Print this page
11546 Track TCP round-trip time in nanoseconds
Portions contributed by: Cody Peter Mello <cody.mello@joyent.com>
Portions contributed by: Brandon Baker <bbaker@delphix.com>
Reviewed by: Jason King <jason.king@joyent.com>
Reviewed by: Robert Mustacchi <rm@joyent.com>
Reviewed by: Dan McDonald <danmcd@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/inet/tcp/tcp_output.c
+++ new/usr/src/uts/common/inet/tcp/tcp_output.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24 - * Copyright (c) 2014 by Delphix. All rights reserved.
24 + * Copyright (c) 2014, 2016 by Delphix. All rights reserved.
25 + * Copyright 2019 Joyent, Inc.
25 26 */
26 27
27 28 /* This file contains all TCP output processing functions. */
28 29
29 30 #include <sys/types.h>
30 31 #include <sys/stream.h>
31 32 #include <sys/strsun.h>
32 33 #include <sys/strsubr.h>
33 34 #include <sys/stropts.h>
34 35 #include <sys/strlog.h>
35 36 #define _SUN_TPI_VERSION 2
36 37 #include <sys/tihdr.h>
37 38 #include <sys/suntpi.h>
38 39 #include <sys/xti_inet.h>
39 40 #include <sys/timod.h>
40 41 #include <sys/pattr.h>
41 42 #include <sys/squeue_impl.h>
42 43 #include <sys/squeue.h>
43 44 #include <sys/sockio.h>
44 45 #include <sys/tsol/tnet.h>
45 46
46 47 #include <inet/common.h>
47 48 #include <inet/ip.h>
48 49 #include <inet/tcp.h>
49 50 #include <inet/tcp_impl.h>
50 51 #include <inet/snmpcom.h>
|
↓ open down ↓ |
16 lines elided |
↑ open up ↑ |
51 52 #include <inet/proto_set.h>
52 53 #include <inet/ipsec_impl.h>
53 54 #include <inet/ip_ndp.h>
54 55
55 56 static mblk_t *tcp_get_seg_mp(tcp_t *, uint32_t, int32_t *);
56 57 static void tcp_wput_cmdblk(queue_t *, mblk_t *);
57 58 static void tcp_wput_flush(tcp_t *, mblk_t *);
58 59 static void tcp_wput_iocdata(tcp_t *tcp, mblk_t *mp);
59 60 static int tcp_xmit_end(tcp_t *);
60 61 static int tcp_send(tcp_t *, const int, const int, const int,
61 - const int, int *, uint_t *, int *, mblk_t **, mblk_t *);
62 + const int, int *, uint32_t *, int *, mblk_t **, mblk_t *);
62 63 static void tcp_xmit_early_reset(char *, mblk_t *, uint32_t, uint32_t,
63 64 int, ip_recv_attr_t *, ip_stack_t *, conn_t *);
64 65 static boolean_t tcp_send_rst_chk(tcp_stack_t *);
65 66 static void tcp_process_shrunk_swnd(tcp_t *, uint32_t);
66 -static void tcp_fill_header(tcp_t *, uchar_t *, clock_t, int);
67 +static void tcp_fill_header(tcp_t *, uchar_t *, int);
67 68
68 69 /*
69 70 * Functions called directly via squeue having a prototype of edesc_t.
70 71 */
71 72 static void tcp_wput_nondata(void *, mblk_t *, void *, ip_recv_attr_t *);
72 73 static void tcp_wput_ioctl(void *, mblk_t *, void *, ip_recv_attr_t *);
73 74 static void tcp_wput_proto(void *, mblk_t *, void *, ip_recv_attr_t *);
74 75
75 76 /*
76 77 * This controls how tiny a write must be before we try to copy it
77 78 * into the mblk on the tail of the transmit queue. Not much
78 79 * speedup is observed for values larger than sixteen. Zero will
79 80 * disable the optimisation.
80 81 */
81 82 static int tcp_tx_pull_len = 16;
82 83
83 84 int
84 85 tcp_wput(queue_t *q, mblk_t *mp)
85 86 {
86 87 conn_t *connp = Q_TO_CONN(q);
87 88 tcp_t *tcp;
88 89 void (*output_proc)();
89 90 t_scalar_t type;
90 91 uchar_t *rptr;
91 92 struct iocblk *iocp;
92 93 size_t size;
93 94
94 95 ASSERT(connp->conn_ref >= 2);
95 96
96 97 switch (DB_TYPE(mp)) {
97 98 case M_DATA:
98 99 tcp = connp->conn_tcp;
99 100 ASSERT(tcp != NULL);
100 101
101 102 size = msgdsize(mp);
102 103
103 104 mutex_enter(&tcp->tcp_non_sq_lock);
104 105 tcp->tcp_squeue_bytes += size;
105 106 if (TCP_UNSENT_BYTES(tcp) > connp->conn_sndbuf) {
106 107 tcp_setqfull(tcp);
107 108 }
108 109 mutex_exit(&tcp->tcp_non_sq_lock);
109 110
110 111 CONN_INC_REF(connp);
111 112 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_output, connp,
112 113 NULL, tcp_squeue_flag, SQTAG_TCP_OUTPUT);
113 114 return (0);
114 115
115 116 case M_CMD:
116 117 tcp_wput_cmdblk(q, mp);
117 118 return (0);
118 119
119 120 case M_PROTO:
120 121 case M_PCPROTO:
121 122 /*
122 123 * if it is a snmp message, don't get behind the squeue
123 124 */
124 125 tcp = connp->conn_tcp;
125 126 rptr = mp->b_rptr;
126 127 if ((mp->b_wptr - rptr) >= sizeof (t_scalar_t)) {
127 128 type = ((union T_primitives *)rptr)->type;
128 129 } else {
129 130 if (connp->conn_debug) {
130 131 (void) strlog(TCP_MOD_ID, 0, 1,
131 132 SL_ERROR|SL_TRACE,
132 133 "tcp_wput_proto, dropping one...");
133 134 }
134 135 freemsg(mp);
135 136 return (0);
136 137 }
137 138 if (type == T_SVR4_OPTMGMT_REQ) {
138 139 /*
139 140 * All Solaris components should pass a db_credp
140 141 * for this TPI message, hence we ASSERT.
141 142 * But in case there is some other M_PROTO that looks
142 143 * like a TPI message sent by some other kernel
143 144 * component, we check and return an error.
144 145 */
145 146 cred_t *cr = msg_getcred(mp, NULL);
146 147
147 148 ASSERT(cr != NULL);
148 149 if (cr == NULL) {
149 150 tcp_err_ack(tcp, mp, TSYSERR, EINVAL);
150 151 return (0);
151 152 }
152 153 if (snmpcom_req(q, mp, tcp_snmp_set, ip_snmp_get,
153 154 cr)) {
154 155 /*
155 156 * This was a SNMP request
156 157 */
157 158 return (0);
158 159 } else {
159 160 output_proc = tcp_wput_proto;
160 161 }
161 162 } else {
162 163 output_proc = tcp_wput_proto;
163 164 }
164 165 break;
165 166 case M_IOCTL:
166 167 /*
167 168 * Most ioctls can be processed right away without going via
168 169 * squeues - process them right here. Those that do require
169 170 * squeue (currently _SIOCSOCKFALLBACK)
170 171 * are processed by tcp_wput_ioctl().
171 172 */
172 173 iocp = (struct iocblk *)mp->b_rptr;
173 174 tcp = connp->conn_tcp;
174 175
175 176 switch (iocp->ioc_cmd) {
176 177 case TCP_IOC_ABORT_CONN:
177 178 tcp_ioctl_abort_conn(q, mp);
178 179 return (0);
179 180 case TI_GETPEERNAME:
180 181 case TI_GETMYNAME:
181 182 mi_copyin(q, mp, NULL,
182 183 SIZEOF_STRUCT(strbuf, iocp->ioc_flag));
183 184 return (0);
184 185
185 186 default:
186 187 output_proc = tcp_wput_ioctl;
187 188 break;
188 189 }
189 190 break;
190 191 default:
191 192 output_proc = tcp_wput_nondata;
192 193 break;
193 194 }
194 195
195 196 CONN_INC_REF(connp);
196 197 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, output_proc, connp,
197 198 NULL, tcp_squeue_flag, SQTAG_TCP_WPUT_OTHER);
198 199 return (0);
199 200 }
200 201
201 202 /*
202 203 * The TCP normal data output path.
203 204 * NOTE: the logic of the fast path is duplicated from this function.
204 205 */
205 206 void
206 207 tcp_wput_data(tcp_t *tcp, mblk_t *mp, boolean_t urgent)
207 208 {
208 209 int len;
209 210 mblk_t *local_time;
210 211 mblk_t *mp1;
211 212 uint32_t snxt;
212 213 int tail_unsent;
213 214 int tcpstate;
214 215 int usable = 0;
215 216 mblk_t *xmit_tail;
216 217 int32_t mss;
217 218 int32_t num_sack_blk = 0;
218 219 int32_t total_hdr_len;
219 220 int32_t tcp_hdr_len;
220 221 int rc;
221 222 tcp_stack_t *tcps = tcp->tcp_tcps;
222 223 conn_t *connp = tcp->tcp_connp;
223 224 clock_t now = LBOLT_FASTPATH;
224 225
225 226 tcpstate = tcp->tcp_state;
226 227 if (mp == NULL) {
227 228 /*
228 229 * tcp_wput_data() with NULL mp should only be called when
229 230 * there is unsent data.
230 231 */
231 232 ASSERT(tcp->tcp_unsent > 0);
232 233 /* Really tacky... but we need this for detached closes. */
233 234 len = tcp->tcp_unsent;
234 235 goto data_null;
235 236 }
236 237
237 238 ASSERT(mp->b_datap->db_type == M_DATA);
238 239 /*
239 240 * Don't allow data after T_ORDREL_REQ or T_DISCON_REQ,
240 241 * or before a connection attempt has begun.
241 242 */
242 243 if (tcpstate < TCPS_SYN_SENT || tcpstate > TCPS_CLOSE_WAIT ||
243 244 (tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
244 245 if ((tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
245 246 #ifdef DEBUG
246 247 cmn_err(CE_WARN,
247 248 "tcp_wput_data: data after ordrel, %s",
248 249 tcp_display(tcp, NULL,
249 250 DISP_ADDR_AND_PORT));
250 251 #else
251 252 if (connp->conn_debug) {
252 253 (void) strlog(TCP_MOD_ID, 0, 1,
253 254 SL_TRACE|SL_ERROR,
254 255 "tcp_wput_data: data after ordrel, %s\n",
255 256 tcp_display(tcp, NULL,
256 257 DISP_ADDR_AND_PORT));
257 258 }
258 259 #endif /* DEBUG */
259 260 }
260 261 if (tcp->tcp_snd_zcopy_aware &&
261 262 (mp->b_datap->db_struioflag & STRUIO_ZCNOTIFY))
262 263 tcp_zcopy_notify(tcp);
263 264 freemsg(mp);
264 265 mutex_enter(&tcp->tcp_non_sq_lock);
265 266 if (tcp->tcp_flow_stopped &&
266 267 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
267 268 tcp_clrqfull(tcp);
268 269 }
269 270 mutex_exit(&tcp->tcp_non_sq_lock);
270 271 return;
271 272 }
272 273
273 274 /* Strip empties */
274 275 for (;;) {
275 276 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
276 277 (uintptr_t)INT_MAX);
277 278 len = (int)(mp->b_wptr - mp->b_rptr);
278 279 if (len > 0)
279 280 break;
280 281 mp1 = mp;
281 282 mp = mp->b_cont;
282 283 freeb(mp1);
283 284 if (mp == NULL) {
284 285 return;
285 286 }
286 287 }
287 288
288 289 /* If we are the first on the list ... */
289 290 if (tcp->tcp_xmit_head == NULL) {
290 291 tcp->tcp_xmit_head = mp;
291 292 tcp->tcp_xmit_tail = mp;
292 293 tcp->tcp_xmit_tail_unsent = len;
293 294 } else {
294 295 /* If tiny tx and room in txq tail, pullup to save mblks. */
295 296 struct datab *dp;
296 297
297 298 mp1 = tcp->tcp_xmit_last;
298 299 if (len < tcp_tx_pull_len &&
299 300 (dp = mp1->b_datap)->db_ref == 1 &&
300 301 dp->db_lim - mp1->b_wptr >= len) {
301 302 ASSERT(len > 0);
302 303 ASSERT(!mp1->b_cont);
303 304 if (len == 1) {
304 305 *mp1->b_wptr++ = *mp->b_rptr;
305 306 } else {
306 307 bcopy(mp->b_rptr, mp1->b_wptr, len);
307 308 mp1->b_wptr += len;
308 309 }
309 310 if (mp1 == tcp->tcp_xmit_tail)
310 311 tcp->tcp_xmit_tail_unsent += len;
311 312 mp1->b_cont = mp->b_cont;
312 313 if (tcp->tcp_snd_zcopy_aware &&
313 314 (mp->b_datap->db_struioflag & STRUIO_ZCNOTIFY))
314 315 mp1->b_datap->db_struioflag |= STRUIO_ZCNOTIFY;
315 316 freeb(mp);
316 317 mp = mp1;
317 318 } else {
318 319 tcp->tcp_xmit_last->b_cont = mp;
319 320 }
320 321 len += tcp->tcp_unsent;
321 322 }
322 323
323 324 /* Tack on however many more positive length mblks we have */
324 325 if ((mp1 = mp->b_cont) != NULL) {
325 326 do {
326 327 int tlen;
327 328 ASSERT((uintptr_t)(mp1->b_wptr - mp1->b_rptr) <=
328 329 (uintptr_t)INT_MAX);
329 330 tlen = (int)(mp1->b_wptr - mp1->b_rptr);
330 331 if (tlen <= 0) {
331 332 mp->b_cont = mp1->b_cont;
332 333 freeb(mp1);
333 334 } else {
334 335 len += tlen;
335 336 mp = mp1;
336 337 }
337 338 } while ((mp1 = mp->b_cont) != NULL);
338 339 }
339 340 tcp->tcp_xmit_last = mp;
340 341 tcp->tcp_unsent = len;
341 342
342 343 if (urgent)
343 344 usable = 1;
344 345
345 346 data_null:
346 347 snxt = tcp->tcp_snxt;
347 348 xmit_tail = tcp->tcp_xmit_tail;
348 349 tail_unsent = tcp->tcp_xmit_tail_unsent;
349 350
350 351 /*
351 352 * Note that tcp_mss has been adjusted to take into account the
352 353 * timestamp option if applicable. Because SACK options do not
353 354 * appear in every TCP segments and they are of variable lengths,
354 355 * they cannot be included in tcp_mss. Thus we need to calculate
355 356 * the actual segment length when we need to send a segment which
356 357 * includes SACK options.
357 358 */
358 359 if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
359 360 int32_t opt_len;
360 361
361 362 num_sack_blk = MIN(tcp->tcp_max_sack_blk,
362 363 tcp->tcp_num_sack_blk);
363 364 opt_len = num_sack_blk * sizeof (sack_blk_t) + TCPOPT_NOP_LEN *
364 365 2 + TCPOPT_HEADER_LEN;
365 366 mss = tcp->tcp_mss - opt_len;
366 367 total_hdr_len = connp->conn_ht_iphc_len + opt_len;
367 368 tcp_hdr_len = connp->conn_ht_ulp_len + opt_len;
368 369 } else {
369 370 mss = tcp->tcp_mss;
370 371 total_hdr_len = connp->conn_ht_iphc_len;
371 372 tcp_hdr_len = connp->conn_ht_ulp_len;
372 373 }
373 374
374 375 if ((tcp->tcp_suna == snxt) && !tcp->tcp_localnet &&
375 376 (TICK_TO_MSEC(now - tcp->tcp_last_recv_time) >= tcp->tcp_rto)) {
376 377 TCP_SET_INIT_CWND(tcp, mss, tcps->tcps_slow_start_after_idle);
377 378 }
378 379 if (tcpstate == TCPS_SYN_RCVD) {
379 380 /*
380 381 * The three-way connection establishment handshake is not
381 382 * complete yet. We want to queue the data for transmission
382 383 * after entering ESTABLISHED state (RFC793). A jump to
383 384 * "done" label effectively leaves data on the queue.
384 385 */
385 386 goto done;
386 387 } else {
387 388 int usable_r;
388 389
389 390 /*
390 391 * In the special case when cwnd is zero, which can only
391 392 * happen if the connection is ECN capable, return now.
392 393 * New segments is sent using tcp_timer(). The timer
393 394 * is set in tcp_input_data().
394 395 */
395 396 if (tcp->tcp_cwnd == 0) {
396 397 /*
397 398 * Note that tcp_cwnd is 0 before 3-way handshake is
398 399 * finished.
399 400 */
400 401 ASSERT(tcp->tcp_ecn_ok ||
401 402 tcp->tcp_state < TCPS_ESTABLISHED);
402 403 return;
403 404 }
404 405
405 406 /* NOTE: trouble if xmitting while SYN not acked? */
406 407 usable_r = snxt - tcp->tcp_suna;
407 408 usable_r = tcp->tcp_swnd - usable_r;
408 409
409 410 /*
410 411 * Check if the receiver has shrunk the window. If
411 412 * tcp_wput_data() with NULL mp is called, tcp_fin_sent
412 413 * cannot be set as there is unsent data, so FIN cannot
413 414 * be sent out. Otherwise, we need to take into account
414 415 * of FIN as it consumes an "invisible" sequence number.
415 416 */
416 417 ASSERT(tcp->tcp_fin_sent == 0);
417 418 if (usable_r < 0) {
418 419 /*
419 420 * The receiver has shrunk the window and we have sent
420 421 * -usable_r date beyond the window, re-adjust.
421 422 *
422 423 * If TCP window scaling is enabled, there can be
423 424 * round down error as the advertised receive window
424 425 * is actually right shifted n bits. This means that
425 426 * the lower n bits info is wiped out. It will look
426 427 * like the window is shrunk. Do a check here to
427 428 * see if the shrunk amount is actually within the
428 429 * error in window calculation. If it is, just
429 430 * return. Note that this check is inside the
430 431 * shrunk window check. This makes sure that even
431 432 * though tcp_process_shrunk_swnd() is not called,
432 433 * we will stop further processing.
433 434 */
434 435 if ((-usable_r >> tcp->tcp_snd_ws) > 0) {
435 436 tcp_process_shrunk_swnd(tcp, -usable_r);
436 437 }
437 438 return;
438 439 }
439 440
440 441 /* usable = MIN(swnd, cwnd) - unacked_bytes */
441 442 if (tcp->tcp_swnd > tcp->tcp_cwnd)
442 443 usable_r -= tcp->tcp_swnd - tcp->tcp_cwnd;
443 444
444 445 /* usable = MIN(usable, unsent) */
445 446 if (usable_r > len)
446 447 usable_r = len;
|
↓ open down ↓ |
370 lines elided |
↑ open up ↑ |
447 448
448 449 /* usable = MAX(usable, {1 for urgent, 0 for data}) */
449 450 if (usable_r > 0) {
450 451 usable = usable_r;
451 452 } else {
452 453 /* Bypass all other unnecessary processing. */
453 454 goto done;
454 455 }
455 456 }
456 457
457 - local_time = (mblk_t *)now;
458 + local_time = (mblk_t *)(intptr_t)gethrtime();
458 459
459 460 /*
460 461 * "Our" Nagle Algorithm. This is not the same as in the old
461 462 * BSD. This is more in line with the true intent of Nagle.
462 463 *
463 464 * The conditions are:
464 465 * 1. The amount of unsent data (or amount of data which can be
465 466 * sent, whichever is smaller) is less than Nagle limit.
466 467 * 2. The last sent size is also less than Nagle limit.
467 468 * 3. There is unack'ed data.
468 469 * 4. Urgent pointer is not set. Send urgent data ignoring the
469 470 * Nagle algorithm. This reduces the probability that urgent
470 471 * bytes get "merged" together.
471 472 * 5. The app has not closed the connection. This eliminates the
472 473 * wait time of the receiving side waiting for the last piece of
473 474 * (small) data.
474 475 *
475 476 * If all are satisified, exit without sending anything. Note
476 477 * that Nagle limit can be smaller than 1 MSS. Nagle limit is
477 478 * the smaller of 1 MSS and global tcp_naglim_def (default to be
478 479 * 4095).
479 480 */
480 481 if (usable < (int)tcp->tcp_naglim &&
481 482 tcp->tcp_naglim > tcp->tcp_last_sent_len &&
482 483 snxt != tcp->tcp_suna &&
483 484 !(tcp->tcp_valid_bits & TCP_URG_VALID) &&
484 485 !(tcp->tcp_valid_bits & TCP_FSS_VALID)) {
485 486 goto done;
486 487 }
487 488
488 489 /*
489 490 * If tcp_zero_win_probe is not set and the tcp->tcp_cork option
490 491 * is set, then we have to force TCP not to send partial segment
491 492 * (smaller than MSS bytes). We are calculating the usable now
492 493 * based on full mss and will save the rest of remaining data for
493 494 * later. When tcp_zero_win_probe is set, TCP needs to send out
494 495 * something to do zero window probe.
