xref: /titanic_51/usr/src/uts/common/inet/tcp/tcp.c (revision 81d9f076db88c1f40c85831ce1ebb444a209c5a8)
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 2010 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 /* Copyright (c) 1990 Mentat Inc. */
27 
28 #include <sys/types.h>
29 #include <sys/stream.h>
30 #include <sys/strsun.h>
31 #include <sys/strsubr.h>
32 #include <sys/stropts.h>
33 #include <sys/strlog.h>
34 #define	_SUN_TPI_VERSION 2
35 #include <sys/tihdr.h>
36 #include <sys/timod.h>
37 #include <sys/ddi.h>
38 #include <sys/sunddi.h>
39 #include <sys/suntpi.h>
40 #include <sys/xti_inet.h>
41 #include <sys/cmn_err.h>
42 #include <sys/debug.h>
43 #include <sys/sdt.h>
44 #include <sys/vtrace.h>
45 #include <sys/kmem.h>
46 #include <sys/ethernet.h>
47 #include <sys/cpuvar.h>
48 #include <sys/dlpi.h>
49 #include <sys/pattr.h>
50 #include <sys/policy.h>
51 #include <sys/priv.h>
52 #include <sys/zone.h>
53 #include <sys/sunldi.h>
54 
55 #include <sys/errno.h>
56 #include <sys/signal.h>
57 #include <sys/socket.h>
58 #include <sys/socketvar.h>
59 #include <sys/sockio.h>
60 #include <sys/isa_defs.h>
61 #include <sys/md5.h>
62 #include <sys/random.h>
63 #include <sys/uio.h>
64 #include <sys/systm.h>
65 #include <netinet/in.h>
66 #include <netinet/tcp.h>
67 #include <netinet/ip6.h>
68 #include <netinet/icmp6.h>
69 #include <net/if.h>
70 #include <net/route.h>
71 #include <inet/ipsec_impl.h>
72 
73 #include <inet/common.h>
74 #include <inet/ip.h>
75 #include <inet/ip_impl.h>
76 #include <inet/ip6.h>
77 #include <inet/ip_ndp.h>
78 #include <inet/proto_set.h>
79 #include <inet/mib2.h>
80 #include <inet/nd.h>
81 #include <inet/optcom.h>
82 #include <inet/snmpcom.h>
83 #include <inet/kstatcom.h>
84 #include <inet/tcp.h>
85 #include <inet/tcp_impl.h>
86 #include <inet/tcp_cluster.h>
87 #include <inet/udp_impl.h>
88 #include <net/pfkeyv2.h>
89 #include <inet/ipdrop.h>
90 
91 #include <inet/ipclassifier.h>
92 #include <inet/ip_ire.h>
93 #include <inet/ip_ftable.h>
94 #include <inet/ip_if.h>
95 #include <inet/ipp_common.h>
96 #include <inet/ip_rts.h>
97 #include <inet/ip_netinfo.h>
98 #include <sys/squeue_impl.h>
99 #include <sys/squeue.h>
100 #include <inet/kssl/ksslapi.h>
101 #include <sys/tsol/label.h>
102 #include <sys/tsol/tnet.h>
103 #include <rpc/pmap_prot.h>
104 #include <sys/callo.h>
105 
106 /*
107  * TCP Notes: aka FireEngine Phase I (PSARC 2002/433)
108  *
109  * (Read the detailed design doc in PSARC case directory)
110  *
111  * The entire tcp state is contained in tcp_t and conn_t structure
112  * which are allocated in tandem using ipcl_conn_create() and passing
113  * IPCL_TCPCONN as a flag. We use 'conn_ref' and 'conn_lock' to protect
114  * the references on the tcp_t. The tcp_t structure is never compressed
115  * and packets always land on the correct TCP perimeter from the time
116  * eager is created till the time tcp_t dies (as such the old mentat
117  * TCP global queue is not used for detached state and no IPSEC checking
118  * is required). The global queue is still allocated to send out resets
119  * for connection which have no listeners and IP directly calls
120  * tcp_xmit_listeners_reset() which does any policy check.
121  *
122  * Protection and Synchronisation mechanism:
123  *
124  * The tcp data structure does not use any kind of lock for protecting
125  * its state but instead uses 'squeues' for mutual exclusion from various
126  * read and write side threads. To access a tcp member, the thread should
127  * always be behind squeue (via squeue_enter with flags as SQ_FILL, SQ_PROCESS,
128  * or SQ_NODRAIN). Since the squeues allow a direct function call, caller
129  * can pass any tcp function having prototype of edesc_t as argument
130  * (different from traditional STREAMs model where packets come in only
131  * designated entry points). The list of functions that can be directly
132  * called via squeue are listed before the usual function prototype.
133  *
134  * Referencing:
135  *
136  * TCP is MT-Hot and we use a reference based scheme to make sure that the
137  * tcp structure doesn't disappear when its needed. When the application
138  * creates an outgoing connection or accepts an incoming connection, we
139  * start out with 2 references on 'conn_ref'. One for TCP and one for IP.
140  * The IP reference is just a symbolic reference since ip_tcpclose()
141  * looks at tcp structure after tcp_close_output() returns which could
142  * have dropped the last TCP reference. So as long as the connection is
143  * in attached state i.e. !TCP_IS_DETACHED, we have 2 references on the
144  * conn_t. The classifier puts its own reference when the connection is
145  * inserted in listen or connected hash. Anytime a thread needs to enter
146  * the tcp connection perimeter, it retrieves the conn/tcp from q->ptr
147  * on write side or by doing a classify on read side and then puts a
148  * reference on the conn before doing squeue_enter/tryenter/fill. For
149  * read side, the classifier itself puts the reference under fanout lock
150  * to make sure that tcp can't disappear before it gets processed. The
151  * squeue will drop this reference automatically so the called function
152  * doesn't have to do a DEC_REF.
153  *
154  * Opening a new connection:
155  *
156  * The outgoing connection open is pretty simple. tcp_open() does the
157  * work in creating the conn/tcp structure and initializing it. The
158  * squeue assignment is done based on the CPU the application
159  * is running on. So for outbound connections, processing is always done
160  * on application CPU which might be different from the incoming CPU
161  * being interrupted by the NIC. An optimal way would be to figure out
162  * the NIC <-> CPU binding at listen time, and assign the outgoing
163  * connection to the squeue attached to the CPU that will be interrupted
164  * for incoming packets (we know the NIC based on the bind IP address).
165  * This might seem like a problem if more data is going out but the
166  * fact is that in most cases the transmit is ACK driven transmit where
167  * the outgoing data normally sits on TCP's xmit queue waiting to be
168  * transmitted.
169  *
170  * Accepting a connection:
171  *
172  * This is a more interesting case because of various races involved in
173  * establishing a eager in its own perimeter. Read the meta comment on
174  * top of tcp_input_listener(). But briefly, the squeue is picked by
175  * ip_fanout based on the ring or the sender (if loopback).
176  *
177  * Closing a connection:
178  *
179  * The close is fairly straight forward. tcp_close() calls tcp_close_output()
180  * via squeue to do the close and mark the tcp as detached if the connection
181  * was in state TCPS_ESTABLISHED or greater. In the later case, TCP keep its
182  * reference but tcp_close() drop IP's reference always. So if tcp was
183  * not killed, it is sitting in time_wait list with 2 reference - 1 for TCP
184  * and 1 because it is in classifier's connected hash. This is the condition
185  * we use to determine that its OK to clean up the tcp outside of squeue
186  * when time wait expires (check the ref under fanout and conn_lock and
187  * if it is 2, remove it from fanout hash and kill it).
188  *
189  * Although close just drops the necessary references and marks the
190  * tcp_detached state, tcp_close needs to know the tcp_detached has been
191  * set (under squeue) before letting the STREAM go away (because a
192  * inbound packet might attempt to go up the STREAM while the close
193  * has happened and tcp_detached is not set). So a special lock and
194  * flag is used along with a condition variable (tcp_closelock, tcp_closed,
195  * and tcp_closecv) to signal tcp_close that tcp_close_out() has marked
196  * tcp_detached.
197  *
198  * Special provisions and fast paths:
199  *
200  * We make special provisions for sockfs by marking tcp_issocket
201  * whenever we have only sockfs on top of TCP. This allows us to skip
202  * putting the tcp in acceptor hash since a sockfs listener can never
203  * become acceptor and also avoid allocating a tcp_t for acceptor STREAM
204  * since eager has already been allocated and the accept now happens
205  * on acceptor STREAM. There is a big blob of comment on top of
206  * tcp_input_listener explaining the new accept. When socket is POP'd,
207  * sockfs sends us an ioctl to mark the fact and we go back to old
208  * behaviour. Once tcp_issocket is unset, its never set for the
209  * life of that connection.
210  *
211  * IPsec notes :
212  *
213  * Since a packet is always executed on the correct TCP perimeter
214  * all IPsec processing is defered to IP including checking new
215  * connections and setting IPSEC policies for new connection. The
216  * only exception is tcp_xmit_listeners_reset() which is called
217  * directly from IP and needs to policy check to see if TH_RST
218  * can be sent out.
219  */
220 
221 /*
222  * Values for squeue switch:
223  * 1: SQ_NODRAIN
224  * 2: SQ_PROCESS
225  * 3: SQ_FILL
226  */
227 int tcp_squeue_wput = 2;	/* /etc/systems */
228 int tcp_squeue_flag;
229 
230 kmem_cache_t	*tcp_sack_info_cache;
231 
232 /*
233  * To prevent memory hog, limit the number of entries in tcp_free_list
234  * to 1% of available memory / number of cpus
235  */
236 uint_t tcp_free_list_max_cnt = 0;
237 
238 #define	TCP_XMIT_LOWATER	4096
239 #define	TCP_XMIT_HIWATER	49152
240 #define	TCP_RECV_LOWATER	2048
241 #define	TCP_RECV_HIWATER	128000
242 
243 #define	TIDUSZ	4096	/* transport interface data unit size */
244 
245 /*
246  * Size of acceptor hash list.  It has to be a power of 2 for hashing.
247  */
248 #define	TCP_ACCEPTOR_FANOUT_SIZE		256
249 
250 #ifdef	_ILP32
251 #define	TCP_ACCEPTOR_HASH(accid)					\
252 		(((uint_t)(accid) >> 8) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
253 #else
254 #define	TCP_ACCEPTOR_HASH(accid)					\
255 		((uint_t)(accid) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
256 #endif	/* _ILP32 */
257 
258 /* Minimum number of connections per listener. */
259 static uint32_t tcp_min_conn_listener = 2;
260 
261 uint32_t tcp_early_abort = 30;
262 
263 /* TCP Timer control structure */
264 typedef struct tcpt_s {
265 	pfv_t	tcpt_pfv;	/* The routine we are to call */
266 	tcp_t	*tcpt_tcp;	/* The parameter we are to pass in */
267 } tcpt_t;
268 
269 /*
270  * Functions called directly via squeue having a prototype of edesc_t.
271  */
272 void		tcp_input_listener(void *arg, mblk_t *mp, void *arg2,
273     ip_recv_attr_t *ira);
274 void		tcp_input_data(void *arg, mblk_t *mp, void *arg2,
275     ip_recv_attr_t *ira);
276 static void	tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2,
277     ip_recv_attr_t *dummy);
278 
279 
280 /* Prototype for TCP functions */
281 static void	tcp_random_init(void);
282 int		tcp_random(void);
283 static int	tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp,
284 		    in_port_t dstport, uint_t srcid);
285 static int	tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp,
286 		    in_port_t dstport, uint32_t flowinfo,
287 		    uint_t srcid, uint32_t scope_id);
288 static int	tcp_extra_priv_ports_get(queue_t *q, mblk_t *mp, caddr_t cp,
289 		    cred_t *cr);
290 static int	tcp_extra_priv_ports_add(queue_t *q, mblk_t *mp,
291 		    char *value, caddr_t cp, cred_t *cr);
292 static int	tcp_extra_priv_ports_del(queue_t *q, mblk_t *mp,
293 		    char *value, caddr_t cp, cred_t *cr);
294 static void	tcp_iss_init(tcp_t *tcp);
295 static int	tcp_param_get(queue_t *q, mblk_t *mp, caddr_t cp, cred_t *cr);
296 static boolean_t tcp_param_register(IDP *ndp, tcpparam_t *tcppa, int cnt,
297     tcp_stack_t *);
298 static int	tcp_param_set(queue_t *q, mblk_t *mp, char *value,
299 		    caddr_t cp, cred_t *cr);
300 static int	tcp_param_set_aligned(queue_t *q, mblk_t *mp, char *value,
301 		    caddr_t cp, cred_t *cr);
302 static void	tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *);
303 static int	tcp_1948_phrase_set(queue_t *q, mblk_t *mp, char *value,
304 		    caddr_t cp, cred_t *cr);
305 static void	tcp_reinit(tcp_t *tcp);
306 static void	tcp_reinit_values(tcp_t *tcp);
307 
308 static void	tcp_wsrv(queue_t *q);
309 static void	tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa);
310 static void	tcp_update_zcopy(tcp_t *tcp);
311 static void	tcp_notify(void *, ip_xmit_attr_t *, ixa_notify_type_t,
312     ixa_notify_arg_t);
313 static void	*tcp_stack_init(netstackid_t stackid, netstack_t *ns);
314 static void	tcp_stack_fini(netstackid_t stackid, void *arg);
315 
316 static int	tcp_squeue_switch(int);
317 
318 static int	tcp_open(queue_t *, dev_t *, int, int, cred_t *, boolean_t);
319 static int	tcp_openv4(queue_t *, dev_t *, int, int, cred_t *);
320 static int	tcp_openv6(queue_t *, dev_t *, int, int, cred_t *);
321 
322 static void	tcp_squeue_add(squeue_t *);
323 
324 struct module_info tcp_rinfo =  {
325 	TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, TCP_RECV_HIWATER, TCP_RECV_LOWATER
326 };
327 
328 static struct module_info tcp_winfo =  {
329 	TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, 127, 16
330 };
331 
332 /*
333  * Entry points for TCP as a device. The normal case which supports
334  * the TCP functionality.
335  * We have separate open functions for the /dev/tcp and /dev/tcp6 devices.
336  */
337 struct qinit tcp_rinitv4 = {
338 	NULL, (pfi_t)tcp_rsrv, tcp_openv4, tcp_tpi_close, NULL, &tcp_rinfo
339 };
340 
341 struct qinit tcp_rinitv6 = {
342 	NULL, (pfi_t)tcp_rsrv, tcp_openv6, tcp_tpi_close, NULL, &tcp_rinfo
343 };
344 
345 struct qinit tcp_winit = {
346 	(pfi_t)tcp_wput, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
347 };
348 
349 /* Initial entry point for TCP in socket mode. */
350 struct qinit tcp_sock_winit = {
351 	(pfi_t)tcp_wput_sock, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
352 };
353 
354 /* TCP entry point during fallback */
355 struct qinit tcp_fallback_sock_winit = {
356 	(pfi_t)tcp_wput_fallback, NULL, NULL, NULL, NULL, &tcp_winfo
357 };
358 
359 /*
360  * Entry points for TCP as a acceptor STREAM opened by sockfs when doing
361  * an accept. Avoid allocating data structures since eager has already
362  * been created.
363  */
364 struct qinit tcp_acceptor_rinit = {
365 	NULL, (pfi_t)tcp_rsrv, NULL, tcp_tpi_close_accept, NULL, &tcp_winfo
366 };
367 
368 struct qinit tcp_acceptor_winit = {
369 	(pfi_t)tcp_tpi_accept, NULL, NULL, NULL, NULL, &tcp_winfo
370 };
371 
372 /* For AF_INET aka /dev/tcp */
373 struct streamtab tcpinfov4 = {
374 	&tcp_rinitv4, &tcp_winit
375 };
376 
377 /* For AF_INET6 aka /dev/tcp6 */
378 struct streamtab tcpinfov6 = {
379 	&tcp_rinitv6, &tcp_winit
380 };
381 
382 /*
383  * Following assumes TPI alignment requirements stay along 32 bit
384  * boundaries
385  */
386 #define	ROUNDUP32(x) \
387 	(((x) + (sizeof (int32_t) - 1)) & ~(sizeof (int32_t) - 1))
388 
389 /* Template for response to info request. */
390 struct T_info_ack tcp_g_t_info_ack = {
391 	T_INFO_ACK,		/* PRIM_type */
392 	0,			/* TSDU_size */
393 	T_INFINITE,		/* ETSDU_size */
394 	T_INVALID,		/* CDATA_size */
395 	T_INVALID,		/* DDATA_size */
396 	sizeof (sin_t),		/* ADDR_size */
397 	0,			/* OPT_size - not initialized here */
398 	TIDUSZ,			/* TIDU_size */
399 	T_COTS_ORD,		/* SERV_type */
400 	TCPS_IDLE,		/* CURRENT_state */
401 	(XPG4_1|EXPINLINE)	/* PROVIDER_flag */
402 };
403 
404 struct T_info_ack tcp_g_t_info_ack_v6 = {
405 	T_INFO_ACK,		/* PRIM_type */
406 	0,			/* TSDU_size */
407 	T_INFINITE,		/* ETSDU_size */
408 	T_INVALID,		/* CDATA_size */
409 	T_INVALID,		/* DDATA_size */
410 	sizeof (sin6_t),	/* ADDR_size */
411 	0,			/* OPT_size - not initialized here */
412 	TIDUSZ,		/* TIDU_size */
413 	T_COTS_ORD,		/* SERV_type */
414 	TCPS_IDLE,		/* CURRENT_state */
415 	(XPG4_1|EXPINLINE)	/* PROVIDER_flag */
416 };
417 
418 #define	PARAM_MAX (~(uint32_t)0)
419 
420 /* Max size IP datagram is 64k - 1 */
421 #define	TCP_MSS_MAX_IPV4 (IP_MAXPACKET - (sizeof (ipha_t) + sizeof (tcpha_t)))
422 #define	TCP_MSS_MAX_IPV6 (IP_MAXPACKET - (sizeof (ip6_t) + sizeof (tcpha_t)))
423 /* Max of the above */
424 #define	TCP_MSS_MAX	TCP_MSS_MAX_IPV4
425 
426 /* Largest TCP port number */
427 #define	TCP_MAX_PORT	(64 * 1024 - 1)
428 
429 /*
430  * tcp_wroff_xtra is the extra space in front of TCP/IP header for link
431  * layer header.  It has to be a multiple of 4.
432  */
433 static tcpparam_t lcl_tcp_wroff_xtra_param = { 0, 256, 32, "tcp_wroff_xtra" };
434 
435 #define	MB	(1024 * 1024)
436 
437 /*
438  * All of these are alterable, within the min/max values given, at run time.
439  * Note that the default value of "tcp_time_wait_interval" is four minutes,
440  * per the TCP spec.
441  */
442 /* BEGIN CSTYLED */
443 static tcpparam_t	lcl_tcp_param_arr[] = {
444  /*min		max		value		name */
445  { 1*SECONDS,	10*MINUTES,	1*MINUTES,	"tcp_time_wait_interval"},
446  { 1,		PARAM_MAX,	128,		"tcp_conn_req_max_q" },
447  { 0,		PARAM_MAX,	1024,		"tcp_conn_req_max_q0" },
448  { 1,		1024,		1,		"tcp_conn_req_min" },
449  { 0*MS,	20*SECONDS,	0*MS,		"tcp_conn_grace_period" },
450  { 128,		(1<<30),	1*MB,		"tcp_cwnd_max" },
451  { 0,		10,		0,		"tcp_debug" },
452  { 1024,	(32*1024),	1024,		"tcp_smallest_nonpriv_port"},
453  { 1*SECONDS,	PARAM_MAX,	3*MINUTES,	"tcp_ip_abort_cinterval"},
454  { 1*SECONDS,	PARAM_MAX,	3*MINUTES,	"tcp_ip_abort_linterval"},
455  { 500*MS,	PARAM_MAX,	5*MINUTES,	"tcp_ip_abort_interval"},
456  { 1*SECONDS,	PARAM_MAX,	10*SECONDS,	"tcp_ip_notify_cinterval"},
457  { 500*MS,	PARAM_MAX,	10*SECONDS,	"tcp_ip_notify_interval"},
458  { 1,		255,		64,		"tcp_ipv4_ttl"},
459  { 10*SECONDS,	10*DAYS,	2*HOURS,	"tcp_keepalive_interval"},
460  { 0,		100,		10,		"tcp_maxpsz_multiplier" },
461  { 1,		TCP_MSS_MAX_IPV4, 536,		"tcp_mss_def_ipv4"},
462  { 1,		TCP_MSS_MAX_IPV4, TCP_MSS_MAX_IPV4, "tcp_mss_max_ipv4"},
463  { 1,		TCP_MSS_MAX,	108,		"tcp_mss_min"},
464  { 1,		(64*1024)-1,	(4*1024)-1,	"tcp_naglim_def"},
465  { 1*MS,	20*SECONDS,	1*SECONDS,	"tcp_rexmit_interval_initial"},
466  { 1*MS,	2*HOURS,	60*SECONDS,	"tcp_rexmit_interval_max"},
467  { 1*MS,	2*HOURS,	400*MS,		"tcp_rexmit_interval_min"},
468  { 1*MS,	1*MINUTES,	100*MS,		"tcp_deferred_ack_interval" },
469  { 0,		16,		0,		"tcp_snd_lowat_fraction" },
470  { 1,		10000,		3,		"tcp_dupack_fast_retransmit" },
471  { 0,		1,		0,		"tcp_ignore_path_mtu" },
472  { 1024,	TCP_MAX_PORT,	32*1024,	"tcp_smallest_anon_port"},
473  { 1024,	TCP_MAX_PORT,	TCP_MAX_PORT,	"tcp_largest_anon_port"},
474  { TCP_XMIT_LOWATER, (1<<30), TCP_XMIT_HIWATER,"tcp_xmit_hiwat"},
475  { TCP_XMIT_LOWATER, (1<<30), TCP_XMIT_LOWATER,"tcp_xmit_lowat"},
476  { TCP_RECV_LOWATER, (1<<30), TCP_RECV_HIWATER,"tcp_recv_hiwat"},
477  { 1,		65536,		4,		"tcp_recv_hiwat_minmss"},
478  { 1*SECONDS,	PARAM_MAX,	675*SECONDS,	"tcp_fin_wait_2_flush_interval"},
479  { 8192,	(1<<30),	1*MB,		"tcp_max_buf"},
480 /*
481  * Question:  What default value should I set for tcp_strong_iss?
482  */
483  { 0,		2,		1,		"tcp_strong_iss"},
484  { 0,		65536,		20,		"tcp_rtt_updates"},
485  { 0,		1,		1,		"tcp_wscale_always"},
486  { 0,		1,		0,		"tcp_tstamp_always"},
487  { 0,		1,		1,		"tcp_tstamp_if_wscale"},
488  { 0*MS,	2*HOURS,	0*MS,		"tcp_rexmit_interval_extra"},
489  { 0,		16,		2,		"tcp_deferred_acks_max"},
490  { 1,		16384,		4,		"tcp_slow_start_after_idle"},
491  { 1,		4,		4,		"tcp_slow_start_initial"},
492  { 0,		2,		2,		"tcp_sack_permitted"},
493  { 0,		IPV6_MAX_HOPS,	IPV6_DEFAULT_HOPS,	"tcp_ipv6_hoplimit"},
494  { 1,		TCP_MSS_MAX_IPV6, 1220,		"tcp_mss_def_ipv6"},
495  { 1,		TCP_MSS_MAX_IPV6, TCP_MSS_MAX_IPV6, "tcp_mss_max_ipv6"},
496  { 0,		1,		0,		"tcp_rev_src_routes"},
497  { 10*MS,	500*MS,		50*MS,		"tcp_local_dack_interval"},
498  { 0,		16,		8,		"tcp_local_dacks_max"},
499  { 0,		2,		1,		"tcp_ecn_permitted"},
500  { 0,		1,		1,		"tcp_rst_sent_rate_enabled"},
501  { 0,		PARAM_MAX,	40,		"tcp_rst_sent_rate"},
502  { 0,		100*MS,		50*MS,		"tcp_push_timer_interval"},
503  { 0,		1,		0,		"tcp_use_smss_as_mss_opt"},
504  { 0,		PARAM_MAX,	8*MINUTES,	"tcp_keepalive_abort_interval"},
505  { 0,		1,		0,		"tcp_dev_flow_ctl"},
506  { 0,		PARAM_MAX,	100*SECONDS,	"tcp_reass_timeout"}
507 };
508 /* END CSTYLED */
509 
510 #define	IS_VMLOANED_MBLK(mp) \
511 	(((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0)
512 
513 uint32_t do_tcpzcopy = 1;		/* 0: disable, 1: enable, 2: force */
514 
515 /*
516  * Forces all connections to obey the value of the tcps_maxpsz_multiplier
517  * tunable settable via NDD.  Otherwise, the per-connection behavior is
518  * determined dynamically during tcp_set_destination(), which is the default.
519  */
520 boolean_t tcp_static_maxpsz = B_FALSE;
521 
522 /*
523  * If the receive buffer size is changed, this function is called to update
524  * the upper socket layer on the new delayed receive wake up threshold.
525  */
526 static void
527 tcp_set_recv_threshold(tcp_t *tcp, uint32_t new_rcvthresh)
528 {
529 	uint32_t default_threshold = SOCKET_RECVHIWATER >> 3;
530 
531 	if (IPCL_IS_NONSTR(tcp->tcp_connp)) {
532 		conn_t *connp = tcp->tcp_connp;
533 		struct sock_proto_props sopp;
534 
535 		/*
536 		 * only increase rcvthresh upto default_threshold
537 		 */
538 		if (new_rcvthresh > default_threshold)
539 			new_rcvthresh = default_threshold;
540 
541 		sopp.sopp_flags = SOCKOPT_RCVTHRESH;
542 		sopp.sopp_rcvthresh = new_rcvthresh;
543 
544 		(*connp->conn_upcalls->su_set_proto_props)
545 		    (connp->conn_upper_handle, &sopp);
546 	}
547 }
548 
549 /*
550  * Figure out the value of window scale opton.  Note that the rwnd is
551  * ASSUMED to be rounded up to the nearest MSS before the calculation.
552  * We cannot find the scale value and then do a round up of tcp_rwnd
553  * because the scale value may not be correct after that.
