsack_list == NULL) {
3326 return;
3327 }
3328 notsack_blk = tcp->tcp_notsack_list;
3329 mss = tcp->tcp_mss;
3330
3331 /*
3332 * Limit the num of outstanding data in the network to be
3333 * tcp_cwnd_ssthresh, which is half of the original congestion wnd.
3334 */
3335 usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
3336
3337 /* At least retransmit 1 MSS of data. */
3338 if (usable_swnd <= 0) {
3339 usable_swnd = mss;
3340 }
3341
3342 /* Make sure no new RTT samples will be taken. */
3343 tcp->tcp_csuna = tcp->tcp_snxt;
3344
3345 notsack_blk = tcp->tcp_notsack_list;
3346 while (usable_swnd > 0) {
3347 mblk_t *snxt_mp, *tmp_mp;
3348 tcp_seq begin = tcp->tcp_sack_snxt;
3349 tcp_seq end;
3350 int32_t off;
3351
3352 for (; notsack_blk != NULL; notsack_blk = notsack_blk->next) {
3353 if (SEQ_GT(notsack_blk->end, begin) &&
3354 (notsack_blk->sack_cnt >=
3355 tcps->tcps_dupack_fast_retransmit)) {
3356 end = notsack_blk->end;
3357 if (SEQ_LT(begin, notsack_blk->begin)) {
3358 begin = notsack_blk->begin;
3359 }
3360 break;
3361 }
3362 }
3363 /*
3364 * All holes are filled. Manipulate tcp_cwnd to send more
3365 * if we can. Note that after the SACK recovery, tcp_cwnd is
3366 * set to tcp_cwnd_ssthresh.
3367 */
3368 if (notsack_blk == NULL) {
3369 usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
3370 if (usable_swnd <= 0 || tcp->tcp_unsent == 0) {
3371 tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna;
3372 ASSERT(tcp->tcp_cwnd > 0);
3373 return;
3374 } else {
3375 usable_swnd = usable_swnd / mss;
3376 tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna +
3377 MAX(usable_swnd * mss, mss);
3378 *flags |= TH_XMIT_NEEDED;
3379 return;
3380 }
3381 }
3382
3383 /*
3384 * Note that we may send more than usable_swnd allows here
3385 * because of round off, but no more than 1 MSS of data.
3386 */
3387 seg_len = end - begin;
3388 if (seg_len > mss)
3389 seg_len = mss;
3390 snxt_mp = tcp_get_seg_mp(tcp, begin, &off);
3391 ASSERT(snxt_mp != NULL);
3392 /* This should not happen. Defensive coding again... */
3393 if (snxt_mp == NULL) {
3394 return;
3395 }
3396
3397 xmit_mp = tcp_xmit_mp(tcp, snxt_mp, seg_len, &off,
3398 &tmp_mp, begin, B_TRUE, &seg_len, B_TRUE);
3399 if (xmit_mp == NULL)
3400 return;
3401
3402 usable_swnd -= seg_len;
3403 tcp->tcp_pipe += seg_len;
3404 tcp->tcp_sack_snxt = begin + seg_len;
3405
3406 tcp_send_data(tcp, xmit_mp);
3407
3408 /*
3409 * Update the send timestamp to avoid false retransmission.
3410 */
3411 snxt_mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
3412
3413 TCPS_BUMP_MIB(tcps, tcpRetransSegs);
3414 TCPS_UPDATE_MIB(tcps, tcpRetransBytes, seg_len);
3415 TCPS_BUMP_MIB(tcps, tcpOutSackRetransSegs);
3416 tcp->tcp_cs.tcp_out_retrans_segs++;
3417 tcp->tcp_cs.tcp_out_retrans_bytes += seg_len;
3418 /*
3419 * Update tcp_rexmit_max to extend this SACK recovery phase.
3420 * This happens when new data sent during fast recovery is
3421 * also lost. If TCP retransmits those new data, it needs
3422 * to extend SACK recover phase to avoid starting another
3423 * fast retransmit/recovery unnecessarily.
3424 */
3425 if (SEQ_GT(tcp->tcp_sack_snxt, tcp->tcp_rexmit_max)) {
3426 tcp->tcp_rexmit_max = tcp->tcp_sack_snxt;
3427 }
3428 }
3429 }
3430
3431 /*
3432 * tcp_ss_rexmit() is called to do slow start retransmission after a timeout
3433 * or ICMP errors.
