Print this page
5295 remove maxburst logic from TCP's send algorithm Reviewed by: Dan McDonald <danmcd@omniti.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/inet/tcp/tcp.c
+++ new/usr/src/uts/common/inet/tcp/tcp.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
|
↓ open down ↓ |
15 lines elided |
↑ open up ↑ |
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 * Copyright (c) 2011, Joyent Inc. All rights reserved.
25 25 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
26 - * Copyright (c) 2013 by Delphix. All rights reserved.
26 + * Copyright (c) 2013,2014 by Delphix. All rights reserved.
27 27 * Copyright 2014, OmniTI Computer Consulting, Inc. All rights reserved.
28 28 */
29 29 /* Copyright (c) 1990 Mentat Inc. */
30 30
31 31 #include <sys/types.h>
32 32 #include <sys/stream.h>
33 33 #include <sys/strsun.h>
34 34 #include <sys/strsubr.h>
35 35 #include <sys/stropts.h>
36 36 #include <sys/strlog.h>
37 37 #define _SUN_TPI_VERSION 2
38 38 #include <sys/tihdr.h>
39 39 #include <sys/timod.h>
40 40 #include <sys/ddi.h>
41 41 #include <sys/sunddi.h>
42 42 #include <sys/suntpi.h>
43 43 #include <sys/xti_inet.h>
44 44 #include <sys/cmn_err.h>
45 45 #include <sys/debug.h>
46 46 #include <sys/sdt.h>
47 47 #include <sys/vtrace.h>
48 48 #include <sys/kmem.h>
49 49 #include <sys/ethernet.h>
50 50 #include <sys/cpuvar.h>
51 51 #include <sys/dlpi.h>
52 52 #include <sys/pattr.h>
53 53 #include <sys/policy.h>
54 54 #include <sys/priv.h>
55 55 #include <sys/zone.h>
56 56 #include <sys/sunldi.h>
57 57
58 58 #include <sys/errno.h>
59 59 #include <sys/signal.h>
60 60 #include <sys/socket.h>
61 61 #include <sys/socketvar.h>
62 62 #include <sys/sockio.h>
63 63 #include <sys/isa_defs.h>
64 64 #include <sys/md5.h>
65 65 #include <sys/random.h>
66 66 #include <sys/uio.h>
67 67 #include <sys/systm.h>
68 68 #include <netinet/in.h>
69 69 #include <netinet/tcp.h>
70 70 #include <netinet/ip6.h>
71 71 #include <netinet/icmp6.h>
72 72 #include <net/if.h>
73 73 #include <net/route.h>
74 74 #include <inet/ipsec_impl.h>
75 75
76 76 #include <inet/common.h>
77 77 #include <inet/ip.h>
78 78 #include <inet/ip_impl.h>
79 79 #include <inet/ip6.h>
80 80 #include <inet/ip_ndp.h>
81 81 #include <inet/proto_set.h>
82 82 #include <inet/mib2.h>
83 83 #include <inet/optcom.h>
84 84 #include <inet/snmpcom.h>
85 85 #include <inet/kstatcom.h>
86 86 #include <inet/tcp.h>
87 87 #include <inet/tcp_impl.h>
88 88 #include <inet/tcp_cluster.h>
89 89 #include <inet/udp_impl.h>
90 90 #include <net/pfkeyv2.h>
91 91 #include <inet/ipdrop.h>
92 92
93 93 #include <inet/ipclassifier.h>
94 94 #include <inet/ip_ire.h>
95 95 #include <inet/ip_ftable.h>
96 96 #include <inet/ip_if.h>
97 97 #include <inet/ipp_common.h>
98 98 #include <inet/ip_rts.h>
99 99 #include <inet/ip_netinfo.h>
100 100 #include <sys/squeue_impl.h>
101 101 #include <sys/squeue.h>
102 102 #include <sys/tsol/label.h>
103 103 #include <sys/tsol/tnet.h>
104 104 #include <rpc/pmap_prot.h>
105 105 #include <sys/callo.h>
106 106
107 107 /*
108 108 * TCP Notes: aka FireEngine Phase I (PSARC 2002/433)
109 109 *
110 110 * (Read the detailed design doc in PSARC case directory)
111 111 *
112 112 * The entire tcp state is contained in tcp_t and conn_t structure
113 113 * which are allocated in tandem using ipcl_conn_create() and passing
114 114 * IPCL_TCPCONN as a flag. We use 'conn_ref' and 'conn_lock' to protect
115 115 * the references on the tcp_t. The tcp_t structure is never compressed
116 116 * and packets always land on the correct TCP perimeter from the time
117 117 * eager is created till the time tcp_t dies (as such the old mentat
118 118 * TCP global queue is not used for detached state and no IPSEC checking
119 119 * is required). The global queue is still allocated to send out resets
120 120 * for connection which have no listeners and IP directly calls
121 121 * tcp_xmit_listeners_reset() which does any policy check.
122 122 *
123 123 * Protection and Synchronisation mechanism:
124 124 *
125 125 * The tcp data structure does not use any kind of lock for protecting
126 126 * its state but instead uses 'squeues' for mutual exclusion from various
127 127 * read and write side threads. To access a tcp member, the thread should
128 128 * always be behind squeue (via squeue_enter with flags as SQ_FILL, SQ_PROCESS,
129 129 * or SQ_NODRAIN). Since the squeues allow a direct function call, caller
130 130 * can pass any tcp function having prototype of edesc_t as argument
131 131 * (different from traditional STREAMs model where packets come in only
132 132 * designated entry points). The list of functions that can be directly
133 133 * called via squeue are listed before the usual function prototype.
134 134 *
135 135 * Referencing:
136 136 *
137 137 * TCP is MT-Hot and we use a reference based scheme to make sure that the
138 138 * tcp structure doesn't disappear when its needed. When the application
139 139 * creates an outgoing connection or accepts an incoming connection, we
140 140 * start out with 2 references on 'conn_ref'. One for TCP and one for IP.
141 141 * The IP reference is just a symbolic reference since ip_tcpclose()
142 142 * looks at tcp structure after tcp_close_output() returns which could
143 143 * have dropped the last TCP reference. So as long as the connection is
144 144 * in attached state i.e. !TCP_IS_DETACHED, we have 2 references on the
145 145 * conn_t. The classifier puts its own reference when the connection is
146 146 * inserted in listen or connected hash. Anytime a thread needs to enter
147 147 * the tcp connection perimeter, it retrieves the conn/tcp from q->ptr
148 148 * on write side or by doing a classify on read side and then puts a
149 149 * reference on the conn before doing squeue_enter/tryenter/fill. For
150 150 * read side, the classifier itself puts the reference under fanout lock
151 151 * to make sure that tcp can't disappear before it gets processed. The
152 152 * squeue will drop this reference automatically so the called function
153 153 * doesn't have to do a DEC_REF.
154 154 *
155 155 * Opening a new connection:
156 156 *
157 157 * The outgoing connection open is pretty simple. tcp_open() does the
158 158 * work in creating the conn/tcp structure and initializing it. The
159 159 * squeue assignment is done based on the CPU the application
160 160 * is running on. So for outbound connections, processing is always done
161 161 * on application CPU which might be different from the incoming CPU
162 162 * being interrupted by the NIC. An optimal way would be to figure out
163 163 * the NIC <-> CPU binding at listen time, and assign the outgoing
164 164 * connection to the squeue attached to the CPU that will be interrupted
165 165 * for incoming packets (we know the NIC based on the bind IP address).
166 166 * This might seem like a problem if more data is going out but the
167 167 * fact is that in most cases the transmit is ACK driven transmit where
168 168 * the outgoing data normally sits on TCP's xmit queue waiting to be
169 169 * transmitted.
170 170 *
171 171 * Accepting a connection:
172 172 *
173 173 * This is a more interesting case because of various races involved in
174 174 * establishing a eager in its own perimeter. Read the meta comment on
175 175 * top of tcp_input_listener(). But briefly, the squeue is picked by
176 176 * ip_fanout based on the ring or the sender (if loopback).
177 177 *
178 178 * Closing a connection:
179 179 *
180 180 * The close is fairly straight forward. tcp_close() calls tcp_close_output()
181 181 * via squeue to do the close and mark the tcp as detached if the connection
182 182 * was in state TCPS_ESTABLISHED or greater. In the later case, TCP keep its
183 183 * reference but tcp_close() drop IP's reference always. So if tcp was
184 184 * not killed, it is sitting in time_wait list with 2 reference - 1 for TCP
185 185 * and 1 because it is in classifier's connected hash. This is the condition
186 186 * we use to determine that its OK to clean up the tcp outside of squeue
187 187 * when time wait expires (check the ref under fanout and conn_lock and
188 188 * if it is 2, remove it from fanout hash and kill it).
189 189 *
190 190 * Although close just drops the necessary references and marks the
191 191 * tcp_detached state, tcp_close needs to know the tcp_detached has been
192 192 * set (under squeue) before letting the STREAM go away (because a
193 193 * inbound packet might attempt to go up the STREAM while the close
194 194 * has happened and tcp_detached is not set). So a special lock and
195 195 * flag is used along with a condition variable (tcp_closelock, tcp_closed,
196 196 * and tcp_closecv) to signal tcp_close that tcp_close_out() has marked
197 197 * tcp_detached.
198 198 *
199 199 * Special provisions and fast paths:
200 200 *
201 201 * We make special provisions for sockfs by marking tcp_issocket
202 202 * whenever we have only sockfs on top of TCP. This allows us to skip
203 203 * putting the tcp in acceptor hash since a sockfs listener can never
204 204 * become acceptor and also avoid allocating a tcp_t for acceptor STREAM
205 205 * since eager has already been allocated and the accept now happens
206 206 * on acceptor STREAM. There is a big blob of comment on top of
207 207 * tcp_input_listener explaining the new accept. When socket is POP'd,
208 208 * sockfs sends us an ioctl to mark the fact and we go back to old
209 209 * behaviour. Once tcp_issocket is unset, its never set for the
210 210 * life of that connection.
211 211 *
212 212 * IPsec notes :
213 213 *
214 214 * Since a packet is always executed on the correct TCP perimeter
215 215 * all IPsec processing is defered to IP including checking new
216 216 * connections and setting IPSEC policies for new connection. The
217 217 * only exception is tcp_xmit_listeners_reset() which is called
218 218 * directly from IP and needs to policy check to see if TH_RST
219 219 * can be sent out.
220 220 */
221 221
222 222 /*
223 223 * Values for squeue switch:
224 224 * 1: SQ_NODRAIN
225 225 * 2: SQ_PROCESS
226 226 * 3: SQ_FILL
227 227 */
228 228 int tcp_squeue_wput = 2; /* /etc/systems */
229 229 int tcp_squeue_flag;
230 230
231 231 /*
232 232 * To prevent memory hog, limit the number of entries in tcp_free_list
233 233 * to 1% of available memory / number of cpus
234 234 */
235 235 uint_t tcp_free_list_max_cnt = 0;
236 236
237 237 #define TIDUSZ 4096 /* transport interface data unit size */
238 238
239 239 /*
240 240 * Size of acceptor hash list. It has to be a power of 2 for hashing.
241 241 */
242 242 #define TCP_ACCEPTOR_FANOUT_SIZE 512
243 243
244 244 #ifdef _ILP32
245 245 #define TCP_ACCEPTOR_HASH(accid) \
246 246 (((uint_t)(accid) >> 8) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
247 247 #else
248 248 #define TCP_ACCEPTOR_HASH(accid) \
249 249 ((uint_t)(accid) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
250 250 #endif /* _ILP32 */
251 251
252 252 /*
253 253 * Minimum number of connections which can be created per listener. Used
254 254 * when the listener connection count is in effect.
255 255 */
256 256 static uint32_t tcp_min_conn_listener = 2;
257 257
258 258 uint32_t tcp_early_abort = 30;
259 259
260 260 /* TCP Timer control structure */
261 261 typedef struct tcpt_s {
262 262 pfv_t tcpt_pfv; /* The routine we are to call */
263 263 tcp_t *tcpt_tcp; /* The parameter we are to pass in */
264 264 } tcpt_t;
265 265
266 266 /*
267 267 * Functions called directly via squeue having a prototype of edesc_t.
268 268 */
269 269 void tcp_input_listener(void *arg, mblk_t *mp, void *arg2,
270 270 ip_recv_attr_t *ira);
271 271 void tcp_input_data(void *arg, mblk_t *mp, void *arg2,
272 272 ip_recv_attr_t *ira);
273 273 static void tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2,
274 274 ip_recv_attr_t *dummy);
275 275
276 276
277 277 /* Prototype for TCP functions */
278 278 static void tcp_random_init(void);
279 279 int tcp_random(void);
280 280 static int tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp,
281 281 in_port_t dstport, uint_t srcid);
282 282 static int tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp,
283 283 in_port_t dstport, uint32_t flowinfo,
284 284 uint_t srcid, uint32_t scope_id);
285 285 static void tcp_iss_init(tcp_t *tcp);
286 286 static void tcp_reinit(tcp_t *tcp);
287 287 static void tcp_reinit_values(tcp_t *tcp);
288 288
289 289 static void tcp_wsrv(queue_t *q);
290 290 static void tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa);
291 291 static void tcp_update_zcopy(tcp_t *tcp);
292 292 static void tcp_notify(void *, ip_xmit_attr_t *, ixa_notify_type_t,
293 293 ixa_notify_arg_t);
294 294 static void *tcp_stack_init(netstackid_t stackid, netstack_t *ns);
295 295 static void tcp_stack_fini(netstackid_t stackid, void *arg);
296 296
297 297 static int tcp_squeue_switch(int);
298 298
299 299 static int tcp_open(queue_t *, dev_t *, int, int, cred_t *, boolean_t);
300 300 static int tcp_openv4(queue_t *, dev_t *, int, int, cred_t *);
301 301 static int tcp_openv6(queue_t *, dev_t *, int, int, cred_t *);
302 302
303 303 static void tcp_squeue_add(squeue_t *);
304 304
305 305 struct module_info tcp_rinfo = {
306 306 TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, TCP_RECV_HIWATER, TCP_RECV_LOWATER
307 307 };
308 308
309 309 static struct module_info tcp_winfo = {
310 310 TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, 127, 16
311 311 };
312 312
313 313 /*
314 314 * Entry points for TCP as a device. The normal case which supports
315 315 * the TCP functionality.
316 316 * We have separate open functions for the /dev/tcp and /dev/tcp6 devices.
317 317 */
318 318 struct qinit tcp_rinitv4 = {
319 319 NULL, (pfi_t)tcp_rsrv, tcp_openv4, tcp_tpi_close, NULL, &tcp_rinfo
320 320 };
321 321
322 322 struct qinit tcp_rinitv6 = {
323 323 NULL, (pfi_t)tcp_rsrv, tcp_openv6, tcp_tpi_close, NULL, &tcp_rinfo
324 324 };
325 325
326 326 struct qinit tcp_winit = {
327 327 (pfi_t)tcp_wput, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
328 328 };
329 329
330 330 /* Initial entry point for TCP in socket mode. */
331 331 struct qinit tcp_sock_winit = {
332 332 (pfi_t)tcp_wput_sock, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
333 333 };
334 334
335 335 /* TCP entry point during fallback */
336 336 struct qinit tcp_fallback_sock_winit = {
337 337 (pfi_t)tcp_wput_fallback, NULL, NULL, NULL, NULL, &tcp_winfo
338 338 };
339 339
340 340 /*
341 341 * Entry points for TCP as a acceptor STREAM opened by sockfs when doing
342 342 * an accept. Avoid allocating data structures since eager has already
343 343 * been created.
344 344 */
345 345 struct qinit tcp_acceptor_rinit = {
346 346 NULL, (pfi_t)tcp_rsrv, NULL, tcp_tpi_close_accept, NULL, &tcp_winfo
347 347 };
348 348
349 349 struct qinit tcp_acceptor_winit = {
350 350 (pfi_t)tcp_tpi_accept, NULL, NULL, NULL, NULL, &tcp_winfo
351 351 };
352 352
353 353 /* For AF_INET aka /dev/tcp */
354 354 struct streamtab tcpinfov4 = {
355 355 &tcp_rinitv4, &tcp_winit
356 356 };
357 357
358 358 /* For AF_INET6 aka /dev/tcp6 */
359 359 struct streamtab tcpinfov6 = {
360 360 &tcp_rinitv6, &tcp_winit
361 361 };
362 362
363 363 /*
364 364 * Following assumes TPI alignment requirements stay along 32 bit
365 365 * boundaries
366 366 */
367 367 #define ROUNDUP32(x) \
368 368 (((x) + (sizeof (int32_t) - 1)) & ~(sizeof (int32_t) - 1))
369 369
370 370 /* Template for response to info request. */
371 371 struct T_info_ack tcp_g_t_info_ack = {
372 372 T_INFO_ACK, /* PRIM_type */
373 373 0, /* TSDU_size */
374 374 T_INFINITE, /* ETSDU_size */
375 375 T_INVALID, /* CDATA_size */
376 376 T_INVALID, /* DDATA_size */
377 377 sizeof (sin_t), /* ADDR_size */
378 378 0, /* OPT_size - not initialized here */
379 379 TIDUSZ, /* TIDU_size */
380 380 T_COTS_ORD, /* SERV_type */
381 381 TCPS_IDLE, /* CURRENT_state */
382 382 (XPG4_1|EXPINLINE) /* PROVIDER_flag */
383 383 };
384 384
385 385 struct T_info_ack tcp_g_t_info_ack_v6 = {
386 386 T_INFO_ACK, /* PRIM_type */
387 387 0, /* TSDU_size */
388 388 T_INFINITE, /* ETSDU_size */
389 389 T_INVALID, /* CDATA_size */
390 390 T_INVALID, /* DDATA_size */
391 391 sizeof (sin6_t), /* ADDR_size */
392 392 0, /* OPT_size - not initialized here */
393 393 TIDUSZ, /* TIDU_size */
394 394 T_COTS_ORD, /* SERV_type */
395 395 TCPS_IDLE, /* CURRENT_state */
396 396 (XPG4_1|EXPINLINE) /* PROVIDER_flag */
397 397 };
398 398
399 399 /*
400 400 * TCP tunables related declarations. Definitions are in tcp_tunables.c
401 401 */
402 402 extern mod_prop_info_t tcp_propinfo_tbl[];
403 403 extern int tcp_propinfo_count;
404 404
405 405 #define IS_VMLOANED_MBLK(mp) \
406 406 (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0)
407 407
408 408 uint32_t do_tcpzcopy = 1; /* 0: disable, 1: enable, 2: force */
409 409
410 410 /*
411 411 * Forces all connections to obey the value of the tcps_maxpsz_multiplier
412 412 * tunable settable via NDD. Otherwise, the per-connection behavior is
413 413 * determined dynamically during tcp_set_destination(), which is the default.