554  *
555  * Set the compiler flag to make this function inline.
556  */
557 void
558 tcp_set_ws_value(tcp_t *tcp)
559 {
560 	int i;
561 	uint32_t rwnd = tcp->tcp_rwnd;
562 
563 	for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT;
564 	    i++, rwnd >>= 1)
565 		;
566 	tcp->tcp_rcv_ws = i;
567 }
568 
569 /*
570  * Remove cached/latched IPsec references.
571  */
572 void
573 tcp_ipsec_cleanup(tcp_t *tcp)
574 {
575 	conn_t		*connp = tcp->tcp_connp;
576 
577 	ASSERT(connp->conn_flags & IPCL_TCPCONN);
578 
579 	if (connp->conn_latch != NULL) {
580 		IPLATCH_REFRELE(connp->conn_latch);
581 		connp->conn_latch = NULL;
582 	}
583 	if (connp->conn_latch_in_policy != NULL) {
584 		IPPOL_REFRELE(connp->conn_latch_in_policy);
585 		connp->conn_latch_in_policy = NULL;
586 	}
587 	if (connp->conn_latch_in_action != NULL) {
588 		IPACT_REFRELE(connp->conn_latch_in_action);
589 		connp->conn_latch_in_action = NULL;
590 	}
591 	if (connp->conn_policy != NULL) {
592 		IPPH_REFRELE(connp->conn_policy, connp->conn_netstack);
593 		connp->conn_policy = NULL;
594 	}
595 }
596 
597 /*
598  * Cleaup before placing on free list.
599  * Disassociate from the netstack/tcp_stack_t since the freelist
600  * is per squeue and not per netstack.
601  */
602 void
603 tcp_cleanup(tcp_t *tcp)
604 {
605 	mblk_t		*mp;
606 	tcp_sack_info_t	*tcp_sack_info;
607 	conn_t		*connp = tcp->tcp_connp;
608 	tcp_stack_t	*tcps = tcp->tcp_tcps;
609 	netstack_t	*ns = tcps->tcps_netstack;
610 	mblk_t		*tcp_rsrv_mp;
611 
612 	tcp_bind_hash_remove(tcp);
613 
614 	/* Cleanup that which needs the netstack first */
615 	tcp_ipsec_cleanup(tcp);
616 	ixa_cleanup(connp->conn_ixa);
617 
618 	if (connp->conn_ht_iphc != NULL) {
619 		kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
620 		connp->conn_ht_iphc = NULL;
621 		connp->conn_ht_iphc_allocated = 0;
622 		connp->conn_ht_iphc_len = 0;
623 		connp->conn_ht_ulp = NULL;
624 		connp->conn_ht_ulp_len = 0;
625 		tcp->tcp_ipha = NULL;
626 		tcp->tcp_ip6h = NULL;
627 		tcp->tcp_tcpha = NULL;
628 	}
629 
630 	/* We clear any IP_OPTIONS and extension headers */
631 	ip_pkt_free(&connp->conn_xmit_ipp);
632 
633 	tcp_free(tcp);
634 
635 	/* Release any SSL context */
636 	if (tcp->tcp_kssl_ent != NULL) {
637 		kssl_release_ent(tcp->tcp_kssl_ent, NULL, KSSL_NO_PROXY);
638 		tcp->tcp_kssl_ent = NULL;
639 	}
640 
641 	if (tcp->tcp_kssl_ctx != NULL) {
642 		kssl_release_ctx(tcp->tcp_kssl_ctx);
643 		tcp->tcp_kssl_ctx = NULL;
644 	}
645 	tcp->tcp_kssl_pending = B_FALSE;
646 
647 	/*
648 	 * Since we will bzero the entire structure, we need to
649 	 * remove it and reinsert it in global hash list. We
650 	 * know the walkers can't get to this conn because we
651 	 * had set CONDEMNED flag earlier and checked reference
652 	 * under conn_lock so walker won't pick it and when we
653 	 * go the ipcl_globalhash_remove() below, no walker
654 	 * can get to it.
655 	 */
656 	ipcl_globalhash_remove(connp);
657 
658 	/* Save some state */
659 	mp = tcp->tcp_timercache;
660 
661 	tcp_sack_info = tcp->tcp_sack_info;
662 	tcp_rsrv_mp = tcp->tcp_rsrv_mp;
663 
664 	if (connp->conn_cred != NULL) {
665 		crfree(connp->conn_cred);
666 		connp->conn_cred = NULL;
667 	}
668 	ipcl_conn_cleanup(connp);
669 	connp->conn_flags = IPCL_TCPCONN;
670 
671 	/*
672 	 * Now it is safe to decrement the reference counts.
673 	 * This might be the last reference on the netstack
674 	 * in which case it will cause the freeing of the IP Instance.
675 	 */
676 	connp->conn_netstack = NULL;
677 	connp->conn_ixa->ixa_ipst = NULL;
678 	netstack_rele(ns);
679 	ASSERT(tcps != NULL);
680 	tcp->tcp_tcps = NULL;
681 
682 	bzero(tcp, sizeof (tcp_t));
683 
684 	/* restore the state */
685 	tcp->tcp_timercache = mp;
686 
687 	tcp->tcp_sack_info = tcp_sack_info;
688 	tcp->tcp_rsrv_mp = tcp_rsrv_mp;
689 
690 	tcp->tcp_connp = connp;
691 
692 	ASSERT(connp->conn_tcp == tcp);
693 	ASSERT(connp->conn_flags & IPCL_TCPCONN);
694 	connp->conn_state_flags = CONN_INCIPIENT;
695 	ASSERT(connp->conn_proto == IPPROTO_TCP);
696 	ASSERT(connp->conn_ref == 1);
697 }
698 
699 /*
700  * Adapt to the information, such as rtt and rtt_sd, provided from the
701  * DCE and IRE maintained by IP.
702  *
703  * Checks for multicast and broadcast destination address.
704  * Returns zero if ok; an errno on failure.
705  *
706  * Note that the MSS calculation here is based on the info given in
707  * the DCE and IRE.  We do not do any calculation based on TCP options.  They
708  * will be handled in tcp_input_data() when TCP knows which options to use.
709  *
710  * Note on how TCP gets its parameters for a connection.
711  *
712  * When a tcp_t structure is allocated, it gets all the default parameters.
713  * In tcp_set_destination(), it gets those metric parameters, like rtt, rtt_sd,
714  * spipe, rpipe, ... from the route metrics.  Route metric overrides the
715  * default.
716  *
717  * An incoming SYN with a multicast or broadcast destination address is dropped
718  * in ip_fanout_v4/v6.
719  *
720  * An incoming SYN with a multicast or broadcast source address is always
721  * dropped in tcp_set_destination, since IPDF_ALLOW_MCBC is not set in
722  * conn_connect.
723  * The same logic in tcp_set_destination also serves to
724  * reject an attempt to connect to a broadcast or multicast (destination)
725  * address.
726  */
727 int
728 tcp_set_destination(tcp_t *tcp)
729 {
730 	uint32_t	mss_max;
731 	uint32_t	mss;
732 	boolean_t	tcp_detached = TCP_IS_DETACHED(tcp);
733 	conn_t		*connp = tcp->tcp_connp;
734 	tcp_stack_t	*tcps = tcp->tcp_tcps;
735 	iulp_t		uinfo;
736 	int		error;
737 	uint32_t	flags;
738 
739 	flags = IPDF_LSO | IPDF_ZCOPY;
740 	/*
741 	 * Make sure we have a dce for the destination to avoid dce_ident
742 	 * contention for connected sockets.
743 	 */
744 	flags |= IPDF_UNIQUE_DCE;
745 
746 	if (!tcps->tcps_ignore_path_mtu)
747 		connp->conn_ixa->ixa_flags |= IXAF_PMTU_DISCOVERY;
748 
749 	/* Use conn_lock to satify ASSERT; tcp is already serialized */
750 	mutex_enter(&connp->conn_lock);
751 	error = conn_connect(connp, &uinfo, flags);
752 	mutex_exit(&connp->conn_lock);
753 	if (error != 0)
754 		return (error);
755 
756 	error = tcp_build_hdrs(tcp);
757 	if (error != 0)
758 		return (error);
759 
760 	tcp->tcp_localnet = uinfo.iulp_localnet;
761 
762 	if (uinfo.iulp_rtt != 0) {
763 		clock_t	rto;
764 
765 		tcp->tcp_rtt_sa = uinfo.iulp_rtt;
766 		tcp->tcp_rtt_sd = uinfo.iulp_rtt_sd;
767 		rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
768 		    tcps->tcps_rexmit_interval_extra +
769 		    (tcp->tcp_rtt_sa >> 5);
770 
771 		if (rto > tcps->tcps_rexmit_interval_max) {
772 			tcp->tcp_rto = tcps->tcps_rexmit_interval_max;
773 		} else if (rto < tcps->tcps_rexmit_interval_min) {
774 			tcp->tcp_rto = tcps->tcps_rexmit_interval_min;
775 		} else {
776 			tcp->tcp_rto = rto;
777 		}
778 	}
779 	if (uinfo.iulp_ssthresh != 0)
780 		tcp->tcp_cwnd_ssthresh = uinfo.iulp_ssthresh;
781 	else
782 		tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
783 	if (uinfo.iulp_spipe > 0) {
784 		connp->conn_sndbuf = MIN(uinfo.iulp_spipe,
785 		    tcps->tcps_max_buf);
786 		if (tcps->tcps_snd_lowat_fraction != 0) {
787 			connp->conn_sndlowat = connp->conn_sndbuf /
788 			    tcps->tcps_snd_lowat_fraction;
789 		}
790 		(void) tcp_maxpsz_set(tcp, B_TRUE);
791 	}
792 	/*
793 	 * Note that up till now, acceptor always inherits receive
794 	 * window from the listener.  But if there is a metrics
795 	 * associated with a host, we should use that instead of
796 	 * inheriting it from listener. Thus we need to pass this
797 	 * info back to the caller.
798 	 */
799 	if (uinfo.iulp_rpipe > 0) {
800 		tcp->tcp_rwnd = MIN(uinfo.iulp_rpipe,
801 		    tcps->tcps_max_buf);
802 	}
803 
804 	if (uinfo.iulp_rtomax > 0) {
805 		tcp->tcp_second_timer_threshold =
806 		    uinfo.iulp_rtomax;
807 	}
808 
809 	/*
810 	 * Use the metric option settings, iulp_tstamp_ok and
811 	 * iulp_wscale_ok, only for active open. What this means
812 	 * is that if the other side uses timestamp or window
813 	 * scale option, TCP will also use those options. That
814 	 * is for passive open.  If the application sets a
815 	 * large window, window scale is enabled regardless of
816 	 * the value in iulp_wscale_ok.  This is the behavior
817 	 * since 2.6.  So we keep it.
818 	 * The only case left in passive open processing is the
819 	 * check for SACK.
820 	 * For ECN, it should probably be like SACK.  But the
821 	 * current value is binary, so we treat it like the other
822 	 * cases.  The metric only controls active open.For passive
823 	 * open, the ndd param, tcp_ecn_permitted, controls the
824 	 * behavior.
825 	 */
826 	if (!tcp_detached) {
827 		/*
828 		 * The if check means that the following can only
829 		 * be turned on by the metrics only IRE, but not off.
830 		 */
831 		if (uinfo.iulp_tstamp_ok)
832 			tcp->tcp_snd_ts_ok = B_TRUE;
833 		if (uinfo.iulp_wscale_ok)
834 			tcp->tcp_snd_ws_ok = B_TRUE;
835 		if (uinfo.iulp_sack == 2)
836 			tcp->tcp_snd_sack_ok = B_TRUE;
837 		if (uinfo.iulp_ecn_ok)
838 			tcp->tcp_ecn_ok = B_TRUE;
839 	} else {
840 		/*
841 		 * Passive open.
842 		 *
843 		 * As above, the if check means that SACK can only be
844 		 * turned on by the metric only IRE.
845 		 */
846 		if (uinfo.iulp_sack > 0) {
847 			tcp->tcp_snd_sack_ok = B_TRUE;
848 		}
849 	}
850 
851 	/*
852 	 * XXX Note that currently, iulp_mtu can be as small as 68
853 	 * because of PMTUd.  So tcp_mss may go to negative if combined
854 	 * length of all those options exceeds 28 bytes.  But because
855 	 * of the tcp_mss_min check below, we may not have a problem if
856 	 * tcp_mss_min is of a reasonable value.  The default is 1 so
857 	 * the negative problem still exists.  And the check defeats PMTUd.
858 	 * In fact, if PMTUd finds that the MSS should be smaller than
859 	 * tcp_mss_min, TCP should turn off PMUTd and use the tcp_mss_min
860 	 * value.
861 	 *
862 	 * We do not deal with that now.  All those problems related to
863 	 * PMTUd will be fixed later.
864 	 */
865 	ASSERT(uinfo.iulp_mtu != 0);
866 	mss = tcp->tcp_initial_pmtu = uinfo.iulp_mtu;
867 
868 	/* Sanity check for MSS value. */
869 	if (connp->conn_ipversion == IPV4_VERSION)
870 		mss_max = tcps->tcps_mss_max_ipv4;
871 	else
872 		mss_max = tcps->tcps_mss_max_ipv6;
873 
874 	if (tcp->tcp_ipsec_overhead == 0)
875 		tcp->tcp_ipsec_overhead = conn_ipsec_length(connp);
876 
877 	mss -= tcp->tcp_ipsec_overhead;
878 
879 	if (mss < tcps->tcps_mss_min)
880 		mss = tcps->tcps_mss_min;
881 	if (mss > mss_max)
882 		mss = mss_max;
883 
884 	/* Note that this is the maximum MSS, excluding all options. */
885 	tcp->tcp_mss = mss;
886 
887 	/*
888 	 * Update the tcp connection with LSO capability.
889 	 */
890 	tcp_update_lso(tcp, connp->conn_ixa);
891 
892 	/*
893 	 * Initialize the ISS here now that we have the full connection ID.
894 	 * The RFC 1948 method of initial sequence number generation requires
895 	 * knowledge of the full connection ID before setting the ISS.
896 	 */
897 	tcp_iss_init(tcp);
898 
899 	tcp->tcp_loopback = (uinfo.iulp_loopback | uinfo.iulp_local);
900 
901 	/*
902 	 * Make sure that conn is not marked incipient
903 	 * for incoming connections. A blind
904 	 * removal of incipient flag is cheaper than
905 	 * check and removal.
906 	 */
907 	mutex_enter(&connp->conn_lock);
908 	connp->conn_state_flags &= ~CONN_INCIPIENT;
909 	mutex_exit(&connp->conn_lock);
910 	return (0);
911 }
912 
913 /*
914  * tcp_clean_death / tcp_close_detached must not be called more than once
915  * on a tcp. Thus every function that potentially calls tcp_clean_death
916  * must check for the tcp state before calling tcp_clean_death.
917  * Eg. tcp_input_data, tcp_eager_kill, tcp_clean_death_wrapper,
918  * tcp_timer_handler, all check for the tcp state.
919  */
920 /* ARGSUSED */
921 void
922 tcp_clean_death_wrapper(void *arg, mblk_t *mp, void *arg2,
923     ip_recv_attr_t *dummy)
924 {
925 	tcp_t	*tcp = ((conn_t *)arg)->conn_tcp;
926 
927 	freemsg(mp);
928 	if (tcp->tcp_state > TCPS_BOUND)
929 		(void) tcp_clean_death(((conn_t *)arg)->conn_tcp, ETIMEDOUT);
930 }
931 
932 /*
933  * We are dying for some reason.  Try to do it gracefully.  (May be called
934  * as writer.)
935  *
936  * Return -1 if the structure was not cleaned up (if the cleanup had to be
937  * done by a service procedure).
938  * TBD - Should the return value distinguish between the tcp_t being
939  * freed and it being reinitialized?
940  */
941 int
942 tcp_clean_death(tcp_t *tcp, int err)
943 {
944 	mblk_t	*mp;
945 	queue_t	*q;
946 	conn_t	*connp = tcp->tcp_connp;
947 	tcp_stack_t	*tcps = tcp->tcp_tcps;
948 
949 	if (tcp->tcp_fused)
950 		tcp_unfuse(tcp);
951 
952 	if (tcp->tcp_linger_tid != 0 &&
953 	    TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
954 		tcp_stop_lingering(tcp);
955 	}
956 
957 	ASSERT(tcp != NULL);
958 	ASSERT((connp->conn_family == AF_INET &&
959 	    connp->conn_ipversion == IPV4_VERSION) ||
960 	    (connp->conn_family == AF_INET6 &&
961 	    (connp->conn_ipversion == IPV4_VERSION ||
962 	    connp->conn_ipversion == IPV6_VERSION)));
963 
964 	if (TCP_IS_DETACHED(tcp)) {
965 		if (tcp->tcp_hard_binding) {
966 			/*
967 			 * Its an eager that we are dealing with. We close the
968 			 * eager but in case a conn_ind has already gone to the
969 			 * listener, let tcp_accept_finish() send a discon_ind
970 			 * to the listener and drop the last reference. If the
971 			 * listener doesn't even know about the eager i.e. the
972 			 * conn_ind hasn't gone up, blow away the eager and drop
973 			 * the last reference as well. If the conn_ind has gone
974 			 * up, state should be BOUND. tcp_accept_finish
975 			 * will figure out that the connection has received a
976 			 * RST and will send a DISCON_IND to the application.
977 			 */
978 			tcp_closei_local(tcp);
979 			if (!tcp->tcp_tconnind_started) {
980 				CONN_DEC_REF(connp);
981 			} else {
982 				tcp->tcp_state = TCPS_BOUND;
983 			}
984 		} else {
985 			tcp_close_detached(tcp);
986 		}
987 		return (0);
988 	}
989 
990 	TCP_STAT(tcps, tcp_clean_death_nondetached);
991 
992 	/*
993 	 * The connection is dead.  Decrement listener connection counter if
994 	 * necessary.
995 	 */
996 	if (tcp->tcp_listen_cnt != NULL)
997 		TCP_DECR_LISTEN_CNT(tcp);
998 
999 	/*
1000 	 * When a connection is moved to TIME_WAIT state, the connection
1001 	 * counter is already decremented.  So no need to decrement here
1002 	 * again.  See SET_TIME_WAIT() macro.
1003 	 */
1004 	if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1005 	    tcp->tcp_state < TCPS_TIME_WAIT) {
1006 		TCPS_CONN_DEC(tcps);
1007 	}
1008 
1009 	q = connp->conn_rq;
1010 
1011 	/* Trash all inbound data */
1012 	if (!IPCL_IS_NONSTR(connp)) {
1013 		ASSERT(q != NULL);
1014 		flushq(q, FLUSHALL);
1015 	}
1016 
1017 	/*
1018 	 * If we are at least part way open and there is error
1019 	 * (err==0 implies no error)
1020 	 * notify our client by a T_DISCON_IND.
1021 	 */
1022 	if ((tcp->tcp_state >= TCPS_SYN_SENT) && err) {
1023 		if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1024 		    !TCP_IS_SOCKET(tcp)) {
1025 			/*
1026 			 * Send M_FLUSH according to TPI. Because sockets will
1027 			 * (and must) ignore FLUSHR we do that only for TPI
1028 			 * endpoints and sockets in STREAMS mode.
1029 			 */
1030 			(void) putnextctl1(q, M_FLUSH, FLUSHR);
1031 		}
1032 		if (connp->conn_debug) {
1033 			(void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE|SL_ERROR,
1034 			    "tcp_clean_death: discon err %d", err);
1035 		}
1036 		if (IPCL_IS_NONSTR(connp)) {
1037 			/* Direct socket, use upcall */
1038 			(*connp->conn_upcalls->su_disconnected)(
1039 			    connp->conn_upper_handle, tcp->tcp_connid, err);
1040 		} else {
1041 			mp = mi_tpi_discon_ind(NULL, err, 0);
1042 			if (mp != NULL) {
1043 				putnext(q, mp);
1044 			} else {
1045 				if (connp->conn_debug) {
1046 					(void) strlog(TCP_MOD_ID, 0, 1,
1047 					    SL_ERROR|SL_TRACE,
1048 					    "tcp_clean_death, sending M_ERROR");
1049 				}
1050 				(void) putnextctl1(q, M_ERROR, EPROTO);
1051 			}
1052 		}
1053 		if (tcp->tcp_state <= TCPS_SYN_RCVD) {
1054 			/* SYN_SENT or SYN_RCVD */
1055 			TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1056 		} else if (tcp->tcp_state <= TCPS_CLOSE_WAIT) {
1057 			/* ESTABLISHED or CLOSE_WAIT */
1058 			TCPS_BUMP_MIB(tcps, tcpEstabResets);
1059 		}
1060 	}
1061 
1062 	tcp_reinit(tcp);
1063 	if (IPCL_IS_NONSTR(connp))
1064 		(void) tcp_do_unbind(connp);
1065 
1066 	return (-1);
1067 }
1068 
1069 /*
1070  * In case tcp is in the "lingering state" and waits for the SO_LINGER timeout
1071  * to expire, stop the wait and finish the close.
1072  */
1073 void
1074 tcp_stop_lingering(tcp_t *tcp)
1075 {
1076 	clock_t	delta = 0;
1077 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1078 	conn_t		*connp = tcp->tcp_connp;
1079 
1080 	tcp->tcp_linger_tid = 0;
1081 	if (tcp->tcp_state > TCPS_LISTEN) {
1082 		tcp_acceptor_hash_remove(tcp);
1083 		mutex_enter(&tcp->tcp_non_sq_lock);
1084 		if (tcp->tcp_flow_stopped) {
1085 			tcp_clrqfull(tcp);
1086 		}
1087 		mutex_exit(&tcp->tcp_non_sq_lock);
1088 
1089 		if (tcp->tcp_timer_tid != 0) {
1090 			delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
1091 			tcp->tcp_timer_tid = 0;
1092 		}
1093 		/*
1094 		 * Need to cancel those timers which will not be used when
1095 		 * TCP is detached.  This has to be done before the conn_wq
1096 		 * is cleared.
1097 		 */
1098 		tcp_timers_stop(tcp);
1099 
1100 		tcp->tcp_detached = B_TRUE;
1101 		connp->conn_rq = NULL;
1102 		connp->conn_wq = NULL;
1103 
1104 		if (tcp->tcp_state == TCPS_TIME_WAIT) {
1105 			tcp_time_wait_append(tcp);
1106 			TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1107 			goto finish;
1108 		}
1109 
1110 		/*
1111 		 * If delta is zero the timer event wasn't executed and was
1112 		 * successfully canceled. In this case we need to restart it
1113 		 * with the minimal delta possible.
1114 		 */
1115 		if (delta >= 0) {
1116 			tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1117 			    delta ? delta : 1);
1118 		}
1119 	} else {
1120 		tcp_closei_local(tcp);
1121 		CONN_DEC_REF(connp);
1122 	}
1123 finish:
1124 	/* Signal closing thread that it can complete close */
1125 	mutex_enter(&tcp->tcp_closelock);
1126 	tcp->tcp_detached = B_TRUE;
1127 	connp->conn_rq = NULL;
1128 	connp->conn_wq = NULL;
1129 
1130 	tcp->tcp_closed = 1;
1131 	cv_signal(&tcp->tcp_closecv);
1132 	mutex_exit(&tcp->tcp_closelock);
1133 }
1134 
1135 void
1136 tcp_close_common(conn_t *connp, int flags)
1137 {
1138 	tcp_t		*tcp = connp->conn_tcp;
1139 	mblk_t 		*mp = &tcp->tcp_closemp;
1140 	boolean_t	conn_ioctl_cleanup_reqd = B_FALSE;
1141 	mblk_t		*bp;
1142 
1143 	ASSERT(connp->conn_ref >= 2);
1144 
1145 	/*
1146 	 * Mark the conn as closing. ipsq_pending_mp_add will not
1147 	 * add any mp to the pending mp list, after this conn has
1148 	 * started closing.
1149 	 */
1150 	mutex_enter(&connp->conn_lock);
1151 	connp->conn_state_flags |= CONN_CLOSING;
1152 	if (connp->conn_oper_pending_ill != NULL)
1153 		conn_ioctl_cleanup_reqd = B_TRUE;
1154 	CONN_INC_REF_LOCKED(connp);
1155 	mutex_exit(&connp->conn_lock);
1156 	tcp->tcp_closeflags = (uint8_t)flags;
1157 	ASSERT(connp->conn_ref >= 3);
1158 
1159 	/*
1160 	 * tcp_closemp_used is used below without any protection of a lock
1161 	 * as we don't expect any one else to use it concurrently at this
1162 	 * point otherwise it would be a major defect.
1163 	 */
1164 
1165 	if (mp->b_prev == NULL)
1166 		tcp->tcp_closemp_used = B_TRUE;
1167 	else
1168 		cmn_err(CE_PANIC, "tcp_close: concurrent use of tcp_closemp: "
1169 		    "connp %p tcp %p\n", (void *)connp, (void *)tcp);
1170 
1171 	TCP_DEBUG_GETPCSTACK(tcp->tcmp_stk, 15);
1172 
1173 	/*
1174 	 * Cleanup any queued ioctls here. This must be done before the wq/rq
1175 	 * are re-written by tcp_close_output().
1176 	 */
1177 	if (conn_ioctl_cleanup_reqd)
1178 		conn_ioctl_cleanup(connp);
1179 
1180 	/*
1181 	 * As CONN_CLOSING is set, no further ioctls should be passed down to
1182 	 * IP for this conn (see the guards in tcp_ioctl, tcp_wput_ioctl and
1183 	 * tcp_wput_iocdata). If the ioctl was queued on an ipsq,
1184 	 * conn_ioctl_cleanup should have found it and removed it. If the ioctl
1185 	 * was still in flight at the time, we wait for it here. See comments
1186 	 * for CONN_INC_IOCTLREF in ip.h for details.
1187 	 */
1188 	mutex_enter(&connp->conn_lock);
1189 	while (connp->conn_ioctlref > 0)
1190 		cv_wait(&connp->conn_cv, &connp->conn_lock);
1191 	ASSERT(connp->conn_ioctlref == 0);
1192 	ASSERT(connp->conn_oper_pending_ill == NULL);
1193 	mutex_exit(&connp->conn_lock);
1194 
1195 	SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_close_output, connp,
1196 	    NULL, tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1197 
1198 	mutex_enter(&tcp->tcp_closelock);
1199 	while (!tcp->tcp_closed) {
1200 		if (!cv_wait_sig(&tcp->tcp_closecv, &tcp->tcp_closelock)) {
1201 			/*
1202 			 * The cv_wait_sig() was interrupted. We now do the
1203 			 * following:
1204 			 *
1205 			 * 1) If the endpoint was lingering, we allow this
1206 			 * to be interrupted by cancelling the linger timeout
1207 			 * and closing normally.
