xref: /illumos-gate/usr/src/uts/common/inet/tcp/tcp.c (revision 861fa149)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25  * Copyright (c) 2013, 2017 by Delphix. All rights reserved.
26  * Copyright 2014, OmniTI Computer Consulting, Inc. All rights reserved.
27  * Copyright 2020 Joyent, Inc.
28  * Copyright 2022 Oxide Computer Company
29  */
30 /* Copyright (c) 1990 Mentat Inc. */
31 
32 #include <sys/types.h>
33 #include <sys/stream.h>
34 #include <sys/strsun.h>
35 #include <sys/strsubr.h>
36 #include <sys/stropts.h>
37 #include <sys/strlog.h>
38 #define	_SUN_TPI_VERSION 2
39 #include <sys/tihdr.h>
40 #include <sys/timod.h>
41 #include <sys/ddi.h>
42 #include <sys/sunddi.h>
43 #include <sys/suntpi.h>
44 #include <sys/xti_inet.h>
45 #include <sys/cmn_err.h>
46 #include <sys/debug.h>
47 #include <sys/sdt.h>
48 #include <sys/vtrace.h>
49 #include <sys/kmem.h>
50 #include <sys/ethernet.h>
51 #include <sys/cpuvar.h>
52 #include <sys/dlpi.h>
53 #include <sys/pattr.h>
54 #include <sys/policy.h>
55 #include <sys/priv.h>
56 #include <sys/zone.h>
57 #include <sys/sunldi.h>
58 
59 #include <sys/errno.h>
60 #include <sys/signal.h>
61 #include <sys/socket.h>
62 #include <sys/socketvar.h>
63 #include <sys/sockio.h>
64 #include <sys/isa_defs.h>
65 #include <sys/md5.h>
66 #include <sys/random.h>
67 #include <sys/uio.h>
68 #include <sys/systm.h>
69 #include <netinet/in.h>
70 #include <netinet/tcp.h>
71 #include <netinet/ip6.h>
72 #include <netinet/icmp6.h>
73 #include <net/if.h>
74 #include <net/route.h>
75 #include <inet/ipsec_impl.h>
76 
77 #include <inet/common.h>
78 #include <inet/cc.h>
79 #include <inet/ip.h>
80 #include <inet/ip_impl.h>
81 #include <inet/ip6.h>
82 #include <inet/ip_ndp.h>
83 #include <inet/proto_set.h>
84 #include <inet/mib2.h>
85 #include <inet/optcom.h>
86 #include <inet/snmpcom.h>
87 #include <inet/kstatcom.h>
88 #include <inet/tcp.h>
89 #include <inet/tcp_impl.h>
90 #include <inet/tcp_cluster.h>
91 #include <inet/udp_impl.h>
92 #include <net/pfkeyv2.h>
93 #include <inet/ipdrop.h>
94 
95 #include <inet/ipclassifier.h>
96 #include <inet/ip_ire.h>
97 #include <inet/ip_ftable.h>
98 #include <inet/ip_if.h>
99 #include <inet/ipp_common.h>
100 #include <inet/ip_rts.h>
101 #include <inet/ip_netinfo.h>
102 #include <sys/squeue_impl.h>
103 #include <sys/squeue.h>
104 #include <sys/tsol/label.h>
105 #include <sys/tsol/tnet.h>
106 #include <rpc/pmap_prot.h>
107 #include <sys/callo.h>
108 
109 /*
110  * TCP Notes: aka FireEngine Phase I (PSARC 2002/433)
111  *
112  * (Read the detailed design doc in PSARC case directory)
113  *
114  * The entire tcp state is contained in tcp_t and conn_t structure
115  * which are allocated in tandem using ipcl_conn_create() and passing
116  * IPCL_TCPCONN as a flag. We use 'conn_ref' and 'conn_lock' to protect
117  * the references on the tcp_t. The tcp_t structure is never compressed
118  * and packets always land on the correct TCP perimeter from the time
119  * eager is created till the time tcp_t dies (as such the old mentat
120  * TCP global queue is not used for detached state and no IPSEC checking
121  * is required). The global queue is still allocated to send out resets
122  * for connection which have no listeners and IP directly calls
123  * tcp_xmit_listeners_reset() which does any policy check.
124  *
125  * Protection and Synchronisation mechanism:
126  *
127  * The tcp data structure does not use any kind of lock for protecting
128  * its state but instead uses 'squeues' for mutual exclusion from various
129  * read and write side threads. To access a tcp member, the thread should
130  * always be behind squeue (via squeue_enter with flags as SQ_FILL, SQ_PROCESS,
131  * or SQ_NODRAIN). Since the squeues allow a direct function call, caller
132  * can pass any tcp function having prototype of edesc_t as argument
133  * (different from traditional STREAMs model where packets come in only
134  * designated entry points). The list of functions that can be directly
135  * called via squeue are listed before the usual function prototype.
136  *
137  * Referencing:
138  *
139  * TCP is MT-Hot and we use a reference based scheme to make sure that the
140  * tcp structure doesn't disappear when its needed. When the application
141  * creates an outgoing connection or accepts an incoming connection, we
142  * start out with 2 references on 'conn_ref'. One for TCP and one for IP.
143  * The IP reference is just a symbolic reference since ip_tcpclose()
144  * looks at tcp structure after tcp_close_output() returns which could
145  * have dropped the last TCP reference. So as long as the connection is
146  * in attached state i.e. !TCP_IS_DETACHED, we have 2 references on the
147  * conn_t. The classifier puts its own reference when the connection is
148  * inserted in listen or connected hash. Anytime a thread needs to enter
149  * the tcp connection perimeter, it retrieves the conn/tcp from q->ptr
150  * on write side or by doing a classify on read side and then puts a
151  * reference on the conn before doing squeue_enter/tryenter/fill. For
152  * read side, the classifier itself puts the reference under fanout lock
153  * to make sure that tcp can't disappear before it gets processed. The
154  * squeue will drop this reference automatically so the called function
155  * doesn't have to do a DEC_REF.
156  *
157  * Opening a new connection:
158  *
159  * The outgoing connection open is pretty simple. tcp_open() does the
160  * work in creating the conn/tcp structure and initializing it. The
161  * squeue assignment is done based on the CPU the application
162  * is running on. So for outbound connections, processing is always done
163  * on application CPU which might be different from the incoming CPU
164  * being interrupted by the NIC. An optimal way would be to figure out
165  * the NIC <-> CPU binding at listen time, and assign the outgoing
166  * connection to the squeue attached to the CPU that will be interrupted
167  * for incoming packets (we know the NIC based on the bind IP address).
168  * This might seem like a problem if more data is going out but the
169  * fact is that in most cases the transmit is ACK driven transmit where
170  * the outgoing data normally sits on TCP's xmit queue waiting to be
171  * transmitted.
172  *
173  * Accepting a connection:
174  *
175  * This is a more interesting case because of various races involved in
176  * establishing a eager in its own perimeter. Read the meta comment on
177  * top of tcp_input_listener(). But briefly, the squeue is picked by
178  * ip_fanout based on the ring or the sender (if loopback).
179  *
180  * Closing a connection:
181  *
182  * The close is fairly straight forward. tcp_close() calls tcp_close_output()
183  * via squeue to do the close and mark the tcp as detached if the connection
184  * was in state TCPS_ESTABLISHED or greater. In the later case, TCP keep its
185  * reference but tcp_close() drop IP's reference always. So if tcp was
186  * not killed, it is sitting in time_wait list with 2 reference - 1 for TCP
187  * and 1 because it is in classifier's connected hash. This is the condition
188  * we use to determine that its OK to clean up the tcp outside of squeue
189  * when time wait expires (check the ref under fanout and conn_lock and
190  * if it is 2, remove it from fanout hash and kill it).
191  *
192  * Although close just drops the necessary references and marks the
193  * tcp_detached state, tcp_close needs to know the tcp_detached has been
194  * set (under squeue) before letting the STREAM go away (because a
195  * inbound packet might attempt to go up the STREAM while the close
196  * has happened and tcp_detached is not set). So a special lock and
197  * flag is used along with a condition variable (tcp_closelock, tcp_closed,
198  * and tcp_closecv) to signal tcp_close that tcp_close_out() has marked
199  * tcp_detached.
200  *
201  * Special provisions and fast paths:
202  *
203  * We make special provisions for sockfs by marking tcp_issocket
204  * whenever we have only sockfs on top of TCP. This allows us to skip
205  * putting the tcp in acceptor hash since a sockfs listener can never
206  * become acceptor and also avoid allocating a tcp_t for acceptor STREAM
207  * since eager has already been allocated and the accept now happens
208  * on acceptor STREAM. There is a big blob of comment on top of
209  * tcp_input_listener explaining the new accept. When socket is POP'd,
210  * sockfs sends us an ioctl to mark the fact and we go back to old
211  * behaviour. Once tcp_issocket is unset, its never set for the
212  * life of that connection.
213  *
214  * IPsec notes :
215  *
216  * Since a packet is always executed on the correct TCP perimeter
217  * all IPsec processing is defered to IP including checking new
218  * connections and setting IPSEC policies for new connection. The
219  * only exception is tcp_xmit_listeners_reset() which is called
220  * directly from IP and needs to policy check to see if TH_RST
221  * can be sent out.
222  */
223 
224 /*
225  * Values for squeue switch:
226  * 1: SQ_NODRAIN
227  * 2: SQ_PROCESS
228  * 3: SQ_FILL
229  */
230 int tcp_squeue_wput = 2;	/* /etc/systems */
231 int tcp_squeue_flag;
232 
233 /*
234  * To prevent memory hog, limit the number of entries in tcp_free_list
235  * to 1% of available memory / number of cpus
236  */
237 uint_t tcp_free_list_max_cnt = 0;
238 
239 #define	TIDUSZ	4096	/* transport interface data unit size */
240 
241 /*
242  * Size of acceptor hash list.  It has to be a power of 2 for hashing.
243  */
244 #define	TCP_ACCEPTOR_FANOUT_SIZE		512
245 
246 #ifdef	_ILP32
247 #define	TCP_ACCEPTOR_HASH(accid)					\
248 		(((uint_t)(accid) >> 8) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
249 #else
250 #define	TCP_ACCEPTOR_HASH(accid)					\
251 		((uint_t)(accid) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
252 #endif	/* _ILP32 */
253 
254 /*
255  * Minimum number of connections which can be created per listener.  Used
256  * when the listener connection count is in effect.
257  */
258 static uint32_t tcp_min_conn_listener = 2;
259 
260 uint32_t tcp_early_abort = 30;
261 
262 /* TCP Timer control structure */
263 typedef struct tcpt_s {
264 	pfv_t	tcpt_pfv;	/* The routine we are to call */
265 	tcp_t	*tcpt_tcp;	/* The parameter we are to pass in */
266 } tcpt_t;
267 
268 /*
269  * Functions called directly via squeue having a prototype of edesc_t.
270  */
271 void		tcp_input_data(void *arg, mblk_t *mp, void *arg2,
272     ip_recv_attr_t *ira);
273 static void	tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2,
274     ip_recv_attr_t *dummy);
275 
276 
277 /* Prototype for TCP functions */
278 static void	tcp_random_init(void);
279 int		tcp_random(void);
280 static int	tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp,
281 		    in_port_t dstport, uint_t srcid);
282 static int	tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp,
283 		    in_port_t dstport, uint32_t flowinfo,
284 		    uint_t srcid, uint32_t scope_id);
285 static void	tcp_iss_init(tcp_t *tcp);
286 static void	tcp_reinit(tcp_t *tcp);
287 static void	tcp_reinit_values(tcp_t *tcp);
288 
289 static int	tcp_wsrv(queue_t *q);
290 static void	tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa);
291 static void	tcp_update_zcopy(tcp_t *tcp);
292 static void	tcp_notify(void *, ip_xmit_attr_t *, ixa_notify_type_t,
293     ixa_notify_arg_t);
294 static void	*tcp_stack_init(netstackid_t stackid, netstack_t *ns);
295 static void	tcp_stack_fini(netstackid_t stackid, void *arg);
296 
297 static int	tcp_squeue_switch(int);
298 
299 static int	tcp_open(queue_t *, dev_t *, int, int, cred_t *, boolean_t);
300 static int	tcp_openv4(queue_t *, dev_t *, int, int, cred_t *);
301 static int	tcp_openv6(queue_t *, dev_t *, int, int, cred_t *);
302 
303 static void	tcp_squeue_add(squeue_t *);
304 
305 struct module_info tcp_rinfo =  {
306 	TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, TCP_RECV_HIWATER, TCP_RECV_LOWATER
307 };
308 
309 static struct module_info tcp_winfo =  {
310 	TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, 127, 16
311 };
312 
313 /*
314  * Entry points for TCP as a device. The normal case which supports
315  * the TCP functionality.
316  * We have separate open functions for the /dev/tcp and /dev/tcp6 devices.
317  */
318 struct qinit tcp_rinitv4 = {
319 	NULL, tcp_rsrv, tcp_openv4, tcp_tpi_close, NULL, &tcp_rinfo
320 };
321 
322 struct qinit tcp_rinitv6 = {
323 	NULL, tcp_rsrv, tcp_openv6, tcp_tpi_close, NULL, &tcp_rinfo
324 };
325 
326 struct qinit tcp_winit = {
327 	tcp_wput, tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
328 };
329 
330 /* Initial entry point for TCP in socket mode. */
331 struct qinit tcp_sock_winit = {
332 	tcp_wput_sock, tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
333 };
334 
335 /* TCP entry point during fallback */
336 struct qinit tcp_fallback_sock_winit = {
337 	tcp_wput_fallback, NULL, NULL, NULL, NULL, &tcp_winfo
338 };
339 
340 /*
341  * Entry points for TCP as a acceptor STREAM opened by sockfs when doing
342  * an accept. Avoid allocating data structures since eager has already
343  * been created.
344  */
345 struct qinit tcp_acceptor_rinit = {
346 	NULL, tcp_rsrv, NULL, tcp_tpi_close_accept, NULL, &tcp_winfo
347 };
348 
349 struct qinit tcp_acceptor_winit = {
350 	tcp_tpi_accept, NULL, NULL, NULL, NULL, &tcp_winfo
351 };
352 
353 /* For AF_INET aka /dev/tcp */
354 struct streamtab tcpinfov4 = {
355 	&tcp_rinitv4, &tcp_winit
356 };
357 
358 /* For AF_INET6 aka /dev/tcp6 */
359 struct streamtab tcpinfov6 = {
360 	&tcp_rinitv6, &tcp_winit
361 };
362 
363 /*
364  * Following assumes TPI alignment requirements stay along 32 bit
365  * boundaries
366  */
367 #define	ROUNDUP32(x) \
368 	(((x) + (sizeof (int32_t) - 1)) & ~(sizeof (int32_t) - 1))
369 
370 /* Template for response to info request. */
371 struct T_info_ack tcp_g_t_info_ack = {
372 	T_INFO_ACK,		/* PRIM_type */
373 	0,			/* TSDU_size */
374 	T_INFINITE,		/* ETSDU_size */
375 	T_INVALID,		/* CDATA_size */
376 	T_INVALID,		/* DDATA_size */
377 	sizeof (sin_t),		/* ADDR_size */
378 	0,			/* OPT_size - not initialized here */
379 	TIDUSZ,			/* TIDU_size */
380 	T_COTS_ORD,		/* SERV_type */
381 	TCPS_IDLE,		/* CURRENT_state */
382 	(XPG4_1|EXPINLINE)	/* PROVIDER_flag */
383 };
384 
385 struct T_info_ack tcp_g_t_info_ack_v6 = {
386 	T_INFO_ACK,		/* PRIM_type */
387 	0,			/* TSDU_size */
388 	T_INFINITE,		/* ETSDU_size */
389 	T_INVALID,		/* CDATA_size */
390 	T_INVALID,		/* DDATA_size */
391 	sizeof (sin6_t),	/* ADDR_size */
392 	0,			/* OPT_size - not initialized here */
393 	TIDUSZ,		/* TIDU_size */
394 	T_COTS_ORD,		/* SERV_type */
395 	TCPS_IDLE,		/* CURRENT_state */
396 	(XPG4_1|EXPINLINE)	/* PROVIDER_flag */
397 };
398 
399 /*
400  * TCP tunables related declarations. Definitions are in tcp_tunables.c
401  */
402 extern mod_prop_info_t tcp_propinfo_tbl[];
403 extern int tcp_propinfo_count;
404 
405 #define	IS_VMLOANED_MBLK(mp) \
406 	(((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0)
407 
408 uint32_t do_tcpzcopy = 1;		/* 0: disable, 1: enable, 2: force */
409 
410 /*
411  * Forces all connections to obey the value of the tcps_maxpsz_multiplier
412  * tunable settable via NDD.  Otherwise, the per-connection behavior is
413  * determined dynamically during tcp_set_destination(), which is the default.
414  */
415 boolean_t tcp_static_maxpsz = B_FALSE;
416 
417 /*
418  * If the receive buffer size is changed, this function is called to update
419  * the upper socket layer on the new delayed receive wake up threshold.
420  */
421 static void
tcp_set_recv_threshold(tcp_t * tcp,uint32_t new_rcvthresh)422 tcp_set_recv_threshold(tcp_t *tcp, uint32_t new_rcvthresh)
423 {
424 	uint32_t default_threshold = SOCKET_RECVHIWATER >> 3;
425 
426 	if (IPCL_IS_NONSTR(tcp->tcp_connp)) {
427 		conn_t *connp = tcp->tcp_connp;
428 		struct sock_proto_props sopp;
429 
430 		/*
431 		 * only increase rcvthresh upto default_threshold
432 		 */
433 		if (new_rcvthresh > default_threshold)
434 			new_rcvthresh = default_threshold;
435 
436 		sopp.sopp_flags = SOCKOPT_RCVTHRESH;
437 		sopp.sopp_rcvthresh = new_rcvthresh;
438 
439 		(*connp->conn_upcalls->su_set_proto_props)
440 		    (connp->conn_upper_handle, &sopp);
441 	}
442 }
443 
444 /*
445  * Figure out the value of window scale opton.  Note that the rwnd is
446  * ASSUMED to be rounded up to the nearest MSS before the calculation.
447  * We cannot find the scale value and then do a round up of tcp_rwnd
448  * because the scale value may not be correct after that.
449  *
450  * Set the compiler flag to make this function inline.
451  */
452 void
tcp_set_ws_value(tcp_t * tcp)453 tcp_set_ws_value(tcp_t *tcp)
454 {
455 	int i;
456 	uint32_t rwnd = tcp->tcp_rwnd;
457 
458 	for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT;
459 	    i++, rwnd >>= 1)
460 		;
461 	tcp->tcp_rcv_ws = i;
462 }
463 
464 /*
465  * Remove cached/latched IPsec references.
466  */
467 void
tcp_ipsec_cleanup(tcp_t * tcp)468 tcp_ipsec_cleanup(tcp_t *tcp)
469 {
470 	conn_t		*connp = tcp->tcp_connp;
471 
472 	ASSERT(connp->conn_flags & IPCL_TCPCONN);
473 
474 	if (connp->conn_latch != NULL) {
475 		IPLATCH_REFRELE(connp->conn_latch);
476 		connp->conn_latch = NULL;
477 	}
478 	if (connp->conn_latch_in_policy != NULL) {
479 		IPPOL_REFRELE(connp->conn_latch_in_policy);
480 		connp->conn_latch_in_policy = NULL;
481 	}
482 	if (connp->conn_latch_in_action != NULL) {
483 		IPACT_REFRELE(connp->conn_latch_in_action);
484 		connp->conn_latch_in_action = NULL;
485 	}
486 	if (connp->conn_policy != NULL) {
487 		IPPH_REFRELE(connp->conn_policy, connp->conn_netstack);
488 		connp->conn_policy = NULL;
489 	}
490 }
491 
492 /*
493  * Cleaup before placing on free list.
494  * Disassociate from the netstack/tcp_stack_t since the freelist
495  * is per squeue and not per netstack.
496  */
497 void
tcp_cleanup(tcp_t * tcp)498 tcp_cleanup(tcp_t *tcp)
499 {
500 	mblk_t		*mp;
501 	conn_t		*connp = tcp->tcp_connp;
502 	tcp_stack_t	*tcps = tcp->tcp_tcps;
503 	netstack_t	*ns = tcps->tcps_netstack;
504 	mblk_t		*tcp_rsrv_mp;
505 
506 	tcp_bind_hash_remove(tcp);
507 
508 	/* Cleanup that which needs the netstack first */
509 	tcp_ipsec_cleanup(tcp);
510 	ixa_cleanup(connp->conn_ixa);
511 
512 	if (connp->conn_ht_iphc != NULL) {
513 		kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
514 		connp->conn_ht_iphc = NULL;
515 		connp->conn_ht_iphc_allocated = 0;
516 		connp->conn_ht_iphc_len = 0;
517 		connp->conn_ht_ulp = NULL;
518 		connp->conn_ht_ulp_len = 0;
519 		tcp->tcp_ipha = NULL;
520 		tcp->tcp_ip6h = NULL;
521 		tcp->tcp_tcpha = NULL;
522 	}
523 
524 	/* We clear any IP_OPTIONS and extension headers */
525 	ip_pkt_free(&connp->conn_xmit_ipp);
526 
527 	tcp_free(tcp);
528 
529 	/*
530 	 * Since we will bzero the entire structure, we need to
531 	 * remove it and reinsert it in global hash list. We
532 	 * know the walkers can't get to this conn because we
533 	 * had set CONDEMNED flag earlier and checked reference
534 	 * under conn_lock so walker won't pick it and when we
535 	 * go the ipcl_globalhash_remove() below, no walker
536 	 * can get to it.
537 	 */
538 	ipcl_globalhash_remove(connp);
539 
540 	/* Save some state */
541 	mp = tcp->tcp_timercache;
542 
543 	tcp_rsrv_mp = tcp->tcp_rsrv_mp;
544 
545 	if (connp->conn_cred != NULL) {
546 		crfree(connp->conn_cred);
547 		connp->conn_cred = NULL;
548 	}
549 	ipcl_conn_cleanup(connp);
550 	connp->conn_flags = IPCL_TCPCONN;
551 
552 	/*
553 	 * Now it is safe to decrement the reference counts.
554 	 * This might be the last reference on the netstack
555 	 * in which case it will cause the freeing of the IP Instance.
556 	 */
557 	connp->conn_netstack = NULL;
558 	connp->conn_ixa->ixa_ipst = NULL;
559 	netstack_rele(ns);
560 	ASSERT(tcps != NULL);
561 	tcp->tcp_tcps = NULL;
562 
563 	bzero(tcp, sizeof (tcp_t));
564 
565 	/* restore the state */
566 	tcp->tcp_timercache = mp;
567 
568 	tcp->tcp_rsrv_mp = tcp_rsrv_mp;
569 
570 	tcp->tcp_connp = connp;
571 
572 	ASSERT(connp->conn_tcp == tcp);
573 	ASSERT(connp->conn_flags & IPCL_TCPCONN);
574 	connp->conn_state_flags = CONN_INCIPIENT;
575 	ASSERT(connp->conn_proto == IPPROTO_TCP);
576 	ASSERT(connp->conn_ref == 1);
577 }
578 
579 /*
580  * Adapt to the information, such as rtt and rtt_sd, provided from the
581  * DCE and IRE maintained by IP.
