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