1208 			 *
1209 			 * 2) Revert to calling cv_wait()
1210 			 *
1211 			 * We revert to using cv_wait() to avoid an
1212 			 * infinite loop which can occur if the calling
1213 			 * thread is higher priority than the squeue worker
1214 			 * thread and is bound to the same cpu.
1215 			 */
1216 			if (connp->conn_linger && connp->conn_lingertime > 0) {
1217 				mutex_exit(&tcp->tcp_closelock);
1218 				/* Entering squeue, bump ref count. */
1219 				CONN_INC_REF(connp);
1220 				bp = allocb_wait(0, BPRI_HI, STR_NOSIG, NULL);
1221 				SQUEUE_ENTER_ONE(connp->conn_sqp, bp,
1222 				    tcp_linger_interrupted, connp, NULL,
1223 				    tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1224 				mutex_enter(&tcp->tcp_closelock);
1225 			}
1226 			break;
1227 		}
1228 	}
1229 	while (!tcp->tcp_closed)
1230 		cv_wait(&tcp->tcp_closecv, &tcp->tcp_closelock);
1231 	mutex_exit(&tcp->tcp_closelock);
1232 
1233 	/*
1234 	 * In the case of listener streams that have eagers in the q or q0
1235 	 * we wait for the eagers to drop their reference to us. conn_rq and
1236 	 * conn_wq of the eagers point to our queues. By waiting for the
1237 	 * refcnt to drop to 1, we are sure that the eagers have cleaned
1238 	 * up their queue pointers and also dropped their references to us.
1239 	 */
1240 	if (tcp->tcp_wait_for_eagers) {
1241 		mutex_enter(&connp->conn_lock);
1242 		while (connp->conn_ref != 1) {
1243 			cv_wait(&connp->conn_cv, &connp->conn_lock);
1244 		}
1245 		mutex_exit(&connp->conn_lock);
1246 	}
1247 
1248 	connp->conn_cpid = NOPID;
1249 }
1250 
1251 /*
1252  * Called by tcp_close() routine via squeue when lingering is
1253  * interrupted by a signal.
1254  */
1255 
1256 /* ARGSUSED */
1257 static void
1258 tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1259 {
1260 	conn_t	*connp = (conn_t *)arg;
1261 	tcp_t	*tcp = connp->conn_tcp;
1262 
1263 	freeb(mp);
1264 	if (tcp->tcp_linger_tid != 0 &&
1265 	    TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
1266 		tcp_stop_lingering(tcp);
1267 		tcp->tcp_client_errno = EINTR;
1268 	}
1269 }
1270 
1271 /*
1272  * Clean up the b_next and b_prev fields of every mblk pointed at by *mpp.
1273  * Some stream heads get upset if they see these later on as anything but NULL.
1274  */
1275 void
1276 tcp_close_mpp(mblk_t **mpp)
1277 {
1278 	mblk_t	*mp;
1279 
1280 	if ((mp = *mpp) != NULL) {
1281 		do {
1282 			mp->b_next = NULL;
1283 			mp->b_prev = NULL;
1284 		} while ((mp = mp->b_cont) != NULL);
1285 
1286 		mp = *mpp;
1287 		*mpp = NULL;
1288 		freemsg(mp);
1289 	}
1290 }
1291 
1292 /* Do detached close. */
1293 void
1294 tcp_close_detached(tcp_t *tcp)
1295 {
1296 	if (tcp->tcp_fused)
1297 		tcp_unfuse(tcp);
1298 
1299 	/*
1300 	 * Clustering code serializes TCP disconnect callbacks and
1301 	 * cluster tcp list walks by blocking a TCP disconnect callback
1302 	 * if a cluster tcp list walk is in progress. This ensures
1303 	 * accurate accounting of TCPs in the cluster code even though
1304 	 * the TCP list walk itself is not atomic.
1305 	 */
1306 	tcp_closei_local(tcp);
1307 	CONN_DEC_REF(tcp->tcp_connp);
1308 }
1309 
1310 /*
1311  * The tcp_t is going away. Remove it from all lists and set it
1312  * to TCPS_CLOSED. The freeing up of memory is deferred until
1313  * tcp_inactive. This is needed since a thread in tcp_rput might have
1314  * done a CONN_INC_REF on this structure before it was removed from the
1315  * hashes.
1316  */
1317 void
1318 tcp_closei_local(tcp_t *tcp)
1319 {
1320 	conn_t		*connp = tcp->tcp_connp;
1321 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1322 
1323 	if (!TCP_IS_SOCKET(tcp))
1324 		tcp_acceptor_hash_remove(tcp);
1325 
1326 	TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1327 	tcp->tcp_ibsegs = 0;
1328 	TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1329 	tcp->tcp_obsegs = 0;
1330 
1331 	/*
1332 	 * This can be called via tcp_time_wait_processing() if TCP gets a
1333 	 * SYN with sequence number outside the TIME-WAIT connection's
1334 	 * window.  So we need to check for TIME-WAIT state here as the
1335 	 * connection counter is already decremented.  See SET_TIME_WAIT()
1336 	 * macro
1337 	 */
1338 	if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1339 	    tcp->tcp_state < TCPS_TIME_WAIT) {
1340 		TCPS_CONN_DEC(tcps);
1341 	}
1342 
1343 	/*
1344 	 * If we are an eager connection hanging off a listener that
1345 	 * hasn't formally accepted the connection yet, get off his
1346 	 * list and blow off any data that we have accumulated.
1347 	 */
1348 	if (tcp->tcp_listener != NULL) {
1349 		tcp_t	*listener = tcp->tcp_listener;
1350 		mutex_enter(&listener->tcp_eager_lock);
1351 		/*
1352 		 * tcp_tconnind_started == B_TRUE means that the
1353 		 * conn_ind has already gone to listener. At
1354 		 * this point, eager will be closed but we
1355 		 * leave it in listeners eager list so that
1356 		 * if listener decides to close without doing
1357 		 * accept, we can clean this up. In tcp_tli_accept
1358 		 * we take care of the case of accept on closed
1359 		 * eager.
1360 		 */
1361 		if (!tcp->tcp_tconnind_started) {
1362 			tcp_eager_unlink(tcp);
1363 			mutex_exit(&listener->tcp_eager_lock);
1364 			/*
1365 			 * We don't want to have any pointers to the
1366 			 * listener queue, after we have released our
1367 			 * reference on the listener
1368 			 */
1369 			ASSERT(tcp->tcp_detached);
1370 			connp->conn_rq = NULL;
1371 			connp->conn_wq = NULL;
1372 			CONN_DEC_REF(listener->tcp_connp);
1373 		} else {
1374 			mutex_exit(&listener->tcp_eager_lock);
1375 		}
1376 	}
1377 
1378 	/* Stop all the timers */
1379 	tcp_timers_stop(tcp);
1380 
1381 	if (tcp->tcp_state == TCPS_LISTEN) {
1382 		if (tcp->tcp_ip_addr_cache) {
1383 			kmem_free((void *)tcp->tcp_ip_addr_cache,
1384 			    IP_ADDR_CACHE_SIZE * sizeof (ipaddr_t));
1385 			tcp->tcp_ip_addr_cache = NULL;
1386 		}
1387 	}
1388 
1389 	/* Decrement listerner connection counter if necessary. */
1390 	if (tcp->tcp_listen_cnt != NULL)
1391 		TCP_DECR_LISTEN_CNT(tcp);
1392 
1393 	mutex_enter(&tcp->tcp_non_sq_lock);
1394 	if (tcp->tcp_flow_stopped)
1395 		tcp_clrqfull(tcp);
1396 	mutex_exit(&tcp->tcp_non_sq_lock);
1397 
1398 	tcp_bind_hash_remove(tcp);
1399 	/*
1400 	 * If the tcp_time_wait_collector (which runs outside the squeue)
1401 	 * is trying to remove this tcp from the time wait list, we will
1402 	 * block in tcp_time_wait_remove while trying to acquire the
1403 	 * tcp_time_wait_lock. The logic in tcp_time_wait_collector also
1404 	 * requires the ipcl_hash_remove to be ordered after the
1405 	 * tcp_time_wait_remove for the refcnt checks to work correctly.
1406 	 */
1407 	if (tcp->tcp_state == TCPS_TIME_WAIT)
1408 		(void) tcp_time_wait_remove(tcp, NULL);
1409 	CL_INET_DISCONNECT(connp);
1410 	ipcl_hash_remove(connp);
1411 	ixa_cleanup(connp->conn_ixa);
1412 
1413 	/*
1414 	 * Mark the conn as CONDEMNED
1415 	 */
1416 	mutex_enter(&connp->conn_lock);
1417 	connp->conn_state_flags |= CONN_CONDEMNED;
1418 	mutex_exit(&connp->conn_lock);
1419 
1420 	ASSERT(tcp->tcp_time_wait_next == NULL);
1421 	ASSERT(tcp->tcp_time_wait_prev == NULL);
1422 	ASSERT(tcp->tcp_time_wait_expire == 0);
1423 	tcp->tcp_state = TCPS_CLOSED;
1424 
1425 	/* Release any SSL context */
1426 	if (tcp->tcp_kssl_ent != NULL) {
1427 		kssl_release_ent(tcp->tcp_kssl_ent, NULL, KSSL_NO_PROXY);
1428 		tcp->tcp_kssl_ent = NULL;
1429 	}
1430 	if (tcp->tcp_kssl_ctx != NULL) {
1431 		kssl_release_ctx(tcp->tcp_kssl_ctx);
1432 		tcp->tcp_kssl_ctx = NULL;
1433 	}
1434 	tcp->tcp_kssl_pending = B_FALSE;
1435 
1436 	tcp_ipsec_cleanup(tcp);
1437 }
1438 
1439 /*
1440  * tcp is dying (called from ipcl_conn_destroy and error cases).
1441  * Free the tcp_t in either case.
1442  */
1443 void
1444 tcp_free(tcp_t *tcp)
1445 {
1446 	mblk_t		*mp;
1447 	conn_t		*connp = tcp->tcp_connp;
1448 
1449 	ASSERT(tcp != NULL);
1450 	ASSERT(tcp->tcp_ptpahn == NULL && tcp->tcp_acceptor_hash == NULL);
1451 
1452 	connp->conn_rq = NULL;
1453 	connp->conn_wq = NULL;
1454 
1455 	tcp_close_mpp(&tcp->tcp_xmit_head);
1456 	tcp_close_mpp(&tcp->tcp_reass_head);
1457 	if (tcp->tcp_rcv_list != NULL) {
1458 		/* Free b_next chain */
1459 		tcp_close_mpp(&tcp->tcp_rcv_list);
1460 	}
1461 	if ((mp = tcp->tcp_urp_mp) != NULL) {
1462 		freemsg(mp);
1463 	}
1464 	if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1465 		freemsg(mp);
1466 	}
1467 
1468 	if (tcp->tcp_fused_sigurg_mp != NULL) {
1469 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1470 		freeb(tcp->tcp_fused_sigurg_mp);
1471 		tcp->tcp_fused_sigurg_mp = NULL;
1472 	}
1473 
1474 	if (tcp->tcp_ordrel_mp != NULL) {
1475 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1476 		freeb(tcp->tcp_ordrel_mp);
1477 		tcp->tcp_ordrel_mp = NULL;
1478 	}
1479 
1480 	if (tcp->tcp_sack_info != NULL) {
1481 		if (tcp->tcp_notsack_list != NULL) {
1482 			TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list,
1483 			    tcp);
1484 		}
1485 		bzero(tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
1486 	}
1487 
1488 	if (tcp->tcp_hopopts != NULL) {
1489 		mi_free(tcp->tcp_hopopts);
1490 		tcp->tcp_hopopts = NULL;
1491 		tcp->tcp_hopoptslen = 0;
1492 	}
1493 	ASSERT(tcp->tcp_hopoptslen == 0);
1494 	if (tcp->tcp_dstopts != NULL) {
1495 		mi_free(tcp->tcp_dstopts);
1496 		tcp->tcp_dstopts = NULL;
1497 		tcp->tcp_dstoptslen = 0;
1498 	}
1499 	ASSERT(tcp->tcp_dstoptslen == 0);
1500 	if (tcp->tcp_rthdrdstopts != NULL) {
1501 		mi_free(tcp->tcp_rthdrdstopts);
1502 		tcp->tcp_rthdrdstopts = NULL;
1503 		tcp->tcp_rthdrdstoptslen = 0;
1504 	}
1505 	ASSERT(tcp->tcp_rthdrdstoptslen == 0);
1506 	if (tcp->tcp_rthdr != NULL) {
1507 		mi_free(tcp->tcp_rthdr);
1508 		tcp->tcp_rthdr = NULL;
1509 		tcp->tcp_rthdrlen = 0;
1510 	}
1511 	ASSERT(tcp->tcp_rthdrlen == 0);
1512 
1513 	/*
1514 	 * Following is really a blowing away a union.
1515 	 * It happens to have exactly two members of identical size
1516 	 * the following code is enough.
1517 	 */
1518 	tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1519 }
1520 
1521 /*
1522  * tcp_get_conn/tcp_free_conn
1523  *
1524  * tcp_get_conn is used to get a clean tcp connection structure.
1525  * It tries to reuse the connections put on the freelist by the
1526  * time_wait_collector failing which it goes to kmem_cache. This
1527  * way has two benefits compared to just allocating from and
1528  * freeing to kmem_cache.
1529  * 1) The time_wait_collector can free (which includes the cleanup)
1530  * outside the squeue. So when the interrupt comes, we have a clean
1531  * connection sitting in the freelist. Obviously, this buys us
1532  * performance.
1533  *
1534  * 2) Defence against DOS attack. Allocating a tcp/conn in tcp_input_listener
1535  * has multiple disadvantages - tying up the squeue during alloc.
1536  * But allocating the conn/tcp in IP land is also not the best since
1537  * we can't check the 'q' and 'q0' which are protected by squeue and
1538  * blindly allocate memory which might have to be freed here if we are
1539  * not allowed to accept the connection. By using the freelist and
1540  * putting the conn/tcp back in freelist, we don't pay a penalty for
1541  * allocating memory without checking 'q/q0' and freeing it if we can't
1542  * accept the connection.
1543  *
1544  * Care should be taken to put the conn back in the same squeue's freelist
1545  * from which it was allocated. Best results are obtained if conn is
1546  * allocated from listener's squeue and freed to the same. Time wait
1547  * collector will free up the freelist is the connection ends up sitting
1548  * there for too long.
1549  */
1550 void *
1551 tcp_get_conn(void *arg, tcp_stack_t *tcps)
1552 {
1553 	tcp_t			*tcp = NULL;
1554 	conn_t			*connp = NULL;
1555 	squeue_t		*sqp = (squeue_t *)arg;
1556 	tcp_squeue_priv_t 	*tcp_time_wait;
1557 	netstack_t		*ns;
1558 	mblk_t			*tcp_rsrv_mp = NULL;
1559 
1560 	tcp_time_wait =
1561 	    *((tcp_squeue_priv_t **)squeue_getprivate(sqp, SQPRIVATE_TCP));
1562 
1563 	mutex_enter(&tcp_time_wait->tcp_time_wait_lock);
1564 	tcp = tcp_time_wait->tcp_free_list;
1565 	ASSERT((tcp != NULL) ^ (tcp_time_wait->tcp_free_list_cnt == 0));
1566 	if (tcp != NULL) {
1567 		tcp_time_wait->tcp_free_list = tcp->tcp_time_wait_next;
1568 		tcp_time_wait->tcp_free_list_cnt--;
1569 		mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1570 		tcp->tcp_time_wait_next = NULL;
1571 		connp = tcp->tcp_connp;
1572 		connp->conn_flags |= IPCL_REUSED;
1573 
1574 		ASSERT(tcp->tcp_tcps == NULL);
1575 		ASSERT(connp->conn_netstack == NULL);
1576 		ASSERT(tcp->tcp_rsrv_mp != NULL);
1577 		ns = tcps->tcps_netstack;
1578 		netstack_hold(ns);
1579 		connp->conn_netstack = ns;
1580 		connp->conn_ixa->ixa_ipst = ns->netstack_ip;
1581 		tcp->tcp_tcps = tcps;
1582 		ipcl_globalhash_insert(connp);
1583 
1584 		connp->conn_ixa->ixa_notify_cookie = tcp;
1585 		ASSERT(connp->conn_ixa->ixa_notify == tcp_notify);
1586 		connp->conn_recv = tcp_input_data;
1587 		ASSERT(connp->conn_recvicmp == tcp_icmp_input);
1588 		ASSERT(connp->conn_verifyicmp == tcp_verifyicmp);
1589 		return ((void *)connp);
1590 	}
1591 	mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1592 	/*
1593 	 * Pre-allocate the tcp_rsrv_mp. This mblk will not be freed until
1594 	 * this conn_t/tcp_t is freed at ipcl_conn_destroy().
1595 	 */
1596 	tcp_rsrv_mp = allocb(0, BPRI_HI);
1597 	if (tcp_rsrv_mp == NULL)
1598 		return (NULL);
1599 
1600 	if ((connp = ipcl_conn_create(IPCL_TCPCONN, KM_NOSLEEP,
1601 	    tcps->tcps_netstack)) == NULL) {
1602 		freeb(tcp_rsrv_mp);
1603 		return (NULL);
1604 	}
1605 
1606 	tcp = connp->conn_tcp;
1607 	tcp->tcp_rsrv_mp = tcp_rsrv_mp;
1608 	mutex_init(&tcp->tcp_rsrv_mp_lock, NULL, MUTEX_DEFAULT, NULL);
1609 
1610 	tcp->tcp_tcps = tcps;
1611 
1612 	connp->conn_recv = tcp_input_data;
1613 	connp->conn_recvicmp = tcp_icmp_input;
1614 	connp->conn_verifyicmp = tcp_verifyicmp;
1615 
1616 	/*
1617 	 * Register tcp_notify to listen to capability changes detected by IP.
1618 	 * This upcall is made in the context of the call to conn_ip_output
1619 	 * thus it is inside the squeue.
1620 	 */
1621 	connp->conn_ixa->ixa_notify = tcp_notify;
1622 	connp->conn_ixa->ixa_notify_cookie = tcp;
1623 
1624 	return ((void *)connp);
1625 }
1626 
1627 /*
1628  * Handle connect to IPv4 destinations, including connections for AF_INET6
1629  * sockets connecting to IPv4 mapped IPv6 destinations.
1630  * Returns zero if OK, a positive errno, or a negative TLI error.
1631  */
1632 static int
1633 tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp, in_port_t dstport,
1634     uint_t srcid)
1635 {
1636 	ipaddr_t 	dstaddr = *dstaddrp;
1637 	uint16_t 	lport;
1638 	conn_t		*connp = tcp->tcp_connp;
1639 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1640 	int		error;
1641 
1642 	ASSERT(connp->conn_ipversion == IPV4_VERSION);
1643 
1644 	/* Check for attempt to connect to INADDR_ANY */
1645 	if (dstaddr == INADDR_ANY)  {
1646 		/*
1647 		 * SunOS 4.x and 4.3 BSD allow an application
1648 		 * to connect a TCP socket to INADDR_ANY.
1649 		 * When they do this, the kernel picks the
1650 		 * address of one interface and uses it
1651 		 * instead.  The kernel usually ends up
1652 		 * picking the address of the loopback
1653 		 * interface.  This is an undocumented feature.
1654 		 * However, we provide the same thing here
1655 		 * in order to have source and binary
1656 		 * compatibility with SunOS 4.x.
1657 		 * Update the T_CONN_REQ (sin/sin6) since it is used to
1658 		 * generate the T_CONN_CON.
1659 		 */
1660 		dstaddr = htonl(INADDR_LOOPBACK);
1661 		*dstaddrp = dstaddr;
1662 	}
1663 
1664 	/* Handle __sin6_src_id if socket not bound to an IP address */
1665 	if (srcid != 0 && connp->conn_laddr_v4 == INADDR_ANY) {
1666 		ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1667 		    IPCL_ZONEID(connp), tcps->tcps_netstack);
1668 		connp->conn_saddr_v6 = connp->conn_laddr_v6;
1669 	}
1670 
1671 	IN6_IPADDR_TO_V4MAPPED(dstaddr, &connp->conn_faddr_v6);
1672 	connp->conn_fport = dstport;
1673 
1674 	/*
1675 	 * At this point the remote destination address and remote port fields
1676 	 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1677 	 * have to see which state tcp was in so we can take appropriate action.
1678 	 */
1679 	if (tcp->tcp_state == TCPS_IDLE) {
1680 		/*
1681 		 * We support a quick connect capability here, allowing
1682 		 * clients to transition directly from IDLE to SYN_SENT
1683 		 * tcp_bindi will pick an unused port, insert the connection
1684 		 * in the bind hash and transition to BOUND state.
1685 		 */
1686 		lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1687 		    tcp, B_TRUE);
1688 		lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1689 		    B_FALSE, B_FALSE);
1690 		if (lport == 0)
1691 			return (-TNOADDR);
1692 	}
1693 
1694 	/*
1695 	 * Lookup the route to determine a source address and the uinfo.
1696 	 * Setup TCP parameters based on the metrics/DCE.
1697 	 */
1698 	error = tcp_set_destination(tcp);
1699 	if (error != 0)
1700 		return (error);
1701 
1702 	/*
1703 	 * Don't let an endpoint connect to itself.
1704 	 */
1705 	if (connp->conn_faddr_v4 == connp->conn_laddr_v4 &&
1706 	    connp->conn_fport == connp->conn_lport)
1707 		return (-TBADADDR);
1708 
1709 	tcp->tcp_state = TCPS_SYN_SENT;
1710 
1711 	return (ipcl_conn_insert_v4(connp));
1712 }
1713 
1714 /*
1715  * Handle connect to IPv6 destinations.
1716  * Returns zero if OK, a positive errno, or a negative TLI error.
1717  */
1718 static int
1719 tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp, in_port_t dstport,
1720     uint32_t flowinfo, uint_t srcid, uint32_t scope_id)
1721 {
1722 	uint16_t 	lport;
1723 	conn_t		*connp = tcp->tcp_connp;
1724 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1725 	int		error;
1726 
1727 	ASSERT(connp->conn_family == AF_INET6);
1728 
1729 	/*
1730 	 * If we're here, it means that the destination address is a native
1731 	 * IPv6 address.  Return an error if conn_ipversion is not IPv6.  A
1732 	 * reason why it might not be IPv6 is if the socket was bound to an
1733 	 * IPv4-mapped IPv6 address.
1734 	 */
1735 	if (connp->conn_ipversion != IPV6_VERSION)
1736 		return (-TBADADDR);
1737 
1738 	/*
1739 	 * Interpret a zero destination to mean loopback.
1740 	 * Update the T_CONN_REQ (sin/sin6) since it is used to
1741 	 * generate the T_CONN_CON.
1742 	 */
1743 	if (IN6_IS_ADDR_UNSPECIFIED(dstaddrp))
1744 		*dstaddrp = ipv6_loopback;
1745 
1746 	/* Handle __sin6_src_id if socket not bound to an IP address */
1747 	if (srcid != 0 && IN6_IS_ADDR_UNSPECIFIED(&connp->conn_laddr_v6)) {
1748 		ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1749 		    IPCL_ZONEID(connp), tcps->tcps_netstack);
1750 		connp->conn_saddr_v6 = connp->conn_laddr_v6;
1751 	}
1752 
1753 	/*
1754 	 * Take care of the scope_id now.
1755 	 */
1756 	if (scope_id != 0 && IN6_IS_ADDR_LINKSCOPE(dstaddrp)) {
1757 		connp->conn_ixa->ixa_flags |= IXAF_SCOPEID_SET;
1758 		connp->conn_ixa->ixa_scopeid = scope_id;
1759 	} else {
1760 		connp->conn_ixa->ixa_flags &= ~IXAF_SCOPEID_SET;
1761 	}
1762 
1763 	connp->conn_flowinfo = flowinfo;
1764 	connp->conn_faddr_v6 = *dstaddrp;
1765 	connp->conn_fport = dstport;
1766 
1767 	/*
1768 	 * At this point the remote destination address and remote port fields
1769 	 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1770 	 * have to see which state tcp was in so we can take appropriate action.
1771 	 */
1772 	if (tcp->tcp_state == TCPS_IDLE) {
1773 		/*
1774 		 * We support a quick connect capability here, allowing
1775 		 * clients to transition directly from IDLE to SYN_SENT
1776 		 * tcp_bindi will pick an unused port, insert the connection
1777 		 * in the bind hash and transition to BOUND state.
1778 		 */
1779 		lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1780 		    tcp, B_TRUE);
1781 		lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1782 		    B_FALSE, B_FALSE);
1783 		if (lport == 0)
1784 			return (-TNOADDR);
1785 	}
1786 
1787 	/*
1788 	 * Lookup the route to determine a source address and the uinfo.
1789 	 * Setup TCP parameters based on the metrics/DCE.
1790 	 */
1791 	error = tcp_set_destination(tcp);
1792 	if (error != 0)
1793 		return (error);
1794 
1795 	/*
1796 	 * Don't let an endpoint connect to itself.
1797 	 */
1798 	if (IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6, &connp->conn_laddr_v6) &&
1799 	    connp->conn_fport == connp->conn_lport)
1800 		return (-TBADADDR);
1801 
1802 	tcp->tcp_state = TCPS_SYN_SENT;
1803 
1804 	return (ipcl_conn_insert_v6(connp));
1805 }
1806 
1807 /*
1808  * Disconnect
1809  * Note that unlike other functions this returns a positive tli error
1810  * when it fails; it never returns an errno.
1811  */
1812 static int
1813 tcp_disconnect_common(tcp_t *tcp, t_scalar_t seqnum)
1814 {
1815 	conn_t		*lconnp;
1816 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1817 	conn_t		*connp = tcp->tcp_connp;
1818 
1819 	/*
1820 	 * Right now, upper modules pass down a T_DISCON_REQ to TCP,
1821 	 * when the stream is in BOUND state. Do not send a reset,
1822 	 * since the destination IP address is not valid, and it can
1823 	 * be the initialized value of all zeros (broadcast address).