495 496 */
496 497 if (tcp->tcp_cork && !tcp->tcp_zero_win_probe) {
497 498 if (usable < mss)
498 499 goto done;
499 500 usable = (usable / mss) * mss;
500 501 }
501 502
502 503 /* Update the latest receive window size in TCP header. */
503 504 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
504 505
505 506 /* Send the packet. */
506 507 rc = tcp_send(tcp, mss, total_hdr_len, tcp_hdr_len,
507 508 num_sack_blk, &usable, &snxt, &tail_unsent, &xmit_tail,
508 509 local_time);
509 510
510 511 /* Pretend that all we were trying to send really got sent */
511 512 if (rc < 0 && tail_unsent < 0) {
512 513 do {
513 514 xmit_tail = xmit_tail->b_cont;
514 515 xmit_tail->b_prev = local_time;
515 516 ASSERT((uintptr_t)(xmit_tail->b_wptr -
516 517 xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
517 518 tail_unsent += (int)(xmit_tail->b_wptr -
518 519 xmit_tail->b_rptr);
519 520 } while (tail_unsent < 0);
520 521 }
521 522 done:;
522 523 tcp->tcp_xmit_tail = xmit_tail;
523 524 tcp->tcp_xmit_tail_unsent = tail_unsent;
524 525 len = tcp->tcp_snxt - snxt;
525 526 if (len) {
526 527 /*
527 528 * If new data was sent, need to update the notsack
528 529 * list, which is, afterall, data blocks that have
529 530 * not been sack'ed by the receiver. New data is
530 531 * not sack'ed.
531 532 */
532 533 if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
533 534 /* len is a negative value. */
534 535 tcp->tcp_pipe -= len;
535 536 tcp_notsack_update(&(tcp->tcp_notsack_list),
536 537 tcp->tcp_snxt, snxt,
537 538 &(tcp->tcp_num_notsack_blk),
538 539 &(tcp->tcp_cnt_notsack_list));
539 540 }
540 541 tcp->tcp_snxt = snxt + tcp->tcp_fin_sent;
541 542 tcp->tcp_rack = tcp->tcp_rnxt;
542 543 tcp->tcp_rack_cnt = 0;
543 544 if ((snxt + len) == tcp->tcp_suna) {
544 545 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
545 546 }
546 547 } else if (snxt == tcp->tcp_suna && tcp->tcp_swnd == 0) {
547 548 /*
548 549 * Didn't send anything. Make sure the timer is running
549 550 * so that we will probe a zero window.
550 551 */
551 552 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
552 553 }
553 554 /* Note that len is the amount we just sent but with a negative sign */
554 555 tcp->tcp_unsent += len;
555 556 mutex_enter(&tcp->tcp_non_sq_lock);
556 557 if (tcp->tcp_flow_stopped) {
557 558 if (TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
558 559 tcp_clrqfull(tcp);
559 560 }
560 561 } else if (TCP_UNSENT_BYTES(tcp) >= connp->conn_sndbuf) {
561 562 if (!(tcp->tcp_detached))
562 563 tcp_setqfull(tcp);
563 564 }
564 565 mutex_exit(&tcp->tcp_non_sq_lock);
565 566 }
566 567
567 568 /*
568 569 * Initial STREAMS write side put() procedure for sockets. It tries to
569 570 * handle the T_CAPABILITY_REQ which sockfs sends down while setting
570 571 * up the socket without using the squeue. Non T_CAPABILITY_REQ messages
571 572 * are handled by tcp_wput() as usual.
572 573 *
573 574 * All further messages will also be handled by tcp_wput() because we cannot
574 575 * be sure that the above short cut is safe later.
575 576 */
576 577 int
577 578 tcp_wput_sock(queue_t *wq, mblk_t *mp)
578 579 {
579 580 conn_t *connp = Q_TO_CONN(wq);
580 581 tcp_t *tcp = connp->conn_tcp;
581 582 struct T_capability_req *car = (struct T_capability_req *)mp->b_rptr;
582 583
583 584 ASSERT(wq->q_qinfo == &tcp_sock_winit);
584 585 wq->q_qinfo = &tcp_winit;
585 586
586 587 ASSERT(IPCL_IS_TCP(connp));
587 588 ASSERT(TCP_IS_SOCKET(tcp));
588 589
589 590 if (DB_TYPE(mp) == M_PCPROTO &&
590 591 MBLKL(mp) == sizeof (struct T_capability_req) &&
591 592 car->PRIM_type == T_CAPABILITY_REQ) {
592 593 tcp_capability_req(tcp, mp);
593 594 return (0);
594 595 }
595 596
596 597 tcp_wput(wq, mp);
597 598 return (0);
598 599 }
599 600
600 601 /* ARGSUSED */
601 602 int
602 603 tcp_wput_fallback(queue_t *wq, mblk_t *mp)
603 604 {
604 605 #ifdef DEBUG
605 606 cmn_err(CE_CONT, "tcp_wput_fallback: Message during fallback \n");
606 607 #endif
607 608 freemsg(mp);
608 609 return (0);
609 610 }
610 611
611 612 /*
612 613 * Call by tcp_wput() to handle misc non M_DATA messages.
613 614 */
614 615 /* ARGSUSED */
615 616 static void
616 617 tcp_wput_nondata(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
617 618 {
618 619 conn_t *connp = (conn_t *)arg;
619 620 tcp_t *tcp = connp->conn_tcp;
620 621
621 622 ASSERT(DB_TYPE(mp) != M_IOCTL);
622 623 /*
623 624 * TCP is D_MP and qprocsoff() is done towards the end of the tcp_close.
624 625 * Once the close starts, streamhead and sockfs will not let any data
625 626 * packets come down (close ensures that there are no threads using the
626 627 * queue and no new threads will come down) but since qprocsoff()
627 628 * hasn't happened yet, a M_FLUSH or some non data message might
628 629 * get reflected back (in response to our own FLUSHRW) and get
629 630 * processed after tcp_close() is done. The conn would still be valid
630 631 * because a ref would have added but we need to check the state
631 632 * before actually processing the packet.
632 633 */
633 634 if (TCP_IS_DETACHED(tcp) || (tcp->tcp_state == TCPS_CLOSED)) {
634 635 freemsg(mp);
635 636 return;
636 637 }
637 638
638 639 switch (DB_TYPE(mp)) {
639 640 case M_IOCDATA:
640 641 tcp_wput_iocdata(tcp, mp);
641 642 break;
642 643 case M_FLUSH:
643 644 tcp_wput_flush(tcp, mp);
644 645 break;
645 646 default:
646 647 ip_wput_nondata(connp->conn_wq, mp);
647 648 break;
648 649 }
649 650 }
650 651
651 652 /* tcp_wput_flush is called by tcp_wput_nondata to handle M_FLUSH messages. */
652 653 static void
653 654 tcp_wput_flush(tcp_t *tcp, mblk_t *mp)
654 655 {
655 656 uchar_t fval = *mp->b_rptr;
656 657 mblk_t *tail;
657 658 conn_t *connp = tcp->tcp_connp;
658 659 queue_t *q = connp->conn_wq;
659 660
660 661 /* TODO: How should flush interact with urgent data? */
661 662 if ((fval & FLUSHW) && tcp->tcp_xmit_head != NULL &&
662 663 !(tcp->tcp_valid_bits & TCP_URG_VALID)) {
663 664 /*
664 665 * Flush only data that has not yet been put on the wire. If
665 666 * we flush data that we have already transmitted, life, as we
666 667 * know it, may come to an end.
667 668 */
668 669 tail = tcp->tcp_xmit_tail;
669 670 tail->b_wptr -= tcp->tcp_xmit_tail_unsent;
670 671 tcp->tcp_xmit_tail_unsent = 0;
671 672 tcp->tcp_unsent = 0;
672 673 if (tail->b_wptr != tail->b_rptr)
673 674 tail = tail->b_cont;
674 675 if (tail) {
675 676 mblk_t **excess = &tcp->tcp_xmit_head;
676 677 for (;;) {
677 678 mblk_t *mp1 = *excess;
678 679 if (mp1 == tail)
679 680 break;
680 681 tcp->tcp_xmit_tail = mp1;
681 682 tcp->tcp_xmit_last = mp1;
682 683 excess = &mp1->b_cont;
683 684 }
684 685 *excess = NULL;
685 686 tcp_close_mpp(&tail);
686 687 if (tcp->tcp_snd_zcopy_aware)
687 688 tcp_zcopy_notify(tcp);
688 689 }
689 690 /*
690 691 * We have no unsent data, so unsent must be less than
691 692 * conn_sndlowat, so re-enable flow.
692 693 */
693 694 mutex_enter(&tcp->tcp_non_sq_lock);
694 695 if (tcp->tcp_flow_stopped) {
695 696 tcp_clrqfull(tcp);
696 697 }
697 698 mutex_exit(&tcp->tcp_non_sq_lock);
698 699 }
699 700 /*
700 701 * TODO: you can't just flush these, you have to increase rwnd for one
701 702 * thing. For another, how should urgent data interact?
702 703 */
703 704 if (fval & FLUSHR) {
704 705 *mp->b_rptr = fval & ~FLUSHW;
705 706 /* XXX */
706 707 qreply(q, mp);
707 708 return;
708 709 }
709 710 freemsg(mp);
710 711 }
711 712
712 713 /*
713 714 * tcp_wput_iocdata is called by tcp_wput_nondata to handle all M_IOCDATA
714 715 * messages.
715 716 */
716 717 static void
717 718 tcp_wput_iocdata(tcp_t *tcp, mblk_t *mp)
718 719 {
719 720 mblk_t *mp1;
720 721 struct iocblk *iocp = (struct iocblk *)mp->b_rptr;
721 722 STRUCT_HANDLE(strbuf, sb);
722 723 uint_t addrlen;
723 724 conn_t *connp = tcp->tcp_connp;
724 725 queue_t *q = connp->conn_wq;
725 726
726 727 /* Make sure it is one of ours. */
727 728 switch (iocp->ioc_cmd) {
728 729 case TI_GETMYNAME:
729 730 case TI_GETPEERNAME:
730 731 break;
731 732 default:
732 733 /*
733 734 * If the conn is closing, then error the ioctl here. Otherwise
734 735 * use the CONN_IOCTLREF_* macros to hold off tcp_close until
735 736 * we're done here.
736 737 */
737 738 mutex_enter(&connp->conn_lock);
738 739 if (connp->conn_state_flags & CONN_CLOSING) {
739 740 mutex_exit(&connp->conn_lock);
740 741 iocp->ioc_error = EINVAL;
741 742 mp->b_datap->db_type = M_IOCNAK;
742 743 iocp->ioc_count = 0;
743 744 qreply(q, mp);
744 745 return;
745 746 }
746 747
747 748 CONN_INC_IOCTLREF_LOCKED(connp);
748 749 ip_wput_nondata(q, mp);
749 750 CONN_DEC_IOCTLREF(connp);
750 751 return;
751 752 }
752 753 switch (mi_copy_state(q, mp, &mp1)) {
753 754 case -1:
754 755 return;
755 756 case MI_COPY_CASE(MI_COPY_IN, 1):
756 757 break;
757 758 case MI_COPY_CASE(MI_COPY_OUT, 1):
758 759 /* Copy out the strbuf. */
759 760 mi_copyout(q, mp);
760 761 return;
761 762 case MI_COPY_CASE(MI_COPY_OUT, 2):
762 763 /* All done. */
763 764 mi_copy_done(q, mp, 0);
764 765 return;
765 766 default:
766 767 mi_copy_done(q, mp, EPROTO);
767 768 return;
768 769 }
769 770 /* Check alignment of the strbuf */
770 771 if (!OK_32PTR(mp1->b_rptr)) {
771 772 mi_copy_done(q, mp, EINVAL);
772 773 return;
773 774 }
774 775
775 776 STRUCT_SET_HANDLE(sb, iocp->ioc_flag, (void *)mp1->b_rptr);
776 777
777 778 if (connp->conn_family == AF_INET)
778 779 addrlen = sizeof (sin_t);
779 780 else
780 781 addrlen = sizeof (sin6_t);
781 782
782 783 if (STRUCT_FGET(sb, maxlen) < addrlen) {
783 784 mi_copy_done(q, mp, EINVAL);
784 785 return;
785 786 }
786 787
787 788 switch (iocp->ioc_cmd) {
788 789 case TI_GETMYNAME:
789 790 break;
790 791 case TI_GETPEERNAME:
791 792 if (tcp->tcp_state < TCPS_SYN_RCVD) {
792 793 mi_copy_done(q, mp, ENOTCONN);
793 794 return;
794 795 }
795 796 break;
796 797 }
797 798 mp1 = mi_copyout_alloc(q, mp, STRUCT_FGETP(sb, buf), addrlen, B_TRUE);
798 799 if (!mp1)
799 800 return;
800 801
801 802 STRUCT_FSET(sb, len, addrlen);
802 803 switch (((struct iocblk *)mp->b_rptr)->ioc_cmd) {
803 804 case TI_GETMYNAME:
804 805 (void) conn_getsockname(connp, (struct sockaddr *)mp1->b_wptr,
805 806 &addrlen);
806 807 break;
807 808 case TI_GETPEERNAME:
808 809 (void) conn_getpeername(connp, (struct sockaddr *)mp1->b_wptr,
809 810 &addrlen);
810 811 break;
811 812 }
812 813 mp1->b_wptr += addrlen;
813 814 /* Copy out the address */
814 815 mi_copyout(q, mp);
815 816 }
816 817
817 818 /*
818 819 * tcp_wput_ioctl is called by tcp_wput_nondata() to handle all M_IOCTL
819 820 * messages.
820 821 */
821 822 /* ARGSUSED */
822 823 static void
823 824 tcp_wput_ioctl(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
824 825 {
825 826 conn_t *connp = (conn_t *)arg;
826 827 tcp_t *tcp = connp->conn_tcp;
827 828 queue_t *q = connp->conn_wq;
828 829 struct iocblk *iocp;
829 830
830 831 ASSERT(DB_TYPE(mp) == M_IOCTL);
831 832 /*
832 833 * Try and ASSERT the minimum possible references on the
833 834 * conn early enough. Since we are executing on write side,
834 835 * the connection is obviously not detached and that means
835 836 * there is a ref each for TCP and IP. Since we are behind
836 837 * the squeue, the minimum references needed are 3. If the
837 838 * conn is in classifier hash list, there should be an
838 839 * extra ref for that (we check both the possibilities).
839 840 */
840 841 ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
841 842 (connp->conn_fanout == NULL && connp->conn_ref >= 3));
842 843
843 844 iocp = (struct iocblk *)mp->b_rptr;
844 845 switch (iocp->ioc_cmd) {
845 846 case _SIOCSOCKFALLBACK:
846 847 /*
847 848 * Either sockmod is about to be popped and the socket
848 849 * would now be treated as a plain stream, or a module
849 850 * is about to be pushed so we could no longer use read-
850 851 * side synchronous streams for fused loopback tcp.
851 852 * Drain any queued data and disable direct sockfs
852 853 * interface from now on.
853 854 */
854 855 if (!tcp->tcp_issocket) {
855 856 DB_TYPE(mp) = M_IOCNAK;
856 857 iocp->ioc_error = EINVAL;
857 858 } else {
858 859 tcp_use_pure_tpi(tcp);
859 860 DB_TYPE(mp) = M_IOCACK;
860 861 iocp->ioc_error = 0;
861 862 }
862 863 iocp->ioc_count = 0;
863 864 iocp->ioc_rval = 0;
864 865 qreply(q, mp);
865 866 return;
866 867 }
867 868
868 869 /*
869 870 * If the conn is closing, then error the ioctl here. Otherwise bump the
870 871 * conn_ioctlref to hold off tcp_close until we're done here.
871 872 */
872 873 mutex_enter(&(connp)->conn_lock);
873 874 if ((connp)->conn_state_flags & CONN_CLOSING) {
874 875 mutex_exit(&(connp)->conn_lock);
875 876 iocp->ioc_error = EINVAL;
876 877 mp->b_datap->db_type = M_IOCNAK;
877 878 iocp->ioc_count = 0;
878 879 qreply(q, mp);
879 880 return;
880 881 }
881 882
882 883 CONN_INC_IOCTLREF_LOCKED(connp);
883 884 ip_wput_nondata(q, mp);
884 885 CONN_DEC_IOCTLREF(connp);
885 886 }
886 887
887 888 /*
888 889 * This routine is called by tcp_wput() to handle all TPI requests.
889 890 */
890 891 /* ARGSUSED */
891 892 static void
892 893 tcp_wput_proto(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
893 894 {
894 895 conn_t *connp = (conn_t *)arg;
895 896 tcp_t *tcp = connp->conn_tcp;
896 897 union T_primitives *tprim = (union T_primitives *)mp->b_rptr;
897 898 uchar_t *rptr;
898 899 t_scalar_t type;
899 900 cred_t *cr;
900 901
901 902 /*
902 903 * Try and ASSERT the minimum possible references on the
903 904 * conn early enough. Since we are executing on write side,
904 905 * the connection is obviously not detached and that means
905 906 * there is a ref each for TCP and IP. Since we are behind
906 907 * the squeue, the minimum references needed are 3. If the
907 908 * conn is in classifier hash list, there should be an
908 909 * extra ref for that (we check both the possibilities).
909 910 */
910 911 ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
911 912 (connp->conn_fanout == NULL && connp->conn_ref >= 3));
912 913
913 914 rptr = mp->b_rptr;
914 915 ASSERT((uintptr_t)(mp->b_wptr - rptr) <= (uintptr_t)INT_MAX);
915 916 if ((mp->b_wptr - rptr) >= sizeof (t_scalar_t)) {
916 917 type = ((union T_primitives *)rptr)->type;
917 918 if (type == T_EXDATA_REQ) {
918 919 tcp_output_urgent(connp, mp, arg2, NULL);
919 920 } else if (type != T_DATA_REQ) {
920 921 goto non_urgent_data;
921 922 } else {
922 923 /* TODO: options, flags, ... from user */
923 924 /* Set length to zero for reclamation below */
924 925 tcp_wput_data(tcp, mp->b_cont, B_TRUE);
925 926 freeb(mp);
926 927 }
927 928 return;
928 929 } else {
929 930 if (connp->conn_debug) {
930 931 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
931 932 "tcp_wput_proto, dropping one...");
932 933 }
933 934 freemsg(mp);
934 935 return;
935 936 }
936 937
937 938 non_urgent_data:
938 939
939 940 switch ((int)tprim->type) {
940 941 case O_T_BIND_REQ: /* bind request */
941 942 case T_BIND_REQ: /* new semantics bind request */
942 943 tcp_tpi_bind(tcp, mp);
943 944 break;
944 945 case T_UNBIND_REQ: /* unbind request */
945 946 tcp_tpi_unbind(tcp, mp);
946 947 break;
947 948 case O_T_CONN_RES: /* old connection response XXX */
948 949 case T_CONN_RES: /* connection response */
949 950 tcp_tli_accept(tcp, mp);
950 951 break;
951 952 case T_CONN_REQ: /* connection request */
952 953 tcp_tpi_connect(tcp, mp);
953 954 break;
954 955 case T_DISCON_REQ: /* disconnect request */
955 956 tcp_disconnect(tcp, mp);
956 957 break;
957 958 case T_CAPABILITY_REQ:
958 959 tcp_capability_req(tcp, mp); /* capability request */
959 960 break;
960 961 case T_INFO_REQ: /* information request */
961 962 tcp_info_req(tcp, mp);
962 963 break;
963 964 case T_SVR4_OPTMGMT_REQ: /* manage options req */
964 965 case T_OPTMGMT_REQ:
965 966 /*
966 967 * Note: no support for snmpcom_req() through new
967 968 * T_OPTMGMT_REQ. See comments in ip.c
968 969 */
969 970
970 971 /*
971 972 * All Solaris components should pass a db_credp
972 973 * for this TPI message, hence we ASSERT.
973 974 * But in case there is some other M_PROTO that looks
974 975 * like a TPI message sent by some other kernel
975 976 * component, we check and return an error.
976 977 */
977 978 cr = msg_getcred(mp, NULL);
978 979 ASSERT(cr != NULL);
979 980 if (cr == NULL) {
980 981 tcp_err_ack(tcp, mp, TSYSERR, EINVAL);
981 982 return;
982 983 }
983 984 /*
984 985 * If EINPROGRESS is returned, the request has been queued
985 986 * for subsequent processing by ip_restart_optmgmt(), which
986 987 * will do the CONN_DEC_REF().
987 988 */
988 989 if ((int)tprim->type == T_SVR4_OPTMGMT_REQ) {
989 990 svr4_optcom_req(connp->conn_wq, mp, cr, &tcp_opt_obj);
990 991 } else {
991 992 tpi_optcom_req(connp->conn_wq, mp, cr, &tcp_opt_obj);
992 993 }
993 994 break;
994 995
995 996 case T_UNITDATA_REQ: /* unitdata request */
996 997 tcp_err_ack(tcp, mp, TNOTSUPPORT, 0);
997 998 break;
998 999 case T_ORDREL_REQ: /* orderly release req */
999 1000 freemsg(mp);
1000 1001
1001 1002 if (tcp->tcp_fused)
1002 1003 tcp_unfuse(tcp);
1003 1004
1004 1005 if (tcp_xmit_end(tcp) != 0) {
1005 1006 /*
1006 1007 * We were crossing FINs and got a reset from
1007 1008 * the other side. Just ignore it.
