xref: /titanic_44/usr/src/stand/lib/tcp/tcp.c (revision 8eea8e29cc4374d1ee24c25a07f45af132db3499)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  *
26  * tcp.c, Code implementing the TCP protocol.
27  */
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 #include <sys/types.h>
32 #include <socket_impl.h>
33 #include <socket_inet.h>
34 #include <sys/sysmacros.h>
35 #include <sys/promif.h>
36 #include <sys/socket.h>
37 #include <netinet/in_systm.h>
38 #include <netinet/in.h>
39 #include <netinet/ip.h>
40 #include <netinet/tcp.h>
41 #include <net/if_types.h>
42 #include <sys/salib.h>
43 
44 #include "ipv4.h"
45 #include "ipv4_impl.h"
46 #include "mac.h"
47 #include "mac_impl.h"
48 #include "v4_sum_impl.h"
49 #include <sys/bootdebug.h>
50 #include "tcp_inet.h"
51 #include "tcp_sack.h"
52 #include <inet/common.h>
53 #include <inet/mib2.h>
54 
55 /*
56  * We need to redefine BUMP_MIB/UPDATE_MIB to not have DTrace probes.
57  */
58 #undef BUMP_MIB
59 #define	BUMP_MIB(x) (x)++
60 
61 #undef UPDATE_MIB
62 #define	UPDATE_MIB(x, y) x += y
63 
64 /*
65  * MIB-2 stuff for SNMP
66  */
67 mib2_tcp_t	tcp_mib;	/* SNMP fixed size info */
68 
69 /* The TCP mib does not include the following errors. */
70 static uint_t tcp_cksum_errors;
71 static uint_t tcp_drops;
72 
73 /* Macros for timestamp comparisons */
74 #define	TSTMP_GEQ(a, b)	((int32_t)((a)-(b)) >= 0)
75 #define	TSTMP_LT(a, b)	((int32_t)((a)-(b)) < 0)
76 
77 /*
78  * Parameters for TCP Initial Send Sequence number (ISS) generation.
79  * The ISS is calculated by adding three components: a time component
80  * which grows by 1 every 4096 nanoseconds (versus every 4 microseconds
81  * suggested by RFC 793, page 27);
82  * a per-connection component which grows by 125000 for every new connection;
83  * and an "extra" component that grows by a random amount centered
84  * approximately on 64000.  This causes the the ISS generator to cycle every
85  * 4.89 hours if no TCP connections are made, and faster if connections are
86  * made.
87  */
88 #define	ISS_INCR	250000
89 #define	ISS_NSEC_SHT	0
90 
91 static uint32_t tcp_iss_incr_extra;	/* Incremented for each connection */
92 
93 #define	TCP_XMIT_LOWATER	4096
94 #define	TCP_XMIT_HIWATER	49152
95 #define	TCP_RECV_LOWATER	2048
96 #define	TCP_RECV_HIWATER	49152
97 
98 /*
99  *  PAWS needs a timer for 24 days.  This is the number of ms in 24 days
100  */
101 #define	PAWS_TIMEOUT	((uint32_t)(24*24*60*60*1000))
102 
103 /*
104  * TCP options struct returned from tcp_parse_options.
105  */
106 typedef struct tcp_opt_s {
107 	uint32_t	tcp_opt_mss;
108 	uint32_t	tcp_opt_wscale;
109 	uint32_t	tcp_opt_ts_val;
110 	uint32_t	tcp_opt_ts_ecr;
111 	tcp_t		*tcp;
112 } tcp_opt_t;
113 
114 /*
115  * RFC1323-recommended phrasing of TSTAMP option, for easier parsing
116  */
117 
118 #ifdef _BIG_ENDIAN
119 #define	TCPOPT_NOP_NOP_TSTAMP ((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
120 	(TCPOPT_TSTAMP << 8) | 10)
121 #else
122 #define	TCPOPT_NOP_NOP_TSTAMP ((10 << 24) | (TCPOPT_TSTAMP << 16) | \
123 	(TCPOPT_NOP << 8) | TCPOPT_NOP)
124 #endif
125 
126 /*
127  * Flags returned from tcp_parse_options.
128  */
129 #define	TCP_OPT_MSS_PRESENT	1
130 #define	TCP_OPT_WSCALE_PRESENT	2
131 #define	TCP_OPT_TSTAMP_PRESENT	4
132 #define	TCP_OPT_SACK_OK_PRESENT	8
133 #define	TCP_OPT_SACK_PRESENT	16
134 
135 /* TCP option length */
136 #define	TCPOPT_NOP_LEN		1
137 #define	TCPOPT_MAXSEG_LEN	4
138 #define	TCPOPT_WS_LEN		3
139 #define	TCPOPT_REAL_WS_LEN	(TCPOPT_WS_LEN+1)
140 #define	TCPOPT_TSTAMP_LEN	10
141 #define	TCPOPT_REAL_TS_LEN	(TCPOPT_TSTAMP_LEN+2)
142 #define	TCPOPT_SACK_OK_LEN	2
143 #define	TCPOPT_REAL_SACK_OK_LEN	(TCPOPT_SACK_OK_LEN+2)
144 #define	TCPOPT_REAL_SACK_LEN	4
145 #define	TCPOPT_MAX_SACK_LEN	36
146 #define	TCPOPT_HEADER_LEN	2
147 
148 /* TCP cwnd burst factor. */
149 #define	TCP_CWND_INFINITE	65535
150 #define	TCP_CWND_SS		3
151 #define	TCP_CWND_NORMAL		5
152 
153 /* Named Dispatch Parameter Management Structure */
154 typedef struct tcpparam_s {
155 	uint32_t	tcp_param_min;
156 	uint32_t	tcp_param_max;
157 	uint32_t	tcp_param_val;
158 	char		*tcp_param_name;
159 } tcpparam_t;
160 
161 /* Max size IP datagram is 64k - 1 */
162 #define	TCP_MSS_MAX_IPV4 (IP_MAXPACKET - (sizeof (struct ip) + \
163 	sizeof (tcph_t)))
164 
165 /* Max of the above */
166 #define	TCP_MSS_MAX	TCP_MSS_MAX_IPV4
167 
168 /* Largest TCP port number */
169 #define	TCP_MAX_PORT	(64 * 1024 - 1)
170 
171 /* Round up the value to the nearest mss. */
172 #define	MSS_ROUNDUP(value, mss)		((((value) - 1) / (mss) + 1) * (mss))
173 
174 #define	MS	1L
175 #define	SECONDS	(1000 * MS)
176 #define	MINUTES	(60 * SECONDS)
177 #define	HOURS	(60 * MINUTES)
178 #define	DAYS	(24 * HOURS)
179 
180 /* All NDD params in the core TCP became static variables. */
181 static int	tcp_time_wait_interval = 1 * MINUTES;
182 static int	tcp_conn_req_max_q = 128;
183 static int	tcp_conn_req_max_q0 = 1024;
184 static int	tcp_conn_req_min = 1;
185 static int	tcp_conn_grace_period = 0 * SECONDS;
186 static int	tcp_cwnd_max_ = 1024 * 1024;
187 static int	tcp_smallest_nonpriv_port = 1024;
188 static int	tcp_ip_abort_cinterval = 3 * MINUTES;
189 static int	tcp_ip_abort_linterval = 3 * MINUTES;
190 static int	tcp_ip_abort_interval = 8 * MINUTES;
191 static int	tcp_ip_notify_cinterval = 10 * SECONDS;
192 static int	tcp_ip_notify_interval = 10 * SECONDS;
193 static int	tcp_ipv4_ttl = 64;
194 static int	tcp_mss_def_ipv4 = 536;
195 static int	tcp_mss_max_ipv4 = TCP_MSS_MAX_IPV4;
196 static int	tcp_mss_min = 108;
197 static int	tcp_naglim_def = (4*1024)-1;
198 static int	tcp_rexmit_interval_initial = 3 * SECONDS;
199 static int	tcp_rexmit_interval_max = 60 * SECONDS;
200 static int	tcp_rexmit_interval_min = 400 * MS;
201 static int	tcp_dupack_fast_retransmit = 3;
202 static int	tcp_smallest_anon_port = 32 * 1024;
203 static int	tcp_largest_anon_port = TCP_MAX_PORT;
204 static int	tcp_xmit_lowat = TCP_XMIT_LOWATER;
205 static int	tcp_recv_hiwat_minmss = 4;
206 static int	tcp_fin_wait_2_flush_interval = 1 * MINUTES;
207 static int	tcp_max_buf = 1024 * 1024;
208 static int	tcp_wscale_always = 1;
209 static int	tcp_tstamp_always = 1;
210 static int	tcp_tstamp_if_wscale = 1;
211 static int	tcp_rexmit_interval_extra = 0;
212 static int	tcp_slow_start_after_idle = 2;
213 static int	tcp_slow_start_initial = 2;
214 static int	tcp_sack_permitted = 2;
215 static int	tcp_ecn_permitted = 2;
216 
217 /* Extra room to fit in headers. */
218 static uint_t	tcp_wroff_xtra;
219 
220 /* Hint for next port to try. */
221 static in_port_t	tcp_next_port_to_try = 32*1024;
222 
223 /*
224  * Figure out the value of window scale opton.  Note that the rwnd is
225  * ASSUMED to be rounded up to the nearest MSS before the calculation.
226  * We cannot find the scale value and then do a round up of tcp_rwnd
227  * because the scale value may not be correct after that.
228  */
229 #define	SET_WS_VALUE(tcp) \
230 { \
231 	int i; \
232 	uint32_t rwnd = (tcp)->tcp_rwnd; \
233 	for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT; \
234 	    i++, rwnd >>= 1) \
235 		; \
236 	(tcp)->tcp_rcv_ws = i; \
237 }
238 
239 /*
240  * Set ECN capable transport (ECT) code point in IP header.
241  *
242  * Note that there are 2 ECT code points '01' and '10', which are called
243  * ECT(1) and ECT(0) respectively.  Here we follow the original ECT code
244  * point ECT(0) for TCP as described in RFC 2481.
245  */
246 #define	SET_ECT(tcp, iph) \
247 	if ((tcp)->tcp_ipversion == IPV4_VERSION) { \
248 		/* We need to clear the code point first. */ \
249 		((struct ip *)(iph))->ip_tos &= 0xFC; \
250 		((struct ip *)(iph))->ip_tos |= IPH_ECN_ECT0; \
251 	}
252 
253 /*
254  * The format argument to pass to tcp_display().
255  * DISP_PORT_ONLY means that the returned string has only port info.
256  * DISP_ADDR_AND_PORT means that the returned string also contains the
257  * remote and local IP address.
258  */
259 #define	DISP_PORT_ONLY		1
260 #define	DISP_ADDR_AND_PORT	2
261 
262 /*
263  * TCP reassembly macros.  We hide starting and ending sequence numbers in
264  * b_next and b_prev of messages on the reassembly queue.  The messages are
265  * chained using b_cont.  These macros are used in tcp_reass() so we don't
266  * have to see the ugly casts and assignments.
267  */
268 #define	TCP_REASS_SEQ(mp)		((uint32_t)((mp)->b_next))
269 #define	TCP_REASS_SET_SEQ(mp, u)	((mp)->b_next = (mblk_t *)(u))
270 #define	TCP_REASS_END(mp)		((uint32_t)((mp)->b_prev))
271 #define	TCP_REASS_SET_END(mp, u)	((mp)->b_prev = (mblk_t *)(u))
272 
273 #define	TCP_TIMER_RESTART(tcp, intvl) \
274 	(tcp)->tcp_rto_timeout = prom_gettime() + intvl; \
275 	(tcp)->tcp_timer_running = B_TRUE;
276 
277 static int tcp_accept_comm(tcp_t *, tcp_t *, mblk_t *, uint_t);
278 static mblk_t *tcp_ack_mp(tcp_t *);
279 static in_port_t tcp_bindi(in_port_t, in_addr_t *, boolean_t, boolean_t);
280 static uint16_t tcp_cksum(uint16_t *, uint32_t);
281 static void tcp_clean_death(int, tcp_t *, int err);
282 static tcp_t *tcp_conn_request(tcp_t *, mblk_t *mp, uint_t, uint_t);
283 static char *tcp_display(tcp_t *, char *, char);
284 static int tcp_drain_input(tcp_t *, int, int);
285 static void tcp_drain_needed(int, tcp_t *);
286 static boolean_t tcp_drop_q0(tcp_t *);
287 static mblk_t *tcp_get_seg_mp(tcp_t *, uint32_t, int32_t *);
288 static int tcp_header_len(struct inetgram *);
289 static in_port_t tcp_report_ports(uint16_t *, enum Ports);
290 static int tcp_input(int);
291 static void tcp_iss_init(tcp_t *);
292 static tcp_t *tcp_lookup_ipv4(struct ip *, tcpha_t *, int, int *);
293 static tcp_t *tcp_lookup_listener_ipv4(in_addr_t, in_port_t, int *);
294 static int tcp_conn_check(tcp_t *);
295 static int tcp_close(int);
296 static void tcp_close_detached(tcp_t *);
297 static void tcp_eager_cleanup(tcp_t *, boolean_t, int);
298 static void tcp_eager_unlink(tcp_t *);
299 static void tcp_free(tcp_t *);
300 static int tcp_header_init_ipv4(tcp_t *);
301 static void tcp_mss_set(tcp_t *, uint32_t);
302 static int tcp_parse_options(tcph_t *, tcp_opt_t *);
303 static boolean_t tcp_paws_check(tcp_t *, tcph_t *, tcp_opt_t *);
304 static void tcp_process_options(tcp_t *, tcph_t *);
305 static int tcp_random(void);
306 static void tcp_random_init(void);
307 static mblk_t *tcp_reass(tcp_t *, mblk_t *, uint32_t);
308 static void tcp_reass_elim_overlap(tcp_t *, mblk_t *);
309 static void tcp_rcv_drain(int sock_id, tcp_t *);
310 static void tcp_rcv_enqueue(tcp_t *, mblk_t *, uint_t);
311 static void tcp_rput_data(tcp_t *, mblk_t *, int);
312 static int tcp_rwnd_set(tcp_t *, uint32_t);
313 static int32_t tcp_sack_rxmit(tcp_t *, int);
314 static void tcp_set_cksum(mblk_t *);
315 static void tcp_set_rto(tcp_t *, int32_t);
316 static void tcp_ss_rexmit(tcp_t *, int);
317 static int tcp_state_wait(int, tcp_t *, int);
318 static void tcp_timer(tcp_t *, int);
319 static void tcp_time_wait_append(tcp_t *);
320 static void tcp_time_wait_collector(void);
321 static void tcp_time_wait_processing(tcp_t *, mblk_t *, uint32_t,
322     uint32_t, int, tcph_t *, int sock_id);
323 static void tcp_time_wait_remove(tcp_t *);
324 static in_port_t tcp_update_next_port(in_port_t);
325 static int tcp_verify_cksum(mblk_t *);
326 static void tcp_wput_data(tcp_t *, mblk_t *, int);
327 static void tcp_xmit_ctl(char *, tcp_t *, mblk_t *, uint32_t, uint32_t,
328     int, uint_t, int);
329 static void tcp_xmit_early_reset(char *, int, mblk_t *, uint32_t, uint32_t,
330     int, uint_t);
331 static int tcp_xmit_end(tcp_t *, int);
332 static void tcp_xmit_listeners_reset(int, mblk_t *, uint_t);
333 static mblk_t *tcp_xmit_mp(tcp_t *, mblk_t *, int32_t, int32_t *,
334     mblk_t **, uint32_t, boolean_t, uint32_t *, boolean_t);
335 static int tcp_init_values(tcp_t *, struct inetboot_socket *);
336 
337 #if DEBUG > 1
338 #define	TCP_DUMP_PACKET(str, mp) \
339 { \
340 	int len = (mp)->b_wptr - (mp)->b_rptr; \
341 \
342 	printf("%s: dump TCP(%d): \n", (str), len); \
343 	hexdump((char *)(mp)->b_rptr, len); \
344 }
345 #else
346 #define	TCP_DUMP_PACKET(str, mp)
347 #endif
348 
349 #ifdef DEBUG
350 #define	DEBUG_1(str, arg)		printf(str, (arg))
351 #define	DEBUG_2(str, arg1, arg2)	printf(str, (arg1), (arg2))
352 #define	DEBUG_3(str, arg1, arg2, arg3)	printf(str, (arg1), (arg2), (arg3))
353 #else
354 #define	DEBUG_1(str, arg)
355 #define	DEBUG_2(str, arg1, arg2)
356 #define	DEBUG_3(str, arg1, arg2, arg3)
357 #endif
358 
359 /* Whether it is the first time TCP is used. */
360 static boolean_t tcp_initialized = B_FALSE;
361 
362 /* TCP time wait list. */
363 static tcp_t *tcp_time_wait_head;
364 static tcp_t *tcp_time_wait_tail;
365 static uint32_t tcp_cum_timewait;
366 /* When the tcp_time_wait_collector is run. */
367 static uint32_t tcp_time_wait_runtime;
368 
369 #define	TCP_RUN_TIME_WAIT_COLLECTOR() \
370 	if (prom_gettime() > tcp_time_wait_runtime) \
371 		tcp_time_wait_collector();
372 
373 /*
374  * Accept will return with an error if there is no connection coming in
375  * after this (in ms).
376  */
377 static int tcp_accept_timeout = 60000;
378 
379 /*
380  * Initialize the TCP-specific parts of a socket.
381  */
382 void
383 tcp_socket_init(struct inetboot_socket *isp)
384 {
385 	/* Do some initializations. */
386 	if (!tcp_initialized) {
387 		tcp_random_init();
388 		/* Extra head room for the MAC layer address. */
389 		if ((tcp_wroff_xtra = mac_get_hdr_len()) & 0x3) {
390 			tcp_wroff_xtra = (tcp_wroff_xtra & ~0x3) + 0x4;
391 		}
392 		/* Schedule the first time wait cleanup time */
393 		tcp_time_wait_runtime = prom_gettime() + tcp_time_wait_interval;
394 		tcp_initialized = B_TRUE;
395 	}
396 	TCP_RUN_TIME_WAIT_COLLECTOR();
397 
398 	isp->proto = IPPROTO_TCP;
399 	isp->input[TRANSPORT_LVL] = tcp_input;
400 	/* Socket layer should call tcp_send() directly. */
401 	isp->output[TRANSPORT_LVL] = NULL;
402 	isp->close[TRANSPORT_LVL] = tcp_close;
403 	isp->headerlen[TRANSPORT_LVL] = tcp_header_len;
404 	isp->ports = tcp_report_ports;
405 	if ((isp->pcb = bkmem_alloc(sizeof (tcp_t))) == NULL) {
406 		errno = ENOBUFS;
407 		return;
408 	}
409 	if ((errno = tcp_init_values((tcp_t *)isp->pcb, isp)) != 0) {
410 		bkmem_free(isp->pcb, sizeof (tcp_t));
411 		return;
412 	}
413 	/*
414 	 * This is set last because this field is used to determine if
415 	 * a socket is in use or not.
416 	 */
417 	isp->type = INETBOOT_STREAM;
418 }
419 
420 /*
421  * Return the size of a TCP header including TCP option.
422  */
423 static int
424 tcp_header_len(struct inetgram *igm)
425 {
426 	mblk_t *pkt;
427 	int ipvers;
428 
429 	/* Just returns the standard TCP header without option */
430 	if (igm == NULL)
431 		return (sizeof (tcph_t));
432 
433 	if ((pkt = igm->igm_mp) == NULL)
434 		return (0);
435 
436 	ipvers = ((struct ip *)pkt->b_rptr)->ip_v;
437 	if (ipvers == IPV4_VERSION) {
438 		return (TCP_HDR_LENGTH((tcph_t *)(pkt + IPH_HDR_LENGTH(pkt))));
439 	} else {
440 		dprintf("tcp_header_len: non-IPv4 packet.\n");
441 		return (0);
442 	}
443 }
444 
445 /*
446  * Return the requested port number in network order.
447  */
448 static in_port_t
449 tcp_report_ports(uint16_t *tcphp, enum Ports request)
450 {
451 	if (request == SOURCE)
452 		return (*(uint16_t *)(((tcph_t *)tcphp)->th_lport));
453 	return (*(uint16_t *)(((tcph_t *)tcphp)->th_fport));
454 }
455 
456 /*
457  * Because inetboot is not interrupt driven, TCP can only poll.  This
458  * means that there can be packets stuck in the NIC buffer waiting to
459  * be processed.  Thus we need to drain them before, for example, sending
460  * anything because an ACK may actually be stuck there.
461  *
462  * The timeout arguments determine how long we should wait for draining.
463  */
464 static int
465 tcp_drain_input(tcp_t *tcp, int sock_id, int timeout)
466 {
467 	struct inetgram *in_gram;
468 	struct inetgram *old_in_gram;
469 	int old_timeout;
470 	mblk_t *mp;
471 	int i;
472 
473 	dprintf("tcp_drain_input(%d): %s\n", sock_id,
474 	    tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
475 
476 	/*
477 	 * Since the driver uses the in_timeout value in the socket
478 	 * structure to determine the timeout value, we need to save
479 	 * the original one so that we can restore that after draining.
480 	 */
481 	old_timeout = sockets[sock_id].in_timeout;
482 	sockets[sock_id].in_timeout = timeout;
483 
484 	/*
485 	 * We do this because the input queue may have some user
486 	 * data already.
487 	 */
488 	old_in_gram = sockets[sock_id].inq;
489 	sockets[sock_id].inq = NULL;
490 
491 	/* Go out and check the wire */
492 	for (i = MEDIA_LVL; i < TRANSPORT_LVL; i++) {
493 		if (sockets[sock_id].input[i] != NULL) {
494 			if (sockets[sock_id].input[i](sock_id) < 0) {
495 				sockets[sock_id].in_timeout = old_timeout;
496 				if (sockets[sock_id].inq != NULL)
497 					nuke_grams(&sockets[sock_id].inq);
498 				sockets[sock_id].inq = old_in_gram;
499 				return (-1);
500 			}
501 		}
502 	}
503 #if DEBUG
504 	printf("tcp_drain_input: done with checking packets\n");
505 #endif
506 	while ((in_gram = sockets[sock_id].inq) != NULL) {
507 		/* Remove unknown inetgrams from the head of inq. */
508 		if (in_gram->igm_level != TRANSPORT_LVL) {
509 #if DEBUG
510 			printf("tcp_drain_input: unexpected packet "
511 			    "level %d frame found\n", in_gram->igm_level);
512 #endif
513 			del_gram(&sockets[sock_id].inq, in_gram, B_TRUE);
514 			continue;
515 		}
516 		mp = in_gram->igm_mp;
517 		del_gram(&sockets[sock_id].inq, in_gram, B_FALSE);
518 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
519 		tcp_rput_data(tcp, mp, sock_id);
520 		sockets[sock_id].in_timeout = old_timeout;
521 
522 		/*
523 		 * The other side may have closed this connection or
524 		 * RST us.  But we need to continue to process other
525 		 * packets in the socket's queue because they may be
526 		 * belong to another TCP connections.
527 		 */
528 		if (sockets[sock_id].pcb == NULL)
529 			tcp = NULL;
530 	}
531 
532 	if (tcp == NULL || sockets[sock_id].pcb == NULL) {
533 		if (sockets[sock_id].so_error != 0)
534 			return (-1);
535 		else
536 			return (0);
537 	}
538 #if DEBUG
539 	printf("tcp_drain_input: done with processing packets\n");
540 #endif
541 	sockets[sock_id].in_timeout = old_timeout;
542 	sockets[sock_id].inq = old_in_gram;
543 
544 	/*
545 	 * Data may have been received so indicate it is available
546 	 */
547 	tcp_drain_needed(sock_id, tcp);
548 	return (0);
549 }
550 
551 /*
552  * The receive entry point for upper layer to call to get data.  Note
553  * that this follows the current architecture that lower layer receive
554  * routines have been called already.  Thus if the inq of socket is
555  * not NULL, the packets must be for us.
556  */
557 static int
558 tcp_input(int sock_id)
559 {
560 	struct inetgram *in_gram;
561 	mblk_t *mp;
562 	tcp_t *tcp;
563 
564 	TCP_RUN_TIME_WAIT_COLLECTOR();
565 
566 	if ((tcp = sockets[sock_id].pcb) == NULL)
567 		return (-1);
568 
569 	while ((in_gram = sockets[sock_id].inq) != NULL) {
570 		/* Remove unknown inetgrams from the head of inq. */
571 		if (in_gram->igm_level != TRANSPORT_LVL) {
572 #ifdef DEBUG
573 			printf("tcp_input: unexpected packet "
574 			    "level %d frame found\n", in_gram->igm_level);
575 #endif
576 			del_gram(&sockets[sock_id].inq, in_gram, B_TRUE);
577 			continue;
578 		}
579 		mp = in_gram->igm_mp;
580 		del_gram(&sockets[sock_id].inq, in_gram, B_FALSE);
581 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
582 		tcp_rput_data(tcp, mp, sock_id);
583 		/* The TCP may be gone because it gets a RST. */
584 		if (sockets[sock_id].pcb == NULL)
585 			return (-1);
586 	}
587 
588 	/* Flush the receive list. */
589 	if (tcp->tcp_rcv_list != NULL) {
590 		tcp_rcv_drain(sock_id, tcp);
591 	} else {
592 		/* The other side has closed the connection, report this up. */
593 		if (tcp->tcp_state == TCPS_CLOSE_WAIT) {
594 			sockets[sock_id].so_state |= SS_CANTRCVMORE;
595 			return (0);
596 		}
597 	}
598 	return (0);
599 }
600 
601 /*
602  * The send entry point for upper layer to call to send data.  In order
603  * to minimize changes to the core TCP code, we need to put the
604  * data into mblks.
605  */
606 int
607 tcp_send(int sock_id, tcp_t *tcp, const void *msg, int len)
608 {
609 	mblk_t *mp;
610 	mblk_t *head = NULL;
611 	mblk_t *tail;
612 	int mss = tcp->tcp_mss;
613 	int cnt = 0;
614 	int win_size;
615 	char *buf = (char *)msg;
616 
617 	TCP_RUN_TIME_WAIT_COLLECTOR();
618 
619 	/* We don't want to append 0 size mblk. */
620 	if (len == 0)
621 		return (0);
622 	while (len > 0) {
623 		if (len < mss) {
624 			mss = len;
625 		}
626 		/*
627 		 * If we cannot allocate more buffer, stop here and
628 		 * the number of bytes buffered will be returned.
629 		 *
630 		 * Note that we follow the core TCP optimization that
631 		 * each mblk contains only MSS bytes data.
632 		 */
633 		if ((mp = allocb(mss + tcp->tcp_ip_hdr_len +
634 		    TCP_MAX_HDR_LENGTH + tcp_wroff_xtra, 0)) == NULL) {
635 			break;
636 		}
637 		mp->b_rptr += tcp->tcp_hdr_len + tcp_wroff_xtra;
638 		bcopy(buf, mp->b_rptr, mss);
639 		mp->b_wptr = mp->b_rptr + mss;
640 		buf += mss;
641 		cnt += mss;
642 		len -= mss;
643 
644 		if (head == NULL) {
645 			head = mp;
646 			tail = mp;
647 		} else {
648 			tail->b_cont = mp;
649 			tail = mp;
650 		}
651 	}
652 
653 	/*
654 	 * Since inetboot is not interrupt driven, there may be
655 	 * some ACKs in the MAC's buffer.  Drain them first,
656 	 * otherwise, we may not be able to send.
657 	 *
658 	 * We expect an ACK in two cases:
659 	 *
660 	 * 1) We have un-ACK'ed data.
661 	 *
662 	 * 2) All ACK's have been received and the sender's window has been
663 	 * closed.  We need an ACK back to open the window so that we can
664 	 * send.  In this case, call tcp_drain_input() if the window size is
665 	 * less than 2 * MSS.
666 	 */
667 
668 	/* window size = MIN(swnd, cwnd) - unacked bytes */
669 	win_size = (tcp->tcp_swnd > tcp->tcp_cwnd) ? tcp->tcp_cwnd :
670 		tcp->tcp_swnd;
671 	win_size -= tcp->tcp_snxt;
672 	win_size += tcp->tcp_suna;
673 	if (win_size < (2 * tcp->tcp_mss))
674 		if (tcp_drain_input(tcp, sock_id, 5) < 0)
675 			return (-1);
676 
677 	tcp_wput_data(tcp, head, sock_id);
678 	return (cnt);
679 }
680 
681 /* Free up all TCP related stuff */
682 static void
683 tcp_free(tcp_t *tcp)
684 {
685 	if (tcp->tcp_iphc != NULL) {
686 		bkmem_free((caddr_t)tcp->tcp_iphc, tcp->tcp_iphc_len);
687 		tcp->tcp_iphc = NULL;
688 	}
689 	if (tcp->tcp_xmit_head != NULL) {
690 		freemsg(tcp->tcp_xmit_head);
691 		tcp->tcp_xmit_head = NULL;
692 	}
693 	if (tcp->tcp_rcv_list != NULL) {
694 		freemsg(tcp->tcp_rcv_list);
695 		tcp->tcp_rcv_list = NULL;
696 	}
697 	if (tcp->tcp_reass_head != NULL) {
698 		freemsg(tcp->tcp_reass_head);
699 		tcp->tcp_reass_head = NULL;
700 	}
701 	if (tcp->tcp_sack_info != NULL) {
702 		bkmem_free((caddr_t)tcp->tcp_sack_info,
703 		    sizeof (tcp_sack_info_t));
704 		tcp->tcp_sack_info = NULL;
705 	}
706 }
707 
708 static void
709 tcp_close_detached(tcp_t *tcp)
710 {
711 	if (tcp->tcp_listener != NULL)
712 		tcp_eager_unlink(tcp);
713 	tcp_free(tcp);
714 	bkmem_free((caddr_t)tcp, sizeof (tcp_t));
715 }
716 
717 /*
718  * If we are an eager connection hanging off a listener that hasn't
719  * formally accepted the connection yet, get off his list and blow off
720  * any data that we have accumulated.
721  */
722 static void
723 tcp_eager_unlink(tcp_t *tcp)
724 {
725 	tcp_t	*listener = tcp->tcp_listener;
726 
727 	assert(listener != NULL);
728 	if (tcp->tcp_eager_next_q0 != NULL) {
729 		assert(tcp->tcp_eager_prev_q0 != NULL);
730 
731 		/* Remove the eager tcp from q0 */
732 		tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
733 		    tcp->tcp_eager_prev_q0;
734 		tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
735 		    tcp->tcp_eager_next_q0;
736 		listener->tcp_conn_req_cnt_q0--;
737 	} else {
738 		tcp_t   **tcpp = &listener->tcp_eager_next_q;
739 		tcp_t	*prev = NULL;
740 
741 		for (; tcpp[0]; tcpp = &tcpp[0]->tcp_eager_next_q) {
742 			if (tcpp[0] == tcp) {
743 				if (listener->tcp_eager_last_q == tcp) {
744 					/*
745 					 * If we are unlinking the last
746 					 * element on the list, adjust
747 					 * tail pointer. Set tail pointer
748 					 * to nil when list is empty.
749 					 */
750 					assert(tcp->tcp_eager_next_q == NULL);
751 					if (listener->tcp_eager_last_q ==
752 					    listener->tcp_eager_next_q) {
753 						listener->tcp_eager_last_q =
754 						NULL;
755 					} else {
756 						/*
757 						 * We won't get here if there
758 						 * is only one eager in the
759 						 * list.
760 						 */
761 						assert(prev != NULL);
762 						listener->tcp_eager_last_q =
763 						    prev;
764 					}
765 				}
766 				tcpp[0] = tcp->tcp_eager_next_q;
767 				tcp->tcp_eager_next_q = NULL;
768 				tcp->tcp_eager_last_q = NULL;
769 				listener->tcp_conn_req_cnt_q--;
770 				break;
771 			}
772 			prev = tcpp[0];
773 		}
774 	}
775 	tcp->tcp_listener = NULL;
776 }
777 
778 /*
779  * Reset any eager connection hanging off this listener
780  * and then reclaim it's resources.
781  */
782 static void
783 tcp_eager_cleanup(tcp_t *listener, boolean_t q0_only, int sock_id)
784 {
785 	tcp_t	*eager;
786 
787 	if (!q0_only) {
788 		/* First cleanup q */
789 		while ((eager = listener->tcp_eager_next_q) != NULL) {
790 			assert(listener->tcp_eager_last_q != NULL);
791 			tcp_xmit_ctl("tcp_eager_cleanup, can't wait",
792 			    eager, NULL, eager->tcp_snxt, 0, TH_RST, 0,
793 			    sock_id);
794 			tcp_close_detached(eager);
795 		}
796 		assert(listener->tcp_eager_last_q == NULL);
797 	}
798 	/* Then cleanup q0 */
799 	while ((eager = listener->tcp_eager_next_q0) != listener) {
800 		tcp_xmit_ctl("tcp_eager_cleanup, can't wait",
801 		    eager, NULL, eager->tcp_snxt, 0, TH_RST, 0, sock_id);
802 		tcp_close_detached(eager);
803 	}
804 }
805 
806 /*
807  * To handle the shutdown request. Called from shutdown()
808  */
809 int
810 tcp_shutdown(int sock_id)
811 {
812 	tcp_t	*tcp;
813 
814 	DEBUG_1("tcp_shutdown: sock_id %x\n", sock_id);
815 
816 	if ((tcp = sockets[sock_id].pcb) == NULL) {
817 		return (-1);
818 	}
819 
820 	/*
821 	 * Since inetboot is not interrupt driven, there may be
822 	 * some ACKs in the MAC's buffer.  Drain them first,
823 	 * otherwise, we may not be able to send.
824 	 */
825 	if (tcp_drain_input(tcp, sock_id, 5) < 0) {
826 		/*
827 		 * If we return now without freeing TCP, there will be
828 		 * a memory leak.
829 		 */
830 		if (sockets[sock_id].pcb != NULL)
831 			tcp_clean_death(sock_id, tcp, 0);
832 		return (-1);
833 	}
834 
835 	DEBUG_1("tcp_shutdown: tcp_state %x\n", tcp->tcp_state);
836 	switch (tcp->tcp_state) {
837 
838 	case TCPS_SYN_RCVD:
839 		/*
840 		 * Shutdown during the connect 3-way handshake
841 		 */
842 	case TCPS_ESTABLISHED:
843 		/*
844 		 * Transmit the FIN
845 		 * wait for the FIN to be ACKed,
846 		 * then remain in FIN_WAIT_2
847 		 */
848 		dprintf("tcp_shutdown: sending fin\n");
849 		if (tcp_xmit_end(tcp, sock_id) == 0 &&
850 			tcp_state_wait(sock_id, tcp, TCPS_FIN_WAIT_2) < 0) {
851 			/* During the wait, TCP may be gone... */
852 			if (sockets[sock_id].pcb == NULL)
853 				return (-1);
854 		}
855 		dprintf("tcp_shutdown: done\n");
856 		break;
857 
858 	default:
859 		break;
860 
861 	}
862 	return (0);
863 }
864 
865 /* To handle closing of the socket */
866 static int
867 tcp_close(int sock_id)
868 {
869 	char	*msg;
870 	tcp_t	*tcp;
871 	int	error = 0;
872 
873 	if ((tcp = sockets[sock_id].pcb) == NULL) {
874 		return (-1);
875 	}
876 
877 	TCP_RUN_TIME_WAIT_COLLECTOR();
878 
879 	/*
880 	 * Since inetboot is not interrupt driven, there may be
881 	 * some ACKs in the MAC's buffer.  Drain them first,
882 	 * otherwise, we may not be able to send.
883 	 */
884 	if (tcp_drain_input(tcp, sock_id, 5) < 0) {
885 		/*
886 		 * If we return now without freeing TCP, there will be
887 		 * a memory leak.
888 		 */
889 		if (sockets[sock_id].pcb != NULL)
890 			tcp_clean_death(sock_id, tcp, 0);
891 		return (-1);
892 	}
893 
894 	if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
895 		/* Cleanup for listener */
896 		tcp_eager_cleanup(tcp, 0, sock_id);
897 	}
898 
899 	msg = NULL;
900 	switch (tcp->tcp_state) {
901 	case TCPS_CLOSED:
902 	case TCPS_IDLE:
903 	case TCPS_BOUND:
904 	case TCPS_LISTEN:
905 		break;
906 	case TCPS_SYN_SENT:
907 		msg = "tcp_close, during connect";
908 		break;
909 	case TCPS_SYN_RCVD:
910 		/*
911 		 * Close during the connect 3-way handshake
912 		 * but here there may or may not be pending data
913 		 * already on queue. Process almost same as in
914 		 * the ESTABLISHED state.
915 		 */
916 		/* FALLTHRU */
917 	default:
918 		/*
919 		 * If SO_LINGER has set a zero linger time, abort the
920 		 * connection with a reset.
921 		 */
922 		if (tcp->tcp_linger && tcp->tcp_lingertime == 0) {
923 			msg = "tcp_close, zero lingertime";
924 			break;
925 		}
926 
927 		/*
928 		 * Abort connection if there is unread data queued.
929 		 */
930 		if (tcp->tcp_rcv_list != NULL ||
931 				tcp->tcp_reass_head != NULL) {
932 			msg = "tcp_close, unread data";
933 			break;
934 		}
935 		if (tcp->tcp_state <= TCPS_LISTEN)
936 			break;
937 
938 		/*
939 		 * Transmit the FIN before detaching the tcp_t.
940 		 * After tcp_detach returns this queue/perimeter
941 		 * no longer owns the tcp_t thus others can modify it.
942 		 * The TCP could be closed in tcp_state_wait called by
943 		 * tcp_wput_data called by tcp_xmit_end.