1824 	 */
1825 	if (tcp->tcp_state <= TCPS_BOUND) {
1826 		if (connp->conn_debug) {
1827 			(void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1828 			    "tcp_disconnect: bad state, %d", tcp->tcp_state);
1829 		}
1830 		return (TOUTSTATE);
1831 	} else if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1832 		TCPS_CONN_DEC(tcps);
1833 	}
1834 
1835 	if (seqnum == -1 || tcp->tcp_conn_req_max == 0) {
1836 
1837 		/*
1838 		 * According to TPI, for non-listeners, ignore seqnum
1839 		 * and disconnect.
1840 		 * Following interpretation of -1 seqnum is historical
1841 		 * and implied TPI ? (TPI only states that for T_CONN_IND,
1842 		 * a valid seqnum should not be -1).
1843 		 *
1844 		 *	-1 means disconnect everything
1845 		 *	regardless even on a listener.
1846 		 */
1847 
1848 		int old_state = tcp->tcp_state;
1849 		ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
1850 
1851 		/*
1852 		 * The connection can't be on the tcp_time_wait_head list
1853 		 * since it is not detached.
1854 		 */
1855 		ASSERT(tcp->tcp_time_wait_next == NULL);
1856 		ASSERT(tcp->tcp_time_wait_prev == NULL);
1857 		ASSERT(tcp->tcp_time_wait_expire == 0);
1858 		/*
1859 		 * If it used to be a listener, check to make sure no one else
1860 		 * has taken the port before switching back to LISTEN state.
1861 		 */
1862 		if (connp->conn_ipversion == IPV4_VERSION) {
1863 			lconnp = ipcl_lookup_listener_v4(connp->conn_lport,
1864 			    connp->conn_laddr_v4, IPCL_ZONEID(connp), ipst);
1865 		} else {
1866 			uint_t ifindex = 0;
1867 
1868 			if (connp->conn_ixa->ixa_flags & IXAF_SCOPEID_SET)
1869 				ifindex = connp->conn_ixa->ixa_scopeid;
1870 
1871 			/* Allow conn_bound_if listeners? */
1872 			lconnp = ipcl_lookup_listener_v6(connp->conn_lport,
1873 			    &connp->conn_laddr_v6, ifindex, IPCL_ZONEID(connp),
1874 			    ipst);
1875 		}
1876 		if (tcp->tcp_conn_req_max && lconnp == NULL) {
1877 			tcp->tcp_state = TCPS_LISTEN;
1878 		} else if (old_state > TCPS_BOUND) {
1879 			tcp->tcp_conn_req_max = 0;
1880 			tcp->tcp_state = TCPS_BOUND;
1881 
1882 			/*
1883 			 * If this end point is not going to become a listener,
1884 			 * decrement the listener connection count if
1885 			 * necessary.  Note that we do not do this if it is
1886 			 * going to be a listner (the above if case) since
1887 			 * then it may remove the counter struct.
1888 			 */
1889 			if (tcp->tcp_listen_cnt != NULL)
1890 				TCP_DECR_LISTEN_CNT(tcp);
1891 		}
1892 		if (lconnp != NULL)
1893 			CONN_DEC_REF(lconnp);
1894 		switch (old_state) {
1895 		case TCPS_SYN_SENT:
1896 		case TCPS_SYN_RCVD:
1897 			TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1898 			break;
1899 		case TCPS_ESTABLISHED:
1900 		case TCPS_CLOSE_WAIT:
1901 			TCPS_BUMP_MIB(tcps, tcpEstabResets);
1902 			break;
1903 		}
1904 
1905 		if (tcp->tcp_fused)
1906 			tcp_unfuse(tcp);
1907 
1908 		mutex_enter(&tcp->tcp_eager_lock);
1909 		if ((tcp->tcp_conn_req_cnt_q0 != 0) ||
1910 		    (tcp->tcp_conn_req_cnt_q != 0)) {
1911 			tcp_eager_cleanup(tcp, 0);
1912 		}
1913 		mutex_exit(&tcp->tcp_eager_lock);
1914 
1915 		tcp_xmit_ctl("tcp_disconnect", tcp, tcp->tcp_snxt,
1916 		    tcp->tcp_rnxt, TH_RST | TH_ACK);
1917 
1918 		tcp_reinit(tcp);
1919 
1920 		return (0);
1921 	} else if (!tcp_eager_blowoff(tcp, seqnum)) {
1922 		return (TBADSEQ);
1923 	}
1924 	return (0);
1925 }
1926 
1927 /*
1928  * Our client hereby directs us to reject the connection request
1929  * that tcp_input_listener() marked with 'seqnum'.  Rejection consists
1930  * of sending the appropriate RST, not an ICMP error.
1931  */
1932 void
1933 tcp_disconnect(tcp_t *tcp, mblk_t *mp)
1934 {
1935 	t_scalar_t seqnum;
1936 	int	error;
1937 	conn_t	*connp = tcp->tcp_connp;
1938 
1939 	ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1940 	if ((mp->b_wptr - mp->b_rptr) < sizeof (struct T_discon_req)) {
1941 		tcp_err_ack(tcp, mp, TPROTO, 0);
1942 		return;
1943 	}
1944 	seqnum = ((struct T_discon_req *)mp->b_rptr)->SEQ_number;
1945 	error = tcp_disconnect_common(tcp, seqnum);
1946 	if (error != 0)
1947 		tcp_err_ack(tcp, mp, error, 0);
1948 	else {
1949 		if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1950 			/* Send M_FLUSH according to TPI */
1951 			(void) putnextctl1(connp->conn_rq, M_FLUSH, FLUSHRW);
1952 		}
1953 		mp = mi_tpi_ok_ack_alloc(mp);
1954 		if (mp != NULL)
1955 			putnext(connp->conn_rq, mp);
1956 	}
1957 }
1958 
1959 /*
1960  * Note: No locks are held when inspecting tcp_g_*epriv_ports
1961  * but instead the code relies on:
1962  * - the fact that the address of the array and its size never changes
1963  * - the atomic assignment of the elements of the array
1964  */
1965 /* ARGSUSED */
1966 static int
1967 tcp_extra_priv_ports_get(queue_t *q, mblk_t *mp, caddr_t cp, cred_t *cr)
1968 {
1969 	int i;
1970 	tcp_stack_t	*tcps = Q_TO_TCP(q)->tcp_tcps;
1971 
1972 	for (i = 0; i < tcps->tcps_g_num_epriv_ports; i++) {
1973 		if (tcps->tcps_g_epriv_ports[i] != 0)
1974 			(void) mi_mpprintf(mp, "%d ",
1975 			    tcps->tcps_g_epriv_ports[i]);
1976 	}
1977 	return (0);
1978 }
1979 
1980 /*
1981  * Hold a lock while changing tcp_g_epriv_ports to prevent multiple
1982  * threads from changing it at the same time.
1983  */
1984 /* ARGSUSED */
1985 static int
1986 tcp_extra_priv_ports_add(queue_t *q, mblk_t *mp, char *value, caddr_t cp,
1987     cred_t *cr)
1988 {
1989 	long	new_value;
1990 	int	i;
1991 	tcp_stack_t	*tcps = Q_TO_TCP(q)->tcp_tcps;
1992 
1993 	/*
1994 	 * Fail the request if the new value does not lie within the
1995 	 * port number limits.
1996 	 */
1997 	if (ddi_strtol(value, NULL, 10, &new_value) != 0 ||
1998 	    new_value <= 0 || new_value >= 65536) {
1999 		return (EINVAL);
2000 	}
2001 
2002 	mutex_enter(&tcps->tcps_epriv_port_lock);
2003 	/* Check if the value is already in the list */
2004 	for (i = 0; i < tcps->tcps_g_num_epriv_ports; i++) {
2005 		if (new_value == tcps->tcps_g_epriv_ports[i]) {
2006 			mutex_exit(&tcps->tcps_epriv_port_lock);
2007 			return (EEXIST);
2008 		}
2009 	}
2010 	/* Find an empty slot */
2011 	for (i = 0; i < tcps->tcps_g_num_epriv_ports; i++) {
2012 		if (tcps->tcps_g_epriv_ports[i] == 0)
2013 			break;
2014 	}
2015 	if (i == tcps->tcps_g_num_epriv_ports) {
2016 		mutex_exit(&tcps->tcps_epriv_port_lock);
2017 		return (EOVERFLOW);
2018 	}
2019 	/* Set the new value */
2020 	tcps->tcps_g_epriv_ports[i] = (uint16_t)new_value;
2021 	mutex_exit(&tcps->tcps_epriv_port_lock);
2022 	return (0);
2023 }
2024 
2025 /*
2026  * Hold a lock while changing tcp_g_epriv_ports to prevent multiple
2027  * threads from changing it at the same time.
2028  */
2029 /* ARGSUSED */
2030 static int
2031 tcp_extra_priv_ports_del(queue_t *q, mblk_t *mp, char *value, caddr_t cp,
2032     cred_t *cr)
2033 {
2034 	long	new_value;
2035 	int	i;
2036 	tcp_stack_t	*tcps = Q_TO_TCP(q)->tcp_tcps;
2037 
2038 	/*
2039 	 * Fail the request if the new value does not lie within the
2040 	 * port number limits.
2041 	 */
2042 	if (ddi_strtol(value, NULL, 10, &new_value) != 0 || new_value <= 0 ||
2043 	    new_value >= 65536) {
2044 		return (EINVAL);
2045 	}
2046 
2047 	mutex_enter(&tcps->tcps_epriv_port_lock);
2048 	/* Check that the value is already in the list */
2049 	for (i = 0; i < tcps->tcps_g_num_epriv_ports; i++) {
2050 		if (tcps->tcps_g_epriv_ports[i] == new_value)
2051 			break;
2052 	}
2053 	if (i == tcps->tcps_g_num_epriv_ports) {
2054 		mutex_exit(&tcps->tcps_epriv_port_lock);
2055 		return (ESRCH);
2056 	}
2057 	/* Clear the value */
2058 	tcps->tcps_g_epriv_ports[i] = 0;
2059 	mutex_exit(&tcps->tcps_epriv_port_lock);
2060 	return (0);
2061 }
2062 
2063 /*
2064  * Handle reinitialization of a tcp structure.
2065  * Maintain "binding state" resetting the state to BOUND, LISTEN, or IDLE.
2066  */
2067 static void
2068 tcp_reinit(tcp_t *tcp)
2069 {
2070 	mblk_t		*mp;
2071 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2072 	conn_t		*connp  = tcp->tcp_connp;
2073 
2074 	/* tcp_reinit should never be called for detached tcp_t's */
2075 	ASSERT(tcp->tcp_listener == NULL);
2076 	ASSERT((connp->conn_family == AF_INET &&
2077 	    connp->conn_ipversion == IPV4_VERSION) ||
2078 	    (connp->conn_family == AF_INET6 &&
2079 	    (connp->conn_ipversion == IPV4_VERSION ||
2080 	    connp->conn_ipversion == IPV6_VERSION)));
2081 
2082 	/* Cancel outstanding timers */
2083 	tcp_timers_stop(tcp);
2084 
2085 	/*
2086 	 * Reset everything in the state vector, after updating global
2087 	 * MIB data from instance counters.
2088 	 */
2089 	TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
2090 	tcp->tcp_ibsegs = 0;
2091 	TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
2092 	tcp->tcp_obsegs = 0;
2093 
2094 	tcp_close_mpp(&tcp->tcp_xmit_head);
2095 	if (tcp->tcp_snd_zcopy_aware)
2096 		tcp_zcopy_notify(tcp);
2097 	tcp->tcp_xmit_last = tcp->tcp_xmit_tail = NULL;
2098 	tcp->tcp_unsent = tcp->tcp_xmit_tail_unsent = 0;
2099 	mutex_enter(&tcp->tcp_non_sq_lock);
2100 	if (tcp->tcp_flow_stopped &&
2101 	    TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
2102 		tcp_clrqfull(tcp);
2103 	}
2104 	mutex_exit(&tcp->tcp_non_sq_lock);
2105 	tcp_close_mpp(&tcp->tcp_reass_head);
2106 	tcp->tcp_reass_tail = NULL;
2107 	if (tcp->tcp_rcv_list != NULL) {
2108 		/* Free b_next chain */
2109 		tcp_close_mpp(&tcp->tcp_rcv_list);
2110 		tcp->tcp_rcv_last_head = NULL;
2111 		tcp->tcp_rcv_last_tail = NULL;
2112 		tcp->tcp_rcv_cnt = 0;
2113 	}
2114 	tcp->tcp_rcv_last_tail = NULL;
2115 
2116 	if ((mp = tcp->tcp_urp_mp) != NULL) {
2117 		freemsg(mp);
2118 		tcp->tcp_urp_mp = NULL;
2119 	}
2120 	if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
2121 		freemsg(mp);
2122 		tcp->tcp_urp_mark_mp = NULL;
2123 	}
2124 	if (tcp->tcp_fused_sigurg_mp != NULL) {
2125 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
2126 		freeb(tcp->tcp_fused_sigurg_mp);
2127 		tcp->tcp_fused_sigurg_mp = NULL;
2128 	}
2129 	if (tcp->tcp_ordrel_mp != NULL) {
2130 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
2131 		freeb(tcp->tcp_ordrel_mp);
2132 		tcp->tcp_ordrel_mp = NULL;
2133 	}
2134 
2135 	/*
2136 	 * Following is a union with two members which are
2137 	 * identical types and size so the following cleanup
2138 	 * is enough.
2139 	 */
2140 	tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
2141 
2142 	CL_INET_DISCONNECT(connp);
2143 
2144 	/*
2145 	 * The connection can't be on the tcp_time_wait_head list
2146 	 * since it is not detached.
2147 	 */
2148 	ASSERT(tcp->tcp_time_wait_next == NULL);
2149 	ASSERT(tcp->tcp_time_wait_prev == NULL);
2150 	ASSERT(tcp->tcp_time_wait_expire == 0);
2151 
2152 	if (tcp->tcp_kssl_pending) {
2153 		tcp->tcp_kssl_pending = B_FALSE;
2154 
2155 		/* Don't reset if the initialized by bind. */
2156 		if (tcp->tcp_kssl_ent != NULL) {
2157 			kssl_release_ent(tcp->tcp_kssl_ent, NULL,
2158 			    KSSL_NO_PROXY);
2159 		}
2160 	}
2161 	if (tcp->tcp_kssl_ctx != NULL) {
2162 		kssl_release_ctx(tcp->tcp_kssl_ctx);
2163 		tcp->tcp_kssl_ctx = NULL;
2164 	}
2165 
2166 	/*
2167 	 * Reset/preserve other values
2168 	 */
2169 	tcp_reinit_values(tcp);
2170 	ipcl_hash_remove(connp);
2171 	/* Note that ixa_cred gets cleared in ixa_cleanup */
2172 	ixa_cleanup(connp->conn_ixa);
2173 	tcp_ipsec_cleanup(tcp);
2174 
2175 	connp->conn_laddr_v6 = connp->conn_bound_addr_v6;
2176 	connp->conn_saddr_v6 = connp->conn_bound_addr_v6;
2177 
2178 	if (tcp->tcp_conn_req_max != 0) {
2179 		/*
2180 		 * This is the case when a TLI program uses the same
2181 		 * transport end point to accept a connection.  This
2182 		 * makes the TCP both a listener and acceptor.  When
2183 		 * this connection is closed, we need to set the state
2184 		 * back to TCPS_LISTEN.  Make sure that the eager list
2185 		 * is reinitialized.
2186 		 *
2187 		 * Note that this stream is still bound to the four
2188 		 * tuples of the previous connection in IP.  If a new
2189 		 * SYN with different foreign address comes in, IP will
2190 		 * not find it and will send it to the global queue.  In
2191 		 * the global queue, TCP will do a tcp_lookup_listener()
2192 		 * to find this stream.  This works because this stream
2193 		 * is only removed from connected hash.
2194 		 *
2195 		 */
2196 		tcp->tcp_state = TCPS_LISTEN;
2197 		tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
2198 		tcp->tcp_eager_next_drop_q0 = tcp;
2199 		tcp->tcp_eager_prev_drop_q0 = tcp;
2200 		/*
2201 		 * Initially set conn_recv to tcp_input_listener_unbound to try
2202 		 * to pick a good squeue for the listener when the first SYN
2203 		 * arrives. tcp_input_listener_unbound sets it to
2204 		 * tcp_input_listener on that first SYN.
2205 		 */
2206 		connp->conn_recv = tcp_input_listener_unbound;
2207 
2208 		connp->conn_proto = IPPROTO_TCP;
2209 		connp->conn_faddr_v6 = ipv6_all_zeros;
2210 		connp->conn_fport = 0;
2211 
2212 		(void) ipcl_bind_insert(connp);
2213 	} else {
2214 		tcp->tcp_state = TCPS_BOUND;
2215 	}
2216 
2217 	/*
2218 	 * Initialize to default values
2219 	 */
2220 	tcp_init_values(tcp);
2221 
2222 	ASSERT(tcp->tcp_ptpbhn != NULL);
2223 	tcp->tcp_rwnd = connp->conn_rcvbuf;
2224 	tcp->tcp_mss = connp->conn_ipversion != IPV4_VERSION ?
2225 	    tcps->tcps_mss_def_ipv6 : tcps->tcps_mss_def_ipv4;
2226 }
2227 
2228 /*
2229  * Force values to zero that need be zero.
2230  * Do not touch values asociated with the BOUND or LISTEN state
2231  * since the connection will end up in that state after the reinit.
2232  * NOTE: tcp_reinit_values MUST have a line for each field in the tcp_t
2233  * structure!
2234  */
2235 static void
2236 tcp_reinit_values(tcp)
2237 	tcp_t *tcp;
2238 {
2239 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2240 	conn_t		*connp = tcp->tcp_connp;
2241 
2242 #ifndef	lint
2243 #define	DONTCARE(x)
2244 #define	PRESERVE(x)
2245 #else
2246 #define	DONTCARE(x)	((x) = (x))
2247 #define	PRESERVE(x)	((x) = (x))
2248 #endif	/* lint */
2249 
2250 	PRESERVE(tcp->tcp_bind_hash_port);
2251 	PRESERVE(tcp->tcp_bind_hash);
2252 	PRESERVE(tcp->tcp_ptpbhn);
2253 	PRESERVE(tcp->tcp_acceptor_hash);
2254 	PRESERVE(tcp->tcp_ptpahn);
2255 
2256 	/* Should be ASSERT NULL on these with new code! */
2257 	ASSERT(tcp->tcp_time_wait_next == NULL);
2258 	ASSERT(tcp->tcp_time_wait_prev == NULL);
2259 	ASSERT(tcp->tcp_time_wait_expire == 0);
2260 	PRESERVE(tcp->tcp_state);
2261 	PRESERVE(connp->conn_rq);
2262 	PRESERVE(connp->conn_wq);
2263 
2264 	ASSERT(tcp->tcp_xmit_head == NULL);
2265 	ASSERT(tcp->tcp_xmit_last == NULL);
2266 	ASSERT(tcp->tcp_unsent == 0);
2267 	ASSERT(tcp->tcp_xmit_tail == NULL);
2268 	ASSERT(tcp->tcp_xmit_tail_unsent == 0);
2269 
2270 	tcp->tcp_snxt = 0;			/* Displayed in mib */
2271 	tcp->tcp_suna = 0;			/* Displayed in mib */
2272 	tcp->tcp_swnd = 0;
2273 	DONTCARE(tcp->tcp_cwnd);	/* Init in tcp_process_options */
2274 
2275 	ASSERT(tcp->tcp_ibsegs == 0);
2276 	ASSERT(tcp->tcp_obsegs == 0);
2277 
2278 	if (connp->conn_ht_iphc != NULL) {
2279 		kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
2280 		connp->conn_ht_iphc = NULL;
2281 		connp->conn_ht_iphc_allocated = 0;
2282 		connp->conn_ht_iphc_len = 0;
2283 		connp->conn_ht_ulp = NULL;
2284 		connp->conn_ht_ulp_len = 0;
2285 		tcp->tcp_ipha = NULL;
2286 		tcp->tcp_ip6h = NULL;
2287 		tcp->tcp_tcpha = NULL;
2288 	}
2289 
2290 	/* We clear any IP_OPTIONS and extension headers */
2291 	ip_pkt_free(&connp->conn_xmit_ipp);
2292 
2293 	DONTCARE(tcp->tcp_naglim);		/* Init in tcp_init_values */
2294 	DONTCARE(tcp->tcp_ipha);
2295 	DONTCARE(tcp->tcp_ip6h);
2296 	DONTCARE(tcp->tcp_tcpha);
2297 	tcp->tcp_valid_bits = 0;
2298 
2299 	DONTCARE(tcp->tcp_timer_backoff);	/* Init in tcp_init_values */
2300 	DONTCARE(tcp->tcp_last_recv_time);	/* Init in tcp_init_values */
2301 	tcp->tcp_last_rcv_lbolt = 0;
2302 
2303 	tcp->tcp_init_cwnd = 0;
2304 
2305 	tcp->tcp_urp_last_valid = 0;
2306 	tcp->tcp_hard_binding = 0;
2307 
2308 	tcp->tcp_fin_acked = 0;
2309 	tcp->tcp_fin_rcvd = 0;
2310 	tcp->tcp_fin_sent = 0;
2311 	tcp->tcp_ordrel_done = 0;
2312 
2313 	tcp->tcp_detached = 0;
2314 
2315 	tcp->tcp_snd_ws_ok = B_FALSE;
2316 	tcp->tcp_snd_ts_ok = B_FALSE;
2317 	tcp->tcp_zero_win_probe = 0;
2318 
2319 	tcp->tcp_loopback = 0;
2320 	tcp->tcp_localnet = 0;
2321 	tcp->tcp_syn_defense = 0;
2322 	tcp->tcp_set_timer = 0;
2323 
2324 	tcp->tcp_active_open = 0;
2325 	tcp->tcp_rexmit = B_FALSE;
2326 	tcp->tcp_xmit_zc_clean = B_FALSE;
2327 
2328 	tcp->tcp_snd_sack_ok = B_FALSE;
2329 	tcp->tcp_hwcksum = B_FALSE;
2330 
2331 	DONTCARE(tcp->tcp_maxpsz_multiplier);	/* Init in tcp_init_values */
2332 
2333 	tcp->tcp_conn_def_q0 = 0;
2334 	tcp->tcp_ip_forward_progress = B_FALSE;
2335 	tcp->tcp_ecn_ok = B_FALSE;
2336 
2337 	tcp->tcp_cwr = B_FALSE;
2338 	tcp->tcp_ecn_echo_on = B_FALSE;
2339 	tcp->tcp_is_wnd_shrnk = B_FALSE;
2340 
2341 	if (tcp->tcp_sack_info != NULL) {
2342 		if (tcp->tcp_notsack_list != NULL) {
2343 			TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list,
2344 			    tcp);
2345 		}
2346 		kmem_cache_free(tcp_sack_info_cache, tcp->tcp_sack_info);
2347 		tcp->tcp_sack_info = NULL;
2348 	}
2349 
2350 	tcp->tcp_rcv_ws = 0;
2351 	tcp->tcp_snd_ws = 0;
2352 	tcp->tcp_ts_recent = 0;
2353 	tcp->tcp_rnxt = 0;			/* Displayed in mib */
2354 	DONTCARE(tcp->tcp_rwnd);		/* Set in tcp_reinit() */
2355 	tcp->tcp_initial_pmtu = 0;
2356 
2357 	ASSERT(tcp->tcp_reass_head == NULL);
2358 	ASSERT(tcp->tcp_reass_tail == NULL);
2359 
2360 	tcp->tcp_cwnd_cnt = 0;
2361 
2362 	ASSERT(tcp->tcp_rcv_list == NULL);
2363 	ASSERT(tcp->tcp_rcv_last_head == NULL);
2364 	ASSERT(tcp->tcp_rcv_last_tail == NULL);
2365 	ASSERT(tcp->tcp_rcv_cnt == 0);
2366 
2367 	DONTCARE(tcp->tcp_cwnd_ssthresh); /* Init in tcp_set_destination */
2368 	DONTCARE(tcp->tcp_cwnd_max);		/* Init in tcp_init_values */
2369 	tcp->tcp_csuna = 0;
2370 
2371 	tcp->tcp_rto = 0;			/* Displayed in MIB */
2372 	DONTCARE(tcp->tcp_rtt_sa);		/* Init in tcp_init_values */
2373 	DONTCARE(tcp->tcp_rtt_sd);		/* Init in tcp_init_values */
2374 	tcp->tcp_rtt_update = 0;
2375 
2376 	DONTCARE(tcp->tcp_swl1); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2377 	DONTCARE(tcp->tcp_swl2); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2378 
2379 	tcp->tcp_rack = 0;			/* Displayed in mib */
2380 	tcp->tcp_rack_cnt = 0;
2381 	tcp->tcp_rack_cur_max = 0;
2382 	tcp->tcp_rack_abs_max = 0;
2383 
2384 	tcp->tcp_max_swnd = 0;
2385 
2386 	ASSERT(tcp->tcp_listener == NULL);
2387 
2388 	DONTCARE(tcp->tcp_irs);			/* tcp_valid_bits cleared */
2389 	DONTCARE(tcp->tcp_iss);			/* tcp_valid_bits cleared */
2390 	DONTCARE(tcp->tcp_fss);			/* tcp_valid_bits cleared */
2391 	DONTCARE(tcp->tcp_urg);			/* tcp_valid_bits cleared */
2392 
2393 	ASSERT(tcp->tcp_conn_req_cnt_q == 0);
2394 	ASSERT(tcp->tcp_conn_req_cnt_q0 == 0);
2395 	PRESERVE(tcp->tcp_conn_req_max);
2396 	PRESERVE(tcp->tcp_conn_req_seqnum);
2397 
2398 	DONTCARE(tcp->tcp_first_timer_threshold); /* Init in tcp_init_values */
2399 	DONTCARE(tcp->tcp_second_timer_threshold); /* Init in tcp_init_values */
2400 	DONTCARE(tcp->tcp_first_ctimer_threshold); /* Init in tcp_init_values */
2401 	DONTCARE(tcp->tcp_second_ctimer_threshold); /* in tcp_init_values */
2402 
2403 	DONTCARE(tcp->tcp_urp_last);	/* tcp_urp_last_valid is cleared */
2404 	ASSERT(tcp->tcp_urp_mp == NULL);
2405 	ASSERT(tcp->tcp_urp_mark_mp == NULL);
2406 	ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
2407 
2408 	ASSERT(tcp->tcp_eager_next_q == NULL);
2409 	ASSERT(tcp->tcp_eager_last_q == NULL);
2410 	ASSERT((tcp->tcp_eager_next_q0 == NULL &&
2411 	    tcp->tcp_eager_prev_q0 == NULL) ||
2412 	    tcp->tcp_eager_next_q0 == tcp->tcp_eager_prev_q0);
2413 	ASSERT(tcp->tcp_conn.tcp_eager_conn_ind == NULL);
2414 
2415 	ASSERT((tcp->tcp_eager_next_drop_q0 == NULL &&
2416 	    tcp->tcp_eager_prev_drop_q0 == NULL) ||
2417 	    tcp->tcp_eager_next_drop_q0 == tcp->tcp_eager_prev_drop_q0);
2418 
2419 	tcp->tcp_client_errno = 0;
2420 
2421 	DONTCARE(connp->conn_sum);		/* Init in tcp_init_values */
2422 
2423 	connp->conn_faddr_v6 = ipv6_all_zeros;	/* Displayed in MIB */
2424 
2425 	PRESERVE(connp->conn_bound_addr_v6);
2426 	tcp->tcp_last_sent_len = 0;
2427 	tcp->tcp_dupack_cnt = 0;
2428 
2429 	connp->conn_fport = 0;			/* Displayed in MIB */
2430 	PRESERVE(connp->conn_lport);
2431 
2432 	PRESERVE(tcp->tcp_acceptor_lockp);
2433 
2434 	ASSERT(tcp->tcp_ordrel_mp == NULL);
2435 	PRESERVE(tcp->tcp_acceptor_id);
2436 	DONTCARE(tcp->tcp_ipsec_overhead);
2437 
2438 	PRESERVE(connp->conn_family);
2439 	/* Remove any remnants of mapped address binding */
2440 	if (connp->conn_family == AF_INET6) {
2441 		connp->conn_ipversion = IPV6_VERSION;
2442 		tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2443 	} else {
2444 		connp->conn_ipversion = IPV4_VERSION;
2445 		tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2446 	}
2447 
2448 	connp->conn_bound_if = 0;
2449 	connp->conn_recv_ancillary.crb_all = 0;
2450 	tcp->tcp_recvifindex = 0;
2451 	tcp->tcp_recvhops = 0;
2452 	tcp->tcp_closed = 0;
2453 	if (tcp->tcp_hopopts != NULL) {
2454 		mi_free(tcp->tcp_hopopts);
2455 		tcp->tcp_hopopts = NULL;
2456 		tcp->tcp_hopoptslen = 0;
2457 	}
2458 	ASSERT(tcp->tcp_hopoptslen == 0);
2459 	if (tcp->tcp_dstopts != NULL) {
2460 		mi_free(tcp->tcp_dstopts);
2461 		tcp->tcp_dstopts = NULL;
2462 		tcp->tcp_dstoptslen = 0;
2463 	}
2464 	ASSERT(tcp->tcp_dstoptslen == 0);
2465 	if (tcp->tcp_rthdrdstopts != NULL) {
2466 		mi_free(tcp->tcp_rthdrdstopts);
2467 		tcp->tcp_rthdrdstopts = NULL;
2468 		tcp->tcp_rthdrdstoptslen = 0;
2469 	}
2470 	ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2471 	if (tcp->tcp_rthdr != NULL) {
2472 		mi_free(tcp->tcp_rthdr);
2473 		tcp->tcp_rthdr = NULL;
2474 		tcp->tcp_rthdrlen = 0;
2475 	}
2476 	ASSERT(tcp->tcp_rthdrlen == 0);
2477 
2478 	/* Reset fusion-related fields */
2479 	tcp->tcp_fused = B_FALSE;
2480 	tcp->tcp_unfusable = B_FALSE;
2481 	tcp->tcp_fused_sigurg = B_FALSE;
2482 	tcp->tcp_loopback_peer = NULL;
2483 
2484 	tcp->tcp_lso = B_FALSE;
2485 
2486 	tcp->tcp_in_ack_unsent = 0;
2487 	tcp->tcp_cork = B_FALSE;
2488 	tcp->tcp_tconnind_started = B_FALSE;
2489 
2490 	PRESERVE(tcp->tcp_squeue_bytes);
2491 
2492 	ASSERT(tcp->tcp_kssl_ctx == NULL);
2493 	ASSERT(!tcp->tcp_kssl_pending);
2494 	PRESERVE(tcp->tcp_kssl_ent);
2495 
2496 	tcp->tcp_closemp_used = B_FALSE;
2497 
2498 	PRESERVE(tcp->tcp_rsrv_mp);
2499 	PRESERVE(tcp->tcp_rsrv_mp_lock);
2500 
2501 #ifdef DEBUG
2502 	DONTCARE(tcp->tcmp_stk[0]);
2503 #endif
2504 
2505 	PRESERVE(tcp->tcp_connid);
2506 
2507 	ASSERT(tcp->tcp_listen_cnt == NULL);
2508 	ASSERT(tcp->tcp_reass_tid == 0);
2509 
2510 #undef	DONTCARE
2511 #undef	PRESERVE
2512 }
2513 
2514 void
2515 tcp_init_values(tcp_t *tcp)
2516 {
2517 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2518 	conn_t		*connp = tcp->tcp_connp;
2519 
2520 	ASSERT((connp->conn_family == AF_INET &&
2521 	    connp->conn_ipversion == IPV4_VERSION) ||
2522 	    (connp->conn_family == AF_INET6 &&
2523 	    (connp->conn_ipversion == IPV4_VERSION ||
2524 	    connp->conn_ipversion == IPV6_VERSION)));
2525 
2526 	/*
2527 	 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2528 	 * will be close to tcp_rexmit_interval_initial.  By doing this, we
2529 	 * allow the algorithm to adjust slowly to large fluctuations of RTT
2530 	 * during first few transmissions of a connection as seen in slow
2531 	 * links.