414 414 */
415 415 boolean_t tcp_static_maxpsz = B_FALSE;
416 416
417 417 /*
418 418 * If the receive buffer size is changed, this function is called to update
419 419 * the upper socket layer on the new delayed receive wake up threshold.
420 420 */
421 421 static void
422 422 tcp_set_recv_threshold(tcp_t *tcp, uint32_t new_rcvthresh)
423 423 {
424 424 uint32_t default_threshold = SOCKET_RECVHIWATER >> 3;
425 425
426 426 if (IPCL_IS_NONSTR(tcp->tcp_connp)) {
427 427 conn_t *connp = tcp->tcp_connp;
428 428 struct sock_proto_props sopp;
429 429
430 430 /*
431 431 * only increase rcvthresh upto default_threshold
432 432 */
433 433 if (new_rcvthresh > default_threshold)
434 434 new_rcvthresh = default_threshold;
435 435
436 436 sopp.sopp_flags = SOCKOPT_RCVTHRESH;
437 437 sopp.sopp_rcvthresh = new_rcvthresh;
438 438
439 439 (*connp->conn_upcalls->su_set_proto_props)
440 440 (connp->conn_upper_handle, &sopp);
441 441 }
442 442 }
443 443
444 444 /*
445 445 * Figure out the value of window scale opton. Note that the rwnd is
446 446 * ASSUMED to be rounded up to the nearest MSS before the calculation.
447 447 * We cannot find the scale value and then do a round up of tcp_rwnd
448 448 * because the scale value may not be correct after that.
449 449 *
450 450 * Set the compiler flag to make this function inline.
451 451 */
452 452 void
453 453 tcp_set_ws_value(tcp_t *tcp)
454 454 {
455 455 int i;
456 456 uint32_t rwnd = tcp->tcp_rwnd;
457 457
458 458 for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT;
459 459 i++, rwnd >>= 1)
460 460 ;
461 461 tcp->tcp_rcv_ws = i;
462 462 }
463 463
464 464 /*
465 465 * Remove cached/latched IPsec references.
466 466 */
467 467 void
468 468 tcp_ipsec_cleanup(tcp_t *tcp)
469 469 {
470 470 conn_t *connp = tcp->tcp_connp;
471 471
472 472 ASSERT(connp->conn_flags & IPCL_TCPCONN);
473 473
474 474 if (connp->conn_latch != NULL) {
475 475 IPLATCH_REFRELE(connp->conn_latch);
476 476 connp->conn_latch = NULL;
477 477 }
478 478 if (connp->conn_latch_in_policy != NULL) {
479 479 IPPOL_REFRELE(connp->conn_latch_in_policy);
480 480 connp->conn_latch_in_policy = NULL;
481 481 }
482 482 if (connp->conn_latch_in_action != NULL) {
483 483 IPACT_REFRELE(connp->conn_latch_in_action);
484 484 connp->conn_latch_in_action = NULL;
485 485 }
486 486 if (connp->conn_policy != NULL) {
487 487 IPPH_REFRELE(connp->conn_policy, connp->conn_netstack);
488 488 connp->conn_policy = NULL;
489 489 }
490 490 }
491 491
492 492 /*
493 493 * Cleaup before placing on free list.
494 494 * Disassociate from the netstack/tcp_stack_t since the freelist
495 495 * is per squeue and not per netstack.
496 496 */
497 497 void
498 498 tcp_cleanup(tcp_t *tcp)
499 499 {
500 500 mblk_t *mp;
501 501 conn_t *connp = tcp->tcp_connp;
502 502 tcp_stack_t *tcps = tcp->tcp_tcps;
503 503 netstack_t *ns = tcps->tcps_netstack;
504 504 mblk_t *tcp_rsrv_mp;
505 505
506 506 tcp_bind_hash_remove(tcp);
507 507
508 508 /* Cleanup that which needs the netstack first */
509 509 tcp_ipsec_cleanup(tcp);
510 510 ixa_cleanup(connp->conn_ixa);
511 511
512 512 if (connp->conn_ht_iphc != NULL) {
513 513 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
514 514 connp->conn_ht_iphc = NULL;
515 515 connp->conn_ht_iphc_allocated = 0;
516 516 connp->conn_ht_iphc_len = 0;
517 517 connp->conn_ht_ulp = NULL;
518 518 connp->conn_ht_ulp_len = 0;
519 519 tcp->tcp_ipha = NULL;
520 520 tcp->tcp_ip6h = NULL;
521 521 tcp->tcp_tcpha = NULL;
522 522 }
523 523
524 524 /* We clear any IP_OPTIONS and extension headers */
525 525 ip_pkt_free(&connp->conn_xmit_ipp);
526 526
527 527 tcp_free(tcp);
528 528
529 529 /*
530 530 * Since we will bzero the entire structure, we need to
531 531 * remove it and reinsert it in global hash list. We
532 532 * know the walkers can't get to this conn because we
533 533 * had set CONDEMNED flag earlier and checked reference
534 534 * under conn_lock so walker won't pick it and when we
535 535 * go the ipcl_globalhash_remove() below, no walker
536 536 * can get to it.
537 537 */
538 538 ipcl_globalhash_remove(connp);
539 539
540 540 /* Save some state */
541 541 mp = tcp->tcp_timercache;
542 542
543 543 tcp_rsrv_mp = tcp->tcp_rsrv_mp;
544 544
545 545 if (connp->conn_cred != NULL) {
546 546 crfree(connp->conn_cred);
547 547 connp->conn_cred = NULL;
548 548 }
549 549 ipcl_conn_cleanup(connp);
550 550 connp->conn_flags = IPCL_TCPCONN;
551 551
552 552 /*
553 553 * Now it is safe to decrement the reference counts.
554 554 * This might be the last reference on the netstack
555 555 * in which case it will cause the freeing of the IP Instance.
556 556 */
557 557 connp->conn_netstack = NULL;
558 558 connp->conn_ixa->ixa_ipst = NULL;
559 559 netstack_rele(ns);
560 560 ASSERT(tcps != NULL);
561 561 tcp->tcp_tcps = NULL;
562 562
563 563 bzero(tcp, sizeof (tcp_t));
564 564
565 565 /* restore the state */
566 566 tcp->tcp_timercache = mp;
567 567
568 568 tcp->tcp_rsrv_mp = tcp_rsrv_mp;
569 569
570 570 tcp->tcp_connp = connp;
571 571
572 572 ASSERT(connp->conn_tcp == tcp);
573 573 ASSERT(connp->conn_flags & IPCL_TCPCONN);
574 574 connp->conn_state_flags = CONN_INCIPIENT;
575 575 ASSERT(connp->conn_proto == IPPROTO_TCP);
576 576 ASSERT(connp->conn_ref == 1);
577 577 }
578 578
579 579 /*
580 580 * Adapt to the information, such as rtt and rtt_sd, provided from the
581 581 * DCE and IRE maintained by IP.
582 582 *
583 583 * Checks for multicast and broadcast destination address.
584 584 * Returns zero if ok; an errno on failure.
585 585 *
586 586 * Note that the MSS calculation here is based on the info given in
587 587 * the DCE and IRE. We do not do any calculation based on TCP options. They
588 588 * will be handled in tcp_input_data() when TCP knows which options to use.
589 589 *
590 590 * Note on how TCP gets its parameters for a connection.
591 591 *
592 592 * When a tcp_t structure is allocated, it gets all the default parameters.
593 593 * In tcp_set_destination(), it gets those metric parameters, like rtt, rtt_sd,
594 594 * spipe, rpipe, ... from the route metrics. Route metric overrides the
595 595 * default.
596 596 *
597 597 * An incoming SYN with a multicast or broadcast destination address is dropped
598 598 * in ip_fanout_v4/v6.
599 599 *
600 600 * An incoming SYN with a multicast or broadcast source address is always
601 601 * dropped in tcp_set_destination, since IPDF_ALLOW_MCBC is not set in
602 602 * conn_connect.
603 603 * The same logic in tcp_set_destination also serves to
604 604 * reject an attempt to connect to a broadcast or multicast (destination)
605 605 * address.
606 606 */
607 607 int
608 608 tcp_set_destination(tcp_t *tcp)
609 609 {
610 610 uint32_t mss_max;
611 611 uint32_t mss;
612 612 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
613 613 conn_t *connp = tcp->tcp_connp;
614 614 tcp_stack_t *tcps = tcp->tcp_tcps;
615 615 iulp_t uinfo;
616 616 int error;
617 617 uint32_t flags;
618 618
619 619 flags = IPDF_LSO | IPDF_ZCOPY;
620 620 /*
621 621 * Make sure we have a dce for the destination to avoid dce_ident
622 622 * contention for connected sockets.
623 623 */
624 624 flags |= IPDF_UNIQUE_DCE;
625 625
626 626 if (!tcps->tcps_ignore_path_mtu)
627 627 connp->conn_ixa->ixa_flags |= IXAF_PMTU_DISCOVERY;
628 628
629 629 /* Use conn_lock to satify ASSERT; tcp is already serialized */
630 630 mutex_enter(&connp->conn_lock);
631 631 error = conn_connect(connp, &uinfo, flags);
632 632 mutex_exit(&connp->conn_lock);
633 633 if (error != 0)
634 634 return (error);
635 635
636 636 error = tcp_build_hdrs(tcp);
637 637 if (error != 0)
638 638 return (error);
639 639
640 640 tcp->tcp_localnet = uinfo.iulp_localnet;
641 641
642 642 if (uinfo.iulp_rtt != 0) {
643 643 clock_t rto;
644 644
645 645 tcp->tcp_rtt_sa = uinfo.iulp_rtt;
646 646 tcp->tcp_rtt_sd = uinfo.iulp_rtt_sd;
647 647 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
648 648 tcps->tcps_rexmit_interval_extra +
649 649 (tcp->tcp_rtt_sa >> 5);
650 650
651 651 TCP_SET_RTO(tcp, rto);
652 652 }
653 653 if (uinfo.iulp_ssthresh != 0)
654 654 tcp->tcp_cwnd_ssthresh = uinfo.iulp_ssthresh;
655 655 else
656 656 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
657 657 if (uinfo.iulp_spipe > 0) {
658 658 connp->conn_sndbuf = MIN(uinfo.iulp_spipe,
659 659 tcps->tcps_max_buf);
660 660 if (tcps->tcps_snd_lowat_fraction != 0) {
661 661 connp->conn_sndlowat = connp->conn_sndbuf /
662 662 tcps->tcps_snd_lowat_fraction;
663 663 }
664 664 (void) tcp_maxpsz_set(tcp, B_TRUE);
665 665 }
666 666 /*
667 667 * Note that up till now, acceptor always inherits receive
668 668 * window from the listener. But if there is a metrics
669 669 * associated with a host, we should use that instead of
670 670 * inheriting it from listener. Thus we need to pass this
671 671 * info back to the caller.
672 672 */
673 673 if (uinfo.iulp_rpipe > 0) {
674 674 tcp->tcp_rwnd = MIN(uinfo.iulp_rpipe,
675 675 tcps->tcps_max_buf);
676 676 }
677 677
678 678 if (uinfo.iulp_rtomax > 0) {
679 679 tcp->tcp_second_timer_threshold =
680 680 uinfo.iulp_rtomax;
681 681 }
682 682
683 683 /*
684 684 * Use the metric option settings, iulp_tstamp_ok and
685 685 * iulp_wscale_ok, only for active open. What this means
686 686 * is that if the other side uses timestamp or window
687 687 * scale option, TCP will also use those options. That
688 688 * is for passive open. If the application sets a
689 689 * large window, window scale is enabled regardless of
690 690 * the value in iulp_wscale_ok. This is the behavior
691 691 * since 2.6. So we keep it.
692 692 * The only case left in passive open processing is the
693 693 * check for SACK.
694 694 * For ECN, it should probably be like SACK. But the
695 695 * current value is binary, so we treat it like the other
696 696 * cases. The metric only controls active open.For passive
697 697 * open, the ndd param, tcp_ecn_permitted, controls the
698 698 * behavior.
699 699 */
700 700 if (!tcp_detached) {
701 701 /*
702 702 * The if check means that the following can only
703 703 * be turned on by the metrics only IRE, but not off.
704 704 */
705 705 if (uinfo.iulp_tstamp_ok)
706 706 tcp->tcp_snd_ts_ok = B_TRUE;
707 707 if (uinfo.iulp_wscale_ok)
708 708 tcp->tcp_snd_ws_ok = B_TRUE;
709 709 if (uinfo.iulp_sack == 2)
710 710 tcp->tcp_snd_sack_ok = B_TRUE;
711 711 if (uinfo.iulp_ecn_ok)
712 712 tcp->tcp_ecn_ok = B_TRUE;
713 713 } else {
714 714 /*
715 715 * Passive open.
716 716 *
717 717 * As above, the if check means that SACK can only be
718 718 * turned on by the metric only IRE.
719 719 */
720 720 if (uinfo.iulp_sack > 0) {
721 721 tcp->tcp_snd_sack_ok = B_TRUE;
722 722 }
723 723 }
724 724
725 725 /*
726 726 * XXX Note that currently, iulp_mtu can be as small as 68
727 727 * because of PMTUd. So tcp_mss may go to negative if combined
728 728 * length of all those options exceeds 28 bytes. But because
729 729 * of the tcp_mss_min check below, we may not have a problem if
730 730 * tcp_mss_min is of a reasonable value. The default is 1 so
731 731 * the negative problem still exists. And the check defeats PMTUd.
732 732 * In fact, if PMTUd finds that the MSS should be smaller than
733 733 * tcp_mss_min, TCP should turn off PMUTd and use the tcp_mss_min
734 734 * value.
735 735 *
736 736 * We do not deal with that now. All those problems related to
737 737 * PMTUd will be fixed later.
738 738 */
739 739 ASSERT(uinfo.iulp_mtu != 0);
740 740 mss = tcp->tcp_initial_pmtu = uinfo.iulp_mtu;
741 741
742 742 /* Sanity check for MSS value. */
743 743 if (connp->conn_ipversion == IPV4_VERSION)
744 744 mss_max = tcps->tcps_mss_max_ipv4;
745 745 else
746 746 mss_max = tcps->tcps_mss_max_ipv6;
747 747
748 748 if (tcp->tcp_ipsec_overhead == 0)
749 749 tcp->tcp_ipsec_overhead = conn_ipsec_length(connp);
750 750
751 751 mss -= tcp->tcp_ipsec_overhead;
752 752
753 753 if (mss < tcps->tcps_mss_min)
754 754 mss = tcps->tcps_mss_min;
755 755 if (mss > mss_max)
756 756 mss = mss_max;
757 757
758 758 /* Note that this is the maximum MSS, excluding all options. */
759 759 tcp->tcp_mss = mss;
760 760
761 761 /*
762 762 * Update the tcp connection with LSO capability.
763 763 */
764 764 tcp_update_lso(tcp, connp->conn_ixa);
765 765
766 766 /*
767 767 * Initialize the ISS here now that we have the full connection ID.
768 768 * The RFC 1948 method of initial sequence number generation requires
769 769 * knowledge of the full connection ID before setting the ISS.
770 770 */
771 771 tcp_iss_init(tcp);
772 772
773 773 tcp->tcp_loopback = (uinfo.iulp_loopback | uinfo.iulp_local);
774 774
775 775 /*
776 776 * Make sure that conn is not marked incipient
777 777 * for incoming connections. A blind
778 778 * removal of incipient flag is cheaper than
779 779 * check and removal.
780 780 */
781 781 mutex_enter(&connp->conn_lock);
782 782 connp->conn_state_flags &= ~CONN_INCIPIENT;
783 783 mutex_exit(&connp->conn_lock);
784 784 return (0);
785 785 }
786 786
787 787 /*
788 788 * tcp_clean_death / tcp_close_detached must not be called more than once
789 789 * on a tcp. Thus every function that potentially calls tcp_clean_death
790 790 * must check for the tcp state before calling tcp_clean_death.
791 791 * Eg. tcp_input_data, tcp_eager_kill, tcp_clean_death_wrapper,
792 792 * tcp_timer_handler, all check for the tcp state.
793 793 */
794 794 /* ARGSUSED */
795 795 void
796 796 tcp_clean_death_wrapper(void *arg, mblk_t *mp, void *arg2,
797 797 ip_recv_attr_t *dummy)
798 798 {
799 799 tcp_t *tcp = ((conn_t *)arg)->conn_tcp;
800 800
801 801 freemsg(mp);
802 802 if (tcp->tcp_state > TCPS_BOUND)
803 803 (void) tcp_clean_death(((conn_t *)arg)->conn_tcp, ETIMEDOUT);
804 804 }
805 805
806 806 /*
807 807 * We are dying for some reason. Try to do it gracefully. (May be called
808 808 * as writer.)
809 809 *
810 810 * Return -1 if the structure was not cleaned up (if the cleanup had to be
811 811 * done by a service procedure).
812 812 * TBD - Should the return value distinguish between the tcp_t being
813 813 * freed and it being reinitialized?
814 814 */
815 815 int
816 816 tcp_clean_death(tcp_t *tcp, int err)
817 817 {
818 818 mblk_t *mp;
819 819 queue_t *q;
820 820 conn_t *connp = tcp->tcp_connp;
821 821 tcp_stack_t *tcps = tcp->tcp_tcps;
822 822
823 823 if (tcp->tcp_fused)
824 824 tcp_unfuse(tcp);
825 825
826 826 if (tcp->tcp_linger_tid != 0 &&
827 827 TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
828 828 tcp_stop_lingering(tcp);
829 829 }
830 830
831 831 ASSERT(tcp != NULL);
832 832 ASSERT((connp->conn_family == AF_INET &&
833 833 connp->conn_ipversion == IPV4_VERSION) ||
834 834 (connp->conn_family == AF_INET6 &&
835 835 (connp->conn_ipversion == IPV4_VERSION ||
836 836 connp->conn_ipversion == IPV6_VERSION)));
837 837
838 838 if (TCP_IS_DETACHED(tcp)) {
839 839 if (tcp->tcp_hard_binding) {
840 840 /*
841 841 * Its an eager that we are dealing with. We close the
842 842 * eager but in case a conn_ind has already gone to the
843 843 * listener, let tcp_accept_finish() send a discon_ind
844 844 * to the listener and drop the last reference. If the
845 845 * listener doesn't even know about the eager i.e. the
846 846 * conn_ind hasn't gone up, blow away the eager and drop
847 847 * the last reference as well. If the conn_ind has gone
848 848 * up, state should be BOUND. tcp_accept_finish
849 849 * will figure out that the connection has received a
850 850 * RST and will send a DISCON_IND to the application.