3434 */
3435 void
3436 tcp_ss_rexmit(tcp_t *tcp)
3437 {
3438 uint32_t snxt;
3439 uint32_t smax;
3440 int32_t win;
3441 int32_t mss;
3442 int32_t off;
3443 mblk_t *snxt_mp;
3444 tcp_stack_t *tcps = tcp->tcp_tcps;
3445
3446 /*
3447 * Note that tcp_rexmit can be set even though TCP has retransmitted
3448 * all unack'ed segments.
3449 */
3450 if (SEQ_LT(tcp->tcp_rexmit_nxt, tcp->tcp_rexmit_max)) {
3451 smax = tcp->tcp_rexmit_max;
3452 snxt = tcp->tcp_rexmit_nxt;
3453 if (SEQ_LT(snxt, tcp->tcp_suna)) {
3454 snxt = tcp->tcp_suna;
3455 }
3456 win = MIN(tcp->tcp_cwnd, tcp->tcp_swnd);
3457 win -= snxt - tcp->tcp_suna;
3458 mss = tcp->tcp_mss;
3459 snxt_mp = tcp_get_seg_mp(tcp, snxt, &off);
3460
3461 while (SEQ_LT(snxt, smax) && (win > 0) && (snxt_mp != NULL)) {
3462 mblk_t *xmit_mp;
3463 mblk_t *old_snxt_mp = snxt_mp;
3464 uint32_t cnt = mss;
3465
3466 if (win < cnt) {
3467 cnt = win;
3468 }
3469 if (SEQ_GT(snxt + cnt, smax)) {
3470 cnt = smax - snxt;
3471 }
3472 xmit_mp = tcp_xmit_mp(tcp, snxt_mp, cnt, &off,
3473 &snxt_mp, snxt, B_TRUE, &cnt, B_TRUE);
3474 if (xmit_mp == NULL)
3475 return;
3476
3477 tcp_send_data(tcp, xmit_mp);
3478
3479 snxt += cnt;
3480 win -= cnt;
3481 /*
3482 * Update the send timestamp to avoid false
3483 * retransmission.
3484 */
3485 old_snxt_mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
3486 TCPS_BUMP_MIB(tcps, tcpRetransSegs);
3487 TCPS_UPDATE_MIB(tcps, tcpRetransBytes, cnt);
3488 tcp->tcp_cs.tcp_out_retrans_segs++;
3489 tcp->tcp_cs.tcp_out_retrans_bytes += cnt;
3490
3491 tcp->tcp_rexmit_nxt = snxt;
3492 }
3493 /*
3494 * If we have transmitted all we have at the time
3495 * we started the retranmission, we can leave
3496 * the rest of the job to tcp_wput_data(). But we
3497 * need to check the send window first. If the
3498 * win is not 0, go on with tcp_wput_data().
3499 */
3500 if (SEQ_LT(snxt, smax) || win == 0) {
3501 return;
3502 }
3503 }
3504 /* Only call tcp_wput_data() if there is data to be sent. */
3505 if (tcp->tcp_unsent) {
3506 tcp_wput_data(tcp, NULL, B_FALSE);
3507 }
3508 }
3509
3510 /*
3511 * Do slow start retransmission after ICMP errors of PMTU changes.
3512 */
3513 void
3514 tcp_rexmit_after_error(tcp_t *tcp)
3515 {
3516 /*
3517 * All sent data has been acknowledged or no data left to send, just
3518 * to return.
3519 */
3520 if (!SEQ_LT(tcp->tcp_suna, tcp->tcp_snxt) ||
3521 (tcp->tcp_xmit_head == NULL))
3522 return;
3523
3524 if ((tcp->tcp_valid_bits & TCP_FSS_VALID) && (tcp->tcp_unsent == 0))
3525 tcp->tcp_rexmit_max = tcp->tcp_fss;
3526 else
3527 tcp->tcp_rexmit_max = tcp->tcp_snxt;
3528
3529 tcp->tcp_rexmit_nxt = tcp->tcp_suna;
3530 tcp->tcp_rexmit = B_TRUE;
3531 tcp->tcp_dupack_cnt = 0;
3532 tcp_ss_rexmit(tcp);
3533 }
3534
3535 /*
3536 * tcp_get_seg_mp() is called to get the pointer to a segment in the
3537 * send queue which starts at the given sequence number. If the given
3538 * sequence number is equal to last valid sequence number (tcp_snxt), the
3539 * returned mblk is the last valid mblk, and off is set to the length of
3540 * that mblk.