582  *
583  * Checks for multicast and broadcast destination address.
584  * Returns zero if ok; an errno on failure.
585  *
586  * Note that the MSS calculation here is based on the info given in
587  * the DCE and IRE.  We do not do any calculation based on TCP options.  They
588  * will be handled in tcp_input_data() when TCP knows which options to use.
589  *
590  * Note on how TCP gets its parameters for a connection.
591  *
592  * When a tcp_t structure is allocated, it gets all the default parameters.
593  * In tcp_set_destination(), it gets those metric parameters, like rtt, rtt_sd,
594  * spipe, rpipe, ... from the route metrics.  Route metric overrides the
595  * default.
596  *
597  * An incoming SYN with a multicast or broadcast destination address is dropped
598  * in ip_fanout_v4/v6.
599  *
600  * An incoming SYN with a multicast or broadcast source address is always
601  * dropped in tcp_set_destination, since IPDF_ALLOW_MCBC is not set in
602  * conn_connect.
603  * The same logic in tcp_set_destination also serves to
604  * reject an attempt to connect to a broadcast or multicast (destination)
605  * address.
606  */
607 int
tcp_set_destination(tcp_t * tcp)608 tcp_set_destination(tcp_t *tcp)
609 {
610 	uint32_t	mss_max;
611 	uint32_t	mss;
612 	boolean_t	tcp_detached = TCP_IS_DETACHED(tcp);
613 	conn_t		*connp = tcp->tcp_connp;
614 	tcp_stack_t	*tcps = tcp->tcp_tcps;
615 	iulp_t		uinfo;
616 	int		error;
617 	uint32_t	flags;
618 
619 	flags = IPDF_LSO | IPDF_ZCOPY;
620 	/*
621 	 * Make sure we have a dce for the destination to avoid dce_ident
622 	 * contention for connected sockets.
623 	 */
624 	flags |= IPDF_UNIQUE_DCE;
625 
626 	if (!tcps->tcps_ignore_path_mtu)
627 		connp->conn_ixa->ixa_flags |= IXAF_PMTU_DISCOVERY;
628 
629 	/* Use conn_lock to satify ASSERT; tcp is already serialized */
630 	mutex_enter(&connp->conn_lock);
631 	error = conn_connect(connp, &uinfo, flags);
632 	mutex_exit(&connp->conn_lock);
633 	if (error != 0)
634 		return (error);
635 
636 	error = tcp_build_hdrs(tcp);
637 	if (error != 0)
638 		return (error);
639 
640 	tcp->tcp_localnet = uinfo.iulp_localnet;
641 
642 	if (uinfo.iulp_rtt != 0) {
643 		tcp->tcp_rtt_sa = MSEC2NSEC(uinfo.iulp_rtt);
644 		tcp->tcp_rtt_sd = MSEC2NSEC(uinfo.iulp_rtt_sd);
645 		tcp->tcp_rto = tcp_calculate_rto(tcp, tcps, 0);
646 	}
647 	if (uinfo.iulp_ssthresh != 0)
648 		tcp->tcp_cwnd_ssthresh = uinfo.iulp_ssthresh;
649 	else
650 		tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
651 	if (uinfo.iulp_spipe > 0) {
652 		connp->conn_sndbuf = MIN(uinfo.iulp_spipe,
653 		    tcps->tcps_max_buf);
654 		if (tcps->tcps_snd_lowat_fraction != 0) {
655 			connp->conn_sndlowat = connp->conn_sndbuf /
656 			    tcps->tcps_snd_lowat_fraction;
657 		}
658 		(void) tcp_maxpsz_set(tcp, B_TRUE);
659 	}
660 	/*
661 	 * Note that up till now, acceptor always inherits receive
662 	 * window from the listener.  But if there is a metrics
663 	 * associated with a host, we should use that instead of
664 	 * inheriting it from listener. Thus we need to pass this
665 	 * info back to the caller.
666 	 */
667 	if (uinfo.iulp_rpipe > 0) {
668 		tcp->tcp_rwnd = MIN(uinfo.iulp_rpipe,
669 		    tcps->tcps_max_buf);
670 	}
671 
672 	if (uinfo.iulp_rtomax > 0) {
673 		tcp->tcp_second_timer_threshold =
674 		    uinfo.iulp_rtomax;
675 	}
676 
677 	/*
678 	 * Use the metric option settings, iulp_tstamp_ok and
679 	 * iulp_wscale_ok, only for active open. What this means
680 	 * is that if the other side uses timestamp or window
681 	 * scale option, TCP will also use those options. That
682 	 * is for passive open.  If the application sets a
683 	 * large window, window scale is enabled regardless of
684 	 * the value in iulp_wscale_ok.  This is the behavior
685 	 * since 2.6.  So we keep it.
686 	 * The only case left in passive open processing is the
687 	 * check for SACK.
688 	 * For ECN, it should probably be like SACK.  But the
689 	 * current value is binary, so we treat it like the other
690 	 * cases.  The metric only controls active open.For passive
691 	 * open, the ndd param, tcp_ecn_permitted, controls the
692 	 * behavior.
693 	 */
694 	if (!tcp_detached) {
695 		/*
696 		 * The if check means that the following can only
697 		 * be turned on by the metrics only IRE, but not off.
698 		 */
699 		if (uinfo.iulp_tstamp_ok)
700 			tcp->tcp_snd_ts_ok = B_TRUE;
701 		if (uinfo.iulp_wscale_ok)
702 			tcp->tcp_snd_ws_ok = B_TRUE;
703 		if (uinfo.iulp_sack == 2)
704 			tcp->tcp_snd_sack_ok = B_TRUE;
705 		if (uinfo.iulp_ecn_ok)
706 			tcp->tcp_ecn_ok = B_TRUE;
707 	} else {
708 		/*
709 		 * Passive open.
710 		 *
711 		 * As above, the if check means that SACK can only be
712 		 * turned on by the metric only IRE.
713 		 */
714 		if (uinfo.iulp_sack > 0) {
715 			tcp->tcp_snd_sack_ok = B_TRUE;
716 		}
717 	}
718 
719 	/*
720 	 * XXX Note that currently, iulp_mtu can be as small as 68
721 	 * because of PMTUd.  So tcp_mss may go to negative if combined
722 	 * length of all those options exceeds 28 bytes.  But because
723 	 * of the tcp_mss_min check below, we may not have a problem if
724 	 * tcp_mss_min is of a reasonable value.  The default is 1 so
725 	 * the negative problem still exists.  And the check defeats PMTUd.
726 	 * In fact, if PMTUd finds that the MSS should be smaller than
727 	 * tcp_mss_min, TCP should turn off PMUTd and use the tcp_mss_min
728 	 * value.
729 	 *
730 	 * We do not deal with that now.  All those problems related to
731 	 * PMTUd will be fixed later.
732 	 */
733 	ASSERT(uinfo.iulp_mtu != 0);
734 	mss = tcp->tcp_initial_pmtu = uinfo.iulp_mtu;
735 
736 	/* Sanity check for MSS value. */
737 	if (connp->conn_ipversion == IPV4_VERSION)
738 		mss_max = tcps->tcps_mss_max_ipv4;
739 	else
740 		mss_max = tcps->tcps_mss_max_ipv6;
741 
742 	if (tcp->tcp_ipsec_overhead == 0)
743 		tcp->tcp_ipsec_overhead = conn_ipsec_length(connp);
744 
745 	mss -= tcp->tcp_ipsec_overhead;
746 
747 	if (mss < tcps->tcps_mss_min)
748 		mss = tcps->tcps_mss_min;
749 	if (mss > mss_max)
750 		mss = mss_max;
751 
752 	/* Note that this is the maximum MSS, excluding all options. */
753 	tcp->tcp_mss = mss;
754 
755 	/*
756 	 * Update the tcp connection with LSO capability.
757 	 */
758 	tcp_update_lso(tcp, connp->conn_ixa);
759 
760 	/*
761 	 * Initialize the ISS here now that we have the full connection ID.
762 	 * The RFC 1948 method of initial sequence number generation requires
763 	 * knowledge of the full connection ID before setting the ISS.
764 	 */
765 	tcp_iss_init(tcp);
766 
767 	tcp->tcp_loopback = (uinfo.iulp_loopback | uinfo.iulp_local);
768 
769 	/*
770 	 * Make sure that conn is not marked incipient
771 	 * for incoming connections. A blind
772 	 * removal of incipient flag is cheaper than
773 	 * check and removal.
774 	 */
775 	mutex_enter(&connp->conn_lock);
776 	connp->conn_state_flags &= ~CONN_INCIPIENT;
777 	mutex_exit(&connp->conn_lock);
778 	return (0);
779 }
780 
781 /*
782  * tcp_clean_death / tcp_close_detached must not be called more than once
783  * on a tcp. Thus every function that potentially calls tcp_clean_death
784  * must check for the tcp state before calling tcp_clean_death.
785  * Eg. tcp_input_data, tcp_eager_kill, tcp_clean_death_wrapper,
786  * tcp_timer_handler, all check for the tcp state.
787  */
788 /* ARGSUSED */
789 void
tcp_clean_death_wrapper(void * arg,mblk_t * mp,void * arg2,ip_recv_attr_t * dummy)790 tcp_clean_death_wrapper(void *arg, mblk_t *mp, void *arg2,
791     ip_recv_attr_t *dummy)
792 {
793 	tcp_t	*tcp = ((conn_t *)arg)->conn_tcp;
794 
795 	freemsg(mp);
796 	if (tcp->tcp_state > TCPS_BOUND)
797 		(void) tcp_clean_death(((conn_t *)arg)->conn_tcp, ETIMEDOUT);
798 }
799 
800 /*
801  * We are dying for some reason.  Try to do it gracefully.  (May be called
802  * as writer.)
803  *
804  * Return -1 if the structure was not cleaned up (if the cleanup had to be
805  * done by a service procedure).
806  * TBD - Should the return value distinguish between the tcp_t being
807  * freed and it being reinitialized?
808  */
809 int
tcp_clean_death(tcp_t * tcp,int err)810 tcp_clean_death(tcp_t *tcp, int err)
811 {
812 	mblk_t	*mp;
813 	queue_t	*q;
814 	conn_t	*connp = tcp->tcp_connp;
815 	tcp_stack_t	*tcps = tcp->tcp_tcps;
816 
817 	if (tcp->tcp_fused)
818 		tcp_unfuse(tcp);
819 
820 	if (tcp->tcp_linger_tid != 0 &&
821 	    TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
822 		tcp_stop_lingering(tcp);
823 	}
824 
825 	ASSERT(tcp != NULL);
826 	ASSERT((connp->conn_family == AF_INET &&
827 	    connp->conn_ipversion == IPV4_VERSION) ||
828 	    (connp->conn_family == AF_INET6 &&
829 	    (connp->conn_ipversion == IPV4_VERSION ||
830 	    connp->conn_ipversion == IPV6_VERSION)));
831 
832 	if (TCP_IS_DETACHED(tcp)) {
833 		if (tcp->tcp_hard_binding) {
834 			/*
835 			 * Its an eager that we are dealing with. We close the
836 			 * eager but in case a conn_ind has already gone to the
837 			 * listener, let tcp_accept_finish() send a discon_ind
838 			 * to the listener and drop the last reference. If the
839 			 * listener doesn't even know about the eager i.e. the
840 			 * conn_ind hasn't gone up, blow away the eager and drop
841 			 * the last reference as well. If the conn_ind has gone
842 			 * up, state should be BOUND. tcp_accept_finish
843 			 * will figure out that the connection has received a
844 			 * RST and will send a DISCON_IND to the application.
845 			 */
846 			tcp_closei_local(tcp);
847 			if (!tcp->tcp_tconnind_started) {
848 				CONN_DEC_REF(connp);
849 			} else {
850 				tcp->tcp_state = TCPS_BOUND;
851 				DTRACE_TCP6(state__change, void, NULL,
852 				    ip_xmit_attr_t *, connp->conn_ixa,
853 				    void, NULL, tcp_t *, tcp, void, NULL,
854 				    int32_t, TCPS_CLOSED);
855 			}
856 		} else {
857 			tcp_close_detached(tcp);
858 		}
859 		return (0);
860 	}
861 
862 	TCP_STAT(tcps, tcp_clean_death_nondetached);
863 
864 	/*
865 	 * The connection is dead.  Decrement listener connection counter if
866 	 * necessary.
867 	 */
868 	if (tcp->tcp_listen_cnt != NULL)
869 		TCP_DECR_LISTEN_CNT(tcp);
870 
871 	/*
872 	 * When a connection is moved to TIME_WAIT state, the connection
873 	 * counter is already decremented.  So no need to decrement here
874 	 * again.  See SET_TIME_WAIT() macro.
875 	 */
876 	if (tcp->tcp_state >= TCPS_ESTABLISHED &&
877 	    tcp->tcp_state < TCPS_TIME_WAIT) {
878 		TCPS_CONN_DEC(tcps);
879 	}
880 
881 	q = connp->conn_rq;
882 
883 	/* Trash all inbound data */
884 	if (!IPCL_IS_NONSTR(connp)) {
885 		ASSERT(q != NULL);
886 		flushq(q, FLUSHALL);
887 	}
888 
889 	/*
890 	 * If we are at least part way open and there is error
891 	 * (err==0 implies no error)
892 	 * notify our client by a T_DISCON_IND.
893 	 */
894 	if ((tcp->tcp_state >= TCPS_SYN_SENT) && err) {
895 		if (tcp->tcp_state >= TCPS_ESTABLISHED &&
896 		    !TCP_IS_SOCKET(tcp)) {
897 			/*
898 			 * Send M_FLUSH according to TPI. Because sockets will
899 			 * (and must) ignore FLUSHR we do that only for TPI
900 			 * endpoints and sockets in STREAMS mode.
901 			 */
902 			(void) putnextctl1(q, M_FLUSH, FLUSHR);
903 		}
904 		if (connp->conn_debug) {
905 			(void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE|SL_ERROR,
906 			    "tcp_clean_death: discon err %d", err);
907 		}
908 		if (IPCL_IS_NONSTR(connp)) {
909 			/* Direct socket, use upcall */
910 			(*connp->conn_upcalls->su_disconnected)(
911 			    connp->conn_upper_handle, tcp->tcp_connid, err);
912 		} else {
913 			mp = mi_tpi_discon_ind(NULL, err, 0);
914 			if (mp != NULL) {
915 				putnext(q, mp);
916 			} else {
917 				if (connp->conn_debug) {
918 					(void) strlog(TCP_MOD_ID, 0, 1,
919 					    SL_ERROR|SL_TRACE,
920 					    "tcp_clean_death, sending M_ERROR");
921 				}
922 				(void) putnextctl1(q, M_ERROR, EPROTO);
923 			}
924 		}
925 		if (tcp->tcp_state <= TCPS_SYN_RCVD) {
926 			/* SYN_SENT or SYN_RCVD */
927 			TCPS_BUMP_MIB(tcps, tcpAttemptFails);
928 		} else if (tcp->tcp_state <= TCPS_CLOSE_WAIT) {
929 			/* ESTABLISHED or CLOSE_WAIT */
930 			TCPS_BUMP_MIB(tcps, tcpEstabResets);
931 		}
932 	}
933 
934 	/*
935 	 * ESTABLISHED non-STREAMS eagers are not 'detached' because
936 	 * an upper handle is obtained when the SYN-ACK comes in. So it
937 	 * should receive the 'disconnected' upcall, but tcp_reinit should
938 	 * not be called since this is an eager.
939 	 */
940 	if (tcp->tcp_listener != NULL && IPCL_IS_NONSTR(connp)) {
941 		tcp_closei_local(tcp);
942 		tcp->tcp_state = TCPS_BOUND;
943 		DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
944 		    connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
945 		    int32_t, TCPS_CLOSED);
946 		return (0);
947 	}
948 
949 	tcp_reinit(tcp);
950 	if (IPCL_IS_NONSTR(connp))
951 		(void) tcp_do_unbind(connp);
952 
953 	return (-1);
954 }
955 
956 /*
957  * In case tcp is in the "lingering state" and waits for the SO_LINGER timeout
958  * to expire, stop the wait and finish the close.
959  */
960 void
tcp_stop_lingering(tcp_t * tcp)961 tcp_stop_lingering(tcp_t *tcp)
962 {
963 	clock_t	delta = 0;
964 	tcp_stack_t	*tcps = tcp->tcp_tcps;
965 	conn_t		*connp = tcp->tcp_connp;
966 
967 	tcp->tcp_linger_tid = 0;
968 	if (tcp->tcp_state > TCPS_LISTEN) {
969 		tcp_acceptor_hash_remove(tcp);
970 		mutex_enter(&tcp->tcp_non_sq_lock);
971 		if (tcp->tcp_flow_stopped) {
972 			tcp_clrqfull(tcp);
973 		}
974 		mutex_exit(&tcp->tcp_non_sq_lock);
975 
976 		if (tcp->tcp_timer_tid != 0) {
977 			delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
978 			tcp->tcp_timer_tid = 0;
979 		}
980 		/*
981 		 * Need to cancel those timers which will not be used when
982 		 * TCP is detached.  This has to be done before the conn_wq
983 		 * is cleared.
984 		 */
985 		tcp_timers_stop(tcp);
986 
987 		tcp->tcp_detached = B_TRUE;
988 		connp->conn_rq = NULL;
989 		connp->conn_wq = NULL;
990 
991 		if (tcp->tcp_state == TCPS_TIME_WAIT) {
992 			tcp_time_wait_append(tcp);
993 			TCP_DBGSTAT(tcps, tcp_detach_time_wait);
994 			goto finish;
995 		}
996 
997 		/*
998 		 * If delta is zero the timer event wasn't executed and was
999 		 * successfully canceled. In this case we need to restart it
1000 		 * with the minimal delta possible.
1001 		 */
1002 		if (delta >= 0) {
1003 			tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1004 			    delta ? delta : 1);
1005 		}
1006 	} else {
1007 		tcp_closei_local(tcp);
1008 		CONN_DEC_REF(connp);
1009 	}
1010 finish:
1011 	tcp->tcp_detached = B_TRUE;
1012 	connp->conn_rq = NULL;
1013 	connp->conn_wq = NULL;
1014 
1015 	/* Signal closing thread that it can complete close */
1016 	mutex_enter(&tcp->tcp_closelock);
1017 	tcp->tcp_closed = 1;
1018 	cv_signal(&tcp->tcp_closecv);
1019 	mutex_exit(&tcp->tcp_closelock);
1020 
1021 	/* If we have an upper handle (socket), release it */
1022 	if (IPCL_IS_NONSTR(connp)) {
1023 		sock_upcalls_t *upcalls = connp->conn_upcalls;
1024 		sock_upper_handle_t handle = connp->conn_upper_handle;
1025 
1026 		ASSERT(upcalls != NULL);
1027 		ASSERT(upcalls->su_closed != NULL);
1028 		ASSERT(handle != NULL);
1029 		/*
1030 		 * Set these to NULL first because closed() will free upper
1031 		 * structures.  Acquire conn_lock because an external caller
1032 		 * like conn_get_socket_info() will upcall if these are
1033 		 * non-NULL.
1034 		 */
1035 		mutex_enter(&connp->conn_lock);
1036 		connp->conn_upper_handle = NULL;
1037 		connp->conn_upcalls = NULL;
1038 		mutex_exit(&connp->conn_lock);
1039 		upcalls->su_closed(handle);
1040 	}
1041 }
1042 
1043 void
tcp_close_common(conn_t * connp,int flags)1044 tcp_close_common(conn_t *connp, int flags)
1045 {
1046 	tcp_t		*tcp = connp->conn_tcp;
1047 	mblk_t		*mp = &tcp->tcp_closemp;
1048 	boolean_t	conn_ioctl_cleanup_reqd = B_FALSE;
1049 	mblk_t		*bp;
1050 
1051 	ASSERT(connp->conn_ref >= 2);
1052 
1053 	/*
1054 	 * Mark the conn as closing. ipsq_pending_mp_add will not
1055 	 * add any mp to the pending mp list, after this conn has
1056 	 * started closing.
1057 	 */
1058 	mutex_enter(&connp->conn_lock);
1059 	connp->conn_state_flags |= CONN_CLOSING;
1060 	if (connp->conn_oper_pending_ill != NULL)
1061 		conn_ioctl_cleanup_reqd = B_TRUE;
1062 	CONN_INC_REF_LOCKED(connp);
1063 	mutex_exit(&connp->conn_lock);
1064 	tcp->tcp_closeflags = (uint8_t)flags;
1065 	ASSERT(connp->conn_ref >= 3);
1066 
1067 	/*
1068 	 * tcp_closemp_used is used below without any protection of a lock
1069 	 * as we don't expect any one else to use it concurrently at this
1070 	 * point otherwise it would be a major defect.
1071 	 */
1072 
1073 	if (mp->b_prev == NULL)
1074 		tcp->tcp_closemp_used = B_TRUE;
1075 	else
1076 		cmn_err(CE_PANIC, "tcp_close: concurrent use of tcp_closemp: "
1077 		    "connp %p tcp %p\n", (void *)connp, (void *)tcp);
1078 
1079 	TCP_DEBUG_GETPCSTACK(tcp->tcmp_stk, 15);
1080 
1081 	/*
1082 	 * Cleanup any queued ioctls here. This must be done before the wq/rq
1083 	 * are re-written by tcp_close_output().
1084 	 */
1085 	if (conn_ioctl_cleanup_reqd)
1086 		conn_ioctl_cleanup(connp);
1087 
1088 	/*
1089 	 * As CONN_CLOSING is set, no further ioctls should be passed down to
1090 	 * IP for this conn (see the guards in tcp_ioctl, tcp_wput_ioctl and
1091 	 * tcp_wput_iocdata). If the ioctl was queued on an ipsq,
1092 	 * conn_ioctl_cleanup should have found it and removed it. If the ioctl
1093 	 * was still in flight at the time, we wait for it here. See comments
1094 	 * for CONN_INC_IOCTLREF in ip.h for details.
1095 	 */
1096 	mutex_enter(&connp->conn_lock);
1097 	while (connp->conn_ioctlref > 0)
1098 		cv_wait(&connp->conn_cv, &connp->conn_lock);
1099 	ASSERT(connp->conn_ioctlref == 0);
1100 	ASSERT(connp->conn_oper_pending_ill == NULL);
1101 	mutex_exit(&connp->conn_lock);
1102 
1103 	SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_close_output, connp,
1104 	    NULL, tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1105 
1106 	/*
1107 	 * For non-STREAMS sockets, the normal case is that the conn makes
1108 	 * an upcall when it's finally closed, so there is no need to wait
1109 	 * in the protocol. But in case of SO_LINGER the thread sleeps here
1110 	 * so it can properly deal with the thread being interrupted.