944 		 */
945 		(void) tcp_xmit_end(tcp, sock_id);
946 		if (sockets[sock_id].pcb == NULL)
947 			return (0);
948 
949 		/*
950 		 * If lingering on close then wait until the fin is acked,
951 		 * the SO_LINGER time passes, or a reset is sent/received.
952 		 */
953 		if (tcp->tcp_linger && tcp->tcp_lingertime > 0 &&
954 		    !(tcp->tcp_fin_acked) &&
955 		    tcp->tcp_state >= TCPS_ESTABLISHED) {
956 			uint32_t stoptime; /* in ms */
957 
958 			tcp->tcp_client_errno = 0;
959 			stoptime = prom_gettime() +
960 			    (tcp->tcp_lingertime * 1000);
961 			while (!(tcp->tcp_fin_acked) &&
962 			    tcp->tcp_state >= TCPS_ESTABLISHED &&
963 			    tcp->tcp_client_errno == 0 &&
964 			    ((int32_t)(stoptime - prom_gettime()) > 0)) {
965 				if (tcp_drain_input(tcp, sock_id, 5) < 0) {
966 					if (sockets[sock_id].pcb != NULL) {
967 						tcp_clean_death(sock_id,
968 						    tcp, 0);
969 					}
970 					return (-1);
971 				}
972 			}
973 			tcp->tcp_client_errno = 0;
974 		}
975 		if (tcp_state_wait(sock_id, tcp, TCPS_TIME_WAIT) < 0) {
976 			/* During the wait, TCP may be gone... */
977 			if (sockets[sock_id].pcb == NULL)
978 				return (0);
979 			msg = "tcp_close, couldn't detach";
980 		} else {
981 			return (0);
982 		}
983 		break;
984 	}
985 
986 	/* Something went wrong...  Send a RST and report the error */
987 	if (msg != NULL) {
988 		if (tcp->tcp_state == TCPS_ESTABLISHED ||
989 		    tcp->tcp_state == TCPS_CLOSE_WAIT)
990 			BUMP_MIB(tcp_mib.tcpEstabResets);
991 		if (tcp->tcp_state == TCPS_SYN_SENT ||
992 		    tcp->tcp_state == TCPS_SYN_RCVD)
993 			BUMP_MIB(tcp_mib.tcpAttemptFails);
994 		tcp_xmit_ctl(msg, tcp, NULL, tcp->tcp_snxt, 0, TH_RST, 0,
995 		    sock_id);
996 	}
997 
998 	tcp_free(tcp);
999 	bkmem_free((caddr_t)tcp, sizeof (tcp_t));
1000 	sockets[sock_id].pcb = NULL;
1001 	return (error);
1002 }
1003 
1004 /* To make an endpoint a listener. */
1005 int
1006 tcp_listen(int sock_id, int backlog)
1007 {
1008 	tcp_t *tcp;
1009 
1010 	if ((tcp = (tcp_t *)(sockets[sock_id].pcb)) == NULL) {
1011 		errno = EINVAL;
1012 		return (-1);
1013 	}
1014 	/* We allow calling listen() multiple times to change the backlog. */
1015 	if (tcp->tcp_state > TCPS_LISTEN || tcp->tcp_state < TCPS_BOUND) {
1016 		errno = EOPNOTSUPP;
1017 		return (-1);
1018 	}
1019 	/* The following initialization should only be done once. */
1020 	if (tcp->tcp_state != TCPS_LISTEN) {
1021 		tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
1022 		tcp->tcp_eager_next_q = NULL;
1023 		tcp->tcp_state = TCPS_LISTEN;
1024 		tcp->tcp_second_ctimer_threshold = tcp_ip_abort_linterval;
1025 	}
1026 	if ((tcp->tcp_conn_req_max = backlog) > tcp_conn_req_max_q) {
1027 		tcp->tcp_conn_req_max = tcp_conn_req_max_q;
1028 	}
1029 	if (tcp->tcp_conn_req_max < tcp_conn_req_min) {
1030 		tcp->tcp_conn_req_max = tcp_conn_req_min;
1031 	}
1032 	return (0);
1033 }
1034 
1035 /* To accept connections. */
1036 int
1037 tcp_accept(int sock_id, struct sockaddr *addr, socklen_t *addr_len)
1038 {
1039 	tcp_t *listener;
1040 	tcp_t *eager;
1041 	int sd, new_sock_id;
1042 	struct sockaddr_in *new_addr = (struct sockaddr_in *)addr;
1043 	int timeout;
1044 
1045 	/* Sanity check. */
1046 	if ((listener = (tcp_t *)(sockets[sock_id].pcb)) == NULL ||
1047 	    new_addr == NULL || addr_len == NULL ||
1048 	    *addr_len < sizeof (struct sockaddr_in) ||
1049 	    listener->tcp_state != TCPS_LISTEN) {
1050 		errno = EINVAL;
1051 		return (-1);
1052 	}
1053 
1054 	if (sockets[sock_id].in_timeout > tcp_accept_timeout)
1055 		timeout = prom_gettime() + sockets[sock_id].in_timeout;
1056 	else
1057 		timeout = prom_gettime() + tcp_accept_timeout;
1058 	while (listener->tcp_eager_next_q == NULL &&
1059 	    timeout > prom_gettime()) {
1060 #if DEBUG
1061 		printf("tcp_accept: Waiting in tcp_accept()\n");
1062 #endif
1063 		if (tcp_drain_input(listener, sock_id, 5) < 0) {
1064 			return (-1);
1065 		}
1066 	}
1067 	/* If there is an eager, don't timeout... */
1068 	if (timeout <= prom_gettime() && listener->tcp_eager_next_q == NULL) {
1069 #if DEBUG
1070 		printf("tcp_accept: timeout\n");
1071 #endif
1072 		errno = ETIMEDOUT;
1073 		return (-1);
1074 	}
1075 #if DEBUG
1076 	printf("tcp_accept: got a connection\n");
1077 #endif
1078 
1079 	/* Now create the socket for this new TCP. */
1080 	if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
1081 		return (-1);
1082 	}
1083 	if ((new_sock_id = so_check_fd(sd, &errno)) == -1)
1084 		/* This should not happen! */
1085 		prom_panic("so_check_fd() fails in tcp_accept()");
1086 	/* Free the TCP PCB in the original socket. */
1087 	bkmem_free((caddr_t)(sockets[new_sock_id].pcb), sizeof (tcp_t));
1088 	/* Dequeue the eager and attach it to the socket. */
1089 	eager = listener->tcp_eager_next_q;
1090 	listener->tcp_eager_next_q = eager->tcp_eager_next_q;
1091 	if (listener->tcp_eager_last_q == eager)
1092 		listener->tcp_eager_last_q = NULL;
1093 	eager->tcp_eager_next_q = NULL;
1094 	sockets[new_sock_id].pcb = eager;
1095 	listener->tcp_conn_req_cnt_q--;
1096 
1097 	/* Copy in the address info. */
1098 	bcopy(&eager->tcp_remote, &new_addr->sin_addr.s_addr,
1099 	    sizeof (in_addr_t));
1100 	bcopy(&eager->tcp_fport, &new_addr->sin_port, sizeof (in_port_t));
1101 	new_addr->sin_family = AF_INET;
1102 
1103 #ifdef DEBUG
1104 	printf("tcp_accept(), new sock_id: %d\n", sd);
1105 #endif
1106 	return (sd);
1107 }
1108 
1109 /* Update the next anonymous port to use.  */
1110 static in_port_t
1111 tcp_update_next_port(in_port_t port)
1112 {
1113 	/* Don't allow the port to fall out of the anonymous port range. */
1114 	if (port < tcp_smallest_anon_port || port > tcp_largest_anon_port)
1115 		port = (in_port_t)tcp_smallest_anon_port;
1116 
1117 	if (port < tcp_smallest_nonpriv_port)
1118 		port = (in_port_t)tcp_smallest_nonpriv_port;
1119 	return (port);
1120 }
1121 
1122 /* To check whether a bind to a port is allowed. */
1123 static in_port_t
1124 tcp_bindi(in_port_t port, in_addr_t *addr, boolean_t reuseaddr,
1125     boolean_t bind_to_req_port_only)
1126 {
1127 	int i, count;
1128 	tcp_t *tcp;
1129 
1130 	count = tcp_largest_anon_port - tcp_smallest_anon_port;
1131 try_again:
1132 	for (i = 0; i < MAXSOCKET; i++) {
1133 		if (sockets[i].type != INETBOOT_STREAM ||
1134 		    ((tcp = (tcp_t *)sockets[i].pcb) == NULL) ||
1135 		    ntohs(tcp->tcp_lport) != port) {
1136 			continue;
1137 		}
1138 		/*
1139 		 * Both TCPs have the same port.  If SO_REUSEDADDR is
1140 		 * set and the bound TCP has a state greater than
1141 		 * TCPS_LISTEN, it is fine.
1142 		 */
1143 		if (reuseaddr && tcp->tcp_state > TCPS_LISTEN) {
1144 			continue;
1145 		}
1146 		if (tcp->tcp_bound_source != INADDR_ANY &&
1147 		    *addr != INADDR_ANY &&
1148 		    tcp->tcp_bound_source != *addr) {
1149 			continue;
1150 		}
1151 		if (bind_to_req_port_only) {
1152 			return (0);
1153 		}
1154 		if (--count > 0) {
1155 			port = tcp_update_next_port(++port);
1156 			goto try_again;
1157 		} else {
1158 			return (0);
1159 		}
1160 	}
1161 	return (port);
1162 }
1163 
1164 /* To handle the bind request. */
1165 int
1166 tcp_bind(int sock_id)
1167 {
1168 	tcp_t *tcp;
1169 	in_port_t requested_port, allocated_port;
1170 	boolean_t bind_to_req_port_only;
1171 	boolean_t reuseaddr;
1172 
1173 	if ((tcp = (tcp_t *)sockets[sock_id].pcb) == NULL) {
1174 		errno = EINVAL;
1175 		return (-1);
1176 	}
1177 
1178 	if (tcp->tcp_state >= TCPS_BOUND) {
1179 		/* We don't allow multiple bind(). */
1180 		errno = EPROTO;
1181 		return (-1);
1182 	}
1183 
1184 	requested_port = ntohs(sockets[sock_id].bind.sin_port);
1185 
1186 	/* The bound source can be INADDR_ANY. */
1187 	tcp->tcp_bound_source = sockets[sock_id].bind.sin_addr.s_addr;
1188 
1189 	tcp->tcp_ipha->ip_src.s_addr = tcp->tcp_bound_source;
1190 
1191 	/* Verify the port is available. */
1192 	if (requested_port == 0)
1193 		bind_to_req_port_only = B_FALSE;
1194 	else			/* T_BIND_REQ and requested_port != 0 */
1195 		bind_to_req_port_only = B_TRUE;
1196 
1197 	if (requested_port == 0) {
1198 		requested_port = tcp_update_next_port(++tcp_next_port_to_try);
1199 	}
1200 	reuseaddr = sockets[sock_id].so_opt & SO_REUSEADDR;
1201 	allocated_port = tcp_bindi(requested_port, &(tcp->tcp_bound_source),
1202 	    reuseaddr, bind_to_req_port_only);
1203 
1204 	if (allocated_port == 0) {
1205 		errno = EADDRINUSE;
1206 		return (-1);
1207 	}
1208 	tcp->tcp_lport = htons(allocated_port);
1209 	*(uint16_t *)tcp->tcp_tcph->th_lport = tcp->tcp_lport;
1210 	sockets[sock_id].bind.sin_port = tcp->tcp_lport;
1211 	tcp->tcp_state = TCPS_BOUND;
1212 	return (0);
1213 }
1214 
1215 /*
1216  * Check for duplicate TCP connections.
1217  */
1218 static int
1219 tcp_conn_check(tcp_t *tcp)
1220 {
1221 	int i;
1222 	tcp_t *tmp_tcp;
1223 
1224 	for (i = 0; i < MAXSOCKET; i++) {
1225 		if (sockets[i].type != INETBOOT_STREAM)
1226 			continue;
1227 		/* Socket may not be closed but the TCP can be gone. */
1228 		if ((tmp_tcp = (tcp_t *)sockets[i].pcb) == NULL)
1229 			continue;
1230 		/* We only care about TCP in states later than SYN_SENT. */
1231 		if (tmp_tcp->tcp_state < TCPS_SYN_SENT)
1232 			continue;
1233 		if (tmp_tcp->tcp_lport != tcp->tcp_lport ||
1234 		    tmp_tcp->tcp_fport != tcp->tcp_fport ||
1235 		    tmp_tcp->tcp_bound_source != tcp->tcp_bound_source ||
1236 		    tmp_tcp->tcp_remote != tcp->tcp_remote) {
1237 			continue;
1238 		} else {
1239 			return (-1);
1240 		}
1241 	}
1242 	return (0);
1243 }
1244 
1245 /* To handle a connect request. */
1246 int
1247 tcp_connect(int sock_id)
1248 {
1249 	tcp_t *tcp;
1250 	in_addr_t dstaddr;
1251 	in_port_t dstport;
1252 	tcph_t	*tcph;
1253 	int mss;
1254 	mblk_t *syn_mp;
1255 
1256 	if ((tcp = (tcp_t *)(sockets[sock_id].pcb)) == NULL) {
1257 		errno = EINVAL;
1258 		return (-1);
1259 	}
1260 
1261 	TCP_RUN_TIME_WAIT_COLLECTOR();
1262 
1263 	dstaddr = sockets[sock_id].remote.sin_addr.s_addr;
1264 	dstport = sockets[sock_id].remote.sin_port;
1265 
1266 	/*
1267 	 * Check for attempt to connect to INADDR_ANY or non-unicast addrress.
1268 	 * We don't have enough info to check for broadcast addr, except
1269 	 * for the all 1 broadcast.
1270 	 */
1271 	if (dstaddr == INADDR_ANY || IN_CLASSD(ntohl(dstaddr)) ||
1272 	    dstaddr == INADDR_BROADCAST)  {
1273 		/*
1274 		 * SunOS 4.x and 4.3 BSD allow an application
1275 		 * to connect a TCP socket to INADDR_ANY.
1276 		 * When they do this, the kernel picks the
1277 		 * address of one interface and uses it
1278 		 * instead.  The kernel usually ends up
1279 		 * picking the address of the loopback
1280 		 * interface.  This is an undocumented feature.
1281 		 * However, we provide the same thing here
1282 		 * in order to have source and binary
1283 		 * compatibility with SunOS 4.x.
1284 		 * Update the T_CONN_REQ (sin/sin6) since it is used to
1285 		 * generate the T_CONN_CON.
1286 		 *
1287 		 * Fail this for inetboot TCP.
1288 		 */
1289 		errno = EINVAL;
1290 		return (-1);
1291 	}
1292 
1293 	/* It is not bound to any address yet... */
1294 	if (tcp->tcp_bound_source == INADDR_ANY) {
1295 		ipv4_getipaddr(&(sockets[sock_id].bind.sin_addr));
1296 		/* We don't have an address! */
1297 		if (ntohl(sockets[sock_id].bind.sin_addr.s_addr) ==
1298 		    INADDR_ANY) {
1299 			errno = EPROTO;
1300 			return (-1);
1301 		}
1302 		tcp->tcp_bound_source = sockets[sock_id].bind.sin_addr.s_addr;
1303 		tcp->tcp_ipha->ip_src.s_addr = tcp->tcp_bound_source;
1304 	}
1305 
1306 	/*
1307 	 * Don't let an endpoint connect to itself.
1308 	 */
1309 	if (dstaddr == tcp->tcp_ipha->ip_src.s_addr &&
1310 	    dstport == tcp->tcp_lport) {
1311 		errno = EINVAL;
1312 		return (-1);
1313 	}
1314 
1315 	tcp->tcp_ipha->ip_dst.s_addr = dstaddr;
1316 	tcp->tcp_remote = dstaddr;
1317 	tcph = tcp->tcp_tcph;
1318 	*(uint16_t *)tcph->th_fport = dstport;
1319 	tcp->tcp_fport = dstport;
1320 
1321 	/*
1322 	 * Don't allow this connection to completely duplicate
1323 	 * an existing connection.
1324 	 */
1325 	if (tcp_conn_check(tcp) < 0) {
1326 		errno = EADDRINUSE;
1327 		return (-1);
1328 	}
1329 
1330 	/*
1331 	 * Just make sure our rwnd is at
1332 	 * least tcp_recv_hiwat_mss * MSS
1333 	 * large, and round up to the nearest
1334 	 * MSS.
1335 	 *
1336 	 * We do the round up here because
1337 	 * we need to get the interface
1338 	 * MTU first before we can do the
1339 	 * round up.
1340 	 */
1341 	mss = tcp->tcp_mss - tcp->tcp_hdr_len;
1342 	tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
1343 	    tcp_recv_hiwat_minmss * mss);
1344 	tcp->tcp_rwnd_max = tcp->tcp_rwnd;
1345 	SET_WS_VALUE(tcp);
1346 	U32_TO_ABE16((tcp->tcp_rwnd >> tcp->tcp_rcv_ws),
1347 	    tcp->tcp_tcph->th_win);
1348 	if (tcp->tcp_rcv_ws > 0 || tcp_wscale_always)
1349 		tcp->tcp_snd_ws_ok = B_TRUE;
1350 
1351 	/*
1352 	 * Set tcp_snd_ts_ok to true
1353 	 * so that tcp_xmit_mp will
1354 	 * include the timestamp
1355 	 * option in the SYN segment.
1356 	 */
1357 	if (tcp_tstamp_always ||
1358 	    (tcp->tcp_rcv_ws && tcp_tstamp_if_wscale)) {
1359 		tcp->tcp_snd_ts_ok = B_TRUE;
1360 	}
1361 
1362 	if (tcp_sack_permitted == 2 ||
1363 	    tcp->tcp_snd_sack_ok) {
1364 		assert(tcp->tcp_sack_info == NULL);
1365 		if ((tcp->tcp_sack_info = (tcp_sack_info_t *)bkmem_zalloc(
1366 		    sizeof (tcp_sack_info_t))) == NULL) {
1367 			tcp->tcp_snd_sack_ok = B_FALSE;
1368 		} else {
1369 			tcp->tcp_snd_sack_ok = B_TRUE;
1370 		}
1371 	}
1372 	/*
1373 	 * Should we use ECN?  Note that the current
1374 	 * default value (SunOS 5.9) of tcp_ecn_permitted
1375 	 * is 2.  The reason for doing this is that there
1376 	 * are equipments out there that will drop ECN
1377 	 * enabled IP packets.  Setting it to 1 avoids
1378 	 * compatibility problems.
1379 	 */
1380 	if (tcp_ecn_permitted == 2)
1381 		tcp->tcp_ecn_ok = B_TRUE;
1382 
1383 	tcp_iss_init(tcp);
1384 	TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1385 	tcp->tcp_active_open = B_TRUE;
1386 
1387 	tcp->tcp_state = TCPS_SYN_SENT;
1388 	syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL, tcp->tcp_iss, B_FALSE,
1389 	    NULL, B_FALSE);
1390 	if (syn_mp != NULL) {
1391 		int ret;
1392 
1393 		/* Dump the packet when debugging. */
1394 		TCP_DUMP_PACKET("tcp_connect", syn_mp);
1395 		/* Send out the SYN packet. */
1396 		ret = ipv4_tcp_output(sock_id, syn_mp);
1397 		freeb(syn_mp);
1398 		if (ret < 0) {
1399 			return (-1);
1400 		}
1401 		/* tcp_state_wait() will finish the 3 way handshake. */
1402 		return (tcp_state_wait(sock_id, tcp, TCPS_ESTABLISHED));
1403 	} else {
1404 		errno = ENOBUFS;
1405 		return (-1);
1406 	}
1407 }
1408 
1409 /*
1410  * Common accept code.  Called by tcp_conn_request.
1411  * cr_pkt is the SYN packet.
1412  */
1413 static int
1414 tcp_accept_comm(tcp_t *listener, tcp_t *acceptor, mblk_t *cr_pkt,
1415     uint_t ip_hdr_len)
1416 {
1417 	tcph_t		*tcph;
1418 
1419 #ifdef DEBUG
1420 	printf("tcp_accept_comm #######################\n");
1421 #endif
1422 
1423 	/*
1424 	 * When we get here, we know that the acceptor header template
1425 	 * has already been initialized.
1426 	 * However, it may not match the listener if the listener
1427 	 * includes options...
1428 	 * It may also not match the listener if the listener is v6 and
1429 	 * and the acceptor is v4
1430 	 */
1431 	acceptor->tcp_lport = listener->tcp_lport;
1432 
1433 	if (listener->tcp_ipversion == acceptor->tcp_ipversion) {
1434 		if (acceptor->tcp_iphc_len != listener->tcp_iphc_len) {
1435 			/*
1436 			 * Listener had options of some sort; acceptor inherits.
1437 			 * Free up the acceptor template and allocate one
1438 			 * of the right size.
1439 			 */
1440 			bkmem_free(acceptor->tcp_iphc, acceptor->tcp_iphc_len);
1441 			acceptor->tcp_iphc = bkmem_zalloc(
1442 			    listener->tcp_iphc_len);
1443 			if (acceptor->tcp_iphc == NULL) {
1444 				acceptor->tcp_iphc_len = 0;
1445 				return (ENOMEM);
1446 			}
1447 			acceptor->tcp_iphc_len = listener->tcp_iphc_len;
1448 		}
1449 		acceptor->tcp_hdr_len = listener->tcp_hdr_len;
1450 		acceptor->tcp_ip_hdr_len = listener->tcp_ip_hdr_len;
1451 		acceptor->tcp_tcp_hdr_len = listener->tcp_tcp_hdr_len;
1452 
1453 		/*
1454 		 * Copy the IP+TCP header template from listener to acceptor
1455 		 */
1456 		bcopy(listener->tcp_iphc, acceptor->tcp_iphc,
1457 		    listener->tcp_hdr_len);
1458 		acceptor->tcp_ipha = (struct ip *)acceptor->tcp_iphc;
1459 		acceptor->tcp_tcph = (tcph_t *)(acceptor->tcp_iphc +
1460 		    acceptor->tcp_ip_hdr_len);
1461 	} else {
1462 		prom_panic("tcp_accept_comm: version not equal");
1463 	}
1464 
1465 	/* Copy our new dest and fport from the connection request packet */
1466 	if (acceptor->tcp_ipversion == IPV4_VERSION) {
1467 		struct ip *ipha;
1468 
1469 		ipha = (struct ip *)cr_pkt->b_rptr;
1470 		acceptor->tcp_ipha->ip_dst = ipha->ip_src;
1471 		acceptor->tcp_remote = ipha->ip_src.s_addr;
1472 		acceptor->tcp_ipha->ip_src = ipha->ip_dst;
1473 		acceptor->tcp_bound_source = ipha->ip_dst.s_addr;
1474 		tcph = (tcph_t *)&cr_pkt->b_rptr[ip_hdr_len];
1475 	} else {
1476 		prom_panic("tcp_accept_comm: not IPv4");
1477 	}
1478 	bcopy(tcph->th_lport, acceptor->tcp_tcph->th_fport, sizeof (in_port_t));
1479 	bcopy(acceptor->tcp_tcph->th_fport, &acceptor->tcp_fport,
1480 	    sizeof (in_port_t));
1481 	/*
1482 	 * For an all-port proxy listener, the local port is determined by
1483 	 * the port number field in the SYN packet.
1484 	 */
1485 	if (listener->tcp_lport == 0) {
1486 		acceptor->tcp_lport = *(in_port_t *)tcph->th_fport;
1487 		bcopy(tcph->th_fport, acceptor->tcp_tcph->th_lport,
1488 		    sizeof (in_port_t));
1489 	}
1490 	/* Inherit various TCP parameters from the listener */
1491 	acceptor->tcp_naglim = listener->tcp_naglim;
1492 	acceptor->tcp_first_timer_threshold =
1493 	    listener->tcp_first_timer_threshold;
1494 	acceptor->tcp_second_timer_threshold =
1495 	    listener->tcp_second_timer_threshold;
1496 
1497 	acceptor->tcp_first_ctimer_threshold =
1498 	    listener->tcp_first_ctimer_threshold;
1499 	acceptor->tcp_second_ctimer_threshold =
1500 	    listener->tcp_second_ctimer_threshold;
1501 
1502 	acceptor->tcp_xmit_hiwater = listener->tcp_xmit_hiwater;
1503 
1504 	acceptor->tcp_state = TCPS_LISTEN;
1505 	tcp_iss_init(acceptor);
1506 
1507 	/* Process all TCP options. */
1508 	tcp_process_options(acceptor, tcph);
1509 
1510 	/* Is the other end ECN capable? */
1511 	if (tcp_ecn_permitted >= 1 &&
1512 	    (tcph->th_flags[0] & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) {
1513 		acceptor->tcp_ecn_ok = B_TRUE;
1514 	}
1515 
1516 	/*
1517 	 * listener->tcp_rq->q_hiwat should be the default window size or a
1518 	 * window size changed via SO_RCVBUF option.  First round up the
1519 	 * acceptor's tcp_rwnd to the nearest MSS.  Then find out the window
1520 	 * scale option value if needed.  Call tcp_rwnd_set() to finish the
1521 	 * setting.
1522 	 *
1523 	 * Note if there is a rpipe metric associated with the remote host,
1524 	 * we should not inherit receive window size from listener.
1525 	 */
1526 	acceptor->tcp_rwnd = MSS_ROUNDUP(
1527 	    (acceptor->tcp_rwnd == 0 ? listener->tcp_rwnd_max :
1528 	    acceptor->tcp_rwnd), acceptor->tcp_mss);
1529 	if (acceptor->tcp_snd_ws_ok)
1530 		SET_WS_VALUE(acceptor);
1531 	/*
1532 	 * Note that this is the only place tcp_rwnd_set() is called for
1533 	 * accepting a connection.  We need to call it here instead of
1534 	 * after the 3-way handshake because we need to tell the other
1535 	 * side our rwnd in the SYN-ACK segment.
1536 	 */
1537 	(void) tcp_rwnd_set(acceptor, acceptor->tcp_rwnd);
1538 
1539 	return (0);
1540 }
1541 
1542 /*
1543  * Defense for the SYN attack -
1544  * 1. When q0 is full, drop from the tail (tcp_eager_prev_q0) the oldest
1545  *    one that doesn't have the dontdrop bit set.
1546  * 2. Don't drop a SYN request before its first timeout. This gives every
1547  *    request at least til the first timeout to complete its 3-way handshake.
1548  * 3. The current threshold is - # of timeout > q0len/4 => SYN alert on
1549  *    # of timeout drops back to <= q0len/32 => SYN alert off
1550  */
1551 static boolean_t
1552 tcp_drop_q0(tcp_t *tcp)
1553 {
1554 	tcp_t	*eager;
1555 
1556 	assert(tcp->tcp_eager_next_q0 != tcp->tcp_eager_prev_q0);
1557 	/*
1558 	 * New one is added after next_q0 so prev_q0 points to the oldest
1559 	 * Also do not drop any established connections that are deferred on
1560 	 * q0 due to q being full
1561 	 */
1562 
1563 	eager = tcp->tcp_eager_prev_q0;
1564 	while (eager->tcp_dontdrop || eager->tcp_conn_def_q0) {
1565 		/* XXX should move the eager to the head */
1566 		eager = eager->tcp_eager_prev_q0;
1567 		if (eager == tcp) {
1568 			eager = tcp->tcp_eager_prev_q0;
1569 			break;
1570 		}
1571 	}
1572 	dprintf("tcp_drop_q0: listen half-open queue (max=%d) overflow"
1573 	    " (%d pending) on %s, drop one", tcp_conn_req_max_q0,
1574 	    tcp->tcp_conn_req_cnt_q0,
1575 	    tcp_display(tcp, NULL, DISP_PORT_ONLY));
1576 
1577 	BUMP_MIB(tcp_mib.tcpHalfOpenDrop);
1578 	bkmem_free((caddr_t)eager, sizeof (tcp_t));
1579 	return (B_TRUE);
1580 }
1581 
1582 /* ARGSUSED */
1583 static tcp_t *
1584 tcp_conn_request(tcp_t *tcp, mblk_t *mp, uint_t sock_id, uint_t ip_hdr_len)
1585 {
1586 	tcp_t	*eager;
1587 	struct ip *ipha;
1588 	int	err;
1589 
1590 #ifdef DEBUG
1591 	printf("tcp_conn_request ###################\n");
1592 #endif
1593 
1594 	if (tcp->tcp_conn_req_cnt_q >= tcp->tcp_conn_req_max) {
1595 		BUMP_MIB(tcp_mib.tcpListenDrop);
1596 		dprintf("tcp_conn_request: listen backlog (max=%d) "
1597 		    "overflow (%d pending) on %s",
1598 		    tcp->tcp_conn_req_max, tcp->tcp_conn_req_cnt_q,
1599 		    tcp_display(tcp, NULL, DISP_PORT_ONLY));
1600 		return (NULL);
1601 	}
1602 
1603 	assert(OK_32PTR(mp->b_rptr));
1604 
1605 	if (tcp->tcp_conn_req_cnt_q0 >=
1606 	    tcp->tcp_conn_req_max + tcp_conn_req_max_q0) {
1607 		/*
1608 		 * Q0 is full. Drop a pending half-open req from the queue
1609 		 * to make room for the new SYN req. Also mark the time we
1610 		 * drop a SYN.
1611 		 */
1612 		tcp->tcp_last_rcv_lbolt = prom_gettime();
1613 		if (!tcp_drop_q0(tcp)) {
1614 			freemsg(mp);
1615 			BUMP_MIB(tcp_mib.tcpListenDropQ0);
1616 			dprintf("tcp_conn_request: listen half-open queue "
1617 			    "(max=%d) full (%d pending) on %s",
1618 			    tcp_conn_req_max_q0,
1619 			    tcp->tcp_conn_req_cnt_q0,
1620 			    tcp_display(tcp, NULL, DISP_PORT_ONLY));
1621 			return (NULL);
1622 		}
1623 	}
1624 
1625 	ipha = (struct ip *)mp->b_rptr;
1626 	if (IN_CLASSD(ntohl(ipha->ip_src.s_addr)) ||
1627 	    ipha->ip_src.s_addr == INADDR_BROADCAST ||
1628 	    ipha->ip_src.s_addr == INADDR_ANY ||
1629 	    ipha->ip_dst.s_addr == INADDR_BROADCAST) {
1630 		freemsg(mp);
1631 		return (NULL);
1632 	}
1633 	/*
1634 	 * We allow the connection to proceed
1635 	 * by generating a detached tcp state vector and put it in
1636 	 * the eager queue.  When an accept happens, it will be
1637 	 * dequeued sequentially.
1638 	 */
1639 	if ((eager = (tcp_t *)bkmem_alloc(sizeof (tcp_t))) == NULL) {
1640 		freemsg(mp);
1641 		errno = ENOBUFS;
1642 		return (NULL);
1643 	}
1644 	if ((errno = tcp_init_values(eager, NULL)) != 0) {
1645 		freemsg(mp);
1646 		bkmem_free((caddr_t)eager, sizeof (tcp_t));
1647 		return (NULL);
1648 	}
1649 
1650 	/*
1651 	 * Eager connection inherits address form from its listener,
1652 	 * but its packet form comes from the version of the received
1653 	 * SYN segment.
1654 	 */
1655 	eager->tcp_family = tcp->tcp_family;
1656 
1657 	err = tcp_accept_comm(tcp, eager, mp, ip_hdr_len);
1658 	if (err) {
1659 		bkmem_free((caddr_t)eager, sizeof (tcp_t));
1660 		return (NULL);
1661 	}
1662 
1663 	tcp->tcp_eager_next_q0->tcp_eager_prev_q0 = eager;
1664 	eager->tcp_eager_next_q0 = tcp->tcp_eager_next_q0;
1665 	tcp->tcp_eager_next_q0 = eager;
1666 	eager->tcp_eager_prev_q0 = tcp;
1667 
1668 	/* Set tcp_listener before adding it to tcp_conn_fanout */
1669 	eager->tcp_listener = tcp;
1670 	tcp->tcp_conn_req_cnt_q0++;
1671 
1672 	return (eager);
1673 }
1674 
1675 /*
1676  * To get around the non-interrupt problem of inetboot.
1677  * Keep on processing packets until a certain state is reached or the
1678  * TCP is destroyed because of getting a RST packet.
1679  */
1680 static int
1681 tcp_state_wait(int sock_id, tcp_t *tcp, int state)
1682 {
1683 	int i;
1684 	struct inetgram *in_gram;
1685 	mblk_t *mp;
1686 	int timeout;
1687 	boolean_t changed = B_FALSE;
1688 
1689 	/*
1690 	 * We need to make sure that the MAC does not wait longer
1691 	 * than RTO for any packet so that TCP can do retransmission.
1692 	 * But if the MAC timeout is less than tcp_rto, we are fine
1693 	 * and do not need to change it.
1694 	 */
1695 	timeout = sockets[sock_id].in_timeout;
1696 	if (timeout > tcp->tcp_rto) {
1697 		sockets[sock_id].in_timeout = tcp->tcp_rto;
1698 		changed = B_TRUE;
1699 	}
1700 retry:
1701 	if (sockets[sock_id].inq == NULL) {
1702 		/* Go out and check the wire */
1703 		for (i = MEDIA_LVL; i < TRANSPORT_LVL; i++) {
1704 			if (sockets[sock_id].input[i] != NULL) {
1705 				if (sockets[sock_id].input[i](sock_id) < 0) {
1706 					if (changed) {
1707 						sockets[sock_id].in_timeout =
1708 						    timeout;
1709 					}
1710 					return (-1);
1711 				}
1712 			}
1713 		}
1714 	}
1715 
1716 	while ((in_gram = sockets[sock_id].inq) != NULL) {
1717 		if (tcp != NULL && tcp->tcp_state == state)
1718 			break;
1719 
1720 		/* Remove unknown inetgrams from the head of inq. */
1721 		if (in_gram->igm_level != TRANSPORT_LVL) {
1722 #ifdef DEBUG
1723 			printf("tcp_state_wait for state %d: unexpected "
1724 			    "packet level %d frame found\n", state,
1725 			    in_gram->igm_level);
1726 #endif
1727 			del_gram(&sockets[sock_id].inq, in_gram, B_TRUE);
1728 			continue;
1729 		}
1730 		mp = in_gram->igm_mp;
1731 		del_gram(&sockets[sock_id].inq, in_gram, B_FALSE);
1732 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
1733 		tcp_rput_data(tcp, mp, sock_id);
1734 
1735 		/*
1736 		 * The other side may have closed this connection or
1737 		 * RST us.  But we need to continue to process other
1738 		 * packets in the socket's queue because they may be
1739 		 * belong to another TCP connections.
1740 		 */
1741 		if (sockets[sock_id].pcb == NULL) {
1742 			tcp = NULL;
1743 		}
1744 	}
1745 
1746 	/* If the other side has closed the connection, just return. */
1747 	if (tcp == NULL || sockets[sock_id].pcb == NULL) {
1748 #ifdef DEBUG
1749 		printf("tcp_state_wait other side dead: state %d "
1750 		    "error %d\n", state, sockets[sock_id].so_error);
1751 #endif
1752 		if (sockets[sock_id].so_error != 0)
1753 			return (-1);
1754 		else
1755 			return (0);
1756 	}
1757 	/*
1758 	 * TCPS_ALL_ACKED is not a valid TCP state, it is just used as an
1759 	 * indicator to tcp_state_wait to mean that it is being called
1760 	 * to wait till we have received acks for all the new segments sent.
1761 	 */
1762 	if ((state == TCPS_ALL_ACKED) && (tcp->tcp_suna == tcp->tcp_snxt)) {
1763 		goto done;
1764 	}
1765 	if (tcp->tcp_state != state) {
1766 		if (prom_gettime() > tcp->tcp_rto_timeout)
1767 			tcp_timer(tcp, sock_id);
1768 		goto retry;
1769 	}
1770 done:
1771 	if (changed)
1772 		sockets[sock_id].in_timeout = timeout;
1773 
1774 	tcp_drain_needed(sock_id, tcp);
1775 	return (0);
1776 }
1777 
1778 /* Verify the checksum of a segment. */
1779 static int
1780 tcp_verify_cksum(mblk_t *mp)
1781 {
1782 	struct ip *iph;
1783 	tcpha_t *tcph;
1784 	int len;
1785 	uint16_t old_sum;
1786 
1787 	iph = (struct ip *)mp->b_rptr;
1788 	tcph = (tcpha_t *)(iph + 1);
1789 	len = ntohs(iph->ip_len);
1790 
1791 	/*
1792 	 * Calculate the TCP checksum.  Need to include the psuedo header,
1793 	 * which is similar to the real IP header starting at the TTL field.