2532 	 */
2533 	tcp->tcp_rtt_sa = tcps->tcps_rexmit_interval_initial << 2;
2534 	tcp->tcp_rtt_sd = tcps->tcps_rexmit_interval_initial >> 1;
2535 	tcp->tcp_rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
2536 	    tcps->tcps_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
2537 	    tcps->tcps_conn_grace_period;
2538 	if (tcp->tcp_rto < tcps->tcps_rexmit_interval_min)
2539 		tcp->tcp_rto = tcps->tcps_rexmit_interval_min;
2540 	tcp->tcp_timer_backoff = 0;
2541 	tcp->tcp_ms_we_have_waited = 0;
2542 	tcp->tcp_last_recv_time = ddi_get_lbolt();
2543 	tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2544 	tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2545 	tcp->tcp_snd_burst = TCP_CWND_INFINITE;
2546 
2547 	tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2548 
2549 	tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2550 	tcp->tcp_first_ctimer_threshold = tcps->tcps_ip_notify_cinterval;
2551 	tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2552 	/*
2553 	 * Fix it to tcp_ip_abort_linterval later if it turns out to be a
2554 	 * passive open.
2555 	 */
2556 	tcp->tcp_second_ctimer_threshold = tcps->tcps_ip_abort_cinterval;
2557 
2558 	tcp->tcp_naglim = tcps->tcps_naglim_def;
2559 
2560 	/* NOTE:  ISS is now set in tcp_set_destination(). */
2561 
2562 	/* Reset fusion-related fields */
2563 	tcp->tcp_fused = B_FALSE;
2564 	tcp->tcp_unfusable = B_FALSE;
2565 	tcp->tcp_fused_sigurg = B_FALSE;
2566 	tcp->tcp_loopback_peer = NULL;
2567 
2568 	/* We rebuild the header template on the next connect/conn_request */
2569 
2570 	connp->conn_mlp_type = mlptSingle;
2571 
2572 	/*
2573 	 * Init the window scale to the max so tcp_rwnd_set() won't pare
2574 	 * down tcp_rwnd. tcp_set_destination() will set the right value later.
2575 	 */
2576 	tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2577 	tcp->tcp_rwnd = connp->conn_rcvbuf;
2578 
2579 	tcp->tcp_cork = B_FALSE;
2580 	/*
2581 	 * Init the tcp_debug option if it wasn't already set.  This value
2582 	 * determines whether TCP
2583 	 * calls strlog() to print out debug messages.  Doing this
2584 	 * initialization here means that this value is not inherited thru
2585 	 * tcp_reinit().
2586 	 */
2587 	if (!connp->conn_debug)
2588 		connp->conn_debug = tcps->tcps_dbg;
2589 
2590 	tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2591 	tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2592 }
2593 
2594 /*
2595  * Update the TCP connection according to change of PMTU.
2596  *
2597  * Path MTU might have changed by either increase or decrease, so need to
2598  * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2599  * or negative MSS, since tcp_mss_set() will do it.
2600  */
2601 void
2602 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2603 {
2604 	uint32_t	pmtu;
2605 	int32_t		mss;
2606 	conn_t		*connp = tcp->tcp_connp;
2607 	ip_xmit_attr_t	*ixa = connp->conn_ixa;
2608 	iaflags_t	ixaflags;
2609 
2610 	if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2611 		return;
2612 
2613 	if (tcp->tcp_state < TCPS_ESTABLISHED)
2614 		return;
2615 
2616 	/*
2617 	 * Always call ip_get_pmtu() to make sure that IP has updated
2618 	 * ixa_flags properly.
2619 	 */
2620 	pmtu = ip_get_pmtu(ixa);
2621 	ixaflags = ixa->ixa_flags;
2622 
2623 	/*
2624 	 * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2625 	 * IPsec overhead if applied. Make sure to use the most recent
2626 	 * IPsec information.
2627 	 */
2628 	mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2629 
2630 	/*
2631 	 * Nothing to change, so just return.
2632 	 */
2633 	if (mss == tcp->tcp_mss)
2634 		return;
2635 
2636 	/*
2637 	 * Currently, for ICMP errors, only PMTU decrease is handled.
2638 	 */
2639 	if (mss > tcp->tcp_mss && decrease_only)
2640 		return;
2641 
2642 	DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2643 
2644 	/*
2645 	 * Update ixa_fragsize and ixa_pmtu.
2646 	 */
2647 	ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2648 
2649 	/*
2650 	 * Adjust MSS and all relevant variables.
2651 	 */
2652 	tcp_mss_set(tcp, mss);
2653 
2654 	/*
2655 	 * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2656 	 * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2657 	 * has a (potentially different) min size we do the same. Make sure to
2658 	 * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2659 	 * fragment the packet.
2660 	 *
2661 	 * LSO over IPv6 can not be fragmented. So need to disable LSO
2662 	 * when IPv6 fragmentation is needed.
2663 	 */
2664 	if (mss < tcp->tcp_tcps->tcps_mss_min)
2665 		ixaflags |= IXAF_PMTU_TOO_SMALL;
2666 
2667 	if (ixaflags & IXAF_PMTU_TOO_SMALL)
2668 		ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2669 
2670 	if ((connp->conn_ipversion == IPV4_VERSION) &&
2671 	    !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2672 		tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2673 	}
2674 	ixa->ixa_flags = ixaflags;
2675 }
2676 
2677 int
2678 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2679 {
2680 	conn_t	*connp = tcp->tcp_connp;
2681 	queue_t	*q = connp->conn_rq;
2682 	int32_t	mss = tcp->tcp_mss;
2683 	int	maxpsz;
2684 
2685 	if (TCP_IS_DETACHED(tcp))
2686 		return (mss);
2687 	if (tcp->tcp_fused) {
2688 		maxpsz = tcp_fuse_maxpsz(tcp);
2689 		mss = INFPSZ;
2690 	} else if (tcp->tcp_maxpsz_multiplier == 0) {
2691 		/*
2692 		 * Set the sd_qn_maxpsz according to the socket send buffer
2693 		 * size, and sd_maxblk to INFPSZ (-1).  This will essentially
2694 		 * instruct the stream head to copyin user data into contiguous
2695 		 * kernel-allocated buffers without breaking it up into smaller
2696 		 * chunks.  We round up the buffer size to the nearest SMSS.
2697 		 */
2698 		maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2699 		if (tcp->tcp_kssl_ctx == NULL)
2700 			mss = INFPSZ;
2701 		else
2702 			mss = SSL3_MAX_RECORD_LEN;
2703 	} else {
2704 		/*
2705 		 * Set sd_qn_maxpsz to approx half the (receivers) buffer
2706 		 * (and a multiple of the mss).  This instructs the stream
2707 		 * head to break down larger than SMSS writes into SMSS-
2708 		 * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2709 		 */
2710 		maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2711 		if (maxpsz > connp->conn_sndbuf / 2) {
2712 			maxpsz = connp->conn_sndbuf / 2;
2713 			/* Round up to nearest mss */
2714 			maxpsz = MSS_ROUNDUP(maxpsz, mss);
2715 		}
2716 	}
2717 
2718 	(void) proto_set_maxpsz(q, connp, maxpsz);
2719 	if (!(IPCL_IS_NONSTR(connp)))
2720 		connp->conn_wq->q_maxpsz = maxpsz;
2721 	if (set_maxblk)
2722 		(void) proto_set_tx_maxblk(q, connp, mss);
2723 	return (mss);
2724 }
2725 
2726 /* For /dev/tcp aka AF_INET open */
2727 static int
2728 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2729 {
2730 	return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2731 }
2732 
2733 /* For /dev/tcp6 aka AF_INET6 open */
2734 static int
2735 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2736 {
2737 	return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2738 }
2739 
2740 conn_t *
2741 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2742     int *errorp)
2743 {
2744 	tcp_t		*tcp = NULL;
2745 	conn_t		*connp;
2746 	zoneid_t	zoneid;
2747 	tcp_stack_t	*tcps;
2748 	squeue_t	*sqp;
2749 
2750 	ASSERT(errorp != NULL);
2751 	/*
2752 	 * Find the proper zoneid and netstack.
2753 	 */
2754 	/*
2755 	 * Special case for install: miniroot needs to be able to
2756 	 * access files via NFS as though it were always in the
2757 	 * global zone.
2758 	 */
2759 	if (credp == kcred && nfs_global_client_only != 0) {
2760 		zoneid = GLOBAL_ZONEID;
2761 		tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2762 		    netstack_tcp;
2763 		ASSERT(tcps != NULL);
2764 	} else {
2765 		netstack_t *ns;
2766 		int err;
2767 
2768 		if ((err = secpolicy_basic_net_access(credp)) != 0) {
2769 			*errorp = err;
2770 			return (NULL);
2771 		}
2772 
2773 		ns = netstack_find_by_cred(credp);
2774 		ASSERT(ns != NULL);
2775 		tcps = ns->netstack_tcp;
2776 		ASSERT(tcps != NULL);
2777 
2778 		/*
2779 		 * For exclusive stacks we set the zoneid to zero
2780 		 * to make TCP operate as if in the global zone.
2781 		 */
2782 		if (tcps->tcps_netstack->netstack_stackid !=
2783 		    GLOBAL_NETSTACKID)
2784 			zoneid = GLOBAL_ZONEID;
2785 		else
2786 			zoneid = crgetzoneid(credp);
2787 	}
2788 
2789 	sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2790 	connp = (conn_t *)tcp_get_conn(sqp, tcps);
2791 	/*
2792 	 * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2793 	 * so we drop it by one.
2794 	 */
2795 	netstack_rele(tcps->tcps_netstack);
2796 	if (connp == NULL) {
2797 		*errorp = ENOSR;
2798 		return (NULL);
2799 	}
2800 	ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2801 
2802 	connp->conn_sqp = sqp;
2803 	connp->conn_initial_sqp = connp->conn_sqp;
2804 	connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2805 	tcp = connp->conn_tcp;
2806 
2807 	/*
2808 	 * Besides asking IP to set the checksum for us, have conn_ip_output
2809 	 * to do the following checks when necessary:
2810 	 *
2811 	 * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2812 	 * IXAF_VERIFY_PMTU: verify PMTU changes
2813 	 * IXAF_VERIFY_LSO: verify LSO capability changes
2814 	 */
2815 	connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2816 	    IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2817 
2818 	if (!tcps->tcps_dev_flow_ctl)
2819 		connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2820 
2821 	if (isv6) {
2822 		connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2823 		connp->conn_ipversion = IPV6_VERSION;
2824 		connp->conn_family = AF_INET6;
2825 		tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2826 		connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2827 	} else {
2828 		connp->conn_ipversion = IPV4_VERSION;
2829 		connp->conn_family = AF_INET;
2830 		tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2831 		connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2832 	}
2833 	connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2834 
2835 	crhold(credp);
2836 	connp->conn_cred = credp;
2837 	connp->conn_cpid = curproc->p_pid;
2838 	connp->conn_open_time = ddi_get_lbolt64();
2839 
2840 	/* Cache things in the ixa without any refhold */
2841 	connp->conn_ixa->ixa_cred = credp;
2842 	connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2843 
2844 	connp->conn_zoneid = zoneid;
2845 	/* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2846 	connp->conn_ixa->ixa_zoneid = zoneid;
2847 	connp->conn_mlp_type = mlptSingle;
2848 	ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2849 	ASSERT(tcp->tcp_tcps == tcps);
2850 
2851 	/*
2852 	 * If the caller has the process-wide flag set, then default to MAC
2853 	 * exempt mode.  This allows read-down to unlabeled hosts.
2854 	 */
2855 	if (getpflags(NET_MAC_AWARE, credp) != 0)
2856 		connp->conn_mac_mode = CONN_MAC_AWARE;
2857 
2858 	connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2859 
2860 	if (issocket) {
2861 		tcp->tcp_issocket = 1;
2862 	}
2863 
2864 	connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2865 	connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2866 	connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2867 	connp->conn_so_type = SOCK_STREAM;
2868 	connp->conn_wroff = connp->conn_ht_iphc_allocated +
2869 	    tcps->tcps_wroff_xtra;
2870 
2871 	SOCK_CONNID_INIT(tcp->tcp_connid);
2872 	tcp->tcp_state = TCPS_IDLE;
2873 	tcp_init_values(tcp);
2874 	return (connp);
2875 }
2876 
2877 static int
2878 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2879     boolean_t isv6)
2880 {
2881 	tcp_t		*tcp = NULL;
2882 	conn_t		*connp = NULL;
2883 	int		err;
2884 	vmem_t		*minor_arena = NULL;
2885 	dev_t		conn_dev;
2886 	boolean_t	issocket;
2887 
2888 	if (q->q_ptr != NULL)
2889 		return (0);
2890 
2891 	if (sflag == MODOPEN)
2892 		return (EINVAL);
2893 
2894 	if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2895 	    ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2896 		minor_arena = ip_minor_arena_la;
2897 	} else {
2898 		/*
2899 		 * Either minor numbers in the large arena were exhausted
2900 		 * or a non socket application is doing the open.
2901 		 * Try to allocate from the small arena.
2902 		 */
2903 		if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2904 			return (EBUSY);
2905 		}
2906 		minor_arena = ip_minor_arena_sa;
2907 	}
2908 
2909 	ASSERT(minor_arena != NULL);
2910 
2911 	*devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2912 
2913 	if (flag & SO_FALLBACK) {
2914 		/*
2915 		 * Non streams socket needs a stream to fallback to
2916 		 */
2917 		RD(q)->q_ptr = (void *)conn_dev;
2918 		WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2919 		WR(q)->q_ptr = (void *)minor_arena;
2920 		qprocson(q);
2921 		return (0);
2922 	} else if (flag & SO_ACCEPTOR) {
2923 		q->q_qinfo = &tcp_acceptor_rinit;
2924 		/*
2925 		 * the conn_dev and minor_arena will be subsequently used by
2926 		 * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2927 		 * the minor device number for this connection from the q_ptr.
2928 		 */
2929 		RD(q)->q_ptr = (void *)conn_dev;
2930 		WR(q)->q_qinfo = &tcp_acceptor_winit;
2931 		WR(q)->q_ptr = (void *)minor_arena;
2932 		qprocson(q);
2933 		return (0);
2934 	}
2935 
2936 	issocket = flag & SO_SOCKSTR;
2937 	connp = tcp_create_common(credp, isv6, issocket, &err);
2938 
2939 	if (connp == NULL) {
2940 		inet_minor_free(minor_arena, conn_dev);
2941 		q->q_ptr = WR(q)->q_ptr = NULL;
2942 		return (err);
2943 	}
2944 
2945 	connp->conn_rq = q;
2946 	connp->conn_wq = WR(q);
2947 	q->q_ptr = WR(q)->q_ptr = connp;
2948 
2949 	connp->conn_dev = conn_dev;
2950 	connp->conn_minor_arena = minor_arena;
2951 
2952 	ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2953 	ASSERT(WR(q)->q_qinfo == &tcp_winit);
2954 
2955 	tcp = connp->conn_tcp;
2956 
2957 	if (issocket) {
2958 		WR(q)->q_qinfo = &tcp_sock_winit;
2959 	} else {
2960 #ifdef  _ILP32
2961 		tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2962 #else
2963 		tcp->tcp_acceptor_id = conn_dev;
2964 #endif  /* _ILP32 */
2965 		tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2966 	}
2967 
2968 	/*
2969 	 * Put the ref for TCP. Ref for IP was already put
2970 	 * by ipcl_conn_create. Also Make the conn_t globally
2971 	 * visible to walkers
2972 	 */
2973 	mutex_enter(&connp->conn_lock);
2974 	CONN_INC_REF_LOCKED(connp);
2975 	ASSERT(connp->conn_ref == 2);
2976 	connp->conn_state_flags &= ~CONN_INCIPIENT;
2977 	mutex_exit(&connp->conn_lock);
2978 
2979 	qprocson(q);
2980 	return (0);
2981 }
2982 
2983 /*
2984  * Build/update the tcp header template (in conn_ht_iphc) based on
2985  * conn_xmit_ipp. The headers include ip6_t, any extension
2986  * headers, and the maximum size tcp header (to avoid reallocation
2987  * on the fly for additional tcp options).
2988  *
2989  * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2990  * Returns failure if can't allocate memory.
2991  */
2992 int
2993 tcp_build_hdrs(tcp_t *tcp)
2994 {
2995 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2996 	conn_t		*connp = tcp->tcp_connp;
2997 	char		buf[TCP_MAX_HDR_LENGTH];
2998 	uint_t		buflen;
2999 	uint_t		ulplen = TCP_MIN_HEADER_LENGTH;
3000 	uint_t		extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
3001 	tcpha_t		*tcpha;
3002 	uint32_t	cksum;
3003 	int		error;
3004 
3005 	/*
3006 	 * We might be called after the connection is set up, and we might
3007 	 * have TS options already in the TCP header. Thus we  save any
3008 	 * existing tcp header.
3009 	 */
3010 	buflen = connp->conn_ht_ulp_len;
3011 	if (buflen != 0) {
3012 		bcopy(connp->conn_ht_ulp, buf, buflen);
3013 		extralen -= buflen - ulplen;
3014 		ulplen = buflen;
3015 	}
3016 
3017 	/* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
3018 	mutex_enter(&connp->conn_lock);
3019 	error = conn_build_hdr_template(connp, ulplen, extralen,
3020 	    &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
3021 	mutex_exit(&connp->conn_lock);
3022 	if (error != 0)
3023 		return (error);
3024 
3025 	/*
3026 	 * Any routing header/option has been massaged. The checksum difference
3027 	 * is stored in conn_sum for later use.
3028 	 */
3029 	tcpha = (tcpha_t *)connp->conn_ht_ulp;
3030 	tcp->tcp_tcpha = tcpha;
3031 
3032 	/* restore any old tcp header */
3033 	if (buflen != 0) {
3034 		bcopy(buf, connp->conn_ht_ulp, buflen);
3035 	} else {
3036 		tcpha->tha_sum = 0;
3037 		tcpha->tha_urp = 0;
3038 		tcpha->tha_ack = 0;
3039 		tcpha->tha_offset_and_reserved = (5 << 4);
3040 		tcpha->tha_lport = connp->conn_lport;
3041 		tcpha->tha_fport = connp->conn_fport;
3042 	}
3043 
3044 	/*
3045 	 * IP wants our header length in the checksum field to
3046 	 * allow it to perform a single pseudo-header+checksum
3047 	 * calculation on behalf of TCP.
3048 	 * Include the adjustment for a source route once IP_OPTIONS is set.
3049 	 */
3050 	cksum = sizeof (tcpha_t) + connp->conn_sum;
3051 	cksum = (cksum >> 16) + (cksum & 0xFFFF);
3052 	ASSERT(cksum < 0x10000);
3053 	tcpha->tha_sum = htons(cksum);
3054 
3055 	if (connp->conn_ipversion == IPV4_VERSION)
3056 		tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
3057 	else
3058 		tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
3059 
3060 	if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
3061 	    connp->conn_wroff) {
3062 		connp->conn_wroff = connp->conn_ht_iphc_allocated +
3063 		    tcps->tcps_wroff_xtra;
3064 		(void) proto_set_tx_wroff(connp->conn_rq, connp,
3065 		    connp->conn_wroff);
3066 	}
3067 	return (0);
3068 }
3069 
3070 /* Get callback routine passed to nd_load by tcp_param_register */
3071 /* ARGSUSED */
3072 static int
3073 tcp_param_get(queue_t *q, mblk_t *mp, caddr_t cp, cred_t *cr)
3074 {
3075 	tcpparam_t	*tcppa = (tcpparam_t *)cp;
3076 
3077 	(void) mi_mpprintf(mp, "%u", tcppa->tcp_param_val);
3078 	return (0);
3079 }
3080 
3081 /*
3082  * Walk through the param array specified registering each element with the
3083  * named dispatch handler.