1111 	 */
1112 	if (IPCL_IS_NONSTR(connp) && connp->conn_linger == 0)
1113 		goto nowait;
1114 
1115 	mutex_enter(&tcp->tcp_closelock);
1116 	while (!tcp->tcp_closed) {
1117 		if (!cv_wait_sig(&tcp->tcp_closecv, &tcp->tcp_closelock)) {
1118 			/*
1119 			 * The cv_wait_sig() was interrupted. We now do the
1120 			 * following:
1121 			 *
1122 			 * 1) If the endpoint was lingering, we allow this
1123 			 * to be interrupted by cancelling the linger timeout
1124 			 * and closing normally.
1125 			 *
1126 			 * 2) Revert to calling cv_wait()
1127 			 *
1128 			 * We revert to using cv_wait() to avoid an
1129 			 * infinite loop which can occur if the calling
1130 			 * thread is higher priority than the squeue worker
1131 			 * thread and is bound to the same cpu.
1132 			 */
1133 			if (connp->conn_linger && connp->conn_lingertime > 0) {
1134 				mutex_exit(&tcp->tcp_closelock);
1135 				/* Entering squeue, bump ref count. */
1136 				CONN_INC_REF(connp);
1137 				bp = allocb_wait(0, BPRI_HI, STR_NOSIG, NULL);
1138 				SQUEUE_ENTER_ONE(connp->conn_sqp, bp,
1139 				    tcp_linger_interrupted, connp, NULL,
1140 				    tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1141 				mutex_enter(&tcp->tcp_closelock);
1142 			}
1143 			break;
1144 		}
1145 	}
1146 	while (!tcp->tcp_closed)
1147 		cv_wait(&tcp->tcp_closecv, &tcp->tcp_closelock);
1148 	mutex_exit(&tcp->tcp_closelock);
1149 
1150 	/*
1151 	 * In the case of listener streams that have eagers in the q or q0
1152 	 * we wait for the eagers to drop their reference to us. conn_rq and
1153 	 * conn_wq of the eagers point to our queues. By waiting for the
1154 	 * refcnt to drop to 1, we are sure that the eagers have cleaned
1155 	 * up their queue pointers and also dropped their references to us.
1156 	 *
1157 	 * For non-STREAMS sockets we do not have to wait here; the
1158 	 * listener will instead make a su_closed upcall when the last
1159 	 * reference is dropped.
1160 	 */
1161 	if (tcp->tcp_wait_for_eagers && !IPCL_IS_NONSTR(connp)) {
1162 		mutex_enter(&connp->conn_lock);
1163 		while (connp->conn_ref != 1) {
1164 			cv_wait(&connp->conn_cv, &connp->conn_lock);
1165 		}
1166 		mutex_exit(&connp->conn_lock);
1167 	}
1168 
1169 nowait:
1170 	connp->conn_cpid = NOPID;
1171 }
1172 
1173 /*
1174  * Called by tcp_close() routine via squeue when lingering is
1175  * interrupted by a signal.
1176  */
1177 
1178 /* ARGSUSED */
1179 static void
tcp_linger_interrupted(void * arg,mblk_t * mp,void * arg2,ip_recv_attr_t * dummy)1180 tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1181 {
1182 	conn_t	*connp = (conn_t *)arg;
1183 	tcp_t	*tcp = connp->conn_tcp;
1184 
1185 	freeb(mp);
1186 	if (tcp->tcp_linger_tid != 0 &&
1187 	    TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
1188 		tcp_stop_lingering(tcp);
1189 		tcp->tcp_client_errno = EINTR;
1190 	}
1191 }
1192 
1193 /*
1194  * Clean up the b_next and b_prev fields of every mblk pointed at by *mpp.
1195  * Some stream heads get upset if they see these later on as anything but NULL.
1196  */
1197 void
tcp_close_mpp(mblk_t ** mpp)1198 tcp_close_mpp(mblk_t **mpp)
1199 {
1200 	mblk_t	*mp;
1201 
1202 	if ((mp = *mpp) != NULL) {
1203 		do {
1204 			mp->b_next = NULL;
1205 			mp->b_prev = NULL;
1206 		} while ((mp = mp->b_cont) != NULL);
1207 
1208 		mp = *mpp;
1209 		*mpp = NULL;
1210 		freemsg(mp);
1211 	}
1212 }
1213 
1214 /* Do detached close. */
1215 void
tcp_close_detached(tcp_t * tcp)1216 tcp_close_detached(tcp_t *tcp)
1217 {
1218 	if (tcp->tcp_fused)
1219 		tcp_unfuse(tcp);
1220 
1221 	/*
1222 	 * Clustering code serializes TCP disconnect callbacks and
1223 	 * cluster tcp list walks by blocking a TCP disconnect callback
1224 	 * if a cluster tcp list walk is in progress. This ensures
1225 	 * accurate accounting of TCPs in the cluster code even though
1226 	 * the TCP list walk itself is not atomic.
1227 	 */
1228 	tcp_closei_local(tcp);
1229 	CONN_DEC_REF(tcp->tcp_connp);
1230 }
1231 
1232 /*
1233  * The tcp_t is going away. Remove it from all lists and set it
1234  * to TCPS_CLOSED. The freeing up of memory is deferred until
1235  * tcp_inactive. This is needed since a thread in tcp_rput might have
1236  * done a CONN_INC_REF on this structure before it was removed from the
1237  * hashes.
1238  */
1239 void
tcp_closei_local(tcp_t * tcp)1240 tcp_closei_local(tcp_t *tcp)
1241 {
1242 	conn_t		*connp = tcp->tcp_connp;
1243 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1244 	int32_t		oldstate;
1245 
1246 	if (!TCP_IS_SOCKET(tcp))
1247 		tcp_acceptor_hash_remove(tcp);
1248 
1249 	/*
1250 	 * This can be called via tcp_time_wait_processing() if TCP gets a
1251 	 * SYN with sequence number outside the TIME-WAIT connection's
1252 	 * window.  So we need to check for TIME-WAIT state here as the
1253 	 * connection counter is already decremented.  See SET_TIME_WAIT()
1254 	 * macro
1255 	 */
1256 	if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1257 	    tcp->tcp_state < TCPS_TIME_WAIT) {
1258 		TCPS_CONN_DEC(tcps);
1259 	}
1260 
1261 	/*
1262 	 * If we are an eager connection hanging off a listener that
1263 	 * hasn't formally accepted the connection yet, get off its
1264 	 * list and blow off any data that we have accumulated.
1265 	 */
1266 	if (tcp->tcp_listener != NULL) {
1267 		tcp_t	*listener = tcp->tcp_listener;
1268 		mutex_enter(&listener->tcp_eager_lock);
1269 		/*
1270 		 * tcp_tconnind_started == B_TRUE means that the
1271 		 * conn_ind has already gone to listener. At
1272 		 * this point, eager will be closed but we
1273 		 * leave it in listeners eager list so that
1274 		 * if listener decides to close without doing
1275 		 * accept, we can clean this up. In tcp_tli_accept
1276 		 * we take care of the case of accept on closed
1277 		 * eager.
1278 		 */
1279 		if (!tcp->tcp_tconnind_started) {
1280 			tcp_eager_unlink(tcp);
1281 			mutex_exit(&listener->tcp_eager_lock);
1282 			/*
1283 			 * We don't want to have any pointers to the
1284 			 * listener queue, after we have released our
1285 			 * reference on the listener
1286 			 */
1287 			ASSERT(tcp->tcp_detached);
1288 			connp->conn_rq = NULL;
1289 			connp->conn_wq = NULL;
1290 			CONN_DEC_REF(listener->tcp_connp);
1291 		} else {
1292 			mutex_exit(&listener->tcp_eager_lock);
1293 		}
1294 	}
1295 
1296 	/* Stop all the timers */
1297 	tcp_timers_stop(tcp);
1298 
1299 	if (tcp->tcp_state == TCPS_LISTEN) {
1300 		if (tcp->tcp_ip_addr_cache) {
1301 			kmem_free((void *)tcp->tcp_ip_addr_cache,
1302 			    IP_ADDR_CACHE_SIZE * sizeof (ipaddr_t));
1303 			tcp->tcp_ip_addr_cache = NULL;
1304 		}
1305 	}
1306 
1307 	/* Decrement listerner connection counter if necessary. */
1308 	if (tcp->tcp_listen_cnt != NULL)
1309 		TCP_DECR_LISTEN_CNT(tcp);
1310 
1311 	mutex_enter(&tcp->tcp_non_sq_lock);
1312 	if (tcp->tcp_flow_stopped)
1313 		tcp_clrqfull(tcp);
1314 	mutex_exit(&tcp->tcp_non_sq_lock);
1315 
1316 	tcp_bind_hash_remove(tcp);
1317 	/*
1318 	 * If the tcp_time_wait_collector (which runs outside the squeue)
1319 	 * is trying to remove this tcp from the time wait list, we will
1320 	 * block in tcp_time_wait_remove while trying to acquire the
1321 	 * tcp_time_wait_lock. The logic in tcp_time_wait_collector also
1322 	 * requires the ipcl_hash_remove to be ordered after the
1323 	 * tcp_time_wait_remove for the refcnt checks to work correctly.
1324 	 */
1325 	if (tcp->tcp_state == TCPS_TIME_WAIT)
1326 		(void) tcp_time_wait_remove(tcp, NULL);
1327 	CL_INET_DISCONNECT(connp);
1328 	ipcl_hash_remove(connp);
1329 	oldstate = tcp->tcp_state;
1330 	tcp->tcp_state = TCPS_CLOSED;
1331 	/* Need to probe before ixa_cleanup() is called */
1332 	DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1333 	    connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
1334 	    int32_t, oldstate);
1335 	ixa_cleanup(connp->conn_ixa);
1336 
1337 	/*
1338 	 * Mark the conn as CONDEMNED
1339 	 */
1340 	mutex_enter(&connp->conn_lock);
1341 	connp->conn_state_flags |= CONN_CONDEMNED;
1342 	mutex_exit(&connp->conn_lock);
1343 
1344 	ASSERT(tcp->tcp_time_wait_next == NULL);
1345 	ASSERT(tcp->tcp_time_wait_prev == NULL);
1346 	ASSERT(tcp->tcp_time_wait_expire == 0);
1347 
1348 	tcp_ipsec_cleanup(tcp);
1349 }
1350 
1351 /*
1352  * tcp is dying (called from ipcl_conn_destroy and error cases).
1353  * Free the tcp_t in either case.
1354  */
1355 void
tcp_free(tcp_t * tcp)1356 tcp_free(tcp_t *tcp)
1357 {
1358 	mblk_t		*mp;
1359 	conn_t		*connp = tcp->tcp_connp;
1360 
1361 	ASSERT(tcp != NULL);
1362 	ASSERT(tcp->tcp_ptpahn == NULL && tcp->tcp_acceptor_hash == NULL);
1363 
1364 	connp->conn_rq = NULL;
1365 	connp->conn_wq = NULL;
1366 
1367 	tcp_close_mpp(&tcp->tcp_xmit_head);
1368 	tcp_close_mpp(&tcp->tcp_reass_head);
1369 	if (tcp->tcp_rcv_list != NULL) {
1370 		/* Free b_next chain */
1371 		tcp_close_mpp(&tcp->tcp_rcv_list);
1372 	}
1373 	if ((mp = tcp->tcp_urp_mp) != NULL) {
1374 		freemsg(mp);
1375 	}
1376 	if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1377 		freemsg(mp);
1378 	}
1379 
1380 	if (tcp->tcp_fused_sigurg_mp != NULL) {
1381 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1382 		freeb(tcp->tcp_fused_sigurg_mp);
1383 		tcp->tcp_fused_sigurg_mp = NULL;
1384 	}
1385 
1386 	if (tcp->tcp_ordrel_mp != NULL) {
1387 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1388 		freeb(tcp->tcp_ordrel_mp);
1389 		tcp->tcp_ordrel_mp = NULL;
1390 	}
1391 
1392 	TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
1393 	bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
1394 
1395 	if (tcp->tcp_hopopts != NULL) {
1396 		mi_free(tcp->tcp_hopopts);
1397 		tcp->tcp_hopopts = NULL;
1398 		tcp->tcp_hopoptslen = 0;
1399 	}
1400 	ASSERT(tcp->tcp_hopoptslen == 0);
1401 	if (tcp->tcp_dstopts != NULL) {
1402 		mi_free(tcp->tcp_dstopts);
1403 		tcp->tcp_dstopts = NULL;
1404 		tcp->tcp_dstoptslen = 0;
1405 	}
1406 	ASSERT(tcp->tcp_dstoptslen == 0);
1407 	if (tcp->tcp_rthdrdstopts != NULL) {
1408 		mi_free(tcp->tcp_rthdrdstopts);
1409 		tcp->tcp_rthdrdstopts = NULL;
1410 		tcp->tcp_rthdrdstoptslen = 0;
1411 	}
1412 	ASSERT(tcp->tcp_rthdrdstoptslen == 0);
1413 	if (tcp->tcp_rthdr != NULL) {
1414 		mi_free(tcp->tcp_rthdr);
1415 		tcp->tcp_rthdr = NULL;
1416 		tcp->tcp_rthdrlen = 0;
1417 	}
1418 	ASSERT(tcp->tcp_rthdrlen == 0);
1419 
1420 	/*
1421 	 * Following is really a blowing away a union.
1422 	 * It happens to have exactly two members of identical size
1423 	 * the following code is enough.
1424 	 */
1425 	tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1426 
1427 	/* Allow the CC algorithm to clean up after itself. */
1428 	if (tcp->tcp_cc_algo != NULL && tcp->tcp_cc_algo->cb_destroy != NULL)
1429 		tcp->tcp_cc_algo->cb_destroy(&tcp->tcp_ccv);
1430 
1431 	/*
1432 	 * If this is a non-STREAM socket still holding on to an upper
1433 	 * handle, release it. As a result of fallback we might also see
1434 	 * STREAMS based conns with upper handles, in which case there is
1435 	 * nothing to do other than clearing the field.
1436 	 */
1437 	if (connp->conn_upper_handle != NULL) {
1438 		sock_upcalls_t *upcalls = connp->conn_upcalls;
1439 		sock_upper_handle_t handle = connp->conn_upper_handle;
1440 
1441 		/*
1442 		 * Set these to NULL first because closed() will free upper
1443 		 * structures.  Acquire conn_lock because an external caller
1444 		 * like conn_get_socket_info() will upcall if these are
1445 		 * non-NULL.
1446 		 */
1447 		mutex_enter(&connp->conn_lock);
1448 		connp->conn_upper_handle = NULL;
1449 		connp->conn_upcalls = NULL;
1450 		mutex_exit(&connp->conn_lock);
1451 		if (IPCL_IS_NONSTR(connp)) {
1452 			ASSERT(upcalls != NULL);
1453 			ASSERT(upcalls->su_closed != NULL);
1454 			ASSERT(handle != NULL);
1455 			upcalls->su_closed(handle);
1456 			tcp->tcp_detached = B_TRUE;
1457 		}
1458 	}
1459 }
1460 
1461 /*
1462  * tcp_get_conn/tcp_free_conn
1463  *
1464  * tcp_get_conn is used to get a clean tcp connection structure.
1465  * It tries to reuse the connections put on the freelist by the
1466  * time_wait_collector failing which it goes to kmem_cache. This
1467  * way has two benefits compared to just allocating from and
1468  * freeing to kmem_cache.
1469  * 1) The time_wait_collector can free (which includes the cleanup)
1470  * outside the squeue. So when the interrupt comes, we have a clean
1471  * connection sitting in the freelist. Obviously, this buys us
1472  * performance.
1473  *
1474  * 2) Defence against DOS attack. Allocating a tcp/conn in tcp_input_listener
1475  * has multiple disadvantages - tying up the squeue during alloc.
1476  * But allocating the conn/tcp in IP land is also not the best since
1477  * we can't check the 'q' and 'q0' which are protected by squeue and
1478  * blindly allocate memory which might have to be freed here if we are
1479  * not allowed to accept the connection. By using the freelist and
1480  * putting the conn/tcp back in freelist, we don't pay a penalty for
1481  * allocating memory without checking 'q/q0' and freeing it if we can't
1482  * accept the connection.
1483  *
1484  * Care should be taken to put the conn back in the same squeue's freelist
1485  * from which it was allocated. Best results are obtained if conn is
1486  * allocated from listener's squeue and freed to the same. Time wait
1487  * collector will free up the freelist is the connection ends up sitting
1488  * there for too long.
1489  */
1490 conn_t *
tcp_get_conn(void * arg,tcp_stack_t * tcps)1491 tcp_get_conn(void *arg, tcp_stack_t *tcps)
1492 {
1493 	tcp_t			*tcp = NULL;
1494 	conn_t			*connp = NULL;
1495 	squeue_t		*sqp = (squeue_t *)arg;
1496 	tcp_squeue_priv_t	*tcp_time_wait;
1497 	netstack_t		*ns;
1498 	mblk_t			*tcp_rsrv_mp = NULL;
1499 
1500 	tcp_time_wait =
1501 	    *((tcp_squeue_priv_t **)squeue_getprivate(sqp, SQPRIVATE_TCP));
1502 
1503 	mutex_enter(&tcp_time_wait->tcp_time_wait_lock);
1504 	tcp = tcp_time_wait->tcp_free_list;
1505 	ASSERT((tcp != NULL) ^ (tcp_time_wait->tcp_free_list_cnt == 0));
1506 	if (tcp != NULL) {
1507 		tcp_time_wait->tcp_free_list = tcp->tcp_time_wait_next;
1508 		tcp_time_wait->tcp_free_list_cnt--;
1509 		mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1510 		tcp->tcp_time_wait_next = NULL;
1511 		connp = tcp->tcp_connp;
1512 		connp->conn_flags |= IPCL_REUSED;
1513 
1514 		ASSERT(tcp->tcp_tcps == NULL);
1515 		ASSERT(connp->conn_netstack == NULL);
1516 		ASSERT(tcp->tcp_rsrv_mp != NULL);
1517 		ns = tcps->tcps_netstack;
1518 		netstack_hold(ns);
1519 		connp->conn_netstack = ns;
1520 		connp->conn_ixa->ixa_ipst = ns->netstack_ip;
1521 		tcp->tcp_tcps = tcps;
1522 		ipcl_globalhash_insert(connp);
1523 
1524 		connp->conn_ixa->ixa_notify_cookie = tcp;
1525 		ASSERT(connp->conn_ixa->ixa_notify == tcp_notify);
1526 		connp->conn_recv = tcp_input_data;
1527 		ASSERT(connp->conn_recvicmp == tcp_icmp_input);
1528 		ASSERT(connp->conn_verifyicmp == tcp_verifyicmp);
1529 		return (connp);
1530 	}
1531 	mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1532 	/*
1533 	 * Pre-allocate the tcp_rsrv_mp. This mblk will not be freed until
1534 	 * this conn_t/tcp_t is freed at ipcl_conn_destroy().
1535 	 */
1536 	tcp_rsrv_mp = allocb(0, BPRI_HI);
1537 	if (tcp_rsrv_mp == NULL)
1538 		return (NULL);
1539 
1540 	if ((connp = ipcl_conn_create(IPCL_TCPCONN, KM_NOSLEEP,
1541 	    tcps->tcps_netstack)) == NULL) {
1542 		freeb(tcp_rsrv_mp);
1543 		return (NULL);
1544 	}
1545 
1546 	tcp = connp->conn_tcp;
1547 	tcp->tcp_rsrv_mp = tcp_rsrv_mp;
1548 	mutex_init(&tcp->tcp_rsrv_mp_lock, NULL, MUTEX_DEFAULT, NULL);
1549 
1550 	tcp->tcp_tcps = tcps;
1551 
1552 	connp->conn_recv = tcp_input_data;
1553 	connp->conn_recvicmp = tcp_icmp_input;
1554 	connp->conn_verifyicmp = tcp_verifyicmp;
1555 
1556 	/*
1557 	 * Register tcp_notify to listen to capability changes detected by IP.
1558 	 * This upcall is made in the context of the call to conn_ip_output
1559 	 * thus it is inside the squeue.
1560 	 */
1561 	connp->conn_ixa->ixa_notify = tcp_notify;
1562 	connp->conn_ixa->ixa_notify_cookie = tcp;
1563 
1564 	return (connp);
1565 }
1566 
1567 /*
1568  * Handle connect to IPv4 destinations, including connections for AF_INET6
1569  * sockets connecting to IPv4 mapped IPv6 destinations.
1570  * Returns zero if OK, a positive errno, or a negative TLI error.
1571  */
1572 static int
tcp_connect_ipv4(tcp_t * tcp,ipaddr_t * dstaddrp,in_port_t dstport,uint_t srcid)1573 tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp, in_port_t dstport,
1574     uint_t srcid)
1575 {
1576 	ipaddr_t	dstaddr = *dstaddrp;
1577 	uint16_t	lport;
1578 	conn_t		*connp = tcp->tcp_connp;
1579 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1580 	int		error;
1581 
1582 	ASSERT(connp->conn_ipversion == IPV4_VERSION);
1583 
1584 	/* Check for attempt to connect to INADDR_ANY */
1585 	if (dstaddr == INADDR_ANY)  {
1586 		/*
1587 		 * SunOS 4.x and 4.3 BSD allow an application
1588 		 * to connect a TCP socket to INADDR_ANY.
1589 		 * When they do this, the kernel picks the
1590 		 * address of one interface and uses it
1591 		 * instead.  The kernel usually ends up
1592 		 * picking the address of the loopback
1593 		 * interface.  This is an undocumented feature.
1594 		 * However, we provide the same thing here
1595 		 * in order to have source and binary
1596 		 * compatibility with SunOS 4.x.
1597 		 * Update the T_CONN_REQ (sin/sin6) since it is used to
1598 		 * generate the T_CONN_CON.
1599 		 */
1600 		dstaddr = htonl(INADDR_LOOPBACK);
1601 		*dstaddrp = dstaddr;
1602 	}
1603 
1604 	/* Handle __sin6_src_id if socket not bound to an IP address */
1605 	if (srcid != 0 && connp->conn_laddr_v4 == INADDR_ANY) {
1606 		if (!ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1607 		    IPCL_ZONEID(connp), B_TRUE, tcps->tcps_netstack)) {
1608 			/* Mismatch - conn_laddr_v6 would be v6 address. */
1609 			return (EADDRNOTAVAIL);
1610 		}
1611 		connp->conn_saddr_v6 = connp->conn_laddr_v6;
1612 	}
1613 
1614 	IN6_IPADDR_TO_V4MAPPED(dstaddr, &connp->conn_faddr_v6);
1615 	connp->conn_fport = dstport;
1616 
1617 	/*
1618 	 * At this point the remote destination address and remote port fields
1619 	 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1620 	 * have to see which state tcp was in so we can take appropriate action.
