xref: /illumos-gate/usr/src/stand/lib/tcp/tcp.c (revision 9b9d39d2a32ff806d2431dbcc50968ef1e6d46b2)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  * Copyright (c) 2016 by Delphix. All rights reserved.
26  *
27  * tcp.c, Code implementing the TCP protocol.
28  */
29 
30 #include <sys/types.h>
31 #include <socket_impl.h>
32 #include <socket_inet.h>
33 #include <sys/sysmacros.h>
34 #include <sys/promif.h>
35 #include <sys/socket.h>
36 #include <netinet/in_systm.h>
37 #include <netinet/in.h>
38 #include <netinet/ip.h>
39 #include <netinet/tcp.h>
40 #include <net/if_types.h>
41 #include <sys/salib.h>
42 
43 #include "ipv4.h"
44 #include "ipv4_impl.h"
45 #include "mac.h"
46 #include "mac_impl.h"
47 #include "v4_sum_impl.h"
48 #include <sys/bootdebug.h>
49 #include "tcp_inet.h"
50 #include "tcp_sack.h"
51 #include <inet/common.h>
52 #include <inet/mib2.h>
53 
54 /*
55  * We need to redefine BUMP_MIB/UPDATE_MIB to not have DTrace probes.
56  */
57 #undef BUMP_MIB
58 #define	BUMP_MIB(x) (x)++
59 
60 #undef UPDATE_MIB
61 #define	UPDATE_MIB(x, y) x += y
62 
63 /*
64  * MIB-2 stuff for SNMP
65  */
66 mib2_tcp_t	tcp_mib;	/* SNMP fixed size info */
67 
68 /* The TCP mib does not include the following errors. */
69 static uint_t tcp_cksum_errors;
70 static uint_t tcp_drops;
71 
72 /* Macros for timestamp comparisons */
73 #define	TSTMP_GEQ(a, b)	((int32_t)((a)-(b)) >= 0)
74 #define	TSTMP_LT(a, b)	((int32_t)((a)-(b)) < 0)
75 
76 /*
77  * Parameters for TCP Initial Send Sequence number (ISS) generation.
78  * The ISS is calculated by adding three components: a time component
79  * which grows by 1 every 4096 nanoseconds (versus every 4 microseconds
80  * suggested by RFC 793, page 27);
81  * a per-connection component which grows by 125000 for every new connection;
82  * and an "extra" component that grows by a random amount centered
83  * approximately on 64000.  This causes the the ISS generator to cycle every
84  * 4.89 hours if no TCP connections are made, and faster if connections are
85  * made.
86  */
87 #define	ISS_INCR	250000
88 #define	ISS_NSEC_SHT	0
89 
90 static uint32_t tcp_iss_incr_extra;	/* Incremented for each connection */
91 
92 #define	TCP_XMIT_LOWATER	4096
93 #define	TCP_XMIT_HIWATER	49152
94 #define	TCP_RECV_LOWATER	2048
95 #define	TCP_RECV_HIWATER	49152
96 
97 /*
98  *  PAWS needs a timer for 24 days.  This is the number of ms in 24 days
99  */
100 #define	PAWS_TIMEOUT	((uint32_t)(24*24*60*60*1000))
101 
102 /*
103  * TCP options struct returned from tcp_parse_options.
104  */
105 typedef struct tcp_opt_s {
106 	uint32_t	tcp_opt_mss;
107 	uint32_t	tcp_opt_wscale;
108 	uint32_t	tcp_opt_ts_val;
109 	uint32_t	tcp_opt_ts_ecr;
110 	tcp_t		*tcp;
111 } tcp_opt_t;
112 
113 /*
114  * RFC1323-recommended phrasing of TSTAMP option, for easier parsing
115  */
116 
117 #ifdef _BIG_ENDIAN
118 #define	TCPOPT_NOP_NOP_TSTAMP ((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
119 	(TCPOPT_TSTAMP << 8) | 10)
120 #else
121 #define	TCPOPT_NOP_NOP_TSTAMP ((10 << 24) | (TCPOPT_TSTAMP << 16) | \
122 	(TCPOPT_NOP << 8) | TCPOPT_NOP)
123 #endif
124 
125 /*
126  * Flags returned from tcp_parse_options.
127  */
128 #define	TCP_OPT_MSS_PRESENT	1
129 #define	TCP_OPT_WSCALE_PRESENT	2
130 #define	TCP_OPT_TSTAMP_PRESENT	4
131 #define	TCP_OPT_SACK_OK_PRESENT	8
132 #define	TCP_OPT_SACK_PRESENT	16
133 
134 /* TCP option length */
135 #define	TCPOPT_NOP_LEN		1
136 #define	TCPOPT_MAXSEG_LEN	4
137 #define	TCPOPT_WS_LEN		3
138 #define	TCPOPT_REAL_WS_LEN	(TCPOPT_WS_LEN+1)
139 #define	TCPOPT_TSTAMP_LEN	10
140 #define	TCPOPT_REAL_TS_LEN	(TCPOPT_TSTAMP_LEN+2)
141 #define	TCPOPT_SACK_OK_LEN	2
142 #define	TCPOPT_REAL_SACK_OK_LEN	(TCPOPT_SACK_OK_LEN+2)
143 #define	TCPOPT_REAL_SACK_LEN	4
144 #define	TCPOPT_MAX_SACK_LEN	36
145 #define	TCPOPT_HEADER_LEN	2
146 
147 /* TCP cwnd burst factor. */
148 #define	TCP_CWND_INFINITE	65535
149 #define	TCP_CWND_SS		3
150 #define	TCP_CWND_NORMAL		5
151 
152 /* Named Dispatch Parameter Management Structure */
153 typedef struct tcpparam_s {
154 	uint32_t	tcp_param_min;
155 	uint32_t	tcp_param_max;
156 	uint32_t	tcp_param_val;
157 	char		*tcp_param_name;
158 } tcpparam_t;
159 
160 /* Max size IP datagram is 64k - 1 */
161 #define	TCP_MSS_MAX_IPV4 (IP_MAXPACKET - (sizeof (struct ip) + \
162 	sizeof (tcph_t)))
163 
164 /* Max of the above */
165 #define	TCP_MSS_MAX	TCP_MSS_MAX_IPV4
166 
167 /* Largest TCP port number */
168 #define	TCP_MAX_PORT	(64 * 1024 - 1)
169 
170 /* Round up the value to the nearest mss. */
171 #define	MSS_ROUNDUP(value, mss)		((((value) - 1) / (mss) + 1) * (mss))
172 
173 #define	MS	1L
174 #define	SECONDS	(1000 * MS)
175 #define	MINUTES	(60 * SECONDS)
176 #define	HOURS	(60 * MINUTES)
177 #define	DAYS	(24 * HOURS)
178 
179 /* All NDD params in the core TCP became static variables. */
180 static int	tcp_time_wait_interval = 1 * MINUTES;
181 static int	tcp_conn_req_max_q = 128;
182 static int	tcp_conn_req_max_q0 = 1024;
183 static int	tcp_conn_req_min = 1;
184 static int	tcp_conn_grace_period = 0 * SECONDS;
185 static int	tcp_cwnd_max_ = 1024 * 1024;
186 static int	tcp_smallest_nonpriv_port = 1024;
187 static int	tcp_ip_abort_cinterval = 3 * MINUTES;
188 static int	tcp_ip_abort_linterval = 3 * MINUTES;
189 static int	tcp_ip_abort_interval = 8 * MINUTES;
190 static int	tcp_ip_notify_cinterval = 10 * SECONDS;
191 static int	tcp_ip_notify_interval = 10 * SECONDS;
192 static int	tcp_ipv4_ttl = 64;
193 static int	tcp_mss_def_ipv4 = 536;
194 static int	tcp_mss_max_ipv4 = TCP_MSS_MAX_IPV4;
195 static int	tcp_mss_min = 108;
196 static int	tcp_naglim_def = (4*1024)-1;
197 static int	tcp_rexmit_interval_initial = 3 * SECONDS;
198 static int	tcp_rexmit_interval_max = 60 * SECONDS;
199 static int	tcp_rexmit_interval_min = 400 * MS;
200 static int	tcp_dupack_fast_retransmit = 3;
201 static int	tcp_smallest_anon_port = 32 * 1024;
202 static int	tcp_largest_anon_port = TCP_MAX_PORT;
203 static int	tcp_xmit_lowat = TCP_XMIT_LOWATER;
204 static int	tcp_recv_hiwat_minmss = 4;
205 static int	tcp_fin_wait_2_flush_interval = 1 * MINUTES;
206 static int	tcp_max_buf = 1024 * 1024;
207 static int	tcp_wscale_always = 1;
208 static int	tcp_tstamp_always = 1;
209 static int	tcp_tstamp_if_wscale = 1;
210 static int	tcp_rexmit_interval_extra = 0;
211 static int	tcp_slow_start_after_idle = 2;
212 static int	tcp_slow_start_initial = 2;
213 static int	tcp_sack_permitted = 2;
214 static int	tcp_ecn_permitted = 2;
215 
216 /* Extra room to fit in headers. */
217 static uint_t	tcp_wroff_xtra;
218 
219 /* Hint for next port to try. */
220 static in_port_t	tcp_next_port_to_try = 32*1024;
221 
222 /*
223  * Figure out the value of window scale opton.  Note that the rwnd is
224  * ASSUMED to be rounded up to the nearest MSS before the calculation.
225  * We cannot find the scale value and then do a round up of tcp_rwnd
226  * because the scale value may not be correct after that.
227  */
228 #define	SET_WS_VALUE(tcp) \
229 { \
230 	int i; \
231 	uint32_t rwnd = (tcp)->tcp_rwnd; \
232 	for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT; \
233 	    i++, rwnd >>= 1) \
234 		; \
235 	(tcp)->tcp_rcv_ws = i; \
236 }
237 
238 /*
239  * Set ECN capable transport (ECT) code point in IP header.
240  *
241  * Note that there are 2 ECT code points '01' and '10', which are called
242  * ECT(1) and ECT(0) respectively.  Here we follow the original ECT code
243  * point ECT(0) for TCP as described in RFC 2481.
244  */
245 #define	SET_ECT(tcp, iph) \
246 	if ((tcp)->tcp_ipversion == IPV4_VERSION) { \
247 		/* We need to clear the code point first. */ \
248 		((struct ip *)(iph))->ip_tos &= 0xFC; \
249 		((struct ip *)(iph))->ip_tos |= IPH_ECN_ECT0; \
250 	}
251 
252 /*
253  * The format argument to pass to tcp_display().
254  * DISP_PORT_ONLY means that the returned string has only port info.
255  * DISP_ADDR_AND_PORT means that the returned string also contains the
256  * remote and local IP address.
257  */
258 #define	DISP_PORT_ONLY		1
259 #define	DISP_ADDR_AND_PORT	2
260 
261 /*
262  * TCP reassembly macros.  We hide starting and ending sequence numbers in
263  * b_next and b_prev of messages on the reassembly queue.  The messages are
264  * chained using b_cont.  These macros are used in tcp_reass() so we don't
265  * have to see the ugly casts and assignments.
266  * Note. use uintptr_t to suppress the gcc warning.
267  */
268 #define	TCP_REASS_SEQ(mp)		((uint32_t)(uintptr_t)((mp)->b_next))
269 #define	TCP_REASS_SET_SEQ(mp, u)	((mp)->b_next = \
270 					    (mblk_t *)((uintptr_t)(u)))
271 #define	TCP_REASS_END(mp)		((uint32_t)(uintptr_t)((mp)->b_prev))
272 #define	TCP_REASS_SET_END(mp, u)	((mp)->b_prev = \
273 					    (mblk_t *)((uintptr_t)(u)))
274 
275 #define	TCP_TIMER_RESTART(tcp, intvl) \
276 	(tcp)->tcp_rto_timeout = prom_gettime() + intvl; \
277 	(tcp)->tcp_timer_running = B_TRUE;
278 
279 static int tcp_accept_comm(tcp_t *, tcp_t *, mblk_t *, uint_t);
280 static mblk_t *tcp_ack_mp(tcp_t *);
281 static in_port_t tcp_bindi(in_port_t, in_addr_t *, boolean_t, boolean_t);
282 static uint16_t tcp_cksum(uint16_t *, uint32_t);
283 static void tcp_clean_death(int, tcp_t *, int err);
284 static tcp_t *tcp_conn_request(tcp_t *, mblk_t *mp, uint_t, uint_t);
285 static char *tcp_display(tcp_t *, char *, char);
286 static int tcp_drain_input(tcp_t *, int, int);
287 static void tcp_drain_needed(int, tcp_t *);
288 static boolean_t tcp_drop_q0(tcp_t *);
289 static mblk_t *tcp_get_seg_mp(tcp_t *, uint32_t, int32_t *);
290 static int tcp_header_len(struct inetgram *);
291 static in_port_t tcp_report_ports(uint16_t *, enum Ports);
292 static int tcp_input(int);
293 static void tcp_iss_init(tcp_t *);
294 static tcp_t *tcp_lookup_ipv4(struct ip *, tcpha_t *, int, int *);
295 static tcp_t *tcp_lookup_listener_ipv4(in_addr_t, in_port_t, int *);
296 static int tcp_conn_check(tcp_t *);
297 static int tcp_close(int);
298 static void tcp_close_detached(tcp_t *);
299 static void tcp_eager_cleanup(tcp_t *, boolean_t, int);
300 static void tcp_eager_unlink(tcp_t *);
301 static void tcp_free(tcp_t *);
302 static int tcp_header_init_ipv4(tcp_t *);
303 static void tcp_mss_set(tcp_t *, uint32_t);
304 static int tcp_parse_options(tcph_t *, tcp_opt_t *);
305 static boolean_t tcp_paws_check(tcp_t *, tcph_t *, tcp_opt_t *);
306 static void tcp_process_options(tcp_t *, tcph_t *);
307 static int tcp_random(void);
308 static void tcp_random_init(void);
309 static mblk_t *tcp_reass(tcp_t *, mblk_t *, uint32_t);
310 static void tcp_reass_elim_overlap(tcp_t *, mblk_t *);
311 static void tcp_rcv_drain(int sock_id, tcp_t *);
312 static void tcp_rcv_enqueue(tcp_t *, mblk_t *, uint_t);
313 static void tcp_rput_data(tcp_t *, mblk_t *, int);
314 static int tcp_rwnd_set(tcp_t *, uint32_t);
315 static int32_t tcp_sack_rxmit(tcp_t *, int);
316 static void tcp_set_cksum(mblk_t *);
317 static void tcp_set_rto(tcp_t *, int32_t);
318 static void tcp_ss_rexmit(tcp_t *, int);
319 static int tcp_state_wait(int, tcp_t *, int);
320 static void tcp_timer(tcp_t *, int);
321 static void tcp_time_wait_append(tcp_t *);
322 static void tcp_time_wait_collector(void);
323 static void tcp_time_wait_processing(tcp_t *, mblk_t *, uint32_t,
324     uint32_t, int, tcph_t *, int sock_id);
325 static void tcp_time_wait_remove(tcp_t *);
326 static in_port_t tcp_update_next_port(in_port_t);
327 static int tcp_verify_cksum(mblk_t *);
328 static void tcp_wput_data(tcp_t *, mblk_t *, int);
329 static void tcp_xmit_ctl(char *, tcp_t *, mblk_t *, uint32_t, uint32_t,
330     int, uint_t, int);
331 static void tcp_xmit_early_reset(char *, int, mblk_t *, uint32_t, uint32_t,
332     int, uint_t);
333 static int tcp_xmit_end(tcp_t *, int);
334 static void tcp_xmit_listeners_reset(int, mblk_t *, uint_t);
335 static mblk_t *tcp_xmit_mp(tcp_t *, mblk_t *, int32_t, int32_t *,
336     mblk_t **, uint32_t, boolean_t, uint32_t *, boolean_t);
337 static int tcp_init_values(tcp_t *, struct inetboot_socket *);
338 
339 #if DEBUG > 1
340 #define	TCP_DUMP_PACKET(str, mp) \
341 { \
342 	int len = (mp)->b_wptr - (mp)->b_rptr; \
343 \
344 	printf("%s: dump TCP(%d): \n", (str), len); \
345 	hexdump((char *)(mp)->b_rptr, len); \
346 }
347 #else
348 #define	TCP_DUMP_PACKET(str, mp)
349 #endif
350 
351 #ifdef DEBUG
352 #define	DEBUG_1(str, arg)		printf(str, (arg))
353 #define	DEBUG_2(str, arg1, arg2)	printf(str, (arg1), (arg2))
354 #define	DEBUG_3(str, arg1, arg2, arg3)	printf(str, (arg1), (arg2), (arg3))
355 #else
356 #define	DEBUG_1(str, arg)
357 #define	DEBUG_2(str, arg1, arg2)
358 #define	DEBUG_3(str, arg1, arg2, arg3)
359 #endif
360 
361 /* Whether it is the first time TCP is used. */
362 static boolean_t tcp_initialized = B_FALSE;
363 
364 /* TCP time wait list. */
365 static tcp_t *tcp_time_wait_head;
366 static tcp_t *tcp_time_wait_tail;
367 static uint32_t tcp_cum_timewait;
368 /* When the tcp_time_wait_collector is run. */
369 static uint32_t tcp_time_wait_runtime;
370 
371 #define	TCP_RUN_TIME_WAIT_COLLECTOR() \
372 	if (prom_gettime() > tcp_time_wait_runtime) \
373 		tcp_time_wait_collector();
374 
375 /*
376  * Accept will return with an error if there is no connection coming in
377  * after this (in ms).
378  */
379 static int tcp_accept_timeout = 60000;
380 
381 /*
382  * Initialize the TCP-specific parts of a socket.
383  */
384 void
385 tcp_socket_init(struct inetboot_socket *isp)
386 {
387 	/* Do some initializations. */
388 	if (!tcp_initialized) {
389 		tcp_random_init();
390 		/* Extra head room for the MAC layer address. */
391 		if ((tcp_wroff_xtra = mac_get_hdr_len()) & 0x3) {
392 			tcp_wroff_xtra = (tcp_wroff_xtra & ~0x3) + 0x4;
393 		}
394 		/* Schedule the first time wait cleanup time */
395 		tcp_time_wait_runtime = prom_gettime() + tcp_time_wait_interval;
396 		tcp_initialized = B_TRUE;
397 	}
398 	TCP_RUN_TIME_WAIT_COLLECTOR();
399 
400 	isp->proto = IPPROTO_TCP;
401 	isp->input[TRANSPORT_LVL] = tcp_input;
402 	/* Socket layer should call tcp_send() directly. */
403 	isp->output[TRANSPORT_LVL] = NULL;
404 	isp->close[TRANSPORT_LVL] = tcp_close;
405 	isp->headerlen[TRANSPORT_LVL] = tcp_header_len;
406 	isp->ports = tcp_report_ports;
407 	if ((isp->pcb = bkmem_alloc(sizeof (tcp_t))) == NULL) {
408 		errno = ENOBUFS;
409 		return;
410 	}
411 	if ((errno = tcp_init_values((tcp_t *)isp->pcb, isp)) != 0) {
412 		bkmem_free(isp->pcb, sizeof (tcp_t));
413 		return;
414 	}
415 	/*
416 	 * This is set last because this field is used to determine if
417 	 * a socket is in use or not.
418 	 */
419 	isp->type = INETBOOT_STREAM;
420 }
421 
422 /*
423  * Return the size of a TCP header including TCP option.
424  */
425 static int
426 tcp_header_len(struct inetgram *igm)
427 {
428 	mblk_t *pkt;
429 	int ipvers;
430 
431 	/* Just returns the standard TCP header without option */
432 	if (igm == NULL)
433 		return (sizeof (tcph_t));
434 
435 	if ((pkt = igm->igm_mp) == NULL)
436 		return (0);
437 
438 	ipvers = ((struct ip *)pkt->b_rptr)->ip_v;
439 	if (ipvers == IPV4_VERSION) {
440 		return (TCP_HDR_LENGTH((tcph_t *)(pkt + IPH_HDR_LENGTH(pkt))));
441 	} else {
442 		dprintf("tcp_header_len: non-IPv4 packet.\n");
443 		return (0);
444 	}
445 }
446 
447 /*
448  * Return the requested port number in network order.
449  */
450 static in_port_t
451 tcp_report_ports(uint16_t *tcphp, enum Ports request)
452 {
453 	if (request == SOURCE)
454 		return (*(uint16_t *)(((tcph_t *)tcphp)->th_lport));
455 	return (*(uint16_t *)(((tcph_t *)tcphp)->th_fport));
456 }
457 
458 /*
459  * Because inetboot is not interrupt driven, TCP can only poll.  This
460  * means that there can be packets stuck in the NIC buffer waiting to
461  * be processed.  Thus we need to drain them before, for example, sending
462  * anything because an ACK may actually be stuck there.
463  *
464  * The timeout arguments determine how long we should wait for draining.
465  */
466 static int
467 tcp_drain_input(tcp_t *tcp, int sock_id, int timeout)
468 {
469 	struct inetgram *in_gram;
470 	struct inetgram *old_in_gram;
471 	int old_timeout;
472 	mblk_t *mp;
473 	int i;
474 
475 	dprintf("tcp_drain_input(%d): %s\n", sock_id,
476 	    tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
477 
478 	/*
479 	 * Since the driver uses the in_timeout value in the socket
480 	 * structure to determine the timeout value, we need to save
481 	 * the original one so that we can restore that after draining.
482 	 */
483 	old_timeout = sockets[sock_id].in_timeout;
484 	sockets[sock_id].in_timeout = timeout;
485 
486 	/*
487 	 * We do this because the input queue may have some user
488 	 * data already.
489 	 */
490 	old_in_gram = sockets[sock_id].inq;
491 	sockets[sock_id].inq = NULL;
492 
493 	/* Go out and check the wire */
494 	for (i = MEDIA_LVL; i < TRANSPORT_LVL; i++) {
495 		if (sockets[sock_id].input[i] != NULL) {
496 			if (sockets[sock_id].input[i](sock_id) < 0) {
497 				sockets[sock_id].in_timeout = old_timeout;
498 				if (sockets[sock_id].inq != NULL)
499 					nuke_grams(&sockets[sock_id].inq);
500 				sockets[sock_id].inq = old_in_gram;
501 				return (-1);
502 			}
503 		}
504 	}
505 #if DEBUG
506 	printf("tcp_drain_input: done with checking packets\n");
507 #endif
508 	while ((in_gram = sockets[sock_id].inq) != NULL) {
509 		/* Remove unknown inetgrams from the head of inq. */
510 		if (in_gram->igm_level != TRANSPORT_LVL) {
511 #if DEBUG
512 			printf("tcp_drain_input: unexpected packet "
513 			    "level %d frame found\n", in_gram->igm_level);
514 #endif
515 			del_gram(&sockets[sock_id].inq, in_gram, B_TRUE);
516 			continue;
517 		}
518 		mp = in_gram->igm_mp;
519 		del_gram(&sockets[sock_id].inq, in_gram, B_FALSE);
520 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
521 		tcp_rput_data(tcp, mp, sock_id);
522 		sockets[sock_id].in_timeout = old_timeout;
523 
524 		/*
525 		 * The other side may have closed this connection or
526 		 * RST us.  But we need to continue to process other
527 		 * packets in the socket's queue because they may be
528 		 * belong to another TCP connections.
529 		 */
530 		if (sockets[sock_id].pcb == NULL)
531 			tcp = NULL;
532 	}
533 
534 	if (tcp == NULL || sockets[sock_id].pcb == NULL) {
535 		if (sockets[sock_id].so_error != 0)
536 			return (-1);
537 		else
538 			return (0);
539 	}
540 #if DEBUG
541 	printf("tcp_drain_input: done with processing packets\n");
542 #endif
543 	sockets[sock_id].in_timeout = old_timeout;
544 	sockets[sock_id].inq = old_in_gram;
545 
546 	/*
547 	 * Data may have been received so indicate it is available
548 	 */
549 	tcp_drain_needed(sock_id, tcp);
550 	return (0);
551 }
552 
553 /*
554  * The receive entry point for upper layer to call to get data.  Note
555  * that this follows the current architecture that lower layer receive
556  * routines have been called already.  Thus if the inq of socket is
557  * not NULL, the packets must be for us.
558  */
559 static int
560 tcp_input(int sock_id)
561 {
562 	struct inetgram *in_gram;
563 	mblk_t *mp;
564 	tcp_t *tcp;
565 
566 	TCP_RUN_TIME_WAIT_COLLECTOR();
567 
568 	if ((tcp = sockets[sock_id].pcb) == NULL)
569 		return (-1);
570 
571 	while ((in_gram = sockets[sock_id].inq) != NULL) {
572 		/* Remove unknown inetgrams from the head of inq. */
573 		if (in_gram->igm_level != TRANSPORT_LVL) {
574 #ifdef DEBUG
575 			printf("tcp_input: unexpected packet "
576 			    "level %d frame found\n", in_gram->igm_level);
577 #endif
578 			del_gram(&sockets[sock_id].inq, in_gram, B_TRUE);
579 			continue;
580 		}
581 		mp = in_gram->igm_mp;
582 		del_gram(&sockets[sock_id].inq, in_gram, B_FALSE);
583 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
584 		tcp_rput_data(tcp, mp, sock_id);
585 		/* The TCP may be gone because it gets a RST. */
586 		if (sockets[sock_id].pcb == NULL)
587 			return (-1);
588 	}
589 
590 	/* Flush the receive list. */
591 	if (tcp->tcp_rcv_list != NULL) {
592 		tcp_rcv_drain(sock_id, tcp);
593 	} else {
594 		/* The other side has closed the connection, report this up. */
595 		if (tcp->tcp_state == TCPS_CLOSE_WAIT) {
596 			sockets[sock_id].so_state |= SS_CANTRCVMORE;
597 			return (0);
598 		}
599 	}
600 	return (0);
601 }
602 
603 /*
604  * The send entry point for upper layer to call to send data.  In order
605  * to minimize changes to the core TCP code, we need to put the
606  * data into mblks.
607  */
608 int
609 tcp_send(int sock_id, tcp_t *tcp, const void *msg, int len)
610 {
611 	mblk_t *mp;
612 	mblk_t *head = NULL;
613 	mblk_t *tail;
614 	int mss = tcp->tcp_mss;
615 	int cnt = 0;
616 	int win_size;
617 	char *buf = (char *)msg;
618 
619 	TCP_RUN_TIME_WAIT_COLLECTOR();
620 
621 	/* We don't want to append 0 size mblk. */
622 	if (len == 0)
623 		return (0);
624 	while (len > 0) {
625 		if (len < mss) {
626 			mss = len;
627 		}
628 		/*
629 		 * If we cannot allocate more buffer, stop here and
630 		 * the number of bytes buffered will be returned.
631 		 *
632 		 * Note that we follow the core TCP optimization that
633 		 * each mblk contains only MSS bytes data.
634 		 */
635 		if ((mp = allocb(mss + tcp->tcp_ip_hdr_len +
636 		    TCP_MAX_HDR_LENGTH + tcp_wroff_xtra, 0)) == NULL) {
637 			break;
638 		}
639 		mp->b_rptr += tcp->tcp_hdr_len + tcp_wroff_xtra;
640 		bcopy(buf, mp->b_rptr, mss);
641 		mp->b_wptr = mp->b_rptr + mss;
642 		buf += mss;
643 		cnt += mss;
644 		len -= mss;
645 
646 		if (head == NULL) {
647 			head = mp;
648 			tail = mp;
649 		} else {
650 			tail->b_cont = mp;
651 			tail = mp;
652 		}
653 	}
654 
655 	/*
656 	 * Since inetboot is not interrupt driven, there may be
657 	 * some ACKs in the MAC's buffer.  Drain them first,
658 	 * otherwise, we may not be able to send.
659 	 *
660 	 * We expect an ACK in two cases:
661 	 *
662 	 * 1) We have un-ACK'ed data.
663 	 *
664 	 * 2) All ACK's have been received and the sender's window has been
665 	 * closed.  We need an ACK back to open the window so that we can
666 	 * send.  In this case, call tcp_drain_input() if the window size is
667 	 * less than 2 * MSS.
668 	 */
669 
670 	/* window size = MIN(swnd, cwnd) - unacked bytes */
671 	win_size = (tcp->tcp_swnd > tcp->tcp_cwnd) ? tcp->tcp_cwnd :
672 		tcp->tcp_swnd;
673 	win_size -= tcp->tcp_snxt;
674 	win_size += tcp->tcp_suna;
675 	if (win_size < (2 * tcp->tcp_mss))
676 		if (tcp_drain_input(tcp, sock_id, 5) < 0)
677 			return (-1);
678 
679 	tcp_wput_data(tcp, head, sock_id);
680 	/*
681 	 * errno should be reset here as it may be
682 	 * set to ETIMEDOUT. This may be set by
683 	 * the MAC driver in case it has timed out
684 	 * waiting for ARP reply. Any segment which
685 	 * was not transmitted because of ARP timeout
686 	 * will be retransmitted by TCP.
687 	 */
688 	if (errno == ETIMEDOUT)
689 		errno = 0;
690 	return (cnt);
691 }
692 
693 /* Free up all TCP related stuff */
694 static void
695 tcp_free(tcp_t *tcp)
696 {
697 	if (tcp->tcp_iphc != NULL) {
698 		bkmem_free((caddr_t)tcp->tcp_iphc, tcp->tcp_iphc_len);
699 		tcp->tcp_iphc = NULL;
700 	}
701 	if (tcp->tcp_xmit_head != NULL) {
702 		freemsg(tcp->tcp_xmit_head);
703 		tcp->tcp_xmit_head = NULL;
704 	}
705 	if (tcp->tcp_rcv_list != NULL) {
706 		freemsg(tcp->tcp_rcv_list);
707 		tcp->tcp_rcv_list = NULL;
708 	}
709 	if (tcp->tcp_reass_head != NULL) {
710 		freemsg(tcp->tcp_reass_head);
711 		tcp->tcp_reass_head = NULL;
712 	}
713 	if (tcp->tcp_sack_info != NULL) {
714 		bkmem_free((caddr_t)tcp->tcp_sack_info,
715 		    sizeof (tcp_sack_info_t));
716 		tcp->tcp_sack_info = NULL;
717 	}
718 }
719 
720 static void
721 tcp_close_detached(tcp_t *tcp)
722 {
723 	if (tcp->tcp_listener != NULL)
724 		tcp_eager_unlink(tcp);
725 	tcp_free(tcp);
726 	bkmem_free((caddr_t)tcp, sizeof (tcp_t));
727 }
728 
729 /*
730  * If we are an eager connection hanging off a listener that hasn't
731  * formally accepted the connection yet, get off its list and blow off
732  * any data that we have accumulated.
733  */
734 static void
735 tcp_eager_unlink(tcp_t *tcp)
736 {
737 	tcp_t	*listener = tcp->tcp_listener;
738 
739 	assert(listener != NULL);
740 	if (tcp->tcp_eager_next_q0 != NULL) {
741 		assert(tcp->tcp_eager_prev_q0 != NULL);
742 
743 		/* Remove the eager tcp from q0 */
744 		tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
745 		    tcp->tcp_eager_prev_q0;
746 		tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
747 		    tcp->tcp_eager_next_q0;
748 		listener->tcp_conn_req_cnt_q0--;
749 	} else {
750 		tcp_t   **tcpp = &listener->tcp_eager_next_q;
751 		tcp_t	*prev = NULL;
752 
753 		for (; tcpp[0]; tcpp = &tcpp[0]->tcp_eager_next_q) {
754 			if (tcpp[0] == tcp) {
755 				if (listener->tcp_eager_last_q == tcp) {
756 					/*
757 					 * If we are unlinking the last
758 					 * element on the list, adjust
759 					 * tail pointer. Set tail pointer
760 					 * to nil when list is empty.
761 					 */
762 					assert(tcp->tcp_eager_next_q == NULL);
763 					if (listener->tcp_eager_last_q ==
764 					    listener->tcp_eager_next_q) {
765 						listener->tcp_eager_last_q =
766 						NULL;
767 					} else {
768 						/*
769 						 * We won't get here if there
770 						 * is only one eager in the
771 						 * list.
772 						 */
773 						assert(prev != NULL);
774 						listener->tcp_eager_last_q =
775 						    prev;
776 					}
777 				}
778 				tcpp[0] = tcp->tcp_eager_next_q;
779 				tcp->tcp_eager_next_q = NULL;
780 				tcp->tcp_eager_last_q = NULL;
781 				listener->tcp_conn_req_cnt_q--;
782 				break;
783 			}
784 			prev = tcpp[0];
785 		}
786 	}
787 	tcp->tcp_listener = NULL;
788 }
789 
790 /*
791  * Reset any eager connection hanging off this listener
792  * and then reclaim it's resources.
793  */
794 static void
795 tcp_eager_cleanup(tcp_t *listener, boolean_t q0_only, int sock_id)
796 {
797 	tcp_t	*eager;
798 
799 	if (!q0_only) {
800 		/* First cleanup q */
801 		while ((eager = listener->tcp_eager_next_q) != NULL) {
802 			assert(listener->tcp_eager_last_q != NULL);
803 			tcp_xmit_ctl("tcp_eager_cleanup, can't wait",
804 			    eager, NULL, eager->tcp_snxt, 0, TH_RST, 0,
805 			    sock_id);
806 			tcp_close_detached(eager);
807 		}
808 		assert(listener->tcp_eager_last_q == NULL);
809 	}
810 	/* Then cleanup q0 */
811 	while ((eager = listener->tcp_eager_next_q0) != listener) {
812 		tcp_xmit_ctl("tcp_eager_cleanup, can't wait",
813 		    eager, NULL, eager->tcp_snxt, 0, TH_RST, 0, sock_id);
814 		tcp_close_detached(eager);
815 	}
816 }
817 
818 /*
819  * To handle the shutdown request. Called from shutdown()
820  */
821 int
822 tcp_shutdown(int sock_id)
823 {
824 	tcp_t	*tcp;
825 
826 	DEBUG_1("tcp_shutdown: sock_id %x\n", sock_id);
827 
828 	if ((tcp = sockets[sock_id].pcb) == NULL) {
829 		return (-1);
830 	}
831 
832 	/*
833 	 * Since inetboot is not interrupt driven, there may be
834 	 * some ACKs in the MAC's buffer.  Drain them first,
835 	 * otherwise, we may not be able to send.
836 	 */
837 	if (tcp_drain_input(tcp, sock_id, 5) < 0) {
838 		/*
839 		 * If we return now without freeing TCP, there will be
840 		 * a memory leak.
841 		 */
842 		if (sockets[sock_id].pcb != NULL)
843 			tcp_clean_death(sock_id, tcp, 0);
844 		return (-1);
845 	}
846 
847 	DEBUG_1("tcp_shutdown: tcp_state %x\n", tcp->tcp_state);
848 	switch (tcp->tcp_state) {
849 
850 	case TCPS_SYN_RCVD:
851 		/*
852 		 * Shutdown during the connect 3-way handshake
853 		 */
854 	case TCPS_ESTABLISHED:
855 		/*
856 		 * Transmit the FIN
857 		 * wait for the FIN to be ACKed,
858 		 * then remain in FIN_WAIT_2
859 		 */
860 		dprintf("tcp_shutdown: sending fin\n");
861 		if (tcp_xmit_end(tcp, sock_id) == 0 &&
862 			tcp_state_wait(sock_id, tcp, TCPS_FIN_WAIT_2) < 0) {
863 			/* During the wait, TCP may be gone... */
864 			if (sockets[sock_id].pcb == NULL)
865 				return (-1);
866 		}
867 		dprintf("tcp_shutdown: done\n");
868 		break;
869 
870 	default:
871 		break;
872 
873 	}
874 	return (0);
875 }
876 
877 /* To handle closing of the socket */
878 static int
879 tcp_close(int sock_id)
880 {
881 	char	*msg;
882 	tcp_t	*tcp;
883 	int	error = 0;
884 
885 	if ((tcp = sockets[sock_id].pcb) == NULL) {
886 		return (-1);
887 	}
888 
889 	TCP_RUN_TIME_WAIT_COLLECTOR();
890 
891 	/*
892 	 * Since inetboot is not interrupt driven, there may be
893 	 * some ACKs in the MAC's buffer.  Drain them first,
894 	 * otherwise, we may not be able to send.
895 	 */
896 	if (tcp_drain_input(tcp, sock_id, 5) < 0) {
897 		/*
898 		 * If we return now without freeing TCP, there will be
899 		 * a memory leak.
900 		 */
901 		if (sockets[sock_id].pcb != NULL)
902 			tcp_clean_death(sock_id, tcp, 0);
903 		return (-1);
904 	}
905 
906 	if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
907 		/* Cleanup for listener */
908 		tcp_eager_cleanup(tcp, 0, sock_id);
909 	}
910 
911 	msg = NULL;
912 	switch (tcp->tcp_state) {
913 	case TCPS_CLOSED:
914 	case TCPS_IDLE:
915 	case TCPS_BOUND:
916 	case TCPS_LISTEN:
917 		break;
918 	case TCPS_SYN_SENT:
919 		msg = "tcp_close, during connect";
920 		break;
921 	case TCPS_SYN_RCVD:
922 		/*
923 		 * Close during the connect 3-way handshake
924 		 * but here there may or may not be pending data
925 		 * already on queue. Process almost same as in
926 		 * the ESTABLISHED state.
927 		 */
928 		/* FALLTHRU */
929 	default:
930 		/*
931 		 * If SO_LINGER has set a zero linger time, abort the
932 		 * connection with a reset.
933 		 */
934 		if (tcp->tcp_linger && tcp->tcp_lingertime == 0) {
935 			msg = "tcp_close, zero lingertime";
936 			break;
937 		}
938 
939 		/*
940 		 * Abort connection if there is unread data queued.
941 		 */
942 		if (tcp->tcp_rcv_list != NULL ||
943 				tcp->tcp_reass_head != NULL) {
944 			msg = "tcp_close, unread data";
945 			break;
946 		}
947 		if (tcp->tcp_state <= TCPS_LISTEN)
948 			break;
949 
950 		/*
951 		 * Transmit the FIN before detaching the tcp_t.