1008 1009 */
1009 1010 if (connp->conn_debug) {
1010 1011 (void) strlog(TCP_MOD_ID, 0, 1,
1011 1012 SL_ERROR|SL_TRACE,
1012 1013 "tcp_wput_proto, T_ORDREL_REQ out of "
1013 1014 "state %s",
1014 1015 tcp_display(tcp, NULL,
1015 1016 DISP_ADDR_AND_PORT));
1016 1017 }
1017 1018 }
1018 1019 break;
1019 1020 case T_ADDR_REQ:
1020 1021 tcp_addr_req(tcp, mp);
1021 1022 break;
1022 1023 default:
1023 1024 if (connp->conn_debug) {
1024 1025 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1025 1026 "tcp_wput_proto, bogus TPI msg, type %d",
1026 1027 tprim->type);
1027 1028 }
1028 1029 /*
1029 1030 * We used to M_ERROR. Sending TNOTSUPPORT gives the user
1030 1031 * to recover.
1031 1032 */
1032 1033 tcp_err_ack(tcp, mp, TNOTSUPPORT, 0);
1033 1034 break;
1034 1035 }
1035 1036 }
1036 1037
1037 1038 /*
1038 1039 * Handle special out-of-band ioctl requests (see PSARC/2008/265).
1039 1040 */
1040 1041 static void
1041 1042 tcp_wput_cmdblk(queue_t *q, mblk_t *mp)
1042 1043 {
1043 1044 void *data;
1044 1045 mblk_t *datamp = mp->b_cont;
1045 1046 conn_t *connp = Q_TO_CONN(q);
1046 1047 tcp_t *tcp = connp->conn_tcp;
1047 1048 cmdblk_t *cmdp = (cmdblk_t *)mp->b_rptr;
1048 1049
1049 1050 if (datamp == NULL || MBLKL(datamp) < cmdp->cb_len) {
1050 1051 cmdp->cb_error = EPROTO;
1051 1052 qreply(q, mp);
1052 1053 return;
1053 1054 }
1054 1055
1055 1056 data = datamp->b_rptr;
1056 1057
1057 1058 switch (cmdp->cb_cmd) {
1058 1059 case TI_GETPEERNAME:
1059 1060 if (tcp->tcp_state < TCPS_SYN_RCVD)
1060 1061 cmdp->cb_error = ENOTCONN;
1061 1062 else
1062 1063 cmdp->cb_error = conn_getpeername(connp, data,
1063 1064 &cmdp->cb_len);
1064 1065 break;
1065 1066 case TI_GETMYNAME:
1066 1067 cmdp->cb_error = conn_getsockname(connp, data, &cmdp->cb_len);
1067 1068 break;
1068 1069 default:
1069 1070 cmdp->cb_error = EINVAL;
1070 1071 break;
1071 1072 }
1072 1073
1073 1074 qreply(q, mp);
1074 1075 }
1075 1076
1076 1077 /*
1077 1078 * The TCP fast path write put procedure.
1078 1079 * NOTE: the logic of the fast path is duplicated from tcp_wput_data()
1079 1080 */
1080 1081 /* ARGSUSED */
1081 1082 void
1082 1083 tcp_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1083 1084 {
1084 1085 int len;
1085 1086 int hdrlen;
1086 1087 int plen;
1087 1088 mblk_t *mp1;
1088 1089 uchar_t *rptr;
1089 1090 uint32_t snxt;
1090 1091 tcpha_t *tcpha;
1091 1092 struct datab *db;
1092 1093 uint32_t suna;
1093 1094 uint32_t mss;
1094 1095 ipaddr_t *dst;
1095 1096 ipaddr_t *src;
1096 1097 uint32_t sum;
1097 1098 int usable;
1098 1099 conn_t *connp = (conn_t *)arg;
1099 1100 tcp_t *tcp = connp->conn_tcp;
1100 1101 uint32_t msize;
1101 1102 tcp_stack_t *tcps = tcp->tcp_tcps;
1102 1103 ip_xmit_attr_t *ixa;
1103 1104 clock_t now;
1104 1105
1105 1106 /*
1106 1107 * Try and ASSERT the minimum possible references on the
1107 1108 * conn early enough. Since we are executing on write side,
1108 1109 * the connection is obviously not detached and that means
1109 1110 * there is a ref each for TCP and IP. Since we are behind
1110 1111 * the squeue, the minimum references needed are 3. If the
1111 1112 * conn is in classifier hash list, there should be an
1112 1113 * extra ref for that (we check both the possibilities).
1113 1114 */
1114 1115 ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
1115 1116 (connp->conn_fanout == NULL && connp->conn_ref >= 3));
1116 1117
1117 1118 ASSERT(DB_TYPE(mp) == M_DATA);
1118 1119 msize = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1119 1120
1120 1121 mutex_enter(&tcp->tcp_non_sq_lock);
1121 1122 tcp->tcp_squeue_bytes -= msize;
1122 1123 mutex_exit(&tcp->tcp_non_sq_lock);
1123 1124
1124 1125 /* Bypass tcp protocol for fused tcp loopback */
1125 1126 if (tcp->tcp_fused && tcp_fuse_output(tcp, mp, msize))
1126 1127 return;
1127 1128
1128 1129 mss = tcp->tcp_mss;
1129 1130 /*
1130 1131 * If ZEROCOPY has turned off, try not to send any zero-copy message
1131 1132 * down. Do backoff, now.
1132 1133 */
1133 1134 if (tcp->tcp_snd_zcopy_aware && !tcp->tcp_snd_zcopy_on)
1134 1135 mp = tcp_zcopy_backoff(tcp, mp, B_FALSE);
1135 1136
1136 1137
1137 1138 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1138 1139 len = (int)(mp->b_wptr - mp->b_rptr);
1139 1140
1140 1141 /*
1141 1142 * Criteria for fast path:
1142 1143 *
1143 1144 * 1. no unsent data
1144 1145 * 2. single mblk in request
1145 1146 * 3. connection established
1146 1147 * 4. data in mblk
1147 1148 * 5. len <= mss
1148 1149 * 6. no tcp_valid bits
1149 1150 */
1150 1151 if ((tcp->tcp_unsent != 0) ||
1151 1152 (tcp->tcp_cork) ||
1152 1153 (mp->b_cont != NULL) ||
1153 1154 (tcp->tcp_state != TCPS_ESTABLISHED) ||
1154 1155 (len == 0) ||
1155 1156 (len > mss) ||
1156 1157 (tcp->tcp_valid_bits != 0)) {
1157 1158 tcp_wput_data(tcp, mp, B_FALSE);
1158 1159 return;
1159 1160 }
1160 1161
1161 1162 ASSERT(tcp->tcp_xmit_tail_unsent == 0);
1162 1163 ASSERT(tcp->tcp_fin_sent == 0);
1163 1164
1164 1165 /* queue new packet onto retransmission queue */
1165 1166 if (tcp->tcp_xmit_head == NULL) {
1166 1167 tcp->tcp_xmit_head = mp;
1167 1168 } else {
1168 1169 tcp->tcp_xmit_last->b_cont = mp;
1169 1170 }
1170 1171 tcp->tcp_xmit_last = mp;
1171 1172 tcp->tcp_xmit_tail = mp;
1172 1173
1173 1174 /* find out how much we can send */
1174 1175 /* BEGIN CSTYLED */
1175 1176 /*
|
↓ open down ↓ |
708 lines elided |
↑ open up ↑ |
1176 1177 * un-acked usable
1177 1178 * |--------------|-----------------|
1178 1179 * tcp_suna tcp_snxt tcp_suna+tcp_swnd
1179 1180 */
1180 1181 /* END CSTYLED */
1181 1182
1182 1183 /* start sending from tcp_snxt */
1183 1184 snxt = tcp->tcp_snxt;
1184 1185
1185 1186 /*
1186 - * Check to see if this connection has been idled for some
1187 - * time and no ACK is expected. If it is, we need to slow
1188 - * start again to get back the connection's "self-clock" as
1189 - * described in VJ's paper.
1187 + * Check to see if this connection has been idle for some time and no
1188 + * ACK is expected. If so, then the congestion window size is no longer
1189 + * meaningfully tied to current network conditions.
1190 1190 *
1191 - * Reinitialize tcp_cwnd after idle.
1191 + * We reinitialize tcp_cwnd, and slow start again to get back the
1192 + * connection's "self-clock" as described in Van Jacobson's 1988 paper
1193 + * "Congestion avoidance and control".
1192 1194 */
1193 1195 now = LBOLT_FASTPATH;
1194 1196 if ((tcp->tcp_suna == snxt) && !tcp->tcp_localnet &&
1195 1197 (TICK_TO_MSEC(now - tcp->tcp_last_recv_time) >= tcp->tcp_rto)) {
1196 1198 TCP_SET_INIT_CWND(tcp, mss, tcps->tcps_slow_start_after_idle);
1197 1199 }
1198 1200
1199 1201 usable = tcp->tcp_swnd; /* tcp window size */
1200 1202 if (usable > tcp->tcp_cwnd)
1201 1203 usable = tcp->tcp_cwnd; /* congestion window smaller */
1202 1204 usable -= snxt; /* subtract stuff already sent */
1203 1205 suna = tcp->tcp_suna;
1204 1206 usable += suna;
1205 1207 /* usable can be < 0 if the congestion window is smaller */
1206 1208 if (len > usable) {
1207 1209 /* Can't send complete M_DATA in one shot */
1208 1210 goto slow;
1209 1211 }
1210 1212
1211 1213 mutex_enter(&tcp->tcp_non_sq_lock);
1212 1214 if (tcp->tcp_flow_stopped &&
1213 1215 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1214 1216 tcp_clrqfull(tcp);
1215 1217 }
1216 1218 mutex_exit(&tcp->tcp_non_sq_lock);
1217 1219
1218 1220 /*
1219 1221 * determine if anything to send (Nagle).
1220 1222 *
1221 1223 * 1. len < tcp_mss (i.e. small)
1222 1224 * 2. unacknowledged data present
1223 1225 * 3. len < nagle limit
1224 1226 * 4. last packet sent < nagle limit (previous packet sent)
1225 1227 */
1226 1228 if ((len < mss) && (snxt != suna) &&
1227 1229 (len < (int)tcp->tcp_naglim) &&
1228 1230 (tcp->tcp_last_sent_len < tcp->tcp_naglim)) {
1229 1231 /*
1230 1232 * This was the first unsent packet and normally
1231 1233 * mss < xmit_hiwater so there is no need to worry
1232 1234 * about flow control. The next packet will go
1233 1235 * through the flow control check in tcp_wput_data().
1234 1236 */
1235 1237 /* leftover work from above */
1236 1238 tcp->tcp_unsent = len;
1237 1239 tcp->tcp_xmit_tail_unsent = len;
1238 1240
1239 1241 return;
1240 1242 }
1241 1243
1242 1244 /*
1243 1245 * len <= tcp->tcp_mss && len == unsent so no sender silly window. Can
1244 1246 * send now.
1245 1247 */
1246 1248
1247 1249 if (snxt == suna) {
1248 1250 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
|
↓ open down ↓ |
47 lines elided |
↑ open up ↑ |
1249 1251 }
1250 1252
1251 1253 /* we have always sent something */
1252 1254 tcp->tcp_rack_cnt = 0;
1253 1255
1254 1256 tcp->tcp_snxt = snxt + len;
1255 1257 tcp->tcp_rack = tcp->tcp_rnxt;
1256 1258
1257 1259 if ((mp1 = dupb(mp)) == 0)
1258 1260 goto no_memory;
1259 - mp->b_prev = (mblk_t *)(uintptr_t)now;
1261 + mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
1260 1262 mp->b_next = (mblk_t *)(uintptr_t)snxt;
1261 1263
1262 1264 /* adjust tcp header information */
1263 1265 tcpha = tcp->tcp_tcpha;
1264 1266 tcpha->tha_flags = (TH_ACK|TH_PUSH);
1265 1267
1266 1268 sum = len + connp->conn_ht_ulp_len + connp->conn_sum;
1267 1269 sum = (sum >> 16) + (sum & 0xFFFF);
1268 1270 tcpha->tha_sum = htons(sum);
1269 1271
1270 1272 tcpha->tha_seq = htonl(snxt);
1271 1273
1272 1274 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1273 1275 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1274 1276 BUMP_LOCAL(tcp->tcp_obsegs);
1275 1277
1276 1278 /* Update the latest receive window size in TCP header. */
1277 1279 tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
1278 1280
1279 1281 tcp->tcp_last_sent_len = (ushort_t)len;
1280 1282
1281 1283 plen = len + connp->conn_ht_iphc_len;
1282 1284
1283 1285 ixa = connp->conn_ixa;
1284 1286 ixa->ixa_pktlen = plen;
1285 1287
1286 1288 if (ixa->ixa_flags & IXAF_IS_IPV4) {
1287 1289 tcp->tcp_ipha->ipha_length = htons(plen);
1288 1290 } else {
1289 1291 tcp->tcp_ip6h->ip6_plen = htons(plen - IPV6_HDR_LEN);
1290 1292 }
1291 1293
1292 1294 /* see if we need to allocate a mblk for the headers */
1293 1295 hdrlen = connp->conn_ht_iphc_len;
1294 1296 rptr = mp1->b_rptr - hdrlen;
1295 1297 db = mp1->b_datap;
1296 1298 if ((db->db_ref != 2) || rptr < db->db_base ||
1297 1299 (!OK_32PTR(rptr))) {
1298 1300 /* NOTE: we assume allocb returns an OK_32PTR */
1299 1301 mp = allocb(hdrlen + tcps->tcps_wroff_xtra, BPRI_MED);
1300 1302 if (!mp) {
1301 1303 freemsg(mp1);
1302 1304 goto no_memory;
1303 1305 }
|
↓ open down ↓ |
34 lines elided |
↑ open up ↑ |
1304 1306 mp->b_cont = mp1;
1305 1307 mp1 = mp;
1306 1308 /* Leave room for Link Level header */
1307 1309 rptr = &mp1->b_rptr[tcps->tcps_wroff_xtra];
1308 1310 mp1->b_wptr = &rptr[hdrlen];
1309 1311 }
1310 1312 mp1->b_rptr = rptr;
1311 1313
1312 1314 /* Fill in the timestamp option. */
1313 1315 if (tcp->tcp_snd_ts_ok) {
1314 - uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
1315 -
1316 - U32_TO_BE32(llbolt,
1317 - (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
1316 + U32_TO_BE32(now,
1317 + (char *)tcpha + TCP_MIN_HEADER_LENGTH + 4);
1318 1318 U32_TO_BE32(tcp->tcp_ts_recent,
1319 - (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
1319 + (char *)tcpha + TCP_MIN_HEADER_LENGTH + 8);
1320 1320 } else {
1321 1321 ASSERT(connp->conn_ht_ulp_len == TCP_MIN_HEADER_LENGTH);
1322 1322 }
1323 1323
1324 1324 /* copy header into outgoing packet */
1325 1325 dst = (ipaddr_t *)rptr;
1326 1326 src = (ipaddr_t *)connp->conn_ht_iphc;
1327 1327 dst[0] = src[0];
1328 1328 dst[1] = src[1];
1329 1329 dst[2] = src[2];
1330 1330 dst[3] = src[3];
1331 1331 dst[4] = src[4];
1332 1332 dst[5] = src[5];
1333 1333 dst[6] = src[6];
1334 1334 dst[7] = src[7];
1335 1335 dst[8] = src[8];
1336 1336 dst[9] = src[9];
1337 1337 if (hdrlen -= 40) {
1338 1338 hdrlen >>= 2;
1339 1339 dst += 10;
1340 1340 src += 10;
1341 1341 do {
1342 1342 *dst++ = *src++;
1343 1343 } while (--hdrlen);
1344 1344 }
1345 1345
1346 1346 /*
1347 1347 * Set the ECN info in the TCP header. Note that this
1348 1348 * is not the template header.
1349 1349 */
1350 1350 if (tcp->tcp_ecn_ok) {
1351 1351 TCP_SET_ECT(tcp, rptr);
1352 1352
1353 1353 tcpha = (tcpha_t *)(rptr + ixa->ixa_ip_hdr_length);
1354 1354 if (tcp->tcp_ecn_echo_on)
1355 1355 tcpha->tha_flags |= TH_ECE;
1356 1356 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
1357 1357 tcpha->tha_flags |= TH_CWR;
1358 1358 tcp->tcp_ecn_cwr_sent = B_TRUE;
1359 1359 }
1360 1360 }
1361 1361
1362 1362 if (tcp->tcp_ip_forward_progress) {
1363 1363 tcp->tcp_ip_forward_progress = B_FALSE;
1364 1364 connp->conn_ixa->ixa_flags |= IXAF_REACH_CONF;
1365 1365 } else {
1366 1366 connp->conn_ixa->ixa_flags &= ~IXAF_REACH_CONF;
1367 1367 }
1368 1368 tcp_send_data(tcp, mp1);
1369 1369 return;
1370 1370
1371 1371 /*
1372 1372 * If we ran out of memory, we pretend to have sent the packet
1373 1373 * and that it was lost on the wire.
1374 1374 */
1375 1375 no_memory:
1376 1376 return;
1377 1377
1378 1378 slow:
1379 1379 /* leftover work from above */
1380 1380 tcp->tcp_unsent = len;
1381 1381 tcp->tcp_xmit_tail_unsent = len;
1382 1382 tcp_wput_data(tcp, NULL, B_FALSE);
1383 1383 }
1384 1384
1385 1385 /* ARGSUSED2 */
1386 1386 void
1387 1387 tcp_output_urgent(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1388 1388 {
1389 1389 int len;
1390 1390 uint32_t msize;
1391 1391 conn_t *connp = (conn_t *)arg;
1392 1392 tcp_t *tcp = connp->conn_tcp;
1393 1393
1394 1394 msize = msgdsize(mp);
1395 1395
1396 1396 len = msize - 1;
1397 1397 if (len < 0) {
1398 1398 freemsg(mp);
1399 1399 return;
1400 1400 }
1401 1401
1402 1402 /*
1403 1403 * Try to force urgent data out on the wire. Even if we have unsent
1404 1404 * data this will at least send the urgent flag.
1405 1405 * XXX does not handle more flag correctly.
1406 1406 */
1407 1407 len += tcp->tcp_unsent;
1408 1408 len += tcp->tcp_snxt;
1409 1409 tcp->tcp_urg = len;
1410 1410 tcp->tcp_valid_bits |= TCP_URG_VALID;
1411 1411
1412 1412 /* Bypass tcp protocol for fused tcp loopback */
1413 1413 if (tcp->tcp_fused && tcp_fuse_output(tcp, mp, msize))
1414 1414 return;
1415 1415
1416 1416 /* Strip off the T_EXDATA_REQ if the data is from TPI */
1417 1417 if (DB_TYPE(mp) != M_DATA) {
1418 1418 mblk_t *mp1 = mp;
1419 1419 ASSERT(!IPCL_IS_NONSTR(connp));
1420 1420 mp = mp->b_cont;
1421 1421 freeb(mp1);
1422 1422 }
1423 1423 tcp_wput_data(tcp, mp, B_TRUE);
1424 1424 }
1425 1425
1426 1426 /*
1427 1427 * Called by streams close routine via squeues when our client blows off its
1428 1428 * descriptor, we take this to mean: "close the stream state NOW, close the tcp
1429 1429 * connection politely" When SO_LINGER is set (with a non-zero linger time and
1430 1430 * it is not a nonblocking socket) then this routine sleeps until the FIN is
1431 1431 * acked.
1432 1432 *
1433 1433 * NOTE: tcp_close potentially returns error when lingering.
1434 1434 * However, the stream head currently does not pass these errors
1435 1435 * to the application. 4.4BSD only returns EINTR and EWOULDBLOCK
1436 1436 * errors to the application (from tsleep()) and not errors
1437 1437 * like ECONNRESET caused by receiving a reset packet.
1438 1438 */
1439 1439
1440 1440 /* ARGSUSED */
1441 1441 void
1442 1442 tcp_close_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1443 1443 {
1444 1444 char *msg;
1445 1445 conn_t *connp = (conn_t *)arg;
1446 1446 tcp_t *tcp = connp->conn_tcp;
1447 1447 clock_t delta = 0;
1448 1448 tcp_stack_t *tcps = tcp->tcp_tcps;
1449 1449
1450 1450 /*
1451 1451 * When a non-STREAMS socket is being closed, it does not always
1452 1452 * stick around waiting for tcp_close_output to run and can therefore
1453 1453 * have dropped a reference already. So adjust the asserts accordingly.
1454 1454 */
1455 1455 ASSERT((connp->conn_fanout != NULL &&
1456 1456 connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 3 : 4)) ||
1457 1457 (connp->conn_fanout == NULL &&
1458 1458 connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 2 : 3)));
1459 1459
1460 1460 mutex_enter(&tcp->tcp_eager_lock);
1461 1461 if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
1462 1462 /*
1463 1463 * Cleanup for listener. For non-STREAM sockets sockfs will
1464 1464 * close all the eagers on 'q', so in that case only deal
1465 1465 * with 'q0'.
1466 1466 */
1467 1467 tcp_eager_cleanup(tcp, IPCL_IS_NONSTR(connp) ? 1 : 0);
1468 1468 tcp->tcp_wait_for_eagers = 1;
1469 1469 }
1470 1470 mutex_exit(&tcp->tcp_eager_lock);
1471 1471
1472 1472 tcp->tcp_lso = B_FALSE;
1473 1473
1474 1474 msg = NULL;
1475 1475 switch (tcp->tcp_state) {
1476 1476 case TCPS_CLOSED:
1477 1477 case TCPS_IDLE:
1478 1478 break;
1479 1479 case TCPS_BOUND:
1480 1480 if (tcp->tcp_listener != NULL) {
1481 1481 ASSERT(IPCL_IS_NONSTR(connp));
1482 1482 /*
1483 1483 * Unlink from the listener and drop the reference
1484 1484 * put on it by the eager. tcp_closei_local will not
1485 1485 * do it because tcp_tconnind_started is TRUE.