1621 	 */
1622 	if (tcp->tcp_state == TCPS_IDLE) {
1623 		/*
1624 		 * We support a quick connect capability here, allowing
1625 		 * clients to transition directly from IDLE to SYN_SENT
1626 		 * tcp_bindi will pick an unused port, insert the connection
1627 		 * in the bind hash and transition to BOUND state.
1628 		 */
1629 		lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1630 		    tcp, B_TRUE);
1631 		lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1632 		    B_FALSE, B_FALSE);
1633 		if (lport == 0)
1634 			return (-TNOADDR);
1635 	}
1636 
1637 	/*
1638 	 * Lookup the route to determine a source address and the uinfo.
1639 	 * Setup TCP parameters based on the metrics/DCE.
1640 	 */
1641 	error = tcp_set_destination(tcp);
1642 	if (error != 0)
1643 		return (error);
1644 
1645 	/*
1646 	 * Don't let an endpoint connect to itself.
1647 	 */
1648 	if (connp->conn_faddr_v4 == connp->conn_laddr_v4 &&
1649 	    connp->conn_fport == connp->conn_lport)
1650 		return (-TBADADDR);
1651 
1652 	tcp->tcp_state = TCPS_SYN_SENT;
1653 
1654 	return (ipcl_conn_insert_v4(connp));
1655 }
1656 
1657 /*
1658  * Handle connect to IPv6 destinations.
1659  * Returns zero if OK, a positive errno, or a negative TLI error.
1660  */
1661 static int
tcp_connect_ipv6(tcp_t * tcp,in6_addr_t * dstaddrp,in_port_t dstport,uint32_t flowinfo,uint_t srcid,uint32_t scope_id)1662 tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp, in_port_t dstport,
1663     uint32_t flowinfo, uint_t srcid, uint32_t scope_id)
1664 {
1665 	uint16_t	lport;
1666 	conn_t		*connp = tcp->tcp_connp;
1667 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1668 	int		error;
1669 
1670 	ASSERT(connp->conn_family == AF_INET6);
1671 
1672 	/*
1673 	 * If we're here, it means that the destination address is a native
1674 	 * IPv6 address.  Return an error if conn_ipversion is not IPv6.  A
1675 	 * reason why it might not be IPv6 is if the socket was bound to an
1676 	 * IPv4-mapped IPv6 address.
1677 	 */
1678 	if (connp->conn_ipversion != IPV6_VERSION)
1679 		return (-TBADADDR);
1680 
1681 	/*
1682 	 * Interpret a zero destination to mean loopback.
1683 	 * Update the T_CONN_REQ (sin/sin6) since it is used to
1684 	 * generate the T_CONN_CON.
1685 	 */
1686 	if (IN6_IS_ADDR_UNSPECIFIED(dstaddrp))
1687 		*dstaddrp = ipv6_loopback;
1688 
1689 	/* Handle __sin6_src_id if socket not bound to an IP address */
1690 	if (srcid != 0 && IN6_IS_ADDR_UNSPECIFIED(&connp->conn_laddr_v6)) {
1691 		if (!ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1692 		    IPCL_ZONEID(connp), B_FALSE, tcps->tcps_netstack)) {
1693 			/* Mismatch - conn_laddr_v6 would be v4-mapped. */
1694 			return (EADDRNOTAVAIL);
1695 		}
1696 		connp->conn_saddr_v6 = connp->conn_laddr_v6;
1697 	}
1698 
1699 	/*
1700 	 * Take care of the scope_id now.
1701 	 */
1702 	if (scope_id != 0 && IN6_IS_ADDR_LINKSCOPE(dstaddrp)) {
1703 		connp->conn_ixa->ixa_flags |= IXAF_SCOPEID_SET;
1704 		connp->conn_ixa->ixa_scopeid = scope_id;
1705 	} else {
1706 		connp->conn_ixa->ixa_flags &= ~IXAF_SCOPEID_SET;
1707 	}
1708 
1709 	connp->conn_flowinfo = flowinfo;
1710 	connp->conn_faddr_v6 = *dstaddrp;
1711 	connp->conn_fport = dstport;
1712 
1713 	/*
1714 	 * At this point the remote destination address and remote port fields
1715 	 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1716 	 * have to see which state tcp was in so we can take appropriate action.
1717 	 */
1718 	if (tcp->tcp_state == TCPS_IDLE) {
1719 		/*
1720 		 * We support a quick connect capability here, allowing
1721 		 * clients to transition directly from IDLE to SYN_SENT
1722 		 * tcp_bindi will pick an unused port, insert the connection
1723 		 * in the bind hash and transition to BOUND state.
1724 		 */
1725 		lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1726 		    tcp, B_TRUE);
1727 		lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1728 		    B_FALSE, B_FALSE);
1729 		if (lport == 0)
1730 			return (-TNOADDR);
1731 	}
1732 
1733 	/*
1734 	 * Lookup the route to determine a source address and the uinfo.
1735 	 * Setup TCP parameters based on the metrics/DCE.
1736 	 */
1737 	error = tcp_set_destination(tcp);
1738 	if (error != 0)
1739 		return (error);
1740 
1741 	/*
1742 	 * Don't let an endpoint connect to itself.
1743 	 */
1744 	if (IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6, &connp->conn_laddr_v6) &&
1745 	    connp->conn_fport == connp->conn_lport)
1746 		return (-TBADADDR);
1747 
1748 	tcp->tcp_state = TCPS_SYN_SENT;
1749 
1750 	return (ipcl_conn_insert_v6(connp));
1751 }
1752 
1753 /*
1754  * Disconnect
1755  * Note that unlike other functions this returns a positive tli error
1756  * when it fails; it never returns an errno.
1757  */
1758 static int
tcp_disconnect_common(tcp_t * tcp,t_scalar_t seqnum)1759 tcp_disconnect_common(tcp_t *tcp, t_scalar_t seqnum)
1760 {
1761 	conn_t		*lconnp;
1762 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1763 	conn_t		*connp = tcp->tcp_connp;
1764 
1765 	/*
1766 	 * Right now, upper modules pass down a T_DISCON_REQ to TCP,
1767 	 * when the stream is in BOUND state. Do not send a reset,
1768 	 * since the destination IP address is not valid, and it can
1769 	 * be the initialized value of all zeros (broadcast address).
1770 	 */
1771 	if (tcp->tcp_state <= TCPS_BOUND) {
1772 		if (connp->conn_debug) {
1773 			(void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1774 			    "tcp_disconnect: bad state, %d", tcp->tcp_state);
1775 		}
1776 		return (TOUTSTATE);
1777 	} else if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1778 		TCPS_CONN_DEC(tcps);
1779 	}
1780 
1781 	if (seqnum == -1 || tcp->tcp_conn_req_max == 0) {
1782 
1783 		/*
1784 		 * According to TPI, for non-listeners, ignore seqnum
1785 		 * and disconnect.
1786 		 * Following interpretation of -1 seqnum is historical
1787 		 * and implied TPI ? (TPI only states that for T_CONN_IND,
1788 		 * a valid seqnum should not be -1).
1789 		 *
1790 		 *	-1 means disconnect everything
1791 		 *	regardless even on a listener.
1792 		 */
1793 
1794 		int old_state = tcp->tcp_state;
1795 		ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
1796 
1797 		/*
1798 		 * The connection can't be on the tcp_time_wait_head list
1799 		 * since it is not detached.
1800 		 */
1801 		ASSERT(tcp->tcp_time_wait_next == NULL);
1802 		ASSERT(tcp->tcp_time_wait_prev == NULL);
1803 		ASSERT(tcp->tcp_time_wait_expire == 0);
1804 		/*
1805 		 * If it used to be a listener, check to make sure no one else
1806 		 * has taken the port before switching back to LISTEN state.
1807 		 */
1808 		if (connp->conn_ipversion == IPV4_VERSION) {
1809 			lconnp = ipcl_lookup_listener_v4(connp->conn_lport,
1810 			    connp->conn_laddr_v4, IPCL_ZONEID(connp), ipst);
1811 		} else {
1812 			uint_t ifindex = 0;
1813 
1814 			if (connp->conn_ixa->ixa_flags & IXAF_SCOPEID_SET)
1815 				ifindex = connp->conn_ixa->ixa_scopeid;
1816 
1817 			/* Allow conn_bound_if listeners? */
1818 			lconnp = ipcl_lookup_listener_v6(connp->conn_lport,
1819 			    &connp->conn_laddr_v6, ifindex, IPCL_ZONEID(connp),
1820 			    ipst);
1821 		}
1822 		if (tcp->tcp_conn_req_max && lconnp == NULL) {
1823 			tcp->tcp_state = TCPS_LISTEN;
1824 			DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1825 			    connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1826 			    NULL, int32_t, old_state);
1827 		} else if (old_state > TCPS_BOUND) {
1828 			tcp->tcp_conn_req_max = 0;
1829 			tcp->tcp_state = TCPS_BOUND;
1830 			DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1831 			    connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1832 			    NULL, int32_t, old_state);
1833 
1834 			/*
1835 			 * If this end point is not going to become a listener,
1836 			 * decrement the listener connection count if
1837 			 * necessary.  Note that we do not do this if it is
1838 			 * going to be a listner (the above if case) since
1839 			 * then it may remove the counter struct.
1840 			 */
1841 			if (tcp->tcp_listen_cnt != NULL)
1842 				TCP_DECR_LISTEN_CNT(tcp);
1843 		}
1844 		if (lconnp != NULL)
1845 			CONN_DEC_REF(lconnp);
1846 		switch (old_state) {
1847 		case TCPS_SYN_SENT:
1848 		case TCPS_SYN_RCVD:
1849 			TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1850 			break;
1851 		case TCPS_ESTABLISHED:
1852 		case TCPS_CLOSE_WAIT:
1853 			TCPS_BUMP_MIB(tcps, tcpEstabResets);
1854 			break;
1855 		}
1856 
1857 		if (tcp->tcp_fused)
1858 			tcp_unfuse(tcp);
1859 
1860 		mutex_enter(&tcp->tcp_eager_lock);
1861 		if ((tcp->tcp_conn_req_cnt_q0 != 0) ||
1862 		    (tcp->tcp_conn_req_cnt_q != 0)) {
1863 			tcp_eager_cleanup(tcp, 0);
1864 		}
1865 		mutex_exit(&tcp->tcp_eager_lock);
1866 
1867 		tcp_xmit_ctl("tcp_disconnect", tcp, tcp->tcp_snxt,
1868 		    tcp->tcp_rnxt, TH_RST | TH_ACK);
1869 
1870 		tcp_reinit(tcp);
1871 
1872 		return (0);
1873 	} else if (!tcp_eager_blowoff(tcp, seqnum)) {
1874 		return (TBADSEQ);
1875 	}
1876 	return (0);
1877 }
1878 
1879 /*
1880  * Our client hereby directs us to reject the connection request
1881  * that tcp_input_listener() marked with 'seqnum'.  Rejection consists
1882  * of sending the appropriate RST, not an ICMP error.
1883  */
1884 void
tcp_disconnect(tcp_t * tcp,mblk_t * mp)1885 tcp_disconnect(tcp_t *tcp, mblk_t *mp)
1886 {
1887 	t_scalar_t seqnum;
1888 	int	error;
1889 	conn_t	*connp = tcp->tcp_connp;
1890 
1891 	ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1892 	if ((mp->b_wptr - mp->b_rptr) < sizeof (struct T_discon_req)) {
1893 		tcp_err_ack(tcp, mp, TPROTO, 0);
1894 		return;
1895 	}
1896 	seqnum = ((struct T_discon_req *)mp->b_rptr)->SEQ_number;
1897 	error = tcp_disconnect_common(tcp, seqnum);
1898 	if (error != 0)
1899 		tcp_err_ack(tcp, mp, error, 0);
1900 	else {
1901 		if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1902 			/* Send M_FLUSH according to TPI */
1903 			(void) putnextctl1(connp->conn_rq, M_FLUSH, FLUSHRW);
1904 		}
1905 		mp = mi_tpi_ok_ack_alloc(mp);
1906 		if (mp != NULL)
1907 			putnext(connp->conn_rq, mp);
1908 	}
1909 }
1910 
1911 /*
1912  * Handle reinitialization of a tcp structure.
1913  * Maintain "binding state" resetting the state to BOUND, LISTEN, or IDLE.
1914  */
1915 static void
tcp_reinit(tcp_t * tcp)1916 tcp_reinit(tcp_t *tcp)
1917 {
1918 	mblk_t		*mp;
1919 	tcp_stack_t	*tcps = tcp->tcp_tcps;
1920 	conn_t		*connp  = tcp->tcp_connp;
1921 	int32_t		oldstate;
1922 
1923 	/* tcp_reinit should never be called for detached tcp_t's */
1924 	ASSERT(tcp->tcp_listener == NULL);
1925 	ASSERT((connp->conn_family == AF_INET &&
1926 	    connp->conn_ipversion == IPV4_VERSION) ||
1927 	    (connp->conn_family == AF_INET6 &&
1928 	    (connp->conn_ipversion == IPV4_VERSION ||
1929 	    connp->conn_ipversion == IPV6_VERSION)));
1930 
1931 	/* Cancel outstanding timers */
1932 	tcp_timers_stop(tcp);
1933 
1934 	tcp_close_mpp(&tcp->tcp_xmit_head);
1935 	if (tcp->tcp_snd_zcopy_aware)
1936 		tcp_zcopy_notify(tcp);
1937 	tcp->tcp_xmit_last = tcp->tcp_xmit_tail = NULL;
1938 	tcp->tcp_unsent = tcp->tcp_xmit_tail_unsent = 0;
1939 	mutex_enter(&tcp->tcp_non_sq_lock);
1940 	if (tcp->tcp_flow_stopped &&
1941 	    TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1942 		tcp_clrqfull(tcp);
1943 	}
1944 	mutex_exit(&tcp->tcp_non_sq_lock);
1945 	tcp_close_mpp(&tcp->tcp_reass_head);
1946 	tcp->tcp_reass_tail = NULL;
1947 	if (tcp->tcp_rcv_list != NULL) {
1948 		/* Free b_next chain */
1949 		tcp_close_mpp(&tcp->tcp_rcv_list);
1950 		tcp->tcp_rcv_last_head = NULL;
1951 		tcp->tcp_rcv_last_tail = NULL;
1952 		tcp->tcp_rcv_cnt = 0;
1953 	}
1954 	tcp->tcp_rcv_last_tail = NULL;
1955 
1956 	if ((mp = tcp->tcp_urp_mp) != NULL) {
1957 		freemsg(mp);
1958 		tcp->tcp_urp_mp = NULL;
1959 	}
1960 	if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1961 		freemsg(mp);
1962 		tcp->tcp_urp_mark_mp = NULL;
1963 	}
1964 	if (tcp->tcp_fused_sigurg_mp != NULL) {
1965 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1966 		freeb(tcp->tcp_fused_sigurg_mp);
1967 		tcp->tcp_fused_sigurg_mp = NULL;
1968 	}
1969 	if (tcp->tcp_ordrel_mp != NULL) {
1970 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1971 		freeb(tcp->tcp_ordrel_mp);
1972 		tcp->tcp_ordrel_mp = NULL;
1973 	}
1974 
1975 	/*
1976 	 * Following is a union with two members which are
1977 	 * identical types and size so the following cleanup
1978 	 * is enough.
1979 	 */
1980 	tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1981 
1982 	CL_INET_DISCONNECT(connp);
1983 
1984 	/*
1985 	 * The connection can't be on the tcp_time_wait_head list
1986 	 * since it is not detached.
1987 	 */
1988 	ASSERT(tcp->tcp_time_wait_next == NULL);
1989 	ASSERT(tcp->tcp_time_wait_prev == NULL);
1990 	ASSERT(tcp->tcp_time_wait_expire == 0);
1991 
1992 	/*
1993 	 * Reset/preserve other values
1994 	 */
1995 	tcp_reinit_values(tcp);
1996 	ipcl_hash_remove(connp);
1997 	/* Note that ixa_cred gets cleared in ixa_cleanup */
1998 	ixa_cleanup(connp->conn_ixa);
1999 	tcp_ipsec_cleanup(tcp);
2000 
2001 	connp->conn_laddr_v6 = connp->conn_bound_addr_v6;
2002 	connp->conn_saddr_v6 = connp->conn_bound_addr_v6;
2003 	oldstate = tcp->tcp_state;
2004 
2005 	if (tcp->tcp_conn_req_max != 0) {
2006 		/*
2007 		 * This is the case when a TLI program uses the same
2008 		 * transport end point to accept a connection.  This
2009 		 * makes the TCP both a listener and acceptor.  When
2010 		 * this connection is closed, we need to set the state
2011 		 * back to TCPS_LISTEN.  Make sure that the eager list
2012 		 * is reinitialized.
2013 		 *
2014 		 * Note that this stream is still bound to the four
2015 		 * tuples of the previous connection in IP.  If a new
2016 		 * SYN with different foreign address comes in, IP will
2017 		 * not find it and will send it to the global queue.  In
2018 		 * the global queue, TCP will do a tcp_lookup_listener()
2019 		 * to find this stream.  This works because this stream
2020 		 * is only removed from connected hash.
2021 		 *
2022 		 */
2023 		tcp->tcp_state = TCPS_LISTEN;
2024 		tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
2025 		tcp->tcp_eager_next_drop_q0 = tcp;
2026 		tcp->tcp_eager_prev_drop_q0 = tcp;
2027 		/*
2028 		 * Initially set conn_recv to tcp_input_listener_unbound to try
2029 		 * to pick a good squeue for the listener when the first SYN
2030 		 * arrives. tcp_input_listener_unbound sets it to
2031 		 * tcp_input_listener on that first SYN.
2032 		 */
2033 		connp->conn_recv = tcp_input_listener_unbound;
2034 
2035 		connp->conn_proto = IPPROTO_TCP;
2036 		connp->conn_faddr_v6 = ipv6_all_zeros;
2037 		connp->conn_fport = 0;
2038 
2039 		(void) ipcl_bind_insert(connp);
2040 	} else {
2041 		tcp->tcp_state = TCPS_BOUND;
2042 	}
2043 
2044 	/*
2045 	 * Initialize to default values
2046 	 */
2047 	tcp_init_values(tcp, NULL);
2048 
2049 	DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
2050 	    connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
2051 	    int32_t, oldstate);
2052 
2053 	ASSERT(tcp->tcp_ptpbhn != NULL);
2054 	tcp->tcp_rwnd = connp->conn_rcvbuf;
2055 	tcp->tcp_mss = connp->conn_ipversion != IPV4_VERSION ?
2056 	    tcps->tcps_mss_def_ipv6 : tcps->tcps_mss_def_ipv4;
2057 }
2058 
2059 /*
2060  * Force values to zero that need be zero.
2061  * Do not touch values asociated with the BOUND or LISTEN state
2062  * since the connection will end up in that state after the reinit.
2063  * NOTE: tcp_reinit_values MUST have a line for each field in the tcp_t
2064  * structure!