3541 *
3542 * send queue which starts at the given seq. no.
3543 *
3544 * Parameters:
3545 * tcp_t *tcp: the tcp instance pointer.
3546 * uint32_t seq: the starting seq. no of the requested segment.
3547 * int32_t *off: after the execution, *off will be the offset to
3548 * the returned mblk which points to the requested seq no.
3549 * It is the caller's responsibility to send in a non-null off.
3550 *
3551 * Return:
3552 * A mblk_t pointer pointing to the requested segment in send queue.
3553 */
3554 static mblk_t *
3555 tcp_get_seg_mp(tcp_t *tcp, uint32_t seq, int32_t *off)
3556 {
3557 int32_t cnt;
3558 mblk_t *mp;
3559
3560 /* Defensive coding. Make sure we don't send incorrect data. */
3561 if (SEQ_LT(seq, tcp->tcp_suna) || SEQ_GT(seq, tcp->tcp_snxt))
3562 return (NULL);
3563
3564 cnt = seq - tcp->tcp_suna;
3565 mp = tcp->tcp_xmit_head;
3566 while (cnt > 0 && mp != NULL) {
3567 cnt -= mp->b_wptr - mp->b_rptr;
3568 if (cnt <= 0) {
3569 cnt += mp->b_wptr - mp->b_rptr;
3570 break;
3571 }
3572 mp = mp->b_cont;
3573 }
3574 ASSERT(mp != NULL);
3575 *off = cnt;
3576 return (mp);
3577 }
3578
3579 /*
3580 * This routine adjusts next-to-send sequence number variables, in the
3581 * case where the reciever has shrunk it's window.
3582 */
3583 void
3584 tcp_update_xmit_tail(tcp_t *tcp, uint32_t snxt)
3585 {
3586 mblk_t *xmit_tail;
3587 int32_t offset;
3588
3589 tcp->tcp_snxt = snxt;
3590
3591 /* Get the mblk, and the offset in it, as per the shrunk window */
3592 xmit_tail = tcp_get_seg_mp(tcp, snxt, &offset);
3593 ASSERT(xmit_tail != NULL);
3594 tcp->tcp_xmit_tail = xmit_tail;
3595 tcp->tcp_xmit_tail_unsent = xmit_tail->b_wptr -
3596 xmit_tail->b_rptr - offset;
3597 }
3598
3599 /*
3600 * This handles the case when the receiver has shrunk its win. Per RFC 1122
3601 * if the receiver shrinks the window, i.e. moves the right window to the
3602 * left, the we should not send new data, but should retransmit normally the
3603 * old unacked data between suna and suna + swnd. We might has sent data
3604 * that is now outside the new window, pretend that we didn't send it.
3605 */
3606 static void
3607 tcp_process_shrunk_swnd(tcp_t *tcp, uint32_t shrunk_count)
3608 {
3609 uint32_t snxt = tcp->tcp_snxt;
3610
3611 ASSERT(shrunk_count > 0);
3612
3613 if (!tcp->tcp_is_wnd_shrnk) {
3614 tcp->tcp_snxt_shrunk = snxt;
3615 tcp->tcp_is_wnd_shrnk = B_TRUE;
3616 } else if (SEQ_GT(snxt, tcp->tcp_snxt_shrunk)) {
3617 tcp->tcp_snxt_shrunk = snxt;
3618 }
3619
3620 /* Pretend we didn't send the data outside the window */
3621 snxt -= shrunk_count;
3622
3623 /* Reset all the values per the now shrunk window */
3624 tcp_update_xmit_tail(tcp, snxt);
3625 tcp->tcp_unsent += shrunk_count;
3626
3627 /*
3628 * If the SACK option is set, delete the entire list of
3629 * notsack'ed blocks.
3630 */
3631 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
3632
3633 if (tcp->tcp_suna == tcp->tcp_snxt && tcp->tcp_swnd == 0)
3634 /*
3635 * Make sure the timer is running so that we will probe a zero
3636 * window.