1794 	 */
1795 	iph->ip_sum = htons(len - IP_SIMPLE_HDR_LENGTH);
1796 	old_sum = tcph->tha_sum;
1797 	tcph->tha_sum = 0;
1798 	iph->ip_ttl = 0;
1799 	if (old_sum == tcp_cksum((uint16_t *)&(iph->ip_ttl),
1800 	    len - IP_SIMPLE_HDR_LENGTH + 12)) {
1801 		return (0);
1802 	} else {
1803 		tcp_cksum_errors++;
1804 		return (-1);
1805 	}
1806 }
1807 
1808 /* To find a TCP connection matching the incoming segment. */
1809 static tcp_t *
1810 tcp_lookup_ipv4(struct ip *iph, tcpha_t *tcph, int min_state, int *sock_id)
1811 {
1812 	int i;
1813 	tcp_t *tcp;
1814 
1815 	for (i = 0; i < MAXSOCKET; i++) {
1816 		if (sockets[i].type == INETBOOT_STREAM &&
1817 		    (tcp = (tcp_t *)sockets[i].pcb) != NULL) {
1818 			if (tcph->tha_lport == tcp->tcp_fport &&
1819 			    tcph->tha_fport == tcp->tcp_lport &&
1820 			    iph->ip_src.s_addr == tcp->tcp_remote &&
1821 			    iph->ip_dst.s_addr == tcp->tcp_bound_source &&
1822 			    tcp->tcp_state >= min_state) {
1823 				*sock_id = i;
1824 				return (tcp);
1825 			}
1826 		}
1827 	}
1828 	/* Find it in the time wait list. */
1829 	for (tcp = tcp_time_wait_head; tcp != NULL;
1830 	    tcp = tcp->tcp_time_wait_next) {
1831 		if (tcph->tha_lport == tcp->tcp_fport &&
1832 		    tcph->tha_fport == tcp->tcp_lport &&
1833 		    iph->ip_src.s_addr == tcp->tcp_remote &&
1834 		    iph->ip_dst.s_addr == tcp->tcp_bound_source &&
1835 		    tcp->tcp_state >= min_state) {
1836 			*sock_id = -1;
1837 			return (tcp);
1838 		}
1839 	}
1840 	return (NULL);
1841 }
1842 
1843 /* To find a TCP listening connection matching the incoming segment. */
1844 static tcp_t *
1845 tcp_lookup_listener_ipv4(in_addr_t addr, in_port_t port, int *sock_id)
1846 {
1847 	int i;
1848 	tcp_t *tcp;
1849 
1850 	for (i = 0; i < MAXSOCKET; i++) {
1851 		if (sockets[i].type == INETBOOT_STREAM &&
1852 		    (tcp = (tcp_t *)sockets[i].pcb) != NULL) {
1853 			if (tcp->tcp_lport == port &&
1854 			    (tcp->tcp_bound_source == addr ||
1855 			    tcp->tcp_bound_source == INADDR_ANY)) {
1856 				*sock_id = i;
1857 				return (tcp);
1858 			}
1859 		}
1860 	}
1861 
1862 	return (NULL);
1863 }
1864 
1865 /* To find a TCP eager matching the incoming segment. */
1866 static tcp_t *
1867 tcp_lookup_eager_ipv4(tcp_t *listener, struct ip *iph, tcpha_t *tcph)
1868 {
1869 	tcp_t *tcp;
1870 
1871 #ifdef DEBUG
1872 	printf("tcp_lookup_eager_ipv4 ###############\n");
1873 #endif
1874 	for (tcp = listener->tcp_eager_next_q; tcp != NULL;
1875 	    tcp = tcp->tcp_eager_next_q) {
1876 		if (tcph->tha_lport == tcp->tcp_fport &&
1877 		    tcph->tha_fport == tcp->tcp_lport &&
1878 		    iph->ip_src.s_addr == tcp->tcp_remote &&
1879 		    iph->ip_dst.s_addr == tcp->tcp_bound_source) {
1880 			return (tcp);
1881 		}
1882 	}
1883 
1884 	for (tcp = listener->tcp_eager_next_q0; tcp != listener;
1885 	    tcp = tcp->tcp_eager_next_q0) {
1886 		if (tcph->tha_lport == tcp->tcp_fport &&
1887 		    tcph->tha_fport == tcp->tcp_lport &&
1888 		    iph->ip_src.s_addr == tcp->tcp_remote &&
1889 		    iph->ip_dst.s_addr == tcp->tcp_bound_source) {
1890 			return (tcp);
1891 		}
1892 	}
1893 #ifdef DEBUG
1894 	printf("No eager found\n");
1895 #endif
1896 	return (NULL);
1897 }
1898 
1899 /* To destroy a TCP control block. */
1900 static void
1901 tcp_clean_death(int sock_id, tcp_t *tcp, int err)
1902 {
1903 	tcp_free(tcp);
1904 	if (tcp->tcp_state == TCPS_TIME_WAIT)
1905 		tcp_time_wait_remove(tcp);
1906 
1907 	if (sock_id >= 0) {
1908 		sockets[sock_id].pcb = NULL;
1909 		if (err != 0)
1910 			sockets[sock_id].so_error = err;
1911 	}
1912 	bkmem_free((caddr_t)tcp, sizeof (tcp_t));
1913 }
1914 
1915 /*
1916  * tcp_rwnd_set() is called to adjust the receive window to a desired value.
1917  * We do not allow the receive window to shrink.  After setting rwnd,
1918  * set the flow control hiwat of the stream.
1919  *
1920  * This function is called in 2 cases:
1921  *
1922  * 1) Before data transfer begins, in tcp_accept_comm() for accepting a
1923  *    connection (passive open) and in tcp_rput_data() for active connect.
1924  *    This is called after tcp_mss_set() when the desired MSS value is known.
1925  *    This makes sure that our window size is a mutiple of the other side's
1926  *    MSS.
1927  * 2) Handling SO_RCVBUF option.
1928  *
1929  * It is ASSUMED that the requested size is a multiple of the current MSS.
1930  *
1931  * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
1932  * user requests so.
1933  */
1934 static int
1935 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
1936 {
1937 	uint32_t	mss = tcp->tcp_mss;
1938 	uint32_t	old_max_rwnd;
1939 	uint32_t	max_transmittable_rwnd;
1940 
1941 	if (tcp->tcp_rwnd_max != 0)
1942 		old_max_rwnd = tcp->tcp_rwnd_max;
1943 	else
1944 		old_max_rwnd = tcp->tcp_rwnd;
1945 
1946 	/*
1947 	 * Insist on a receive window that is at least
1948 	 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
1949 	 * funny TCP interactions of Nagle algorithm, SWS avoidance
1950 	 * and delayed acknowledgement.
1951 	 */
1952 	rwnd = MAX(rwnd, tcp_recv_hiwat_minmss * mss);
1953 
1954 	/*
1955 	 * If window size info has already been exchanged, TCP should not
1956 	 * shrink the window.  Shrinking window is doable if done carefully.
1957 	 * We may add that support later.  But so far there is not a real
1958 	 * need to do that.
1959 	 */
1960 	if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
1961 		/* MSS may have changed, do a round up again. */
1962 		rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
1963 	}
1964 
1965 	/*
1966 	 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
1967 	 * can be applied even before the window scale option is decided.
1968 	 */
1969 	max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
1970 	if (rwnd > max_transmittable_rwnd) {
1971 		rwnd = max_transmittable_rwnd -
1972 		    (max_transmittable_rwnd % mss);
1973 		if (rwnd < mss)
1974 			rwnd = max_transmittable_rwnd;
1975 		/*
1976 		 * If we're over the limit we may have to back down tcp_rwnd.
1977 		 * The increment below won't work for us. So we set all three
1978 		 * here and the increment below will have no effect.
1979 		 */
1980 		tcp->tcp_rwnd = old_max_rwnd = rwnd;
1981 	}
1982 
1983 	/*
1984 	 * Increment the current rwnd by the amount the maximum grew (we
1985 	 * can not overwrite it since we might be in the middle of a
1986 	 * connection.)
1987 	 */
1988 	tcp->tcp_rwnd += rwnd - old_max_rwnd;
1989 	U32_TO_ABE16(tcp->tcp_rwnd >> tcp->tcp_rcv_ws, tcp->tcp_tcph->th_win);
1990 	if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
1991 		tcp->tcp_cwnd_max = rwnd;
1992 	tcp->tcp_rwnd_max = rwnd;
1993 
1994 	return (rwnd);
1995 }
1996 
1997 /*
1998  * Extract option values from a tcp header.  We put any found values into the
1999  * tcpopt struct and return a bitmask saying which options were found.
2000  */
2001 static int
2002 tcp_parse_options(tcph_t *tcph, tcp_opt_t *tcpopt)
2003 {
2004 	uchar_t		*endp;
2005 	int		len;
2006 	uint32_t	mss;
2007 	uchar_t		*up = (uchar_t *)tcph;
2008 	int		found = 0;
2009 	int32_t		sack_len;
2010 	tcp_seq		sack_begin, sack_end;
2011 	tcp_t		*tcp;
2012 
2013 	endp = up + TCP_HDR_LENGTH(tcph);
2014 	up += TCP_MIN_HEADER_LENGTH;
2015 	while (up < endp) {
2016 		len = endp - up;
2017 		switch (*up) {
2018 		case TCPOPT_EOL:
2019 			break;
2020 
2021 		case TCPOPT_NOP:
2022 			up++;
2023 			continue;
2024 
2025 		case TCPOPT_MAXSEG:
2026 			if (len < TCPOPT_MAXSEG_LEN ||
2027 			    up[1] != TCPOPT_MAXSEG_LEN)
2028 				break;
2029 
2030 			mss = BE16_TO_U16(up+2);
2031 			/* Caller must handle tcp_mss_min and tcp_mss_max_* */
2032 			tcpopt->tcp_opt_mss = mss;
2033 			found |= TCP_OPT_MSS_PRESENT;
2034 
2035 			up += TCPOPT_MAXSEG_LEN;
2036 			continue;
2037 
2038 		case TCPOPT_WSCALE:
2039 			if (len < TCPOPT_WS_LEN || up[1] != TCPOPT_WS_LEN)
2040 				break;
2041 
2042 			if (up[2] > TCP_MAX_WINSHIFT)
2043 				tcpopt->tcp_opt_wscale = TCP_MAX_WINSHIFT;
2044 			else
2045 				tcpopt->tcp_opt_wscale = up[2];
2046 			found |= TCP_OPT_WSCALE_PRESENT;
2047 
2048 			up += TCPOPT_WS_LEN;
2049 			continue;
2050 
2051 		case TCPOPT_SACK_PERMITTED:
2052 			if (len < TCPOPT_SACK_OK_LEN ||
2053 			    up[1] != TCPOPT_SACK_OK_LEN)
2054 				break;
2055 			found |= TCP_OPT_SACK_OK_PRESENT;
2056 			up += TCPOPT_SACK_OK_LEN;
2057 			continue;
2058 
2059 		case TCPOPT_SACK:
2060 			if (len <= 2 || up[1] <= 2 || len < up[1])
2061 				break;
2062 
2063 			/* If TCP is not interested in SACK blks... */
2064 			if ((tcp = tcpopt->tcp) == NULL) {
2065 				up += up[1];
2066 				continue;
2067 			}
2068 			sack_len = up[1] - TCPOPT_HEADER_LEN;
2069 			up += TCPOPT_HEADER_LEN;
2070 
2071 			/*
2072 			 * If the list is empty, allocate one and assume
2073 			 * nothing is sack'ed.
2074 			 */
2075 			assert(tcp->tcp_sack_info != NULL);
2076 			if (tcp->tcp_notsack_list == NULL) {
2077 				tcp_notsack_update(&(tcp->tcp_notsack_list),
2078 				    tcp->tcp_suna, tcp->tcp_snxt,
2079 				    &(tcp->tcp_num_notsack_blk),
2080 				    &(tcp->tcp_cnt_notsack_list));
2081 
2082 				/*
2083 				 * Make sure tcp_notsack_list is not NULL.
2084 				 * This happens when kmem_alloc(KM_NOSLEEP)
2085 				 * returns NULL.
2086 				 */
2087 				if (tcp->tcp_notsack_list == NULL) {
2088 					up += sack_len;
2089 					continue;
2090 				}
2091 				tcp->tcp_fack = tcp->tcp_suna;
2092 			}
2093 
2094 			while (sack_len > 0) {
2095 				if (up + 8 > endp) {
2096 					up = endp;
2097 					break;
2098 				}
2099 				sack_begin = BE32_TO_U32(up);
2100 				up += 4;
2101 				sack_end = BE32_TO_U32(up);
2102 				up += 4;
2103 				sack_len -= 8;
2104 				/*
2105 				 * Bounds checking.  Make sure the SACK
2106 				 * info is within tcp_suna and tcp_snxt.
2107 				 * If this SACK blk is out of bound, ignore
2108 				 * it but continue to parse the following
2109 				 * blks.
2110 				 */
2111 				if (SEQ_LEQ(sack_end, sack_begin) ||
2112 				    SEQ_LT(sack_begin, tcp->tcp_suna) ||
2113 				    SEQ_GT(sack_end, tcp->tcp_snxt)) {
2114 					continue;
2115 				}
2116 				tcp_notsack_insert(&(tcp->tcp_notsack_list),
2117 				    sack_begin, sack_end,
2118 				    &(tcp->tcp_num_notsack_blk),
2119 				    &(tcp->tcp_cnt_notsack_list));
2120 				if (SEQ_GT(sack_end, tcp->tcp_fack)) {
2121 					tcp->tcp_fack = sack_end;
2122 				}
2123 			}
2124 			found |= TCP_OPT_SACK_PRESENT;
2125 			continue;
2126 
2127 		case TCPOPT_TSTAMP:
2128 			if (len < TCPOPT_TSTAMP_LEN ||
2129 			    up[1] != TCPOPT_TSTAMP_LEN)
2130 				break;
2131 
2132 			tcpopt->tcp_opt_ts_val = BE32_TO_U32(up+2);
2133 			tcpopt->tcp_opt_ts_ecr = BE32_TO_U32(up+6);
2134 
2135 			found |= TCP_OPT_TSTAMP_PRESENT;
2136 
2137 			up += TCPOPT_TSTAMP_LEN;
2138 			continue;
2139 
2140 		default:
2141 			if (len <= 1 || len < (int)up[1] || up[1] == 0)
2142 				break;
2143 			up += up[1];
2144 			continue;
2145 		}
2146 		break;
2147 	}
2148 	return (found);
2149 }
2150 
2151 /*
2152  * Set the mss associated with a particular tcp based on its current value,
2153  * and a new one passed in. Observe minimums and maximums, and reset
2154  * other state variables that we want to view as multiples of mss.
2155  *
2156  * This function is called in various places mainly because
2157  * 1) Various stuffs, tcp_mss, tcp_cwnd, ... need to be adjusted when the
2158  *    other side's SYN/SYN-ACK packet arrives.
2159  * 2) PMTUd may get us a new MSS.
2160  * 3) If the other side stops sending us timestamp option, we need to
2161  *    increase the MSS size to use the extra bytes available.
2162  */
2163 static void
2164 tcp_mss_set(tcp_t *tcp, uint32_t mss)
2165 {
2166 	uint32_t	mss_max;
2167 
2168 	mss_max = tcp_mss_max_ipv4;
2169 
2170 	if (mss < tcp_mss_min)
2171 		mss = tcp_mss_min;
2172 	if (mss > mss_max)
2173 		mss = mss_max;
2174 	/*
2175 	 * Unless naglim has been set by our client to
2176 	 * a non-mss value, force naglim to track mss.
2177 	 * This can help to aggregate small writes.
2178 	 */
2179 	if (mss < tcp->tcp_naglim || tcp->tcp_mss == tcp->tcp_naglim)
2180 		tcp->tcp_naglim = mss;
2181 	/*
2182 	 * TCP should be able to buffer at least 4 MSS data for obvious
2183 	 * performance reason.
2184 	 */
2185 	if ((mss << 2) > tcp->tcp_xmit_hiwater)
2186 		tcp->tcp_xmit_hiwater = mss << 2;
2187 	tcp->tcp_mss = mss;
2188 	/*
2189 	 * Initialize cwnd according to draft-floyd-incr-init-win-01.txt.
2190 	 * Previously, we use tcp_slow_start_initial to control the size
2191 	 * of the initial cwnd.  Now, when tcp_slow_start_initial * mss
2192 	 * is smaller than the cwnd calculated from the formula suggested in
2193 	 * the draft, we use tcp_slow_start_initial * mss as the cwnd.
2194 	 * Otherwise, use the cwnd from the draft's formula.  The default
2195 	 * of tcp_slow_start_initial is 2.
2196 	 */
2197 	tcp->tcp_cwnd = MIN(tcp_slow_start_initial * mss,
2198 	    MIN(4 * mss, MAX(2 * mss, 4380 / mss * mss)));
2199 	tcp->tcp_cwnd_cnt = 0;
2200 }
2201 
2202 /*
2203  * Process all TCP option in SYN segment.
2204  *
2205  * This function sets up the correct tcp_mss value according to the
2206  * MSS option value and our header size.  It also sets up the window scale
2207  * and timestamp values, and initialize SACK info blocks.  But it does not
2208  * change receive window size after setting the tcp_mss value.  The caller
2209  * should do the appropriate change.
2210  */
2211 void
2212 tcp_process_options(tcp_t *tcp, tcph_t *tcph)
2213 {
2214 	int options;
2215 	tcp_opt_t tcpopt;
2216 	uint32_t mss_max;
2217 	char *tmp_tcph;
2218 
2219 	tcpopt.tcp = NULL;
2220 	options = tcp_parse_options(tcph, &tcpopt);
2221 
2222 	/*
2223 	 * Process MSS option.  Note that MSS option value does not account
2224 	 * for IP or TCP options.  This means that it is equal to MTU - minimum
2225 	 * IP+TCP header size, which is 40 bytes for IPv4 and 60 bytes for
2226 	 * IPv6.
2227 	 */
2228 	if (!(options & TCP_OPT_MSS_PRESENT)) {
2229 		tcpopt.tcp_opt_mss = tcp_mss_def_ipv4;
2230 	} else {
2231 		if (tcp->tcp_ipversion == IPV4_VERSION)
2232 			mss_max = tcp_mss_max_ipv4;
2233 		if (tcpopt.tcp_opt_mss < tcp_mss_min)
2234 			tcpopt.tcp_opt_mss = tcp_mss_min;
2235 		else if (tcpopt.tcp_opt_mss > mss_max)
2236 			tcpopt.tcp_opt_mss = mss_max;
2237 	}
2238 
2239 	/* Process Window Scale option. */
2240 	if (options & TCP_OPT_WSCALE_PRESENT) {
2241 		tcp->tcp_snd_ws = tcpopt.tcp_opt_wscale;
2242 		tcp->tcp_snd_ws_ok = B_TRUE;
2243 	} else {
2244 		tcp->tcp_snd_ws = B_FALSE;
2245 		tcp->tcp_snd_ws_ok = B_FALSE;
2246 		tcp->tcp_rcv_ws = B_FALSE;
2247 	}
2248 
2249 	/* Process Timestamp option. */
2250 	if ((options & TCP_OPT_TSTAMP_PRESENT) &&
2251 	    (tcp->tcp_snd_ts_ok || !tcp->tcp_active_open)) {
2252 		tmp_tcph = (char *)tcp->tcp_tcph;
2253 
2254 		tcp->tcp_snd_ts_ok = B_TRUE;
2255 		tcp->tcp_ts_recent = tcpopt.tcp_opt_ts_val;
2256 		tcp->tcp_last_rcv_lbolt = prom_gettime();
2257 		assert(OK_32PTR(tmp_tcph));
2258 		assert(tcp->tcp_tcp_hdr_len == TCP_MIN_HEADER_LENGTH);
2259 
2260 		/* Fill in our template header with basic timestamp option. */
2261 		tmp_tcph += tcp->tcp_tcp_hdr_len;
2262 		tmp_tcph[0] = TCPOPT_NOP;
2263 		tmp_tcph[1] = TCPOPT_NOP;
2264 		tmp_tcph[2] = TCPOPT_TSTAMP;
2265 		tmp_tcph[3] = TCPOPT_TSTAMP_LEN;
2266 		tcp->tcp_hdr_len += TCPOPT_REAL_TS_LEN;
2267 		tcp->tcp_tcp_hdr_len += TCPOPT_REAL_TS_LEN;
2268 		tcp->tcp_tcph->th_offset_and_rsrvd[0] += (3 << 4);
2269 	} else {
2270 		tcp->tcp_snd_ts_ok = B_FALSE;
2271 	}
2272 
2273 	/*
2274 	 * Process SACK options.  If SACK is enabled for this connection,
2275 	 * then allocate the SACK info structure.
2276 	 */
2277 	if ((options & TCP_OPT_SACK_OK_PRESENT) &&
2278 	    (tcp->tcp_snd_sack_ok ||
2279 	    (tcp_sack_permitted != 0 && !tcp->tcp_active_open))) {
2280 		/* This should be true only in the passive case. */
2281 		if (tcp->tcp_sack_info == NULL) {
2282 			tcp->tcp_sack_info = (tcp_sack_info_t *)bkmem_zalloc(
2283 			    sizeof (tcp_sack_info_t));
2284 		}
2285 		if (tcp->tcp_sack_info == NULL) {
2286 			tcp->tcp_snd_sack_ok = B_FALSE;
2287 		} else {
2288 			tcp->tcp_snd_sack_ok = B_TRUE;
2289 			if (tcp->tcp_snd_ts_ok) {
2290 				tcp->tcp_max_sack_blk = 3;
2291 			} else {
2292 				tcp->tcp_max_sack_blk = 4;
2293 			}
2294 		}
2295 	} else {
2296 		/*
2297 		 * Resetting tcp_snd_sack_ok to B_FALSE so that
2298 		 * no SACK info will be used for this
2299 		 * connection.  This assumes that SACK usage
2300 		 * permission is negotiated.  This may need
2301 		 * to be changed once this is clarified.
2302 		 */
2303 		if (tcp->tcp_sack_info != NULL) {
2304 			bkmem_free((caddr_t)tcp->tcp_sack_info,
2305 			    sizeof (tcp_sack_info_t));
2306 			tcp->tcp_sack_info = NULL;
2307 		}
2308 		tcp->tcp_snd_sack_ok = B_FALSE;
2309 	}
2310 
2311 	/*
2312 	 * Now we know the exact TCP/IP header length, subtract
2313 	 * that from tcp_mss to get our side's MSS.
2314 	 */
2315 	tcp->tcp_mss -= tcp->tcp_hdr_len;
2316 	/*
2317 	 * Here we assume that the other side's header size will be equal to
2318 	 * our header size.  We calculate the real MSS accordingly.  Need to
2319 	 * take into additional stuffs IPsec puts in.
2320 	 *
2321 	 * Real MSS = Opt.MSS - (our TCP/IP header - min TCP/IP header)
2322 	 */
2323 	tcpopt.tcp_opt_mss -= tcp->tcp_hdr_len -
2324 	    (IP_SIMPLE_HDR_LENGTH + TCP_MIN_HEADER_LENGTH);
2325 
2326 	/*
2327 	 * Set MSS to the smaller one of both ends of the connection.
2328 	 * We should not have called tcp_mss_set() before, but our
2329 	 * side of the MSS should have been set to a proper value
2330 	 * by tcp_adapt_ire().  tcp_mss_set() will also set up the
2331 	 * STREAM head parameters properly.
2332 	 *
2333 	 * If we have a larger-than-16-bit window but the other side
2334 	 * didn't want to do window scale, tcp_rwnd_set() will take
2335 	 * care of that.
2336 	 */
2337 	tcp_mss_set(tcp, MIN(tcpopt.tcp_opt_mss, tcp->tcp_mss));
2338 }
2339 
2340 /*
2341  * This function does PAWS protection check.  Returns B_TRUE if the
2342  * segment passes the PAWS test, else returns B_FALSE.
2343  */
2344 boolean_t
2345 tcp_paws_check(tcp_t *tcp, tcph_t *tcph, tcp_opt_t *tcpoptp)
2346 {
2347 	uint8_t	flags;
2348 	int	options;
2349 	uint8_t *up;
2350 
2351 	flags = (unsigned int)tcph->th_flags[0] & 0xFF;
2352 	/*
2353 	 * If timestamp option is aligned nicely, get values inline,
2354 	 * otherwise call general routine to parse.  Only do that
2355 	 * if timestamp is the only option.
2356 	 */
2357 	if (TCP_HDR_LENGTH(tcph) == (uint32_t)TCP_MIN_HEADER_LENGTH +
2358 	    TCPOPT_REAL_TS_LEN &&
2359 	    OK_32PTR((up = ((uint8_t *)tcph) +
2360 	    TCP_MIN_HEADER_LENGTH)) &&
2361 	    *(uint32_t *)up == TCPOPT_NOP_NOP_TSTAMP) {
2362 		tcpoptp->tcp_opt_ts_val = ABE32_TO_U32((up+4));
2363 		tcpoptp->tcp_opt_ts_ecr = ABE32_TO_U32((up+8));
2364 
2365 		options = TCP_OPT_TSTAMP_PRESENT;
2366 	} else {
2367 		if (tcp->tcp_snd_sack_ok) {
2368 			tcpoptp->tcp = tcp;
2369 		} else {
2370 			tcpoptp->tcp = NULL;
2371 		}
2372 		options = tcp_parse_options(tcph, tcpoptp);
2373 	}
2374 
2375 	if (options & TCP_OPT_TSTAMP_PRESENT) {
2376 		/*
2377 		 * Do PAWS per RFC 1323 section 4.2.  Accept RST
2378 		 * regardless of the timestamp, page 18 RFC 1323.bis.
2379 		 */
2380 		if ((flags & TH_RST) == 0 &&
2381 		    TSTMP_LT(tcpoptp->tcp_opt_ts_val,
2382 		    tcp->tcp_ts_recent)) {
2383 			if (TSTMP_LT(prom_gettime(),
2384 			    tcp->tcp_last_rcv_lbolt + PAWS_TIMEOUT)) {
2385 				/* This segment is not acceptable. */
2386 				return (B_FALSE);
2387 			} else {
2388 				/*
2389 				 * Connection has been idle for
2390 				 * too long.  Reset the timestamp
2391 				 * and assume the segment is valid.
2392 				 */
2393 				tcp->tcp_ts_recent =
2394 				    tcpoptp->tcp_opt_ts_val;
2395 			}
2396 		}
2397 	} else {
2398 		/*
2399 		 * If we don't get a timestamp on every packet, we
2400 		 * figure we can't really trust 'em, so we stop sending
2401 		 * and parsing them.
2402 		 */
2403 		tcp->tcp_snd_ts_ok = B_FALSE;
2404 
2405 		tcp->tcp_hdr_len -= TCPOPT_REAL_TS_LEN;
2406 		tcp->tcp_tcp_hdr_len -= TCPOPT_REAL_TS_LEN;
2407 		tcp->tcp_tcph->th_offset_and_rsrvd[0] -= (3 << 4);
2408 		tcp_mss_set(tcp, tcp->tcp_mss + TCPOPT_REAL_TS_LEN);
2409 		if (tcp->tcp_snd_sack_ok) {
2410 			assert(tcp->tcp_sack_info != NULL);
2411 			tcp->tcp_max_sack_blk = 4;
2412 		}
2413 	}
2414 	return (B_TRUE);
2415 }
2416 
2417 /*
2418  * tcp_get_seg_mp() is called to get the pointer to a segment in the
2419  * send queue which starts at the given seq. no.
2420  *
2421  * Parameters:
2422  *	tcp_t *tcp: the tcp instance pointer.
2423  *	uint32_t seq: the starting seq. no of the requested segment.
2424  *	int32_t *off: after the execution, *off will be the offset to
2425  *		the returned mblk which points to the requested seq no.
2426  *
2427  * Return:
2428  *	A mblk_t pointer pointing to the requested segment in send queue.
2429  */
2430 static mblk_t *
2431 tcp_get_seg_mp(tcp_t *tcp, uint32_t seq, int32_t *off)
2432 {
2433 	int32_t	cnt;
2434 	mblk_t	*mp;
2435 
2436 	/* Defensive coding.  Make sure we don't send incorrect data. */
2437 	if (SEQ_LT(seq, tcp->tcp_suna) || SEQ_GEQ(seq, tcp->tcp_snxt) ||
2438 	    off == NULL) {
2439 		return (NULL);
2440 	}
2441 	cnt = seq - tcp->tcp_suna;
2442 	mp = tcp->tcp_xmit_head;
2443 	while (cnt > 0 && mp) {
2444 		cnt -= mp->b_wptr - mp->b_rptr;
2445 		if (cnt < 0) {
2446 			cnt += mp->b_wptr - mp->b_rptr;
2447 			break;
2448 		}
2449 		mp = mp->b_cont;
2450 	}
2451 	assert(mp != NULL);
2452 	*off = cnt;
2453 	return (mp);
2454 }
2455 
2456 /*
2457  * This function handles all retransmissions if SACK is enabled for this
2458  * connection.  First it calculates how many segments can be retransmitted
2459  * based on tcp_pipe.  Then it goes thru the notsack list to find eligible
2460  * segments.  A segment is eligible if sack_cnt for that segment is greater
2461  * than or equal tcp_dupack_fast_retransmit.  After it has retransmitted
2462  * all eligible segments, it checks to see if TCP can send some new segments
2463  * (fast recovery).  If it can, it returns 1.  Otherwise it returns 0.
2464  *
2465  * Parameters:
2466  *	tcp_t *tcp: the tcp structure of the connection.
2467  *
2468  * Return:
2469  *	1 if the pipe is not full (new data can be sent), 0 otherwise
2470  */
2471 static int32_t
2472 tcp_sack_rxmit(tcp_t *tcp, int sock_id)
2473 {
2474 	notsack_blk_t	*notsack_blk;
2475 	int32_t		usable_swnd;
2476 	int32_t		mss;
2477 	uint32_t	seg_len;
2478 	mblk_t		*xmit_mp;
2479 
2480 	assert(tcp->tcp_sack_info != NULL);
2481 	assert(tcp->tcp_notsack_list != NULL);
2482 	assert(tcp->tcp_rexmit == B_FALSE);
2483 
2484 	/* Defensive coding in case there is a bug... */
2485 	if (tcp->tcp_notsack_list == NULL) {
2486 		return (0);
2487 	}
2488 	notsack_blk = tcp->tcp_notsack_list;
2489 	mss = tcp->tcp_mss;
2490 
2491 	/*
2492 	 * Limit the num of outstanding data in the network to be
2493 	 * tcp_cwnd_ssthresh, which is half of the original congestion wnd.
2494 	 */
2495 	usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
2496 
2497 	/* At least retransmit 1 MSS of data. */
2498 	if (usable_swnd <= 0) {
2499 		usable_swnd = mss;
2500 	}
2501 
2502 	/* Make sure no new RTT samples will be taken. */
2503 	tcp->tcp_csuna = tcp->tcp_snxt;
2504 
2505 	notsack_blk = tcp->tcp_notsack_list;
2506 	while (usable_swnd > 0) {
2507 		mblk_t		*snxt_mp, *tmp_mp;
2508 		tcp_seq		begin = tcp->tcp_sack_snxt;
2509 		tcp_seq		end;
2510 		int32_t		off;
2511 
2512 		for (; notsack_blk != NULL; notsack_blk = notsack_blk->next) {
2513 			if (SEQ_GT(notsack_blk->end, begin) &&
2514 			    (notsack_blk->sack_cnt >=
2515 			    tcp_dupack_fast_retransmit)) {
2516 				end = notsack_blk->end;
2517 				if (SEQ_LT(begin, notsack_blk->begin)) {
2518 					begin = notsack_blk->begin;
2519 				}
2520 				break;
2521 			}
2522 		}
2523 		/*
2524 		 * All holes are filled.  Manipulate tcp_cwnd to send more
2525 		 * if we can.  Note that after the SACK recovery, tcp_cwnd is
2526 		 * set to tcp_cwnd_ssthresh.
2527 		 */
2528 		if (notsack_blk == NULL) {
2529 			usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
2530 			if (usable_swnd <= 0) {
2531 				tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna;
2532 				assert(tcp->tcp_cwnd > 0);
2533 				return (0);
2534 			} else {
2535 				usable_swnd = usable_swnd / mss;
2536 				tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna +
2537 				    MAX(usable_swnd * mss, mss);
2538 				return (1);
2539 			}
2540 		}
2541 
2542 		/*
2543 		 * Note that we may send more than usable_swnd allows here
2544 		 * because of round off, but no more than 1 MSS of data.
2545 		 */
2546 		seg_len = end - begin;
2547 		if (seg_len > mss)
2548 			seg_len = mss;
2549 		snxt_mp = tcp_get_seg_mp(tcp, begin, &off);
2550 		assert(snxt_mp != NULL);
2551 		/* This should not happen.  Defensive coding again... */
2552 		if (snxt_mp == NULL) {
2553 			return (0);
2554 		}
2555 
2556 		xmit_mp = tcp_xmit_mp(tcp, snxt_mp, seg_len, &off,
2557 		    &tmp_mp, begin, B_TRUE, &seg_len, B_TRUE);
2558 
2559 		if (xmit_mp == NULL)
2560 			return (0);
2561 
2562 		usable_swnd -= seg_len;
2563 		tcp->tcp_pipe += seg_len;
2564 		tcp->tcp_sack_snxt = begin + seg_len;
2565 		TCP_DUMP_PACKET("tcp_sack_rxmit", xmit_mp);
2566 		(void) ipv4_tcp_output(sock_id, xmit_mp);
2567 		freeb(xmit_mp);
2568 
2569 		/*
2570 		 * Update the send timestamp to avoid false retransmission.
2571 		 */
2572 		snxt_mp->b_prev = (mblk_t *)prom_gettime();
2573 
2574 		BUMP_MIB(tcp_mib.tcpRetransSegs);
2575 		UPDATE_MIB(tcp_mib.tcpRetransBytes, seg_len);
2576 		BUMP_MIB(tcp_mib.tcpOutSackRetransSegs);
2577 		/*
2578 		 * Update tcp_rexmit_max to extend this SACK recovery phase.
2579 		 * This happens when new data sent during fast recovery is
2580 		 * also lost.  If TCP retransmits those new data, it needs
2581 		 * to extend SACK recover phase to avoid starting another
2582 		 * fast retransmit/recovery unnecessarily.
2583 		 */
2584 		if (SEQ_GT(tcp->tcp_sack_snxt, tcp->tcp_rexmit_max)) {
2585 			tcp->tcp_rexmit_max = tcp->tcp_sack_snxt;
2586 		}
2587 	}
2588 	return (0);
2589 }
2590 
2591 static void
2592 tcp_rput_data(tcp_t *tcp, mblk_t *mp, int sock_id)
2593 {
2594 	uchar_t		*rptr;
2595 	struct ip	*iph;
2596 	tcp_t		*tcp1;
2597 	tcpha_t		*tcph;
2598 	uint32_t	seg_ack;
2599 	int		seg_len;
2600 	uint_t		ip_hdr_len;
2601 	uint32_t	seg_seq;
2602 	mblk_t		*mp1;
2603 	uint_t		flags;
2604 	uint32_t	new_swnd = 0;
2605 	int		mss;
2606 	boolean_t	ofo_seg = B_FALSE; /* Out of order segment */
2607 	int32_t		gap;
2608 	int32_t		rgap;
2609 	tcp_opt_t	tcpopt;
2610 	int32_t		bytes_acked;
2611 	int		npkt;
2612 	uint32_t	cwnd;
2613 	uint32_t	add;
2614 
2615 #ifdef DEBUG
2616 	printf("tcp_rput_data sock %d mp %x mp_datap %x #################\n",
2617 	    sock_id, mp, mp->b_datap);
2618 #endif
2619 
2620 	/* Dump the packet when debugging. */
2621 	TCP_DUMP_PACKET("tcp_rput_data", mp);
2622 
2623 	assert(OK_32PTR(mp->b_rptr));
2624 
2625 	rptr = mp->b_rptr;
2626 	iph = (struct ip *)rptr;
2627 	ip_hdr_len = IPH_HDR_LENGTH(rptr);
2628 	if (ip_hdr_len != IP_SIMPLE_HDR_LENGTH) {
2629 #ifdef DEBUG
2630 		printf("Not simple IP header\n");
2631 #endif
2632 		/* We cannot handle IP option yet... */
2633 		tcp_drops++;
2634 		freeb(mp);
2635 		return;
2636 	}
2637 	/* The TCP header must be aligned. */
2638 	tcph = (tcpha_t *)&rptr[ip_hdr_len];
2639 	seg_seq = ntohl(tcph->tha_seq);
2640 	seg_ack = ntohl(tcph->tha_ack);
2641 	assert((uintptr_t)(mp->b_wptr - rptr) <= (uintptr_t)INT_MAX);
2642 	seg_len = (int)(mp->b_wptr - rptr) -
2643 	    (ip_hdr_len + TCP_HDR_LENGTH(((tcph_t *)tcph)));
2644 	/* In inetboot, b_cont should always be NULL. */
2645 	assert(mp->b_cont == NULL);
2646 
2647 	/* Verify the checksum. */
2648 	if (tcp_verify_cksum(mp) < 0) {
2649 #ifdef DEBUG
2650 		printf("tcp_rput_data: wrong cksum\n");
2651 #endif
2652 		freemsg(mp);
2653 		return;
2654 	}
2655 
2656 	/*
2657 	 * This segment is not for us, try to find its
2658 	 * intended receiver.
2659 	 */
2660 	if (tcp == NULL ||
2661 	    tcph->tha_lport != tcp->tcp_fport ||
2662 	    tcph->tha_fport != tcp->tcp_lport ||
2663 	    iph->ip_src.s_addr != tcp->tcp_remote ||
2664 	    iph->ip_dst.s_addr != tcp->tcp_bound_source) {
2665 #ifdef DEBUG
2666 		printf("tcp_rput_data: not for us, state %d\n",
2667 		    tcp->tcp_state);
2668 #endif
2669 		/*
2670 		 * First try to find a established connection.  If none
2671 		 * is found, look for a listener.
2672 		 *
2673 		 * If a listener is found, we need to check to see if the
2674 		 * incoming segment is for one of its eagers.  If it is,
2675 		 * give it to the eager.  If not, listener should take care
2676 		 * of it.