851 851 */
852 852 tcp_closei_local(tcp);
853 853 if (!tcp->tcp_tconnind_started) {
854 854 CONN_DEC_REF(connp);
855 855 } else {
856 856 tcp->tcp_state = TCPS_BOUND;
857 857 DTRACE_TCP6(state__change, void, NULL,
858 858 ip_xmit_attr_t *, connp->conn_ixa,
859 859 void, NULL, tcp_t *, tcp, void, NULL,
860 860 int32_t, TCPS_CLOSED);
861 861 }
862 862 } else {
863 863 tcp_close_detached(tcp);
864 864 }
865 865 return (0);
866 866 }
867 867
868 868 TCP_STAT(tcps, tcp_clean_death_nondetached);
869 869
870 870 /*
871 871 * The connection is dead. Decrement listener connection counter if
872 872 * necessary.
873 873 */
874 874 if (tcp->tcp_listen_cnt != NULL)
875 875 TCP_DECR_LISTEN_CNT(tcp);
876 876
877 877 /*
878 878 * When a connection is moved to TIME_WAIT state, the connection
879 879 * counter is already decremented. So no need to decrement here
880 880 * again. See SET_TIME_WAIT() macro.
881 881 */
882 882 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
883 883 tcp->tcp_state < TCPS_TIME_WAIT) {
884 884 TCPS_CONN_DEC(tcps);
885 885 }
886 886
887 887 q = connp->conn_rq;
888 888
889 889 /* Trash all inbound data */
890 890 if (!IPCL_IS_NONSTR(connp)) {
891 891 ASSERT(q != NULL);
892 892 flushq(q, FLUSHALL);
893 893 }
894 894
895 895 /*
896 896 * If we are at least part way open and there is error
897 897 * (err==0 implies no error)
898 898 * notify our client by a T_DISCON_IND.
899 899 */
900 900 if ((tcp->tcp_state >= TCPS_SYN_SENT) && err) {
901 901 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
902 902 !TCP_IS_SOCKET(tcp)) {
903 903 /*
904 904 * Send M_FLUSH according to TPI. Because sockets will
905 905 * (and must) ignore FLUSHR we do that only for TPI
906 906 * endpoints and sockets in STREAMS mode.
907 907 */
908 908 (void) putnextctl1(q, M_FLUSH, FLUSHR);
909 909 }
910 910 if (connp->conn_debug) {
911 911 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE|SL_ERROR,
912 912 "tcp_clean_death: discon err %d", err);
913 913 }
914 914 if (IPCL_IS_NONSTR(connp)) {
915 915 /* Direct socket, use upcall */
916 916 (*connp->conn_upcalls->su_disconnected)(
917 917 connp->conn_upper_handle, tcp->tcp_connid, err);
918 918 } else {
919 919 mp = mi_tpi_discon_ind(NULL, err, 0);
920 920 if (mp != NULL) {
921 921 putnext(q, mp);
922 922 } else {
923 923 if (connp->conn_debug) {
924 924 (void) strlog(TCP_MOD_ID, 0, 1,
925 925 SL_ERROR|SL_TRACE,
926 926 "tcp_clean_death, sending M_ERROR");
927 927 }
928 928 (void) putnextctl1(q, M_ERROR, EPROTO);
929 929 }
930 930 }
931 931 if (tcp->tcp_state <= TCPS_SYN_RCVD) {
932 932 /* SYN_SENT or SYN_RCVD */
933 933 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
934 934 } else if (tcp->tcp_state <= TCPS_CLOSE_WAIT) {
935 935 /* ESTABLISHED or CLOSE_WAIT */
936 936 TCPS_BUMP_MIB(tcps, tcpEstabResets);
937 937 }
938 938 }
939 939
940 940 /*
941 941 * ESTABLISHED non-STREAMS eagers are not 'detached' because
942 942 * an upper handle is obtained when the SYN-ACK comes in. So it
943 943 * should receive the 'disconnected' upcall, but tcp_reinit should
944 944 * not be called since this is an eager.
945 945 */
946 946 if (tcp->tcp_listener != NULL && IPCL_IS_NONSTR(connp)) {
947 947 tcp_closei_local(tcp);
948 948 tcp->tcp_state = TCPS_BOUND;
949 949 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
950 950 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
951 951 int32_t, TCPS_CLOSED);
952 952 return (0);
953 953 }
954 954
955 955 tcp_reinit(tcp);
956 956 if (IPCL_IS_NONSTR(connp))
957 957 (void) tcp_do_unbind(connp);
958 958
959 959 return (-1);
960 960 }
961 961
962 962 /*
963 963 * In case tcp is in the "lingering state" and waits for the SO_LINGER timeout
964 964 * to expire, stop the wait and finish the close.
965 965 */
966 966 void
967 967 tcp_stop_lingering(tcp_t *tcp)
968 968 {
969 969 clock_t delta = 0;
970 970 tcp_stack_t *tcps = tcp->tcp_tcps;
971 971 conn_t *connp = tcp->tcp_connp;
972 972
973 973 tcp->tcp_linger_tid = 0;
974 974 if (tcp->tcp_state > TCPS_LISTEN) {
975 975 tcp_acceptor_hash_remove(tcp);
976 976 mutex_enter(&tcp->tcp_non_sq_lock);
977 977 if (tcp->tcp_flow_stopped) {
978 978 tcp_clrqfull(tcp);
979 979 }
980 980 mutex_exit(&tcp->tcp_non_sq_lock);
981 981
982 982 if (tcp->tcp_timer_tid != 0) {
983 983 delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
984 984 tcp->tcp_timer_tid = 0;
985 985 }
986 986 /*
987 987 * Need to cancel those timers which will not be used when
988 988 * TCP is detached. This has to be done before the conn_wq
989 989 * is cleared.
990 990 */
991 991 tcp_timers_stop(tcp);
992 992
993 993 tcp->tcp_detached = B_TRUE;
994 994 connp->conn_rq = NULL;
995 995 connp->conn_wq = NULL;
996 996
997 997 if (tcp->tcp_state == TCPS_TIME_WAIT) {
998 998 tcp_time_wait_append(tcp);
999 999 TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1000 1000 goto finish;
1001 1001 }
1002 1002
1003 1003 /*
1004 1004 * If delta is zero the timer event wasn't executed and was
1005 1005 * successfully canceled. In this case we need to restart it
1006 1006 * with the minimal delta possible.
1007 1007 */
1008 1008 if (delta >= 0) {
1009 1009 tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1010 1010 delta ? delta : 1);
1011 1011 }
1012 1012 } else {
1013 1013 tcp_closei_local(tcp);
1014 1014 CONN_DEC_REF(connp);
1015 1015 }
1016 1016 finish:
1017 1017 tcp->tcp_detached = B_TRUE;
1018 1018 connp->conn_rq = NULL;
1019 1019 connp->conn_wq = NULL;
1020 1020
1021 1021 /* Signal closing thread that it can complete close */
1022 1022 mutex_enter(&tcp->tcp_closelock);
1023 1023 tcp->tcp_closed = 1;
1024 1024 cv_signal(&tcp->tcp_closecv);
1025 1025 mutex_exit(&tcp->tcp_closelock);
1026 1026
1027 1027 /* If we have an upper handle (socket), release it */
1028 1028 if (IPCL_IS_NONSTR(connp)) {
1029 1029 ASSERT(connp->conn_upper_handle != NULL);
1030 1030 (*connp->conn_upcalls->su_closed)(connp->conn_upper_handle);
1031 1031 connp->conn_upper_handle = NULL;
1032 1032 connp->conn_upcalls = NULL;
1033 1033 }
1034 1034 }
1035 1035
1036 1036 void
1037 1037 tcp_close_common(conn_t *connp, int flags)
1038 1038 {
1039 1039 tcp_t *tcp = connp->conn_tcp;
1040 1040 mblk_t *mp = &tcp->tcp_closemp;
1041 1041 boolean_t conn_ioctl_cleanup_reqd = B_FALSE;
1042 1042 mblk_t *bp;
1043 1043
1044 1044 ASSERT(connp->conn_ref >= 2);
1045 1045
1046 1046 /*
1047 1047 * Mark the conn as closing. ipsq_pending_mp_add will not
1048 1048 * add any mp to the pending mp list, after this conn has
1049 1049 * started closing.
1050 1050 */
1051 1051 mutex_enter(&connp->conn_lock);
1052 1052 connp->conn_state_flags |= CONN_CLOSING;
1053 1053 if (connp->conn_oper_pending_ill != NULL)
1054 1054 conn_ioctl_cleanup_reqd = B_TRUE;
1055 1055 CONN_INC_REF_LOCKED(connp);
1056 1056 mutex_exit(&connp->conn_lock);
1057 1057 tcp->tcp_closeflags = (uint8_t)flags;
1058 1058 ASSERT(connp->conn_ref >= 3);
1059 1059
1060 1060 /*
1061 1061 * tcp_closemp_used is used below without any protection of a lock
1062 1062 * as we don't expect any one else to use it concurrently at this
1063 1063 * point otherwise it would be a major defect.
1064 1064 */
1065 1065
1066 1066 if (mp->b_prev == NULL)
1067 1067 tcp->tcp_closemp_used = B_TRUE;
1068 1068 else
1069 1069 cmn_err(CE_PANIC, "tcp_close: concurrent use of tcp_closemp: "
1070 1070 "connp %p tcp %p\n", (void *)connp, (void *)tcp);
1071 1071
1072 1072 TCP_DEBUG_GETPCSTACK(tcp->tcmp_stk, 15);
1073 1073
1074 1074 /*
1075 1075 * Cleanup any queued ioctls here. This must be done before the wq/rq
1076 1076 * are re-written by tcp_close_output().
1077 1077 */
1078 1078 if (conn_ioctl_cleanup_reqd)
1079 1079 conn_ioctl_cleanup(connp);
1080 1080
1081 1081 /*
1082 1082 * As CONN_CLOSING is set, no further ioctls should be passed down to
1083 1083 * IP for this conn (see the guards in tcp_ioctl, tcp_wput_ioctl and
1084 1084 * tcp_wput_iocdata). If the ioctl was queued on an ipsq,
1085 1085 * conn_ioctl_cleanup should have found it and removed it. If the ioctl
1086 1086 * was still in flight at the time, we wait for it here. See comments
1087 1087 * for CONN_INC_IOCTLREF in ip.h for details.
1088 1088 */
1089 1089 mutex_enter(&connp->conn_lock);
1090 1090 while (connp->conn_ioctlref > 0)
1091 1091 cv_wait(&connp->conn_cv, &connp->conn_lock);
1092 1092 ASSERT(connp->conn_ioctlref == 0);
1093 1093 ASSERT(connp->conn_oper_pending_ill == NULL);
1094 1094 mutex_exit(&connp->conn_lock);
1095 1095
1096 1096 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_close_output, connp,
1097 1097 NULL, tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1098 1098
1099 1099 /*
1100 1100 * For non-STREAMS sockets, the normal case is that the conn makes
1101 1101 * an upcall when it's finally closed, so there is no need to wait
1102 1102 * in the protocol. But in case of SO_LINGER the thread sleeps here
1103 1103 * so it can properly deal with the thread being interrupted.
1104 1104 */
1105 1105 if (IPCL_IS_NONSTR(connp) && connp->conn_linger == 0)
1106 1106 goto nowait;
1107 1107
1108 1108 mutex_enter(&tcp->tcp_closelock);
1109 1109 while (!tcp->tcp_closed) {
1110 1110 if (!cv_wait_sig(&tcp->tcp_closecv, &tcp->tcp_closelock)) {
1111 1111 /*
1112 1112 * The cv_wait_sig() was interrupted. We now do the
1113 1113 * following:
1114 1114 *
1115 1115 * 1) If the endpoint was lingering, we allow this
1116 1116 * to be interrupted by cancelling the linger timeout
1117 1117 * and closing normally.
1118 1118 *
1119 1119 * 2) Revert to calling cv_wait()
1120 1120 *
1121 1121 * We revert to using cv_wait() to avoid an
1122 1122 * infinite loop which can occur if the calling
1123 1123 * thread is higher priority than the squeue worker
1124 1124 * thread and is bound to the same cpu.
1125 1125 */
1126 1126 if (connp->conn_linger && connp->conn_lingertime > 0) {
1127 1127 mutex_exit(&tcp->tcp_closelock);
1128 1128 /* Entering squeue, bump ref count. */
1129 1129 CONN_INC_REF(connp);
1130 1130 bp = allocb_wait(0, BPRI_HI, STR_NOSIG, NULL);
1131 1131 SQUEUE_ENTER_ONE(connp->conn_sqp, bp,
1132 1132 tcp_linger_interrupted, connp, NULL,
1133 1133 tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1134 1134 mutex_enter(&tcp->tcp_closelock);
1135 1135 }
1136 1136 break;
1137 1137 }
1138 1138 }
1139 1139 while (!tcp->tcp_closed)
1140 1140 cv_wait(&tcp->tcp_closecv, &tcp->tcp_closelock);
1141 1141 mutex_exit(&tcp->tcp_closelock);
1142 1142
1143 1143 /*
1144 1144 * In the case of listener streams that have eagers in the q or q0
1145 1145 * we wait for the eagers to drop their reference to us. conn_rq and
1146 1146 * conn_wq of the eagers point to our queues. By waiting for the
1147 1147 * refcnt to drop to 1, we are sure that the eagers have cleaned
1148 1148 * up their queue pointers and also dropped their references to us.
1149 1149 *
1150 1150 * For non-STREAMS sockets we do not have to wait here; the
1151 1151 * listener will instead make a su_closed upcall when the last
1152 1152 * reference is dropped.
1153 1153 */
1154 1154 if (tcp->tcp_wait_for_eagers && !IPCL_IS_NONSTR(connp)) {
1155 1155 mutex_enter(&connp->conn_lock);
1156 1156 while (connp->conn_ref != 1) {
1157 1157 cv_wait(&connp->conn_cv, &connp->conn_lock);
1158 1158 }
1159 1159 mutex_exit(&connp->conn_lock);
1160 1160 }
1161 1161
1162 1162 nowait:
1163 1163 connp->conn_cpid = NOPID;
1164 1164 }
1165 1165
1166 1166 /*
1167 1167 * Called by tcp_close() routine via squeue when lingering is
1168 1168 * interrupted by a signal.
1169 1169 */
1170 1170
1171 1171 /* ARGSUSED */
1172 1172 static void
1173 1173 tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1174 1174 {
1175 1175 conn_t *connp = (conn_t *)arg;
1176 1176 tcp_t *tcp = connp->conn_tcp;
1177 1177
1178 1178 freeb(mp);
1179 1179 if (tcp->tcp_linger_tid != 0 &&
1180 1180 TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
1181 1181 tcp_stop_lingering(tcp);
1182 1182 tcp->tcp_client_errno = EINTR;
1183 1183 }
1184 1184 }
1185 1185
1186 1186 /*
1187 1187 * Clean up the b_next and b_prev fields of every mblk pointed at by *mpp.
1188 1188 * Some stream heads get upset if they see these later on as anything but NULL.
1189 1189 */
1190 1190 void
1191 1191 tcp_close_mpp(mblk_t **mpp)
1192 1192 {
1193 1193 mblk_t *mp;
1194 1194
1195 1195 if ((mp = *mpp) != NULL) {
1196 1196 do {
1197 1197 mp->b_next = NULL;
1198 1198 mp->b_prev = NULL;
1199 1199 } while ((mp = mp->b_cont) != NULL);
1200 1200
1201 1201 mp = *mpp;
1202 1202 *mpp = NULL;
1203 1203 freemsg(mp);
1204 1204 }
1205 1205 }
1206 1206
1207 1207 /* Do detached close. */
1208 1208 void
1209 1209 tcp_close_detached(tcp_t *tcp)
1210 1210 {
1211 1211 if (tcp->tcp_fused)
1212 1212 tcp_unfuse(tcp);
1213 1213
1214 1214 /*
1215 1215 * Clustering code serializes TCP disconnect callbacks and
1216 1216 * cluster tcp list walks by blocking a TCP disconnect callback
1217 1217 * if a cluster tcp list walk is in progress. This ensures
1218 1218 * accurate accounting of TCPs in the cluster code even though
1219 1219 * the TCP list walk itself is not atomic.
1220 1220 */
1221 1221 tcp_closei_local(tcp);
1222 1222 CONN_DEC_REF(tcp->tcp_connp);
1223 1223 }
1224 1224
1225 1225 /*
1226 1226 * The tcp_t is going away. Remove it from all lists and set it
1227 1227 * to TCPS_CLOSED. The freeing up of memory is deferred until
1228 1228 * tcp_inactive. This is needed since a thread in tcp_rput might have
1229 1229 * done a CONN_INC_REF on this structure before it was removed from the
1230 1230 * hashes.
1231 1231 */
1232 1232 void
1233 1233 tcp_closei_local(tcp_t *tcp)
1234 1234 {
1235 1235 conn_t *connp = tcp->tcp_connp;
1236 1236 tcp_stack_t *tcps = tcp->tcp_tcps;
1237 1237 int32_t oldstate;
1238 1238
1239 1239 if (!TCP_IS_SOCKET(tcp))
1240 1240 tcp_acceptor_hash_remove(tcp);
1241 1241
1242 1242 TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1243 1243 tcp->tcp_ibsegs = 0;
1244 1244 TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1245 1245 tcp->tcp_obsegs = 0;
1246 1246
1247 1247 /*
1248 1248 * This can be called via tcp_time_wait_processing() if TCP gets a
1249 1249 * SYN with sequence number outside the TIME-WAIT connection's
1250 1250 * window. So we need to check for TIME-WAIT state here as the
1251 1251 * connection counter is already decremented. See SET_TIME_WAIT()
1252 1252 * macro
1253 1253 */
1254 1254 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1255 1255 tcp->tcp_state < TCPS_TIME_WAIT) {
1256 1256 TCPS_CONN_DEC(tcps);
1257 1257 }
1258 1258
1259 1259 /*
1260 1260 * If we are an eager connection hanging off a listener that
1261 1261 * hasn't formally accepted the connection yet, get off his
1262 1262 * list and blow off any data that we have accumulated.