1486 1486 */
1487 1487 mutex_enter(&tcp->tcp_saved_listener->tcp_eager_lock);
1488 1488 tcp_eager_unlink(tcp);
1489 1489 mutex_exit(&tcp->tcp_saved_listener->tcp_eager_lock);
1490 1490 CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
1491 1491 }
1492 1492 break;
1493 1493 case TCPS_LISTEN:
1494 1494 break;
1495 1495 case TCPS_SYN_SENT:
1496 1496 msg = "tcp_close, during connect";
1497 1497 break;
1498 1498 case TCPS_SYN_RCVD:
1499 1499 /*
1500 1500 * Close during the connect 3-way handshake
1501 1501 * but here there may or may not be pending data
1502 1502 * already on queue. Process almost same as in
1503 1503 * the ESTABLISHED state.
1504 1504 */
1505 1505 /* FALLTHRU */
1506 1506 default:
1507 1507 if (tcp->tcp_fused)
1508 1508 tcp_unfuse(tcp);
1509 1509
1510 1510 /*
1511 1511 * If SO_LINGER has set a zero linger time, abort the
1512 1512 * connection with a reset.
1513 1513 */
1514 1514 if (connp->conn_linger && connp->conn_lingertime == 0) {
1515 1515 msg = "tcp_close, zero lingertime";
1516 1516 break;
1517 1517 }
1518 1518
1519 1519 /*
1520 1520 * Abort connection if there is unread data queued.
1521 1521 */
1522 1522 if (tcp->tcp_rcv_list || tcp->tcp_reass_head) {
1523 1523 msg = "tcp_close, unread data";
1524 1524 break;
1525 1525 }
1526 1526
1527 1527 /*
1528 1528 * Abort connection if it is being closed without first
1529 1529 * being accepted. This can happen if a listening non-STREAM
1530 1530 * socket wants to get rid of the socket, for example, if the
1531 1531 * listener is closing.
1532 1532 */
1533 1533 if (tcp->tcp_listener != NULL) {
1534 1534 ASSERT(IPCL_IS_NONSTR(connp));
1535 1535 msg = "tcp_close, close before accept";
1536 1536
1537 1537 /*
1538 1538 * Unlink from the listener and drop the reference
1539 1539 * put on it by the eager. tcp_closei_local will not
1540 1540 * do it because tcp_tconnind_started is TRUE.
1541 1541 */
1542 1542 mutex_enter(&tcp->tcp_saved_listener->tcp_eager_lock);
1543 1543 tcp_eager_unlink(tcp);
1544 1544 mutex_exit(&tcp->tcp_saved_listener->tcp_eager_lock);
1545 1545 CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
1546 1546 break;
1547 1547 }
1548 1548
1549 1549 /*
1550 1550 * Transmit the FIN before detaching the tcp_t.
1551 1551 * After tcp_detach returns this queue/perimeter
1552 1552 * no longer owns the tcp_t thus others can modify it.
1553 1553 */
1554 1554 (void) tcp_xmit_end(tcp);
1555 1555
1556 1556 /*
1557 1557 * If lingering on close then wait until the fin is acked,
1558 1558 * the SO_LINGER time passes, or a reset is sent/received.
1559 1559 */
1560 1560 if (connp->conn_linger && connp->conn_lingertime > 0 &&
1561 1561 !(tcp->tcp_fin_acked) &&
1562 1562 tcp->tcp_state >= TCPS_ESTABLISHED) {
1563 1563 if (tcp->tcp_closeflags & (FNDELAY|FNONBLOCK)) {
1564 1564 tcp->tcp_client_errno = EWOULDBLOCK;
1565 1565 } else if (tcp->tcp_client_errno == 0) {
1566 1566
1567 1567 ASSERT(tcp->tcp_linger_tid == 0);
1568 1568
1569 1569 /* conn_lingertime is in sec. */
1570 1570 tcp->tcp_linger_tid = TCP_TIMER(tcp,
1571 1571 tcp_close_linger_timeout,
1572 1572 connp->conn_lingertime * MILLISEC);
1573 1573
1574 1574 /* tcp_close_linger_timeout will finish close */
1575 1575 if (tcp->tcp_linger_tid == 0)
1576 1576 tcp->tcp_client_errno = ENOSR;
1577 1577 else
1578 1578 return;
1579 1579 }
1580 1580
1581 1581 /*
1582 1582 * Check if we need to detach or just close
1583 1583 * the instance.
1584 1584 */
1585 1585 if (tcp->tcp_state <= TCPS_LISTEN)
1586 1586 break;
1587 1587 }
1588 1588
1589 1589 /*
1590 1590 * Make sure that no other thread will access the conn_rq of
1591 1591 * this instance (through lookups etc.) as conn_rq will go
1592 1592 * away shortly.
1593 1593 */
1594 1594 tcp_acceptor_hash_remove(tcp);
1595 1595
1596 1596 mutex_enter(&tcp->tcp_non_sq_lock);
1597 1597 if (tcp->tcp_flow_stopped) {
1598 1598 tcp_clrqfull(tcp);
1599 1599 }
1600 1600 mutex_exit(&tcp->tcp_non_sq_lock);
1601 1601
1602 1602 if (tcp->tcp_timer_tid != 0) {
1603 1603 delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
1604 1604 tcp->tcp_timer_tid = 0;
1605 1605 }
1606 1606 /*
1607 1607 * Need to cancel those timers which will not be used when
1608 1608 * TCP is detached. This has to be done before the conn_wq
1609 1609 * is set to NULL.
1610 1610 */
1611 1611 tcp_timers_stop(tcp);
1612 1612
1613 1613 tcp->tcp_detached = B_TRUE;
1614 1614 if (tcp->tcp_state == TCPS_TIME_WAIT) {
1615 1615 tcp_time_wait_append(tcp);
1616 1616 TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1617 1617 ASSERT(connp->conn_ref >=
1618 1618 (IPCL_IS_NONSTR(connp) ? 2 : 3));
1619 1619 goto finish;
1620 1620 }
1621 1621
1622 1622 /*
1623 1623 * If delta is zero the timer event wasn't executed and was
1624 1624 * successfully canceled. In this case we need to restart it
1625 1625 * with the minimal delta possible.
1626 1626 */
1627 1627 if (delta >= 0)
1628 1628 tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1629 1629 delta ? delta : 1);
1630 1630
1631 1631 ASSERT(connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 2 : 3));
1632 1632 goto finish;
1633 1633 }
1634 1634
1635 1635 /* Detach did not complete. Still need to remove q from stream. */
1636 1636 if (msg) {
1637 1637 if (tcp->tcp_state == TCPS_ESTABLISHED ||
1638 1638 tcp->tcp_state == TCPS_CLOSE_WAIT)
1639 1639 TCPS_BUMP_MIB(tcps, tcpEstabResets);
1640 1640 if (tcp->tcp_state == TCPS_SYN_SENT ||
1641 1641 tcp->tcp_state == TCPS_SYN_RCVD)
1642 1642 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1643 1643 tcp_xmit_ctl(msg, tcp, tcp->tcp_snxt, 0, TH_RST);
1644 1644 }
1645 1645
1646 1646 tcp_closei_local(tcp);
1647 1647 CONN_DEC_REF(connp);
1648 1648 ASSERT(connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 1 : 2));
1649 1649
1650 1650 finish:
1651 1651 /*
1652 1652 * Don't change the queues in the case of a listener that has
1653 1653 * eagers in its q or q0. It could surprise the eagers.
1654 1654 * Instead wait for the eagers outside the squeue.
1655 1655 *
1656 1656 * For non-STREAMS sockets tcp_wait_for_eagers implies that
1657 1657 * we should delay the su_closed upcall until all eagers have
1658 1658 * dropped their references.
1659 1659 */
1660 1660 if (!tcp->tcp_wait_for_eagers) {
1661 1661 tcp->tcp_detached = B_TRUE;
1662 1662 connp->conn_rq = NULL;
1663 1663 connp->conn_wq = NULL;
1664 1664
1665 1665 /* non-STREAM socket, release the upper handle */
1666 1666 if (IPCL_IS_NONSTR(connp)) {
1667 1667 ASSERT(connp->conn_upper_handle != NULL);
1668 1668 (*connp->conn_upcalls->su_closed)
1669 1669 (connp->conn_upper_handle);
1670 1670 connp->conn_upper_handle = NULL;
1671 1671 connp->conn_upcalls = NULL;
1672 1672 }
1673 1673 }
1674 1674
1675 1675 /* Signal tcp_close() to finish closing. */
1676 1676 mutex_enter(&tcp->tcp_closelock);
1677 1677 tcp->tcp_closed = 1;
1678 1678 cv_signal(&tcp->tcp_closecv);
1679 1679 mutex_exit(&tcp->tcp_closelock);
1680 1680 }
1681 1681
1682 1682 /* ARGSUSED */
1683 1683 void
1684 1684 tcp_shutdown_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1685 1685 {
1686 1686 conn_t *connp = (conn_t *)arg;
1687 1687 tcp_t *tcp = connp->conn_tcp;
1688 1688
1689 1689 freemsg(mp);
1690 1690
1691 1691 if (tcp->tcp_fused)
1692 1692 tcp_unfuse(tcp);
1693 1693
1694 1694 if (tcp_xmit_end(tcp) != 0) {
1695 1695 /*
1696 1696 * We were crossing FINs and got a reset from
1697 1697 * the other side. Just ignore it.
1698 1698 */
1699 1699 if (connp->conn_debug) {
1700 1700 (void) strlog(TCP_MOD_ID, 0, 1,
1701 1701 SL_ERROR|SL_TRACE,
1702 1702 "tcp_shutdown_output() out of state %s",
1703 1703 tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
1704 1704 }
1705 1705 }
1706 1706 }
1707 1707
1708 1708 #pragma inline(tcp_send_data)
1709 1709
1710 1710 void
1711 1711 tcp_send_data(tcp_t *tcp, mblk_t *mp)
1712 1712 {
1713 1713 conn_t *connp = tcp->tcp_connp;
1714 1714
1715 1715 /*
1716 1716 * Check here to avoid sending zero-copy message down to IP when
1717 1717 * ZEROCOPY capability has turned off. We only need to deal with
1718 1718 * the race condition between sockfs and the notification here.
1719 1719 * Since we have tried to backoff the tcp_xmit_head when turning
1720 1720 * zero-copy off and new messages in tcp_output(), we simply drop
1721 1721 * the dup'ed packet here and let tcp retransmit, if tcp_xmit_zc_clean
1722 1722 * is not true.
1723 1723 */
1724 1724 if (tcp->tcp_snd_zcopy_aware && !tcp->tcp_snd_zcopy_on &&
1725 1725 !tcp->tcp_xmit_zc_clean) {
1726 1726 ip_drop_output("TCP ZC was disabled but not clean", mp, NULL);
1727 1727 freemsg(mp);
1728 1728 return;
1729 1729 }
1730 1730
1731 1731 DTRACE_TCP5(send, mblk_t *, NULL, ip_xmit_attr_t *, connp->conn_ixa,
1732 1732 __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, tcp,
1733 1733 __dtrace_tcp_tcph_t *,
1734 1734 &mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
1735 1735
1736 1736 ASSERT(connp->conn_ixa->ixa_notify_cookie == connp->conn_tcp);
1737 1737 (void) conn_ip_output(mp, connp->conn_ixa);
1738 1738 }
1739 1739
1740 1740 /* ARGSUSED2 */
1741 1741 void
1742 1742 tcp_send_synack(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1743 1743 {
1744 1744 conn_t *econnp = (conn_t *)arg;
1745 1745 tcp_t *tcp = econnp->conn_tcp;
1746 1746 ip_xmit_attr_t *ixa = econnp->conn_ixa;
1747 1747
1748 1748 /* Guard against a RST having blown it away while on the squeue */
1749 1749 if (tcp->tcp_state == TCPS_CLOSED) {
1750 1750 freemsg(mp);
1751 1751 return;
1752 1752 }
1753 1753
1754 1754 /*
1755 1755 * In the off-chance that the eager received and responded to
1756 1756 * some other packet while the SYN|ACK was queued, we recalculate
1757 1757 * the ixa_pktlen. It would be better to fix the SYN/accept
1758 1758 * multithreading scheme to avoid this complexity.
1759 1759 */
1760 1760 ixa->ixa_pktlen = msgdsize(mp);
1761 1761 (void) conn_ip_output(mp, ixa);
1762 1762 }
1763 1763
|
↓ open down ↓ |
434 lines elided |
↑ open up ↑ |
1764 1764 /*
1765 1765 * tcp_send() is called by tcp_wput_data() and returns one of the following:
1766 1766 *
1767 1767 * -1 = failed allocation.
1768 1768 * 0 = We've either successfully sent data, or our usable send window is too
1769 1769 * small and we'd rather wait until later before sending again.
1770 1770 */
1771 1771 static int
1772 1772 tcp_send(tcp_t *tcp, const int mss, const int total_hdr_len,
1773 1773 const int tcp_hdr_len, const int num_sack_blk, int *usable,
1774 - uint_t *snxt, int *tail_unsent, mblk_t **xmit_tail, mblk_t *local_time)
1774 + uint32_t *snxt, int *tail_unsent, mblk_t **xmit_tail, mblk_t *local_time)
1775 1775 {
1776 1776 int num_lso_seg = 1;
1777 1777 uint_t lso_usable;
1778 1778 boolean_t do_lso_send = B_FALSE;
1779 1779 tcp_stack_t *tcps = tcp->tcp_tcps;
1780 1780 conn_t *connp = tcp->tcp_connp;
1781 1781 ip_xmit_attr_t *ixa = connp->conn_ixa;
1782 1782
1783 1783 /*
1784 1784 * Check LSO possibility. The value of tcp->tcp_lso indicates whether
1785 1785 * the underlying connection is LSO capable. Will check whether having
1786 1786 * enough available data to initiate LSO transmission in the for(){}
1787 1787 * loops.
1788 1788 */
1789 1789 if (tcp->tcp_lso && (tcp->tcp_valid_bits & ~TCP_FSS_VALID) == 0)
1790 1790 do_lso_send = B_TRUE;
1791 1791
1792 1792 for (;;) {
1793 1793 struct datab *db;
1794 1794 tcpha_t *tcpha;
1795 1795 uint32_t sum;
1796 1796 mblk_t *mp, *mp1;
1797 1797 uchar_t *rptr;
1798 1798 int len;
1799 1799
1800 1800 /*
1801 1801 * Calculate the maximum payload length we can send at one
1802 1802 * time.
1803 1803 */
1804 1804 if (do_lso_send) {
1805 1805 /*
1806 1806 * Determine whether or not it's possible to do LSO,
1807 1807 * and if so, how much data we can send.
1808 1808 */
1809 1809 if ((*usable - 1) / mss >= 1) {
1810 1810 lso_usable = MIN(tcp->tcp_lso_max, *usable);
1811 1811 num_lso_seg = lso_usable / mss;
1812 1812 if (lso_usable % mss) {
1813 1813 num_lso_seg++;
1814 1814 tcp->tcp_last_sent_len = (ushort_t)
1815 1815 (lso_usable % mss);
1816 1816 } else {
1817 1817 tcp->tcp_last_sent_len = (ushort_t)mss;
1818 1818 }
1819 1819 } else {
1820 1820 do_lso_send = B_FALSE;
1821 1821 num_lso_seg = 1;
1822 1822 lso_usable = mss;
1823 1823 }
1824 1824 }
1825 1825
1826 1826 ASSERT(num_lso_seg <= IP_MAXPACKET / mss + 1);
1827 1827
1828 1828 len = mss;
1829 1829 if (len > *usable) {
1830 1830 ASSERT(do_lso_send == B_FALSE);
1831 1831
1832 1832 len = *usable;
1833 1833 if (len <= 0) {
1834 1834 /* Terminate the loop */
1835 1835 break; /* success; too small */
1836 1836 }
1837 1837 /*
1838 1838 * Sender silly-window avoidance.
1839 1839 * Ignore this if we are going to send a
1840 1840 * zero window probe out.
1841 1841 *
1842 1842 * TODO: force data into microscopic window?
1843 1843 * ==> (!pushed || (unsent > usable))
1844 1844 */
1845 1845 if (len < (tcp->tcp_max_swnd >> 1) &&
1846 1846 (tcp->tcp_unsent - (*snxt - tcp->tcp_snxt)) > len &&
1847 1847 !((tcp->tcp_valid_bits & TCP_URG_VALID) &&
1848 1848 len == 1) && (! tcp->tcp_zero_win_probe)) {
1849 1849 /*
1850 1850 * If the retransmit timer is not running
1851 1851 * we start it so that we will retransmit
1852 1852 * in the case when the receiver has
1853 1853 * decremented the window.
1854 1854 */
1855 1855 if (*snxt == tcp->tcp_snxt &&
1856 1856 *snxt == tcp->tcp_suna) {
1857 1857 /*
1858 1858 * We are not supposed to send
1859 1859 * anything. So let's wait a little
1860 1860 * bit longer before breaking SWS
1861 1861 * avoidance.
1862 1862 *
1863 1863 * What should the value be?
1864 1864 * Suggestion: MAX(init rexmit time,
1865 1865 * tcp->tcp_rto)
1866 1866 */
1867 1867 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1868 1868 }
1869 1869 break; /* success; too small */
1870 1870 }
1871 1871 }
1872 1872
1873 1873 tcpha = tcp->tcp_tcpha;
1874 1874
1875 1875 /*
1876 1876 * The reason to adjust len here is that we need to set flags
1877 1877 * and calculate checksum.
1878 1878 */
1879 1879 if (do_lso_send)
1880 1880 len = lso_usable;
1881 1881
1882 1882 *usable -= len; /* Approximate - can be adjusted later */
1883 1883 if (*usable > 0)
1884 1884 tcpha->tha_flags = TH_ACK;
1885 1885 else
1886 1886 tcpha->tha_flags = (TH_ACK | TH_PUSH);
1887 1887
1888 1888 /*
1889 1889 * Prime pump for IP's checksumming on our behalf.
1890 1890 * Include the adjustment for a source route if any.
1891 1891 * In case of LSO, the partial pseudo-header checksum should
1892 1892 * exclusive TCP length, so zero tha_sum before IP calculate
1893 1893 * pseudo-header checksum for partial checksum offload.
1894 1894 */
1895 1895 if (do_lso_send) {
1896 1896 sum = 0;
1897 1897 } else {
1898 1898 sum = len + tcp_hdr_len + connp->conn_sum;
1899 1899 sum = (sum >> 16) + (sum & 0xFFFF);
1900 1900 }
1901 1901 tcpha->tha_sum = htons(sum);
1902 1902 tcpha->tha_seq = htonl(*snxt);
1903 1903
1904 1904 /*
1905 1905 * Branch off to tcp_xmit_mp() if any of the VALID bits is
1906 1906 * set. For the case when TCP_FSS_VALID is the only valid
1907 1907 * bit (normal active close), branch off only when we think
1908 1908 * that the FIN flag needs to be set. Note for this case,
1909 1909 * that (snxt + len) may not reflect the actual seg_len,
1910 1910 * as len may be further reduced in tcp_xmit_mp(). If len
1911 1911 * gets modified, we will end up here again.
1912 1912 */
1913 1913 if (tcp->tcp_valid_bits != 0 &&
1914 1914 (tcp->tcp_valid_bits != TCP_FSS_VALID ||
1915 1915 ((*snxt + len) == tcp->tcp_fss))) {
1916 1916 uchar_t *prev_rptr;
1917 1917 uint32_t prev_snxt = tcp->tcp_snxt;
1918 1918
1919 1919 if (*tail_unsent == 0) {
1920 1920 ASSERT((*xmit_tail)->b_cont != NULL);
1921 1921 *xmit_tail = (*xmit_tail)->b_cont;
1922 1922 prev_rptr = (*xmit_tail)->b_rptr;
1923 1923 *tail_unsent = (int)((*xmit_tail)->b_wptr -
1924 1924 (*xmit_tail)->b_rptr);
1925 1925 } else {
1926 1926 prev_rptr = (*xmit_tail)->b_rptr;
1927 1927 (*xmit_tail)->b_rptr = (*xmit_tail)->b_wptr -
1928 1928 *tail_unsent;
1929 1929 }
1930 1930 mp = tcp_xmit_mp(tcp, *xmit_tail, len, NULL, NULL,
1931 1931 *snxt, B_FALSE, (uint32_t *)&len, B_FALSE);
1932 1932 /* Restore tcp_snxt so we get amount sent right. */
1933 1933 tcp->tcp_snxt = prev_snxt;
1934 1934 if (prev_rptr == (*xmit_tail)->b_rptr) {
1935 1935 /*
1936 1936 * If the previous timestamp is still in use,
1937 1937 * don't stomp on it.