2677 		 */
2678 		if ((tcp1 = tcp_lookup_ipv4(iph, tcph, TCPS_SYN_SENT,
2679 		    &sock_id)) != NULL ||
2680 		    (tcp1 = tcp_lookup_listener_ipv4(iph->ip_dst.s_addr,
2681 		    tcph->tha_fport, &sock_id)) != NULL) {
2682 			if (tcp1->tcp_state == TCPS_LISTEN) {
2683 				if ((tcp = tcp_lookup_eager_ipv4(tcp1,
2684 				    iph, tcph)) == NULL) {
2685 					/* No eager... sent to listener */
2686 #ifdef DEBUG
2687 					printf("found the listener: %s\n",
2688 					    tcp_display(tcp1, NULL,
2689 					    DISP_ADDR_AND_PORT));
2690 #endif
2691 					tcp = tcp1;
2692 				}
2693 #ifdef DEBUG
2694 				else {
2695 					printf("found the eager: %s\n",
2696 					    tcp_display(tcp, NULL,
2697 					    DISP_ADDR_AND_PORT));
2698 				}
2699 #endif
2700 			} else {
2701 				/* Non listener found... */
2702 #ifdef DEBUG
2703 				printf("found the connection: %s\n",
2704 				    tcp_display(tcp1, NULL,
2705 				    DISP_ADDR_AND_PORT));
2706 #endif
2707 				tcp = tcp1;
2708 			}
2709 		} else {
2710 			/*
2711 			 * No connection for this segment...
2712 			 * Send a RST to the other side.
2713 			 */
2714 			tcp_xmit_listeners_reset(sock_id, mp, ip_hdr_len);
2715 			return;
2716 		}
2717 	}
2718 
2719 	flags = tcph->tha_flags & 0xFF;
2720 	BUMP_MIB(tcp_mib.tcpInSegs);
2721 	if (tcp->tcp_state == TCPS_TIME_WAIT) {
2722 		tcp_time_wait_processing(tcp, mp, seg_seq, seg_ack,
2723 		    seg_len, (tcph_t *)tcph, sock_id);
2724 		return;
2725 	}
2726 	/*
2727 	 * From this point we can assume that the tcp is not compressed,
2728 	 * since we would have branched off to tcp_time_wait_processing()
2729 	 * in such a case.
2730 	 */
2731 	assert(tcp != NULL && tcp->tcp_state != TCPS_TIME_WAIT);
2732 
2733 	/*
2734 	 * After this point, we know we have the correct TCP, so update
2735 	 * the receive time.
2736 	 */
2737 	tcp->tcp_last_recv_time = prom_gettime();
2738 
2739 	/* In inetboot, we do not handle urgent pointer... */
2740 	if (flags & TH_URG) {
2741 		freemsg(mp);
2742 		DEBUG_1("tcp_rput_data(%d): received segment with urgent "
2743 		    "pointer\n", sock_id);
2744 		tcp_drops++;
2745 		return;
2746 	}
2747 
2748 	switch (tcp->tcp_state) {
2749 	case TCPS_LISTEN:
2750 		if ((flags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
2751 			if (flags & TH_RST) {
2752 				freemsg(mp);
2753 				return;
2754 			}
2755 			if (flags & TH_ACK) {
2756 				tcp_xmit_early_reset("TCPS_LISTEN-TH_ACK",
2757 				    sock_id, mp, seg_ack, 0, TH_RST,
2758 				    ip_hdr_len);
2759 				return;
2760 			}
2761 			if (!(flags & TH_SYN)) {
2762 				freemsg(mp);
2763 				return;
2764 			}
2765 			printf("tcp_rput_data: %d\n", __LINE__);
2766 			prom_panic("inetboot");
2767 		}
2768 		if (tcp->tcp_conn_req_max > 0) {
2769 			tcp = tcp_conn_request(tcp, mp, sock_id, ip_hdr_len);
2770 			if (tcp == NULL) {
2771 				freemsg(mp);
2772 				return;
2773 			}
2774 #ifdef DEBUG
2775 			printf("tcp_rput_data: new tcp created\n");
2776 #endif
2777 		}
2778 		tcp->tcp_irs = seg_seq;
2779 		tcp->tcp_rack = seg_seq;
2780 		tcp->tcp_rnxt = seg_seq + 1;
2781 		U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
2782 		BUMP_MIB(tcp_mib.tcpPassiveOpens);
2783 		goto syn_rcvd;
2784 	case TCPS_SYN_SENT:
2785 		if (flags & TH_ACK) {
2786 			/*
2787 			 * Note that our stack cannot send data before a
2788 			 * connection is established, therefore the
2789 			 * following check is valid.  Otherwise, it has
2790 			 * to be changed.
2791 			 */
2792 			if (SEQ_LEQ(seg_ack, tcp->tcp_iss) ||
2793 			    SEQ_GT(seg_ack, tcp->tcp_snxt)) {
2794 				if (flags & TH_RST) {
2795 					freemsg(mp);
2796 					return;
2797 				}
2798 				tcp_xmit_ctl("TCPS_SYN_SENT-Bad_seq",
2799 				    tcp, mp, seg_ack, 0, TH_RST,
2800 				    ip_hdr_len, sock_id);
2801 				return;
2802 			}
2803 			assert(tcp->tcp_suna + 1 == seg_ack);
2804 		}
2805 		if (flags & TH_RST) {
2806 			freemsg(mp);
2807 			if (flags & TH_ACK) {
2808 				tcp_clean_death(sock_id, tcp, ECONNREFUSED);
2809 			}
2810 			return;
2811 		}
2812 		if (!(flags & TH_SYN)) {
2813 			freemsg(mp);
2814 			return;
2815 		}
2816 
2817 		/* Process all TCP options. */
2818 		tcp_process_options(tcp, (tcph_t *)tcph);
2819 		/*
2820 		 * The following changes our rwnd to be a multiple of the
2821 		 * MIN(peer MSS, our MSS) for performance reason.
2822 		 */
2823 		(void) tcp_rwnd_set(tcp, MSS_ROUNDUP(tcp->tcp_rwnd,
2824 		    tcp->tcp_mss));
2825 
2826 		/* Is the other end ECN capable? */
2827 		if (tcp->tcp_ecn_ok) {
2828 			if ((flags & (TH_ECE|TH_CWR)) != TH_ECE) {
2829 				tcp->tcp_ecn_ok = B_FALSE;
2830 			}
2831 		}
2832 		/*
2833 		 * Clear ECN flags because it may interfere with later
2834 		 * processing.
2835 		 */
2836 		flags &= ~(TH_ECE|TH_CWR);
2837 
2838 		tcp->tcp_irs = seg_seq;
2839 		tcp->tcp_rack = seg_seq;
2840 		tcp->tcp_rnxt = seg_seq + 1;
2841 		U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
2842 
2843 		if (flags & TH_ACK) {
2844 			/* One for the SYN */
2845 			tcp->tcp_suna = tcp->tcp_iss + 1;
2846 			tcp->tcp_valid_bits &= ~TCP_ISS_VALID;
2847 			tcp->tcp_state = TCPS_ESTABLISHED;
2848 
2849 			/*
2850 			 * If SYN was retransmitted, need to reset all
2851 			 * retransmission info.  This is because this
2852 			 * segment will be treated as a dup ACK.
2853 			 */
2854 			if (tcp->tcp_rexmit) {
2855 				tcp->tcp_rexmit = B_FALSE;
2856 				tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
2857 				tcp->tcp_rexmit_max = tcp->tcp_snxt;
2858 				tcp->tcp_snd_burst = TCP_CWND_NORMAL;
2859 
2860 				/*
2861 				 * Set tcp_cwnd back to 1 MSS, per
2862 				 * recommendation from
2863 				 * draft-floyd-incr-init-win-01.txt,
2864 				 * Increasing TCP's Initial Window.
2865 				 */
2866 				tcp->tcp_cwnd = tcp->tcp_mss;
2867 			}
2868 
2869 			tcp->tcp_swl1 = seg_seq;
2870 			tcp->tcp_swl2 = seg_ack;
2871 
2872 			new_swnd = BE16_TO_U16(((tcph_t *)tcph)->th_win);
2873 			tcp->tcp_swnd = new_swnd;
2874 			if (new_swnd > tcp->tcp_max_swnd)
2875 				tcp->tcp_max_swnd = new_swnd;
2876 
2877 			/*
2878 			 * Always send the three-way handshake ack immediately
2879 			 * in order to make the connection complete as soon as
2880 			 * possible on the accepting host.
2881 			 */
2882 			flags |= TH_ACK_NEEDED;
2883 			/*
2884 			 * Check to see if there is data to be sent.  If
2885 			 * yes, set the transmit flag.  Then check to see
2886 			 * if received data processing needs to be done.
2887 			 * If not, go straight to xmit_check.  This short
2888 			 * cut is OK as we don't support T/TCP.
2889 			 */
2890 			if (tcp->tcp_unsent)
2891 				flags |= TH_XMIT_NEEDED;
2892 
2893 			if (seg_len == 0) {
2894 				freemsg(mp);
2895 				goto xmit_check;
2896 			}
2897 
2898 			flags &= ~TH_SYN;
2899 			seg_seq++;
2900 			break;
2901 		}
2902 		syn_rcvd:
2903 		tcp->tcp_state = TCPS_SYN_RCVD;
2904 		mp1 = tcp_xmit_mp(tcp, tcp->tcp_xmit_head, tcp->tcp_mss,
2905 		    NULL, NULL, tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
2906 		if (mp1 != NULL) {
2907 			TCP_DUMP_PACKET("tcp_rput_data replying SYN", mp1);
2908 			(void) ipv4_tcp_output(sock_id, mp1);
2909 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
2910 			freeb(mp1);
2911 			/*
2912 			 * Let's wait till our SYN has been ACKED since we
2913 			 * don't have a timer.
2914 			 */
2915 			if (tcp_state_wait(sock_id, tcp, TCPS_ALL_ACKED) < 0) {
2916 				freemsg(mp);
2917 				return;
2918 			}
2919 		}
2920 		freemsg(mp);
2921 		return;
2922 	default:
2923 		break;
2924 	}
2925 	mp->b_rptr = (uchar_t *)tcph + TCP_HDR_LENGTH((tcph_t *)tcph);
2926 	new_swnd = ntohs(tcph->tha_win) <<
2927 	    ((flags & TH_SYN) ? 0 : tcp->tcp_snd_ws);
2928 	mss = tcp->tcp_mss;
2929 
2930 	if (tcp->tcp_snd_ts_ok) {
2931 		if (!tcp_paws_check(tcp, (tcph_t *)tcph, &tcpopt)) {
2932 			/*
2933 			 * This segment is not acceptable.
2934 			 * Drop it and send back an ACK.
2935 			 */
2936 			freemsg(mp);
2937 			flags |= TH_ACK_NEEDED;
2938 			goto ack_check;
2939 		}
2940 	} else if (tcp->tcp_snd_sack_ok) {
2941 		assert(tcp->tcp_sack_info != NULL);
2942 		tcpopt.tcp = tcp;
2943 		/*
2944 		 * SACK info in already updated in tcp_parse_options.  Ignore
2945 		 * all other TCP options...
2946 		 */
2947 		(void) tcp_parse_options((tcph_t *)tcph, &tcpopt);
2948 	}
2949 try_again:;
2950 	gap = seg_seq - tcp->tcp_rnxt;
2951 	rgap = tcp->tcp_rwnd - (gap + seg_len);
2952 	/*
2953 	 * gap is the amount of sequence space between what we expect to see
2954 	 * and what we got for seg_seq.  A positive value for gap means
2955 	 * something got lost.  A negative value means we got some old stuff.
2956 	 */
2957 	if (gap < 0) {
2958 		/* Old stuff present.  Is the SYN in there? */
2959 		if (seg_seq == tcp->tcp_irs && (flags & TH_SYN) &&
2960 		    (seg_len != 0)) {
2961 			flags &= ~TH_SYN;
2962 			seg_seq++;
2963 			/* Recompute the gaps after noting the SYN. */
2964 			goto try_again;
2965 		}
2966 		BUMP_MIB(tcp_mib.tcpInDataDupSegs);
2967 		UPDATE_MIB(tcp_mib.tcpInDataDupBytes,
2968 		    (seg_len > -gap ? -gap : seg_len));
2969 		/* Remove the old stuff from seg_len. */
2970 		seg_len += gap;
2971 		/*
2972 		 * Anything left?
2973 		 * Make sure to check for unack'd FIN when rest of data
2974 		 * has been previously ack'd.
2975 		 */
2976 		if (seg_len < 0 || (seg_len == 0 && !(flags & TH_FIN))) {
2977 			/*
2978 			 * Resets are only valid if they lie within our offered
2979 			 * window.  If the RST bit is set, we just ignore this
2980 			 * segment.
2981 			 */
2982 			if (flags & TH_RST) {
2983 				freemsg(mp);
2984 				return;
2985 			}
2986 
2987 			/*
2988 			 * This segment is "unacceptable".  None of its
2989 			 * sequence space lies within our advertized window.
2990 			 *
2991 			 * Adjust seg_len to the original value for tracing.
2992 			 */
2993 			seg_len -= gap;
2994 #ifdef DEBUG
2995 			printf("tcp_rput: unacceptable, gap %d, rgap "
2996 			    "%d, flags 0x%x, seg_seq %u, seg_ack %u, "
2997 			    "seg_len %d, rnxt %u, snxt %u, %s",
2998 			    gap, rgap, flags, seg_seq, seg_ack,
2999 			    seg_len, tcp->tcp_rnxt, tcp->tcp_snxt,
3000 			    tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
3001 #endif
3002 
3003 			/*
3004 			 * Arrange to send an ACK in response to the
3005 			 * unacceptable segment per RFC 793 page 69. There
3006 			 * is only one small difference between ours and the
3007 			 * acceptability test in the RFC - we accept ACK-only
3008 			 * packet with SEG.SEQ = RCV.NXT+RCV.WND and no ACK
3009 			 * will be generated.
3010 			 *
3011 			 * Note that we have to ACK an ACK-only packet at least
3012 			 * for stacks that send 0-length keep-alives with
3013 			 * SEG.SEQ = SND.NXT-1 as recommended by RFC1122,
3014 			 * section 4.2.3.6. As long as we don't ever generate
3015 			 * an unacceptable packet in response to an incoming
3016 			 * packet that is unacceptable, it should not cause
3017 			 * "ACK wars".
3018 			 */
3019 			flags |=  TH_ACK_NEEDED;
3020 
3021 			/*
3022 			 * Continue processing this segment in order to use the
3023 			 * ACK information it contains, but skip all other
3024 			 * sequence-number processing.	Processing the ACK
3025 			 * information is necessary in order to
3026 			 * re-synchronize connections that may have lost
3027 			 * synchronization.
3028 			 *
3029 			 * We clear seg_len and flag fields related to
3030 			 * sequence number processing as they are not
3031 			 * to be trusted for an unacceptable segment.
3032 			 */
3033 			seg_len = 0;
3034 			flags &= ~(TH_SYN | TH_FIN | TH_URG);
3035 			goto process_ack;
3036 		}
3037 
3038 		/* Fix seg_seq, and chew the gap off the front. */
3039 		seg_seq = tcp->tcp_rnxt;
3040 		do {
3041 			mblk_t	*mp2;
3042 			assert((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
3043 			    (uintptr_t)UINT_MAX);
3044 			gap += (uint_t)(mp->b_wptr - mp->b_rptr);
3045 			if (gap > 0) {
3046 				mp->b_rptr = mp->b_wptr - gap;
3047 				break;
3048 			}
3049 			mp2 = mp;
3050 			mp = mp->b_cont;
3051 			freeb(mp2);
3052 		} while (gap < 0);
3053 	}
3054 	/*
3055 	 * rgap is the amount of stuff received out of window.  A negative
3056 	 * value is the amount out of window.
3057 	 */
3058 	if (rgap < 0) {
3059 		mblk_t	*mp2;
3060 
3061 		if (tcp->tcp_rwnd == 0)
3062 			BUMP_MIB(tcp_mib.tcpInWinProbe);
3063 		else {
3064 			BUMP_MIB(tcp_mib.tcpInDataPastWinSegs);
3065 			UPDATE_MIB(tcp_mib.tcpInDataPastWinBytes, -rgap);
3066 		}
3067 
3068 		/*
3069 		 * seg_len does not include the FIN, so if more than
3070 		 * just the FIN is out of window, we act like we don't
3071 		 * see it.  (If just the FIN is out of window, rgap
3072 		 * will be zero and we will go ahead and acknowledge
3073 		 * the FIN.)
3074 		 */
3075 		flags &= ~TH_FIN;
3076 
3077 		/* Fix seg_len and make sure there is something left. */
3078 		seg_len += rgap;
3079 		if (seg_len <= 0) {
3080 			/*
3081 			 * Resets are only valid if they lie within our offered
3082 			 * window.  If the RST bit is set, we just ignore this
3083 			 * segment.
3084 			 */
3085 			if (flags & TH_RST) {
3086 				freemsg(mp);
3087 				return;
3088 			}
3089 
3090 			/* Per RFC 793, we need to send back an ACK. */
3091 			flags |= TH_ACK_NEEDED;
3092 
3093 			/*
3094 			 * If this is a zero window probe, continue to
3095 			 * process the ACK part.  But we need to set seg_len
3096 			 * to 0 to avoid data processing.  Otherwise just
3097 			 * drop the segment and send back an ACK.
3098 			 */
3099 			if (tcp->tcp_rwnd == 0 && seg_seq == tcp->tcp_rnxt) {
3100 				flags &= ~(TH_SYN | TH_URG);
3101 				seg_len = 0;
3102 				/* Let's see if we can update our rwnd */
3103 				tcp_rcv_drain(sock_id, tcp);
3104 				goto process_ack;
3105 			} else {
3106 				freemsg(mp);
3107 				goto ack_check;
3108 			}
3109 		}
3110 		/* Pitch out of window stuff off the end. */
3111 		rgap = seg_len;
3112 		mp2 = mp;
3113 		do {
3114 			assert((uintptr_t)(mp2->b_wptr -
3115 			    mp2->b_rptr) <= (uintptr_t)INT_MAX);
3116 			rgap -= (int)(mp2->b_wptr - mp2->b_rptr);
3117 			if (rgap < 0) {
3118 				mp2->b_wptr += rgap;
3119 				if ((mp1 = mp2->b_cont) != NULL) {
3120 					mp2->b_cont = NULL;
3121 					freemsg(mp1);
3122 				}
3123 				break;
3124 			}
3125 		} while ((mp2 = mp2->b_cont) != NULL);
3126 	}
3127 ok:;
3128 	/*
3129 	 * TCP should check ECN info for segments inside the window only.
3130 	 * Therefore the check should be done here.
3131 	 */
3132 	if (tcp->tcp_ecn_ok) {
3133 		uchar_t tos = ((struct ip *)rptr)->ip_tos;
3134 
3135 		if (flags & TH_CWR) {
3136 			tcp->tcp_ecn_echo_on = B_FALSE;
3137 		}
3138 		/*
3139 		 * Note that both ECN_CE and CWR can be set in the
3140 		 * same segment.  In this case, we once again turn
3141 		 * on ECN_ECHO.
3142 		 */
3143 		if ((tos & IPH_ECN_CE) == IPH_ECN_CE) {
3144 			tcp->tcp_ecn_echo_on = B_TRUE;
3145 		}
3146 	}
3147 
3148 	/*
3149 	 * Check whether we can update tcp_ts_recent.  This test is
3150 	 * NOT the one in RFC 1323 3.4.  It is from Braden, 1993, "TCP
3151 	 * Extensions for High Performance: An Update", Internet Draft.
3152 	 */
3153 	if (tcp->tcp_snd_ts_ok &&
3154 	    TSTMP_GEQ(tcpopt.tcp_opt_ts_val, tcp->tcp_ts_recent) &&
3155 	    SEQ_LEQ(seg_seq, tcp->tcp_rack)) {
3156 		tcp->tcp_ts_recent = tcpopt.tcp_opt_ts_val;
3157 		tcp->tcp_last_rcv_lbolt = prom_gettime();
3158 	}
3159 
3160 	if (seg_seq != tcp->tcp_rnxt || tcp->tcp_reass_head) {
3161 		/*
3162 		 * FIN in an out of order segment.  We record this in
3163 		 * tcp_valid_bits and the seq num of FIN in tcp_ofo_fin_seq.
3164 		 * Clear the FIN so that any check on FIN flag will fail.
3165 		 * Remember that FIN also counts in the sequence number
3166 		 * space.  So we need to ack out of order FIN only segments.
3167 		 */
3168 		if (flags & TH_FIN) {
3169 			tcp->tcp_valid_bits |= TCP_OFO_FIN_VALID;
3170 			tcp->tcp_ofo_fin_seq = seg_seq + seg_len;
3171 			flags &= ~TH_FIN;
3172 			flags |= TH_ACK_NEEDED;
3173 		}
3174 		if (seg_len > 0) {
3175 			/* Fill in the SACK blk list. */
3176 			if (tcp->tcp_snd_sack_ok) {
3177 				assert(tcp->tcp_sack_info != NULL);
3178 				tcp_sack_insert(tcp->tcp_sack_list,
3179 				    seg_seq, seg_seq + seg_len,
3180 				    &(tcp->tcp_num_sack_blk));
3181 			}
3182 
3183 			/*
3184 			 * Attempt reassembly and see if we have something
3185 			 * ready to go.
3186 			 */
3187 			mp = tcp_reass(tcp, mp, seg_seq);
3188 			/* Always ack out of order packets */
3189 			flags |= TH_ACK_NEEDED | TH_PUSH;
3190 			if (mp != NULL) {
3191 				assert((uintptr_t)(mp->b_wptr -
3192 				    mp->b_rptr) <= (uintptr_t)INT_MAX);
3193 				seg_len = mp->b_cont ? msgdsize(mp) :
3194 					(int)(mp->b_wptr - mp->b_rptr);
3195 				seg_seq = tcp->tcp_rnxt;
3196 				/*
3197 				 * A gap is filled and the seq num and len
3198 				 * of the gap match that of a previously
3199 				 * received FIN, put the FIN flag back in.
3200 				 */
3201 				if ((tcp->tcp_valid_bits & TCP_OFO_FIN_VALID) &&
3202 				    seg_seq + seg_len == tcp->tcp_ofo_fin_seq) {
3203 					flags |= TH_FIN;
3204 					tcp->tcp_valid_bits &=
3205 					    ~TCP_OFO_FIN_VALID;
3206 				}
3207 			} else {
3208 				/*
3209 				 * Keep going even with NULL mp.
3210 				 * There may be a useful ACK or something else
3211 				 * we don't want to miss.
3212 				 *
3213 				 * But TCP should not perform fast retransmit
3214 				 * because of the ack number.  TCP uses
3215 				 * seg_len == 0 to determine if it is a pure
3216 				 * ACK.  And this is not a pure ACK.
3217 				 */
3218 				seg_len = 0;
3219 				ofo_seg = B_TRUE;
3220 			}
3221 		}
3222 	} else if (seg_len > 0) {
3223 		BUMP_MIB(tcp_mib.tcpInDataInorderSegs);
3224 		UPDATE_MIB(tcp_mib.tcpInDataInorderBytes, seg_len);
3225 		/*
3226 		 * If an out of order FIN was received before, and the seq
3227 		 * num and len of the new segment match that of the FIN,
3228 		 * put the FIN flag back in.
3229 		 */
3230 		if ((tcp->tcp_valid_bits & TCP_OFO_FIN_VALID) &&
3231 		    seg_seq + seg_len == tcp->tcp_ofo_fin_seq) {
3232 			flags |= TH_FIN;
3233 			tcp->tcp_valid_bits &= ~TCP_OFO_FIN_VALID;
3234 		}
3235 	}
3236 	if ((flags & (TH_RST | TH_SYN | TH_URG | TH_ACK)) != TH_ACK) {
3237 	if (flags & TH_RST) {
3238 		freemsg(mp);
3239 		switch (tcp->tcp_state) {
3240 		case TCPS_SYN_RCVD:
3241 			(void) tcp_clean_death(sock_id, tcp, ECONNREFUSED);
3242 			break;
3243 		case TCPS_ESTABLISHED:
3244 		case TCPS_FIN_WAIT_1:
3245 		case TCPS_FIN_WAIT_2:
3246 		case TCPS_CLOSE_WAIT:
3247 			(void) tcp_clean_death(sock_id, tcp, ECONNRESET);
3248 			break;
3249 		case TCPS_CLOSING:
3250 		case TCPS_LAST_ACK:
3251 			(void) tcp_clean_death(sock_id, tcp, 0);
3252 			break;
3253 		default:
3254 			assert(tcp->tcp_state != TCPS_TIME_WAIT);
3255 			(void) tcp_clean_death(sock_id, tcp, ENXIO);
3256 			break;
3257 		}
3258 		return;
3259 	}
3260 	if (flags & TH_SYN) {
3261 		/*
3262 		 * See RFC 793, Page 71
3263 		 *
3264 		 * The seq number must be in the window as it should
3265 		 * be "fixed" above.  If it is outside window, it should
3266 		 * be already rejected.  Note that we allow seg_seq to be
3267 		 * rnxt + rwnd because we want to accept 0 window probe.
3268 		 */
3269 		assert(SEQ_GEQ(seg_seq, tcp->tcp_rnxt) &&
3270 		    SEQ_LEQ(seg_seq, tcp->tcp_rnxt + tcp->tcp_rwnd));
3271 		freemsg(mp);
3272 		/*
3273 		 * If the ACK flag is not set, just use our snxt as the
3274 		 * seq number of the RST segment.
3275 		 */
3276 		if (!(flags & TH_ACK)) {
3277 			seg_ack = tcp->tcp_snxt;
3278 		}
3279 		tcp_xmit_ctl("TH_SYN", tcp, NULL, seg_ack,
3280 		    seg_seq + 1, TH_RST|TH_ACK, 0, sock_id);
3281 		assert(tcp->tcp_state != TCPS_TIME_WAIT);
3282 		(void) tcp_clean_death(sock_id, tcp, ECONNRESET);
3283 		return;
3284 	}
3285 
3286 process_ack:
3287 	if (!(flags & TH_ACK)) {
3288 #ifdef DEBUG
3289 		printf("No ack in segment, dropped it, seq:%x\n", seg_seq);
3290 #endif
3291 		freemsg(mp);
3292 		goto xmit_check;
3293 	}
3294 	}
3295 	bytes_acked = (int)(seg_ack - tcp->tcp_suna);
3296 
3297 	if (tcp->tcp_state == TCPS_SYN_RCVD) {
3298 		tcp_t	*listener = tcp->tcp_listener;
3299 #ifdef DEBUG
3300 		printf("Done with eager 3-way handshake\n");
3301 #endif
3302 		/*
3303 		 * NOTE: RFC 793 pg. 72 says this should be 'bytes_acked < 0'
3304 		 * but that would mean we have an ack that ignored our SYN.
3305 		 */
3306 		if (bytes_acked < 1 || SEQ_GT(seg_ack, tcp->tcp_snxt)) {
3307 			freemsg(mp);
3308 			tcp_xmit_ctl("TCPS_SYN_RCVD-bad_ack",
3309 			    tcp, NULL, seg_ack, 0, TH_RST, 0, sock_id);
3310 			return;
3311 		}
3312 
3313 		/*
3314 		 * if the conn_req_q is full defer processing
3315 		 * until space is availabe after accept()
3316 		 * processing
3317 		 */
3318 		if (listener->tcp_conn_req_cnt_q <
3319 		    listener->tcp_conn_req_max) {
3320 			tcp_t *tail;
3321 
3322 			listener->tcp_conn_req_cnt_q0--;
3323 			listener->tcp_conn_req_cnt_q++;
3324 
3325 			/* Move from SYN_RCVD to ESTABLISHED list  */
3326 			tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
3327 				tcp->tcp_eager_prev_q0;
3328 			tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
3329 				tcp->tcp_eager_next_q0;
3330 			tcp->tcp_eager_prev_q0 = NULL;
3331 			tcp->tcp_eager_next_q0 = NULL;
3332 
3333 			/*
3334 			 * Insert at end of the queue because sockfs
3335 			 * sends down T_CONN_RES in chronological
3336 			 * order. Leaving the older conn indications
3337 			 * at front of the queue helps reducing search
3338 			 * time.
3339 			 */
3340 			tail = listener->tcp_eager_last_q;
3341 			if (tail != NULL) {
3342 				tail->tcp_eager_next_q = tcp;
3343 			} else {
3344 				listener->tcp_eager_next_q = tcp;
3345 			}
3346 			listener->tcp_eager_last_q = tcp;
3347 			tcp->tcp_eager_next_q = NULL;
3348 		} else {
3349 			/*
3350 			 * Defer connection on q0 and set deferred
3351 			 * connection bit true
3352 			 */
3353 			tcp->tcp_conn_def_q0 = B_TRUE;
3354 
3355 			/* take tcp out of q0 ... */
3356 			tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
3357 			    tcp->tcp_eager_next_q0;
3358 			tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
3359 			    tcp->tcp_eager_prev_q0;
3360 
3361 			/* ... and place it at the end of q0 */
3362 			tcp->tcp_eager_prev_q0 = listener->tcp_eager_prev_q0;
3363 			tcp->tcp_eager_next_q0 = listener;
3364 			listener->tcp_eager_prev_q0->tcp_eager_next_q0 = tcp;
3365 			listener->tcp_eager_prev_q0 = tcp;
3366 		}
3367 
3368 		tcp->tcp_suna = tcp->tcp_iss + 1;	/* One for the SYN */
3369 		bytes_acked--;
3370 
3371 		/*
3372 		 * If SYN was retransmitted, need to reset all
3373 		 * retransmission info as this segment will be
3374 		 * treated as a dup ACK.
3375 		 */
3376 		if (tcp->tcp_rexmit) {
3377 			tcp->tcp_rexmit = B_FALSE;
3378 			tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3379 			tcp->tcp_rexmit_max = tcp->tcp_snxt;
3380 			tcp->tcp_snd_burst = TCP_CWND_NORMAL;
3381 			tcp->tcp_ms_we_have_waited = 0;
3382 			tcp->tcp_cwnd = mss;
3383 		}
3384 
3385 		/*
3386 		 * We set the send window to zero here.
3387 		 * This is needed if there is data to be
3388 		 * processed already on the queue.
3389 		 * Later (at swnd_update label), the
3390 		 * "new_swnd > tcp_swnd" condition is satisfied
3391 		 * the XMIT_NEEDED flag is set in the current
3392 		 * (SYN_RCVD) state. This ensures tcp_wput_data() is
3393 		 * called if there is already data on queue in
3394 		 * this state.
3395 		 */
3396 		tcp->tcp_swnd = 0;
3397 
3398 		if (new_swnd > tcp->tcp_max_swnd)
3399 			tcp->tcp_max_swnd = new_swnd;
3400 		tcp->tcp_swl1 = seg_seq;
3401 		tcp->tcp_swl2 = seg_ack;
3402 		tcp->tcp_state = TCPS_ESTABLISHED;
3403 		tcp->tcp_valid_bits &= ~TCP_ISS_VALID;
3404 	}
3405 	/* This code follows 4.4BSD-Lite2 mostly. */
3406 	if (bytes_acked < 0)
3407 		goto est;
3408 
3409 	/*
3410 	 * If TCP is ECN capable and the congestion experience bit is
3411 	 * set, reduce tcp_cwnd and tcp_ssthresh.  But this should only be
3412 	 * done once per window (or more loosely, per RTT).
3413 	 */
3414 	if (tcp->tcp_cwr && SEQ_GT(seg_ack, tcp->tcp_cwr_snd_max))
3415 		tcp->tcp_cwr = B_FALSE;
3416 	if (tcp->tcp_ecn_ok && (flags & TH_ECE)) {
3417 		if (!tcp->tcp_cwr) {
3418 			npkt = (MIN(tcp->tcp_cwnd, tcp->tcp_swnd) >> 1) / mss;
3419 			tcp->tcp_cwnd_ssthresh = MAX(npkt, 2) * mss;
3420 			tcp->tcp_cwnd = npkt * mss;
3421 			/*
3422 			 * If the cwnd is 0, use the timer to clock out
3423 			 * new segments.  This is required by the ECN spec.
3424 			 */
3425 			if (npkt == 0) {
3426 				TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3427 				/*
3428 				 * This makes sure that when the ACK comes
3429 				 * back, we will increase tcp_cwnd by 1 MSS.
3430 				 */
3431 				tcp->tcp_cwnd_cnt = 0;
3432 			}
3433 			tcp->tcp_cwr = B_TRUE;
3434 			/*
3435 			 * This marks the end of the current window of in
3436 			 * flight data.  That is why we don't use
3437 			 * tcp_suna + tcp_swnd.  Only data in flight can
3438 			 * provide ECN info.
3439 			 */
3440 			tcp->tcp_cwr_snd_max = tcp->tcp_snxt;
3441 			tcp->tcp_ecn_cwr_sent = B_FALSE;
3442 		}
3443 	}
3444 
3445 	mp1 = tcp->tcp_xmit_head;
3446 	if (bytes_acked == 0) {
3447 		if (!ofo_seg && seg_len == 0 && new_swnd == tcp->tcp_swnd) {
3448 			int dupack_cnt;
3449 
3450 			BUMP_MIB(tcp_mib.tcpInDupAck);
3451 			/*
3452 			 * Fast retransmit.  When we have seen exactly three
3453 			 * identical ACKs while we have unacked data
3454 			 * outstanding we take it as a hint that our peer
3455 			 * dropped something.
3456 			 *
3457 			 * If TCP is retransmitting, don't do fast retransmit.
3458 			 */
3459 			if (mp1 != NULL && tcp->tcp_suna != tcp->tcp_snxt &&
3460 			    ! tcp->tcp_rexmit) {
3461 				/* Do Limited Transmit */
3462 				if ((dupack_cnt = ++tcp->tcp_dupack_cnt) <
3463 				    tcp_dupack_fast_retransmit) {
3464 					/*
3465 					 * RFC 3042
3466 					 *
3467 					 * What we need to do is temporarily
3468 					 * increase tcp_cwnd so that new
3469 					 * data can be sent if it is allowed
3470 					 * by the receive window (tcp_rwnd).
3471 					 * tcp_wput_data() will take care of
3472 					 * the rest.
3473 					 *
3474 					 * If the connection is SACK capable,
3475 					 * only do limited xmit when there
3476 					 * is SACK info.
3477 					 *
3478 					 * Note how tcp_cwnd is incremented.
3479 					 * The first dup ACK will increase
3480 					 * it by 1 MSS.  The second dup ACK
3481 					 * will increase it by 2 MSS.  This
3482 					 * means that only 1 new segment will
3483 					 * be sent for each dup ACK.
3484 					 */
3485 					if (tcp->tcp_unsent > 0 &&
3486 					    (!tcp->tcp_snd_sack_ok ||
3487 					    (tcp->tcp_snd_sack_ok &&
3488 					    tcp->tcp_notsack_list != NULL))) {
3489 						tcp->tcp_cwnd += mss <<
3490 						    (tcp->tcp_dupack_cnt - 1);
3491 						flags |= TH_LIMIT_XMIT;
3492 					}
3493 				} else if (dupack_cnt ==
3494 				    tcp_dupack_fast_retransmit) {
3495 
3496 				BUMP_MIB(tcp_mib.tcpOutFastRetrans);
3497 				/*
3498 				 * If we have reduced tcp_ssthresh
3499 				 * because of ECN, do not reduce it again
3500 				 * unless it is already one window of data
3501 				 * away.  After one window of data, tcp_cwr
3502 				 * should then be cleared.  Note that
3503 				 * for non ECN capable connection, tcp_cwr
3504 				 * should always be false.
3505 				 *
3506 				 * Adjust cwnd since the duplicate
3507 				 * ack indicates that a packet was
3508 				 * dropped (due to congestion.)
3509 				 */
3510 				if (!tcp->tcp_cwr) {
3511 					npkt = (MIN(tcp->tcp_cwnd,
3512 					    tcp->tcp_swnd) >> 1) / mss;
3513 					if (npkt < 2)
3514 						npkt = 2;
3515 					tcp->tcp_cwnd_ssthresh = npkt * mss;
3516 					tcp->tcp_cwnd = (npkt +
3517 					    tcp->tcp_dupack_cnt) * mss;
3518 				}
3519 				if (tcp->tcp_ecn_ok) {
3520 					tcp->tcp_cwr = B_TRUE;
3521 					tcp->tcp_cwr_snd_max = tcp->tcp_snxt;
3522 					tcp->tcp_ecn_cwr_sent = B_FALSE;
3523 				}
3524 
3525 				/*
3526 				 * We do Hoe's algorithm.  Refer to her
3527 				 * paper "Improving the Start-up Behavior
3528 				 * of a Congestion Control Scheme for TCP,"
3529 				 * appeared in SIGCOMM'96.
3530 				 *
3531 				 * Save highest seq no we have sent so far.
3532 				 * Be careful about the invisible FIN byte.
3533 				 */
3534 				if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
3535 				    (tcp->tcp_unsent == 0)) {
3536 					tcp->tcp_rexmit_max = tcp->tcp_fss;
3537 				} else {
3538 					tcp->tcp_rexmit_max = tcp->tcp_snxt;
3539 				}
3540 
3541 				/*
3542 				 * Do not allow bursty traffic during.
3543 				 * fast recovery.  Refer to Fall and Floyd's
3544 				 * paper "Simulation-based Comparisons of
3545 				 * Tahoe, Reno and SACK TCP" (in CCR ??)
3546 				 * This is a best current practise.
3547 				 */
3548 				tcp->tcp_snd_burst = TCP_CWND_SS;
3549 
3550 				/*
3551 				 * For SACK:
3552 				 * Calculate tcp_pipe, which is the
3553 				 * estimated number of bytes in
3554 				 * network.
3555 				 *
3556 				 * tcp_fack is the highest sack'ed seq num
3557 				 * TCP has received.