1263 1263 */
1264 1264 if (tcp->tcp_listener != NULL) {
1265 1265 tcp_t *listener = tcp->tcp_listener;
1266 1266 mutex_enter(&listener->tcp_eager_lock);
1267 1267 /*
1268 1268 * tcp_tconnind_started == B_TRUE means that the
1269 1269 * conn_ind has already gone to listener. At
1270 1270 * this point, eager will be closed but we
1271 1271 * leave it in listeners eager list so that
1272 1272 * if listener decides to close without doing
1273 1273 * accept, we can clean this up. In tcp_tli_accept
1274 1274 * we take care of the case of accept on closed
1275 1275 * eager.
1276 1276 */
1277 1277 if (!tcp->tcp_tconnind_started) {
1278 1278 tcp_eager_unlink(tcp);
1279 1279 mutex_exit(&listener->tcp_eager_lock);
1280 1280 /*
1281 1281 * We don't want to have any pointers to the
1282 1282 * listener queue, after we have released our
1283 1283 * reference on the listener
1284 1284 */
1285 1285 ASSERT(tcp->tcp_detached);
1286 1286 connp->conn_rq = NULL;
1287 1287 connp->conn_wq = NULL;
1288 1288 CONN_DEC_REF(listener->tcp_connp);
1289 1289 } else {
1290 1290 mutex_exit(&listener->tcp_eager_lock);
1291 1291 }
1292 1292 }
1293 1293
1294 1294 /* Stop all the timers */
1295 1295 tcp_timers_stop(tcp);
1296 1296
1297 1297 if (tcp->tcp_state == TCPS_LISTEN) {
1298 1298 if (tcp->tcp_ip_addr_cache) {
1299 1299 kmem_free((void *)tcp->tcp_ip_addr_cache,
1300 1300 IP_ADDR_CACHE_SIZE * sizeof (ipaddr_t));
1301 1301 tcp->tcp_ip_addr_cache = NULL;
1302 1302 }
1303 1303 }
1304 1304
1305 1305 /* Decrement listerner connection counter if necessary. */
1306 1306 if (tcp->tcp_listen_cnt != NULL)
1307 1307 TCP_DECR_LISTEN_CNT(tcp);
1308 1308
1309 1309 mutex_enter(&tcp->tcp_non_sq_lock);
1310 1310 if (tcp->tcp_flow_stopped)
1311 1311 tcp_clrqfull(tcp);
1312 1312 mutex_exit(&tcp->tcp_non_sq_lock);
1313 1313
1314 1314 tcp_bind_hash_remove(tcp);
1315 1315 /*
1316 1316 * If the tcp_time_wait_collector (which runs outside the squeue)
1317 1317 * is trying to remove this tcp from the time wait list, we will
1318 1318 * block in tcp_time_wait_remove while trying to acquire the
1319 1319 * tcp_time_wait_lock. The logic in tcp_time_wait_collector also
1320 1320 * requires the ipcl_hash_remove to be ordered after the
1321 1321 * tcp_time_wait_remove for the refcnt checks to work correctly.
1322 1322 */
1323 1323 if (tcp->tcp_state == TCPS_TIME_WAIT)
1324 1324 (void) tcp_time_wait_remove(tcp, NULL);
1325 1325 CL_INET_DISCONNECT(connp);
1326 1326 ipcl_hash_remove(connp);
1327 1327 oldstate = tcp->tcp_state;
1328 1328 tcp->tcp_state = TCPS_CLOSED;
1329 1329 /* Need to probe before ixa_cleanup() is called */
1330 1330 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1331 1331 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
1332 1332 int32_t, oldstate);
1333 1333 ixa_cleanup(connp->conn_ixa);
1334 1334
1335 1335 /*
1336 1336 * Mark the conn as CONDEMNED
1337 1337 */
1338 1338 mutex_enter(&connp->conn_lock);
1339 1339 connp->conn_state_flags |= CONN_CONDEMNED;
1340 1340 mutex_exit(&connp->conn_lock);
1341 1341
1342 1342 ASSERT(tcp->tcp_time_wait_next == NULL);
1343 1343 ASSERT(tcp->tcp_time_wait_prev == NULL);
1344 1344 ASSERT(tcp->tcp_time_wait_expire == 0);
1345 1345
1346 1346 tcp_ipsec_cleanup(tcp);
1347 1347 }
1348 1348
1349 1349 /*
1350 1350 * tcp is dying (called from ipcl_conn_destroy and error cases).
1351 1351 * Free the tcp_t in either case.
1352 1352 */
1353 1353 void
1354 1354 tcp_free(tcp_t *tcp)
1355 1355 {
1356 1356 mblk_t *mp;
1357 1357 conn_t *connp = tcp->tcp_connp;
1358 1358
1359 1359 ASSERT(tcp != NULL);
1360 1360 ASSERT(tcp->tcp_ptpahn == NULL && tcp->tcp_acceptor_hash == NULL);
1361 1361
1362 1362 connp->conn_rq = NULL;
1363 1363 connp->conn_wq = NULL;
1364 1364
1365 1365 tcp_close_mpp(&tcp->tcp_xmit_head);
1366 1366 tcp_close_mpp(&tcp->tcp_reass_head);
1367 1367 if (tcp->tcp_rcv_list != NULL) {
1368 1368 /* Free b_next chain */
1369 1369 tcp_close_mpp(&tcp->tcp_rcv_list);
1370 1370 }
1371 1371 if ((mp = tcp->tcp_urp_mp) != NULL) {
1372 1372 freemsg(mp);
1373 1373 }
1374 1374 if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1375 1375 freemsg(mp);
1376 1376 }
1377 1377
1378 1378 if (tcp->tcp_fused_sigurg_mp != NULL) {
1379 1379 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1380 1380 freeb(tcp->tcp_fused_sigurg_mp);
1381 1381 tcp->tcp_fused_sigurg_mp = NULL;
1382 1382 }
1383 1383
1384 1384 if (tcp->tcp_ordrel_mp != NULL) {
1385 1385 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1386 1386 freeb(tcp->tcp_ordrel_mp);
1387 1387 tcp->tcp_ordrel_mp = NULL;
1388 1388 }
1389 1389
1390 1390 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
1391 1391 bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
1392 1392
1393 1393 if (tcp->tcp_hopopts != NULL) {
1394 1394 mi_free(tcp->tcp_hopopts);
1395 1395 tcp->tcp_hopopts = NULL;
1396 1396 tcp->tcp_hopoptslen = 0;
1397 1397 }
1398 1398 ASSERT(tcp->tcp_hopoptslen == 0);
1399 1399 if (tcp->tcp_dstopts != NULL) {
1400 1400 mi_free(tcp->tcp_dstopts);
1401 1401 tcp->tcp_dstopts = NULL;
1402 1402 tcp->tcp_dstoptslen = 0;
1403 1403 }
1404 1404 ASSERT(tcp->tcp_dstoptslen == 0);
1405 1405 if (tcp->tcp_rthdrdstopts != NULL) {
1406 1406 mi_free(tcp->tcp_rthdrdstopts);
1407 1407 tcp->tcp_rthdrdstopts = NULL;
1408 1408 tcp->tcp_rthdrdstoptslen = 0;
1409 1409 }
1410 1410 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
1411 1411 if (tcp->tcp_rthdr != NULL) {
1412 1412 mi_free(tcp->tcp_rthdr);
1413 1413 tcp->tcp_rthdr = NULL;
1414 1414 tcp->tcp_rthdrlen = 0;
1415 1415 }
1416 1416 ASSERT(tcp->tcp_rthdrlen == 0);
1417 1417
1418 1418 /*
1419 1419 * Following is really a blowing away a union.
1420 1420 * It happens to have exactly two members of identical size
1421 1421 * the following code is enough.
1422 1422 */
1423 1423 tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1424 1424
1425 1425 /*
1426 1426 * If this is a non-STREAM socket still holding on to an upper
1427 1427 * handle, release it. As a result of fallback we might also see
1428 1428 * STREAMS based conns with upper handles, in which case there is
1429 1429 * nothing to do other than clearing the field.
1430 1430 */
1431 1431 if (connp->conn_upper_handle != NULL) {
1432 1432 if (IPCL_IS_NONSTR(connp)) {
1433 1433 (*connp->conn_upcalls->su_closed)(
1434 1434 connp->conn_upper_handle);
1435 1435 tcp->tcp_detached = B_TRUE;
1436 1436 }
1437 1437 connp->conn_upper_handle = NULL;
1438 1438 connp->conn_upcalls = NULL;
1439 1439 }
1440 1440 }
1441 1441
1442 1442 /*
1443 1443 * tcp_get_conn/tcp_free_conn
1444 1444 *
1445 1445 * tcp_get_conn is used to get a clean tcp connection structure.
1446 1446 * It tries to reuse the connections put on the freelist by the
1447 1447 * time_wait_collector failing which it goes to kmem_cache. This
1448 1448 * way has two benefits compared to just allocating from and
1449 1449 * freeing to kmem_cache.
1450 1450 * 1) The time_wait_collector can free (which includes the cleanup)
1451 1451 * outside the squeue. So when the interrupt comes, we have a clean
1452 1452 * connection sitting in the freelist. Obviously, this buys us
1453 1453 * performance.
1454 1454 *
1455 1455 * 2) Defence against DOS attack. Allocating a tcp/conn in tcp_input_listener
1456 1456 * has multiple disadvantages - tying up the squeue during alloc.
1457 1457 * But allocating the conn/tcp in IP land is also not the best since
1458 1458 * we can't check the 'q' and 'q0' which are protected by squeue and
1459 1459 * blindly allocate memory which might have to be freed here if we are
1460 1460 * not allowed to accept the connection. By using the freelist and
1461 1461 * putting the conn/tcp back in freelist, we don't pay a penalty for
1462 1462 * allocating memory without checking 'q/q0' and freeing it if we can't
1463 1463 * accept the connection.
1464 1464 *
1465 1465 * Care should be taken to put the conn back in the same squeue's freelist
1466 1466 * from which it was allocated. Best results are obtained if conn is
1467 1467 * allocated from listener's squeue and freed to the same. Time wait
1468 1468 * collector will free up the freelist is the connection ends up sitting
1469 1469 * there for too long.
1470 1470 */
1471 1471 void *
1472 1472 tcp_get_conn(void *arg, tcp_stack_t *tcps)
1473 1473 {
1474 1474 tcp_t *tcp = NULL;
1475 1475 conn_t *connp = NULL;
1476 1476 squeue_t *sqp = (squeue_t *)arg;
1477 1477 tcp_squeue_priv_t *tcp_time_wait;
1478 1478 netstack_t *ns;
1479 1479 mblk_t *tcp_rsrv_mp = NULL;
1480 1480
1481 1481 tcp_time_wait =
1482 1482 *((tcp_squeue_priv_t **)squeue_getprivate(sqp, SQPRIVATE_TCP));
1483 1483
1484 1484 mutex_enter(&tcp_time_wait->tcp_time_wait_lock);
1485 1485 tcp = tcp_time_wait->tcp_free_list;
1486 1486 ASSERT((tcp != NULL) ^ (tcp_time_wait->tcp_free_list_cnt == 0));
1487 1487 if (tcp != NULL) {
1488 1488 tcp_time_wait->tcp_free_list = tcp->tcp_time_wait_next;
1489 1489 tcp_time_wait->tcp_free_list_cnt--;
1490 1490 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1491 1491 tcp->tcp_time_wait_next = NULL;
1492 1492 connp = tcp->tcp_connp;
1493 1493 connp->conn_flags |= IPCL_REUSED;
1494 1494
1495 1495 ASSERT(tcp->tcp_tcps == NULL);
1496 1496 ASSERT(connp->conn_netstack == NULL);
1497 1497 ASSERT(tcp->tcp_rsrv_mp != NULL);
1498 1498 ns = tcps->tcps_netstack;
1499 1499 netstack_hold(ns);
1500 1500 connp->conn_netstack = ns;
1501 1501 connp->conn_ixa->ixa_ipst = ns->netstack_ip;
1502 1502 tcp->tcp_tcps = tcps;
1503 1503 ipcl_globalhash_insert(connp);
1504 1504
1505 1505 connp->conn_ixa->ixa_notify_cookie = tcp;
1506 1506 ASSERT(connp->conn_ixa->ixa_notify == tcp_notify);
1507 1507 connp->conn_recv = tcp_input_data;
1508 1508 ASSERT(connp->conn_recvicmp == tcp_icmp_input);
1509 1509 ASSERT(connp->conn_verifyicmp == tcp_verifyicmp);
1510 1510 return ((void *)connp);
1511 1511 }
1512 1512 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1513 1513 /*
1514 1514 * Pre-allocate the tcp_rsrv_mp. This mblk will not be freed until
1515 1515 * this conn_t/tcp_t is freed at ipcl_conn_destroy().
1516 1516 */
1517 1517 tcp_rsrv_mp = allocb(0, BPRI_HI);
1518 1518 if (tcp_rsrv_mp == NULL)
1519 1519 return (NULL);
1520 1520
1521 1521 if ((connp = ipcl_conn_create(IPCL_TCPCONN, KM_NOSLEEP,
1522 1522 tcps->tcps_netstack)) == NULL) {
1523 1523 freeb(tcp_rsrv_mp);
1524 1524 return (NULL);
1525 1525 }
1526 1526
1527 1527 tcp = connp->conn_tcp;
1528 1528 tcp->tcp_rsrv_mp = tcp_rsrv_mp;
1529 1529 mutex_init(&tcp->tcp_rsrv_mp_lock, NULL, MUTEX_DEFAULT, NULL);
1530 1530
1531 1531 tcp->tcp_tcps = tcps;
1532 1532
1533 1533 connp->conn_recv = tcp_input_data;
1534 1534 connp->conn_recvicmp = tcp_icmp_input;
1535 1535 connp->conn_verifyicmp = tcp_verifyicmp;
1536 1536
1537 1537 /*
1538 1538 * Register tcp_notify to listen to capability changes detected by IP.
1539 1539 * This upcall is made in the context of the call to conn_ip_output
1540 1540 * thus it is inside the squeue.
1541 1541 */
1542 1542 connp->conn_ixa->ixa_notify = tcp_notify;
1543 1543 connp->conn_ixa->ixa_notify_cookie = tcp;
1544 1544
1545 1545 return ((void *)connp);
1546 1546 }
1547 1547
1548 1548 /*
1549 1549 * Handle connect to IPv4 destinations, including connections for AF_INET6
1550 1550 * sockets connecting to IPv4 mapped IPv6 destinations.
1551 1551 * Returns zero if OK, a positive errno, or a negative TLI error.
1552 1552 */
1553 1553 static int
1554 1554 tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp, in_port_t dstport,
1555 1555 uint_t srcid)
1556 1556 {
1557 1557 ipaddr_t dstaddr = *dstaddrp;
1558 1558 uint16_t lport;
1559 1559 conn_t *connp = tcp->tcp_connp;
1560 1560 tcp_stack_t *tcps = tcp->tcp_tcps;
1561 1561 int error;
1562 1562
1563 1563 ASSERT(connp->conn_ipversion == IPV4_VERSION);
1564 1564
1565 1565 /* Check for attempt to connect to INADDR_ANY */
1566 1566 if (dstaddr == INADDR_ANY) {
1567 1567 /*
1568 1568 * SunOS 4.x and 4.3 BSD allow an application
1569 1569 * to connect a TCP socket to INADDR_ANY.
1570 1570 * When they do this, the kernel picks the
1571 1571 * address of one interface and uses it
1572 1572 * instead. The kernel usually ends up
1573 1573 * picking the address of the loopback
1574 1574 * interface. This is an undocumented feature.
1575 1575 * However, we provide the same thing here
1576 1576 * in order to have source and binary
1577 1577 * compatibility with SunOS 4.x.
1578 1578 * Update the T_CONN_REQ (sin/sin6) since it is used to
1579 1579 * generate the T_CONN_CON.
1580 1580 */
1581 1581 dstaddr = htonl(INADDR_LOOPBACK);
1582 1582 *dstaddrp = dstaddr;
1583 1583 }
1584 1584
1585 1585 /* Handle __sin6_src_id if socket not bound to an IP address */
1586 1586 if (srcid != 0 && connp->conn_laddr_v4 == INADDR_ANY) {
1587 1587 if (!ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1588 1588 IPCL_ZONEID(connp), B_TRUE, tcps->tcps_netstack)) {
1589 1589 /* Mismatch - conn_laddr_v6 would be v6 address. */
1590 1590 return (EADDRNOTAVAIL);
1591 1591 }
1592 1592 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1593 1593 }
1594 1594
1595 1595 IN6_IPADDR_TO_V4MAPPED(dstaddr, &connp->conn_faddr_v6);
1596 1596 connp->conn_fport = dstport;
1597 1597
1598 1598 /*
1599 1599 * At this point the remote destination address and remote port fields
1600 1600 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1601 1601 * have to see which state tcp was in so we can take appropriate action.
1602 1602 */
1603 1603 if (tcp->tcp_state == TCPS_IDLE) {
1604 1604 /*
1605 1605 * We support a quick connect capability here, allowing
1606 1606 * clients to transition directly from IDLE to SYN_SENT
1607 1607 * tcp_bindi will pick an unused port, insert the connection
1608 1608 * in the bind hash and transition to BOUND state.
1609 1609 */
1610 1610 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1611 1611 tcp, B_TRUE);
1612 1612 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1613 1613 B_FALSE, B_FALSE);
1614 1614 if (lport == 0)
1615 1615 return (-TNOADDR);
1616 1616 }
1617 1617
1618 1618 /*
1619 1619 * Lookup the route to determine a source address and the uinfo.
1620 1620 * Setup TCP parameters based on the metrics/DCE.
1621 1621 */
1622 1622 error = tcp_set_destination(tcp);
1623 1623 if (error != 0)
1624 1624 return (error);
1625 1625
1626 1626 /*
1627 1627 * Don't let an endpoint connect to itself.
1628 1628 */
1629 1629 if (connp->conn_faddr_v4 == connp->conn_laddr_v4 &&
1630 1630 connp->conn_fport == connp->conn_lport)
1631 1631 return (-TBADADDR);
1632 1632
1633 1633 tcp->tcp_state = TCPS_SYN_SENT;
1634 1634
1635 1635 return (ipcl_conn_insert_v4(connp));
1636 1636 }
1637 1637
1638 1638 /*
1639 1639 * Handle connect to IPv6 destinations.
1640 1640 * Returns zero if OK, a positive errno, or a negative TLI error.