1938 1938 */
1939 1939 if ((*xmit_tail)->b_next == NULL) {
1940 1940 (*xmit_tail)->b_prev = local_time;
1941 1941 (*xmit_tail)->b_next =
1942 1942 (mblk_t *)(uintptr_t)(*snxt);
1943 1943 }
1944 1944 } else
1945 1945 (*xmit_tail)->b_rptr = prev_rptr;
1946 1946
1947 1947 if (mp == NULL) {
1948 1948 return (-1);
1949 1949 }
1950 1950 mp1 = mp->b_cont;
1951 1951
1952 1952 if (len <= mss) /* LSO is unusable (!do_lso_send) */
1953 1953 tcp->tcp_last_sent_len = (ushort_t)len;
1954 1954 while (mp1->b_cont) {
1955 1955 *xmit_tail = (*xmit_tail)->b_cont;
1956 1956 (*xmit_tail)->b_prev = local_time;
1957 1957 (*xmit_tail)->b_next =
1958 1958 (mblk_t *)(uintptr_t)(*snxt);
1959 1959 mp1 = mp1->b_cont;
1960 1960 }
1961 1961 *snxt += len;
1962 1962 *tail_unsent = (*xmit_tail)->b_wptr - mp1->b_wptr;
1963 1963 BUMP_LOCAL(tcp->tcp_obsegs);
1964 1964 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1965 1965 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1966 1966 tcp_send_data(tcp, mp);
1967 1967 continue;
1968 1968 }
1969 1969
1970 1970 *snxt += len; /* Adjust later if we don't send all of len */
1971 1971 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1972 1972 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1973 1973
1974 1974 if (*tail_unsent) {
1975 1975 /* Are the bytes above us in flight? */
1976 1976 rptr = (*xmit_tail)->b_wptr - *tail_unsent;
1977 1977 if (rptr != (*xmit_tail)->b_rptr) {
1978 1978 *tail_unsent -= len;
1979 1979 if (len <= mss) /* LSO is unusable */
1980 1980 tcp->tcp_last_sent_len = (ushort_t)len;
1981 1981 len += total_hdr_len;
1982 1982 ixa->ixa_pktlen = len;
1983 1983
1984 1984 if (ixa->ixa_flags & IXAF_IS_IPV4) {
1985 1985 tcp->tcp_ipha->ipha_length = htons(len);
1986 1986 } else {
1987 1987 tcp->tcp_ip6h->ip6_plen =
1988 1988 htons(len - IPV6_HDR_LEN);
1989 1989 }
1990 1990
1991 1991 mp = dupb(*xmit_tail);
1992 1992 if (mp == NULL) {
1993 1993 return (-1); /* out_of_mem */
1994 1994 }
1995 1995 mp->b_rptr = rptr;
1996 1996 /*
1997 1997 * If the old timestamp is no longer in use,
1998 1998 * sample a new timestamp now.
1999 1999 */
2000 2000 if ((*xmit_tail)->b_next == NULL) {
2001 2001 (*xmit_tail)->b_prev = local_time;
2002 2002 (*xmit_tail)->b_next =
2003 2003 (mblk_t *)(uintptr_t)(*snxt-len);
2004 2004 }
2005 2005 goto must_alloc;
2006 2006 }
2007 2007 } else {
2008 2008 *xmit_tail = (*xmit_tail)->b_cont;
2009 2009 ASSERT((uintptr_t)((*xmit_tail)->b_wptr -
2010 2010 (*xmit_tail)->b_rptr) <= (uintptr_t)INT_MAX);
2011 2011 *tail_unsent = (int)((*xmit_tail)->b_wptr -
2012 2012 (*xmit_tail)->b_rptr);
2013 2013 }
2014 2014
2015 2015 (*xmit_tail)->b_prev = local_time;
2016 2016 (*xmit_tail)->b_next = (mblk_t *)(uintptr_t)(*snxt - len);
2017 2017
2018 2018 *tail_unsent -= len;
2019 2019 if (len <= mss) /* LSO is unusable (!do_lso_send) */
2020 2020 tcp->tcp_last_sent_len = (ushort_t)len;
2021 2021
2022 2022 len += total_hdr_len;
2023 2023 ixa->ixa_pktlen = len;
2024 2024
2025 2025 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2026 2026 tcp->tcp_ipha->ipha_length = htons(len);
2027 2027 } else {
2028 2028 tcp->tcp_ip6h->ip6_plen = htons(len - IPV6_HDR_LEN);
2029 2029 }
2030 2030
2031 2031 mp = dupb(*xmit_tail);
2032 2032 if (mp == NULL) {
2033 2033 return (-1); /* out_of_mem */
2034 2034 }
2035 2035
2036 2036 len = total_hdr_len;
2037 2037 /*
2038 2038 * There are four reasons to allocate a new hdr mblk:
2039 2039 * 1) The bytes above us are in use by another packet
2040 2040 * 2) We don't have good alignment
2041 2041 * 3) The mblk is being shared
2042 2042 * 4) We don't have enough room for a header
2043 2043 */
2044 2044 rptr = mp->b_rptr - len;
2045 2045 if (!OK_32PTR(rptr) ||
2046 2046 ((db = mp->b_datap), db->db_ref != 2) ||
2047 2047 rptr < db->db_base) {
2048 2048 /* NOTE: we assume allocb returns an OK_32PTR */
2049 2049
2050 2050 must_alloc:;
2051 2051 mp1 = allocb(connp->conn_ht_iphc_allocated +
2052 2052 tcps->tcps_wroff_xtra, BPRI_MED);
2053 2053 if (mp1 == NULL) {
2054 2054 freemsg(mp);
2055 2055 return (-1); /* out_of_mem */
2056 2056 }
2057 2057 mp1->b_cont = mp;
2058 2058 mp = mp1;
|
↓ open down ↓ |
274 lines elided |
↑ open up ↑ |
2059 2059 /* Leave room for Link Level header */
2060 2060 len = total_hdr_len;
2061 2061 rptr = &mp->b_rptr[tcps->tcps_wroff_xtra];
2062 2062 mp->b_wptr = &rptr[len];
2063 2063 }
2064 2064
2065 2065 /*
2066 2066 * Fill in the header using the template header, and add
2067 2067 * options such as time-stamp, ECN and/or SACK, as needed.
2068 2068 */
2069 - tcp_fill_header(tcp, rptr, (clock_t)local_time, num_sack_blk);
2069 + tcp_fill_header(tcp, rptr, num_sack_blk);
2070 2070
2071 2071 mp->b_rptr = rptr;
2072 2072
2073 2073 if (*tail_unsent) {
2074 2074 int spill = *tail_unsent;
2075 2075
2076 2076 mp1 = mp->b_cont;
2077 2077 if (mp1 == NULL)
2078 2078 mp1 = mp;
2079 2079
2080 2080 /*
2081 2081 * If we're a little short, tack on more mblks until
2082 2082 * there is no more spillover.
2083 2083 */
2084 2084 while (spill < 0) {
2085 2085 mblk_t *nmp;
2086 2086 int nmpsz;
2087 2087
2088 2088 nmp = (*xmit_tail)->b_cont;
2089 2089 nmpsz = MBLKL(nmp);
2090 2090
2091 2091 /*
2092 2092 * Excess data in mblk; can we split it?
2093 2093 * If LSO is enabled for the connection,
2094 2094 * keep on splitting as this is a transient
2095 2095 * send path.
2096 2096 */
2097 2097 if (!do_lso_send && (spill + nmpsz > 0)) {
2098 2098 /*
2099 2099 * Don't split if stream head was
2100 2100 * told to break up larger writes
2101 2101 * into smaller ones.
2102 2102 */
2103 2103 if (tcp->tcp_maxpsz_multiplier > 0)
2104 2104 break;
2105 2105
2106 2106 /*
2107 2107 * Next mblk is less than SMSS/2
2108 2108 * rounded up to nearest 64-byte;
2109 2109 * let it get sent as part of the
2110 2110 * next segment.
2111 2111 */
2112 2112 if (tcp->tcp_localnet &&
2113 2113 !tcp->tcp_cork &&
2114 2114 (nmpsz < roundup((mss >> 1), 64)))
2115 2115 break;
2116 2116 }
2117 2117
2118 2118 *xmit_tail = nmp;
2119 2119 ASSERT((uintptr_t)nmpsz <= (uintptr_t)INT_MAX);
2120 2120 /* Stash for rtt use later */
2121 2121 (*xmit_tail)->b_prev = local_time;
2122 2122 (*xmit_tail)->b_next =
2123 2123 (mblk_t *)(uintptr_t)(*snxt - len);
2124 2124 mp1->b_cont = dupb(*xmit_tail);
2125 2125 mp1 = mp1->b_cont;
2126 2126
2127 2127 spill += nmpsz;
2128 2128 if (mp1 == NULL) {
2129 2129 *tail_unsent = spill;
2130 2130 freemsg(mp);
2131 2131 return (-1); /* out_of_mem */
2132 2132 }
2133 2133 }
2134 2134
2135 2135 /* Trim back any surplus on the last mblk */
2136 2136 if (spill >= 0) {
2137 2137 mp1->b_wptr -= spill;
2138 2138 *tail_unsent = spill;
2139 2139 } else {
2140 2140 /*
2141 2141 * We did not send everything we could in
2142 2142 * order to remain within the b_cont limit.
2143 2143 */
2144 2144 *usable -= spill;
2145 2145 *snxt += spill;
2146 2146 tcp->tcp_last_sent_len += spill;
2147 2147 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, spill);
2148 2148 /*
2149 2149 * Adjust the checksum
2150 2150 */
2151 2151 tcpha = (tcpha_t *)(rptr +
2152 2152 ixa->ixa_ip_hdr_length);
2153 2153 sum += spill;
2154 2154 sum = (sum >> 16) + (sum & 0xFFFF);
2155 2155 tcpha->tha_sum = htons(sum);
2156 2156 if (connp->conn_ipversion == IPV4_VERSION) {
2157 2157 sum = ntohs(
2158 2158 ((ipha_t *)rptr)->ipha_length) +
2159 2159 spill;
2160 2160 ((ipha_t *)rptr)->ipha_length =
2161 2161 htons(sum);
2162 2162 } else {
2163 2163 sum = ntohs(
2164 2164 ((ip6_t *)rptr)->ip6_plen) +
2165 2165 spill;
2166 2166 ((ip6_t *)rptr)->ip6_plen =
2167 2167 htons(sum);
2168 2168 }
2169 2169 ixa->ixa_pktlen += spill;
2170 2170 *tail_unsent = 0;
2171 2171 }
2172 2172 }
2173 2173 if (tcp->tcp_ip_forward_progress) {
2174 2174 tcp->tcp_ip_forward_progress = B_FALSE;
2175 2175 ixa->ixa_flags |= IXAF_REACH_CONF;
2176 2176 } else {
2177 2177 ixa->ixa_flags &= ~IXAF_REACH_CONF;
2178 2178 }
2179 2179
2180 2180 if (do_lso_send) {
2181 2181 /* Append LSO information to the mp. */
2182 2182 lso_info_set(mp, mss, HW_LSO);
2183 2183 ixa->ixa_fragsize = IP_MAXPACKET;
2184 2184 ixa->ixa_extra_ident = num_lso_seg - 1;
2185 2185
2186 2186 DTRACE_PROBE2(tcp_send_lso, int, num_lso_seg,
2187 2187 boolean_t, B_TRUE);
2188 2188
2189 2189 tcp_send_data(tcp, mp);
2190 2190
2191 2191 /*
2192 2192 * Restore values of ixa_fragsize and ixa_extra_ident.
2193 2193 */
2194 2194 ixa->ixa_fragsize = ixa->ixa_pmtu;
2195 2195 ixa->ixa_extra_ident = 0;
2196 2196 tcp->tcp_obsegs += num_lso_seg;
2197 2197 TCP_STAT(tcps, tcp_lso_times);
2198 2198 TCP_STAT_UPDATE(tcps, tcp_lso_pkt_out, num_lso_seg);
2199 2199 } else {
2200 2200 /*
2201 2201 * Make sure to clean up LSO information. Wherever a
2202 2202 * new mp uses the prepended header room after dupb(),
2203 2203 * lso_info_cleanup() should be called.
2204 2204 */
2205 2205 lso_info_cleanup(mp);
2206 2206 tcp_send_data(tcp, mp);
2207 2207 BUMP_LOCAL(tcp->tcp_obsegs);
2208 2208 }
2209 2209 }
2210 2210
2211 2211 return (0);
2212 2212 }
2213 2213
2214 2214 /*
2215 2215 * Initiate closedown sequence on an active connection. (May be called as
2216 2216 * writer.) Return value zero for OK return, non-zero for error return.
2217 2217 */
2218 2218 static int
2219 2219 tcp_xmit_end(tcp_t *tcp)
2220 2220 {
2221 2221 mblk_t *mp;
2222 2222 tcp_stack_t *tcps = tcp->tcp_tcps;
2223 2223 iulp_t uinfo;
2224 2224 ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
2225 2225 conn_t *connp = tcp->tcp_connp;
2226 2226
2227 2227 if (tcp->tcp_state < TCPS_SYN_RCVD ||
2228 2228 tcp->tcp_state > TCPS_CLOSE_WAIT) {
2229 2229 /*
2230 2230 * Invalid state, only states TCPS_SYN_RCVD,
2231 2231 * TCPS_ESTABLISHED and TCPS_CLOSE_WAIT are valid
2232 2232 */
2233 2233 return (-1);
2234 2234 }
2235 2235
2236 2236 tcp->tcp_fss = tcp->tcp_snxt + tcp->tcp_unsent;
2237 2237 tcp->tcp_valid_bits |= TCP_FSS_VALID;
2238 2238 /*
2239 2239 * If there is nothing more unsent, send the FIN now.
2240 2240 * Otherwise, it will go out with the last segment.
2241 2241 */
2242 2242 if (tcp->tcp_unsent == 0) {
2243 2243 mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
2244 2244 tcp->tcp_fss, B_FALSE, NULL, B_FALSE);
2245 2245
2246 2246 if (mp) {
2247 2247 tcp_send_data(tcp, mp);
2248 2248 } else {
2249 2249 /*
2250 2250 * Couldn't allocate msg. Pretend we got it out.
2251 2251 * Wait for rexmit timeout.
2252 2252 */
2253 2253 tcp->tcp_snxt = tcp->tcp_fss + 1;
2254 2254 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
2255 2255 }
2256 2256
2257 2257 /*
2258 2258 * If needed, update tcp_rexmit_snxt as tcp_snxt is
2259 2259 * changed.
2260 2260 */
2261 2261 if (tcp->tcp_rexmit && tcp->tcp_rexmit_nxt == tcp->tcp_fss) {
2262 2262 tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
2263 2263 }
2264 2264 } else {
2265 2265 /*
2266 2266 * If tcp->tcp_cork is set, then the data will not get sent,
2267 2267 * so we have to check that and unset it first.
2268 2268 */
2269 2269 if (tcp->tcp_cork)
2270 2270 tcp->tcp_cork = B_FALSE;
2271 2271 tcp_wput_data(tcp, NULL, B_FALSE);
2272 2272 }
2273 2273
2274 2274 /*
2275 2275 * If TCP does not get enough samples of RTT or tcp_rtt_updates
2276 2276 * is 0, don't update the cache.
|
↓ open down ↓ |
197 lines elided |
↑ open up ↑ |
2277 2277 */
2278 2278 if (tcps->tcps_rtt_updates == 0 ||
2279 2279 tcp->tcp_rtt_update < tcps->tcps_rtt_updates)
2280 2280 return (0);
2281 2281
2282 2282 /*
2283 2283 * We do not have a good algorithm to update ssthresh at this time.
2284 2284 * So don't do any update.
2285 2285 */
2286 2286 bzero(&uinfo, sizeof (uinfo));
2287 - uinfo.iulp_rtt = tcp->tcp_rtt_sa;
2288 - uinfo.iulp_rtt_sd = tcp->tcp_rtt_sd;
2287 + uinfo.iulp_rtt = NSEC2MSEC(tcp->tcp_rtt_sa);
2288 + uinfo.iulp_rtt_sd = NSEC2MSEC(tcp->tcp_rtt_sd);
2289 2289
2290 2290 /*
2291 2291 * Note that uinfo is kept for conn_faddr in the DCE. Could update even
2292 2292 * if source routed but we don't.
2293 2293 */
2294 2294 if (connp->conn_ipversion == IPV4_VERSION) {
2295 2295 if (connp->conn_faddr_v4 != tcp->tcp_ipha->ipha_dst) {
2296 2296 return (0);
2297 2297 }
2298 2298 (void) dce_update_uinfo_v4(connp->conn_faddr_v4, &uinfo, ipst);
2299 2299 } else {
2300 2300 uint_t ifindex;
2301 2301
2302 2302 if (!(IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6,
2303 2303 &tcp->tcp_ip6h->ip6_dst))) {
2304 2304 return (0);
2305 2305 }
2306 2306 ifindex = 0;
2307 2307 if (IN6_IS_ADDR_LINKSCOPE(&connp->conn_faddr_v6)) {
2308 2308 ip_xmit_attr_t *ixa = connp->conn_ixa;
2309 2309
2310 2310 /*
2311 2311 * If we are going to create a DCE we'd better have
2312 2312 * an ifindex
2313 2313 */
2314 2314 if (ixa->ixa_nce != NULL) {
2315 2315 ifindex = ixa->ixa_nce->nce_common->ncec_ill->
2316 2316 ill_phyint->phyint_ifindex;
2317 2317 } else {
2318 2318 return (0);
2319 2319 }
2320 2320 }
2321 2321
2322 2322 (void) dce_update_uinfo(&connp->conn_faddr_v6, ifindex, &uinfo,
2323 2323 ipst);
2324 2324 }
2325 2325 return (0);
2326 2326 }
2327 2327
2328 2328 /*
2329 2329 * Send out a control packet on the tcp connection specified. This routine
2330 2330 * is typically called where we need a simple ACK or RST generated.
2331 2331 */
2332 2332 void
2333 2333 tcp_xmit_ctl(char *str, tcp_t *tcp, uint32_t seq, uint32_t ack, int ctl)
2334 2334 {
2335 2335 uchar_t *rptr;
2336 2336 tcpha_t *tcpha;
2337 2337 ipha_t *ipha = NULL;
2338 2338 ip6_t *ip6h = NULL;
2339 2339 uint32_t sum;
2340 2340 int total_hdr_len;
2341 2341 int ip_hdr_len;
2342 2342 mblk_t *mp;
2343 2343 tcp_stack_t *tcps = tcp->tcp_tcps;
2344 2344 conn_t *connp = tcp->tcp_connp;
2345 2345 ip_xmit_attr_t *ixa = connp->conn_ixa;
2346 2346
2347 2347 /*
2348 2348 * Save sum for use in source route later.
2349 2349 */
2350 2350 sum = connp->conn_ht_ulp_len + connp->conn_sum;
2351 2351 total_hdr_len = connp->conn_ht_iphc_len;
2352 2352 ip_hdr_len = ixa->ixa_ip_hdr_length;
2353 2353
2354 2354 /* If a text string is passed in with the request, pass it to strlog. */
2355 2355 if (str != NULL && connp->conn_debug) {
2356 2356 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE,
2357 2357 "tcp_xmit_ctl: '%s', seq 0x%x, ack 0x%x, ctl 0x%x",
2358 2358 str, seq, ack, ctl);
2359 2359 }
2360 2360 mp = allocb(connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra,
2361 2361 BPRI_MED);
2362 2362 if (mp == NULL) {
2363 2363 return;
2364 2364 }
2365 2365 rptr = &mp->b_rptr[tcps->tcps_wroff_xtra];
2366 2366 mp->b_rptr = rptr;
2367 2367 mp->b_wptr = &rptr[total_hdr_len];
2368 2368 bcopy(connp->conn_ht_iphc, rptr, total_hdr_len);
2369 2369
2370 2370 ixa->ixa_pktlen = total_hdr_len;
2371 2371
2372 2372 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2373 2373 ipha = (ipha_t *)rptr;
2374 2374 ipha->ipha_length = htons(total_hdr_len);
2375 2375 } else {
2376 2376 ip6h = (ip6_t *)rptr;
2377 2377 ip6h->ip6_plen = htons(total_hdr_len - IPV6_HDR_LEN);
2378 2378 }
2379 2379 tcpha = (tcpha_t *)&rptr[ip_hdr_len];
2380 2380 tcpha->tha_flags = (uint8_t)ctl;
2381 2381 if (ctl & TH_RST) {
2382 2382 TCPS_BUMP_MIB(tcps, tcpOutRsts);
2383 2383 TCPS_BUMP_MIB(tcps, tcpOutControl);
2384 2384 /*
2385 2385 * Don't send TSopt w/ TH_RST packets per RFC 1323.
2386 2386 */
2387 2387 if (tcp->tcp_snd_ts_ok &&
2388 2388 tcp->tcp_state > TCPS_SYN_SENT) {
2389 2389 mp->b_wptr = &rptr[total_hdr_len - TCPOPT_REAL_TS_LEN];
2390 2390 *(mp->b_wptr) = TCPOPT_EOL;
2391 2391
2392 2392 ixa->ixa_pktlen = total_hdr_len - TCPOPT_REAL_TS_LEN;
2393 2393
2394 2394 if (connp->conn_ipversion == IPV4_VERSION) {
2395 2395 ipha->ipha_length = htons(total_hdr_len -
2396 2396 TCPOPT_REAL_TS_LEN);
2397 2397 } else {
2398 2398 ip6h->ip6_plen = htons(total_hdr_len -
2399 2399 IPV6_HDR_LEN - TCPOPT_REAL_TS_LEN);
2400 2400 }
2401 2401 tcpha->tha_offset_and_reserved -= (3 << 4);
2402 2402 sum -= TCPOPT_REAL_TS_LEN;
2403 2403 }
2404 2404 }
2405 2405 if (ctl & TH_ACK) {
2406 2406 if (tcp->tcp_snd_ts_ok) {
2407 2407 uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
2408 2408
2409 2409 U32_TO_BE32(llbolt,
2410 2410 (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
2411 2411 U32_TO_BE32(tcp->tcp_ts_recent,
2412 2412 (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
2413 2413 }
2414 2414
2415 2415 /* Update the latest receive window size in TCP header. */
2416 2416 tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2417 2417 /* Track what we sent to the peer */
2418 2418 tcp->tcp_tcpha->tha_win = tcpha->tha_win;
2419 2419 tcp->tcp_rack = ack;
2420 2420 tcp->tcp_rack_cnt = 0;
2421 2421 TCPS_BUMP_MIB(tcps, tcpOutAck);
2422 2422 }
2423 2423 BUMP_LOCAL(tcp->tcp_obsegs);
2424 2424 tcpha->tha_seq = htonl(seq);
2425 2425 tcpha->tha_ack = htonl(ack);
2426 2426 /*
2427 2427 * Include the adjustment for a source route if any.