3084  */
3085 static boolean_t
3086 tcp_param_register(IDP *ndp, tcpparam_t *tcppa, int cnt, tcp_stack_t *tcps)
3087 {
3088 	for (; cnt-- > 0; tcppa++) {
3089 		if (tcppa->tcp_param_name && tcppa->tcp_param_name[0]) {
3090 			if (!nd_load(ndp, tcppa->tcp_param_name,
3091 			    tcp_param_get, tcp_param_set,
3092 			    (caddr_t)tcppa)) {
3093 				nd_free(ndp);
3094 				return (B_FALSE);
3095 			}
3096 		}
3097 	}
3098 	tcps->tcps_wroff_xtra_param = kmem_zalloc(sizeof (tcpparam_t),
3099 	    KM_SLEEP);
3100 	bcopy(&lcl_tcp_wroff_xtra_param, tcps->tcps_wroff_xtra_param,
3101 	    sizeof (tcpparam_t));
3102 	if (!nd_load(ndp, tcps->tcps_wroff_xtra_param->tcp_param_name,
3103 	    tcp_param_get, tcp_param_set_aligned,
3104 	    (caddr_t)tcps->tcps_wroff_xtra_param)) {
3105 		nd_free(ndp);
3106 		return (B_FALSE);
3107 	}
3108 	if (!nd_load(ndp, "tcp_extra_priv_ports",
3109 	    tcp_extra_priv_ports_get, NULL, NULL)) {
3110 		nd_free(ndp);
3111 		return (B_FALSE);
3112 	}
3113 	if (!nd_load(ndp, "tcp_extra_priv_ports_add",
3114 	    NULL, tcp_extra_priv_ports_add, NULL)) {
3115 		nd_free(ndp);
3116 		return (B_FALSE);
3117 	}
3118 	if (!nd_load(ndp, "tcp_extra_priv_ports_del",
3119 	    NULL, tcp_extra_priv_ports_del, NULL)) {
3120 		nd_free(ndp);
3121 		return (B_FALSE);
3122 	}
3123 	if (!nd_load(ndp, "tcp_1948_phrase", NULL,
3124 	    tcp_1948_phrase_set, NULL)) {
3125 		nd_free(ndp);
3126 		return (B_FALSE);
3127 	}
3128 
3129 
3130 	if (!nd_load(ndp, "tcp_listener_limit_conf",
3131 	    tcp_listener_conf_get, NULL, NULL)) {
3132 		nd_free(ndp);
3133 		return (B_FALSE);
3134 	}
3135 	if (!nd_load(ndp, "tcp_listener_limit_conf_add",
3136 	    NULL, tcp_listener_conf_add, NULL)) {
3137 		nd_free(ndp);
3138 		return (B_FALSE);
3139 	}
3140 	if (!nd_load(ndp, "tcp_listener_limit_conf_del",
3141 	    NULL, tcp_listener_conf_del, NULL)) {
3142 		nd_free(ndp);
3143 		return (B_FALSE);
3144 	}
3145 
3146 	/*
3147 	 * Dummy ndd variables - only to convey obsolescence information
3148 	 * through printing of their name (no get or set routines)
3149 	 * XXX Remove in future releases ?
3150 	 */
3151 	if (!nd_load(ndp,
3152 	    "tcp_close_wait_interval(obsoleted - "
3153 	    "use tcp_time_wait_interval)", NULL, NULL, NULL)) {
3154 		nd_free(ndp);
3155 		return (B_FALSE);
3156 	}
3157 	return (B_TRUE);
3158 }
3159 
3160 /* ndd set routine for tcp_wroff_xtra. */
3161 /* ARGSUSED */
3162 static int
3163 tcp_param_set_aligned(queue_t *q, mblk_t *mp, char *value, caddr_t cp,
3164     cred_t *cr)
3165 {
3166 	long new_value;
3167 	tcpparam_t *tcppa = (tcpparam_t *)cp;
3168 
3169 	if (ddi_strtol(value, NULL, 10, &new_value) != 0 ||
3170 	    new_value < tcppa->tcp_param_min ||
3171 	    new_value > tcppa->tcp_param_max) {
3172 		return (EINVAL);
3173 	}
3174 	/*
3175 	 * Need to make sure new_value is a multiple of 4.  If it is not,
3176 	 * round it up.  For future 64 bit requirement, we actually make it
3177 	 * a multiple of 8.
3178 	 */
3179 	if (new_value & 0x7) {
3180 		new_value = (new_value & ~0x7) + 0x8;
3181 	}
3182 	tcppa->tcp_param_val = new_value;
3183 	return (0);
3184 }
3185 
3186 /* Set callback routine passed to nd_load by tcp_param_register */
3187 /* ARGSUSED */
3188 static int
3189 tcp_param_set(queue_t *q, mblk_t *mp, char *value, caddr_t cp, cred_t *cr)
3190 {
3191 	long	new_value;
3192 	tcpparam_t	*tcppa = (tcpparam_t *)cp;
3193 
3194 	if (ddi_strtol(value, NULL, 10, &new_value) != 0 ||
3195 	    new_value < tcppa->tcp_param_min ||
3196 	    new_value > tcppa->tcp_param_max) {
3197 		return (EINVAL);
3198 	}
3199 	tcppa->tcp_param_val = new_value;
3200 	return (0);
3201 }
3202 
3203 /*
3204  * tcp_rwnd_set() is called to adjust the receive window to a desired value.
3205  * We do not allow the receive window to shrink.  After setting rwnd,
3206  * set the flow control hiwat of the stream.
3207  *
3208  * This function is called in 2 cases:
3209  *
3210  * 1) Before data transfer begins, in tcp_input_listener() for accepting a
3211  *    connection (passive open) and in tcp_input_data() for active connect.
3212  *    This is called after tcp_mss_set() when the desired MSS value is known.
3213  *    This makes sure that our window size is a mutiple of the other side's
3214  *    MSS.
3215  * 2) Handling SO_RCVBUF option.
3216  *
3217  * It is ASSUMED that the requested size is a multiple of the current MSS.
3218  *
3219  * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
3220  * user requests so.
3221  */
3222 int
3223 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
3224 {
3225 	uint32_t	mss = tcp->tcp_mss;
3226 	uint32_t	old_max_rwnd;
3227 	uint32_t	max_transmittable_rwnd;
3228 	boolean_t	tcp_detached = TCP_IS_DETACHED(tcp);
3229 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3230 	conn_t		*connp = tcp->tcp_connp;
3231 
3232 	/*
3233 	 * Insist on a receive window that is at least
3234 	 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
3235 	 * funny TCP interactions of Nagle algorithm, SWS avoidance
3236 	 * and delayed acknowledgement.
3237 	 */
3238 	rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
3239 
3240 	if (tcp->tcp_fused) {
3241 		size_t sth_hiwat;
3242 		tcp_t *peer_tcp = tcp->tcp_loopback_peer;
3243 
3244 		ASSERT(peer_tcp != NULL);
3245 		sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
3246 		if (!tcp_detached) {
3247 			(void) proto_set_rx_hiwat(connp->conn_rq, connp,
3248 			    sth_hiwat);
3249 			tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
3250 		}
3251 
3252 		/* Caller could have changed tcp_rwnd; update tha_win */
3253 		if (tcp->tcp_tcpha != NULL) {
3254 			tcp->tcp_tcpha->tha_win =
3255 			    htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3256 		}
3257 		if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3258 			tcp->tcp_cwnd_max = rwnd;
3259 
3260 		/*
3261 		 * In the fusion case, the maxpsz stream head value of
3262 		 * our peer is set according to its send buffer size
3263 		 * and our receive buffer size; since the latter may
3264 		 * have changed we need to update the peer's maxpsz.
3265 		 */
3266 		(void) tcp_maxpsz_set(peer_tcp, B_TRUE);
3267 		return (sth_hiwat);
3268 	}
3269 
3270 	if (tcp_detached)
3271 		old_max_rwnd = tcp->tcp_rwnd;
3272 	else
3273 		old_max_rwnd = connp->conn_rcvbuf;
3274 
3275 
3276 	/*
3277 	 * If window size info has already been exchanged, TCP should not
3278 	 * shrink the window.  Shrinking window is doable if done carefully.
3279 	 * We may add that support later.  But so far there is not a real
3280 	 * need to do that.
3281 	 */
3282 	if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
3283 		/* MSS may have changed, do a round up again. */
3284 		rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
3285 	}
3286 
3287 	/*
3288 	 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3289 	 * can be applied even before the window scale option is decided.
3290 	 */
3291 	max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3292 	if (rwnd > max_transmittable_rwnd) {
3293 		rwnd = max_transmittable_rwnd -
3294 		    (max_transmittable_rwnd % mss);
3295 		if (rwnd < mss)
3296 			rwnd = max_transmittable_rwnd;
3297 		/*
3298 		 * If we're over the limit we may have to back down tcp_rwnd.
3299 		 * The increment below won't work for us. So we set all three
3300 		 * here and the increment below will have no effect.
3301 		 */
3302 		tcp->tcp_rwnd = old_max_rwnd = rwnd;
3303 	}
3304 	if (tcp->tcp_localnet) {
3305 		tcp->tcp_rack_abs_max =
3306 		    MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3307 	} else {
3308 		/*
3309 		 * For a remote host on a different subnet (through a router),
3310 		 * we ack every other packet to be conforming to RFC1122.
3311 		 * tcp_deferred_acks_max is default to 2.
3312 		 */
3313 		tcp->tcp_rack_abs_max =
3314 		    MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3315 	}
3316 	if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3317 		tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3318 	else
3319 		tcp->tcp_rack_cur_max = 0;
3320 	/*
3321 	 * Increment the current rwnd by the amount the maximum grew (we
3322 	 * can not overwrite it since we might be in the middle of a
3323 	 * connection.)
3324 	 */
3325 	tcp->tcp_rwnd += rwnd - old_max_rwnd;
3326 	connp->conn_rcvbuf = rwnd;
3327 
3328 	/* Are we already connected? */
3329 	if (tcp->tcp_tcpha != NULL) {
3330 		tcp->tcp_tcpha->tha_win =
3331 		    htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3332 	}
3333 
3334 	if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3335 		tcp->tcp_cwnd_max = rwnd;
3336 
3337 	if (tcp_detached)
3338 		return (rwnd);
3339 
3340 	tcp_set_recv_threshold(tcp, rwnd >> 3);
3341 
3342 	(void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3343 	return (rwnd);
3344 }
3345 
3346 int
3347 tcp_do_unbind(conn_t *connp)
3348 {
3349 	tcp_t *tcp = connp->conn_tcp;
3350 
3351 	switch (tcp->tcp_state) {
3352 	case TCPS_BOUND:
3353 	case TCPS_LISTEN:
3354 		break;
3355 	default:
3356 		return (-TOUTSTATE);
3357 	}
3358 
3359 	/*
3360 	 * Need to clean up all the eagers since after the unbind, segments
3361 	 * will no longer be delivered to this listener stream.
3362 	 */
3363 	mutex_enter(&tcp->tcp_eager_lock);
3364 	if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3365 		tcp_eager_cleanup(tcp, 0);
3366 	}
3367 	mutex_exit(&tcp->tcp_eager_lock);
3368 
3369 	/* Clean up the listener connection counter if necessary. */
3370 	if (tcp->tcp_listen_cnt != NULL)
3371 		TCP_DECR_LISTEN_CNT(tcp);
3372 	connp->conn_laddr_v6 = ipv6_all_zeros;
3373 	connp->conn_saddr_v6 = ipv6_all_zeros;
3374 	tcp_bind_hash_remove(tcp);
3375 	tcp->tcp_state = TCPS_IDLE;
3376 
3377 	ip_unbind(connp);
3378 	bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3379 
3380 	return (0);
3381 }
3382 
3383 /*
3384  * This runs at the tail end of accept processing on the squeue of the
3385  * new connection.
3386  */
3387 /* ARGSUSED */
3388 void
3389 tcp_accept_finish(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
3390 {
3391 	conn_t			*connp = (conn_t *)arg;
3392 	tcp_t			*tcp = connp->conn_tcp;
3393 	queue_t			*q = connp->conn_rq;
3394 	tcp_stack_t		*tcps = tcp->tcp_tcps;
3395 	/* socket options */
3396 	struct sock_proto_props	sopp;
3397 
3398 	/* We should just receive a single mblk that fits a T_discon_ind */
3399 	ASSERT(mp->b_cont == NULL);
3400 
3401 	/*
3402 	 * Drop the eager's ref on the listener, that was placed when
3403 	 * this eager began life in tcp_input_listener.
3404 	 */
3405 	CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
3406 	if (IPCL_IS_NONSTR(connp)) {
3407 		/* Safe to free conn_ind message */
3408 		freemsg(tcp->tcp_conn.tcp_eager_conn_ind);
3409 		tcp->tcp_conn.tcp_eager_conn_ind = NULL;
3410 	}
3411 
3412 	tcp->tcp_detached = B_FALSE;
3413 
3414 	if (tcp->tcp_state <= TCPS_BOUND || tcp->tcp_accept_error) {
3415 		/*
3416 		 * Someone blewoff the eager before we could finish
3417 		 * the accept.
3418 		 *
3419 		 * The only reason eager exists it because we put in
3420 		 * a ref on it when conn ind went up. We need to send
3421 		 * a disconnect indication up while the last reference
3422 		 * on the eager will be dropped by the squeue when we
3423 		 * return.
3424 		 */
3425 		ASSERT(tcp->tcp_listener == NULL);
3426 		if (tcp->tcp_issocket || tcp->tcp_send_discon_ind) {
3427 			if (IPCL_IS_NONSTR(connp)) {
3428 				ASSERT(tcp->tcp_issocket);
3429 				(*connp->conn_upcalls->su_disconnected)(
3430 				    connp->conn_upper_handle, tcp->tcp_connid,
3431 				    ECONNREFUSED);
3432 				freemsg(mp);
3433 			} else {
3434 				struct	T_discon_ind	*tdi;
3435 
3436 				(void) putnextctl1(q, M_FLUSH, FLUSHRW);
3437 				/*
3438 				 * Let us reuse the incoming mblk to avoid
3439 				 * memory allocation failure problems. We know
3440 				 * that the size of the incoming mblk i.e.
3441 				 * stroptions is greater than sizeof
3442 				 * T_discon_ind.
3443 				 */
3444 				ASSERT(DB_REF(mp) == 1);
3445 				ASSERT(MBLKSIZE(mp) >=
3446 				    sizeof (struct T_discon_ind));
3447 
3448 				DB_TYPE(mp) = M_PROTO;
3449 				((union T_primitives *)mp->b_rptr)->type =
3450 				    T_DISCON_IND;
3451 				tdi = (struct T_discon_ind *)mp->b_rptr;
3452 				if (tcp->tcp_issocket) {
3453 					tdi->DISCON_reason = ECONNREFUSED;
3454 					tdi->SEQ_number = 0;
3455 				} else {
3456 					tdi->DISCON_reason = ENOPROTOOPT;
3457 					tdi->SEQ_number =
3458 					    tcp->tcp_conn_req_seqnum;
3459 				}
3460 				mp->b_wptr = mp->b_rptr +
3461 				    sizeof (struct T_discon_ind);
3462 				putnext(q, mp);
3463 			}
3464 		}
3465 		tcp->tcp_hard_binding = B_FALSE;
3466 		return;
3467 	}
3468 
3469 	/*
3470 	 * This is the first time we run on the correct
3471 	 * queue after tcp_accept. So fix all the q parameters
3472 	 * here.
3473 	 */
3474 	sopp.sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3475 	sopp.sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3476 
3477 	sopp.sopp_rxhiwat = tcp->tcp_fused ?
3478 	    tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3479 	    connp->conn_rcvbuf;
3480 
3481 	/*
3482 	 * Determine what write offset value to use depending on SACK and
3483 	 * whether the endpoint is fused or not.
3484 	 */
3485 	if (tcp->tcp_fused) {
3486 		ASSERT(tcp->tcp_loopback);
3487 		ASSERT(tcp->tcp_loopback_peer != NULL);
3488 		/*
3489 		 * For fused tcp loopback, set the stream head's write
3490 		 * offset value to zero since we won't be needing any room
3491 		 * for TCP/IP headers.  This would also improve performance
3492 		 * since it would reduce the amount of work done by kmem.
3493 		 * Non-fused tcp loopback case is handled separately below.
3494 		 */
3495 		sopp.sopp_wroff = 0;
3496 		/*
3497 		 * Update the peer's transmit parameters according to
3498 		 * our recently calculated high water mark value.
3499 		 */
3500 		(void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3501 	} else if (tcp->tcp_snd_sack_ok) {
3502 		sopp.sopp_wroff = connp->conn_ht_iphc_allocated +
3503 		    (tcp->tcp_loopback ? 0 : tcps->tcps_wroff_xtra);
3504 	} else {
3505 		sopp.sopp_wroff = connp->conn_ht_iphc_len +
3506 		    (tcp->tcp_loopback ? 0 : tcps->tcps_wroff_xtra);
3507 	}
3508 
3509 	/*
3510 	 * If this is endpoint is handling SSL, then reserve extra
3511 	 * offset and space at the end.
3512 	 * Also have the stream head allocate SSL3_MAX_RECORD_LEN packets,
3513 	 * overriding the previous setting. The extra cost of signing and
3514 	 * encrypting multiple MSS-size records (12 of them with Ethernet),
3515 	 * instead of a single contiguous one by the stream head
3516 	 * largely outweighs the statistical reduction of ACKs, when
3517 	 * applicable. The peer will also save on decryption and verification
3518 	 * costs.
3519 	 */
3520 	if (tcp->tcp_kssl_ctx != NULL) {
3521 		sopp.sopp_wroff += SSL3_WROFFSET;
3522 
3523 		sopp.sopp_flags |= SOCKOPT_TAIL;
3524 		sopp.sopp_tail = SSL3_MAX_TAIL_LEN;
3525 
3526 		sopp.sopp_flags |= SOCKOPT_ZCOPY;
3527 		sopp.sopp_zcopyflag = ZCVMUNSAFE;
3528 
3529 		sopp.sopp_maxblk = SSL3_MAX_RECORD_LEN;
3530 	}
3531 
3532 	/* Send the options up */
3533 	if (IPCL_IS_NONSTR(connp)) {
3534 		if (sopp.sopp_flags & SOCKOPT_TAIL) {
3535 			ASSERT(tcp->tcp_kssl_ctx != NULL);
3536 			ASSERT(sopp.sopp_flags & SOCKOPT_ZCOPY);
3537 		}
3538 		if (tcp->tcp_loopback) {
3539 			sopp.sopp_flags |= SOCKOPT_LOOPBACK;
3540 			sopp.sopp_loopback = B_TRUE;
3541 		}
3542 		(*connp->conn_upcalls->su_set_proto_props)
3543 		    (connp->conn_upper_handle, &sopp);
3544 		freemsg(mp);
3545 	} else {
3546 		/*
3547 		 * Let us reuse the incoming mblk to avoid
3548 		 * memory allocation failure problems. We know
3549 		 * that the size of the incoming mblk is at least
3550 		 * stroptions
3551 		 */
3552 		struct stroptions *stropt;
3553 
3554 		ASSERT(DB_REF(mp) == 1);
3555 		ASSERT(MBLKSIZE(mp) >= sizeof (struct stroptions));
3556 
3557 		DB_TYPE(mp) = M_SETOPTS;
3558 		stropt = (struct stroptions *)mp->b_rptr;
3559 		mp->b_wptr = mp->b_rptr + sizeof (struct stroptions);
3560 		stropt = (struct stroptions *)mp->b_rptr;
3561 		stropt->so_flags = SO_HIWAT | SO_WROFF | SO_MAXBLK;
3562 		stropt->so_hiwat = sopp.sopp_rxhiwat;
3563 		stropt->so_wroff = sopp.sopp_wroff;
3564 		stropt->so_maxblk = sopp.sopp_maxblk;
3565 
3566 		if (sopp.sopp_flags & SOCKOPT_TAIL) {
3567 			ASSERT(tcp->tcp_kssl_ctx != NULL);
3568 
3569 			stropt->so_flags |= SO_TAIL | SO_COPYOPT;
3570 			stropt->so_tail = sopp.sopp_tail;
3571 			stropt->so_copyopt = sopp.sopp_zcopyflag;
3572 		}
3573 
3574 		/* Send the options up */
3575 		putnext(q, mp);
3576 	}
3577 
3578 	/*
3579 	 * Pass up any data and/or a fin that has been received.
3580 	 *
3581 	 * Adjust receive window in case it had decreased
3582 	 * (because there is data <=> tcp_rcv_list != NULL)
3583 	 * while the connection was detached. Note that
3584 	 * in case the eager was flow-controlled, w/o this
3585 	 * code, the rwnd may never open up again!
3586 	 */
3587 	if (tcp->tcp_rcv_list != NULL) {
3588 		if (IPCL_IS_NONSTR(connp)) {
3589 			mblk_t *mp;
3590 			int space_left;
3591 			int error;
3592 			boolean_t push = B_TRUE;
3593 
3594 			if (!tcp->tcp_fused && (*connp->conn_upcalls->su_recv)
3595 			    (connp->conn_upper_handle, NULL, 0, 0, &error,
3596 			    &push) >= 0) {
3597 				tcp->tcp_rwnd = connp->conn_rcvbuf;
3598 				if (tcp->tcp_state >= TCPS_ESTABLISHED &&
3599 				    tcp_rwnd_reopen(tcp) == TH_ACK_NEEDED) {
3600 					tcp_xmit_ctl(NULL,
3601 					    tcp, (tcp->tcp_swnd == 0) ?
3602 					    tcp->tcp_suna : tcp->tcp_snxt,
3603 					    tcp->tcp_rnxt, TH_ACK);
3604 				}
3605 			}
3606 			while ((mp = tcp->tcp_rcv_list) != NULL) {
3607 				push = B_TRUE;
3608 				tcp->tcp_rcv_list = mp->b_next;
3609 				mp->b_next = NULL;
3610 				space_left = (*connp->conn_upcalls->su_recv)
3611 				    (connp->conn_upper_handle, mp, msgdsize(mp),
3612 				    0, &error, &push);
3613 				if (space_left < 0) {
3614 					/*
3615 					 * We should never be in middle of a
3616 					 * fallback, the squeue guarantees that.
3617 					 */
3618 					ASSERT(error != EOPNOTSUPP);
3619 				}
3620 			}
3621 			tcp->tcp_rcv_last_head = NULL;
3622 			tcp->tcp_rcv_last_tail = NULL;
3623 			tcp->tcp_rcv_cnt = 0;
3624 		} else {
3625 			/* We drain directly in case of fused tcp loopback */
3626 
3627 			if (!tcp->tcp_fused && canputnext(q)) {
3628 				tcp->tcp_rwnd = connp->conn_rcvbuf;
3629 				if (tcp->tcp_state >= TCPS_ESTABLISHED &&
3630 				    tcp_rwnd_reopen(tcp) == TH_ACK_NEEDED) {
3631 					tcp_xmit_ctl(NULL,
3632 					    tcp, (tcp->tcp_swnd == 0) ?
3633 					    tcp->tcp_suna : tcp->tcp_snxt,
3634 					    tcp->tcp_rnxt, TH_ACK);
3635 				}
3636 			}
3637 
3638 			(void) tcp_rcv_drain(tcp);
3639 		}
3640 
3641 		/*
3642 		 * For fused tcp loopback, back-enable peer endpoint
3643 		 * if it's currently flow-controlled.
3644 		 */
3645 		if (tcp->tcp_fused) {
3646 			tcp_t *peer_tcp = tcp->tcp_loopback_peer;
3647 
3648 			ASSERT(peer_tcp != NULL);
3649 			ASSERT(peer_tcp->tcp_fused);
3650 
3651 			mutex_enter(&peer_tcp->tcp_non_sq_lock);
3652 			if (peer_tcp->tcp_flow_stopped) {
3653 				tcp_clrqfull(peer_tcp);
3654 				TCP_STAT(tcps, tcp_fusion_backenabled);
3655 			}
3656 			mutex_exit(&peer_tcp->tcp_non_sq_lock);
3657 		}
3658 	}
3659 	ASSERT(tcp->tcp_rcv_list == NULL || tcp->tcp_fused_sigurg);
3660 	if (tcp->tcp_fin_rcvd && !tcp->tcp_ordrel_done) {
3661 		tcp->tcp_ordrel_done = B_TRUE;
3662 		if (IPCL_IS_NONSTR(connp)) {
3663 			ASSERT(tcp->tcp_ordrel_mp == NULL);
3664 			(*connp->conn_upcalls->su_opctl)(
3665 			    connp->conn_upper_handle,
3666 			    SOCK_OPCTL_SHUT_RECV, 0);
3667 		} else {
3668 			mp = tcp->tcp_ordrel_mp;
3669 			tcp->tcp_ordrel_mp = NULL;
3670 			putnext(q, mp);
3671 		}
3672 	}
3673 	tcp->tcp_hard_binding = B_FALSE;
3674 
3675 	if (connp->conn_keepalive) {
3676 		tcp->tcp_ka_last_intrvl = 0;
3677 		tcp->tcp_ka_tid = TCP_TIMER(tcp, tcp_keepalive_timer,
3678 		    MSEC_TO_TICK(tcp->tcp_ka_interval));
3679 	}
3680 
3681 	/*
3682 	 * At this point, eager is fully established and will
3683 	 * have the following references -
3684 	 *
3685 	 * 2 references for connection to exist (1 for TCP and 1 for IP).
3686 	 * 1 reference for the squeue which will be dropped by the squeue as
3687 	 *	soon as this function returns.
3688 	 * There will be 1 additonal reference for being in classifier
3689 	 *	hash list provided something bad hasn't happened.
3690 	 */
3691 	ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
3692 	    (connp->conn_fanout == NULL && connp->conn_ref >= 3));
3693 }
3694 
3695 /*
3696  * Common to TPI and sockfs accept code.
3697  */
3698 /* ARGSUSED2 */
3699 int
3700 tcp_accept_common(conn_t *lconnp, conn_t *econnp, cred_t *cr)
3701 {
3702 	tcp_t *listener, *eager;
3703 	mblk_t *discon_mp;
3704 
3705 	listener = lconnp->conn_tcp;
3706 	ASSERT(listener->tcp_state == TCPS_LISTEN);
3707 	eager = econnp->conn_tcp;
3708 	ASSERT(eager->tcp_listener != NULL);
3709 
3710 	/*
3711 	 * Pre allocate the discon_ind mblk also. tcp_accept_finish will
3712 	 * use it if something failed.
