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, Joyent Inc. All rights reserved.
25 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
26 */
27 /* Copyright (c) 1990 Mentat Inc. */
28
29 #include <sys/types.h>
30 #include <sys/stream.h>
31 #include <sys/strsun.h>
32 #include <sys/strsubr.h>
33 #include <sys/stropts.h>
34 #include <sys/strlog.h>
35 #define _SUN_TPI_VERSION 2
36 #include <sys/tihdr.h>
37 #include <sys/timod.h>
38 #include <sys/ddi.h>
39 #include <sys/sunddi.h>
40 #include <sys/suntpi.h>
41 #include <sys/xti_inet.h>
42 #include <sys/cmn_err.h>
43 #include <sys/debug.h>
44 #include <sys/sdt.h>
45 #include <sys/vtrace.h>
46 #include <sys/kmem.h>
47 #include <sys/ethernet.h>
48 #include <sys/cpuvar.h>
49 #include <sys/dlpi.h>
50 #include <sys/pattr.h>
51 #include <sys/policy.h>
52 #include <sys/priv.h>
53 #include <sys/zone.h>
54 #include <sys/sunldi.h>
55
56 #include <sys/errno.h>
57 #include <sys/signal.h>
58 #include <sys/socket.h>
59 #include <sys/socketvar.h>
60 #include <sys/sockio.h>
61 #include <sys/isa_defs.h>
62 #include <sys/md5.h>
63 #include <sys/random.h>
64 #include <sys/uio.h>
65 #include <sys/systm.h>
66 #include <netinet/in.h>
67 #include <netinet/tcp.h>
68 #include <netinet/ip6.h>
69 #include <netinet/icmp6.h>
70 #include <net/if.h>
71 #include <net/route.h>
72 #include <inet/ipsec_impl.h>
73
74 #include <inet/common.h>
75 #include <inet/ip.h>
76 #include <inet/ip_impl.h>
77 #include <inet/ip6.h>
78 #include <inet/ip_ndp.h>
79 #include <inet/proto_set.h>
80 #include <inet/mib2.h>
81 #include <inet/optcom.h>
82 #include <inet/snmpcom.h>
83 #include <inet/kstatcom.h>
84 #include <inet/tcp.h>
85 #include <inet/tcp_impl.h>
86 #include <inet/tcp_cluster.h>
87 #include <inet/udp_impl.h>
88 #include <net/pfkeyv2.h>
89 #include <inet/ipdrop.h>
90
91 #include <inet/ipclassifier.h>
92 #include <inet/ip_ire.h>
93 #include <inet/ip_ftable.h>
94 #include <inet/ip_if.h>
95 #include <inet/ipp_common.h>
96 #include <inet/ip_rts.h>
97 #include <inet/ip_netinfo.h>
98 #include <sys/squeue_impl.h>
99 #include <sys/squeue.h>
100 #include <sys/tsol/label.h>
101 #include <sys/tsol/tnet.h>
102 #include <rpc/pmap_prot.h>
103 #include <sys/callo.h>
104
105 /*
106 * TCP Notes: aka FireEngine Phase I (PSARC 2002/433)
107 *
108 * (Read the detailed design doc in PSARC case directory)
109 *
110 * The entire tcp state is contained in tcp_t and conn_t structure
111 * which are allocated in tandem using ipcl_conn_create() and passing
112 * IPCL_TCPCONN as a flag. We use 'conn_ref' and 'conn_lock' to protect
113 * the references on the tcp_t. The tcp_t structure is never compressed
114 * and packets always land on the correct TCP perimeter from the time
115 * eager is created till the time tcp_t dies (as such the old mentat
116 * TCP global queue is not used for detached state and no IPSEC checking
117 * is required). The global queue is still allocated to send out resets
118 * for connection which have no listeners and IP directly calls
119 * tcp_xmit_listeners_reset() which does any policy check.
120 *
121 * Protection and Synchronisation mechanism:
122 *
123 * The tcp data structure does not use any kind of lock for protecting
124 * its state but instead uses 'squeues' for mutual exclusion from various
125 * read and write side threads. To access a tcp member, the thread should
126 * always be behind squeue (via squeue_enter with flags as SQ_FILL, SQ_PROCESS,
127 * or SQ_NODRAIN). Since the squeues allow a direct function call, caller
128 * can pass any tcp function having prototype of edesc_t as argument
129 * (different from traditional STREAMs model where packets come in only
130 * designated entry points). The list of functions that can be directly
131 * called via squeue are listed before the usual function prototype.
132 *
133 * Referencing:
134 *
135 * TCP is MT-Hot and we use a reference based scheme to make sure that the
136 * tcp structure doesn't disappear when its needed. When the application
137 * creates an outgoing connection or accepts an incoming connection, we
138 * start out with 2 references on 'conn_ref'. One for TCP and one for IP.
139 * The IP reference is just a symbolic reference since ip_tcpclose()
140 * looks at tcp structure after tcp_close_output() returns which could
141 * have dropped the last TCP reference. So as long as the connection is
142 * in attached state i.e. !TCP_IS_DETACHED, we have 2 references on the
143 * conn_t. The classifier puts its own reference when the connection is
144 * inserted in listen or connected hash. Anytime a thread needs to enter
145 * the tcp connection perimeter, it retrieves the conn/tcp from q->ptr
146 * on write side or by doing a classify on read side and then puts a
147 * reference on the conn before doing squeue_enter/tryenter/fill. For
148 * read side, the classifier itself puts the reference under fanout lock
149 * to make sure that tcp can't disappear before it gets processed. The
150 * squeue will drop this reference automatically so the called function
151 * doesn't have to do a DEC_REF.
152 *
153 * Opening a new connection:
154 *
155 * The outgoing connection open is pretty simple. tcp_open() does the
156 * work in creating the conn/tcp structure and initializing it. The
157 * squeue assignment is done based on the CPU the application
158 * is running on. So for outbound connections, processing is always done
159 * on application CPU which might be different from the incoming CPU
160 * being interrupted by the NIC. An optimal way would be to figure out
161 * the NIC <-> CPU binding at listen time, and assign the outgoing
162 * connection to the squeue attached to the CPU that will be interrupted
163 * for incoming packets (we know the NIC based on the bind IP address).
164 * This might seem like a problem if more data is going out but the
165 * fact is that in most cases the transmit is ACK driven transmit where
166 * the outgoing data normally sits on TCP's xmit queue waiting to be
167 * transmitted.
168 *
169 * Accepting a connection:
170 *
171 * This is a more interesting case because of various races involved in
172 * establishing a eager in its own perimeter. Read the meta comment on
173 * top of tcp_input_listener(). But briefly, the squeue is picked by
174 * ip_fanout based on the ring or the sender (if loopback).
175 *
176 * Closing a connection:
177 *
178 * The close is fairly straight forward. tcp_close() calls tcp_close_output()
179 * via squeue to do the close and mark the tcp as detached if the connection
180 * was in state TCPS_ESTABLISHED or greater. In the later case, TCP keep its
181 * reference but tcp_close() drop IP's reference always. So if tcp was
182 * not killed, it is sitting in time_wait list with 2 reference - 1 for TCP
183 * and 1 because it is in classifier's connected hash. This is the condition
184 * we use to determine that its OK to clean up the tcp outside of squeue
185 * when time wait expires (check the ref under fanout and conn_lock and
186 * if it is 2, remove it from fanout hash and kill it).
187 *
188 * Although close just drops the necessary references and marks the
189 * tcp_detached state, tcp_close needs to know the tcp_detached has been
190 * set (under squeue) before letting the STREAM go away (because a
191 * inbound packet might attempt to go up the STREAM while the close
192 * has happened and tcp_detached is not set). So a special lock and
193 * flag is used along with a condition variable (tcp_closelock, tcp_closed,
194 * and tcp_closecv) to signal tcp_close that tcp_close_out() has marked
195 * tcp_detached.
196 *
197 * Special provisions and fast paths:
198 *
199 * We make special provisions for sockfs by marking tcp_issocket
200 * whenever we have only sockfs on top of TCP. This allows us to skip
201 * putting the tcp in acceptor hash since a sockfs listener can never
202 * become acceptor and also avoid allocating a tcp_t for acceptor STREAM
203 * since eager has already been allocated and the accept now happens
204 * on acceptor STREAM. There is a big blob of comment on top of
205 * tcp_input_listener explaining the new accept. When socket is POP'd,
206 * sockfs sends us an ioctl to mark the fact and we go back to old
207 * behaviour. Once tcp_issocket is unset, its never set for the
208 * life of that connection.
209 *
210 * IPsec notes :
211 *
212 * Since a packet is always executed on the correct TCP perimeter
213 * all IPsec processing is defered to IP including checking new
214 * connections and setting IPSEC policies for new connection. The
215 * only exception is tcp_xmit_listeners_reset() which is called
216 * directly from IP and needs to policy check to see if TH_RST
217 * can be sent out.
218 */
219
220 /*
221 * Values for squeue switch:
222 * 1: SQ_NODRAIN
223 * 2: SQ_PROCESS
224 * 3: SQ_FILL
225 */
226 int tcp_squeue_wput = 2; /* /etc/systems */
227 int tcp_squeue_flag;
228
229 /*
230 * To prevent memory hog, limit the number of entries in tcp_free_list
231 * to 1% of available memory / number of cpus
232 */
233 uint_t tcp_free_list_max_cnt = 0;
234
235 #define TCP_XMIT_LOWATER 4096
236 #define TCP_XMIT_HIWATER 49152
237 #define TCP_RECV_LOWATER 2048
238 #define TCP_RECV_HIWATER 128000
239
240 #define TIDUSZ 4096 /* transport interface data unit size */
241
242 /*
243 * Size of acceptor hash list. It has to be a power of 2 for hashing.
244 */
245 #define TCP_ACCEPTOR_FANOUT_SIZE 512
246
247 #ifdef _ILP32
248 #define TCP_ACCEPTOR_HASH(accid) \
249 (((uint_t)(accid) >> 8) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
250 #else
251 #define TCP_ACCEPTOR_HASH(accid) \
252 ((uint_t)(accid) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
253 #endif /* _ILP32 */
254
255 /*
256 * Minimum number of connections which can be created per listener. Used
257 * when the listener connection count is in effect.
258 */
259 static uint32_t tcp_min_conn_listener = 2;
260
261 uint32_t tcp_early_abort = 30;
262
263 /* TCP Timer control structure */
264 typedef struct tcpt_s {
265 pfv_t tcpt_pfv; /* The routine we are to call */
266 tcp_t *tcpt_tcp; /* The parameter we are to pass in */
267 } tcpt_t;
268
269 /*
270 * Functions called directly via squeue having a prototype of edesc_t.
271 */
272 void tcp_input_listener(void *arg, mblk_t *mp, void *arg2,
273 ip_recv_attr_t *ira);
274 void tcp_input_data(void *arg, mblk_t *mp, void *arg2,
275 ip_recv_attr_t *ira);
276 static void tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2,
277 ip_recv_attr_t *dummy);
278
279
280 /* Prototype for TCP functions */
281 static void tcp_random_init(void);
282 int tcp_random(void);
283 static int tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp,
284 in_port_t dstport, uint_t srcid);
285 static int tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp,
286 in_port_t dstport, uint32_t flowinfo,
287 uint_t srcid, uint32_t scope_id);
288 static void tcp_iss_init(tcp_t *tcp);
289 static void tcp_reinit(tcp_t *tcp);
290 static void tcp_reinit_values(tcp_t *tcp);
291
292 static void tcp_wsrv(queue_t *q);
293 static void tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa);
294 static void tcp_update_zcopy(tcp_t *tcp);
295 static void tcp_notify(void *, ip_xmit_attr_t *, ixa_notify_type_t,
296 ixa_notify_arg_t);
297 static void *tcp_stack_init(netstackid_t stackid, netstack_t *ns);
298 static void tcp_stack_fini(netstackid_t stackid, void *arg);
299
300 static int tcp_squeue_switch(int);
301
302 static int tcp_open(queue_t *, dev_t *, int, int, cred_t *, boolean_t);
303 static int tcp_openv4(queue_t *, dev_t *, int, int, cred_t *);
304 static int tcp_openv6(queue_t *, dev_t *, int, int, cred_t *);
305
306 static void tcp_squeue_add(squeue_t *);
307
308 struct module_info tcp_rinfo = {
309 TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, TCP_RECV_HIWATER, TCP_RECV_LOWATER
310 };
311
312 static struct module_info tcp_winfo = {
313 TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, 127, 16
314 };
315
316 /*
317 * Entry points for TCP as a device. The normal case which supports
318 * the TCP functionality.
319 * We have separate open functions for the /dev/tcp and /dev/tcp6 devices.
320 */
321 struct qinit tcp_rinitv4 = {
322 NULL, (pfi_t)tcp_rsrv, tcp_openv4, tcp_tpi_close, NULL, &tcp_rinfo
323 };
324
325 struct qinit tcp_rinitv6 = {
326 NULL, (pfi_t)tcp_rsrv, tcp_openv6, tcp_tpi_close, NULL, &tcp_rinfo
327 };
328
329 struct qinit tcp_winit = {
330 (pfi_t)tcp_wput, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
331 };
332
333 /* Initial entry point for TCP in socket mode. */
334 struct qinit tcp_sock_winit = {
335 (pfi_t)tcp_wput_sock, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
336 };
337
338 /* TCP entry point during fallback */
339 struct qinit tcp_fallback_sock_winit = {
340 (pfi_t)tcp_wput_fallback, NULL, NULL, NULL, NULL, &tcp_winfo
341 };
342
343 /*
344 * Entry points for TCP as a acceptor STREAM opened by sockfs when doing
345 * an accept. Avoid allocating data structures since eager has already
346 * been created.
347 */
348 struct qinit tcp_acceptor_rinit = {
349 NULL, (pfi_t)tcp_rsrv, NULL, tcp_tpi_close_accept, NULL, &tcp_winfo
350 };
351
352 struct qinit tcp_acceptor_winit = {
353 (pfi_t)tcp_tpi_accept, NULL, NULL, NULL, NULL, &tcp_winfo
354 };
355
356 /* For AF_INET aka /dev/tcp */
357 struct streamtab tcpinfov4 = {
358 &tcp_rinitv4, &tcp_winit
359 };
360
361 /* For AF_INET6 aka /dev/tcp6 */
362 struct streamtab tcpinfov6 = {
363 &tcp_rinitv6, &tcp_winit
364 };
365
366 /*
367 * Following assumes TPI alignment requirements stay along 32 bit
368 * boundaries
369 */
370 #define ROUNDUP32(x) \
371 (((x) + (sizeof (int32_t) - 1)) & ~(sizeof (int32_t) - 1))
372
373 /* Template for response to info request. */
374 struct T_info_ack tcp_g_t_info_ack = {
375 T_INFO_ACK, /* PRIM_type */
376 0, /* TSDU_size */
377 T_INFINITE, /* ETSDU_size */
378 T_INVALID, /* CDATA_size */
379 T_INVALID, /* DDATA_size */
380 sizeof (sin_t), /* ADDR_size */
381 0, /* OPT_size - not initialized here */
382 TIDUSZ, /* TIDU_size */
383 T_COTS_ORD, /* SERV_type */
384 TCPS_IDLE, /* CURRENT_state */
385 (XPG4_1|EXPINLINE) /* PROVIDER_flag */
386 };
387
388 struct T_info_ack tcp_g_t_info_ack_v6 = {
389 T_INFO_ACK, /* PRIM_type */
390 0, /* TSDU_size */
391 T_INFINITE, /* ETSDU_size */
392 T_INVALID, /* CDATA_size */
393 T_INVALID, /* DDATA_size */
394 sizeof (sin6_t), /* ADDR_size */
395 0, /* OPT_size - not initialized here */
396 TIDUSZ, /* TIDU_size */
397 T_COTS_ORD, /* SERV_type */
398 TCPS_IDLE, /* CURRENT_state */
399 (XPG4_1|EXPINLINE) /* PROVIDER_flag */
400 };
401
402 /*
403 * TCP tunables related declarations. Definitions are in tcp_tunables.c
404 */
405 extern mod_prop_info_t tcp_propinfo_tbl[];
406 extern int tcp_propinfo_count;
407
408 #define IS_VMLOANED_MBLK(mp) \
409 (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0)
410
411 uint32_t do_tcpzcopy = 1; /* 0: disable, 1: enable, 2: force */
412
413 /*
414 * Forces all connections to obey the value of the tcps_maxpsz_multiplier
415 * tunable settable via NDD. Otherwise, the per-connection behavior is
416 * determined dynamically during tcp_set_destination(), which is the default.
417 */
418 boolean_t tcp_static_maxpsz = B_FALSE;
419
420 /*
421 * If the receive buffer size is changed, this function is called to update
422 * the upper socket layer on the new delayed receive wake up threshold.
423 */
424 static void
425 tcp_set_recv_threshold(tcp_t *tcp, uint32_t new_rcvthresh)
426 {
427 uint32_t default_threshold = SOCKET_RECVHIWATER >> 3;
428
429 if (IPCL_IS_NONSTR(tcp->tcp_connp)) {
430 conn_t *connp = tcp->tcp_connp;
431 struct sock_proto_props sopp;
432
433 /*
434 * only increase rcvthresh upto default_threshold
435 */
436 if (new_rcvthresh > default_threshold)
437 new_rcvthresh = default_threshold;
438
439 sopp.sopp_flags = SOCKOPT_RCVTHRESH;
440 sopp.sopp_rcvthresh = new_rcvthresh;
441
442 (*connp->conn_upcalls->su_set_proto_props)
443 (connp->conn_upper_handle, &sopp);
444 }
445 }
446
447 /*
448 * Figure out the value of window scale opton. Note that the rwnd is
449 * ASSUMED to be rounded up to the nearest MSS before the calculation.
450 * We cannot find the scale value and then do a round up of tcp_rwnd
451 * because the scale value may not be correct after that.
452 *
453 * Set the compiler flag to make this function inline.
454 */
455 void
456 tcp_set_ws_value(tcp_t *tcp)
457 {
458 int i;
459 uint32_t rwnd = tcp->tcp_rwnd;
460
461 for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT;
462 i++, rwnd >>= 1)
463 ;
464 tcp->tcp_rcv_ws = i;
465 }
466
467 /*
468 * Remove cached/latched IPsec references.
469 */
470 void
471 tcp_ipsec_cleanup(tcp_t *tcp)
472 {
473 conn_t *connp = tcp->tcp_connp;
474
475 ASSERT(connp->conn_flags & IPCL_TCPCONN);
476
477 if (connp->conn_latch != NULL) {
478 IPLATCH_REFRELE(connp->conn_latch);
479 connp->conn_latch = NULL;
480 }
481 if (connp->conn_latch_in_policy != NULL) {
482 IPPOL_REFRELE(connp->conn_latch_in_policy);
483 connp->conn_latch_in_policy = NULL;
484 }
485 if (connp->conn_latch_in_action != NULL) {
486 IPACT_REFRELE(connp->conn_latch_in_action);
487 connp->conn_latch_in_action = NULL;
488 }
489 if (connp->conn_policy != NULL) {
490 IPPH_REFRELE(connp->conn_policy, connp->conn_netstack);
491 connp->conn_policy = NULL;
492 }
493 }
494
495 /*
496 * Cleaup before placing on free list.
497 * Disassociate from the netstack/tcp_stack_t since the freelist
498 * is per squeue and not per netstack.
499 */
500 void
501 tcp_cleanup(tcp_t *tcp)
502 {
503 mblk_t *mp;
504 conn_t *connp = tcp->tcp_connp;
505 tcp_stack_t *tcps = tcp->tcp_tcps;
506 netstack_t *ns = tcps->tcps_netstack;
507 mblk_t *tcp_rsrv_mp;
508
509 tcp_bind_hash_remove(tcp);
510
511 /* Cleanup that which needs the netstack first */
512 tcp_ipsec_cleanup(tcp);
513 ixa_cleanup(connp->conn_ixa);
514
515 if (connp->conn_ht_iphc != NULL) {
516 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
517 connp->conn_ht_iphc = NULL;
518 connp->conn_ht_iphc_allocated = 0;
519 connp->conn_ht_iphc_len = 0;
520 connp->conn_ht_ulp = NULL;
521 connp->conn_ht_ulp_len = 0;
522 tcp->tcp_ipha = NULL;
523 tcp->tcp_ip6h = NULL;
524 tcp->tcp_tcpha = NULL;
525 }
526
527 /* We clear any IP_OPTIONS and extension headers */
528 ip_pkt_free(&connp->conn_xmit_ipp);
529
530 tcp_free(tcp);
531
532 /*
533 * Since we will bzero the entire structure, we need to
534 * remove it and reinsert it in global hash list. We
535 * know the walkers can't get to this conn because we
536 * had set CONDEMNED flag earlier and checked reference
537 * under conn_lock so walker won't pick it and when we
538 * go the ipcl_globalhash_remove() below, no walker
539 * can get to it.
540 */
541 ipcl_globalhash_remove(connp);
542
543 /* Save some state */
544 mp = tcp->tcp_timercache;
545
546 tcp_rsrv_mp = tcp->tcp_rsrv_mp;
547
548 if (connp->conn_cred != NULL) {
549 crfree(connp->conn_cred);
550 connp->conn_cred = NULL;
551 }
552 ipcl_conn_cleanup(connp);
553 connp->conn_flags = IPCL_TCPCONN;
554
555 /*
556 * Now it is safe to decrement the reference counts.
557 * This might be the last reference on the netstack
558 * in which case it will cause the freeing of the IP Instance.
559 */
560 connp->conn_netstack = NULL;
561 connp->conn_ixa->ixa_ipst = NULL;
562 netstack_rele(ns);
563 ASSERT(tcps != NULL);
564 tcp->tcp_tcps = NULL;
565
566 bzero(tcp, sizeof (tcp_t));
567
568 /* restore the state */
569 tcp->tcp_timercache = mp;
570
571 tcp->tcp_rsrv_mp = tcp_rsrv_mp;
572
573 tcp->tcp_connp = connp;
574
575 ASSERT(connp->conn_tcp == tcp);
576 ASSERT(connp->conn_flags & IPCL_TCPCONN);
577 connp->conn_state_flags = CONN_INCIPIENT;
578 ASSERT(connp->conn_proto == IPPROTO_TCP);
579 ASSERT(connp->conn_ref == 1);
580 }
581
582 /*
583 * Adapt to the information, such as rtt and rtt_sd, provided from the
584 * DCE and IRE maintained by IP.
585 *
586 * Checks for multicast and broadcast destination address.
587 * Returns zero if ok; an errno on failure.
588 *
589 * Note that the MSS calculation here is based on the info given in
590 * the DCE and IRE. We do not do any calculation based on TCP options. They
591 * will be handled in tcp_input_data() when TCP knows which options to use.
592 *
593 * Note on how TCP gets its parameters for a connection.
594 *
595 * When a tcp_t structure is allocated, it gets all the default parameters.
596 * In tcp_set_destination(), it gets those metric parameters, like rtt, rtt_sd,
597 * spipe, rpipe, ... from the route metrics. Route metric overrides the
598 * default.
599 *
600 * An incoming SYN with a multicast or broadcast destination address is dropped
601 * in ip_fanout_v4/v6.
602 *
603 * An incoming SYN with a multicast or broadcast source address is always
604 * dropped in tcp_set_destination, since IPDF_ALLOW_MCBC is not set in
605 * conn_connect.
606 * The same logic in tcp_set_destination also serves to
607 * reject an attempt to connect to a broadcast or multicast (destination)
608 * address.
609 */
610 int
611 tcp_set_destination(tcp_t *tcp)
612 {
613 uint32_t mss_max;
614 uint32_t mss;
615 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
616 conn_t *connp = tcp->tcp_connp;
617 tcp_stack_t *tcps = tcp->tcp_tcps;
618 iulp_t uinfo;
619 int error;
620 uint32_t flags;
621
622 flags = IPDF_LSO | IPDF_ZCOPY;
623 /*
624 * Make sure we have a dce for the destination to avoid dce_ident
625 * contention for connected sockets.
626 */
627 flags |= IPDF_UNIQUE_DCE;
628
629 if (!tcps->tcps_ignore_path_mtu)
630 connp->conn_ixa->ixa_flags |= IXAF_PMTU_DISCOVERY;
631
632 /* Use conn_lock to satify ASSERT; tcp is already serialized */
633 mutex_enter(&connp->conn_lock);
634 error = conn_connect(connp, &uinfo, flags);
635 mutex_exit(&connp->conn_lock);
636 if (error != 0)
637 return (error);
638
639 error = tcp_build_hdrs(tcp);
640 if (error != 0)
641 return (error);
642
643 tcp->tcp_localnet = uinfo.iulp_localnet;
644
645 if (uinfo.iulp_rtt != 0) {
646 clock_t rto;
647
648 tcp->tcp_rtt_sa = uinfo.iulp_rtt;
649 tcp->tcp_rtt_sd = uinfo.iulp_rtt_sd;
650 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
651 tcps->tcps_rexmit_interval_extra +
652 (tcp->tcp_rtt_sa >> 5);
653
654 TCP_SET_RTO(tcp, rto);
655 }
656 if (uinfo.iulp_ssthresh != 0)
657 tcp->tcp_cwnd_ssthresh = uinfo.iulp_ssthresh;
658 else
659 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
660 if (uinfo.iulp_spipe > 0) {
661 connp->conn_sndbuf = MIN(uinfo.iulp_spipe,
662 tcps->tcps_max_buf);
663 if (tcps->tcps_snd_lowat_fraction != 0) {
664 connp->conn_sndlowat = connp->conn_sndbuf /
665 tcps->tcps_snd_lowat_fraction;
666 }
667 (void) tcp_maxpsz_set(tcp, B_TRUE);
668 }
669 /*
670 * Note that up till now, acceptor always inherits receive
671 * window from the listener. But if there is a metrics
672 * associated with a host, we should use that instead of
673 * inheriting it from listener. Thus we need to pass this
674 * info back to the caller.
675 */
676 if (uinfo.iulp_rpipe > 0) {
677 tcp->tcp_rwnd = MIN(uinfo.iulp_rpipe,
678 tcps->tcps_max_buf);
679 }
680
681 if (uinfo.iulp_rtomax > 0) {
682 tcp->tcp_second_timer_threshold =
683 uinfo.iulp_rtomax;
684 }
685
686 /*
687 * Use the metric option settings, iulp_tstamp_ok and
688 * iulp_wscale_ok, only for active open. What this means
689 * is that if the other side uses timestamp or window
690 * scale option, TCP will also use those options. That
691 * is for passive open. If the application sets a
692 * large window, window scale is enabled regardless of
693 * the value in iulp_wscale_ok. This is the behavior
694 * since 2.6. So we keep it.
695 * The only case left in passive open processing is the
696 * check for SACK.
697 * For ECN, it should probably be like SACK. But the
698 * current value is binary, so we treat it like the other
699 * cases. The metric only controls active open.For passive
700 * open, the ndd param, tcp_ecn_permitted, controls the
701 * behavior.
702 */
703 if (!tcp_detached) {
704 /*
705 * The if check means that the following can only
706 * be turned on by the metrics only IRE, but not off.
707 */
708 if (uinfo.iulp_tstamp_ok)
709 tcp->tcp_snd_ts_ok = B_TRUE;
710 if (uinfo.iulp_wscale_ok)
711 tcp->tcp_snd_ws_ok = B_TRUE;
712 if (uinfo.iulp_sack == 2)
713 tcp->tcp_snd_sack_ok = B_TRUE;
714 if (uinfo.iulp_ecn_ok)
715 tcp->tcp_ecn_ok = B_TRUE;
716 } else {
717 /*
718 * Passive open.
719 *
720 * As above, the if check means that SACK can only be
721 * turned on by the metric only IRE.
722 */
723 if (uinfo.iulp_sack > 0) {
724 tcp->tcp_snd_sack_ok = B_TRUE;
725 }
726 }
727
728 /*
729 * XXX Note that currently, iulp_mtu can be as small as 68
730 * because of PMTUd. So tcp_mss may go to negative if combined
731 * length of all those options exceeds 28 bytes. But because
732 * of the tcp_mss_min check below, we may not have a problem if
733 * tcp_mss_min is of a reasonable value. The default is 1 so
734 * the negative problem still exists. And the check defeats PMTUd.
735 * In fact, if PMTUd finds that the MSS should be smaller than
736 * tcp_mss_min, TCP should turn off PMUTd and use the tcp_mss_min
737 * value.
738 *
739 * We do not deal with that now. All those problems related to
740 * PMTUd will be fixed later.
741 */
742 ASSERT(uinfo.iulp_mtu != 0);
743 mss = tcp->tcp_initial_pmtu = uinfo.iulp_mtu;
744
745 /* Sanity check for MSS value. */
746 if (connp->conn_ipversion == IPV4_VERSION)
747 mss_max = tcps->tcps_mss_max_ipv4;
748 else
749 mss_max = tcps->tcps_mss_max_ipv6;
750
751 if (tcp->tcp_ipsec_overhead == 0)
752 tcp->tcp_ipsec_overhead = conn_ipsec_length(connp);
753
754 mss -= tcp->tcp_ipsec_overhead;
755
756 if (mss < tcps->tcps_mss_min)
757 mss = tcps->tcps_mss_min;
758 if (mss > mss_max)
759 mss = mss_max;
760
761 /* Note that this is the maximum MSS, excluding all options. */
762 tcp->tcp_mss = mss;
763
764 /*
765 * Update the tcp connection with LSO capability.
766 */
767 tcp_update_lso(tcp, connp->conn_ixa);
768
769 /*
770 * Initialize the ISS here now that we have the full connection ID.
771 * The RFC 1948 method of initial sequence number generation requires
772 * knowledge of the full connection ID before setting the ISS.
773 */
774 tcp_iss_init(tcp);
775
776 tcp->tcp_loopback = (uinfo.iulp_loopback | uinfo.iulp_local);
777
778 /*
779 * Make sure that conn is not marked incipient
780 * for incoming connections. A blind
781 * removal of incipient flag is cheaper than
782 * check and removal.
783 */
784 mutex_enter(&connp->conn_lock);
785 connp->conn_state_flags &= ~CONN_INCIPIENT;
786 mutex_exit(&connp->conn_lock);
787 return (0);
788 }
789
790 /*
791 * tcp_clean_death / tcp_close_detached must not be called more than once
792 * on a tcp. Thus every function that potentially calls tcp_clean_death
793 * must check for the tcp state before calling tcp_clean_death.
794 * Eg. tcp_input_data, tcp_eager_kill, tcp_clean_death_wrapper,
795 * tcp_timer_handler, all check for the tcp state.
796 */
797 /* ARGSUSED */
798 void
799 tcp_clean_death_wrapper(void *arg, mblk_t *mp, void *arg2,
800 ip_recv_attr_t *dummy)
801 {
802 tcp_t *tcp = ((conn_t *)arg)->conn_tcp;
803
804 freemsg(mp);
805 if (tcp->tcp_state > TCPS_BOUND)
806 (void) tcp_clean_death(((conn_t *)arg)->conn_tcp, ETIMEDOUT);
807 }
808
809 /*
810 * We are dying for some reason. Try to do it gracefully. (May be called
811 * as writer.)
812 *
813 * Return -1 if the structure was not cleaned up (if the cleanup had to be
814 * done by a service procedure).
815 * TBD - Should the return value distinguish between the tcp_t being
816 * freed and it being reinitialized?
817 */
818 int
819 tcp_clean_death(tcp_t *tcp, int err)
820 {
821 mblk_t *mp;
822 queue_t *q;
823 conn_t *connp = tcp->tcp_connp;
824 tcp_stack_t *tcps = tcp->tcp_tcps;
825
826 if (tcp->tcp_fused)
827 tcp_unfuse(tcp);
828
829 if (tcp->tcp_linger_tid != 0 &&
830 TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
831 tcp_stop_lingering(tcp);
832 }
833
834 ASSERT(tcp != NULL);
835 ASSERT((connp->conn_family == AF_INET &&
836 connp->conn_ipversion == IPV4_VERSION) ||
837 (connp->conn_family == AF_INET6 &&
838 (connp->conn_ipversion == IPV4_VERSION ||
839 connp->conn_ipversion == IPV6_VERSION)));
840
841 if (TCP_IS_DETACHED(tcp)) {
842 if (tcp->tcp_hard_binding) {
843 /*
844 * Its an eager that we are dealing with. We close the
845 * eager but in case a conn_ind has already gone to the
846 * listener, let tcp_accept_finish() send a discon_ind
847 * to the listener and drop the last reference. If the
848 * listener doesn't even know about the eager i.e. the
849 * conn_ind hasn't gone up, blow away the eager and drop
850 * the last reference as well. If the conn_ind has gone
851 * up, state should be BOUND. tcp_accept_finish
852 * will figure out that the connection has received a
853 * RST and will send a DISCON_IND to the application.
854 */
855 tcp_closei_local(tcp);
856 if (!tcp->tcp_tconnind_started) {
857 CONN_DEC_REF(connp);
858 } else {
859 tcp->tcp_state = TCPS_BOUND;
860 DTRACE_TCP6(state__change, void, NULL,
861 ip_xmit_attr_t *, connp->conn_ixa,
862 void, NULL, tcp_t *, tcp, void, NULL,
863 int32_t, TCPS_CLOSED);
864 }
865 } else {
866 tcp_close_detached(tcp);
867 }
868 return (0);
869 }
870
871 TCP_STAT(tcps, tcp_clean_death_nondetached);
872
873 /*
874 * The connection is dead. Decrement listener connection counter if
875 * necessary.
876 */
877 if (tcp->tcp_listen_cnt != NULL)
878 TCP_DECR_LISTEN_CNT(tcp);
879
880 /*
881 * When a connection is moved to TIME_WAIT state, the connection
882 * counter is already decremented. So no need to decrement here
883 * again. See SET_TIME_WAIT() macro.
884 */
885 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
886 tcp->tcp_state < TCPS_TIME_WAIT) {
887 TCPS_CONN_DEC(tcps);
888 }
889
890 q = connp->conn_rq;
891
892 /* Trash all inbound data */
893 if (!IPCL_IS_NONSTR(connp)) {
894 ASSERT(q != NULL);
895 flushq(q, FLUSHALL);
896 }
897
898 /*
899 * If we are at least part way open and there is error
900 * (err==0 implies no error)
901 * notify our client by a T_DISCON_IND.
902 */
903 if ((tcp->tcp_state >= TCPS_SYN_SENT) && err) {
904 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
905 !TCP_IS_SOCKET(tcp)) {
906 /*
907 * Send M_FLUSH according to TPI. Because sockets will
908 * (and must) ignore FLUSHR we do that only for TPI
909 * endpoints and sockets in STREAMS mode.
910 */
911 (void) putnextctl1(q, M_FLUSH, FLUSHR);
912 }
913 if (connp->conn_debug) {
914 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE|SL_ERROR,
915 "tcp_clean_death: discon err %d", err);
916 }
917 if (IPCL_IS_NONSTR(connp)) {
918 /* Direct socket, use upcall */
919 (*connp->conn_upcalls->su_disconnected)(
920 connp->conn_upper_handle, tcp->tcp_connid, err);
921 } else {
922 mp = mi_tpi_discon_ind(NULL, err, 0);
923 if (mp != NULL) {
924 putnext(q, mp);
925 } else {
926 if (connp->conn_debug) {
927 (void) strlog(TCP_MOD_ID, 0, 1,
928 SL_ERROR|SL_TRACE,
929 "tcp_clean_death, sending M_ERROR");
930 }
931 (void) putnextctl1(q, M_ERROR, EPROTO);
932 }
933 }
934 if (tcp->tcp_state <= TCPS_SYN_RCVD) {
935 /* SYN_SENT or SYN_RCVD */
936 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
937 } else if (tcp->tcp_state <= TCPS_CLOSE_WAIT) {
938 /* ESTABLISHED or CLOSE_WAIT */
939 TCPS_BUMP_MIB(tcps, tcpEstabResets);
940 }
941 }
942
943 /*
944 * ESTABLISHED non-STREAMS eagers are not 'detached' because
945 * an upper handle is obtained when the SYN-ACK comes in. So it
946 * should receive the 'disconnected' upcall, but tcp_reinit should
947 * not be called since this is an eager.
948 */
949 if (tcp->tcp_listener != NULL && IPCL_IS_NONSTR(connp)) {
950 tcp_closei_local(tcp);
951 tcp->tcp_state = TCPS_BOUND;
952 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
953 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
954 int32_t, TCPS_CLOSED);
955 return (0);
956 }
957
958 tcp_reinit(tcp);
959 if (IPCL_IS_NONSTR(connp))
960 (void) tcp_do_unbind(connp);
961
962 return (-1);
963 }
964
965 /*
966 * In case tcp is in the "lingering state" and waits for the SO_LINGER timeout
967 * to expire, stop the wait and finish the close.
968 */
969 void
970 tcp_stop_lingering(tcp_t *tcp)
971 {
972 clock_t delta = 0;
973 tcp_stack_t *tcps = tcp->tcp_tcps;
974 conn_t *connp = tcp->tcp_connp;
975
976 tcp->tcp_linger_tid = 0;
977 if (tcp->tcp_state > TCPS_LISTEN) {
978 tcp_acceptor_hash_remove(tcp);
979 mutex_enter(&tcp->tcp_non_sq_lock);
980 if (tcp->tcp_flow_stopped) {
981 tcp_clrqfull(tcp);
982 }
983 mutex_exit(&tcp->tcp_non_sq_lock);
984
985 if (tcp->tcp_timer_tid != 0) {
986 delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
987 tcp->tcp_timer_tid = 0;
988 }
989 /*
990 * Need to cancel those timers which will not be used when
991 * TCP is detached. This has to be done before the conn_wq
992 * is cleared.
993 */
994 tcp_timers_stop(tcp);
995
996 tcp->tcp_detached = B_TRUE;
997 connp->conn_rq = NULL;
998 connp->conn_wq = NULL;
999
1000 if (tcp->tcp_state == TCPS_TIME_WAIT) {
1001 tcp_time_wait_append(tcp);
1002 TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1003 goto finish;
1004 }
1005
1006 /*
1007 * If delta is zero the timer event wasn't executed and was
1008 * successfully canceled. In this case we need to restart it
1009 * with the minimal delta possible.
1010 */
1011 if (delta >= 0) {
1012 tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1013 delta ? delta : 1);
1014 }
1015 } else {
1016 tcp_closei_local(tcp);
1017 CONN_DEC_REF(connp);
1018 }
1019 finish:
1020 tcp->tcp_detached = B_TRUE;
1021 connp->conn_rq = NULL;
1022 connp->conn_wq = NULL;
1023
1024 /* Signal closing thread that it can complete close */
1025 mutex_enter(&tcp->tcp_closelock);
1026 tcp->tcp_closed = 1;
1027 cv_signal(&tcp->tcp_closecv);
1028 mutex_exit(&tcp->tcp_closelock);
1029
1030 /* If we have an upper handle (socket), release it */
1031 if (IPCL_IS_NONSTR(connp)) {
1032 ASSERT(connp->conn_upper_handle != NULL);
1033 (*connp->conn_upcalls->su_closed)(connp->conn_upper_handle);
1034 connp->conn_upper_handle = NULL;
1035 connp->conn_upcalls = NULL;
1036 }
1037 }
1038
1039 void
1040 tcp_close_common(conn_t *connp, int flags)
1041 {
1042 tcp_t *tcp = connp->conn_tcp;
1043 mblk_t *mp = &tcp->tcp_closemp;
1044 boolean_t conn_ioctl_cleanup_reqd = B_FALSE;
1045 mblk_t *bp;
1046
1047 ASSERT(connp->conn_ref >= 2);
1048
1049 /*
1050 * Mark the conn as closing. ipsq_pending_mp_add will not
1051 * add any mp to the pending mp list, after this conn has
1052 * started closing.
1053 */
1054 mutex_enter(&connp->conn_lock);
1055 connp->conn_state_flags |= CONN_CLOSING;
1056 if (connp->conn_oper_pending_ill != NULL)
1057 conn_ioctl_cleanup_reqd = B_TRUE;
1058 CONN_INC_REF_LOCKED(connp);
1059 mutex_exit(&connp->conn_lock);
1060 tcp->tcp_closeflags = (uint8_t)flags;
1061 ASSERT(connp->conn_ref >= 3);
1062
1063 /*
1064 * tcp_closemp_used is used below without any protection of a lock
1065 * as we don't expect any one else to use it concurrently at this
1066 * point otherwise it would be a major defect.
1067 */
1068
1069 if (mp->b_prev == NULL)
1070 tcp->tcp_closemp_used = B_TRUE;
1071 else
1072 cmn_err(CE_PANIC, "tcp_close: concurrent use of tcp_closemp: "
1073 "connp %p tcp %p\n", (void *)connp, (void *)tcp);
1074
1075 TCP_DEBUG_GETPCSTACK(tcp->tcmp_stk, 15);
1076
1077 /*
1078 * Cleanup any queued ioctls here. This must be done before the wq/rq
1079 * are re-written by tcp_close_output().
1080 */
1081 if (conn_ioctl_cleanup_reqd)
1082 conn_ioctl_cleanup(connp);
1083
1084 /*
1085 * As CONN_CLOSING is set, no further ioctls should be passed down to
1086 * IP for this conn (see the guards in tcp_ioctl, tcp_wput_ioctl and
1087 * tcp_wput_iocdata). If the ioctl was queued on an ipsq,
1088 * conn_ioctl_cleanup should have found it and removed it. If the ioctl
1089 * was still in flight at the time, we wait for it here. See comments
1090 * for CONN_INC_IOCTLREF in ip.h for details.
1091 */
1092 mutex_enter(&connp->conn_lock);
1093 while (connp->conn_ioctlref > 0)
1094 cv_wait(&connp->conn_cv, &connp->conn_lock);
1095 ASSERT(connp->conn_ioctlref == 0);
1096 ASSERT(connp->conn_oper_pending_ill == NULL);
1097 mutex_exit(&connp->conn_lock);
1098
1099 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_close_output, connp,
1100 NULL, tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1101
1102 /*
1103 * For non-STREAMS sockets, the normal case is that the conn makes
1104 * an upcall when it's finally closed, so there is no need to wait
1105 * in the protocol. But in case of SO_LINGER the thread sleeps here
1106 * so it can properly deal with the thread being interrupted.
1107 */
1108 if (IPCL_IS_NONSTR(connp) && connp->conn_linger == 0)
1109 goto nowait;
1110
1111 mutex_enter(&tcp->tcp_closelock);
1112 while (!tcp->tcp_closed) {
1113 if (!cv_wait_sig(&tcp->tcp_closecv, &tcp->tcp_closelock)) {
1114 /*
1115 * The cv_wait_sig() was interrupted. We now do the
1116 * following:
1117 *
1118 * 1) If the endpoint was lingering, we allow this
1119 * to be interrupted by cancelling the linger timeout
1120 * and closing normally.
1121 *
1122 * 2) Revert to calling cv_wait()
1123 *
1124 * We revert to using cv_wait() to avoid an
1125 * infinite loop which can occur if the calling
1126 * thread is higher priority than the squeue worker
1127 * thread and is bound to the same cpu.
1128 */
1129 if (connp->conn_linger && connp->conn_lingertime > 0) {
1130 mutex_exit(&tcp->tcp_closelock);
1131 /* Entering squeue, bump ref count. */
1132 CONN_INC_REF(connp);
1133 bp = allocb_wait(0, BPRI_HI, STR_NOSIG, NULL);
1134 SQUEUE_ENTER_ONE(connp->conn_sqp, bp,
1135 tcp_linger_interrupted, connp, NULL,
1136 tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1137 mutex_enter(&tcp->tcp_closelock);
1138 }
1139 break;
1140 }
1141 }
1142 while (!tcp->tcp_closed)
1143 cv_wait(&tcp->tcp_closecv, &tcp->tcp_closelock);
1144 mutex_exit(&tcp->tcp_closelock);
1145
1146 /*
1147 * In the case of listener streams that have eagers in the q or q0
1148 * we wait for the eagers to drop their reference to us. conn_rq and
1149 * conn_wq of the eagers point to our queues. By waiting for the
1150 * refcnt to drop to 1, we are sure that the eagers have cleaned
1151 * up their queue pointers and also dropped their references to us.
1152 *
1153 * For non-STREAMS sockets we do not have to wait here; the
1154 * listener will instead make a su_closed upcall when the last
1155 * reference is dropped.
1156 */
1157 if (tcp->tcp_wait_for_eagers && !IPCL_IS_NONSTR(connp)) {
1158 mutex_enter(&connp->conn_lock);
1159 while (connp->conn_ref != 1) {
1160 cv_wait(&connp->conn_cv, &connp->conn_lock);
1161 }
1162 mutex_exit(&connp->conn_lock);
1163 }
1164
1165 nowait:
1166 connp->conn_cpid = NOPID;
1167 }
1168
1169 /*
1170 * Called by tcp_close() routine via squeue when lingering is
1171 * interrupted by a signal.
1172 */
1173
1174 /* ARGSUSED */
1175 static void
1176 tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1177 {
1178 conn_t *connp = (conn_t *)arg;
1179 tcp_t *tcp = connp->conn_tcp;
1180
1181 freeb(mp);
1182 if (tcp->tcp_linger_tid != 0 &&
1183 TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
1184 tcp_stop_lingering(tcp);
1185 tcp->tcp_client_errno = EINTR;
1186 }
1187 }
1188
1189 /*
1190 * Clean up the b_next and b_prev fields of every mblk pointed at by *mpp.
1191 * Some stream heads get upset if they see these later on as anything but NULL.
1192 */
1193 void
1194 tcp_close_mpp(mblk_t **mpp)
1195 {
1196 mblk_t *mp;
1197
1198 if ((mp = *mpp) != NULL) {
1199 do {
1200 mp->b_next = NULL;
1201 mp->b_prev = NULL;
1202 } while ((mp = mp->b_cont) != NULL);
1203
1204 mp = *mpp;
1205 *mpp = NULL;
1206 freemsg(mp);
1207 }
1208 }
1209
1210 /* Do detached close. */
1211 void
1212 tcp_close_detached(tcp_t *tcp)
1213 {
1214 if (tcp->tcp_fused)
1215 tcp_unfuse(tcp);
1216
1217 /*
1218 * Clustering code serializes TCP disconnect callbacks and
1219 * cluster tcp list walks by blocking a TCP disconnect callback
1220 * if a cluster tcp list walk is in progress. This ensures
1221 * accurate accounting of TCPs in the cluster code even though
1222 * the TCP list walk itself is not atomic.
1223 */
1224 tcp_closei_local(tcp);
1225 CONN_DEC_REF(tcp->tcp_connp);
1226 }
1227
1228 /*
1229 * The tcp_t is going away. Remove it from all lists and set it
1230 * to TCPS_CLOSED. The freeing up of memory is deferred until
1231 * tcp_inactive. This is needed since a thread in tcp_rput might have
1232 * done a CONN_INC_REF on this structure before it was removed from the
1233 * hashes.
1234 */
1235 void
1236 tcp_closei_local(tcp_t *tcp)
1237 {
1238 conn_t *connp = tcp->tcp_connp;
1239 tcp_stack_t *tcps = tcp->tcp_tcps;
1240 int32_t oldstate;
1241
1242 if (!TCP_IS_SOCKET(tcp))
1243 tcp_acceptor_hash_remove(tcp);
1244
1245 TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1246 tcp->tcp_ibsegs = 0;
1247 TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1248 tcp->tcp_obsegs = 0;
1249
1250 /*
1251 * This can be called via tcp_time_wait_processing() if TCP gets a
1252 * SYN with sequence number outside the TIME-WAIT connection's
1253 * window. So we need to check for TIME-WAIT state here as the
1254 * connection counter is already decremented. See SET_TIME_WAIT()
1255 * macro
1256 */
1257 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1258 tcp->tcp_state < TCPS_TIME_WAIT) {
1259 TCPS_CONN_DEC(tcps);
1260 }
1261
1262 /*
1263 * If we are an eager connection hanging off a listener that
1264 * hasn't formally accepted the connection yet, get off his
1265 * list and blow off any data that we have accumulated.
1266 */
1267 if (tcp->tcp_listener != NULL) {
1268 tcp_t *listener = tcp->tcp_listener;
1269 mutex_enter(&listener->tcp_eager_lock);
1270 /*
1271 * tcp_tconnind_started == B_TRUE means that the
1272 * conn_ind has already gone to listener. At
1273 * this point, eager will be closed but we
1274 * leave it in listeners eager list so that
1275 * if listener decides to close without doing
1276 * accept, we can clean this up. In tcp_tli_accept
1277 * we take care of the case of accept on closed
1278 * eager.
1279 */
1280 if (!tcp->tcp_tconnind_started) {
1281 tcp_eager_unlink(tcp);
1282 mutex_exit(&listener->tcp_eager_lock);
1283 /*
1284 * We don't want to have any pointers to the
1285 * listener queue, after we have released our
1286 * reference on the listener
1287 */
1288 ASSERT(tcp->tcp_detached);
1289 connp->conn_rq = NULL;
1290 connp->conn_wq = NULL;
1291 CONN_DEC_REF(listener->tcp_connp);
1292 } else {
1293 mutex_exit(&listener->tcp_eager_lock);
1294 }
1295 }
1296
1297 /* Stop all the timers */
1298 tcp_timers_stop(tcp);
1299
1300 if (tcp->tcp_state == TCPS_LISTEN) {
1301 if (tcp->tcp_ip_addr_cache) {
1302 kmem_free((void *)tcp->tcp_ip_addr_cache,
1303 IP_ADDR_CACHE_SIZE * sizeof (ipaddr_t));
1304 tcp->tcp_ip_addr_cache = NULL;
1305 }
1306 }
1307
1308 /* Decrement listerner connection counter if necessary. */
1309 if (tcp->tcp_listen_cnt != NULL)
1310 TCP_DECR_LISTEN_CNT(tcp);
1311
1312 mutex_enter(&tcp->tcp_non_sq_lock);
1313 if (tcp->tcp_flow_stopped)
1314 tcp_clrqfull(tcp);
1315 mutex_exit(&tcp->tcp_non_sq_lock);
1316
1317 tcp_bind_hash_remove(tcp);
1318 /*
1319 * If the tcp_time_wait_collector (which runs outside the squeue)
1320 * is trying to remove this tcp from the time wait list, we will
1321 * block in tcp_time_wait_remove while trying to acquire the
1322 * tcp_time_wait_lock. The logic in tcp_time_wait_collector also
1323 * requires the ipcl_hash_remove to be ordered after the
1324 * tcp_time_wait_remove for the refcnt checks to work correctly.
1325 */
1326 if (tcp->tcp_state == TCPS_TIME_WAIT)
1327 (void) tcp_time_wait_remove(tcp, NULL);
1328 CL_INET_DISCONNECT(connp);
1329 ipcl_hash_remove(connp);
1330 oldstate = tcp->tcp_state;
1331 tcp->tcp_state = TCPS_CLOSED;
1332 /* Need to probe before ixa_cleanup() is called */
1333 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1334 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
1335 int32_t, oldstate);
1336 ixa_cleanup(connp->conn_ixa);
1337
1338 /*
1339 * Mark the conn as CONDEMNED
1340 */
1341 mutex_enter(&connp->conn_lock);
1342 connp->conn_state_flags |= CONN_CONDEMNED;
1343 mutex_exit(&connp->conn_lock);
1344
1345 ASSERT(tcp->tcp_time_wait_next == NULL);
1346 ASSERT(tcp->tcp_time_wait_prev == NULL);
1347 ASSERT(tcp->tcp_time_wait_expire == 0);
1348
1349 tcp_ipsec_cleanup(tcp);
1350 }
1351
1352 /*
1353 * tcp is dying (called from ipcl_conn_destroy and error cases).
1354 * Free the tcp_t in either case.
1355 */
1356 void
1357 tcp_free(tcp_t *tcp)
1358 {
1359 mblk_t *mp;
1360 conn_t *connp = tcp->tcp_connp;
1361
1362 ASSERT(tcp != NULL);
1363 ASSERT(tcp->tcp_ptpahn == NULL && tcp->tcp_acceptor_hash == NULL);
1364
1365 connp->conn_rq = NULL;
1366 connp->conn_wq = NULL;
1367
1368 tcp_close_mpp(&tcp->tcp_xmit_head);
1369 tcp_close_mpp(&tcp->tcp_reass_head);
1370 if (tcp->tcp_rcv_list != NULL) {
1371 /* Free b_next chain */
1372 tcp_close_mpp(&tcp->tcp_rcv_list);
1373 }
1374 if ((mp = tcp->tcp_urp_mp) != NULL) {
1375 freemsg(mp);
1376 }
1377 if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1378 freemsg(mp);
1379 }
1380
1381 if (tcp->tcp_fused_sigurg_mp != NULL) {
1382 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1383 freeb(tcp->tcp_fused_sigurg_mp);
1384 tcp->tcp_fused_sigurg_mp = NULL;
1385 }
1386
1387 if (tcp->tcp_ordrel_mp != NULL) {
1388 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1389 freeb(tcp->tcp_ordrel_mp);
1390 tcp->tcp_ordrel_mp = NULL;
1391 }
1392
1393 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
1394 bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
1395
1396 if (tcp->tcp_hopopts != NULL) {
1397 mi_free(tcp->tcp_hopopts);
1398 tcp->tcp_hopopts = NULL;
1399 tcp->tcp_hopoptslen = 0;
1400 }
1401 ASSERT(tcp->tcp_hopoptslen == 0);
1402 if (tcp->tcp_dstopts != NULL) {
1403 mi_free(tcp->tcp_dstopts);
1404 tcp->tcp_dstopts = NULL;
1405 tcp->tcp_dstoptslen = 0;
1406 }
1407 ASSERT(tcp->tcp_dstoptslen == 0);
1408 if (tcp->tcp_rthdrdstopts != NULL) {
1409 mi_free(tcp->tcp_rthdrdstopts);
1410 tcp->tcp_rthdrdstopts = NULL;
1411 tcp->tcp_rthdrdstoptslen = 0;
1412 }
1413 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
1414 if (tcp->tcp_rthdr != NULL) {
1415 mi_free(tcp->tcp_rthdr);
1416 tcp->tcp_rthdr = NULL;
1417 tcp->tcp_rthdrlen = 0;
1418 }
1419 ASSERT(tcp->tcp_rthdrlen == 0);
1420
1421 /*
1422 * Following is really a blowing away a union.
1423 * It happens to have exactly two members of identical size
1424 * the following code is enough.
1425 */
1426 tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1427
1428 /*
1429 * If this is a non-STREAM socket still holding on to an upper
1430 * handle, release it. As a result of fallback we might also see
1431 * STREAMS based conns with upper handles, in which case there is
1432 * nothing to do other than clearing the field.
1433 */
1434 if (connp->conn_upper_handle != NULL) {
1435 if (IPCL_IS_NONSTR(connp)) {
1436 (*connp->conn_upcalls->su_closed)(
1437 connp->conn_upper_handle);
1438 tcp->tcp_detached = B_TRUE;
1439 }
1440 connp->conn_upper_handle = NULL;
1441 connp->conn_upcalls = NULL;
1442 }
1443 }
1444
1445 /*
1446 * tcp_get_conn/tcp_free_conn
1447 *
1448 * tcp_get_conn is used to get a clean tcp connection structure.
1449 * It tries to reuse the connections put on the freelist by the
1450 * time_wait_collector failing which it goes to kmem_cache. This
1451 * way has two benefits compared to just allocating from and
1452 * freeing to kmem_cache.
1453 * 1) The time_wait_collector can free (which includes the cleanup)
1454 * outside the squeue. So when the interrupt comes, we have a clean
1455 * connection sitting in the freelist. Obviously, this buys us
1456 * performance.
1457 *
1458 * 2) Defence against DOS attack. Allocating a tcp/conn in tcp_input_listener
1459 * has multiple disadvantages - tying up the squeue during alloc.
1460 * But allocating the conn/tcp in IP land is also not the best since
1461 * we can't check the 'q' and 'q0' which are protected by squeue and
1462 * blindly allocate memory which might have to be freed here if we are
1463 * not allowed to accept the connection. By using the freelist and
1464 * putting the conn/tcp back in freelist, we don't pay a penalty for
1465 * allocating memory without checking 'q/q0' and freeing it if we can't
1466 * accept the connection.
1467 *
1468 * Care should be taken to put the conn back in the same squeue's freelist
1469 * from which it was allocated. Best results are obtained if conn is
1470 * allocated from listener's squeue and freed to the same. Time wait
1471 * collector will free up the freelist is the connection ends up sitting
1472 * there for too long.
1473 */
1474 void *
1475 tcp_get_conn(void *arg, tcp_stack_t *tcps)
1476 {
1477 tcp_t *tcp = NULL;
1478 conn_t *connp = NULL;
1479 squeue_t *sqp = (squeue_t *)arg;
1480 tcp_squeue_priv_t *tcp_time_wait;
1481 netstack_t *ns;
1482 mblk_t *tcp_rsrv_mp = NULL;
1483
1484 tcp_time_wait =
1485 *((tcp_squeue_priv_t **)squeue_getprivate(sqp, SQPRIVATE_TCP));
1486
1487 mutex_enter(&tcp_time_wait->tcp_time_wait_lock);
1488 tcp = tcp_time_wait->tcp_free_list;
1489 ASSERT((tcp != NULL) ^ (tcp_time_wait->tcp_free_list_cnt == 0));
1490 if (tcp != NULL) {
1491 tcp_time_wait->tcp_free_list = tcp->tcp_time_wait_next;
1492 tcp_time_wait->tcp_free_list_cnt--;
1493 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1494 tcp->tcp_time_wait_next = NULL;
1495 connp = tcp->tcp_connp;
1496 connp->conn_flags |= IPCL_REUSED;
1497
1498 ASSERT(tcp->tcp_tcps == NULL);
1499 ASSERT(connp->conn_netstack == NULL);
1500 ASSERT(tcp->tcp_rsrv_mp != NULL);
1501 ns = tcps->tcps_netstack;
1502 netstack_hold(ns);
1503 connp->conn_netstack = ns;
1504 connp->conn_ixa->ixa_ipst = ns->netstack_ip;
1505 tcp->tcp_tcps = tcps;
1506 ipcl_globalhash_insert(connp);
1507
1508 connp->conn_ixa->ixa_notify_cookie = tcp;
1509 ASSERT(connp->conn_ixa->ixa_notify == tcp_notify);
1510 connp->conn_recv = tcp_input_data;
1511 ASSERT(connp->conn_recvicmp == tcp_icmp_input);
1512 ASSERT(connp->conn_verifyicmp == tcp_verifyicmp);
1513 return ((void *)connp);
1514 }
1515 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1516 /*
1517 * Pre-allocate the tcp_rsrv_mp. This mblk will not be freed until
1518 * this conn_t/tcp_t is freed at ipcl_conn_destroy().
1519 */
1520 tcp_rsrv_mp = allocb(0, BPRI_HI);
1521 if (tcp_rsrv_mp == NULL)
1522 return (NULL);
1523
1524 if ((connp = ipcl_conn_create(IPCL_TCPCONN, KM_NOSLEEP,
1525 tcps->tcps_netstack)) == NULL) {
1526 freeb(tcp_rsrv_mp);
1527 return (NULL);
1528 }
1529
1530 tcp = connp->conn_tcp;
1531 tcp->tcp_rsrv_mp = tcp_rsrv_mp;
1532 mutex_init(&tcp->tcp_rsrv_mp_lock, NULL, MUTEX_DEFAULT, NULL);
1533
1534 tcp->tcp_tcps = tcps;
1535
1536 connp->conn_recv = tcp_input_data;
1537 connp->conn_recvicmp = tcp_icmp_input;
1538 connp->conn_verifyicmp = tcp_verifyicmp;
1539
1540 /*
1541 * Register tcp_notify to listen to capability changes detected by IP.
1542 * This upcall is made in the context of the call to conn_ip_output
1543 * thus it is inside the squeue.
1544 */
1545 connp->conn_ixa->ixa_notify = tcp_notify;
1546 connp->conn_ixa->ixa_notify_cookie = tcp;
1547
1548 return ((void *)connp);
1549 }
1550
1551 /*
1552 * Handle connect to IPv4 destinations, including connections for AF_INET6
1553 * sockets connecting to IPv4 mapped IPv6 destinations.
1554 * Returns zero if OK, a positive errno, or a negative TLI error.
1555 */
1556 static int
1557 tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp, in_port_t dstport,
1558 uint_t srcid)
1559 {
1560 ipaddr_t dstaddr = *dstaddrp;
1561 uint16_t lport;
1562 conn_t *connp = tcp->tcp_connp;
1563 tcp_stack_t *tcps = tcp->tcp_tcps;
1564 int error;
1565
1566 ASSERT(connp->conn_ipversion == IPV4_VERSION);
1567
1568 /* Check for attempt to connect to INADDR_ANY */
1569 if (dstaddr == INADDR_ANY) {
1570 /*
1571 * SunOS 4.x and 4.3 BSD allow an application
1572 * to connect a TCP socket to INADDR_ANY.
1573 * When they do this, the kernel picks the
1574 * address of one interface and uses it
1575 * instead. The kernel usually ends up
1576 * picking the address of the loopback
1577 * interface. This is an undocumented feature.
1578 * However, we provide the same thing here
1579 * in order to have source and binary
1580 * compatibility with SunOS 4.x.
1581 * Update the T_CONN_REQ (sin/sin6) since it is used to
1582 * generate the T_CONN_CON.
1583 */
1584 dstaddr = htonl(INADDR_LOOPBACK);
1585 *dstaddrp = dstaddr;
1586 }
1587
1588 /* Handle __sin6_src_id if socket not bound to an IP address */
1589 if (srcid != 0 && connp->conn_laddr_v4 == INADDR_ANY) {
1590 ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1591 IPCL_ZONEID(connp), tcps->tcps_netstack);
1592 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1593 }
1594
1595 IN6_IPADDR_TO_V4MAPPED(dstaddr, &connp->conn_faddr_v6);
1596 connp->conn_fport = dstport;
1597
1598 /*
1599 * At this point the remote destination address and remote port fields
1600 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1601 * have to see which state tcp was in so we can take appropriate action.
1602 */
1603 if (tcp->tcp_state == TCPS_IDLE) {
1604 /*
1605 * We support a quick connect capability here, allowing
1606 * clients to transition directly from IDLE to SYN_SENT
1607 * tcp_bindi will pick an unused port, insert the connection
1608 * in the bind hash and transition to BOUND state.
1609 */
1610 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1611 tcp, B_TRUE);
1612 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1613 B_FALSE, B_FALSE);
1614 if (lport == 0)
1615 return (-TNOADDR);
1616 }
1617
1618 /*
1619 * Lookup the route to determine a source address and the uinfo.
1620 * Setup TCP parameters based on the metrics/DCE.
1621 */
1622 error = tcp_set_destination(tcp);
1623 if (error != 0)
1624 return (error);
1625
1626 /*
1627 * Don't let an endpoint connect to itself.
1628 */
1629 if (connp->conn_faddr_v4 == connp->conn_laddr_v4 &&
1630 connp->conn_fport == connp->conn_lport)
1631 return (-TBADADDR);
1632
1633 tcp->tcp_state = TCPS_SYN_SENT;
1634
1635 return (ipcl_conn_insert_v4(connp));
1636 }
1637
1638 /*
1639 * Handle connect to IPv6 destinations.
1640 * Returns zero if OK, a positive errno, or a negative TLI error.
1641 */
1642 static int
1643 tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp, in_port_t dstport,
1644 uint32_t flowinfo, uint_t srcid, uint32_t scope_id)
1645 {
1646 uint16_t lport;
1647 conn_t *connp = tcp->tcp_connp;
1648 tcp_stack_t *tcps = tcp->tcp_tcps;
1649 int error;
1650
1651 ASSERT(connp->conn_family == AF_INET6);
1652
1653 /*
1654 * If we're here, it means that the destination address is a native
1655 * IPv6 address. Return an error if conn_ipversion is not IPv6. A
1656 * reason why it might not be IPv6 is if the socket was bound to an
1657 * IPv4-mapped IPv6 address.
1658 */
1659 if (connp->conn_ipversion != IPV6_VERSION)
1660 return (-TBADADDR);
1661
1662 /*
1663 * Interpret a zero destination to mean loopback.
1664 * Update the T_CONN_REQ (sin/sin6) since it is used to
1665 * generate the T_CONN_CON.
1666 */
1667 if (IN6_IS_ADDR_UNSPECIFIED(dstaddrp))
1668 *dstaddrp = ipv6_loopback;
1669
1670 /* Handle __sin6_src_id if socket not bound to an IP address */
1671 if (srcid != 0 && IN6_IS_ADDR_UNSPECIFIED(&connp->conn_laddr_v6)) {
1672 ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1673 IPCL_ZONEID(connp), tcps->tcps_netstack);
1674 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1675 }
1676
1677 /*
1678 * Take care of the scope_id now.
1679 */
1680 if (scope_id != 0 && IN6_IS_ADDR_LINKSCOPE(dstaddrp)) {
1681 connp->conn_ixa->ixa_flags |= IXAF_SCOPEID_SET;
1682 connp->conn_ixa->ixa_scopeid = scope_id;
1683 } else {
1684 connp->conn_ixa->ixa_flags &= ~IXAF_SCOPEID_SET;
1685 }
1686
1687 connp->conn_flowinfo = flowinfo;
1688 connp->conn_faddr_v6 = *dstaddrp;
1689 connp->conn_fport = dstport;
1690
1691 /*
1692 * At this point the remote destination address and remote port fields
1693 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1694 * have to see which state tcp was in so we can take appropriate action.
1695 */
1696 if (tcp->tcp_state == TCPS_IDLE) {
1697 /*
1698 * We support a quick connect capability here, allowing
1699 * clients to transition directly from IDLE to SYN_SENT
1700 * tcp_bindi will pick an unused port, insert the connection
1701 * in the bind hash and transition to BOUND state.
1702 */
1703 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1704 tcp, B_TRUE);
1705 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1706 B_FALSE, B_FALSE);
1707 if (lport == 0)
1708 return (-TNOADDR);
1709 }
1710
1711 /*
1712 * Lookup the route to determine a source address and the uinfo.
1713 * Setup TCP parameters based on the metrics/DCE.
1714 */
1715 error = tcp_set_destination(tcp);
1716 if (error != 0)
1717 return (error);
1718
1719 /*
1720 * Don't let an endpoint connect to itself.
1721 */
1722 if (IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6, &connp->conn_laddr_v6) &&
1723 connp->conn_fport == connp->conn_lport)
1724 return (-TBADADDR);
1725
1726 tcp->tcp_state = TCPS_SYN_SENT;
1727
1728 return (ipcl_conn_insert_v6(connp));
1729 }
1730
1731 /*
1732 * Disconnect
1733 * Note that unlike other functions this returns a positive tli error
1734 * when it fails; it never returns an errno.
1735 */
1736 static int
1737 tcp_disconnect_common(tcp_t *tcp, t_scalar_t seqnum)
1738 {
1739 conn_t *lconnp;
1740 tcp_stack_t *tcps = tcp->tcp_tcps;
1741 conn_t *connp = tcp->tcp_connp;
1742
1743 /*
1744 * Right now, upper modules pass down a T_DISCON_REQ to TCP,
1745 * when the stream is in BOUND state. Do not send a reset,
1746 * since the destination IP address is not valid, and it can
1747 * be the initialized value of all zeros (broadcast address).
1748 */
1749 if (tcp->tcp_state <= TCPS_BOUND) {
1750 if (connp->conn_debug) {
1751 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1752 "tcp_disconnect: bad state, %d", tcp->tcp_state);
1753 }
1754 return (TOUTSTATE);
1755 } else if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1756 TCPS_CONN_DEC(tcps);
1757 }
1758
1759 if (seqnum == -1 || tcp->tcp_conn_req_max == 0) {
1760
1761 /*
1762 * According to TPI, for non-listeners, ignore seqnum
1763 * and disconnect.
1764 * Following interpretation of -1 seqnum is historical
1765 * and implied TPI ? (TPI only states that for T_CONN_IND,
1766 * a valid seqnum should not be -1).
1767 *
1768 * -1 means disconnect everything
1769 * regardless even on a listener.
1770 */
1771
1772 int old_state = tcp->tcp_state;
1773 ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
1774
1775 /*
1776 * The connection can't be on the tcp_time_wait_head list
1777 * since it is not detached.
1778 */
1779 ASSERT(tcp->tcp_time_wait_next == NULL);
1780 ASSERT(tcp->tcp_time_wait_prev == NULL);
1781 ASSERT(tcp->tcp_time_wait_expire == 0);
1782 /*
1783 * If it used to be a listener, check to make sure no one else
1784 * has taken the port before switching back to LISTEN state.
1785 */
1786 if (connp->conn_ipversion == IPV4_VERSION) {
1787 lconnp = ipcl_lookup_listener_v4(connp->conn_lport,
1788 connp->conn_laddr_v4, IPCL_ZONEID(connp), ipst);
1789 } else {
1790 uint_t ifindex = 0;
1791
1792 if (connp->conn_ixa->ixa_flags & IXAF_SCOPEID_SET)
1793 ifindex = connp->conn_ixa->ixa_scopeid;
1794
1795 /* Allow conn_bound_if listeners? */
1796 lconnp = ipcl_lookup_listener_v6(connp->conn_lport,
1797 &connp->conn_laddr_v6, ifindex, IPCL_ZONEID(connp),
1798 ipst);
1799 }
1800 if (tcp->tcp_conn_req_max && lconnp == NULL) {
1801 tcp->tcp_state = TCPS_LISTEN;
1802 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1803 connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1804 NULL, int32_t, old_state);
1805 } else if (old_state > TCPS_BOUND) {
1806 tcp->tcp_conn_req_max = 0;
1807 tcp->tcp_state = TCPS_BOUND;
1808 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1809 connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1810 NULL, int32_t, old_state);
1811
1812 /*
1813 * If this end point is not going to become a listener,
1814 * decrement the listener connection count if
1815 * necessary. Note that we do not do this if it is
1816 * going to be a listner (the above if case) since
1817 * then it may remove the counter struct.
1818 */
1819 if (tcp->tcp_listen_cnt != NULL)
1820 TCP_DECR_LISTEN_CNT(tcp);
1821 }
1822 if (lconnp != NULL)
1823 CONN_DEC_REF(lconnp);
1824 switch (old_state) {
1825 case TCPS_SYN_SENT:
1826 case TCPS_SYN_RCVD:
1827 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1828 break;
1829 case TCPS_ESTABLISHED:
1830 case TCPS_CLOSE_WAIT:
1831 TCPS_BUMP_MIB(tcps, tcpEstabResets);
1832 break;
1833 }
1834
1835 if (tcp->tcp_fused)
1836 tcp_unfuse(tcp);
1837
1838 mutex_enter(&tcp->tcp_eager_lock);
1839 if ((tcp->tcp_conn_req_cnt_q0 != 0) ||
1840 (tcp->tcp_conn_req_cnt_q != 0)) {
1841 tcp_eager_cleanup(tcp, 0);
1842 }
1843 mutex_exit(&tcp->tcp_eager_lock);
1844
1845 tcp_xmit_ctl("tcp_disconnect", tcp, tcp->tcp_snxt,
1846 tcp->tcp_rnxt, TH_RST | TH_ACK);
1847
1848 tcp_reinit(tcp);
1849
1850 return (0);
1851 } else if (!tcp_eager_blowoff(tcp, seqnum)) {
1852 return (TBADSEQ);
1853 }
1854 return (0);
1855 }
1856
1857 /*
1858 * Our client hereby directs us to reject the connection request
1859 * that tcp_input_listener() marked with 'seqnum'. Rejection consists
1860 * of sending the appropriate RST, not an ICMP error.
1861 */
1862 void
1863 tcp_disconnect(tcp_t *tcp, mblk_t *mp)
1864 {
1865 t_scalar_t seqnum;
1866 int error;
1867 conn_t *connp = tcp->tcp_connp;
1868
1869 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1870 if ((mp->b_wptr - mp->b_rptr) < sizeof (struct T_discon_req)) {
1871 tcp_err_ack(tcp, mp, TPROTO, 0);
1872 return;
1873 }
1874 seqnum = ((struct T_discon_req *)mp->b_rptr)->SEQ_number;
1875 error = tcp_disconnect_common(tcp, seqnum);
1876 if (error != 0)
1877 tcp_err_ack(tcp, mp, error, 0);
1878 else {
1879 if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1880 /* Send M_FLUSH according to TPI */
1881 (void) putnextctl1(connp->conn_rq, M_FLUSH, FLUSHRW);
1882 }
1883 mp = mi_tpi_ok_ack_alloc(mp);
1884 if (mp != NULL)
1885 putnext(connp->conn_rq, mp);
1886 }
1887 }
1888
1889 /*
1890 * Handle reinitialization of a tcp structure.
1891 * Maintain "binding state" resetting the state to BOUND, LISTEN, or IDLE.
1892 */
1893 static void
1894 tcp_reinit(tcp_t *tcp)
1895 {
1896 mblk_t *mp;
1897 tcp_stack_t *tcps = tcp->tcp_tcps;
1898 conn_t *connp = tcp->tcp_connp;
1899 int32_t oldstate;
1900
1901 /* tcp_reinit should never be called for detached tcp_t's */
1902 ASSERT(tcp->tcp_listener == NULL);
1903 ASSERT((connp->conn_family == AF_INET &&
1904 connp->conn_ipversion == IPV4_VERSION) ||
1905 (connp->conn_family == AF_INET6 &&
1906 (connp->conn_ipversion == IPV4_VERSION ||
1907 connp->conn_ipversion == IPV6_VERSION)));
1908
1909 /* Cancel outstanding timers */
1910 tcp_timers_stop(tcp);
1911
1912 /*
1913 * Reset everything in the state vector, after updating global
1914 * MIB data from instance counters.
1915 */
1916 TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1917 tcp->tcp_ibsegs = 0;
1918 TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1919 tcp->tcp_obsegs = 0;
1920
1921 tcp_close_mpp(&tcp->tcp_xmit_head);
1922 if (tcp->tcp_snd_zcopy_aware)
1923 tcp_zcopy_notify(tcp);
1924 tcp->tcp_xmit_last = tcp->tcp_xmit_tail = NULL;
1925 tcp->tcp_unsent = tcp->tcp_xmit_tail_unsent = 0;
1926 mutex_enter(&tcp->tcp_non_sq_lock);
1927 if (tcp->tcp_flow_stopped &&
1928 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1929 tcp_clrqfull(tcp);
1930 }
1931 mutex_exit(&tcp->tcp_non_sq_lock);
1932 tcp_close_mpp(&tcp->tcp_reass_head);
1933 tcp->tcp_reass_tail = NULL;
1934 if (tcp->tcp_rcv_list != NULL) {
1935 /* Free b_next chain */
1936 tcp_close_mpp(&tcp->tcp_rcv_list);
1937 tcp->tcp_rcv_last_head = NULL;
1938 tcp->tcp_rcv_last_tail = NULL;
1939 tcp->tcp_rcv_cnt = 0;
1940 }
1941 tcp->tcp_rcv_last_tail = NULL;
1942
1943 if ((mp = tcp->tcp_urp_mp) != NULL) {
1944 freemsg(mp);
1945 tcp->tcp_urp_mp = NULL;
1946 }
1947 if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1948 freemsg(mp);
1949 tcp->tcp_urp_mark_mp = NULL;
1950 }
1951 if (tcp->tcp_fused_sigurg_mp != NULL) {
1952 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1953 freeb(tcp->tcp_fused_sigurg_mp);
1954 tcp->tcp_fused_sigurg_mp = NULL;
1955 }
1956 if (tcp->tcp_ordrel_mp != NULL) {
1957 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1958 freeb(tcp->tcp_ordrel_mp);
1959 tcp->tcp_ordrel_mp = NULL;
1960 }
1961
1962 /*
1963 * Following is a union with two members which are
1964 * identical types and size so the following cleanup
1965 * is enough.
1966 */
1967 tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1968
1969 CL_INET_DISCONNECT(connp);
1970
1971 /*
1972 * The connection can't be on the tcp_time_wait_head list
1973 * since it is not detached.
1974 */
1975 ASSERT(tcp->tcp_time_wait_next == NULL);
1976 ASSERT(tcp->tcp_time_wait_prev == NULL);
1977 ASSERT(tcp->tcp_time_wait_expire == 0);
1978
1979 /*
1980 * Reset/preserve other values
1981 */
1982 tcp_reinit_values(tcp);
1983 ipcl_hash_remove(connp);
1984 /* Note that ixa_cred gets cleared in ixa_cleanup */
1985 ixa_cleanup(connp->conn_ixa);
1986 tcp_ipsec_cleanup(tcp);
1987
1988 connp->conn_laddr_v6 = connp->conn_bound_addr_v6;
1989 connp->conn_saddr_v6 = connp->conn_bound_addr_v6;
1990 oldstate = tcp->tcp_state;
1991
1992 if (tcp->tcp_conn_req_max != 0) {
1993 /*
1994 * This is the case when a TLI program uses the same
1995 * transport end point to accept a connection. This
1996 * makes the TCP both a listener and acceptor. When
1997 * this connection is closed, we need to set the state
1998 * back to TCPS_LISTEN. Make sure that the eager list
1999 * is reinitialized.
2000 *
2001 * Note that this stream is still bound to the four
2002 * tuples of the previous connection in IP. If a new
2003 * SYN with different foreign address comes in, IP will
2004 * not find it and will send it to the global queue. In
2005 * the global queue, TCP will do a tcp_lookup_listener()
2006 * to find this stream. This works because this stream
2007 * is only removed from connected hash.
2008 *
2009 */
2010 tcp->tcp_state = TCPS_LISTEN;
2011 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
2012 tcp->tcp_eager_next_drop_q0 = tcp;
2013 tcp->tcp_eager_prev_drop_q0 = tcp;
2014 /*
2015 * Initially set conn_recv to tcp_input_listener_unbound to try
2016 * to pick a good squeue for the listener when the first SYN
2017 * arrives. tcp_input_listener_unbound sets it to
2018 * tcp_input_listener on that first SYN.
2019 */
2020 connp->conn_recv = tcp_input_listener_unbound;
2021
2022 connp->conn_proto = IPPROTO_TCP;
2023 connp->conn_faddr_v6 = ipv6_all_zeros;
2024 connp->conn_fport = 0;
2025
2026 (void) ipcl_bind_insert(connp);
2027 } else {
2028 tcp->tcp_state = TCPS_BOUND;
2029 }
2030
2031 /*
2032 * Initialize to default values
2033 */
2034 tcp_init_values(tcp, NULL);
2035
2036 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
2037 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
2038 int32_t, oldstate);
2039
2040 ASSERT(tcp->tcp_ptpbhn != NULL);
2041 tcp->tcp_rwnd = connp->conn_rcvbuf;
2042 tcp->tcp_mss = connp->conn_ipversion != IPV4_VERSION ?
2043 tcps->tcps_mss_def_ipv6 : tcps->tcps_mss_def_ipv4;
2044 }
2045
2046 /*
2047 * Force values to zero that need be zero.
2048 * Do not touch values asociated with the BOUND or LISTEN state
2049 * since the connection will end up in that state after the reinit.
2050 * NOTE: tcp_reinit_values MUST have a line for each field in the tcp_t
2051 * structure!
2052 */
2053 static void
2054 tcp_reinit_values(tcp)
2055 tcp_t *tcp;
2056 {
2057 tcp_stack_t *tcps = tcp->tcp_tcps;
2058 conn_t *connp = tcp->tcp_connp;
2059
2060 #ifndef lint
2061 #define DONTCARE(x)
2062 #define PRESERVE(x)
2063 #else
2064 #define DONTCARE(x) ((x) = (x))
2065 #define PRESERVE(x) ((x) = (x))
2066 #endif /* lint */
2067
2068 PRESERVE(tcp->tcp_bind_hash_port);
2069 PRESERVE(tcp->tcp_bind_hash);
2070 PRESERVE(tcp->tcp_ptpbhn);
2071 PRESERVE(tcp->tcp_acceptor_hash);
2072 PRESERVE(tcp->tcp_ptpahn);
2073
2074 /* Should be ASSERT NULL on these with new code! */
2075 ASSERT(tcp->tcp_time_wait_next == NULL);
2076 ASSERT(tcp->tcp_time_wait_prev == NULL);
2077 ASSERT(tcp->tcp_time_wait_expire == 0);
2078 PRESERVE(tcp->tcp_state);
2079 PRESERVE(connp->conn_rq);
2080 PRESERVE(connp->conn_wq);
2081
2082 ASSERT(tcp->tcp_xmit_head == NULL);
2083 ASSERT(tcp->tcp_xmit_last == NULL);
2084 ASSERT(tcp->tcp_unsent == 0);
2085 ASSERT(tcp->tcp_xmit_tail == NULL);
2086 ASSERT(tcp->tcp_xmit_tail_unsent == 0);
2087
2088 tcp->tcp_snxt = 0; /* Displayed in mib */
2089 tcp->tcp_suna = 0; /* Displayed in mib */
2090 tcp->tcp_swnd = 0;
2091 DONTCARE(tcp->tcp_cwnd); /* Init in tcp_process_options */
2092
2093 ASSERT(tcp->tcp_ibsegs == 0);
2094 ASSERT(tcp->tcp_obsegs == 0);
2095
2096 if (connp->conn_ht_iphc != NULL) {
2097 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
2098 connp->conn_ht_iphc = NULL;
2099 connp->conn_ht_iphc_allocated = 0;
2100 connp->conn_ht_iphc_len = 0;
2101 connp->conn_ht_ulp = NULL;
2102 connp->conn_ht_ulp_len = 0;
2103 tcp->tcp_ipha = NULL;
2104 tcp->tcp_ip6h = NULL;
2105 tcp->tcp_tcpha = NULL;
2106 }
2107
2108 /* We clear any IP_OPTIONS and extension headers */
2109 ip_pkt_free(&connp->conn_xmit_ipp);
2110
2111 DONTCARE(tcp->tcp_naglim); /* Init in tcp_init_values */
2112 DONTCARE(tcp->tcp_ipha);
2113 DONTCARE(tcp->tcp_ip6h);
2114 DONTCARE(tcp->tcp_tcpha);
2115 tcp->tcp_valid_bits = 0;
2116
2117 DONTCARE(tcp->tcp_timer_backoff); /* Init in tcp_init_values */
2118 DONTCARE(tcp->tcp_last_recv_time); /* Init in tcp_init_values */
2119 tcp->tcp_last_rcv_lbolt = 0;
2120
2121 tcp->tcp_init_cwnd = 0;
2122
2123 tcp->tcp_urp_last_valid = 0;
2124 tcp->tcp_hard_binding = 0;
2125
2126 tcp->tcp_fin_acked = 0;
2127 tcp->tcp_fin_rcvd = 0;
2128 tcp->tcp_fin_sent = 0;
2129 tcp->tcp_ordrel_done = 0;
2130
2131 tcp->tcp_detached = 0;
2132
2133 tcp->tcp_snd_ws_ok = B_FALSE;
2134 tcp->tcp_snd_ts_ok = B_FALSE;
2135 tcp->tcp_zero_win_probe = 0;
2136
2137 tcp->tcp_loopback = 0;
2138 tcp->tcp_localnet = 0;
2139 tcp->tcp_syn_defense = 0;
2140 tcp->tcp_set_timer = 0;
2141
2142 tcp->tcp_active_open = 0;
2143 tcp->tcp_rexmit = B_FALSE;
2144 tcp->tcp_xmit_zc_clean = B_FALSE;
2145
2146 tcp->tcp_snd_sack_ok = B_FALSE;
2147 tcp->tcp_hwcksum = B_FALSE;
2148
2149 DONTCARE(tcp->tcp_maxpsz_multiplier); /* Init in tcp_init_values */
2150
2151 tcp->tcp_conn_def_q0 = 0;
2152 tcp->tcp_ip_forward_progress = B_FALSE;
2153 tcp->tcp_ecn_ok = B_FALSE;
2154
2155 tcp->tcp_cwr = B_FALSE;
2156 tcp->tcp_ecn_echo_on = B_FALSE;
2157 tcp->tcp_is_wnd_shrnk = B_FALSE;
2158
2159 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
2160 bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
2161
2162 tcp->tcp_rcv_ws = 0;
2163 tcp->tcp_snd_ws = 0;
2164 tcp->tcp_ts_recent = 0;
2165 tcp->tcp_rnxt = 0; /* Displayed in mib */
2166 DONTCARE(tcp->tcp_rwnd); /* Set in tcp_reinit() */
2167 tcp->tcp_initial_pmtu = 0;
2168
2169 ASSERT(tcp->tcp_reass_head == NULL);
2170 ASSERT(tcp->tcp_reass_tail == NULL);
2171
2172 tcp->tcp_cwnd_cnt = 0;
2173
2174 ASSERT(tcp->tcp_rcv_list == NULL);
2175 ASSERT(tcp->tcp_rcv_last_head == NULL);
2176 ASSERT(tcp->tcp_rcv_last_tail == NULL);
2177 ASSERT(tcp->tcp_rcv_cnt == 0);
2178
2179 DONTCARE(tcp->tcp_cwnd_ssthresh); /* Init in tcp_set_destination */
2180 DONTCARE(tcp->tcp_cwnd_max); /* Init in tcp_init_values */
2181 tcp->tcp_csuna = 0;
2182
2183 tcp->tcp_rto = 0; /* Displayed in MIB */
2184 DONTCARE(tcp->tcp_rtt_sa); /* Init in tcp_init_values */
2185 DONTCARE(tcp->tcp_rtt_sd); /* Init in tcp_init_values */
2186 tcp->tcp_rtt_update = 0;
2187
2188 DONTCARE(tcp->tcp_swl1); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2189 DONTCARE(tcp->tcp_swl2); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2190
2191 tcp->tcp_rack = 0; /* Displayed in mib */
2192 tcp->tcp_rack_cnt = 0;
2193 tcp->tcp_rack_cur_max = 0;
2194 tcp->tcp_rack_abs_max = 0;
2195
2196 tcp->tcp_max_swnd = 0;
2197
2198 ASSERT(tcp->tcp_listener == NULL);
2199
2200 DONTCARE(tcp->tcp_irs); /* tcp_valid_bits cleared */
2201 DONTCARE(tcp->tcp_iss); /* tcp_valid_bits cleared */
2202 DONTCARE(tcp->tcp_fss); /* tcp_valid_bits cleared */
2203 DONTCARE(tcp->tcp_urg); /* tcp_valid_bits cleared */
2204
2205 ASSERT(tcp->tcp_conn_req_cnt_q == 0);
2206 ASSERT(tcp->tcp_conn_req_cnt_q0 == 0);
2207 PRESERVE(tcp->tcp_conn_req_max);
2208 PRESERVE(tcp->tcp_conn_req_seqnum);
2209
2210 DONTCARE(tcp->tcp_first_timer_threshold); /* Init in tcp_init_values */
2211 DONTCARE(tcp->tcp_second_timer_threshold); /* Init in tcp_init_values */
2212 DONTCARE(tcp->tcp_first_ctimer_threshold); /* Init in tcp_init_values */
2213 DONTCARE(tcp->tcp_second_ctimer_threshold); /* in tcp_init_values */
2214
2215 DONTCARE(tcp->tcp_urp_last); /* tcp_urp_last_valid is cleared */
2216 ASSERT(tcp->tcp_urp_mp == NULL);
2217 ASSERT(tcp->tcp_urp_mark_mp == NULL);
2218 ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
2219
2220 ASSERT(tcp->tcp_eager_next_q == NULL);
2221 ASSERT(tcp->tcp_eager_last_q == NULL);
2222 ASSERT((tcp->tcp_eager_next_q0 == NULL &&
2223 tcp->tcp_eager_prev_q0 == NULL) ||
2224 tcp->tcp_eager_next_q0 == tcp->tcp_eager_prev_q0);
2225 ASSERT(tcp->tcp_conn.tcp_eager_conn_ind == NULL);
2226
2227 ASSERT((tcp->tcp_eager_next_drop_q0 == NULL &&
2228 tcp->tcp_eager_prev_drop_q0 == NULL) ||
2229 tcp->tcp_eager_next_drop_q0 == tcp->tcp_eager_prev_drop_q0);
2230
2231 tcp->tcp_client_errno = 0;
2232
2233 DONTCARE(connp->conn_sum); /* Init in tcp_init_values */
2234
2235 connp->conn_faddr_v6 = ipv6_all_zeros; /* Displayed in MIB */
2236
2237 PRESERVE(connp->conn_bound_addr_v6);
2238 tcp->tcp_last_sent_len = 0;
2239 tcp->tcp_dupack_cnt = 0;
2240
2241 connp->conn_fport = 0; /* Displayed in MIB */
2242 PRESERVE(connp->conn_lport);
2243
2244 PRESERVE(tcp->tcp_acceptor_lockp);
2245
2246 ASSERT(tcp->tcp_ordrel_mp == NULL);
2247 PRESERVE(tcp->tcp_acceptor_id);
2248 DONTCARE(tcp->tcp_ipsec_overhead);
2249
2250 PRESERVE(connp->conn_family);
2251 /* Remove any remnants of mapped address binding */
2252 if (connp->conn_family == AF_INET6) {
2253 connp->conn_ipversion = IPV6_VERSION;
2254 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2255 } else {
2256 connp->conn_ipversion = IPV4_VERSION;
2257 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2258 }
2259
2260 connp->conn_bound_if = 0;
2261 connp->conn_recv_ancillary.crb_all = 0;
2262 tcp->tcp_recvifindex = 0;
2263 tcp->tcp_recvhops = 0;
2264 tcp->tcp_closed = 0;
2265 if (tcp->tcp_hopopts != NULL) {
2266 mi_free(tcp->tcp_hopopts);
2267 tcp->tcp_hopopts = NULL;
2268 tcp->tcp_hopoptslen = 0;
2269 }
2270 ASSERT(tcp->tcp_hopoptslen == 0);
2271 if (tcp->tcp_dstopts != NULL) {
2272 mi_free(tcp->tcp_dstopts);
2273 tcp->tcp_dstopts = NULL;
2274 tcp->tcp_dstoptslen = 0;
2275 }
2276 ASSERT(tcp->tcp_dstoptslen == 0);
2277 if (tcp->tcp_rthdrdstopts != NULL) {
2278 mi_free(tcp->tcp_rthdrdstopts);
2279 tcp->tcp_rthdrdstopts = NULL;
2280 tcp->tcp_rthdrdstoptslen = 0;
2281 }
2282 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2283 if (tcp->tcp_rthdr != NULL) {
2284 mi_free(tcp->tcp_rthdr);
2285 tcp->tcp_rthdr = NULL;
2286 tcp->tcp_rthdrlen = 0;
2287 }
2288 ASSERT(tcp->tcp_rthdrlen == 0);
2289
2290 /* Reset fusion-related fields */
2291 tcp->tcp_fused = B_FALSE;
2292 tcp->tcp_unfusable = B_FALSE;
2293 tcp->tcp_fused_sigurg = B_FALSE;
2294 tcp->tcp_loopback_peer = NULL;
2295
2296 tcp->tcp_lso = B_FALSE;
2297
2298 tcp->tcp_in_ack_unsent = 0;
2299 tcp->tcp_cork = B_FALSE;
2300 tcp->tcp_tconnind_started = B_FALSE;
2301
2302 PRESERVE(tcp->tcp_squeue_bytes);
2303
2304 tcp->tcp_closemp_used = B_FALSE;
2305
2306 PRESERVE(tcp->tcp_rsrv_mp);
2307 PRESERVE(tcp->tcp_rsrv_mp_lock);
2308
2309 #ifdef DEBUG
2310 DONTCARE(tcp->tcmp_stk[0]);
2311 #endif
2312
2313 PRESERVE(tcp->tcp_connid);
2314
2315 ASSERT(tcp->tcp_listen_cnt == NULL);
2316 ASSERT(tcp->tcp_reass_tid == 0);
2317
2318 #undef DONTCARE
2319 #undef PRESERVE
2320 }
2321
2322 /*
2323 * Initialize the various fields in tcp_t. If parent (the listener) is non
2324 * NULL, certain values will be inheritted from it.
2325 */
2326 void
2327 tcp_init_values(tcp_t *tcp, tcp_t *parent)
2328 {
2329 tcp_stack_t *tcps = tcp->tcp_tcps;
2330 conn_t *connp = tcp->tcp_connp;
2331 clock_t rto;
2332
2333 ASSERT((connp->conn_family == AF_INET &&
2334 connp->conn_ipversion == IPV4_VERSION) ||
2335 (connp->conn_family == AF_INET6 &&
2336 (connp->conn_ipversion == IPV4_VERSION ||
2337 connp->conn_ipversion == IPV6_VERSION)));
2338
2339 if (parent == NULL) {
2340 tcp->tcp_naglim = tcps->tcps_naglim_def;
2341
2342 tcp->tcp_rto_initial = tcps->tcps_rexmit_interval_initial;
2343 tcp->tcp_rto_min = tcps->tcps_rexmit_interval_min;
2344 tcp->tcp_rto_max = tcps->tcps_rexmit_interval_max;
2345
2346 tcp->tcp_first_ctimer_threshold =
2347 tcps->tcps_ip_notify_cinterval;
2348 tcp->tcp_second_ctimer_threshold =
2349 tcps->tcps_ip_abort_cinterval;
2350 tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2351 tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2352
2353 tcp->tcp_fin_wait_2_flush_interval =
2354 tcps->tcps_fin_wait_2_flush_interval;
2355
2356 tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2357 tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2358 tcp->tcp_ka_cnt = 0;
2359 tcp->tcp_ka_rinterval = 0;
2360
2361 /*
2362 * Default value of tcp_init_cwnd is 0, so no need to set here
2363 * if parent is NULL. But we need to inherit it from parent.
2364 */
2365 } else {
2366 /* Inherit various TCP parameters from the parent. */
2367 tcp->tcp_naglim = parent->tcp_naglim;
2368
2369 tcp->tcp_rto_initial = parent->tcp_rto_initial;
2370 tcp->tcp_rto_min = parent->tcp_rto_min;
2371 tcp->tcp_rto_max = parent->tcp_rto_max;
2372
2373 tcp->tcp_first_ctimer_threshold =
2374 parent->tcp_first_ctimer_threshold;
2375 tcp->tcp_second_ctimer_threshold =
2376 parent->tcp_second_ctimer_threshold;
2377 tcp->tcp_first_timer_threshold =
2378 parent->tcp_first_timer_threshold;
2379 tcp->tcp_second_timer_threshold =
2380 parent->tcp_second_timer_threshold;
2381
2382 tcp->tcp_fin_wait_2_flush_interval =
2383 parent->tcp_fin_wait_2_flush_interval;
2384
2385 tcp->tcp_ka_interval = parent->tcp_ka_interval;
2386 tcp->tcp_ka_abort_thres = parent->tcp_ka_abort_thres;
2387
2388 tcp->tcp_init_cwnd = parent->tcp_init_cwnd;
2389 }
2390
2391 /*
2392 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2393 * will be close to tcp_rexmit_interval_initial. By doing this, we
2394 * allow the algorithm to adjust slowly to large fluctuations of RTT
2395 * during first few transmissions of a connection as seen in slow
2396 * links.
2397 */
2398 tcp->tcp_rtt_sa = tcp->tcp_rto_initial << 2;
2399 tcp->tcp_rtt_sd = tcp->tcp_rto_initial >> 1;
2400 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
2401 tcps->tcps_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
2402 tcps->tcps_conn_grace_period;
2403 TCP_SET_RTO(tcp, rto);
2404
2405 tcp->tcp_timer_backoff = 0;
2406 tcp->tcp_ms_we_have_waited = 0;
2407 tcp->tcp_last_recv_time = ddi_get_lbolt();
2408 tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2409 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2410 tcp->tcp_snd_burst = TCP_CWND_INFINITE;
2411
2412 tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2413
2414 /* NOTE: ISS is now set in tcp_set_destination(). */
2415
2416 /* Reset fusion-related fields */
2417 tcp->tcp_fused = B_FALSE;
2418 tcp->tcp_unfusable = B_FALSE;
2419 tcp->tcp_fused_sigurg = B_FALSE;
2420 tcp->tcp_loopback_peer = NULL;
2421
2422 /* We rebuild the header template on the next connect/conn_request */
2423
2424 connp->conn_mlp_type = mlptSingle;
2425
2426 /*
2427 * Init the window scale to the max so tcp_rwnd_set() won't pare
2428 * down tcp_rwnd. tcp_set_destination() will set the right value later.
2429 */
2430 tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2431 tcp->tcp_rwnd = connp->conn_rcvbuf;
2432
2433 tcp->tcp_cork = B_FALSE;
2434 /*
2435 * Init the tcp_debug option if it wasn't already set. This value
2436 * determines whether TCP
2437 * calls strlog() to print out debug messages. Doing this
2438 * initialization here means that this value is not inherited thru
2439 * tcp_reinit().
2440 */
2441 if (!connp->conn_debug)
2442 connp->conn_debug = tcps->tcps_dbg;
2443 }
2444
2445 /*
2446 * Update the TCP connection according to change of PMTU.
2447 *
2448 * Path MTU might have changed by either increase or decrease, so need to
2449 * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2450 * or negative MSS, since tcp_mss_set() will do it.
2451 */
2452 void
2453 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2454 {
2455 uint32_t pmtu;
2456 int32_t mss;
2457 conn_t *connp = tcp->tcp_connp;
2458 ip_xmit_attr_t *ixa = connp->conn_ixa;
2459 iaflags_t ixaflags;
2460
2461 if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2462 return;
2463
2464 if (tcp->tcp_state < TCPS_ESTABLISHED)
2465 return;
2466
2467 /*
2468 * Always call ip_get_pmtu() to make sure that IP has updated
2469 * ixa_flags properly.
2470 */
2471 pmtu = ip_get_pmtu(ixa);
2472 ixaflags = ixa->ixa_flags;
2473
2474 /*
2475 * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2476 * IPsec overhead if applied. Make sure to use the most recent
2477 * IPsec information.
2478 */
2479 mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2480
2481 /*
2482 * Nothing to change, so just return.
2483 */
2484 if (mss == tcp->tcp_mss)
2485 return;
2486
2487 /*
2488 * Currently, for ICMP errors, only PMTU decrease is handled.
2489 */
2490 if (mss > tcp->tcp_mss && decrease_only)
2491 return;
2492
2493 DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2494
2495 /*
2496 * Update ixa_fragsize and ixa_pmtu.
2497 */
2498 ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2499
2500 /*
2501 * Adjust MSS and all relevant variables.
2502 */
2503 tcp_mss_set(tcp, mss);
2504
2505 /*
2506 * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2507 * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2508 * has a (potentially different) min size we do the same. Make sure to
2509 * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2510 * fragment the packet.
2511 *
2512 * LSO over IPv6 can not be fragmented. So need to disable LSO
2513 * when IPv6 fragmentation is needed.
2514 */
2515 if (mss < tcp->tcp_tcps->tcps_mss_min)
2516 ixaflags |= IXAF_PMTU_TOO_SMALL;
2517
2518 if (ixaflags & IXAF_PMTU_TOO_SMALL)
2519 ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2520
2521 if ((connp->conn_ipversion == IPV4_VERSION) &&
2522 !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2523 tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2524 }
2525 ixa->ixa_flags = ixaflags;
2526 }
2527
2528 int
2529 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2530 {
2531 conn_t *connp = tcp->tcp_connp;
2532 queue_t *q = connp->conn_rq;
2533 int32_t mss = tcp->tcp_mss;
2534 int maxpsz;
2535
2536 if (TCP_IS_DETACHED(tcp))
2537 return (mss);
2538 if (tcp->tcp_fused) {
2539 maxpsz = tcp_fuse_maxpsz(tcp);
2540 mss = INFPSZ;
2541 } else if (tcp->tcp_maxpsz_multiplier == 0) {
2542 /*
2543 * Set the sd_qn_maxpsz according to the socket send buffer
2544 * size, and sd_maxblk to INFPSZ (-1). This will essentially
2545 * instruct the stream head to copyin user data into contiguous
2546 * kernel-allocated buffers without breaking it up into smaller
2547 * chunks. We round up the buffer size to the nearest SMSS.
2548 */
2549 maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2550 mss = INFPSZ;
2551 } else {
2552 /*
2553 * Set sd_qn_maxpsz to approx half the (receivers) buffer
2554 * (and a multiple of the mss). This instructs the stream
2555 * head to break down larger than SMSS writes into SMSS-
2556 * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2557 */
2558 maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2559 if (maxpsz > connp->conn_sndbuf / 2) {
2560 maxpsz = connp->conn_sndbuf / 2;
2561 /* Round up to nearest mss */
2562 maxpsz = MSS_ROUNDUP(maxpsz, mss);
2563 }
2564 }
2565
2566 (void) proto_set_maxpsz(q, connp, maxpsz);
2567 if (!(IPCL_IS_NONSTR(connp)))
2568 connp->conn_wq->q_maxpsz = maxpsz;
2569 if (set_maxblk)
2570 (void) proto_set_tx_maxblk(q, connp, mss);
2571 return (mss);
2572 }
2573
2574 /* For /dev/tcp aka AF_INET open */
2575 static int
2576 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2577 {
2578 return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2579 }
2580
2581 /* For /dev/tcp6 aka AF_INET6 open */
2582 static int
2583 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2584 {
2585 return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2586 }
2587
2588 conn_t *
2589 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2590 int *errorp)
2591 {
2592 tcp_t *tcp = NULL;
2593 conn_t *connp;
2594 zoneid_t zoneid;
2595 tcp_stack_t *tcps;
2596 squeue_t *sqp;
2597
2598 ASSERT(errorp != NULL);
2599 /*
2600 * Find the proper zoneid and netstack.
2601 */
2602 /*
2603 * Special case for install: miniroot needs to be able to
2604 * access files via NFS as though it were always in the
2605 * global zone.
2606 */
2607 if (credp == kcred && nfs_global_client_only != 0) {
2608 zoneid = GLOBAL_ZONEID;
2609 tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2610 netstack_tcp;
2611 ASSERT(tcps != NULL);
2612 } else {
2613 netstack_t *ns;
2614 int err;
2615
2616 if ((err = secpolicy_basic_net_access(credp)) != 0) {
2617 *errorp = err;
2618 return (NULL);
2619 }
2620
2621 ns = netstack_find_by_cred(credp);
2622 ASSERT(ns != NULL);
2623 tcps = ns->netstack_tcp;
2624 ASSERT(tcps != NULL);
2625
2626 /*
2627 * For exclusive stacks we set the zoneid to zero
2628 * to make TCP operate as if in the global zone.
2629 */
2630 if (tcps->tcps_netstack->netstack_stackid !=
2631 GLOBAL_NETSTACKID)
2632 zoneid = GLOBAL_ZONEID;
2633 else
2634 zoneid = crgetzoneid(credp);
2635 }
2636
2637 sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2638 connp = (conn_t *)tcp_get_conn(sqp, tcps);
2639 /*
2640 * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2641 * so we drop it by one.
2642 */
2643 netstack_rele(tcps->tcps_netstack);
2644 if (connp == NULL) {
2645 *errorp = ENOSR;
2646 return (NULL);
2647 }
2648 ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2649
2650 connp->conn_sqp = sqp;
2651 connp->conn_initial_sqp = connp->conn_sqp;
2652 connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2653 tcp = connp->conn_tcp;
2654
2655 /*
2656 * Besides asking IP to set the checksum for us, have conn_ip_output
2657 * to do the following checks when necessary:
2658 *
2659 * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2660 * IXAF_VERIFY_PMTU: verify PMTU changes
2661 * IXAF_VERIFY_LSO: verify LSO capability changes
2662 */
2663 connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2664 IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2665
2666 if (!tcps->tcps_dev_flow_ctl)
2667 connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2668
2669 if (isv6) {
2670 connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2671 connp->conn_ipversion = IPV6_VERSION;
2672 connp->conn_family = AF_INET6;
2673 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2674 connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2675 } else {
2676 connp->conn_ipversion = IPV4_VERSION;
2677 connp->conn_family = AF_INET;
2678 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2679 connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2680 }
2681 connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2682
2683 crhold(credp);
2684 connp->conn_cred = credp;
2685 connp->conn_cpid = curproc->p_pid;
2686 connp->conn_open_time = ddi_get_lbolt64();
2687
2688 /* Cache things in the ixa without any refhold */
2689 ASSERT(!(connp->conn_ixa->ixa_free_flags & IXA_FREE_CRED));
2690 connp->conn_ixa->ixa_cred = credp;
2691 connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2692
2693 connp->conn_zoneid = zoneid;
2694 /* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2695 connp->conn_ixa->ixa_zoneid = zoneid;
2696 connp->conn_mlp_type = mlptSingle;
2697 ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2698 ASSERT(tcp->tcp_tcps == tcps);
2699
2700 /*
2701 * If the caller has the process-wide flag set, then default to MAC
2702 * exempt mode. This allows read-down to unlabeled hosts.
2703 */
2704 if (getpflags(NET_MAC_AWARE, credp) != 0)
2705 connp->conn_mac_mode = CONN_MAC_AWARE;
2706
2707 connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2708
2709 if (issocket) {
2710 tcp->tcp_issocket = 1;
2711 }
2712
2713 connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2714 connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2715 connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2716 connp->conn_so_type = SOCK_STREAM;
2717 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2718 tcps->tcps_wroff_xtra;
2719
2720 SOCK_CONNID_INIT(tcp->tcp_connid);
2721 /* DTrace ignores this - it isn't a tcp:::state-change */
2722 tcp->tcp_state = TCPS_IDLE;
2723 tcp_init_values(tcp, NULL);
2724 return (connp);
2725 }
2726
2727 static int
2728 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2729 boolean_t isv6)
2730 {
2731 tcp_t *tcp = NULL;
2732 conn_t *connp = NULL;
2733 int err;
2734 vmem_t *minor_arena = NULL;
2735 dev_t conn_dev;
2736 boolean_t issocket;
2737
2738 if (q->q_ptr != NULL)
2739 return (0);
2740
2741 if (sflag == MODOPEN)
2742 return (EINVAL);
2743
2744 if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2745 ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2746 minor_arena = ip_minor_arena_la;
2747 } else {
2748 /*
2749 * Either minor numbers in the large arena were exhausted
2750 * or a non socket application is doing the open.
2751 * Try to allocate from the small arena.
2752 */
2753 if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2754 return (EBUSY);
2755 }
2756 minor_arena = ip_minor_arena_sa;
2757 }
2758
2759 ASSERT(minor_arena != NULL);
2760
2761 *devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2762
2763 if (flag & SO_FALLBACK) {
2764 /*
2765 * Non streams socket needs a stream to fallback to
2766 */
2767 RD(q)->q_ptr = (void *)conn_dev;
2768 WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2769 WR(q)->q_ptr = (void *)minor_arena;
2770 qprocson(q);
2771 return (0);
2772 } else if (flag & SO_ACCEPTOR) {
2773 q->q_qinfo = &tcp_acceptor_rinit;
2774 /*
2775 * the conn_dev and minor_arena will be subsequently used by
2776 * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2777 * the minor device number for this connection from the q_ptr.
2778 */
2779 RD(q)->q_ptr = (void *)conn_dev;
2780 WR(q)->q_qinfo = &tcp_acceptor_winit;
2781 WR(q)->q_ptr = (void *)minor_arena;
2782 qprocson(q);
2783 return (0);
2784 }
2785
2786 issocket = flag & SO_SOCKSTR;
2787 connp = tcp_create_common(credp, isv6, issocket, &err);
2788
2789 if (connp == NULL) {
2790 inet_minor_free(minor_arena, conn_dev);
2791 q->q_ptr = WR(q)->q_ptr = NULL;
2792 return (err);
2793 }
2794
2795 connp->conn_rq = q;
2796 connp->conn_wq = WR(q);
2797 q->q_ptr = WR(q)->q_ptr = connp;
2798
2799 connp->conn_dev = conn_dev;
2800 connp->conn_minor_arena = minor_arena;
2801
2802 ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2803 ASSERT(WR(q)->q_qinfo == &tcp_winit);
2804
2805 tcp = connp->conn_tcp;
2806
2807 if (issocket) {
2808 WR(q)->q_qinfo = &tcp_sock_winit;
2809 } else {
2810 #ifdef _ILP32
2811 tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2812 #else
2813 tcp->tcp_acceptor_id = conn_dev;
2814 #endif /* _ILP32 */
2815 tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2816 }
2817
2818 /*
2819 * Put the ref for TCP. Ref for IP was already put
2820 * by ipcl_conn_create. Also Make the conn_t globally
2821 * visible to walkers
2822 */
2823 mutex_enter(&connp->conn_lock);
2824 CONN_INC_REF_LOCKED(connp);
2825 ASSERT(connp->conn_ref == 2);
2826 connp->conn_state_flags &= ~CONN_INCIPIENT;
2827 mutex_exit(&connp->conn_lock);
2828
2829 qprocson(q);
2830 return (0);
2831 }
2832
2833 /*
2834 * Build/update the tcp header template (in conn_ht_iphc) based on
2835 * conn_xmit_ipp. The headers include ip6_t, any extension
2836 * headers, and the maximum size tcp header (to avoid reallocation
2837 * on the fly for additional tcp options).
2838 *
2839 * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2840 * Returns failure if can't allocate memory.
2841 */
2842 int
2843 tcp_build_hdrs(tcp_t *tcp)
2844 {
2845 tcp_stack_t *tcps = tcp->tcp_tcps;
2846 conn_t *connp = tcp->tcp_connp;
2847 char buf[TCP_MAX_HDR_LENGTH];
2848 uint_t buflen;
2849 uint_t ulplen = TCP_MIN_HEADER_LENGTH;
2850 uint_t extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
2851 tcpha_t *tcpha;
2852 uint32_t cksum;
2853 int error;
2854
2855 /*
2856 * We might be called after the connection is set up, and we might
2857 * have TS options already in the TCP header. Thus we save any
2858 * existing tcp header.
2859 */
2860 buflen = connp->conn_ht_ulp_len;
2861 if (buflen != 0) {
2862 bcopy(connp->conn_ht_ulp, buf, buflen);
2863 extralen -= buflen - ulplen;
2864 ulplen = buflen;
2865 }
2866
2867 /* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
2868 mutex_enter(&connp->conn_lock);
2869 error = conn_build_hdr_template(connp, ulplen, extralen,
2870 &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
2871 mutex_exit(&connp->conn_lock);
2872 if (error != 0)
2873 return (error);
2874
2875 /*
2876 * Any routing header/option has been massaged. The checksum difference
2877 * is stored in conn_sum for later use.
2878 */
2879 tcpha = (tcpha_t *)connp->conn_ht_ulp;
2880 tcp->tcp_tcpha = tcpha;
2881
2882 /* restore any old tcp header */
2883 if (buflen != 0) {
2884 bcopy(buf, connp->conn_ht_ulp, buflen);
2885 } else {
2886 tcpha->tha_sum = 0;
2887 tcpha->tha_urp = 0;
2888 tcpha->tha_ack = 0;
2889 tcpha->tha_offset_and_reserved = (5 << 4);
2890 tcpha->tha_lport = connp->conn_lport;
2891 tcpha->tha_fport = connp->conn_fport;
2892 }
2893
2894 /*
2895 * IP wants our header length in the checksum field to
2896 * allow it to perform a single pseudo-header+checksum
2897 * calculation on behalf of TCP.
2898 * Include the adjustment for a source route once IP_OPTIONS is set.
2899 */
2900 cksum = sizeof (tcpha_t) + connp->conn_sum;
2901 cksum = (cksum >> 16) + (cksum & 0xFFFF);
2902 ASSERT(cksum < 0x10000);
2903 tcpha->tha_sum = htons(cksum);
2904
2905 if (connp->conn_ipversion == IPV4_VERSION)
2906 tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
2907 else
2908 tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
2909
2910 if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
2911 connp->conn_wroff) {
2912 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2913 tcps->tcps_wroff_xtra;
2914 (void) proto_set_tx_wroff(connp->conn_rq, connp,
2915 connp->conn_wroff);
2916 }
2917 return (0);
2918 }
2919
2920 /*
2921 * tcp_rwnd_set() is called to adjust the receive window to a desired value.
2922 * We do not allow the receive window to shrink. After setting rwnd,
2923 * set the flow control hiwat of the stream.
2924 *
2925 * This function is called in 2 cases:
2926 *
2927 * 1) Before data transfer begins, in tcp_input_listener() for accepting a
2928 * connection (passive open) and in tcp_input_data() for active connect.
2929 * This is called after tcp_mss_set() when the desired MSS value is known.
2930 * This makes sure that our window size is a mutiple of the other side's
2931 * MSS.
2932 * 2) Handling SO_RCVBUF option.
2933 *
2934 * It is ASSUMED that the requested size is a multiple of the current MSS.
2935 *
2936 * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
2937 * user requests so.
2938 */
2939 int
2940 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
2941 {
2942 uint32_t mss = tcp->tcp_mss;
2943 uint32_t old_max_rwnd;
2944 uint32_t max_transmittable_rwnd;
2945 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
2946 tcp_stack_t *tcps = tcp->tcp_tcps;
2947 conn_t *connp = tcp->tcp_connp;
2948
2949 /*
2950 * Insist on a receive window that is at least
2951 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
2952 * funny TCP interactions of Nagle algorithm, SWS avoidance
2953 * and delayed acknowledgement.
2954 */
2955 rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
2956
2957 if (tcp->tcp_fused) {
2958 size_t sth_hiwat;
2959 tcp_t *peer_tcp = tcp->tcp_loopback_peer;
2960
2961 ASSERT(peer_tcp != NULL);
2962 sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
2963 if (!tcp_detached) {
2964 (void) proto_set_rx_hiwat(connp->conn_rq, connp,
2965 sth_hiwat);
2966 tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
2967 }
2968
2969 /* Caller could have changed tcp_rwnd; update tha_win */
2970 if (tcp->tcp_tcpha != NULL) {
2971 tcp->tcp_tcpha->tha_win =
2972 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2973 }
2974 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
2975 tcp->tcp_cwnd_max = rwnd;
2976
2977 /*
2978 * In the fusion case, the maxpsz stream head value of
2979 * our peer is set according to its send buffer size
2980 * and our receive buffer size; since the latter may
2981 * have changed we need to update the peer's maxpsz.
2982 */
2983 (void) tcp_maxpsz_set(peer_tcp, B_TRUE);
2984 return (sth_hiwat);
2985 }
2986
2987 if (tcp_detached)
2988 old_max_rwnd = tcp->tcp_rwnd;
2989 else
2990 old_max_rwnd = connp->conn_rcvbuf;
2991
2992
2993 /*
2994 * If window size info has already been exchanged, TCP should not
2995 * shrink the window. Shrinking window is doable if done carefully.
2996 * We may add that support later. But so far there is not a real
2997 * need to do that.
2998 */
2999 if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
3000 /* MSS may have changed, do a round up again. */
3001 rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
3002 }
3003
3004 /*
3005 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3006 * can be applied even before the window scale option is decided.
3007 */
3008 max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3009 if (rwnd > max_transmittable_rwnd) {
3010 rwnd = max_transmittable_rwnd -
3011 (max_transmittable_rwnd % mss);
3012 if (rwnd < mss)
3013 rwnd = max_transmittable_rwnd;
3014 /*
3015 * If we're over the limit we may have to back down tcp_rwnd.
3016 * The increment below won't work for us. So we set all three
3017 * here and the increment below will have no effect.
3018 */
3019 tcp->tcp_rwnd = old_max_rwnd = rwnd;
3020 }
3021 if (tcp->tcp_localnet) {
3022 tcp->tcp_rack_abs_max =
3023 MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3024 } else {
3025 /*
3026 * For a remote host on a different subnet (through a router),
3027 * we ack every other packet to be conforming to RFC1122.
3028 * tcp_deferred_acks_max is default to 2.
3029 */
3030 tcp->tcp_rack_abs_max =
3031 MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3032 }
3033 if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3034 tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3035 else
3036 tcp->tcp_rack_cur_max = 0;
3037 /*
3038 * Increment the current rwnd by the amount the maximum grew (we
3039 * can not overwrite it since we might be in the middle of a
3040 * connection.)
3041 */
3042 tcp->tcp_rwnd += rwnd - old_max_rwnd;
3043 connp->conn_rcvbuf = rwnd;
3044
3045 /* Are we already connected? */
3046 if (tcp->tcp_tcpha != NULL) {
3047 tcp->tcp_tcpha->tha_win =
3048 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3049 }
3050
3051 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3052 tcp->tcp_cwnd_max = rwnd;
3053
3054 if (tcp_detached)
3055 return (rwnd);
3056
3057 tcp_set_recv_threshold(tcp, rwnd >> 3);
3058
3059 (void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3060 return (rwnd);
3061 }
3062
3063 int
3064 tcp_do_unbind(conn_t *connp)
3065 {
3066 tcp_t *tcp = connp->conn_tcp;
3067 int32_t oldstate;
3068
3069 switch (tcp->tcp_state) {
3070 case TCPS_BOUND:
3071 case TCPS_LISTEN:
3072 break;
3073 default:
3074 return (-TOUTSTATE);
3075 }
3076
3077 /*
3078 * Need to clean up all the eagers since after the unbind, segments
3079 * will no longer be delivered to this listener stream.
3080 */
3081 mutex_enter(&tcp->tcp_eager_lock);
3082 if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3083 tcp_eager_cleanup(tcp, 0);
3084 }
3085 mutex_exit(&tcp->tcp_eager_lock);
3086
3087 /* Clean up the listener connection counter if necessary. */
3088 if (tcp->tcp_listen_cnt != NULL)
3089 TCP_DECR_LISTEN_CNT(tcp);
3090 connp->conn_laddr_v6 = ipv6_all_zeros;
3091 connp->conn_saddr_v6 = ipv6_all_zeros;
3092 tcp_bind_hash_remove(tcp);
3093 oldstate = tcp->tcp_state;
3094 tcp->tcp_state = TCPS_IDLE;
3095 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
3096 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
3097 int32_t, oldstate);
3098
3099 ip_unbind(connp);
3100 bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3101
3102 return (0);
3103 }
3104
3105 /*
3106 * Collect protocol properties to send to the upper handle.
3107 */
3108 void
3109 tcp_get_proto_props(tcp_t *tcp, struct sock_proto_props *sopp)
3110 {
3111 conn_t *connp = tcp->tcp_connp;
3112
3113 sopp->sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3114 sopp->sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3115
3116 sopp->sopp_rxhiwat = tcp->tcp_fused ?
3117 tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3118 connp->conn_rcvbuf;
3119 /*
3120 * Determine what write offset value to use depending on SACK and
3121 * whether the endpoint is fused or not.
3122 */
3123 if (tcp->tcp_fused) {
3124 ASSERT(tcp->tcp_loopback);
3125 ASSERT(tcp->tcp_loopback_peer != NULL);
3126 /*
3127 * For fused tcp loopback, set the stream head's write
3128 * offset value to zero since we won't be needing any room
3129 * for TCP/IP headers. This would also improve performance
3130 * since it would reduce the amount of work done by kmem.
3131 * Non-fused tcp loopback case is handled separately below.
3132 */
3133 sopp->sopp_wroff = 0;
3134 /*
3135 * Update the peer's transmit parameters according to
3136 * our recently calculated high water mark value.
3137 */
3138 (void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3139 } else if (tcp->tcp_snd_sack_ok) {
3140 sopp->sopp_wroff = connp->conn_ht_iphc_allocated +
3141 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3142 } else {
3143 sopp->sopp_wroff = connp->conn_ht_iphc_len +
3144 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3145 }
3146
3147 if (tcp->tcp_loopback) {
3148 sopp->sopp_flags |= SOCKOPT_LOOPBACK;
3149 sopp->sopp_loopback = B_TRUE;
3150 }
3151 }
3152
3153 /*
3154 * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3155 */
3156 boolean_t
3157 tcp_zcopy_check(tcp_t *tcp)
3158 {
3159 conn_t *connp = tcp->tcp_connp;
3160 ip_xmit_attr_t *ixa = connp->conn_ixa;
3161 boolean_t zc_enabled = B_FALSE;
3162 tcp_stack_t *tcps = tcp->tcp_tcps;
3163
3164 if (do_tcpzcopy == 2)
3165 zc_enabled = B_TRUE;
3166 else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3167 zc_enabled = B_TRUE;
3168
3169 tcp->tcp_snd_zcopy_on = zc_enabled;
3170 if (!TCP_IS_DETACHED(tcp)) {
3171 if (zc_enabled) {
3172 ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3173 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3174 ZCVMSAFE);
3175 TCP_STAT(tcps, tcp_zcopy_on);
3176 } else {
3177 ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3178 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3179 ZCVMUNSAFE);
3180 TCP_STAT(tcps, tcp_zcopy_off);
3181 }
3182 }
3183 return (zc_enabled);
3184 }
3185
3186 /*
3187 * Backoff from a zero-copy message by copying data to a new allocated
3188 * message and freeing the original desballoca'ed segmapped message.
3189 *
3190 * This function is called by following two callers:
3191 * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3192 * the origial desballoca'ed message and notify sockfs. This is in re-
3193 * transmit state.
3194 * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3195 * to be copied to new message.
3196 */
3197 mblk_t *
3198 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3199 {
3200 mblk_t *nbp;
3201 mblk_t *head = NULL;
3202 mblk_t *tail = NULL;
3203 tcp_stack_t *tcps = tcp->tcp_tcps;
3204
3205 ASSERT(bp != NULL);
3206 while (bp != NULL) {
3207 if (IS_VMLOANED_MBLK(bp)) {
3208 TCP_STAT(tcps, tcp_zcopy_backoff);
3209 if ((nbp = copyb(bp)) == NULL) {
3210 tcp->tcp_xmit_zc_clean = B_FALSE;
3211 if (tail != NULL)
3212 tail->b_cont = bp;
3213 return ((head == NULL) ? bp : head);
3214 }
3215
3216 if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3217 if (fix_xmitlist)
3218 tcp_zcopy_notify(tcp);
3219 else
3220 nbp->b_datap->db_struioflag |=
3221 STRUIO_ZCNOTIFY;
3222 }
3223 nbp->b_cont = bp->b_cont;
3224
3225 /*
3226 * Copy saved information and adjust tcp_xmit_tail
3227 * if needed.
3228 */
3229 if (fix_xmitlist) {
3230 nbp->b_prev = bp->b_prev;
3231 nbp->b_next = bp->b_next;
3232
3233 if (tcp->tcp_xmit_tail == bp)
3234 tcp->tcp_xmit_tail = nbp;
3235 }
3236
3237 /* Free the original message. */
3238 bp->b_prev = NULL;
3239 bp->b_next = NULL;
3240 freeb(bp);
3241
3242 bp = nbp;
3243 }
3244
3245 if (head == NULL) {
3246 head = bp;
3247 }
3248 if (tail == NULL) {
3249 tail = bp;
3250 } else {
3251 tail->b_cont = bp;
3252 tail = bp;
3253 }
3254
3255 /* Move forward. */
3256 bp = bp->b_cont;
3257 }
3258
3259 if (fix_xmitlist) {
3260 tcp->tcp_xmit_last = tail;
3261 tcp->tcp_xmit_zc_clean = B_TRUE;
3262 }
3263
3264 return (head);
3265 }
3266
3267 void
3268 tcp_zcopy_notify(tcp_t *tcp)
3269 {
3270 struct stdata *stp;
3271 conn_t *connp;
3272
3273 if (tcp->tcp_detached)
3274 return;
3275 connp = tcp->tcp_connp;
3276 if (IPCL_IS_NONSTR(connp)) {
3277 (*connp->conn_upcalls->su_zcopy_notify)
3278 (connp->conn_upper_handle);
3279 return;
3280 }
3281 stp = STREAM(connp->conn_rq);
3282 mutex_enter(&stp->sd_lock);
3283 stp->sd_flag |= STZCNOTIFY;
3284 cv_broadcast(&stp->sd_zcopy_wait);
3285 mutex_exit(&stp->sd_lock);
3286 }
3287
3288 /*
3289 * Update the TCP connection according to change of LSO capability.
3290 */
3291 static void
3292 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3293 {
3294 /*
3295 * We check against IPv4 header length to preserve the old behavior
3296 * of only enabling LSO when there are no IP options.
3297 * But this restriction might not be necessary at all. Before removing
3298 * it, need to verify how LSO is handled for source routing case, with
3299 * which IP does software checksum.
3300 *
3301 * For IPv6, whenever any extension header is needed, LSO is supressed.
3302 */
3303 if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3304 IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3305 return;
3306
3307 /*
3308 * Either the LSO capability newly became usable, or it has changed.
3309 */
3310 if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3311 ill_lso_capab_t *lsoc = &ixa->ixa_lso_capab;
3312
3313 ASSERT(lsoc->ill_lso_max > 0);
3314 tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lsoc->ill_lso_max);
3315
3316 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3317 boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3318
3319 /*
3320 * If LSO to be enabled, notify the STREAM header with larger
3321 * data block.
3322 */
3323 if (!tcp->tcp_lso)
3324 tcp->tcp_maxpsz_multiplier = 0;
3325
3326 tcp->tcp_lso = B_TRUE;
3327 TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3328 } else { /* LSO capability is not usable any more. */
3329 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3330 boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
3331
3332 /*
3333 * If LSO to be disabled, notify the STREAM header with smaller
3334 * data block. And need to restore fragsize to PMTU.
3335 */
3336 if (tcp->tcp_lso) {
3337 tcp->tcp_maxpsz_multiplier =
3338 tcp->tcp_tcps->tcps_maxpsz_multiplier;
3339 ixa->ixa_fragsize = ixa->ixa_pmtu;
3340 tcp->tcp_lso = B_FALSE;
3341 TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
3342 }
3343 }
3344
3345 (void) tcp_maxpsz_set(tcp, B_TRUE);
3346 }
3347
3348 /*
3349 * Update the TCP connection according to change of ZEROCOPY capability.
3350 */
3351 static void
3352 tcp_update_zcopy(tcp_t *tcp)
3353 {
3354 conn_t *connp = tcp->tcp_connp;
3355 tcp_stack_t *tcps = tcp->tcp_tcps;
3356
3357 if (tcp->tcp_snd_zcopy_on) {
3358 tcp->tcp_snd_zcopy_on = B_FALSE;
3359 if (!TCP_IS_DETACHED(tcp)) {
3360 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3361 ZCVMUNSAFE);
3362 TCP_STAT(tcps, tcp_zcopy_off);
3363 }
3364 } else {
3365 tcp->tcp_snd_zcopy_on = B_TRUE;
3366 if (!TCP_IS_DETACHED(tcp)) {
3367 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3368 ZCVMSAFE);
3369 TCP_STAT(tcps, tcp_zcopy_on);
3370 }
3371 }
3372 }
3373
3374 /*
3375 * Notify function registered with ip_xmit_attr_t. It's called in the squeue
3376 * so it's safe to update the TCP connection.
3377 */
3378 /* ARGSUSED1 */
3379 static void
3380 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
3381 ixa_notify_arg_t narg)
3382 {
3383 tcp_t *tcp = (tcp_t *)arg;
3384 conn_t *connp = tcp->tcp_connp;
3385
3386 switch (ntype) {
3387 case IXAN_LSO:
3388 tcp_update_lso(tcp, connp->conn_ixa);
3389 break;
3390 case IXAN_PMTU:
3391 tcp_update_pmtu(tcp, B_FALSE);
3392 break;
3393 case IXAN_ZCOPY:
3394 tcp_update_zcopy(tcp);
3395 break;
3396 default:
3397 break;
3398 }
3399 }
3400
3401 /*
3402 * The TCP write service routine should never be called...
3403 */
3404 /* ARGSUSED */
3405 static void
3406 tcp_wsrv(queue_t *q)
3407 {
3408 tcp_stack_t *tcps = Q_TO_TCP(q)->tcp_tcps;
3409
3410 TCP_STAT(tcps, tcp_wsrv_called);
3411 }
3412
3413 /*
3414 * Hash list lookup routine for tcp_t structures.
3415 * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
3416 */
3417 tcp_t *
3418 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
3419 {
3420 tf_t *tf;
3421 tcp_t *tcp;
3422
3423 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3424 mutex_enter(&tf->tf_lock);
3425 for (tcp = tf->tf_tcp; tcp != NULL;
3426 tcp = tcp->tcp_acceptor_hash) {
3427 if (tcp->tcp_acceptor_id == id) {
3428 CONN_INC_REF(tcp->tcp_connp);
3429 mutex_exit(&tf->tf_lock);
3430 return (tcp);
3431 }
3432 }
3433 mutex_exit(&tf->tf_lock);
3434 return (NULL);
3435 }
3436
3437 /*
3438 * Hash list insertion routine for tcp_t structures.
3439 */
3440 void
3441 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
3442 {
3443 tf_t *tf;
3444 tcp_t **tcpp;
3445 tcp_t *tcpnext;
3446 tcp_stack_t *tcps = tcp->tcp_tcps;
3447
3448 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3449
3450 if (tcp->tcp_ptpahn != NULL)
3451 tcp_acceptor_hash_remove(tcp);
3452 tcpp = &tf->tf_tcp;
3453 mutex_enter(&tf->tf_lock);
3454 tcpnext = tcpp[0];
3455 if (tcpnext)
3456 tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
3457 tcp->tcp_acceptor_hash = tcpnext;
3458 tcp->tcp_ptpahn = tcpp;
3459 tcpp[0] = tcp;
3460 tcp->tcp_acceptor_lockp = &tf->tf_lock; /* For tcp_*_hash_remove */
3461 mutex_exit(&tf->tf_lock);
3462 }
3463
3464 /*
3465 * Hash list removal routine for tcp_t structures.
3466 */
3467 void
3468 tcp_acceptor_hash_remove(tcp_t *tcp)
3469 {
3470 tcp_t *tcpnext;
3471 kmutex_t *lockp;
3472
3473 /*
3474 * Extract the lock pointer in case there are concurrent
3475 * hash_remove's for this instance.
3476 */
3477 lockp = tcp->tcp_acceptor_lockp;
3478
3479 if (tcp->tcp_ptpahn == NULL)
3480 return;
3481
3482 ASSERT(lockp != NULL);
3483 mutex_enter(lockp);
3484 if (tcp->tcp_ptpahn) {
3485 tcpnext = tcp->tcp_acceptor_hash;
3486 if (tcpnext) {
3487 tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
3488 tcp->tcp_acceptor_hash = NULL;
3489 }
3490 *tcp->tcp_ptpahn = tcpnext;
3491 tcp->tcp_ptpahn = NULL;
3492 }
3493 mutex_exit(lockp);
3494 tcp->tcp_acceptor_lockp = NULL;
3495 }
3496
3497 /*
3498 * Type three generator adapted from the random() function in 4.4 BSD:
3499 */
3500
3501 /*
3502 * Copyright (c) 1983, 1993
3503 * The Regents of the University of California. All rights reserved.
3504 *
3505 * Redistribution and use in source and binary forms, with or without
3506 * modification, are permitted provided that the following conditions
3507 * are met:
3508 * 1. Redistributions of source code must retain the above copyright
3509 * notice, this list of conditions and the following disclaimer.
3510 * 2. Redistributions in binary form must reproduce the above copyright
3511 * notice, this list of conditions and the following disclaimer in the
3512 * documentation and/or other materials provided with the distribution.
3513 * 3. All advertising materials mentioning features or use of this software
3514 * must display the following acknowledgement:
3515 * This product includes software developed by the University of
3516 * California, Berkeley and its contributors.
3517 * 4. Neither the name of the University nor the names of its contributors
3518 * may be used to endorse or promote products derived from this software
3519 * without specific prior written permission.
3520 *
3521 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
3522 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
3523 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
3524 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
3525 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
3526 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
3527 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
3528 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
3529 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
3530 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
3531 * SUCH DAMAGE.
3532 */
3533
3534 /* Type 3 -- x**31 + x**3 + 1 */
3535 #define DEG_3 31
3536 #define SEP_3 3
3537
3538
3539 /* Protected by tcp_random_lock */
3540 static int tcp_randtbl[DEG_3 + 1];
3541
3542 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
3543 static int *tcp_random_rptr = &tcp_randtbl[1];
3544
3545 static int *tcp_random_state = &tcp_randtbl[1];
3546 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
3547
3548 kmutex_t tcp_random_lock;
3549
3550 void
3551 tcp_random_init(void)
3552 {
3553 int i;
3554 hrtime_t hrt;
3555 time_t wallclock;
3556 uint64_t result;
3557
3558 /*
3559 * Use high-res timer and current time for seed. Gethrtime() returns
3560 * a longlong, which may contain resolution down to nanoseconds.
3561 * The current time will either be a 32-bit or a 64-bit quantity.
3562 * XOR the two together in a 64-bit result variable.
3563 * Convert the result to a 32-bit value by multiplying the high-order
3564 * 32-bits by the low-order 32-bits.
3565 */
3566
3567 hrt = gethrtime();
3568 (void) drv_getparm(TIME, &wallclock);
3569 result = (uint64_t)wallclock ^ (uint64_t)hrt;
3570 mutex_enter(&tcp_random_lock);
3571 tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
3572 (result & 0xffffffff);
3573
3574 for (i = 1; i < DEG_3; i++)
3575 tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
3576 + 12345;
3577 tcp_random_fptr = &tcp_random_state[SEP_3];
3578 tcp_random_rptr = &tcp_random_state[0];
3579 mutex_exit(&tcp_random_lock);
3580 for (i = 0; i < 10 * DEG_3; i++)
3581 (void) tcp_random();
3582 }
3583
3584 /*
3585 * tcp_random: Return a random number in the range [1 - (128K + 1)].
3586 * This range is selected to be approximately centered on TCP_ISS / 2,
3587 * and easy to compute. We get this value by generating a 32-bit random
3588 * number, selecting out the high-order 17 bits, and then adding one so
3589 * that we never return zero.
3590 */
3591 int
3592 tcp_random(void)
3593 {
3594 int i;
3595
3596 mutex_enter(&tcp_random_lock);
3597 *tcp_random_fptr += *tcp_random_rptr;
3598
3599 /*
3600 * The high-order bits are more random than the low-order bits,
3601 * so we select out the high-order 17 bits and add one so that
3602 * we never return zero.
3603 */
3604 i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
3605 if (++tcp_random_fptr >= tcp_random_end_ptr) {
3606 tcp_random_fptr = tcp_random_state;
3607 ++tcp_random_rptr;
3608 } else if (++tcp_random_rptr >= tcp_random_end_ptr)
3609 tcp_random_rptr = tcp_random_state;
3610
3611 mutex_exit(&tcp_random_lock);
3612 return (i);
3613 }
3614
3615 /*
3616 * Split this function out so that if the secret changes, I'm okay.
3617 *
3618 * Initialize the tcp_iss_cookie and tcp_iss_key.
3619 */
3620
3621 #define PASSWD_SIZE 16 /* MUST be multiple of 4 */
3622
3623 void
3624 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
3625 {
3626 struct {
3627 int32_t current_time;
3628 uint32_t randnum;
3629 uint16_t pad;
3630 uint8_t ether[6];
3631 uint8_t passwd[PASSWD_SIZE];
3632 } tcp_iss_cookie;
3633 time_t t;
3634
3635 /*
3636 * Start with the current absolute time.
3637 */
3638 (void) drv_getparm(TIME, &t);
3639 tcp_iss_cookie.current_time = t;
3640
3641 /*
3642 * XXX - Need a more random number per RFC 1750, not this crap.
3643 * OTOH, if what follows is pretty random, then I'm in better shape.
3644 */
3645 tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
3646 tcp_iss_cookie.pad = 0x365c; /* Picked from HMAC pad values. */
3647
3648 /*
3649 * The cpu_type_info is pretty non-random. Ugggh. It does serve
3650 * as a good template.
3651 */
3652 bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
3653 min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
3654
3655 /*
3656 * The pass-phrase. Normally this is supplied by user-called NDD.
3657 */
3658 bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
3659
3660 /*
3661 * See 4010593 if this section becomes a problem again,
3662 * but the local ethernet address is useful here.
3663 */
3664 (void) localetheraddr(NULL,
3665 (struct ether_addr *)&tcp_iss_cookie.ether);
3666
3667 /*
3668 * Hash 'em all together. The MD5Final is called per-connection.
3669 */
3670 mutex_enter(&tcps->tcps_iss_key_lock);
3671 MD5Init(&tcps->tcps_iss_key);
3672 MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
3673 sizeof (tcp_iss_cookie));
3674 mutex_exit(&tcps->tcps_iss_key_lock);
3675 }
3676
3677 /*
3678 * Called by IP when IP is loaded into the kernel
3679 */
3680 void
3681 tcp_ddi_g_init(void)
3682 {
3683 tcp_timercache = kmem_cache_create("tcp_timercache",
3684 sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
3685 NULL, NULL, NULL, NULL, NULL, 0);
3686
3687 tcp_notsack_blk_cache = kmem_cache_create("tcp_notsack_blk_cache",
3688 sizeof (notsack_blk_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3689
3690 mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
3691
3692 /* Initialize the random number generator */
3693 tcp_random_init();
3694
3695 /* A single callback independently of how many netstacks we have */
3696 ip_squeue_init(tcp_squeue_add);
3697
3698 tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
3699
3700 tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
3701
3702 /*
3703 * We want to be informed each time a stack is created or
3704 * destroyed in the kernel, so we can maintain the
3705 * set of tcp_stack_t's.
3706 */
3707 netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
3708 }
3709
3710
3711 #define INET_NAME "ip"
3712
3713 /*
3714 * Initialize the TCP stack instance.
3715 */
3716 static void *
3717 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
3718 {
3719 tcp_stack_t *tcps;
3720 int i;
3721 int error = 0;
3722 major_t major;
3723 size_t arrsz;
3724
3725 tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
3726 tcps->tcps_netstack = ns;
3727
3728 /* Initialize locks */
3729 mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
3730 mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
3731
3732 tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
3733 tcps->tcps_g_epriv_ports[0] = ULP_DEF_EPRIV_PORT1;
3734 tcps->tcps_g_epriv_ports[1] = ULP_DEF_EPRIV_PORT2;
3735 tcps->tcps_min_anonpriv_port = 512;
3736
3737 tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
3738 TCP_BIND_FANOUT_SIZE, KM_SLEEP);
3739 tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
3740 TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
3741
3742 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3743 mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
3744 MUTEX_DEFAULT, NULL);
3745 }
3746
3747 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3748 mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
3749 MUTEX_DEFAULT, NULL);
3750 }
3751
3752 /* TCP's IPsec code calls the packet dropper. */
3753 ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
3754
3755 arrsz = tcp_propinfo_count * sizeof (mod_prop_info_t);
3756 tcps->tcps_propinfo_tbl = (mod_prop_info_t *)kmem_alloc(arrsz,
3757 KM_SLEEP);
3758 bcopy(tcp_propinfo_tbl, tcps->tcps_propinfo_tbl, arrsz);
3759
3760 /*
3761 * Note: To really walk the device tree you need the devinfo
3762 * pointer to your device which is only available after probe/attach.
3763 * The following is safe only because it uses ddi_root_node()
3764 */
3765 tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
3766 tcp_opt_obj.odb_opt_arr_cnt);
3767
3768 /*
3769 * Initialize RFC 1948 secret values. This will probably be reset once
3770 * by the boot scripts.
3771 *
3772 * Use NULL name, as the name is caught by the new lockstats.
3773 *
3774 * Initialize with some random, non-guessable string, like the global
3775 * T_INFO_ACK.
3776 */
3777
3778 tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
3779 sizeof (tcp_g_t_info_ack), tcps);
3780
3781 tcps->tcps_kstat = tcp_kstat2_init(stackid);
3782 tcps->tcps_mibkp = tcp_kstat_init(stackid);
3783
3784 major = mod_name_to_major(INET_NAME);
3785 error = ldi_ident_from_major(major, &tcps->tcps_ldi_ident);
3786 ASSERT(error == 0);
3787 tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
3788 ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
3789 cv_init(&tcps->tcps_ixa_cleanup_cv, NULL, CV_DEFAULT, NULL);
3790 mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
3791
3792 mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
3793 tcps->tcps_reclaim = B_FALSE;
3794 tcps->tcps_reclaim_tid = 0;
3795 tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
3796
3797 /*
3798 * ncpus is the current number of CPUs, which can be bigger than
3799 * boot_ncpus. But we don't want to use ncpus to allocate all the
3800 * tcp_stats_cpu_t at system boot up time since it will be 1. While
3801 * we handle adding CPU in tcp_cpu_update(), it will be slow if
3802 * there are many CPUs as we will be adding them 1 by 1.
3803 *
3804 * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
3805 * are not freed until the stack is going away. So there is no need
3806 * to grab a lock to access the per CPU tcps_sc[x] pointer.
3807 */
3808 mutex_enter(&cpu_lock);
3809 tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
3810 mutex_exit(&cpu_lock);
3811 tcps->tcps_sc = kmem_zalloc(max_ncpus * sizeof (tcp_stats_cpu_t *),
3812 KM_SLEEP);
3813 for (i = 0; i < tcps->tcps_sc_cnt; i++) {
3814 tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
3815 KM_SLEEP);
3816 }
3817
3818 mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
3819 list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
3820 offsetof(tcp_listener_t, tl_link));
3821
3822 return (tcps);
3823 }
3824
3825 /*
3826 * Called when the IP module is about to be unloaded.
3827 */
3828 void
3829 tcp_ddi_g_destroy(void)
3830 {
3831 tcp_g_kstat_fini(tcp_g_kstat);
3832 tcp_g_kstat = NULL;
3833 bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
3834
3835 mutex_destroy(&tcp_random_lock);
3836
3837 kmem_cache_destroy(tcp_timercache);
3838 kmem_cache_destroy(tcp_notsack_blk_cache);
3839
3840 netstack_unregister(NS_TCP);
3841 }
3842
3843 /*
3844 * Free the TCP stack instance.
3845 */
3846 static void
3847 tcp_stack_fini(netstackid_t stackid, void *arg)
3848 {
3849 tcp_stack_t *tcps = (tcp_stack_t *)arg;
3850 int i;
3851
3852 freeb(tcps->tcps_ixa_cleanup_mp);
3853 tcps->tcps_ixa_cleanup_mp = NULL;
3854 cv_destroy(&tcps->tcps_ixa_cleanup_cv);
3855 mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
3856
3857 /*
3858 * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
3859 * the timer.
3860 */
3861 mutex_enter(&tcps->tcps_reclaim_lock);
3862 tcps->tcps_reclaim = B_FALSE;
3863 mutex_exit(&tcps->tcps_reclaim_lock);
3864 if (tcps->tcps_reclaim_tid != 0)
3865 (void) untimeout(tcps->tcps_reclaim_tid);
3866 mutex_destroy(&tcps->tcps_reclaim_lock);
3867
3868 tcp_listener_conf_cleanup(tcps);
3869
3870 for (i = 0; i < tcps->tcps_sc_cnt; i++)
3871 kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
3872 kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
3873
3874 kmem_free(tcps->tcps_propinfo_tbl,
3875 tcp_propinfo_count * sizeof (mod_prop_info_t));
3876 tcps->tcps_propinfo_tbl = NULL;
3877
3878 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3879 ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
3880 mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
3881 }
3882
3883 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3884 ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
3885 mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
3886 }
3887
3888 kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
3889 tcps->tcps_bind_fanout = NULL;
3890
3891 kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
3892 TCP_ACCEPTOR_FANOUT_SIZE);
3893 tcps->tcps_acceptor_fanout = NULL;
3894
3895 mutex_destroy(&tcps->tcps_iss_key_lock);
3896 mutex_destroy(&tcps->tcps_epriv_port_lock);
3897
3898 ip_drop_unregister(&tcps->tcps_dropper);
3899
3900 tcp_kstat2_fini(stackid, tcps->tcps_kstat);
3901 tcps->tcps_kstat = NULL;
3902
3903 tcp_kstat_fini(stackid, tcps->tcps_mibkp);
3904 tcps->tcps_mibkp = NULL;
3905
3906 ldi_ident_release(tcps->tcps_ldi_ident);
3907 kmem_free(tcps, sizeof (*tcps));
3908 }
3909
3910 /*
3911 * Generate ISS, taking into account NDD changes may happen halfway through.
3912 * (If the iss is not zero, set it.)
3913 */
3914
3915 static void
3916 tcp_iss_init(tcp_t *tcp)
3917 {
3918 MD5_CTX context;
3919 struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
3920 uint32_t answer[4];
3921 tcp_stack_t *tcps = tcp->tcp_tcps;
3922 conn_t *connp = tcp->tcp_connp;
3923
3924 tcps->tcps_iss_incr_extra += (ISS_INCR >> 1);
3925 tcp->tcp_iss = tcps->tcps_iss_incr_extra;
3926 switch (tcps->tcps_strong_iss) {
3927 case 2:
3928 mutex_enter(&tcps->tcps_iss_key_lock);
3929 context = tcps->tcps_iss_key;
3930 mutex_exit(&tcps->tcps_iss_key_lock);
3931 arg.ports = connp->conn_ports;
3932 arg.src = connp->conn_laddr_v6;
3933 arg.dst = connp->conn_faddr_v6;
3934 MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
3935 MD5Final((uchar_t *)answer, &context);
3936 tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
3937 /*
3938 * Now that we've hashed into a unique per-connection sequence
3939 * space, add a random increment per strong_iss == 1. So I
3940 * guess we'll have to...
3941 */
3942 /* FALLTHRU */
3943 case 1:
3944 tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
3945 break;
3946 default:
3947 tcp->tcp_iss += (uint32_t)gethrestime_sec() * ISS_INCR;
3948 break;
3949 }
3950 tcp->tcp_valid_bits = TCP_ISS_VALID;
3951 tcp->tcp_fss = tcp->tcp_iss - 1;
3952 tcp->tcp_suna = tcp->tcp_iss;
3953 tcp->tcp_snxt = tcp->tcp_iss + 1;
3954 tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3955 tcp->tcp_csuna = tcp->tcp_snxt;
3956 }
3957
3958 /*
3959 * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
3960 * on the specified backing STREAMS q. Note, the caller may make the
3961 * decision to call based on the tcp_t.tcp_flow_stopped value which
3962 * when check outside the q's lock is only an advisory check ...
3963 */
3964 void
3965 tcp_setqfull(tcp_t *tcp)
3966 {
3967 tcp_stack_t *tcps = tcp->tcp_tcps;
3968 conn_t *connp = tcp->tcp_connp;
3969
3970 if (tcp->tcp_closed)
3971 return;
3972
3973 conn_setqfull(connp, &tcp->tcp_flow_stopped);
3974 if (tcp->tcp_flow_stopped)
3975 TCP_STAT(tcps, tcp_flwctl_on);
3976 }
3977
3978 void
3979 tcp_clrqfull(tcp_t *tcp)
3980 {
3981 conn_t *connp = tcp->tcp_connp;
3982
3983 if (tcp->tcp_closed)
3984 return;
3985 conn_clrqfull(connp, &tcp->tcp_flow_stopped);
3986 }
3987
3988 static int
3989 tcp_squeue_switch(int val)
3990 {
3991 int rval = SQ_FILL;
3992
3993 switch (val) {
3994 case 1:
3995 rval = SQ_NODRAIN;
3996 break;
3997 case 2:
3998 rval = SQ_PROCESS;
3999 break;
4000 default:
4001 break;
4002 }
4003 return (rval);
4004 }
4005
4006 /*
4007 * This is called once for each squeue - globally for all stack
4008 * instances.
4009 */
4010 static void
4011 tcp_squeue_add(squeue_t *sqp)
4012 {
4013 tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4014 sizeof (tcp_squeue_priv_t), KM_SLEEP);
4015
4016 *squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4017 if (tcp_free_list_max_cnt == 0) {
4018 int tcp_ncpus = ((boot_max_ncpus == -1) ?
4019 max_ncpus : boot_max_ncpus);
4020
4021 /*
4022 * Limit number of entries to 1% of availble memory / tcp_ncpus
4023 */
4024 tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4025 (tcp_ncpus * sizeof (tcp_t) * 100);
4026 }
4027 tcp_time_wait->tcp_free_list_cnt = 0;
4028 }
4029 /*
4030 * Return unix error is tli error is TSYSERR, otherwise return a negative
4031 * tli error.
4032 */
4033 int
4034 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4035 boolean_t bind_to_req_port_only)
4036 {
4037 int error;
4038 tcp_t *tcp = connp->conn_tcp;
4039
4040 if (tcp->tcp_state >= TCPS_BOUND) {
4041 if (connp->conn_debug) {
4042 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4043 "tcp_bind: bad state, %d", tcp->tcp_state);
4044 }
4045 return (-TOUTSTATE);
4046 }
4047
4048 error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4049 if (error != 0)
4050 return (error);
4051
4052 ASSERT(tcp->tcp_state == TCPS_BOUND);
4053 tcp->tcp_conn_req_max = 0;
4054 return (0);
4055 }
4056
4057 /*
4058 * If the return value from this function is positive, it's a UNIX error.
4059 * Otherwise, if it's negative, then the absolute value is a TLI error.
4060 * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4061 */
4062 int
4063 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4064 cred_t *cr, pid_t pid)
4065 {
4066 tcp_t *tcp = connp->conn_tcp;
4067 sin_t *sin = (sin_t *)sa;
4068 sin6_t *sin6 = (sin6_t *)sa;
4069 ipaddr_t *dstaddrp;
4070 in_port_t dstport;
4071 uint_t srcid;
4072 int error;
4073 uint32_t mss;
4074 mblk_t *syn_mp;
4075 tcp_stack_t *tcps = tcp->tcp_tcps;
4076 int32_t oldstate;
4077 ip_xmit_attr_t *ixa = connp->conn_ixa;
4078
4079 oldstate = tcp->tcp_state;
4080
4081 switch (len) {
4082 default:
4083 /*
4084 * Should never happen
4085 */
4086 return (EINVAL);
4087
4088 case sizeof (sin_t):
4089 sin = (sin_t *)sa;
4090 if (sin->sin_port == 0) {
4091 return (-TBADADDR);
4092 }
4093 if (connp->conn_ipv6_v6only) {
4094 return (EAFNOSUPPORT);
4095 }
4096 break;
4097
4098 case sizeof (sin6_t):
4099 sin6 = (sin6_t *)sa;
4100 if (sin6->sin6_port == 0) {
4101 return (-TBADADDR);
4102 }
4103 break;
4104 }
4105 /*
4106 * If we're connecting to an IPv4-mapped IPv6 address, we need to
4107 * make sure that the conn_ipversion is IPV4_VERSION. We
4108 * need to this before we call tcp_bindi() so that the port lookup
4109 * code will look for ports in the correct port space (IPv4 and
4110 * IPv6 have separate port spaces).
4111 */
4112 if (connp->conn_family == AF_INET6 &&
4113 connp->conn_ipversion == IPV6_VERSION &&
4114 IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4115 if (connp->conn_ipv6_v6only)
4116 return (EADDRNOTAVAIL);
4117
4118 connp->conn_ipversion = IPV4_VERSION;
4119 }
4120
4121 switch (tcp->tcp_state) {
4122 case TCPS_LISTEN:
4123 /*
4124 * Listening sockets are not allowed to issue connect().
4125 */
4126 if (IPCL_IS_NONSTR(connp))
4127 return (EOPNOTSUPP);
4128 /* FALLTHRU */
4129 case TCPS_IDLE:
4130 /*
4131 * We support quick connect, refer to comments in
4132 * tcp_connect_*()
4133 */
4134 /* FALLTHRU */
4135 case TCPS_BOUND:
4136 break;
4137 default:
4138 return (-TOUTSTATE);
4139 }
4140
4141 /*
4142 * We update our cred/cpid based on the caller of connect
4143 */
4144 if (connp->conn_cred != cr) {
4145 crhold(cr);
4146 crfree(connp->conn_cred);
4147 connp->conn_cred = cr;
4148 }
4149 connp->conn_cpid = pid;
4150
4151 /* Cache things in the ixa without any refhold */
4152 ASSERT(!(ixa->ixa_free_flags & IXA_FREE_CRED));
4153 ixa->ixa_cred = cr;
4154 ixa->ixa_cpid = pid;
4155 if (is_system_labeled()) {
4156 /* We need to restart with a label based on the cred */
4157 ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4158 }
4159
4160 if (connp->conn_family == AF_INET6) {
4161 if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4162 error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4163 sin6->sin6_port, sin6->sin6_flowinfo,
4164 sin6->__sin6_src_id, sin6->sin6_scope_id);
4165 } else {
4166 /*
4167 * Destination adress is mapped IPv6 address.
4168 * Source bound address should be unspecified or
4169 * IPv6 mapped address as well.
4170 */
4171 if (!IN6_IS_ADDR_UNSPECIFIED(
4172 &connp->conn_bound_addr_v6) &&
4173 !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4174 return (EADDRNOTAVAIL);
4175 }
4176 dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4177 dstport = sin6->sin6_port;
4178 srcid = sin6->__sin6_src_id;
4179 error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4180 srcid);
4181 }
4182 } else {
4183 dstaddrp = &sin->sin_addr.s_addr;
4184 dstport = sin->sin_port;
4185 srcid = 0;
4186 error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4187 }
4188
4189 if (error != 0)
4190 goto connect_failed;
4191
4192 CL_INET_CONNECT(connp, B_TRUE, error);
4193 if (error != 0)
4194 goto connect_failed;
4195
4196 /* connect succeeded */
4197 TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4198 tcp->tcp_active_open = 1;
4199
4200 /*
4201 * tcp_set_destination() does not adjust for TCP/IP header length.
4202 */
4203 mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4204
4205 /*
4206 * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4207 * to the nearest MSS.
4208 *
4209 * We do the round up here because we need to get the interface MTU
4210 * first before we can do the round up.
4211 */
4212 tcp->tcp_rwnd = connp->conn_rcvbuf;
4213 tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4214 tcps->tcps_recv_hiwat_minmss * mss);
4215 connp->conn_rcvbuf = tcp->tcp_rwnd;
4216 tcp_set_ws_value(tcp);
4217 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4218 if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4219 tcp->tcp_snd_ws_ok = B_TRUE;
4220
4221 /*
4222 * Set tcp_snd_ts_ok to true
4223 * so that tcp_xmit_mp will
4224 * include the timestamp
4225 * option in the SYN segment.
4226 */
4227 if (tcps->tcps_tstamp_always ||
4228 (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4229 tcp->tcp_snd_ts_ok = B_TRUE;
4230 }
4231
4232 /*
4233 * Note that tcp_snd_sack_ok can be set in tcp_set_destination() if
4234 * the SACK metric is set. So here we just check the per stack SACK
4235 * permitted param.
4236 */
4237 if (tcps->tcps_sack_permitted == 2) {
4238 ASSERT(tcp->tcp_num_sack_blk == 0);
4239 ASSERT(tcp->tcp_notsack_list == NULL);
4240 tcp->tcp_snd_sack_ok = B_TRUE;
4241 }
4242
4243 /*
4244 * Should we use ECN? Note that the current
4245 * default value (SunOS 5.9) of tcp_ecn_permitted
4246 * is 1. The reason for doing this is that there
4247 * are equipments out there that will drop ECN
4248 * enabled IP packets. Setting it to 1 avoids
4249 * compatibility problems.
4250 */
4251 if (tcps->tcps_ecn_permitted == 2)
4252 tcp->tcp_ecn_ok = B_TRUE;
4253
4254 /* Trace change from BOUND -> SYN_SENT here */
4255 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4256 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4257 int32_t, TCPS_BOUND);
4258
4259 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4260 syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4261 tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4262 if (syn_mp != NULL) {
4263 /*
4264 * We must bump the generation before sending the syn
4265 * to ensure that we use the right generation in case
4266 * this thread issues a "connected" up call.
4267 */
4268 SOCK_CONNID_BUMP(tcp->tcp_connid);
4269 /*
4270 * DTrace sending the first SYN as a
4271 * tcp:::connect-request event.
4272 */
4273 DTRACE_TCP5(connect__request, mblk_t *, NULL,
4274 ip_xmit_attr_t *, connp->conn_ixa,
4275 void_ip_t *, syn_mp->b_rptr, tcp_t *, tcp,
4276 tcph_t *,
4277 &syn_mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
4278 tcp_send_data(tcp, syn_mp);
4279 }
4280
4281 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4282 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4283 return (0);
4284
4285 connect_failed:
4286 connp->conn_faddr_v6 = ipv6_all_zeros;
4287 connp->conn_fport = 0;
4288 tcp->tcp_state = oldstate;
4289 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4290 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4291 return (error);
4292 }
4293
4294 int
4295 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4296 int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4297 {
4298 tcp_t *tcp = connp->conn_tcp;
4299 int error = 0;
4300 tcp_stack_t *tcps = tcp->tcp_tcps;
4301 int32_t oldstate;
4302
4303 /* All Solaris components should pass a cred for this operation. */
4304 ASSERT(cr != NULL);
4305
4306 if (tcp->tcp_state >= TCPS_BOUND) {
4307 if ((tcp->tcp_state == TCPS_BOUND ||
4308 tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
4309 /*
4310 * Handle listen() increasing backlog.
4311 * This is more "liberal" then what the TPI spec
4312 * requires but is needed to avoid a t_unbind
4313 * when handling listen() since the port number
4314 * might be "stolen" between the unbind and bind.
4315 */
4316 goto do_listen;
4317 }
4318 if (connp->conn_debug) {
4319 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4320 "tcp_listen: bad state, %d", tcp->tcp_state);
4321 }
4322 return (-TOUTSTATE);
4323 } else {
4324 if (sa == NULL) {
4325 sin6_t addr;
4326 sin_t *sin;
4327 sin6_t *sin6;
4328
4329 ASSERT(IPCL_IS_NONSTR(connp));
4330 /* Do an implicit bind: Request for a generic port. */
4331 if (connp->conn_family == AF_INET) {
4332 len = sizeof (sin_t);
4333 sin = (sin_t *)&addr;
4334 *sin = sin_null;
4335 sin->sin_family = AF_INET;
4336 } else {
4337 ASSERT(connp->conn_family == AF_INET6);
4338 len = sizeof (sin6_t);
4339 sin6 = (sin6_t *)&addr;
4340 *sin6 = sin6_null;
4341 sin6->sin6_family = AF_INET6;
4342 }
4343 sa = (struct sockaddr *)&addr;
4344 }
4345
4346 error = tcp_bind_check(connp, sa, len, cr,
4347 bind_to_req_port_only);
4348 if (error)
4349 return (error);
4350 /* Fall through and do the fanout insertion */
4351 }
4352
4353 do_listen:
4354 ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
4355 tcp->tcp_conn_req_max = backlog;
4356 if (tcp->tcp_conn_req_max) {
4357 if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
4358 tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
4359 if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
4360 tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
4361 /*
4362 * If this is a listener, do not reset the eager list
4363 * and other stuffs. Note that we don't check if the
4364 * existing eager list meets the new tcp_conn_req_max
4365 * requirement.
4366 */
4367 if (tcp->tcp_state != TCPS_LISTEN) {
4368 tcp->tcp_state = TCPS_LISTEN;
4369 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4370 connp->conn_ixa, void, NULL, tcp_t *, tcp,
4371 void, NULL, int32_t, TCPS_BOUND);
4372 /* Initialize the chain. Don't need the eager_lock */
4373 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
4374 tcp->tcp_eager_next_drop_q0 = tcp;
4375 tcp->tcp_eager_prev_drop_q0 = tcp;
4376 tcp->tcp_second_ctimer_threshold =
4377 tcps->tcps_ip_abort_linterval;
4378 }
4379 }
4380
4381 /*
4382 * We need to make sure that the conn_recv is set to a non-null
4383 * value before we insert the conn into the classifier table.
4384 * This is to avoid a race with an incoming packet which does an
4385 * ipcl_classify().
4386 * We initially set it to tcp_input_listener_unbound to try to
4387 * pick a good squeue for the listener when the first SYN arrives.
4388 * tcp_input_listener_unbound sets it to tcp_input_listener on that
4389 * first SYN.
4390 */
4391 connp->conn_recv = tcp_input_listener_unbound;
4392
4393 /* Insert the listener in the classifier table */
4394 error = ip_laddr_fanout_insert(connp);
4395 if (error != 0) {
4396 /* Undo the bind - release the port number */
4397 oldstate = tcp->tcp_state;
4398 tcp->tcp_state = TCPS_IDLE;
4399 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4400 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4401 int32_t, oldstate);
4402 connp->conn_bound_addr_v6 = ipv6_all_zeros;
4403
4404 connp->conn_laddr_v6 = ipv6_all_zeros;
4405 connp->conn_saddr_v6 = ipv6_all_zeros;
4406 connp->conn_ports = 0;
4407
4408 if (connp->conn_anon_port) {
4409 zone_t *zone;
4410
4411 zone = crgetzone(cr);
4412 connp->conn_anon_port = B_FALSE;
4413 (void) tsol_mlp_anon(zone, connp->conn_mlp_type,
4414 connp->conn_proto, connp->conn_lport, B_FALSE);
4415 }
4416 connp->conn_mlp_type = mlptSingle;
4417
4418 tcp_bind_hash_remove(tcp);
4419 return (error);
4420 } else {
4421 /*
4422 * If there is a connection limit, allocate and initialize
4423 * the counter struct. Note that since listen can be called
4424 * multiple times, the struct may have been allready allocated.
4425 */
4426 if (!list_is_empty(&tcps->tcps_listener_conf) &&
4427 tcp->tcp_listen_cnt == NULL) {
4428 tcp_listen_cnt_t *tlc;
4429 uint32_t ratio;
4430
4431 ratio = tcp_find_listener_conf(tcps,
4432 ntohs(connp->conn_lport));
4433 if (ratio != 0) {
4434 uint32_t mem_ratio, tot_buf;
4435
4436 tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
4437 KM_SLEEP);
4438 /*
4439 * Calculate the connection limit based on
4440 * the configured ratio and maxusers. Maxusers
4441 * are calculated based on memory size,
4442 * ~ 1 user per MB. Note that the conn_rcvbuf
4443 * and conn_sndbuf may change after a
4444 * connection is accepted. So what we have
4445 * is only an approximation.
4446 */
4447 if ((tot_buf = connp->conn_rcvbuf +
4448 connp->conn_sndbuf) < MB) {
4449 mem_ratio = MB / tot_buf;
4450 tlc->tlc_max = maxusers / ratio *
4451 mem_ratio;
4452 } else {
4453 mem_ratio = tot_buf / MB;
4454 tlc->tlc_max = maxusers / ratio /
4455 mem_ratio;
4456 }
4457 /* At least we should allow two connections! */
4458 if (tlc->tlc_max <= tcp_min_conn_listener)
4459 tlc->tlc_max = tcp_min_conn_listener;
4460 tlc->tlc_cnt = 1;
4461 tlc->tlc_drop = 0;
4462 tcp->tcp_listen_cnt = tlc;
4463 }
4464 }
4465 }
4466 return (error);
4467 }