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