3713 	 */
3714 	discon_mp = allocb(MAX(sizeof (struct T_discon_ind),
3715 	    sizeof (struct stroptions)), BPRI_HI);
3716 
3717 	if (discon_mp == NULL) {
3718 		return (-TPROTO);
3719 	}
3720 	eager->tcp_issocket = B_TRUE;
3721 
3722 	econnp->conn_zoneid = listener->tcp_connp->conn_zoneid;
3723 	econnp->conn_allzones = listener->tcp_connp->conn_allzones;
3724 	ASSERT(econnp->conn_netstack ==
3725 	    listener->tcp_connp->conn_netstack);
3726 	ASSERT(eager->tcp_tcps == listener->tcp_tcps);
3727 
3728 	/* Put the ref for IP */
3729 	CONN_INC_REF(econnp);
3730 
3731 	/*
3732 	 * We should have minimum of 3 references on the conn
3733 	 * at this point. One each for TCP and IP and one for
3734 	 * the T_conn_ind that was sent up when the 3-way handshake
3735 	 * completed. In the normal case we would also have another
3736 	 * reference (making a total of 4) for the conn being in the
3737 	 * classifier hash list. However the eager could have received
3738 	 * an RST subsequently and tcp_closei_local could have removed
3739 	 * the eager from the classifier hash list, hence we can't
3740 	 * assert that reference.
3741 	 */
3742 	ASSERT(econnp->conn_ref >= 3);
3743 
3744 	mutex_enter(&listener->tcp_eager_lock);
3745 	if (listener->tcp_eager_prev_q0->tcp_conn_def_q0) {
3746 
3747 		tcp_t *tail;
3748 		tcp_t *tcp;
3749 		mblk_t *mp1;
3750 
3751 		tcp = listener->tcp_eager_prev_q0;
3752 		/*
3753 		 * listener->tcp_eager_prev_q0 points to the TAIL of the
3754 		 * deferred T_conn_ind queue. We need to get to the head
3755 		 * of the queue in order to send up T_conn_ind the same
3756 		 * order as how the 3WHS is completed.
3757 		 */
3758 		while (tcp != listener) {
3759 			if (!tcp->tcp_eager_prev_q0->tcp_conn_def_q0 &&
3760 			    !tcp->tcp_kssl_pending)
3761 				break;
3762 			else
3763 				tcp = tcp->tcp_eager_prev_q0;
3764 		}
3765 		/* None of the pending eagers can be sent up now */
3766 		if (tcp == listener)
3767 			goto no_more_eagers;
3768 
3769 		mp1 = tcp->tcp_conn.tcp_eager_conn_ind;
3770 		tcp->tcp_conn.tcp_eager_conn_ind = NULL;
3771 		/* Move from q0 to q */
3772 		ASSERT(listener->tcp_conn_req_cnt_q0 > 0);
3773 		listener->tcp_conn_req_cnt_q0--;
3774 		listener->tcp_conn_req_cnt_q++;
3775 		tcp->tcp_eager_next_q0->tcp_eager_prev_q0 =
3776 		    tcp->tcp_eager_prev_q0;
3777 		tcp->tcp_eager_prev_q0->tcp_eager_next_q0 =
3778 		    tcp->tcp_eager_next_q0;
3779 		tcp->tcp_eager_prev_q0 = NULL;
3780 		tcp->tcp_eager_next_q0 = NULL;
3781 		tcp->tcp_conn_def_q0 = B_FALSE;
3782 
3783 		/* Make sure the tcp isn't in the list of droppables */
3784 		ASSERT(tcp->tcp_eager_next_drop_q0 == NULL &&
3785 		    tcp->tcp_eager_prev_drop_q0 == NULL);
3786 
3787 		/*
3788 		 * Insert at end of the queue because sockfs sends
3789 		 * down T_CONN_RES in chronological order. Leaving
3790 		 * the older conn indications at front of the queue
3791 		 * helps reducing search time.
3792 		 */
3793 		tail = listener->tcp_eager_last_q;
3794 		if (tail != NULL) {
3795 			tail->tcp_eager_next_q = tcp;
3796 		} else {
3797 			listener->tcp_eager_next_q = tcp;
3798 		}
3799 		listener->tcp_eager_last_q = tcp;
3800 		tcp->tcp_eager_next_q = NULL;
3801 
3802 		/* Need to get inside the listener perimeter */
3803 		CONN_INC_REF(listener->tcp_connp);
3804 		SQUEUE_ENTER_ONE(listener->tcp_connp->conn_sqp, mp1,
3805 		    tcp_send_pending, listener->tcp_connp, NULL, SQ_FILL,
3806 		    SQTAG_TCP_SEND_PENDING);
3807 	}
3808 no_more_eagers:
3809 	tcp_eager_unlink(eager);
3810 	mutex_exit(&listener->tcp_eager_lock);
3811 
3812 	/*
3813 	 * At this point, the eager is detached from the listener
3814 	 * but we still have an extra refs on eager (apart from the
3815 	 * usual tcp references). The ref was placed in tcp_input_data
3816 	 * before sending the conn_ind in tcp_send_conn_ind.
3817 	 * The ref will be dropped in tcp_accept_finish().
3818 	 */
3819 	SQUEUE_ENTER_ONE(econnp->conn_sqp, discon_mp, tcp_accept_finish,
3820 	    econnp, NULL, SQ_NODRAIN, SQTAG_TCP_ACCEPT_FINISH_Q0);
3821 	return (0);
3822 }
3823 
3824 /*
3825  * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3826  */
3827 boolean_t
3828 tcp_zcopy_check(tcp_t *tcp)
3829 {
3830 	conn_t		*connp = tcp->tcp_connp;
3831 	ip_xmit_attr_t	*ixa = connp->conn_ixa;
3832 	boolean_t	zc_enabled = B_FALSE;
3833 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3834 
3835 	if (do_tcpzcopy == 2)
3836 		zc_enabled = B_TRUE;
3837 	else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3838 		zc_enabled = B_TRUE;
3839 
3840 	tcp->tcp_snd_zcopy_on = zc_enabled;
3841 	if (!TCP_IS_DETACHED(tcp)) {
3842 		if (zc_enabled) {
3843 			ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3844 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
3845 			    ZCVMSAFE);
3846 			TCP_STAT(tcps, tcp_zcopy_on);
3847 		} else {
3848 			ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3849 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
3850 			    ZCVMUNSAFE);
3851 			TCP_STAT(tcps, tcp_zcopy_off);
3852 		}
3853 	}
3854 	return (zc_enabled);
3855 }
3856 
3857 /*
3858  * Backoff from a zero-copy message by copying data to a new allocated
3859  * message and freeing the original desballoca'ed segmapped message.
3860  *
3861  * This function is called by following two callers:
3862  * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3863  *    the origial desballoca'ed message and notify sockfs. This is in re-
3864  *    transmit state.
3865  * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3866  *    to be copied to new message.
3867  */
3868 mblk_t *
3869 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3870 {
3871 	mblk_t		*nbp;
3872 	mblk_t		*head = NULL;
3873 	mblk_t		*tail = NULL;
3874 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3875 
3876 	ASSERT(bp != NULL);
3877 	while (bp != NULL) {
3878 		if (IS_VMLOANED_MBLK(bp)) {
3879 			TCP_STAT(tcps, tcp_zcopy_backoff);
3880 			if ((nbp = copyb(bp)) == NULL) {
3881 				tcp->tcp_xmit_zc_clean = B_FALSE;
3882 				if (tail != NULL)
3883 					tail->b_cont = bp;
3884 				return ((head == NULL) ? bp : head);
3885 			}
3886 
3887 			if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3888 				if (fix_xmitlist)
3889 					tcp_zcopy_notify(tcp);
3890 				else
3891 					nbp->b_datap->db_struioflag |=
3892 					    STRUIO_ZCNOTIFY;
3893 			}
3894 			nbp->b_cont = bp->b_cont;
3895 
3896 			/*
3897 			 * Copy saved information and adjust tcp_xmit_tail
3898 			 * if needed.
3899 			 */
3900 			if (fix_xmitlist) {
3901 				nbp->b_prev = bp->b_prev;
3902 				nbp->b_next = bp->b_next;
3903 
3904 				if (tcp->tcp_xmit_tail == bp)
3905 					tcp->tcp_xmit_tail = nbp;
3906 			}
3907 
3908 			/* Free the original message. */
3909 			bp->b_prev = NULL;
3910 			bp->b_next = NULL;
3911 			freeb(bp);
3912 
3913 			bp = nbp;
3914 		}
3915 
3916 		if (head == NULL) {
3917 			head = bp;
3918 		}
3919 		if (tail == NULL) {
3920 			tail = bp;
3921 		} else {
3922 			tail->b_cont = bp;
3923 			tail = bp;
3924 		}
3925 
3926 		/* Move forward. */
3927 		bp = bp->b_cont;
3928 	}
3929 
3930 	if (fix_xmitlist) {
3931 		tcp->tcp_xmit_last = tail;
3932 		tcp->tcp_xmit_zc_clean = B_TRUE;
3933 	}
3934 
3935 	return (head);
3936 }
3937 
3938 void
3939 tcp_zcopy_notify(tcp_t *tcp)
3940 {
3941 	struct stdata	*stp;
3942 	conn_t		*connp;
3943 
3944 	if (tcp->tcp_detached)
3945 		return;
3946 	connp = tcp->tcp_connp;
3947 	if (IPCL_IS_NONSTR(connp)) {
3948 		(*connp->conn_upcalls->su_zcopy_notify)
3949 		    (connp->conn_upper_handle);
3950 		return;
3951 	}
3952 	stp = STREAM(connp->conn_rq);
3953 	mutex_enter(&stp->sd_lock);
3954 	stp->sd_flag |= STZCNOTIFY;
3955 	cv_broadcast(&stp->sd_zcopy_wait);
3956 	mutex_exit(&stp->sd_lock);
3957 }
3958 
3959 /*
3960  * Update the TCP connection according to change of LSO capability.
3961  */
3962 static void
3963 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3964 {
3965 	/*
3966 	 * We check against IPv4 header length to preserve the old behavior
3967 	 * of only enabling LSO when there are no IP options.
3968 	 * But this restriction might not be necessary at all. Before removing
3969 	 * it, need to verify how LSO is handled for source routing case, with
3970 	 * which IP does software checksum.
3971 	 *
3972 	 * For IPv6, whenever any extension header is needed, LSO is supressed.
3973 	 */
3974 	if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3975 	    IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3976 		return;
3977 
3978 	/*
3979 	 * Either the LSO capability newly became usable, or it has changed.
3980 	 */
3981 	if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3982 		ill_lso_capab_t	*lsoc = &ixa->ixa_lso_capab;
3983 
3984 		ASSERT(lsoc->ill_lso_max > 0);
3985 		tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lsoc->ill_lso_max);
3986 
3987 		DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3988 		    boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3989 
3990 		/*
3991 		 * If LSO to be enabled, notify the STREAM header with larger
3992 		 * data block.
3993 		 */
3994 		if (!tcp->tcp_lso)
3995 			tcp->tcp_maxpsz_multiplier = 0;
3996 
3997 		tcp->tcp_lso = B_TRUE;
3998 		TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3999 	} else { /* LSO capability is not usable any more. */
4000 		DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
4001 		    boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
4002 
4003 		/*
4004 		 * If LSO to be disabled, notify the STREAM header with smaller
4005 		 * data block. And need to restore fragsize to PMTU.
4006 		 */
4007 		if (tcp->tcp_lso) {
4008 			tcp->tcp_maxpsz_multiplier =
4009 			    tcp->tcp_tcps->tcps_maxpsz_multiplier;
4010 			ixa->ixa_fragsize = ixa->ixa_pmtu;
4011 			tcp->tcp_lso = B_FALSE;
4012 			TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
4013 		}
4014 	}
4015 
4016 	(void) tcp_maxpsz_set(tcp, B_TRUE);
4017 }
4018 
4019 /*
4020  * Update the TCP connection according to change of ZEROCOPY capability.
4021  */
4022 static void
4023 tcp_update_zcopy(tcp_t *tcp)
4024 {
4025 	conn_t		*connp = tcp->tcp_connp;
4026 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4027 
4028 	if (tcp->tcp_snd_zcopy_on) {
4029 		tcp->tcp_snd_zcopy_on = B_FALSE;
4030 		if (!TCP_IS_DETACHED(tcp)) {
4031 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
4032 			    ZCVMUNSAFE);
4033 			TCP_STAT(tcps, tcp_zcopy_off);
4034 		}
4035 	} else {
4036 		tcp->tcp_snd_zcopy_on = B_TRUE;
4037 		if (!TCP_IS_DETACHED(tcp)) {
4038 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
4039 			    ZCVMSAFE);
4040 			TCP_STAT(tcps, tcp_zcopy_on);
4041 		}
4042 	}
4043 }
4044 
4045 /*
4046  * Notify function registered with ip_xmit_attr_t. It's called in the squeue
4047  * so it's safe to update the TCP connection.
4048  */
4049 /* ARGSUSED1 */
4050 static void
4051 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
4052     ixa_notify_arg_t narg)
4053 {
4054 	tcp_t		*tcp = (tcp_t *)arg;
4055 	conn_t		*connp = tcp->tcp_connp;
4056 
4057 	switch (ntype) {
4058 	case IXAN_LSO:
4059 		tcp_update_lso(tcp, connp->conn_ixa);
4060 		break;
4061 	case IXAN_PMTU:
4062 		tcp_update_pmtu(tcp, B_FALSE);
4063 		break;
4064 	case IXAN_ZCOPY:
4065 		tcp_update_zcopy(tcp);
4066 		break;
4067 	default:
4068 		break;
4069 	}
4070 }
4071 
4072 /*
4073  * The TCP write service routine should never be called...
4074  */
4075 /* ARGSUSED */
4076 static void
4077 tcp_wsrv(queue_t *q)
4078 {
4079 	tcp_stack_t	*tcps = Q_TO_TCP(q)->tcp_tcps;
4080 
4081 	TCP_STAT(tcps, tcp_wsrv_called);
4082 }
4083 
4084 /*
4085  * Hash list lookup routine for tcp_t structures.
4086  * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
4087  */
4088 tcp_t *
4089 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
4090 {
4091 	tf_t	*tf;
4092 	tcp_t	*tcp;
4093 
4094 	tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
4095 	mutex_enter(&tf->tf_lock);
4096 	for (tcp = tf->tf_tcp; tcp != NULL;
4097 	    tcp = tcp->tcp_acceptor_hash) {
4098 		if (tcp->tcp_acceptor_id == id) {
4099 			CONN_INC_REF(tcp->tcp_connp);
4100 			mutex_exit(&tf->tf_lock);
4101 			return (tcp);
4102 		}
4103 	}
4104 	mutex_exit(&tf->tf_lock);
4105 	return (NULL);
4106 }
4107 
4108 /*
4109  * Hash list insertion routine for tcp_t structures.
4110  */
4111 void
4112 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
4113 {
4114 	tf_t	*tf;
4115 	tcp_t	**tcpp;
4116 	tcp_t	*tcpnext;
4117 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4118 
4119 	tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
4120 
4121 	if (tcp->tcp_ptpahn != NULL)
4122 		tcp_acceptor_hash_remove(tcp);
4123 	tcpp = &tf->tf_tcp;
4124 	mutex_enter(&tf->tf_lock);
4125 	tcpnext = tcpp[0];
4126 	if (tcpnext)
4127 		tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
4128 	tcp->tcp_acceptor_hash = tcpnext;
4129 	tcp->tcp_ptpahn = tcpp;
4130 	tcpp[0] = tcp;
4131 	tcp->tcp_acceptor_lockp = &tf->tf_lock;	/* For tcp_*_hash_remove */
4132 	mutex_exit(&tf->tf_lock);
4133 }
4134 
4135 /*
4136  * Hash list removal routine for tcp_t structures.
4137  */
4138 void
4139 tcp_acceptor_hash_remove(tcp_t *tcp)
4140 {
4141 	tcp_t	*tcpnext;
4142 	kmutex_t *lockp;
4143 
4144 	/*
4145 	 * Extract the lock pointer in case there are concurrent
4146 	 * hash_remove's for this instance.
4147 	 */
4148 	lockp = tcp->tcp_acceptor_lockp;
4149 
4150 	if (tcp->tcp_ptpahn == NULL)
4151 		return;
4152 
4153 	ASSERT(lockp != NULL);
4154 	mutex_enter(lockp);
4155 	if (tcp->tcp_ptpahn) {
4156 		tcpnext = tcp->tcp_acceptor_hash;
4157 		if (tcpnext) {
4158 			tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
4159 			tcp->tcp_acceptor_hash = NULL;
4160 		}
4161 		*tcp->tcp_ptpahn = tcpnext;
4162 		tcp->tcp_ptpahn = NULL;
4163 	}
4164 	mutex_exit(lockp);
4165 	tcp->tcp_acceptor_lockp = NULL;
4166 }
4167 
4168 /*
4169  * Type three generator adapted from the random() function in 4.4 BSD:
4170  */
4171 
4172 /*
4173  * Copyright (c) 1983, 1993
4174  *	The Regents of the University of California.  All rights reserved.
4175  *
4176  * Redistribution and use in source and binary forms, with or without
4177  * modification, are permitted provided that the following conditions
4178  * are met:
4179  * 1. Redistributions of source code must retain the above copyright
4180  *    notice, this list of conditions and the following disclaimer.
4181  * 2. Redistributions in binary form must reproduce the above copyright
4182  *    notice, this list of conditions and the following disclaimer in the
4183  *    documentation and/or other materials provided with the distribution.
4184  * 3. All advertising materials mentioning features or use of this software
4185  *    must display the following acknowledgement:
4186  *	This product includes software developed by the University of
4187  *	California, Berkeley and its contributors.
4188  * 4. Neither the name of the University nor the names of its contributors
4189  *    may be used to endorse or promote products derived from this software
4190  *    without specific prior written permission.
4191  *
4192  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
4193  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
4194  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
4195  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
4196  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
4197  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
4198  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
4199  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
4200  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
4201  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
4202  * SUCH DAMAGE.
4203  */
4204 
4205 /* Type 3 -- x**31 + x**3 + 1 */
4206 #define	DEG_3		31
4207 #define	SEP_3		3
4208 
4209 
4210 /* Protected by tcp_random_lock */
4211 static int tcp_randtbl[DEG_3 + 1];
4212 
4213 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
4214 static int *tcp_random_rptr = &tcp_randtbl[1];
4215 
4216 static int *tcp_random_state = &tcp_randtbl[1];
4217 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
4218 
4219 kmutex_t tcp_random_lock;
4220 
4221 void
4222 tcp_random_init(void)
4223 {
4224 	int i;
4225 	hrtime_t hrt;
4226 	time_t wallclock;
4227 	uint64_t result;
4228 
4229 	/*
4230 	 * Use high-res timer and current time for seed.  Gethrtime() returns
4231 	 * a longlong, which may contain resolution down to nanoseconds.
4232 	 * The current time will either be a 32-bit or a 64-bit quantity.
4233 	 * XOR the two together in a 64-bit result variable.
4234 	 * Convert the result to a 32-bit value by multiplying the high-order
4235 	 * 32-bits by the low-order 32-bits.
4236 	 */
4237 
4238 	hrt = gethrtime();
4239 	(void) drv_getparm(TIME, &wallclock);
4240 	result = (uint64_t)wallclock ^ (uint64_t)hrt;
4241 	mutex_enter(&tcp_random_lock);
4242 	tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
4243 	    (result & 0xffffffff);
4244 
4245 	for (i = 1; i < DEG_3; i++)
4246 		tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
4247 		    + 12345;
4248 	tcp_random_fptr = &tcp_random_state[SEP_3];
4249 	tcp_random_rptr = &tcp_random_state[0];
4250 	mutex_exit(&tcp_random_lock);
4251 	for (i = 0; i < 10 * DEG_3; i++)
4252 		(void) tcp_random();
4253 }
4254 
4255 /*
4256  * tcp_random: Return a random number in the range [1 - (128K + 1)].
4257  * This range is selected to be approximately centered on TCP_ISS / 2,
4258  * and easy to compute. We get this value by generating a 32-bit random
4259  * number, selecting out the high-order 17 bits, and then adding one so
4260  * that we never return zero.
4261  */
4262 int
4263 tcp_random(void)
4264 {
4265 	int i;
4266 
4267 	mutex_enter(&tcp_random_lock);
4268 	*tcp_random_fptr += *tcp_random_rptr;
4269 
4270 	/*
4271 	 * The high-order bits are more random than the low-order bits,
4272 	 * so we select out the high-order 17 bits and add one so that
4273 	 * we never return zero.
4274 	 */
4275 	i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
4276 	if (++tcp_random_fptr >= tcp_random_end_ptr) {
4277 		tcp_random_fptr = tcp_random_state;
4278 		++tcp_random_rptr;
4279 	} else if (++tcp_random_rptr >= tcp_random_end_ptr)
4280 		tcp_random_rptr = tcp_random_state;
4281 
4282 	mutex_exit(&tcp_random_lock);
4283 	return (i);
4284 }
4285 
4286 /*
4287  * Split this function out so that if the secret changes, I'm okay.
4288  *
4289  * Initialize the tcp_iss_cookie and tcp_iss_key.
4290  */
4291 
4292 #define	PASSWD_SIZE 16  /* MUST be multiple of 4 */
4293 
4294 static void
4295 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
4296 {
4297 	struct {
4298 		int32_t current_time;
4299 		uint32_t randnum;
4300 		uint16_t pad;
4301 		uint8_t ether[6];
4302 		uint8_t passwd[PASSWD_SIZE];
4303 	} tcp_iss_cookie;
4304 	time_t t;
4305 
4306 	/*
4307 	 * Start with the current absolute time.
4308 	 */
4309 	(void) drv_getparm(TIME, &t);
4310 	tcp_iss_cookie.current_time = t;
4311 
4312 	/*
4313 	 * XXX - Need a more random number per RFC 1750, not this crap.
4314 	 * OTOH, if what follows is pretty random, then I'm in better shape.
4315 	 */
4316 	tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
4317 	tcp_iss_cookie.pad = 0x365c;  /* Picked from HMAC pad values. */
4318 
4319 	/*
4320 	 * The cpu_type_info is pretty non-random.  Ugggh.  It does serve
4321 	 * as a good template.
4322 	 */
4323 	bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
4324 	    min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
4325 
4326 	/*
4327 	 * The pass-phrase.  Normally this is supplied by user-called NDD.
4328 	 */
4329 	bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
4330 
4331 	/*
4332 	 * See 4010593 if this section becomes a problem again,
4333 	 * but the local ethernet address is useful here.
4334 	 */
4335 	(void) localetheraddr(NULL,
4336 	    (struct ether_addr *)&tcp_iss_cookie.ether);
4337 
4338 	/*
4339 	 * Hash 'em all together.  The MD5Final is called per-connection.
4340 	 */
4341 	mutex_enter(&tcps->tcps_iss_key_lock);
4342 	MD5Init(&tcps->tcps_iss_key);
4343 	MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
4344 	    sizeof (tcp_iss_cookie));
4345 	mutex_exit(&tcps->tcps_iss_key_lock);
4346 }
4347 
4348 /*
4349  * Set the RFC 1948 pass phrase
4350  */
4351 /* ARGSUSED */
4352 static int
4353 tcp_1948_phrase_set(queue_t *q, mblk_t *mp, char *value, caddr_t cp,
4354     cred_t *cr)
4355 {
4356 	tcp_stack_t	*tcps = Q_TO_TCP(q)->tcp_tcps;
4357 
4358 	/*
4359 	 * Basically, value contains a new pass phrase.  Pass it along!
4360 	 */
4361 	tcp_iss_key_init((uint8_t *)value, strlen(value), tcps);
4362 	return (0);
4363 }
4364 
4365 /* ARGSUSED */
4366 static int
4367 tcp_sack_info_constructor(void *buf, void *cdrarg, int kmflags)
4368 {
4369 	bzero(buf, sizeof (tcp_sack_info_t));
4370 	return (0);
4371 }
4372 
4373 /*
4374  * Called by IP when IP is loaded into the kernel
4375  */
4376 void
4377 tcp_ddi_g_init(void)
4378 {
4379 	tcp_timercache = kmem_cache_create("tcp_timercache",
4380 	    sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
4381 	    NULL, NULL, NULL, NULL, NULL, 0);
4382 
4383 	tcp_sack_info_cache = kmem_cache_create("tcp_sack_info_cache",
4384 	    sizeof (tcp_sack_info_t), 0,
4385 	    tcp_sack_info_constructor, NULL, NULL, NULL, NULL, 0);
4386 
4387 	mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
4388 
4389 	/* Initialize the random number generator */
4390 	tcp_random_init();
4391 
4392 	/* A single callback independently of how many netstacks we have */
4393 	ip_squeue_init(tcp_squeue_add);
4394 
4395 	tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
4396 
4397 	tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
4398 
4399 	/*
4400 	 * We want to be informed each time a stack is created or
4401 	 * destroyed in the kernel, so we can maintain the
4402 	 * set of tcp_stack_t's.
4403 	 */
4404 	netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
4405 
4406 	mutex_enter(&cpu_lock);
4407 	register_cpu_setup_func(tcp_cpu_update, NULL);
4408 	mutex_exit(&cpu_lock);
4409 }
4410 
4411 
4412 #define	INET_NAME	"ip"
4413 
4414 /*
4415  * Initialize the TCP stack instance.
4416  */
4417 static void *
4418 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
4419 {
4420 	tcp_stack_t	*tcps;
4421 	tcpparam_t	*pa;
4422 	int		i;
4423 	int		error = 0;
4424 	major_t		major;
4425 
4426 	tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
4427 	tcps->tcps_netstack = ns;
4428 
4429 	/* Initialize locks */
4430 	mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
4431 	mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
4432 
4433 	tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
4434 	tcps->tcps_g_epriv_ports[0] = 2049;
4435 	tcps->tcps_g_epriv_ports[1] = 4045;
4436 	tcps->tcps_min_anonpriv_port = 512;
4437 
4438 	tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
4439 	    TCP_BIND_FANOUT_SIZE, KM_SLEEP);
4440 	tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
4441 	    TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
4442 
4443 	for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
4444 		mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
4445 		    MUTEX_DEFAULT, NULL);
4446 	}
4447 
4448 	for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
4449 		mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
4450 		    MUTEX_DEFAULT, NULL);
4451 	}
4452 
4453 	/* TCP's IPsec code calls the packet dropper. */
4454 	ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
4455 
4456 	pa = (tcpparam_t *)kmem_alloc(sizeof (lcl_tcp_param_arr), KM_SLEEP);
4457 	tcps->tcps_params = pa;
4458 	bcopy(lcl_tcp_param_arr, tcps->tcps_params, sizeof (lcl_tcp_param_arr));
4459 
4460 	(void) tcp_param_register(&tcps->tcps_g_nd, tcps->tcps_params,
4461 	    A_CNT(lcl_tcp_param_arr), tcps);
4462 
4463 	/*
4464 	 * Note: To really walk the device tree you need the devinfo
4465 	 * pointer to your device which is only available after probe/attach.
