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