2065  */
2066 static void
tcp_reinit_values(tcp_t * tcp)2067 tcp_reinit_values(tcp_t *tcp)
2068 {
2069 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2070 	conn_t		*connp = tcp->tcp_connp;
2071 
2072 #ifndef	lint
2073 #define	DONTCARE(x)
2074 #define	PRESERVE(x)
2075 #else
2076 #define	DONTCARE(x)	((x) = (x))
2077 #define	PRESERVE(x)	((x) = (x))
2078 #endif	/* lint */
2079 
2080 	PRESERVE(tcp->tcp_bind_hash_port);
2081 	PRESERVE(tcp->tcp_bind_hash);
2082 	PRESERVE(tcp->tcp_ptpbhn);
2083 	PRESERVE(tcp->tcp_acceptor_hash);
2084 	PRESERVE(tcp->tcp_ptpahn);
2085 
2086 	/* Should be ASSERT NULL on these with new code! */
2087 	ASSERT(tcp->tcp_time_wait_next == NULL);
2088 	ASSERT(tcp->tcp_time_wait_prev == NULL);
2089 	ASSERT(tcp->tcp_time_wait_expire == 0);
2090 	PRESERVE(tcp->tcp_state);
2091 	PRESERVE(connp->conn_rq);
2092 	PRESERVE(connp->conn_wq);
2093 
2094 	ASSERT(tcp->tcp_xmit_head == NULL);
2095 	ASSERT(tcp->tcp_xmit_last == NULL);
2096 	ASSERT(tcp->tcp_unsent == 0);
2097 	ASSERT(tcp->tcp_xmit_tail == NULL);
2098 	ASSERT(tcp->tcp_xmit_tail_unsent == 0);
2099 
2100 	tcp->tcp_snxt = 0;			/* Displayed in mib */
2101 	tcp->tcp_suna = 0;			/* Displayed in mib */
2102 	tcp->tcp_swnd = 0;
2103 	DONTCARE(tcp->tcp_cwnd);	/* Init in tcp_process_options */
2104 
2105 	if (connp->conn_ht_iphc != NULL) {
2106 		kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
2107 		connp->conn_ht_iphc = NULL;
2108 		connp->conn_ht_iphc_allocated = 0;
2109 		connp->conn_ht_iphc_len = 0;
2110 		connp->conn_ht_ulp = NULL;
2111 		connp->conn_ht_ulp_len = 0;
2112 		tcp->tcp_ipha = NULL;
2113 		tcp->tcp_ip6h = NULL;
2114 		tcp->tcp_tcpha = NULL;
2115 	}
2116 
2117 	/* We clear any IP_OPTIONS and extension headers */
2118 	ip_pkt_free(&connp->conn_xmit_ipp);
2119 
2120 	DONTCARE(tcp->tcp_naglim);		/* Init in tcp_init_values */
2121 	DONTCARE(tcp->tcp_ipha);
2122 	DONTCARE(tcp->tcp_ip6h);
2123 	DONTCARE(tcp->tcp_tcpha);
2124 	tcp->tcp_valid_bits = 0;
2125 
2126 	DONTCARE(tcp->tcp_timer_backoff);	/* Init in tcp_init_values */
2127 	DONTCARE(tcp->tcp_last_recv_time);	/* Init in tcp_init_values */
2128 	tcp->tcp_last_rcv_lbolt = 0;
2129 
2130 	tcp->tcp_init_cwnd = 0;
2131 
2132 	tcp->tcp_urp_last_valid = 0;
2133 	tcp->tcp_hard_binding = 0;
2134 
2135 	tcp->tcp_fin_acked = 0;
2136 	tcp->tcp_fin_rcvd = 0;
2137 	tcp->tcp_fin_sent = 0;
2138 	tcp->tcp_ordrel_done = 0;
2139 
2140 	tcp->tcp_detached = 0;
2141 
2142 	tcp->tcp_snd_ws_ok = B_FALSE;
2143 	tcp->tcp_snd_ts_ok = B_FALSE;
2144 	tcp->tcp_zero_win_probe = 0;
2145 
2146 	tcp->tcp_loopback = 0;
2147 	tcp->tcp_localnet = 0;
2148 	tcp->tcp_syn_defense = 0;
2149 	tcp->tcp_set_timer = 0;
2150 
2151 	tcp->tcp_active_open = 0;
2152 	tcp->tcp_rexmit = B_FALSE;
2153 	tcp->tcp_xmit_zc_clean = B_FALSE;
2154 
2155 	tcp->tcp_snd_sack_ok = B_FALSE;
2156 	tcp->tcp_hwcksum = B_FALSE;
2157 
2158 	DONTCARE(tcp->tcp_maxpsz_multiplier);	/* Init in tcp_init_values */
2159 
2160 	tcp->tcp_conn_def_q0 = 0;
2161 	tcp->tcp_ip_forward_progress = B_FALSE;
2162 	tcp->tcp_ecn_ok = B_FALSE;
2163 
2164 	tcp->tcp_cwr = B_FALSE;
2165 	tcp->tcp_ecn_echo_on = B_FALSE;
2166 	tcp->tcp_is_wnd_shrnk = B_FALSE;
2167 
2168 	TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
2169 	bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
2170 
2171 	tcp->tcp_rcv_ws = 0;
2172 	tcp->tcp_snd_ws = 0;
2173 	tcp->tcp_ts_recent = 0;
2174 	tcp->tcp_rnxt = 0;			/* Displayed in mib */
2175 	DONTCARE(tcp->tcp_rwnd);		/* Set in tcp_reinit() */
2176 	tcp->tcp_initial_pmtu = 0;
2177 
2178 	ASSERT(tcp->tcp_reass_head == NULL);
2179 	ASSERT(tcp->tcp_reass_tail == NULL);
2180 
2181 	tcp->tcp_cwnd_cnt = 0;
2182 
2183 	ASSERT(tcp->tcp_rcv_list == NULL);
2184 	ASSERT(tcp->tcp_rcv_last_head == NULL);
2185 	ASSERT(tcp->tcp_rcv_last_tail == NULL);
2186 	ASSERT(tcp->tcp_rcv_cnt == 0);
2187 
2188 	DONTCARE(tcp->tcp_cwnd_ssthresh); /* Init in tcp_set_destination */
2189 	DONTCARE(tcp->tcp_cwnd_max);		/* Init in tcp_init_values */
2190 	tcp->tcp_csuna = 0;
2191 
2192 	tcp->tcp_rto = 0;			/* Displayed in MIB */
2193 	DONTCARE(tcp->tcp_rtt_sa);		/* Init in tcp_init_values */
2194 	DONTCARE(tcp->tcp_rtt_sd);		/* Init in tcp_init_values */
2195 	tcp->tcp_rtt_update = 0;
2196 	tcp->tcp_rtt_sum = 0;
2197 	tcp->tcp_rtt_cnt = 0;
2198 
2199 	DONTCARE(tcp->tcp_swl1); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2200 	DONTCARE(tcp->tcp_swl2); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2201 
2202 	tcp->tcp_rack = 0;			/* Displayed in mib */
2203 	tcp->tcp_rack_cnt = 0;
2204 	tcp->tcp_rack_cur_max = 0;
2205 	tcp->tcp_rack_abs_max = 0;
2206 
2207 	tcp->tcp_max_swnd = 0;
2208 
2209 	ASSERT(tcp->tcp_listener == NULL);
2210 
2211 	DONTCARE(tcp->tcp_irs);			/* tcp_valid_bits cleared */
2212 	DONTCARE(tcp->tcp_iss);			/* tcp_valid_bits cleared */
2213 	DONTCARE(tcp->tcp_fss);			/* tcp_valid_bits cleared */
2214 	DONTCARE(tcp->tcp_urg);			/* tcp_valid_bits cleared */
2215 
2216 	ASSERT(tcp->tcp_conn_req_cnt_q == 0);
2217 	ASSERT(tcp->tcp_conn_req_cnt_q0 == 0);
2218 	PRESERVE(tcp->tcp_conn_req_max);
2219 	PRESERVE(tcp->tcp_conn_req_seqnum);
2220 
2221 	DONTCARE(tcp->tcp_first_timer_threshold); /* Init in tcp_init_values */
2222 	DONTCARE(tcp->tcp_second_timer_threshold); /* Init in tcp_init_values */
2223 	DONTCARE(tcp->tcp_first_ctimer_threshold); /* Init in tcp_init_values */
2224 	DONTCARE(tcp->tcp_second_ctimer_threshold); /* in tcp_init_values */
2225 
2226 	DONTCARE(tcp->tcp_urp_last);	/* tcp_urp_last_valid is cleared */
2227 	ASSERT(tcp->tcp_urp_mp == NULL);
2228 	ASSERT(tcp->tcp_urp_mark_mp == NULL);
2229 	ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
2230 
2231 	ASSERT(tcp->tcp_eager_next_q == NULL);
2232 	ASSERT(tcp->tcp_eager_last_q == NULL);
2233 	ASSERT((tcp->tcp_eager_next_q0 == NULL &&
2234 	    tcp->tcp_eager_prev_q0 == NULL) ||
2235 	    tcp->tcp_eager_next_q0 == tcp->tcp_eager_prev_q0);
2236 	ASSERT(tcp->tcp_conn.tcp_eager_conn_ind == NULL);
2237 
2238 	ASSERT((tcp->tcp_eager_next_drop_q0 == NULL &&
2239 	    tcp->tcp_eager_prev_drop_q0 == NULL) ||
2240 	    tcp->tcp_eager_next_drop_q0 == tcp->tcp_eager_prev_drop_q0);
2241 
2242 	DONTCARE(tcp->tcp_ka_rinterval);	/* Init in tcp_init_values */
2243 	DONTCARE(tcp->tcp_ka_abort_thres);	/* Init in tcp_init_values */
2244 	DONTCARE(tcp->tcp_ka_cnt);		/* Init in tcp_init_values */
2245 
2246 	tcp->tcp_client_errno = 0;
2247 
2248 	DONTCARE(connp->conn_sum);		/* Init in tcp_init_values */
2249 
2250 	connp->conn_faddr_v6 = ipv6_all_zeros;	/* Displayed in MIB */
2251 
2252 	PRESERVE(connp->conn_bound_addr_v6);
2253 	tcp->tcp_last_sent_len = 0;
2254 	tcp->tcp_dupack_cnt = 0;
2255 
2256 	connp->conn_fport = 0;			/* Displayed in MIB */
2257 	PRESERVE(connp->conn_lport);
2258 
2259 	PRESERVE(tcp->tcp_acceptor_lockp);
2260 
2261 	ASSERT(tcp->tcp_ordrel_mp == NULL);
2262 	PRESERVE(tcp->tcp_acceptor_id);
2263 	DONTCARE(tcp->tcp_ipsec_overhead);
2264 
2265 	PRESERVE(connp->conn_family);
2266 	/* Remove any remnants of mapped address binding */
2267 	if (connp->conn_family == AF_INET6) {
2268 		connp->conn_ipversion = IPV6_VERSION;
2269 		tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2270 	} else {
2271 		connp->conn_ipversion = IPV4_VERSION;
2272 		tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2273 	}
2274 
2275 	connp->conn_bound_if = 0;
2276 	connp->conn_recv_ancillary.crb_all = 0;
2277 	tcp->tcp_recvifindex = 0;
2278 	tcp->tcp_recvhops = 0;
2279 	tcp->tcp_closed = 0;
2280 	if (tcp->tcp_hopopts != NULL) {
2281 		mi_free(tcp->tcp_hopopts);
2282 		tcp->tcp_hopopts = NULL;
2283 		tcp->tcp_hopoptslen = 0;
2284 	}
2285 	ASSERT(tcp->tcp_hopoptslen == 0);
2286 	if (tcp->tcp_dstopts != NULL) {
2287 		mi_free(tcp->tcp_dstopts);
2288 		tcp->tcp_dstopts = NULL;
2289 		tcp->tcp_dstoptslen = 0;
2290 	}
2291 	ASSERT(tcp->tcp_dstoptslen == 0);
2292 	if (tcp->tcp_rthdrdstopts != NULL) {
2293 		mi_free(tcp->tcp_rthdrdstopts);
2294 		tcp->tcp_rthdrdstopts = NULL;
2295 		tcp->tcp_rthdrdstoptslen = 0;
2296 	}
2297 	ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2298 	if (tcp->tcp_rthdr != NULL) {
2299 		mi_free(tcp->tcp_rthdr);
2300 		tcp->tcp_rthdr = NULL;
2301 		tcp->tcp_rthdrlen = 0;
2302 	}
2303 	ASSERT(tcp->tcp_rthdrlen == 0);
2304 
2305 	/* Reset fusion-related fields */
2306 	tcp->tcp_fused = B_FALSE;
2307 	tcp->tcp_unfusable = B_FALSE;
2308 	tcp->tcp_fused_sigurg = B_FALSE;
2309 	tcp->tcp_loopback_peer = NULL;
2310 
2311 	tcp->tcp_lso = B_FALSE;
2312 
2313 	tcp->tcp_in_ack_unsent = 0;
2314 	tcp->tcp_cork = B_FALSE;
2315 	tcp->tcp_tconnind_started = B_FALSE;
2316 
2317 	PRESERVE(tcp->tcp_squeue_bytes);
2318 
2319 	tcp->tcp_closemp_used = B_FALSE;
2320 
2321 	PRESERVE(tcp->tcp_rsrv_mp);
2322 	PRESERVE(tcp->tcp_rsrv_mp_lock);
2323 
2324 #ifdef DEBUG
2325 	DONTCARE(tcp->tcmp_stk[0]);
2326 #endif
2327 
2328 	PRESERVE(tcp->tcp_connid);
2329 
2330 	ASSERT(tcp->tcp_listen_cnt == NULL);
2331 	ASSERT(tcp->tcp_reass_tid == 0);
2332 
2333 	/* Allow the CC algorithm to clean up after itself. */
2334 	if (tcp->tcp_cc_algo->cb_destroy != NULL)
2335 		tcp->tcp_cc_algo->cb_destroy(&tcp->tcp_ccv);
2336 	tcp->tcp_cc_algo = NULL;
2337 
2338 #undef	DONTCARE
2339 #undef	PRESERVE
2340 }
2341 
2342 /*
2343  * Initialize the various fields in tcp_t.  If parent (the listener) is non
2344  * NULL, certain values will be inheritted from it.
2345  */
2346 void
tcp_init_values(tcp_t * tcp,tcp_t * parent)2347 tcp_init_values(tcp_t *tcp, tcp_t *parent)
2348 {
2349 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2350 	conn_t		*connp = tcp->tcp_connp;
2351 
2352 	ASSERT((connp->conn_family == AF_INET &&
2353 	    connp->conn_ipversion == IPV4_VERSION) ||
2354 	    (connp->conn_family == AF_INET6 &&
2355 	    (connp->conn_ipversion == IPV4_VERSION ||
2356 	    connp->conn_ipversion == IPV6_VERSION)));
2357 
2358 	tcp->tcp_ccv.type = IPPROTO_TCP;
2359 	tcp->tcp_ccv.ccvc.tcp = tcp;
2360 
2361 	if (parent == NULL) {
2362 		tcp->tcp_cc_algo = tcps->tcps_default_cc_algo;
2363 
2364 		tcp->tcp_naglim = tcps->tcps_naglim_def;
2365 
2366 		tcp->tcp_rto_initial = tcps->tcps_rexmit_interval_initial;
2367 		tcp->tcp_rto_min = tcps->tcps_rexmit_interval_min;
2368 		tcp->tcp_rto_max = tcps->tcps_rexmit_interval_max;
2369 
2370 		tcp->tcp_first_ctimer_threshold =
2371 		    tcps->tcps_ip_notify_cinterval;
2372 		tcp->tcp_second_ctimer_threshold =
2373 		    tcps->tcps_ip_abort_cinterval;
2374 		tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2375 		tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2376 
2377 		tcp->tcp_fin_wait_2_flush_interval =
2378 		    tcps->tcps_fin_wait_2_flush_interval;
2379 
2380 		tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2381 		tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2382 		tcp->tcp_ka_cnt = 0;
2383 		tcp->tcp_ka_rinterval = 0;
2384 
2385 		/*
2386 		 * Default value of tcp_init_cwnd is 0, so no need to set here
2387 		 * if parent is NULL.  But we need to inherit it from parent.
2388 		 */
2389 	} else {
2390 		/* Inherit various TCP parameters from the parent. */
2391 		tcp->tcp_cc_algo = parent->tcp_cc_algo;
2392 
2393 		tcp->tcp_naglim = parent->tcp_naglim;
2394 
2395 		tcp->tcp_rto_initial = parent->tcp_rto_initial;
2396 		tcp->tcp_rto_min = parent->tcp_rto_min;
2397 		tcp->tcp_rto_max = parent->tcp_rto_max;
2398 
2399 		tcp->tcp_first_ctimer_threshold =
2400 		    parent->tcp_first_ctimer_threshold;
2401 		tcp->tcp_second_ctimer_threshold =
2402 		    parent->tcp_second_ctimer_threshold;
2403 		tcp->tcp_first_timer_threshold =
2404 		    parent->tcp_first_timer_threshold;
2405 		tcp->tcp_second_timer_threshold =
2406 		    parent->tcp_second_timer_threshold;
2407 
2408 		tcp->tcp_fin_wait_2_flush_interval =
2409 		    parent->tcp_fin_wait_2_flush_interval;
2410 		tcp->tcp_quickack = parent->tcp_quickack;
2411 
2412 		tcp->tcp_ka_interval = parent->tcp_ka_interval;
2413 		tcp->tcp_ka_abort_thres = parent->tcp_ka_abort_thres;
2414 		tcp->tcp_ka_cnt = parent->tcp_ka_cnt;
2415 		tcp->tcp_ka_rinterval = parent->tcp_ka_rinterval;
2416 
2417 		tcp->tcp_init_cwnd = parent->tcp_init_cwnd;
2418 	}
2419 
2420 	if (tcp->tcp_cc_algo->cb_init != NULL)
2421 		VERIFY(tcp->tcp_cc_algo->cb_init(&tcp->tcp_ccv) == 0);
2422 
2423 	/*
2424 	 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2425 	 * will be close to tcp_rexmit_interval_initial.  By doing this, we
2426 	 * allow the algorithm to adjust slowly to large fluctuations of RTT
2427 	 * during first few transmissions of a connection as seen in slow
2428 	 * links.
2429 	 */
2430 	tcp->tcp_rtt_sa = MSEC2NSEC(tcp->tcp_rto_initial) << 2;
2431 	tcp->tcp_rtt_sd = MSEC2NSEC(tcp->tcp_rto_initial) >> 1;
2432 	tcp->tcp_rto = tcp_calculate_rto(tcp, tcps,
2433 	    tcps->tcps_conn_grace_period);
2434 
2435 	tcp->tcp_timer_backoff = 0;
2436 	tcp->tcp_ms_we_have_waited = 0;
2437 	tcp->tcp_last_recv_time = ddi_get_lbolt();
2438 	tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2439 	tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2440 
2441 	tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2442 
2443 	/* NOTE:  ISS is now set in tcp_set_destination(). */
2444 
2445 	/* Reset fusion-related fields */
2446 	tcp->tcp_fused = B_FALSE;
2447 	tcp->tcp_unfusable = B_FALSE;
2448 	tcp->tcp_fused_sigurg = B_FALSE;
2449 	tcp->tcp_loopback_peer = NULL;
2450 
2451 	/* We rebuild the header template on the next connect/conn_request */
2452 
2453 	connp->conn_mlp_type = mlptSingle;
2454 
2455 	/*
2456 	 * Init the window scale to the max so tcp_rwnd_set() won't pare
2457 	 * down tcp_rwnd. tcp_set_destination() will set the right value later.
2458 	 */
2459 	tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2460 	tcp->tcp_rwnd = connp->conn_rcvbuf;
2461 
2462 	tcp->tcp_cork = B_FALSE;
2463 	/*
2464 	 * Init the tcp_debug option if it wasn't already set.  This value
2465 	 * determines whether TCP
2466 	 * calls strlog() to print out debug messages.  Doing this
2467 	 * initialization here means that this value is not inherited thru
2468 	 * tcp_reinit().
2469 	 */
2470 	if (!connp->conn_debug)
2471 		connp->conn_debug = tcps->tcps_dbg;
2472 }
2473 
2474 /*
2475  * Update the TCP connection according to change of PMTU.
2476  *
2477  * Path MTU might have changed by either increase or decrease, so need to
2478  * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2479  * or negative MSS, since tcp_mss_set() will do it.
2480  */
2481 void
tcp_update_pmtu(tcp_t * tcp,boolean_t decrease_only)2482 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2483 {
2484 	uint32_t	pmtu;
2485 	int32_t		mss;
2486 	conn_t		*connp = tcp->tcp_connp;
2487 	ip_xmit_attr_t	*ixa = connp->conn_ixa;
2488 	iaflags_t	ixaflags;
2489 
2490 	if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2491 		return;
2492 
2493 	if (tcp->tcp_state < TCPS_ESTABLISHED)
2494 		return;
2495 
2496 	/*
2497 	 * Always call ip_get_pmtu() to make sure that IP has updated
2498 	 * ixa_flags properly.
2499 	 */
2500 	pmtu = ip_get_pmtu(ixa);
2501 	ixaflags = ixa->ixa_flags;
2502 
2503 	/*
2504 	 * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2505 	 * IPsec overhead if applied. Make sure to use the most recent
2506 	 * IPsec information.
2507 	 */
2508 	mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2509 
2510 	/*
2511 	 * Nothing to change, so just return.
2512 	 */
2513 	if (mss == tcp->tcp_mss)
2514 		return;
2515 
2516 	/*
2517 	 * Currently, for ICMP errors, only PMTU decrease is handled.
2518 	 */
2519 	if (mss > tcp->tcp_mss && decrease_only)
2520 		return;
2521 
2522 	DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2523 
2524 	/*
2525 	 * Update ixa_fragsize and ixa_pmtu.
2526 	 */
2527 	ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2528 
2529 	/*
2530 	 * Adjust MSS and all relevant variables.
2531 	 */
2532 	tcp_mss_set(tcp, mss);
2533 
2534 	/*
2535 	 * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2536 	 * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2537 	 * has a (potentially different) min size we do the same. Make sure to
2538 	 * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2539 	 * fragment the packet.
2540 	 *
2541 	 * LSO over IPv6 can not be fragmented. So need to disable LSO
2542 	 * when IPv6 fragmentation is needed.
2543 	 */
2544 	if (mss < tcp->tcp_tcps->tcps_mss_min)
2545 		ixaflags |= IXAF_PMTU_TOO_SMALL;
2546 
2547 	if (ixaflags & IXAF_PMTU_TOO_SMALL)
2548 		ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2549 
2550 	if ((connp->conn_ipversion == IPV4_VERSION) &&
2551 	    !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2552 		tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2553 	}
2554 	ixa->ixa_flags = ixaflags;
2555 }
2556 
2557 int
tcp_maxpsz_set(tcp_t * tcp,boolean_t set_maxblk)2558 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2559 {
2560 	conn_t	*connp = tcp->tcp_connp;
2561 	queue_t	*q = connp->conn_rq;
2562 	int32_t	mss = tcp->tcp_mss;
2563 	int	maxpsz;
2564 
2565 	if (TCP_IS_DETACHED(tcp))
2566 		return (mss);
2567 	if (tcp->tcp_fused) {
2568 		maxpsz = tcp_fuse_maxpsz(tcp);
2569 		mss = INFPSZ;
2570 	} else if (tcp->tcp_maxpsz_multiplier == 0) {
2571 		/*
2572 		 * Set the sd_qn_maxpsz according to the socket send buffer
2573 		 * size, and sd_maxblk to INFPSZ (-1).  This will essentially
2574 		 * instruct the stream head to copyin user data into contiguous
2575 		 * kernel-allocated buffers without breaking it up into smaller
2576 		 * chunks.  We round up the buffer size to the nearest SMSS.
2577 		 */
2578 		maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2579 		mss = INFPSZ;
2580 	} else {
2581 		/*
2582 		 * Set sd_qn_maxpsz to approx half the (receivers) buffer
2583 		 * (and a multiple of the mss).  This instructs the stream
2584 		 * head to break down larger than SMSS writes into SMSS-
2585 		 * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2586 		 */
2587 		maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2588 		if (maxpsz > connp->conn_sndbuf / 2) {
2589 			maxpsz = connp->conn_sndbuf / 2;
2590 			/* Round up to nearest mss */
2591 			maxpsz = MSS_ROUNDUP(maxpsz, mss);
2592 		}
2593 	}
2594 
2595 	(void) proto_set_maxpsz(q, connp, maxpsz);
2596 	if (!(IPCL_IS_NONSTR(connp)))
2597 		connp->conn_wq->q_maxpsz = maxpsz;
2598 	if (set_maxblk)
2599 		(void) proto_set_tx_maxblk(q, connp, mss);
2600 	return (mss);
2601 }
2602 
2603 /* For /dev/tcp aka AF_INET open */
2604 static int
tcp_openv4(queue_t * q,dev_t * devp,int flag,int sflag,cred_t * credp)2605 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2606 {
2607 	return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2608 }
2609 
2610 /* For /dev/tcp6 aka AF_INET6 open */
2611 static int
tcp_openv6(queue_t * q,dev_t * devp,int flag,int sflag,cred_t * credp)2612 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2613 {
2614 	return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2615 }
2616 
2617 conn_t *
tcp_create_common(cred_t * credp,boolean_t isv6,boolean_t issocket,int * errorp)2618 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2619     int *errorp)
2620 {
2621 	tcp_t		*tcp = NULL;
2622 	conn_t		*connp;
2623 	zoneid_t	zoneid;
2624 	tcp_stack_t	*tcps;
2625 	squeue_t	*sqp;
2626 
2627 	ASSERT(errorp != NULL);
2628 	/*
2629 	 * Find the proper zoneid and netstack.
2630 	 */
2631 	/*
2632 	 * Special case for install: miniroot needs to be able to
2633 	 * access files via NFS as though it were always in the
2634 	 * global zone.
2635 	 */
2636 	if (credp == kcred && nfs_global_client_only != 0) {
2637 		zoneid = GLOBAL_ZONEID;
2638 		tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2639 		    netstack_tcp;
2640 		ASSERT(tcps != NULL);
2641 	} else {
2642 		netstack_t *ns;
2643 		int err;
2644 
2645 		if ((err = secpolicy_basic_net_access(credp)) != 0) {
2646 			*errorp = err;
2647 			return (NULL);
2648 		}
2649 
2650 		ns = netstack_find_by_cred(credp);
2651 		ASSERT(ns != NULL);
2652 		tcps = ns->netstack_tcp;
2653 		ASSERT(tcps != NULL);
2654 
2655 		/*
2656 		 * For exclusive stacks we set the zoneid to zero
2657 		 * to make TCP operate as if in the global zone.