4466 	 * The following is safe only because it uses ddi_root_node()
4467 	 */
4468 	tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
4469 	    tcp_opt_obj.odb_opt_arr_cnt);
4470 
4471 	/*
4472 	 * Initialize RFC 1948 secret values.  This will probably be reset once
4473 	 * by the boot scripts.
4474 	 *
4475 	 * Use NULL name, as the name is caught by the new lockstats.
4476 	 *
4477 	 * Initialize with some random, non-guessable string, like the global
4478 	 * T_INFO_ACK.
4479 	 */
4480 
4481 	tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
4482 	    sizeof (tcp_g_t_info_ack), tcps);
4483 
4484 	tcps->tcps_kstat = tcp_kstat2_init(stackid);
4485 	tcps->tcps_mibkp = tcp_kstat_init(stackid);
4486 
4487 	major = mod_name_to_major(INET_NAME);
4488 	error = ldi_ident_from_major(major, &tcps->tcps_ldi_ident);
4489 	ASSERT(error == 0);
4490 	tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
4491 	ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
4492 	cv_init(&tcps->tcps_ixa_cleanup_cv, NULL, CV_DEFAULT, NULL);
4493 	mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
4494 
4495 	mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
4496 	tcps->tcps_reclaim = B_FALSE;
4497 	tcps->tcps_reclaim_tid = 0;
4498 	tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
4499 
4500 	/*
4501 	 * ncpus is the current number of CPUs, which can be bigger than
4502 	 * boot_ncpus.  But we don't want to use ncpus to allocate all the
4503 	 * tcp_stats_cpu_t at system boot up time since it will be 1.  While
4504 	 * we handle adding CPU in tcp_cpu_update(), it will be slow if
4505 	 * there are many CPUs as we will be adding them 1 by 1.
4506 	 *
4507 	 * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
4508 	 * are not freed until the stack is going away.  So there is no need
4509 	 * to grab a lock to access the per CPU tcps_sc[x] pointer.
4510 	 */
4511 	tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
4512 	tcps->tcps_sc = kmem_zalloc(max_ncpus  * sizeof (tcp_stats_cpu_t *),
4513 	    KM_SLEEP);
4514 	for (i = 0; i < tcps->tcps_sc_cnt; i++) {
4515 		tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
4516 		    KM_SLEEP);
4517 	}
4518 
4519 	mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
4520 	list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
4521 	    offsetof(tcp_listener_t, tl_link));
4522 
4523 	return (tcps);
4524 }
4525 
4526 /*
4527  * Called when the IP module is about to be unloaded.
4528  */
4529 void
4530 tcp_ddi_g_destroy(void)
4531 {
4532 	mutex_enter(&cpu_lock);
4533 	unregister_cpu_setup_func(tcp_cpu_update, NULL);
4534 	mutex_exit(&cpu_lock);
4535 
4536 	tcp_g_kstat_fini(tcp_g_kstat);
4537 	tcp_g_kstat = NULL;
4538 	bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
4539 
4540 	mutex_destroy(&tcp_random_lock);
4541 
4542 	kmem_cache_destroy(tcp_timercache);
4543 	kmem_cache_destroy(tcp_sack_info_cache);
4544 
4545 	netstack_unregister(NS_TCP);
4546 }
4547 
4548 /*
4549  * Free the TCP stack instance.
4550  */
4551 static void
4552 tcp_stack_fini(netstackid_t stackid, void *arg)
4553 {
4554 	tcp_stack_t *tcps = (tcp_stack_t *)arg;
4555 	int i;
4556 
4557 	freeb(tcps->tcps_ixa_cleanup_mp);
4558 	tcps->tcps_ixa_cleanup_mp = NULL;
4559 	cv_destroy(&tcps->tcps_ixa_cleanup_cv);
4560 	mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
4561 
4562 	/*
4563 	 * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
4564 	 * the timer.
4565 	 */
4566 	mutex_enter(&tcps->tcps_reclaim_lock);
4567 	tcps->tcps_reclaim = B_FALSE;
4568 	mutex_exit(&tcps->tcps_reclaim_lock);
4569 	if (tcps->tcps_reclaim_tid != 0)
4570 		(void) untimeout(tcps->tcps_reclaim_tid);
4571 	mutex_destroy(&tcps->tcps_reclaim_lock);
4572 
4573 	tcp_listener_conf_cleanup(tcps);
4574 
4575 	for (i = 0; i < tcps->tcps_sc_cnt; i++)
4576 		kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
4577 	kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
4578 
4579 	nd_free(&tcps->tcps_g_nd);
4580 	kmem_free(tcps->tcps_params, sizeof (lcl_tcp_param_arr));
4581 	tcps->tcps_params = NULL;
4582 	kmem_free(tcps->tcps_wroff_xtra_param, sizeof (tcpparam_t));
4583 	tcps->tcps_wroff_xtra_param = NULL;
4584 
4585 	for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
4586 		ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
4587 		mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
4588 	}
4589 
4590 	for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
4591 		ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
4592 		mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
4593 	}
4594 
4595 	kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
4596 	tcps->tcps_bind_fanout = NULL;
4597 
4598 	kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
4599 	    TCP_ACCEPTOR_FANOUT_SIZE);
4600 	tcps->tcps_acceptor_fanout = NULL;
4601 
4602 	mutex_destroy(&tcps->tcps_iss_key_lock);
4603 	mutex_destroy(&tcps->tcps_epriv_port_lock);
4604 
4605 	ip_drop_unregister(&tcps->tcps_dropper);
4606 
4607 	tcp_kstat2_fini(stackid, tcps->tcps_kstat);
4608 	tcps->tcps_kstat = NULL;
4609 
4610 	tcp_kstat_fini(stackid, tcps->tcps_mibkp);
4611 	tcps->tcps_mibkp = NULL;
4612 
4613 	ldi_ident_release(tcps->tcps_ldi_ident);
4614 	kmem_free(tcps, sizeof (*tcps));
4615 }
4616 
4617 /*
4618  * Generate ISS, taking into account NDD changes may happen halfway through.
4619  * (If the iss is not zero, set it.)
4620  */
4621 
4622 static void
4623 tcp_iss_init(tcp_t *tcp)
4624 {
4625 	MD5_CTX context;
4626 	struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
4627 	uint32_t answer[4];
4628 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4629 	conn_t		*connp = tcp->tcp_connp;
4630 
4631 	tcps->tcps_iss_incr_extra += (ISS_INCR >> 1);
4632 	tcp->tcp_iss = tcps->tcps_iss_incr_extra;
4633 	switch (tcps->tcps_strong_iss) {
4634 	case 2:
4635 		mutex_enter(&tcps->tcps_iss_key_lock);
4636 		context = tcps->tcps_iss_key;
4637 		mutex_exit(&tcps->tcps_iss_key_lock);
4638 		arg.ports = connp->conn_ports;
4639 		arg.src = connp->conn_laddr_v6;
4640 		arg.dst = connp->conn_faddr_v6;
4641 		MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
4642 		MD5Final((uchar_t *)answer, &context);
4643 		tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
4644 		/*
4645 		 * Now that we've hashed into a unique per-connection sequence
4646 		 * space, add a random increment per strong_iss == 1.  So I
4647 		 * guess we'll have to...
4648 		 */
4649 		/* FALLTHRU */
4650 	case 1:
4651 		tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
4652 		break;
4653 	default:
4654 		tcp->tcp_iss += (uint32_t)gethrestime_sec() * ISS_INCR;
4655 		break;
4656 	}
4657 	tcp->tcp_valid_bits = TCP_ISS_VALID;
4658 	tcp->tcp_fss = tcp->tcp_iss - 1;
4659 	tcp->tcp_suna = tcp->tcp_iss;
4660 	tcp->tcp_snxt = tcp->tcp_iss + 1;
4661 	tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
4662 	tcp->tcp_csuna = tcp->tcp_snxt;
4663 }
4664 
4665 /*
4666  * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
4667  * on the specified backing STREAMS q. Note, the caller may make the
4668  * decision to call based on the tcp_t.tcp_flow_stopped value which
4669  * when check outside the q's lock is only an advisory check ...
4670  */
4671 void
4672 tcp_setqfull(tcp_t *tcp)
4673 {
4674 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4675 	conn_t	*connp = tcp->tcp_connp;
4676 
4677 	if (tcp->tcp_closed)
4678 		return;
4679 
4680 	conn_setqfull(connp, &tcp->tcp_flow_stopped);
4681 	if (tcp->tcp_flow_stopped)
4682 		TCP_STAT(tcps, tcp_flwctl_on);
4683 }
4684 
4685 void
4686 tcp_clrqfull(tcp_t *tcp)
4687 {
4688 	conn_t  *connp = tcp->tcp_connp;
4689 
4690 	if (tcp->tcp_closed)
4691 		return;
4692 	conn_clrqfull(connp, &tcp->tcp_flow_stopped);
4693 }
4694 
4695 static int
4696 tcp_squeue_switch(int val)
4697 {
4698 	int rval = SQ_FILL;
4699 
4700 	switch (val) {
4701 	case 1:
4702 		rval = SQ_NODRAIN;
4703 		break;
4704 	case 2:
4705 		rval = SQ_PROCESS;
4706 		break;
4707 	default:
4708 		break;
4709 	}
4710 	return (rval);
4711 }
4712 
4713 /*
4714  * This is called once for each squeue - globally for all stack
4715  * instances.
4716  */
4717 static void
4718 tcp_squeue_add(squeue_t *sqp)
4719 {
4720 	tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4721 	    sizeof (tcp_squeue_priv_t), KM_SLEEP);
4722 
4723 	*squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4724 	/* Kick start the periodic TIME WAIT collector. */
4725 	tcp_time_wait->tcp_time_wait_tid =
4726 	    timeout_generic(CALLOUT_NORMAL, tcp_time_wait_collector, sqp,
4727 	    (hrtime_t)10 * NANOSEC, CALLOUT_TCP_RESOLUTION,
4728 	    CALLOUT_FLAG_ROUNDUP);
4729 	if (tcp_free_list_max_cnt == 0) {
4730 		int tcp_ncpus = ((boot_max_ncpus == -1) ?
4731 		    max_ncpus : boot_max_ncpus);
4732 
4733 		/*
4734 		 * Limit number of entries to 1% of availble memory / tcp_ncpus
4735 		 */
4736 		tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4737 		    (tcp_ncpus * sizeof (tcp_t) * 100);
4738 	}
4739 	tcp_time_wait->tcp_free_list_cnt = 0;
4740 }
4741 /*
4742  * Return unix error is tli error is TSYSERR, otherwise return a negative
4743  * tli error.
4744  */
4745 int
4746 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4747     boolean_t bind_to_req_port_only)
4748 {
4749 	int error;
4750 	tcp_t *tcp = connp->conn_tcp;
4751 
4752 	if (tcp->tcp_state >= TCPS_BOUND) {
4753 		if (connp->conn_debug) {
4754 			(void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4755 			    "tcp_bind: bad state, %d", tcp->tcp_state);
4756 		}
4757 		return (-TOUTSTATE);
4758 	}
4759 
4760 	error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4761 	if (error != 0)
4762 		return (error);
4763 
4764 	ASSERT(tcp->tcp_state == TCPS_BOUND);
4765 	tcp->tcp_conn_req_max = 0;
4766 	return (0);
4767 }
4768 
4769 /*
4770  * If the return value from this function is positive, it's a UNIX error.
4771  * Otherwise, if it's negative, then the absolute value is a TLI error.
4772  * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4773  */
4774 int
4775 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4776     cred_t *cr, pid_t pid)
4777 {
4778 	tcp_t		*tcp = connp->conn_tcp;
4779 	sin_t		*sin = (sin_t *)sa;
4780 	sin6_t		*sin6 = (sin6_t *)sa;
4781 	ipaddr_t	*dstaddrp;
4782 	in_port_t	dstport;
4783 	uint_t		srcid;
4784 	int		error;
4785 	uint32_t	mss;
4786 	mblk_t		*syn_mp;
4787 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4788 	int32_t		oldstate;
4789 	ip_xmit_attr_t	*ixa = connp->conn_ixa;
4790 
4791 	oldstate = tcp->tcp_state;
4792 
4793 	switch (len) {
4794 	default:
4795 		/*
4796 		 * Should never happen
4797 		 */
4798 		return (EINVAL);
4799 
4800 	case sizeof (sin_t):
4801 		sin = (sin_t *)sa;
4802 		if (sin->sin_port == 0) {
4803 			return (-TBADADDR);
4804 		}
4805 		if (connp->conn_ipv6_v6only) {
4806 			return (EAFNOSUPPORT);
4807 		}
4808 		break;
4809 
4810 	case sizeof (sin6_t):
4811 		sin6 = (sin6_t *)sa;
4812 		if (sin6->sin6_port == 0) {
4813 			return (-TBADADDR);
4814 		}
4815 		break;
4816 	}
4817 	/*
4818 	 * If we're connecting to an IPv4-mapped IPv6 address, we need to
4819 	 * make sure that the conn_ipversion is IPV4_VERSION.  We
4820 	 * need to this before we call tcp_bindi() so that the port lookup
4821 	 * code will look for ports in the correct port space (IPv4 and
4822 	 * IPv6 have separate port spaces).
4823 	 */
4824 	if (connp->conn_family == AF_INET6 &&
4825 	    connp->conn_ipversion == IPV6_VERSION &&
4826 	    IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4827 		if (connp->conn_ipv6_v6only)
4828 			return (EADDRNOTAVAIL);
4829 
4830 		connp->conn_ipversion = IPV4_VERSION;
4831 	}
4832 
4833 	switch (tcp->tcp_state) {
4834 	case TCPS_LISTEN:
4835 		/*
4836 		 * Listening sockets are not allowed to issue connect().
4837 		 */
4838 		if (IPCL_IS_NONSTR(connp))
4839 			return (EOPNOTSUPP);
4840 		/* FALLTHRU */
4841 	case TCPS_IDLE:
4842 		/*
4843 		 * We support quick connect, refer to comments in
4844 		 * tcp_connect_*()
4845 		 */
4846 		/* FALLTHRU */
4847 	case TCPS_BOUND:
4848 		break;
4849 	default:
4850 		return (-TOUTSTATE);
4851 	}
4852 
4853 	/*
4854 	 * We update our cred/cpid based on the caller of connect
4855 	 */
4856 	if (connp->conn_cred != cr) {
4857 		crhold(cr);
4858 		crfree(connp->conn_cred);
4859 		connp->conn_cred = cr;
4860 	}
4861 	connp->conn_cpid = pid;
4862 
4863 	/* Cache things in the ixa without any refhold */
4864 	ixa->ixa_cred = cr;
4865 	ixa->ixa_cpid = pid;
4866 	if (is_system_labeled()) {
4867 		/* We need to restart with a label based on the cred */
4868 		ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4869 	}
4870 
4871 	if (connp->conn_family == AF_INET6) {
4872 		if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4873 			error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4874 			    sin6->sin6_port, sin6->sin6_flowinfo,
4875 			    sin6->__sin6_src_id, sin6->sin6_scope_id);
4876 		} else {
4877 			/*
4878 			 * Destination adress is mapped IPv6 address.
4879 			 * Source bound address should be unspecified or
4880 			 * IPv6 mapped address as well.
4881 			 */
4882 			if (!IN6_IS_ADDR_UNSPECIFIED(
4883 			    &connp->conn_bound_addr_v6) &&
4884 			    !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4885 				return (EADDRNOTAVAIL);
4886 			}
4887 			dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4888 			dstport = sin6->sin6_port;
4889 			srcid = sin6->__sin6_src_id;
4890 			error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4891 			    srcid);
4892 		}
4893 	} else {
4894 		dstaddrp = &sin->sin_addr.s_addr;
4895 		dstport = sin->sin_port;
4896 		srcid = 0;
4897 		error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4898 	}
4899 
4900 	if (error != 0)
4901 		goto connect_failed;
4902 
4903 	CL_INET_CONNECT(connp, B_TRUE, error);
4904 	if (error != 0)
4905 		goto connect_failed;
4906 
4907 	/* connect succeeded */
4908 	TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4909 	tcp->tcp_active_open = 1;
4910 
4911 	/*
4912 	 * tcp_set_destination() does not adjust for TCP/IP header length.
4913 	 */
4914 	mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4915 
4916 	/*
4917 	 * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4918 	 * to the nearest MSS.
4919 	 *
4920 	 * We do the round up here because we need to get the interface MTU
4921 	 * first before we can do the round up.
4922 	 */
4923 	tcp->tcp_rwnd = connp->conn_rcvbuf;
4924 	tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4925 	    tcps->tcps_recv_hiwat_minmss * mss);
4926 	connp->conn_rcvbuf = tcp->tcp_rwnd;
4927 	tcp_set_ws_value(tcp);
4928 	tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4929 	if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4930 		tcp->tcp_snd_ws_ok = B_TRUE;
4931 
4932 	/*
4933 	 * Set tcp_snd_ts_ok to true
4934 	 * so that tcp_xmit_mp will
4935 	 * include the timestamp
4936 	 * option in the SYN segment.
4937 	 */
4938 	if (tcps->tcps_tstamp_always ||
4939 	    (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4940 		tcp->tcp_snd_ts_ok = B_TRUE;
4941 	}
4942 
4943 	/*
4944 	 * tcp_snd_sack_ok can be set in
4945 	 * tcp_set_destination() if the sack metric
4946 	 * is set.  So check it here also.
4947 	 */
4948 	if (tcps->tcps_sack_permitted == 2 ||
4949 	    tcp->tcp_snd_sack_ok) {
4950 		if (tcp->tcp_sack_info == NULL) {
4951 			tcp->tcp_sack_info = kmem_cache_alloc(
4952 			    tcp_sack_info_cache, KM_SLEEP);
4953 		}
4954 		tcp->tcp_snd_sack_ok = B_TRUE;
4955 	}
4956 
4957 	/*
4958 	 * Should we use ECN?  Note that the current
4959 	 * default value (SunOS 5.9) of tcp_ecn_permitted
4960 	 * is 1.  The reason for doing this is that there
4961 	 * are equipments out there that will drop ECN
4962 	 * enabled IP packets.  Setting it to 1 avoids
4963 	 * compatibility problems.
4964 	 */
4965 	if (tcps->tcps_ecn_permitted == 2)
4966 		tcp->tcp_ecn_ok = B_TRUE;
4967 
4968 	TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4969 	syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4970 	    tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4971 	if (syn_mp != NULL) {
4972 		/*
4973 		 * We must bump the generation before sending the syn
4974 		 * to ensure that we use the right generation in case
4975 		 * this thread issues a "connected" up call.
4976 		 */
4977 		SOCK_CONNID_BUMP(tcp->tcp_connid);
4978 		tcp_send_data(tcp, syn_mp);
4979 	}
4980 
4981 	if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4982 		tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4983 	return (0);
4984 
4985 connect_failed:
4986 	connp->conn_faddr_v6 = ipv6_all_zeros;
4987 	connp->conn_fport = 0;
4988 	tcp->tcp_state = oldstate;
4989 	if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4990 		tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4991 	return (error);
4992 }
4993 
4994 int
4995 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4996     int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4997 {
4998 	tcp_t		*tcp = connp->conn_tcp;
4999 	int		error = 0;
5000 	tcp_stack_t	*tcps = tcp->tcp_tcps;
5001 
5002 	/* All Solaris components should pass a cred for this operation. */
5003 	ASSERT(cr != NULL);
5004 
5005 	if (tcp->tcp_state >= TCPS_BOUND) {
5006 		if ((tcp->tcp_state == TCPS_BOUND ||
5007 		    tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
5008 			/*
5009 			 * Handle listen() increasing backlog.
5010 			 * This is more "liberal" then what the TPI spec
5011 			 * requires but is needed to avoid a t_unbind
5012 			 * when handling listen() since the port number
5013 			 * might be "stolen" between the unbind and bind.
5014 			 */
5015 			goto do_listen;
5016 		}
5017 		if (connp->conn_debug) {
5018 			(void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
5019 			    "tcp_listen: bad state, %d", tcp->tcp_state);
5020 		}
5021 		return (-TOUTSTATE);
5022 	} else {
5023 		if (sa == NULL) {
5024 			sin6_t	addr;
5025 			sin_t *sin;
5026 			sin6_t *sin6;
5027 
5028 			ASSERT(IPCL_IS_NONSTR(connp));
5029 			/* Do an implicit bind: Request for a generic port. */
5030 			if (connp->conn_family == AF_INET) {
5031 				len = sizeof (sin_t);
5032 				sin = (sin_t *)&addr;
5033 				*sin = sin_null;
5034 				sin->sin_family = AF_INET;
5035 			} else {
5036 				ASSERT(connp->conn_family == AF_INET6);
5037 				len = sizeof (sin6_t);
5038 				sin6 = (sin6_t *)&addr;
5039 				*sin6 = sin6_null;
5040 				sin6->sin6_family = AF_INET6;
5041 			}
5042 			sa = (struct sockaddr *)&addr;
5043 		}
5044 
5045 		error = tcp_bind_check(connp, sa, len, cr,
5046 		    bind_to_req_port_only);
5047 		if (error)
5048 			return (error);
5049 		/* Fall through and do the fanout insertion */
5050 	}
5051 
5052 do_listen:
5053 	ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
5054 	tcp->tcp_conn_req_max = backlog;
5055 	if (tcp->tcp_conn_req_max) {
5056 		if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
5057 			tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
5058 		if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
5059 			tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
5060 		/*
5061 		 * If this is a listener, do not reset the eager list
5062 		 * and other stuffs.  Note that we don't check if the
5063 		 * existing eager list meets the new tcp_conn_req_max
5064 		 * requirement.
5065 		 */
5066 		if (tcp->tcp_state != TCPS_LISTEN) {
5067 			tcp->tcp_state = TCPS_LISTEN;
5068 			/* Initialize the chain. Don't need the eager_lock */
5069 			tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
5070 			tcp->tcp_eager_next_drop_q0 = tcp;
5071 			tcp->tcp_eager_prev_drop_q0 = tcp;
5072 			tcp->tcp_second_ctimer_threshold =
5073 			    tcps->tcps_ip_abort_linterval;
5074 		}
5075 	}
5076 
5077 	/*
5078 	 * We need to make sure that the conn_recv is set to a non-null
5079 	 * value before we insert the conn into the classifier table.
5080 	 * This is to avoid a race with an incoming packet which does an
5081 	 * ipcl_classify().
5082 	 * We initially set it to tcp_input_listener_unbound to try to
5083 	 * pick a good squeue for the listener when the first SYN arrives.
5084 	 * tcp_input_listener_unbound sets it to tcp_input_listener on that
5085 	 * first SYN.
5086 	 */
5087 	connp->conn_recv = tcp_input_listener_unbound;
5088 
5089 	/* Insert the listener in the classifier table */
5090 	error = ip_laddr_fanout_insert(connp);
5091 	if (error != 0) {
5092 		/* Undo the bind - release the port number */
5093 		tcp->tcp_state = TCPS_IDLE;
5094 		connp->conn_bound_addr_v6 = ipv6_all_zeros;
5095 
5096 		connp->conn_laddr_v6 = ipv6_all_zeros;
5097 		connp->conn_saddr_v6 = ipv6_all_zeros;
5098 		connp->conn_ports = 0;
5099 
5100 		if (connp->conn_anon_port) {
5101 			zone_t		*zone;
5102 
5103 			zone = crgetzone(cr);
5104 			connp->conn_anon_port = B_FALSE;
5105 			(void) tsol_mlp_anon(zone, connp->conn_mlp_type,
5106 			    connp->conn_proto, connp->conn_lport, B_FALSE);
5107 		}
5108 		connp->conn_mlp_type = mlptSingle;
5109 
5110 		tcp_bind_hash_remove(tcp);
5111 		return (error);
5112 	} else {
5113 		/*
5114 		 * If there is a connection limit, allocate and initialize
5115 		 * the counter struct.  Note that since listen can be called
5116 		 * multiple times, the struct may have been allready allocated.
5117 		 */
5118 		if (!list_is_empty(&tcps->tcps_listener_conf) &&
5119 		    tcp->tcp_listen_cnt == NULL) {
5120 			tcp_listen_cnt_t *tlc;
5121 			uint32_t ratio;
5122 
5123 			ratio = tcp_find_listener_conf(tcps,
5124 			    ntohs(connp->conn_lport));
5125 			if (ratio != 0) {
5126 				uint32_t mem_ratio, tot_buf;
5127 
5128 				tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
5129 				    KM_SLEEP);
5130 				/*
5131 				 * Calculate the connection limit based on
5132 				 * the configured ratio and maxusers.  Maxusers
5133 				 * are calculated based on memory size,
5134 				 * ~ 1 user per MB.  Note that the conn_rcvbuf
5135 				 * and conn_sndbuf may change after a
5136 				 * connection is accepted.  So what we have
5137 				 * is only an approximation.
5138 				 */
5139 				if ((tot_buf = connp->conn_rcvbuf +
5140 				    connp->conn_sndbuf) < MB) {
5141 					mem_ratio = MB / tot_buf;
5142 					tlc->tlc_max = maxusers / ratio *
5143 					    mem_ratio;
5144 				} else {
5145 					mem_ratio = tot_buf / MB;
5146 					tlc->tlc_max = maxusers / ratio /
5147 					    mem_ratio;
5148 				}
5149 				/* At least we should allow two connections! */
5150 				if (tlc->tlc_max <= tcp_min_conn_listener)
5151 					tlc->tlc_max = tcp_min_conn_listener;
5152 				tlc->tlc_cnt = 1;
5153 				tlc->tlc_drop = 0;
5154 				tcp->tcp_listen_cnt = tlc;
5155 			}
5156 		}
5157 	}
5158 	return (error);
5159 }
5160