952 		 * After tcp_detach returns this queue/perimeter
953 		 * no longer owns the tcp_t thus others can modify it.
954 		 * The TCP could be closed in tcp_state_wait called by
955 		 * tcp_wput_data called by tcp_xmit_end.
956 		 */
957 		(void) tcp_xmit_end(tcp, sock_id);
958 		if (sockets[sock_id].pcb == NULL)
959 			return (0);
960 
961 		/*
962 		 * If lingering on close then wait until the fin is acked,
963 		 * the SO_LINGER time passes, or a reset is sent/received.
964 		 */
965 		if (tcp->tcp_linger && tcp->tcp_lingertime > 0 &&
966 		    !(tcp->tcp_fin_acked) &&
967 		    tcp->tcp_state >= TCPS_ESTABLISHED) {
968 			uint32_t stoptime; /* in ms */
969 
970 			tcp->tcp_client_errno = 0;
971 			stoptime = prom_gettime() +
972 			    (tcp->tcp_lingertime * 1000);
973 			while (!(tcp->tcp_fin_acked) &&
974 			    tcp->tcp_state >= TCPS_ESTABLISHED &&
975 			    tcp->tcp_client_errno == 0 &&
976 			    ((int32_t)(stoptime - prom_gettime()) > 0)) {
977 				if (tcp_drain_input(tcp, sock_id, 5) < 0) {
978 					if (sockets[sock_id].pcb != NULL) {
979 						tcp_clean_death(sock_id,
980 						    tcp, 0);
981 					}
982 					return (-1);
983 				}
984 			}
985 			tcp->tcp_client_errno = 0;
986 		}
987 		if (tcp_state_wait(sock_id, tcp, TCPS_TIME_WAIT) < 0) {
988 			/* During the wait, TCP may be gone... */
989 			if (sockets[sock_id].pcb == NULL)
990 				return (0);
991 			msg = "tcp_close, couldn't detach";
992 		} else {
993 			return (0);
994 		}
995 		break;
996 	}
997 
998 	/* Something went wrong...  Send a RST and report the error */
999 	if (msg != NULL) {
1000 		if (tcp->tcp_state == TCPS_ESTABLISHED ||
1001 		    tcp->tcp_state == TCPS_CLOSE_WAIT)
1002 			BUMP_MIB(tcp_mib.tcpEstabResets);
1003 		if (tcp->tcp_state == TCPS_SYN_SENT ||
1004 		    tcp->tcp_state == TCPS_SYN_RCVD)
1005 			BUMP_MIB(tcp_mib.tcpAttemptFails);
1006 		tcp_xmit_ctl(msg, tcp, NULL, tcp->tcp_snxt, 0, TH_RST, 0,
1007 		    sock_id);
1008 	}
1009 
1010 	tcp_free(tcp);
1011 	bkmem_free((caddr_t)tcp, sizeof (tcp_t));
1012 	sockets[sock_id].pcb = NULL;
1013 	return (error);
1014 }
1015 
1016 /* To make an endpoint a listener. */
1017 int
1018 tcp_listen(int sock_id, int backlog)
1019 {
1020 	tcp_t *tcp;
1021 
1022 	if ((tcp = (tcp_t *)(sockets[sock_id].pcb)) == NULL) {
1023 		errno = EINVAL;
1024 		return (-1);
1025 	}
1026 	/* We allow calling listen() multiple times to change the backlog. */
1027 	if (tcp->tcp_state > TCPS_LISTEN || tcp->tcp_state < TCPS_BOUND) {
1028 		errno = EOPNOTSUPP;
1029 		return (-1);
1030 	}
1031 	/* The following initialization should only be done once. */
1032 	if (tcp->tcp_state != TCPS_LISTEN) {
1033 		tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
1034 		tcp->tcp_eager_next_q = NULL;
1035 		tcp->tcp_state = TCPS_LISTEN;
1036 		tcp->tcp_second_ctimer_threshold = tcp_ip_abort_linterval;
1037 	}
1038 	if ((tcp->tcp_conn_req_max = backlog) > tcp_conn_req_max_q) {
1039 		tcp->tcp_conn_req_max = tcp_conn_req_max_q;
1040 	}
1041 	if (tcp->tcp_conn_req_max < tcp_conn_req_min) {
1042 		tcp->tcp_conn_req_max = tcp_conn_req_min;
1043 	}
1044 	return (0);
1045 }
1046 
1047 /* To accept connections. */
1048 int
1049 tcp_accept(int sock_id, struct sockaddr *addr, socklen_t *addr_len)
1050 {
1051 	tcp_t *listener;
1052 	tcp_t *eager;
1053 	int sd, new_sock_id;
1054 	struct sockaddr_in *new_addr = (struct sockaddr_in *)addr;
1055 	int timeout;
1056 
1057 	/* Sanity check. */
1058 	if ((listener = (tcp_t *)(sockets[sock_id].pcb)) == NULL ||
1059 	    new_addr == NULL || addr_len == NULL ||
1060 	    *addr_len < sizeof (struct sockaddr_in) ||
1061 	    listener->tcp_state != TCPS_LISTEN) {
1062 		errno = EINVAL;
1063 		return (-1);
1064 	}
1065 
1066 	if (sockets[sock_id].in_timeout > tcp_accept_timeout)
1067 		timeout = prom_gettime() + sockets[sock_id].in_timeout;
1068 	else
1069 		timeout = prom_gettime() + tcp_accept_timeout;
1070 	while (listener->tcp_eager_next_q == NULL &&
1071 	    timeout > prom_gettime()) {
1072 #if DEBUG
1073 		printf("tcp_accept: Waiting in tcp_accept()\n");
1074 #endif
1075 		if (tcp_drain_input(listener, sock_id, 5) < 0) {
1076 			return (-1);
1077 		}
1078 	}
1079 	/* If there is an eager, don't timeout... */
1080 	if (timeout <= prom_gettime() && listener->tcp_eager_next_q == NULL) {
1081 #if DEBUG
1082 		printf("tcp_accept: timeout\n");
1083 #endif
1084 		errno = ETIMEDOUT;
1085 		return (-1);
1086 	}
1087 #if DEBUG
1088 	printf("tcp_accept: got a connection\n");
1089 #endif
1090 
1091 	/* Now create the socket for this new TCP. */
1092 	if ((sd = socket(AF_INET, SOCK_STREAM, 0)) < 0) {
1093 		return (-1);
1094 	}
1095 	if ((new_sock_id = so_check_fd(sd, &errno)) == -1)
1096 		/* This should not happen! */
1097 		prom_panic("so_check_fd() fails in tcp_accept()");
1098 	/* Free the TCP PCB in the original socket. */
1099 	bkmem_free((caddr_t)(sockets[new_sock_id].pcb), sizeof (tcp_t));
1100 	/* Dequeue the eager and attach it to the socket. */
1101 	eager = listener->tcp_eager_next_q;
1102 	listener->tcp_eager_next_q = eager->tcp_eager_next_q;
1103 	if (listener->tcp_eager_last_q == eager)
1104 		listener->tcp_eager_last_q = NULL;
1105 	eager->tcp_eager_next_q = NULL;
1106 	sockets[new_sock_id].pcb = eager;
1107 	listener->tcp_conn_req_cnt_q--;
1108 
1109 	/* Copy in the address info. */
1110 	bcopy(&eager->tcp_remote, &new_addr->sin_addr.s_addr,
1111 	    sizeof (in_addr_t));
1112 	bcopy(&eager->tcp_fport, &new_addr->sin_port, sizeof (in_port_t));
1113 	new_addr->sin_family = AF_INET;
1114 
1115 #ifdef DEBUG
1116 	printf("tcp_accept(), new sock_id: %d\n", sd);
1117 #endif
1118 	return (sd);
1119 }
1120 
1121 /* Update the next anonymous port to use.  */
1122 static in_port_t
1123 tcp_update_next_port(in_port_t port)
1124 {
1125 	/* Don't allow the port to fall out of the anonymous port range. */
1126 	if (port < tcp_smallest_anon_port || port > tcp_largest_anon_port)
1127 		port = (in_port_t)tcp_smallest_anon_port;
1128 
1129 	if (port < tcp_smallest_nonpriv_port)
1130 		port = (in_port_t)tcp_smallest_nonpriv_port;
1131 	return (port);
1132 }
1133 
1134 /* To check whether a bind to a port is allowed. */
1135 static in_port_t
1136 tcp_bindi(in_port_t port, in_addr_t *addr, boolean_t reuseaddr,
1137     boolean_t bind_to_req_port_only)
1138 {
1139 	int i, count;
1140 	tcp_t *tcp;
1141 
1142 	count = tcp_largest_anon_port - tcp_smallest_anon_port;
1143 try_again:
1144 	for (i = 0; i < MAXSOCKET; i++) {
1145 		if (sockets[i].type != INETBOOT_STREAM ||
1146 		    ((tcp = (tcp_t *)sockets[i].pcb) == NULL) ||
1147 		    ntohs(tcp->tcp_lport) != port) {
1148 			continue;
1149 		}
1150 		/*
1151 		 * Both TCPs have the same port.  If SO_REUSEDADDR is
1152 		 * set and the bound TCP has a state greater than
1153 		 * TCPS_LISTEN, it is fine.
1154 		 */
1155 		if (reuseaddr && tcp->tcp_state > TCPS_LISTEN) {
1156 			continue;
1157 		}
1158 		if (tcp->tcp_bound_source != INADDR_ANY &&
1159 		    *addr != INADDR_ANY &&
1160 		    tcp->tcp_bound_source != *addr) {
1161 			continue;
1162 		}
1163 		if (bind_to_req_port_only) {
1164 			return (0);
1165 		}
1166 		if (--count > 0) {
1167 			port = tcp_update_next_port(++port);
1168 			goto try_again;
1169 		} else {
1170 			return (0);
1171 		}
1172 	}
1173 	return (port);
1174 }
1175 
1176 /* To handle the bind request. */
1177 int
1178 tcp_bind(int sock_id)
1179 {
1180 	tcp_t *tcp;
1181 	in_port_t requested_port, allocated_port;
1182 	boolean_t bind_to_req_port_only;
1183 	boolean_t reuseaddr;
1184 
1185 	if ((tcp = (tcp_t *)sockets[sock_id].pcb) == NULL) {
1186 		errno = EINVAL;
1187 		return (-1);
1188 	}
1189 
1190 	if (tcp->tcp_state >= TCPS_BOUND) {
1191 		/* We don't allow multiple bind(). */
1192 		errno = EPROTO;
1193 		return (-1);
1194 	}
1195 
1196 	requested_port = ntohs(sockets[sock_id].bind.sin_port);
1197 
1198 	/* The bound source can be INADDR_ANY. */
1199 	tcp->tcp_bound_source = sockets[sock_id].bind.sin_addr.s_addr;
1200 
1201 	tcp->tcp_ipha->ip_src.s_addr = tcp->tcp_bound_source;
1202 
1203 	/* Verify the port is available. */
1204 	if (requested_port == 0)
1205 		bind_to_req_port_only = B_FALSE;
1206 	else			/* T_BIND_REQ and requested_port != 0 */
1207 		bind_to_req_port_only = B_TRUE;
1208 
1209 	if (requested_port == 0) {
1210 		requested_port = tcp_update_next_port(++tcp_next_port_to_try);
1211 	}
1212 	reuseaddr = sockets[sock_id].so_opt & SO_REUSEADDR;
1213 	allocated_port = tcp_bindi(requested_port, &(tcp->tcp_bound_source),
1214 	    reuseaddr, bind_to_req_port_only);
1215 
1216 	if (allocated_port == 0) {
1217 		errno = EADDRINUSE;
1218 		return (-1);
1219 	}
1220 	tcp->tcp_lport = htons(allocated_port);
1221 	*(uint16_t *)tcp->tcp_tcph->th_lport = tcp->tcp_lport;
1222 	sockets[sock_id].bind.sin_port = tcp->tcp_lport;
1223 	tcp->tcp_state = TCPS_BOUND;
1224 	return (0);
1225 }
1226 
1227 /*
1228  * Check for duplicate TCP connections.
1229  */
1230 static int
1231 tcp_conn_check(tcp_t *tcp)
1232 {
1233 	int i;
1234 	tcp_t *tmp_tcp;
1235 
1236 	for (i = 0; i < MAXSOCKET; i++) {
1237 		if (sockets[i].type != INETBOOT_STREAM)
1238 			continue;
1239 		/* Socket may not be closed but the TCP can be gone. */
1240 		if ((tmp_tcp = (tcp_t *)sockets[i].pcb) == NULL)
1241 			continue;
1242 		/* We only care about TCP in states later than SYN_SENT. */
1243 		if (tmp_tcp->tcp_state < TCPS_SYN_SENT)
1244 			continue;
1245 		if (tmp_tcp->tcp_lport != tcp->tcp_lport ||
1246 		    tmp_tcp->tcp_fport != tcp->tcp_fport ||
1247 		    tmp_tcp->tcp_bound_source != tcp->tcp_bound_source ||
1248 		    tmp_tcp->tcp_remote != tcp->tcp_remote) {
1249 			continue;
1250 		} else {
1251 			return (-1);
1252 		}
1253 	}
1254 	return (0);
1255 }
1256 
1257 /* To handle a connect request. */
1258 int
1259 tcp_connect(int sock_id)
1260 {
1261 	tcp_t *tcp;
1262 	in_addr_t dstaddr;
1263 	in_port_t dstport;
1264 	tcph_t	*tcph;
1265 	int mss;
1266 	mblk_t *syn_mp;
1267 
1268 	if ((tcp = (tcp_t *)(sockets[sock_id].pcb)) == NULL) {
1269 		errno = EINVAL;
1270 		return (-1);
1271 	}
1272 
1273 	TCP_RUN_TIME_WAIT_COLLECTOR();
1274 
1275 	dstaddr = sockets[sock_id].remote.sin_addr.s_addr;
1276 	dstport = sockets[sock_id].remote.sin_port;
1277 
1278 	/*
1279 	 * Check for attempt to connect to INADDR_ANY or non-unicast addrress.
1280 	 * We don't have enough info to check for broadcast addr, except
1281 	 * for the all 1 broadcast.
1282 	 */
1283 	if (dstaddr == INADDR_ANY || IN_CLASSD(ntohl(dstaddr)) ||
1284 	    dstaddr == INADDR_BROADCAST)  {
1285 		/*
1286 		 * SunOS 4.x and 4.3 BSD allow an application
1287 		 * to connect a TCP socket to INADDR_ANY.
1288 		 * When they do this, the kernel picks the
1289 		 * address of one interface and uses it
1290 		 * instead.  The kernel usually ends up
1291 		 * picking the address of the loopback
1292 		 * interface.  This is an undocumented feature.
1293 		 * However, we provide the same thing here
1294 		 * in order to have source and binary
1295 		 * compatibility with SunOS 4.x.
1296 		 * Update the T_CONN_REQ (sin/sin6) since it is used to
1297 		 * generate the T_CONN_CON.
1298 		 *
1299 		 * Fail this for inetboot TCP.
1300 		 */
1301 		errno = EINVAL;
1302 		return (-1);
1303 	}
1304 
1305 	/* It is not bound to any address yet... */
1306 	if (tcp->tcp_bound_source == INADDR_ANY) {
1307 		ipv4_getipaddr(&(sockets[sock_id].bind.sin_addr));
1308 		/* We don't have an address! */
1309 		if (ntohl(sockets[sock_id].bind.sin_addr.s_addr) ==
1310 		    INADDR_ANY) {
1311 			errno = EPROTO;
1312 			return (-1);
1313 		}
1314 		tcp->tcp_bound_source = sockets[sock_id].bind.sin_addr.s_addr;
1315 		tcp->tcp_ipha->ip_src.s_addr = tcp->tcp_bound_source;
1316 	}
1317 
1318 	/*
1319 	 * Don't let an endpoint connect to itself.
1320 	 */
1321 	if (dstaddr == tcp->tcp_ipha->ip_src.s_addr &&
1322 	    dstport == tcp->tcp_lport) {
1323 		errno = EINVAL;
1324 		return (-1);
1325 	}
1326 
1327 	tcp->tcp_ipha->ip_dst.s_addr = dstaddr;
1328 	tcp->tcp_remote = dstaddr;
1329 	tcph = tcp->tcp_tcph;
1330 	*(uint16_t *)tcph->th_fport = dstport;
1331 	tcp->tcp_fport = dstport;
1332 
1333 	/*
1334 	 * Don't allow this connection to completely duplicate
1335 	 * an existing connection.
1336 	 */
1337 	if (tcp_conn_check(tcp) < 0) {
1338 		errno = EADDRINUSE;
1339 		return (-1);
1340 	}
1341 
1342 	/*
1343 	 * Just make sure our rwnd is at
1344 	 * least tcp_recv_hiwat_mss * MSS
1345 	 * large, and round up to the nearest
1346 	 * MSS.
1347 	 *
1348 	 * We do the round up here because
1349 	 * we need to get the interface
1350 	 * MTU first before we can do the
1351 	 * round up.
1352 	 */
1353 	mss = tcp->tcp_mss - tcp->tcp_hdr_len;
1354 	tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
1355 	    tcp_recv_hiwat_minmss * mss);
1356 	tcp->tcp_rwnd_max = tcp->tcp_rwnd;
1357 	SET_WS_VALUE(tcp);
1358 	U32_TO_ABE16((tcp->tcp_rwnd >> tcp->tcp_rcv_ws),
1359 	    tcp->tcp_tcph->th_win);
1360 	if (tcp->tcp_rcv_ws > 0 || tcp_wscale_always)
1361 		tcp->tcp_snd_ws_ok = B_TRUE;
1362 
1363 	/*
1364 	 * Set tcp_snd_ts_ok to true
1365 	 * so that tcp_xmit_mp will
1366 	 * include the timestamp
1367 	 * option in the SYN segment.
1368 	 */
1369 	if (tcp_tstamp_always ||
1370 	    (tcp->tcp_rcv_ws && tcp_tstamp_if_wscale)) {
1371 		tcp->tcp_snd_ts_ok = B_TRUE;
1372 	}
1373 
1374 	if (tcp_sack_permitted == 2 ||
1375 	    tcp->tcp_snd_sack_ok) {
1376 		assert(tcp->tcp_sack_info == NULL);
1377 		if ((tcp->tcp_sack_info = (tcp_sack_info_t *)bkmem_zalloc(
1378 		    sizeof (tcp_sack_info_t))) == NULL) {
1379 			tcp->tcp_snd_sack_ok = B_FALSE;
1380 		} else {
1381 			tcp->tcp_snd_sack_ok = B_TRUE;
1382 		}
1383 	}
1384 	/*
1385 	 * Should we use ECN?  Note that the current
1386 	 * default value (SunOS 5.9) of tcp_ecn_permitted
1387 	 * is 2.  The reason for doing this is that there
1388 	 * are equipments out there that will drop ECN
1389 	 * enabled IP packets.  Setting it to 1 avoids
1390 	 * compatibility problems.
1391 	 */
1392 	if (tcp_ecn_permitted == 2)
1393 		tcp->tcp_ecn_ok = B_TRUE;
1394 
1395 	tcp_iss_init(tcp);
1396 	TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1397 	tcp->tcp_active_open = B_TRUE;
1398 
1399 	tcp->tcp_state = TCPS_SYN_SENT;
1400 	syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL, tcp->tcp_iss, B_FALSE,
1401 	    NULL, B_FALSE);
1402 	if (syn_mp != NULL) {
1403 		int ret;
1404 
1405 		/* Dump the packet when debugging. */
1406 		TCP_DUMP_PACKET("tcp_connect", syn_mp);
1407 		/* Send out the SYN packet. */
1408 		ret = ipv4_tcp_output(sock_id, syn_mp);
1409 		freeb(syn_mp);
1410 		/*
1411 		 * errno ETIMEDOUT is set by the mac driver
1412 		 * in case it is not able to receive ARP reply.
1413 		 * TCP will retransmit this segment so we can
1414 		 * ignore the ARP timeout.
1415 		 */
1416 		if ((ret < 0) && (errno != ETIMEDOUT)) {
1417 			return (-1);
1418 		}
1419 		/* tcp_state_wait() will finish the 3 way handshake. */
1420 		return (tcp_state_wait(sock_id, tcp, TCPS_ESTABLISHED));
1421 	} else {
1422 		errno = ENOBUFS;
1423 		return (-1);
1424 	}
1425 }
1426 
1427 /*
1428  * Common accept code.  Called by tcp_conn_request.
1429  * cr_pkt is the SYN packet.
1430  */
1431 static int
1432 tcp_accept_comm(tcp_t *listener, tcp_t *acceptor, mblk_t *cr_pkt,
1433     uint_t ip_hdr_len)
1434 {
1435 	tcph_t		*tcph;
1436 
1437 #ifdef DEBUG
1438 	printf("tcp_accept_comm #######################\n");
1439 #endif
1440 
1441 	/*
1442 	 * When we get here, we know that the acceptor header template
1443 	 * has already been initialized.
1444 	 * However, it may not match the listener if the listener
1445 	 * includes options...
1446 	 * It may also not match the listener if the listener is v6 and
1447 	 * and the acceptor is v4
1448 	 */
1449 	acceptor->tcp_lport = listener->tcp_lport;
1450 
1451 	if (listener->tcp_ipversion == acceptor->tcp_ipversion) {
1452 		if (acceptor->tcp_iphc_len != listener->tcp_iphc_len) {
1453 			/*
1454 			 * Listener had options of some sort; acceptor inherits.
1455 			 * Free up the acceptor template and allocate one
1456 			 * of the right size.
1457 			 */
1458 			bkmem_free(acceptor->tcp_iphc, acceptor->tcp_iphc_len);
1459 			acceptor->tcp_iphc = bkmem_zalloc(
1460 			    listener->tcp_iphc_len);
1461 			if (acceptor->tcp_iphc == NULL) {
1462 				acceptor->tcp_iphc_len = 0;
1463 				return (ENOMEM);
1464 			}
1465 			acceptor->tcp_iphc_len = listener->tcp_iphc_len;
1466 		}
1467 		acceptor->tcp_hdr_len = listener->tcp_hdr_len;
1468 		acceptor->tcp_ip_hdr_len = listener->tcp_ip_hdr_len;
1469 		acceptor->tcp_tcp_hdr_len = listener->tcp_tcp_hdr_len;
1470 
1471 		/*
1472 		 * Copy the IP+TCP header template from listener to acceptor
1473 		 */
1474 		bcopy(listener->tcp_iphc, acceptor->tcp_iphc,
1475 		    listener->tcp_hdr_len);
1476 		acceptor->tcp_ipha = (struct ip *)acceptor->tcp_iphc;
1477 		acceptor->tcp_tcph = (tcph_t *)(acceptor->tcp_iphc +
1478 		    acceptor->tcp_ip_hdr_len);
1479 	} else {
1480 		prom_panic("tcp_accept_comm: version not equal");
1481 	}
1482 
1483 	/* Copy our new dest and fport from the connection request packet */
1484 	if (acceptor->tcp_ipversion == IPV4_VERSION) {
1485 		struct ip *ipha;
1486 
1487 		ipha = (struct ip *)cr_pkt->b_rptr;
1488 		acceptor->tcp_ipha->ip_dst = ipha->ip_src;
1489 		acceptor->tcp_remote = ipha->ip_src.s_addr;
1490 		acceptor->tcp_ipha->ip_src = ipha->ip_dst;
1491 		acceptor->tcp_bound_source = ipha->ip_dst.s_addr;
1492 		tcph = (tcph_t *)&cr_pkt->b_rptr[ip_hdr_len];
1493 	} else {
1494 		prom_panic("tcp_accept_comm: not IPv4");
1495 	}
1496 	bcopy(tcph->th_lport, acceptor->tcp_tcph->th_fport, sizeof (in_port_t));
1497 	bcopy(acceptor->tcp_tcph->th_fport, &acceptor->tcp_fport,
1498 	    sizeof (in_port_t));
1499 	/*
1500 	 * For an all-port proxy listener, the local port is determined by
1501 	 * the port number field in the SYN packet.
1502 	 */
1503 	if (listener->tcp_lport == 0) {
1504 		acceptor->tcp_lport = *(in_port_t *)tcph->th_fport;
1505 		bcopy(tcph->th_fport, acceptor->tcp_tcph->th_lport,
1506 		    sizeof (in_port_t));
1507 	}
1508 	/* Inherit various TCP parameters from the listener */
1509 	acceptor->tcp_naglim = listener->tcp_naglim;
1510 	acceptor->tcp_first_timer_threshold =
1511 	    listener->tcp_first_timer_threshold;
1512 	acceptor->tcp_second_timer_threshold =
1513 	    listener->tcp_second_timer_threshold;
1514 
1515 	acceptor->tcp_first_ctimer_threshold =
1516 	    listener->tcp_first_ctimer_threshold;
1517 	acceptor->tcp_second_ctimer_threshold =
1518 	    listener->tcp_second_ctimer_threshold;
1519 
1520 	acceptor->tcp_xmit_hiwater = listener->tcp_xmit_hiwater;
1521 
1522 	acceptor->tcp_state = TCPS_LISTEN;
1523 	tcp_iss_init(acceptor);
1524 
1525 	/* Process all TCP options. */
1526 	tcp_process_options(acceptor, tcph);
1527 
1528 	/* Is the other end ECN capable? */
1529 	if (tcp_ecn_permitted >= 1 &&
1530 	    (tcph->th_flags[0] & (TH_ECE|TH_CWR)) == (TH_ECE|TH_CWR)) {
1531 		acceptor->tcp_ecn_ok = B_TRUE;
1532 	}
1533 
1534 	/*
1535 	 * listener->tcp_rq->q_hiwat should be the default window size or a
1536 	 * window size changed via SO_RCVBUF option.  First round up the
1537 	 * acceptor's tcp_rwnd to the nearest MSS.  Then find out the window
1538 	 * scale option value if needed.  Call tcp_rwnd_set() to finish the
1539 	 * setting.
1540 	 *
1541 	 * Note if there is a rpipe metric associated with the remote host,
1542 	 * we should not inherit receive window size from listener.
1543 	 */
1544 	acceptor->tcp_rwnd = MSS_ROUNDUP(
1545 	    (acceptor->tcp_rwnd == 0 ? listener->tcp_rwnd_max :
1546 	    acceptor->tcp_rwnd), acceptor->tcp_mss);
1547 	if (acceptor->tcp_snd_ws_ok)
1548 		SET_WS_VALUE(acceptor);
1549 	/*
1550 	 * Note that this is the only place tcp_rwnd_set() is called for
1551 	 * accepting a connection.  We need to call it here instead of
1552 	 * after the 3-way handshake because we need to tell the other
1553 	 * side our rwnd in the SYN-ACK segment.
1554 	 */
1555 	(void) tcp_rwnd_set(acceptor, acceptor->tcp_rwnd);
1556 
1557 	return (0);
1558 }
1559 
1560 /*
1561  * Defense for the SYN attack -
1562  * 1. When q0 is full, drop from the tail (tcp_eager_prev_q0) the oldest
1563  *    one that doesn't have the dontdrop bit set.
1564  * 2. Don't drop a SYN request before its first timeout. This gives every
1565  *    request at least til the first timeout to complete its 3-way handshake.
1566  * 3. The current threshold is - # of timeout > q0len/4 => SYN alert on
1567  *    # of timeout drops back to <= q0len/32 => SYN alert off
1568  */
1569 static boolean_t
1570 tcp_drop_q0(tcp_t *tcp)
1571 {
1572 	tcp_t	*eager;
1573 
1574 	assert(tcp->tcp_eager_next_q0 != tcp->tcp_eager_prev_q0);
1575 	/*
1576 	 * New one is added after next_q0 so prev_q0 points to the oldest
1577 	 * Also do not drop any established connections that are deferred on
1578 	 * q0 due to q being full
1579 	 */
1580 
1581 	eager = tcp->tcp_eager_prev_q0;
1582 	while (eager->tcp_dontdrop || eager->tcp_conn_def_q0) {
1583 		/* XXX should move the eager to the head */
1584 		eager = eager->tcp_eager_prev_q0;
1585 		if (eager == tcp) {
1586 			eager = tcp->tcp_eager_prev_q0;
1587 			break;
1588 		}
1589 	}
1590 	dprintf("tcp_drop_q0: listen half-open queue (max=%d) overflow"
1591 	    " (%d pending) on %s, drop one", tcp_conn_req_max_q0,
1592 	    tcp->tcp_conn_req_cnt_q0,
1593 	    tcp_display(tcp, NULL, DISP_PORT_ONLY));
1594 
1595 	BUMP_MIB(tcp_mib.tcpHalfOpenDrop);
1596 	bkmem_free((caddr_t)eager, sizeof (tcp_t));
1597 	return (B_TRUE);
1598 }
1599 
1600 /* ARGSUSED */
1601 static tcp_t *
1602 tcp_conn_request(tcp_t *tcp, mblk_t *mp, uint_t sock_id, uint_t ip_hdr_len)
1603 {
1604 	tcp_t	*eager;
1605 	struct ip *ipha;
1606 	int	err;
1607 
1608 #ifdef DEBUG
1609 	printf("tcp_conn_request ###################\n");
1610 #endif
1611 
1612 	if (tcp->tcp_conn_req_cnt_q >= tcp->tcp_conn_req_max) {
1613 		BUMP_MIB(tcp_mib.tcpListenDrop);
1614 		dprintf("tcp_conn_request: listen backlog (max=%d) "
1615 		    "overflow (%d pending) on %s",
1616 		    tcp->tcp_conn_req_max, tcp->tcp_conn_req_cnt_q,
1617 		    tcp_display(tcp, NULL, DISP_PORT_ONLY));
1618 		return (NULL);
1619 	}
1620 
1621 	assert(OK_32PTR(mp->b_rptr));
1622 
1623 	if (tcp->tcp_conn_req_cnt_q0 >=
1624 	    tcp->tcp_conn_req_max + tcp_conn_req_max_q0) {
1625 		/*
1626 		 * Q0 is full. Drop a pending half-open req from the queue
1627 		 * to make room for the new SYN req. Also mark the time we
1628 		 * drop a SYN.
1629 		 */
1630 		tcp->tcp_last_rcv_lbolt = prom_gettime();
1631 		if (!tcp_drop_q0(tcp)) {
1632 			freemsg(mp);
1633 			BUMP_MIB(tcp_mib.tcpListenDropQ0);
1634 			dprintf("tcp_conn_request: listen half-open queue "
1635 			    "(max=%d) full (%d pending) on %s",
1636 			    tcp_conn_req_max_q0,
1637 			    tcp->tcp_conn_req_cnt_q0,
1638 			    tcp_display(tcp, NULL, DISP_PORT_ONLY));
1639 			return (NULL);
1640 		}
1641 	}
1642 
1643 	ipha = (struct ip *)mp->b_rptr;
1644 	if (IN_CLASSD(ntohl(ipha->ip_src.s_addr)) ||
1645 	    ipha->ip_src.s_addr == INADDR_BROADCAST ||
1646 	    ipha->ip_src.s_addr == INADDR_ANY ||
1647 	    ipha->ip_dst.s_addr == INADDR_BROADCAST) {
1648 		freemsg(mp);
1649 		return (NULL);
1650 	}
1651 	/*
1652 	 * We allow the connection to proceed
1653 	 * by generating a detached tcp state vector and put it in
1654 	 * the eager queue.  When an accept happens, it will be
1655 	 * dequeued sequentially.
1656 	 */
1657 	if ((eager = (tcp_t *)bkmem_alloc(sizeof (tcp_t))) == NULL) {
1658 		freemsg(mp);
1659 		errno = ENOBUFS;
1660 		return (NULL);
1661 	}
1662 	if ((errno = tcp_init_values(eager, NULL)) != 0) {
1663 		freemsg(mp);
1664 		bkmem_free((caddr_t)eager, sizeof (tcp_t));
1665 		return (NULL);
1666 	}
1667 
1668 	/*
1669 	 * Eager connection inherits address form from its listener,
1670 	 * but its packet form comes from the version of the received
1671 	 * SYN segment.
1672 	 */
1673 	eager->tcp_family = tcp->tcp_family;
1674 
1675 	err = tcp_accept_comm(tcp, eager, mp, ip_hdr_len);
1676 	if (err) {
1677 		bkmem_free((caddr_t)eager, sizeof (tcp_t));
1678 		return (NULL);
1679 	}
1680 
1681 	tcp->tcp_eager_next_q0->tcp_eager_prev_q0 = eager;
1682 	eager->tcp_eager_next_q0 = tcp->tcp_eager_next_q0;
1683 	tcp->tcp_eager_next_q0 = eager;
1684 	eager->tcp_eager_prev_q0 = tcp;
1685 
1686 	/* Set tcp_listener before adding it to tcp_conn_fanout */
1687 	eager->tcp_listener = tcp;
1688 	tcp->tcp_conn_req_cnt_q0++;
1689 
1690 	return (eager);
1691 }
1692 
1693 /*
1694  * To get around the non-interrupt problem of inetboot.
1695  * Keep on processing packets until a certain state is reached or the
1696  * TCP is destroyed because of getting a RST packet.
1697  */
1698 static int
1699 tcp_state_wait(int sock_id, tcp_t *tcp, int state)
1700 {
1701 	int i;
1702 	struct inetgram *in_gram;
1703 	mblk_t *mp;
1704 	int timeout;
1705 	boolean_t changed = B_FALSE;
1706 
1707 	/*
1708 	 * We need to make sure that the MAC does not wait longer
1709 	 * than RTO for any packet so that TCP can do retransmission.
1710 	 * But if the MAC timeout is less than tcp_rto, we are fine
1711 	 * and do not need to change it.
1712 	 */
1713 	timeout = sockets[sock_id].in_timeout;
1714 	if (timeout > tcp->tcp_rto) {
1715 		sockets[sock_id].in_timeout = tcp->tcp_rto;
1716 		changed = B_TRUE;
1717 	}
1718 retry:
1719 	if (sockets[sock_id].inq == NULL) {
1720 		/* Go out and check the wire */
1721 		for (i = MEDIA_LVL; i < TRANSPORT_LVL; i++) {
1722 			if (sockets[sock_id].input[i] != NULL) {
1723 				if (sockets[sock_id].input[i](sock_id) < 0) {
1724 					if (changed) {
1725 						sockets[sock_id].in_timeout =
1726 						    timeout;
1727 					}
1728 					return (-1);
1729 				}
1730 			}
1731 		}
1732 	}
1733 
1734 	while ((in_gram = sockets[sock_id].inq) != NULL) {
1735 		if (tcp->tcp_state == state)
1736 			break;
1737 
1738 		/* Remove unknown inetgrams from the head of inq. */
1739 		if (in_gram->igm_level != TRANSPORT_LVL) {
1740 #ifdef DEBUG
1741 			printf("tcp_state_wait for state %d: unexpected "
1742 			    "packet level %d frame found\n", state,
1743 			    in_gram->igm_level);
1744 #endif
1745 			del_gram(&sockets[sock_id].inq, in_gram, B_TRUE);
1746 			continue;
1747 		}
1748 		mp = in_gram->igm_mp;
1749 		del_gram(&sockets[sock_id].inq, in_gram, B_FALSE);
1750 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
1751 		tcp_rput_data(tcp, mp, sock_id);
1752 
1753 		/*
1754 		 * The other side may have closed this connection or
1755 		 * RST us.  But we need to continue to process other
1756 		 * packets in the socket's queue because they may be
1757 		 * belong to another TCP connections.
1758 		 */
1759 		if (sockets[sock_id].pcb == NULL) {
1760 			tcp = NULL;
1761 		}
1762 	}
1763 
1764 	/* If the other side has closed the connection, just return. */
1765 	if (tcp == NULL || sockets[sock_id].pcb == NULL) {
1766 #ifdef DEBUG
1767 		printf("tcp_state_wait other side dead: state %d "
1768 		    "error %d\n", state, sockets[sock_id].so_error);
1769 #endif
1770 		if (sockets[sock_id].so_error != 0)
1771 			return (-1);
1772 		else
1773 			return (0);
1774 	}
1775 	/*
1776 	 * TCPS_ALL_ACKED is not a valid TCP state, it is just used as an
1777 	 * indicator to tcp_state_wait to mean that it is being called
1778 	 * to wait till we have received acks for all the new segments sent.
1779 	 */
1780 	if ((state == TCPS_ALL_ACKED) && (tcp->tcp_suna == tcp->tcp_snxt)) {
1781 		goto done;
1782 	}
1783 	if (tcp->tcp_state != state) {
1784 		if (prom_gettime() > tcp->tcp_rto_timeout)
1785 			tcp_timer(tcp, sock_id);
1786 		goto retry;
1787 	}
1788 done:
1789 	if (changed)
1790 		sockets[sock_id].in_timeout = timeout;
1791 
1792 	tcp_drain_needed(sock_id, tcp);
1793 	return (0);
1794 }
1795 
1796 /* Verify the checksum of a segment. */
1797 static int
1798 tcp_verify_cksum(mblk_t *mp)
1799 {
1800 	struct ip *iph;
1801 	tcpha_t *tcph;
1802 	int len;
1803 	uint16_t old_sum;
1804 
1805 	iph = (struct ip *)mp->b_rptr;
1806 	tcph = (tcpha_t *)(iph + 1);
1807 	len = ntohs(iph->ip_len);
1808 
1809 	/*
1810 	 * Calculate the TCP checksum.  Need to include the psuedo header,
1811 	 * which is similar to the real IP header starting at the TTL field.