2428 2428 */
2429 2429 sum = (sum >> 16) + (sum & 0xFFFF);
2430 2430 tcpha->tha_sum = htons(sum);
2431 2431 tcp_send_data(tcp, mp);
2432 2432 }
2433 2433
2434 2434 /*
2435 2435 * Generate a reset based on an inbound packet, connp is set by caller
2436 2436 * when RST is in response to an unexpected inbound packet for which
2437 2437 * there is active tcp state in the system.
2438 2438 *
2439 2439 * IPSEC NOTE : Try to send the reply with the same protection as it came
2440 2440 * in. We have the ip_recv_attr_t which is reversed to form the ip_xmit_attr_t.
2441 2441 * That way the packet will go out at the same level of protection as it
2442 2442 * came in with.
2443 2443 */
2444 2444 static void
2445 2445 tcp_xmit_early_reset(char *str, mblk_t *mp, uint32_t seq, uint32_t ack, int ctl,
2446 2446 ip_recv_attr_t *ira, ip_stack_t *ipst, conn_t *connp)
2447 2447 {
2448 2448 ipha_t *ipha = NULL;
2449 2449 ip6_t *ip6h = NULL;
2450 2450 ushort_t len;
2451 2451 tcpha_t *tcpha;
2452 2452 int i;
2453 2453 ipaddr_t v4addr;
2454 2454 in6_addr_t v6addr;
2455 2455 netstack_t *ns = ipst->ips_netstack;
2456 2456 tcp_stack_t *tcps = ns->netstack_tcp;
2457 2457 ip_xmit_attr_t ixas, *ixa;
2458 2458 uint_t ip_hdr_len = ira->ira_ip_hdr_length;
2459 2459 boolean_t need_refrele = B_FALSE; /* ixa_refrele(ixa) */
2460 2460 ushort_t port;
2461 2461
2462 2462 if (!tcp_send_rst_chk(tcps)) {
2463 2463 TCP_STAT(tcps, tcp_rst_unsent);
2464 2464 freemsg(mp);
2465 2465 return;
2466 2466 }
2467 2467
2468 2468 /*
2469 2469 * If connp != NULL we use conn_ixa to keep IP_NEXTHOP and other
2470 2470 * options from the listener. In that case the caller must ensure that
2471 2471 * we are running on the listener = connp squeue.
2472 2472 *
2473 2473 * We get a safe copy of conn_ixa so we don't need to restore anything
2474 2474 * we or ip_output_simple might change in the ixa.
2475 2475 */
2476 2476 if (connp != NULL) {
2477 2477 ASSERT(connp->conn_on_sqp);
2478 2478
2479 2479 ixa = conn_get_ixa_exclusive(connp);
2480 2480 if (ixa == NULL) {
2481 2481 TCP_STAT(tcps, tcp_rst_unsent);
2482 2482 freemsg(mp);
2483 2483 return;
2484 2484 }
2485 2485 need_refrele = B_TRUE;
2486 2486 } else {
2487 2487 bzero(&ixas, sizeof (ixas));
2488 2488 ixa = &ixas;
2489 2489 /*
2490 2490 * IXAF_VERIFY_SOURCE is overkill since we know the
2491 2491 * packet was for us.
2492 2492 */
2493 2493 ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE;
2494 2494 ixa->ixa_protocol = IPPROTO_TCP;
2495 2495 ixa->ixa_zoneid = ira->ira_zoneid;
2496 2496 ixa->ixa_ifindex = 0;
2497 2497 ixa->ixa_ipst = ipst;
2498 2498 ixa->ixa_cred = kcred;
2499 2499 ixa->ixa_cpid = NOPID;
2500 2500 }
2501 2501
2502 2502 if (str && tcps->tcps_dbg) {
2503 2503 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE,
2504 2504 "tcp_xmit_early_reset: '%s', seq 0x%x, ack 0x%x, "
2505 2505 "flags 0x%x",
2506 2506 str, seq, ack, ctl);
2507 2507 }
2508 2508 if (mp->b_datap->db_ref != 1) {
2509 2509 mblk_t *mp1 = copyb(mp);
2510 2510 freemsg(mp);
2511 2511 mp = mp1;
2512 2512 if (mp == NULL)
2513 2513 goto done;
2514 2514 } else if (mp->b_cont) {
2515 2515 freemsg(mp->b_cont);
2516 2516 mp->b_cont = NULL;
2517 2517 DB_CKSUMFLAGS(mp) = 0;
2518 2518 }
2519 2519 /*
2520 2520 * We skip reversing source route here.
2521 2521 * (for now we replace all IP options with EOL)
2522 2522 */
2523 2523 if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2524 2524 ipha = (ipha_t *)mp->b_rptr;
2525 2525 for (i = IP_SIMPLE_HDR_LENGTH; i < (int)ip_hdr_len; i++)
2526 2526 mp->b_rptr[i] = IPOPT_EOL;
2527 2527 /*
2528 2528 * Make sure that src address isn't flagrantly invalid.
2529 2529 * Not all broadcast address checking for the src address
2530 2530 * is possible, since we don't know the netmask of the src
2531 2531 * addr. No check for destination address is done, since
2532 2532 * IP will not pass up a packet with a broadcast dest
2533 2533 * address to TCP. Similar checks are done below for IPv6.
2534 2534 */
2535 2535 if (ipha->ipha_src == 0 || ipha->ipha_src == INADDR_BROADCAST ||
2536 2536 CLASSD(ipha->ipha_src)) {
2537 2537 BUMP_MIB(&ipst->ips_ip_mib, ipIfStatsInDiscards);
2538 2538 ip_drop_input("ipIfStatsInDiscards", mp, NULL);
2539 2539 freemsg(mp);
2540 2540 goto done;
2541 2541 }
2542 2542 } else {
2543 2543 ip6h = (ip6_t *)mp->b_rptr;
2544 2544
2545 2545 if (IN6_IS_ADDR_UNSPECIFIED(&ip6h->ip6_src) ||
2546 2546 IN6_IS_ADDR_MULTICAST(&ip6h->ip6_src)) {
2547 2547 BUMP_MIB(&ipst->ips_ip6_mib, ipIfStatsInDiscards);
2548 2548 ip_drop_input("ipIfStatsInDiscards", mp, NULL);
2549 2549 freemsg(mp);
2550 2550 goto done;
2551 2551 }
2552 2552
2553 2553 /* Remove any extension headers assuming partial overlay */
2554 2554 if (ip_hdr_len > IPV6_HDR_LEN) {
2555 2555 uint8_t *to;
2556 2556
2557 2557 to = mp->b_rptr + ip_hdr_len - IPV6_HDR_LEN;
2558 2558 ovbcopy(ip6h, to, IPV6_HDR_LEN);
2559 2559 mp->b_rptr += ip_hdr_len - IPV6_HDR_LEN;
2560 2560 ip_hdr_len = IPV6_HDR_LEN;
2561 2561 ip6h = (ip6_t *)mp->b_rptr;
2562 2562 ip6h->ip6_nxt = IPPROTO_TCP;
2563 2563 }
2564 2564 }
2565 2565 tcpha = (tcpha_t *)&mp->b_rptr[ip_hdr_len];
2566 2566 if (tcpha->tha_flags & TH_RST) {
2567 2567 freemsg(mp);
2568 2568 goto done;
2569 2569 }
2570 2570 tcpha->tha_offset_and_reserved = (5 << 4);
2571 2571 len = ip_hdr_len + sizeof (tcpha_t);
2572 2572 mp->b_wptr = &mp->b_rptr[len];
2573 2573 if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2574 2574 ipha->ipha_length = htons(len);
2575 2575 /* Swap addresses */
2576 2576 v4addr = ipha->ipha_src;
2577 2577 ipha->ipha_src = ipha->ipha_dst;
2578 2578 ipha->ipha_dst = v4addr;
2579 2579 ipha->ipha_ident = 0;
2580 2580 ipha->ipha_ttl = (uchar_t)tcps->tcps_ipv4_ttl;
2581 2581 ixa->ixa_flags |= IXAF_IS_IPV4;
2582 2582 ixa->ixa_ip_hdr_length = ip_hdr_len;
2583 2583 } else {
2584 2584 ip6h->ip6_plen = htons(len - IPV6_HDR_LEN);
2585 2585 /* Swap addresses */
2586 2586 v6addr = ip6h->ip6_src;
2587 2587 ip6h->ip6_src = ip6h->ip6_dst;
2588 2588 ip6h->ip6_dst = v6addr;
2589 2589 ip6h->ip6_hops = (uchar_t)tcps->tcps_ipv6_hoplimit;
2590 2590 ixa->ixa_flags &= ~IXAF_IS_IPV4;
2591 2591
2592 2592 if (IN6_IS_ADDR_LINKSCOPE(&ip6h->ip6_dst)) {
2593 2593 ixa->ixa_flags |= IXAF_SCOPEID_SET;
2594 2594 ixa->ixa_scopeid = ira->ira_ruifindex;
2595 2595 }
2596 2596 ixa->ixa_ip_hdr_length = IPV6_HDR_LEN;
2597 2597 }
2598 2598 ixa->ixa_pktlen = len;
2599 2599
2600 2600 /* Swap the ports */
2601 2601 port = tcpha->tha_fport;
2602 2602 tcpha->tha_fport = tcpha->tha_lport;
2603 2603 tcpha->tha_lport = port;
2604 2604
2605 2605 tcpha->tha_ack = htonl(ack);
2606 2606 tcpha->tha_seq = htonl(seq);
2607 2607 tcpha->tha_win = 0;
2608 2608 tcpha->tha_sum = htons(sizeof (tcpha_t));
2609 2609 tcpha->tha_flags = (uint8_t)ctl;
2610 2610 if (ctl & TH_RST) {
2611 2611 if (ctl & TH_ACK) {
2612 2612 /*
2613 2613 * Probe connection rejection here.
2614 2614 * tcp_xmit_listeners_reset() drops non-SYN segments
2615 2615 * that do not specify TH_ACK in their flags without
2616 2616 * calling this function. As a consequence, if this
2617 2617 * function is called with a TH_RST|TH_ACK ctl argument,
2618 2618 * it is being called in response to a SYN segment
2619 2619 * and thus the tcp:::accept-refused probe point
2620 2620 * is valid here.
2621 2621 */
2622 2622 DTRACE_TCP5(accept__refused, mblk_t *, NULL,
2623 2623 void, NULL, void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2624 2624 tcph_t *, tcpha);
2625 2625 }
2626 2626 TCPS_BUMP_MIB(tcps, tcpOutRsts);
2627 2627 TCPS_BUMP_MIB(tcps, tcpOutControl);
2628 2628 }
2629 2629
2630 2630 /* Discard any old label */
2631 2631 if (ixa->ixa_free_flags & IXA_FREE_TSL) {
2632 2632 ASSERT(ixa->ixa_tsl != NULL);
2633 2633 label_rele(ixa->ixa_tsl);
2634 2634 ixa->ixa_free_flags &= ~IXA_FREE_TSL;
2635 2635 }
2636 2636 ixa->ixa_tsl = ira->ira_tsl; /* Behave as a multi-level responder */
2637 2637
2638 2638 if (ira->ira_flags & IRAF_IPSEC_SECURE) {
2639 2639 /*
2640 2640 * Apply IPsec based on how IPsec was applied to
2641 2641 * the packet that caused the RST.
2642 2642 */
2643 2643 if (!ipsec_in_to_out(ira, ixa, mp, ipha, ip6h)) {
2644 2644 BUMP_MIB(&ipst->ips_ip_mib, ipIfStatsOutDiscards);
2645 2645 /* Note: mp already consumed and ip_drop_packet done */
2646 2646 goto done;
2647 2647 }
2648 2648 } else {
2649 2649 /*
2650 2650 * This is in clear. The RST message we are building
2651 2651 * here should go out in clear, independent of our policy.
2652 2652 */
2653 2653 ixa->ixa_flags |= IXAF_NO_IPSEC;
2654 2654 }
2655 2655
2656 2656 DTRACE_TCP5(send, mblk_t *, NULL, ip_xmit_attr_t *, ixa,
2657 2657 __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2658 2658 __dtrace_tcp_tcph_t *, tcpha);
2659 2659
2660 2660 /*
2661 2661 * NOTE: one might consider tracing a TCP packet here, but
2662 2662 * this function has no active TCP state and no tcp structure
2663 2663 * that has a trace buffer. If we traced here, we would have
2664 2664 * to keep a local trace buffer in tcp_record_trace().
2665 2665 */
2666 2666
2667 2667 (void) ip_output_simple(mp, ixa);
2668 2668 done:
2669 2669 ixa_cleanup(ixa);
2670 2670 if (need_refrele) {
2671 2671 ASSERT(ixa != &ixas);
2672 2672 ixa_refrele(ixa);
2673 2673 }
2674 2674 }
2675 2675
2676 2676 /*
2677 2677 * Generate a "no listener here" RST in response to an "unknown" segment.
2678 2678 * connp is set by caller when RST is in response to an unexpected
2679 2679 * inbound packet for which there is active tcp state in the system.
2680 2680 * Note that we are reusing the incoming mp to construct the outgoing RST.
2681 2681 */
2682 2682 void
2683 2683 tcp_xmit_listeners_reset(mblk_t *mp, ip_recv_attr_t *ira, ip_stack_t *ipst,
2684 2684 conn_t *connp)
2685 2685 {
2686 2686 uchar_t *rptr;
2687 2687 uint32_t seg_len;
2688 2688 tcpha_t *tcpha;
2689 2689 uint32_t seg_seq;
2690 2690 uint32_t seg_ack;
2691 2691 uint_t flags;
2692 2692 ipha_t *ipha;
2693 2693 ip6_t *ip6h;
2694 2694 boolean_t policy_present;
2695 2695 netstack_t *ns = ipst->ips_netstack;
2696 2696 tcp_stack_t *tcps = ns->netstack_tcp;
2697 2697 ipsec_stack_t *ipss = tcps->tcps_netstack->netstack_ipsec;
2698 2698 uint_t ip_hdr_len = ira->ira_ip_hdr_length;
2699 2699
2700 2700 TCP_STAT(tcps, tcp_no_listener);
2701 2701
2702 2702 /*
2703 2703 * DTrace this "unknown" segment as a tcp:::receive, as we did
2704 2704 * just receive something that was TCP.
2705 2705 */
2706 2706 DTRACE_TCP5(receive, mblk_t *, NULL, ip_xmit_attr_t *, NULL,
2707 2707 __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2708 2708 __dtrace_tcp_tcph_t *, &mp->b_rptr[ip_hdr_len]);
2709 2709
2710 2710 if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2711 2711 policy_present = ipss->ipsec_inbound_v4_policy_present;
2712 2712 ipha = (ipha_t *)mp->b_rptr;
2713 2713 ip6h = NULL;
2714 2714 } else {
2715 2715 policy_present = ipss->ipsec_inbound_v6_policy_present;
2716 2716 ipha = NULL;
2717 2717 ip6h = (ip6_t *)mp->b_rptr;
2718 2718 }
2719 2719
2720 2720 if (policy_present) {
2721 2721 /*
2722 2722 * The conn_t parameter is NULL because we already know
2723 2723 * nobody's home.
2724 2724 */
2725 2725 mp = ipsec_check_global_policy(mp, (conn_t *)NULL, ipha, ip6h,
2726 2726 ira, ns);
2727 2727 if (mp == NULL)
2728 2728 return;
2729 2729 }
2730 2730 if (is_system_labeled() && !tsol_can_reply_error(mp, ira)) {
2731 2731 DTRACE_PROBE2(
2732 2732 tx__ip__log__error__nolistener__tcp,
2733 2733 char *, "Could not reply with RST to mp(1)",
2734 2734 mblk_t *, mp);
2735 2735 ip2dbg(("tcp_xmit_listeners_reset: not permitted to reply\n"));
2736 2736 freemsg(mp);
2737 2737 return;
2738 2738 }
2739 2739
2740 2740 rptr = mp->b_rptr;
2741 2741
2742 2742 tcpha = (tcpha_t *)&rptr[ip_hdr_len];
2743 2743 seg_seq = ntohl(tcpha->tha_seq);
2744 2744 seg_ack = ntohl(tcpha->tha_ack);
2745 2745 flags = tcpha->tha_flags;
2746 2746
2747 2747 seg_len = msgdsize(mp) - (TCP_HDR_LENGTH(tcpha) + ip_hdr_len);
2748 2748 if (flags & TH_RST) {
2749 2749 freemsg(mp);
2750 2750 } else if (flags & TH_ACK) {
2751 2751 tcp_xmit_early_reset("no tcp, reset", mp, seg_ack, 0, TH_RST,
2752 2752 ira, ipst, connp);
2753 2753 } else {
2754 2754 if (flags & TH_SYN) {
2755 2755 seg_len++;
2756 2756 } else {
2757 2757 /*
2758 2758 * Here we violate the RFC. Note that a normal
2759 2759 * TCP will never send a segment without the ACK
2760 2760 * flag, except for RST or SYN segment. This
2761 2761 * segment is neither. Just drop it on the
2762 2762 * floor.
2763 2763 */
2764 2764 freemsg(mp);
2765 2765 TCP_STAT(tcps, tcp_rst_unsent);
2766 2766 return;
2767 2767 }
2768 2768
2769 2769 tcp_xmit_early_reset("no tcp, reset/ack", mp, 0,
2770 2770 seg_seq + seg_len, TH_RST | TH_ACK, ira, ipst, connp);
2771 2771 }
2772 2772 }
2773 2773
2774 2774 /*
2775 2775 * Helper function for tcp_xmit_mp() in handling connection set up flag
2776 2776 * options setting.
2777 2777 */
2778 2778 static void
2779 2779 tcp_xmit_mp_aux_iss(tcp_t *tcp, conn_t *connp, tcpha_t *tcpha, mblk_t *mp,
2780 2780 uint_t *flags)
2781 2781 {
2782 2782 uint32_t u1;
2783 2783 uint8_t *wptr = mp->b_wptr;
2784 2784 tcp_stack_t *tcps = tcp->tcp_tcps;
2785 2785 boolean_t add_sack = B_FALSE;
2786 2786
2787 2787 /*
2788 2788 * If TCP_ISS_VALID and the seq number is tcp_iss,
2789 2789 * TCP can only be in SYN-SENT, SYN-RCVD or
2790 2790 * FIN-WAIT-1 state. It can be FIN-WAIT-1 if
2791 2791 * our SYN is not ack'ed but the app closes this
2792 2792 * TCP connection.
2793 2793 */
2794 2794 ASSERT(tcp->tcp_state == TCPS_SYN_SENT ||
2795 2795 tcp->tcp_state == TCPS_SYN_RCVD ||
2796 2796 tcp->tcp_state == TCPS_FIN_WAIT_1);
2797 2797
2798 2798 /*
2799 2799 * Tack on the MSS option. It is always needed
2800 2800 * for both active and passive open.
2801 2801 *
2802 2802 * MSS option value should be interface MTU - MIN
2803 2803 * TCP/IP header according to RFC 793 as it means
2804 2804 * the maximum segment size TCP can receive. But
2805 2805 * to get around some broken middle boxes/end hosts
2806 2806 * out there, we allow the option value to be the
2807 2807 * same as the MSS option size on the peer side.
2808 2808 * In this way, the other side will not send
2809 2809 * anything larger than they can receive.
2810 2810 *
2811 2811 * Note that for SYN_SENT state, the ndd param
2812 2812 * tcp_use_smss_as_mss_opt has no effect as we
2813 2813 * don't know the peer's MSS option value. So
2814 2814 * the only case we need to take care of is in
2815 2815 * SYN_RCVD state, which is done later.
2816 2816 */
2817 2817 wptr[0] = TCPOPT_MAXSEG;
2818 2818 wptr[1] = TCPOPT_MAXSEG_LEN;
2819 2819 wptr += 2;
2820 2820 u1 = tcp->tcp_initial_pmtu - (connp->conn_ipversion == IPV4_VERSION ?