2658 		 */
2659 		if (tcps->tcps_netstack->netstack_stackid !=
2660 		    GLOBAL_NETSTACKID)
2661 			zoneid = GLOBAL_ZONEID;
2662 		else
2663 			zoneid = crgetzoneid(credp);
2664 	}
2665 
2666 	sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2667 	connp = tcp_get_conn(sqp, tcps);
2668 	/*
2669 	 * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2670 	 * so we drop it by one.
2671 	 */
2672 	netstack_rele(tcps->tcps_netstack);
2673 	if (connp == NULL) {
2674 		*errorp = ENOSR;
2675 		return (NULL);
2676 	}
2677 	ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2678 
2679 	connp->conn_sqp = sqp;
2680 	connp->conn_initial_sqp = connp->conn_sqp;
2681 	connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2682 	tcp = connp->conn_tcp;
2683 
2684 	/*
2685 	 * Besides asking IP to set the checksum for us, have conn_ip_output
2686 	 * to do the following checks when necessary:
2687 	 *
2688 	 * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2689 	 * IXAF_VERIFY_PMTU: verify PMTU changes
2690 	 * IXAF_VERIFY_LSO: verify LSO capability changes
2691 	 */
2692 	connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2693 	    IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2694 
2695 	if (!tcps->tcps_dev_flow_ctl)
2696 		connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2697 
2698 	if (isv6) {
2699 		connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2700 		connp->conn_ipversion = IPV6_VERSION;
2701 		connp->conn_family = AF_INET6;
2702 		tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2703 		connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2704 	} else {
2705 		connp->conn_ipversion = IPV4_VERSION;
2706 		connp->conn_family = AF_INET;
2707 		tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2708 		connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2709 	}
2710 	connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2711 
2712 	crhold(credp);
2713 	connp->conn_cred = credp;
2714 	connp->conn_cpid = curproc->p_pid;
2715 	connp->conn_open_time = ddi_get_lbolt64();
2716 
2717 	/* Cache things in the ixa without any refhold */
2718 	ASSERT(!(connp->conn_ixa->ixa_free_flags & IXA_FREE_CRED));
2719 	connp->conn_ixa->ixa_cred = credp;
2720 	connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2721 
2722 	connp->conn_zoneid = zoneid;
2723 	/* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2724 	connp->conn_ixa->ixa_zoneid = zoneid;
2725 	connp->conn_mlp_type = mlptSingle;
2726 	ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2727 	ASSERT(tcp->tcp_tcps == tcps);
2728 
2729 	/*
2730 	 * If the caller has the process-wide flag set, then default to MAC
2731 	 * exempt mode.  This allows read-down to unlabeled hosts.
2732 	 */
2733 	if (getpflags(NET_MAC_AWARE, credp) != 0)
2734 		connp->conn_mac_mode = CONN_MAC_AWARE;
2735 
2736 	connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2737 
2738 	if (issocket) {
2739 		tcp->tcp_issocket = 1;
2740 	}
2741 
2742 	connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2743 	connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2744 	if (tcps->tcps_snd_lowat_fraction != 0) {
2745 		connp->conn_sndlowat = connp->conn_sndbuf /
2746 		    tcps->tcps_snd_lowat_fraction;
2747 	} else {
2748 		connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2749 	}
2750 	connp->conn_so_type = SOCK_STREAM;
2751 	connp->conn_wroff = connp->conn_ht_iphc_allocated +
2752 	    tcps->tcps_wroff_xtra;
2753 
2754 	SOCK_CONNID_INIT(tcp->tcp_connid);
2755 	/* DTrace ignores this - it isn't a tcp:::state-change */
2756 	tcp->tcp_state = TCPS_IDLE;
2757 	tcp_init_values(tcp, NULL);
2758 	return (connp);
2759 }
2760 
2761 static int
tcp_open(queue_t * q,dev_t * devp,int flag,int sflag,cred_t * credp,boolean_t isv6)2762 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2763     boolean_t isv6)
2764 {
2765 	tcp_t		*tcp = NULL;
2766 	conn_t		*connp = NULL;
2767 	int		err;
2768 	vmem_t		*minor_arena = NULL;
2769 	dev_t		conn_dev;
2770 	boolean_t	issocket;
2771 
2772 	if (q->q_ptr != NULL)
2773 		return (0);
2774 
2775 	if (sflag == MODOPEN)
2776 		return (EINVAL);
2777 
2778 	if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2779 	    ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2780 		minor_arena = ip_minor_arena_la;
2781 	} else {
2782 		/*
2783 		 * Either minor numbers in the large arena were exhausted
2784 		 * or a non socket application is doing the open.
2785 		 * Try to allocate from the small arena.
2786 		 */
2787 		if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2788 			return (EBUSY);
2789 		}
2790 		minor_arena = ip_minor_arena_sa;
2791 	}
2792 
2793 	ASSERT(minor_arena != NULL);
2794 
2795 	*devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2796 
2797 	if (flag & SO_FALLBACK) {
2798 		/*
2799 		 * Non streams socket needs a stream to fallback to
2800 		 */
2801 		RD(q)->q_ptr = (void *)conn_dev;
2802 		WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2803 		WR(q)->q_ptr = (void *)minor_arena;
2804 		qprocson(q);
2805 		return (0);
2806 	} else if (flag & SO_ACCEPTOR) {
2807 		q->q_qinfo = &tcp_acceptor_rinit;
2808 		/*
2809 		 * the conn_dev and minor_arena will be subsequently used by
2810 		 * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2811 		 * the minor device number for this connection from the q_ptr.
2812 		 */
2813 		RD(q)->q_ptr = (void *)conn_dev;
2814 		WR(q)->q_qinfo = &tcp_acceptor_winit;
2815 		WR(q)->q_ptr = (void *)minor_arena;
2816 		qprocson(q);
2817 		return (0);
2818 	}
2819 
2820 	issocket = flag & SO_SOCKSTR;
2821 	connp = tcp_create_common(credp, isv6, issocket, &err);
2822 
2823 	if (connp == NULL) {
2824 		inet_minor_free(minor_arena, conn_dev);
2825 		q->q_ptr = WR(q)->q_ptr = NULL;
2826 		return (err);
2827 	}
2828 
2829 	connp->conn_rq = q;
2830 	connp->conn_wq = WR(q);
2831 	q->q_ptr = WR(q)->q_ptr = connp;
2832 
2833 	connp->conn_dev = conn_dev;
2834 	connp->conn_minor_arena = minor_arena;
2835 
2836 	ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2837 	ASSERT(WR(q)->q_qinfo == &tcp_winit);
2838 
2839 	tcp = connp->conn_tcp;
2840 
2841 	if (issocket) {
2842 		WR(q)->q_qinfo = &tcp_sock_winit;
2843 	} else {
2844 #ifdef  _ILP32
2845 		tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2846 #else
2847 		tcp->tcp_acceptor_id = conn_dev;
2848 #endif  /* _ILP32 */
2849 		tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2850 	}
2851 
2852 	/*
2853 	 * Put the ref for TCP. Ref for IP was already put
2854 	 * by ipcl_conn_create. Also Make the conn_t globally
2855 	 * visible to walkers
2856 	 */
2857 	mutex_enter(&connp->conn_lock);
2858 	CONN_INC_REF_LOCKED(connp);
2859 	ASSERT(connp->conn_ref == 2);
2860 	connp->conn_state_flags &= ~CONN_INCIPIENT;
2861 	mutex_exit(&connp->conn_lock);
2862 
2863 	qprocson(q);
2864 	return (0);
2865 }
2866 
2867 /*
2868  * Build/update the tcp header template (in conn_ht_iphc) based on
2869  * conn_xmit_ipp. The headers include ip6_t, any extension
2870  * headers, and the maximum size tcp header (to avoid reallocation
2871  * on the fly for additional tcp options).
2872  *
2873  * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2874  * Returns failure if can't allocate memory.
2875  */
2876 int
tcp_build_hdrs(tcp_t * tcp)2877 tcp_build_hdrs(tcp_t *tcp)
2878 {
2879 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2880 	conn_t		*connp = tcp->tcp_connp;
2881 	char		buf[TCP_MAX_HDR_LENGTH];
2882 	uint_t		buflen;
2883 	uint_t		ulplen = TCP_MIN_HEADER_LENGTH;
2884 	uint_t		extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
2885 	tcpha_t		*tcpha;
2886 	uint32_t	cksum;
2887 	int		error;
2888 
2889 	/*
2890 	 * We might be called after the connection is set up, and we might
2891 	 * have TS options already in the TCP header. Thus we  save any
2892 	 * existing tcp header.
2893 	 */
2894 	buflen = connp->conn_ht_ulp_len;
2895 	if (buflen != 0) {
2896 		bcopy(connp->conn_ht_ulp, buf, buflen);
2897 		extralen -= buflen - ulplen;
2898 		ulplen = buflen;
2899 	}
2900 
2901 	/* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
2902 	mutex_enter(&connp->conn_lock);
2903 	error = conn_build_hdr_template(connp, ulplen, extralen,
2904 	    &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
2905 	mutex_exit(&connp->conn_lock);
2906 	if (error != 0)
2907 		return (error);
2908 
2909 	/*
2910 	 * Any routing header/option has been massaged. The checksum difference
2911 	 * is stored in conn_sum for later use.
2912 	 */
2913 	tcpha = (tcpha_t *)connp->conn_ht_ulp;
2914 	tcp->tcp_tcpha = tcpha;
2915 
2916 	/* restore any old tcp header */
2917 	if (buflen != 0) {
2918 		bcopy(buf, connp->conn_ht_ulp, buflen);
2919 	} else {
2920 		tcpha->tha_sum = 0;
2921 		tcpha->tha_urp = 0;
2922 		tcpha->tha_ack = 0;
2923 		tcpha->tha_offset_and_reserved = (5 << 4);
2924 		tcpha->tha_lport = connp->conn_lport;
2925 		tcpha->tha_fport = connp->conn_fport;
2926 	}
2927 
2928 	/*
2929 	 * IP wants our header length in the checksum field to
2930 	 * allow it to perform a single pseudo-header+checksum
2931 	 * calculation on behalf of TCP.
2932 	 * Include the adjustment for a source route once IP_OPTIONS is set.
2933 	 */
2934 	cksum = sizeof (tcpha_t) + connp->conn_sum;
2935 	cksum = (cksum >> 16) + (cksum & 0xFFFF);
2936 	ASSERT(cksum < 0x10000);
2937 	tcpha->tha_sum = htons(cksum);
2938 
2939 	if (connp->conn_ipversion == IPV4_VERSION)
2940 		tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
2941 	else
2942 		tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
2943 
2944 	if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
2945 	    connp->conn_wroff) {
2946 		connp->conn_wroff = connp->conn_ht_iphc_allocated +
2947 		    tcps->tcps_wroff_xtra;
2948 		(void) proto_set_tx_wroff(connp->conn_rq, connp,
2949 		    connp->conn_wroff);
2950 	}
2951 	return (0);
2952 }
2953 
2954 /*
2955  * tcp_rwnd_set() is called to adjust the receive window to a desired value.
2956  * We do not allow the receive window to shrink.  After setting rwnd,
2957  * set the flow control hiwat of the stream.
2958  *
2959  * This function is called in 2 cases:
2960  *
2961  * 1) Before data transfer begins, in tcp_input_listener() for accepting a
2962  *    connection (passive open) and in tcp_input_data() for active connect.
2963  *    This is called after tcp_mss_set() when the desired MSS value is known.
2964  *    This makes sure that our window size is a mutiple of the other side's
2965  *    MSS.
2966  * 2) Handling SO_RCVBUF option.
2967  *
2968  * It is ASSUMED that the requested size is a multiple of the current MSS.
2969  *
2970  * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
2971  * user requests so.
2972  */
2973 int
tcp_rwnd_set(tcp_t * tcp,uint32_t rwnd)2974 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
2975 {
2976 	uint32_t	mss = tcp->tcp_mss;
2977 	uint32_t	old_max_rwnd;
2978 	uint32_t	max_transmittable_rwnd;
2979 	boolean_t	tcp_detached = TCP_IS_DETACHED(tcp);
2980 	tcp_stack_t	*tcps = tcp->tcp_tcps;
2981 	conn_t		*connp = tcp->tcp_connp;
2982 
2983 	/*
2984 	 * Insist on a receive window that is at least
2985 	 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
2986 	 * funny TCP interactions of Nagle algorithm, SWS avoidance
2987 	 * and delayed acknowledgement.
2988 	 */
2989 	rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
2990 
2991 	if (tcp->tcp_fused) {
2992 		size_t sth_hiwat;
2993 		tcp_t *peer_tcp = tcp->tcp_loopback_peer;
2994 
2995 		ASSERT(peer_tcp != NULL);
2996 		sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
2997 		if (!tcp_detached) {
2998 			(void) proto_set_rx_hiwat(connp->conn_rq, connp,
2999 			    sth_hiwat);
3000 			tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
3001 		}
3002 
3003 		/* Caller could have changed tcp_rwnd; update tha_win */
3004 		if (tcp->tcp_tcpha != NULL) {
3005 			tcp->tcp_tcpha->tha_win =
3006 			    htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3007 		}
3008 		if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3009 			tcp->tcp_cwnd_max = rwnd;
3010 
3011 		/*
3012 		 * In the fusion case, the maxpsz stream head value of
3013 		 * our peer is set according to its send buffer size
3014 		 * and our receive buffer size; since the latter may
3015 		 * have changed we need to update the peer's maxpsz.
3016 		 */
3017 		(void) tcp_maxpsz_set(peer_tcp, B_TRUE);
3018 		return (sth_hiwat);
3019 	}
3020 
3021 	if (tcp_detached)
3022 		old_max_rwnd = tcp->tcp_rwnd;
3023 	else
3024 		old_max_rwnd = connp->conn_rcvbuf;
3025 
3026 
3027 	/*
3028 	 * If window size info has already been exchanged, TCP should not
3029 	 * shrink the window.  Shrinking window is doable if done carefully.
3030 	 * We may add that support later.  But so far there is not a real
3031 	 * need to do that.
3032 	 */
3033 	if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
3034 		/* MSS may have changed, do a round up again. */
3035 		rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
3036 	}
3037 
3038 	/*
3039 	 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3040 	 * can be applied even before the window scale option is decided.
3041 	 */
3042 	max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3043 	if (rwnd > max_transmittable_rwnd) {
3044 		rwnd = max_transmittable_rwnd -
3045 		    (max_transmittable_rwnd % mss);
3046 		if (rwnd < mss)
3047 			rwnd = max_transmittable_rwnd;
3048 		/*
3049 		 * If we're over the limit we may have to back down tcp_rwnd.
3050 		 * The increment below won't work for us. So we set all three
3051 		 * here and the increment below will have no effect.
3052 		 */
3053 		tcp->tcp_rwnd = old_max_rwnd = rwnd;
3054 	}
3055 	if (tcp->tcp_localnet) {
3056 		tcp->tcp_rack_abs_max =
3057 		    MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3058 	} else {
3059 		/*
3060 		 * For a remote host on a different subnet (through a router),
3061 		 * we ack every other packet to be conforming to RFC1122.
3062 		 * tcp_deferred_acks_max is default to 2.
3063 		 */
3064 		tcp->tcp_rack_abs_max =
3065 		    MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3066 	}
3067 	if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3068 		tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3069 	else
3070 		tcp->tcp_rack_cur_max = 0;
3071 	/*
3072 	 * Increment the current rwnd by the amount the maximum grew (we
3073 	 * can not overwrite it since we might be in the middle of a
3074 	 * connection.)
3075 	 */
3076 	tcp->tcp_rwnd += rwnd - old_max_rwnd;
3077 	connp->conn_rcvbuf = rwnd;
3078 
3079 	/* Are we already connected? */
3080 	if (tcp->tcp_tcpha != NULL) {
3081 		tcp->tcp_tcpha->tha_win =
3082 		    htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3083 	}
3084 
3085 	if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3086 		tcp->tcp_cwnd_max = rwnd;
3087 
3088 	if (tcp_detached)
3089 		return (rwnd);
3090 
3091 	tcp_set_recv_threshold(tcp, rwnd >> 3);
3092 
3093 	(void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3094 	return (rwnd);
3095 }
3096 
3097 int
tcp_do_unbind(conn_t * connp)3098 tcp_do_unbind(conn_t *connp)
3099 {
3100 	tcp_t *tcp = connp->conn_tcp;
3101 	int32_t oldstate;
3102 
3103 	switch (tcp->tcp_state) {
3104 	case TCPS_BOUND:
3105 	case TCPS_LISTEN:
3106 		break;
3107 	default:
3108 		return (-TOUTSTATE);
3109 	}
3110 
3111 	/*
3112 	 * Need to clean up all the eagers since after the unbind, segments
3113 	 * will no longer be delivered to this listener stream.
3114 	 */
3115 	mutex_enter(&tcp->tcp_eager_lock);
3116 	if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3117 		tcp_eager_cleanup(tcp, 0);
3118 	}
3119 	mutex_exit(&tcp->tcp_eager_lock);
3120 
3121 	/* Clean up the listener connection counter if necessary. */
3122 	if (tcp->tcp_listen_cnt != NULL)
3123 		TCP_DECR_LISTEN_CNT(tcp);
3124 	connp->conn_laddr_v6 = ipv6_all_zeros;
3125 	connp->conn_saddr_v6 = ipv6_all_zeros;
3126 	tcp_bind_hash_remove(tcp);
3127 	oldstate = tcp->tcp_state;
3128 	tcp->tcp_state = TCPS_IDLE;
3129 	DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
3130 	    connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
3131 	    int32_t, oldstate);
3132 
3133 	ip_unbind(connp);
3134 	bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3135 
3136 	return (0);
3137 }
3138 
3139 /*
3140  * Collect protocol properties to send to the upper handle.
3141  */
3142 void
tcp_get_proto_props(tcp_t * tcp,struct sock_proto_props * sopp)3143 tcp_get_proto_props(tcp_t *tcp, struct sock_proto_props *sopp)
3144 {
3145 	conn_t *connp = tcp->tcp_connp;
3146 
3147 	sopp->sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3148 	sopp->sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3149 
3150 	sopp->sopp_rxhiwat = tcp->tcp_fused ?
3151 	    tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3152 	    connp->conn_rcvbuf;
3153 	/*
3154 	 * Determine what write offset value to use depending on SACK and
3155 	 * whether the endpoint is fused or not.
3156 	 */
3157 	if (tcp->tcp_fused) {
3158 		ASSERT(tcp->tcp_loopback);
3159 		ASSERT(tcp->tcp_loopback_peer != NULL);
3160 		/*
3161 		 * For fused tcp loopback, set the stream head's write
3162 		 * offset value to zero since we won't be needing any room
3163 		 * for TCP/IP headers.  This would also improve performance
3164 		 * since it would reduce the amount of work done by kmem.
3165 		 * Non-fused tcp loopback case is handled separately below.
3166 		 */
3167 		sopp->sopp_wroff = 0;
3168 		/*
3169 		 * Update the peer's transmit parameters according to
3170 		 * our recently calculated high water mark value.
3171 		 */
3172 		(void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3173 	} else if (tcp->tcp_snd_sack_ok) {
3174 		sopp->sopp_wroff = connp->conn_ht_iphc_allocated +
3175 		    (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3176 	} else {
3177 		sopp->sopp_wroff = connp->conn_ht_iphc_len +
3178 		    (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3179 	}
3180 
3181 	if (tcp->tcp_loopback) {
3182 		sopp->sopp_flags |= SOCKOPT_LOOPBACK;
3183 		sopp->sopp_loopback = B_TRUE;
3184 	}
3185 }
3186 
3187 /*
3188  * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3189  */
3190 boolean_t
tcp_zcopy_check(tcp_t * tcp)3191 tcp_zcopy_check(tcp_t *tcp)
3192 {
3193 	conn_t		*connp = tcp->tcp_connp;
3194 	ip_xmit_attr_t	*ixa = connp->conn_ixa;
3195 	boolean_t	zc_enabled = B_FALSE;
3196 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3197 
3198 	if (do_tcpzcopy == 2)
3199 		zc_enabled = B_TRUE;
3200 	else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3201 		zc_enabled = B_TRUE;
3202 
3203 	tcp->tcp_snd_zcopy_on = zc_enabled;
3204 	if (!TCP_IS_DETACHED(tcp)) {
3205 		if (zc_enabled) {
3206 			ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3207 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
3208 			    ZCVMSAFE);
3209 			TCP_STAT(tcps, tcp_zcopy_on);
3210 		} else {
3211 			ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3212 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
3213 			    ZCVMUNSAFE);
3214 			TCP_STAT(tcps, tcp_zcopy_off);
3215 		}
3216 	}
3217 	return (zc_enabled);
3218 }
3219 
3220 /*
3221  * Backoff from a zero-copy message by copying data to a new allocated
3222  * message and freeing the original desballoca'ed segmapped message.
3223  *
3224  * This function is called by following two callers:
3225  * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3226  *    the origial desballoca'ed message and notify sockfs. This is in re-
3227  *    transmit state.
3228  * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3229  *    to be copied to new message.
3230  */
3231 mblk_t *
tcp_zcopy_backoff(tcp_t * tcp,mblk_t * bp,boolean_t fix_xmitlist)3232 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3233 {
3234 	mblk_t		*nbp;
3235 	mblk_t		*head = NULL;
3236 	mblk_t		*tail = NULL;
3237 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3238 
3239 	ASSERT(bp != NULL);
3240 	while (bp != NULL) {
3241 		if (IS_VMLOANED_MBLK(bp)) {
3242 			TCP_STAT(tcps, tcp_zcopy_backoff);
3243 			if ((nbp = copyb(bp)) == NULL) {
3244 				tcp->tcp_xmit_zc_clean = B_FALSE;
3245 				if (tail != NULL)
3246 					tail->b_cont = bp;
3247 				return ((head == NULL) ? bp : head);
3248 			}
3249 
3250 			if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3251 				if (fix_xmitlist)
3252 					tcp_zcopy_notify(tcp);
3253 				else
3254 					nbp->b_datap->db_struioflag |=
3255 					    STRUIO_ZCNOTIFY;
3256 			}
3257 			nbp->b_cont = bp->b_cont;
3258 
3259 			/*
3260 			 * Copy saved information and adjust tcp_xmit_tail
3261 			 * if needed.