1812 	 */
1813 	iph->ip_sum = htons(len - IP_SIMPLE_HDR_LENGTH);
1814 	old_sum = tcph->tha_sum;
1815 	tcph->tha_sum = 0;
1816 	iph->ip_ttl = 0;
1817 	if (old_sum == tcp_cksum((uint16_t *)&(iph->ip_ttl),
1818 	    len - IP_SIMPLE_HDR_LENGTH + 12)) {
1819 		return (0);
1820 	} else {
1821 		tcp_cksum_errors++;
1822 		return (-1);
1823 	}
1824 }
1825 
1826 /* To find a TCP connection matching the incoming segment. */
1827 static tcp_t *
1828 tcp_lookup_ipv4(struct ip *iph, tcpha_t *tcph, int min_state, int *sock_id)
1829 {
1830 	int i;
1831 	tcp_t *tcp;
1832 
1833 	for (i = 0; i < MAXSOCKET; i++) {
1834 		if (sockets[i].type == INETBOOT_STREAM &&
1835 		    (tcp = (tcp_t *)sockets[i].pcb) != NULL) {
1836 			if (tcph->tha_lport == tcp->tcp_fport &&
1837 			    tcph->tha_fport == tcp->tcp_lport &&
1838 			    iph->ip_src.s_addr == tcp->tcp_remote &&
1839 			    iph->ip_dst.s_addr == tcp->tcp_bound_source &&
1840 			    tcp->tcp_state >= min_state) {
1841 				*sock_id = i;
1842 				return (tcp);
1843 			}
1844 		}
1845 	}
1846 	/* Find it in the time wait list. */
1847 	for (tcp = tcp_time_wait_head; tcp != NULL;
1848 	    tcp = tcp->tcp_time_wait_next) {
1849 		if (tcph->tha_lport == tcp->tcp_fport &&
1850 		    tcph->tha_fport == tcp->tcp_lport &&
1851 		    iph->ip_src.s_addr == tcp->tcp_remote &&
1852 		    iph->ip_dst.s_addr == tcp->tcp_bound_source &&
1853 		    tcp->tcp_state >= min_state) {
1854 			*sock_id = -1;
1855 			return (tcp);
1856 		}
1857 	}
1858 	return (NULL);
1859 }
1860 
1861 /* To find a TCP listening connection matching the incoming segment. */
1862 static tcp_t *
1863 tcp_lookup_listener_ipv4(in_addr_t addr, in_port_t port, int *sock_id)
1864 {
1865 	int i;
1866 	tcp_t *tcp;
1867 
1868 	for (i = 0; i < MAXSOCKET; i++) {
1869 		if (sockets[i].type == INETBOOT_STREAM &&
1870 		    (tcp = (tcp_t *)sockets[i].pcb) != NULL) {
1871 			if (tcp->tcp_lport == port &&
1872 			    (tcp->tcp_bound_source == addr ||
1873 			    tcp->tcp_bound_source == INADDR_ANY)) {
1874 				*sock_id = i;
1875 				return (tcp);
1876 			}
1877 		}
1878 	}
1879 
1880 	return (NULL);
1881 }
1882 
1883 /* To find a TCP eager matching the incoming segment. */
1884 static tcp_t *
1885 tcp_lookup_eager_ipv4(tcp_t *listener, struct ip *iph, tcpha_t *tcph)
1886 {
1887 	tcp_t *tcp;
1888 
1889 #ifdef DEBUG
1890 	printf("tcp_lookup_eager_ipv4 ###############\n");
1891 #endif
1892 	for (tcp = listener->tcp_eager_next_q; tcp != NULL;
1893 	    tcp = tcp->tcp_eager_next_q) {
1894 		if (tcph->tha_lport == tcp->tcp_fport &&
1895 		    tcph->tha_fport == tcp->tcp_lport &&
1896 		    iph->ip_src.s_addr == tcp->tcp_remote &&
1897 		    iph->ip_dst.s_addr == tcp->tcp_bound_source) {
1898 			return (tcp);
1899 		}
1900 	}
1901 
1902 	for (tcp = listener->tcp_eager_next_q0; tcp != listener;
1903 	    tcp = tcp->tcp_eager_next_q0) {
1904 		if (tcph->tha_lport == tcp->tcp_fport &&
1905 		    tcph->tha_fport == tcp->tcp_lport &&
1906 		    iph->ip_src.s_addr == tcp->tcp_remote &&
1907 		    iph->ip_dst.s_addr == tcp->tcp_bound_source) {
1908 			return (tcp);
1909 		}
1910 	}
1911 #ifdef DEBUG
1912 	printf("No eager found\n");
1913 #endif
1914 	return (NULL);
1915 }
1916 
1917 /* To destroy a TCP control block. */
1918 static void
1919 tcp_clean_death(int sock_id, tcp_t *tcp, int err)
1920 {
1921 	tcp_free(tcp);
1922 	if (tcp->tcp_state == TCPS_TIME_WAIT)
1923 		tcp_time_wait_remove(tcp);
1924 
1925 	if (sock_id >= 0) {
1926 		sockets[sock_id].pcb = NULL;
1927 		if (err != 0)
1928 			sockets[sock_id].so_error = err;
1929 	}
1930 	bkmem_free((caddr_t)tcp, sizeof (tcp_t));
1931 }
1932 
1933 /*
1934  * tcp_rwnd_set() is called to adjust the receive window to a desired value.
1935  * We do not allow the receive window to shrink.  After setting rwnd,
1936  * set the flow control hiwat of the stream.
1937  *
1938  * This function is called in 2 cases:
1939  *
1940  * 1) Before data transfer begins, in tcp_accept_comm() for accepting a
1941  *    connection (passive open) and in tcp_rput_data() for active connect.
1942  *    This is called after tcp_mss_set() when the desired MSS value is known.
1943  *    This makes sure that our window size is a mutiple of the other side's
1944  *    MSS.
1945  * 2) Handling SO_RCVBUF option.
1946  *
1947  * It is ASSUMED that the requested size is a multiple of the current MSS.
1948  *
1949  * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
1950  * user requests so.
1951  */
1952 static int
1953 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
1954 {
1955 	uint32_t	mss = tcp->tcp_mss;
1956 	uint32_t	old_max_rwnd;
1957 	uint32_t	max_transmittable_rwnd;
1958 
1959 	if (tcp->tcp_rwnd_max != 0)
1960 		old_max_rwnd = tcp->tcp_rwnd_max;
1961 	else
1962 		old_max_rwnd = tcp->tcp_rwnd;
1963 
1964 	/*
1965 	 * Insist on a receive window that is at least
1966 	 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
1967 	 * funny TCP interactions of Nagle algorithm, SWS avoidance
1968 	 * and delayed acknowledgement.
1969 	 */
1970 	rwnd = MAX(rwnd, tcp_recv_hiwat_minmss * mss);
1971 
1972 	/*
1973 	 * If window size info has already been exchanged, TCP should not
1974 	 * shrink the window.  Shrinking window is doable if done carefully.
1975 	 * We may add that support later.  But so far there is not a real
1976 	 * need to do that.
1977 	 */
1978 	if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
1979 		/* MSS may have changed, do a round up again. */
1980 		rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
1981 	}
1982 
1983 	/*
1984 	 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
1985 	 * can be applied even before the window scale option is decided.
1986 	 */
1987 	max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
1988 	if (rwnd > max_transmittable_rwnd) {
1989 		rwnd = max_transmittable_rwnd -
1990 		    (max_transmittable_rwnd % mss);
1991 		if (rwnd < mss)
1992 			rwnd = max_transmittable_rwnd;
1993 		/*
1994 		 * If we're over the limit we may have to back down tcp_rwnd.
1995 		 * The increment below won't work for us. So we set all three
1996 		 * here and the increment below will have no effect.
1997 		 */
1998 		tcp->tcp_rwnd = old_max_rwnd = rwnd;
1999 	}
2000 
2001 	/*
2002 	 * Increment the current rwnd by the amount the maximum grew (we
2003 	 * can not overwrite it since we might be in the middle of a
2004 	 * connection.)
2005 	 */
2006 	tcp->tcp_rwnd += rwnd - old_max_rwnd;
2007 	U32_TO_ABE16(tcp->tcp_rwnd >> tcp->tcp_rcv_ws, tcp->tcp_tcph->th_win);
2008 	if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
2009 		tcp->tcp_cwnd_max = rwnd;
2010 	tcp->tcp_rwnd_max = rwnd;
2011 
2012 	return (rwnd);
2013 }
2014 
2015 /*
2016  * Extract option values from a tcp header.  We put any found values into the
2017  * tcpopt struct and return a bitmask saying which options were found.
2018  */
2019 static int
2020 tcp_parse_options(tcph_t *tcph, tcp_opt_t *tcpopt)
2021 {
2022 	uchar_t		*endp;
2023 	int		len;
2024 	uint32_t	mss;
2025 	uchar_t		*up = (uchar_t *)tcph;
2026 	int		found = 0;
2027 	int32_t		sack_len;
2028 	tcp_seq		sack_begin, sack_end;
2029 	tcp_t		*tcp;
2030 
2031 	endp = up + TCP_HDR_LENGTH(tcph);
2032 	up += TCP_MIN_HEADER_LENGTH;
2033 	while (up < endp) {
2034 		len = endp - up;
2035 		switch (*up) {
2036 		case TCPOPT_EOL:
2037 			break;
2038 
2039 		case TCPOPT_NOP:
2040 			up++;
2041 			continue;
2042 
2043 		case TCPOPT_MAXSEG:
2044 			if (len < TCPOPT_MAXSEG_LEN ||
2045 			    up[1] != TCPOPT_MAXSEG_LEN)
2046 				break;
2047 
2048 			mss = BE16_TO_U16(up+2);
2049 			/* Caller must handle tcp_mss_min and tcp_mss_max_* */
2050 			tcpopt->tcp_opt_mss = mss;
2051 			found |= TCP_OPT_MSS_PRESENT;
2052 
2053 			up += TCPOPT_MAXSEG_LEN;
2054 			continue;
2055 
2056 		case TCPOPT_WSCALE:
2057 			if (len < TCPOPT_WS_LEN || up[1] != TCPOPT_WS_LEN)
2058 				break;
2059 
2060 			if (up[2] > TCP_MAX_WINSHIFT)
2061 				tcpopt->tcp_opt_wscale = TCP_MAX_WINSHIFT;
2062 			else
2063 				tcpopt->tcp_opt_wscale = up[2];
2064 			found |= TCP_OPT_WSCALE_PRESENT;
2065 
2066 			up += TCPOPT_WS_LEN;
2067 			continue;
2068 
2069 		case TCPOPT_SACK_PERMITTED:
2070 			if (len < TCPOPT_SACK_OK_LEN ||
2071 			    up[1] != TCPOPT_SACK_OK_LEN)
2072 				break;
2073 			found |= TCP_OPT_SACK_OK_PRESENT;
2074 			up += TCPOPT_SACK_OK_LEN;
2075 			continue;
2076 
2077 		case TCPOPT_SACK:
2078 			if (len <= 2 || up[1] <= 2 || len < up[1])
2079 				break;
2080 
2081 			/* If TCP is not interested in SACK blks... */
2082 			if ((tcp = tcpopt->tcp) == NULL) {
2083 				up += up[1];
2084 				continue;
2085 			}
2086 			sack_len = up[1] - TCPOPT_HEADER_LEN;
2087 			up += TCPOPT_HEADER_LEN;
2088 
2089 			/*
2090 			 * If the list is empty, allocate one and assume
2091 			 * nothing is sack'ed.
2092 			 */
2093 			assert(tcp->tcp_sack_info != NULL);
2094 			if (tcp->tcp_notsack_list == NULL) {
2095 				tcp_notsack_update(&(tcp->tcp_notsack_list),
2096 				    tcp->tcp_suna, tcp->tcp_snxt,
2097 				    &(tcp->tcp_num_notsack_blk),
2098 				    &(tcp->tcp_cnt_notsack_list));
2099 
2100 				/*
2101 				 * Make sure tcp_notsack_list is not NULL.
2102 				 * This happens when kmem_alloc(KM_NOSLEEP)
2103 				 * returns NULL.
2104 				 */
2105 				if (tcp->tcp_notsack_list == NULL) {
2106 					up += sack_len;
2107 					continue;
2108 				}
2109 				tcp->tcp_fack = tcp->tcp_suna;
2110 			}
2111 
2112 			while (sack_len > 0) {
2113 				if (up + 8 > endp) {
2114 					up = endp;
2115 					break;
2116 				}
2117 				sack_begin = BE32_TO_U32(up);
2118 				up += 4;
2119 				sack_end = BE32_TO_U32(up);
2120 				up += 4;
2121 				sack_len -= 8;
2122 				/*
2123 				 * Bounds checking.  Make sure the SACK
2124 				 * info is within tcp_suna and tcp_snxt.
2125 				 * If this SACK blk is out of bound, ignore
2126 				 * it but continue to parse the following
2127 				 * blks.
2128 				 */
2129 				if (SEQ_LEQ(sack_end, sack_begin) ||
2130 				    SEQ_LT(sack_begin, tcp->tcp_suna) ||
2131 				    SEQ_GT(sack_end, tcp->tcp_snxt)) {
2132 					continue;
2133 				}
2134 				tcp_notsack_insert(&(tcp->tcp_notsack_list),
2135 				    sack_begin, sack_end,
2136 				    &(tcp->tcp_num_notsack_blk),
2137 				    &(tcp->tcp_cnt_notsack_list));
2138 				if (SEQ_GT(sack_end, tcp->tcp_fack)) {
2139 					tcp->tcp_fack = sack_end;
2140 				}
2141 			}
2142 			found |= TCP_OPT_SACK_PRESENT;
2143 			continue;
2144 
2145 		case TCPOPT_TSTAMP:
2146 			if (len < TCPOPT_TSTAMP_LEN ||
2147 			    up[1] != TCPOPT_TSTAMP_LEN)
2148 				break;
2149 
2150 			tcpopt->tcp_opt_ts_val = BE32_TO_U32(up+2);
2151 			tcpopt->tcp_opt_ts_ecr = BE32_TO_U32(up+6);
2152 
2153 			found |= TCP_OPT_TSTAMP_PRESENT;
2154 
2155 			up += TCPOPT_TSTAMP_LEN;
2156 			continue;
2157 
2158 		default:
2159 			if (len <= 1 || len < (int)up[1] || up[1] == 0)
2160 				break;
2161 			up += up[1];
2162 			continue;
2163 		}
2164 		break;
2165 	}
2166 	return (found);
2167 }
2168 
2169 /*
2170  * Set the mss associated with a particular tcp based on its current value,
2171  * and a new one passed in. Observe minimums and maximums, and reset
2172  * other state variables that we want to view as multiples of mss.
2173  *
2174  * This function is called in various places mainly because
2175  * 1) Various stuffs, tcp_mss, tcp_cwnd, ... need to be adjusted when the
2176  *    other side's SYN/SYN-ACK packet arrives.
2177  * 2) PMTUd may get us a new MSS.
2178  * 3) If the other side stops sending us timestamp option, we need to
2179  *    increase the MSS size to use the extra bytes available.
2180  */
2181 static void
2182 tcp_mss_set(tcp_t *tcp, uint32_t mss)
2183 {
2184 	uint32_t	mss_max;
2185 
2186 	mss_max = tcp_mss_max_ipv4;
2187 
2188 	if (mss < tcp_mss_min)
2189 		mss = tcp_mss_min;
2190 	if (mss > mss_max)
2191 		mss = mss_max;
2192 	/*
2193 	 * Unless naglim has been set by our client to
2194 	 * a non-mss value, force naglim to track mss.
2195 	 * This can help to aggregate small writes.
2196 	 */
2197 	if (mss < tcp->tcp_naglim || tcp->tcp_mss == tcp->tcp_naglim)
2198 		tcp->tcp_naglim = mss;
2199 	/*
2200 	 * TCP should be able to buffer at least 4 MSS data for obvious
2201 	 * performance reason.
2202 	 */
2203 	if ((mss << 2) > tcp->tcp_xmit_hiwater)
2204 		tcp->tcp_xmit_hiwater = mss << 2;
2205 	tcp->tcp_mss = mss;
2206 	/*
2207 	 * Initialize cwnd according to draft-floyd-incr-init-win-01.txt.
2208 	 * Previously, we use tcp_slow_start_initial to control the size
2209 	 * of the initial cwnd.  Now, when tcp_slow_start_initial * mss
2210 	 * is smaller than the cwnd calculated from the formula suggested in
2211 	 * the draft, we use tcp_slow_start_initial * mss as the cwnd.
2212 	 * Otherwise, use the cwnd from the draft's formula.  The default
2213 	 * of tcp_slow_start_initial is 2.
2214 	 */
2215 	tcp->tcp_cwnd = MIN(tcp_slow_start_initial * mss,
2216 	    MIN(4 * mss, MAX(2 * mss, 4380 / mss * mss)));
2217 	tcp->tcp_cwnd_cnt = 0;
2218 }
2219 
2220 /*
2221  * Process all TCP option in SYN segment.
2222  *
2223  * This function sets up the correct tcp_mss value according to the
2224  * MSS option value and our header size.  It also sets up the window scale
2225  * and timestamp values, and initialize SACK info blocks.  But it does not
2226  * change receive window size after setting the tcp_mss value.  The caller
2227  * should do the appropriate change.
2228  */
2229 void
2230 tcp_process_options(tcp_t *tcp, tcph_t *tcph)
2231 {
2232 	int options;
2233 	tcp_opt_t tcpopt;
2234 	uint32_t mss_max;
2235 	char *tmp_tcph;
2236 
2237 	tcpopt.tcp = NULL;
2238 	options = tcp_parse_options(tcph, &tcpopt);
2239 
2240 	/*
2241 	 * Process MSS option.  Note that MSS option value does not account
2242 	 * for IP or TCP options.  This means that it is equal to MTU - minimum
2243 	 * IP+TCP header size, which is 40 bytes for IPv4 and 60 bytes for
2244 	 * IPv6.
2245 	 */
2246 	if (!(options & TCP_OPT_MSS_PRESENT)) {
2247 		tcpopt.tcp_opt_mss = tcp_mss_def_ipv4;
2248 	} else {
2249 		if (tcp->tcp_ipversion == IPV4_VERSION)
2250 			mss_max = tcp_mss_max_ipv4;
2251 		if (tcpopt.tcp_opt_mss < tcp_mss_min)
2252 			tcpopt.tcp_opt_mss = tcp_mss_min;
2253 		else if (tcpopt.tcp_opt_mss > mss_max)
2254 			tcpopt.tcp_opt_mss = mss_max;
2255 	}
2256 
2257 	/* Process Window Scale option. */
2258 	if (options & TCP_OPT_WSCALE_PRESENT) {
2259 		tcp->tcp_snd_ws = tcpopt.tcp_opt_wscale;
2260 		tcp->tcp_snd_ws_ok = B_TRUE;
2261 	} else {
2262 		tcp->tcp_snd_ws = B_FALSE;
2263 		tcp->tcp_snd_ws_ok = B_FALSE;
2264 		tcp->tcp_rcv_ws = B_FALSE;
2265 	}
2266 
2267 	/* Process Timestamp option. */
2268 	if ((options & TCP_OPT_TSTAMP_PRESENT) &&
2269 	    (tcp->tcp_snd_ts_ok || !tcp->tcp_active_open)) {
2270 		tmp_tcph = (char *)tcp->tcp_tcph;
2271 
2272 		tcp->tcp_snd_ts_ok = B_TRUE;
2273 		tcp->tcp_ts_recent = tcpopt.tcp_opt_ts_val;
2274 		tcp->tcp_last_rcv_lbolt = prom_gettime();
2275 		assert(OK_32PTR(tmp_tcph));
2276 		assert(tcp->tcp_tcp_hdr_len == TCP_MIN_HEADER_LENGTH);
2277 
2278 		/* Fill in our template header with basic timestamp option. */
2279 		tmp_tcph += tcp->tcp_tcp_hdr_len;
2280 		tmp_tcph[0] = TCPOPT_NOP;
2281 		tmp_tcph[1] = TCPOPT_NOP;
2282 		tmp_tcph[2] = TCPOPT_TSTAMP;
2283 		tmp_tcph[3] = TCPOPT_TSTAMP_LEN;
2284 		tcp->tcp_hdr_len += TCPOPT_REAL_TS_LEN;
2285 		tcp->tcp_tcp_hdr_len += TCPOPT_REAL_TS_LEN;
2286 		tcp->tcp_tcph->th_offset_and_rsrvd[0] += (3 << 4);
2287 	} else {
2288 		tcp->tcp_snd_ts_ok = B_FALSE;
2289 	}
2290 
2291 	/*
2292 	 * Process SACK options.  If SACK is enabled for this connection,
2293 	 * then allocate the SACK info structure.
2294 	 */
2295 	if ((options & TCP_OPT_SACK_OK_PRESENT) &&
2296 	    (tcp->tcp_snd_sack_ok ||
2297 	    (tcp_sack_permitted != 0 && !tcp->tcp_active_open))) {
2298 		/* This should be true only in the passive case. */
2299 		if (tcp->tcp_sack_info == NULL) {
2300 			tcp->tcp_sack_info = (tcp_sack_info_t *)bkmem_zalloc(
2301 			    sizeof (tcp_sack_info_t));
2302 		}
2303 		if (tcp->tcp_sack_info == NULL) {
2304 			tcp->tcp_snd_sack_ok = B_FALSE;
2305 		} else {
2306 			tcp->tcp_snd_sack_ok = B_TRUE;
2307 			if (tcp->tcp_snd_ts_ok) {
2308 				tcp->tcp_max_sack_blk = 3;
2309 			} else {
2310 				tcp->tcp_max_sack_blk = 4;
2311 			}
2312 		}
2313 	} else {
2314 		/*
2315 		 * Resetting tcp_snd_sack_ok to B_FALSE so that
2316 		 * no SACK info will be used for this
2317 		 * connection.  This assumes that SACK usage
2318 		 * permission is negotiated.  This may need
2319 		 * to be changed once this is clarified.
2320 		 */
2321 		if (tcp->tcp_sack_info != NULL) {
2322 			bkmem_free((caddr_t)tcp->tcp_sack_info,
2323 			    sizeof (tcp_sack_info_t));
2324 			tcp->tcp_sack_info = NULL;
2325 		}
2326 		tcp->tcp_snd_sack_ok = B_FALSE;
2327 	}
2328 
2329 	/*
2330 	 * Now we know the exact TCP/IP header length, subtract
2331 	 * that from tcp_mss to get our side's MSS.
2332 	 */
2333 	tcp->tcp_mss -= tcp->tcp_hdr_len;
2334 	/*
2335 	 * Here we assume that the other side's header size will be equal to
2336 	 * our header size.  We calculate the real MSS accordingly.  Need to
2337 	 * take into additional stuffs IPsec puts in.
2338 	 *
2339 	 * Real MSS = Opt.MSS - (our TCP/IP header - min TCP/IP header)
2340 	 */
2341 	tcpopt.tcp_opt_mss -= tcp->tcp_hdr_len -
2342 	    (IP_SIMPLE_HDR_LENGTH + TCP_MIN_HEADER_LENGTH);
2343 
2344 	/*
2345 	 * Set MSS to the smaller one of both ends of the connection.
2346 	 * We should not have called tcp_mss_set() before, but our
2347 	 * side of the MSS should have been set to a proper value
2348 	 * by tcp_adapt_ire().  tcp_mss_set() will also set up the
2349 	 * STREAM head parameters properly.
2350 	 *
2351 	 * If we have a larger-than-16-bit window but the other side
2352 	 * didn't want to do window scale, tcp_rwnd_set() will take
2353 	 * care of that.
2354 	 */
2355 	tcp_mss_set(tcp, MIN(tcpopt.tcp_opt_mss, tcp->tcp_mss));
2356 }
2357 
2358 /*
2359  * This function does PAWS protection check.  Returns B_TRUE if the
2360  * segment passes the PAWS test, else returns B_FALSE.
2361  */
2362 boolean_t
2363 tcp_paws_check(tcp_t *tcp, tcph_t *tcph, tcp_opt_t *tcpoptp)
2364 {
2365 	uint8_t	flags;
2366 	int	options;
2367 	uint8_t *up;
2368 
2369 	flags = (unsigned int)tcph->th_flags[0] & 0xFF;
2370 	/*
2371 	 * If timestamp option is aligned nicely, get values inline,
2372 	 * otherwise call general routine to parse.  Only do that
2373 	 * if timestamp is the only option.
2374 	 */
2375 	if (TCP_HDR_LENGTH(tcph) == (uint32_t)TCP_MIN_HEADER_LENGTH +
2376 	    TCPOPT_REAL_TS_LEN &&
2377 	    OK_32PTR((up = ((uint8_t *)tcph) +
2378 	    TCP_MIN_HEADER_LENGTH)) &&
2379 	    *(uint32_t *)up == TCPOPT_NOP_NOP_TSTAMP) {
2380 		tcpoptp->tcp_opt_ts_val = ABE32_TO_U32((up+4));
2381 		tcpoptp->tcp_opt_ts_ecr = ABE32_TO_U32((up+8));
2382 
2383 		options = TCP_OPT_TSTAMP_PRESENT;
2384 	} else {
2385 		if (tcp->tcp_snd_sack_ok) {
2386 			tcpoptp->tcp = tcp;
2387 		} else {
2388 			tcpoptp->tcp = NULL;
2389 		}
2390 		options = tcp_parse_options(tcph, tcpoptp);
2391 	}
2392 
2393 	if (options & TCP_OPT_TSTAMP_PRESENT) {
2394 		/*
2395 		 * Do PAWS per RFC 1323 section 4.2.  Accept RST
2396 		 * regardless of the timestamp, page 18 RFC 1323.bis.
2397 		 */
2398 		if ((flags & TH_RST) == 0 &&
2399 		    TSTMP_LT(tcpoptp->tcp_opt_ts_val,
2400 		    tcp->tcp_ts_recent)) {
2401 			if (TSTMP_LT(prom_gettime(),
2402 			    tcp->tcp_last_rcv_lbolt + PAWS_TIMEOUT)) {
2403 				/* This segment is not acceptable. */
2404 				return (B_FALSE);
2405 			} else {
2406 				/*
2407 				 * Connection has been idle for
2408 				 * too long.  Reset the timestamp
2409 				 * and assume the segment is valid.
2410 				 */
2411 				tcp->tcp_ts_recent =
2412 				    tcpoptp->tcp_opt_ts_val;
2413 			}
2414 		}
2415 	} else {
2416 		/*
2417 		 * If we don't get a timestamp on every packet, we
2418 		 * figure we can't really trust 'em, so we stop sending
2419 		 * and parsing them.
2420 		 */
2421 		tcp->tcp_snd_ts_ok = B_FALSE;
2422 
2423 		tcp->tcp_hdr_len -= TCPOPT_REAL_TS_LEN;
2424 		tcp->tcp_tcp_hdr_len -= TCPOPT_REAL_TS_LEN;
2425 		tcp->tcp_tcph->th_offset_and_rsrvd[0] -= (3 << 4);
2426 		tcp_mss_set(tcp, tcp->tcp_mss + TCPOPT_REAL_TS_LEN);
2427 		if (tcp->tcp_snd_sack_ok) {
2428 			assert(tcp->tcp_sack_info != NULL);
2429 			tcp->tcp_max_sack_blk = 4;
2430 		}
2431 	}
2432 	return (B_TRUE);
2433 }
2434 
2435 /*
2436  * tcp_get_seg_mp() is called to get the pointer to a segment in the
2437  * send queue which starts at the given seq. no.
2438  *
2439  * Parameters:
2440  *	tcp_t *tcp: the tcp instance pointer.
2441  *	uint32_t seq: the starting seq. no of the requested segment.
2442  *	int32_t *off: after the execution, *off will be the offset to
2443  *		the returned mblk which points to the requested seq no.
2444  *
2445  * Return:
2446  *	A mblk_t pointer pointing to the requested segment in send queue.
2447  */
2448 static mblk_t *
2449 tcp_get_seg_mp(tcp_t *tcp, uint32_t seq, int32_t *off)
2450 {
2451 	int32_t	cnt;
2452 	mblk_t	*mp;
2453 
2454 	/* Defensive coding.  Make sure we don't send incorrect data. */
2455 	if (SEQ_LT(seq, tcp->tcp_suna) || SEQ_GEQ(seq, tcp->tcp_snxt) ||
2456 	    off == NULL) {
2457 		return (NULL);
2458 	}
2459 	cnt = seq - tcp->tcp_suna;
2460 	mp = tcp->tcp_xmit_head;
2461 	while (cnt > 0 && mp) {
2462 		cnt -= mp->b_wptr - mp->b_rptr;
2463 		if (cnt < 0) {
2464 			cnt += mp->b_wptr - mp->b_rptr;
2465 			break;
2466 		}
2467 		mp = mp->b_cont;
2468 	}
2469 	assert(mp != NULL);
2470 	*off = cnt;
2471 	return (mp);
2472 }
2473 
2474 /*
2475  * This function handles all retransmissions if SACK is enabled for this
2476  * connection.  First it calculates how many segments can be retransmitted
2477  * based on tcp_pipe.  Then it goes thru the notsack list to find eligible
2478  * segments.  A segment is eligible if sack_cnt for that segment is greater
2479  * than or equal tcp_dupack_fast_retransmit.  After it has retransmitted
2480  * all eligible segments, it checks to see if TCP can send some new segments
2481  * (fast recovery).  If it can, it returns 1.  Otherwise it returns 0.
2482  *
2483  * Parameters:
2484  *	tcp_t *tcp: the tcp structure of the connection.
2485  *
2486  * Return:
2487  *	1 if the pipe is not full (new data can be sent), 0 otherwise
2488  */
2489 static int32_t
2490 tcp_sack_rxmit(tcp_t *tcp, int sock_id)
2491 {
2492 	notsack_blk_t	*notsack_blk;
2493 	int32_t		usable_swnd;
2494 	int32_t		mss;
2495 	uint32_t	seg_len;
2496 	mblk_t		*xmit_mp;
2497 
2498 	assert(tcp->tcp_sack_info != NULL);
2499 	assert(tcp->tcp_notsack_list != NULL);
2500 	assert(tcp->tcp_rexmit == B_FALSE);
2501 
2502 	/* Defensive coding in case there is a bug... */
2503 	if (tcp->tcp_notsack_list == NULL) {
2504 		return (0);
2505 	}
2506 	notsack_blk = tcp->tcp_notsack_list;
2507 	mss = tcp->tcp_mss;
2508 
2509 	/*
2510 	 * Limit the num of outstanding data in the network to be
2511 	 * tcp_cwnd_ssthresh, which is half of the original congestion wnd.
2512 	 */
2513 	usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
2514 
2515 	/* At least retransmit 1 MSS of data. */
2516 	if (usable_swnd <= 0) {
2517 		usable_swnd = mss;
2518 	}
2519 
2520 	/* Make sure no new RTT samples will be taken. */
2521 	tcp->tcp_csuna = tcp->tcp_snxt;
2522 
2523 	notsack_blk = tcp->tcp_notsack_list;
2524 	while (usable_swnd > 0) {
2525 		mblk_t		*snxt_mp, *tmp_mp;
2526 		tcp_seq		begin = tcp->tcp_sack_snxt;
2527 		tcp_seq		end;
2528 		int32_t		off;
2529 
2530 		for (; notsack_blk != NULL; notsack_blk = notsack_blk->next) {
2531 			if (SEQ_GT(notsack_blk->end, begin) &&
2532 			    (notsack_blk->sack_cnt >=
2533 			    tcp_dupack_fast_retransmit)) {
2534 				end = notsack_blk->end;
2535 				if (SEQ_LT(begin, notsack_blk->begin)) {
2536 					begin = notsack_blk->begin;
2537 				}
2538 				break;
2539 			}
2540 		}
2541 		/*
2542 		 * All holes are filled.  Manipulate tcp_cwnd to send more
2543 		 * if we can.  Note that after the SACK recovery, tcp_cwnd is
2544 		 * set to tcp_cwnd_ssthresh.
2545 		 */
2546 		if (notsack_blk == NULL) {
2547 			usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
2548 			if (usable_swnd <= 0) {
2549 				tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna;
2550 				assert(tcp->tcp_cwnd > 0);
2551 				return (0);
2552 			} else {
2553 				usable_swnd = usable_swnd / mss;
2554 				tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna +
2555 				    MAX(usable_swnd * mss, mss);
2556 				return (1);
2557 			}
2558 		}
2559 
2560 		/*
2561 		 * Note that we may send more than usable_swnd allows here
2562 		 * because of round off, but no more than 1 MSS of data.
2563 		 */
2564 		seg_len = end - begin;
2565 		if (seg_len > mss)
2566 			seg_len = mss;
2567 		snxt_mp = tcp_get_seg_mp(tcp, begin, &off);
2568 		assert(snxt_mp != NULL);
2569 		/* This should not happen.  Defensive coding again... */
2570 		if (snxt_mp == NULL) {
2571 			return (0);
2572 		}
2573 
2574 		xmit_mp = tcp_xmit_mp(tcp, snxt_mp, seg_len, &off,
2575 		    &tmp_mp, begin, B_TRUE, &seg_len, B_TRUE);
2576 
2577 		if (xmit_mp == NULL)
2578 			return (0);
2579 
2580 		usable_swnd -= seg_len;
2581 		tcp->tcp_pipe += seg_len;
2582 		tcp->tcp_sack_snxt = begin + seg_len;
2583 		TCP_DUMP_PACKET("tcp_sack_rxmit", xmit_mp);
2584 		(void) ipv4_tcp_output(sock_id, xmit_mp);
2585 		freeb(xmit_mp);
2586 
2587 		/*
2588 		 * Update the send timestamp to avoid false retransmission.
2589 		 * Note. use uintptr_t to suppress the gcc warning.
2590 		 */
2591 		snxt_mp->b_prev = (mblk_t *)(uintptr_t)prom_gettime();
2592 
2593 		BUMP_MIB(tcp_mib.tcpRetransSegs);
2594 		UPDATE_MIB(tcp_mib.tcpRetransBytes, seg_len);
2595 		BUMP_MIB(tcp_mib.tcpOutSackRetransSegs);
2596 		/*
2597 		 * Update tcp_rexmit_max to extend this SACK recovery phase.
2598 		 * This happens when new data sent during fast recovery is
2599 		 * also lost.  If TCP retransmits those new data, it needs
2600 		 * to extend SACK recover phase to avoid starting another
2601 		 * fast retransmit/recovery unnecessarily.
2602 		 */
2603 		if (SEQ_GT(tcp->tcp_sack_snxt, tcp->tcp_rexmit_max)) {
2604 			tcp->tcp_rexmit_max = tcp->tcp_sack_snxt;
2605 		}
2606 	}
2607 	return (0);
2608 }
2609 
2610 static void
2611 tcp_rput_data(tcp_t *tcp, mblk_t *mp, int sock_id)
2612 {
2613 	uchar_t		*rptr;
2614 	struct ip	*iph;
2615 	tcp_t		*tcp1;
2616 	tcpha_t		*tcph;
2617 	uint32_t	seg_ack;
2618 	int		seg_len;
2619 	uint_t		ip_hdr_len;
2620 	uint32_t	seg_seq;
2621 	mblk_t		*mp1;
2622 	uint_t		flags;
2623 	uint32_t	new_swnd = 0;
2624 	int		mss;
2625 	boolean_t	ofo_seg = B_FALSE; /* Out of order segment */
2626 	int32_t		gap;
2627 	int32_t		rgap;
2628 	tcp_opt_t	tcpopt;
2629 	int32_t		bytes_acked;
2630 	int		npkt;
2631 	uint32_t	cwnd;
2632 	uint32_t	add;
2633 
2634 #ifdef DEBUG
2635 	printf("tcp_rput_data sock %d mp %x mp_datap %x #################\n",
2636 	    sock_id, mp, mp->b_datap);
2637 #endif
2638 
2639 	/* Dump the packet when debugging. */
2640 	TCP_DUMP_PACKET("tcp_rput_data", mp);
2641 
2642 	assert(OK_32PTR(mp->b_rptr));
2643 
2644 	rptr = mp->b_rptr;
2645 	iph = (struct ip *)rptr;
2646 	ip_hdr_len = IPH_HDR_LENGTH(rptr);
2647 	if (ip_hdr_len != IP_SIMPLE_HDR_LENGTH) {
2648 #ifdef DEBUG
2649 		printf("Not simple IP header\n");
2650 #endif
2651 		/* We cannot handle IP option yet... */
2652 		tcp_drops++;
2653 		freeb(mp);
2654 		return;
2655 	}
2656 	/* The TCP header must be aligned. */
2657 	tcph = (tcpha_t *)&rptr[ip_hdr_len];
2658 	seg_seq = ntohl(tcph->tha_seq);
2659 	seg_ack = ntohl(tcph->tha_ack);
2660 	assert((uintptr_t)(mp->b_wptr - rptr) <= (uintptr_t)INT_MAX);
2661 	seg_len = (int)(mp->b_wptr - rptr) -
2662 	    (ip_hdr_len + TCP_HDR_LENGTH(((tcph_t *)tcph)));
2663 	/* In inetboot, b_cont should always be NULL. */
2664 	assert(mp->b_cont == NULL);
2665 
2666 	/* Verify the checksum. */
2667 	if (tcp_verify_cksum(mp) < 0) {
2668 #ifdef DEBUG
2669 		printf("tcp_rput_data: wrong cksum\n");
2670 #endif
2671 		freemsg(mp);
2672 		return;
2673 	}
2674 
2675 	/*
2676 	 * This segment is not for us, try to find its
2677 	 * intended receiver.
2678 	 */
2679 	if (tcp == NULL ||
2680 	    tcph->tha_lport != tcp->tcp_fport ||
2681 	    tcph->tha_fport != tcp->tcp_lport ||
2682 	    iph->ip_src.s_addr != tcp->tcp_remote ||
2683 	    iph->ip_dst.s_addr != tcp->tcp_bound_source) {
2684 #ifdef DEBUG
2685 		printf("tcp_rput_data: not for us, state %d\n",
2686 		    tcp->tcp_state);
2687 #endif
2688 		/*
2689 		 * First try to find a established connection.  If none
2690 		 * is found, look for a listener.
2691 		 *
2692 		 * If a listener is found, we need to check to see if the
2693 		 * incoming segment is for one of its eagers.  If it is,
2694 		 * give it to the eager.  If not, listener should take care
2695 		 * of it.