2821 2821 IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN) - TCP_MIN_HEADER_LENGTH;
2822 2822 U16_TO_BE16(u1, wptr);
2823 2823 wptr += 2;
2824 2824
2825 2825 /* Update the offset to cover the additional word */
2826 2826 tcpha->tha_offset_and_reserved += (1 << 4);
2827 2827
2828 2828 switch (tcp->tcp_state) {
2829 2829 case TCPS_SYN_SENT:
2830 2830 *flags = TH_SYN;
2831 2831
2832 2832 if (tcp->tcp_snd_sack_ok)
2833 2833 add_sack = B_TRUE;
2834 2834
2835 2835 if (tcp->tcp_snd_ts_ok) {
2836 2836 uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
2837 2837
2838 2838 if (add_sack) {
2839 2839 wptr[0] = TCPOPT_SACK_PERMITTED;
2840 2840 wptr[1] = TCPOPT_SACK_OK_LEN;
2841 2841 add_sack = B_FALSE;
2842 2842 } else {
2843 2843 wptr[0] = TCPOPT_NOP;
2844 2844 wptr[1] = TCPOPT_NOP;
2845 2845 }
2846 2846 wptr[2] = TCPOPT_TSTAMP;
2847 2847 wptr[3] = TCPOPT_TSTAMP_LEN;
2848 2848 wptr += 4;
2849 2849 U32_TO_BE32(llbolt, wptr);
2850 2850 wptr += 4;
2851 2851 ASSERT(tcp->tcp_ts_recent == 0);
2852 2852 U32_TO_BE32(0L, wptr);
2853 2853 wptr += 4;
2854 2854 tcpha->tha_offset_and_reserved += (3 << 4);
2855 2855 }
2856 2856
2857 2857 /*
2858 2858 * Set up all the bits to tell other side
2859 2859 * we are ECN capable.
2860 2860 */
2861 2861 if (tcp->tcp_ecn_ok)
2862 2862 *flags |= (TH_ECE | TH_CWR);
2863 2863
2864 2864 break;
2865 2865
2866 2866 case TCPS_SYN_RCVD:
2867 2867 *flags |= TH_SYN;
2868 2868
2869 2869 /*
2870 2870 * Reset the MSS option value to be SMSS
2871 2871 * We should probably add back the bytes
2872 2872 * for timestamp option and IPsec. We
2873 2873 * don't do that as this is a workaround
2874 2874 * for broken middle boxes/end hosts, it
2875 2875 * is better for us to be more cautious.
2876 2876 * They may not take these things into
2877 2877 * account in their SMSS calculation. Thus
2878 2878 * the peer's calculated SMSS may be smaller
2879 2879 * than what it can be. This should be OK.
2880 2880 */
2881 2881 if (tcps->tcps_use_smss_as_mss_opt) {
2882 2882 u1 = tcp->tcp_mss;
2883 2883 /*
2884 2884 * Note that wptr points just past the MSS
2885 2885 * option value.
2886 2886 */
2887 2887 U16_TO_BE16(u1, wptr - 2);
2888 2888 }
2889 2889
2890 2890 /*
2891 2891 * tcp_snd_ts_ok can only be set in TCPS_SYN_RCVD
2892 2892 * when the peer also uses timestamps option. And
2893 2893 * the TCP header template must have already been
2894 2894 * updated to include the timestamps option.
2895 2895 */
2896 2896 if (tcp->tcp_snd_sack_ok) {
2897 2897 if (tcp->tcp_snd_ts_ok) {
2898 2898 uint8_t *tmp_wptr;
2899 2899
2900 2900 /*
2901 2901 * Use the NOP in the header just
2902 2902 * before timestamps opton.
2903 2903 */
2904 2904 tmp_wptr = (uint8_t *)tcpha +
2905 2905 TCP_MIN_HEADER_LENGTH;
2906 2906 ASSERT(tmp_wptr[0] == TCPOPT_NOP &&
2907 2907 tmp_wptr[1] == TCPOPT_NOP);
2908 2908 tmp_wptr[0] = TCPOPT_SACK_PERMITTED;
2909 2909 tmp_wptr[1] = TCPOPT_SACK_OK_LEN;
2910 2910 } else {
2911 2911 add_sack = B_TRUE;
2912 2912 }
2913 2913 }
2914 2914
2915 2915
2916 2916 /*
2917 2917 * If the other side is ECN capable, reply
2918 2918 * that we are also ECN capable.
2919 2919 */
2920 2920 if (tcp->tcp_ecn_ok)
2921 2921 *flags |= TH_ECE;
2922 2922 break;
2923 2923
2924 2924 default:
2925 2925 /*
2926 2926 * The above ASSERT() makes sure that this
2927 2927 * must be FIN-WAIT-1 state. Our SYN has
2928 2928 * not been ack'ed so retransmit it.
2929 2929 */
2930 2930 *flags |= TH_SYN;
2931 2931 break;
2932 2932 }
2933 2933
2934 2934 if (add_sack) {
2935 2935 wptr[0] = TCPOPT_NOP;
2936 2936 wptr[1] = TCPOPT_NOP;
2937 2937 wptr[2] = TCPOPT_SACK_PERMITTED;
2938 2938 wptr[3] = TCPOPT_SACK_OK_LEN;
2939 2939 wptr += TCPOPT_REAL_SACK_OK_LEN;
2940 2940 tcpha->tha_offset_and_reserved += (1 << 4);
2941 2941 }
2942 2942
2943 2943 if (tcp->tcp_snd_ws_ok) {
2944 2944 wptr[0] = TCPOPT_NOP;
2945 2945 wptr[1] = TCPOPT_WSCALE;
2946 2946 wptr[2] = TCPOPT_WS_LEN;
2947 2947 wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
2948 2948 wptr += TCPOPT_REAL_WS_LEN;
2949 2949 tcpha->tha_offset_and_reserved += (1 << 4);
2950 2950 }
2951 2951
2952 2952 mp->b_wptr = wptr;
2953 2953 u1 = (int)(mp->b_wptr - mp->b_rptr);
2954 2954 /*
2955 2955 * Get IP set to checksum on our behalf
2956 2956 * Include the adjustment for a source route if any.
2957 2957 */
2958 2958 u1 += connp->conn_sum;
2959 2959 u1 = (u1 >> 16) + (u1 & 0xFFFF);
2960 2960 tcpha->tha_sum = htons(u1);
2961 2961 TCPS_BUMP_MIB(tcps, tcpOutControl);
2962 2962 }
2963 2963
2964 2964 /*
2965 2965 * Helper function for tcp_xmit_mp() in handling connection tear down
2966 2966 * flag setting and state changes.
2967 2967 */
2968 2968 static void
2969 2969 tcp_xmit_mp_aux_fss(tcp_t *tcp, ip_xmit_attr_t *ixa, uint_t *flags)
2970 2970 {
2971 2971 if (!tcp->tcp_fin_acked) {
2972 2972 *flags |= TH_FIN;
2973 2973 TCPS_BUMP_MIB(tcp->tcp_tcps, tcpOutControl);
2974 2974 }
2975 2975 if (!tcp->tcp_fin_sent) {
2976 2976 tcp->tcp_fin_sent = B_TRUE;
2977 2977 switch (tcp->tcp_state) {
2978 2978 case TCPS_SYN_RCVD:
2979 2979 tcp->tcp_state = TCPS_FIN_WAIT_1;
2980 2980 DTRACE_TCP6(state__change, void, NULL,
2981 2981 ip_xmit_attr_t *, ixa, void, NULL,
2982 2982 tcp_t *, tcp, void, NULL,
2983 2983 int32_t, TCPS_SYN_RCVD);
2984 2984 break;
2985 2985 case TCPS_ESTABLISHED:
2986 2986 tcp->tcp_state = TCPS_FIN_WAIT_1;
2987 2987 DTRACE_TCP6(state__change, void, NULL,
2988 2988 ip_xmit_attr_t *, ixa, void, NULL,
2989 2989 tcp_t *, tcp, void, NULL,
2990 2990 int32_t, TCPS_ESTABLISHED);
2991 2991 break;
2992 2992 case TCPS_CLOSE_WAIT:
2993 2993 tcp->tcp_state = TCPS_LAST_ACK;
2994 2994 DTRACE_TCP6(state__change, void, NULL,
2995 2995 ip_xmit_attr_t *, ixa, void, NULL,
2996 2996 tcp_t *, tcp, void, NULL,
2997 2997 int32_t, TCPS_CLOSE_WAIT);
2998 2998 break;
2999 2999 }
3000 3000 if (tcp->tcp_suna == tcp->tcp_snxt)
3001 3001 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3002 3002 tcp->tcp_snxt = tcp->tcp_fss + 1;
3003 3003 }
3004 3004 }
3005 3005
3006 3006 /*
3007 3007 * tcp_xmit_mp is called to return a pointer to an mblk chain complete with
3008 3008 * ip and tcp header ready to pass down to IP. If the mp passed in is
3009 3009 * non-NULL, then up to max_to_send bytes of data will be dup'ed off that
3010 3010 * mblk. (If sendall is not set the dup'ing will stop at an mblk boundary
3011 3011 * otherwise it will dup partial mblks.)
3012 3012 * Otherwise, an appropriate ACK packet will be generated. This
3013 3013 * routine is not usually called to send new data for the first time. It
3014 3014 * is mostly called out of the timer for retransmits, and to generate ACKs.
3015 3015 *
3016 3016 * If offset is not NULL, the returned mblk chain's first mblk's b_rptr will
3017 3017 * be adjusted by *offset. And after dupb(), the offset and the ending mblk
3018 3018 * of the original mblk chain will be returned in *offset and *end_mp.
3019 3019 */
3020 3020 mblk_t *
3021 3021 tcp_xmit_mp(tcp_t *tcp, mblk_t *mp, int32_t max_to_send, int32_t *offset,
3022 3022 mblk_t **end_mp, uint32_t seq, boolean_t sendall, uint32_t *seg_len,
3023 3023 boolean_t rexmit)
3024 3024 {
3025 3025 int data_length;
3026 3026 int32_t off = 0;
3027 3027 uint_t flags;
3028 3028 mblk_t *mp1;
3029 3029 mblk_t *mp2;
3030 3030 uchar_t *rptr;
3031 3031 tcpha_t *tcpha;
3032 3032 int32_t num_sack_blk = 0;
3033 3033 int32_t sack_opt_len = 0;
3034 3034 tcp_stack_t *tcps = tcp->tcp_tcps;
3035 3035 conn_t *connp = tcp->tcp_connp;
3036 3036 ip_xmit_attr_t *ixa = connp->conn_ixa;
3037 3037
3038 3038 /* Allocate for our maximum TCP header + link-level */
3039 3039 mp1 = allocb(connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra,
3040 3040 BPRI_MED);
3041 3041 if (mp1 == NULL)
3042 3042 return (NULL);
3043 3043 data_length = 0;
3044 3044
3045 3045 /*
3046 3046 * Note that tcp_mss has been adjusted to take into account the
3047 3047 * timestamp option if applicable. Because SACK options do not
3048 3048 * appear in every TCP segments and they are of variable lengths,
3049 3049 * they cannot be included in tcp_mss. Thus we need to calculate
3050 3050 * the actual segment length when we need to send a segment which
3051 3051 * includes SACK options.
3052 3052 */
3053 3053 if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
3054 3054 num_sack_blk = MIN(tcp->tcp_max_sack_blk,
3055 3055 tcp->tcp_num_sack_blk);
3056 3056 sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
3057 3057 TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
3058 3058 if (max_to_send + sack_opt_len > tcp->tcp_mss)
3059 3059 max_to_send -= sack_opt_len;
3060 3060 }
3061 3061
3062 3062 if (offset != NULL) {
3063 3063 off = *offset;
3064 3064 /* We use offset as an indicator that end_mp is not NULL. */
3065 3065 *end_mp = NULL;
3066 3066 }
3067 3067 for (mp2 = mp1; mp && data_length != max_to_send; mp = mp->b_cont) {
3068 3068 /* This could be faster with cooperation from downstream */
3069 3069 if (mp2 != mp1 && !sendall &&
3070 3070 data_length + (int)(mp->b_wptr - mp->b_rptr) >
3071 3071 max_to_send)
3072 3072 /*
3073 3073 * Don't send the next mblk since the whole mblk
3074 3074 * does not fit.
3075 3075 */
3076 3076 break;
3077 3077 mp2->b_cont = dupb(mp);
3078 3078 mp2 = mp2->b_cont;
3079 3079 if (!mp2) {
3080 3080 freemsg(mp1);
3081 3081 return (NULL);
3082 3082 }
3083 3083 mp2->b_rptr += off;
3084 3084 ASSERT((uintptr_t)(mp2->b_wptr - mp2->b_rptr) <=
3085 3085 (uintptr_t)INT_MAX);
3086 3086
3087 3087 data_length += (int)(mp2->b_wptr - mp2->b_rptr);
3088 3088 if (data_length > max_to_send) {
3089 3089 mp2->b_wptr -= data_length - max_to_send;
3090 3090 data_length = max_to_send;
3091 3091 off = mp2->b_wptr - mp->b_rptr;
3092 3092 break;
3093 3093 } else {
3094 3094 off = 0;
3095 3095 }
3096 3096 }
3097 3097 if (offset != NULL) {
3098 3098 *offset = off;
3099 3099 *end_mp = mp;
3100 3100 }
3101 3101 if (seg_len != NULL) {
3102 3102 *seg_len = data_length;
3103 3103 }
3104 3104
3105 3105 /* Update the latest receive window size in TCP header. */
3106 3106 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3107 3107
3108 3108 rptr = mp1->b_rptr + tcps->tcps_wroff_xtra;
3109 3109 mp1->b_rptr = rptr;
3110 3110 mp1->b_wptr = rptr + connp->conn_ht_iphc_len + sack_opt_len;
3111 3111 bcopy(connp->conn_ht_iphc, rptr, connp->conn_ht_iphc_len);
3112 3112 tcpha = (tcpha_t *)&rptr[ixa->ixa_ip_hdr_length];
3113 3113 tcpha->tha_seq = htonl(seq);
3114 3114
3115 3115 /*
3116 3116 * Use tcp_unsent to determine if the PUSH bit should be used assumes
3117 3117 * that this function was called from tcp_wput_data. Thus, when called
3118 3118 * to retransmit data the setting of the PUSH bit may appear some
3119 3119 * what random in that it might get set when it should not. This
3120 3120 * should not pose any performance issues.
3121 3121 */
3122 3122 if (data_length != 0 && (tcp->tcp_unsent == 0 ||
3123 3123 tcp->tcp_unsent == data_length)) {
3124 3124 flags = TH_ACK | TH_PUSH;
3125 3125 } else {
3126 3126 flags = TH_ACK;
3127 3127 }
3128 3128
3129 3129 if (tcp->tcp_ecn_ok) {
3130 3130 if (tcp->tcp_ecn_echo_on)
3131 3131 flags |= TH_ECE;
3132 3132
3133 3133 /*
3134 3134 * Only set ECT bit and ECN_CWR if a segment contains new data.
3135 3135 * There is no TCP flow control for non-data segments, and
3136 3136 * only data segment is transmitted reliably.
3137 3137 */
3138 3138 if (data_length > 0 && !rexmit) {
3139 3139 TCP_SET_ECT(tcp, rptr);
3140 3140 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
3141 3141 flags |= TH_CWR;
3142 3142 tcp->tcp_ecn_cwr_sent = B_TRUE;
3143 3143 }
3144 3144 }
3145 3145 }
3146 3146
3147 3147 /* Check if there is any special processing needs to be done. */
3148 3148 if (tcp->tcp_valid_bits) {
3149 3149 uint32_t u1;
3150 3150
3151 3151 /* We don't allow having SYN and FIN in the same segment... */
3152 3152 if ((tcp->tcp_valid_bits & TCP_ISS_VALID) &&
3153 3153 seq == tcp->tcp_iss) {
3154 3154 /* Need to do connection set up processing. */
3155 3155 tcp_xmit_mp_aux_iss(tcp, connp, tcpha, mp1, &flags);
3156 3156 } else if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
3157 3157 (seq + data_length) == tcp->tcp_fss) {
3158 3158 /* Need to do connection tear down processing. */
3159 3159 tcp_xmit_mp_aux_fss(tcp, ixa, &flags);
3160 3160 }
3161 3161
3162 3162 /*
3163 3163 * Need to do urgent pointer processing.
3164 3164 *
3165 3165 * Note the trick here. u1 is unsigned. When tcp_urg
3166 3166 * is smaller than seq, u1 will become a very huge value.
3167 3167 * So the comparison will fail. Also note that tcp_urp
3168 3168 * should be positive, see RFC 793 page 17.
3169 3169 */
3170 3170 u1 = tcp->tcp_urg - seq + TCP_OLD_URP_INTERPRETATION;
3171 3171 if ((tcp->tcp_valid_bits & TCP_URG_VALID) && u1 != 0 &&
3172 3172 u1 < (uint32_t)(64 * 1024)) {
3173 3173 flags |= TH_URG;
3174 3174 TCPS_BUMP_MIB(tcps, tcpOutUrg);
3175 3175 tcpha->tha_urp = htons(u1);
3176 3176 }
3177 3177 }
3178 3178 tcpha->tha_flags = (uchar_t)flags;
3179 3179 tcp->tcp_rack = tcp->tcp_rnxt;
3180 3180 tcp->tcp_rack_cnt = 0;
3181 3181
3182 3182 /* Fill in the current value of timestamps option. */
3183 3183 if (tcp->tcp_snd_ts_ok) {
3184 3184 if (tcp->tcp_state != TCPS_SYN_SENT) {
3185 3185 uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
3186 3186
3187 3187 U32_TO_BE32(llbolt,
3188 3188 (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
3189 3189 U32_TO_BE32(tcp->tcp_ts_recent,
3190 3190 (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
3191 3191 }
3192 3192 }
3193 3193
3194 3194 /* Fill in the SACK blocks. */
3195 3195 if (num_sack_blk > 0) {
3196 3196 uchar_t *wptr = (uchar_t *)tcpha + connp->conn_ht_ulp_len;
3197 3197 sack_blk_t *tmp;
3198 3198 int32_t i;
3199 3199
3200 3200 wptr[0] = TCPOPT_NOP;
3201 3201 wptr[1] = TCPOPT_NOP;
3202 3202 wptr[2] = TCPOPT_SACK;
3203 3203 wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
3204 3204 sizeof (sack_blk_t);
3205 3205 wptr += TCPOPT_REAL_SACK_LEN;
3206 3206
3207 3207 tmp = tcp->tcp_sack_list;
3208 3208 for (i = 0; i < num_sack_blk; i++) {
3209 3209 U32_TO_BE32(tmp[i].begin, wptr);
3210 3210 wptr += sizeof (tcp_seq);
3211 3211 U32_TO_BE32(tmp[i].end, wptr);
3212 3212 wptr += sizeof (tcp_seq);
3213 3213 }
3214 3214 tcpha->tha_offset_and_reserved += ((num_sack_blk * 2 + 1) << 4);
3215 3215 }
3216 3216 ASSERT((uintptr_t)(mp1->b_wptr - rptr) <= (uintptr_t)INT_MAX);
3217 3217 data_length += (int)(mp1->b_wptr - rptr);
3218 3218
3219 3219 ixa->ixa_pktlen = data_length;
3220 3220
3221 3221 if (ixa->ixa_flags & IXAF_IS_IPV4) {
3222 3222 ((ipha_t *)rptr)->ipha_length = htons(data_length);
3223 3223 } else {
3224 3224 ip6_t *ip6 = (ip6_t *)rptr;
3225 3225
3226 3226 ip6->ip6_plen = htons(data_length - IPV6_HDR_LEN);
3227 3227 }
3228 3228
3229 3229 /*
3230 3230 * Prime pump for IP
3231 3231 * Include the adjustment for a source route if any.
3232 3232 */
3233 3233 data_length -= ixa->ixa_ip_hdr_length;
3234 3234 data_length += connp->conn_sum;
3235 3235 data_length = (data_length >> 16) + (data_length & 0xFFFF);
3236 3236 tcpha->tha_sum = htons(data_length);
3237 3237 if (tcp->tcp_ip_forward_progress) {
3238 3238 tcp->tcp_ip_forward_progress = B_FALSE;
3239 3239 connp->conn_ixa->ixa_flags |= IXAF_REACH_CONF;
3240 3240 } else {
3241 3241 connp->conn_ixa->ixa_flags &= ~IXAF_REACH_CONF;
3242 3242 }
3243 3243 return (mp1);
3244 3244 }
3245 3245
3246 3246 /*
3247 3247 * If this routine returns B_TRUE, TCP can generate a RST in response
3248 3248 * to a segment. If it returns B_FALSE, TCP should not respond.
3249 3249 */
3250 3250 static boolean_t
3251 3251 tcp_send_rst_chk(tcp_stack_t *tcps)
3252 3252 {
3253 3253 int64_t now;
3254 3254
3255 3255 /*
3256 3256 * TCP needs to protect itself from generating too many RSTs.
3257 3257 * This can be a DoS attack by sending us random segments
3258 3258 * soliciting RSTs.
3259 3259 *
3260 3260 * What we do here is to have a limit of tcp_rst_sent_rate RSTs
3261 3261 * in each 1 second interval. In this way, TCP still generate
3262 3262 * RSTs in normal cases but when under attack, the impact is
3263 3263 * limited.
3264 3264 */
3265 3265 if (tcps->tcps_rst_sent_rate_enabled != 0) {
3266 3266 now = ddi_get_lbolt64();
3267 3267 if (TICK_TO_MSEC(now - tcps->tcps_last_rst_intrvl) >
3268 3268 1*SECONDS) {
3269 3269 tcps->tcps_last_rst_intrvl = now;
3270 3270 tcps->tcps_rst_cnt = 1;
3271 3271 } else if (++tcps->tcps_rst_cnt > tcps->tcps_rst_sent_rate) {
3272 3272 return (B_FALSE);
3273 3273 }
3274 3274 }
3275 3275 return (B_TRUE);
3276 3276 }
3277 3277
3278 3278 /*
3279 3279 * This function handles all retransmissions if SACK is enabled for this
3280 3280 * connection. First it calculates how many segments can be retransmitted
3281 3281 * based on tcp_pipe. Then it goes thru the notsack list to find eligible
3282 3282 * segments. A segment is eligible if sack_cnt for that segment is greater
3283 3283 * than or equal tcp_dupack_fast_retransmit. After it has retransmitted
3284 3284 * all eligible segments, it checks to see if TCP can send some new segments
3285 3285 * (fast recovery). If it can, set the appropriate flag for tcp_input_data().