3262 			 */
3263 			if (fix_xmitlist) {
3264 				nbp->b_prev = bp->b_prev;
3265 				nbp->b_next = bp->b_next;
3266 
3267 				if (tcp->tcp_xmit_tail == bp)
3268 					tcp->tcp_xmit_tail = nbp;
3269 			}
3270 
3271 			/* Free the original message. */
3272 			bp->b_prev = NULL;
3273 			bp->b_next = NULL;
3274 			freeb(bp);
3275 
3276 			bp = nbp;
3277 		}
3278 
3279 		if (head == NULL) {
3280 			head = bp;
3281 		}
3282 		if (tail == NULL) {
3283 			tail = bp;
3284 		} else {
3285 			tail->b_cont = bp;
3286 			tail = bp;
3287 		}
3288 
3289 		/* Move forward. */
3290 		bp = bp->b_cont;
3291 	}
3292 
3293 	if (fix_xmitlist) {
3294 		tcp->tcp_xmit_last = tail;
3295 		tcp->tcp_xmit_zc_clean = B_TRUE;
3296 	}
3297 
3298 	return (head);
3299 }
3300 
3301 void
tcp_zcopy_notify(tcp_t * tcp)3302 tcp_zcopy_notify(tcp_t *tcp)
3303 {
3304 	struct stdata	*stp;
3305 	conn_t		*connp;
3306 
3307 	if (tcp->tcp_detached)
3308 		return;
3309 	connp = tcp->tcp_connp;
3310 	if (IPCL_IS_NONSTR(connp)) {
3311 		(*connp->conn_upcalls->su_zcopy_notify)
3312 		    (connp->conn_upper_handle);
3313 		return;
3314 	}
3315 	stp = STREAM(connp->conn_rq);
3316 	mutex_enter(&stp->sd_lock);
3317 	stp->sd_flag |= STZCNOTIFY;
3318 	cv_broadcast(&stp->sd_zcopy_wait);
3319 	mutex_exit(&stp->sd_lock);
3320 }
3321 
3322 /*
3323  * Update the TCP connection according to change of LSO capability.
3324  */
3325 static void
tcp_update_lso(tcp_t * tcp,ip_xmit_attr_t * ixa)3326 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3327 {
3328 	/*
3329 	 * We check against IPv4 header length to preserve the old behavior
3330 	 * of only enabling LSO when there are no IP options.
3331 	 * But this restriction might not be necessary at all. Before removing
3332 	 * it, need to verify how LSO is handled for source routing case, with
3333 	 * which IP does software checksum.
3334 	 *
3335 	 * For IPv6, whenever any extension header is needed, LSO is supressed.
3336 	 */
3337 	if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3338 	    IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3339 		return;
3340 
3341 	/*
3342 	 * Either the LSO capability newly became usable, or it has changed.
3343 	 */
3344 	if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3345 		ill_lso_capab_t	*lsoc = &ixa->ixa_lso_capab;
3346 		uint_t lso_max = (ixa->ixa_flags & IXAF_IS_IPV4) ?
3347 		    lsoc->ill_lso_max_tcpv4 : lsoc->ill_lso_max_tcpv6;
3348 
3349 		ASSERT3U(lso_max, >, 0);
3350 		tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lso_max);
3351 
3352 		DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3353 		    boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3354 
3355 		/*
3356 		 * If LSO to be enabled, notify the STREAM header with larger
3357 		 * data block.
3358 		 */
3359 		if (!tcp->tcp_lso)
3360 			tcp->tcp_maxpsz_multiplier = 0;
3361 
3362 		tcp->tcp_lso = B_TRUE;
3363 		TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3364 	} else { /* LSO capability is not usable any more. */
3365 		DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3366 		    boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
3367 
3368 		/*
3369 		 * If LSO to be disabled, notify the STREAM header with smaller
3370 		 * data block. And need to restore fragsize to PMTU.
3371 		 */
3372 		if (tcp->tcp_lso) {
3373 			tcp->tcp_maxpsz_multiplier =
3374 			    tcp->tcp_tcps->tcps_maxpsz_multiplier;
3375 			ixa->ixa_fragsize = ixa->ixa_pmtu;
3376 			tcp->tcp_lso = B_FALSE;
3377 			TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
3378 		}
3379 	}
3380 
3381 	(void) tcp_maxpsz_set(tcp, B_TRUE);
3382 }
3383 
3384 /*
3385  * Update the TCP connection according to change of ZEROCOPY capability.
3386  */
3387 static void
tcp_update_zcopy(tcp_t * tcp)3388 tcp_update_zcopy(tcp_t *tcp)
3389 {
3390 	conn_t		*connp = tcp->tcp_connp;
3391 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3392 
3393 	if (tcp->tcp_snd_zcopy_on) {
3394 		tcp->tcp_snd_zcopy_on = B_FALSE;
3395 		if (!TCP_IS_DETACHED(tcp)) {
3396 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
3397 			    ZCVMUNSAFE);
3398 			TCP_STAT(tcps, tcp_zcopy_off);
3399 		}
3400 	} else {
3401 		tcp->tcp_snd_zcopy_on = B_TRUE;
3402 		if (!TCP_IS_DETACHED(tcp)) {
3403 			(void) proto_set_tx_copyopt(connp->conn_rq, connp,
3404 			    ZCVMSAFE);
3405 			TCP_STAT(tcps, tcp_zcopy_on);
3406 		}
3407 	}
3408 }
3409 
3410 /*
3411  * Notify function registered with ip_xmit_attr_t. It's called in the squeue
3412  * so it's safe to update the TCP connection.
3413  */
3414 /* ARGSUSED1 */
3415 static void
tcp_notify(void * arg,ip_xmit_attr_t * ixa,ixa_notify_type_t ntype,ixa_notify_arg_t narg)3416 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
3417     ixa_notify_arg_t narg)
3418 {
3419 	tcp_t		*tcp = (tcp_t *)arg;
3420 	conn_t		*connp = tcp->tcp_connp;
3421 
3422 	switch (ntype) {
3423 	case IXAN_LSO:
3424 		tcp_update_lso(tcp, connp->conn_ixa);
3425 		break;
3426 	case IXAN_PMTU:
3427 		tcp_update_pmtu(tcp, B_FALSE);
3428 		break;
3429 	case IXAN_ZCOPY:
3430 		tcp_update_zcopy(tcp);
3431 		break;
3432 	default:
3433 		break;
3434 	}
3435 }
3436 
3437 /*
3438  * The TCP write service routine should never be called...
3439  */
3440 /* ARGSUSED */
3441 static int
tcp_wsrv(queue_t * q)3442 tcp_wsrv(queue_t *q)
3443 {
3444 	tcp_stack_t	*tcps = Q_TO_TCP(q)->tcp_tcps;
3445 
3446 	TCP_STAT(tcps, tcp_wsrv_called);
3447 	return (0);
3448 }
3449 
3450 /*
3451  * Hash list lookup routine for tcp_t structures.
3452  * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
3453  */
3454 tcp_t *
tcp_acceptor_hash_lookup(t_uscalar_t id,tcp_stack_t * tcps)3455 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
3456 {
3457 	tf_t	*tf;
3458 	tcp_t	*tcp;
3459 
3460 	tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3461 	mutex_enter(&tf->tf_lock);
3462 	for (tcp = tf->tf_tcp; tcp != NULL;
3463 	    tcp = tcp->tcp_acceptor_hash) {
3464 		if (tcp->tcp_acceptor_id == id) {
3465 			CONN_INC_REF(tcp->tcp_connp);
3466 			mutex_exit(&tf->tf_lock);
3467 			return (tcp);
3468 		}
3469 	}
3470 	mutex_exit(&tf->tf_lock);
3471 	return (NULL);
3472 }
3473 
3474 /*
3475  * Hash list insertion routine for tcp_t structures.
3476  */
3477 void
tcp_acceptor_hash_insert(t_uscalar_t id,tcp_t * tcp)3478 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
3479 {
3480 	tf_t	*tf;
3481 	tcp_t	**tcpp;
3482 	tcp_t	*tcpnext;
3483 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3484 
3485 	tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3486 
3487 	if (tcp->tcp_ptpahn != NULL)
3488 		tcp_acceptor_hash_remove(tcp);
3489 	tcpp = &tf->tf_tcp;
3490 	mutex_enter(&tf->tf_lock);
3491 	tcpnext = tcpp[0];
3492 	if (tcpnext)
3493 		tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
3494 	tcp->tcp_acceptor_hash = tcpnext;
3495 	tcp->tcp_ptpahn = tcpp;
3496 	tcpp[0] = tcp;
3497 	tcp->tcp_acceptor_lockp = &tf->tf_lock;	/* For tcp_*_hash_remove */
3498 	mutex_exit(&tf->tf_lock);
3499 }
3500 
3501 /*
3502  * Hash list removal routine for tcp_t structures.
3503  */
3504 void
tcp_acceptor_hash_remove(tcp_t * tcp)3505 tcp_acceptor_hash_remove(tcp_t *tcp)
3506 {
3507 	tcp_t	*tcpnext;
3508 	kmutex_t *lockp;
3509 
3510 	/*
3511 	 * Extract the lock pointer in case there are concurrent
3512 	 * hash_remove's for this instance.
3513 	 */
3514 	lockp = tcp->tcp_acceptor_lockp;
3515 
3516 	if (tcp->tcp_ptpahn == NULL)
3517 		return;
3518 
3519 	ASSERT(lockp != NULL);
3520 	mutex_enter(lockp);
3521 	if (tcp->tcp_ptpahn) {
3522 		tcpnext = tcp->tcp_acceptor_hash;
3523 		if (tcpnext) {
3524 			tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
3525 			tcp->tcp_acceptor_hash = NULL;
3526 		}
3527 		*tcp->tcp_ptpahn = tcpnext;
3528 		tcp->tcp_ptpahn = NULL;
3529 	}
3530 	mutex_exit(lockp);
3531 	tcp->tcp_acceptor_lockp = NULL;
3532 }
3533 
3534 /*
3535  * Type three generator adapted from the random() function in 4.4 BSD:
3536  */
3537 
3538 /*
3539  * Copyright (c) 1983, 1993
3540  *	The Regents of the University of California.  All rights reserved.
3541  *
3542  * Redistribution and use in source and binary forms, with or without
3543  * modification, are permitted provided that the following conditions
3544  * are met:
3545  * 1. Redistributions of source code must retain the above copyright
3546  *    notice, this list of conditions and the following disclaimer.
3547  * 2. Redistributions in binary form must reproduce the above copyright
3548  *    notice, this list of conditions and the following disclaimer in the
3549  *    documentation and/or other materials provided with the distribution.
3550  * 3. All advertising materials mentioning features or use of this software
3551  *    must display the following acknowledgement:
3552  *	This product includes software developed by the University of
3553  *	California, Berkeley and its contributors.
3554  * 4. Neither the name of the University nor the names of its contributors
3555  *    may be used to endorse or promote products derived from this software
3556  *    without specific prior written permission.
3557  *
3558  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
3559  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
3560  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
3561  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
3562  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
3563  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
3564  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
3565  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
3566  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
3567  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
3568  * SUCH DAMAGE.
3569  */
3570 
3571 /* Type 3 -- x**31 + x**3 + 1 */
3572 #define	DEG_3		31
3573 #define	SEP_3		3
3574 
3575 
3576 /* Protected by tcp_random_lock */
3577 static int tcp_randtbl[DEG_3 + 1];
3578 
3579 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
3580 static int *tcp_random_rptr = &tcp_randtbl[1];
3581 
3582 static int *tcp_random_state = &tcp_randtbl[1];
3583 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
3584 
3585 kmutex_t tcp_random_lock;
3586 
3587 void
tcp_random_init(void)3588 tcp_random_init(void)
3589 {
3590 	int i;
3591 	hrtime_t hrt;
3592 	time_t wallclock;
3593 	uint64_t result;
3594 
3595 	/*
3596 	 * Use high-res timer and current time for seed.  Gethrtime() returns
3597 	 * a longlong, which may contain resolution down to nanoseconds.
3598 	 * The current time will either be a 32-bit or a 64-bit quantity.
3599 	 * XOR the two together in a 64-bit result variable.
3600 	 * Convert the result to a 32-bit value by multiplying the high-order
3601 	 * 32-bits by the low-order 32-bits.
3602 	 */
3603 
3604 	hrt = gethrtime();
3605 	(void) drv_getparm(TIME, &wallclock);
3606 	result = (uint64_t)wallclock ^ (uint64_t)hrt;
3607 	mutex_enter(&tcp_random_lock);
3608 	tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
3609 	    (result & 0xffffffff);
3610 
3611 	for (i = 1; i < DEG_3; i++)
3612 		tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
3613 		    + 12345;
3614 	tcp_random_fptr = &tcp_random_state[SEP_3];
3615 	tcp_random_rptr = &tcp_random_state[0];
3616 	mutex_exit(&tcp_random_lock);
3617 	for (i = 0; i < 10 * DEG_3; i++)
3618 		(void) tcp_random();
3619 }
3620 
3621 /*
3622  * tcp_random: Return a random number in the range [1 - (128K + 1)].
3623  * This range is selected to be approximately centered on TCP_ISS / 2,
3624  * and easy to compute. We get this value by generating a 32-bit random
3625  * number, selecting out the high-order 17 bits, and then adding one so
3626  * that we never return zero.
3627  */
3628 int
tcp_random(void)3629 tcp_random(void)
3630 {
3631 	int i;
3632 
3633 	mutex_enter(&tcp_random_lock);
3634 	*tcp_random_fptr += *tcp_random_rptr;
3635 
3636 	/*
3637 	 * The high-order bits are more random than the low-order bits,
3638 	 * so we select out the high-order 17 bits and add one so that
3639 	 * we never return zero.
3640 	 */
3641 	i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
3642 	if (++tcp_random_fptr >= tcp_random_end_ptr) {
3643 		tcp_random_fptr = tcp_random_state;
3644 		++tcp_random_rptr;
3645 	} else if (++tcp_random_rptr >= tcp_random_end_ptr)
3646 		tcp_random_rptr = tcp_random_state;
3647 
3648 	mutex_exit(&tcp_random_lock);
3649 	return (i);
3650 }
3651 
3652 /*
3653  * Split this function out so that if the secret changes, I'm okay.
3654  *
3655  * Initialize the tcp_iss_cookie and tcp_iss_key.
3656  */
3657 
3658 #define	PASSWD_SIZE 16  /* MUST be multiple of 4 */
3659 
3660 void
tcp_iss_key_init(uint8_t * phrase,int len,tcp_stack_t * tcps)3661 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
3662 {
3663 	struct {
3664 		int32_t current_time;
3665 		uint32_t randnum;
3666 		uint16_t pad;
3667 		uint8_t ether[6];
3668 		uint8_t passwd[PASSWD_SIZE];
3669 	} tcp_iss_cookie;
3670 	time_t t;
3671 
3672 	/*
3673 	 * Start with the current absolute time.
3674 	 */
3675 	(void) drv_getparm(TIME, &t);
3676 	tcp_iss_cookie.current_time = t;
3677 
3678 	/*
3679 	 * XXX - Need a more random number per RFC 1750, not this crap.
3680 	 * OTOH, if what follows is pretty random, then I'm in better shape.
3681 	 */
3682 	tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
3683 	tcp_iss_cookie.pad = 0x365c;  /* Picked from HMAC pad values. */
3684 
3685 	/*
3686 	 * The cpu_type_info is pretty non-random.  Ugggh.  It does serve
3687 	 * as a good template.
3688 	 */
3689 	bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
3690 	    min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
3691 
3692 	/*
3693 	 * The pass-phrase.  Normally this is supplied by user-called NDD.
3694 	 */
3695 	bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
3696 
3697 	/*
3698 	 * See 4010593 if this section becomes a problem again,
3699 	 * but the local ethernet address is useful here.
3700 	 */
3701 	(void) localetheraddr(NULL,
3702 	    (struct ether_addr *)&tcp_iss_cookie.ether);
3703 
3704 	/*
3705 	 * Hash 'em all together.  The MD5Final is called per-connection.
3706 	 */
3707 	mutex_enter(&tcps->tcps_iss_key_lock);
3708 	MD5Init(&tcps->tcps_iss_key);
3709 	MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
3710 	    sizeof (tcp_iss_cookie));
3711 	mutex_exit(&tcps->tcps_iss_key_lock);
3712 }
3713 
3714 /*
3715  * Called by IP when IP is loaded into the kernel
3716  */
3717 void
tcp_ddi_g_init(void)3718 tcp_ddi_g_init(void)
3719 {
3720 	tcp_timercache = kmem_cache_create("tcp_timercache",
3721 	    sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
3722 	    NULL, NULL, NULL, NULL, NULL, 0);
3723 
3724 	tcp_notsack_blk_cache = kmem_cache_create("tcp_notsack_blk_cache",
3725 	    sizeof (notsack_blk_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3726 
3727 	mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
3728 
3729 	/* Initialize the random number generator */
3730 	tcp_random_init();
3731 
3732 	/* A single callback independently of how many netstacks we have */
3733 	ip_squeue_init(tcp_squeue_add);
3734 
3735 	tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
3736 
3737 	tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
3738 
3739 	/*
3740 	 * We want to be informed each time a stack is created or
3741 	 * destroyed in the kernel, so we can maintain the
3742 	 * set of tcp_stack_t's.
3743 	 */
3744 	netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
3745 }
3746 
3747 
3748 #define	INET_NAME	"ip"
3749 
3750 /*
3751  * Initialize the TCP stack instance.
3752  */
3753 static void *
tcp_stack_init(netstackid_t stackid,netstack_t * ns)3754 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
3755 {
3756 	tcp_stack_t	*tcps;
3757 	int		i;
3758 	int		error = 0;
3759 	major_t		major;
3760 	size_t		arrsz;
3761 
3762 	tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
3763 	tcps->tcps_netstack = ns;
3764 
3765 	/* Initialize locks */
3766 	mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
3767 	mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
3768 
3769 	tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
3770 	tcps->tcps_g_epriv_ports[0] = ULP_DEF_EPRIV_PORT1;
3771 	tcps->tcps_g_epriv_ports[1] = ULP_DEF_EPRIV_PORT2;
3772 	tcps->tcps_min_anonpriv_port = 512;
3773 
3774 	tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
3775 	    TCP_BIND_FANOUT_SIZE, KM_SLEEP);
3776 	tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
3777 	    TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
3778 
3779 	for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3780 		mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
3781 		    MUTEX_DEFAULT, NULL);
3782 	}
3783 
3784 	for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3785 		mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
3786 		    MUTEX_DEFAULT, NULL);
3787 	}
3788 
3789 	/* TCP's IPsec code calls the packet dropper. */
3790 	ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
3791 
3792 	arrsz = tcp_propinfo_count * sizeof (mod_prop_info_t);
3793 	tcps->tcps_propinfo_tbl = (mod_prop_info_t *)kmem_alloc(arrsz,
3794 	    KM_SLEEP);
3795 	bcopy(tcp_propinfo_tbl, tcps->tcps_propinfo_tbl, arrsz);
3796 
3797 	/*
3798 	 * Note: To really walk the device tree you need the devinfo
3799 	 * pointer to your device which is only available after probe/attach.
3800 	 * The following is safe only because it uses ddi_root_node()
3801 	 */
3802 	tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
3803 	    tcp_opt_obj.odb_opt_arr_cnt);
3804 
3805 	/*
3806 	 * Initialize RFC 1948 secret values.  This will probably be reset once
3807 	 * by the boot scripts.
3808 	 *
3809 	 * Use NULL name, as the name is caught by the new lockstats.
3810 	 *
3811 	 * Initialize with some random, non-guessable string, like the global
3812 	 * T_INFO_ACK.
3813 	 */
3814 
3815 	tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
3816 	    sizeof (tcp_g_t_info_ack), tcps);
3817 
3818 	tcps->tcps_kstat = tcp_kstat2_init(stackid);
3819 	tcps->tcps_mibkp = tcp_kstat_init(stackid);
3820 
3821 	major = mod_name_to_major(INET_NAME);
3822 	error = ldi_ident_from_major(major, &tcps->tcps_ldi_ident);
3823 	ASSERT(error == 0);
3824 	tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
3825 	ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
3826 	cv_init(&tcps->tcps_ixa_cleanup_ready_cv, NULL, CV_DEFAULT, NULL);
3827 	cv_init(&tcps->tcps_ixa_cleanup_done_cv, NULL, CV_DEFAULT, NULL);
3828 	mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
3829 
3830 	mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
3831 	tcps->tcps_reclaim = B_FALSE;
3832 	tcps->tcps_reclaim_tid = 0;
3833 	tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
3834 
3835 	/*
3836 	 * ncpus is the current number of CPUs, which can be bigger than
3837 	 * boot_ncpus.  But we don't want to use ncpus to allocate all the
3838 	 * tcp_stats_cpu_t at system boot up time since it will be 1.  While
3839 	 * we handle adding CPU in tcp_cpu_update(), it will be slow if
3840 	 * there are many CPUs as we will be adding them 1 by 1.
3841 	 *
3842 	 * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
3843 	 * are not freed until the stack is going away.  So there is no need
3844 	 * to grab a lock to access the per CPU tcps_sc[x] pointer.
3845 	 */
3846 	mutex_enter(&cpu_lock);
3847 	tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
3848 	mutex_exit(&cpu_lock);
3849 	tcps->tcps_sc = kmem_zalloc(max_ncpus  * sizeof (tcp_stats_cpu_t *),
3850 	    KM_SLEEP);
3851 	for (i = 0; i < tcps->tcps_sc_cnt; i++) {
3852 		tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
3853 		    KM_SLEEP);
3854 	}
3855 
3856 	mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
3857 	list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
3858 	    offsetof(tcp_listener_t, tl_link));
3859 
3860 	tcps->tcps_default_cc_algo = cc_load_algo(CC_DEFAULT_ALGO_NAME);
3861 	VERIFY3P(tcps->tcps_default_cc_algo, !=, NULL);
3862 
3863 	return (tcps);
3864 }
3865 
3866 /*
3867  * Called when the IP module is about to be unloaded.
3868  */
3869 void
tcp_ddi_g_destroy(void)3870 tcp_ddi_g_destroy(void)
3871 {
3872 	tcp_g_kstat_fini(tcp_g_kstat);
3873 	tcp_g_kstat = NULL;
3874 	bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
3875 
3876 	mutex_destroy(&tcp_random_lock);
3877 
3878 	kmem_cache_destroy(tcp_timercache);
3879 	kmem_cache_destroy(tcp_notsack_blk_cache);
3880 
3881 	netstack_unregister(NS_TCP);
3882 }
3883 
3884 /*
3885  * Free the TCP stack instance.
3886  */
3887 static void
tcp_stack_fini(netstackid_t stackid,void * arg)3888 tcp_stack_fini(netstackid_t stackid, void *arg)
3889 {
3890 	tcp_stack_t *tcps = (tcp_stack_t *)arg;
3891 	int i;
3892 
3893 	freeb(tcps->tcps_ixa_cleanup_mp);
3894 	tcps->tcps_ixa_cleanup_mp = NULL;
3895 	cv_destroy(&tcps->tcps_ixa_cleanup_ready_cv);
3896 	cv_destroy(&tcps->tcps_ixa_cleanup_done_cv);
3897 	mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
3898 
3899 	/*
3900 	 * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
3901 	 * the timer.
3902 	 */
3903 	mutex_enter(&tcps->tcps_reclaim_lock);
3904 	tcps->tcps_reclaim = B_FALSE;
3905 	mutex_exit(&tcps->tcps_reclaim_lock);
3906 	if (tcps->tcps_reclaim_tid != 0)
3907 		(void) untimeout(tcps->tcps_reclaim_tid);
3908 	mutex_destroy(&tcps->tcps_reclaim_lock);
3909 
3910 	tcp_listener_conf_cleanup(tcps);
3911 
3912 	for (i = 0; i < tcps->tcps_sc_cnt; i++)
3913 		kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
3914 	kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
3915 
3916 	kmem_free(tcps->tcps_propinfo_tbl,
3917 	    tcp_propinfo_count * sizeof (mod_prop_info_t));
3918 	tcps->tcps_propinfo_tbl = NULL;
3919 
3920 	for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3921 		ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
3922 		mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
3923 	}
3924 
3925 	for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3926 		ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
3927 		mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
3928 	}
3929 
3930 	kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
3931 	tcps->tcps_bind_fanout = NULL;
3932 
3933 	kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
3934 	    TCP_ACCEPTOR_FANOUT_SIZE);
3935 	tcps->tcps_acceptor_fanout = NULL;
3936 
3937 	mutex_destroy(&tcps->tcps_iss_key_lock);
3938 	mutex_destroy(&tcps->tcps_epriv_port_lock);
3939 
3940 	ip_drop_unregister(&tcps->tcps_dropper);
3941 
3942 	tcp_kstat2_fini(stackid, tcps->tcps_kstat);
3943 	tcps->tcps_kstat = NULL;
3944 
3945 	tcp_kstat_fini(stackid, tcps->tcps_mibkp);
3946 	tcps->tcps_mibkp = NULL;
3947 
3948 	ldi_ident_release(tcps->tcps_ldi_ident);
3949 	kmem_free(tcps, sizeof (*tcps));
3950 }
3951 
3952 /*
3953  * Generate ISS, taking into account NDD changes may happen halfway through.
3954  * (If the iss is not zero, set it.)
3955  */
3956 
3957 static void
tcp_iss_init(tcp_t * tcp)3958 tcp_iss_init(tcp_t *tcp)
3959 {
3960 	MD5_CTX context;
3961 	struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
3962 	uint32_t answer[4];
3963 	tcp_stack_t	*tcps = tcp->tcp_tcps;
3964 	conn_t		*connp = tcp->tcp_connp;
3965 
3966 	tcps->tcps_iss_incr_extra += (tcps->tcps_iss_incr >> 1);
3967 	tcp->tcp_iss = tcps->tcps_iss_incr_extra;
3968 	switch (tcps->tcps_strong_iss) {
3969 	case 2:
3970 		mutex_enter(&tcps->tcps_iss_key_lock);
3971 		context = tcps->tcps_iss_key;
3972 		mutex_exit(&tcps->tcps_iss_key_lock);
3973 		arg.ports = connp->conn_ports;
3974 		arg.src = connp->conn_laddr_v6;
3975 		arg.dst = connp->conn_faddr_v6;
3976 		MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
3977 		MD5Final((uchar_t *)answer, &context);
3978 		tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
3979 		/*
3980 		 * Now that we've hashed into a unique per-connection sequence
3981 		 * space, add a random increment per strong_iss == 1.  So I
3982 		 * guess we'll have to...
3983 		 */
3984 		/* FALLTHRU */
3985 	case 1:
3986 		tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
3987 		break;
3988 	default:
3989 		tcp->tcp_iss += (uint32_t)gethrestime_sec() *
3990 		    tcps->tcps_iss_incr;
3991 		break;
3992 	}
3993 	tcp->tcp_valid_bits = TCP_ISS_VALID;
3994 	tcp->tcp_fss = tcp->tcp_iss - 1;
3995 	tcp->tcp_suna = tcp->tcp_iss;
3996 	tcp->tcp_snxt = tcp->tcp_iss + 1;
3997 	tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3998 	tcp->tcp_csuna = tcp->tcp_snxt;
3999 }
4000 
4001 /*
4002  * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
4003  * on the specified backing STREAMS q. Note, the caller may make the
4004  * decision to call based on the tcp_t.tcp_flow_stopped value which
4005  * when check outside the q's lock is only an advisory check ...
4006  */
4007 void
tcp_setqfull(tcp_t * tcp)4008 tcp_setqfull(tcp_t *tcp)
4009 {
4010 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4011 	conn_t	*connp = tcp->tcp_connp;
4012 
4013 	if (tcp->tcp_closed)
4014 		return;
4015 
4016 	conn_setqfull(connp, &tcp->tcp_flow_stopped);
4017 	if (tcp->tcp_flow_stopped)
4018 		TCP_STAT(tcps, tcp_flwctl_on);
4019 }
4020 
4021 void
tcp_clrqfull(tcp_t * tcp)4022 tcp_clrqfull(tcp_t *tcp)
4023 {
4024 	conn_t  *connp = tcp->tcp_connp;
4025 
4026 	if (tcp->tcp_closed)
4027 		return;
4028 	conn_clrqfull(connp, &tcp->tcp_flow_stopped);
4029 }
4030 
4031 static int
tcp_squeue_switch(int val)4032 tcp_squeue_switch(int val)
4033 {
4034 	int rval = SQ_FILL;
4035 
4036 	switch (val) {
4037 	case 1:
4038 		rval = SQ_NODRAIN;
4039 		break;
4040 	case 2:
4041 		rval = SQ_PROCESS;
4042 		break;
4043 	default:
4044 		break;
4045 	}
4046 	return (rval);
4047 }
4048 
4049 /*
4050  * This is called once for each squeue - globally for all stack
4051  * instances.
4052  */
4053 static void
tcp_squeue_add(squeue_t * sqp)4054 tcp_squeue_add(squeue_t *sqp)
4055 {
4056 	tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4057 	    sizeof (tcp_squeue_priv_t), KM_SLEEP);
4058 
4059 	*squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4060 	if (tcp_free_list_max_cnt == 0) {
4061 		int tcp_ncpus = ((boot_max_ncpus == -1) ?
4062 		    max_ncpus : boot_max_ncpus);
4063 
4064 		/*
4065 		 * Limit number of entries to 1% of availble memory / tcp_ncpus
4066 		 */
4067 		tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4068 		    (tcp_ncpus * sizeof (tcp_t) * 100);
4069 	}
4070 	tcp_time_wait->tcp_free_list_cnt = 0;
4071 }
4072 /*
4073  * Return unix error is tli error is TSYSERR, otherwise return a negative
4074  * tli error.
4075  */
4076 int
tcp_do_bind(conn_t * connp,struct sockaddr * sa,socklen_t len,cred_t * cr,boolean_t bind_to_req_port_only)4077 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4078     boolean_t bind_to_req_port_only)
4079 {
4080 	int error;
4081 	tcp_t *tcp = connp->conn_tcp;
4082 
4083 	if (tcp->tcp_state >= TCPS_BOUND) {
4084 		if (connp->conn_debug) {
4085 			(void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4086 			    "tcp_bind: bad state, %d", tcp->tcp_state);
4087 		}
4088 		return (-TOUTSTATE);
4089 	}
4090 
4091 	error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4092 	if (error != 0)
4093 		return (error);
4094 
4095 	ASSERT(tcp->tcp_state == TCPS_BOUND);
4096 	tcp->tcp_conn_req_max = 0;
4097 	return (0);
4098 }
4099 
4100 /*
4101  * If the return value from this function is positive, it's a UNIX error.
4102  * Otherwise, if it's negative, then the absolute value is a TLI error.
4103  * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4104  */
4105 int
tcp_do_connect(conn_t * connp,const struct sockaddr * sa,socklen_t len,cred_t * cr,pid_t pid)4106 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4107     cred_t *cr, pid_t pid)
4108 {
4109 	tcp_t		*tcp = connp->conn_tcp;
4110 	sin_t		*sin = (sin_t *)sa;
4111 	sin6_t		*sin6 = (sin6_t *)sa;
4112 	ipaddr_t	*dstaddrp;
4113 	in_port_t	dstport;
4114 	uint_t		srcid;
4115 	int		error;
4116 	uint32_t	mss;
4117 	mblk_t		*syn_mp;
4118 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4119 	int32_t		oldstate;
4120 	ip_xmit_attr_t	*ixa = connp->conn_ixa;
4121 
4122 	oldstate = tcp->tcp_state;
4123 
4124 	switch (len) {
4125 	default:
4126 		/*
4127 		 * Should never happen
4128 		 */
4129 		return (EINVAL);
4130 
4131 	case sizeof (sin_t):
4132 		sin = (sin_t *)sa;
4133 		if (sin->sin_port == 0) {
4134 			return (-TBADADDR);
4135 		}
4136 		if (connp->conn_ipv6_v6only) {
4137 			return (EAFNOSUPPORT);
4138 		}
4139 		break;
4140 
4141 	case sizeof (sin6_t):
4142 		sin6 = (sin6_t *)sa;
4143 		if (sin6->sin6_port == 0) {
4144 			return (-TBADADDR);
4145 		}
4146 		break;
4147 	}
4148 	/*
4149 	 * If we're connecting to an IPv4-mapped IPv6 address, we need to
4150 	 * make sure that the conn_ipversion is IPV4_VERSION.  We
4151 	 * need to this before we call tcp_bindi() so that the port lookup
4152 	 * code will look for ports in the correct port space (IPv4 and
4153 	 * IPv6 have separate port spaces).
4154 	 */
4155 	if (connp->conn_family == AF_INET6 &&
4156 	    connp->conn_ipversion == IPV6_VERSION &&
4157 	    IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4158 		if (connp->conn_ipv6_v6only)
4159 			return (EADDRNOTAVAIL);
4160 
4161 		connp->conn_ipversion = IPV4_VERSION;
4162 	}
4163 
4164 	switch (tcp->tcp_state) {
4165 	case TCPS_LISTEN:
4166 		/*
4167 		 * Listening sockets are not allowed to issue connect().
4168 		 */
4169 		if (IPCL_IS_NONSTR(connp))
4170 			return (EOPNOTSUPP);
4171 		/* FALLTHRU */
4172 	case TCPS_IDLE:
4173 		/*
4174 		 * We support quick connect, refer to comments in
4175 		 * tcp_connect_*()
4176 		 */
4177 		/* FALLTHRU */
4178 	case TCPS_BOUND:
4179 		break;
4180 	default:
4181 		return (-TOUTSTATE);
4182 	}
4183 
4184 	/*
4185 	 * We update our cred/cpid based on the caller of connect
4186 	 */
4187 	if (connp->conn_cred != cr) {
4188 		crhold(cr);
4189 		crfree(connp->conn_cred);
4190 		connp->conn_cred = cr;
4191 	}
4192 	connp->conn_cpid = pid;
4193 
4194 	/* Cache things in the ixa without any refhold */
4195 	ASSERT(!(ixa->ixa_free_flags & IXA_FREE_CRED));
4196 	ixa->ixa_cred = cr;
4197 	ixa->ixa_cpid = pid;
4198 	if (is_system_labeled()) {
4199 		/* We need to restart with a label based on the cred */
4200 		ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4201 	}
4202 
4203 	if (connp->conn_family == AF_INET6) {
4204 		if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4205 			error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4206 			    sin6->sin6_port, sin6->sin6_flowinfo,
4207 			    sin6->__sin6_src_id, sin6->sin6_scope_id);
4208 		} else {
4209 			/*
4210 			 * Destination adress is mapped IPv6 address.
4211 			 * Source bound address should be unspecified or
4212 			 * IPv6 mapped address as well.
4213 			 */
4214 			if (!IN6_IS_ADDR_UNSPECIFIED(
4215 			    &connp->conn_bound_addr_v6) &&
4216 			    !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4217 				return (EADDRNOTAVAIL);
4218 			}
4219 			dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4220 			dstport = sin6->sin6_port;
4221 			srcid = sin6->__sin6_src_id;
4222 			error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4223 			    srcid);
4224 		}
4225 	} else {
4226 		dstaddrp = &sin->sin_addr.s_addr;
4227 		dstport = sin->sin_port;
4228 		srcid = 0;
4229 		error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4230 	}
4231 
4232 	if (error != 0)
4233 		goto connect_failed;
4234 
4235 	CL_INET_CONNECT(connp, B_TRUE, error);
4236 	if (error != 0)
4237 		goto connect_failed;
4238 
4239 	/* connect succeeded */
4240 	TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4241 	tcp->tcp_active_open = 1;
4242 
4243 	/*
4244 	 * tcp_set_destination() does not adjust for TCP/IP header length.
4245 	 */
4246 	mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4247 
4248 	/*
4249 	 * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4250 	 * to the nearest MSS.
4251 	 *
4252 	 * We do the round up here because we need to get the interface MTU
4253 	 * first before we can do the round up.
4254 	 */
4255 	tcp->tcp_rwnd = connp->conn_rcvbuf;
4256 	tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4257 	    tcps->tcps_recv_hiwat_minmss * mss);
4258 	connp->conn_rcvbuf = tcp->tcp_rwnd;
4259 	tcp_set_ws_value(tcp);
4260 	tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4261 	if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4262 		tcp->tcp_snd_ws_ok = B_TRUE;
4263 
4264 	/*
4265 	 * Set tcp_snd_ts_ok to true
4266 	 * so that tcp_xmit_mp will
4267 	 * include the timestamp
4268 	 * option in the SYN segment.
4269 	 */
4270 	if (tcps->tcps_tstamp_always ||
4271 	    (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4272 		tcp->tcp_snd_ts_ok = B_TRUE;
4273 	}
4274 
4275 	/*
4276 	 * Note that tcp_snd_sack_ok can be set in tcp_set_destination() if
4277 	 * the SACK metric is set.  So here we just check the per stack SACK
4278 	 * permitted param.
4279 	 */
4280 	if (tcps->tcps_sack_permitted == 2) {
4281 		ASSERT(tcp->tcp_num_sack_blk == 0);
4282 		ASSERT(tcp->tcp_notsack_list == NULL);
4283 		tcp->tcp_snd_sack_ok = B_TRUE;
4284 	}
4285 
4286 	/*
4287 	 * Should we use ECN?  Note that the current
4288 	 * default value (SunOS 5.9) of tcp_ecn_permitted
4289 	 * is 1.  The reason for doing this is that there
4290 	 * are equipments out there that will drop ECN
4291 	 * enabled IP packets.  Setting it to 1 avoids
4292 	 * compatibility problems.
4293 	 */
4294 	if (tcps->tcps_ecn_permitted == 2)
4295 		tcp->tcp_ecn_ok = B_TRUE;
4296 
4297 	/* Trace change from BOUND -> SYN_SENT here */
4298 	DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4299 	    connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4300 	    int32_t, TCPS_BOUND);
4301 
4302 	TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4303 	syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4304 	    tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4305 	if (syn_mp != NULL) {
4306 		/*
4307 		 * We must bump the generation before sending the syn
4308 		 * to ensure that we use the right generation in case
4309 		 * this thread issues a "connected" up call.
4310 		 */
4311 		SOCK_CONNID_BUMP(tcp->tcp_connid);
4312 		/*
4313 		 * DTrace sending the first SYN as a
4314 		 * tcp:::connect-request event.
4315 		 */
4316 		DTRACE_TCP5(connect__request, mblk_t *, NULL,
4317 		    ip_xmit_attr_t *, connp->conn_ixa,
4318 		    void_ip_t *, syn_mp->b_rptr, tcp_t *, tcp,
4319 		    tcph_t *,
4320 		    &syn_mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
4321 		tcp_send_data(tcp, syn_mp);
4322 	}
4323 
4324 	if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4325 		tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4326 	return (0);
4327 
4328 connect_failed:
4329 	connp->conn_faddr_v6 = ipv6_all_zeros;
4330 	connp->conn_fport = 0;
4331 	tcp->tcp_state = oldstate;
4332 	if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4333 		tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4334 	return (error);
4335 }
4336 
4337 int
tcp_do_listen(conn_t * connp,struct sockaddr * sa,socklen_t len,int backlog,cred_t * cr,boolean_t bind_to_req_port_only)4338 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4339     int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4340 {
4341 	tcp_t		*tcp = connp->conn_tcp;
4342 	int		error = 0;
4343 	tcp_stack_t	*tcps = tcp->tcp_tcps;
4344 	int32_t		oldstate;
4345 
4346 	/* All Solaris components should pass a cred for this operation. */
4347 	ASSERT(cr != NULL);
4348 
4349 	if (tcp->tcp_state >= TCPS_BOUND) {
4350 		if ((tcp->tcp_state == TCPS_BOUND ||
4351 		    tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
4352 			/*
4353 			 * Handle listen() increasing backlog.
4354 			 * This is more "liberal" then what the TPI spec
4355 			 * requires but is needed to avoid a t_unbind
4356 			 * when handling listen() since the port number
4357 			 * might be "stolen" between the unbind and bind.
4358 			 */
4359 			goto do_listen;
4360 		}
4361 		if (connp->conn_debug) {
4362 			(void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4363 			    "tcp_listen: bad state, %d", tcp->tcp_state);
4364 		}
4365 		return (-TOUTSTATE);
4366 	} else {
4367 		sin6_t	addr;
4368 		sin_t *sin;
4369 		sin6_t *sin6;
4370 
4371 		if (sa == NULL) {
4372 			ASSERT(IPCL_IS_NONSTR(connp));
4373 			/* Do an implicit bind: Request for a generic port. */
4374 			if (connp->conn_family == AF_INET) {
4375 				len = sizeof (sin_t);
4376 				sin = (sin_t *)&addr;
4377 				*sin = sin_null;
4378 				sin->sin_family = AF_INET;
4379 			} else {
4380 				ASSERT(connp->conn_family == AF_INET6);
4381 				len = sizeof (sin6_t);
4382 				sin6 = (sin6_t *)&addr;
4383 				*sin6 = sin6_null;
4384 				sin6->sin6_family = AF_INET6;
4385 			}
4386 			sa = (struct sockaddr *)&addr;
4387 		}
4388 
4389 		error = tcp_bind_check(connp, sa, len, cr,
4390 		    bind_to_req_port_only);
4391 		if (error)
4392 			return (error);
4393 		/* Fall through and do the fanout insertion */
4394 	}
4395 
4396 do_listen:
4397 	ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
4398 	tcp->tcp_conn_req_max = backlog;
4399 	if (tcp->tcp_conn_req_max) {
4400 		if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
4401 			tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
4402 		if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
4403 			tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
4404 		/*
4405 		 * If this is a listener, do not reset the eager list
4406 		 * and other stuffs.  Note that we don't check if the
4407 		 * existing eager list meets the new tcp_conn_req_max
4408 		 * requirement.
4409 		 */
4410 		if (tcp->tcp_state != TCPS_LISTEN) {
4411 			tcp->tcp_state = TCPS_LISTEN;
4412 			DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4413 			    connp->conn_ixa, void, NULL, tcp_t *, tcp,
4414 			    void, NULL, int32_t, TCPS_BOUND);
4415 			/* Initialize the chain. Don't need the eager_lock */
4416 			tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
4417 			tcp->tcp_eager_next_drop_q0 = tcp;
4418 			tcp->tcp_eager_prev_drop_q0 = tcp;
4419 			tcp->tcp_second_ctimer_threshold =
4420 			    tcps->tcps_ip_abort_linterval;
4421 		}
4422 	}
4423 
4424 	/*
4425 	 * We need to make sure that the conn_recv is set to a non-null
4426 	 * value before we insert the conn into the classifier table.
4427 	 * This is to avoid a race with an incoming packet which does an
4428 	 * ipcl_classify().
4429 	 * We initially set it to tcp_input_listener_unbound to try to
4430 	 * pick a good squeue for the listener when the first SYN arrives.
4431 	 * tcp_input_listener_unbound sets it to tcp_input_listener on that
4432 	 * first SYN.
4433 	 */
4434 	connp->conn_recv = tcp_input_listener_unbound;
4435 
4436 	/* Insert the listener in the classifier table */
4437 	error = ip_laddr_fanout_insert(connp);
4438 	if (error != 0) {
4439 		/* Undo the bind - release the port number */
4440 		oldstate = tcp->tcp_state;
4441 		tcp->tcp_state = TCPS_IDLE;
4442 		DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4443 		    connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4444 		    int32_t, oldstate);
4445 		connp->conn_bound_addr_v6 = ipv6_all_zeros;
4446 
4447 		connp->conn_laddr_v6 = ipv6_all_zeros;
4448 		connp->conn_saddr_v6 = ipv6_all_zeros;
4449 		connp->conn_ports = 0;
4450 
4451 		if (connp->conn_anon_port) {
4452 			zone_t		*zone;
4453 
4454 			zone = crgetzone(cr);
4455 			connp->conn_anon_port = B_FALSE;
4456 			(void) tsol_mlp_anon(zone, connp->conn_mlp_type,
4457 			    connp->conn_proto, connp->conn_lport, B_FALSE);
4458 		}
4459 		connp->conn_mlp_type = mlptSingle;
4460 
4461 		tcp_bind_hash_remove(tcp);
4462 		return (error);
4463 	} else {
4464 		/*
4465 		 * If there is a connection limit, allocate and initialize
4466 		 * the counter struct.  Note that since listen can be called
4467 		 * multiple times, the struct may have been allready allocated.
4468 		 */
4469 		if (!list_is_empty(&tcps->tcps_listener_conf) &&
4470 		    tcp->tcp_listen_cnt == NULL) {
4471 			tcp_listen_cnt_t *tlc;
4472 			uint32_t ratio;
4473 
4474 			ratio = tcp_find_listener_conf(tcps,
4475 			    ntohs(connp->conn_lport));
4476 			if (ratio != 0) {
4477 				uint32_t mem_ratio, tot_buf;
4478 
4479 				tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
4480 				    KM_SLEEP);
4481 				/*
4482 				 * Calculate the connection limit based on
4483 				 * the configured ratio and maxusers.  Maxusers
4484 				 * are calculated based on memory size,
4485 				 * ~ 1 user per MB.  Note that the conn_rcvbuf
4486 				 * and conn_sndbuf may change after a
4487 				 * connection is accepted.  So what we have
4488 				 * is only an approximation.
4489 				 */
4490 				if ((tot_buf = connp->conn_rcvbuf +
4491 				    connp->conn_sndbuf) < MB) {
4492 					mem_ratio = MB / tot_buf;
4493 					tlc->tlc_max = maxusers / ratio *
4494 					    mem_ratio;
4495 				} else {
4496 					mem_ratio = tot_buf / MB;
4497 					tlc->tlc_max = maxusers / ratio /
4498 					    mem_ratio;
4499 				}
4500 				/* At least we should allow two connections! */
4501 				if (tlc->tlc_max <= tcp_min_conn_listener)
4502 					tlc->tlc_max = tcp_min_conn_listener;
4503 				tlc->tlc_cnt = 1;
4504 				tlc->tlc_drop = 0;
4505 				tcp->tcp_listen_cnt = tlc;
4506 			}
4507 		}
4508 	}
4509 	return (error);
4510 }
4511