2696 		 */
2697 		if ((tcp1 = tcp_lookup_ipv4(iph, tcph, TCPS_SYN_SENT,
2698 		    &sock_id)) != NULL ||
2699 		    (tcp1 = tcp_lookup_listener_ipv4(iph->ip_dst.s_addr,
2700 		    tcph->tha_fport, &sock_id)) != NULL) {
2701 			if (tcp1->tcp_state == TCPS_LISTEN) {
2702 				if ((tcp = tcp_lookup_eager_ipv4(tcp1,
2703 				    iph, tcph)) == NULL) {
2704 					/* No eager... sent to listener */
2705 #ifdef DEBUG
2706 					printf("found the listener: %s\n",
2707 					    tcp_display(tcp1, NULL,
2708 					    DISP_ADDR_AND_PORT));
2709 #endif
2710 					tcp = tcp1;
2711 				}
2712 #ifdef DEBUG
2713 				else {
2714 					printf("found the eager: %s\n",
2715 					    tcp_display(tcp, NULL,
2716 					    DISP_ADDR_AND_PORT));
2717 				}
2718 #endif
2719 			} else {
2720 				/* Non listener found... */
2721 #ifdef DEBUG
2722 				printf("found the connection: %s\n",
2723 				    tcp_display(tcp1, NULL,
2724 				    DISP_ADDR_AND_PORT));
2725 #endif
2726 				tcp = tcp1;
2727 			}
2728 		} else {
2729 			/*
2730 			 * No connection for this segment...
2731 			 * Send a RST to the other side.
2732 			 */
2733 			tcp_xmit_listeners_reset(sock_id, mp, ip_hdr_len);
2734 			return;
2735 		}
2736 	}
2737 
2738 	flags = tcph->tha_flags & 0xFF;
2739 	BUMP_MIB(tcp_mib.tcpInSegs);
2740 	if (tcp->tcp_state == TCPS_TIME_WAIT) {
2741 		tcp_time_wait_processing(tcp, mp, seg_seq, seg_ack,
2742 		    seg_len, (tcph_t *)tcph, sock_id);
2743 		return;
2744 	}
2745 	/*
2746 	 * From this point we can assume that the tcp is not compressed,
2747 	 * since we would have branched off to tcp_time_wait_processing()
2748 	 * in such a case.
2749 	 */
2750 	assert(tcp != NULL && tcp->tcp_state != TCPS_TIME_WAIT);
2751 
2752 	/*
2753 	 * After this point, we know we have the correct TCP, so update
2754 	 * the receive time.
2755 	 */
2756 	tcp->tcp_last_recv_time = prom_gettime();
2757 
2758 	/* In inetboot, we do not handle urgent pointer... */
2759 	if (flags & TH_URG) {
2760 		freemsg(mp);
2761 		DEBUG_1("tcp_rput_data(%d): received segment with urgent "
2762 		    "pointer\n", sock_id);
2763 		tcp_drops++;
2764 		return;
2765 	}
2766 
2767 	switch (tcp->tcp_state) {
2768 	case TCPS_LISTEN:
2769 		if ((flags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
2770 			if (flags & TH_RST) {
2771 				freemsg(mp);
2772 				return;
2773 			}
2774 			if (flags & TH_ACK) {
2775 				tcp_xmit_early_reset("TCPS_LISTEN-TH_ACK",
2776 				    sock_id, mp, seg_ack, 0, TH_RST,
2777 				    ip_hdr_len);
2778 				return;
2779 			}
2780 			if (!(flags & TH_SYN)) {
2781 				freemsg(mp);
2782 				return;
2783 			}
2784 			printf("tcp_rput_data: %d\n", __LINE__);
2785 			prom_panic("inetboot");
2786 		}
2787 		if (tcp->tcp_conn_req_max > 0) {
2788 			tcp = tcp_conn_request(tcp, mp, sock_id, ip_hdr_len);
2789 			if (tcp == NULL) {
2790 				freemsg(mp);
2791 				return;
2792 			}
2793 #ifdef DEBUG
2794 			printf("tcp_rput_data: new tcp created\n");
2795 #endif
2796 		}
2797 		tcp->tcp_irs = seg_seq;
2798 		tcp->tcp_rack = seg_seq;
2799 		tcp->tcp_rnxt = seg_seq + 1;
2800 		U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
2801 		BUMP_MIB(tcp_mib.tcpPassiveOpens);
2802 		goto syn_rcvd;
2803 	case TCPS_SYN_SENT:
2804 		if (flags & TH_ACK) {
2805 			/*
2806 			 * Note that our stack cannot send data before a
2807 			 * connection is established, therefore the
2808 			 * following check is valid.  Otherwise, it has
2809 			 * to be changed.
2810 			 */
2811 			if (SEQ_LEQ(seg_ack, tcp->tcp_iss) ||
2812 			    SEQ_GT(seg_ack, tcp->tcp_snxt)) {
2813 				if (flags & TH_RST) {
2814 					freemsg(mp);
2815 					return;
2816 				}
2817 				tcp_xmit_ctl("TCPS_SYN_SENT-Bad_seq",
2818 				    tcp, mp, seg_ack, 0, TH_RST,
2819 				    ip_hdr_len, sock_id);
2820 				return;
2821 			}
2822 			assert(tcp->tcp_suna + 1 == seg_ack);
2823 		}
2824 		if (flags & TH_RST) {
2825 			freemsg(mp);
2826 			if (flags & TH_ACK) {
2827 				tcp_clean_death(sock_id, tcp, ECONNREFUSED);
2828 			}
2829 			return;
2830 		}
2831 		if (!(flags & TH_SYN)) {
2832 			freemsg(mp);
2833 			return;
2834 		}
2835 
2836 		/* Process all TCP options. */
2837 		tcp_process_options(tcp, (tcph_t *)tcph);
2838 		/*
2839 		 * The following changes our rwnd to be a multiple of the
2840 		 * MIN(peer MSS, our MSS) for performance reason.
2841 		 */
2842 		(void) tcp_rwnd_set(tcp, MSS_ROUNDUP(tcp->tcp_rwnd,
2843 		    tcp->tcp_mss));
2844 
2845 		/* Is the other end ECN capable? */
2846 		if (tcp->tcp_ecn_ok) {
2847 			if ((flags & (TH_ECE|TH_CWR)) != TH_ECE) {
2848 				tcp->tcp_ecn_ok = B_FALSE;
2849 			}
2850 		}
2851 		/*
2852 		 * Clear ECN flags because it may interfere with later
2853 		 * processing.
2854 		 */
2855 		flags &= ~(TH_ECE|TH_CWR);
2856 
2857 		tcp->tcp_irs = seg_seq;
2858 		tcp->tcp_rack = seg_seq;
2859 		tcp->tcp_rnxt = seg_seq + 1;
2860 		U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
2861 
2862 		if (flags & TH_ACK) {
2863 			/* One for the SYN */
2864 			tcp->tcp_suna = tcp->tcp_iss + 1;
2865 			tcp->tcp_valid_bits &= ~TCP_ISS_VALID;
2866 			tcp->tcp_state = TCPS_ESTABLISHED;
2867 
2868 			/*
2869 			 * If SYN was retransmitted, need to reset all
2870 			 * retransmission info.  This is because this
2871 			 * segment will be treated as a dup ACK.
2872 			 */
2873 			if (tcp->tcp_rexmit) {
2874 				tcp->tcp_rexmit = B_FALSE;
2875 				tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
2876 				tcp->tcp_rexmit_max = tcp->tcp_snxt;
2877 				tcp->tcp_snd_burst = TCP_CWND_NORMAL;
2878 
2879 				/*
2880 				 * Set tcp_cwnd back to 1 MSS, per
2881 				 * recommendation from
2882 				 * draft-floyd-incr-init-win-01.txt,
2883 				 * Increasing TCP's Initial Window.
2884 				 */
2885 				tcp->tcp_cwnd = tcp->tcp_mss;
2886 			}
2887 
2888 			tcp->tcp_swl1 = seg_seq;
2889 			tcp->tcp_swl2 = seg_ack;
2890 
2891 			new_swnd = BE16_TO_U16(((tcph_t *)tcph)->th_win);
2892 			tcp->tcp_swnd = new_swnd;
2893 			if (new_swnd > tcp->tcp_max_swnd)
2894 				tcp->tcp_max_swnd = new_swnd;
2895 
2896 			/*
2897 			 * Always send the three-way handshake ack immediately
2898 			 * in order to make the connection complete as soon as
2899 			 * possible on the accepting host.
2900 			 */
2901 			flags |= TH_ACK_NEEDED;
2902 			/*
2903 			 * Check to see if there is data to be sent.  If
2904 			 * yes, set the transmit flag.  Then check to see
2905 			 * if received data processing needs to be done.
2906 			 * If not, go straight to xmit_check.  This short
2907 			 * cut is OK as we don't support T/TCP.
2908 			 */
2909 			if (tcp->tcp_unsent)
2910 				flags |= TH_XMIT_NEEDED;
2911 
2912 			if (seg_len == 0) {
2913 				freemsg(mp);
2914 				goto xmit_check;
2915 			}
2916 
2917 			flags &= ~TH_SYN;
2918 			seg_seq++;
2919 			break;
2920 		}
2921 		syn_rcvd:
2922 		tcp->tcp_state = TCPS_SYN_RCVD;
2923 		mp1 = tcp_xmit_mp(tcp, tcp->tcp_xmit_head, tcp->tcp_mss,
2924 		    NULL, NULL, tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
2925 		if (mp1 != NULL) {
2926 			TCP_DUMP_PACKET("tcp_rput_data replying SYN", mp1);
2927 			(void) ipv4_tcp_output(sock_id, mp1);
2928 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
2929 			freeb(mp1);
2930 			/*
2931 			 * Let's wait till our SYN has been ACKED since we
2932 			 * don't have a timer.
2933 			 */
2934 			if (tcp_state_wait(sock_id, tcp, TCPS_ALL_ACKED) < 0) {
2935 				freemsg(mp);
2936 				return;
2937 			}
2938 		}
2939 		freemsg(mp);
2940 		return;
2941 	default:
2942 		break;
2943 	}
2944 	mp->b_rptr = (uchar_t *)tcph + TCP_HDR_LENGTH((tcph_t *)tcph);
2945 	new_swnd = ntohs(tcph->tha_win) <<
2946 	    ((flags & TH_SYN) ? 0 : tcp->tcp_snd_ws);
2947 	mss = tcp->tcp_mss;
2948 
2949 	if (tcp->tcp_snd_ts_ok) {
2950 		if (!tcp_paws_check(tcp, (tcph_t *)tcph, &tcpopt)) {
2951 			/*
2952 			 * This segment is not acceptable.
2953 			 * Drop it and send back an ACK.
2954 			 */
2955 			freemsg(mp);
2956 			flags |= TH_ACK_NEEDED;
2957 			goto ack_check;
2958 		}
2959 	} else if (tcp->tcp_snd_sack_ok) {
2960 		assert(tcp->tcp_sack_info != NULL);
2961 		tcpopt.tcp = tcp;
2962 		/*
2963 		 * SACK info in already updated in tcp_parse_options.  Ignore
2964 		 * all other TCP options...
2965 		 */
2966 		(void) tcp_parse_options((tcph_t *)tcph, &tcpopt);
2967 	}
2968 try_again:;
2969 	gap = seg_seq - tcp->tcp_rnxt;
2970 	rgap = tcp->tcp_rwnd - (gap + seg_len);
2971 	/*
2972 	 * gap is the amount of sequence space between what we expect to see
2973 	 * and what we got for seg_seq.  A positive value for gap means
2974 	 * something got lost.  A negative value means we got some old stuff.
2975 	 */
2976 	if (gap < 0) {
2977 		/* Old stuff present.  Is the SYN in there? */
2978 		if (seg_seq == tcp->tcp_irs && (flags & TH_SYN) &&
2979 		    (seg_len != 0)) {
2980 			flags &= ~TH_SYN;
2981 			seg_seq++;
2982 			/* Recompute the gaps after noting the SYN. */
2983 			goto try_again;
2984 		}
2985 		BUMP_MIB(tcp_mib.tcpInDataDupSegs);
2986 		UPDATE_MIB(tcp_mib.tcpInDataDupBytes,
2987 		    (seg_len > -gap ? -gap : seg_len));
2988 		/* Remove the old stuff from seg_len. */
2989 		seg_len += gap;
2990 		/*
2991 		 * Anything left?
2992 		 * Make sure to check for unack'd FIN when rest of data
2993 		 * has been previously ack'd.
2994 		 */
2995 		if (seg_len < 0 || (seg_len == 0 && !(flags & TH_FIN))) {
2996 			/*
2997 			 * Resets are only valid if they lie within our offered
2998 			 * window.  If the RST bit is set, we just ignore this
2999 			 * segment.
3000 			 */
3001 			if (flags & TH_RST) {
3002 				freemsg(mp);
3003 				return;
3004 			}
3005 
3006 			/*
3007 			 * This segment is "unacceptable".  None of its
3008 			 * sequence space lies within our advertized window.
3009 			 *
3010 			 * Adjust seg_len to the original value for tracing.
3011 			 */
3012 			seg_len -= gap;
3013 #ifdef DEBUG
3014 			printf("tcp_rput: unacceptable, gap %d, rgap "
3015 			    "%d, flags 0x%x, seg_seq %u, seg_ack %u, "
3016 			    "seg_len %d, rnxt %u, snxt %u, %s",
3017 			    gap, rgap, flags, seg_seq, seg_ack,
3018 			    seg_len, tcp->tcp_rnxt, tcp->tcp_snxt,
3019 			    tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
3020 #endif
3021 
3022 			/*
3023 			 * Arrange to send an ACK in response to the
3024 			 * unacceptable segment per RFC 793 page 69. There
3025 			 * is only one small difference between ours and the
3026 			 * acceptability test in the RFC - we accept ACK-only
3027 			 * packet with SEG.SEQ = RCV.NXT+RCV.WND and no ACK
3028 			 * will be generated.
3029 			 *
3030 			 * Note that we have to ACK an ACK-only packet at least
3031 			 * for stacks that send 0-length keep-alives with
3032 			 * SEG.SEQ = SND.NXT-1 as recommended by RFC1122,
3033 			 * section 4.2.3.6. As long as we don't ever generate
3034 			 * an unacceptable packet in response to an incoming
3035 			 * packet that is unacceptable, it should not cause
3036 			 * "ACK wars".
3037 			 */
3038 			flags |=  TH_ACK_NEEDED;
3039 
3040 			/*
3041 			 * Continue processing this segment in order to use the
3042 			 * ACK information it contains, but skip all other
3043 			 * sequence-number processing.	Processing the ACK
3044 			 * information is necessary in order to
3045 			 * re-synchronize connections that may have lost
3046 			 * synchronization.
3047 			 *
3048 			 * We clear seg_len and flag fields related to
3049 			 * sequence number processing as they are not
3050 			 * to be trusted for an unacceptable segment.
3051 			 */
3052 			seg_len = 0;
3053 			flags &= ~(TH_SYN | TH_FIN | TH_URG);
3054 			goto process_ack;
3055 		}
3056 
3057 		/* Fix seg_seq, and chew the gap off the front. */
3058 		seg_seq = tcp->tcp_rnxt;
3059 		do {
3060 			mblk_t	*mp2;
3061 			assert((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
3062 			    (uintptr_t)UINT_MAX);
3063 			gap += (uint_t)(mp->b_wptr - mp->b_rptr);
3064 			if (gap > 0) {
3065 				mp->b_rptr = mp->b_wptr - gap;
3066 				break;
3067 			}
3068 			mp2 = mp;
3069 			mp = mp->b_cont;
3070 			freeb(mp2);
3071 		} while (gap < 0);
3072 	}
3073 	/*
3074 	 * rgap is the amount of stuff received out of window.  A negative
3075 	 * value is the amount out of window.
3076 	 */
3077 	if (rgap < 0) {
3078 		mblk_t	*mp2;
3079 
3080 		if (tcp->tcp_rwnd == 0)
3081 			BUMP_MIB(tcp_mib.tcpInWinProbe);
3082 		else {
3083 			BUMP_MIB(tcp_mib.tcpInDataPastWinSegs);
3084 			UPDATE_MIB(tcp_mib.tcpInDataPastWinBytes, -rgap);
3085 		}
3086 
3087 		/*
3088 		 * seg_len does not include the FIN, so if more than
3089 		 * just the FIN is out of window, we act like we don't
3090 		 * see it.  (If just the FIN is out of window, rgap
3091 		 * will be zero and we will go ahead and acknowledge
3092 		 * the FIN.)
3093 		 */
3094 		flags &= ~TH_FIN;
3095 
3096 		/* Fix seg_len and make sure there is something left. */
3097 		seg_len += rgap;
3098 		if (seg_len <= 0) {
3099 			/*
3100 			 * Resets are only valid if they lie within our offered
3101 			 * window.  If the RST bit is set, we just ignore this
3102 			 * segment.
3103 			 */
3104 			if (flags & TH_RST) {
3105 				freemsg(mp);
3106 				return;
3107 			}
3108 
3109 			/* Per RFC 793, we need to send back an ACK. */
3110 			flags |= TH_ACK_NEEDED;
3111 
3112 			/*
3113 			 * If this is a zero window probe, continue to
3114 			 * process the ACK part.  But we need to set seg_len
3115 			 * to 0 to avoid data processing.  Otherwise just
3116 			 * drop the segment and send back an ACK.
3117 			 */
3118 			if (tcp->tcp_rwnd == 0 && seg_seq == tcp->tcp_rnxt) {
3119 				flags &= ~(TH_SYN | TH_URG);
3120 				seg_len = 0;
3121 				/* Let's see if we can update our rwnd */
3122 				tcp_rcv_drain(sock_id, tcp);
3123 				goto process_ack;
3124 			} else {
3125 				freemsg(mp);
3126 				goto ack_check;
3127 			}
3128 		}
3129 		/* Pitch out of window stuff off the end. */
3130 		rgap = seg_len;
3131 		mp2 = mp;
3132 		do {
3133 			assert((uintptr_t)(mp2->b_wptr -
3134 			    mp2->b_rptr) <= (uintptr_t)INT_MAX);
3135 			rgap -= (int)(mp2->b_wptr - mp2->b_rptr);
3136 			if (rgap < 0) {
3137 				mp2->b_wptr += rgap;
3138 				if ((mp1 = mp2->b_cont) != NULL) {
3139 					mp2->b_cont = NULL;
3140 					freemsg(mp1);
3141 				}
3142 				break;
3143 			}
3144 		} while ((mp2 = mp2->b_cont) != NULL);
3145 	}
3146 ok:;
3147 	/*
3148 	 * TCP should check ECN info for segments inside the window only.
3149 	 * Therefore the check should be done here.
3150 	 */
3151 	if (tcp->tcp_ecn_ok) {
3152 		uchar_t tos = ((struct ip *)rptr)->ip_tos;
3153 
3154 		if (flags & TH_CWR) {
3155 			tcp->tcp_ecn_echo_on = B_FALSE;
3156 		}
3157 		/*
3158 		 * Note that both ECN_CE and CWR can be set in the
3159 		 * same segment.  In this case, we once again turn
3160 		 * on ECN_ECHO.
3161 		 */
3162 		if ((tos & IPH_ECN_CE) == IPH_ECN_CE) {
3163 			tcp->tcp_ecn_echo_on = B_TRUE;
3164 		}
3165 	}
3166 
3167 	/*
3168 	 * Check whether we can update tcp_ts_recent.  This test is
3169 	 * NOT the one in RFC 1323 3.4.  It is from Braden, 1993, "TCP
3170 	 * Extensions for High Performance: An Update", Internet Draft.
3171 	 */
3172 	if (tcp->tcp_snd_ts_ok &&
3173 	    TSTMP_GEQ(tcpopt.tcp_opt_ts_val, tcp->tcp_ts_recent) &&
3174 	    SEQ_LEQ(seg_seq, tcp->tcp_rack)) {
3175 		tcp->tcp_ts_recent = tcpopt.tcp_opt_ts_val;
3176 		tcp->tcp_last_rcv_lbolt = prom_gettime();
3177 	}
3178 
3179 	if (seg_seq != tcp->tcp_rnxt || tcp->tcp_reass_head) {
3180 		/*
3181 		 * FIN in an out of order segment.  We record this in
3182 		 * tcp_valid_bits and the seq num of FIN in tcp_ofo_fin_seq.
3183 		 * Clear the FIN so that any check on FIN flag will fail.
3184 		 * Remember that FIN also counts in the sequence number
3185 		 * space.  So we need to ack out of order FIN only segments.
3186 		 */
3187 		if (flags & TH_FIN) {
3188 			tcp->tcp_valid_bits |= TCP_OFO_FIN_VALID;
3189 			tcp->tcp_ofo_fin_seq = seg_seq + seg_len;
3190 			flags &= ~TH_FIN;
3191 			flags |= TH_ACK_NEEDED;
3192 		}
3193 		if (seg_len > 0) {
3194 			/* Fill in the SACK blk list. */
3195 			if (tcp->tcp_snd_sack_ok) {
3196 				assert(tcp->tcp_sack_info != NULL);
3197 				tcp_sack_insert(tcp->tcp_sack_list,
3198 				    seg_seq, seg_seq + seg_len,
3199 				    &(tcp->tcp_num_sack_blk));
3200 			}
3201 
3202 			/*
3203 			 * Attempt reassembly and see if we have something
3204 			 * ready to go.
3205 			 */
3206 			mp = tcp_reass(tcp, mp, seg_seq);
3207 			/* Always ack out of order packets */
3208 			flags |= TH_ACK_NEEDED | TH_PUSH;
3209 			if (mp != NULL) {
3210 				assert((uintptr_t)(mp->b_wptr -
3211 				    mp->b_rptr) <= (uintptr_t)INT_MAX);
3212 				seg_len = mp->b_cont ? msgdsize(mp) :
3213 					(int)(mp->b_wptr - mp->b_rptr);
3214 				seg_seq = tcp->tcp_rnxt;
3215 				/*
3216 				 * A gap is filled and the seq num and len
3217 				 * of the gap match that of a previously
3218 				 * received FIN, put the FIN flag back in.
3219 				 */
3220 				if ((tcp->tcp_valid_bits & TCP_OFO_FIN_VALID) &&
3221 				    seg_seq + seg_len == tcp->tcp_ofo_fin_seq) {
3222 					flags |= TH_FIN;
3223 					tcp->tcp_valid_bits &=
3224 					    ~TCP_OFO_FIN_VALID;
3225 				}
3226 			} else {
3227 				/*
3228 				 * Keep going even with NULL mp.
3229 				 * There may be a useful ACK or something else
3230 				 * we don't want to miss.
3231 				 *
3232 				 * But TCP should not perform fast retransmit
3233 				 * because of the ack number.  TCP uses
3234 				 * seg_len == 0 to determine if it is a pure
3235 				 * ACK.  And this is not a pure ACK.
3236 				 */
3237 				seg_len = 0;
3238 				ofo_seg = B_TRUE;
3239 			}
3240 		}
3241 	} else if (seg_len > 0) {
3242 		BUMP_MIB(tcp_mib.tcpInDataInorderSegs);
3243 		UPDATE_MIB(tcp_mib.tcpInDataInorderBytes, seg_len);
3244 		/*
3245 		 * If an out of order FIN was received before, and the seq
3246 		 * num and len of the new segment match that of the FIN,
3247 		 * put the FIN flag back in.
3248 		 */
3249 		if ((tcp->tcp_valid_bits & TCP_OFO_FIN_VALID) &&
3250 		    seg_seq + seg_len == tcp->tcp_ofo_fin_seq) {
3251 			flags |= TH_FIN;
3252 			tcp->tcp_valid_bits &= ~TCP_OFO_FIN_VALID;
3253 		}
3254 	}
3255 	if ((flags & (TH_RST | TH_SYN | TH_URG | TH_ACK)) != TH_ACK) {
3256 	if (flags & TH_RST) {
3257 		freemsg(mp);
3258 		switch (tcp->tcp_state) {
3259 		case TCPS_SYN_RCVD:
3260 			(void) tcp_clean_death(sock_id, tcp, ECONNREFUSED);
3261 			break;
3262 		case TCPS_ESTABLISHED:
3263 		case TCPS_FIN_WAIT_1:
3264 		case TCPS_FIN_WAIT_2:
3265 		case TCPS_CLOSE_WAIT:
3266 			(void) tcp_clean_death(sock_id, tcp, ECONNRESET);
3267 			break;
3268 		case TCPS_CLOSING:
3269 		case TCPS_LAST_ACK:
3270 			(void) tcp_clean_death(sock_id, tcp, 0);
3271 			break;
3272 		default:
3273 			assert(tcp->tcp_state != TCPS_TIME_WAIT);
3274 			(void) tcp_clean_death(sock_id, tcp, ENXIO);
3275 			break;
3276 		}
3277 		return;
3278 	}
3279 	if (flags & TH_SYN) {
3280 		/*
3281 		 * See RFC 793, Page 71
3282 		 *
3283 		 * The seq number must be in the window as it should
3284 		 * be "fixed" above.  If it is outside window, it should
3285 		 * be already rejected.  Note that we allow seg_seq to be
3286 		 * rnxt + rwnd because we want to accept 0 window probe.
3287 		 */
3288 		assert(SEQ_GEQ(seg_seq, tcp->tcp_rnxt) &&
3289 		    SEQ_LEQ(seg_seq, tcp->tcp_rnxt + tcp->tcp_rwnd));
3290 		freemsg(mp);
3291 		/*
3292 		 * If the ACK flag is not set, just use our snxt as the
3293 		 * seq number of the RST segment.
3294 		 */
3295 		if (!(flags & TH_ACK)) {
3296 			seg_ack = tcp->tcp_snxt;
3297 		}
3298 		tcp_xmit_ctl("TH_SYN", tcp, NULL, seg_ack,
3299 		    seg_seq + 1, TH_RST|TH_ACK, 0, sock_id);
3300 		assert(tcp->tcp_state != TCPS_TIME_WAIT);
3301 		(void) tcp_clean_death(sock_id, tcp, ECONNRESET);
3302 		return;
3303 	}
3304 
3305 process_ack:
3306 	if (!(flags & TH_ACK)) {
3307 #ifdef DEBUG
3308 		printf("No ack in segment, dropped it, seq:%x\n", seg_seq);
3309 #endif
3310 		freemsg(mp);
3311 		goto xmit_check;
3312 	}
3313 	}
3314 	bytes_acked = (int)(seg_ack - tcp->tcp_suna);
3315 
3316 	if (tcp->tcp_state == TCPS_SYN_RCVD) {
3317 		tcp_t	*listener = tcp->tcp_listener;
3318 #ifdef DEBUG
3319 		printf("Done with eager 3-way handshake\n");
3320 #endif
3321 		/*
3322 		 * NOTE: RFC 793 pg. 72 says this should be 'bytes_acked < 0'
3323 		 * but that would mean we have an ack that ignored our SYN.
3324 		 */
3325 		if (bytes_acked < 1 || SEQ_GT(seg_ack, tcp->tcp_snxt)) {
3326 			freemsg(mp);
3327 			tcp_xmit_ctl("TCPS_SYN_RCVD-bad_ack",
3328 			    tcp, NULL, seg_ack, 0, TH_RST, 0, sock_id);
3329 			return;
3330 		}
3331 
3332 		/*
3333 		 * if the conn_req_q is full defer processing
3334 		 * until space is availabe after accept()
3335 		 * processing
3336 		 */
3337 		if (listener->tcp_conn_req_cnt_q <
3338 		    listener->tcp_conn_req_max) {
3339 			tcp_t *tail;
3340 
3341 			listener->tcp_conn_req_cnt_q0--;
3342 			listener->tcp_conn_req_cnt_q++;
3343 
3344 			/* Move from SYN_RCVD to ESTABLISHED list  */
3345 			tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
3346 				tcp->tcp_eager_prev_q0;
3347 			tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
3348 				tcp->tcp_eager_next_q0;
3349 			tcp->tcp_eager_prev_q0 = NULL;
3350 			tcp->tcp_eager_next_q0 = NULL;
3351 
3352 			/*
3353 			 * Insert at end of the queue because sockfs
3354 			 * sends down T_CONN_RES in chronological
3355 			 * order. Leaving the older conn indications
3356 			 * at front of the queue helps reducing search
3357 			 * time.
3358 			 */
3359 			tail = listener->tcp_eager_last_q;
3360 			if (tail != NULL) {
3361 				tail->tcp_eager_next_q = tcp;
3362 			} else {
3363 				listener->tcp_eager_next_q = tcp;
3364 			}
3365 			listener->tcp_eager_last_q = tcp;
3366 			tcp->tcp_eager_next_q = NULL;
3367 		} else {
3368 			/*
3369 			 * Defer connection on q0 and set deferred
3370 			 * connection bit true
3371 			 */
3372 			tcp->tcp_conn_def_q0 = B_TRUE;
3373 
3374 			/* take tcp out of q0 ... */
3375 			tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
3376 			    tcp->tcp_eager_next_q0;
3377 			tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
3378 			    tcp->tcp_eager_prev_q0;
3379 
3380 			/* ... and place it at the end of q0 */
3381 			tcp->tcp_eager_prev_q0 = listener->tcp_eager_prev_q0;
3382 			tcp->tcp_eager_next_q0 = listener;
3383 			listener->tcp_eager_prev_q0->tcp_eager_next_q0 = tcp;
3384 			listener->tcp_eager_prev_q0 = tcp;
3385 		}
3386 
3387 		tcp->tcp_suna = tcp->tcp_iss + 1;	/* One for the SYN */
3388 		bytes_acked--;
3389 
3390 		/*
3391 		 * If SYN was retransmitted, need to reset all
3392 		 * retransmission info as this segment will be
3393 		 * treated as a dup ACK.
3394 		 */
3395 		if (tcp->tcp_rexmit) {
3396 			tcp->tcp_rexmit = B_FALSE;
3397 			tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3398 			tcp->tcp_rexmit_max = tcp->tcp_snxt;
3399 			tcp->tcp_snd_burst = TCP_CWND_NORMAL;
3400 			tcp->tcp_ms_we_have_waited = 0;
3401 			tcp->tcp_cwnd = mss;
3402 		}
3403 
3404 		/*
3405 		 * We set the send window to zero here.
3406 		 * This is needed if there is data to be
3407 		 * processed already on the queue.
3408 		 * Later (at swnd_update label), the
3409 		 * "new_swnd > tcp_swnd" condition is satisfied
3410 		 * the XMIT_NEEDED flag is set in the current
3411 		 * (SYN_RCVD) state. This ensures tcp_wput_data() is
3412 		 * called if there is already data on queue in
3413 		 * this state.
3414 		 */
3415 		tcp->tcp_swnd = 0;
3416 
3417 		if (new_swnd > tcp->tcp_max_swnd)
3418 			tcp->tcp_max_swnd = new_swnd;
3419 		tcp->tcp_swl1 = seg_seq;
3420 		tcp->tcp_swl2 = seg_ack;
3421 		tcp->tcp_state = TCPS_ESTABLISHED;
3422 		tcp->tcp_valid_bits &= ~TCP_ISS_VALID;
3423 	}
3424 	/* This code follows 4.4BSD-Lite2 mostly. */
3425 	if (bytes_acked < 0)
3426 		goto est;
3427 
3428 	/*
3429 	 * If TCP is ECN capable and the congestion experience bit is
3430 	 * set, reduce tcp_cwnd and tcp_ssthresh.  But this should only be
3431 	 * done once per window (or more loosely, per RTT).
3432 	 */
3433 	if (tcp->tcp_cwr && SEQ_GT(seg_ack, tcp->tcp_cwr_snd_max))
3434 		tcp->tcp_cwr = B_FALSE;
3435 	if (tcp->tcp_ecn_ok && (flags & TH_ECE)) {
3436 		if (!tcp->tcp_cwr) {
3437 			npkt = (MIN(tcp->tcp_cwnd, tcp->tcp_swnd) >> 1) / mss;
3438 			tcp->tcp_cwnd_ssthresh = MAX(npkt, 2) * mss;
3439 			tcp->tcp_cwnd = npkt * mss;
3440 			/*
3441 			 * If the cwnd is 0, use the timer to clock out
3442 			 * new segments.  This is required by the ECN spec.
3443 			 */
3444 			if (npkt == 0) {
3445 				TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3446 				/*
3447 				 * This makes sure that when the ACK comes
3448 				 * back, we will increase tcp_cwnd by 1 MSS.
3449 				 */
3450 				tcp->tcp_cwnd_cnt = 0;
3451 			}
3452 			tcp->tcp_cwr = B_TRUE;
3453 			/*
3454 			 * This marks the end of the current window of in
3455 			 * flight data.  That is why we don't use
3456 			 * tcp_suna + tcp_swnd.  Only data in flight can
3457 			 * provide ECN info.
3458 			 */
3459 			tcp->tcp_cwr_snd_max = tcp->tcp_snxt;
3460 			tcp->tcp_ecn_cwr_sent = B_FALSE;
3461 		}
3462 	}
3463 
3464 	mp1 = tcp->tcp_xmit_head;
3465 	if (bytes_acked == 0) {
3466 		if (!ofo_seg && seg_len == 0 && new_swnd == tcp->tcp_swnd) {
3467 			int dupack_cnt;
3468 
3469 			BUMP_MIB(tcp_mib.tcpInDupAck);
3470 			/*
3471 			 * Fast retransmit.  When we have seen exactly three
3472 			 * identical ACKs while we have unacked data
3473 			 * outstanding we take it as a hint that our peer
3474 			 * dropped something.
3475 			 *
3476 			 * If TCP is retransmitting, don't do fast retransmit.
3477 			 */
3478 			if (mp1 != NULL && tcp->tcp_suna != tcp->tcp_snxt &&
3479 			    ! tcp->tcp_rexmit) {
3480 				/* Do Limited Transmit */
3481 				if ((dupack_cnt = ++tcp->tcp_dupack_cnt) <
3482 				    tcp_dupack_fast_retransmit) {
3483 					/*
3484 					 * RFC 3042
3485 					 *
3486 					 * What we need to do is temporarily
3487 					 * increase tcp_cwnd so that new
3488 					 * data can be sent if it is allowed
3489 					 * by the receive window (tcp_rwnd).
3490 					 * tcp_wput_data() will take care of
3491 					 * the rest.
3492 					 *
3493 					 * If the connection is SACK capable,
3494 					 * only do limited xmit when there
3495 					 * is SACK info.
3496 					 *
3497 					 * Note how tcp_cwnd is incremented.
3498 					 * The first dup ACK will increase
3499 					 * it by 1 MSS.  The second dup ACK
3500 					 * will increase it by 2 MSS.  This
3501 					 * means that only 1 new segment will
3502 					 * be sent for each dup ACK.
3503 					 */
3504 					if (tcp->tcp_unsent > 0 &&
3505 					    (!tcp->tcp_snd_sack_ok ||
3506 					    (tcp->tcp_snd_sack_ok &&
3507 					    tcp->tcp_notsack_list != NULL))) {
3508 						tcp->tcp_cwnd += mss <<
3509 						    (tcp->tcp_dupack_cnt - 1);
3510 						flags |= TH_LIMIT_XMIT;
3511 					}
3512 				} else if (dupack_cnt ==
3513 				    tcp_dupack_fast_retransmit) {
3514 
3515 				BUMP_MIB(tcp_mib.tcpOutFastRetrans);
3516 				/*
3517 				 * If we have reduced tcp_ssthresh
3518 				 * because of ECN, do not reduce it again
3519 				 * unless it is already one window of data
3520 				 * away.  After one window of data, tcp_cwr
3521 				 * should then be cleared.  Note that
3522 				 * for non ECN capable connection, tcp_cwr
3523 				 * should always be false.
3524 				 *
3525 				 * Adjust cwnd since the duplicate
3526 				 * ack indicates that a packet was
3527 				 * dropped (due to congestion.)
3528 				 */
3529 				if (!tcp->tcp_cwr) {
3530 					npkt = (MIN(tcp->tcp_cwnd,
3531 					    tcp->tcp_swnd) >> 1) / mss;
3532 					if (npkt < 2)
3533 						npkt = 2;
3534 					tcp->tcp_cwnd_ssthresh = npkt * mss;
3535 					tcp->tcp_cwnd = (npkt +
3536 					    tcp->tcp_dupack_cnt) * mss;
3537 				}
3538 				if (tcp->tcp_ecn_ok) {
3539 					tcp->tcp_cwr = B_TRUE;
3540 					tcp->tcp_cwr_snd_max = tcp->tcp_snxt;
3541 					tcp->tcp_ecn_cwr_sent = B_FALSE;
3542 				}
3543 
3544 				/*
3545 				 * We do Hoe's algorithm.  Refer to her
3546 				 * paper "Improving the Start-up Behavior
3547 				 * of a Congestion Control Scheme for TCP,"
3548 				 * appeared in SIGCOMM'96.
3549 				 *
3550 				 * Save highest seq no we have sent so far.
3551 				 * Be careful about the invisible FIN byte.
3552 				 */
3553 				if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
3554 				    (tcp->tcp_unsent == 0)) {
3555 					tcp->tcp_rexmit_max = tcp->tcp_fss;
3556 				} else {
3557 					tcp->tcp_rexmit_max = tcp->tcp_snxt;
3558 				}
3559 
3560 				/*
3561 				 * Do not allow bursty traffic during.
3562 				 * fast recovery.  Refer to Fall and Floyd's
3563 				 * paper "Simulation-based Comparisons of
3564 				 * Tahoe, Reno and SACK TCP" (in CCR ??)
3565 				 * This is a best current practise.
3566 				 */
3567 				tcp->tcp_snd_burst = TCP_CWND_SS;
3568 
3569 				/*
3570 				 * For SACK:
3571 				 * Calculate tcp_pipe, which is the
3572 				 * estimated number of bytes in
3573 				 * network.
3574 				 *
3575 				 * tcp_fack is the highest sack'ed seq num
3576 				 * TCP has received.
3577 				 *
3578 				 * tcp_pipe is explained in the above quoted
3579 				 * Fall and Floyd's paper.  tcp_fack is
3580 				 * explained in Mathis and Mahdavi's
3581 				 * "Forward Acknowledgment: Refining TCP
3582 				 * Congestion Control" in SIGCOMM '96.
3583 				 */
3584 				if (tcp->tcp_snd_sack_ok) {
3585 					assert(tcp->tcp_sack_info != NULL);
3586 					if (tcp->tcp_notsack_list != NULL) {
3587 						tcp->tcp_pipe = tcp->tcp_snxt -
3588 						    tcp->tcp_fack;
3589 						tcp->tcp_sack_snxt = seg_ack;
3590 						flags |= TH_NEED_SACK_REXMIT;
3591 					} else {
3592 						/*
3593 						 * Always initialize tcp_pipe
3594 						 * even though we don't have
3595 						 * any SACK info.  If later
3596 						 * we get SACK info and
3597 						 * tcp_pipe is not initialized,
3598 						 * funny things will happen.
3599 						 */
3600 						tcp->tcp_pipe =
3601 						    tcp->tcp_cwnd_ssthresh;
3602 					}
3603 				} else {
3604 					flags |= TH_REXMIT_NEEDED;
3605 				} /* tcp_snd_sack_ok */
3606 
3607 				} else {
3608 					/*
3609 					 * Here we perform congestion
3610 					 * avoidance, but NOT slow start.
3611 					 * This is known as the Fast
3612 					 * Recovery Algorithm.
3613 					 */
3614 					if (tcp->tcp_snd_sack_ok &&
3615 					    tcp->tcp_notsack_list != NULL) {
3616 						flags |= TH_NEED_SACK_REXMIT;
3617 						tcp->tcp_pipe -= mss;
3618 						if (tcp->tcp_pipe < 0)
3619 							tcp->tcp_pipe = 0;
3620 					} else {
3621 					/*
3622 					 * We know that one more packet has
3623 					 * left the pipe thus we can update
3624 					 * cwnd.
3625 					 */
3626 					cwnd = tcp->tcp_cwnd + mss;
3627 					if (cwnd > tcp->tcp_cwnd_max)
3628 						cwnd = tcp->tcp_cwnd_max;
3629 					tcp->tcp_cwnd = cwnd;
3630 					flags |= TH_XMIT_NEEDED;
3631 					}
3632 				}
3633 			}
3634 		} else if (tcp->tcp_zero_win_probe) {
3635 			/*
3636 			 * If the window has opened, need to arrange
3637 			 * to send additional data.
3638 			 */
3639 			if (new_swnd != 0) {
3640 				/* tcp_suna != tcp_snxt */
3641 				/* Packet contains a window update */
3642 				BUMP_MIB(tcp_mib.tcpInWinUpdate);
3643 				tcp->tcp_zero_win_probe = 0;
3644 				tcp->tcp_timer_backoff = 0;
3645 				tcp->tcp_ms_we_have_waited = 0;
3646 
3647 				/*
3648 				 * Transmit starting with tcp_suna since
3649 				 * the one byte probe is not ack'ed.
3650 				 * If TCP has sent more than one identical
3651 				 * probe, tcp_rexmit will be set.  That means
3652 				 * tcp_ss_rexmit() will send out the one
3653 				 * byte along with new data.  Otherwise,
3654 				 * fake the retransmission.
3655 				 */
3656 				flags |= TH_XMIT_NEEDED;
3657 				if (!tcp->tcp_rexmit) {
3658 					tcp->tcp_rexmit = B_TRUE;
3659 					tcp->tcp_dupack_cnt = 0;
3660 					tcp->tcp_rexmit_nxt = tcp->tcp_suna;
3661 					tcp->tcp_rexmit_max = tcp->tcp_suna + 1;
3662 				}
3663 			}
3664 		}
3665 		goto swnd_update;
3666 	}
3667 
3668 	/*
3669 	 * Check for "acceptability" of ACK value per RFC 793, pages 72 - 73.
3670 	 * If the ACK value acks something that we have not yet sent, it might
3671 	 * be an old duplicate segment.  Send an ACK to re-synchronize the
3672 	 * other side.
3673 	 * Note: reset in response to unacceptable ACK in SYN_RECEIVE
3674 	 * state is handled above, so we can always just drop the segment and
3675 	 * send an ACK here.
3676 	 *
3677 	 * Should we send ACKs in response to ACK only segments?
3678 	 */
3679 	if (SEQ_GT(seg_ack, tcp->tcp_snxt)) {
3680 		BUMP_MIB(tcp_mib.tcpInAckUnsent);
3681 		/* drop the received segment */
3682 		freemsg(mp);
3683 
3684 		/* Send back an ACK. */
3685 		mp = tcp_ack_mp(tcp);
3686 
3687 		if (mp == NULL) {
3688 			return;
3689 		}
3690 		BUMP_MIB(tcp_mib.tcpOutAck);
3691 		(void) ipv4_tcp_output(sock_id, mp);
3692 		freeb(mp);
3693 		return;
3694 	}
3695 
3696 	/*
3697 	 * TCP gets a new ACK, update the notsack'ed list to delete those
3698 	 * blocks that are covered by this ACK.
3699 	 */
3700 	if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
3701 		tcp_notsack_remove(&(tcp->tcp_notsack_list), seg_ack,
3702 		    &(tcp->tcp_num_notsack_blk), &(tcp->tcp_cnt_notsack_list));
3703 	}
3704 
3705 	/*
3706 	 * If we got an ACK after fast retransmit, check to see
3707 	 * if it is a partial ACK.  If it is not and the congestion
3708 	 * window was inflated to account for the other side's
3709 	 * cached packets, retract it.  If it is, do Hoe's algorithm.
3710 	 */
3711 	if (tcp->tcp_dupack_cnt >= tcp_dupack_fast_retransmit) {
3712 		assert(tcp->tcp_rexmit == B_FALSE);
3713 		if (SEQ_GEQ(seg_ack, tcp->tcp_rexmit_max)) {
3714 			tcp->tcp_dupack_cnt = 0;
3715 			/*
3716 			 * Restore the orig tcp_cwnd_ssthresh after
3717 			 * fast retransmit phase.
3718 			 */
3719 			if (tcp->tcp_cwnd > tcp->tcp_cwnd_ssthresh) {
3720 				tcp->tcp_cwnd = tcp->tcp_cwnd_ssthresh;
3721 			}
3722 			tcp->tcp_rexmit_max = seg_ack;
3723 			tcp->tcp_cwnd_cnt = 0;
3724 			tcp->tcp_snd_burst = TCP_CWND_NORMAL;
3725 
3726 			/*
3727 			 * Remove all notsack info to avoid confusion with
3728 			 * the next fast retrasnmit/recovery phase.
3729 			 */
3730 			if (tcp->tcp_snd_sack_ok &&
3731 			    tcp->tcp_notsack_list != NULL) {
3732 				TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list);
3733 			}
3734 		} else {
3735 			if (tcp->tcp_snd_sack_ok &&
3736 			    tcp->tcp_notsack_list != NULL) {
3737 				flags |= TH_NEED_SACK_REXMIT;
3738 				tcp->tcp_pipe -= mss;
3739 				if (tcp->tcp_pipe < 0)
3740 					tcp->tcp_pipe = 0;
3741 			} else {
3742 				/*
3743 				 * Hoe's algorithm:
3744 				 *
3745 				 * Retransmit the unack'ed segment and
3746 				 * restart fast recovery.  Note that we
3747 				 * need to scale back tcp_cwnd to the
3748 				 * original value when we started fast
3749 				 * recovery.  This is to prevent overly
3750 				 * aggressive behaviour in sending new
3751 				 * segments.
3752 				 */
3753 				tcp->tcp_cwnd = tcp->tcp_cwnd_ssthresh +
3754 					tcp_dupack_fast_retransmit * mss;
3755 				tcp->tcp_cwnd_cnt = tcp->tcp_cwnd;
3756 				BUMP_MIB(tcp_mib.tcpOutFastRetrans);
3757 				flags |= TH_REXMIT_NEEDED;
3758 			}
3759 		}
3760 	} else {
3761 		tcp->tcp_dupack_cnt = 0;
3762 		if (tcp->tcp_rexmit) {
3763 			/*
3764 			 * TCP is retranmitting.  If the ACK ack's all
3765 			 * outstanding data, update tcp_rexmit_max and
3766 			 * tcp_rexmit_nxt.  Otherwise, update tcp_rexmit_nxt
3767 			 * to the correct value.
3768 			 *
3769 			 * Note that SEQ_LEQ() is used.  This is to avoid
3770 			 * unnecessary fast retransmit caused by dup ACKs
3771 			 * received when TCP does slow start retransmission
3772 			 * after a time out.  During this phase, TCP may
3773 			 * send out segments which are already received.
3774 			 * This causes dup ACKs to be sent back.
3775 			 */
3776 			if (SEQ_LEQ(seg_ack, tcp->tcp_rexmit_max)) {
3777 				if (SEQ_GT(seg_ack, tcp->tcp_rexmit_nxt)) {
3778 					tcp->tcp_rexmit_nxt = seg_ack;
3779 				}
3780 				if (seg_ack != tcp->tcp_rexmit_max) {
3781 					flags |= TH_XMIT_NEEDED;
3782 				}
3783 			} else {
3784 				tcp->tcp_rexmit = B_FALSE;
3785 				tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3786 				tcp->tcp_snd_burst = TCP_CWND_NORMAL;
3787 			}
3788 			tcp->tcp_ms_we_have_waited = 0;
3789 		}
3790 	}
3791 
3792 	BUMP_MIB(tcp_mib.tcpInAckSegs);
3793 	UPDATE_MIB(tcp_mib.tcpInAckBytes, bytes_acked);
3794 	tcp->tcp_suna = seg_ack;
3795 	if (tcp->tcp_zero_win_probe != 0) {
3796 		tcp->tcp_zero_win_probe = 0;
3797 		tcp->tcp_timer_backoff = 0;
3798 	}
3799 
3800 	/*
3801 	 * If tcp_xmit_head is NULL, then it must be the FIN being ack'ed.
3802 	 * Note that it cannot be the SYN being ack'ed.  The code flow
3803 	 * will not reach here.
3804 	 */
3805 	if (mp1 == NULL) {
3806 		goto fin_acked;
3807 	}
3808 
3809 	/*
3810 	 * Update the congestion window.
3811 	 *
3812 	 * If TCP is not ECN capable or TCP is ECN capable but the
3813 	 * congestion experience bit is not set, increase the tcp_cwnd as
3814 	 * usual.
3815 	 */
3816 	if (!tcp->tcp_ecn_ok || !(flags & TH_ECE)) {
3817 		cwnd = tcp->tcp_cwnd;
3818 		add = mss;
3819 
3820 		if (cwnd >= tcp->tcp_cwnd_ssthresh) {
3821 			/*
3822 			 * This is to prevent an increase of less than 1 MSS of
3823 			 * tcp_cwnd.  With partial increase, tcp_wput_data()
3824 			 * may send out tinygrams in order to preserve mblk
3825 			 * boundaries.
3826 			 *
3827 			 * By initializing tcp_cwnd_cnt to new tcp_cwnd and
3828 			 * decrementing it by 1 MSS for every ACKs, tcp_cwnd is
3829 			 * increased by 1 MSS for every RTTs.
3830 			 */
3831 			if (tcp->tcp_cwnd_cnt <= 0) {
3832 				tcp->tcp_cwnd_cnt = cwnd + add;
3833 			} else {
3834 				tcp->tcp_cwnd_cnt -= add;
3835 				add = 0;
3836 			}
3837 		}
3838 		tcp->tcp_cwnd = MIN(cwnd + add, tcp->tcp_cwnd_max);
3839 	}
3840 
3841 	/* Can we update the RTT estimates? */
3842 	if (tcp->tcp_snd_ts_ok) {
3843 		/* Ignore zero timestamp echo-reply. */
3844 		if (tcpopt.tcp_opt_ts_ecr != 0) {
3845 			tcp_set_rto(tcp, (int32_t)(prom_gettime() -
3846 			    tcpopt.tcp_opt_ts_ecr));
3847 		}
3848 
3849 		/* If needed, restart the timer. */
3850 		if (tcp->tcp_set_timer == 1) {
3851 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3852 			tcp->tcp_set_timer = 0;
3853 		}
3854 		/*
3855 		 * Update tcp_csuna in case the other side stops sending
3856 		 * us timestamps.
3857 		 */
3858 		tcp->tcp_csuna = tcp->tcp_snxt;
3859 	} else if (SEQ_GT(seg_ack, tcp->tcp_csuna)) {
3860 		/*
3861 		 * An ACK sequence we haven't seen before, so get the RTT
3862 		 * and update the RTO.
3863 		 * Note. use uintptr_t to suppress the gcc warning.
3864 		 */
3865 		tcp_set_rto(tcp, (int32_t)(prom_gettime() -
3866 		    (uint32_t)(uintptr_t)mp1->b_prev));
3867 
3868 		/* Remeber the last sequence to be ACKed */
3869 		tcp->tcp_csuna = seg_ack;
3870 		if (tcp->tcp_set_timer == 1) {
3871 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3872 			tcp->tcp_set_timer = 0;
3873 		}
3874 	} else {
3875 		BUMP_MIB(tcp_mib.tcpRttNoUpdate);
3876 	}
3877 
3878 	/* Eat acknowledged bytes off the xmit queue. */
3879 	for (;;) {
3880 		mblk_t	*mp2;
3881 		uchar_t	*wptr;
3882 
3883 		wptr = mp1->b_wptr;
3884 		assert((uintptr_t)(wptr - mp1->b_rptr) <= (uintptr_t)INT_MAX);
3885 		bytes_acked -= (int)(wptr - mp1->b_rptr);
3886 		if (bytes_acked < 0) {
3887 			mp1->b_rptr = wptr + bytes_acked;
3888 			break;
3889 		}
3890 		mp1->b_prev = NULL;
3891 		mp2 = mp1;
3892 		mp1 = mp1->b_cont;
3893 		freeb(mp2);
3894 		if (bytes_acked == 0) {
3895 			if (mp1 == NULL) {
3896 				/* Everything is ack'ed, clear the tail. */
3897 				tcp->tcp_xmit_tail = NULL;
3898 				goto pre_swnd_update;
3899 			}
3900 			if (mp2 != tcp->tcp_xmit_tail)
3901 				break;
3902 			tcp->tcp_xmit_tail = mp1;
3903 			assert((uintptr_t)(mp1->b_wptr -
3904 			    mp1->b_rptr) <= (uintptr_t)INT_MAX);
3905 			tcp->tcp_xmit_tail_unsent = (int)(mp1->b_wptr -
3906 			    mp1->b_rptr);
3907 			break;
3908 		}
3909 		if (mp1 == NULL) {
3910 			/*
3911 			 * More was acked but there is nothing more
3912 			 * outstanding.  This means that the FIN was
3913 			 * just acked or that we're talking to a clown.
3914 			 */
3915 fin_acked:
3916 			assert(tcp->tcp_fin_sent);
3917 			tcp->tcp_xmit_tail = NULL;
3918 			if (tcp->tcp_fin_sent) {
3919 				tcp->tcp_fin_acked = B_TRUE;
3920 			} else {
3921 				/*
3922 				 * We should never got here because
3923 				 * we have already checked that the
3924 				 * number of bytes ack'ed should be
3925 				 * smaller than or equal to what we
3926 				 * have sent so far (it is the
3927 				 * acceptability check of the ACK).
3928 				 * We can only get here if the send
3929 				 * queue is corrupted.
3930 				 *
3931 				 * Terminate the connection and
3932 				 * panic the system.  It is better
3933 				 * for us to panic instead of
3934 				 * continuing to avoid other disaster.
3935 				 */
3936 				tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
3937 				    tcp->tcp_rnxt, TH_RST|TH_ACK, 0, sock_id);
3938 				printf("Memory corruption "
3939 				    "detected for connection %s.\n",
3940 				    tcp_display(tcp, NULL,
3941 					DISP_ADDR_AND_PORT));
3942 				/* We should never get here... */
3943 				prom_panic("tcp_rput_data");
3944 			}
3945 			goto pre_swnd_update;
3946 		}
3947 		assert(mp2 != tcp->tcp_xmit_tail);
3948 	}
3949 	if (tcp->tcp_unsent) {
3950 		flags |= TH_XMIT_NEEDED;
3951 	}
3952 pre_swnd_update:
3953 	tcp->tcp_xmit_head = mp1;
3954 swnd_update:
3955 	/*
3956 	 * The following check is different from most other implementations.
3957 	 * For bi-directional transfer, when segments are dropped, the
3958 	 * "normal" check will not accept a window update in those
3959 	 * retransmitted segemnts.  Failing to do that, TCP may send out
3960 	 * segments which are outside receiver's window.  As TCP accepts
3961 	 * the ack in those retransmitted segments, if the window update in
3962 	 * the same segment is not accepted, TCP will incorrectly calculates
3963 	 * that it can send more segments.  This can create a deadlock
3964 	 * with the receiver if its window becomes zero.
3965 	 */
3966 	if (SEQ_LT(tcp->tcp_swl2, seg_ack) ||
3967 	    SEQ_LT(tcp->tcp_swl1, seg_seq) ||
3968 	    (tcp->tcp_swl1 == seg_seq && new_swnd > tcp->tcp_swnd)) {
3969 		/*
3970 		 * The criteria for update is:
3971 		 *
3972 		 * 1. the segment acknowledges some data.  Or
3973 		 * 2. the segment is new, i.e. it has a higher seq num. Or
3974 		 * 3. the segment is not old and the advertised window is
3975 		 * larger than the previous advertised window.
3976 		 */
3977 		if (tcp->tcp_unsent && new_swnd > tcp->tcp_swnd)
3978 			flags |= TH_XMIT_NEEDED;
3979 		tcp->tcp_swnd = new_swnd;
3980 		if (new_swnd > tcp->tcp_max_swnd)
3981 			tcp->tcp_max_swnd = new_swnd;
3982 		tcp->tcp_swl1 = seg_seq;
3983 		tcp->tcp_swl2 = seg_ack;
3984 	}
3985 est:
3986 	if (tcp->tcp_state > TCPS_ESTABLISHED) {
3987 		switch (tcp->tcp_state) {
3988 		case TCPS_FIN_WAIT_1:
3989 			if (tcp->tcp_fin_acked) {
3990 				tcp->tcp_state = TCPS_FIN_WAIT_2;
3991 				/*
3992 				 * We implement the non-standard BSD/SunOS
3993 				 * FIN_WAIT_2 flushing algorithm.
3994 				 * If there is no user attached to this
3995 				 * TCP endpoint, then this TCP struct
3996 				 * could hang around forever in FIN_WAIT_2
3997 				 * state if the peer forgets to send us
3998 				 * a FIN.  To prevent this, we wait only
3999 				 * 2*MSL (a convenient time value) for
4000 				 * the FIN to arrive.  If it doesn't show up,
4001 				 * we flush the TCP endpoint.  This algorithm,
4002 				 * though a violation of RFC-793, has worked
4003 				 * for over 10 years in BSD systems.
4004 				 * Note: SunOS 4.x waits 675 seconds before
4005 				 * flushing the FIN_WAIT_2 connection.
4006 				 */
4007 				TCP_TIMER_RESTART(tcp,
4008 				    tcp_fin_wait_2_flush_interval);
4009 			}
4010 			break;
4011 		case TCPS_FIN_WAIT_2:
4012 			break;	/* Shutdown hook? */
4013 		case TCPS_LAST_ACK:
4014 			freemsg(mp);
4015 			if (tcp->tcp_fin_acked) {
4016 				(void) tcp_clean_death(sock_id, tcp, 0);
4017 				return;
4018 			}
4019 			goto xmit_check;
4020 		case TCPS_CLOSING:
4021 			if (tcp->tcp_fin_acked) {
4022 				tcp->tcp_state = TCPS_TIME_WAIT;
4023 				tcp_time_wait_append(tcp);
4024 				TCP_TIMER_RESTART(tcp, tcp_time_wait_interval);
4025 			}
4026 			/*FALLTHRU*/
4027 		case TCPS_CLOSE_WAIT:
4028 			freemsg(mp);
4029 			goto xmit_check;
4030 		default:
4031 			assert(tcp->tcp_state != TCPS_TIME_WAIT);
4032 			break;
4033 		}
4034 	}
4035 	if (flags & TH_FIN) {
4036 		/* Make sure we ack the fin */
4037 		flags |= TH_ACK_NEEDED;
4038 		if (!tcp->tcp_fin_rcvd) {
4039 			tcp->tcp_fin_rcvd = B_TRUE;
4040 			tcp->tcp_rnxt++;
4041 			U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
4042 
4043 			switch (tcp->tcp_state) {
4044 			case TCPS_SYN_RCVD:
4045 			case TCPS_ESTABLISHED:
4046 				tcp->tcp_state = TCPS_CLOSE_WAIT;
4047 				/* Keepalive? */
4048 				break;
4049 			case TCPS_FIN_WAIT_1:
4050 				if (!tcp->tcp_fin_acked) {
4051 					tcp->tcp_state = TCPS_CLOSING;
4052 					break;
4053 				}
4054 				/* FALLTHRU */
4055 			case TCPS_FIN_WAIT_2:
4056 				tcp->tcp_state = TCPS_TIME_WAIT;
4057 				tcp_time_wait_append(tcp);
4058 				TCP_TIMER_RESTART(tcp, tcp_time_wait_interval);
4059 				if (seg_len) {
4060 					/*
4061 					 * implies data piggybacked on FIN.
4062 					 * break to handle data.
4063 					 */
4064 					break;
4065 				}
4066 				freemsg(mp);
4067 				goto ack_check;
4068 			}
4069 		}
4070 	}
4071 	if (mp == NULL)
4072 		goto xmit_check;
4073 	if (seg_len == 0) {
4074 		freemsg(mp);
4075 		goto xmit_check;
4076 	}
4077 	if (mp->b_rptr == mp->b_wptr) {
4078 		/*
4079 		 * The header has been consumed, so we remove the
4080 		 * zero-length mblk here.
4081 		 */
4082 		mp1 = mp;
4083 		mp = mp->b_cont;
4084 		freeb(mp1);
4085 	}
4086 	/*
4087 	 * ACK every other segments, unless the input queue is empty
4088 	 * as we don't have a timer available.
4089 	 */
4090 	if (++tcp->tcp_rack_cnt == 2 || sockets[sock_id].inq == NULL) {
4091 		flags |= TH_ACK_NEEDED;
4092 		tcp->tcp_rack_cnt = 0;
4093 	}
4094 	tcp->tcp_rnxt += seg_len;
4095 	U32_TO_ABE32(tcp->tcp_rnxt, tcp->tcp_tcph->th_ack);
4096 
4097 	/* Update SACK list */
4098 	if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
4099 		tcp_sack_remove(tcp->tcp_sack_list, tcp->tcp_rnxt,
4100 		    &(tcp->tcp_num_sack_blk));
4101 	}
4102 
4103 	if (tcp->tcp_listener) {
4104 		/*
4105 		 * Side queue inbound data until the accept happens.
4106 		 * tcp_accept/tcp_rput drains this when the accept happens.
4107 		 */
4108 		tcp_rcv_enqueue(tcp, mp, seg_len);
4109 	} else {
4110 		/* Just queue the data until the app calls read. */
4111 		tcp_rcv_enqueue(tcp, mp, seg_len);
4112 		/*
4113 		 * Make sure the timer is running if we have data waiting
4114 		 * for a push bit. This provides resiliency against
4115 		 * implementations that do not correctly generate push bits.
4116 		 */
4117 		if (tcp->tcp_rcv_list != NULL)
4118 			flags |= TH_TIMER_NEEDED;
4119 	}
4120 
4121 xmit_check:
4122 	/* Is there anything left to do? */
4123 	if ((flags & (TH_REXMIT_NEEDED|TH_XMIT_NEEDED|TH_ACK_NEEDED|
4124 	    TH_NEED_SACK_REXMIT|TH_LIMIT_XMIT|TH_TIMER_NEEDED)) == 0)
4125 		return;
4126 
4127 	/* Any transmit work to do and a non-zero window? */
4128 	if ((flags & (TH_REXMIT_NEEDED|TH_XMIT_NEEDED|TH_NEED_SACK_REXMIT|
4129 	    TH_LIMIT_XMIT)) && tcp->tcp_swnd != 0) {
4130 		if (flags & TH_REXMIT_NEEDED) {
4131 			uint32_t snd_size = tcp->tcp_snxt - tcp->tcp_suna;
4132 
4133 			if (snd_size > mss)
4134 				snd_size = mss;
4135 			if (snd_size > tcp->tcp_swnd)
4136 				snd_size = tcp->tcp_swnd;
4137 			mp1 = tcp_xmit_mp(tcp, tcp->tcp_xmit_head, snd_size,
4138 			    NULL, NULL, tcp->tcp_suna, B_TRUE, &snd_size,
4139 			    B_TRUE);
4140 
4141 			if (mp1 != NULL) {
4142 				/* use uintptr_t to suppress the gcc warning */
4143 				tcp->tcp_xmit_head->b_prev =
4144 				    (mblk_t *)(uintptr_t)prom_gettime();
4145 				tcp->tcp_csuna = tcp->tcp_snxt;
4146 				BUMP_MIB(tcp_mib.tcpRetransSegs);
4147 				UPDATE_MIB(tcp_mib.tcpRetransBytes, snd_size);
4148 				(void) ipv4_tcp_output(sock_id, mp1);
4149 				freeb(mp1);
4150 			}
4151 		}
4152 		if (flags & TH_NEED_SACK_REXMIT) {
4153 			if (tcp_sack_rxmit(tcp, sock_id) != 0) {
4154 				flags |= TH_XMIT_NEEDED;
4155 			}
4156 		}
4157 		/*
4158 		 * For TH_LIMIT_XMIT, tcp_wput_data() is called to send
4159 		 * out new segment.  Note that tcp_rexmit should not be
4160 		 * set, otherwise TH_LIMIT_XMIT should not be set.
4161 		 */
4162 		if (flags & (TH_XMIT_NEEDED|TH_LIMIT_XMIT)) {
4163 			if (!tcp->tcp_rexmit) {
4164 				tcp_wput_data(tcp, NULL, sock_id);
4165 			} else {
4166 				tcp_ss_rexmit(tcp, sock_id);
4167 			}
4168 			/*
4169 			 * The TCP could be closed in tcp_state_wait via
4170 			 * tcp_wput_data (tcp_ss_rexmit could call
4171 			 * tcp_wput_data as well).
4172 			 */
4173 			if (sockets[sock_id].pcb == NULL)
4174 				return;
4175 		}
4176 		/*
4177 		 * Adjust tcp_cwnd back to normal value after sending
4178 		 * new data segments.
4179 		 */
4180 		if (flags & TH_LIMIT_XMIT) {
4181 			tcp->tcp_cwnd -= mss << (tcp->tcp_dupack_cnt - 1);
4182 		}
4183 
4184 		/* Anything more to do? */
4185 		if ((flags & (TH_ACK_NEEDED|TH_TIMER_NEEDED)) == 0)
4186 			return;
4187 	}
4188 ack_check:
4189 	if (flags & TH_ACK_NEEDED) {
4190 		/*
4191 		 * Time to send an ack for some reason.
4192 		 */
4193 		if ((mp1 = tcp_ack_mp(tcp)) != NULL) {
4194 			TCP_DUMP_PACKET("tcp_rput_data: ack mp", mp1);
4195 			(void) ipv4_tcp_output(sock_id, mp1);
4196 			BUMP_MIB(tcp_mib.tcpOutAck);
4197 			freeb(mp1);
4198 		}
4199 	}
4200 }
4201 
4202 /*
4203  * tcp_ss_rexmit() is called in tcp_rput_data() to do slow start
4204  * retransmission after a timeout.
4205  *
4206  * To limit the number of duplicate segments, we limit the number of segment
4207  * to be sent in one time to tcp_snd_burst, the burst variable.
4208  */
4209 static void
4210 tcp_ss_rexmit(tcp_t *tcp, int sock_id)
4211 {
4212 	uint32_t	snxt;
4213 	uint32_t	smax;
4214 	int32_t		win;
4215 	int32_t		mss;
4216 	int32_t		off;
4217 	int32_t		burst = tcp->tcp_snd_burst;
4218 	mblk_t		*snxt_mp;
4219 
4220 	/*
4221 	 * Note that tcp_rexmit can be set even though TCP has retransmitted
4222 	 * all unack'ed segments.
4223 	 */
4224 	if (SEQ_LT(tcp->tcp_rexmit_nxt, tcp->tcp_rexmit_max)) {
4225 		smax = tcp->tcp_rexmit_max;
4226 		snxt = tcp->tcp_rexmit_nxt;
4227 		if (SEQ_LT(snxt, tcp->tcp_suna)) {
4228 			snxt = tcp->tcp_suna;
4229 		}
4230 		win = MIN(tcp->tcp_cwnd, tcp->tcp_swnd);
4231 		win -= snxt - tcp->tcp_suna;
4232 		mss = tcp->tcp_mss;
4233 		snxt_mp = tcp_get_seg_mp(tcp, snxt, &off);
4234 
4235 		while (SEQ_LT(snxt, smax) && (win > 0) &&
4236 		    (burst > 0) && (snxt_mp != NULL)) {
4237 			mblk_t	*xmit_mp;
4238 			mblk_t	*old_snxt_mp = snxt_mp;
4239 			uint32_t cnt = mss;
4240 
4241 			if (win < cnt) {
4242 				cnt = win;
4243 			}
4244 			if (SEQ_GT(snxt + cnt, smax)) {
4245 				cnt = smax - snxt;
4246 			}
4247 			xmit_mp = tcp_xmit_mp(tcp, snxt_mp, cnt, &off,
4248 			    &snxt_mp, snxt, B_TRUE, &cnt, B_TRUE);
4249 
4250 			if (xmit_mp == NULL)
4251 				return;
4252 
4253 			(void) ipv4_tcp_output(sock_id, xmit_mp);
4254 			freeb(xmit_mp);
4255 
4256 			snxt += cnt;
4257 			win -= cnt;
4258 			/*
4259 			 * Update the send timestamp to avoid false
4260 			 * retransmission.
4261 			 * Note. use uintptr_t to suppress the gcc warning.
4262 			 */
4263 			old_snxt_mp->b_prev =
4264 			    (mblk_t *)(uintptr_t)prom_gettime();
4265 			BUMP_MIB(tcp_mib.tcpRetransSegs);
4266 			UPDATE_MIB(tcp_mib.tcpRetransBytes, cnt);
4267 
4268 			tcp->tcp_rexmit_nxt = snxt;
4269 			burst--;
4270 		}
4271 		/*
4272 		 * If we have transmitted all we have at the time
4273 		 * we started the retranmission, we can leave
4274 		 * the rest of the job to tcp_wput_data().  But we
4275 		 * need to check the send window first.  If the
4276 		 * win is not 0, go on with tcp_wput_data().
4277 		 */
4278 		if (SEQ_LT(snxt, smax) || win == 0) {
4279 			return;
4280 		}
4281 	}
4282 	/* Only call tcp_wput_data() if there is data to be sent. */
4283 	if (tcp->tcp_unsent) {
4284 		tcp_wput_data(tcp, NULL, sock_id);
4285 	}
4286 }
4287 
4288 /*
4289  * tcp_timer is the timer service routine.  It handles all timer events for
4290  * a tcp instance except keepalives.  It figures out from the state of the
4291  * tcp instance what kind of action needs to be done at the time it is called.
4292  */
4293 static void
4294 tcp_timer(tcp_t	*tcp, int sock_id)
4295 {
4296 	mblk_t		*mp;
4297 	uint32_t	first_threshold;
4298 	uint32_t	second_threshold;
4299 	uint32_t	ms;
4300 	uint32_t	mss;
4301 
4302 	first_threshold =  tcp->tcp_first_timer_threshold;
4303 	second_threshold = tcp->tcp_second_timer_threshold;
4304 	switch (tcp->tcp_state) {
4305 	case TCPS_IDLE:
4306 	case TCPS_BOUND:
4307 	case TCPS_LISTEN:
4308 		return;
4309 	case TCPS_SYN_RCVD:
4310 	case TCPS_SYN_SENT:
4311 		first_threshold =  tcp->tcp_first_ctimer_threshold;
4312 		second_threshold = tcp->tcp_second_ctimer_threshold;
4313 		break;
4314 	case TCPS_ESTABLISHED:
4315 	case TCPS_FIN_WAIT_1:
4316 	case TCPS_CLOSING:
4317 	case TCPS_CLOSE_WAIT:
4318 	case TCPS_LAST_ACK:
4319 		/* If we have data to rexmit */
4320 		if (tcp->tcp_suna != tcp->tcp_snxt) {
4321 			int32_t time_to_wait;
4322 
4323 			BUMP_MIB(tcp_mib.tcpTimRetrans);
4324 			if (tcp->tcp_xmit_head == NULL)
4325 				break;
4326 			/* use uintptr_t to suppress the gcc warning */
4327 			time_to_wait = (int32_t)(prom_gettime() -
4328 			    (uint32_t)(uintptr_t)tcp->tcp_xmit_head->b_prev);
4329 			time_to_wait = tcp->tcp_rto - time_to_wait;
4330 			if (time_to_wait > 0) {
4331 				/*
4332 				 * Timer fired too early, so restart it.
4333 				 */
4334 				TCP_TIMER_RESTART(tcp, time_to_wait);
4335 				return;
4336 			}
4337 			/*
4338 			 * When we probe zero windows, we force the swnd open.
4339 			 * If our peer acks with a closed window swnd will be
4340 			 * set to zero by tcp_rput(). As long as we are
4341 			 * receiving acks tcp_rput will
4342 			 * reset 'tcp_ms_we_have_waited' so as not to trip the
4343 			 * first and second interval actions.  NOTE: the timer
4344 			 * interval is allowed to continue its exponential
4345 			 * backoff.
4346 			 */
4347 			if (tcp->tcp_swnd == 0 || tcp->tcp_zero_win_probe) {
4348 				DEBUG_1("tcp_timer (%d): zero win", sock_id);
4349 				break;
4350 			} else {
4351 				/*
4352 				 * After retransmission, we need to do
4353 				 * slow start.  Set the ssthresh to one
4354 				 * half of current effective window and
4355 				 * cwnd to one MSS.  Also reset
4356 				 * tcp_cwnd_cnt.
4357 				 *
4358 				 * Note that if tcp_ssthresh is reduced because
4359 				 * of ECN, do not reduce it again unless it is
4360 				 * already one window of data away (tcp_cwr
4361 				 * should then be cleared) or this is a
4362 				 * timeout for a retransmitted segment.
4363 				 */
4364 				uint32_t npkt;
4365 
4366 				if (!tcp->tcp_cwr || tcp->tcp_rexmit) {
4367 					npkt = (MIN((tcp->tcp_timer_backoff ?
4368 					    tcp->tcp_cwnd_ssthresh :
4369 					    tcp->tcp_cwnd),
4370 					    tcp->tcp_swnd) >> 1) /
4371 					    tcp->tcp_mss;
4372 					if (npkt < 2)
4373 						npkt = 2;
4374 					tcp->tcp_cwnd_ssthresh = npkt *
4375 					    tcp->tcp_mss;
4376 				}
4377 				tcp->tcp_cwnd = tcp->tcp_mss;
4378 				tcp->tcp_cwnd_cnt = 0;
4379 				if (tcp->tcp_ecn_ok) {
4380 					tcp->tcp_cwr = B_TRUE;
4381 					tcp->tcp_cwr_snd_max = tcp->tcp_snxt;
4382 					tcp->tcp_ecn_cwr_sent = B_FALSE;
4383 				}
4384 			}
4385 			break;
4386 		}
4387 		/*
4388 		 * We have something to send yet we cannot send.  The
4389 		 * reason can be:
4390 		 *
4391 		 * 1. Zero send window: we need to do zero window probe.
4392 		 * 2. Zero cwnd: because of ECN, we need to "clock out
4393 		 * segments.
4394 		 * 3. SWS avoidance: receiver may have shrunk window,
4395 		 * reset our knowledge.
4396 		 *
4397 		 * Note that condition 2 can happen with either 1 or
4398 		 * 3.  But 1 and 3 are exclusive.
4399 		 */
4400 		if (tcp->tcp_unsent != 0) {
4401 			if (tcp->tcp_cwnd == 0) {
4402 				/*
4403 				 * Set tcp_cwnd to 1 MSS so that a
4404 				 * new segment can be sent out.  We
4405 				 * are "clocking out" new data when
4406 				 * the network is really congested.
4407 				 */
4408 				assert(tcp->tcp_ecn_ok);
4409 				tcp->tcp_cwnd = tcp->tcp_mss;
4410 			}
4411 			if (tcp->tcp_swnd == 0) {
4412 				/* Extend window for zero window probe */
4413 				tcp->tcp_swnd++;
4414 				tcp->tcp_zero_win_probe = B_TRUE;
4415 				BUMP_MIB(tcp_mib.tcpOutWinProbe);
4416 			} else {
4417 				/*
4418 				 * Handle timeout from sender SWS avoidance.
4419 				 * Reset our knowledge of the max send window
4420 				 * since the receiver might have reduced its
4421 				 * receive buffer.  Avoid setting tcp_max_swnd
4422 				 * to one since that will essentially disable
4423 				 * the SWS checks.
4424 				 *
4425 				 * Note that since we don't have a SWS
4426 				 * state variable, if the timeout is set
4427 				 * for ECN but not for SWS, this
4428 				 * code will also be executed.  This is
4429 				 * fine as tcp_max_swnd is updated
4430 				 * constantly and it will not affect
4431 				 * anything.
4432 				 */
4433 				tcp->tcp_max_swnd = MAX(tcp->tcp_swnd, 2);
4434 			}
4435 			tcp_wput_data(tcp, NULL, sock_id);
4436 			return;
4437 		}
4438 		/* Is there a FIN that needs to be to re retransmitted? */
4439 		if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
4440 		    !tcp->tcp_fin_acked)
4441 			break;
4442 		/* Nothing to do, return without restarting timer. */
4443 		return;
4444 	case TCPS_FIN_WAIT_2:
4445 		/*
4446 		 * User closed the TCP endpoint and peer ACK'ed our FIN.
4447 		 * We waited some time for for peer's FIN, but it hasn't
4448 		 * arrived.  We flush the connection now to avoid
4449 		 * case where the peer has rebooted.
4450 		 */
4451 		/* FALLTHRU */
4452 	case TCPS_TIME_WAIT:
4453 		(void) tcp_clean_death(sock_id, tcp, 0);
4454 		return;
4455 	default:
4456 		DEBUG_3("tcp_timer (%d): strange state (%d) %s", sock_id,
4457 		    tcp->tcp_state, tcp_display(tcp, NULL,
4458 		    DISP_PORT_ONLY));
4459 		return;
4460 	}
4461 	if ((ms = tcp->tcp_ms_we_have_waited) > second_threshold) {
4462 		/*
4463 		 * For zero window probe, we need to send indefinitely,
4464 		 * unless we have not heard from the other side for some
4465 		 * time...
4466 		 */
4467 		if ((tcp->tcp_zero_win_probe == 0) ||
4468 		    ((prom_gettime() - tcp->tcp_last_recv_time) >
4469 		    second_threshold)) {
4470 			BUMP_MIB(tcp_mib.tcpTimRetransDrop);
4471 			/*
4472 			 * If TCP is in SYN_RCVD state, send back a
4473 			 * RST|ACK as BSD does.  Note that tcp_zero_win_probe
4474 			 * should be zero in TCPS_SYN_RCVD state.
4475 			 */
4476 			if (tcp->tcp_state == TCPS_SYN_RCVD) {
4477 				tcp_xmit_ctl("tcp_timer: RST sent on timeout "
4478 				    "in SYN_RCVD",
4479 				    tcp, NULL, tcp->tcp_snxt,
4480 				    tcp->tcp_rnxt, TH_RST | TH_ACK, 0, sock_id);
4481 			}
4482 			(void) tcp_clean_death(sock_id, tcp,
4483 			    tcp->tcp_client_errno ?
4484 			    tcp->tcp_client_errno : ETIMEDOUT);
4485 			return;
4486 		} else {
4487 			/*
4488 			 * Set tcp_ms_we_have_waited to second_threshold
4489 			 * so that in next timeout, we will do the above
4490 			 * check (lbolt - tcp_last_recv_time).  This is
4491 			 * also to avoid overflow.
4492 			 *
4493 			 * We don't need to decrement tcp_timer_backoff
4494 			 * to avoid overflow because it will be decremented
4495 			 * later if new timeout value is greater than
4496 			 * tcp_rexmit_interval_max.  In the case when
4497 			 * tcp_rexmit_interval_max is greater than
4498 			 * second_threshold, it means that we will wait
4499 			 * longer than second_threshold to send the next
4500 			 * window probe.
4501 			 */
4502 			tcp->tcp_ms_we_have_waited = second_threshold;
4503 		}
4504 	} else if (ms > first_threshold && tcp->tcp_rtt_sa != 0) {
4505 		/*
4506 		 * We have been retransmitting for too long...  The RTT
4507 		 * we calculated is probably incorrect.  Reinitialize it.
4508 		 * Need to compensate for 0 tcp_rtt_sa.  Reset
4509 		 * tcp_rtt_update so that we won't accidentally cache a
4510 		 * bad value.  But only do this if this is not a zero
4511 		 * window probe.
4512 		 */
4513 		if (tcp->tcp_zero_win_probe == 0) {
4514 			tcp->tcp_rtt_sd += (tcp->tcp_rtt_sa >> 3) +
4515 			    (tcp->tcp_rtt_sa >> 5);
4516 			tcp->tcp_rtt_sa = 0;
4517 			tcp->tcp_rtt_update = 0;
4518 		}
4519 	}
4520 	tcp->tcp_timer_backoff++;
4521 	if ((ms = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
4522 	    tcp_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5)) <
4523 	    tcp_rexmit_interval_min) {
4524 		/*
4525 		 * This means the original RTO is tcp_rexmit_interval_min.
4526 		 * So we will use tcp_rexmit_interval_min as the RTO value
4527 		 * and do the backoff.
4528 		 */
4529 		ms = tcp_rexmit_interval_min << tcp->tcp_timer_backoff;
4530 	} else {
4531 		ms <<= tcp->tcp_timer_backoff;
4532 	}
4533 	if (ms > tcp_rexmit_interval_max) {
4534 		ms = tcp_rexmit_interval_max;
4535 		/*
4536 		 * ms is at max, decrement tcp_timer_backoff to avoid
4537 		 * overflow.
4538 		 */
4539 		tcp->tcp_timer_backoff--;
4540 	}
4541 	tcp->tcp_ms_we_have_waited += ms;
4542 	if (tcp->tcp_zero_win_probe == 0) {
4543 		tcp->tcp_rto = ms;
4544 	}
4545 	TCP_TIMER_RESTART(tcp, ms);
4546 	/*
4547 	 * This is after a timeout and tcp_rto is backed off.  Set
4548 	 * tcp_set_timer to 1 so that next time RTO is updated, we will
4549 	 * restart the timer with a correct value.
4550 	 */
4551 	tcp->tcp_set_timer = 1;
4552 	mss = tcp->tcp_snxt - tcp->tcp_suna;
4553 	if (mss > tcp->tcp_mss)
4554 		mss = tcp->tcp_mss;
4555 	if (mss > tcp->tcp_swnd && tcp->tcp_swnd != 0)
4556 		mss = tcp->tcp_swnd;
4557 
4558 	if ((mp = tcp->tcp_xmit_head) != NULL) {
4559 		/* use uintptr_t to suppress the gcc warning */
4560 		mp->b_prev = (mblk_t *)(uintptr_t)prom_gettime();
4561 	}
4562 	mp = tcp_xmit_mp(tcp, mp, mss, NULL, NULL, tcp->tcp_suna, B_TRUE, &mss,
4563 	    B_TRUE);
4564 	if (mp == NULL)
4565 		return;
4566 	tcp->tcp_csuna = tcp->tcp_snxt;
4567 	BUMP_MIB(tcp_mib.tcpRetransSegs);
4568 	UPDATE_MIB(tcp_mib.tcpRetransBytes, mss);
4569 	/* Dump the packet when debugging. */
4570 	TCP_DUMP_PACKET("tcp_timer", mp);
4571 
4572 	(void) ipv4_tcp_output(sock_id, mp);
4573 	freeb(mp);
4574 
4575 	/*
4576 	 * When slow start after retransmission begins, start with
4577 	 * this seq no.  tcp_rexmit_max marks the end of special slow
4578 	 * start phase.  tcp_snd_burst controls how many segments
4579 	 * can be sent because of an ack.
4580 	 */
4581 	tcp->tcp_rexmit_nxt = tcp->tcp_suna;
4582 	tcp->tcp_snd_burst = TCP_CWND_SS;
4583 	if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
4584 	    (tcp->tcp_unsent == 0)) {
4585 		tcp->tcp_rexmit_max = tcp->tcp_fss;
4586 	} else {
4587 		tcp->tcp_rexmit_max = tcp->tcp_snxt;
4588 	}
4589 	tcp->tcp_rexmit = B_TRUE;
4590 	tcp->tcp_dupack_cnt = 0;
4591 
4592 	/*
4593 	 * Remove all rexmit SACK blk to start from fresh.
4594 	 */
4595 	if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
4596 		TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list);
4597 		tcp->tcp_num_notsack_blk = 0;
4598 		tcp->tcp_cnt_notsack_list = 0;
4599 	}
4600 }
4601 
4602 /*
4603  * The TCP normal data output path.
4604  * NOTE: the logic of the fast path is duplicated from this function.
4605  */
4606 static void
4607 tcp_wput_data(tcp_t *tcp, mblk_t *mp, int sock_id)
4608 {
4609 	int		len;
4610 	mblk_t		*local_time;
4611 	mblk_t		*mp1;
4612 	uchar_t		*rptr;
4613 	uint32_t	snxt;
4614 	int		tail_unsent;
4615 	int		tcpstate;
4616 	int		usable = 0;
4617 	mblk_t		*xmit_tail;
4618 	int32_t		num_burst_seg;
4619 	int32_t		mss;
4620 	int32_t		num_sack_blk = 0;
4621 	int32_t		tcp_hdr_len;
4622 	ipaddr_t	*dst;
4623 	ipaddr_t	*src;
4624 
4625 #ifdef DEBUG
4626 	printf("tcp_wput_data(%d) ##############################\n", sock_id);
4627 #endif
4628 	tcpstate = tcp->tcp_state;
4629 	if (mp == NULL) {
4630 		/* Really tacky... but we need this for detached closes. */
4631 		len = tcp->tcp_unsent;
4632 		goto data_null;
4633 	}
4634 
4635 	/*
4636 	 * Don't allow data after T_ORDREL_REQ or T_DISCON_REQ,
4637 	 * or before a connection attempt has begun.
4638 	 *
4639 	 * The following should not happen in inetboot....
4640 	 */
4641 	if (tcpstate < TCPS_SYN_SENT || tcpstate > TCPS_CLOSE_WAIT ||
4642 	    (tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
4643 		if ((tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
4644 			printf("tcp_wput_data: data after ordrel, %s\n",
4645 			    tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
4646 		}
4647 		freemsg(mp);
4648 		return;
4649 	}
4650 
4651 	/* Strip empties */
4652 	for (;;) {
4653 		assert((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
4654 		    (uintptr_t)INT_MAX);
4655 		len = (int)(mp->b_wptr - mp->b_rptr);
4656 		if (len > 0)
4657 			break;
4658 		mp1 = mp;
4659 		mp = mp->b_cont;
4660 		freeb(mp1);
4661 		if (mp == NULL) {
4662 			return;
4663 		}
4664 	}
4665 
4666 	/* If we are the first on the list ... */
4667 	if (tcp->tcp_xmit_head == NULL) {
4668 		tcp->tcp_xmit_head = mp;
4669 		tcp->tcp_xmit_tail = mp;
4670 		tcp->tcp_xmit_tail_unsent = len;
4671 	} else {
4672 		tcp->tcp_xmit_last->b_cont = mp;
4673 		len += tcp->tcp_unsent;
4674 	}
4675 
4676 	/* Tack on however many more positive length mblks we have */
4677 	if ((mp1 = mp->b_cont) != NULL) {
4678 		do {
4679 			int tlen;
4680 			assert((uintptr_t)(mp1->b_wptr -
4681 			    mp1->b_rptr) <= (uintptr_t)INT_MAX);
4682 			tlen = (int)(mp1->b_wptr - mp1->b_rptr);
4683 			if (tlen <= 0) {
4684 				mp->b_cont = mp1->b_cont;
4685 				freeb(mp1);
4686 			} else {
4687 				len += tlen;
4688 				mp = mp1;
4689 			}
4690 		} while ((mp1 = mp->b_cont) != NULL);
4691 	}
4692 	tcp->tcp_xmit_last = mp;
4693 	tcp->tcp_unsent = len;
4694 
4695 data_null:
4696 	snxt = tcp->tcp_snxt;
4697 	xmit_tail = tcp->tcp_xmit_tail;
4698 	tail_unsent = tcp->tcp_xmit_tail_unsent;
4699 
4700 	/*
4701 	 * Note that tcp_mss has been adjusted to take into account the
4702 	 * timestamp option if applicable.  Because SACK options do not
4703 	 * appear in every TCP segments and they are of variable lengths,
4704 	 * they cannot be included in tcp_mss.  Thus we need to calculate
4705 	 * the actual segment length when we need to send a segment which
4706 	 * includes SACK options.
4707 	 */
4708 	if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
4709 		int32_t	opt_len;
4710 
4711 		num_sack_blk = MIN(tcp->tcp_max_sack_blk,
4712 		    tcp->tcp_num_sack_blk);
4713 		opt_len = num_sack_blk * sizeof (sack_blk_t) + TCPOPT_NOP_LEN *
4714 		    2 + TCPOPT_HEADER_LEN;
4715 		mss = tcp->tcp_mss - opt_len;
4716 		tcp_hdr_len = tcp->tcp_hdr_len + opt_len;
4717 	} else {
4718 		mss = tcp->tcp_mss;
4719 		tcp_hdr_len = tcp->tcp_hdr_len;
4720 	}
4721 
4722 	if ((tcp->tcp_suna == snxt) &&
4723 	    (prom_gettime() - tcp->tcp_last_recv_time) >= tcp->tcp_rto) {
4724 		tcp->tcp_cwnd = MIN(tcp_slow_start_after_idle * mss,
4725 		    MIN(4 * mss, MAX(2 * mss, 4380 / mss * mss)));
4726 	}
4727 	if (tcpstate == TCPS_SYN_RCVD) {
4728 		/*
4729 		 * The three-way connection establishment handshake is not
4730 		 * complete yet. We want to queue the data for transmission
4731 		 * after entering ESTABLISHED state (RFC793). Setting usable to
4732 		 * zero cause a jump to "done" label effectively leaving data
4733 		 * on the queue.
4734 		 */
4735 
4736 		usable = 0;
4737 	} else {
4738 		int usable_r = tcp->tcp_swnd;
4739 
4740 		/*
4741 		 * In the special case when cwnd is zero, which can only
4742 		 * happen if the connection is ECN capable, return now.
4743 		 * New segments is sent using tcp_timer().  The timer
4744 		 * is set in tcp_rput_data().
4745 		 */
4746 		if (tcp->tcp_cwnd == 0) {
4747 			/*
4748 			 * Note that tcp_cwnd is 0 before 3-way handshake is
4749 			 * finished.
4750 			 */
4751 			assert(tcp->tcp_ecn_ok ||
4752 			    tcp->tcp_state < TCPS_ESTABLISHED);
4753 			return;
4754 		}
4755 
4756 		/* usable = MIN(swnd, cwnd) - unacked_bytes */
4757 		if (usable_r > tcp->tcp_cwnd)
4758 			usable_r = tcp->tcp_cwnd;
4759 
4760 		/* NOTE: trouble if xmitting while SYN not acked? */
4761 		usable_r -= snxt;
4762 		usable_r += tcp->tcp_suna;
4763 
4764 		/* usable = MIN(usable, unsent) */
4765 		if (usable_r > len)
4766 			usable_r = len;
4767 
4768 		/* usable = MAX(usable, {1 for urgent, 0 for data}) */
4769 		if (usable_r != 0)
4770 			usable = usable_r;
4771 	}
4772 
4773 	/* use uintptr_t to suppress the gcc warning */
4774 	local_time = (mblk_t *)(uintptr_t)prom_gettime();
4775 
4776 	/*
4777 	 * "Our" Nagle Algorithm.  This is not the same as in the old
4778 	 * BSD.  This is more in line with the true intent of Nagle.
4779 	 *
4780 	 * The conditions are:
4781 	 * 1. The amount of unsent data (or amount of data which can be
4782 	 *    sent, whichever is smaller) is less than Nagle limit.
4783 	 * 2. The last sent size is also less than Nagle limit.
4784 	 * 3. There is unack'ed data.
4785 	 * 4. Urgent pointer is not set.  Send urgent data ignoring the
4786 	 *    Nagle algorithm.  This reduces the probability that urgent
4787 	 *    bytes get "merged" together.
4788 	 * 5. The app has not closed the connection.  This eliminates the
4789 	 *    wait time of the receiving side waiting for the last piece of
4790 	 *    (small) data.
4791 	 *
4792 	 * If all are satisified, exit without sending anything.  Note
4793 	 * that Nagle limit can be smaller than 1 MSS.  Nagle limit is
4794 	 * the smaller of 1 MSS and global tcp_naglim_def (default to be
4795 	 * 4095).
4796 	 */
4797 	if (usable < (int)tcp->tcp_naglim &&
4798 	    tcp->tcp_naglim > tcp->tcp_last_sent_len &&
4799 	    snxt != tcp->tcp_suna &&
4800 	    !(tcp->tcp_valid_bits & TCP_URG_VALID))
4801 		goto done;
4802 
4803 	num_burst_seg = tcp->tcp_snd_burst;
4804 	for (;;) {
4805 		tcph_t		*tcph;
4806 		mblk_t		*new_mp;
4807 
4808 		if (num_burst_seg-- == 0)
4809 			goto done;
4810 
4811 		len = mss;
4812 		if (len > usable) {
4813 			len = usable;
4814 			if (len <= 0) {
4815 				/* Terminate the loop */
4816 				goto done;
4817 			}
4818 			/*
4819 			 * Sender silly-window avoidance.
4820 			 * Ignore this if we are going to send a
4821 			 * zero window probe out.
4822 			 *
4823 			 * TODO: force data into microscopic window ??
4824 			 *	==> (!pushed || (unsent > usable))
4825 			 */
4826 			if (len < (tcp->tcp_max_swnd >> 1) &&
4827 			    (tcp->tcp_unsent - (snxt - tcp->tcp_snxt)) > len &&
4828 			    !((tcp->tcp_valid_bits & TCP_URG_VALID) &&
4829 			    len == 1) && (! tcp->tcp_zero_win_probe)) {
4830 				/*
4831 				 * If the retransmit timer is not running
4832 				 * we start it so that we will retransmit
4833 				 * in the case when the the receiver has
4834 				 * decremented the window.
4835 				 */
4836 				if (snxt == tcp->tcp_snxt &&
4837 				    snxt == tcp->tcp_suna) {
4838 					/*
4839 					 * We are not supposed to send
4840 					 * anything.  So let's wait a little
4841 					 * bit longer before breaking SWS
4842 					 * avoidance.
4843 					 *
4844 					 * What should the value be?
4845 					 * Suggestion: MAX(init rexmit time,
4846 					 * tcp->tcp_rto)
4847 					 */
4848 					TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4849 				}
4850 				goto done;
4851 			}
4852 		}
4853 
4854 		tcph = tcp->tcp_tcph;
4855 
4856 		usable -= len;	/* Approximate - can be adjusted later */
4857 		if (usable > 0)
4858 			tcph->th_flags[0] = TH_ACK;
4859 		else
4860 			tcph->th_flags[0] = (TH_ACK | TH_PUSH);
4861 
4862 		U32_TO_ABE32(snxt, tcph->th_seq);
4863 
4864 		if (tcp->tcp_valid_bits) {
4865 			uchar_t		*prev_rptr = xmit_tail->b_rptr;
4866 			uint32_t	prev_snxt = tcp->tcp_snxt;
4867 
4868 			if (tail_unsent == 0) {
4869 				assert(xmit_tail->b_cont != NULL);
4870 				xmit_tail = xmit_tail->b_cont;
4871 				prev_rptr = xmit_tail->b_rptr;
4872 				tail_unsent = (int)(xmit_tail->b_wptr -
4873 				    xmit_tail->b_rptr);
4874 			} else {
4875 				xmit_tail->b_rptr = xmit_tail->b_wptr -
4876 				    tail_unsent;
4877 			}
4878 			mp = tcp_xmit_mp(tcp, xmit_tail, len, NULL, NULL,
4879 			    snxt, B_FALSE, (uint32_t *)&len, B_FALSE);
4880 			/* Restore tcp_snxt so we get amount sent right. */
4881 			tcp->tcp_snxt = prev_snxt;
4882 			if (prev_rptr == xmit_tail->b_rptr)
4883 				xmit_tail->b_prev = local_time;
4884 			else
4885 				xmit_tail->b_rptr = prev_rptr;
4886 
4887 			if (mp == NULL)
4888 				break;
4889 
4890 			mp1 = mp->b_cont;
4891 
4892 			snxt += len;
4893 			tcp->tcp_last_sent_len = (ushort_t)len;
4894 			while (mp1->b_cont) {
4895 				xmit_tail = xmit_tail->b_cont;
4896 				xmit_tail->b_prev = local_time;
4897 				mp1 = mp1->b_cont;
4898 			}
4899 			tail_unsent = xmit_tail->b_wptr - mp1->b_wptr;
4900 			BUMP_MIB(tcp_mib.tcpOutDataSegs);
4901 			UPDATE_MIB(tcp_mib.tcpOutDataBytes, len);
4902 			/* Dump the packet when debugging. */
4903 			TCP_DUMP_PACKET("tcp_wput_data (valid bits)", mp);
4904 			(void) ipv4_tcp_output(sock_id, mp);
4905 			freeb(mp);
4906 			continue;
4907 		}
4908 
4909 		snxt += len;	/* Adjust later if we don't send all of len */
4910 		BUMP_MIB(tcp_mib.tcpOutDataSegs);
4911 		UPDATE_MIB(tcp_mib.tcpOutDataBytes, len);
4912 
4913 		if (tail_unsent) {
4914 			/* Are the bytes above us in flight? */
4915 			rptr = xmit_tail->b_wptr - tail_unsent;
4916 			if (rptr != xmit_tail->b_rptr) {
4917 				tail_unsent -= len;
4918 				len += tcp_hdr_len;
4919 				tcp->tcp_ipha->ip_len = htons(len);
4920 				mp = dupb(xmit_tail);
4921 				if (!mp)
4922 					break;
4923 				mp->b_rptr = rptr;
4924 				goto must_alloc;
4925 			}
4926 		} else {
4927 			xmit_tail = xmit_tail->b_cont;
4928 			assert((uintptr_t)(xmit_tail->b_wptr -
4929 			    xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
4930 			tail_unsent = (int)(xmit_tail->b_wptr -
4931 			    xmit_tail->b_rptr);
4932 		}
4933 
4934 		tail_unsent -= len;
4935 		tcp->tcp_last_sent_len = (ushort_t)len;
4936 
4937 		len += tcp_hdr_len;
4938 		if (tcp->tcp_ipversion == IPV4_VERSION)
4939 			tcp->tcp_ipha->ip_len = htons(len);
4940 
4941 		xmit_tail->b_prev = local_time;
4942 
4943 		mp = dupb(xmit_tail);
4944 		if (mp == NULL)
4945 			goto out_of_mem;
4946 
4947 		len = tcp_hdr_len;
4948 		/*
4949 		 * There are four reasons to allocate a new hdr mblk:
4950 		 *  1) The bytes above us are in use by another packet
4951 		 *  2) We don't have good alignment
4952 		 *  3) The mblk is being shared
4953 		 *  4) We don't have enough room for a header
4954 		 */
4955 		rptr = mp->b_rptr - len;
4956 		if (!OK_32PTR(rptr) ||
4957 		    rptr < mp->b_datap) {
4958 			/* NOTE: we assume allocb returns an OK_32PTR */
4959 
4960 		must_alloc:;
4961 			mp1 = allocb(tcp->tcp_ip_hdr_len + TCP_MAX_HDR_LENGTH +
4962 			    tcp_wroff_xtra, 0);
4963 			if (mp1 == NULL) {
4964 				freemsg(mp);
4965 				goto out_of_mem;
4966 			}
4967 			mp1->b_cont = mp;
4968 			mp = mp1;
4969 			/* Leave room for Link Level header */
4970 			len = tcp_hdr_len;
4971 			rptr = &mp->b_rptr[tcp_wroff_xtra];
4972 			mp->b_wptr = &rptr[len];
4973 		}
4974 
4975 		if (tcp->tcp_snd_ts_ok) {
4976 			/* use uintptr_t to suppress the gcc warning */
4977 			U32_TO_BE32((uint32_t)(uintptr_t)local_time,
4978 				(char *)tcph+TCP_MIN_HEADER_LENGTH+4);
4979 			U32_TO_BE32(tcp->tcp_ts_recent,
4980 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
4981 		} else {
4982 			assert(tcp->tcp_tcp_hdr_len == TCP_MIN_HEADER_LENGTH);
4983 		}
4984 
4985 		mp->b_rptr = rptr;
4986 
4987 		/* Copy the template header. */
4988 		dst = (ipaddr_t *)rptr;
4989 		src = (ipaddr_t *)tcp->tcp_iphc;
4990 		dst[0] = src[0];
4991 		dst[1] = src[1];
4992 		dst[2] = src[2];
4993 		dst[3] = src[3];
4994 		dst[4] = src[4];
4995 		dst[5] = src[5];
4996 		dst[6] = src[6];
4997 		dst[7] = src[7];
4998 		dst[8] = src[8];
4999 		dst[9] = src[9];
5000 		len = tcp->tcp_hdr_len;
5001 		if (len -= 40) {
5002 			len >>= 2;
5003 			dst += 10;
5004 			src += 10;
5005 			do {
5006 				*dst++ = *src++;
5007 			} while (--len);
5008 		}
5009 
5010 		/*
5011 		 * Set tcph to point to the header of the outgoing packet,
5012 		 * not to the template header.
5013 		 */
5014 		tcph = (tcph_t *)(rptr + tcp->tcp_ip_hdr_len);
5015 
5016 		/*
5017 		 * Set the ECN info in the TCP header if it is not a zero
5018 		 * window probe.  Zero window probe is only sent in
5019 		 * tcp_wput_data() and tcp_timer().
5020 		 */
5021 		if (tcp->tcp_ecn_ok && !tcp->tcp_zero_win_probe) {
5022 			SET_ECT(tcp, rptr);
5023 
5024 			if (tcp->tcp_ecn_echo_on)
5025 				tcph->th_flags[0] |= TH_ECE;
5026 			if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
5027 				tcph->th_flags[0] |= TH_CWR;
5028 				tcp->tcp_ecn_cwr_sent = B_TRUE;
5029 			}
5030 		}
5031 
5032 		/* Fill in SACK options */
5033 		if (num_sack_blk > 0) {
5034 			uchar_t *wptr = rptr + tcp->tcp_hdr_len;
5035 			sack_blk_t *tmp;
5036 			int32_t	i;
5037 
5038 			wptr[0] = TCPOPT_NOP;
5039 			wptr[1] = TCPOPT_NOP;
5040 			wptr[2] = TCPOPT_SACK;
5041 			wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
5042 			    sizeof (sack_blk_t);
5043 			wptr += TCPOPT_REAL_SACK_LEN;
5044 
5045 			tmp = tcp->tcp_sack_list;
5046 			for (i = 0; i < num_sack_blk; i++) {
5047 				U32_TO_BE32(tmp[i].begin, wptr);
5048 				wptr += sizeof (tcp_seq);
5049 				U32_TO_BE32(tmp[i].end, wptr);
5050 				wptr += sizeof (tcp_seq);
5051 			}
5052 			tcph->th_offset_and_rsrvd[0] += ((num_sack_blk * 2 + 1)
5053 			    << 4);
5054 		}
5055 
5056 		if (tail_unsent) {
5057 			mp1 = mp->b_cont;
5058 			if (mp1 == NULL)
5059 				mp1 = mp;
5060 			/*
5061 			 * If we're a little short, tack on more mblks
5062 			 * as long as we don't need to split an mblk.
5063 			 */
5064 			while (tail_unsent < 0 &&
5065 			    tail_unsent + (int)(xmit_tail->b_cont->b_wptr -
5066 			    xmit_tail->b_cont->b_rptr) <= 0) {
5067 				xmit_tail = xmit_tail->b_cont;
5068 				/* Stash for rtt use later */
5069 				xmit_tail->b_prev = local_time;
5070 				mp1->b_cont = dupb(xmit_tail);
5071 				mp1 = mp1->b_cont;
5072 				assert((uintptr_t)(xmit_tail->b_wptr -
5073 				    xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
5074 				tail_unsent += (int)(xmit_tail->b_wptr -
5075 				    xmit_tail->b_rptr);
5076 				if (mp1 == NULL) {
5077 					freemsg(mp);
5078 					goto out_of_mem;
5079 				}
5080 			}
5081 			/* Trim back any surplus on the last mblk */
5082 			if (tail_unsent > 0)
5083 				mp1->b_wptr -= tail_unsent;
5084 			if (tail_unsent < 0) {
5085 				uint32_t ip_len;
5086 
5087 				/*
5088 				 * We did not send everything we could in
5089 				 * order to preserve mblk boundaries.
5090 				 */
5091 				usable -= tail_unsent;
5092 				snxt += tail_unsent;
5093 				tcp->tcp_last_sent_len += tail_unsent;
5094 				UPDATE_MIB(tcp_mib.tcpOutDataBytes,
5095 				    tail_unsent);
5096 				/* Adjust the IP length field. */
5097 				ip_len = ntohs(((struct ip *)rptr)->ip_len) +
5098 				    tail_unsent;
5099 				((struct ip *)rptr)->ip_len = htons(ip_len);
5100 				tail_unsent = 0;
5101 			}
5102 		}
5103 
5104 		if (mp == NULL)
5105 			goto out_of_mem;
5106 
5107 		/*
5108 		 * Performance hit!  We need to pullup the whole message
5109 		 * in order to do checksum and for the MAC output routine.
5110 		 */
5111 		if (mp->b_cont != NULL) {
5112 			int mp_size;
5113 #ifdef	DEBUG
5114 			printf("Multiple mblk %d\n", msgdsize(mp));
5115 #endif
5116 			new_mp = allocb(msgdsize(mp) + tcp_wroff_xtra, 0);
5117 			new_mp->b_rptr += tcp_wroff_xtra;
5118 			new_mp->b_wptr = new_mp->b_rptr;
5119 			while (mp != NULL) {
5120 				mp_size = mp->b_wptr - mp->b_rptr;
5121 				bcopy(mp->b_rptr, new_mp->b_wptr, mp_size);
5122 				new_mp->b_wptr += mp_size;
5123 				mp = mp->b_cont;
5124 			}
5125 			freemsg(mp);
5126 			mp = new_mp;
5127 		}
5128 		tcp_set_cksum(mp);
5129 		((struct ip *)mp->b_rptr)->ip_ttl = (uint8_t)tcp_ipv4_ttl;
5130 		TCP_DUMP_PACKET("tcp_wput_data", mp);
5131 		(void) ipv4_tcp_output(sock_id, mp);
5132 		freemsg(mp);
5133 	}
5134 out_of_mem:;
5135 	/* Pretend that all we were trying to send really got sent */
5136 	if (tail_unsent < 0) {
5137 		do {
5138 			xmit_tail = xmit_tail->b_cont;
5139 			xmit_tail->b_prev = local_time;
5140 			assert((uintptr_t)(xmit_tail->b_wptr -
5141 			    xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
5142 			tail_unsent += (int)(xmit_tail->b_wptr -
5143 			    xmit_tail->b_rptr);
5144 		} while (tail_unsent < 0);
5145 	}
5146 done:;
5147 	tcp->tcp_xmit_tail = xmit_tail;
5148 	tcp->tcp_xmit_tail_unsent = tail_unsent;
5149 	len = tcp->tcp_snxt - snxt;
5150 	if (len) {
5151 		/*
5152 		 * If new data was sent, need to update the notsack
5153 		 * list, which is, afterall, data blocks that have
5154 		 * not been sack'ed by the receiver.  New data is
5155 		 * not sack'ed.
5156 		 */
5157 		if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
5158 			/* len is a negative value. */
5159 			tcp->tcp_pipe -= len;
5160 			tcp_notsack_update(&(tcp->tcp_notsack_list),
5161 			    tcp->tcp_snxt, snxt,
5162 			    &(tcp->tcp_num_notsack_blk),
5163 			    &(tcp->tcp_cnt_notsack_list));
5164 		}
5165 		tcp->tcp_snxt = snxt + tcp->tcp_fin_sent;
5166 		tcp->tcp_rack = tcp->tcp_rnxt;
5167 		tcp->tcp_rack_cnt = 0;
5168 		if ((snxt + len) == tcp->tcp_suna) {
5169 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
5170 		}
5171 		/*
5172 		 * Note that len is the amount we just sent but with a negative
5173 		 * sign. We update tcp_unsent here since we may come back to
5174 		 * tcp_wput_data from tcp_state_wait.
5175 		 */
5176 		len += tcp->tcp_unsent;
5177 		tcp->tcp_unsent = len;
5178 
5179 		/*
5180 		 * Let's wait till all the segments have been acked, since we
5181 		 * don't have a timer.
5182 		 */
5183 		(void) tcp_state_wait(sock_id, tcp, TCPS_ALL_ACKED);
5184 		return;
5185 	} else if (snxt == tcp->tcp_suna && tcp->tcp_swnd == 0) {
5186 		/*
5187 		 * Didn't send anything. Make sure the timer is running
5188 		 * so that we will probe a zero window.
5189 		 */
5190 		TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
5191 	}
5192 
5193 	/* Note that len is the amount we just sent but with a negative sign */
5194 	len += tcp->tcp_unsent;
5195 	tcp->tcp_unsent = len;
5196 
5197 }
5198 
5199 static void
5200 tcp_time_wait_processing(tcp_t *tcp, mblk_t *mp,
5201     uint32_t seg_seq, uint32_t seg_ack, int seg_len, tcph_t *tcph,
5202     int sock_id)
5203 {
5204 	int32_t		bytes_acked;
5205 	int32_t		gap;
5206 	int32_t		rgap;
5207 	tcp_opt_t	tcpopt;
5208 	uint_t		flags;
5209 	uint32_t	new_swnd = 0;
5210 
5211 #ifdef DEBUG
5212 	printf("Time wait processing called ###############3\n");
5213 #endif
5214 
5215 	/* Just make sure we send the right sock_id to tcp_clean_death */
5216 	if ((sockets[sock_id].pcb == NULL) || (sockets[sock_id].pcb != tcp))
5217 		sock_id = -1;
5218 
5219 	flags = (unsigned int)tcph->th_flags[0] & 0xFF;
5220 	new_swnd = BE16_TO_U16(tcph->th_win) <<
5221 	    ((tcph->th_flags[0] & TH_SYN) ? 0 : tcp->tcp_snd_ws);
5222 	if (tcp->tcp_snd_ts_ok) {
5223 		if (!tcp_paws_check(tcp, tcph, &tcpopt)) {
5224 			freemsg(mp);
5225 			tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
5226 			    tcp->tcp_rnxt, TH_ACK, 0, -1);
5227 			return;
5228 		}
5229 	}
5230 	gap = seg_seq - tcp->tcp_rnxt;
5231 	rgap = tcp->tcp_rwnd - (gap + seg_len);
5232 	if (gap < 0) {
5233 		BUMP_MIB(tcp_mib.tcpInDataDupSegs);
5234 		UPDATE_MIB(tcp_mib.tcpInDataDupBytes,
5235 		    (seg_len > -gap ? -gap : seg_len));
5236 		seg_len += gap;
5237 		if (seg_len < 0 || (seg_len == 0 && !(flags & TH_FIN))) {
5238 			if (flags & TH_RST) {
5239 				freemsg(mp);
5240 				return;
5241 			}
5242 			if ((flags & TH_FIN) && seg_len == -1) {
5243 				/*
5244 				 * When TCP receives a duplicate FIN in
5245 				 * TIME_WAIT state, restart the 2 MSL timer.
5246 				 * See page 73 in RFC 793. Make sure this TCP
5247 				 * is already on the TIME_WAIT list. If not,
5248 				 * just restart the timer.
5249 				 */
5250 				tcp_time_wait_remove(tcp);
5251 				tcp_time_wait_append(tcp);
5252 				TCP_TIMER_RESTART(tcp, tcp_time_wait_interval);
5253 				tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
5254 				    tcp->tcp_rnxt, TH_ACK, 0, -1);
5255 				freemsg(mp);
5256 				return;
5257 			}
5258 			flags |=  TH_ACK_NEEDED;
5259 			seg_len = 0;
5260 			goto process_ack;
5261 		}
5262 
5263 		/* Fix seg_seq, and chew the gap off the front. */
5264 		seg_seq = tcp->tcp_rnxt;
5265 	}
5266 
5267 	if ((flags & TH_SYN) && gap > 0 && rgap < 0) {
5268 		/*
5269 		 * Make sure that when we accept the connection, pick
5270 		 * an ISS greater than (tcp_snxt + ISS_INCR/2) for the
5271 		 * old connection.
5272 		 *
5273 		 * The next ISS generated is equal to tcp_iss_incr_extra
5274 		 * + ISS_INCR/2 + other components depending on the
5275 		 * value of tcp_strong_iss.  We pre-calculate the new
5276 		 * ISS here and compare with tcp_snxt to determine if
5277 		 * we need to make adjustment to tcp_iss_incr_extra.
5278 		 *
5279 		 * Note that since we are now in the global queue
5280 		 * perimeter and need to do a lateral_put() to the
5281 		 * listener queue, there can be other connection requests/
5282 		 * attempts while the lateral_put() is going on.  That
5283 		 * means what we calculate here may not be correct.  This
5284 		 * is extremely difficult to solve unless TCP and IP
5285 		 * modules are merged and there is no perimeter, but just
5286 		 * locks.  The above calculation is ugly and is a
5287 		 * waste of CPU cycles...
5288 		 */
5289 		uint32_t new_iss = tcp_iss_incr_extra;
5290 		int32_t adj;
5291 
5292 		/* Add time component and min random (i.e. 1). */
5293 		new_iss += (prom_gettime() >> ISS_NSEC_SHT) + 1;
5294 		if ((adj = (int32_t)(tcp->tcp_snxt - new_iss)) > 0) {
5295 			/*
5296 			 * New ISS not guaranteed to be ISS_INCR/2
5297 			 * ahead of the current tcp_snxt, so add the
5298 			 * difference to tcp_iss_incr_extra.
5299 			 */
5300 			tcp_iss_incr_extra += adj;
5301 		}
5302 		tcp_clean_death(sock_id, tcp, 0);
5303 
5304 		/*
5305 		 * This is a passive open.  Right now we do not
5306 		 * do anything...
5307 		 */
5308 		freemsg(mp);
5309 		return;
5310 	}
5311 
5312 	/*
5313 	 * rgap is the amount of stuff received out of window.  A negative
5314 	 * value is the amount out of window.
5315 	 */
5316 	if (rgap < 0) {
5317 		BUMP_MIB(tcp_mib.tcpInDataPastWinSegs);
5318 		UPDATE_MIB(tcp_mib.tcpInDataPastWinBytes, -rgap);
5319 		/* Fix seg_len and make sure there is something left. */
5320 		seg_len += rgap;
5321 		if (seg_len <= 0) {
5322 			if (flags & TH_RST) {
5323 				freemsg(mp);
5324 				return;
5325 			}
5326 			flags |=  TH_ACK_NEEDED;
5327 			seg_len = 0;
5328 			goto process_ack;
5329 		}
5330 	}
5331 	/*
5332 	 * Check whether we can update tcp_ts_recent.  This test is
5333 	 * NOT the one in RFC 1323 3.4.  It is from Braden, 1993, "TCP
5334 	 * Extensions for High Performance: An Update", Internet Draft.
5335 	 */
5336 	if (tcp->tcp_snd_ts_ok &&
5337 	    TSTMP_GEQ(tcpopt.tcp_opt_ts_val, tcp->tcp_ts_recent) &&
5338 	    SEQ_LEQ(seg_seq, tcp->tcp_rack)) {
5339 		tcp->tcp_ts_recent = tcpopt.tcp_opt_ts_val;
5340 		tcp->tcp_last_rcv_lbolt = prom_gettime();
5341 	}
5342 
5343 	if (seg_seq != tcp->tcp_rnxt && seg_len > 0) {
5344 		/* Always ack out of order packets */
5345 		flags |= TH_ACK_NEEDED;
5346 		seg_len = 0;
5347 	} else if (seg_len > 0) {
5348 		BUMP_MIB(tcp_mib.tcpInDataInorderSegs);
5349 		UPDATE_MIB(tcp_mib.tcpInDataInorderBytes, seg_len);
5350 	}
5351 	if (flags & TH_RST) {
5352 		freemsg(mp);
5353 		(void) tcp_clean_death(sock_id, tcp, 0);
5354 		return;
5355 	}
5356 	if (flags & TH_SYN) {
5357 		freemsg(mp);
5358 		tcp_xmit_ctl("TH_SYN", tcp, NULL, seg_ack, seg_seq + 1,
5359 		    TH_RST|TH_ACK, 0, -1);
5360 		/*
5361 		 * Do not delete the TCP structure if it is in
5362 		 * TIME_WAIT state.  Refer to RFC 1122, 4.2.2.13.
5363 		 */
5364 		return;
5365 	}
5366 process_ack:
5367 	if (flags & TH_ACK) {
5368 		bytes_acked = (int)(seg_ack - tcp->tcp_suna);
5369 		if (bytes_acked <= 0) {
5370 			if (bytes_acked == 0 && seg_len == 0 &&
5371 			    new_swnd == tcp->tcp_swnd)
5372 				BUMP_MIB(tcp_mib.tcpInDupAck);
5373 		} else {
5374 			/* Acks something not sent */
5375 			flags |= TH_ACK_NEEDED;
5376 		}
5377 	}
5378 	freemsg(mp);
5379 	if (flags & TH_ACK_NEEDED) {
5380 		/*
5381 		 * Time to send an ack for some reason.
5382 		 */
5383 		tcp_xmit_ctl(NULL, tcp, NULL, tcp->tcp_snxt,
5384 		    tcp->tcp_rnxt, TH_ACK, 0, -1);
5385 	}
5386 }
5387 
5388 static int
5389 tcp_init_values(tcp_t *tcp, struct inetboot_socket *isp)
5390 {
5391 	int	err;
5392 
5393 	tcp->tcp_family = AF_INET;
5394 	tcp->tcp_ipversion = IPV4_VERSION;
5395 
5396 	/*
5397 	 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
5398 	 * will be close to tcp_rexmit_interval_initial.  By doing this, we
5399 	 * allow the algorithm to adjust slowly to large fluctuations of RTT
5400 	 * during first few transmissions of a connection as seen in slow
5401 	 * links.
5402 	 */
5403 	tcp->tcp_rtt_sa = tcp_rexmit_interval_initial << 2;
5404 	tcp->tcp_rtt_sd = tcp_rexmit_interval_initial >> 1;
5405 	tcp->tcp_rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
5406 	    tcp_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
5407 	    tcp_conn_grace_period;
5408 	if (tcp->tcp_rto < tcp_rexmit_interval_min)
5409 		tcp->tcp_rto = tcp_rexmit_interval_min;
5410 	tcp->tcp_timer_backoff = 0;
5411 	tcp->tcp_ms_we_have_waited = 0;
5412 	tcp->tcp_last_recv_time = prom_gettime();
5413 	tcp->tcp_cwnd_max = tcp_cwnd_max_;
5414 	tcp->tcp_snd_burst = TCP_CWND_INFINITE;
5415 	tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
5416 	/* For Ethernet, the mtu returned is actually 1550... */
5417 	if (mac_get_type() == IFT_ETHER) {
5418 		tcp->tcp_if_mtu = mac_get_mtu() - 50;
5419 	} else {
5420 		tcp->tcp_if_mtu = mac_get_mtu();
5421 	}
5422 	tcp->tcp_mss = tcp->tcp_if_mtu;
5423 
5424 	tcp->tcp_first_timer_threshold = tcp_ip_notify_interval;
5425 	tcp->tcp_first_ctimer_threshold = tcp_ip_notify_cinterval;
5426 	tcp->tcp_second_timer_threshold = tcp_ip_abort_interval;
5427 	/*
5428 	 * Fix it to tcp_ip_abort_linterval later if it turns out to be a
5429 	 * passive open.
5430 	 */
5431 	tcp->tcp_second_ctimer_threshold = tcp_ip_abort_cinterval;
5432 
5433 	tcp->tcp_naglim = tcp_naglim_def;
5434 
5435 	/* NOTE:  ISS is now set in tcp_adapt_ire(). */
5436 
5437 	/* Initialize the header template */
5438 	if (tcp->tcp_ipversion == IPV4_VERSION) {
5439 		err = tcp_header_init_ipv4(tcp);
5440 	}
5441 	if (err)
5442 		return (err);
5443 
5444 	/*
5445 	 * Init the window scale to the max so tcp_rwnd_set() won't pare
5446 	 * down tcp_rwnd. tcp_adapt_ire() will set the right value later.
5447 	 */
5448 	tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
5449 	tcp->tcp_xmit_lowater = tcp_xmit_lowat;
5450 	if (isp != NULL) {
5451 		tcp->tcp_xmit_hiwater = isp->so_sndbuf;
5452 		tcp->tcp_rwnd = isp->so_rcvbuf;
5453 		tcp->tcp_rwnd_max = isp->so_rcvbuf;
5454 	}
5455 	tcp->tcp_state = TCPS_IDLE;
5456 	return (0);
5457 }
5458 
5459 /*
5460  * Initialize the IPv4 header. Loses any record of any IP options.
5461  */
5462 static int
5463 tcp_header_init_ipv4(tcp_t *tcp)
5464 {
5465 	tcph_t		*tcph;
5466 
5467 	/*
5468 	 * This is a simple initialization. If there's
5469 	 * already a template, it should never be too small,
5470 	 * so reuse it.  Otherwise, allocate space for the new one.
5471 	 */
5472 	if (tcp->tcp_iphc != NULL) {
5473 		assert(tcp->tcp_iphc_len >= TCP_MAX_COMBINED_HEADER_LENGTH);
5474 		bzero(tcp->tcp_iphc, tcp->tcp_iphc_len);
5475 	} else {
5476 		tcp->tcp_iphc_len = TCP_MAX_COMBINED_HEADER_LENGTH;
5477 		tcp->tcp_iphc = bkmem_zalloc(tcp->tcp_iphc_len);
5478 		if (tcp->tcp_iphc == NULL) {
5479 			tcp->tcp_iphc_len = 0;
5480 			return (ENOMEM);
5481 		}
5482 	}
5483 	tcp->tcp_ipha = (struct ip *)tcp->tcp_iphc;
5484 	tcp->tcp_ipversion = IPV4_VERSION;
5485 
5486 	/*
5487 	 * Note that it does not include TCP options yet.  It will
5488 	 * after the connection is established.
5489 	 */
5490 	tcp->tcp_hdr_len = sizeof (struct ip) + sizeof (tcph_t);
5491 	tcp->tcp_tcp_hdr_len = sizeof (tcph_t);
5492 	tcp->tcp_ip_hdr_len = sizeof (struct ip);
5493 	tcp->tcp_ipha->ip_v = IP_VERSION;
5494 	/* We don't support IP options... */
5495 	tcp->tcp_ipha->ip_hl = IP_SIMPLE_HDR_LENGTH_IN_WORDS;
5496 	tcp->tcp_ipha->ip_p = IPPROTO_TCP;
5497 	/* We are not supposed to do PMTU discovery... */
5498 	tcp->tcp_ipha->ip_sum = 0;
5499 
5500 	tcph = (tcph_t *)(tcp->tcp_iphc + sizeof (struct ip));
5501 	tcp->tcp_tcph = tcph;
5502 	tcph->th_offset_and_rsrvd[0] = (5 << 4);
5503 	return (0);
5504 }
5505 
5506 /*
5507  * Send out a control packet on the tcp connection specified.  This routine
5508  * is typically called where we need a simple ACK or RST generated.
5509  *
5510  * This function is called with or without a mp.
5511  */
5512 static void
5513 tcp_xmit_ctl(char *str, tcp_t *tcp, mblk_t *mp, uint32_t seq,
5514     uint32_t ack, int ctl, uint_t ip_hdr_len, int sock_id)
5515 {
5516 	uchar_t		*rptr;
5517 	tcph_t		*tcph;
5518 	struct ip	*iph = NULL;
5519 	int		tcp_hdr_len;
5520 	int		tcp_ip_hdr_len;
5521 
5522 	tcp_hdr_len = tcp->tcp_hdr_len;
5523 	tcp_ip_hdr_len = tcp->tcp_ip_hdr_len;
5524 
5525 	if (mp) {
5526 		assert(ip_hdr_len != 0);
5527 		rptr = mp->b_rptr;
5528 		tcph = (tcph_t *)(rptr + ip_hdr_len);
5529 		/* Don't reply to a RST segment. */
5530 		if (tcph->th_flags[0] & TH_RST) {
5531 			freeb(mp);
5532 			return;
5533 		}
5534 		freemsg(mp);
5535 		rptr = NULL;
5536 	} else {
5537 		assert(ip_hdr_len == 0);
5538 	}
5539 	/* If a text string is passed in with the request, print it out. */
5540 	if (str != NULL) {
5541 		dprintf("tcp_xmit_ctl(%d): '%s', seq 0x%x, ack 0x%x, "
5542 		    "ctl 0x%x\n", sock_id, str, seq, ack, ctl);
5543 	}
5544 	mp = allocb(tcp_ip_hdr_len + TCP_MAX_HDR_LENGTH + tcp_wroff_xtra, 0);
5545 	if (mp == NULL) {
5546 		dprintf("tcp_xmit_ctl(%d): Cannot allocate memory\n", sock_id);
5547 		return;
5548 	}
5549 	rptr = &mp->b_rptr[tcp_wroff_xtra];
5550 	mp->b_rptr = rptr;
5551 	mp->b_wptr = &rptr[tcp_hdr_len];
5552 	bcopy(tcp->tcp_iphc, rptr, tcp_hdr_len);
5553 
5554 	iph = (struct ip *)rptr;
5555 	iph->ip_len = htons(tcp_hdr_len);
5556 
5557 	tcph = (tcph_t *)&rptr[tcp_ip_hdr_len];
5558 	tcph->th_flags[0] = (uint8_t)ctl;
5559 	if (ctl & TH_RST) {
5560 		BUMP_MIB(tcp_mib.tcpOutRsts);
5561 		BUMP_MIB(tcp_mib.tcpOutControl);
5562 		/*
5563 		 * Don't send TSopt w/ TH_RST packets per RFC 1323.
5564 		 */
5565 		if (tcp->tcp_snd_ts_ok && tcp->tcp_state > TCPS_SYN_SENT) {
5566 			mp->b_wptr = &rptr[tcp_hdr_len - TCPOPT_REAL_TS_LEN];
5567 			*(mp->b_wptr) = TCPOPT_EOL;
5568 			iph->ip_len = htons(tcp_hdr_len -
5569 			    TCPOPT_REAL_TS_LEN);
5570 			tcph->th_offset_and_rsrvd[0] -= (3 << 4);
5571 		}
5572 	}
5573 	if (ctl & TH_ACK) {
5574 		uint32_t now = prom_gettime();
5575 
5576 		if (tcp->tcp_snd_ts_ok) {
5577 			U32_TO_BE32(now,
5578 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+4);
5579 			U32_TO_BE32(tcp->tcp_ts_recent,
5580 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
5581 		}
5582 		tcp->tcp_rack = ack;
5583 		tcp->tcp_rack_cnt = 0;
5584 		BUMP_MIB(tcp_mib.tcpOutAck);
5585 	}
5586 	BUMP_MIB(tcp_mib.tcpOutSegs);
5587 	U32_TO_BE32(seq, tcph->th_seq);
5588 	U32_TO_BE32(ack, tcph->th_ack);
5589 
5590 	tcp_set_cksum(mp);
5591 	iph->ip_ttl = (uint8_t)tcp_ipv4_ttl;
5592 	TCP_DUMP_PACKET("tcp_xmit_ctl", mp);
5593 	(void) ipv4_tcp_output(sock_id, mp);
5594 	freeb(mp);
5595 }
5596 
5597 /* Generate an ACK-only (no data) segment for a TCP endpoint */
5598 static mblk_t *
5599 tcp_ack_mp(tcp_t *tcp)
5600 {
5601 	if (tcp->tcp_valid_bits) {
5602 		/*
5603 		 * For the complex case where we have to send some
5604 		 * controls (FIN or SYN), let tcp_xmit_mp do it.
5605 		 * When sending an ACK-only segment (no data)
5606 		 * into a zero window, always set the seq number to
5607 		 * suna, since snxt will be extended past the window.
5608 		 * If we used snxt, the receiver might consider the ACK
5609 		 * unacceptable.
5610 		 */
5611 		return (tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
5612 		    (tcp->tcp_zero_win_probe) ?
5613 		    tcp->tcp_suna :
5614 		    tcp->tcp_snxt, B_FALSE, NULL, B_FALSE));
5615 	} else {
5616 		/* Generate a simple ACK */
5617 		uchar_t	*rptr;
5618 		tcph_t	*tcph;
5619 		mblk_t	*mp1;
5620 		int32_t	tcp_hdr_len;
5621 		int32_t	num_sack_blk = 0;
5622 		int32_t sack_opt_len;
5623 
5624 		/*
5625 		 * Allocate space for TCP + IP headers
5626 		 * and link-level header
5627 		 */
5628 		if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
5629 			num_sack_blk = MIN(tcp->tcp_max_sack_blk,
5630 			    tcp->tcp_num_sack_blk);
5631 			sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
5632 			    TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
5633 			tcp_hdr_len = tcp->tcp_hdr_len + sack_opt_len;
5634 		} else {
5635 			tcp_hdr_len = tcp->tcp_hdr_len;
5636 		}
5637 		mp1 = allocb(tcp_hdr_len + tcp_wroff_xtra, 0);
5638 		if (mp1 == NULL)
5639 			return (NULL);
5640 
5641 		/* copy in prototype TCP + IP header */
5642 		rptr = mp1->b_rptr + tcp_wroff_xtra;
5643 		mp1->b_rptr = rptr;
5644 		mp1->b_wptr = rptr + tcp_hdr_len;
5645 		bcopy(tcp->tcp_iphc, rptr, tcp->tcp_hdr_len);
5646 
5647 		tcph = (tcph_t *)&rptr[tcp->tcp_ip_hdr_len];
5648 
5649 		/*
5650 		 * Set the TCP sequence number.
5651 		 * When sending an ACK-only segment (no data)
5652 		 * into a zero window, always set the seq number to
5653 		 * suna, since snxt will be extended past the window.
5654 		 * If we used snxt, the receiver might consider the ACK
5655 		 * unacceptable.
5656 		 */
5657 		U32_TO_ABE32((tcp->tcp_zero_win_probe) ?
5658 		    tcp->tcp_suna : tcp->tcp_snxt, tcph->th_seq);
5659 
5660 		/* Set up the TCP flag field. */
5661 		tcph->th_flags[0] = (uchar_t)TH_ACK;
5662 		if (tcp->tcp_ecn_echo_on)
5663 			tcph->th_flags[0] |= TH_ECE;
5664 
5665 		tcp->tcp_rack = tcp->tcp_rnxt;
5666 		tcp->tcp_rack_cnt = 0;
5667 
5668 		/* fill in timestamp option if in use */
5669 		if (tcp->tcp_snd_ts_ok) {
5670 			uint32_t llbolt = (uint32_t)prom_gettime();
5671 
5672 			U32_TO_BE32(llbolt,
5673 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+4);
5674 			U32_TO_BE32(tcp->tcp_ts_recent,
5675 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
5676 		}
5677 
5678 		/* Fill in SACK options */
5679 		if (num_sack_blk > 0) {
5680 			uchar_t *wptr = (uchar_t *)tcph + tcp->tcp_tcp_hdr_len;
5681 			sack_blk_t *tmp;
5682 			int32_t	i;
5683 
5684 			wptr[0] = TCPOPT_NOP;
5685 			wptr[1] = TCPOPT_NOP;
5686 			wptr[2] = TCPOPT_SACK;
5687 			wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
5688 			    sizeof (sack_blk_t);
5689 			wptr += TCPOPT_REAL_SACK_LEN;
5690 
5691 			tmp = tcp->tcp_sack_list;
5692 			for (i = 0; i < num_sack_blk; i++) {
5693 				U32_TO_BE32(tmp[i].begin, wptr);
5694 				wptr += sizeof (tcp_seq);
5695 				U32_TO_BE32(tmp[i].end, wptr);
5696 				wptr += sizeof (tcp_seq);
5697 			}
5698 			tcph->th_offset_and_rsrvd[0] += ((num_sack_blk * 2 + 1)
5699 			    << 4);
5700 		}
5701 
5702 		((struct ip *)rptr)->ip_len = htons(tcp_hdr_len);
5703 		tcp_set_cksum(mp1);
5704 		((struct ip *)rptr)->ip_ttl = (uint8_t)tcp_ipv4_ttl;
5705 		return (mp1);
5706 	}
5707 }
5708 
5709 /*
5710  * tcp_xmit_mp is called to return a pointer to an mblk chain complete with
5711  * ip and tcp header ready to pass down to IP.  If the mp passed in is
5712  * non-NULL, then up to max_to_send bytes of data will be dup'ed off that
5713  * mblk. (If sendall is not set the dup'ing will stop at an mblk boundary
5714  * otherwise it will dup partial mblks.)
5715  * Otherwise, an appropriate ACK packet will be generated.  This
5716  * routine is not usually called to send new data for the first time.  It
5717  * is mostly called out of the timer for retransmits, and to generate ACKs.
5718  *
5719  * If offset is not NULL, the returned mblk chain's first mblk's b_rptr will
5720  * be adjusted by *offset.  And after dupb(), the offset and the ending mblk
5721  * of the original mblk chain will be returned in *offset and *end_mp.
5722  */
5723 static mblk_t *
5724 tcp_xmit_mp(tcp_t *tcp, mblk_t *mp, int32_t max_to_send, int32_t *offset,
5725     mblk_t **end_mp, uint32_t seq, boolean_t sendall, uint32_t *seg_len,
5726     boolean_t rexmit)
5727 {
5728 	int	data_length;
5729 	int32_t	off = 0;
5730 	uint_t	flags;
5731 	mblk_t	*mp1;
5732 	mblk_t	*mp2;
5733 	mblk_t	*new_mp;
5734 	uchar_t	*rptr;
5735 	tcph_t	*tcph;
5736 	int32_t	num_sack_blk = 0;
5737 	int32_t	sack_opt_len = 0;
5738 
5739 	/* Allocate for our maximum TCP header + link-level */
5740 	mp1 = allocb(tcp->tcp_ip_hdr_len + TCP_MAX_HDR_LENGTH +
5741 	    tcp_wroff_xtra, 0);
5742 	if (mp1 == NULL)
5743 		return (NULL);
5744 	data_length = 0;
5745 
5746 	/*
5747 	 * Note that tcp_mss has been adjusted to take into account the
5748 	 * timestamp option if applicable.  Because SACK options do not
5749 	 * appear in every TCP segments and they are of variable lengths,
5750 	 * they cannot be included in tcp_mss.  Thus we need to calculate
5751 	 * the actual segment length when we need to send a segment which
5752 	 * includes SACK options.
5753 	 */
5754 	if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
5755 		num_sack_blk = MIN(tcp->tcp_max_sack_blk,
5756 		    tcp->tcp_num_sack_blk);
5757 		sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
5758 		    TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
5759 		if (max_to_send + sack_opt_len > tcp->tcp_mss)
5760 			max_to_send -= sack_opt_len;
5761 	}
5762 
5763 	if (offset != NULL) {
5764 		off = *offset;
5765 		/* We use offset as an indicator that end_mp is not NULL. */
5766 		*end_mp = NULL;
5767 	}
5768 	for (mp2 = mp1; mp && data_length != max_to_send; mp = mp->b_cont) {
5769 		/* This could be faster with cooperation from downstream */
5770 		if (mp2 != mp1 && !sendall &&
5771 		    data_length + (int)(mp->b_wptr - mp->b_rptr) >
5772 		    max_to_send)
5773 			/*
5774 			 * Don't send the next mblk since the whole mblk
5775 			 * does not fit.
5776 			 */
5777 			break;
5778 		mp2->b_cont = dupb(mp);
5779 		mp2 = mp2->b_cont;
5780 		if (mp2 == NULL) {
5781 			freemsg(mp1);
5782 			return (NULL);
5783 		}
5784 		mp2->b_rptr += off;
5785 		assert((uintptr_t)(mp2->b_wptr - mp2->b_rptr) <=
5786 		    (uintptr_t)INT_MAX);
5787 
5788 		data_length += (int)(mp2->b_wptr - mp2->b_rptr);
5789 		if (data_length > max_to_send) {
5790 			mp2->b_wptr -= data_length - max_to_send;
5791 			data_length = max_to_send;
5792 			off = mp2->b_wptr - mp->b_rptr;
5793 			break;
5794 		} else {
5795 			off = 0;
5796 		}
5797 	}
5798 	if (offset != NULL) {
5799 		*offset = off;
5800 		*end_mp = mp;
5801 	}
5802 	if (seg_len != NULL) {
5803 		*seg_len = data_length;
5804 	}
5805 
5806 	rptr = mp1->b_rptr + tcp_wroff_xtra;
5807 	mp1->b_rptr = rptr;
5808 	mp1->b_wptr = rptr + tcp->tcp_hdr_len + sack_opt_len;
5809 	bcopy(tcp->tcp_iphc, rptr, tcp->tcp_hdr_len);
5810 	tcph = (tcph_t *)&rptr[tcp->tcp_ip_hdr_len];
5811 	U32_TO_ABE32(seq, tcph->th_seq);
5812 
5813 	/*
5814 	 * Use tcp_unsent to determine if the PUSH bit should be used assumes
5815 	 * that this function was called from tcp_wput_data. Thus, when called
5816 	 * to retransmit data the setting of the PUSH bit may appear some
5817 	 * what random in that it might get set when it should not. This
5818 	 * should not pose any performance issues.
5819 	 */
5820 	if (data_length != 0 && (tcp->tcp_unsent == 0 ||
5821 	    tcp->tcp_unsent == data_length)) {
5822 		flags = TH_ACK | TH_PUSH;
5823 	} else {
5824 		flags = TH_ACK;
5825 	}
5826 
5827 	if (tcp->tcp_ecn_ok) {
5828 		if (tcp->tcp_ecn_echo_on)
5829 			flags |= TH_ECE;
5830 
5831 		/*
5832 		 * Only set ECT bit and ECN_CWR if a segment contains new data.
5833 		 * There is no TCP flow control for non-data segments, and
5834 		 * only data segment is transmitted reliably.
5835 		 */
5836 		if (data_length > 0 && !rexmit) {
5837 			SET_ECT(tcp, rptr);
5838 			if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
5839 				flags |= TH_CWR;
5840 				tcp->tcp_ecn_cwr_sent = B_TRUE;
5841 			}
5842 		}
5843 	}
5844 
5845 	if (tcp->tcp_valid_bits) {
5846 		uint32_t u1;
5847 
5848 		if ((tcp->tcp_valid_bits & TCP_ISS_VALID) &&
5849 		    seq == tcp->tcp_iss) {
5850 			uchar_t	*wptr;
5851 
5852 			/*
5853 			 * Tack on the MSS option.  It is always needed
5854 			 * for both active and passive open.
5855 			 */
5856 			wptr = mp1->b_wptr;
5857 			wptr[0] = TCPOPT_MAXSEG;
5858 			wptr[1] = TCPOPT_MAXSEG_LEN;
5859 			wptr += 2;
5860 			/*
5861 			 * MSS option value should be interface MTU - MIN
5862 			 * TCP/IP header.
5863 			 */
5864 			u1 = tcp->tcp_if_mtu - IP_SIMPLE_HDR_LENGTH -
5865 			    TCP_MIN_HEADER_LENGTH;
5866 			U16_TO_BE16(u1, wptr);
5867 			mp1->b_wptr = wptr + 2;
5868 			/* Update the offset to cover the additional word */
5869 			tcph->th_offset_and_rsrvd[0] += (1 << 4);
5870 
5871 			/*
5872 			 * Note that the following way of filling in
5873 			 * TCP options are not optimal.  Some NOPs can
5874 			 * be saved.  But there is no need at this time
5875 			 * to optimize it.  When it is needed, we will
5876 			 * do it.
5877 			 */
5878 			switch (tcp->tcp_state) {
5879 			case TCPS_SYN_SENT:
5880 				flags = TH_SYN;
5881 
5882 				if (tcp->tcp_snd_ws_ok) {
5883 					wptr = mp1->b_wptr;
5884 					wptr[0] =  TCPOPT_NOP;
5885 					wptr[1] =  TCPOPT_WSCALE;
5886 					wptr[2] =  TCPOPT_WS_LEN;
5887 					wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
5888 					mp1->b_wptr += TCPOPT_REAL_WS_LEN;
5889 					tcph->th_offset_and_rsrvd[0] +=
5890 					    (1 << 4);
5891 				}
5892 
5893 				if (tcp->tcp_snd_ts_ok) {
5894 					uint32_t llbolt;
5895 
5896 					llbolt = prom_gettime();
5897 					wptr = mp1->b_wptr;
5898 					wptr[0] = TCPOPT_NOP;
5899 					wptr[1] = TCPOPT_NOP;
5900 					wptr[2] = TCPOPT_TSTAMP;
5901 					wptr[3] = TCPOPT_TSTAMP_LEN;
5902 					wptr += 4;
5903 					U32_TO_BE32(llbolt, wptr);
5904 					wptr += 4;
5905 					assert(tcp->tcp_ts_recent == 0);
5906 					U32_TO_BE32(0L, wptr);
5907 					mp1->b_wptr += TCPOPT_REAL_TS_LEN;
5908 					tcph->th_offset_and_rsrvd[0] +=
5909 					    (3 << 4);
5910 				}
5911 
5912 				if (tcp->tcp_snd_sack_ok) {
5913 					wptr = mp1->b_wptr;
5914 					wptr[0] = TCPOPT_NOP;
5915 					wptr[1] = TCPOPT_NOP;
5916 					wptr[2] = TCPOPT_SACK_PERMITTED;
5917 					wptr[3] = TCPOPT_SACK_OK_LEN;
5918 					mp1->b_wptr += TCPOPT_REAL_SACK_OK_LEN;
5919 					tcph->th_offset_and_rsrvd[0] +=
5920 					    (1 << 4);
5921 				}
5922 
5923 				/*
5924 				 * Set up all the bits to tell other side
5925 				 * we are ECN capable.
5926 				 */
5927 				if (tcp->tcp_ecn_ok) {
5928 					flags |= (TH_ECE | TH_CWR);
5929 				}
5930 				break;
5931 			case TCPS_SYN_RCVD:
5932 				flags |= TH_SYN;
5933 
5934 				if (tcp->tcp_snd_ws_ok) {
5935 				    wptr = mp1->b_wptr;
5936 				    wptr[0] =  TCPOPT_NOP;
5937 				    wptr[1] =  TCPOPT_WSCALE;
5938 				    wptr[2] =  TCPOPT_WS_LEN;
5939 				    wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
5940 				    mp1->b_wptr += TCPOPT_REAL_WS_LEN;
5941 				    tcph->th_offset_and_rsrvd[0] += (1 << 4);
5942 				}
5943 
5944 				if (tcp->tcp_snd_sack_ok) {
5945 					wptr = mp1->b_wptr;
5946 					wptr[0] = TCPOPT_NOP;
5947 					wptr[1] = TCPOPT_NOP;
5948 					wptr[2] = TCPOPT_SACK_PERMITTED;
5949 					wptr[3] = TCPOPT_SACK_OK_LEN;
5950 					mp1->b_wptr += TCPOPT_REAL_SACK_OK_LEN;
5951 					tcph->th_offset_and_rsrvd[0] +=
5952 					    (1 << 4);
5953 				}
5954 
5955 				/*
5956 				 * If the other side is ECN capable, reply
5957 				 * that we are also ECN capable.
5958 				 */
5959 				if (tcp->tcp_ecn_ok) {
5960 					flags |= TH_ECE;
5961 				}
5962 				break;
5963 			default:
5964 				break;
5965 			}
5966 			/* allocb() of adequate mblk assures space */
5967 			assert((uintptr_t)(mp1->b_wptr -
5968 			    mp1->b_rptr) <= (uintptr_t)INT_MAX);
5969 			if (flags & TH_SYN)
5970 				BUMP_MIB(tcp_mib.tcpOutControl);
5971 		}
5972 		if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
5973 		    (seq + data_length) == tcp->tcp_fss) {
5974 			if (!tcp->tcp_fin_acked) {
5975 				flags |= TH_FIN;
5976 				BUMP_MIB(tcp_mib.tcpOutControl);
5977 			}
5978 			if (!tcp->tcp_fin_sent) {
5979 				tcp->tcp_fin_sent = B_TRUE;
5980 				switch (tcp->tcp_state) {
5981 				case TCPS_SYN_RCVD:
5982 				case TCPS_ESTABLISHED:
5983 					tcp->tcp_state = TCPS_FIN_WAIT_1;
5984 					break;
5985 				case TCPS_CLOSE_WAIT:
5986 					tcp->tcp_state = TCPS_LAST_ACK;
5987 					break;
5988 				}
5989 				if (tcp->tcp_suna == tcp->tcp_snxt) {
5990 					TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
5991 				}
5992 				tcp->tcp_snxt = tcp->tcp_fss + 1;
5993 			}
5994 		}
5995 	}
5996 	tcph->th_flags[0] = (uchar_t)flags;
5997 	tcp->tcp_rack = tcp->tcp_rnxt;
5998 	tcp->tcp_rack_cnt = 0;
5999 
6000 	if (tcp->tcp_snd_ts_ok) {
6001 		if (tcp->tcp_state != TCPS_SYN_SENT) {
6002 			uint32_t llbolt = prom_gettime();
6003 
6004 			U32_TO_BE32(llbolt,
6005 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+4);
6006 			U32_TO_BE32(tcp->tcp_ts_recent,
6007 			    (char *)tcph+TCP_MIN_HEADER_LENGTH+8);
6008 		}
6009 	}
6010 
6011 	if (num_sack_blk > 0) {
6012 		uchar_t *wptr = (uchar_t *)tcph + tcp->tcp_tcp_hdr_len;
6013 		sack_blk_t *tmp;
6014 		int32_t	i;
6015 
6016 		wptr[0] = TCPOPT_NOP;
6017 		wptr[1] = TCPOPT_NOP;
6018 		wptr[2] = TCPOPT_SACK;
6019 		wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
6020 		    sizeof (sack_blk_t);
6021 		wptr += TCPOPT_REAL_SACK_LEN;
6022 
6023 		tmp = tcp->tcp_sack_list;
6024 		for (i = 0; i < num_sack_blk; i++) {
6025 			U32_TO_BE32(tmp[i].begin, wptr);
6026 			wptr += sizeof (tcp_seq);
6027 			U32_TO_BE32(tmp[i].end, wptr);
6028 			wptr += sizeof (tcp_seq);
6029 		}
6030 		tcph->th_offset_and_rsrvd[0] += ((num_sack_blk * 2 + 1) << 4);
6031 	}
6032 	assert((uintptr_t)(mp1->b_wptr - rptr) <= (uintptr_t)INT_MAX);
6033 	data_length += (int)(mp1->b_wptr - rptr);
6034 	if (tcp->tcp_ipversion == IPV4_VERSION)
6035 		((struct ip *)rptr)->ip_len = htons(data_length);
6036 
6037 	/*
6038 	 * Performance hit!  We need to pullup the whole message
6039 	 * in order to do checksum and for the MAC output routine.
6040 	 */
6041 	if (mp1->b_cont != NULL) {
6042 		int mp_size;
6043 #ifdef DEBUG
6044 		printf("Multiple mblk %d\n", msgdsize(mp1));
6045 #endif
6046 		mp2 = mp1;
6047 		new_mp = allocb(msgdsize(mp1) + tcp_wroff_xtra, 0);
6048 		new_mp->b_rptr += tcp_wroff_xtra;
6049 		new_mp->b_wptr = new_mp->b_rptr;
6050 		while (mp1 != NULL) {
6051 			mp_size = mp1->b_wptr - mp1->b_rptr;
6052 			bcopy(mp1->b_rptr, new_mp->b_wptr, mp_size);
6053 			new_mp->b_wptr += mp_size;
6054 			mp1 = mp1->b_cont;
6055 		}
6056 		freemsg(mp2);
6057 		mp1 = new_mp;
6058 	}
6059 	tcp_set_cksum(mp1);
6060 	/* Fill in the TTL field as it is 0 in the header template. */
6061 	((struct ip *)mp1->b_rptr)->ip_ttl = (uint8_t)tcp_ipv4_ttl;
6062 
6063 	return (mp1);
6064 }
6065 
6066 /*
6067  * Generate a "no listener here" reset in response to the
6068  * connection request contained within 'mp'
6069  */
6070 static void
6071 tcp_xmit_listeners_reset(int sock_id, mblk_t *mp, uint_t ip_hdr_len)
6072 {
6073 	uchar_t		*rptr;
6074 	uint32_t	seg_len;
6075 	tcph_t		*tcph;
6076 	uint32_t	seg_seq;
6077 	uint32_t	seg_ack;
6078 	uint_t		flags;
6079 
6080 	rptr = mp->b_rptr;
6081 
6082 	tcph = (tcph_t *)&rptr[ip_hdr_len];
6083 	seg_seq = BE32_TO_U32(tcph->th_seq);
6084 	seg_ack = BE32_TO_U32(tcph->th_ack);
6085 	flags = tcph->th_flags[0];
6086 
6087 	seg_len = msgdsize(mp) - (TCP_HDR_LENGTH(tcph) + ip_hdr_len);
6088 	if (flags & TH_RST) {
6089 		freeb(mp);
6090 	} else if (flags & TH_ACK) {
6091 		tcp_xmit_early_reset("no tcp, reset",
6092 		    sock_id, mp, seg_ack, 0, TH_RST, ip_hdr_len);
6093 	} else {
6094 		if (flags & TH_SYN)
6095 			seg_len++;
6096 		tcp_xmit_early_reset("no tcp, reset/ack", sock_id,
6097 		    mp, 0, seg_seq + seg_len,
6098 		    TH_RST | TH_ACK, ip_hdr_len);
6099 	}
6100 }
6101 
6102 /* Non overlapping byte exchanger */
6103 static void
6104 tcp_xchg(uchar_t *a, uchar_t *b, int len)
6105 {
6106 	uchar_t	uch;
6107 
6108 	while (len-- > 0) {
6109 		uch = a[len];
6110 		a[len] = b[len];
6111 		b[len] = uch;
6112 	}
6113 }
6114 
6115 /*
6116  * Generate a reset based on an inbound packet for which there is no active
6117  * tcp state that we can find.
6118  */
6119 static void
6120 tcp_xmit_early_reset(char *str, int sock_id, mblk_t *mp, uint32_t seq,
6121     uint32_t ack, int ctl, uint_t ip_hdr_len)
6122 {
6123 	struct ip	*iph = NULL;
6124 	ushort_t	len;
6125 	tcph_t		*tcph;
6126 	int		i;
6127 	ipaddr_t	addr;
6128 	mblk_t		*new_mp;
6129 
6130 	if (str != NULL) {
6131 		dprintf("tcp_xmit_early_reset: '%s', seq 0x%x, ack 0x%x, "
6132 		    "flags 0x%x\n", str, seq, ack, ctl);
6133 	}
6134 
6135 	/*
6136 	 * We skip reversing source route here.
6137 	 * (for now we replace all IP options with EOL)
6138 	 */
6139 	iph = (struct ip *)mp->b_rptr;
6140 	for (i = IP_SIMPLE_HDR_LENGTH; i < (int)ip_hdr_len; i++)
6141 		mp->b_rptr[i] = IPOPT_EOL;
6142 	/*
6143 	 * Make sure that src address is not a limited broadcast
6144 	 * address. Not all broadcast address checking for the
6145 	 * src address is possible, since we don't know the
6146 	 * netmask of the src addr.
6147 	 * No check for destination address is done, since
6148 	 * IP will not pass up a packet with a broadcast dest address
6149 	 * to TCP.
6150 	 */
6151 	if (iph->ip_src.s_addr == INADDR_ANY ||
6152 	    iph->ip_src.s_addr == INADDR_BROADCAST) {
6153 		freemsg(mp);
6154 		return;
6155 	}
6156 
6157 	tcph = (tcph_t *)&mp->b_rptr[ip_hdr_len];
6158 	if (tcph->th_flags[0] & TH_RST) {
6159 		freemsg(mp);
6160 		return;
6161 	}
6162 	/*
6163 	 * Now copy the original header to a new buffer.  The reason
6164 	 * for doing this is that we need to put extra room before
6165 	 * the header for the MAC layer address.  The original mblk
6166 	 * does not have this extra head room.
6167 	 */
6168 	len = ip_hdr_len + sizeof (tcph_t);
6169 	if ((new_mp = allocb(len + tcp_wroff_xtra, 0)) == NULL) {
6170 		freemsg(mp);
6171 		return;
6172 	}
6173 	new_mp->b_rptr += tcp_wroff_xtra;
6174 	bcopy(mp->b_rptr, new_mp->b_rptr, len);
6175 	new_mp->b_wptr = new_mp->b_rptr + len;
6176 	freemsg(mp);
6177 	mp = new_mp;
6178 	iph = (struct ip *)mp->b_rptr;
6179 	tcph = (tcph_t *)&mp->b_rptr[ip_hdr_len];
6180 
6181 	tcph->th_offset_and_rsrvd[0] = (5 << 4);
6182 	tcp_xchg(tcph->th_fport, tcph->th_lport, 2);
6183 	U32_TO_BE32(ack, tcph->th_ack);
6184 	U32_TO_BE32(seq, tcph->th_seq);
6185 	U16_TO_BE16(0, tcph->th_win);
6186 	bzero(tcph->th_sum, sizeof (int16_t));
6187 	tcph->th_flags[0] = (uint8_t)ctl;
6188 	if (ctl & TH_RST) {
6189 		BUMP_MIB(tcp_mib.tcpOutRsts);
6190 		BUMP_MIB(tcp_mib.tcpOutControl);
6191 	}
6192 
6193 	iph->ip_len = htons(len);
6194 	/* Swap addresses */
6195 	addr = iph->ip_src.s_addr;
6196 	iph->ip_src = iph->ip_dst;
6197 	iph->ip_dst.s_addr = addr;
6198 	iph->ip_id = 0;
6199 	iph->ip_ttl = 0;
6200 	tcp_set_cksum(mp);
6201 	iph->ip_ttl = (uint8_t)tcp_ipv4_ttl;
6202 
6203 	/* Dump the packet when debugging. */
6204 	TCP_DUMP_PACKET("tcp_xmit_early_reset", mp);
6205 	(void) ipv4_tcp_output(sock_id, mp);
6206 	freemsg(mp);
6207 }
6208 
6209 static void
6210 tcp_set_cksum(mblk_t *mp)
6211 {
6212 	struct ip *iph;
6213 	tcpha_t *tcph;
6214 	int len;
6215 
6216 	iph = (struct ip *)mp->b_rptr;
6217 	tcph = (tcpha_t *)(iph + 1);
6218 	len = ntohs(iph->ip_len);
6219 	/*
6220 	 * Calculate the TCP checksum.  Need to include the psuedo header,
6221 	 * which is similar to the real IP header starting at the TTL field.
6222 	 */
6223 	iph->ip_sum = htons(len - IP_SIMPLE_HDR_LENGTH);
6224 	tcph->tha_sum = 0;
6225 	tcph->tha_sum = tcp_cksum((uint16_t *)&(iph->ip_ttl),
6226 	    len - IP_SIMPLE_HDR_LENGTH + 12);
6227 	iph->ip_sum = 0;
6228 }
6229 
6230 static uint16_t
6231 tcp_cksum(uint16_t *buf, uint32_t len)
6232 {
6233 	/*
6234 	 * Compute Internet Checksum for "count" bytes
6235 	 * beginning at location "addr".
6236 	 */
6237 	int32_t sum = 0;
6238 
6239 	while (len > 1) {
6240 		/*  This is the inner loop */
6241 		sum += *buf++;
6242 		len -= 2;
6243 	}
6244 
6245 	/*  Add left-over byte, if any */
6246 	if (len > 0)
6247 		sum += *(unsigned char *)buf * 256;
6248 
6249 	/*  Fold 32-bit sum to 16 bits */
6250 	while (sum >> 16)
6251 		sum = (sum & 0xffff) + (sum >> 16);
6252 
6253 	return ((uint16_t)~sum);
6254 }
6255 
6256 /*
6257  * Type three generator adapted from the random() function in 4.4 BSD:
6258  */
6259 
6260 /*
6261  * Copyright (c) 1983, 1993
6262  *	The Regents of the University of California.  All rights reserved.
6263  *
6264  * Redistribution and use in source and binary forms, with or without
6265  * modification, are permitted provided that the following conditions
6266  * are met:
6267  * 1. Redistributions of source code must retain the above copyright
6268  *    notice, this list of conditions and the following disclaimer.
6269  * 2. Redistributions in binary form must reproduce the above copyright
6270  *    notice, this list of conditions and the following disclaimer in the
6271  *    documentation and/or other materials provided with the distribution.
6272  * 3. All advertising materials mentioning features or use of this software
6273  *    must display the following acknowledgement:
6274  *	This product includes software developed by the University of
6275  *	California, Berkeley and its contributors.
6276  * 4. Neither the name of the University nor the names of its contributors
6277  *    may be used to endorse or promote products derived from this software
6278  *    without specific prior written permission.
6279  *
6280  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
6281  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
6282  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
6283  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
6284  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
6285  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
6286  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
6287  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
6288  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
6289  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
6290  * SUCH DAMAGE.
6291  */
6292 
6293 /* Type 3 -- x**31 + x**3 + 1 */
6294 #define	DEG_3		31
6295 #define	SEP_3		3
6296 
6297 
6298 /* Protected by tcp_random_lock */
6299 static int tcp_randtbl[DEG_3 + 1];
6300 
6301 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
6302 static int *tcp_random_rptr = &tcp_randtbl[1];
6303 
6304 static int *tcp_random_state = &tcp_randtbl[1];
6305 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
6306 
6307 static void
6308 tcp_random_init(void)
6309 {
6310 	int i;
6311 	uint32_t hrt;
6312 	uint32_t wallclock;
6313 	uint32_t result;
6314 
6315 	/*
6316 	 *
6317 	 * XXX We don't have high resolution time in standalone...  The
6318 	 * following is just some approximation on the comment below.
6319 	 *
6320 	 * Use high-res timer and current time for seed.  Gethrtime() returns
6321 	 * a longlong, which may contain resolution down to nanoseconds.
6322 	 * The current time will either be a 32-bit or a 64-bit quantity.
6323 	 * XOR the two together in a 64-bit result variable.
6324 	 * Convert the result to a 32-bit value by multiplying the high-order
6325 	 * 32-bits by the low-order 32-bits.
6326 	 *
6327 	 * XXX We don't have gethrtime() in prom and the wallclock....
6328 	 */
6329 
6330 	hrt = prom_gettime();
6331 	wallclock = (uint32_t)time(NULL);
6332 	result = wallclock ^ hrt;
6333 	tcp_random_state[0] = result;
6334 
6335 	for (i = 1; i < DEG_3; i++)
6336 		tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
6337 			+ 12345;
6338 	tcp_random_fptr = &tcp_random_state[SEP_3];
6339 	tcp_random_rptr = &tcp_random_state[0];
6340 	for (i = 0; i < 10 * DEG_3; i++)
6341 		(void) tcp_random();
6342 }
6343 
6344 /*
6345  * tcp_random: Return a random number in the range [1 - (128K + 1)].
6346  * This range is selected to be approximately centered on TCP_ISS / 2,
6347  * and easy to compute. We get this value by generating a 32-bit random
6348  * number, selecting out the high-order 17 bits, and then adding one so
6349  * that we never return zero.
6350  */
6351 static int
6352 tcp_random(void)
6353 {
6354 	int i;
6355 
6356 	*tcp_random_fptr += *tcp_random_rptr;
6357 
6358 	/*
6359 	 * The high-order bits are more random than the low-order bits,
6360 	 * so we select out the high-order 17 bits and add one so that
6361 	 * we never return zero.
6362 	 */
6363 	i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
6364 	if (++tcp_random_fptr >= tcp_random_end_ptr) {
6365 		tcp_random_fptr = tcp_random_state;
6366 		++tcp_random_rptr;
6367 	} else if (++tcp_random_rptr >= tcp_random_end_ptr)
6368 		tcp_random_rptr = tcp_random_state;
6369 
6370 	return (i);
6371 }
6372 
6373 /*
6374  * Generate ISS, taking into account NDD changes may happen halfway through.
6375  * (If the iss is not zero, set it.)
6376  */
6377 static void
6378 tcp_iss_init(tcp_t *tcp)
6379 {
6380 	tcp_iss_incr_extra += (ISS_INCR >> 1);
6381 	tcp->tcp_iss = tcp_iss_incr_extra;
6382 	tcp->tcp_iss += (prom_gettime() >> ISS_NSEC_SHT) + tcp_random();
6383 	tcp->tcp_valid_bits = TCP_ISS_VALID;
6384 	tcp->tcp_fss = tcp->tcp_iss - 1;
6385 	tcp->tcp_suna = tcp->tcp_iss;
6386 	tcp->tcp_snxt = tcp->tcp_iss + 1;
6387 	tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
6388 	tcp->tcp_csuna = tcp->tcp_snxt;
6389 }
6390 
6391 /*
6392  * Diagnostic routine used to return a string associated with the tcp state.
6393  * Note that if the caller does not supply a buffer, it will use an internal
6394  * static string.  This means that if multiple threads call this function at
6395  * the same time, output can be corrupted...  Note also that this function
6396  * does not check the size of the supplied buffer.  The caller has to make
6397  * sure that it is big enough.
6398  */
6399 static char *
6400 tcp_display(tcp_t *tcp, char *sup_buf, char format)
6401 {
6402 	char		buf1[30];
6403 	static char	priv_buf[INET_ADDRSTRLEN * 2 + 80];
6404 	char		*buf;
6405 	char		*cp;
6406 	char		local_addrbuf[INET_ADDRSTRLEN];
6407 	char		remote_addrbuf[INET_ADDRSTRLEN];
6408 	struct in_addr	addr;
6409 
6410 	if (sup_buf != NULL)
6411 		buf = sup_buf;
6412 	else
6413 		buf = priv_buf;
6414 
6415 	if (tcp == NULL)
6416 		return ("NULL_TCP");
6417 	switch (tcp->tcp_state) {
6418 	case TCPS_CLOSED:
6419 		cp = "TCP_CLOSED";
6420 		break;
6421 	case TCPS_IDLE:
6422 		cp = "TCP_IDLE";
6423 		break;
6424 	case TCPS_BOUND:
6425 		cp = "TCP_BOUND";
6426 		break;
6427 	case TCPS_LISTEN:
6428 		cp = "TCP_LISTEN";
6429 		break;
6430 	case TCPS_SYN_SENT:
6431 		cp = "TCP_SYN_SENT";
6432 		break;
6433 	case TCPS_SYN_RCVD:
6434 		cp = "TCP_SYN_RCVD";
6435 		break;
6436 	case TCPS_ESTABLISHED:
6437 		cp = "TCP_ESTABLISHED";
6438 		break;
6439 	case TCPS_CLOSE_WAIT:
6440 		cp = "TCP_CLOSE_WAIT";
6441 		break;
6442 	case TCPS_FIN_WAIT_1:
6443 		cp = "TCP_FIN_WAIT_1";
6444 		break;
6445 	case TCPS_CLOSING:
6446 		cp = "TCP_CLOSING";
6447 		break;
6448 	case TCPS_LAST_ACK:
6449 		cp = "TCP_LAST_ACK";
6450 		break;
6451 	case TCPS_FIN_WAIT_2:
6452 		cp = "TCP_FIN_WAIT_2";
6453 		break;
6454 	case TCPS_TIME_WAIT:
6455 		cp = "TCP_TIME_WAIT";
6456 		break;
6457 	default:
6458 		(void) sprintf(buf1, "TCPUnkState(%d)", tcp->tcp_state);
6459 		cp = buf1;
6460 		break;
6461 	}
6462 	switch (format) {
6463 	case DISP_ADDR_AND_PORT:
6464 		/*
6465 		 * Note that we use the remote address in the tcp_b
6466 		 * structure.  This means that it will print out
6467 		 * the real destination address, not the next hop's
6468 		 * address if source routing is used.
6469 		 */
6470 		addr.s_addr = tcp->tcp_bound_source;
6471 		bcopy(inet_ntoa(addr), local_addrbuf, sizeof (local_addrbuf));
6472 		addr.s_addr = tcp->tcp_remote;
6473 		bcopy(inet_ntoa(addr), remote_addrbuf, sizeof (remote_addrbuf));
6474 		(void) snprintf(buf, sizeof (priv_buf), "[%s.%u, %s.%u] %s",
6475 		    local_addrbuf, ntohs(tcp->tcp_lport), remote_addrbuf,
6476 		    ntohs(tcp->tcp_fport), cp);
6477 		break;
6478 	case DISP_PORT_ONLY:
6479 	default:
6480 		(void) snprintf(buf, sizeof (priv_buf), "[%u, %u] %s",
6481 		    ntohs(tcp->tcp_lport), ntohs(tcp->tcp_fport), cp);
6482 		break;
6483 	}
6484 
6485 	return (buf);
6486 }
6487 
6488 /*
6489  * Add a new piece to the tcp reassembly queue.  If the gap at the beginning
6490  * is filled, return as much as we can.  The message passed in may be
6491  * multi-part, chained using b_cont.  "start" is the starting sequence
6492  * number for this piece.
6493  */
6494 static mblk_t *
6495 tcp_reass(tcp_t *tcp, mblk_t *mp, uint32_t start)
6496 {
6497 	uint32_t	end;
6498 	mblk_t		*mp1;
6499 	mblk_t		*mp2;
6500 	mblk_t		*next_mp;
6501 	uint32_t	u1;
6502 
6503 	/* Walk through all the new pieces. */
6504 	do {
6505 		assert((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
6506 		    (uintptr_t)INT_MAX);
6507 		end = start + (int)(mp->b_wptr - mp->b_rptr);
6508 		next_mp = mp->b_cont;
6509 		if (start == end) {
6510 			/* Empty.  Blast it. */
6511 			freeb(mp);
6512 			continue;
6513 		}
6514 		mp->b_cont = NULL;
6515 		TCP_REASS_SET_SEQ(mp, start);
6516 		TCP_REASS_SET_END(mp, end);
6517 		mp1 = tcp->tcp_reass_tail;
6518 		if (!mp1) {
6519 			tcp->tcp_reass_tail = mp;
6520 			tcp->tcp_reass_head = mp;
6521 			BUMP_MIB(tcp_mib.tcpInDataUnorderSegs);
6522 			UPDATE_MIB(tcp_mib.tcpInDataUnorderBytes, end - start);
6523 			continue;
6524 		}
6525 		/* New stuff completely beyond tail? */
6526 		if (SEQ_GEQ(start, TCP_REASS_END(mp1))) {
6527 			/* Link it on end. */
6528 			mp1->b_cont = mp;
6529 			tcp->tcp_reass_tail = mp;
6530 			BUMP_MIB(tcp_mib.tcpInDataUnorderSegs);
6531 			UPDATE_MIB(tcp_mib.tcpInDataUnorderBytes, end - start);
6532 			continue;
6533 		}
6534 		mp1 = tcp->tcp_reass_head;
6535 		u1 = TCP_REASS_SEQ(mp1);
6536 		/* New stuff at the front? */
6537 		if (SEQ_LT(start, u1)) {
6538 			/* Yes... Check for overlap. */
6539 			mp->b_cont = mp1;
6540 			tcp->tcp_reass_head = mp;
6541 			tcp_reass_elim_overlap(tcp, mp);
6542 			continue;
6543 		}
6544 		/*
6545 		 * The new piece fits somewhere between the head and tail.
6546 		 * We find our slot, where mp1 precedes us and mp2 trails.
6547 		 */
6548 		for (; (mp2 = mp1->b_cont) != NULL; mp1 = mp2) {
6549 			u1 = TCP_REASS_SEQ(mp2);
6550 			if (SEQ_LEQ(start, u1))
6551 				break;
6552 		}
6553 		/* Link ourselves in */
6554 		mp->b_cont = mp2;
6555 		mp1->b_cont = mp;
6556 
6557 		/* Trim overlap with following mblk(s) first */
6558 		tcp_reass_elim_overlap(tcp, mp);
6559 
6560 		/* Trim overlap with preceding mblk */
6561 		tcp_reass_elim_overlap(tcp, mp1);
6562 
6563 	} while (start = end, mp = next_mp);
6564 	mp1 = tcp->tcp_reass_head;
6565 	/* Anything ready to go? */
6566 	if (TCP_REASS_SEQ(mp1) != tcp->tcp_rnxt)
6567 		return (NULL);
6568 	/* Eat what we can off the queue */
6569 	for (;;) {
6570 		mp = mp1->b_cont;
6571 		end = TCP_REASS_END(mp1);
6572 		TCP_REASS_SET_SEQ(mp1, 0);
6573 		TCP_REASS_SET_END(mp1, 0);
6574 		if (!mp) {
6575 			tcp->tcp_reass_tail = NULL;
6576 			break;
6577 		}
6578 		if (end != TCP_REASS_SEQ(mp)) {
6579 			mp1->b_cont = NULL;
6580 			break;
6581 		}
6582 		mp1 = mp;
6583 	}
6584 	mp1 = tcp->tcp_reass_head;
6585 	tcp->tcp_reass_head = mp;
6586 	return (mp1);
6587 }
6588 
6589 /* Eliminate any overlap that mp may have over later mblks */
6590 static void
6591 tcp_reass_elim_overlap(tcp_t *tcp, mblk_t *mp)
6592 {
6593 	uint32_t	end;
6594 	mblk_t		*mp1;
6595 	uint32_t	u1;
6596 
6597 	end = TCP_REASS_END(mp);
6598 	while ((mp1 = mp->b_cont) != NULL) {
6599 		u1 = TCP_REASS_SEQ(mp1);
6600 		if (!SEQ_GT(end, u1))
6601 			break;
6602 		if (!SEQ_GEQ(end, TCP_REASS_END(mp1))) {
6603 			mp->b_wptr -= end - u1;
6604 			TCP_REASS_SET_END(mp, u1);
6605 			BUMP_MIB(tcp_mib.tcpInDataPartDupSegs);
6606 			UPDATE_MIB(tcp_mib.tcpInDataPartDupBytes, end - u1);
6607 			break;
6608 		}
6609 		mp->b_cont = mp1->b_cont;
6610 		freeb(mp1);
6611 		BUMP_MIB(tcp_mib.tcpInDataDupSegs);
6612 		UPDATE_MIB(tcp_mib.tcpInDataDupBytes, end - u1);
6613 	}
6614 	if (!mp1)
6615 		tcp->tcp_reass_tail = mp;
6616 }
6617 
6618 /*
6619  * Remove a connection from the list of detached TIME_WAIT connections.
6620  */
6621 static void
6622 tcp_time_wait_remove(tcp_t *tcp)
6623 {
6624 	if (tcp->tcp_time_wait_expire == 0) {
6625 		assert(tcp->tcp_time_wait_next == NULL);
6626 		assert(tcp->tcp_time_wait_prev == NULL);
6627 		return;
6628 	}
6629 	assert(tcp->tcp_state == TCPS_TIME_WAIT);
6630 	if (tcp == tcp_time_wait_head) {
6631 		assert(tcp->tcp_time_wait_prev == NULL);
6632 		tcp_time_wait_head = tcp->tcp_time_wait_next;
6633 		if (tcp_time_wait_head != NULL) {
6634 			tcp_time_wait_head->tcp_time_wait_prev = NULL;
6635 		} else {
6636 			tcp_time_wait_tail = NULL;
6637 		}
6638 	} else if (tcp == tcp_time_wait_tail) {
6639 		assert(tcp != tcp_time_wait_head);
6640 		assert(tcp->tcp_time_wait_next == NULL);
6641 		tcp_time_wait_tail = tcp->tcp_time_wait_prev;
6642 		assert(tcp_time_wait_tail != NULL);
6643 		tcp_time_wait_tail->tcp_time_wait_next = NULL;
6644 	} else {
6645 		assert(tcp->tcp_time_wait_prev->tcp_time_wait_next == tcp);
6646 		assert(tcp->tcp_time_wait_next->tcp_time_wait_prev == tcp);
6647 		tcp->tcp_time_wait_prev->tcp_time_wait_next =
6648 		    tcp->tcp_time_wait_next;
6649 		tcp->tcp_time_wait_next->tcp_time_wait_prev =
6650 		    tcp->tcp_time_wait_prev;
6651 	}
6652 	tcp->tcp_time_wait_next = NULL;
6653 	tcp->tcp_time_wait_prev = NULL;
6654 	tcp->tcp_time_wait_expire = 0;
6655 }
6656 
6657 /*
6658  * Add a connection to the list of detached TIME_WAIT connections
6659  * and set its time to expire ...
6660  */
6661 static void
6662 tcp_time_wait_append(tcp_t *tcp)
6663 {
6664 	tcp->tcp_time_wait_expire = prom_gettime() + tcp_time_wait_interval;
6665 	if (tcp->tcp_time_wait_expire == 0)
6666 		tcp->tcp_time_wait_expire = 1;
6667 
6668 	if (tcp_time_wait_head == NULL) {
6669 		assert(tcp_time_wait_tail == NULL);
6670 		tcp_time_wait_head = tcp;
6671 	} else {
6672 		assert(tcp_time_wait_tail != NULL);
6673 		assert(tcp_time_wait_tail->tcp_state == TCPS_TIME_WAIT);
6674 		tcp_time_wait_tail->tcp_time_wait_next = tcp;
6675 		tcp->tcp_time_wait_prev = tcp_time_wait_tail;
6676 	}
6677 	tcp_time_wait_tail = tcp;
6678 
6679 	/* for ndd stats about compression */
6680 	tcp_cum_timewait++;
6681 }
6682 
6683 /*
6684  * Periodic qtimeout routine run on the default queue.
6685  * Performs 2 functions.
6686  * 	1.  Does TIME_WAIT compression on all recently added tcps. List
6687  *	    traversal is done backwards from the tail.
6688  *	2.  Blows away all tcps whose TIME_WAIT has expired. List traversal
6689  *	    is done forwards from the head.
6690  */
6691 void
6692 tcp_time_wait_collector(void)
6693 {
6694 	tcp_t *tcp;
6695 	uint32_t now;
6696 
6697 	/*
6698 	 * In order to reap time waits reliably, we should use a
6699 	 * source of time that is not adjustable by the user
6700 	 */
6701 	now = prom_gettime();
6702 	while ((tcp = tcp_time_wait_head) != NULL) {
6703 		/*
6704 		 * Compare times using modular arithmetic, since
6705 		 * lbolt can wrapover.
6706 		 */
6707 		if ((int32_t)(now - tcp->tcp_time_wait_expire) < 0) {
6708 			break;
6709 		}
6710 		/*
6711 		 * Note that the err must be 0 as there is no socket
6712 		 * associated with this TCP...
6713 		 */
6714 		(void) tcp_clean_death(-1, tcp, 0);
6715 	}
6716 	/* Schedule next run time. */
6717 	tcp_time_wait_runtime = prom_gettime() + 10000;
6718 }
6719 
6720 void
6721 tcp_time_wait_report(void)
6722 {
6723 	tcp_t *tcp;
6724 
6725 	printf("Current time %u\n", prom_gettime());
6726 	for (tcp = tcp_time_wait_head; tcp != NULL;
6727 	    tcp = tcp->tcp_time_wait_next) {
6728 		printf("%s expires at %u\n", tcp_display(tcp, NULL,
6729 		    DISP_ADDR_AND_PORT), tcp->tcp_time_wait_expire);
6730 	}
6731 }
6732 
6733 /*
6734  * Send up all messages queued on tcp_rcv_list.
6735  * Have to set tcp_co_norm since we use putnext.
6736  */
6737 static void
6738 tcp_rcv_drain(int sock_id, tcp_t *tcp)
6739 {
6740 	mblk_t *mp;
6741 	struct inetgram *in_gram;
6742 	mblk_t *in_mp;
6743 	int len;
6744 
6745 	/* Don't drain if the app has not finished reading all the data. */
6746 	if (sockets[sock_id].so_rcvbuf <= 0)
6747 		return;
6748 
6749 	/* We might have come here just to updated the rwnd */
6750 	if (tcp->tcp_rcv_list == NULL)
6751 		goto win_update;
6752 
6753 	if ((in_gram = (struct inetgram *)bkmem_zalloc(
6754 	    sizeof (struct inetgram))) == NULL) {
6755 		return;
6756 	}
6757 	if ((in_mp = allocb(tcp->tcp_rcv_cnt, 0)) == NULL) {
6758 		bkmem_free((caddr_t)in_gram, sizeof (struct inetgram));
6759 		return;
6760 	}
6761 	in_gram->igm_level = APP_LVL;
6762 	in_gram->igm_mp = in_mp;
6763 	in_gram->igm_id = 0;
6764 
6765 	while ((mp = tcp->tcp_rcv_list) != NULL) {
6766 		tcp->tcp_rcv_list = mp->b_cont;
6767 		len = mp->b_wptr - mp->b_rptr;
6768 		bcopy(mp->b_rptr, in_mp->b_wptr, len);
6769 		in_mp->b_wptr += len;
6770 		freeb(mp);
6771 	}
6772 
6773 	tcp->tcp_rcv_last_tail = NULL;
6774 	tcp->tcp_rcv_cnt = 0;
6775 	add_grams(&sockets[sock_id].inq, in_gram);
6776 
6777 	/* This means that so_rcvbuf can be less than 0. */
6778 	sockets[sock_id].so_rcvbuf -= in_mp->b_wptr - in_mp->b_rptr;
6779 win_update:
6780 	/*
6781 	 * Increase the receive window to max.  But we need to do receiver
6782 	 * SWS avoidance.  This means that we need to check the increase of
6783 	 * of receive window is at least 1 MSS.
6784 	 */
6785 	if (sockets[sock_id].so_rcvbuf > 0 &&
6786 	    (tcp->tcp_rwnd_max - tcp->tcp_rwnd >= tcp->tcp_mss)) {
6787 		tcp->tcp_rwnd = tcp->tcp_rwnd_max;
6788 		U32_TO_ABE16(tcp->tcp_rwnd >> tcp->tcp_rcv_ws,
6789 		    tcp->tcp_tcph->th_win);
6790 	}
6791 }
6792 
6793 /*
6794  * Wrapper for recvfrom to call
6795  */
6796 void
6797 tcp_rcv_drain_sock(int sock_id)
6798 {
6799 	tcp_t *tcp;
6800 	if ((tcp = sockets[sock_id].pcb) == NULL)
6801 		return;
6802 	tcp_rcv_drain(sock_id, tcp);
6803 }
6804 
6805 /*
6806  * If the inq == NULL and the tcp_rcv_list != NULL, we have data that
6807  * recvfrom could read. Place a magic message in the inq to let recvfrom
6808  * know that it needs to call tcp_rcv_drain_sock to pullup the data.
6809  */
6810 static void
6811 tcp_drain_needed(int sock_id, tcp_t *tcp)
6812 {
6813 	struct inetgram *in_gram;
6814 #ifdef DEBUG
6815 	printf("tcp_drain_needed: inq %x, tcp_rcv_list %x\n",
6816 		sockets[sock_id].inq, tcp->tcp_rcv_list);
6817 #endif
6818 	if ((sockets[sock_id].inq != NULL) ||
6819 		(tcp->tcp_rcv_list == NULL))
6820 		return;
6821 
6822 	if ((in_gram = (struct inetgram *)bkmem_zalloc(
6823 		sizeof (struct inetgram))) == NULL)
6824 		return;
6825 
6826 	in_gram->igm_level = APP_LVL;
6827 	in_gram->igm_mp = NULL;
6828 	in_gram->igm_id = TCP_CALLB_MAGIC_ID;
6829 
6830 	add_grams(&sockets[sock_id].inq, in_gram);
6831 }
6832 
6833 /*
6834  * Queue data on tcp_rcv_list which is a b_next chain.
6835  * Each element of the chain is a b_cont chain.
6836  *
6837  * M_DATA messages are added to the current element.
6838  * Other messages are added as new (b_next) elements.
6839  */
6840 static void
6841 tcp_rcv_enqueue(tcp_t *tcp, mblk_t *mp, uint_t seg_len)
6842 {
6843 	assert(seg_len == msgdsize(mp));
6844 	if (tcp->tcp_rcv_list == NULL) {
6845 		tcp->tcp_rcv_list = mp;
6846 	} else {
6847 		tcp->tcp_rcv_last_tail->b_cont = mp;
6848 	}
6849 	while (mp->b_cont)
6850 		mp = mp->b_cont;
6851 	tcp->tcp_rcv_last_tail = mp;
6852 	tcp->tcp_rcv_cnt += seg_len;
6853 	tcp->tcp_rwnd -= seg_len;
6854 #ifdef DEBUG
6855 	printf("tcp_rcv_enqueue rwnd %d\n", tcp->tcp_rwnd);
6856 #endif
6857 	U32_TO_ABE16(tcp->tcp_rwnd >> tcp->tcp_rcv_ws, tcp->tcp_tcph->th_win);
6858 }
6859 
6860 /* The minimum of smoothed mean deviation in RTO calculation. */
6861 #define	TCP_SD_MIN	400
6862 
6863 /*
6864  * Set RTO for this connection.  The formula is from Jacobson and Karels'
6865  * "Congestion Avoidance and Control" in SIGCOMM '88.  The variable names
6866  * are the same as those in Appendix A.2 of that paper.
6867  *
6868  * m = new measurement
6869  * sa = smoothed RTT average (8 * average estimates).
6870  * sv = smoothed mean deviation (mdev) of RTT (4 * deviation estimates).
6871  */
6872 static void
6873 tcp_set_rto(tcp_t *tcp, int32_t rtt)
6874 {
6875 	int32_t m = rtt;
6876 	uint32_t sa = tcp->tcp_rtt_sa;
6877 	uint32_t sv = tcp->tcp_rtt_sd;
6878 	uint32_t rto;
6879 
6880 	BUMP_MIB(tcp_mib.tcpRttUpdate);
6881 	tcp->tcp_rtt_update++;
6882 
6883 	/* tcp_rtt_sa is not 0 means this is a new sample. */
6884 	if (sa != 0) {
6885 		/*
6886 		 * Update average estimator:
6887 		 *	new rtt = 7/8 old rtt + 1/8 Error
6888 		 */
6889 
6890 		/* m is now Error in estimate. */
6891 		m -= sa >> 3;
6892 		if ((int32_t)(sa += m) <= 0) {
6893 			/*
6894 			 * Don't allow the smoothed average to be negative.
6895 			 * We use 0 to denote reinitialization of the
6896 			 * variables.
6897 			 */
6898 			sa = 1;
6899 		}
6900 
6901 		/*
6902 		 * Update deviation estimator:
6903 		 *	new mdev = 3/4 old mdev + 1/4 (abs(Error) - old mdev)
6904 		 */
6905 		if (m < 0)
6906 			m = -m;
6907 		m -= sv >> 2;
6908 		sv += m;
6909 	} else {
6910 		/*
6911 		 * This follows BSD's implementation.  So the reinitialized
6912 		 * RTO is 3 * m.  We cannot go less than 2 because if the
6913 		 * link is bandwidth dominated, doubling the window size
6914 		 * during slow start means doubling the RTT.  We want to be
6915 		 * more conservative when we reinitialize our estimates.  3
6916 		 * is just a convenient number.
6917 		 */
6918 		sa = m << 3;
6919 		sv = m << 1;
6920 	}
6921 	if (sv < TCP_SD_MIN) {
6922 		/*
6923 		 * We do not know that if sa captures the delay ACK
6924 		 * effect as in a long train of segments, a receiver
6925 		 * does not delay its ACKs.  So set the minimum of sv
6926 		 * to be TCP_SD_MIN, which is default to 400 ms, twice
6927 		 * of BSD DATO.  That means the minimum of mean
6928 		 * deviation is 100 ms.
6929 		 *
6930 		 */
6931 		sv = TCP_SD_MIN;
6932 	}
6933 	tcp->tcp_rtt_sa = sa;
6934 	tcp->tcp_rtt_sd = sv;
6935 	/*
6936 	 * RTO = average estimates (sa / 8) + 4 * deviation estimates (sv)
6937 	 *
6938 	 * Add tcp_rexmit_interval extra in case of extreme environment
6939 	 * where the algorithm fails to work.  The default value of
6940 	 * tcp_rexmit_interval_extra should be 0.
6941 	 *
6942 	 * As we use a finer grained clock than BSD and update
6943 	 * RTO for every ACKs, add in another .25 of RTT to the
6944 	 * deviation of RTO to accomodate burstiness of 1/4 of
6945 	 * window size.
6946 	 */
6947 	rto = (sa >> 3) + sv + tcp_rexmit_interval_extra + (sa >> 5);
6948 
6949 	if (rto > tcp_rexmit_interval_max) {
6950 		tcp->tcp_rto = tcp_rexmit_interval_max;
6951 	} else if (rto < tcp_rexmit_interval_min) {
6952 		tcp->tcp_rto = tcp_rexmit_interval_min;
6953 	} else {
6954 		tcp->tcp_rto = rto;
6955 	}
6956 
6957 	/* Now, we can reset tcp_timer_backoff to use the new RTO... */
6958 	tcp->tcp_timer_backoff = 0;
6959 }
6960 
6961 /*
6962  * Initiate closedown sequence on an active connection.
6963  * Return value zero for OK return, non-zero for error return.
6964  */
6965 static int
6966 tcp_xmit_end(tcp_t *tcp, int sock_id)
6967 {
6968 	mblk_t	*mp;
6969 
6970 	if (tcp->tcp_state < TCPS_SYN_RCVD ||
6971 	    tcp->tcp_state > TCPS_CLOSE_WAIT) {
6972 		/*
6973 		 * Invalid state, only states TCPS_SYN_RCVD,
6974 		 * TCPS_ESTABLISHED and TCPS_CLOSE_WAIT are valid
6975 		 */
6976 		return (-1);
6977 	}
6978 
6979 	tcp->tcp_fss = tcp->tcp_snxt + tcp->tcp_unsent;
6980 	tcp->tcp_valid_bits |= TCP_FSS_VALID;
6981 	/*
6982 	 * If there is nothing more unsent, send the FIN now.
6983 	 * Otherwise, it will go out with the last segment.
6984 	 */
6985 	if (tcp->tcp_unsent == 0) {
6986 		mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
6987 		    tcp->tcp_fss, B_FALSE, NULL, B_FALSE);
6988 
6989 		if (mp != NULL) {
6990 			/* Dump the packet when debugging. */
6991 			TCP_DUMP_PACKET("tcp_xmit_end", mp);
6992 			(void) ipv4_tcp_output(sock_id, mp);
6993 			freeb(mp);
6994 		} else {
6995 			/*
6996 			 * Couldn't allocate msg.  Pretend we got it out.
6997 			 * Wait for rexmit timeout.
6998 			 */
6999 			tcp->tcp_snxt = tcp->tcp_fss + 1;
7000 			TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
7001 		}
7002 
7003 		/*
7004 		 * If needed, update tcp_rexmit_snxt as tcp_snxt is
7005 		 * changed.
7006 		 */
7007 		if (tcp->tcp_rexmit && tcp->tcp_rexmit_nxt == tcp->tcp_fss) {
7008 			tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
7009 		}
7010 	} else {
7011 		tcp_wput_data(tcp, NULL, B_FALSE);
7012 	}
7013 
7014 	return (0);
7015 }
7016 
7017 int
7018 tcp_opt_set(tcp_t *tcp, int level, int option, const void *optval,
7019     socklen_t optlen)
7020 {
7021 	switch (level) {
7022 	case SOL_SOCKET: {
7023 		switch (option) {
7024 		case SO_RCVBUF:
7025 			if (optlen == sizeof (int)) {
7026 				int val = *(int *)optval;
7027 
7028 				if (val > tcp_max_buf) {
7029 					errno = ENOBUFS;
7030 					break;
7031 				}
7032 				/* Silently ignore zero */
7033 				if (val != 0) {
7034 					val = MSS_ROUNDUP(val, tcp->tcp_mss);
7035 					(void) tcp_rwnd_set(tcp, val);
7036 				}
7037 			} else {
7038 				errno = EINVAL;
7039 			}
7040 			break;
7041 		case SO_SNDBUF:
7042 			if (optlen == sizeof (int)) {
7043 				tcp->tcp_xmit_hiwater = *(int *)optval;
7044 				if (tcp->tcp_xmit_hiwater > tcp_max_buf)
7045 					tcp->tcp_xmit_hiwater = tcp_max_buf;
7046 			} else {
7047 				errno = EINVAL;
7048 			}
7049 			break;
7050 		case SO_LINGER:
7051 			if (optlen == sizeof (struct linger)) {
7052 				struct linger *lgr = (struct linger *)optval;
7053 
7054 				if (lgr->l_onoff) {
7055 					tcp->tcp_linger = 1;
7056 					tcp->tcp_lingertime = lgr->l_linger;
7057 				} else {
7058 					tcp->tcp_linger = 0;
7059 					tcp->tcp_lingertime = 0;
7060 				}
7061 			} else {
7062 				errno = EINVAL;
7063 			}
7064 			break;
7065 		default:
7066 			errno = ENOPROTOOPT;
7067 			break;
7068 		}
7069 		break;
7070 	} /* case SOL_SOCKET */
7071 	case IPPROTO_TCP: {
7072 		switch (option) {
7073 		default:
7074 			errno = ENOPROTOOPT;
7075 			break;
7076 		}
7077 		break;
7078 	} /* case IPPROTO_TCP */
7079 	case IPPROTO_IP: {
7080 		switch (option) {
7081 		default:
7082 			errno = ENOPROTOOPT;
7083 			break;
7084 		}
7085 		break;
7086 	} /* case IPPROTO_IP */
7087 	default:
7088 		errno = ENOPROTOOPT;
7089 		break;
7090 	} /* switch (level) */
7091 
7092 	if (errno != 0)
7093 		return (-1);
7094 	else
7095 		return (0);
7096 }
7097