3286 3286 *
3287 3287 * Parameters:
3288 3288 * tcp_t *tcp: the tcp structure of the connection.
3289 3289 * uint_t *flags: in return, appropriate value will be set for
3290 3290 * tcp_input_data().
3291 3291 */
3292 3292 void
3293 3293 tcp_sack_rexmit(tcp_t *tcp, uint_t *flags)
3294 3294 {
3295 3295 notsack_blk_t *notsack_blk;
3296 3296 int32_t usable_swnd;
3297 3297 int32_t mss;
3298 3298 uint32_t seg_len;
3299 3299 mblk_t *xmit_mp;
3300 3300 tcp_stack_t *tcps = tcp->tcp_tcps;
3301 3301
3302 3302 ASSERT(tcp->tcp_notsack_list != NULL);
3303 3303 ASSERT(tcp->tcp_rexmit == B_FALSE);
3304 3304
3305 3305 /* Defensive coding in case there is a bug... */
3306 3306 if (tcp->tcp_notsack_list == NULL) {
3307 3307 return;
3308 3308 }
3309 3309 notsack_blk = tcp->tcp_notsack_list;
3310 3310 mss = tcp->tcp_mss;
3311 3311
3312 3312 /*
3313 3313 * Limit the num of outstanding data in the network to be
3314 3314 * tcp_cwnd_ssthresh, which is half of the original congestion wnd.
3315 3315 */
3316 3316 usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
3317 3317
3318 3318 /* At least retransmit 1 MSS of data. */
3319 3319 if (usable_swnd <= 0) {
3320 3320 usable_swnd = mss;
3321 3321 }
3322 3322
3323 3323 /* Make sure no new RTT samples will be taken. */
3324 3324 tcp->tcp_csuna = tcp->tcp_snxt;
3325 3325
3326 3326 notsack_blk = tcp->tcp_notsack_list;
3327 3327 while (usable_swnd > 0) {
3328 3328 mblk_t *snxt_mp, *tmp_mp;
3329 3329 tcp_seq begin = tcp->tcp_sack_snxt;
3330 3330 tcp_seq end;
3331 3331 int32_t off;
3332 3332
3333 3333 for (; notsack_blk != NULL; notsack_blk = notsack_blk->next) {
3334 3334 if (SEQ_GT(notsack_blk->end, begin) &&
3335 3335 (notsack_blk->sack_cnt >=
3336 3336 tcps->tcps_dupack_fast_retransmit)) {
3337 3337 end = notsack_blk->end;
3338 3338 if (SEQ_LT(begin, notsack_blk->begin)) {
3339 3339 begin = notsack_blk->begin;
3340 3340 }
3341 3341 break;
3342 3342 }
3343 3343 }
3344 3344 /*
3345 3345 * All holes are filled. Manipulate tcp_cwnd to send more
3346 3346 * if we can. Note that after the SACK recovery, tcp_cwnd is
3347 3347 * set to tcp_cwnd_ssthresh.
3348 3348 */
3349 3349 if (notsack_blk == NULL) {
3350 3350 usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
3351 3351 if (usable_swnd <= 0 || tcp->tcp_unsent == 0) {
3352 3352 tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna;
3353 3353 ASSERT(tcp->tcp_cwnd > 0);
3354 3354 return;
3355 3355 } else {
3356 3356 usable_swnd = usable_swnd / mss;
3357 3357 tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna +
3358 3358 MAX(usable_swnd * mss, mss);
3359 3359 *flags |= TH_XMIT_NEEDED;
3360 3360 return;
3361 3361 }
3362 3362 }
3363 3363
3364 3364 /*
3365 3365 * Note that we may send more than usable_swnd allows here
3366 3366 * because of round off, but no more than 1 MSS of data.
3367 3367 */
3368 3368 seg_len = end - begin;
3369 3369 if (seg_len > mss)
3370 3370 seg_len = mss;
3371 3371 snxt_mp = tcp_get_seg_mp(tcp, begin, &off);
3372 3372 ASSERT(snxt_mp != NULL);
3373 3373 /* This should not happen. Defensive coding again... */
3374 3374 if (snxt_mp == NULL) {
3375 3375 return;
3376 3376 }
3377 3377
3378 3378 xmit_mp = tcp_xmit_mp(tcp, snxt_mp, seg_len, &off,
3379 3379 &tmp_mp, begin, B_TRUE, &seg_len, B_TRUE);
3380 3380 if (xmit_mp == NULL)
3381 3381 return;
|
↓ open down ↓ |
1083 lines elided |
↑ open up ↑ |
3382 3382
3383 3383 usable_swnd -= seg_len;
3384 3384 tcp->tcp_pipe += seg_len;
3385 3385 tcp->tcp_sack_snxt = begin + seg_len;
3386 3386
3387 3387 tcp_send_data(tcp, xmit_mp);
3388 3388
3389 3389 /*
3390 3390 * Update the send timestamp to avoid false retransmission.
3391 3391 */
3392 - snxt_mp->b_prev = (mblk_t *)ddi_get_lbolt();
3392 + snxt_mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
3393 3393
3394 3394 TCPS_BUMP_MIB(tcps, tcpRetransSegs);
3395 3395 TCPS_UPDATE_MIB(tcps, tcpRetransBytes, seg_len);
3396 3396 TCPS_BUMP_MIB(tcps, tcpOutSackRetransSegs);
3397 3397 /*
3398 3398 * Update tcp_rexmit_max to extend this SACK recovery phase.
3399 3399 * This happens when new data sent during fast recovery is
3400 3400 * also lost. If TCP retransmits those new data, it needs
3401 3401 * to extend SACK recover phase to avoid starting another
3402 3402 * fast retransmit/recovery unnecessarily.
3403 3403 */
3404 3404 if (SEQ_GT(tcp->tcp_sack_snxt, tcp->tcp_rexmit_max)) {
3405 3405 tcp->tcp_rexmit_max = tcp->tcp_sack_snxt;
3406 3406 }
3407 3407 }
3408 3408 }
3409 3409
3410 3410 /*
3411 3411 * tcp_ss_rexmit() is called to do slow start retransmission after a timeout
3412 3412 * or ICMP errors.
3413 3413 */
3414 3414 void
3415 3415 tcp_ss_rexmit(tcp_t *tcp)
3416 3416 {
3417 3417 uint32_t snxt;
3418 3418 uint32_t smax;
3419 3419 int32_t win;
3420 3420 int32_t mss;
3421 3421 int32_t off;
3422 3422 mblk_t *snxt_mp;
3423 3423 tcp_stack_t *tcps = tcp->tcp_tcps;
3424 3424
3425 3425 /*
3426 3426 * Note that tcp_rexmit can be set even though TCP has retransmitted
3427 3427 * all unack'ed segments.
3428 3428 */
3429 3429 if (SEQ_LT(tcp->tcp_rexmit_nxt, tcp->tcp_rexmit_max)) {
3430 3430 smax = tcp->tcp_rexmit_max;
3431 3431 snxt = tcp->tcp_rexmit_nxt;
3432 3432 if (SEQ_LT(snxt, tcp->tcp_suna)) {
3433 3433 snxt = tcp->tcp_suna;
3434 3434 }
3435 3435 win = MIN(tcp->tcp_cwnd, tcp->tcp_swnd);
3436 3436 win -= snxt - tcp->tcp_suna;
3437 3437 mss = tcp->tcp_mss;
3438 3438 snxt_mp = tcp_get_seg_mp(tcp, snxt, &off);
3439 3439
3440 3440 while (SEQ_LT(snxt, smax) && (win > 0) && (snxt_mp != NULL)) {
3441 3441 mblk_t *xmit_mp;
3442 3442 mblk_t *old_snxt_mp = snxt_mp;
3443 3443 uint32_t cnt = mss;
3444 3444
3445 3445 if (win < cnt) {
3446 3446 cnt = win;
3447 3447 }
3448 3448 if (SEQ_GT(snxt + cnt, smax)) {
3449 3449 cnt = smax - snxt;
3450 3450 }
3451 3451 xmit_mp = tcp_xmit_mp(tcp, snxt_mp, cnt, &off,
3452 3452 &snxt_mp, snxt, B_TRUE, &cnt, B_TRUE);
3453 3453 if (xmit_mp == NULL)
|
↓ open down ↓ |
51 lines elided |
↑ open up ↑ |
3454 3454 return;
3455 3455
3456 3456 tcp_send_data(tcp, xmit_mp);
3457 3457
3458 3458 snxt += cnt;
3459 3459 win -= cnt;
3460 3460 /*
3461 3461 * Update the send timestamp to avoid false
3462 3462 * retransmission.
3463 3463 */
3464 - old_snxt_mp->b_prev = (mblk_t *)ddi_get_lbolt();
3464 + old_snxt_mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
3465 3465 TCPS_BUMP_MIB(tcps, tcpRetransSegs);
3466 3466 TCPS_UPDATE_MIB(tcps, tcpRetransBytes, cnt);
3467 3467
3468 3468 tcp->tcp_rexmit_nxt = snxt;
3469 3469 }
3470 3470 /*
3471 3471 * If we have transmitted all we have at the time
3472 3472 * we started the retranmission, we can leave
3473 3473 * the rest of the job to tcp_wput_data(). But we
3474 3474 * need to check the send window first. If the
3475 3475 * win is not 0, go on with tcp_wput_data().
3476 3476 */
3477 3477 if (SEQ_LT(snxt, smax) || win == 0) {
3478 3478 return;
3479 3479 }
3480 3480 }
3481 3481 /* Only call tcp_wput_data() if there is data to be sent. */
3482 3482 if (tcp->tcp_unsent) {
3483 3483 tcp_wput_data(tcp, NULL, B_FALSE);
3484 3484 }
3485 3485 }
3486 3486
3487 3487 /*
3488 3488 * Do slow start retransmission after ICMP errors of PMTU changes.
3489 3489 */
3490 3490 void
3491 3491 tcp_rexmit_after_error(tcp_t *tcp)
3492 3492 {
3493 3493 /*
3494 3494 * All sent data has been acknowledged or no data left to send, just
3495 3495 * to return.
3496 3496 */
3497 3497 if (!SEQ_LT(tcp->tcp_suna, tcp->tcp_snxt) ||
3498 3498 (tcp->tcp_xmit_head == NULL))
3499 3499 return;
3500 3500
3501 3501 if ((tcp->tcp_valid_bits & TCP_FSS_VALID) && (tcp->tcp_unsent == 0))
3502 3502 tcp->tcp_rexmit_max = tcp->tcp_fss;
3503 3503 else
3504 3504 tcp->tcp_rexmit_max = tcp->tcp_snxt;
3505 3505
3506 3506 tcp->tcp_rexmit_nxt = tcp->tcp_suna;
3507 3507 tcp->tcp_rexmit = B_TRUE;
3508 3508 tcp->tcp_dupack_cnt = 0;
3509 3509 tcp_ss_rexmit(tcp);
3510 3510 }
3511 3511
3512 3512 /*
3513 3513 * tcp_get_seg_mp() is called to get the pointer to a segment in the
3514 3514 * send queue which starts at the given sequence number. If the given
3515 3515 * sequence number is equal to last valid sequence number (tcp_snxt), the
3516 3516 * returned mblk is the last valid mblk, and off is set to the length of
3517 3517 * that mblk.
3518 3518 *
3519 3519 * send queue which starts at the given seq. no.
3520 3520 *
3521 3521 * Parameters:
3522 3522 * tcp_t *tcp: the tcp instance pointer.
3523 3523 * uint32_t seq: the starting seq. no of the requested segment.
3524 3524 * int32_t *off: after the execution, *off will be the offset to
3525 3525 * the returned mblk which points to the requested seq no.
3526 3526 * It is the caller's responsibility to send in a non-null off.
3527 3527 *
3528 3528 * Return:
3529 3529 * A mblk_t pointer pointing to the requested segment in send queue.
3530 3530 */
3531 3531 static mblk_t *
3532 3532 tcp_get_seg_mp(tcp_t *tcp, uint32_t seq, int32_t *off)
3533 3533 {
3534 3534 int32_t cnt;
3535 3535 mblk_t *mp;
3536 3536
3537 3537 /* Defensive coding. Make sure we don't send incorrect data. */
3538 3538 if (SEQ_LT(seq, tcp->tcp_suna) || SEQ_GT(seq, tcp->tcp_snxt))
3539 3539 return (NULL);
3540 3540
3541 3541 cnt = seq - tcp->tcp_suna;
3542 3542 mp = tcp->tcp_xmit_head;
3543 3543 while (cnt > 0 && mp != NULL) {
3544 3544 cnt -= mp->b_wptr - mp->b_rptr;
3545 3545 if (cnt <= 0) {
3546 3546 cnt += mp->b_wptr - mp->b_rptr;
3547 3547 break;
3548 3548 }
3549 3549 mp = mp->b_cont;
3550 3550 }
3551 3551 ASSERT(mp != NULL);
3552 3552 *off = cnt;
3553 3553 return (mp);
3554 3554 }
3555 3555
3556 3556 /*
3557 3557 * This routine adjusts next-to-send sequence number variables, in the
3558 3558 * case where the reciever has shrunk it's window.
3559 3559 */
3560 3560 void
3561 3561 tcp_update_xmit_tail(tcp_t *tcp, uint32_t snxt)
3562 3562 {
3563 3563 mblk_t *xmit_tail;
3564 3564 int32_t offset;
3565 3565
3566 3566 tcp->tcp_snxt = snxt;
3567 3567
3568 3568 /* Get the mblk, and the offset in it, as per the shrunk window */
3569 3569 xmit_tail = tcp_get_seg_mp(tcp, snxt, &offset);
3570 3570 ASSERT(xmit_tail != NULL);
3571 3571 tcp->tcp_xmit_tail = xmit_tail;
3572 3572 tcp->tcp_xmit_tail_unsent = xmit_tail->b_wptr -
3573 3573 xmit_tail->b_rptr - offset;
3574 3574 }
3575 3575
3576 3576 /*
3577 3577 * This handles the case when the receiver has shrunk its win. Per RFC 1122
3578 3578 * if the receiver shrinks the window, i.e. moves the right window to the
3579 3579 * left, the we should not send new data, but should retransmit normally the
3580 3580 * old unacked data between suna and suna + swnd. We might has sent data
3581 3581 * that is now outside the new window, pretend that we didn't send it.
3582 3582 */
3583 3583 static void
3584 3584 tcp_process_shrunk_swnd(tcp_t *tcp, uint32_t shrunk_count)
3585 3585 {
3586 3586 uint32_t snxt = tcp->tcp_snxt;
3587 3587
3588 3588 ASSERT(shrunk_count > 0);
3589 3589
3590 3590 if (!tcp->tcp_is_wnd_shrnk) {
3591 3591 tcp->tcp_snxt_shrunk = snxt;
3592 3592 tcp->tcp_is_wnd_shrnk = B_TRUE;
3593 3593 } else if (SEQ_GT(snxt, tcp->tcp_snxt_shrunk)) {
3594 3594 tcp->tcp_snxt_shrunk = snxt;
3595 3595 }
3596 3596
3597 3597 /* Pretend we didn't send the data outside the window */
3598 3598 snxt -= shrunk_count;
3599 3599
3600 3600 /* Reset all the values per the now shrunk window */
3601 3601 tcp_update_xmit_tail(tcp, snxt);
3602 3602 tcp->tcp_unsent += shrunk_count;
3603 3603
3604 3604 /*
3605 3605 * If the SACK option is set, delete the entire list of
3606 3606 * notsack'ed blocks.
3607 3607 */
3608 3608 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
3609 3609
3610 3610 if (tcp->tcp_suna == tcp->tcp_snxt && tcp->tcp_swnd == 0)
3611 3611 /*
3612 3612 * Make sure the timer is running so that we will probe a zero
3613 3613 * window.
|
↓ open down ↓ |
139 lines elided |
↑ open up ↑ |
3614 3614 */
3615 3615 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3616 3616 }
3617 3617
3618 3618 /*
3619 3619 * tcp_fill_header is called by tcp_send() to fill the outgoing TCP header
3620 3620 * with the template header, as well as other options such as time-stamp,
3621 3621 * ECN and/or SACK.
3622 3622 */
3623 3623 static void
3624 -tcp_fill_header(tcp_t *tcp, uchar_t *rptr, clock_t now, int num_sack_blk)
3624 +tcp_fill_header(tcp_t *tcp, uchar_t *rptr, int num_sack_blk)
3625 3625 {
3626 3626 tcpha_t *tcp_tmpl, *tcpha;
3627 3627 uint32_t *dst, *src;
3628 3628 int hdrlen;
3629 3629 conn_t *connp = tcp->tcp_connp;
3630 3630
3631 3631 ASSERT(OK_32PTR(rptr));
3632 3632
3633 3633 /* Template header */
3634 3634 tcp_tmpl = tcp->tcp_tcpha;
3635 3635
|
↓ open down ↓ |
1 lines elided |
↑ open up ↑ |
3636 3636 /* Header of outgoing packet */
3637 3637 tcpha = (tcpha_t *)(rptr + connp->conn_ixa->ixa_ip_hdr_length);
3638 3638
3639 3639 /* dst and src are opaque 32-bit fields, used for copying */
3640 3640 dst = (uint32_t *)rptr;
3641 3641 src = (uint32_t *)connp->conn_ht_iphc;
3642 3642 hdrlen = connp->conn_ht_iphc_len;
3643 3643
3644 3644 /* Fill time-stamp option if needed */
3645 3645 if (tcp->tcp_snd_ts_ok) {
3646 - U32_TO_BE32((uint32_t)now,
3646 + U32_TO_BE32(LBOLT_FASTPATH,
3647 3647 (char *)tcp_tmpl + TCP_MIN_HEADER_LENGTH + 4);
3648 3648 U32_TO_BE32(tcp->tcp_ts_recent,
3649 3649 (char *)tcp_tmpl + TCP_MIN_HEADER_LENGTH + 8);
3650 3650 } else {
3651 3651 ASSERT(connp->conn_ht_ulp_len == TCP_MIN_HEADER_LENGTH);
3652 3652 }
3653 3653
3654 3654 /*
3655 3655 * Copy the template header; is this really more efficient than
3656 3656 * calling bcopy()? For simple IPv4/TCP, it may be the case,
3657 3657 * but perhaps not for other scenarios.
3658 3658 */
3659 3659 dst[0] = src[0];
3660 3660 dst[1] = src[1];
3661 3661 dst[2] = src[2];
3662 3662 dst[3] = src[3];
3663 3663 dst[4] = src[4];
3664 3664 dst[5] = src[5];
3665 3665 dst[6] = src[6];
3666 3666 dst[7] = src[7];
3667 3667 dst[8] = src[8];
3668 3668 dst[9] = src[9];
3669 3669 if (hdrlen -= 40) {
3670 3670 hdrlen >>= 2;
3671 3671 dst += 10;
3672 3672 src += 10;
3673 3673 do {
3674 3674 *dst++ = *src++;
3675 3675 } while (--hdrlen);
3676 3676 }
3677 3677
3678 3678 /*
3679 3679 * Set the ECN info in the TCP header if it is not a zero
3680 3680 * window probe. Zero window probe is only sent in
3681 3681 * tcp_wput_data() and tcp_timer().
3682 3682 */
3683 3683 if (tcp->tcp_ecn_ok && !tcp->tcp_zero_win_probe) {
3684 3684 TCP_SET_ECT(tcp, rptr);
3685 3685
3686 3686 if (tcp->tcp_ecn_echo_on)
3687 3687 tcpha->tha_flags |= TH_ECE;
3688 3688 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
3689 3689 tcpha->tha_flags |= TH_CWR;
3690 3690 tcp->tcp_ecn_cwr_sent = B_TRUE;
3691 3691 }
3692 3692 }
3693 3693
3694 3694 /* Fill in SACK options */
3695 3695 if (num_sack_blk > 0) {
3696 3696 uchar_t *wptr = rptr + connp->conn_ht_iphc_len;
3697 3697 sack_blk_t *tmp;
3698 3698 int32_t i;
3699 3699
3700 3700 wptr[0] = TCPOPT_NOP;
3701 3701 wptr[1] = TCPOPT_NOP;
3702 3702 wptr[2] = TCPOPT_SACK;
3703 3703 wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
3704 3704 sizeof (sack_blk_t);
3705 3705 wptr += TCPOPT_REAL_SACK_LEN;
3706 3706
3707 3707 tmp = tcp->tcp_sack_list;
3708 3708 for (i = 0; i < num_sack_blk; i++) {
3709 3709 U32_TO_BE32(tmp[i].begin, wptr);
3710 3710 wptr += sizeof (tcp_seq);
3711 3711 U32_TO_BE32(tmp[i].end, wptr);
3712 3712 wptr += sizeof (tcp_seq);
3713 3713 }
3714 3714 tcpha->tha_offset_and_reserved +=
3715 3715 ((num_sack_blk * 2 + 1) << 4);
3716 3716 }
3717 3717 }
|
↓ open down ↓ |
61 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX