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 DONTCARE(tcp->tcp_ka_rinterval); /* Init in tcp_init_values */
2232 DONTCARE(tcp->tcp_ka_abort_thres); /* Init in tcp_init_values */
2233 DONTCARE(tcp->tcp_ka_cnt); /* Init in tcp_init_values */
2234
2235 tcp->tcp_client_errno = 0;
2236
2237 DONTCARE(connp->conn_sum); /* Init in tcp_init_values */
2238
2239 connp->conn_faddr_v6 = ipv6_all_zeros; /* Displayed in MIB */
2240
2241 PRESERVE(connp->conn_bound_addr_v6);
2242 tcp->tcp_last_sent_len = 0;
2243 tcp->tcp_dupack_cnt = 0;
2244
2245 connp->conn_fport = 0; /* Displayed in MIB */
2246 PRESERVE(connp->conn_lport);
2247
2248 PRESERVE(tcp->tcp_acceptor_lockp);
2249
2250 ASSERT(tcp->tcp_ordrel_mp == NULL);
2251 PRESERVE(tcp->tcp_acceptor_id);
2252 DONTCARE(tcp->tcp_ipsec_overhead);
2253
2254 PRESERVE(connp->conn_family);
2255 /* Remove any remnants of mapped address binding */
2256 if (connp->conn_family == AF_INET6) {
2257 connp->conn_ipversion = IPV6_VERSION;
2258 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2259 } else {
2260 connp->conn_ipversion = IPV4_VERSION;
2261 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2262 }
2263
2264 connp->conn_bound_if = 0;
2265 connp->conn_recv_ancillary.crb_all = 0;
2266 tcp->tcp_recvifindex = 0;
2267 tcp->tcp_recvhops = 0;
2268 tcp->tcp_closed = 0;
2269 if (tcp->tcp_hopopts != NULL) {
2270 mi_free(tcp->tcp_hopopts);
2271 tcp->tcp_hopopts = NULL;
2272 tcp->tcp_hopoptslen = 0;
2273 }
2274 ASSERT(tcp->tcp_hopoptslen == 0);
2275 if (tcp->tcp_dstopts != NULL) {
2276 mi_free(tcp->tcp_dstopts);
2277 tcp->tcp_dstopts = NULL;
2278 tcp->tcp_dstoptslen = 0;
2279 }
2280 ASSERT(tcp->tcp_dstoptslen == 0);
2281 if (tcp->tcp_rthdrdstopts != NULL) {
2282 mi_free(tcp->tcp_rthdrdstopts);
2283 tcp->tcp_rthdrdstopts = NULL;
2284 tcp->tcp_rthdrdstoptslen = 0;
2285 }
2286 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2287 if (tcp->tcp_rthdr != NULL) {
2288 mi_free(tcp->tcp_rthdr);
2289 tcp->tcp_rthdr = NULL;
2290 tcp->tcp_rthdrlen = 0;
2291 }
2292 ASSERT(tcp->tcp_rthdrlen == 0);
2293
2294 /* Reset fusion-related fields */
2295 tcp->tcp_fused = B_FALSE;
2296 tcp->tcp_unfusable = B_FALSE;
2297 tcp->tcp_fused_sigurg = B_FALSE;
2298 tcp->tcp_loopback_peer = NULL;
2299
2300 tcp->tcp_lso = B_FALSE;
2301
2302 tcp->tcp_in_ack_unsent = 0;
2303 tcp->tcp_cork = B_FALSE;
2304 tcp->tcp_tconnind_started = B_FALSE;
2305
2306 PRESERVE(tcp->tcp_squeue_bytes);
2307
2308 tcp->tcp_closemp_used = B_FALSE;
2309
2310 PRESERVE(tcp->tcp_rsrv_mp);
2311 PRESERVE(tcp->tcp_rsrv_mp_lock);
2312
2313 #ifdef DEBUG
2314 DONTCARE(tcp->tcmp_stk[0]);
2315 #endif
2316
2317 PRESERVE(tcp->tcp_connid);
2318
2319 ASSERT(tcp->tcp_listen_cnt == NULL);
2320 ASSERT(tcp->tcp_reass_tid == 0);
2321
2322 #undef DONTCARE
2323 #undef PRESERVE
2324 }
2325
2326 /*
2327 * Initialize the various fields in tcp_t. If parent (the listener) is non
2328 * NULL, certain values will be inheritted from it.
2329 */
2330 void
2331 tcp_init_values(tcp_t *tcp, tcp_t *parent)
2332 {
2333 tcp_stack_t *tcps = tcp->tcp_tcps;
2334 conn_t *connp = tcp->tcp_connp;
2335 clock_t rto;
2336
2337 ASSERT((connp->conn_family == AF_INET &&
2338 connp->conn_ipversion == IPV4_VERSION) ||
2339 (connp->conn_family == AF_INET6 &&
2340 (connp->conn_ipversion == IPV4_VERSION ||
2341 connp->conn_ipversion == IPV6_VERSION)));
2342
2343 if (parent == NULL) {
2344 tcp->tcp_naglim = tcps->tcps_naglim_def;
2345
2346 tcp->tcp_rto_initial = tcps->tcps_rexmit_interval_initial;
2347 tcp->tcp_rto_min = tcps->tcps_rexmit_interval_min;
2348 tcp->tcp_rto_max = tcps->tcps_rexmit_interval_max;
2349
2350 tcp->tcp_first_ctimer_threshold =
2351 tcps->tcps_ip_notify_cinterval;
2352 tcp->tcp_second_ctimer_threshold =
2353 tcps->tcps_ip_abort_cinterval;
2354 tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2355 tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2356
2357 tcp->tcp_fin_wait_2_flush_interval =
2358 tcps->tcps_fin_wait_2_flush_interval;
2359
2360 tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2361 tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2362 tcp->tcp_ka_cnt = 0;
2363 tcp->tcp_ka_rinterval = 0;
2364
2365 /*
2366 * Default value of tcp_init_cwnd is 0, so no need to set here
2367 * if parent is NULL. But we need to inherit it from parent.
2368 */
2369 } else {
2370 /* Inherit various TCP parameters from the parent. */
2371 tcp->tcp_naglim = parent->tcp_naglim;
2372
2373 tcp->tcp_rto_initial = parent->tcp_rto_initial;
2374 tcp->tcp_rto_min = parent->tcp_rto_min;
2375 tcp->tcp_rto_max = parent->tcp_rto_max;
2376
2377 tcp->tcp_first_ctimer_threshold =
2378 parent->tcp_first_ctimer_threshold;
2379 tcp->tcp_second_ctimer_threshold =
2380 parent->tcp_second_ctimer_threshold;
2381 tcp->tcp_first_timer_threshold =
2382 parent->tcp_first_timer_threshold;
2383 tcp->tcp_second_timer_threshold =
2384 parent->tcp_second_timer_threshold;
2385
2386 tcp->tcp_fin_wait_2_flush_interval =
2387 parent->tcp_fin_wait_2_flush_interval;
2388
2389 tcp->tcp_ka_interval = parent->tcp_ka_interval;
2390 tcp->tcp_ka_abort_thres = parent->tcp_ka_abort_thres;
2391 tcp->tcp_ka_cnt = parent->tcp_ka_cnt;
2392 tcp->tcp_ka_rinterval = parent->tcp_ka_rinterval;
2393
2394 tcp->tcp_init_cwnd = parent->tcp_init_cwnd;
2395 }
2396
2397 /*
2398 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2399 * will be close to tcp_rexmit_interval_initial. By doing this, we
2400 * allow the algorithm to adjust slowly to large fluctuations of RTT
2401 * during first few transmissions of a connection as seen in slow
2402 * links.
2403 */
2404 tcp->tcp_rtt_sa = tcp->tcp_rto_initial << 2;
2405 tcp->tcp_rtt_sd = tcp->tcp_rto_initial >> 1;
2406 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
2407 tcps->tcps_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
2408 tcps->tcps_conn_grace_period;
2409 TCP_SET_RTO(tcp, rto);
2410
2411 tcp->tcp_timer_backoff = 0;
2412 tcp->tcp_ms_we_have_waited = 0;
2413 tcp->tcp_last_recv_time = ddi_get_lbolt();
2414 tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2415 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2416 tcp->tcp_snd_burst = TCP_CWND_INFINITE;
2417
2418 tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2419
2420 /* NOTE: ISS is now set in tcp_set_destination(). */
2421
2422 /* Reset fusion-related fields */
2423 tcp->tcp_fused = B_FALSE;
2424 tcp->tcp_unfusable = B_FALSE;
2425 tcp->tcp_fused_sigurg = B_FALSE;
2426 tcp->tcp_loopback_peer = NULL;
2427
2428 /* We rebuild the header template on the next connect/conn_request */
2429
2430 connp->conn_mlp_type = mlptSingle;
2431
2432 /*
2433 * Init the window scale to the max so tcp_rwnd_set() won't pare
2434 * down tcp_rwnd. tcp_set_destination() will set the right value later.
2435 */
2436 tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2437 tcp->tcp_rwnd = connp->conn_rcvbuf;
2438
2439 tcp->tcp_cork = B_FALSE;
2440 /*
2441 * Init the tcp_debug option if it wasn't already set. This value
2442 * determines whether TCP
2443 * calls strlog() to print out debug messages. Doing this
2444 * initialization here means that this value is not inherited thru
2445 * tcp_reinit().
2446 */
2447 if (!connp->conn_debug)
2448 connp->conn_debug = tcps->tcps_dbg;
2449 }
2450
2451 /*
2452 * Update the TCP connection according to change of PMTU.
2453 *
2454 * Path MTU might have changed by either increase or decrease, so need to
2455 * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2456 * or negative MSS, since tcp_mss_set() will do it.
2457 */
2458 void
2459 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2460 {
2461 uint32_t pmtu;
2462 int32_t mss;
2463 conn_t *connp = tcp->tcp_connp;
2464 ip_xmit_attr_t *ixa = connp->conn_ixa;
2465 iaflags_t ixaflags;
2466
2467 if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2468 return;
2469
2470 if (tcp->tcp_state < TCPS_ESTABLISHED)
2471 return;
2472
2473 /*
2474 * Always call ip_get_pmtu() to make sure that IP has updated
2475 * ixa_flags properly.
2476 */
2477 pmtu = ip_get_pmtu(ixa);
2478 ixaflags = ixa->ixa_flags;
2479
2480 /*
2481 * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2482 * IPsec overhead if applied. Make sure to use the most recent
2483 * IPsec information.
2484 */
2485 mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2486
2487 /*
2488 * Nothing to change, so just return.
2489 */
2490 if (mss == tcp->tcp_mss)
2491 return;
2492
2493 /*
2494 * Currently, for ICMP errors, only PMTU decrease is handled.
2495 */
2496 if (mss > tcp->tcp_mss && decrease_only)
2497 return;
2498
2499 DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2500
2501 /*
2502 * Update ixa_fragsize and ixa_pmtu.
2503 */
2504 ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2505
2506 /*
2507 * Adjust MSS and all relevant variables.
2508 */
2509 tcp_mss_set(tcp, mss);
2510
2511 /*
2512 * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2513 * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2514 * has a (potentially different) min size we do the same. Make sure to
2515 * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2516 * fragment the packet.
2517 *
2518 * LSO over IPv6 can not be fragmented. So need to disable LSO
2519 * when IPv6 fragmentation is needed.
2520 */
2521 if (mss < tcp->tcp_tcps->tcps_mss_min)
2522 ixaflags |= IXAF_PMTU_TOO_SMALL;
2523
2524 if (ixaflags & IXAF_PMTU_TOO_SMALL)
2525 ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2526
2527 if ((connp->conn_ipversion == IPV4_VERSION) &&
2528 !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2529 tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2530 }
2531 ixa->ixa_flags = ixaflags;
2532 }
2533
2534 int
2535 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2536 {
2537 conn_t *connp = tcp->tcp_connp;
2538 queue_t *q = connp->conn_rq;
2539 int32_t mss = tcp->tcp_mss;
2540 int maxpsz;
2541
2542 if (TCP_IS_DETACHED(tcp))
2543 return (mss);
2544 if (tcp->tcp_fused) {
2545 maxpsz = tcp_fuse_maxpsz(tcp);
2546 mss = INFPSZ;
2547 } else if (tcp->tcp_maxpsz_multiplier == 0) {
2548 /*
2549 * Set the sd_qn_maxpsz according to the socket send buffer
2550 * size, and sd_maxblk to INFPSZ (-1). This will essentially
2551 * instruct the stream head to copyin user data into contiguous
2552 * kernel-allocated buffers without breaking it up into smaller
2553 * chunks. We round up the buffer size to the nearest SMSS.
2554 */
2555 maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2556 mss = INFPSZ;
2557 } else {
2558 /*
2559 * Set sd_qn_maxpsz to approx half the (receivers) buffer
2560 * (and a multiple of the mss). This instructs the stream
2561 * head to break down larger than SMSS writes into SMSS-
2562 * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2563 */
2564 maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2565 if (maxpsz > connp->conn_sndbuf / 2) {
2566 maxpsz = connp->conn_sndbuf / 2;
2567 /* Round up to nearest mss */
2568 maxpsz = MSS_ROUNDUP(maxpsz, mss);
2569 }
2570 }
2571
2572 (void) proto_set_maxpsz(q, connp, maxpsz);
2573 if (!(IPCL_IS_NONSTR(connp)))
2574 connp->conn_wq->q_maxpsz = maxpsz;
2575 if (set_maxblk)
2576 (void) proto_set_tx_maxblk(q, connp, mss);
2577 return (mss);
2578 }
2579
2580 /* For /dev/tcp aka AF_INET open */
2581 static int
2582 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2583 {
2584 return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2585 }
2586
2587 /* For /dev/tcp6 aka AF_INET6 open */
2588 static int
2589 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2590 {
2591 return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2592 }
2593
2594 conn_t *
2595 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2596 int *errorp)
2597 {
2598 tcp_t *tcp = NULL;
2599 conn_t *connp;
2600 zoneid_t zoneid;
2601 tcp_stack_t *tcps;
2602 squeue_t *sqp;
2603
2604 ASSERT(errorp != NULL);
2605 /*
2606 * Find the proper zoneid and netstack.
2607 */
2608 /*
2609 * Special case for install: miniroot needs to be able to
2610 * access files via NFS as though it were always in the
2611 * global zone.
2612 */
2613 if (credp == kcred && nfs_global_client_only != 0) {
2614 zoneid = GLOBAL_ZONEID;
2615 tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2616 netstack_tcp;
2617 ASSERT(tcps != NULL);
2618 } else {
2619 netstack_t *ns;
2620 int err;
2621
2622 if ((err = secpolicy_basic_net_access(credp)) != 0) {
2623 *errorp = err;
2624 return (NULL);
2625 }
2626
2627 ns = netstack_find_by_cred(credp);
2628 ASSERT(ns != NULL);
2629 tcps = ns->netstack_tcp;
2630 ASSERT(tcps != NULL);
2631
2632 /*
2633 * For exclusive stacks we set the zoneid to zero
2634 * to make TCP operate as if in the global zone.
2635 */
2636 if (tcps->tcps_netstack->netstack_stackid !=
2637 GLOBAL_NETSTACKID)
2638 zoneid = GLOBAL_ZONEID;
2639 else
2640 zoneid = crgetzoneid(credp);
2641 }
2642
2643 sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2644 connp = (conn_t *)tcp_get_conn(sqp, tcps);
2645 /*
2646 * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2647 * so we drop it by one.
2648 */
2649 netstack_rele(tcps->tcps_netstack);
2650 if (connp == NULL) {
2651 *errorp = ENOSR;
2652 return (NULL);
2653 }
2654 ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2655
2656 connp->conn_sqp = sqp;
2657 connp->conn_initial_sqp = connp->conn_sqp;
2658 connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2659 tcp = connp->conn_tcp;
2660
2661 /*
2662 * Besides asking IP to set the checksum for us, have conn_ip_output
2663 * to do the following checks when necessary:
2664 *
2665 * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2666 * IXAF_VERIFY_PMTU: verify PMTU changes
2667 * IXAF_VERIFY_LSO: verify LSO capability changes
2668 */
2669 connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2670 IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2671
2672 if (!tcps->tcps_dev_flow_ctl)
2673 connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2674
2675 if (isv6) {
2676 connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2677 connp->conn_ipversion = IPV6_VERSION;
2678 connp->conn_family = AF_INET6;
2679 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2680 connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2681 } else {
2682 connp->conn_ipversion = IPV4_VERSION;
2683 connp->conn_family = AF_INET;
2684 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2685 connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2686 }
2687 connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2688
2689 crhold(credp);
2690 connp->conn_cred = credp;
2691 connp->conn_cpid = curproc->p_pid;
2692 connp->conn_open_time = ddi_get_lbolt64();
2693
2694 /* Cache things in the ixa without any refhold */
2695 ASSERT(!(connp->conn_ixa->ixa_free_flags & IXA_FREE_CRED));
2696 connp->conn_ixa->ixa_cred = credp;
2697 connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2698
2699 connp->conn_zoneid = zoneid;
2700 /* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2701 connp->conn_ixa->ixa_zoneid = zoneid;
2702 connp->conn_mlp_type = mlptSingle;
2703 ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2704 ASSERT(tcp->tcp_tcps == tcps);
2705
2706 /*
2707 * If the caller has the process-wide flag set, then default to MAC
2708 * exempt mode. This allows read-down to unlabeled hosts.
2709 */
2710 if (getpflags(NET_MAC_AWARE, credp) != 0)
2711 connp->conn_mac_mode = CONN_MAC_AWARE;
2712
2713 connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2714
2715 if (issocket) {
2716 tcp->tcp_issocket = 1;
2717 }
2718
2719 connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2720 connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2721 connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2722 connp->conn_so_type = SOCK_STREAM;
2723 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2724 tcps->tcps_wroff_xtra;
2725
2726 SOCK_CONNID_INIT(tcp->tcp_connid);
2727 /* DTrace ignores this - it isn't a tcp:::state-change */
2728 tcp->tcp_state = TCPS_IDLE;
2729 tcp_init_values(tcp, NULL);
2730 return (connp);
2731 }
2732
2733 static int
2734 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2735 boolean_t isv6)
2736 {
2737 tcp_t *tcp = NULL;
2738 conn_t *connp = NULL;
2739 int err;
2740 vmem_t *minor_arena = NULL;
2741 dev_t conn_dev;
2742 boolean_t issocket;
2743
2744 if (q->q_ptr != NULL)
2745 return (0);
2746
2747 if (sflag == MODOPEN)
2748 return (EINVAL);
2749
2750 if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2751 ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2752 minor_arena = ip_minor_arena_la;
2753 } else {
2754 /*
2755 * Either minor numbers in the large arena were exhausted
2756 * or a non socket application is doing the open.
2757 * Try to allocate from the small arena.
2758 */
2759 if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2760 return (EBUSY);
2761 }
2762 minor_arena = ip_minor_arena_sa;
2763 }
2764
2765 ASSERT(minor_arena != NULL);
2766
2767 *devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2768
2769 if (flag & SO_FALLBACK) {
2770 /*
2771 * Non streams socket needs a stream to fallback to
2772 */
2773 RD(q)->q_ptr = (void *)conn_dev;
2774 WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2775 WR(q)->q_ptr = (void *)minor_arena;
2776 qprocson(q);
2777 return (0);
2778 } else if (flag & SO_ACCEPTOR) {
2779 q->q_qinfo = &tcp_acceptor_rinit;
2780 /*
2781 * the conn_dev and minor_arena will be subsequently used by
2782 * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2783 * the minor device number for this connection from the q_ptr.
2784 */
2785 RD(q)->q_ptr = (void *)conn_dev;
2786 WR(q)->q_qinfo = &tcp_acceptor_winit;
2787 WR(q)->q_ptr = (void *)minor_arena;
2788 qprocson(q);
2789 return (0);
2790 }
2791
2792 issocket = flag & SO_SOCKSTR;
2793 connp = tcp_create_common(credp, isv6, issocket, &err);
2794
2795 if (connp == NULL) {
2796 inet_minor_free(minor_arena, conn_dev);
2797 q->q_ptr = WR(q)->q_ptr = NULL;
2798 return (err);
2799 }
2800
2801 connp->conn_rq = q;
2802 connp->conn_wq = WR(q);
2803 q->q_ptr = WR(q)->q_ptr = connp;
2804
2805 connp->conn_dev = conn_dev;
2806 connp->conn_minor_arena = minor_arena;
2807
2808 ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2809 ASSERT(WR(q)->q_qinfo == &tcp_winit);
2810
2811 tcp = connp->conn_tcp;
2812
2813 if (issocket) {
2814 WR(q)->q_qinfo = &tcp_sock_winit;
2815 } else {
2816 #ifdef _ILP32
2817 tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2818 #else
2819 tcp->tcp_acceptor_id = conn_dev;
2820 #endif /* _ILP32 */
2821 tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2822 }
2823
2824 /*
2825 * Put the ref for TCP. Ref for IP was already put
2826 * by ipcl_conn_create. Also Make the conn_t globally
2827 * visible to walkers
2828 */
2829 mutex_enter(&connp->conn_lock);
2830 CONN_INC_REF_LOCKED(connp);
2831 ASSERT(connp->conn_ref == 2);
2832 connp->conn_state_flags &= ~CONN_INCIPIENT;
2833 mutex_exit(&connp->conn_lock);
2834
2835 qprocson(q);
2836 return (0);
2837 }
2838
2839 /*
2840 * Build/update the tcp header template (in conn_ht_iphc) based on
2841 * conn_xmit_ipp. The headers include ip6_t, any extension
2842 * headers, and the maximum size tcp header (to avoid reallocation
2843 * on the fly for additional tcp options).
2844 *
2845 * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2846 * Returns failure if can't allocate memory.
2847 */
2848 int
2849 tcp_build_hdrs(tcp_t *tcp)
2850 {
2851 tcp_stack_t *tcps = tcp->tcp_tcps;
2852 conn_t *connp = tcp->tcp_connp;
2853 char buf[TCP_MAX_HDR_LENGTH];
2854 uint_t buflen;
2855 uint_t ulplen = TCP_MIN_HEADER_LENGTH;
2856 uint_t extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
2857 tcpha_t *tcpha;
2858 uint32_t cksum;
2859 int error;
2860
2861 /*
2862 * We might be called after the connection is set up, and we might
2863 * have TS options already in the TCP header. Thus we save any
2864 * existing tcp header.
2865 */
2866 buflen = connp->conn_ht_ulp_len;
2867 if (buflen != 0) {
2868 bcopy(connp->conn_ht_ulp, buf, buflen);
2869 extralen -= buflen - ulplen;
2870 ulplen = buflen;
2871 }
2872
2873 /* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
2874 mutex_enter(&connp->conn_lock);
2875 error = conn_build_hdr_template(connp, ulplen, extralen,
2876 &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
2877 mutex_exit(&connp->conn_lock);
2878 if (error != 0)
2879 return (error);
2880
2881 /*
2882 * Any routing header/option has been massaged. The checksum difference
2883 * is stored in conn_sum for later use.
2884 */
2885 tcpha = (tcpha_t *)connp->conn_ht_ulp;
2886 tcp->tcp_tcpha = tcpha;
2887
2888 /* restore any old tcp header */
2889 if (buflen != 0) {
2890 bcopy(buf, connp->conn_ht_ulp, buflen);
2891 } else {
2892 tcpha->tha_sum = 0;
2893 tcpha->tha_urp = 0;
2894 tcpha->tha_ack = 0;
2895 tcpha->tha_offset_and_reserved = (5 << 4);
2896 tcpha->tha_lport = connp->conn_lport;
2897 tcpha->tha_fport = connp->conn_fport;
2898 }
2899
2900 /*
2901 * IP wants our header length in the checksum field to
2902 * allow it to perform a single pseudo-header+checksum
2903 * calculation on behalf of TCP.
2904 * Include the adjustment for a source route once IP_OPTIONS is set.
2905 */
2906 cksum = sizeof (tcpha_t) + connp->conn_sum;
2907 cksum = (cksum >> 16) + (cksum & 0xFFFF);
2908 ASSERT(cksum < 0x10000);
2909 tcpha->tha_sum = htons(cksum);
2910
2911 if (connp->conn_ipversion == IPV4_VERSION)
2912 tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
2913 else
2914 tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
2915
2916 if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
2917 connp->conn_wroff) {
2918 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2919 tcps->tcps_wroff_xtra;
2920 (void) proto_set_tx_wroff(connp->conn_rq, connp,
2921 connp->conn_wroff);
2922 }
2923 return (0);
2924 }
2925
2926 /*
2927 * tcp_rwnd_set() is called to adjust the receive window to a desired value.
2928 * We do not allow the receive window to shrink. After setting rwnd,
2929 * set the flow control hiwat of the stream.
2930 *
2931 * This function is called in 2 cases:
2932 *
2933 * 1) Before data transfer begins, in tcp_input_listener() for accepting a
2934 * connection (passive open) and in tcp_input_data() for active connect.
2935 * This is called after tcp_mss_set() when the desired MSS value is known.
2936 * This makes sure that our window size is a mutiple of the other side's
2937 * MSS.
2938 * 2) Handling SO_RCVBUF option.
2939 *
2940 * It is ASSUMED that the requested size is a multiple of the current MSS.
2941 *
2942 * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
2943 * user requests so.
2944 */
2945 int
2946 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
2947 {
2948 uint32_t mss = tcp->tcp_mss;
2949 uint32_t old_max_rwnd;
2950 uint32_t max_transmittable_rwnd;
2951 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
2952 tcp_stack_t *tcps = tcp->tcp_tcps;
2953 conn_t *connp = tcp->tcp_connp;
2954
2955 /*
2956 * Insist on a receive window that is at least
2957 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
2958 * funny TCP interactions of Nagle algorithm, SWS avoidance
2959 * and delayed acknowledgement.
2960 */
2961 rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
2962
2963 if (tcp->tcp_fused) {
2964 size_t sth_hiwat;
2965 tcp_t *peer_tcp = tcp->tcp_loopback_peer;
2966
2967 ASSERT(peer_tcp != NULL);
2968 sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
2969 if (!tcp_detached) {
2970 (void) proto_set_rx_hiwat(connp->conn_rq, connp,
2971 sth_hiwat);
2972 tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
2973 }
2974
2975 /* Caller could have changed tcp_rwnd; update tha_win */
2976 if (tcp->tcp_tcpha != NULL) {
2977 tcp->tcp_tcpha->tha_win =
2978 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2979 }
2980 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
2981 tcp->tcp_cwnd_max = rwnd;
2982
2983 /*
2984 * In the fusion case, the maxpsz stream head value of
2985 * our peer is set according to its send buffer size
2986 * and our receive buffer size; since the latter may
2987 * have changed we need to update the peer's maxpsz.
2988 */
2989 (void) tcp_maxpsz_set(peer_tcp, B_TRUE);
2990 return (sth_hiwat);
2991 }
2992
2993 if (tcp_detached)
2994 old_max_rwnd = tcp->tcp_rwnd;
2995 else
2996 old_max_rwnd = connp->conn_rcvbuf;
2997
2998
2999 /*
3000 * If window size info has already been exchanged, TCP should not
3001 * shrink the window. Shrinking window is doable if done carefully.
3002 * We may add that support later. But so far there is not a real
3003 * need to do that.
3004 */
3005 if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
3006 /* MSS may have changed, do a round up again. */
3007 rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
3008 }
3009
3010 /*
3011 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3012 * can be applied even before the window scale option is decided.
3013 */
3014 max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3015 if (rwnd > max_transmittable_rwnd) {
3016 rwnd = max_transmittable_rwnd -
3017 (max_transmittable_rwnd % mss);
3018 if (rwnd < mss)
3019 rwnd = max_transmittable_rwnd;
3020 /*
3021 * If we're over the limit we may have to back down tcp_rwnd.
3022 * The increment below won't work for us. So we set all three
3023 * here and the increment below will have no effect.
3024 */
3025 tcp->tcp_rwnd = old_max_rwnd = rwnd;
3026 }
3027 if (tcp->tcp_localnet) {
3028 tcp->tcp_rack_abs_max =
3029 MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3030 } else {
3031 /*
3032 * For a remote host on a different subnet (through a router),
3033 * we ack every other packet to be conforming to RFC1122.
3034 * tcp_deferred_acks_max is default to 2.
3035 */
3036 tcp->tcp_rack_abs_max =
3037 MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3038 }
3039 if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3040 tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3041 else
3042 tcp->tcp_rack_cur_max = 0;
3043 /*
3044 * Increment the current rwnd by the amount the maximum grew (we
3045 * can not overwrite it since we might be in the middle of a
3046 * connection.)
3047 */
3048 tcp->tcp_rwnd += rwnd - old_max_rwnd;
3049 connp->conn_rcvbuf = rwnd;
3050
3051 /* Are we already connected? */
3052 if (tcp->tcp_tcpha != NULL) {
3053 tcp->tcp_tcpha->tha_win =
3054 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3055 }
3056
3057 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3058 tcp->tcp_cwnd_max = rwnd;
3059
3060 if (tcp_detached)
3061 return (rwnd);
3062
3063 tcp_set_recv_threshold(tcp, rwnd >> 3);
3064
3065 (void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3066 return (rwnd);
3067 }
3068
3069 int
3070 tcp_do_unbind(conn_t *connp)
3071 {
3072 tcp_t *tcp = connp->conn_tcp;
3073 int32_t oldstate;
3074
3075 switch (tcp->tcp_state) {
3076 case TCPS_BOUND:
3077 case TCPS_LISTEN:
3078 break;
3079 default:
3080 return (-TOUTSTATE);
3081 }
3082
3083 /*
3084 * Need to clean up all the eagers since after the unbind, segments
3085 * will no longer be delivered to this listener stream.
3086 */
3087 mutex_enter(&tcp->tcp_eager_lock);
3088 if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3089 tcp_eager_cleanup(tcp, 0);
3090 }
3091 mutex_exit(&tcp->tcp_eager_lock);
3092
3093 /* Clean up the listener connection counter if necessary. */
3094 if (tcp->tcp_listen_cnt != NULL)
3095 TCP_DECR_LISTEN_CNT(tcp);
3096 connp->conn_laddr_v6 = ipv6_all_zeros;
3097 connp->conn_saddr_v6 = ipv6_all_zeros;
3098 tcp_bind_hash_remove(tcp);
3099 oldstate = tcp->tcp_state;
3100 tcp->tcp_state = TCPS_IDLE;
3101 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
3102 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
3103 int32_t, oldstate);
3104
3105 ip_unbind(connp);
3106 bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3107
3108 return (0);
3109 }
3110
3111 /*
3112 * Collect protocol properties to send to the upper handle.
3113 */
3114 void
3115 tcp_get_proto_props(tcp_t *tcp, struct sock_proto_props *sopp)
3116 {
3117 conn_t *connp = tcp->tcp_connp;
3118
3119 sopp->sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3120 sopp->sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3121
3122 sopp->sopp_rxhiwat = tcp->tcp_fused ?
3123 tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3124 connp->conn_rcvbuf;
3125 /*
3126 * Determine what write offset value to use depending on SACK and
3127 * whether the endpoint is fused or not.
3128 */
3129 if (tcp->tcp_fused) {
3130 ASSERT(tcp->tcp_loopback);
3131 ASSERT(tcp->tcp_loopback_peer != NULL);
3132 /*
3133 * For fused tcp loopback, set the stream head's write
3134 * offset value to zero since we won't be needing any room
3135 * for TCP/IP headers. This would also improve performance
3136 * since it would reduce the amount of work done by kmem.
3137 * Non-fused tcp loopback case is handled separately below.
3138 */
3139 sopp->sopp_wroff = 0;
3140 /*
3141 * Update the peer's transmit parameters according to
3142 * our recently calculated high water mark value.
3143 */
3144 (void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3145 } else if (tcp->tcp_snd_sack_ok) {
3146 sopp->sopp_wroff = connp->conn_ht_iphc_allocated +
3147 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3148 } else {
3149 sopp->sopp_wroff = connp->conn_ht_iphc_len +
3150 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3151 }
3152
3153 if (tcp->tcp_loopback) {
3154 sopp->sopp_flags |= SOCKOPT_LOOPBACK;
3155 sopp->sopp_loopback = B_TRUE;
3156 }
3157 }
3158
3159 /*
3160 * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3161 */
3162 boolean_t
3163 tcp_zcopy_check(tcp_t *tcp)
3164 {
3165 conn_t *connp = tcp->tcp_connp;
3166 ip_xmit_attr_t *ixa = connp->conn_ixa;
3167 boolean_t zc_enabled = B_FALSE;
3168 tcp_stack_t *tcps = tcp->tcp_tcps;
3169
3170 if (do_tcpzcopy == 2)
3171 zc_enabled = B_TRUE;
3172 else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3173 zc_enabled = B_TRUE;
3174
3175 tcp->tcp_snd_zcopy_on = zc_enabled;
3176 if (!TCP_IS_DETACHED(tcp)) {
3177 if (zc_enabled) {
3178 ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3179 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3180 ZCVMSAFE);
3181 TCP_STAT(tcps, tcp_zcopy_on);
3182 } else {
3183 ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3184 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3185 ZCVMUNSAFE);
3186 TCP_STAT(tcps, tcp_zcopy_off);
3187 }
3188 }
3189 return (zc_enabled);
3190 }
3191
3192 /*
3193 * Backoff from a zero-copy message by copying data to a new allocated
3194 * message and freeing the original desballoca'ed segmapped message.
3195 *
3196 * This function is called by following two callers:
3197 * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3198 * the origial desballoca'ed message and notify sockfs. This is in re-
3199 * transmit state.
3200 * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3201 * to be copied to new message.
3202 */
3203 mblk_t *
3204 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3205 {
3206 mblk_t *nbp;
3207 mblk_t *head = NULL;
3208 mblk_t *tail = NULL;
3209 tcp_stack_t *tcps = tcp->tcp_tcps;
3210
3211 ASSERT(bp != NULL);
3212 while (bp != NULL) {
3213 if (IS_VMLOANED_MBLK(bp)) {
3214 TCP_STAT(tcps, tcp_zcopy_backoff);
3215 if ((nbp = copyb(bp)) == NULL) {
3216 tcp->tcp_xmit_zc_clean = B_FALSE;
3217 if (tail != NULL)
3218 tail->b_cont = bp;
3219 return ((head == NULL) ? bp : head);
3220 }
3221
3222 if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3223 if (fix_xmitlist)
3224 tcp_zcopy_notify(tcp);
3225 else
3226 nbp->b_datap->db_struioflag |=
3227 STRUIO_ZCNOTIFY;
3228 }
3229 nbp->b_cont = bp->b_cont;
3230
3231 /*
3232 * Copy saved information and adjust tcp_xmit_tail
3233 * if needed.
3234 */
3235 if (fix_xmitlist) {
3236 nbp->b_prev = bp->b_prev;
3237 nbp->b_next = bp->b_next;
3238
3239 if (tcp->tcp_xmit_tail == bp)
3240 tcp->tcp_xmit_tail = nbp;
3241 }
3242
3243 /* Free the original message. */
3244 bp->b_prev = NULL;
3245 bp->b_next = NULL;
3246 freeb(bp);
3247
3248 bp = nbp;
3249 }
3250
3251 if (head == NULL) {
3252 head = bp;
3253 }
3254 if (tail == NULL) {
3255 tail = bp;
3256 } else {
3257 tail->b_cont = bp;
3258 tail = bp;
3259 }
3260
3261 /* Move forward. */
3262 bp = bp->b_cont;
3263 }
3264
3265 if (fix_xmitlist) {
3266 tcp->tcp_xmit_last = tail;
3267 tcp->tcp_xmit_zc_clean = B_TRUE;
3268 }
3269
3270 return (head);
3271 }
3272
3273 void
3274 tcp_zcopy_notify(tcp_t *tcp)
3275 {
3276 struct stdata *stp;
3277 conn_t *connp;
3278
3279 if (tcp->tcp_detached)
3280 return;
3281 connp = tcp->tcp_connp;
3282 if (IPCL_IS_NONSTR(connp)) {
3283 (*connp->conn_upcalls->su_zcopy_notify)
3284 (connp->conn_upper_handle);
3285 return;
3286 }
3287 stp = STREAM(connp->conn_rq);
3288 mutex_enter(&stp->sd_lock);
3289 stp->sd_flag |= STZCNOTIFY;
3290 cv_broadcast(&stp->sd_zcopy_wait);
3291 mutex_exit(&stp->sd_lock);
3292 }
3293
3294 /*
3295 * Update the TCP connection according to change of LSO capability.
3296 */
3297 static void
3298 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3299 {
3300 /*
3301 * We check against IPv4 header length to preserve the old behavior
3302 * of only enabling LSO when there are no IP options.
3303 * But this restriction might not be necessary at all. Before removing
3304 * it, need to verify how LSO is handled for source routing case, with
3305 * which IP does software checksum.
3306 *
3307 * For IPv6, whenever any extension header is needed, LSO is supressed.
3308 */
3309 if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3310 IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3311 return;
3312
3313 /*
3314 * Either the LSO capability newly became usable, or it has changed.
3315 */
3316 if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3317 ill_lso_capab_t *lsoc = &ixa->ixa_lso_capab;
3318
3319 ASSERT(lsoc->ill_lso_max > 0);
3320 tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lsoc->ill_lso_max);
3321
3322 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3323 boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3324
3325 /*
3326 * If LSO to be enabled, notify the STREAM header with larger
3327 * data block.
3328 */
3329 if (!tcp->tcp_lso)
3330 tcp->tcp_maxpsz_multiplier = 0;
3331
3332 tcp->tcp_lso = B_TRUE;
3333 TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3334 } else { /* LSO capability is not usable any more. */
3335 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3336 boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
3337
3338 /*
3339 * If LSO to be disabled, notify the STREAM header with smaller
3340 * data block. And need to restore fragsize to PMTU.
3341 */
3342 if (tcp->tcp_lso) {
3343 tcp->tcp_maxpsz_multiplier =
3344 tcp->tcp_tcps->tcps_maxpsz_multiplier;
3345 ixa->ixa_fragsize = ixa->ixa_pmtu;
3346 tcp->tcp_lso = B_FALSE;
3347 TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
3348 }
3349 }
3350
3351 (void) tcp_maxpsz_set(tcp, B_TRUE);
3352 }
3353
3354 /*
3355 * Update the TCP connection according to change of ZEROCOPY capability.
3356 */
3357 static void
3358 tcp_update_zcopy(tcp_t *tcp)
3359 {
3360 conn_t *connp = tcp->tcp_connp;
3361 tcp_stack_t *tcps = tcp->tcp_tcps;
3362
3363 if (tcp->tcp_snd_zcopy_on) {
3364 tcp->tcp_snd_zcopy_on = B_FALSE;
3365 if (!TCP_IS_DETACHED(tcp)) {
3366 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3367 ZCVMUNSAFE);
3368 TCP_STAT(tcps, tcp_zcopy_off);
3369 }
3370 } else {
3371 tcp->tcp_snd_zcopy_on = B_TRUE;
3372 if (!TCP_IS_DETACHED(tcp)) {
3373 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3374 ZCVMSAFE);
3375 TCP_STAT(tcps, tcp_zcopy_on);
3376 }
3377 }
3378 }
3379
3380 /*
3381 * Notify function registered with ip_xmit_attr_t. It's called in the squeue
3382 * so it's safe to update the TCP connection.
3383 */
3384 /* ARGSUSED1 */
3385 static void
3386 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
3387 ixa_notify_arg_t narg)
3388 {
3389 tcp_t *tcp = (tcp_t *)arg;
3390 conn_t *connp = tcp->tcp_connp;
3391
3392 switch (ntype) {
3393 case IXAN_LSO:
3394 tcp_update_lso(tcp, connp->conn_ixa);
3395 break;
3396 case IXAN_PMTU:
3397 tcp_update_pmtu(tcp, B_FALSE);
3398 break;
3399 case IXAN_ZCOPY:
3400 tcp_update_zcopy(tcp);
3401 break;
3402 default:
3403 break;
3404 }
3405 }
3406
3407 /*
3408 * The TCP write service routine should never be called...
3409 */
3410 /* ARGSUSED */
3411 static void
3412 tcp_wsrv(queue_t *q)
3413 {
3414 tcp_stack_t *tcps = Q_TO_TCP(q)->tcp_tcps;
3415
3416 TCP_STAT(tcps, tcp_wsrv_called);
3417 }
3418
3419 /*
3420 * Hash list lookup routine for tcp_t structures.
3421 * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
3422 */
3423 tcp_t *
3424 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
3425 {
3426 tf_t *tf;
3427 tcp_t *tcp;
3428
3429 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3430 mutex_enter(&tf->tf_lock);
3431 for (tcp = tf->tf_tcp; tcp != NULL;
3432 tcp = tcp->tcp_acceptor_hash) {
3433 if (tcp->tcp_acceptor_id == id) {
3434 CONN_INC_REF(tcp->tcp_connp);
3435 mutex_exit(&tf->tf_lock);
3436 return (tcp);
3437 }
3438 }
3439 mutex_exit(&tf->tf_lock);
3440 return (NULL);
3441 }
3442
3443 /*
3444 * Hash list insertion routine for tcp_t structures.
3445 */
3446 void
3447 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
3448 {
3449 tf_t *tf;
3450 tcp_t **tcpp;
3451 tcp_t *tcpnext;
3452 tcp_stack_t *tcps = tcp->tcp_tcps;
3453
3454 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3455
3456 if (tcp->tcp_ptpahn != NULL)
3457 tcp_acceptor_hash_remove(tcp);
3458 tcpp = &tf->tf_tcp;
3459 mutex_enter(&tf->tf_lock);
3460 tcpnext = tcpp[0];
3461 if (tcpnext)
3462 tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
3463 tcp->tcp_acceptor_hash = tcpnext;
3464 tcp->tcp_ptpahn = tcpp;
3465 tcpp[0] = tcp;
3466 tcp->tcp_acceptor_lockp = &tf->tf_lock; /* For tcp_*_hash_remove */
3467 mutex_exit(&tf->tf_lock);
3468 }
3469
3470 /*
3471 * Hash list removal routine for tcp_t structures.
3472 */
3473 void
3474 tcp_acceptor_hash_remove(tcp_t *tcp)
3475 {
3476 tcp_t *tcpnext;
3477 kmutex_t *lockp;
3478
3479 /*
3480 * Extract the lock pointer in case there are concurrent
3481 * hash_remove's for this instance.
3482 */
3483 lockp = tcp->tcp_acceptor_lockp;
3484
3485 if (tcp->tcp_ptpahn == NULL)
3486 return;
3487
3488 ASSERT(lockp != NULL);
3489 mutex_enter(lockp);
3490 if (tcp->tcp_ptpahn) {
3491 tcpnext = tcp->tcp_acceptor_hash;
3492 if (tcpnext) {
3493 tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
3494 tcp->tcp_acceptor_hash = NULL;
3495 }
3496 *tcp->tcp_ptpahn = tcpnext;
3497 tcp->tcp_ptpahn = NULL;
3498 }
3499 mutex_exit(lockp);
3500 tcp->tcp_acceptor_lockp = NULL;
3501 }
3502
3503 /*
3504 * Type three generator adapted from the random() function in 4.4 BSD:
3505 */
3506
3507 /*
3508 * Copyright (c) 1983, 1993
3509 * The Regents of the University of California. All rights reserved.
3510 *
3511 * Redistribution and use in source and binary forms, with or without
3512 * modification, are permitted provided that the following conditions
3513 * are met:
3514 * 1. Redistributions of source code must retain the above copyright
3515 * notice, this list of conditions and the following disclaimer.
3516 * 2. Redistributions in binary form must reproduce the above copyright
3517 * notice, this list of conditions and the following disclaimer in the
3518 * documentation and/or other materials provided with the distribution.
3519 * 3. All advertising materials mentioning features or use of this software
3520 * must display the following acknowledgement:
3521 * This product includes software developed by the University of
3522 * California, Berkeley and its contributors.
3523 * 4. Neither the name of the University nor the names of its contributors
3524 * may be used to endorse or promote products derived from this software
3525 * without specific prior written permission.
3526 *
3527 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
3528 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
3529 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
3530 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
3531 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
3532 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
3533 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
3534 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
3535 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
3536 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
3537 * SUCH DAMAGE.
3538 */
3539
3540 /* Type 3 -- x**31 + x**3 + 1 */
3541 #define DEG_3 31
3542 #define SEP_3 3
3543
3544
3545 /* Protected by tcp_random_lock */
3546 static int tcp_randtbl[DEG_3 + 1];
3547
3548 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
3549 static int *tcp_random_rptr = &tcp_randtbl[1];
3550
3551 static int *tcp_random_state = &tcp_randtbl[1];
3552 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
3553
3554 kmutex_t tcp_random_lock;
3555
3556 void
3557 tcp_random_init(void)
3558 {
3559 int i;
3560 hrtime_t hrt;
3561 time_t wallclock;
3562 uint64_t result;
3563
3564 /*
3565 * Use high-res timer and current time for seed. Gethrtime() returns
3566 * a longlong, which may contain resolution down to nanoseconds.
3567 * The current time will either be a 32-bit or a 64-bit quantity.
3568 * XOR the two together in a 64-bit result variable.
3569 * Convert the result to a 32-bit value by multiplying the high-order
3570 * 32-bits by the low-order 32-bits.
3571 */
3572
3573 hrt = gethrtime();
3574 (void) drv_getparm(TIME, &wallclock);
3575 result = (uint64_t)wallclock ^ (uint64_t)hrt;
3576 mutex_enter(&tcp_random_lock);
3577 tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
3578 (result & 0xffffffff);
3579
3580 for (i = 1; i < DEG_3; i++)
3581 tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
3582 + 12345;
3583 tcp_random_fptr = &tcp_random_state[SEP_3];
3584 tcp_random_rptr = &tcp_random_state[0];
3585 mutex_exit(&tcp_random_lock);
3586 for (i = 0; i < 10 * DEG_3; i++)
3587 (void) tcp_random();
3588 }
3589
3590 /*
3591 * tcp_random: Return a random number in the range [1 - (128K + 1)].
3592 * This range is selected to be approximately centered on TCP_ISS / 2,
3593 * and easy to compute. We get this value by generating a 32-bit random
3594 * number, selecting out the high-order 17 bits, and then adding one so
3595 * that we never return zero.
3596 */
3597 int
3598 tcp_random(void)
3599 {
3600 int i;
3601
3602 mutex_enter(&tcp_random_lock);
3603 *tcp_random_fptr += *tcp_random_rptr;
3604
3605 /*
3606 * The high-order bits are more random than the low-order bits,
3607 * so we select out the high-order 17 bits and add one so that
3608 * we never return zero.
3609 */
3610 i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
3611 if (++tcp_random_fptr >= tcp_random_end_ptr) {
3612 tcp_random_fptr = tcp_random_state;
3613 ++tcp_random_rptr;
3614 } else if (++tcp_random_rptr >= tcp_random_end_ptr)
3615 tcp_random_rptr = tcp_random_state;
3616
3617 mutex_exit(&tcp_random_lock);
3618 return (i);
3619 }
3620
3621 /*
3622 * Split this function out so that if the secret changes, I'm okay.
3623 *
3624 * Initialize the tcp_iss_cookie and tcp_iss_key.
3625 */
3626
3627 #define PASSWD_SIZE 16 /* MUST be multiple of 4 */
3628
3629 void
3630 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
3631 {
3632 struct {
3633 int32_t current_time;
3634 uint32_t randnum;
3635 uint16_t pad;
3636 uint8_t ether[6];
3637 uint8_t passwd[PASSWD_SIZE];
3638 } tcp_iss_cookie;
3639 time_t t;
3640
3641 /*
3642 * Start with the current absolute time.
3643 */
3644 (void) drv_getparm(TIME, &t);
3645 tcp_iss_cookie.current_time = t;
3646
3647 /*
3648 * XXX - Need a more random number per RFC 1750, not this crap.
3649 * OTOH, if what follows is pretty random, then I'm in better shape.
3650 */
3651 tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
3652 tcp_iss_cookie.pad = 0x365c; /* Picked from HMAC pad values. */
3653
3654 /*
3655 * The cpu_type_info is pretty non-random. Ugggh. It does serve
3656 * as a good template.
3657 */
3658 bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
3659 min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
3660
3661 /*
3662 * The pass-phrase. Normally this is supplied by user-called NDD.
3663 */
3664 bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
3665
3666 /*
3667 * See 4010593 if this section becomes a problem again,
3668 * but the local ethernet address is useful here.
3669 */
3670 (void) localetheraddr(NULL,
3671 (struct ether_addr *)&tcp_iss_cookie.ether);
3672
3673 /*
3674 * Hash 'em all together. The MD5Final is called per-connection.
3675 */
3676 mutex_enter(&tcps->tcps_iss_key_lock);
3677 MD5Init(&tcps->tcps_iss_key);
3678 MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
3679 sizeof (tcp_iss_cookie));
3680 mutex_exit(&tcps->tcps_iss_key_lock);
3681 }
3682
3683 /*
3684 * Called by IP when IP is loaded into the kernel
3685 */
3686 void
3687 tcp_ddi_g_init(void)
3688 {
3689 tcp_timercache = kmem_cache_create("tcp_timercache",
3690 sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
3691 NULL, NULL, NULL, NULL, NULL, 0);
3692
3693 tcp_notsack_blk_cache = kmem_cache_create("tcp_notsack_blk_cache",
3694 sizeof (notsack_blk_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3695
3696 mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
3697
3698 /* Initialize the random number generator */
3699 tcp_random_init();
3700
3701 /* A single callback independently of how many netstacks we have */
3702 ip_squeue_init(tcp_squeue_add);
3703
3704 tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
3705
3706 tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
3707
3708 /*
3709 * We want to be informed each time a stack is created or
3710 * destroyed in the kernel, so we can maintain the
3711 * set of tcp_stack_t's.
3712 */
3713 netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
3714 }
3715
3716
3717 #define INET_NAME "ip"
3718
3719 /*
3720 * Initialize the TCP stack instance.
3721 */
3722 static void *
3723 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
3724 {
3725 tcp_stack_t *tcps;
3726 int i;
3727 int error = 0;
3728 major_t major;
3729 size_t arrsz;
3730
3731 tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
3732 tcps->tcps_netstack = ns;
3733
3734 /* Initialize locks */
3735 mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
3736 mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
3737
3738 tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
3739 tcps->tcps_g_epriv_ports[0] = ULP_DEF_EPRIV_PORT1;
3740 tcps->tcps_g_epriv_ports[1] = ULP_DEF_EPRIV_PORT2;
3741 tcps->tcps_min_anonpriv_port = 512;
3742
3743 tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
3744 TCP_BIND_FANOUT_SIZE, KM_SLEEP);
3745 tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
3746 TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
3747
3748 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3749 mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
3750 MUTEX_DEFAULT, NULL);
3751 }
3752
3753 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3754 mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
3755 MUTEX_DEFAULT, NULL);
3756 }
3757
3758 /* TCP's IPsec code calls the packet dropper. */
3759 ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
3760
3761 arrsz = tcp_propinfo_count * sizeof (mod_prop_info_t);
3762 tcps->tcps_propinfo_tbl = (mod_prop_info_t *)kmem_alloc(arrsz,
3763 KM_SLEEP);
3764 bcopy(tcp_propinfo_tbl, tcps->tcps_propinfo_tbl, arrsz);
3765
3766 /*
3767 * Note: To really walk the device tree you need the devinfo
3768 * pointer to your device which is only available after probe/attach.
3769 * The following is safe only because it uses ddi_root_node()
3770 */
3771 tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
3772 tcp_opt_obj.odb_opt_arr_cnt);
3773
3774 /*
3775 * Initialize RFC 1948 secret values. This will probably be reset once
3776 * by the boot scripts.
3777 *
3778 * Use NULL name, as the name is caught by the new lockstats.
3779 *
3780 * Initialize with some random, non-guessable string, like the global
3781 * T_INFO_ACK.
3782 */
3783
3784 tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
3785 sizeof (tcp_g_t_info_ack), tcps);
3786
3787 tcps->tcps_kstat = tcp_kstat2_init(stackid);
3788 tcps->tcps_mibkp = tcp_kstat_init(stackid);
3789
3790 major = mod_name_to_major(INET_NAME);
3791 error = ldi_ident_from_major(major, &tcps->tcps_ldi_ident);
3792 ASSERT(error == 0);
3793 tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
3794 ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
3795 cv_init(&tcps->tcps_ixa_cleanup_cv, NULL, CV_DEFAULT, NULL);
3796 mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
3797
3798 mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
3799 tcps->tcps_reclaim = B_FALSE;
3800 tcps->tcps_reclaim_tid = 0;
3801 tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
3802
3803 /*
3804 * ncpus is the current number of CPUs, which can be bigger than
3805 * boot_ncpus. But we don't want to use ncpus to allocate all the
3806 * tcp_stats_cpu_t at system boot up time since it will be 1. While
3807 * we handle adding CPU in tcp_cpu_update(), it will be slow if
3808 * there are many CPUs as we will be adding them 1 by 1.
3809 *
3810 * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
3811 * are not freed until the stack is going away. So there is no need
3812 * to grab a lock to access the per CPU tcps_sc[x] pointer.
3813 */
3814 mutex_enter(&cpu_lock);
3815 tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
3816 mutex_exit(&cpu_lock);
3817 tcps->tcps_sc = kmem_zalloc(max_ncpus * sizeof (tcp_stats_cpu_t *),
3818 KM_SLEEP);
3819 for (i = 0; i < tcps->tcps_sc_cnt; i++) {
3820 tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
3821 KM_SLEEP);
3822 }
3823
3824 mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
3825 list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
3826 offsetof(tcp_listener_t, tl_link));
3827
3828 return (tcps);
3829 }
3830
3831 /*
3832 * Called when the IP module is about to be unloaded.
3833 */
3834 void
3835 tcp_ddi_g_destroy(void)
3836 {
3837 tcp_g_kstat_fini(tcp_g_kstat);
3838 tcp_g_kstat = NULL;
3839 bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
3840
3841 mutex_destroy(&tcp_random_lock);
3842
3843 kmem_cache_destroy(tcp_timercache);
3844 kmem_cache_destroy(tcp_notsack_blk_cache);
3845
3846 netstack_unregister(NS_TCP);
3847 }
3848
3849 /*
3850 * Free the TCP stack instance.
3851 */
3852 static void
3853 tcp_stack_fini(netstackid_t stackid, void *arg)
3854 {
3855 tcp_stack_t *tcps = (tcp_stack_t *)arg;
3856 int i;
3857
3858 freeb(tcps->tcps_ixa_cleanup_mp);
3859 tcps->tcps_ixa_cleanup_mp = NULL;
3860 cv_destroy(&tcps->tcps_ixa_cleanup_cv);
3861 mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
3862
3863 /*
3864 * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
3865 * the timer.
3866 */
3867 mutex_enter(&tcps->tcps_reclaim_lock);
3868 tcps->tcps_reclaim = B_FALSE;
3869 mutex_exit(&tcps->tcps_reclaim_lock);
3870 if (tcps->tcps_reclaim_tid != 0)
3871 (void) untimeout(tcps->tcps_reclaim_tid);
3872 mutex_destroy(&tcps->tcps_reclaim_lock);
3873
3874 tcp_listener_conf_cleanup(tcps);
3875
3876 for (i = 0; i < tcps->tcps_sc_cnt; i++)
3877 kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
3878 kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
3879
3880 kmem_free(tcps->tcps_propinfo_tbl,
3881 tcp_propinfo_count * sizeof (mod_prop_info_t));
3882 tcps->tcps_propinfo_tbl = NULL;
3883
3884 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3885 ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
3886 mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
3887 }
3888
3889 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3890 ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
3891 mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
3892 }
3893
3894 kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
3895 tcps->tcps_bind_fanout = NULL;
3896
3897 kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
3898 TCP_ACCEPTOR_FANOUT_SIZE);
3899 tcps->tcps_acceptor_fanout = NULL;
3900
3901 mutex_destroy(&tcps->tcps_iss_key_lock);
3902 mutex_destroy(&tcps->tcps_epriv_port_lock);
3903
3904 ip_drop_unregister(&tcps->tcps_dropper);
3905
3906 tcp_kstat2_fini(stackid, tcps->tcps_kstat);
3907 tcps->tcps_kstat = NULL;
3908
3909 tcp_kstat_fini(stackid, tcps->tcps_mibkp);
3910 tcps->tcps_mibkp = NULL;
3911
3912 ldi_ident_release(tcps->tcps_ldi_ident);
3913 kmem_free(tcps, sizeof (*tcps));
3914 }
3915
3916 /*
3917 * Generate ISS, taking into account NDD changes may happen halfway through.
3918 * (If the iss is not zero, set it.)
3919 */
3920
3921 static void
3922 tcp_iss_init(tcp_t *tcp)
3923 {
3924 MD5_CTX context;
3925 struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
3926 uint32_t answer[4];
3927 tcp_stack_t *tcps = tcp->tcp_tcps;
3928 conn_t *connp = tcp->tcp_connp;
3929
3930 tcps->tcps_iss_incr_extra += (tcps->tcps_iss_incr >> 1);
3931 tcp->tcp_iss = tcps->tcps_iss_incr_extra;
3932 switch (tcps->tcps_strong_iss) {
3933 case 2:
3934 mutex_enter(&tcps->tcps_iss_key_lock);
3935 context = tcps->tcps_iss_key;
3936 mutex_exit(&tcps->tcps_iss_key_lock);
3937 arg.ports = connp->conn_ports;
3938 arg.src = connp->conn_laddr_v6;
3939 arg.dst = connp->conn_faddr_v6;
3940 MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
3941 MD5Final((uchar_t *)answer, &context);
3942 tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
3943 /*
3944 * Now that we've hashed into a unique per-connection sequence
3945 * space, add a random increment per strong_iss == 1. So I
3946 * guess we'll have to...
3947 */
3948 /* FALLTHRU */
3949 case 1:
3950 tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
3951 break;
3952 default:
3953 tcp->tcp_iss += (uint32_t)gethrestime_sec() *
3954 tcps->tcps_iss_incr;
3955 break;
3956 }
3957 tcp->tcp_valid_bits = TCP_ISS_VALID;
3958 tcp->tcp_fss = tcp->tcp_iss - 1;
3959 tcp->tcp_suna = tcp->tcp_iss;
3960 tcp->tcp_snxt = tcp->tcp_iss + 1;
3961 tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3962 tcp->tcp_csuna = tcp->tcp_snxt;
3963 }
3964
3965 /*
3966 * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
3967 * on the specified backing STREAMS q. Note, the caller may make the
3968 * decision to call based on the tcp_t.tcp_flow_stopped value which
3969 * when check outside the q's lock is only an advisory check ...
3970 */
3971 void
3972 tcp_setqfull(tcp_t *tcp)
3973 {
3974 tcp_stack_t *tcps = tcp->tcp_tcps;
3975 conn_t *connp = tcp->tcp_connp;
3976
3977 if (tcp->tcp_closed)
3978 return;
3979
3980 conn_setqfull(connp, &tcp->tcp_flow_stopped);
3981 if (tcp->tcp_flow_stopped)
3982 TCP_STAT(tcps, tcp_flwctl_on);
3983 }
3984
3985 void
3986 tcp_clrqfull(tcp_t *tcp)
3987 {
3988 conn_t *connp = tcp->tcp_connp;
3989
3990 if (tcp->tcp_closed)
3991 return;
3992 conn_clrqfull(connp, &tcp->tcp_flow_stopped);
3993 }
3994
3995 static int
3996 tcp_squeue_switch(int val)
3997 {
3998 int rval = SQ_FILL;
3999
4000 switch (val) {
4001 case 1:
4002 rval = SQ_NODRAIN;
4003 break;
4004 case 2:
4005 rval = SQ_PROCESS;
4006 break;
4007 default:
4008 break;
4009 }
4010 return (rval);
4011 }
4012
4013 /*
4014 * This is called once for each squeue - globally for all stack
4015 * instances.
4016 */
4017 static void
4018 tcp_squeue_add(squeue_t *sqp)
4019 {
4020 tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4021 sizeof (tcp_squeue_priv_t), KM_SLEEP);
4022
4023 *squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4024 if (tcp_free_list_max_cnt == 0) {
4025 int tcp_ncpus = ((boot_max_ncpus == -1) ?
4026 max_ncpus : boot_max_ncpus);
4027
4028 /*
4029 * Limit number of entries to 1% of availble memory / tcp_ncpus
4030 */
4031 tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4032 (tcp_ncpus * sizeof (tcp_t) * 100);
4033 }
4034 tcp_time_wait->tcp_free_list_cnt = 0;
4035 }
4036 /*
4037 * Return unix error is tli error is TSYSERR, otherwise return a negative
4038 * tli error.
4039 */
4040 int
4041 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4042 boolean_t bind_to_req_port_only)
4043 {
4044 int error;
4045 tcp_t *tcp = connp->conn_tcp;
4046
4047 if (tcp->tcp_state >= TCPS_BOUND) {
4048 if (connp->conn_debug) {
4049 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4050 "tcp_bind: bad state, %d", tcp->tcp_state);
4051 }
4052 return (-TOUTSTATE);
4053 }
4054
4055 error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4056 if (error != 0)
4057 return (error);
4058
4059 ASSERT(tcp->tcp_state == TCPS_BOUND);
4060 tcp->tcp_conn_req_max = 0;
4061 return (0);
4062 }
4063
4064 /*
4065 * If the return value from this function is positive, it's a UNIX error.
4066 * Otherwise, if it's negative, then the absolute value is a TLI error.
4067 * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4068 */
4069 int
4070 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4071 cred_t *cr, pid_t pid)
4072 {
4073 tcp_t *tcp = connp->conn_tcp;
4074 sin_t *sin = (sin_t *)sa;
4075 sin6_t *sin6 = (sin6_t *)sa;
4076 ipaddr_t *dstaddrp;
4077 in_port_t dstport;
4078 uint_t srcid;
4079 int error;
4080 uint32_t mss;
4081 mblk_t *syn_mp;
4082 tcp_stack_t *tcps = tcp->tcp_tcps;
4083 int32_t oldstate;
4084 ip_xmit_attr_t *ixa = connp->conn_ixa;
4085
4086 oldstate = tcp->tcp_state;
4087
4088 switch (len) {
4089 default:
4090 /*
4091 * Should never happen
4092 */
4093 return (EINVAL);
4094
4095 case sizeof (sin_t):
4096 sin = (sin_t *)sa;
4097 if (sin->sin_port == 0) {
4098 return (-TBADADDR);
4099 }
4100 if (connp->conn_ipv6_v6only) {
4101 return (EAFNOSUPPORT);
4102 }
4103 break;
4104
4105 case sizeof (sin6_t):
4106 sin6 = (sin6_t *)sa;
4107 if (sin6->sin6_port == 0) {
4108 return (-TBADADDR);
4109 }
4110 break;
4111 }
4112 /*
4113 * If we're connecting to an IPv4-mapped IPv6 address, we need to
4114 * make sure that the conn_ipversion is IPV4_VERSION. We
4115 * need to this before we call tcp_bindi() so that the port lookup
4116 * code will look for ports in the correct port space (IPv4 and
4117 * IPv6 have separate port spaces).
4118 */
4119 if (connp->conn_family == AF_INET6 &&
4120 connp->conn_ipversion == IPV6_VERSION &&
4121 IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4122 if (connp->conn_ipv6_v6only)
4123 return (EADDRNOTAVAIL);
4124
4125 connp->conn_ipversion = IPV4_VERSION;
4126 }
4127
4128 switch (tcp->tcp_state) {
4129 case TCPS_LISTEN:
4130 /*
4131 * Listening sockets are not allowed to issue connect().
4132 */
4133 if (IPCL_IS_NONSTR(connp))
4134 return (EOPNOTSUPP);
4135 /* FALLTHRU */
4136 case TCPS_IDLE:
4137 /*
4138 * We support quick connect, refer to comments in
4139 * tcp_connect_*()
4140 */
4141 /* FALLTHRU */
4142 case TCPS_BOUND:
4143 break;
4144 default:
4145 return (-TOUTSTATE);
4146 }
4147
4148 /*
4149 * We update our cred/cpid based on the caller of connect
4150 */
4151 if (connp->conn_cred != cr) {
4152 crhold(cr);
4153 crfree(connp->conn_cred);
4154 connp->conn_cred = cr;
4155 }
4156 connp->conn_cpid = pid;
4157
4158 /* Cache things in the ixa without any refhold */
4159 ASSERT(!(ixa->ixa_free_flags & IXA_FREE_CRED));
4160 ixa->ixa_cred = cr;
4161 ixa->ixa_cpid = pid;
4162 if (is_system_labeled()) {
4163 /* We need to restart with a label based on the cred */
4164 ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4165 }
4166
4167 if (connp->conn_family == AF_INET6) {
4168 if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4169 error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4170 sin6->sin6_port, sin6->sin6_flowinfo,
4171 sin6->__sin6_src_id, sin6->sin6_scope_id);
4172 } else {
4173 /*
4174 * Destination adress is mapped IPv6 address.
4175 * Source bound address should be unspecified or
4176 * IPv6 mapped address as well.
4177 */
4178 if (!IN6_IS_ADDR_UNSPECIFIED(
4179 &connp->conn_bound_addr_v6) &&
4180 !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4181 return (EADDRNOTAVAIL);
4182 }
4183 dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4184 dstport = sin6->sin6_port;
4185 srcid = sin6->__sin6_src_id;
4186 error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4187 srcid);
4188 }
4189 } else {
4190 dstaddrp = &sin->sin_addr.s_addr;
4191 dstport = sin->sin_port;
4192 srcid = 0;
4193 error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4194 }
4195
4196 if (error != 0)
4197 goto connect_failed;
4198
4199 CL_INET_CONNECT(connp, B_TRUE, error);
4200 if (error != 0)
4201 goto connect_failed;
4202
4203 /* connect succeeded */
4204 TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4205 tcp->tcp_active_open = 1;
4206
4207 /*
4208 * tcp_set_destination() does not adjust for TCP/IP header length.
4209 */
4210 mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4211
4212 /*
4213 * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4214 * to the nearest MSS.
4215 *
4216 * We do the round up here because we need to get the interface MTU
4217 * first before we can do the round up.
4218 */
4219 tcp->tcp_rwnd = connp->conn_rcvbuf;
4220 tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4221 tcps->tcps_recv_hiwat_minmss * mss);
4222 connp->conn_rcvbuf = tcp->tcp_rwnd;
4223 tcp_set_ws_value(tcp);
4224 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4225 if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4226 tcp->tcp_snd_ws_ok = B_TRUE;
4227
4228 /*
4229 * Set tcp_snd_ts_ok to true
4230 * so that tcp_xmit_mp will
4231 * include the timestamp
4232 * option in the SYN segment.
4233 */
4234 if (tcps->tcps_tstamp_always ||
4235 (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4236 tcp->tcp_snd_ts_ok = B_TRUE;
4237 }
4238
4239 /*
4240 * Note that tcp_snd_sack_ok can be set in tcp_set_destination() if
4241 * the SACK metric is set. So here we just check the per stack SACK
4242 * permitted param.
4243 */
4244 if (tcps->tcps_sack_permitted == 2) {
4245 ASSERT(tcp->tcp_num_sack_blk == 0);
4246 ASSERT(tcp->tcp_notsack_list == NULL);
4247 tcp->tcp_snd_sack_ok = B_TRUE;
4248 }
4249
4250 /*
4251 * Should we use ECN? Note that the current
4252 * default value (SunOS 5.9) of tcp_ecn_permitted
4253 * is 1. The reason for doing this is that there
4254 * are equipments out there that will drop ECN
4255 * enabled IP packets. Setting it to 1 avoids
4256 * compatibility problems.
4257 */
4258 if (tcps->tcps_ecn_permitted == 2)
4259 tcp->tcp_ecn_ok = B_TRUE;
4260
4261 /* Trace change from BOUND -> SYN_SENT here */
4262 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4263 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4264 int32_t, TCPS_BOUND);
4265
4266 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4267 syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4268 tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4269 if (syn_mp != NULL) {
4270 /*
4271 * We must bump the generation before sending the syn
4272 * to ensure that we use the right generation in case
4273 * this thread issues a "connected" up call.
4274 */
4275 SOCK_CONNID_BUMP(tcp->tcp_connid);
4276 /*
4277 * DTrace sending the first SYN as a
4278 * tcp:::connect-request event.
4279 */
4280 DTRACE_TCP5(connect__request, mblk_t *, NULL,
4281 ip_xmit_attr_t *, connp->conn_ixa,
4282 void_ip_t *, syn_mp->b_rptr, tcp_t *, tcp,
4283 tcph_t *,
4284 &syn_mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
4285 tcp_send_data(tcp, syn_mp);
4286 }
4287
4288 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4289 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4290 return (0);
4291
4292 connect_failed:
4293 connp->conn_faddr_v6 = ipv6_all_zeros;
4294 connp->conn_fport = 0;
4295 tcp->tcp_state = oldstate;
4296 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4297 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4298 return (error);
4299 }
4300
4301 int
4302 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4303 int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4304 {
4305 tcp_t *tcp = connp->conn_tcp;
4306 int error = 0;
4307 tcp_stack_t *tcps = tcp->tcp_tcps;
4308 int32_t oldstate;
4309
4310 /* All Solaris components should pass a cred for this operation. */
4311 ASSERT(cr != NULL);
4312
4313 if (tcp->tcp_state >= TCPS_BOUND) {
4314 if ((tcp->tcp_state == TCPS_BOUND ||
4315 tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
4316 /*
4317 * Handle listen() increasing backlog.
4318 * This is more "liberal" then what the TPI spec
4319 * requires but is needed to avoid a t_unbind
4320 * when handling listen() since the port number
4321 * might be "stolen" between the unbind and bind.
4322 */
4323 goto do_listen;
4324 }
4325 if (connp->conn_debug) {
4326 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4327 "tcp_listen: bad state, %d", tcp->tcp_state);
4328 }
4329 return (-TOUTSTATE);
4330 } else {
4331 if (sa == NULL) {
4332 sin6_t addr;
4333 sin_t *sin;
4334 sin6_t *sin6;
4335
4336 ASSERT(IPCL_IS_NONSTR(connp));
4337 /* Do an implicit bind: Request for a generic port. */
4338 if (connp->conn_family == AF_INET) {
4339 len = sizeof (sin_t);
4340 sin = (sin_t *)&addr;
4341 *sin = sin_null;
4342 sin->sin_family = AF_INET;
4343 } else {
4344 ASSERT(connp->conn_family == AF_INET6);
4345 len = sizeof (sin6_t);
4346 sin6 = (sin6_t *)&addr;
4347 *sin6 = sin6_null;
4348 sin6->sin6_family = AF_INET6;
4349 }
4350 sa = (struct sockaddr *)&addr;
4351 }
4352
4353 error = tcp_bind_check(connp, sa, len, cr,
4354 bind_to_req_port_only);
4355 if (error)
4356 return (error);
4357 /* Fall through and do the fanout insertion */
4358 }
4359
4360 do_listen:
4361 ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
4362 tcp->tcp_conn_req_max = backlog;
4363 if (tcp->tcp_conn_req_max) {
4364 if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
4365 tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
4366 if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
4367 tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
4368 /*
4369 * If this is a listener, do not reset the eager list
4370 * and other stuffs. Note that we don't check if the
4371 * existing eager list meets the new tcp_conn_req_max
4372 * requirement.
4373 */
4374 if (tcp->tcp_state != TCPS_LISTEN) {
4375 tcp->tcp_state = TCPS_LISTEN;
4376 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4377 connp->conn_ixa, void, NULL, tcp_t *, tcp,
4378 void, NULL, int32_t, TCPS_BOUND);
4379 /* Initialize the chain. Don't need the eager_lock */
4380 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
4381 tcp->tcp_eager_next_drop_q0 = tcp;
4382 tcp->tcp_eager_prev_drop_q0 = tcp;
4383 tcp->tcp_second_ctimer_threshold =
4384 tcps->tcps_ip_abort_linterval;
4385 }
4386 }
4387
4388 /*
4389 * We need to make sure that the conn_recv is set to a non-null
4390 * value before we insert the conn into the classifier table.
4391 * This is to avoid a race with an incoming packet which does an
4392 * ipcl_classify().
4393 * We initially set it to tcp_input_listener_unbound to try to
4394 * pick a good squeue for the listener when the first SYN arrives.
4395 * tcp_input_listener_unbound sets it to tcp_input_listener on that
4396 * first SYN.
4397 */
4398 connp->conn_recv = tcp_input_listener_unbound;
4399
4400 /* Insert the listener in the classifier table */
4401 error = ip_laddr_fanout_insert(connp);
4402 if (error != 0) {
4403 /* Undo the bind - release the port number */
4404 oldstate = tcp->tcp_state;
4405 tcp->tcp_state = TCPS_IDLE;
4406 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4407 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4408 int32_t, oldstate);
4409 connp->conn_bound_addr_v6 = ipv6_all_zeros;
4410
4411 connp->conn_laddr_v6 = ipv6_all_zeros;
4412 connp->conn_saddr_v6 = ipv6_all_zeros;
4413 connp->conn_ports = 0;
4414
4415 if (connp->conn_anon_port) {
4416 zone_t *zone;
4417
4418 zone = crgetzone(cr);
4419 connp->conn_anon_port = B_FALSE;
4420 (void) tsol_mlp_anon(zone, connp->conn_mlp_type,
4421 connp->conn_proto, connp->conn_lport, B_FALSE);
4422 }
4423 connp->conn_mlp_type = mlptSingle;
4424
4425 tcp_bind_hash_remove(tcp);
4426 return (error);
4427 } else {
4428 /*
4429 * If there is a connection limit, allocate and initialize
4430 * the counter struct. Note that since listen can be called
4431 * multiple times, the struct may have been allready allocated.
4432 */
4433 if (!list_is_empty(&tcps->tcps_listener_conf) &&
4434 tcp->tcp_listen_cnt == NULL) {
4435 tcp_listen_cnt_t *tlc;
4436 uint32_t ratio;
4437
4438 ratio = tcp_find_listener_conf(tcps,
4439 ntohs(connp->conn_lport));
4440 if (ratio != 0) {
4441 uint32_t mem_ratio, tot_buf;
4442
4443 tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
4444 KM_SLEEP);
4445 /*
4446 * Calculate the connection limit based on
4447 * the configured ratio and maxusers. Maxusers
4448 * are calculated based on memory size,
4449 * ~ 1 user per MB. Note that the conn_rcvbuf
4450 * and conn_sndbuf may change after a
4451 * connection is accepted. So what we have
4452 * is only an approximation.
4453 */
4454 if ((tot_buf = connp->conn_rcvbuf +
4455 connp->conn_sndbuf) < MB) {
4456 mem_ratio = MB / tot_buf;
4457 tlc->tlc_max = maxusers / ratio *
4458 mem_ratio;
4459 } else {
4460 mem_ratio = tot_buf / MB;
4461 tlc->tlc_max = maxusers / ratio /
4462 mem_ratio;
4463 }
4464 /* At least we should allow two connections! */
4465 if (tlc->tlc_max <= tcp_min_conn_listener)
4466 tlc->tlc_max = tcp_min_conn_listener;
4467 tlc->tlc_cnt = 1;
4468 tlc->tlc_drop = 0;
4469 tcp->tcp_listen_cnt = tlc;
4470 }
4471 }
4472 }
4473 return (error);
4474 }