3558 				 *
3559 				 * tcp_pipe is explained in the above quoted
3560 				 * Fall and Floyd's paper.  tcp_fack is
3561 				 * explained in Mathis and Mahdavi's
3562 				 * "Forward Acknowledgment: Refining TCP
3563 				 * Congestion Control" in SIGCOMM '96.
3564 				 */
3565 				if (tcp->tcp_snd_sack_ok) {
3566 					assert(tcp->tcp_sack_info != NULL);
3567 					if (tcp->tcp_notsack_list != NULL) {
3568 						tcp->tcp_pipe = tcp->tcp_snxt -
3569 						    tcp->tcp_fack;
3570 						tcp->tcp_sack_snxt = seg_ack;
3571 						flags |= TH_NEED_SACK_REXMIT;
3572 					} else {
3573 						/*
3574 						 * Always initialize tcp_pipe
3575 						 * even though we don't have
3576 						 * any SACK info.  If later
3577 						 * we get SACK info and
3578 						 * tcp_pipe is not initialized,
3579 						 * funny things will happen.
3580 						 */
3581 						tcp->tcp_pipe =
3582 						    tcp->tcp_cwnd_ssthresh;
3583 					}
3584 				} else {
3585 					flags |= TH_REXMIT_NEEDED;
3586 				} /* tcp_snd_sack_ok */
3587 
3588 				} else {
3589 					/*
3590 					 * Here we perform congestion
3591 					 * avoidance, but NOT slow start.
3592 					 * This is known as the Fast
3593 					 * Recovery Algorithm.
3594 					 */
3595 					if (tcp->tcp_snd_sack_ok &&
3596 					    tcp->tcp_notsack_list != NULL) {
3597 						flags |= TH_NEED_SACK_REXMIT;
3598 						tcp->tcp_pipe -= mss;
3599 						if (tcp->tcp_pipe < 0)
3600 							tcp->tcp_pipe = 0;
3601 					} else {
3602 					/*
3603 					 * We know that one more packet has
3604 					 * left the pipe thus we can update
3605 					 * cwnd.
3606 					 */
3607 					cwnd = tcp->tcp_cwnd + mss;
3608 					if (cwnd > tcp->tcp_cwnd_max)
3609 						cwnd = tcp->tcp_cwnd_max;
3610 					tcp->tcp_cwnd = cwnd;
3611 					flags |= TH_XMIT_NEEDED;
3612 					}
3613 				}
3614 			}
3615 		} else if (tcp->tcp_zero_win_probe) {
3616 			/*
3617 			 * If the window has opened, need to arrange
3618 			 * to send additional data.
3619 			 */
3620 			if (new_swnd != 0) {
3621 				/* tcp_suna != tcp_snxt */
3622 				/* Packet contains a window update */
3623 				BUMP_MIB(tcp_mib.tcpInWinUpdate);
3624 				tcp->tcp_zero_win_probe = 0;
3625 				tcp->tcp_timer_backoff = 0;
3626 				tcp->tcp_ms_we_have_waited = 0;
3627 
3628 				/*
3629 				 * Transmit starting with tcp_suna since
3630 				 * the one byte probe is not ack'ed.
3631 				 * If TCP has sent more than one identical
3632 				 * probe, tcp_rexmit will be set.  That means
3633 				 * tcp_ss_rexmit() will send out the one
3634 				 * byte along with new data.  Otherwise,
3635 				 * fake the retransmission.
3636 				 */
3637 				flags |= TH_XMIT_NEEDED;
3638 				if (!tcp->tcp_rexmit) {
3639 					tcp->tcp_rexmit = B_TRUE;
3640 					tcp->tcp_dupack_cnt = 0;
3641 					tcp->tcp_rexmit_nxt = tcp->tcp_suna;
3642 					tcp->tcp_rexmit_max = tcp->tcp_suna + 1;
3643 				}
3644 			}
3645 		}
3646 		goto swnd_update;
3647 	}
3648 
3649 	/*
3650 	 * Check for "acceptability" of ACK value per RFC 793, pages 72 - 73.
3651 	 * If the ACK value acks something that we have not yet sent, it might
3652 	 * be an old duplicate segment.  Send an ACK to re-synchronize the
3653 	 * other side.
3654 	 * Note: reset in response to unacceptable ACK in SYN_RECEIVE
3655 	 * state is handled above, so we can always just drop the segment and
3656 	 * send an ACK here.
3657 	 *
3658 	 * Should we send ACKs in response to ACK only segments?
3659 	 */
3660 	if (SEQ_GT(seg_ack, tcp->tcp_snxt)) {
3661 		BUMP_MIB(tcp_mib.tcpInAckUnsent);
3662 		/* drop the received segment */
3663 		freemsg(mp);
3664 
3665 		/* Send back an ACK. */
3666 		mp = tcp_ack_mp(tcp);
3667 
3668 		if (mp == NULL) {
3669 			return;
3670 		}
3671 		BUMP_MIB(tcp_mib.tcpOutAck);
3672 		(void) ipv4_tcp_output(sock_id, mp);
3673 		freeb(mp);
3674 		return;
3675 	}
3676 
3677 	/*
3678 	 * TCP gets a new ACK, update the notsack'ed list to delete those
3679 	 * blocks that are covered by this ACK.
3680 	 */
3681 	if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
3682 		tcp_notsack_remove(&(tcp->tcp_notsack_list), seg_ack,
3683 		    &(tcp->tcp_num_notsack_blk), &(tcp->tcp_cnt_notsack_list));
3684 	}
3685 
3686 	/*
3687 	 * If we got an ACK after fast retransmit, check to see
3688 	 * if it is a partial ACK.  If it is not and the congestion
3689 	 * window was inflated to account for the other side's
3690 	 * cached packets, retract it.  If it is, do Hoe's algorithm.
3691 	 */
3692 	if (tcp->tcp_dupack_cnt >= tcp_dupack_fast_retransmit) {
3693 		assert(tcp->tcp_rexmit == B_FALSE);
3694 		if (SEQ_GEQ(seg_ack, tcp->tcp_rexmit_max)) {
3695 			tcp->tcp_dupack_cnt = 0;
3696 			/*
3697 			 * Restore the orig tcp_cwnd_ssthresh after
3698 			 * fast retransmit phase.
3699 			 */
3700 			if (tcp->tcp_cwnd > tcp->tcp_cwnd_ssthresh) {
3701 				tcp->tcp_cwnd = tcp->tcp_cwnd_ssthresh;
3702 			}
3703 			tcp->tcp_rexmit_max = seg_ack;
3704 			tcp->tcp_cwnd_cnt = 0;
3705 			tcp->tcp_snd_burst = TCP_CWND_NORMAL;
3706 
3707 			/*
3708 			 * Remove all notsack info to avoid confusion with
3709 			 * the next fast retrasnmit/recovery phase.
3710 			 */
3711 			if (tcp->tcp_snd_sack_ok &&
3712 			    tcp->tcp_notsack_list != NULL) {
3713 				TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list);
3714 			}
3715 		} else {
3716 			if (tcp->tcp_snd_sack_ok &&
3717 			    tcp->tcp_notsack_list != NULL) {
3718 				flags |= TH_NEED_SACK_REXMIT;
3719 				tcp->tcp_pipe -= mss;
3720 				if (tcp->tcp_pipe < 0)
3721 					tcp->tcp_pipe = 0;
3722 			} else {
3723 				/*
3724 				 * Hoe's algorithm:
3725 				 *
3726 				 * Retransmit the unack'ed segment and
3727 				 * restart fast recovery.  Note that we
3728 				 * need to scale back tcp_cwnd to the
3729 				 * original value when we started fast
3730 				 * recovery.  This is to prevent overly
3731 				 * aggressive behaviour in sending new
3732 				 * segments.
3733 				 */
3734 				tcp->tcp_cwnd = tcp->tcp_cwnd_ssthresh +
3735 					tcp_dupack_fast_retransmit * mss;
3736 				tcp->tcp_cwnd_cnt = tcp->tcp_cwnd;
3737 				BUMP_MIB(tcp_mib.tcpOutFastRetrans);
3738 				flags |= TH_REXMIT_NEEDED;
3739 			}
3740 		}
3741 	} else {
3742 		tcp->tcp_dupack_cnt = 0;
3743 		if (tcp->tcp_rexmit) {
3744 			/*
3745 			 * TCP is retranmitting.  If the ACK ack's all
3746 			 * outstanding data, update tcp_rexmit_max and
3747 			 * tcp_rexmit_nxt.  Otherwise, update tcp_rexmit_nxt
3748 			 * to the correct value.
3749 			 *
3750 			 * Note that SEQ_LEQ() is used.  This is to avoid
3751 			 * unnecessary fast retransmit caused by dup ACKs
3752 			 * received when TCP does slow start retransmission
3753 			 * after a time out.  During this phase, TCP may
3754 			 * send out segments which are already received.
3755 			 * This causes dup ACKs to be sent back.
3756 			 */
3757 			if (SEQ_LEQ(seg_ack, tcp->tcp_rexmit_max)) {
3758 				if (SEQ_GT(seg_ack, tcp->tcp_rexmit_nxt)) {
3759 					tcp->tcp_rexmit_nxt = seg_ack;
3760 				}
3761 				if (seg_ack != tcp->tcp_rexmit_max) {
3762 					flags |= TH_XMIT_NEEDED;
3763 				}
3764 			} else {
3765 				tcp->tcp_rexmit = B_FALSE;
3766 				tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3767 				tcp->tcp_snd_burst = TCP_CWND_NORMAL;
3768 			}
3769 			tcp->tcp_ms_we_have_waited = 0;
3770 		}
3771 	}
3772 
3773 	BUMP_MIB(tcp_mib.tcpInAckSegs);
3774 	UPDATE_MIB(tcp_mib.tcpInAckBytes, bytes_acked);
3775 	tcp->tcp_suna = seg_ack;
3776 	if (tcp->tcp_zero_win_probe != 0) {
3777 		tcp->tcp_zero_win_probe = 0;
3778 		tcp->tcp_timer_backoff = 0;
3779 	}
3780 
3781 	/*
3782 	 * If tcp_xmit_head is NULL, then it must be the FIN being ack'ed.
3783 	 * Note that it cannot be the SYN being ack'ed.  The code flow
3784 	 * will not reach here.
3785 	 */
3786 	if (mp1 == NULL) {
3787 		goto fin_acked;
3788 	}
3789 
3790 	/*
3791 	 * Update the congestion window.
3792 	 *
3793 	 * If TCP is not ECN capable or TCP is ECN capable but the
3794 	 * congestion experience bit is not set, increase the tcp_cwnd as
3795 	 * usual.
3796 	 */
3797 	if (!tcp->tcp_ecn_ok || !(flags & TH_ECE)) {
3798 		cwnd = tcp->tcp_cwnd;
3799 		add = mss;
3800 
3801 		if (cwnd >= tcp->tcp_cwnd_ssthresh) {
3802 			/*
3803 			 * This is to prevent an increase of less than 1 MSS of
3804 			 * tcp_cwnd.  With partial increase, tcp_wput_data()
3805 			 * may send out tinygrams in order to preserve mblk
3806 			 * boundaries.
3807 			 *
3808 			 * By initializing tcp_cwnd_cnt to new tcp_cwnd and
3809 			 * decrementing it by 1 MSS for every ACKs, tcp_cwnd is
3810 			 * increased by 1 MSS for every RTTs.
3811 			 */
3812 			if (tcp->tcp_cwnd_cnt <= 0) {
3813 				tcp->tcp_cwnd_cnt = cwnd + add;
3814 			} else {
3815 				tcp->tcp_cwnd_cnt -= add;
3816 				add = 0;
3817 			}
3818 		}
3819 		tcp->tcp_cwnd = MIN(cwnd + add, tcp->tcp_cwnd_max);
3820 	}
3821 
3822 	/* Can we update the RTT estimates? */
3823 	if (tcp->tcp_snd_ts_ok) {
3824 		/* Ignore zero timestamp echo-reply. */
3825 		if (tcpopt.tcp_opt_ts_ecr != 0) {
3826 			tcp_set_rto(tcp, (int32_t)(prom_gettime() -
3827 			    tcpopt.tcp_opt_ts_ecr));
3828 		}
3829 
3830 		/* If needed, restart the timer. */
3831 		if (tcp->tcp_set_timer == 1) {
3832 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3833 			tcp->tcp_set_timer = 0;
3834 		}
3835 		/*
3836 		 * Update tcp_csuna in case the other side stops sending
3837 		 * us timestamps.
3838 		 */
3839 		tcp->tcp_csuna = tcp->tcp_snxt;
3840 	} else if (SEQ_GT(seg_ack, tcp->tcp_csuna)) {
3841 		/*
3842 		 * An ACK sequence we haven't seen before, so get the RTT
3843 		 * and update the RTO.
3844 		 */
3845 		tcp_set_rto(tcp, (int32_t)(prom_gettime() -
3846 		    (uint32_t)mp1->b_prev));
3847 
3848 		/* Remeber the last sequence to be ACKed */
3849 		tcp->tcp_csuna = seg_ack;
3850 		if (tcp->tcp_set_timer == 1) {
3851 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3852 			tcp->tcp_set_timer = 0;
3853 		}
3854 	} else {
3855 		BUMP_MIB(tcp_mib.tcpRttNoUpdate);
3856 	}
3857 
3858 	/* Eat acknowledged bytes off the xmit queue. */
3859 	for (;;) {
3860 		mblk_t	*mp2;
3861 		uchar_t	*wptr;
3862 
3863 		wptr = mp1->b_wptr;
3864 		assert((uintptr_t)(wptr - mp1->b_rptr) <= (uintptr_t)INT_MAX);
3865 		bytes_acked -= (int)(wptr - mp1->b_rptr);
3866 		if (bytes_acked < 0) {
3867 			mp1->b_rptr = wptr + bytes_acked;
3868 			break;
3869 		}
3870 		mp1->b_prev = NULL;
3871 		mp2 = mp1;
3872 		mp1 = mp1->b_cont;
3873 		freeb(mp2);
3874 		if (bytes_acked == 0) {
3875 			if (mp1 == NULL) {
3876 				/* Everything is ack'ed, clear the tail. */
3877 				tcp->tcp_xmit_tail = NULL;
3878 				goto pre_swnd_update;
3879 			}
3880 			if (mp2 != tcp->tcp_xmit_tail)
3881 				break;
3882 			tcp->tcp_xmit_tail = mp1;
3883 			assert((uintptr_t)(mp1->b_wptr -
3884 			    mp1->b_rptr) <= (uintptr_t)INT_MAX);
3885 			tcp->tcp_xmit_tail_unsent = (int)(mp1->b_wptr -
3886 			    mp1->b_rptr);
3887 			break;
3888 		}
3889 		if (mp1 == NULL) {
3890 			/*
3891 			 * More was acked but there is nothing more
3892 			 * outstanding.  This means that the FIN was
3893 			 * just acked or that we're talking to a clown.
3894 			 */
3895 fin_acked:
3896 			assert(tcp->tcp_fin_sent);
3897 			tcp->tcp_xmit_tail = NULL;
3898 			if (tcp->tcp_fin_sent) {
3899 				tcp->tcp_fin_acked = B_TRUE;
3900 			} else {
3901 				/*
3902 				 * We should never got here because
3903 				 * we have already checked that the
3904 				 * number of bytes ack'ed should be
3905 				 * smaller than or equal to what we
3906 				 * have sent so far (it is the
3907 				 * acceptability check of the ACK).
3908 				 * We can only get here if the send
3909 				 * queue is corrupted.
3910 				 *
3911 				 * Terminate the connection and
3912 				 * panic the system.  It is better
3913 				 * for us to panic instead of
3914 				 * continuing to avoid other disaster.
3915 				 */
3916 				tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
3917 				    tcp->tcp_rnxt, TH_RST|TH_ACK, 0, sock_id);
3918 				printf("Memory corruption "
3919 				    "detected for connection %s.\n",
3920 				    tcp_display(tcp, NULL,
3921 					DISP_ADDR_AND_PORT));
3922 				/* We should never get here... */
3923 				prom_panic("tcp_rput_data");
3924 				return;
3925 			}
3926 			goto pre_swnd_update;
3927 		}
3928 		assert(mp2 != tcp->tcp_xmit_tail);
3929 	}
3930 	if (tcp->tcp_unsent) {
3931 		flags |= TH_XMIT_NEEDED;
3932 	}
3933 pre_swnd_update:
3934 	tcp->tcp_xmit_head = mp1;
3935 swnd_update:
3936 	/*
3937 	 * The following check is different from most other implementations.
3938 	 * For bi-directional transfer, when segments are dropped, the
3939 	 * "normal" check will not accept a window update in those
3940 	 * retransmitted segemnts.  Failing to do that, TCP may send out
3941 	 * segments which are outside receiver's window.  As TCP accepts
3942 	 * the ack in those retransmitted segments, if the window update in
3943 	 * the same segment is not accepted, TCP will incorrectly calculates
3944 	 * that it can send more segments.  This can create a deadlock
3945 	 * with the receiver if its window becomes zero.
3946 	 */
3947 	if (SEQ_LT(tcp->tcp_swl2, seg_ack) ||
3948 	    SEQ_LT(tcp->tcp_swl1, seg_seq) ||
3949 	    (tcp->tcp_swl1 == seg_seq && new_swnd > tcp->tcp_swnd)) {
3950 		/*
3951 		 * The criteria for update is:
3952 		 *
3953 		 * 1. the segment acknowledges some data.  Or
3954 		 * 2. the segment is new, i.e. it has a higher seq num. Or
3955 		 * 3. the segment is not old and the advertised window is
3956 		 * larger than the previous advertised window.
3957 		 */
3958 		if (tcp->tcp_unsent && new_swnd > tcp->tcp_swnd)
3959 			flags |= TH_XMIT_NEEDED;
3960 		tcp->tcp_swnd = new_swnd;
3961 		if (new_swnd > tcp->tcp_max_swnd)
3962 			tcp->tcp_max_swnd = new_swnd;
3963 		tcp->tcp_swl1 = seg_seq;
3964 		tcp->tcp_swl2 = seg_ack;
3965 	}
3966 est:
3967 	if (tcp->tcp_state > TCPS_ESTABLISHED) {
3968 		switch (tcp->tcp_state) {
3969 		case TCPS_FIN_WAIT_1:
3970 			if (tcp->tcp_fin_acked) {
3971 				tcp->tcp_state = TCPS_FIN_WAIT_2;
3972 				/*
3973 				 * We implement the non-standard BSD/SunOS
3974 				 * FIN_WAIT_2 flushing algorithm.
3975 				 * If there is no user attached to this
3976 				 * TCP endpoint, then this TCP struct
3977 				 * could hang around forever in FIN_WAIT_2
3978 				 * state if the peer forgets to send us
3979 				 * a FIN.  To prevent this, we wait only
3980 				 * 2*MSL (a convenient time value) for
3981 				 * the FIN to arrive.  If it doesn't show up,
3982 				 * we flush the TCP endpoint.  This algorithm,
3983 				 * though a violation of RFC-793, has worked
3984 				 * for over 10 years in BSD systems.
3985 				 * Note: SunOS 4.x waits 675 seconds before
3986 				 * flushing the FIN_WAIT_2 connection.
3987 				 */
3988 				TCP_TIMER_RESTART(tcp,
3989 				    tcp_fin_wait_2_flush_interval);
3990 			}
3991 			break;
3992 		case TCPS_FIN_WAIT_2:
3993 			break;	/* Shutdown hook? */
3994 		case TCPS_LAST_ACK:
3995 			freemsg(mp);
3996 			if (tcp->tcp_fin_acked) {
3997 				(void) tcp_clean_death(sock_id, tcp, 0);
3998 				return;
3999 			}
4000 			goto xmit_check;
4001 		case TCPS_CLOSING:
4002 			if (tcp->tcp_fin_acked) {
4003 				tcp->tcp_state = TCPS_TIME_WAIT;
4004 				tcp_time_wait_append(tcp);
4005 				TCP_TIMER_RESTART(tcp, tcp_time_wait_interval);
4006 			}
4007 			/*FALLTHRU*/
4008 		case TCPS_CLOSE_WAIT:
4009 			freemsg(mp);
4010 			goto xmit_check;
4011 		default:
4012 			assert(tcp->tcp_state != TCPS_TIME_WAIT);
4013 			break;
4014 		}
4015 	}
4016 	if (flags & TH_FIN) {
4017 		/* Make sure we ack the fin */
4018 		flags |= TH_ACK_NEEDED;
4019 		if (!tcp->tcp_fin_rcvd) {
4020 			tcp->tcp_fin_rcvd = B_TRUE;
4021 			tcp->tcp_rnxt++;
4022 			U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
4023 
4024 			switch (tcp->tcp_state) {
4025 			case TCPS_SYN_RCVD:
4026 			case TCPS_ESTABLISHED:
4027 				tcp->tcp_state = TCPS_CLOSE_WAIT;
4028 				/* Keepalive? */
4029 				break;
4030 			case TCPS_FIN_WAIT_1:
4031 				if (!tcp->tcp_fin_acked) {
4032 					tcp->tcp_state = TCPS_CLOSING;
4033 					break;
4034 				}
4035 				/* FALLTHRU */
4036 			case TCPS_FIN_WAIT_2:
4037 				tcp->tcp_state = TCPS_TIME_WAIT;
4038 				tcp_time_wait_append(tcp);
4039 				TCP_TIMER_RESTART(tcp, tcp_time_wait_interval);
4040 				if (seg_len) {
4041 					/*
4042 					 * implies data piggybacked on FIN.
4043 					 * break to handle data.
4044 					 */
4045 					break;
4046 				}
4047 				freemsg(mp);
4048 				goto ack_check;
4049 			}
4050 		}
4051 	}
4052 	if (mp == NULL)
4053 		goto xmit_check;
4054 	if (seg_len == 0) {
4055 		freemsg(mp);
4056 		goto xmit_check;
4057 	}
4058 	if (mp->b_rptr == mp->b_wptr) {
4059 		/*
4060 		 * The header has been consumed, so we remove the
4061 		 * zero-length mblk here.
4062 		 */
4063 		mp1 = mp;
4064 		mp = mp->b_cont;
4065 		freeb(mp1);
4066 	}
4067 	/*
4068 	 * ACK every other segments, unless the input queue is empty
4069 	 * as we don't have a timer available.
4070 	 */
4071 	if (++tcp->tcp_rack_cnt == 2 || sockets[sock_id].inq == NULL) {
4072 		flags |= TH_ACK_NEEDED;
4073 		tcp->tcp_rack_cnt = 0;
4074 	}
4075 	tcp->tcp_rnxt += seg_len;
4076 	U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
4077 
4078 	/* Update SACK list */
4079 	if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
4080 		tcp_sack_remove(tcp->tcp_sack_list, tcp->tcp_rnxt,
4081 		    &(tcp->tcp_num_sack_blk));
4082 	}
4083 
4084 	if (tcp->tcp_listener) {
4085 		/*
4086 		 * Side queue inbound data until the accept happens.
4087 		 * tcp_accept/tcp_rput drains this when the accept happens.
4088 		 */
4089 		tcp_rcv_enqueue(tcp, mp, seg_len);
4090 	} else {
4091 		/* Just queue the data until the app calls read. */
4092 		tcp_rcv_enqueue(tcp, mp, seg_len);
4093 		/*
4094 		 * Make sure the timer is running if we have data waiting
4095 		 * for a push bit. This provides resiliency against
4096 		 * implementations that do not correctly generate push bits.
4097 		 */
4098 		if (tcp->tcp_rcv_list != NULL)
4099 			flags |= TH_TIMER_NEEDED;
4100 	}
4101 
4102 xmit_check:
4103 	/* Is there anything left to do? */
4104 	if ((flags & (TH_REXMIT_NEEDED|TH_XMIT_NEEDED|TH_ACK_NEEDED|
4105 	    TH_NEED_SACK_REXMIT|TH_LIMIT_XMIT|TH_TIMER_NEEDED)) == 0)
4106 		return;
4107 
4108 	/* Any transmit work to do and a non-zero window? */
4109 	if ((flags & (TH_REXMIT_NEEDED|TH_XMIT_NEEDED|TH_NEED_SACK_REXMIT|
4110 	    TH_LIMIT_XMIT)) && tcp->tcp_swnd != 0) {
4111 		if (flags & TH_REXMIT_NEEDED) {
4112 			uint32_t snd_size = tcp->tcp_snxt - tcp->tcp_suna;
4113 
4114 			if (snd_size > mss)
4115 				snd_size = mss;
4116 			if (snd_size > tcp->tcp_swnd)
4117 				snd_size = tcp->tcp_swnd;
4118 			mp1 = tcp_xmit_mp(tcp, tcp->tcp_xmit_head, snd_size,
4119 			    NULL, NULL, tcp->tcp_suna, B_TRUE, &snd_size,
4120 			    B_TRUE);
4121 
4122 			if (mp1 != NULL) {
4123 				tcp->tcp_xmit_head->b_prev =
4124 				    (mblk_t *)prom_gettime();
4125 				tcp->tcp_csuna = tcp->tcp_snxt;
4126 				BUMP_MIB(tcp_mib.tcpRetransSegs);
4127 				UPDATE_MIB(tcp_mib.tcpRetransBytes, snd_size);
4128 				(void) ipv4_tcp_output(sock_id, mp1);
4129 				freeb(mp1);
4130 			}
4131 		}
4132 		if (flags & TH_NEED_SACK_REXMIT) {
4133 			if (tcp_sack_rxmit(tcp, sock_id) != 0) {
4134 				flags |= TH_XMIT_NEEDED;
4135 			}
4136 		}
4137 		/*
4138 		 * For TH_LIMIT_XMIT, tcp_wput_data() is called to send
4139 		 * out new segment.  Note that tcp_rexmit should not be
4140 		 * set, otherwise TH_LIMIT_XMIT should not be set.
4141 		 */
4142 		if (flags & (TH_XMIT_NEEDED|TH_LIMIT_XMIT)) {
4143 			if (!tcp->tcp_rexmit) {
4144 				tcp_wput_data(tcp, NULL, sock_id);
4145 			} else {
4146 				tcp_ss_rexmit(tcp, sock_id);
4147 			}
4148 			/*
4149 			 * The TCP could be closed in tcp_state_wait via
4150 			 * tcp_wput_data (tcp_ss_rexmit could call
4151 			 * tcp_wput_data as well).
4152 			 */
4153 			if (sockets[sock_id].pcb == NULL)
4154 				return;
4155 		}
4156 		/*
4157 		 * Adjust tcp_cwnd back to normal value after sending
4158 		 * new data segments.
4159 		 */
4160 		if (flags & TH_LIMIT_XMIT) {
4161 			tcp->tcp_cwnd -= mss << (tcp->tcp_dupack_cnt - 1);
4162 		}
4163 
4164 		/* Anything more to do? */
4165 		if ((flags & (TH_ACK_NEEDED|TH_TIMER_NEEDED)) == 0)
4166 			return;
4167 	}
4168 ack_check:
4169 	if (flags & TH_ACK_NEEDED) {
4170 		/*
4171 		 * Time to send an ack for some reason.
4172 		 */
4173 		if ((mp1 = tcp_ack_mp(tcp)) != NULL) {
4174 			TCP_DUMP_PACKET("tcp_rput_data: ack mp", mp1);
4175 			(void) ipv4_tcp_output(sock_id, mp1);
4176 			BUMP_MIB(tcp_mib.tcpOutAck);
4177 			freeb(mp1);
4178 		}
4179 	}
4180 }
4181 
4182 /*
4183  * tcp_ss_rexmit() is called in tcp_rput_data() to do slow start
4184  * retransmission after a timeout.
4185  *
4186  * To limit the number of duplicate segments, we limit the number of segment
4187  * to be sent in one time to tcp_snd_burst, the burst variable.
4188  */
4189 static void
4190 tcp_ss_rexmit(tcp_t *tcp, int sock_id)
4191 {
4192 	uint32_t	snxt;
4193 	uint32_t	smax;
4194 	int32_t		win;
4195 	int32_t		mss;
4196 	int32_t		off;
4197 	int32_t		burst = tcp->tcp_snd_burst;
4198 	mblk_t		*snxt_mp;
4199 
4200 	/*
4201 	 * Note that tcp_rexmit can be set even though TCP has retransmitted
4202 	 * all unack'ed segments.
4203 	 */
4204 	if (SEQ_LT(tcp->tcp_rexmit_nxt, tcp->tcp_rexmit_max)) {
4205 		smax = tcp->tcp_rexmit_max;
4206 		snxt = tcp->tcp_rexmit_nxt;
4207 		if (SEQ_LT(snxt, tcp->tcp_suna)) {
4208 			snxt = tcp->tcp_suna;
4209 		}
4210 		win = MIN(tcp->tcp_cwnd, tcp->tcp_swnd);
4211 		win -= snxt - tcp->tcp_suna;
4212 		mss = tcp->tcp_mss;
4213 		snxt_mp = tcp_get_seg_mp(tcp, snxt, &off);
4214 
4215 		while (SEQ_LT(snxt, smax) && (win > 0) &&
4216 		    (burst > 0) && (snxt_mp != NULL)) {
4217 			mblk_t	*xmit_mp;
4218 			mblk_t	*old_snxt_mp = snxt_mp;
4219 			uint32_t cnt = mss;
4220 
4221 			if (win < cnt) {
4222 				cnt = win;
4223 			}
4224 			if (SEQ_GT(snxt + cnt, smax)) {
4225 				cnt = smax - snxt;
4226 			}
4227 			xmit_mp = tcp_xmit_mp(tcp, snxt_mp, cnt, &off,
4228 			    &snxt_mp, snxt, B_TRUE, &cnt, B_TRUE);
4229 
4230 			if (xmit_mp == NULL)
4231 				return;
4232 
4233 			(void) ipv4_tcp_output(sock_id, xmit_mp);
4234 			freeb(xmit_mp);
4235 
4236 			snxt += cnt;
4237 			win -= cnt;
4238 			/*
4239 			 * Update the send timestamp to avoid false
4240 			 * retransmission.
4241 			 */
4242 			old_snxt_mp->b_prev = (mblk_t *)prom_gettime();
4243 			BUMP_MIB(tcp_mib.tcpRetransSegs);
4244 			UPDATE_MIB(tcp_mib.tcpRetransBytes, cnt);
4245 
4246 			tcp->tcp_rexmit_nxt = snxt;
4247 			burst--;
4248 		}
4249 		/*
4250 		 * If we have transmitted all we have at the time
4251 		 * we started the retranmission, we can leave
4252 		 * the rest of the job to tcp_wput_data().  But we
4253 		 * need to check the send window first.  If the
4254 		 * win is not 0, go on with tcp_wput_data().
4255 		 */
4256 		if (SEQ_LT(snxt, smax) || win == 0) {
4257 			return;
4258 		}
4259 	}
4260 	/* Only call tcp_wput_data() if there is data to be sent. */
4261 	if (tcp->tcp_unsent) {
4262 		tcp_wput_data(tcp, NULL, sock_id);
4263 	}
4264 }
4265 
4266 /*
4267  * tcp_timer is the timer service routine.  It handles all timer events for
4268  * a tcp instance except keepalives.  It figures out from the state of the
4269  * tcp instance what kind of action needs to be done at the time it is called.
4270  */
4271 static void
4272 tcp_timer(tcp_t	*tcp, int sock_id)
4273 {
4274 	mblk_t		*mp;
4275 	uint32_t	first_threshold;
4276 	uint32_t	second_threshold;
4277 	uint32_t	ms;
4278 	uint32_t	mss;
4279 
4280 	first_threshold =  tcp->tcp_first_timer_threshold;
4281 	second_threshold = tcp->tcp_second_timer_threshold;
4282 	switch (tcp->tcp_state) {
4283 	case TCPS_IDLE:
4284 	case TCPS_BOUND:
4285 	case TCPS_LISTEN:
4286 		return;
4287 	case TCPS_SYN_RCVD:
4288 	case TCPS_SYN_SENT:
4289 		first_threshold =  tcp->tcp_first_ctimer_threshold;
4290 		second_threshold = tcp->tcp_second_ctimer_threshold;
4291 		break;
4292 	case TCPS_ESTABLISHED:
4293 	case TCPS_FIN_WAIT_1:
4294 	case TCPS_CLOSING:
4295 	case TCPS_CLOSE_WAIT:
4296 	case TCPS_LAST_ACK:
4297 		/* If we have data to rexmit */
4298 		if (tcp->tcp_suna != tcp->tcp_snxt) {
4299 			int32_t time_to_wait;
4300 
4301 			BUMP_MIB(tcp_mib.tcpTimRetrans);
4302 			if (tcp->tcp_xmit_head == NULL)
4303 				break;
4304 			time_to_wait = (int32_t)(prom_gettime() -
4305 			    (uint32_t)tcp->tcp_xmit_head->b_prev);
4306 			time_to_wait = tcp->tcp_rto - time_to_wait;
4307 			if (time_to_wait > 0) {
4308 				/*
4309 				 * Timer fired too early, so restart it.
4310 				 */
4311 				TCP_TIMER_RESTART(tcp, time_to_wait);
4312 				return;
4313 			}
4314 			/*
4315 			 * When we probe zero windows, we force the swnd open.
4316 			 * If our peer acks with a closed window swnd will be
4317 			 * set to zero by tcp_rput(). As long as we are
4318 			 * receiving acks tcp_rput will
4319 			 * reset 'tcp_ms_we_have_waited' so as not to trip the
4320 			 * first and second interval actions.  NOTE: the timer
4321 			 * interval is allowed to continue its exponential
4322 			 * backoff.
4323 			 */
4324 			if (tcp->tcp_swnd == 0 || tcp->tcp_zero_win_probe) {
4325 				DEBUG_1("tcp_timer (%d): zero win", sock_id);
4326 				break;
4327 			} else {
4328 				/*
4329 				 * After retransmission, we need to do
4330 				 * slow start.  Set the ssthresh to one
4331 				 * half of current effective window and
4332 				 * cwnd to one MSS.  Also reset
4333 				 * tcp_cwnd_cnt.
4334 				 *
4335 				 * Note that if tcp_ssthresh is reduced because
4336 				 * of ECN, do not reduce it again unless it is
4337 				 * already one window of data away (tcp_cwr
4338 				 * should then be cleared) or this is a
4339 				 * timeout for a retransmitted segment.
4340 				 */
4341 				uint32_t npkt;
4342 
4343 				if (!tcp->tcp_cwr || tcp->tcp_rexmit) {
4344 					npkt = (MIN((tcp->tcp_timer_backoff ?
4345 					    tcp->tcp_cwnd_ssthresh :
4346 					    tcp->tcp_cwnd),
4347 					    tcp->tcp_swnd) >> 1) /
4348 					    tcp->tcp_mss;
4349 					if (npkt < 2)
4350 						npkt = 2;
4351 					tcp->tcp_cwnd_ssthresh = npkt *
4352 					    tcp->tcp_mss;
4353 				}
4354 				tcp->tcp_cwnd = tcp->tcp_mss;
4355 				tcp->tcp_cwnd_cnt = 0;
4356 				if (tcp->tcp_ecn_ok) {
4357 					tcp->tcp_cwr = B_TRUE;
4358 					tcp->tcp_cwr_snd_max = tcp->tcp_snxt;
4359 					tcp->tcp_ecn_cwr_sent = B_FALSE;
4360 				}
4361 			}
4362 			break;
4363 		}
4364 		/*
4365 		 * We have something to send yet we cannot send.  The
4366 		 * reason can be:
4367 		 *
4368 		 * 1. Zero send window: we need to do zero window probe.
4369 		 * 2. Zero cwnd: because of ECN, we need to "clock out
4370 		 * segments.
4371 		 * 3. SWS avoidance: receiver may have shrunk window,
4372 		 * reset our knowledge.
4373 		 *
4374 		 * Note that condition 2 can happen with either 1 or
4375 		 * 3.  But 1 and 3 are exclusive.
4376 		 */
4377 		if (tcp->tcp_unsent != 0) {
4378 			if (tcp->tcp_cwnd == 0) {
4379 				/*
4380 				 * Set tcp_cwnd to 1 MSS so that a
4381 				 * new segment can be sent out.  We
4382 				 * are "clocking out" new data when
4383 				 * the network is really congested.
4384 				 */
4385 				assert(tcp->tcp_ecn_ok);
4386 				tcp->tcp_cwnd = tcp->tcp_mss;
4387 			}
4388 			if (tcp->tcp_swnd == 0) {
4389 				/* Extend window for zero window probe */
4390 				tcp->tcp_swnd++;
4391 				tcp->tcp_zero_win_probe = B_TRUE;
4392 				BUMP_MIB(tcp_mib.tcpOutWinProbe);
4393 			} else {
4394 				/*
4395 				 * Handle timeout from sender SWS avoidance.
4396 				 * Reset our knowledge of the max send window
4397 				 * since the receiver might have reduced its
4398 				 * receive buffer.  Avoid setting tcp_max_swnd
4399 				 * to one since that will essentially disable
4400 				 * the SWS checks.
4401 				 *
4402 				 * Note that since we don't have a SWS
4403 				 * state variable, if the timeout is set
4404 				 * for ECN but not for SWS, this
4405 				 * code will also be executed.  This is
4406 				 * fine as tcp_max_swnd is updated
4407 				 * constantly and it will not affect
4408 				 * anything.
4409 				 */
4410 				tcp->tcp_max_swnd = MAX(tcp->tcp_swnd, 2);
4411 			}
4412 			tcp_wput_data(tcp, NULL, sock_id);
4413 			return;
4414 		}
4415 		/* Is there a FIN that needs to be to re retransmitted? */
4416 		if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
4417 		    !tcp->tcp_fin_acked)
4418 			break;
4419 		/* Nothing to do, return without restarting timer. */
4420 		return;
4421 	case TCPS_FIN_WAIT_2:
4422 		/*
4423 		 * User closed the TCP endpoint and peer ACK'ed our FIN.
4424 		 * We waited some time for for peer's FIN, but it hasn't
4425 		 * arrived.  We flush the connection now to avoid
4426 		 * case where the peer has rebooted.
4427 		 */
4428 		/* FALLTHRU */
4429 	case TCPS_TIME_WAIT:
4430 		(void) tcp_clean_death(sock_id, tcp, 0);
4431 		return;
4432 	default:
4433 		DEBUG_3("tcp_timer (%d): strange state (%d) %s", sock_id,
4434 		    tcp->tcp_state, tcp_display(tcp, NULL,
4435 		    DISP_PORT_ONLY));
4436 		return;
4437 	}
4438 	if ((ms = tcp->tcp_ms_we_have_waited) > second_threshold) {
4439 		/*
4440 		 * For zero window probe, we need to send indefinitely,
4441 		 * unless we have not heard from the other side for some
4442 		 * time...
4443 		 */
4444 		if ((tcp->tcp_zero_win_probe == 0) ||
4445 		    ((prom_gettime() - tcp->tcp_last_recv_time) >
4446 		    second_threshold)) {
4447 			BUMP_MIB(tcp_mib.tcpTimRetransDrop);
4448 			/*
4449 			 * If TCP is in SYN_RCVD state, send back a
4450 			 * RST|ACK as BSD does.  Note that tcp_zero_win_probe
4451 			 * should be zero in TCPS_SYN_RCVD state.
4452 			 */
4453 			if (tcp->tcp_state == TCPS_SYN_RCVD) {
4454 				tcp_xmit_ctl("tcp_timer: RST sent on timeout "
4455 				    "in SYN_RCVD",
4456 				    tcp, NULL, tcp->tcp_snxt,
4457 				    tcp->tcp_rnxt, TH_RST | TH_ACK, 0, sock_id);
4458 			}
4459 			(void) tcp_clean_death(sock_id, tcp,
4460 			    tcp->tcp_client_errno ?
4461 			    tcp->tcp_client_errno : ETIMEDOUT);
4462 			return;
4463 		} else {
4464 			/*
4465 			 * Set tcp_ms_we_have_waited to second_threshold
4466 			 * so that in next timeout, we will do the above
4467 			 * check (lbolt - tcp_last_recv_time).  This is
4468 			 * also to avoid overflow.
4469 			 *
4470 			 * We don't need to decrement tcp_timer_backoff
4471 			 * to avoid overflow because it will be decremented
4472 			 * later if new timeout value is greater than
4473 			 * tcp_rexmit_interval_max.  In the case when
4474 			 * tcp_rexmit_interval_max is greater than
4475 			 * second_threshold, it means that we will wait
4476 			 * longer than second_threshold to send the next
4477 			 * window probe.
4478 			 */
4479 			tcp->tcp_ms_we_have_waited = second_threshold;
4480 		}
4481 	} else if (ms > first_threshold && tcp->tcp_rtt_sa != 0) {
4482 		/*
4483 		 * We have been retransmitting for too long...  The RTT
4484 		 * we calculated is probably incorrect.  Reinitialize it.
4485 		 * Need to compensate for 0 tcp_rtt_sa.  Reset
4486 		 * tcp_rtt_update so that we won't accidentally cache a
4487 		 * bad value.  But only do this if this is not a zero
4488 		 * window probe.
4489 		 */
4490 		if (tcp->tcp_zero_win_probe == 0) {
4491 			tcp->tcp_rtt_sd += (tcp->tcp_rtt_sa >> 3) +
4492 			    (tcp->tcp_rtt_sa >> 5);
4493 			tcp->tcp_rtt_sa = 0;
4494 			tcp->tcp_rtt_update = 0;
4495 		}
4496 	}
4497 	tcp->tcp_timer_backoff++;
4498 	if ((ms = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
4499 	    tcp_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5)) <
4500 	    tcp_rexmit_interval_min) {
4501 		/*
4502 		 * This means the original RTO is tcp_rexmit_interval_min.
4503 		 * So we will use tcp_rexmit_interval_min as the RTO value
4504 		 * and do the backoff.
4505 		 */
4506 		ms = tcp_rexmit_interval_min << tcp->tcp_timer_backoff;
4507 	} else {
4508 		ms <<= tcp->tcp_timer_backoff;
4509 	}
4510 	if (ms > tcp_rexmit_interval_max) {
4511 		ms = tcp_rexmit_interval_max;
4512 		/*
4513 		 * ms is at max, decrement tcp_timer_backoff to avoid
4514 		 * overflow.
4515 		 */
4516 		tcp->tcp_timer_backoff--;
4517 	}
4518 	tcp->tcp_ms_we_have_waited += ms;
4519 	if (tcp->tcp_zero_win_probe == 0) {
4520 		tcp->tcp_rto = ms;
4521 	}
4522 	TCP_TIMER_RESTART(tcp, ms);
4523 	/*
4524 	 * This is after a timeout and tcp_rto is backed off.  Set
4525 	 * tcp_set_timer to 1 so that next time RTO is updated, we will
4526 	 * restart the timer with a correct value.
4527 	 */
4528 	tcp->tcp_set_timer = 1;
4529 	mss = tcp->tcp_snxt - tcp->tcp_suna;
4530 	if (mss > tcp->tcp_mss)
4531 		mss = tcp->tcp_mss;
4532 	if (mss > tcp->tcp_swnd && tcp->tcp_swnd != 0)
4533 		mss = tcp->tcp_swnd;
4534 
4535 	if ((mp = tcp->tcp_xmit_head) != NULL)
4536 		mp->b_prev = (mblk_t *)prom_gettime();
4537 	mp = tcp_xmit_mp(tcp, mp, mss, NULL, NULL, tcp->tcp_suna, B_TRUE, &mss,
4538 	    B_TRUE);
4539 	if (mp == NULL)
4540 		return;
4541 	tcp->tcp_csuna = tcp->tcp_snxt;
4542 	BUMP_MIB(tcp_mib.tcpRetransSegs);
4543 	UPDATE_MIB(tcp_mib.tcpRetransBytes, mss);
4544 	/* Dump the packet when debugging. */
4545 	TCP_DUMP_PACKET("tcp_timer", mp);
4546 
4547 	(void) ipv4_tcp_output(sock_id, mp);
4548 	freeb(mp);
4549 
4550 	/*
4551 	 * When slow start after retransmission begins, start with
4552 	 * this seq no.  tcp_rexmit_max marks the end of special slow
4553 	 * start phase.  tcp_snd_burst controls how many segments
4554 	 * can be sent because of an ack.
4555 	 */
4556 	tcp->tcp_rexmit_nxt = tcp->tcp_suna;
4557 	tcp->tcp_snd_burst = TCP_CWND_SS;
4558 	if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
4559 	    (tcp->tcp_unsent == 0)) {
4560 		tcp->tcp_rexmit_max = tcp->tcp_fss;
4561 	} else {
4562 		tcp->tcp_rexmit_max = tcp->tcp_snxt;
4563 	}
4564 	tcp->tcp_rexmit = B_TRUE;
4565 	tcp->tcp_dupack_cnt = 0;
4566 
4567 	/*
4568 	 * Remove all rexmit SACK blk to start from fresh.
4569 	 */
4570 	if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
4571 		TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list);
4572 		tcp->tcp_num_notsack_blk = 0;
4573 		tcp->tcp_cnt_notsack_list = 0;
4574 	}
4575 }
4576 
4577 /*
4578  * The TCP normal data output path.
4579  * NOTE: the logic of the fast path is duplicated from this function.
4580  */
4581 static void
4582 tcp_wput_data(tcp_t *tcp, mblk_t *mp, int sock_id)
4583 {
4584 	int		len;
4585 	mblk_t		*local_time;
4586 	mblk_t		*mp1;
4587 	uchar_t		*rptr;
4588 	uint32_t	snxt;
4589 	int		tail_unsent;
4590 	int		tcpstate;
4591 	int		usable = 0;
4592 	mblk_t		*xmit_tail;
4593 	int32_t		num_burst_seg;
4594 	int32_t		mss;
4595 	int32_t		num_sack_blk = 0;
4596 	int32_t		tcp_hdr_len;
4597 	ipaddr_t	*dst;
4598 	ipaddr_t	*src;
4599 
4600 #ifdef DEBUG
4601 	printf("tcp_wput_data(%d) ##############################\n", sock_id);
4602 #endif
4603 	tcpstate = tcp->tcp_state;
4604 	if (mp == NULL) {
4605 		/* Really tacky... but we need this for detached closes. */
4606 		len = tcp->tcp_unsent;
4607 		goto data_null;
4608 	}
4609 
4610 	/*
4611 	 * Don't allow data after T_ORDREL_REQ or T_DISCON_REQ,
4612 	 * or before a connection attempt has begun.
4613 	 *
4614 	 * The following should not happen in inetboot....
4615 	 */
4616 	if (tcpstate < TCPS_SYN_SENT || tcpstate > TCPS_CLOSE_WAIT ||
4617 	    (tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
4618 		if ((tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
4619 			printf("tcp_wput_data: data after ordrel, %s\n",
4620 			    tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
4621 		}
4622 		freemsg(mp);
4623 		return;
4624 	}
4625 
4626 	/* Strip empties */
4627 	for (;;) {
4628 		assert((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
4629 		    (uintptr_t)INT_MAX);
4630 		len = (int)(mp->b_wptr - mp->b_rptr);
4631 		if (len > 0)
4632 			break;
4633 		mp1 = mp;
4634 		mp = mp->b_cont;
4635 		freeb(mp1);
4636 		if (mp == NULL) {
4637 			return;
4638 		}
4639 	}
4640 
4641 	/* If we are the first on the list ... */
4642 	if (tcp->tcp_xmit_head == NULL) {
4643 		tcp->tcp_xmit_head = mp;
4644 		tcp->tcp_xmit_tail = mp;
4645 		tcp->tcp_xmit_tail_unsent = len;
4646 	} else {
4647 		tcp->tcp_xmit_last->b_cont = mp;
4648 		len += tcp->tcp_unsent;
4649 	}
4650 
4651 	/* Tack on however many more positive length mblks we have */
4652 	if ((mp1 = mp->b_cont) != NULL) {
4653 		do {
4654 			int tlen;
4655 			assert((uintptr_t)(mp1->b_wptr -
4656 			    mp1->b_rptr) <= (uintptr_t)INT_MAX);
4657 			tlen = (int)(mp1->b_wptr - mp1->b_rptr);
4658 			if (tlen <= 0) {
4659 				mp->b_cont = mp1->b_cont;
4660 				freeb(mp1);
4661 			} else {
4662 				len += tlen;
4663 				mp = mp1;
4664 			}
4665 		} while ((mp1 = mp->b_cont) != NULL);
4666 	}
4667 	tcp->tcp_xmit_last = mp;
4668 	tcp->tcp_unsent = len;
4669 
4670 data_null:
4671 	snxt = tcp->tcp_snxt;
4672 	xmit_tail = tcp->tcp_xmit_tail;
4673 	tail_unsent = tcp->tcp_xmit_tail_unsent;
4674 
4675 	/*
4676 	 * Note that tcp_mss has been adjusted to take into account the
4677 	 * timestamp option if applicable.  Because SACK options do not
4678 	 * appear in every TCP segments and they are of variable lengths,
4679 	 * they cannot be included in tcp_mss.  Thus we need to calculate
4680 	 * the actual segment length when we need to send a segment which
4681 	 * includes SACK options.
4682 	 */
4683 	if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
4684 		int32_t	opt_len;
4685 
4686 		num_sack_blk = MIN(tcp->tcp_max_sack_blk,
4687 		    tcp->tcp_num_sack_blk);
4688 		opt_len = num_sack_blk * sizeof (sack_blk_t) + TCPOPT_NOP_LEN *
4689 		    2 + TCPOPT_HEADER_LEN;
4690 		mss = tcp->tcp_mss - opt_len;
4691 		tcp_hdr_len = tcp->tcp_hdr_len + opt_len;
4692 	} else {
4693 		mss = tcp->tcp_mss;
4694 		tcp_hdr_len = tcp->tcp_hdr_len;
4695 	}
4696 
4697 	if ((tcp->tcp_suna == snxt) &&
4698 	    (prom_gettime() - tcp->tcp_last_recv_time) >= tcp->tcp_rto) {
4699 		tcp->tcp_cwnd = MIN(tcp_slow_start_after_idle * mss,
4700 		    MIN(4 * mss, MAX(2 * mss, 4380 / mss * mss)));
4701 	}
4702 	if (tcpstate == TCPS_SYN_RCVD) {
4703 		/*
4704 		 * The three-way connection establishment handshake is not
4705 		 * complete yet. We want to queue the data for transmission
4706 		 * after entering ESTABLISHED state (RFC793). Setting usable to
4707 		 * zero cause a jump to "done" label effectively leaving data
4708 		 * on the queue.
4709 		 */
4710 
4711 		usable = 0;
4712 	} else {
4713 		int usable_r = tcp->tcp_swnd;
4714 
4715 		/*
4716 		 * In the special case when cwnd is zero, which can only
4717 		 * happen if the connection is ECN capable, return now.
4718 		 * New segments is sent using tcp_timer().  The timer
4719 		 * is set in tcp_rput_data().
4720 		 */
4721 		if (tcp->tcp_cwnd == 0) {
4722 			/*
4723 			 * Note that tcp_cwnd is 0 before 3-way handshake is
4724 			 * finished.
4725 			 */
4726 			assert(tcp->tcp_ecn_ok ||
4727 			    tcp->tcp_state < TCPS_ESTABLISHED);
4728 			return;
4729 		}
4730 
4731 		/* usable = MIN(swnd, cwnd) - unacked_bytes */
4732 		if (usable_r > tcp->tcp_cwnd)
4733 			usable_r = tcp->tcp_cwnd;
4734 
4735 		/* NOTE: trouble if xmitting while SYN not acked? */
4736 		usable_r -= snxt;
4737 		usable_r += tcp->tcp_suna;
4738 
4739 		/* usable = MIN(usable, unsent) */
4740 		if (usable_r > len)
4741 			usable_r = len;
4742 
4743 		/* usable = MAX(usable, {1 for urgent, 0 for data}) */
4744 		if (usable_r != 0)
4745 			usable = usable_r;
4746 	}
4747 
4748 	local_time = (mblk_t *)prom_gettime();
4749 
4750 	/*
4751 	 * "Our" Nagle Algorithm.  This is not the same as in the old
4752 	 * BSD.  This is more in line with the true intent of Nagle.
4753 	 *
4754 	 * The conditions are:
4755 	 * 1. The amount of unsent data (or amount of data which can be
4756 	 *    sent, whichever is smaller) is less than Nagle limit.
4757 	 * 2. The last sent size is also less than Nagle limit.
4758 	 * 3. There is unack'ed data.
4759 	 * 4. Urgent pointer is not set.  Send urgent data ignoring the
4760 	 *    Nagle algorithm.  This reduces the probability that urgent
4761 	 *    bytes get "merged" together.
4762 	 * 5. The app has not closed the connection.  This eliminates the
4763 	 *    wait time of the receiving side waiting for the last piece of
4764 	 *    (small) data.
4765 	 *
4766 	 * If all are satisified, exit without sending anything.  Note
4767 	 * that Nagle limit can be smaller than 1 MSS.  Nagle limit is
4768 	 * the smaller of 1 MSS and global tcp_naglim_def (default to be
4769 	 * 4095).
4770 	 */
4771 	if (usable < (int)tcp->tcp_naglim &&
4772 	    tcp->tcp_naglim > tcp->tcp_last_sent_len &&
4773 	    snxt != tcp->tcp_suna &&
4774 	    !(tcp->tcp_valid_bits & TCP_URG_VALID))
4775 		goto done;
4776 
4777 	num_burst_seg = tcp->tcp_snd_burst;
4778 	for (;;) {
4779 		tcph_t		*tcph;
4780 		mblk_t		*new_mp;
4781 
4782 		if (num_burst_seg-- == 0)
4783 			goto done;
4784 
4785 		len = mss;
4786 		if (len > usable) {
4787 			len = usable;
4788 			if (len <= 0) {
4789 				/* Terminate the loop */
4790 				goto done;
4791 			}
4792 			/*
4793 			 * Sender silly-window avoidance.
4794 			 * Ignore this if we are going to send a
4795 			 * zero window probe out.
4796 			 *
4797 			 * TODO: force data into microscopic window ??
4798 			 *	==> (!pushed || (unsent > usable))
4799 			 */
4800 			if (len < (tcp->tcp_max_swnd >> 1) &&
4801 			    (tcp->tcp_unsent - (snxt - tcp->tcp_snxt)) > len &&
4802 			    !((tcp->tcp_valid_bits & TCP_URG_VALID) &&
4803 			    len == 1) && (! tcp->tcp_zero_win_probe)) {
4804 				/*
4805 				 * If the retransmit timer is not running
4806 				 * we start it so that we will retransmit
4807 				 * in the case when the the receiver has
4808 				 * decremented the window.
4809 				 */
4810 				if (snxt == tcp->tcp_snxt &&
4811 				    snxt == tcp->tcp_suna) {
4812 					/*
4813 					 * We are not supposed to send
4814 					 * anything.  So let's wait a little
4815 					 * bit longer before breaking SWS
4816 					 * avoidance.
4817 					 *
4818 					 * What should the value be?
4819 					 * Suggestion: MAX(init rexmit time,
4820 					 * tcp->tcp_rto)
4821 					 */
4822 					TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4823 				}
4824 				goto done;
4825 			}
4826 		}
4827 
4828 		tcph = tcp->tcp_tcph;
4829 
4830 		usable -= len;	/* Approximate - can be adjusted later */
4831 		if (usable > 0)
4832 			tcph->th_flags[0] = TH_ACK;
4833 		else
4834 			tcph->th_flags[0] = (TH_ACK | TH_PUSH);
4835 
4836 		U32_TO_ABE32(snxt, tcph->th_seq);
4837 
4838 		if (tcp->tcp_valid_bits) {
4839 			uchar_t		*prev_rptr = xmit_tail->b_rptr;
4840 			uint32_t	prev_snxt = tcp->tcp_snxt;
4841 
4842 			if (tail_unsent == 0) {
4843 				assert(xmit_tail->b_cont != NULL);
4844 				xmit_tail = xmit_tail->b_cont;
4845 				prev_rptr = xmit_tail->b_rptr;
4846 				tail_unsent = (int)(xmit_tail->b_wptr -
4847 				    xmit_tail->b_rptr);
4848 			} else {
4849 				xmit_tail->b_rptr = xmit_tail->b_wptr -
4850 				    tail_unsent;
4851 			}
4852 			mp = tcp_xmit_mp(tcp, xmit_tail, len, NULL, NULL,
4853 			    snxt, B_FALSE, (uint32_t *)&len, B_FALSE);
4854 			/* Restore tcp_snxt so we get amount sent right. */
4855 			tcp->tcp_snxt = prev_snxt;
4856 			if (prev_rptr == xmit_tail->b_rptr)
4857 				xmit_tail->b_prev = local_time;
4858 			else
4859 				xmit_tail->b_rptr = prev_rptr;
4860 
4861 			if (mp == NULL)
4862 				break;
4863 
4864 			mp1 = mp->b_cont;
4865 
4866 			snxt += len;
4867 			tcp->tcp_last_sent_len = (ushort_t)len;
4868 			while (mp1->b_cont) {
4869 				xmit_tail = xmit_tail->b_cont;
4870 				xmit_tail->b_prev = local_time;
4871 				mp1 = mp1->b_cont;
4872 			}
4873 			tail_unsent = xmit_tail->b_wptr - mp1->b_wptr;
4874 			BUMP_MIB(tcp_mib.tcpOutDataSegs);
4875 			UPDATE_MIB(tcp_mib.tcpOutDataBytes, len);
4876 			/* Dump the packet when debugging. */
4877 			TCP_DUMP_PACKET("tcp_wput_data (valid bits)", mp);
4878 			(void) ipv4_tcp_output(sock_id, mp);
4879 			freeb(mp);
4880 			continue;
4881 		}
4882 
4883 		snxt += len;	/* Adjust later if we don't send all of len */
4884 		BUMP_MIB(tcp_mib.tcpOutDataSegs);
4885 		UPDATE_MIB(tcp_mib.tcpOutDataBytes, len);
4886 
4887 		if (tail_unsent) {
4888 			/* Are the bytes above us in flight? */
4889 			rptr = xmit_tail->b_wptr - tail_unsent;
4890 			if (rptr != xmit_tail->b_rptr) {
4891 				tail_unsent -= len;
4892 				len += tcp_hdr_len;
4893 				tcp->tcp_ipha->ip_len = htons(len);
4894 				mp = dupb(xmit_tail);
4895 				if (!mp)
4896 					break;
4897 				mp->b_rptr = rptr;
4898 				goto must_alloc;
4899 			}
4900 		} else {
4901 			xmit_tail = xmit_tail->b_cont;
4902 			assert((uintptr_t)(xmit_tail->b_wptr -
4903 			    xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
4904 			tail_unsent = (int)(xmit_tail->b_wptr -
4905 			    xmit_tail->b_rptr);
4906 		}
4907 
4908 		tail_unsent -= len;
4909 		tcp->tcp_last_sent_len = (ushort_t)len;
4910 
4911 		len += tcp_hdr_len;
4912 		if (tcp->tcp_ipversion == IPV4_VERSION)
4913 			tcp->tcp_ipha->ip_len = htons(len);
4914 
4915 		xmit_tail->b_prev = local_time;
4916 
4917 		mp = dupb(xmit_tail);
4918 		if (mp == NULL)
4919 			goto out_of_mem;
4920 
4921 		len = tcp_hdr_len;
4922 		/*
4923 		 * There are four reasons to allocate a new hdr mblk:
4924 		 *  1) The bytes above us are in use by another packet
4925 		 *  2) We don't have good alignment
4926 		 *  3) The mblk is being shared
4927 		 *  4) We don't have enough room for a header
4928 		 */
4929 		rptr = mp->b_rptr - len;
4930 		if (!OK_32PTR(rptr) ||
4931 		    rptr < mp->b_datap) {
4932 			/* NOTE: we assume allocb returns an OK_32PTR */
4933 
4934 		must_alloc:;
4935 			mp1 = allocb(tcp->tcp_ip_hdr_len + TCP_MAX_HDR_LENGTH +
4936 			    tcp_wroff_xtra, 0);
4937 			if (mp1 == NULL) {
4938 				freemsg(mp);
4939 				goto out_of_mem;
4940 			}
4941 			mp1->b_cont = mp;
4942 			mp = mp1;
4943 			/* Leave room for Link Level header */
4944 			len = tcp_hdr_len;
4945 			rptr = &mp->b_rptr[tcp_wroff_xtra];
4946 			mp->b_wptr = &rptr[len];
4947 		}
4948 
4949 		if (tcp->tcp_snd_ts_ok) {
4950 			U32_TO_BE32((uint32_t)local_time,
4951 				(char *)tcph+TCP_MIN_HEADER_LENGTH+4);
4952 			U32_TO_BE32(tcp->tcp_ts_recent,
4953 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
4954 		} else {
4955 			assert(tcp->tcp_tcp_hdr_len == TCP_MIN_HEADER_LENGTH);
4956 		}
4957 
4958 		mp->b_rptr = rptr;
4959 
4960 		/* Copy the template header. */
4961 		dst = (ipaddr_t *)rptr;
4962 		src = (ipaddr_t *)tcp->tcp_iphc;
4963 		dst[0] = src[0];
4964 		dst[1] = src[1];
4965 		dst[2] = src[2];
4966 		dst[3] = src[3];
4967 		dst[4] = src[4];
4968 		dst[5] = src[5];
4969 		dst[6] = src[6];
4970 		dst[7] = src[7];
4971 		dst[8] = src[8];
4972 		dst[9] = src[9];
4973 		len = tcp->tcp_hdr_len;
4974 		if (len -= 40) {
4975 			len >>= 2;
4976 			dst += 10;
4977 			src += 10;
4978 			do {
4979 				*dst++ = *src++;
4980 			} while (--len);
4981 		}
4982 
4983 		/*
4984 		 * Set tcph to point to the header of the outgoing packet,
4985 		 * not to the template header.
4986 		 */
4987 		tcph = (tcph_t *)(rptr + tcp->tcp_ip_hdr_len);
4988 
4989 		/*
4990 		 * Set the ECN info in the TCP header if it is not a zero
4991 		 * window probe.  Zero window probe is only sent in
4992 		 * tcp_wput_data() and tcp_timer().
4993 		 */
4994 		if (tcp->tcp_ecn_ok && !tcp->tcp_zero_win_probe) {
4995 			SET_ECT(tcp, rptr);
4996 
4997 			if (tcp->tcp_ecn_echo_on)
4998 				tcph->th_flags[0] |= TH_ECE;
4999 			if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
5000 				tcph->th_flags[0] |= TH_CWR;
5001 				tcp->tcp_ecn_cwr_sent = B_TRUE;
5002 			}
5003 		}
5004 
5005 		/* Fill in SACK options */
5006 		if (num_sack_blk > 0) {
5007 			uchar_t *wptr = rptr + tcp->tcp_hdr_len;
5008 			sack_blk_t *tmp;
5009 			int32_t	i;
5010 
5011 			wptr[0] = TCPOPT_NOP;
5012 			wptr[1] = TCPOPT_NOP;
5013 			wptr[2] = TCPOPT_SACK;
5014 			wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
5015 			    sizeof (sack_blk_t);
5016 			wptr += TCPOPT_REAL_SACK_LEN;
5017 
5018 			tmp = tcp->tcp_sack_list;
5019 			for (i = 0; i < num_sack_blk; i++) {
5020 				U32_TO_BE32(tmp[i].begin, wptr);
5021 				wptr += sizeof (tcp_seq);
5022 				U32_TO_BE32(tmp[i].end, wptr);
5023 				wptr += sizeof (tcp_seq);
5024 			}
5025 			tcph->th_offset_and_rsrvd[0] += ((num_sack_blk * 2 + 1)
5026 			    << 4);
5027 		}
5028 
5029 		if (tail_unsent) {
5030 			mp1 = mp->b_cont;
5031 			if (mp1 == NULL)
5032 				mp1 = mp;
5033 			/*
5034 			 * If we're a little short, tack on more mblks
5035 			 * as long as we don't need to split an mblk.
5036 			 */
5037 			while (tail_unsent < 0 &&
5038 			    tail_unsent + (int)(xmit_tail->b_cont->b_wptr -
5039 			    xmit_tail->b_cont->b_rptr) <= 0) {
5040 				xmit_tail = xmit_tail->b_cont;
5041 				/* Stash for rtt use later */
5042 				xmit_tail->b_prev = local_time;
5043 				mp1->b_cont = dupb(xmit_tail);
5044 				mp1 = mp1->b_cont;
5045 				assert((uintptr_t)(xmit_tail->b_wptr -
5046 				    xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
5047 				tail_unsent += (int)(xmit_tail->b_wptr -
5048 				    xmit_tail->b_rptr);
5049 				if (mp1 == NULL) {
5050 					freemsg(mp);
5051 					goto out_of_mem;
5052 				}
5053 			}
5054 			/* Trim back any surplus on the last mblk */
5055 			if (tail_unsent > 0)
5056 				mp1->b_wptr -= tail_unsent;
5057 			if (tail_unsent < 0) {
5058 				uint32_t ip_len;
5059 
5060 				/*
5061 				 * We did not send everything we could in
5062 				 * order to preserve mblk boundaries.
5063 				 */
5064 				usable -= tail_unsent;
5065 				snxt += tail_unsent;
5066 				tcp->tcp_last_sent_len += tail_unsent;
5067 				UPDATE_MIB(tcp_mib.tcpOutDataBytes,
5068 				    tail_unsent);
5069 				/* Adjust the IP length field. */
5070 				ip_len = ntohs(((struct ip *)rptr)->ip_len) +
5071 				    tail_unsent;
5072 				((struct ip *)rptr)->ip_len = htons(ip_len);
5073 				tail_unsent = 0;
5074 			}
5075 		}
5076 
5077 		if (mp == NULL)
5078 			goto out_of_mem;
5079 
5080 		/*
5081 		 * Performance hit!  We need to pullup the whole message
5082 		 * in order to do checksum and for the MAC output routine.
5083 		 */
5084 		if (mp->b_cont != NULL) {
5085 			int mp_size;
5086 #ifdef	DEBUG
5087 			printf("Multiple mblk %d\n", msgdsize(mp));
5088 #endif
5089 			new_mp = allocb(msgdsize(mp) + tcp_wroff_xtra, 0);
5090 			new_mp->b_rptr += tcp_wroff_xtra;
5091 			new_mp->b_wptr = new_mp->b_rptr;
5092 			while (mp != NULL) {
5093 				mp_size = mp->b_wptr - mp->b_rptr;
5094 				bcopy(mp->b_rptr, new_mp->b_wptr, mp_size);
5095 				new_mp->b_wptr += mp_size;
5096 				mp = mp->b_cont;
5097 			}
5098 			freemsg(mp);
5099 			mp = new_mp;
5100 		}
5101 		tcp_set_cksum(mp);
5102 		((struct ip *)mp->b_rptr)->ip_ttl = (uint8_t)tcp_ipv4_ttl;
5103 		TCP_DUMP_PACKET("tcp_wput_data", mp);
5104 		(void) ipv4_tcp_output(sock_id, mp);
5105 		freemsg(mp);
5106 	}
5107 out_of_mem:;
5108 	/* Pretend that all we were trying to send really got sent */
5109 	if (tail_unsent < 0) {
5110 		do {
5111 			xmit_tail = xmit_tail->b_cont;
5112 			xmit_tail->b_prev = local_time;
5113 			assert((uintptr_t)(xmit_tail->b_wptr -
5114 			    xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
5115 			tail_unsent += (int)(xmit_tail->b_wptr -
5116 			    xmit_tail->b_rptr);
5117 		} while (tail_unsent < 0);
5118 	}
5119 done:;
5120 	tcp->tcp_xmit_tail = xmit_tail;
5121 	tcp->tcp_xmit_tail_unsent = tail_unsent;
5122 	len = tcp->tcp_snxt - snxt;
5123 	if (len) {
5124 		/*
5125 		 * If new data was sent, need to update the notsack
5126 		 * list, which is, afterall, data blocks that have
5127 		 * not been sack'ed by the receiver.  New data is
5128 		 * not sack'ed.
5129 		 */
5130 		if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
5131 			/* len is a negative value. */
5132 			tcp->tcp_pipe -= len;
5133 			tcp_notsack_update(&(tcp->tcp_notsack_list),
5134 			    tcp->tcp_snxt, snxt,
5135 			    &(tcp->tcp_num_notsack_blk),
5136 			    &(tcp->tcp_cnt_notsack_list));
5137 		}
5138 		tcp->tcp_snxt = snxt + tcp->tcp_fin_sent;
5139 		tcp->tcp_rack = tcp->tcp_rnxt;
5140 		tcp->tcp_rack_cnt = 0;
5141 		if ((snxt + len) == tcp->tcp_suna) {
5142 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
5143 		}
5144 		/*
5145 		 * Note that len is the amount we just sent but with a negative
5146 		 * sign. We update tcp_unsent here since we may come back to
5147 		 * tcp_wput_data from tcp_state_wait.
5148 		 */
5149 		len += tcp->tcp_unsent;
5150 		tcp->tcp_unsent = len;
5151 
5152 		/*
5153 		 * Let's wait till all the segments have been acked, since we
5154 		 * don't have a timer.
5155 		 */
5156 		(void) tcp_state_wait(sock_id, tcp, TCPS_ALL_ACKED);
5157 		return;
5158 	} else if (snxt == tcp->tcp_suna && tcp->tcp_swnd == 0) {
5159 		/*
5160 		 * Didn't send anything. Make sure the timer is running
5161 		 * so that we will probe a zero window.
5162 		 */
5163 		TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
5164 	}
5165 
5166 	/* Note that len is the amount we just sent but with a negative sign */
5167 	len += tcp->tcp_unsent;
5168 	tcp->tcp_unsent = len;
5169 
5170 }
5171 
5172 static void
5173 tcp_time_wait_processing(tcp_t *tcp, mblk_t *mp,
5174     uint32_t seg_seq, uint32_t seg_ack, int seg_len, tcph_t *tcph,
5175     int sock_id)
5176 {
5177 	int32_t		bytes_acked;
5178 	int32_t		gap;
5179 	int32_t		rgap;
5180 	tcp_opt_t	tcpopt;
5181 	uint_t		flags;
5182 	uint32_t	new_swnd = 0;
5183 
5184 #ifdef DEBUG
5185 	printf("Time wait processing called ###############3\n");
5186 #endif
5187 
5188 	/* Just make sure we send the right sock_id to tcp_clean_death */
5189 	if ((sockets[sock_id].pcb == NULL) || (sockets[sock_id].pcb != tcp))
5190 		sock_id = -1;
5191 
5192 	flags = (unsigned int)tcph->th_flags[0] & 0xFF;
5193 	new_swnd = BE16_TO_U16(tcph->th_win) <<
5194 	    ((tcph->th_flags[0] & TH_SYN) ? 0 : tcp->tcp_snd_ws);
5195 	if (tcp->tcp_snd_ts_ok) {
5196 		if (!tcp_paws_check(tcp, tcph, &tcpopt)) {
5197 			freemsg(mp);
5198 			tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
5199 			    tcp->tcp_rnxt, TH_ACK, 0, -1);
5200 			return;
5201 		}
5202 	}
5203 	gap = seg_seq - tcp->tcp_rnxt;
5204 	rgap = tcp->tcp_rwnd - (gap + seg_len);
5205 	if (gap < 0) {
5206 		BUMP_MIB(tcp_mib.tcpInDataDupSegs);
5207 		UPDATE_MIB(tcp_mib.tcpInDataDupBytes,
5208 		    (seg_len > -gap ? -gap : seg_len));
5209 		seg_len += gap;
5210 		if (seg_len < 0 || (seg_len == 0 && !(flags & TH_FIN))) {
5211 			if (flags & TH_RST) {
5212 				freemsg(mp);
5213 				return;
5214 			}
5215 			if ((flags & TH_FIN) && seg_len == -1) {
5216 				/*
5217 				 * When TCP receives a duplicate FIN in
5218 				 * TIME_WAIT state, restart the 2 MSL timer.
5219 				 * See page 73 in RFC 793. Make sure this TCP
5220 				 * is already on the TIME_WAIT list. If not,
5221 				 * just restart the timer.
5222 				 */
5223 				tcp_time_wait_remove(tcp);
5224 				tcp_time_wait_append(tcp);
5225 				TCP_TIMER_RESTART(tcp, tcp_time_wait_interval);
5226 				tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
5227 				    tcp->tcp_rnxt, TH_ACK, 0, -1);
5228 				freemsg(mp);
5229 				return;
5230 			}
5231 			flags |=  TH_ACK_NEEDED;
5232 			seg_len = 0;
5233 			goto process_ack;
5234 		}
5235 
5236 		/* Fix seg_seq, and chew the gap off the front. */
5237 		seg_seq = tcp->tcp_rnxt;
5238 	}
5239 
5240 	if ((flags & TH_SYN) && gap > 0 && rgap < 0) {
5241 		/*
5242 		 * Make sure that when we accept the connection, pick
5243 		 * an ISS greater than (tcp_snxt + ISS_INCR/2) for the
5244 		 * old connection.
5245 		 *
5246 		 * The next ISS generated is equal to tcp_iss_incr_extra
5247 		 * + ISS_INCR/2 + other components depending on the
5248 		 * value of tcp_strong_iss.  We pre-calculate the new
5249 		 * ISS here and compare with tcp_snxt to determine if
5250 		 * we need to make adjustment to tcp_iss_incr_extra.
5251 		 *
5252 		 * Note that since we are now in the global queue
5253 		 * perimeter and need to do a lateral_put() to the
5254 		 * listener queue, there can be other connection requests/
5255 		 * attempts while the lateral_put() is going on.  That
5256 		 * means what we calculate here may not be correct.  This
5257 		 * is extremely difficult to solve unless TCP and IP
5258 		 * modules are merged and there is no perimeter, but just
5259 		 * locks.  The above calculation is ugly and is a
5260 		 * waste of CPU cycles...
5261 		 */
5262 		uint32_t new_iss = tcp_iss_incr_extra;
5263 		int32_t adj;
5264 
5265 		/* Add time component and min random (i.e. 1). */
5266 		new_iss += (prom_gettime() >> ISS_NSEC_SHT) + 1;
5267 		if ((adj = (int32_t)(tcp->tcp_snxt - new_iss)) > 0) {
5268 			/*
5269 			 * New ISS not guaranteed to be ISS_INCR/2
5270 			 * ahead of the current tcp_snxt, so add the
5271 			 * difference to tcp_iss_incr_extra.
5272 			 */
5273 			tcp_iss_incr_extra += adj;
5274 		}
5275 		tcp_clean_death(sock_id, tcp, 0);
5276 
5277 		/*
5278 		 * This is a passive open.  Right now we do not
5279 		 * do anything...
5280 		 */
5281 		freemsg(mp);
5282 		return;
5283 	}
5284 
5285 	/*
5286 	 * rgap is the amount of stuff received out of window.  A negative
5287 	 * value is the amount out of window.
5288 	 */
5289 	if (rgap < 0) {
5290 		BUMP_MIB(tcp_mib.tcpInDataPastWinSegs);
5291 		UPDATE_MIB(tcp_mib.tcpInDataPastWinBytes, -rgap);
5292 		/* Fix seg_len and make sure there is something left. */
5293 		seg_len += rgap;
5294 		if (seg_len <= 0) {
5295 			if (flags & TH_RST) {
5296 				freemsg(mp);
5297 				return;
5298 			}
5299 			flags |=  TH_ACK_NEEDED;
5300 			seg_len = 0;
5301 			goto process_ack;
5302 		}
5303 	}
5304 	/*
5305 	 * Check whether we can update tcp_ts_recent.  This test is
5306 	 * NOT the one in RFC 1323 3.4.  It is from Braden, 1993, "TCP
5307 	 * Extensions for High Performance: An Update", Internet Draft.
5308 	 */
5309 	if (tcp->tcp_snd_ts_ok &&
5310 	    TSTMP_GEQ(tcpopt.tcp_opt_ts_val, tcp->tcp_ts_recent) &&
5311 	    SEQ_LEQ(seg_seq, tcp->tcp_rack)) {
5312 		tcp->tcp_ts_recent = tcpopt.tcp_opt_ts_val;
5313 		tcp->tcp_last_rcv_lbolt = prom_gettime();
5314 	}
5315 
5316 	if (seg_seq != tcp->tcp_rnxt && seg_len > 0) {
5317 		/* Always ack out of order packets */
5318 		flags |= TH_ACK_NEEDED;
5319 		seg_len = 0;
5320 	} else if (seg_len > 0) {
5321 		BUMP_MIB(tcp_mib.tcpInDataInorderSegs);
5322 		UPDATE_MIB(tcp_mib.tcpInDataInorderBytes, seg_len);
5323 	}
5324 	if (flags & TH_RST) {
5325 		freemsg(mp);
5326 		(void) tcp_clean_death(sock_id, tcp, 0);
5327 		return;
5328 	}
5329 	if (flags & TH_SYN) {
5330 		freemsg(mp);
5331 		tcp_xmit_ctl("TH_SYN", tcp, NULL, seg_ack, seg_seq + 1,
5332 		    TH_RST|TH_ACK, 0, -1);
5333 		/*
5334 		 * Do not delete the TCP structure if it is in
5335 		 * TIME_WAIT state.  Refer to RFC 1122, 4.2.2.13.
5336 		 */
5337 		return;
5338 	}
5339 process_ack:
5340 	if (flags & TH_ACK) {
5341 		bytes_acked = (int)(seg_ack - tcp->tcp_suna);
5342 		if (bytes_acked <= 0) {
5343 			if (bytes_acked == 0 && seg_len == 0 &&
5344 			    new_swnd == tcp->tcp_swnd)
5345 				BUMP_MIB(tcp_mib.tcpInDupAck);
5346 		} else {
5347 			/* Acks something not sent */
5348 			flags |= TH_ACK_NEEDED;
5349 		}
5350 	}
5351 	freemsg(mp);
5352 	if (flags & TH_ACK_NEEDED) {
5353 		/*
5354 		 * Time to send an ack for some reason.
5355 		 */
5356 		tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
5357 		    tcp->tcp_rnxt, TH_ACK, 0, -1);
5358 	}
5359 }
5360 
5361 static int
5362 tcp_init_values(tcp_t *tcp, struct inetboot_socket *isp)
5363 {
5364 	int	err;
5365 
5366 	tcp->tcp_family = AF_INET;
5367 	tcp->tcp_ipversion = IPV4_VERSION;
5368 
5369 	/*
5370 	 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
5371 	 * will be close to tcp_rexmit_interval_initial.  By doing this, we
5372 	 * allow the algorithm to adjust slowly to large fluctuations of RTT
5373 	 * during first few transmissions of a connection as seen in slow
5374 	 * links.
5375 	 */
5376 	tcp->tcp_rtt_sa = tcp_rexmit_interval_initial << 2;
5377 	tcp->tcp_rtt_sd = tcp_rexmit_interval_initial >> 1;
5378 	tcp->tcp_rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
5379 	    tcp_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
5380 	    tcp_conn_grace_period;
5381 	if (tcp->tcp_rto < tcp_rexmit_interval_min)
5382 		tcp->tcp_rto = tcp_rexmit_interval_min;
5383 	tcp->tcp_timer_backoff = 0;
5384 	tcp->tcp_ms_we_have_waited = 0;
5385 	tcp->tcp_last_recv_time = prom_gettime();
5386 	tcp->tcp_cwnd_max = tcp_cwnd_max_;
5387 	tcp->tcp_snd_burst = TCP_CWND_INFINITE;
5388 	tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
5389 	/* For Ethernet, the mtu returned is actually 1550... */
5390 	if (mac_get_type() == IFT_ETHER) {
5391 		tcp->tcp_if_mtu = mac_get_mtu() - 50;
5392 	} else {
5393 		tcp->tcp_if_mtu = mac_get_mtu();
5394 	}
5395 	tcp->tcp_mss = tcp->tcp_if_mtu;
5396 
5397 	tcp->tcp_first_timer_threshold = tcp_ip_notify_interval;
5398 	tcp->tcp_first_ctimer_threshold = tcp_ip_notify_cinterval;
5399 	tcp->tcp_second_timer_threshold = tcp_ip_abort_interval;
5400 	/*
5401 	 * Fix it to tcp_ip_abort_linterval later if it turns out to be a
5402 	 * passive open.
5403 	 */
5404 	tcp->tcp_second_ctimer_threshold = tcp_ip_abort_cinterval;
5405 
5406 	tcp->tcp_naglim = tcp_naglim_def;
5407 
5408 	/* NOTE:  ISS is now set in tcp_adapt_ire(). */
5409 
5410 	/* Initialize the header template */
5411 	if (tcp->tcp_ipversion == IPV4_VERSION) {
5412 		err = tcp_header_init_ipv4(tcp);
5413 	}
5414 	if (err)
5415 		return (err);
5416 
5417 	/*
5418 	 * Init the window scale to the max so tcp_rwnd_set() won't pare
5419 	 * down tcp_rwnd. tcp_adapt_ire() will set the right value later.
5420 	 */
5421 	tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
5422 	tcp->tcp_xmit_lowater = tcp_xmit_lowat;
5423 	if (isp != NULL) {
5424 		tcp->tcp_xmit_hiwater = isp->so_sndbuf;
5425 		tcp->tcp_rwnd = isp->so_rcvbuf;
5426 		tcp->tcp_rwnd_max = isp->so_rcvbuf;
5427 	}
5428 	tcp->tcp_state = TCPS_IDLE;
5429 	return (0);
5430 }
5431 
5432 /*
5433  * Initialize the IPv4 header. Loses any record of any IP options.
5434  */
5435 static int
5436 tcp_header_init_ipv4(tcp_t *tcp)
5437 {
5438 	tcph_t		*tcph;
5439 
5440 	/*
5441 	 * This is a simple initialization. If there's
5442 	 * already a template, it should never be too small,
5443 	 * so reuse it.  Otherwise, allocate space for the new one.
5444 	 */
5445 	if (tcp->tcp_iphc != NULL) {
5446 		assert(tcp->tcp_iphc_len >= TCP_MAX_COMBINED_HEADER_LENGTH);
5447 		bzero(tcp->tcp_iphc, tcp->tcp_iphc_len);
5448 	} else {
5449 		tcp->tcp_iphc_len = TCP_MAX_COMBINED_HEADER_LENGTH;
5450 		tcp->tcp_iphc = bkmem_zalloc(tcp->tcp_iphc_len);
5451 		if (tcp->tcp_iphc == NULL) {
5452 			tcp->tcp_iphc_len = 0;
5453 			return (ENOMEM);
5454 		}
5455 	}
5456 	tcp->tcp_ipha = (struct ip *)tcp->tcp_iphc;
5457 	tcp->tcp_ipversion = IPV4_VERSION;
5458 
5459 	/*
5460 	 * Note that it does not include TCP options yet.  It will
5461 	 * after the connection is established.
5462 	 */
5463 	tcp->tcp_hdr_len = sizeof (struct ip) + sizeof (tcph_t);
5464 	tcp->tcp_tcp_hdr_len = sizeof (tcph_t);
5465 	tcp->tcp_ip_hdr_len = sizeof (struct ip);
5466 	tcp->tcp_ipha->ip_v = IP_VERSION;
5467 	/* We don't support IP options... */
5468 	tcp->tcp_ipha->ip_hl = IP_SIMPLE_HDR_LENGTH_IN_WORDS;
5469 	tcp->tcp_ipha->ip_p = IPPROTO_TCP;
5470 	/* We are not supposed to do PMTU discovery... */
5471 	tcp->tcp_ipha->ip_sum = 0;
5472 
5473 	tcph = (tcph_t *)(tcp->tcp_iphc + sizeof (struct ip));
5474 	tcp->tcp_tcph = tcph;
5475 	tcph->th_offset_and_rsrvd[0] = (5 << 4);
5476 	return (0);
5477 }
5478 
5479 /*
5480  * Send out a control packet on the tcp connection specified.  This routine
5481  * is typically called where we need a simple ACK or RST generated.
5482  *
5483  * This function is called with or without a mp.
5484  */
5485 static void
5486 tcp_xmit_ctl(char *str, tcp_t *tcp, mblk_t *mp, uint32_t seq,
5487     uint32_t ack, int ctl, uint_t ip_hdr_len, int sock_id)
5488 {
5489 	uchar_t		*rptr;
5490 	tcph_t		*tcph;
5491 	struct ip	*iph = NULL;
5492 	int		tcp_hdr_len;
5493 	int		tcp_ip_hdr_len;
5494 
5495 	tcp_hdr_len = tcp->tcp_hdr_len;
5496 	tcp_ip_hdr_len = tcp->tcp_ip_hdr_len;
5497 
5498 	if (mp) {
5499 		assert(ip_hdr_len != 0);
5500 		rptr = mp->b_rptr;
5501 		tcph = (tcph_t *)(rptr + ip_hdr_len);
5502 		/* Don't reply to a RST segment. */
5503 		if (tcph->th_flags[0] & TH_RST) {
5504 			freeb(mp);
5505 			return;
5506 		}
5507 		freemsg(mp);
5508 		rptr = NULL;
5509 	} else {
5510 		assert(ip_hdr_len == 0);
5511 	}
5512 	/* If a text string is passed in with the request, print it out. */
5513 	if (str != NULL) {
5514 		dprintf("tcp_xmit_ctl(%d): '%s', seq 0x%x, ack 0x%x, "
5515 		    "ctl 0x%x\n", sock_id, str, seq, ack, ctl);
5516 	}
5517 	mp = allocb(tcp_ip_hdr_len + TCP_MAX_HDR_LENGTH + tcp_wroff_xtra, 0);
5518 	if (mp == NULL) {
5519 		dprintf("tcp_xmit_ctl(%d): Cannot allocate memory\n", sock_id);
5520 		return;
5521 	}
5522 	rptr = &mp->b_rptr[tcp_wroff_xtra];
5523 	mp->b_rptr = rptr;
5524 	mp->b_wptr = &rptr[tcp_hdr_len];
5525 	bcopy(tcp->tcp_iphc, rptr, tcp_hdr_len);
5526 
5527 	iph = (struct ip *)rptr;
5528 	iph->ip_len = htons(tcp_hdr_len);
5529 
5530 	tcph = (tcph_t *)&rptr[tcp_ip_hdr_len];
5531 	tcph->th_flags[0] = (uint8_t)ctl;
5532 	if (ctl & TH_RST) {
5533 		BUMP_MIB(tcp_mib.tcpOutRsts);
5534 		BUMP_MIB(tcp_mib.tcpOutControl);
5535 		/*
5536 		 * Don't send TSopt w/ TH_RST packets per RFC 1323.
5537 		 */
5538 		if (tcp->tcp_snd_ts_ok && tcp->tcp_state > TCPS_SYN_SENT) {
5539 			mp->b_wptr = &rptr[tcp_hdr_len - TCPOPT_REAL_TS_LEN];
5540 			*(mp->b_wptr) = TCPOPT_EOL;
5541 			iph->ip_len = htons(tcp_hdr_len -
5542 			    TCPOPT_REAL_TS_LEN);
5543 			tcph->th_offset_and_rsrvd[0] -= (3 << 4);
5544 		}
5545 	}
5546 	if (ctl & TH_ACK) {
5547 		uint32_t now = prom_gettime();
5548 
5549 		if (tcp->tcp_snd_ts_ok) {
5550 			U32_TO_BE32(now,
5551 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+4);
5552 			U32_TO_BE32(tcp->tcp_ts_recent,
5553 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
5554 		}
5555 		tcp->tcp_rack = ack;
5556 		tcp->tcp_rack_cnt = 0;
5557 		BUMP_MIB(tcp_mib.tcpOutAck);
5558 	}
5559 	BUMP_MIB(tcp_mib.tcpOutSegs);
5560 	U32_TO_BE32(seq, tcph->th_seq);
5561 	U32_TO_BE32(ack, tcph->th_ack);
5562 
5563 	tcp_set_cksum(mp);
5564 	iph->ip_ttl = (uint8_t)tcp_ipv4_ttl;
5565 	TCP_DUMP_PACKET("tcp_xmit_ctl", mp);
5566 	(void) ipv4_tcp_output(sock_id, mp);
5567 	freeb(mp);
5568 }
5569 
5570 /* Generate an ACK-only (no data) segment for a TCP endpoint */
5571 static mblk_t *
5572 tcp_ack_mp(tcp_t *tcp)
5573 {
5574 	if (tcp->tcp_valid_bits) {
5575 		/*
5576 		 * For the complex case where we have to send some
5577 		 * controls (FIN or SYN), let tcp_xmit_mp do it.
5578 		 * When sending an ACK-only segment (no data)
5579 		 * into a zero window, always set the seq number to
5580 		 * suna, since snxt will be extended past the window.
5581 		 * If we used snxt, the receiver might consider the ACK
5582 		 * unacceptable.
5583 		 */
5584 		return (tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
5585 		    (tcp->tcp_zero_win_probe) ?
5586 		    tcp->tcp_suna :
5587 		    tcp->tcp_snxt, B_FALSE, NULL, B_FALSE));
5588 	} else {
5589 		/* Generate a simple ACK */
5590 		uchar_t	*rptr;
5591 		tcph_t	*tcph;
5592 		mblk_t	*mp1;
5593 		int32_t	tcp_hdr_len;
5594 		int32_t	num_sack_blk = 0;
5595 		int32_t sack_opt_len;
5596 
5597 		/*
5598 		 * Allocate space for TCP + IP headers
5599 		 * and link-level header
5600 		 */
5601 		if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
5602 			num_sack_blk = MIN(tcp->tcp_max_sack_blk,
5603 			    tcp->tcp_num_sack_blk);
5604 			sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
5605 			    TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
5606 			tcp_hdr_len = tcp->tcp_hdr_len + sack_opt_len;
5607 		} else {
5608 			tcp_hdr_len = tcp->tcp_hdr_len;
5609 		}
5610 		mp1 = allocb(tcp_hdr_len + tcp_wroff_xtra, 0);
5611 		if (mp1 == NULL)
5612 			return (NULL);
5613 
5614 		/* copy in prototype TCP + IP header */
5615 		rptr = mp1->b_rptr + tcp_wroff_xtra;
5616 		mp1->b_rptr = rptr;
5617 		mp1->b_wptr = rptr + tcp_hdr_len;
5618 		bcopy(tcp->tcp_iphc, rptr, tcp->tcp_hdr_len);
5619 
5620 		tcph = (tcph_t *)&rptr[tcp->tcp_ip_hdr_len];
5621 
5622 		/*
5623 		 * Set the TCP sequence number.
5624 		 * When sending an ACK-only segment (no data)
5625 		 * into a zero window, always set the seq number to
5626 		 * suna, since snxt will be extended past the window.
5627 		 * If we used snxt, the receiver might consider the ACK
5628 		 * unacceptable.
5629 		 */
5630 		U32_TO_ABE32((tcp->tcp_zero_win_probe) ?
5631 		    tcp->tcp_suna : tcp->tcp_snxt, tcph->th_seq);
5632 
5633 		/* Set up the TCP flag field. */
5634 		tcph->th_flags[0] = (uchar_t)TH_ACK;
5635 		if (tcp->tcp_ecn_echo_on)
5636 			tcph->th_flags[0] |= TH_ECE;
5637 
5638 		tcp->tcp_rack = tcp->tcp_rnxt;
5639 		tcp->tcp_rack_cnt = 0;
5640 
5641 		/* fill in timestamp option if in use */
5642 		if (tcp->tcp_snd_ts_ok) {
5643 			uint32_t llbolt = (uint32_t)prom_gettime();
5644 
5645 			U32_TO_BE32(llbolt,
5646 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+4);
5647 			U32_TO_BE32(tcp->tcp_ts_recent,
5648 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
5649 		}
5650 
5651 		/* Fill in SACK options */
5652 		if (num_sack_blk > 0) {
5653 			uchar_t *wptr = (uchar_t *)tcph + tcp->tcp_tcp_hdr_len;
5654 			sack_blk_t *tmp;
5655 			int32_t	i;
5656 
5657 			wptr[0] = TCPOPT_NOP;
5658 			wptr[1] = TCPOPT_NOP;
5659 			wptr[2] = TCPOPT_SACK;
5660 			wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
5661 			    sizeof (sack_blk_t);
5662 			wptr += TCPOPT_REAL_SACK_LEN;
5663 
5664 			tmp = tcp->tcp_sack_list;
5665 			for (i = 0; i < num_sack_blk; i++) {
5666 				U32_TO_BE32(tmp[i].begin, wptr);
5667 				wptr += sizeof (tcp_seq);
5668 				U32_TO_BE32(tmp[i].end, wptr);
5669 				wptr += sizeof (tcp_seq);
5670 			}
5671 			tcph->th_offset_and_rsrvd[0] += ((num_sack_blk * 2 + 1)
5672 			    << 4);
5673 		}
5674 
5675 		((struct ip *)rptr)->ip_len = htons(tcp_hdr_len);
5676 		tcp_set_cksum(mp1);
5677 		((struct ip *)rptr)->ip_ttl = (uint8_t)tcp_ipv4_ttl;
5678 		return (mp1);
5679 	}
5680 }
5681 
5682 /*
5683  * tcp_xmit_mp is called to return a pointer to an mblk chain complete with
5684  * ip and tcp header ready to pass down to IP.  If the mp passed in is
5685  * non-NULL, then up to max_to_send bytes of data will be dup'ed off that
5686  * mblk. (If sendall is not set the dup'ing will stop at an mblk boundary
5687  * otherwise it will dup partial mblks.)
5688  * Otherwise, an appropriate ACK packet will be generated.  This
5689  * routine is not usually called to send new data for the first time.  It
5690  * is mostly called out of the timer for retransmits, and to generate ACKs.
5691  *
5692  * If offset is not NULL, the returned mblk chain's first mblk's b_rptr will
5693  * be adjusted by *offset.  And after dupb(), the offset and the ending mblk
5694  * of the original mblk chain will be returned in *offset and *end_mp.
5695  */
5696 static mblk_t *
5697 tcp_xmit_mp(tcp_t *tcp, mblk_t *mp, int32_t max_to_send, int32_t *offset,
5698     mblk_t **end_mp, uint32_t seq, boolean_t sendall, uint32_t *seg_len,
5699     boolean_t rexmit)
5700 {
5701 	int	data_length;
5702 	int32_t	off = 0;
5703 	uint_t	flags;
5704 	mblk_t	*mp1;
5705 	mblk_t	*mp2;
5706 	mblk_t	*new_mp;
5707 	uchar_t	*rptr;
5708 	tcph_t	*tcph;
5709 	int32_t	num_sack_blk = 0;
5710 	int32_t	sack_opt_len = 0;
5711 
5712 	/* Allocate for our maximum TCP header + link-level */
5713 	mp1 = allocb(tcp->tcp_ip_hdr_len + TCP_MAX_HDR_LENGTH +
5714 	    tcp_wroff_xtra, 0);
5715 	if (mp1 == NULL)
5716 		return (NULL);
5717 	data_length = 0;
5718 
5719 	/*
5720 	 * Note that tcp_mss has been adjusted to take into account the
5721 	 * timestamp option if applicable.  Because SACK options do not
5722 	 * appear in every TCP segments and they are of variable lengths,
5723 	 * they cannot be included in tcp_mss.  Thus we need to calculate
5724 	 * the actual segment length when we need to send a segment which
5725 	 * includes SACK options.
5726 	 */
5727 	if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
5728 		num_sack_blk = MIN(tcp->tcp_max_sack_blk,
5729 		    tcp->tcp_num_sack_blk);
5730 		sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
5731 		    TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
5732 		if (max_to_send + sack_opt_len > tcp->tcp_mss)
5733 			max_to_send -= sack_opt_len;
5734 	}
5735 
5736 	if (offset != NULL) {
5737 		off = *offset;
5738 		/* We use offset as an indicator that end_mp is not NULL. */
5739 		*end_mp = NULL;
5740 	}
5741 	for (mp2 = mp1; mp && data_length != max_to_send; mp = mp->b_cont) {
5742 		/* This could be faster with cooperation from downstream */
5743 		if (mp2 != mp1 && !sendall &&
5744 		    data_length + (int)(mp->b_wptr - mp->b_rptr) >
5745 		    max_to_send)
5746 			/*
5747 			 * Don't send the next mblk since the whole mblk
5748 			 * does not fit.
5749 			 */
5750 			break;
5751 		mp2->b_cont = dupb(mp);
5752 		mp2 = mp2->b_cont;
5753 		if (mp2 == NULL) {
5754 			freemsg(mp1);
5755 			return (NULL);
5756 		}
5757 		mp2->b_rptr += off;
5758 		assert((uintptr_t)(mp2->b_wptr - mp2->b_rptr) <=
5759 		    (uintptr_t)INT_MAX);
5760 
5761 		data_length += (int)(mp2->b_wptr - mp2->b_rptr);
5762 		if (data_length > max_to_send) {
5763 			mp2->b_wptr -= data_length - max_to_send;
5764 			data_length = max_to_send;
5765 			off = mp2->b_wptr - mp->b_rptr;
5766 			break;
5767 		} else {
5768 			off = 0;
5769 		}
5770 	}
5771 	if (offset != NULL) {
5772 		*offset = off;
5773 		*end_mp = mp;
5774 	}
5775 	if (seg_len != NULL) {
5776 		*seg_len = data_length;
5777 	}
5778 
5779 	rptr = mp1->b_rptr + tcp_wroff_xtra;
5780 	mp1->b_rptr = rptr;
5781 	mp1->b_wptr = rptr + tcp->tcp_hdr_len + sack_opt_len;
5782 	bcopy(tcp->tcp_iphc, rptr, tcp->tcp_hdr_len);
5783 	tcph = (tcph_t *)&rptr[tcp->tcp_ip_hdr_len];
5784 	U32_TO_ABE32(seq, tcph->th_seq);
5785 
5786 	/*
5787 	 * Use tcp_unsent to determine if the PUSH bit should be used assumes
5788 	 * that this function was called from tcp_wput_data. Thus, when called
5789 	 * to retransmit data the setting of the PUSH bit may appear some
5790 	 * what random in that it might get set when it should not. This
5791 	 * should not pose any performance issues.
5792 	 */
5793 	if (data_length != 0 && (tcp->tcp_unsent == 0 ||
5794 	    tcp->tcp_unsent == data_length)) {
5795 		flags = TH_ACK | TH_PUSH;
5796 	} else {
5797 		flags = TH_ACK;
5798 	}
5799 
5800 	if (tcp->tcp_ecn_ok) {
5801 		if (tcp->tcp_ecn_echo_on)
5802 			flags |= TH_ECE;
5803 
5804 		/*
5805 		 * Only set ECT bit and ECN_CWR if a segment contains new data.
5806 		 * There is no TCP flow control for non-data segments, and
5807 		 * only data segment is transmitted reliably.
5808 		 */
5809 		if (data_length > 0 && !rexmit) {
5810 			SET_ECT(tcp, rptr);
5811 			if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
5812 				flags |= TH_CWR;
5813 				tcp->tcp_ecn_cwr_sent = B_TRUE;
5814 			}
5815 		}
5816 	}
5817 
5818 	if (tcp->tcp_valid_bits) {
5819 		uint32_t u1;
5820 
5821 		if ((tcp->tcp_valid_bits & TCP_ISS_VALID) &&
5822 		    seq == tcp->tcp_iss) {
5823 			uchar_t	*wptr;
5824 
5825 			/*
5826 			 * Tack on the MSS option.  It is always needed
5827 			 * for both active and passive open.
5828 			 */
5829 			wptr = mp1->b_wptr;
5830 			wptr[0] = TCPOPT_MAXSEG;
5831 			wptr[1] = TCPOPT_MAXSEG_LEN;
5832 			wptr += 2;
5833 			/*
5834 			 * MSS option value should be interface MTU - MIN
5835 			 * TCP/IP header.
5836 			 */
5837 			u1 = tcp->tcp_if_mtu - IP_SIMPLE_HDR_LENGTH -
5838 			    TCP_MIN_HEADER_LENGTH;
5839 			U16_TO_BE16(u1, wptr);
5840 			mp1->b_wptr = wptr + 2;
5841 			/* Update the offset to cover the additional word */
5842 			tcph->th_offset_and_rsrvd[0] += (1 << 4);
5843 
5844 			/*
5845 			 * Note that the following way of filling in
5846 			 * TCP options are not optimal.  Some NOPs can
5847 			 * be saved.  But there is no need at this time
5848 			 * to optimize it.  When it is needed, we will
5849 			 * do it.
5850 			 */
5851 			switch (tcp->tcp_state) {
5852 			case TCPS_SYN_SENT:
5853 				flags = TH_SYN;
5854 
5855 				if (tcp->tcp_snd_ws_ok) {
5856 					wptr = mp1->b_wptr;
5857 					wptr[0] =  TCPOPT_NOP;
5858 					wptr[1] =  TCPOPT_WSCALE;
5859 					wptr[2] =  TCPOPT_WS_LEN;
5860 					wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
5861 					mp1->b_wptr += TCPOPT_REAL_WS_LEN;
5862 					tcph->th_offset_and_rsrvd[0] +=
5863 					    (1 << 4);
5864 				}
5865 
5866 				if (tcp->tcp_snd_ts_ok) {
5867 					uint32_t llbolt;
5868 
5869 					llbolt = prom_gettime();
5870 					wptr = mp1->b_wptr;
5871 					wptr[0] = TCPOPT_NOP;
5872 					wptr[1] = TCPOPT_NOP;
5873 					wptr[2] = TCPOPT_TSTAMP;
5874 					wptr[3] = TCPOPT_TSTAMP_LEN;
5875 					wptr += 4;
5876 					U32_TO_BE32(llbolt, wptr);
5877 					wptr += 4;
5878 					assert(tcp->tcp_ts_recent == 0);
5879 					U32_TO_BE32(0L, wptr);
5880 					mp1->b_wptr += TCPOPT_REAL_TS_LEN;
5881 					tcph->th_offset_and_rsrvd[0] +=
5882 					    (3 << 4);
5883 				}
5884 
5885 				if (tcp->tcp_snd_sack_ok) {
5886 					wptr = mp1->b_wptr;
5887 					wptr[0] = TCPOPT_NOP;
5888 					wptr[1] = TCPOPT_NOP;
5889 					wptr[2] = TCPOPT_SACK_PERMITTED;
5890 					wptr[3] = TCPOPT_SACK_OK_LEN;
5891 					mp1->b_wptr += TCPOPT_REAL_SACK_OK_LEN;
5892 					tcph->th_offset_and_rsrvd[0] +=
5893 					    (1 << 4);
5894 				}
5895 
5896 				/*
5897 				 * Set up all the bits to tell other side
5898 				 * we are ECN capable.
5899 				 */
5900 				if (tcp->tcp_ecn_ok) {
5901 					flags |= (TH_ECE | TH_CWR);
5902 				}
5903 				break;
5904 			case TCPS_SYN_RCVD:
5905 				flags |= TH_SYN;
5906 
5907 				if (tcp->tcp_snd_ws_ok) {
5908 				    wptr = mp1->b_wptr;
5909 				    wptr[0] =  TCPOPT_NOP;
5910 				    wptr[1] =  TCPOPT_WSCALE;
5911 				    wptr[2] =  TCPOPT_WS_LEN;
5912 				    wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
5913 				    mp1->b_wptr += TCPOPT_REAL_WS_LEN;
5914 				    tcph->th_offset_and_rsrvd[0] += (1 << 4);
5915 				}
5916 
5917 				if (tcp->tcp_snd_sack_ok) {
5918 					wptr = mp1->b_wptr;
5919 					wptr[0] = TCPOPT_NOP;
5920 					wptr[1] = TCPOPT_NOP;
5921 					wptr[2] = TCPOPT_SACK_PERMITTED;
5922 					wptr[3] = TCPOPT_SACK_OK_LEN;
5923 					mp1->b_wptr += TCPOPT_REAL_SACK_OK_LEN;
5924 					tcph->th_offset_and_rsrvd[0] +=
5925 					    (1 << 4);
5926 				}
5927 
5928 				/*
5929 				 * If the other side is ECN capable, reply
5930 				 * that we are also ECN capable.
5931 				 */
5932 				if (tcp->tcp_ecn_ok) {
5933 					flags |= TH_ECE;
5934 				}
5935 				break;
5936 			default:
5937 				break;
5938 			}
5939 			/* allocb() of adequate mblk assures space */
5940 			assert((uintptr_t)(mp1->b_wptr -
5941 			    mp1->b_rptr) <= (uintptr_t)INT_MAX);
5942 			if (flags & TH_SYN)
5943 				BUMP_MIB(tcp_mib.tcpOutControl);
5944 		}
5945 		if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
5946 		    (seq + data_length) == tcp->tcp_fss) {
5947 			if (!tcp->tcp_fin_acked) {
5948 				flags |= TH_FIN;
5949 				BUMP_MIB(tcp_mib.tcpOutControl);
5950 			}
5951 			if (!tcp->tcp_fin_sent) {
5952 				tcp->tcp_fin_sent = B_TRUE;
5953 				switch (tcp->tcp_state) {
5954 				case TCPS_SYN_RCVD:
5955 				case TCPS_ESTABLISHED:
5956 					tcp->tcp_state = TCPS_FIN_WAIT_1;
5957 					break;
5958 				case TCPS_CLOSE_WAIT:
5959 					tcp->tcp_state = TCPS_LAST_ACK;
5960 					break;
5961 				}
5962 				if (tcp->tcp_suna == tcp->tcp_snxt)
5963 					TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
5964 				tcp->tcp_snxt = tcp->tcp_fss + 1;
5965 			}
5966 		}
5967 	}
5968 	tcph->th_flags[0] = (uchar_t)flags;
5969 	tcp->tcp_rack = tcp->tcp_rnxt;
5970 	tcp->tcp_rack_cnt = 0;
5971 
5972 	if (tcp->tcp_snd_ts_ok) {
5973 		if (tcp->tcp_state != TCPS_SYN_SENT) {
5974 			uint32_t llbolt = prom_gettime();
5975 
5976 			U32_TO_BE32(llbolt,
5977 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+4);
5978 			U32_TO_BE32(tcp->tcp_ts_recent,
5979 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
5980 		}
5981 	}
5982 
5983 	if (num_sack_blk > 0) {
5984 		uchar_t *wptr = (uchar_t *)tcph + tcp->tcp_tcp_hdr_len;
5985 		sack_blk_t *tmp;
5986 		int32_t	i;
5987 
5988 		wptr[0] = TCPOPT_NOP;
5989 		wptr[1] = TCPOPT_NOP;
5990 		wptr[2] = TCPOPT_SACK;
5991 		wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
5992 		    sizeof (sack_blk_t);
5993 		wptr += TCPOPT_REAL_SACK_LEN;
5994 
5995 		tmp = tcp->tcp_sack_list;
5996 		for (i = 0; i < num_sack_blk; i++) {
5997 			U32_TO_BE32(tmp[i].begin, wptr);
5998 			wptr += sizeof (tcp_seq);
5999 			U32_TO_BE32(tmp[i].end, wptr);
6000 			wptr += sizeof (tcp_seq);
6001 		}
6002 		tcph->th_offset_and_rsrvd[0] += ((num_sack_blk * 2 + 1) << 4);
6003 	}
6004 	assert((uintptr_t)(mp1->b_wptr - rptr) <= (uintptr_t)INT_MAX);
6005 	data_length += (int)(mp1->b_wptr - rptr);
6006 	if (tcp->tcp_ipversion == IPV4_VERSION)
6007 		((struct ip *)rptr)->ip_len = htons(data_length);
6008 
6009 	/*
6010 	 * Performance hit!  We need to pullup the whole message
6011 	 * in order to do checksum and for the MAC output routine.
6012 	 */
6013 	if (mp1->b_cont != NULL) {
6014 		int mp_size;
6015 #ifdef DEBUG
6016 		printf("Multiple mblk %d\n", msgdsize(mp1));
6017 #endif
6018 		new_mp = allocb(msgdsize(mp1) + tcp_wroff_xtra, 0);
6019 		new_mp->b_rptr += tcp_wroff_xtra;
6020 		new_mp->b_wptr = new_mp->b_rptr;
6021 		while (mp1 != NULL) {
6022 			mp_size = mp1->b_wptr - mp1->b_rptr;
6023 			bcopy(mp1->b_rptr, new_mp->b_wptr, mp_size);
6024 			new_mp->b_wptr += mp_size;
6025 			mp1 = mp1->b_cont;
6026 		}
6027 		freemsg(mp1);
6028 		mp1 = new_mp;
6029 	}
6030 	tcp_set_cksum(mp1);
6031 	/* Fill in the TTL field as it is 0 in the header template. */
6032 	((struct ip *)mp1->b_rptr)->ip_ttl = (uint8_t)tcp_ipv4_ttl;
6033 
6034 	return (mp1);
6035 }
6036 
6037 /*
6038  * Generate a "no listener here" reset in response to the
6039  * connection request contained within 'mp'
6040  */
6041 static void
6042 tcp_xmit_listeners_reset(int sock_id, mblk_t *mp, uint_t ip_hdr_len)
6043 {
6044 	uchar_t		*rptr;
6045 	uint32_t	seg_len;
6046 	tcph_t		*tcph;
6047 	uint32_t	seg_seq;
6048 	uint32_t	seg_ack;
6049 	uint_t		flags;
6050 
6051 	rptr = mp->b_rptr;
6052 
6053 	tcph = (tcph_t *)&rptr[ip_hdr_len];
6054 	seg_seq = BE32_TO_U32(tcph->th_seq);
6055 	seg_ack = BE32_TO_U32(tcph->th_ack);
6056 	flags = tcph->th_flags[0];
6057 
6058 	seg_len = msgdsize(mp) - (TCP_HDR_LENGTH(tcph) + ip_hdr_len);
6059 	if (flags & TH_RST) {
6060 		freeb(mp);
6061 	} else if (flags & TH_ACK) {
6062 		tcp_xmit_early_reset("no tcp, reset",
6063 		    sock_id, mp, seg_ack, 0, TH_RST, ip_hdr_len);
6064 	} else {
6065 		if (flags & TH_SYN)
6066 			seg_len++;
6067 		tcp_xmit_early_reset("no tcp, reset/ack", sock_id,
6068 		    mp, 0, seg_seq + seg_len,
6069 		    TH_RST | TH_ACK, ip_hdr_len);
6070 	}
6071 }
6072 
6073 /* Non overlapping byte exchanger */
6074 static void
6075 tcp_xchg(uchar_t *a, uchar_t *b, int len)
6076 {
6077 	uchar_t	uch;
6078 
6079 	while (len-- > 0) {
6080 		uch = a[len];
6081 		a[len] = b[len];
6082 		b[len] = uch;
6083 	}
6084 }
6085 
6086 /*
6087  * Generate a reset based on an inbound packet for which there is no active
6088  * tcp state that we can find.
6089  */
6090 static void
6091 tcp_xmit_early_reset(char *str, int sock_id, mblk_t *mp, uint32_t seq,
6092     uint32_t ack, int ctl, uint_t ip_hdr_len)
6093 {
6094 	struct ip	*iph = NULL;
6095 	ushort_t	len;
6096 	tcph_t		*tcph;
6097 	int		i;
6098 	ipaddr_t	addr;
6099 	mblk_t		*new_mp;
6100 
6101 	if (str != NULL) {
6102 		dprintf("tcp_xmit_early_reset: '%s', seq 0x%x, ack 0x%x, "
6103 		    "flags 0x%x\n", str, seq, ack, ctl);
6104 	}
6105 
6106 	/*
6107 	 * We skip reversing source route here.
6108 	 * (for now we replace all IP options with EOL)
6109 	 */
6110 	iph = (struct ip *)mp->b_rptr;
6111 	for (i = IP_SIMPLE_HDR_LENGTH; i < (int)ip_hdr_len; i++)
6112 		mp->b_rptr[i] = IPOPT_EOL;
6113 	/*
6114 	 * Make sure that src address is not a limited broadcast
6115 	 * address. Not all broadcast address checking for the
6116 	 * src address is possible, since we don't know the
6117 	 * netmask of the src addr.
6118 	 * No check for destination address is done, since
6119 	 * IP will not pass up a packet with a broadcast dest address
6120 	 * to TCP.
6121 	 */
6122 	if (iph->ip_src.s_addr == INADDR_ANY ||
6123 	    iph->ip_src.s_addr == INADDR_BROADCAST) {
6124 		freemsg(mp);
6125 		return;
6126 	}
6127 
6128 	tcph = (tcph_t *)&mp->b_rptr[ip_hdr_len];
6129 	if (tcph->th_flags[0] & TH_RST) {
6130 		freemsg(mp);
6131 		return;
6132 	}
6133 	/*
6134 	 * Now copy the original header to a new buffer.  The reason
6135 	 * for doing this is that we need to put extra room before
6136 	 * the header for the MAC layer address.  The original mblk
6137 	 * does not have this extra head room.
6138 	 */
6139 	len = ip_hdr_len + sizeof (tcph_t);
6140 	if ((new_mp = allocb(len + tcp_wroff_xtra, 0)) == NULL) {
6141 		freemsg(mp);
6142 		return;
6143 	}
6144 	new_mp->b_rptr += tcp_wroff_xtra;
6145 	bcopy(mp->b_rptr, new_mp->b_rptr, len);
6146 	new_mp->b_wptr = new_mp->b_rptr + len;
6147 	freemsg(mp);
6148 	mp = new_mp;
6149 	iph = (struct ip *)mp->b_rptr;
6150 	tcph = (tcph_t *)&mp->b_rptr[ip_hdr_len];
6151 
6152 	tcph->th_offset_and_rsrvd[0] = (5 << 4);
6153 	tcp_xchg(tcph->th_fport, tcph->th_lport, 2);
6154 	U32_TO_BE32(ack, tcph->th_ack);
6155 	U32_TO_BE32(seq, tcph->th_seq);
6156 	U16_TO_BE16(0, tcph->th_win);
6157 	bzero(tcph->th_sum, sizeof (int16_t));
6158 	tcph->th_flags[0] = (uint8_t)ctl;
6159 	if (ctl & TH_RST) {
6160 		BUMP_MIB(tcp_mib.tcpOutRsts);
6161 		BUMP_MIB(tcp_mib.tcpOutControl);
6162 	}
6163 
6164 	iph->ip_len = htons(len);
6165 	/* Swap addresses */
6166 	addr = iph->ip_src.s_addr;
6167 	iph->ip_src = iph->ip_dst;
6168 	iph->ip_dst.s_addr = addr;
6169 	iph->ip_id = 0;
6170 	iph->ip_ttl = 0;
6171 	tcp_set_cksum(mp);
6172 	iph->ip_ttl = (uint8_t)tcp_ipv4_ttl;
6173 
6174 	/* Dump the packet when debugging. */
6175 	TCP_DUMP_PACKET("tcp_xmit_early_reset", mp);
6176 	(void) ipv4_tcp_output(sock_id, mp);
6177 	freemsg(mp);
6178 }
6179 
6180 static void
6181 tcp_set_cksum(mblk_t *mp)
6182 {
6183 	struct ip *iph;
6184 	tcpha_t *tcph;
6185 	int len;
6186 
6187 	iph = (struct ip *)mp->b_rptr;
6188 	tcph = (tcpha_t *)(iph + 1);
6189 	len = ntohs(iph->ip_len);
6190 	/*
6191 	 * Calculate the TCP checksum.  Need to include the psuedo header,
6192 	 * which is similar to the real IP header starting at the TTL field.
6193 	 */
6194 	iph->ip_sum = htons(len - IP_SIMPLE_HDR_LENGTH);
6195 	tcph->tha_sum = 0;
6196 	tcph->tha_sum = tcp_cksum((uint16_t *)&(iph->ip_ttl),
6197 	    len - IP_SIMPLE_HDR_LENGTH + 12);
6198 	iph->ip_sum = 0;
6199 }
6200 
6201 static uint16_t
6202 tcp_cksum(uint16_t *buf, uint32_t len)
6203 {
6204 	/*
6205 	 * Compute Internet Checksum for "count" bytes
6206 	 * beginning at location "addr".
6207 	 */
6208 	int32_t sum = 0;
6209 
6210 	while (len > 1) {
6211 		/*  This is the inner loop */
6212 		sum += *buf++;
6213 		len -= 2;
6214 	}
6215 
6216 	/*  Add left-over byte, if any */
6217 	if (len > 0)
6218 		sum += *(unsigned char *)buf * 256;
6219 
6220 	/*  Fold 32-bit sum to 16 bits */
6221 	while (sum >> 16)
6222 		sum = (sum & 0xffff) + (sum >> 16);
6223 
6224 	return ((uint16_t)~sum);
6225 }
6226 
6227 /*
6228  * Type three generator adapted from the random() function in 4.4 BSD:
6229  */
6230 
6231 /*
6232  * Copyright (c) 1983, 1993
6233  *	The Regents of the University of California.  All rights reserved.
6234  *
6235  * Redistribution and use in source and binary forms, with or without
6236  * modification, are permitted provided that the following conditions
6237  * are met:
6238  * 1. Redistributions of source code must retain the above copyright
6239  *    notice, this list of conditions and the following disclaimer.
6240  * 2. Redistributions in binary form must reproduce the above copyright
6241  *    notice, this list of conditions and the following disclaimer in the
6242  *    documentation and/or other materials provided with the distribution.
6243  * 3. All advertising materials mentioning features or use of this software
6244  *    must display the following acknowledgement:
6245  *	This product includes software developed by the University of
6246  *	California, Berkeley and its contributors.
6247  * 4. Neither the name of the University nor the names of its contributors
6248  *    may be used to endorse or promote products derived from this software
6249  *    without specific prior written permission.
6250  *
6251  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
6252  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
6253  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
6254  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
6255  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
6256  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
6257  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
6258  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
6259  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
6260  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
6261  * SUCH DAMAGE.
6262  */
6263 
6264 /* Type 3 -- x**31 + x**3 + 1 */
6265 #define	DEG_3		31
6266 #define	SEP_3		3
6267 
6268 
6269 /* Protected by tcp_random_lock */
6270 static int tcp_randtbl[DEG_3 + 1];
6271 
6272 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
6273 static int *tcp_random_rptr = &tcp_randtbl[1];
6274 
6275 static int *tcp_random_state = &tcp_randtbl[1];
6276 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
6277 
6278 static void
6279 tcp_random_init(void)
6280 {
6281 	int i;
6282 	uint32_t hrt;
6283 	uint32_t wallclock;
6284 	uint32_t result;
6285 
6286 	/*
6287 	 *
6288 	 * XXX We don't have high resolution time in standalone...  The
6289 	 * following is just some approximation on the comment below.
6290 	 *
6291 	 * Use high-res timer and current time for seed.  Gethrtime() returns
6292 	 * a longlong, which may contain resolution down to nanoseconds.
6293 	 * The current time will either be a 32-bit or a 64-bit quantity.
6294 	 * XOR the two together in a 64-bit result variable.
6295 	 * Convert the result to a 32-bit value by multiplying the high-order
6296 	 * 32-bits by the low-order 32-bits.
6297 	 *
6298 	 * XXX We don't have gethrtime() in prom and the wallclock....
6299 	 */
6300 
6301 	hrt = prom_gettime();
6302 	wallclock = (uint32_t)time(NULL);
6303 	result = wallclock ^ hrt;
6304 	tcp_random_state[0] = result;
6305 
6306 	for (i = 1; i < DEG_3; i++)
6307 		tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
6308 			+ 12345;
6309 	tcp_random_fptr = &tcp_random_state[SEP_3];
6310 	tcp_random_rptr = &tcp_random_state[0];
6311 	for (i = 0; i < 10 * DEG_3; i++)
6312 		(void) tcp_random();
6313 }
6314 
6315 /*
6316  * tcp_random: Return a random number in the range [1 - (128K + 1)].
6317  * This range is selected to be approximately centered on TCP_ISS / 2,
6318  * and easy to compute. We get this value by generating a 32-bit random
6319  * number, selecting out the high-order 17 bits, and then adding one so
6320  * that we never return zero.
6321  */
6322 static int
6323 tcp_random(void)
6324 {
6325 	int i;
6326 
6327 	*tcp_random_fptr += *tcp_random_rptr;
6328 
6329 	/*
6330 	 * The high-order bits are more random than the low-order bits,
6331 	 * so we select out the high-order 17 bits and add one so that
6332 	 * we never return zero.
6333 	 */
6334 	i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
6335 	if (++tcp_random_fptr >= tcp_random_end_ptr) {
6336 		tcp_random_fptr = tcp_random_state;
6337 		++tcp_random_rptr;
6338 	} else if (++tcp_random_rptr >= tcp_random_end_ptr)
6339 		tcp_random_rptr = tcp_random_state;
6340 
6341 	return (i);
6342 }
6343 
6344 /*
6345  * Generate ISS, taking into account NDD changes may happen halfway through.
6346  * (If the iss is not zero, set it.)
6347  */
6348 static void
6349 tcp_iss_init(tcp_t *tcp)
6350 {
6351 	tcp_iss_incr_extra += (ISS_INCR >> 1);
6352 	tcp->tcp_iss = tcp_iss_incr_extra;
6353 	tcp->tcp_iss += (prom_gettime() >> ISS_NSEC_SHT) + tcp_random();
6354 	tcp->tcp_valid_bits = TCP_ISS_VALID;
6355 	tcp->tcp_fss = tcp->tcp_iss - 1;
6356 	tcp->tcp_suna = tcp->tcp_iss;
6357 	tcp->tcp_snxt = tcp->tcp_iss + 1;
6358 	tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
6359 	tcp->tcp_csuna = tcp->tcp_snxt;
6360 }
6361 
6362 /*
6363  * Diagnostic routine used to return a string associated with the tcp state.
6364  * Note that if the caller does not supply a buffer, it will use an internal
6365  * static string.  This means that if multiple threads call this function at
6366  * the same time, output can be corrupted...  Note also that this function
6367  * does not check the size of the supplied buffer.  The caller has to make
6368  * sure that it is big enough.
6369  */
6370 static char *
6371 tcp_display(tcp_t *tcp, char *sup_buf, char format)
6372 {
6373 	char		buf1[30];
6374 	static char	priv_buf[INET_ADDRSTRLEN * 2 + 80];
6375 	char		*buf;
6376 	char		*cp;
6377 	char		local_addrbuf[INET_ADDRSTRLEN];
6378 	char		remote_addrbuf[INET_ADDRSTRLEN];
6379 	struct in_addr	addr;
6380 
6381 	if (sup_buf != NULL)
6382 		buf = sup_buf;
6383 	else
6384 		buf = priv_buf;
6385 
6386 	if (tcp == NULL)
6387 		return ("NULL_TCP");
6388 	switch (tcp->tcp_state) {
6389 	case TCPS_CLOSED:
6390 		cp = "TCP_CLOSED";
6391 		break;
6392 	case TCPS_IDLE:
6393 		cp = "TCP_IDLE";
6394 		break;
6395 	case TCPS_BOUND:
6396 		cp = "TCP_BOUND";
6397 		break;
6398 	case TCPS_LISTEN:
6399 		cp = "TCP_LISTEN";
6400 		break;
6401 	case TCPS_SYN_SENT:
6402 		cp = "TCP_SYN_SENT";
6403 		break;
6404 	case TCPS_SYN_RCVD:
6405 		cp = "TCP_SYN_RCVD";
6406 		break;
6407 	case TCPS_ESTABLISHED:
6408 		cp = "TCP_ESTABLISHED";
6409 		break;
6410 	case TCPS_CLOSE_WAIT:
6411 		cp = "TCP_CLOSE_WAIT";
6412 		break;
6413 	case TCPS_FIN_WAIT_1:
6414 		cp = "TCP_FIN_WAIT_1";
6415 		break;
6416 	case TCPS_CLOSING:
6417 		cp = "TCP_CLOSING";
6418 		break;
6419 	case TCPS_LAST_ACK:
6420 		cp = "TCP_LAST_ACK";
6421 		break;
6422 	case TCPS_FIN_WAIT_2:
6423 		cp = "TCP_FIN_WAIT_2";
6424 		break;
6425 	case TCPS_TIME_WAIT:
6426 		cp = "TCP_TIME_WAIT";
6427 		break;
6428 	default:
6429 		(void) sprintf(buf1, "TCPUnkState(%d)", tcp->tcp_state);
6430 		cp = buf1;
6431 		break;
6432 	}
6433 	switch (format) {
6434 	case DISP_ADDR_AND_PORT:
6435 		/*
6436 		 * Note that we use the remote address in the tcp_b
6437 		 * structure.  This means that it will print out
6438 		 * the real destination address, not the next hop's
6439 		 * address if source routing is used.
6440 		 */
6441 		addr.s_addr = tcp->tcp_bound_source;
6442 		bcopy(inet_ntoa(addr), local_addrbuf, sizeof (local_addrbuf));
6443 		addr.s_addr = tcp->tcp_remote;
6444 		bcopy(inet_ntoa(addr), remote_addrbuf, sizeof (remote_addrbuf));
6445 		(void) snprintf(buf, sizeof (priv_buf), "[%s.%u, %s.%u] %s",
6446 		    local_addrbuf, ntohs(tcp->tcp_lport), remote_addrbuf,
6447 		    ntohs(tcp->tcp_fport), cp);
6448 		break;
6449 	case DISP_PORT_ONLY:
6450 	default:
6451 		(void) snprintf(buf, sizeof (priv_buf), "[%u, %u] %s",
6452 		    ntohs(tcp->tcp_lport), ntohs(tcp->tcp_fport), cp);
6453 		break;
6454 	}
6455 
6456 	return (buf);
6457 }
6458 
6459 /*
6460  * Add a new piece to the tcp reassembly queue.  If the gap at the beginning
6461  * is filled, return as much as we can.  The message passed in may be
6462  * multi-part, chained using b_cont.  "start" is the starting sequence
6463  * number for this piece.
6464  */
6465 static mblk_t *
6466 tcp_reass(tcp_t *tcp, mblk_t *mp, uint32_t start)
6467 {
6468 	uint32_t	end;
6469 	mblk_t		*mp1;
6470 	mblk_t		*mp2;
6471 	mblk_t		*next_mp;
6472 	uint32_t	u1;
6473 
6474 	/* Walk through all the new pieces. */
6475 	do {
6476 		assert((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
6477 		    (uintptr_t)INT_MAX);
6478 		end = start + (int)(mp->b_wptr - mp->b_rptr);
6479 		next_mp = mp->b_cont;
6480 		if (start == end) {
6481 			/* Empty.  Blast it. */
6482 			freeb(mp);
6483 			continue;
6484 		}
6485 		mp->b_cont = NULL;
6486 		TCP_REASS_SET_SEQ(mp, start);
6487 		TCP_REASS_SET_END(mp, end);
6488 		mp1 = tcp->tcp_reass_tail;
6489 		if (!mp1) {
6490 			tcp->tcp_reass_tail = mp;
6491 			tcp->tcp_reass_head = mp;
6492 			BUMP_MIB(tcp_mib.tcpInDataUnorderSegs);
6493 			UPDATE_MIB(tcp_mib.tcpInDataUnorderBytes, end - start);
6494 			continue;
6495 		}
6496 		/* New stuff completely beyond tail? */
6497 		if (SEQ_GEQ(start, TCP_REASS_END(mp1))) {
6498 			/* Link it on end. */
6499 			mp1->b_cont = mp;
6500 			tcp->tcp_reass_tail = mp;
6501 			BUMP_MIB(tcp_mib.tcpInDataUnorderSegs);
6502 			UPDATE_MIB(tcp_mib.tcpInDataUnorderBytes, end - start);
6503 			continue;
6504 		}
6505 		mp1 = tcp->tcp_reass_head;
6506 		u1 = TCP_REASS_SEQ(mp1);
6507 		/* New stuff at the front? */
6508 		if (SEQ_LT(start, u1)) {
6509 			/* Yes... Check for overlap. */
6510 			mp->b_cont = mp1;
6511 			tcp->tcp_reass_head = mp;
6512 			tcp_reass_elim_overlap(tcp, mp);
6513 			continue;
6514 		}
6515 		/*
6516 		 * The new piece fits somewhere between the head and tail.
6517 		 * We find our slot, where mp1 precedes us and mp2 trails.
6518 		 */
6519 		for (; (mp2 = mp1->b_cont) != NULL; mp1 = mp2) {
6520 			u1 = TCP_REASS_SEQ(mp2);
6521 			if (SEQ_LEQ(start, u1))
6522 				break;
6523 		}
6524 		/* Link ourselves in */
6525 		mp->b_cont = mp2;
6526 		mp1->b_cont = mp;
6527 
6528 		/* Trim overlap with following mblk(s) first */
6529 		tcp_reass_elim_overlap(tcp, mp);
6530 
6531 		/* Trim overlap with preceding mblk */
6532 		tcp_reass_elim_overlap(tcp, mp1);
6533 
6534 	} while (start = end, mp = next_mp);
6535 	mp1 = tcp->tcp_reass_head;
6536 	/* Anything ready to go? */
6537 	if (TCP_REASS_SEQ(mp1) != tcp->tcp_rnxt)
6538 		return (NULL);
6539 	/* Eat what we can off the queue */
6540 	for (;;) {
6541 		mp = mp1->b_cont;
6542 		end = TCP_REASS_END(mp1);
6543 		TCP_REASS_SET_SEQ(mp1, 0);
6544 		TCP_REASS_SET_END(mp1, 0);
6545 		if (!mp) {
6546 			tcp->tcp_reass_tail = NULL;
6547 			break;
6548 		}
6549 		if (end != TCP_REASS_SEQ(mp)) {
6550 			mp1->b_cont = NULL;
6551 			break;
6552 		}
6553 		mp1 = mp;
6554 	}
6555 	mp1 = tcp->tcp_reass_head;
6556 	tcp->tcp_reass_head = mp;
6557 	return (mp1);
6558 }
6559 
6560 /* Eliminate any overlap that mp may have over later mblks */
6561 static void
6562 tcp_reass_elim_overlap(tcp_t *tcp, mblk_t *mp)
6563 {
6564 	uint32_t	end;
6565 	mblk_t		*mp1;
6566 	uint32_t	u1;
6567 
6568 	end = TCP_REASS_END(mp);
6569 	while ((mp1 = mp->b_cont) != NULL) {
6570 		u1 = TCP_REASS_SEQ(mp1);
6571 		if (!SEQ_GT(end, u1))
6572 			break;
6573 		if (!SEQ_GEQ(end, TCP_REASS_END(mp1))) {
6574 			mp->b_wptr -= end - u1;
6575 			TCP_REASS_SET_END(mp, u1);
6576 			BUMP_MIB(tcp_mib.tcpInDataPartDupSegs);
6577 			UPDATE_MIB(tcp_mib.tcpInDataPartDupBytes, end - u1);
6578 			break;
6579 		}
6580 		mp->b_cont = mp1->b_cont;
6581 		freeb(mp1);
6582 		BUMP_MIB(tcp_mib.tcpInDataDupSegs);
6583 		UPDATE_MIB(tcp_mib.tcpInDataDupBytes, end - u1);
6584 	}
6585 	if (!mp1)
6586 		tcp->tcp_reass_tail = mp;
6587 }
6588 
6589 /*
6590  * Remove a connection from the list of detached TIME_WAIT connections.
6591  */
6592 static void
6593 tcp_time_wait_remove(tcp_t *tcp)
6594 {
6595 	if (tcp->tcp_time_wait_expire == 0) {
6596 		assert(tcp->tcp_time_wait_next == NULL);
6597 		assert(tcp->tcp_time_wait_prev == NULL);
6598 		return;
6599 	}
6600 	assert(tcp->tcp_state == TCPS_TIME_WAIT);
6601 	if (tcp == tcp_time_wait_head) {
6602 		assert(tcp->tcp_time_wait_prev == NULL);
6603 		tcp_time_wait_head = tcp->tcp_time_wait_next;
6604 		if (tcp_time_wait_head != NULL) {
6605 			tcp_time_wait_head->tcp_time_wait_prev = NULL;
6606 		} else {
6607 			tcp_time_wait_tail = NULL;
6608 		}
6609 	} else if (tcp == tcp_time_wait_tail) {
6610 		assert(tcp != tcp_time_wait_head);
6611 		assert(tcp->tcp_time_wait_next == NULL);
6612 		tcp_time_wait_tail = tcp->tcp_time_wait_prev;
6613 		assert(tcp_time_wait_tail != NULL);
6614 		tcp_time_wait_tail->tcp_time_wait_next = NULL;
6615 	} else {
6616 		assert(tcp->tcp_time_wait_prev->tcp_time_wait_next == tcp);
6617 		assert(tcp->tcp_time_wait_next->tcp_time_wait_prev == tcp);
6618 		tcp->tcp_time_wait_prev->tcp_time_wait_next =
6619 		    tcp->tcp_time_wait_next;
6620 		tcp->tcp_time_wait_next->tcp_time_wait_prev =
6621 		    tcp->tcp_time_wait_prev;
6622 	}
6623 	tcp->tcp_time_wait_next = NULL;
6624 	tcp->tcp_time_wait_prev = NULL;
6625 	tcp->tcp_time_wait_expire = 0;
6626 }
6627 
6628 /*
6629  * Add a connection to the list of detached TIME_WAIT connections
6630  * and set its time to expire ...
6631  */
6632 static void
6633 tcp_time_wait_append(tcp_t *tcp)
6634 {
6635 	tcp->tcp_time_wait_expire = prom_gettime() + tcp_time_wait_interval;
6636 	if (tcp->tcp_time_wait_expire == 0)
6637 		tcp->tcp_time_wait_expire = 1;
6638 
6639 	if (tcp_time_wait_head == NULL) {
6640 		assert(tcp_time_wait_tail == NULL);
6641 		tcp_time_wait_head = tcp;
6642 	} else {
6643 		assert(tcp_time_wait_tail != NULL);
6644 		assert(tcp_time_wait_tail->tcp_state == TCPS_TIME_WAIT);
6645 		tcp_time_wait_tail->tcp_time_wait_next = tcp;
6646 		tcp->tcp_time_wait_prev = tcp_time_wait_tail;
6647 	}
6648 	tcp_time_wait_tail = tcp;
6649 
6650 	/* for ndd stats about compression */
6651 	tcp_cum_timewait++;
6652 }
6653 
6654 /*
6655  * Periodic qtimeout routine run on the default queue.
6656  * Performs 2 functions.
6657  * 	1.  Does TIME_WAIT compression on all recently added tcps. List
6658  *	    traversal is done backwards from the tail.
6659  *	2.  Blows away all tcps whose TIME_WAIT has expired. List traversal
6660  *	    is done forwards from the head.
6661  */
6662 void
6663 tcp_time_wait_collector(void)
6664 {
6665 	tcp_t *tcp;
6666 	uint32_t now;
6667 
6668 	/*
6669 	 * In order to reap time waits reliably, we should use a
6670 	 * source of time that is not adjustable by the user
6671 	 */
6672 	now = prom_gettime();
6673 	while ((tcp = tcp_time_wait_head) != NULL) {
6674 		/*
6675 		 * Compare times using modular arithmetic, since
6676 		 * lbolt can wrapover.
6677 		 */
6678 		if ((int32_t)(now - tcp->tcp_time_wait_expire) < 0) {
6679 			break;
6680 		}
6681 		/*
6682 		 * Note that the err must be 0 as there is no socket
6683 		 * associated with this TCP...
6684 		 */
6685 		(void) tcp_clean_death(-1, tcp, 0);
6686 	}
6687 	/* Schedule next run time. */
6688 	tcp_time_wait_runtime = prom_gettime() + 10000;
6689 }
6690 
6691 void
6692 tcp_time_wait_report(void)
6693 {
6694 	tcp_t *tcp;
6695 
6696 	printf("Current time %u\n", prom_gettime());
6697 	for (tcp = tcp_time_wait_head; tcp != NULL;
6698 	    tcp = tcp->tcp_time_wait_next) {
6699 		printf("%s expires at %u\n", tcp_display(tcp, NULL,
6700 		    DISP_ADDR_AND_PORT), tcp->tcp_time_wait_expire);
6701 	}
6702 }
6703 
6704 /*
6705  * Send up all messages queued on tcp_rcv_list.
6706  * Have to set tcp_co_norm since we use putnext.
6707  */
6708 static void
6709 tcp_rcv_drain(int sock_id, tcp_t *tcp)
6710 {
6711 	mblk_t *mp;
6712 	struct inetgram *in_gram;
6713 	mblk_t *in_mp;
6714 	int len;
6715 
6716 	/* Don't drain if the app has not finished reading all the data. */
6717 	if (sockets[sock_id].so_rcvbuf <= 0)
6718 		return;
6719 
6720 	/* We might have come here just to updated the rwnd */
6721 	if (tcp->tcp_rcv_list == NULL)
6722 		goto win_update;
6723 
6724 	if ((in_gram = (struct inetgram *)bkmem_zalloc(
6725 	    sizeof (struct inetgram))) == NULL) {
6726 		return;
6727 	}
6728 	if ((in_mp = allocb(tcp->tcp_rcv_cnt, 0)) == NULL) {
6729 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
6730 		return;
6731 	}
6732 	in_gram->igm_level = APP_LVL;
6733 	in_gram->igm_mp = in_mp;
6734 	in_gram->igm_id = 0;
6735 
6736 	while ((mp = tcp->tcp_rcv_list) != NULL) {
6737 		tcp->tcp_rcv_list = mp->b_cont;
6738 		len = mp->b_wptr - mp->b_rptr;
6739 		bcopy(mp->b_rptr, in_mp->b_wptr, len);
6740 		in_mp->b_wptr += len;
6741 		freeb(mp);
6742 	}
6743 
6744 	tcp->tcp_rcv_last_tail = NULL;
6745 	tcp->tcp_rcv_cnt = 0;
6746 	add_grams(&sockets[sock_id].inq, in_gram);
6747 
6748 	/* This means that so_rcvbuf can be less than 0. */
6749 	sockets[sock_id].so_rcvbuf -= in_mp->b_wptr - in_mp->b_rptr;
6750 win_update:
6751 	/*
6752 	 * Increase the receive window to max.  But we need to do receiver
6753 	 * SWS avoidance.  This means that we need to check the increase of
6754 	 * of receive window is at least 1 MSS.
6755 	 */
6756 	if (sockets[sock_id].so_rcvbuf > 0 &&
6757 	    (tcp->tcp_rwnd_max - tcp->tcp_rwnd >= tcp->tcp_mss)) {
6758 		tcp->tcp_rwnd = tcp->tcp_rwnd_max;
6759 		U32_TO_ABE16(tcp->tcp_rwnd >> tcp->tcp_rcv_ws,
6760 		    tcp->tcp_tcph->th_win);
6761 	}
6762 }
6763 
6764 /*
6765  * Wrapper for recvfrom to call
6766  */
6767 void
6768 tcp_rcv_drain_sock(int sock_id)
6769 {
6770 	tcp_t *tcp;
6771 	if ((tcp = sockets[sock_id].pcb) == NULL)
6772 		return;
6773 	tcp_rcv_drain(sock_id, tcp);
6774 }
6775 
6776 /*
6777  * If the inq == NULL and the tcp_rcv_list != NULL, we have data that
6778  * recvfrom could read. Place a magic message in the inq to let recvfrom
6779  * know that it needs to call tcp_rcv_drain_sock to pullup the data.
6780  */
6781 static void
6782 tcp_drain_needed(int sock_id, tcp_t *tcp)
6783 {
6784 	struct inetgram *in_gram;
6785 #ifdef DEBUG
6786 	printf("tcp_drain_needed: inq %x, tcp_rcv_list %x\n",
6787 		sockets[sock_id].inq, tcp->tcp_rcv_list);
6788 #endif
6789 	if ((sockets[sock_id].inq != NULL) ||
6790 		(tcp->tcp_rcv_list == NULL))
6791 		return;
6792 
6793 	if ((in_gram = (struct inetgram *)bkmem_zalloc(
6794 		sizeof (struct inetgram))) == NULL)
6795 		return;
6796 
6797 	in_gram->igm_level = APP_LVL;
6798 	in_gram->igm_mp = NULL;
6799 	in_gram->igm_id = TCP_CALLB_MAGIC_ID;
6800 
6801 	add_grams(&sockets[sock_id].inq, in_gram);
6802 }
6803 
6804 /*
6805  * Queue data on tcp_rcv_list which is a b_next chain.
6806  * Each element of the chain is a b_cont chain.
6807  *
6808  * M_DATA messages are added to the current element.
6809  * Other messages are added as new (b_next) elements.
6810  */
6811 static void
6812 tcp_rcv_enqueue(tcp_t *tcp, mblk_t *mp, uint_t seg_len)
6813 {
6814 	assert(seg_len == msgdsize(mp));
6815 	if (tcp->tcp_rcv_list == NULL) {
6816 		tcp->tcp_rcv_list = mp;
6817 	} else {
6818 		tcp->tcp_rcv_last_tail->b_cont = mp;
6819 	}
6820 	while (mp->b_cont)
6821 		mp = mp->b_cont;
6822 	tcp->tcp_rcv_last_tail = mp;
6823 	tcp->tcp_rcv_cnt += seg_len;
6824 	tcp->tcp_rwnd -= seg_len;
6825 #ifdef DEBUG
6826 	printf("tcp_rcv_enqueue rwnd %d\n", tcp->tcp_rwnd);
6827 #endif
6828 	U32_TO_ABE16(tcp->tcp_rwnd >> tcp->tcp_rcv_ws, tcp->tcp_tcph->th_win);
6829 }
6830 
6831 /* The minimum of smoothed mean deviation in RTO calculation. */
6832 #define	TCP_SD_MIN	400
6833 
6834 /*
6835  * Set RTO for this connection.  The formula is from Jacobson and Karels'
6836  * "Congestion Avoidance and Control" in SIGCOMM '88.  The variable names
6837  * are the same as those in Appendix A.2 of that paper.
6838  *
6839  * m = new measurement
6840  * sa = smoothed RTT average (8 * average estimates).
6841  * sv = smoothed mean deviation (mdev) of RTT (4 * deviation estimates).
6842  */
6843 static void
6844 tcp_set_rto(tcp_t *tcp, int32_t rtt)
6845 {
6846 	int32_t m = rtt;
6847 	uint32_t sa = tcp->tcp_rtt_sa;
6848 	uint32_t sv = tcp->tcp_rtt_sd;
6849 	uint32_t rto;
6850 
6851 	BUMP_MIB(tcp_mib.tcpRttUpdate);
6852 	tcp->tcp_rtt_update++;
6853 
6854 	/* tcp_rtt_sa is not 0 means this is a new sample. */
6855 	if (sa != 0) {
6856 		/*
6857 		 * Update average estimator:
6858 		 *	new rtt = 7/8 old rtt + 1/8 Error
6859 		 */
6860 
6861 		/* m is now Error in estimate. */
6862 		m -= sa >> 3;
6863 		if ((int32_t)(sa += m) <= 0) {
6864 			/*
6865 			 * Don't allow the smoothed average to be negative.
6866 			 * We use 0 to denote reinitialization of the
6867 			 * variables.
6868 			 */
6869 			sa = 1;
6870 		}
6871 
6872 		/*
6873 		 * Update deviation estimator:
6874 		 *	new mdev = 3/4 old mdev + 1/4 (abs(Error) - old mdev)
6875 		 */
6876 		if (m < 0)
6877 			m = -m;
6878 		m -= sv >> 2;
6879 		sv += m;
6880 	} else {
6881 		/*
6882 		 * This follows BSD's implementation.  So the reinitialized
6883 		 * RTO is 3 * m.  We cannot go less than 2 because if the
6884 		 * link is bandwidth dominated, doubling the window size
6885 		 * during slow start means doubling the RTT.  We want to be
6886 		 * more conservative when we reinitialize our estimates.  3
6887 		 * is just a convenient number.
6888 		 */
6889 		sa = m << 3;
6890 		sv = m << 1;
6891 	}
6892 	if (sv < TCP_SD_MIN) {
6893 		/*
6894 		 * We do not know that if sa captures the delay ACK
6895 		 * effect as in a long train of segments, a receiver
6896 		 * does not delay its ACKs.  So set the minimum of sv
6897 		 * to be TCP_SD_MIN, which is default to 400 ms, twice
6898 		 * of BSD DATO.  That means the minimum of mean
6899 		 * deviation is 100 ms.
6900 		 *
6901 		 */
6902 		sv = TCP_SD_MIN;
6903 	}
6904 	tcp->tcp_rtt_sa = sa;
6905 	tcp->tcp_rtt_sd = sv;
6906 	/*
6907 	 * RTO = average estimates (sa / 8) + 4 * deviation estimates (sv)
6908 	 *
6909 	 * Add tcp_rexmit_interval extra in case of extreme environment
6910 	 * where the algorithm fails to work.  The default value of
6911 	 * tcp_rexmit_interval_extra should be 0.
6912 	 *
6913 	 * As we use a finer grained clock than BSD and update
6914 	 * RTO for every ACKs, add in another .25 of RTT to the
6915 	 * deviation of RTO to accomodate burstiness of 1/4 of
6916 	 * window size.
6917 	 */
6918 	rto = (sa >> 3) + sv + tcp_rexmit_interval_extra + (sa >> 5);
6919 
6920 	if (rto > tcp_rexmit_interval_max) {
6921 		tcp->tcp_rto = tcp_rexmit_interval_max;
6922 	} else if (rto < tcp_rexmit_interval_min) {
6923 		tcp->tcp_rto = tcp_rexmit_interval_min;
6924 	} else {
6925 		tcp->tcp_rto = rto;
6926 	}
6927 
6928 	/* Now, we can reset tcp_timer_backoff to use the new RTO... */
6929 	tcp->tcp_timer_backoff = 0;
6930 }
6931 
6932 /*
6933  * Initiate closedown sequence on an active connection.
6934  * Return value zero for OK return, non-zero for error return.
6935  */
6936 static int
6937 tcp_xmit_end(tcp_t *tcp, int sock_id)
6938 {
6939 	mblk_t	*mp;
6940 
6941 	if (tcp->tcp_state < TCPS_SYN_RCVD ||
6942 	    tcp->tcp_state > TCPS_CLOSE_WAIT) {
6943 		/*
6944 		 * Invalid state, only states TCPS_SYN_RCVD,
6945 		 * TCPS_ESTABLISHED and TCPS_CLOSE_WAIT are valid
6946 		 */
6947 		return (-1);
6948 	}
6949 
6950 	tcp->tcp_fss = tcp->tcp_snxt + tcp->tcp_unsent;
6951 	tcp->tcp_valid_bits |= TCP_FSS_VALID;
6952 	/*
6953 	 * If there is nothing more unsent, send the FIN now.
6954 	 * Otherwise, it will go out with the last segment.
6955 	 */
6956 	if (tcp->tcp_unsent == 0) {
6957 		mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
6958 		    tcp->tcp_fss, B_FALSE, NULL, B_FALSE);
6959 
6960 		if (mp != NULL) {
6961 			/* Dump the packet when debugging. */
6962 			TCP_DUMP_PACKET("tcp_xmit_end", mp);
6963 			(void) ipv4_tcp_output(sock_id, mp);
6964 			freeb(mp);
6965 		} else {
6966 			/*
6967 			 * Couldn't allocate msg.  Pretend we got it out.
6968 			 * Wait for rexmit timeout.
6969 			 */
6970 			tcp->tcp_snxt = tcp->tcp_fss + 1;
6971 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
6972 		}
6973 
6974 		/*
6975 		 * If needed, update tcp_rexmit_snxt as tcp_snxt is
6976 		 * changed.
6977 		 */
6978 		if (tcp->tcp_rexmit && tcp->tcp_rexmit_nxt == tcp->tcp_fss) {
6979 			tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
6980 		}
6981 	} else {
6982 		tcp_wput_data(tcp, NULL, B_FALSE);
6983 	}
6984 
6985 	return (0);
6986 }
6987 
6988 int
6989 tcp_opt_set(tcp_t *tcp, int level, int option, const void *optval,
6990     socklen_t optlen)
6991 {
6992 	switch (level) {
6993 	case SOL_SOCKET: {
6994 		switch (option) {
6995 		case SO_RCVBUF:
6996 			if (optlen == sizeof (int)) {
6997 				int val = *(int *)optval;
6998 
6999 				if (val > tcp_max_buf) {
7000 					errno = ENOBUFS;
7001 					break;
7002 				}
7003 				/* Silently ignore zero */
7004 				if (val != 0) {
7005 					val = MSS_ROUNDUP(val, tcp->tcp_mss);
7006 					(void) tcp_rwnd_set(tcp, val);
7007 				}
7008 			} else {
7009 				errno = EINVAL;
7010 			}
7011 			break;
7012 		case SO_SNDBUF:
7013 			if (optlen == sizeof (int)) {
7014 				tcp->tcp_xmit_hiwater = *(int *)optval;
7015 				if (tcp->tcp_xmit_hiwater > tcp_max_buf)
7016 					tcp->tcp_xmit_hiwater = tcp_max_buf;
7017 			} else {
7018 				errno = EINVAL;
7019 			}
7020 			break;
7021 		case SO_LINGER:
7022 			if (optlen == sizeof (struct linger)) {
7023 				struct linger *lgr = (struct linger *)optval;
7024 
7025 				if (lgr->l_onoff) {
7026 					tcp->tcp_linger = 1;
7027 					tcp->tcp_lingertime = lgr->l_linger;
7028 				} else {
7029 					tcp->tcp_linger = 0;
7030 					tcp->tcp_lingertime = 0;
7031 				}
7032 			} else {
7033 				errno = EINVAL;
7034 			}
7035 			break;
7036 		default:
7037 			errno = ENOPROTOOPT;
7038 			break;
7039 		}
7040 		break;
7041 	} /* case SOL_SOCKET */
7042 	case IPPROTO_TCP: {
7043 		switch (option) {
7044 		default:
7045 			errno = ENOPROTOOPT;
7046 			break;
7047 		}
7048 		break;
7049 	} /* case IPPROTO_TCP */
7050 	case IPPROTO_IP: {
7051 		switch (option) {
7052 		default:
7053 			errno = ENOPROTOOPT;
7054 			break;
7055 		}
7056 		break;
7057 	} /* case IPPROTO_IP */
7058 	default:
7059 		errno = ENOPROTOOPT;
7060 		break;
7061 	} /* switch (level) */
7062 
7063 	if (errno != 0)
7064 		return (-1);
7065 	else
7066 		return (0);
7067 }
7068