3637 */
3638 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3639 }
3640
3641 /*
3642 * tcp_fill_header is called by tcp_send() to fill the outgoing TCP header
3643 * with the template header, as well as other options such as time-stamp,
3644 * ECN and/or SACK.
3645 */
3646 static void
3647 tcp_fill_header(tcp_t *tcp, uchar_t *rptr, int num_sack_blk)
3648 {
3649 tcpha_t *tcp_tmpl, *tcpha;
3650 uint32_t *dst, *src;
3651 int hdrlen;
3652 conn_t *connp = tcp->tcp_connp;
3653
3654 ASSERT(OK_32PTR(rptr));
3655
3656 /* Template header */
3657 tcp_tmpl = tcp->tcp_tcpha;
3658
3659 /* Header of outgoing packet */
3660 tcpha = (tcpha_t *)(rptr + connp->conn_ixa->ixa_ip_hdr_length);
3661
3662 /* dst and src are opaque 32-bit fields, used for copying */
3663 dst = (uint32_t *)rptr;
3664 src = (uint32_t *)connp->conn_ht_iphc;
3665 hdrlen = connp->conn_ht_iphc_len;
3666
3667 /* Fill time-stamp option if needed */
3668 if (tcp->tcp_snd_ts_ok) {
3669 U32_TO_BE32(LBOLT_FASTPATH,
3670 (char *)tcp_tmpl + TCP_MIN_HEADER_LENGTH + 4);
3671 U32_TO_BE32(tcp->tcp_ts_recent,
3672 (char *)tcp_tmpl + TCP_MIN_HEADER_LENGTH + 8);
3673 } else {
3674 ASSERT(connp->conn_ht_ulp_len == TCP_MIN_HEADER_LENGTH);
3675 }
3676
3677 /*
3678 * Copy the template header; is this really more efficient than
3679 * calling bcopy()? For simple IPv4/TCP, it may be the case,
3680 * but perhaps not for other scenarios.
3681 */
3682 dst[0] = src[0];
3683 dst[1] = src[1];
3684 dst[2] = src[2];
3685 dst[3] = src[3];
3686 dst[4] = src[4];
3687 dst[5] = src[5];
3688 dst[6] = src[6];
3689 dst[7] = src[7];
3690 dst[8] = src[8];
3691 dst[9] = src[9];
3692 if (hdrlen -= 40) {
3693 hdrlen >>= 2;
3694 dst += 10;
3695 src += 10;
3696 do {
3697 *dst++ = *src++;
3698 } while (--hdrlen);
3699 }
3700
3701 /*
3702 * Set the ECN info in the TCP header if it is not a zero
3703 * window probe. Zero window probe is only sent in
3704 * tcp_wput_data() and tcp_timer().
3705 */
3706 if (tcp->tcp_ecn_ok && !tcp->tcp_zero_win_probe) {
3707 TCP_SET_ECT(tcp, rptr);
3708
3709 if (tcp->tcp_ecn_echo_on)
3710 tcpha->tha_flags |= TH_ECE;
3711 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
3712 tcpha->tha_flags |= TH_CWR;
3713 tcp->tcp_ecn_cwr_sent = B_TRUE;
3714 }
3715 }
3716
3717 /* Fill in SACK options */
3718 if (num_sack_blk > 0) {
3719 uchar_t *wptr = rptr + connp->conn_ht_iphc_len;
3720 sack_blk_t *tmp;
3721 int32_t i;
3722
3723 wptr[0] = TCPOPT_NOP;
3724 wptr[1] = TCPOPT_NOP;
3725 wptr[2] = TCPOPT_SACK;
3726 wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
3727 sizeof (sack_blk_t);
3728 wptr += TCPOPT_REAL_SACK_LEN;
3729
3730 tmp = tcp->tcp_sack_list;
3731 for (i = 0; i < num_sack_blk; i++) {
3732 U32_TO_BE32(tmp[i].begin, wptr);
3733 wptr += sizeof (tcp_seq);
3734 U32_TO_BE32(tmp[i].end, wptr);
3735 wptr += sizeof (tcp_seq);
3736 }
3737 tcpha->tha_offset_and_reserved +=
3738 ((num_sack_blk * 2 + 1) << 4);
3739 }
3740 }