1641 1641 */
1642 1642 static int
1643 1643 tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp, in_port_t dstport,
1644 1644 uint32_t flowinfo, uint_t srcid, uint32_t scope_id)
1645 1645 {
1646 1646 uint16_t lport;
1647 1647 conn_t *connp = tcp->tcp_connp;
1648 1648 tcp_stack_t *tcps = tcp->tcp_tcps;
1649 1649 int error;
1650 1650
1651 1651 ASSERT(connp->conn_family == AF_INET6);
1652 1652
1653 1653 /*
1654 1654 * If we're here, it means that the destination address is a native
1655 1655 * IPv6 address. Return an error if conn_ipversion is not IPv6. A
1656 1656 * reason why it might not be IPv6 is if the socket was bound to an
1657 1657 * IPv4-mapped IPv6 address.
1658 1658 */
1659 1659 if (connp->conn_ipversion != IPV6_VERSION)
1660 1660 return (-TBADADDR);
1661 1661
1662 1662 /*
1663 1663 * Interpret a zero destination to mean loopback.
1664 1664 * Update the T_CONN_REQ (sin/sin6) since it is used to
1665 1665 * generate the T_CONN_CON.
1666 1666 */
1667 1667 if (IN6_IS_ADDR_UNSPECIFIED(dstaddrp))
1668 1668 *dstaddrp = ipv6_loopback;
1669 1669
1670 1670 /* Handle __sin6_src_id if socket not bound to an IP address */
1671 1671 if (srcid != 0 && IN6_IS_ADDR_UNSPECIFIED(&connp->conn_laddr_v6)) {
1672 1672 if (!ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1673 1673 IPCL_ZONEID(connp), B_FALSE, tcps->tcps_netstack)) {
1674 1674 /* Mismatch - conn_laddr_v6 would be v4-mapped. */
1675 1675 return (EADDRNOTAVAIL);
1676 1676 }
1677 1677 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1678 1678 }
1679 1679
1680 1680 /*
1681 1681 * Take care of the scope_id now.
1682 1682 */
1683 1683 if (scope_id != 0 && IN6_IS_ADDR_LINKSCOPE(dstaddrp)) {
1684 1684 connp->conn_ixa->ixa_flags |= IXAF_SCOPEID_SET;
1685 1685 connp->conn_ixa->ixa_scopeid = scope_id;
1686 1686 } else {
1687 1687 connp->conn_ixa->ixa_flags &= ~IXAF_SCOPEID_SET;
1688 1688 }
1689 1689
1690 1690 connp->conn_flowinfo = flowinfo;
1691 1691 connp->conn_faddr_v6 = *dstaddrp;
1692 1692 connp->conn_fport = dstport;
1693 1693
1694 1694 /*
1695 1695 * At this point the remote destination address and remote port fields
1696 1696 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1697 1697 * have to see which state tcp was in so we can take appropriate action.
1698 1698 */
1699 1699 if (tcp->tcp_state == TCPS_IDLE) {
1700 1700 /*
1701 1701 * We support a quick connect capability here, allowing
1702 1702 * clients to transition directly from IDLE to SYN_SENT
1703 1703 * tcp_bindi will pick an unused port, insert the connection
1704 1704 * in the bind hash and transition to BOUND state.
1705 1705 */
1706 1706 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1707 1707 tcp, B_TRUE);
1708 1708 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1709 1709 B_FALSE, B_FALSE);
1710 1710 if (lport == 0)
1711 1711 return (-TNOADDR);
1712 1712 }
1713 1713
1714 1714 /*
1715 1715 * Lookup the route to determine a source address and the uinfo.
1716 1716 * Setup TCP parameters based on the metrics/DCE.
1717 1717 */
1718 1718 error = tcp_set_destination(tcp);
1719 1719 if (error != 0)
1720 1720 return (error);
1721 1721
1722 1722 /*
1723 1723 * Don't let an endpoint connect to itself.
1724 1724 */
1725 1725 if (IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6, &connp->conn_laddr_v6) &&
1726 1726 connp->conn_fport == connp->conn_lport)
1727 1727 return (-TBADADDR);
1728 1728
1729 1729 tcp->tcp_state = TCPS_SYN_SENT;
1730 1730
1731 1731 return (ipcl_conn_insert_v6(connp));
1732 1732 }
1733 1733
1734 1734 /*
1735 1735 * Disconnect
1736 1736 * Note that unlike other functions this returns a positive tli error
1737 1737 * when it fails; it never returns an errno.
1738 1738 */
1739 1739 static int
1740 1740 tcp_disconnect_common(tcp_t *tcp, t_scalar_t seqnum)
1741 1741 {
1742 1742 conn_t *lconnp;
1743 1743 tcp_stack_t *tcps = tcp->tcp_tcps;
1744 1744 conn_t *connp = tcp->tcp_connp;
1745 1745
1746 1746 /*
1747 1747 * Right now, upper modules pass down a T_DISCON_REQ to TCP,
1748 1748 * when the stream is in BOUND state. Do not send a reset,
1749 1749 * since the destination IP address is not valid, and it can
1750 1750 * be the initialized value of all zeros (broadcast address).
1751 1751 */
1752 1752 if (tcp->tcp_state <= TCPS_BOUND) {
1753 1753 if (connp->conn_debug) {
1754 1754 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1755 1755 "tcp_disconnect: bad state, %d", tcp->tcp_state);
1756 1756 }
1757 1757 return (TOUTSTATE);
1758 1758 } else if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1759 1759 TCPS_CONN_DEC(tcps);
1760 1760 }
1761 1761
1762 1762 if (seqnum == -1 || tcp->tcp_conn_req_max == 0) {
1763 1763
1764 1764 /*
1765 1765 * According to TPI, for non-listeners, ignore seqnum
1766 1766 * and disconnect.
1767 1767 * Following interpretation of -1 seqnum is historical
1768 1768 * and implied TPI ? (TPI only states that for T_CONN_IND,
1769 1769 * a valid seqnum should not be -1).
1770 1770 *
1771 1771 * -1 means disconnect everything
1772 1772 * regardless even on a listener.
1773 1773 */
1774 1774
1775 1775 int old_state = tcp->tcp_state;
1776 1776 ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
1777 1777
1778 1778 /*
1779 1779 * The connection can't be on the tcp_time_wait_head list
1780 1780 * since it is not detached.
1781 1781 */
1782 1782 ASSERT(tcp->tcp_time_wait_next == NULL);
1783 1783 ASSERT(tcp->tcp_time_wait_prev == NULL);
1784 1784 ASSERT(tcp->tcp_time_wait_expire == 0);
1785 1785 /*
1786 1786 * If it used to be a listener, check to make sure no one else
1787 1787 * has taken the port before switching back to LISTEN state.
1788 1788 */
1789 1789 if (connp->conn_ipversion == IPV4_VERSION) {
1790 1790 lconnp = ipcl_lookup_listener_v4(connp->conn_lport,
1791 1791 connp->conn_laddr_v4, IPCL_ZONEID(connp), ipst);
1792 1792 } else {
1793 1793 uint_t ifindex = 0;
1794 1794
1795 1795 if (connp->conn_ixa->ixa_flags & IXAF_SCOPEID_SET)
1796 1796 ifindex = connp->conn_ixa->ixa_scopeid;
1797 1797
1798 1798 /* Allow conn_bound_if listeners? */
1799 1799 lconnp = ipcl_lookup_listener_v6(connp->conn_lport,
1800 1800 &connp->conn_laddr_v6, ifindex, IPCL_ZONEID(connp),
1801 1801 ipst);
1802 1802 }
1803 1803 if (tcp->tcp_conn_req_max && lconnp == NULL) {
1804 1804 tcp->tcp_state = TCPS_LISTEN;
1805 1805 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1806 1806 connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1807 1807 NULL, int32_t, old_state);
1808 1808 } else if (old_state > TCPS_BOUND) {
1809 1809 tcp->tcp_conn_req_max = 0;
1810 1810 tcp->tcp_state = TCPS_BOUND;
1811 1811 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1812 1812 connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1813 1813 NULL, int32_t, old_state);
1814 1814
1815 1815 /*
1816 1816 * If this end point is not going to become a listener,
1817 1817 * decrement the listener connection count if
1818 1818 * necessary. Note that we do not do this if it is
1819 1819 * going to be a listner (the above if case) since
1820 1820 * then it may remove the counter struct.
1821 1821 */
1822 1822 if (tcp->tcp_listen_cnt != NULL)
1823 1823 TCP_DECR_LISTEN_CNT(tcp);
1824 1824 }
1825 1825 if (lconnp != NULL)
1826 1826 CONN_DEC_REF(lconnp);
1827 1827 switch (old_state) {
1828 1828 case TCPS_SYN_SENT:
1829 1829 case TCPS_SYN_RCVD:
1830 1830 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1831 1831 break;
1832 1832 case TCPS_ESTABLISHED:
1833 1833 case TCPS_CLOSE_WAIT:
1834 1834 TCPS_BUMP_MIB(tcps, tcpEstabResets);
1835 1835 break;
1836 1836 }
1837 1837
1838 1838 if (tcp->tcp_fused)
1839 1839 tcp_unfuse(tcp);
1840 1840
1841 1841 mutex_enter(&tcp->tcp_eager_lock);
1842 1842 if ((tcp->tcp_conn_req_cnt_q0 != 0) ||
1843 1843 (tcp->tcp_conn_req_cnt_q != 0)) {
1844 1844 tcp_eager_cleanup(tcp, 0);
1845 1845 }
1846 1846 mutex_exit(&tcp->tcp_eager_lock);
1847 1847
1848 1848 tcp_xmit_ctl("tcp_disconnect", tcp, tcp->tcp_snxt,
1849 1849 tcp->tcp_rnxt, TH_RST | TH_ACK);
1850 1850
1851 1851 tcp_reinit(tcp);
1852 1852
1853 1853 return (0);
1854 1854 } else if (!tcp_eager_blowoff(tcp, seqnum)) {
1855 1855 return (TBADSEQ);
1856 1856 }
1857 1857 return (0);
1858 1858 }
1859 1859
1860 1860 /*
1861 1861 * Our client hereby directs us to reject the connection request
1862 1862 * that tcp_input_listener() marked with 'seqnum'. Rejection consists
1863 1863 * of sending the appropriate RST, not an ICMP error.
1864 1864 */
1865 1865 void
1866 1866 tcp_disconnect(tcp_t *tcp, mblk_t *mp)
1867 1867 {
1868 1868 t_scalar_t seqnum;
1869 1869 int error;
1870 1870 conn_t *connp = tcp->tcp_connp;
1871 1871
1872 1872 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1873 1873 if ((mp->b_wptr - mp->b_rptr) < sizeof (struct T_discon_req)) {
1874 1874 tcp_err_ack(tcp, mp, TPROTO, 0);
1875 1875 return;
1876 1876 }
1877 1877 seqnum = ((struct T_discon_req *)mp->b_rptr)->SEQ_number;
1878 1878 error = tcp_disconnect_common(tcp, seqnum);
1879 1879 if (error != 0)
1880 1880 tcp_err_ack(tcp, mp, error, 0);
1881 1881 else {
1882 1882 if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1883 1883 /* Send M_FLUSH according to TPI */
1884 1884 (void) putnextctl1(connp->conn_rq, M_FLUSH, FLUSHRW);
1885 1885 }
1886 1886 mp = mi_tpi_ok_ack_alloc(mp);
1887 1887 if (mp != NULL)
1888 1888 putnext(connp->conn_rq, mp);
1889 1889 }
1890 1890 }
1891 1891
1892 1892 /*
1893 1893 * Handle reinitialization of a tcp structure.
1894 1894 * Maintain "binding state" resetting the state to BOUND, LISTEN, or IDLE.
1895 1895 */
1896 1896 static void
1897 1897 tcp_reinit(tcp_t *tcp)
1898 1898 {
1899 1899 mblk_t *mp;
1900 1900 tcp_stack_t *tcps = tcp->tcp_tcps;
1901 1901 conn_t *connp = tcp->tcp_connp;
1902 1902 int32_t oldstate;
1903 1903
1904 1904 /* tcp_reinit should never be called for detached tcp_t's */
1905 1905 ASSERT(tcp->tcp_listener == NULL);
1906 1906 ASSERT((connp->conn_family == AF_INET &&
1907 1907 connp->conn_ipversion == IPV4_VERSION) ||
1908 1908 (connp->conn_family == AF_INET6 &&
1909 1909 (connp->conn_ipversion == IPV4_VERSION ||
1910 1910 connp->conn_ipversion == IPV6_VERSION)));
1911 1911
1912 1912 /* Cancel outstanding timers */
1913 1913 tcp_timers_stop(tcp);
1914 1914
1915 1915 /*
1916 1916 * Reset everything in the state vector, after updating global
1917 1917 * MIB data from instance counters.
1918 1918 */
1919 1919 TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1920 1920 tcp->tcp_ibsegs = 0;
1921 1921 TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1922 1922 tcp->tcp_obsegs = 0;
1923 1923
1924 1924 tcp_close_mpp(&tcp->tcp_xmit_head);
1925 1925 if (tcp->tcp_snd_zcopy_aware)
1926 1926 tcp_zcopy_notify(tcp);
1927 1927 tcp->tcp_xmit_last = tcp->tcp_xmit_tail = NULL;
1928 1928 tcp->tcp_unsent = tcp->tcp_xmit_tail_unsent = 0;
1929 1929 mutex_enter(&tcp->tcp_non_sq_lock);
1930 1930 if (tcp->tcp_flow_stopped &&
1931 1931 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1932 1932 tcp_clrqfull(tcp);
1933 1933 }
1934 1934 mutex_exit(&tcp->tcp_non_sq_lock);
1935 1935 tcp_close_mpp(&tcp->tcp_reass_head);
1936 1936 tcp->tcp_reass_tail = NULL;
1937 1937 if (tcp->tcp_rcv_list != NULL) {
1938 1938 /* Free b_next chain */
1939 1939 tcp_close_mpp(&tcp->tcp_rcv_list);
1940 1940 tcp->tcp_rcv_last_head = NULL;
1941 1941 tcp->tcp_rcv_last_tail = NULL;
1942 1942 tcp->tcp_rcv_cnt = 0;
1943 1943 }
1944 1944 tcp->tcp_rcv_last_tail = NULL;
1945 1945
1946 1946 if ((mp = tcp->tcp_urp_mp) != NULL) {
1947 1947 freemsg(mp);
1948 1948 tcp->tcp_urp_mp = NULL;
1949 1949 }
1950 1950 if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1951 1951 freemsg(mp);
1952 1952 tcp->tcp_urp_mark_mp = NULL;
1953 1953 }
1954 1954 if (tcp->tcp_fused_sigurg_mp != NULL) {
1955 1955 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1956 1956 freeb(tcp->tcp_fused_sigurg_mp);
1957 1957 tcp->tcp_fused_sigurg_mp = NULL;
1958 1958 }
1959 1959 if (tcp->tcp_ordrel_mp != NULL) {
1960 1960 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1961 1961 freeb(tcp->tcp_ordrel_mp);
1962 1962 tcp->tcp_ordrel_mp = NULL;
1963 1963 }
1964 1964
1965 1965 /*
1966 1966 * Following is a union with two members which are
1967 1967 * identical types and size so the following cleanup
1968 1968 * is enough.
1969 1969 */
1970 1970 tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1971 1971
1972 1972 CL_INET_DISCONNECT(connp);
1973 1973
1974 1974 /*
1975 1975 * The connection can't be on the tcp_time_wait_head list
1976 1976 * since it is not detached.
1977 1977 */
1978 1978 ASSERT(tcp->tcp_time_wait_next == NULL);
1979 1979 ASSERT(tcp->tcp_time_wait_prev == NULL);
1980 1980 ASSERT(tcp->tcp_time_wait_expire == 0);
1981 1981
1982 1982 /*
1983 1983 * Reset/preserve other values
1984 1984 */
1985 1985 tcp_reinit_values(tcp);
1986 1986 ipcl_hash_remove(connp);
1987 1987 /* Note that ixa_cred gets cleared in ixa_cleanup */
1988 1988 ixa_cleanup(connp->conn_ixa);
1989 1989 tcp_ipsec_cleanup(tcp);
1990 1990
1991 1991 connp->conn_laddr_v6 = connp->conn_bound_addr_v6;
1992 1992 connp->conn_saddr_v6 = connp->conn_bound_addr_v6;
1993 1993 oldstate = tcp->tcp_state;
1994 1994
1995 1995 if (tcp->tcp_conn_req_max != 0) {
1996 1996 /*
1997 1997 * This is the case when a TLI program uses the same
1998 1998 * transport end point to accept a connection. This
1999 1999 * makes the TCP both a listener and acceptor. When
2000 2000 * this connection is closed, we need to set the state
2001 2001 * back to TCPS_LISTEN. Make sure that the eager list
2002 2002 * is reinitialized.
2003 2003 *
2004 2004 * Note that this stream is still bound to the four
2005 2005 * tuples of the previous connection in IP. If a new
2006 2006 * SYN with different foreign address comes in, IP will
2007 2007 * not find it and will send it to the global queue. In
2008 2008 * the global queue, TCP will do a tcp_lookup_listener()
2009 2009 * to find this stream. This works because this stream
2010 2010 * is only removed from connected hash.
2011 2011 *
2012 2012 */
2013 2013 tcp->tcp_state = TCPS_LISTEN;
2014 2014 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
2015 2015 tcp->tcp_eager_next_drop_q0 = tcp;
2016 2016 tcp->tcp_eager_prev_drop_q0 = tcp;
2017 2017 /*
2018 2018 * Initially set conn_recv to tcp_input_listener_unbound to try
2019 2019 * to pick a good squeue for the listener when the first SYN
2020 2020 * arrives. tcp_input_listener_unbound sets it to
2021 2021 * tcp_input_listener on that first SYN.
2022 2022 */
2023 2023 connp->conn_recv = tcp_input_listener_unbound;
2024 2024
2025 2025 connp->conn_proto = IPPROTO_TCP;
2026 2026 connp->conn_faddr_v6 = ipv6_all_zeros;
2027 2027 connp->conn_fport = 0;
2028 2028
2029 2029 (void) ipcl_bind_insert(connp);
2030 2030 } else {
2031 2031 tcp->tcp_state = TCPS_BOUND;
2032 2032 }
2033 2033
2034 2034 /*
2035 2035 * Initialize to default values
2036 2036 */
2037 2037 tcp_init_values(tcp, NULL);
2038 2038
2039 2039 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
2040 2040 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
2041 2041 int32_t, oldstate);
2042 2042
2043 2043 ASSERT(tcp->tcp_ptpbhn != NULL);
2044 2044 tcp->tcp_rwnd = connp->conn_rcvbuf;
2045 2045 tcp->tcp_mss = connp->conn_ipversion != IPV4_VERSION ?
2046 2046 tcps->tcps_mss_def_ipv6 : tcps->tcps_mss_def_ipv4;
2047 2047 }
2048 2048
2049 2049 /*
2050 2050 * Force values to zero that need be zero.
2051 2051 * Do not touch values asociated with the BOUND or LISTEN state
2052 2052 * since the connection will end up in that state after the reinit.
2053 2053 * NOTE: tcp_reinit_values MUST have a line for each field in the tcp_t
2054 2054 * structure!
2055 2055 */
2056 2056 static void
2057 2057 tcp_reinit_values(tcp)
2058 2058 tcp_t *tcp;
2059 2059 {
2060 2060 tcp_stack_t *tcps = tcp->tcp_tcps;
2061 2061 conn_t *connp = tcp->tcp_connp;
2062 2062
2063 2063 #ifndef lint
2064 2064 #define DONTCARE(x)
2065 2065 #define PRESERVE(x)
2066 2066 #else
2067 2067 #define DONTCARE(x) ((x) = (x))
2068 2068 #define PRESERVE(x) ((x) = (x))
2069 2069 #endif /* lint */
2070 2070
2071 2071 PRESERVE(tcp->tcp_bind_hash_port);
2072 2072 PRESERVE(tcp->tcp_bind_hash);
2073 2073 PRESERVE(tcp->tcp_ptpbhn);
2074 2074 PRESERVE(tcp->tcp_acceptor_hash);
2075 2075 PRESERVE(tcp->tcp_ptpahn);
2076 2076
2077 2077 /* Should be ASSERT NULL on these with new code! */
2078 2078 ASSERT(tcp->tcp_time_wait_next == NULL);
2079 2079 ASSERT(tcp->tcp_time_wait_prev == NULL);
2080 2080 ASSERT(tcp->tcp_time_wait_expire == 0);
2081 2081 PRESERVE(tcp->tcp_state);
2082 2082 PRESERVE(connp->conn_rq);
2083 2083 PRESERVE(connp->conn_wq);
2084 2084
2085 2085 ASSERT(tcp->tcp_xmit_head == NULL);
2086 2086 ASSERT(tcp->tcp_xmit_last == NULL);
2087 2087 ASSERT(tcp->tcp_unsent == 0);
2088 2088 ASSERT(tcp->tcp_xmit_tail == NULL);
2089 2089 ASSERT(tcp->tcp_xmit_tail_unsent == 0);
2090 2090
2091 2091 tcp->tcp_snxt = 0; /* Displayed in mib */
2092 2092 tcp->tcp_suna = 0; /* Displayed in mib */
2093 2093 tcp->tcp_swnd = 0;
2094 2094 DONTCARE(tcp->tcp_cwnd); /* Init in tcp_process_options */
2095 2095
2096 2096 ASSERT(tcp->tcp_ibsegs == 0);
2097 2097 ASSERT(tcp->tcp_obsegs == 0);
2098 2098
2099 2099 if (connp->conn_ht_iphc != NULL) {
2100 2100 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
2101 2101 connp->conn_ht_iphc = NULL;
2102 2102 connp->conn_ht_iphc_allocated = 0;
2103 2103 connp->conn_ht_iphc_len = 0;
2104 2104 connp->conn_ht_ulp = NULL;
2105 2105 connp->conn_ht_ulp_len = 0;
2106 2106 tcp->tcp_ipha = NULL;
2107 2107 tcp->tcp_ip6h = NULL;
2108 2108 tcp->tcp_tcpha = NULL;
2109 2109 }
2110 2110
2111 2111 /* We clear any IP_OPTIONS and extension headers */
2112 2112 ip_pkt_free(&connp->conn_xmit_ipp);
2113 2113
2114 2114 DONTCARE(tcp->tcp_naglim); /* Init in tcp_init_values */
2115 2115 DONTCARE(tcp->tcp_ipha);
2116 2116 DONTCARE(tcp->tcp_ip6h);
2117 2117 DONTCARE(tcp->tcp_tcpha);
2118 2118 tcp->tcp_valid_bits = 0;
2119 2119
2120 2120 DONTCARE(tcp->tcp_timer_backoff); /* Init in tcp_init_values */
2121 2121 DONTCARE(tcp->tcp_last_recv_time); /* Init in tcp_init_values */
2122 2122 tcp->tcp_last_rcv_lbolt = 0;
2123 2123
2124 2124 tcp->tcp_init_cwnd = 0;
2125 2125
2126 2126 tcp->tcp_urp_last_valid = 0;
2127 2127 tcp->tcp_hard_binding = 0;
2128 2128
2129 2129 tcp->tcp_fin_acked = 0;
2130 2130 tcp->tcp_fin_rcvd = 0;
2131 2131 tcp->tcp_fin_sent = 0;
2132 2132 tcp->tcp_ordrel_done = 0;
2133 2133
2134 2134 tcp->tcp_detached = 0;
2135 2135
2136 2136 tcp->tcp_snd_ws_ok = B_FALSE;
2137 2137 tcp->tcp_snd_ts_ok = B_FALSE;
2138 2138 tcp->tcp_zero_win_probe = 0;
2139 2139
2140 2140 tcp->tcp_loopback = 0;
2141 2141 tcp->tcp_localnet = 0;
2142 2142 tcp->tcp_syn_defense = 0;
2143 2143 tcp->tcp_set_timer = 0;
2144 2144
2145 2145 tcp->tcp_active_open = 0;
2146 2146 tcp->tcp_rexmit = B_FALSE;
2147 2147 tcp->tcp_xmit_zc_clean = B_FALSE;
2148 2148
2149 2149 tcp->tcp_snd_sack_ok = B_FALSE;
2150 2150 tcp->tcp_hwcksum = B_FALSE;
2151 2151
2152 2152 DONTCARE(tcp->tcp_maxpsz_multiplier); /* Init in tcp_init_values */
2153 2153
2154 2154 tcp->tcp_conn_def_q0 = 0;
2155 2155 tcp->tcp_ip_forward_progress = B_FALSE;
2156 2156 tcp->tcp_ecn_ok = B_FALSE;
2157 2157
2158 2158 tcp->tcp_cwr = B_FALSE;
2159 2159 tcp->tcp_ecn_echo_on = B_FALSE;
2160 2160 tcp->tcp_is_wnd_shrnk = B_FALSE;
2161 2161
2162 2162 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
2163 2163 bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
2164 2164
2165 2165 tcp->tcp_rcv_ws = 0;
2166 2166 tcp->tcp_snd_ws = 0;
2167 2167 tcp->tcp_ts_recent = 0;
2168 2168 tcp->tcp_rnxt = 0; /* Displayed in mib */
2169 2169 DONTCARE(tcp->tcp_rwnd); /* Set in tcp_reinit() */
2170 2170 tcp->tcp_initial_pmtu = 0;
2171 2171
2172 2172 ASSERT(tcp->tcp_reass_head == NULL);
2173 2173 ASSERT(tcp->tcp_reass_tail == NULL);
2174 2174
2175 2175 tcp->tcp_cwnd_cnt = 0;
2176 2176
2177 2177 ASSERT(tcp->tcp_rcv_list == NULL);
2178 2178 ASSERT(tcp->tcp_rcv_last_head == NULL);
2179 2179 ASSERT(tcp->tcp_rcv_last_tail == NULL);
2180 2180 ASSERT(tcp->tcp_rcv_cnt == 0);
2181 2181
2182 2182 DONTCARE(tcp->tcp_cwnd_ssthresh); /* Init in tcp_set_destination */
2183 2183 DONTCARE(tcp->tcp_cwnd_max); /* Init in tcp_init_values */
2184 2184 tcp->tcp_csuna = 0;
2185 2185
2186 2186 tcp->tcp_rto = 0; /* Displayed in MIB */
2187 2187 DONTCARE(tcp->tcp_rtt_sa); /* Init in tcp_init_values */
2188 2188 DONTCARE(tcp->tcp_rtt_sd); /* Init in tcp_init_values */
2189 2189 tcp->tcp_rtt_update = 0;
2190 2190
2191 2191 DONTCARE(tcp->tcp_swl1); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2192 2192 DONTCARE(tcp->tcp_swl2); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2193 2193
2194 2194 tcp->tcp_rack = 0; /* Displayed in mib */
2195 2195 tcp->tcp_rack_cnt = 0;
2196 2196 tcp->tcp_rack_cur_max = 0;
2197 2197 tcp->tcp_rack_abs_max = 0;
2198 2198
2199 2199 tcp->tcp_max_swnd = 0;
2200 2200
2201 2201 ASSERT(tcp->tcp_listener == NULL);
2202 2202
2203 2203 DONTCARE(tcp->tcp_irs); /* tcp_valid_bits cleared */
2204 2204 DONTCARE(tcp->tcp_iss); /* tcp_valid_bits cleared */
2205 2205 DONTCARE(tcp->tcp_fss); /* tcp_valid_bits cleared */
2206 2206 DONTCARE(tcp->tcp_urg); /* tcp_valid_bits cleared */
2207 2207
2208 2208 ASSERT(tcp->tcp_conn_req_cnt_q == 0);
2209 2209 ASSERT(tcp->tcp_conn_req_cnt_q0 == 0);
2210 2210 PRESERVE(tcp->tcp_conn_req_max);
2211 2211 PRESERVE(tcp->tcp_conn_req_seqnum);
2212 2212
2213 2213 DONTCARE(tcp->tcp_first_timer_threshold); /* Init in tcp_init_values */
2214 2214 DONTCARE(tcp->tcp_second_timer_threshold); /* Init in tcp_init_values */
2215 2215 DONTCARE(tcp->tcp_first_ctimer_threshold); /* Init in tcp_init_values */
2216 2216 DONTCARE(tcp->tcp_second_ctimer_threshold); /* in tcp_init_values */
2217 2217
2218 2218 DONTCARE(tcp->tcp_urp_last); /* tcp_urp_last_valid is cleared */
2219 2219 ASSERT(tcp->tcp_urp_mp == NULL);
2220 2220 ASSERT(tcp->tcp_urp_mark_mp == NULL);
2221 2221 ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
2222 2222
2223 2223 ASSERT(tcp->tcp_eager_next_q == NULL);
2224 2224 ASSERT(tcp->tcp_eager_last_q == NULL);
2225 2225 ASSERT((tcp->tcp_eager_next_q0 == NULL &&
2226 2226 tcp->tcp_eager_prev_q0 == NULL) ||
2227 2227 tcp->tcp_eager_next_q0 == tcp->tcp_eager_prev_q0);
2228 2228 ASSERT(tcp->tcp_conn.tcp_eager_conn_ind == NULL);
2229 2229
2230 2230 ASSERT((tcp->tcp_eager_next_drop_q0 == NULL &&
2231 2231 tcp->tcp_eager_prev_drop_q0 == NULL) ||
2232 2232 tcp->tcp_eager_next_drop_q0 == tcp->tcp_eager_prev_drop_q0);
2233 2233
2234 2234 DONTCARE(tcp->tcp_ka_rinterval); /* Init in tcp_init_values */
2235 2235 DONTCARE(tcp->tcp_ka_abort_thres); /* Init in tcp_init_values */
2236 2236 DONTCARE(tcp->tcp_ka_cnt); /* Init in tcp_init_values */
2237 2237
2238 2238 tcp->tcp_client_errno = 0;
2239 2239
2240 2240 DONTCARE(connp->conn_sum); /* Init in tcp_init_values */
2241 2241
2242 2242 connp->conn_faddr_v6 = ipv6_all_zeros; /* Displayed in MIB */
2243 2243
2244 2244 PRESERVE(connp->conn_bound_addr_v6);
2245 2245 tcp->tcp_last_sent_len = 0;
2246 2246 tcp->tcp_dupack_cnt = 0;
2247 2247
2248 2248 connp->conn_fport = 0; /* Displayed in MIB */
2249 2249 PRESERVE(connp->conn_lport);
2250 2250
2251 2251 PRESERVE(tcp->tcp_acceptor_lockp);
2252 2252
2253 2253 ASSERT(tcp->tcp_ordrel_mp == NULL);
2254 2254 PRESERVE(tcp->tcp_acceptor_id);
2255 2255 DONTCARE(tcp->tcp_ipsec_overhead);
2256 2256
2257 2257 PRESERVE(connp->conn_family);
2258 2258 /* Remove any remnants of mapped address binding */
2259 2259 if (connp->conn_family == AF_INET6) {
2260 2260 connp->conn_ipversion = IPV6_VERSION;
2261 2261 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2262 2262 } else {
2263 2263 connp->conn_ipversion = IPV4_VERSION;
2264 2264 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2265 2265 }
2266 2266
2267 2267 connp->conn_bound_if = 0;
2268 2268 connp->conn_recv_ancillary.crb_all = 0;
2269 2269 tcp->tcp_recvifindex = 0;
2270 2270 tcp->tcp_recvhops = 0;
2271 2271 tcp->tcp_closed = 0;
2272 2272 if (tcp->tcp_hopopts != NULL) {
2273 2273 mi_free(tcp->tcp_hopopts);
2274 2274 tcp->tcp_hopopts = NULL;
2275 2275 tcp->tcp_hopoptslen = 0;
2276 2276 }
2277 2277 ASSERT(tcp->tcp_hopoptslen == 0);
2278 2278 if (tcp->tcp_dstopts != NULL) {
2279 2279 mi_free(tcp->tcp_dstopts);
2280 2280 tcp->tcp_dstopts = NULL;
2281 2281 tcp->tcp_dstoptslen = 0;
2282 2282 }
2283 2283 ASSERT(tcp->tcp_dstoptslen == 0);
2284 2284 if (tcp->tcp_rthdrdstopts != NULL) {
2285 2285 mi_free(tcp->tcp_rthdrdstopts);
2286 2286 tcp->tcp_rthdrdstopts = NULL;
2287 2287 tcp->tcp_rthdrdstoptslen = 0;
2288 2288 }
2289 2289 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2290 2290 if (tcp->tcp_rthdr != NULL) {
2291 2291 mi_free(tcp->tcp_rthdr);
2292 2292 tcp->tcp_rthdr = NULL;
2293 2293 tcp->tcp_rthdrlen = 0;
2294 2294 }
2295 2295 ASSERT(tcp->tcp_rthdrlen == 0);
2296 2296
2297 2297 /* Reset fusion-related fields */
2298 2298 tcp->tcp_fused = B_FALSE;
2299 2299 tcp->tcp_unfusable = B_FALSE;
2300 2300 tcp->tcp_fused_sigurg = B_FALSE;
2301 2301 tcp->tcp_loopback_peer = NULL;
2302 2302
2303 2303 tcp->tcp_lso = B_FALSE;
2304 2304
2305 2305 tcp->tcp_in_ack_unsent = 0;
2306 2306 tcp->tcp_cork = B_FALSE;
2307 2307 tcp->tcp_tconnind_started = B_FALSE;
2308 2308
2309 2309 PRESERVE(tcp->tcp_squeue_bytes);
2310 2310
2311 2311 tcp->tcp_closemp_used = B_FALSE;
2312 2312
2313 2313 PRESERVE(tcp->tcp_rsrv_mp);
2314 2314 PRESERVE(tcp->tcp_rsrv_mp_lock);
2315 2315
2316 2316 #ifdef DEBUG
2317 2317 DONTCARE(tcp->tcmp_stk[0]);
2318 2318 #endif
2319 2319
2320 2320 PRESERVE(tcp->tcp_connid);
2321 2321
2322 2322 ASSERT(tcp->tcp_listen_cnt == NULL);
2323 2323 ASSERT(tcp->tcp_reass_tid == 0);
2324 2324
2325 2325 #undef DONTCARE
2326 2326 #undef PRESERVE
2327 2327 }
2328 2328
2329 2329 /*
2330 2330 * Initialize the various fields in tcp_t. If parent (the listener) is non
2331 2331 * NULL, certain values will be inheritted from it.
2332 2332 */
2333 2333 void
2334 2334 tcp_init_values(tcp_t *tcp, tcp_t *parent)
2335 2335 {
2336 2336 tcp_stack_t *tcps = tcp->tcp_tcps;
2337 2337 conn_t *connp = tcp->tcp_connp;
2338 2338 clock_t rto;
2339 2339
2340 2340 ASSERT((connp->conn_family == AF_INET &&
2341 2341 connp->conn_ipversion == IPV4_VERSION) ||
2342 2342 (connp->conn_family == AF_INET6 &&
2343 2343 (connp->conn_ipversion == IPV4_VERSION ||
2344 2344 connp->conn_ipversion == IPV6_VERSION)));
2345 2345
2346 2346 if (parent == NULL) {
2347 2347 tcp->tcp_naglim = tcps->tcps_naglim_def;
2348 2348
2349 2349 tcp->tcp_rto_initial = tcps->tcps_rexmit_interval_initial;
2350 2350 tcp->tcp_rto_min = tcps->tcps_rexmit_interval_min;
2351 2351 tcp->tcp_rto_max = tcps->tcps_rexmit_interval_max;
2352 2352
2353 2353 tcp->tcp_first_ctimer_threshold =
2354 2354 tcps->tcps_ip_notify_cinterval;
2355 2355 tcp->tcp_second_ctimer_threshold =
2356 2356 tcps->tcps_ip_abort_cinterval;
2357 2357 tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2358 2358 tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2359 2359
2360 2360 tcp->tcp_fin_wait_2_flush_interval =
2361 2361 tcps->tcps_fin_wait_2_flush_interval;
2362 2362
2363 2363 tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2364 2364 tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2365 2365 tcp->tcp_ka_cnt = 0;
2366 2366 tcp->tcp_ka_rinterval = 0;
2367 2367
2368 2368 /*
2369 2369 * Default value of tcp_init_cwnd is 0, so no need to set here
2370 2370 * if parent is NULL. But we need to inherit it from parent.
2371 2371 */
2372 2372 } else {
2373 2373 /* Inherit various TCP parameters from the parent. */
2374 2374 tcp->tcp_naglim = parent->tcp_naglim;
2375 2375
2376 2376 tcp->tcp_rto_initial = parent->tcp_rto_initial;
2377 2377 tcp->tcp_rto_min = parent->tcp_rto_min;
2378 2378 tcp->tcp_rto_max = parent->tcp_rto_max;
2379 2379
2380 2380 tcp->tcp_first_ctimer_threshold =
2381 2381 parent->tcp_first_ctimer_threshold;
2382 2382 tcp->tcp_second_ctimer_threshold =
2383 2383 parent->tcp_second_ctimer_threshold;
2384 2384 tcp->tcp_first_timer_threshold =
2385 2385 parent->tcp_first_timer_threshold;
2386 2386 tcp->tcp_second_timer_threshold =
2387 2387 parent->tcp_second_timer_threshold;
2388 2388
2389 2389 tcp->tcp_fin_wait_2_flush_interval =
2390 2390 parent->tcp_fin_wait_2_flush_interval;
2391 2391
2392 2392 tcp->tcp_ka_interval = parent->tcp_ka_interval;
2393 2393 tcp->tcp_ka_abort_thres = parent->tcp_ka_abort_thres;
2394 2394 tcp->tcp_ka_cnt = parent->tcp_ka_cnt;
2395 2395 tcp->tcp_ka_rinterval = parent->tcp_ka_rinterval;
2396 2396
2397 2397 tcp->tcp_init_cwnd = parent->tcp_init_cwnd;
2398 2398 }
2399 2399
2400 2400 /*
2401 2401 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2402 2402 * will be close to tcp_rexmit_interval_initial. By doing this, we
2403 2403 * allow the algorithm to adjust slowly to large fluctuations of RTT
2404 2404 * during first few transmissions of a connection as seen in slow
2405 2405 * links.
2406 2406 */
2407 2407 tcp->tcp_rtt_sa = tcp->tcp_rto_initial << 2;
2408 2408 tcp->tcp_rtt_sd = tcp->tcp_rto_initial >> 1;
|
↓ open down ↓ |
2372 lines elided |
↑ open up ↑ |
2409 2409 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
2410 2410 tcps->tcps_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
2411 2411 tcps->tcps_conn_grace_period;
2412 2412 TCP_SET_RTO(tcp, rto);
2413 2413
2414 2414 tcp->tcp_timer_backoff = 0;
2415 2415 tcp->tcp_ms_we_have_waited = 0;
2416 2416 tcp->tcp_last_recv_time = ddi_get_lbolt();
2417 2417 tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2418 2418 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2419 - tcp->tcp_snd_burst = TCP_CWND_INFINITE;
2420 2419
2421 2420 tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2422 2421
2423 2422 /* NOTE: ISS is now set in tcp_set_destination(). */
2424 2423
2425 2424 /* Reset fusion-related fields */
2426 2425 tcp->tcp_fused = B_FALSE;
2427 2426 tcp->tcp_unfusable = B_FALSE;
2428 2427 tcp->tcp_fused_sigurg = B_FALSE;
2429 2428 tcp->tcp_loopback_peer = NULL;
2430 2429
2431 2430 /* We rebuild the header template on the next connect/conn_request */
2432 2431
2433 2432 connp->conn_mlp_type = mlptSingle;
2434 2433
2435 2434 /*
2436 2435 * Init the window scale to the max so tcp_rwnd_set() won't pare
2437 2436 * down tcp_rwnd. tcp_set_destination() will set the right value later.
2438 2437 */
2439 2438 tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2440 2439 tcp->tcp_rwnd = connp->conn_rcvbuf;
2441 2440
2442 2441 tcp->tcp_cork = B_FALSE;
2443 2442 /*
2444 2443 * Init the tcp_debug option if it wasn't already set. This value
2445 2444 * determines whether TCP
2446 2445 * calls strlog() to print out debug messages. Doing this
2447 2446 * initialization here means that this value is not inherited thru
2448 2447 * tcp_reinit().
2449 2448 */
2450 2449 if (!connp->conn_debug)
2451 2450 connp->conn_debug = tcps->tcps_dbg;
2452 2451 }
2453 2452
2454 2453 /*
2455 2454 * Update the TCP connection according to change of PMTU.
2456 2455 *
2457 2456 * Path MTU might have changed by either increase or decrease, so need to
2458 2457 * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2459 2458 * or negative MSS, since tcp_mss_set() will do it.
2460 2459 */
2461 2460 void
2462 2461 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2463 2462 {
2464 2463 uint32_t pmtu;
2465 2464 int32_t mss;
2466 2465 conn_t *connp = tcp->tcp_connp;
2467 2466 ip_xmit_attr_t *ixa = connp->conn_ixa;
2468 2467 iaflags_t ixaflags;
2469 2468
2470 2469 if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2471 2470 return;
2472 2471
2473 2472 if (tcp->tcp_state < TCPS_ESTABLISHED)
2474 2473 return;
2475 2474
2476 2475 /*
2477 2476 * Always call ip_get_pmtu() to make sure that IP has updated
2478 2477 * ixa_flags properly.
2479 2478 */
2480 2479 pmtu = ip_get_pmtu(ixa);
2481 2480 ixaflags = ixa->ixa_flags;
2482 2481
2483 2482 /*
2484 2483 * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2485 2484 * IPsec overhead if applied. Make sure to use the most recent
2486 2485 * IPsec information.
2487 2486 */
2488 2487 mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2489 2488
2490 2489 /*
2491 2490 * Nothing to change, so just return.
2492 2491 */
2493 2492 if (mss == tcp->tcp_mss)
2494 2493 return;
2495 2494
2496 2495 /*
2497 2496 * Currently, for ICMP errors, only PMTU decrease is handled.
2498 2497 */
2499 2498 if (mss > tcp->tcp_mss && decrease_only)
2500 2499 return;
2501 2500
2502 2501 DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2503 2502
2504 2503 /*
2505 2504 * Update ixa_fragsize and ixa_pmtu.
2506 2505 */
2507 2506 ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2508 2507
2509 2508 /*
2510 2509 * Adjust MSS and all relevant variables.
2511 2510 */
2512 2511 tcp_mss_set(tcp, mss);
2513 2512
2514 2513 /*
2515 2514 * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2516 2515 * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2517 2516 * has a (potentially different) min size we do the same. Make sure to
2518 2517 * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2519 2518 * fragment the packet.
2520 2519 *
2521 2520 * LSO over IPv6 can not be fragmented. So need to disable LSO
2522 2521 * when IPv6 fragmentation is needed.
2523 2522 */
2524 2523 if (mss < tcp->tcp_tcps->tcps_mss_min)
2525 2524 ixaflags |= IXAF_PMTU_TOO_SMALL;
2526 2525
2527 2526 if (ixaflags & IXAF_PMTU_TOO_SMALL)
2528 2527 ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2529 2528
2530 2529 if ((connp->conn_ipversion == IPV4_VERSION) &&
2531 2530 !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2532 2531 tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2533 2532 }
2534 2533 ixa->ixa_flags = ixaflags;
2535 2534 }
2536 2535
2537 2536 int
2538 2537 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2539 2538 {
2540 2539 conn_t *connp = tcp->tcp_connp;
2541 2540 queue_t *q = connp->conn_rq;
2542 2541 int32_t mss = tcp->tcp_mss;
2543 2542 int maxpsz;
2544 2543
2545 2544 if (TCP_IS_DETACHED(tcp))
2546 2545 return (mss);
2547 2546 if (tcp->tcp_fused) {
2548 2547 maxpsz = tcp_fuse_maxpsz(tcp);
2549 2548 mss = INFPSZ;
2550 2549 } else if (tcp->tcp_maxpsz_multiplier == 0) {
2551 2550 /*
2552 2551 * Set the sd_qn_maxpsz according to the socket send buffer
2553 2552 * size, and sd_maxblk to INFPSZ (-1). This will essentially
2554 2553 * instruct the stream head to copyin user data into contiguous
2555 2554 * kernel-allocated buffers without breaking it up into smaller
2556 2555 * chunks. We round up the buffer size to the nearest SMSS.
2557 2556 */
2558 2557 maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2559 2558 mss = INFPSZ;
2560 2559 } else {
2561 2560 /*
2562 2561 * Set sd_qn_maxpsz to approx half the (receivers) buffer
2563 2562 * (and a multiple of the mss). This instructs the stream
2564 2563 * head to break down larger than SMSS writes into SMSS-
2565 2564 * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2566 2565 */
2567 2566 maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2568 2567 if (maxpsz > connp->conn_sndbuf / 2) {
2569 2568 maxpsz = connp->conn_sndbuf / 2;
2570 2569 /* Round up to nearest mss */
2571 2570 maxpsz = MSS_ROUNDUP(maxpsz, mss);
2572 2571 }
2573 2572 }
2574 2573
2575 2574 (void) proto_set_maxpsz(q, connp, maxpsz);
2576 2575 if (!(IPCL_IS_NONSTR(connp)))
2577 2576 connp->conn_wq->q_maxpsz = maxpsz;
2578 2577 if (set_maxblk)
2579 2578 (void) proto_set_tx_maxblk(q, connp, mss);
2580 2579 return (mss);
2581 2580 }
2582 2581
2583 2582 /* For /dev/tcp aka AF_INET open */
2584 2583 static int
2585 2584 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2586 2585 {
2587 2586 return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2588 2587 }
2589 2588
2590 2589 /* For /dev/tcp6 aka AF_INET6 open */
2591 2590 static int
2592 2591 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2593 2592 {
2594 2593 return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2595 2594 }
2596 2595
2597 2596 conn_t *
2598 2597 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2599 2598 int *errorp)
2600 2599 {
2601 2600 tcp_t *tcp = NULL;
2602 2601 conn_t *connp;
2603 2602 zoneid_t zoneid;
2604 2603 tcp_stack_t *tcps;
2605 2604 squeue_t *sqp;
2606 2605
2607 2606 ASSERT(errorp != NULL);
2608 2607 /*
2609 2608 * Find the proper zoneid and netstack.
2610 2609 */
2611 2610 /*
2612 2611 * Special case for install: miniroot needs to be able to
2613 2612 * access files via NFS as though it were always in the
2614 2613 * global zone.
2615 2614 */
2616 2615 if (credp == kcred && nfs_global_client_only != 0) {
2617 2616 zoneid = GLOBAL_ZONEID;
2618 2617 tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2619 2618 netstack_tcp;
2620 2619 ASSERT(tcps != NULL);
2621 2620 } else {
2622 2621 netstack_t *ns;
2623 2622 int err;
2624 2623
2625 2624 if ((err = secpolicy_basic_net_access(credp)) != 0) {
2626 2625 *errorp = err;
2627 2626 return (NULL);
2628 2627 }
2629 2628
2630 2629 ns = netstack_find_by_cred(credp);
2631 2630 ASSERT(ns != NULL);
2632 2631 tcps = ns->netstack_tcp;
2633 2632 ASSERT(tcps != NULL);
2634 2633
2635 2634 /*
2636 2635 * For exclusive stacks we set the zoneid to zero
2637 2636 * to make TCP operate as if in the global zone.
2638 2637 */
2639 2638 if (tcps->tcps_netstack->netstack_stackid !=
2640 2639 GLOBAL_NETSTACKID)
2641 2640 zoneid = GLOBAL_ZONEID;
2642 2641 else
2643 2642 zoneid = crgetzoneid(credp);
2644 2643 }
2645 2644
2646 2645 sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2647 2646 connp = (conn_t *)tcp_get_conn(sqp, tcps);
2648 2647 /*
2649 2648 * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2650 2649 * so we drop it by one.
2651 2650 */
2652 2651 netstack_rele(tcps->tcps_netstack);
2653 2652 if (connp == NULL) {
2654 2653 *errorp = ENOSR;
2655 2654 return (NULL);
2656 2655 }
2657 2656 ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2658 2657
2659 2658 connp->conn_sqp = sqp;
2660 2659 connp->conn_initial_sqp = connp->conn_sqp;
2661 2660 connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2662 2661 tcp = connp->conn_tcp;
2663 2662
2664 2663 /*
2665 2664 * Besides asking IP to set the checksum for us, have conn_ip_output
2666 2665 * to do the following checks when necessary:
2667 2666 *
2668 2667 * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2669 2668 * IXAF_VERIFY_PMTU: verify PMTU changes
2670 2669 * IXAF_VERIFY_LSO: verify LSO capability changes
2671 2670 */
2672 2671 connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2673 2672 IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2674 2673
2675 2674 if (!tcps->tcps_dev_flow_ctl)
2676 2675 connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2677 2676
2678 2677 if (isv6) {
2679 2678 connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2680 2679 connp->conn_ipversion = IPV6_VERSION;
2681 2680 connp->conn_family = AF_INET6;
2682 2681 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2683 2682 connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2684 2683 } else {
2685 2684 connp->conn_ipversion = IPV4_VERSION;
2686 2685 connp->conn_family = AF_INET;
2687 2686 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2688 2687 connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2689 2688 }
2690 2689 connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2691 2690
2692 2691 crhold(credp);
2693 2692 connp->conn_cred = credp;
2694 2693 connp->conn_cpid = curproc->p_pid;
2695 2694 connp->conn_open_time = ddi_get_lbolt64();
2696 2695
2697 2696 /* Cache things in the ixa without any refhold */
2698 2697 ASSERT(!(connp->conn_ixa->ixa_free_flags & IXA_FREE_CRED));
2699 2698 connp->conn_ixa->ixa_cred = credp;
2700 2699 connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2701 2700
2702 2701 connp->conn_zoneid = zoneid;
2703 2702 /* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2704 2703 connp->conn_ixa->ixa_zoneid = zoneid;
2705 2704 connp->conn_mlp_type = mlptSingle;
2706 2705 ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2707 2706 ASSERT(tcp->tcp_tcps == tcps);
2708 2707
2709 2708 /*
2710 2709 * If the caller has the process-wide flag set, then default to MAC
2711 2710 * exempt mode. This allows read-down to unlabeled hosts.
2712 2711 */
2713 2712 if (getpflags(NET_MAC_AWARE, credp) != 0)
2714 2713 connp->conn_mac_mode = CONN_MAC_AWARE;
2715 2714
2716 2715 connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2717 2716
2718 2717 if (issocket) {
2719 2718 tcp->tcp_issocket = 1;
2720 2719 }
2721 2720
2722 2721 connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2723 2722 connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2724 2723 if (tcps->tcps_snd_lowat_fraction != 0) {
2725 2724 connp->conn_sndlowat = connp->conn_sndbuf /
2726 2725 tcps->tcps_snd_lowat_fraction;
2727 2726 } else {
2728 2727 connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2729 2728 }
2730 2729 connp->conn_so_type = SOCK_STREAM;
2731 2730 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2732 2731 tcps->tcps_wroff_xtra;
2733 2732
2734 2733 SOCK_CONNID_INIT(tcp->tcp_connid);
2735 2734 /* DTrace ignores this - it isn't a tcp:::state-change */
2736 2735 tcp->tcp_state = TCPS_IDLE;
2737 2736 tcp_init_values(tcp, NULL);
2738 2737 return (connp);
2739 2738 }
2740 2739
2741 2740 static int
2742 2741 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2743 2742 boolean_t isv6)
2744 2743 {
2745 2744 tcp_t *tcp = NULL;
2746 2745 conn_t *connp = NULL;
2747 2746 int err;
2748 2747 vmem_t *minor_arena = NULL;
2749 2748 dev_t conn_dev;
2750 2749 boolean_t issocket;
2751 2750
2752 2751 if (q->q_ptr != NULL)
2753 2752 return (0);
2754 2753
2755 2754 if (sflag == MODOPEN)
2756 2755 return (EINVAL);
2757 2756
2758 2757 if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2759 2758 ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2760 2759 minor_arena = ip_minor_arena_la;
2761 2760 } else {
2762 2761 /*
2763 2762 * Either minor numbers in the large arena were exhausted
2764 2763 * or a non socket application is doing the open.
2765 2764 * Try to allocate from the small arena.
2766 2765 */
2767 2766 if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2768 2767 return (EBUSY);
2769 2768 }
2770 2769 minor_arena = ip_minor_arena_sa;
2771 2770 }
2772 2771
2773 2772 ASSERT(minor_arena != NULL);
2774 2773
2775 2774 *devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2776 2775
2777 2776 if (flag & SO_FALLBACK) {
2778 2777 /*
2779 2778 * Non streams socket needs a stream to fallback to
2780 2779 */
2781 2780 RD(q)->q_ptr = (void *)conn_dev;
2782 2781 WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2783 2782 WR(q)->q_ptr = (void *)minor_arena;
2784 2783 qprocson(q);
2785 2784 return (0);
2786 2785 } else if (flag & SO_ACCEPTOR) {
2787 2786 q->q_qinfo = &tcp_acceptor_rinit;
2788 2787 /*
2789 2788 * the conn_dev and minor_arena will be subsequently used by
2790 2789 * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2791 2790 * the minor device number for this connection from the q_ptr.
2792 2791 */
2793 2792 RD(q)->q_ptr = (void *)conn_dev;
2794 2793 WR(q)->q_qinfo = &tcp_acceptor_winit;
2795 2794 WR(q)->q_ptr = (void *)minor_arena;
2796 2795 qprocson(q);
2797 2796 return (0);
2798 2797 }
2799 2798
2800 2799 issocket = flag & SO_SOCKSTR;
2801 2800 connp = tcp_create_common(credp, isv6, issocket, &err);
2802 2801
2803 2802 if (connp == NULL) {
2804 2803 inet_minor_free(minor_arena, conn_dev);
2805 2804 q->q_ptr = WR(q)->q_ptr = NULL;
2806 2805 return (err);
2807 2806 }
2808 2807
2809 2808 connp->conn_rq = q;
2810 2809 connp->conn_wq = WR(q);
2811 2810 q->q_ptr = WR(q)->q_ptr = connp;
2812 2811
2813 2812 connp->conn_dev = conn_dev;
2814 2813 connp->conn_minor_arena = minor_arena;
2815 2814
2816 2815 ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2817 2816 ASSERT(WR(q)->q_qinfo == &tcp_winit);
2818 2817
2819 2818 tcp = connp->conn_tcp;
2820 2819
2821 2820 if (issocket) {
2822 2821 WR(q)->q_qinfo = &tcp_sock_winit;
2823 2822 } else {
2824 2823 #ifdef _ILP32
2825 2824 tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2826 2825 #else
2827 2826 tcp->tcp_acceptor_id = conn_dev;
2828 2827 #endif /* _ILP32 */
2829 2828 tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2830 2829 }
2831 2830
2832 2831 /*
2833 2832 * Put the ref for TCP. Ref for IP was already put
2834 2833 * by ipcl_conn_create. Also Make the conn_t globally
2835 2834 * visible to walkers
2836 2835 */
2837 2836 mutex_enter(&connp->conn_lock);
2838 2837 CONN_INC_REF_LOCKED(connp);
2839 2838 ASSERT(connp->conn_ref == 2);
2840 2839 connp->conn_state_flags &= ~CONN_INCIPIENT;
2841 2840 mutex_exit(&connp->conn_lock);
2842 2841
2843 2842 qprocson(q);
2844 2843 return (0);
2845 2844 }
2846 2845
2847 2846 /*
2848 2847 * Build/update the tcp header template (in conn_ht_iphc) based on
2849 2848 * conn_xmit_ipp. The headers include ip6_t, any extension
2850 2849 * headers, and the maximum size tcp header (to avoid reallocation
2851 2850 * on the fly for additional tcp options).
2852 2851 *
2853 2852 * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2854 2853 * Returns failure if can't allocate memory.
2855 2854 */
2856 2855 int
2857 2856 tcp_build_hdrs(tcp_t *tcp)
2858 2857 {
2859 2858 tcp_stack_t *tcps = tcp->tcp_tcps;
2860 2859 conn_t *connp = tcp->tcp_connp;
2861 2860 char buf[TCP_MAX_HDR_LENGTH];
2862 2861 uint_t buflen;
2863 2862 uint_t ulplen = TCP_MIN_HEADER_LENGTH;
2864 2863 uint_t extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
2865 2864 tcpha_t *tcpha;
2866 2865 uint32_t cksum;
2867 2866 int error;
2868 2867
2869 2868 /*
2870 2869 * We might be called after the connection is set up, and we might
2871 2870 * have TS options already in the TCP header. Thus we save any
2872 2871 * existing tcp header.
2873 2872 */
2874 2873 buflen = connp->conn_ht_ulp_len;
2875 2874 if (buflen != 0) {
2876 2875 bcopy(connp->conn_ht_ulp, buf, buflen);
2877 2876 extralen -= buflen - ulplen;
2878 2877 ulplen = buflen;
2879 2878 }
2880 2879
2881 2880 /* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
2882 2881 mutex_enter(&connp->conn_lock);
2883 2882 error = conn_build_hdr_template(connp, ulplen, extralen,
2884 2883 &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
2885 2884 mutex_exit(&connp->conn_lock);
2886 2885 if (error != 0)
2887 2886 return (error);
2888 2887
2889 2888 /*
2890 2889 * Any routing header/option has been massaged. The checksum difference
2891 2890 * is stored in conn_sum for later use.
2892 2891 */
2893 2892 tcpha = (tcpha_t *)connp->conn_ht_ulp;
2894 2893 tcp->tcp_tcpha = tcpha;
2895 2894
2896 2895 /* restore any old tcp header */
2897 2896 if (buflen != 0) {
2898 2897 bcopy(buf, connp->conn_ht_ulp, buflen);
2899 2898 } else {
2900 2899 tcpha->tha_sum = 0;
2901 2900 tcpha->tha_urp = 0;
2902 2901 tcpha->tha_ack = 0;
2903 2902 tcpha->tha_offset_and_reserved = (5 << 4);
2904 2903 tcpha->tha_lport = connp->conn_lport;
2905 2904 tcpha->tha_fport = connp->conn_fport;
2906 2905 }
2907 2906
2908 2907 /*
2909 2908 * IP wants our header length in the checksum field to
2910 2909 * allow it to perform a single pseudo-header+checksum
2911 2910 * calculation on behalf of TCP.
2912 2911 * Include the adjustment for a source route once IP_OPTIONS is set.
2913 2912 */
2914 2913 cksum = sizeof (tcpha_t) + connp->conn_sum;
2915 2914 cksum = (cksum >> 16) + (cksum & 0xFFFF);
2916 2915 ASSERT(cksum < 0x10000);
2917 2916 tcpha->tha_sum = htons(cksum);
2918 2917
2919 2918 if (connp->conn_ipversion == IPV4_VERSION)
2920 2919 tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
2921 2920 else
2922 2921 tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
2923 2922
2924 2923 if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
2925 2924 connp->conn_wroff) {
2926 2925 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2927 2926 tcps->tcps_wroff_xtra;
2928 2927 (void) proto_set_tx_wroff(connp->conn_rq, connp,
2929 2928 connp->conn_wroff);
2930 2929 }
2931 2930 return (0);
2932 2931 }
2933 2932
2934 2933 /*
2935 2934 * tcp_rwnd_set() is called to adjust the receive window to a desired value.
2936 2935 * We do not allow the receive window to shrink. After setting rwnd,
2937 2936 * set the flow control hiwat of the stream.
2938 2937 *
2939 2938 * This function is called in 2 cases:
2940 2939 *
2941 2940 * 1) Before data transfer begins, in tcp_input_listener() for accepting a
2942 2941 * connection (passive open) and in tcp_input_data() for active connect.
2943 2942 * This is called after tcp_mss_set() when the desired MSS value is known.
2944 2943 * This makes sure that our window size is a mutiple of the other side's
2945 2944 * MSS.
2946 2945 * 2) Handling SO_RCVBUF option.
2947 2946 *
2948 2947 * It is ASSUMED that the requested size is a multiple of the current MSS.
2949 2948 *
2950 2949 * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
2951 2950 * user requests so.
2952 2951 */
2953 2952 int
2954 2953 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
2955 2954 {
2956 2955 uint32_t mss = tcp->tcp_mss;
2957 2956 uint32_t old_max_rwnd;
2958 2957 uint32_t max_transmittable_rwnd;
2959 2958 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
2960 2959 tcp_stack_t *tcps = tcp->tcp_tcps;
2961 2960 conn_t *connp = tcp->tcp_connp;
2962 2961
2963 2962 /*
2964 2963 * Insist on a receive window that is at least
2965 2964 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
2966 2965 * funny TCP interactions of Nagle algorithm, SWS avoidance
2967 2966 * and delayed acknowledgement.
2968 2967 */
2969 2968 rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
2970 2969
2971 2970 if (tcp->tcp_fused) {
2972 2971 size_t sth_hiwat;
2973 2972 tcp_t *peer_tcp = tcp->tcp_loopback_peer;
2974 2973
2975 2974 ASSERT(peer_tcp != NULL);
2976 2975 sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
2977 2976 if (!tcp_detached) {
2978 2977 (void) proto_set_rx_hiwat(connp->conn_rq, connp,
2979 2978 sth_hiwat);
2980 2979 tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
2981 2980 }
2982 2981
2983 2982 /* Caller could have changed tcp_rwnd; update tha_win */
2984 2983 if (tcp->tcp_tcpha != NULL) {
2985 2984 tcp->tcp_tcpha->tha_win =
2986 2985 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2987 2986 }
2988 2987 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
2989 2988 tcp->tcp_cwnd_max = rwnd;
2990 2989
2991 2990 /*
2992 2991 * In the fusion case, the maxpsz stream head value of
2993 2992 * our peer is set according to its send buffer size
2994 2993 * and our receive buffer size; since the latter may
2995 2994 * have changed we need to update the peer's maxpsz.
2996 2995 */
2997 2996 (void) tcp_maxpsz_set(peer_tcp, B_TRUE);
2998 2997 return (sth_hiwat);
2999 2998 }
3000 2999
3001 3000 if (tcp_detached)
3002 3001 old_max_rwnd = tcp->tcp_rwnd;
3003 3002 else
3004 3003 old_max_rwnd = connp->conn_rcvbuf;
3005 3004
3006 3005
3007 3006 /*
3008 3007 * If window size info has already been exchanged, TCP should not
3009 3008 * shrink the window. Shrinking window is doable if done carefully.
3010 3009 * We may add that support later. But so far there is not a real
3011 3010 * need to do that.
3012 3011 */
3013 3012 if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
3014 3013 /* MSS may have changed, do a round up again. */
3015 3014 rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
3016 3015 }
3017 3016
3018 3017 /*
3019 3018 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3020 3019 * can be applied even before the window scale option is decided.
3021 3020 */
3022 3021 max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3023 3022 if (rwnd > max_transmittable_rwnd) {
3024 3023 rwnd = max_transmittable_rwnd -
3025 3024 (max_transmittable_rwnd % mss);
3026 3025 if (rwnd < mss)
3027 3026 rwnd = max_transmittable_rwnd;
3028 3027 /*
3029 3028 * If we're over the limit we may have to back down tcp_rwnd.
3030 3029 * The increment below won't work for us. So we set all three
3031 3030 * here and the increment below will have no effect.
3032 3031 */
3033 3032 tcp->tcp_rwnd = old_max_rwnd = rwnd;
3034 3033 }
3035 3034 if (tcp->tcp_localnet) {
3036 3035 tcp->tcp_rack_abs_max =
3037 3036 MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3038 3037 } else {
3039 3038 /*
3040 3039 * For a remote host on a different subnet (through a router),
3041 3040 * we ack every other packet to be conforming to RFC1122.
3042 3041 * tcp_deferred_acks_max is default to 2.
3043 3042 */
3044 3043 tcp->tcp_rack_abs_max =
3045 3044 MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3046 3045 }
3047 3046 if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3048 3047 tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3049 3048 else
3050 3049 tcp->tcp_rack_cur_max = 0;
3051 3050 /*
3052 3051 * Increment the current rwnd by the amount the maximum grew (we
3053 3052 * can not overwrite it since we might be in the middle of a
3054 3053 * connection.)
3055 3054 */
3056 3055 tcp->tcp_rwnd += rwnd - old_max_rwnd;
3057 3056 connp->conn_rcvbuf = rwnd;
3058 3057
3059 3058 /* Are we already connected? */
3060 3059 if (tcp->tcp_tcpha != NULL) {
3061 3060 tcp->tcp_tcpha->tha_win =
3062 3061 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3063 3062 }
3064 3063
3065 3064 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3066 3065 tcp->tcp_cwnd_max = rwnd;
3067 3066
3068 3067 if (tcp_detached)
3069 3068 return (rwnd);
3070 3069
3071 3070 tcp_set_recv_threshold(tcp, rwnd >> 3);
3072 3071
3073 3072 (void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3074 3073 return (rwnd);
3075 3074 }
3076 3075
3077 3076 int
3078 3077 tcp_do_unbind(conn_t *connp)
3079 3078 {
3080 3079 tcp_t *tcp = connp->conn_tcp;
3081 3080 int32_t oldstate;
3082 3081
3083 3082 switch (tcp->tcp_state) {
3084 3083 case TCPS_BOUND:
3085 3084 case TCPS_LISTEN:
3086 3085 break;
3087 3086 default:
3088 3087 return (-TOUTSTATE);
3089 3088 }
3090 3089
3091 3090 /*
3092 3091 * Need to clean up all the eagers since after the unbind, segments
3093 3092 * will no longer be delivered to this listener stream.
3094 3093 */
3095 3094 mutex_enter(&tcp->tcp_eager_lock);
3096 3095 if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3097 3096 tcp_eager_cleanup(tcp, 0);
3098 3097 }
3099 3098 mutex_exit(&tcp->tcp_eager_lock);
3100 3099
3101 3100 /* Clean up the listener connection counter if necessary. */
3102 3101 if (tcp->tcp_listen_cnt != NULL)
3103 3102 TCP_DECR_LISTEN_CNT(tcp);
3104 3103 connp->conn_laddr_v6 = ipv6_all_zeros;
3105 3104 connp->conn_saddr_v6 = ipv6_all_zeros;
3106 3105 tcp_bind_hash_remove(tcp);
3107 3106 oldstate = tcp->tcp_state;
3108 3107 tcp->tcp_state = TCPS_IDLE;
3109 3108 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
3110 3109 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
3111 3110 int32_t, oldstate);
3112 3111
3113 3112 ip_unbind(connp);
3114 3113 bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3115 3114
3116 3115 return (0);
3117 3116 }
3118 3117
3119 3118 /*
3120 3119 * Collect protocol properties to send to the upper handle.
3121 3120 */
3122 3121 void
3123 3122 tcp_get_proto_props(tcp_t *tcp, struct sock_proto_props *sopp)
3124 3123 {
3125 3124 conn_t *connp = tcp->tcp_connp;
3126 3125
3127 3126 sopp->sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3128 3127 sopp->sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3129 3128
3130 3129 sopp->sopp_rxhiwat = tcp->tcp_fused ?
3131 3130 tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3132 3131 connp->conn_rcvbuf;
3133 3132 /*
3134 3133 * Determine what write offset value to use depending on SACK and
3135 3134 * whether the endpoint is fused or not.
3136 3135 */
3137 3136 if (tcp->tcp_fused) {
3138 3137 ASSERT(tcp->tcp_loopback);
3139 3138 ASSERT(tcp->tcp_loopback_peer != NULL);
3140 3139 /*
3141 3140 * For fused tcp loopback, set the stream head's write
3142 3141 * offset value to zero since we won't be needing any room
3143 3142 * for TCP/IP headers. This would also improve performance
3144 3143 * since it would reduce the amount of work done by kmem.
3145 3144 * Non-fused tcp loopback case is handled separately below.
3146 3145 */
3147 3146 sopp->sopp_wroff = 0;
3148 3147 /*
3149 3148 * Update the peer's transmit parameters according to
3150 3149 * our recently calculated high water mark value.
3151 3150 */
3152 3151 (void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3153 3152 } else if (tcp->tcp_snd_sack_ok) {
3154 3153 sopp->sopp_wroff = connp->conn_ht_iphc_allocated +
3155 3154 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3156 3155 } else {
3157 3156 sopp->sopp_wroff = connp->conn_ht_iphc_len +
3158 3157 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3159 3158 }
3160 3159
3161 3160 if (tcp->tcp_loopback) {
3162 3161 sopp->sopp_flags |= SOCKOPT_LOOPBACK;
3163 3162 sopp->sopp_loopback = B_TRUE;
3164 3163 }
3165 3164 }
3166 3165
3167 3166 /*
3168 3167 * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3169 3168 */
3170 3169 boolean_t
3171 3170 tcp_zcopy_check(tcp_t *tcp)
3172 3171 {
3173 3172 conn_t *connp = tcp->tcp_connp;
3174 3173 ip_xmit_attr_t *ixa = connp->conn_ixa;
3175 3174 boolean_t zc_enabled = B_FALSE;
3176 3175 tcp_stack_t *tcps = tcp->tcp_tcps;
3177 3176
3178 3177 if (do_tcpzcopy == 2)
3179 3178 zc_enabled = B_TRUE;
3180 3179 else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3181 3180 zc_enabled = B_TRUE;
3182 3181
3183 3182 tcp->tcp_snd_zcopy_on = zc_enabled;
3184 3183 if (!TCP_IS_DETACHED(tcp)) {
3185 3184 if (zc_enabled) {
3186 3185 ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3187 3186 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3188 3187 ZCVMSAFE);
3189 3188 TCP_STAT(tcps, tcp_zcopy_on);
3190 3189 } else {
3191 3190 ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3192 3191 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3193 3192 ZCVMUNSAFE);
3194 3193 TCP_STAT(tcps, tcp_zcopy_off);
3195 3194 }
3196 3195 }
3197 3196 return (zc_enabled);
3198 3197 }
3199 3198
3200 3199 /*
3201 3200 * Backoff from a zero-copy message by copying data to a new allocated
3202 3201 * message and freeing the original desballoca'ed segmapped message.
3203 3202 *
3204 3203 * This function is called by following two callers:
3205 3204 * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3206 3205 * the origial desballoca'ed message and notify sockfs. This is in re-
3207 3206 * transmit state.
3208 3207 * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3209 3208 * to be copied to new message.
3210 3209 */
3211 3210 mblk_t *
3212 3211 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3213 3212 {
3214 3213 mblk_t *nbp;
3215 3214 mblk_t *head = NULL;
3216 3215 mblk_t *tail = NULL;
3217 3216 tcp_stack_t *tcps = tcp->tcp_tcps;
3218 3217
3219 3218 ASSERT(bp != NULL);
3220 3219 while (bp != NULL) {
3221 3220 if (IS_VMLOANED_MBLK(bp)) {
3222 3221 TCP_STAT(tcps, tcp_zcopy_backoff);
3223 3222 if ((nbp = copyb(bp)) == NULL) {
3224 3223 tcp->tcp_xmit_zc_clean = B_FALSE;
3225 3224 if (tail != NULL)
3226 3225 tail->b_cont = bp;
3227 3226 return ((head == NULL) ? bp : head);
3228 3227 }
3229 3228
3230 3229 if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3231 3230 if (fix_xmitlist)
3232 3231 tcp_zcopy_notify(tcp);
3233 3232 else
3234 3233 nbp->b_datap->db_struioflag |=
3235 3234 STRUIO_ZCNOTIFY;
3236 3235 }
3237 3236 nbp->b_cont = bp->b_cont;
3238 3237
3239 3238 /*
3240 3239 * Copy saved information and adjust tcp_xmit_tail
3241 3240 * if needed.
3242 3241 */
3243 3242 if (fix_xmitlist) {
3244 3243 nbp->b_prev = bp->b_prev;
3245 3244 nbp->b_next = bp->b_next;
3246 3245
3247 3246 if (tcp->tcp_xmit_tail == bp)
3248 3247 tcp->tcp_xmit_tail = nbp;
3249 3248 }
3250 3249
3251 3250 /* Free the original message. */
3252 3251 bp->b_prev = NULL;
3253 3252 bp->b_next = NULL;
3254 3253 freeb(bp);
3255 3254
3256 3255 bp = nbp;
3257 3256 }
3258 3257
3259 3258 if (head == NULL) {
3260 3259 head = bp;
3261 3260 }
3262 3261 if (tail == NULL) {
3263 3262 tail = bp;
3264 3263 } else {
3265 3264 tail->b_cont = bp;
3266 3265 tail = bp;
3267 3266 }
3268 3267
3269 3268 /* Move forward. */
3270 3269 bp = bp->b_cont;
3271 3270 }
3272 3271
3273 3272 if (fix_xmitlist) {
3274 3273 tcp->tcp_xmit_last = tail;
3275 3274 tcp->tcp_xmit_zc_clean = B_TRUE;
3276 3275 }
3277 3276
3278 3277 return (head);
3279 3278 }
3280 3279
3281 3280 void
3282 3281 tcp_zcopy_notify(tcp_t *tcp)
3283 3282 {
3284 3283 struct stdata *stp;
3285 3284 conn_t *connp;
3286 3285
3287 3286 if (tcp->tcp_detached)
3288 3287 return;
3289 3288 connp = tcp->tcp_connp;
3290 3289 if (IPCL_IS_NONSTR(connp)) {
3291 3290 (*connp->conn_upcalls->su_zcopy_notify)
3292 3291 (connp->conn_upper_handle);
3293 3292 return;
3294 3293 }
3295 3294 stp = STREAM(connp->conn_rq);
3296 3295 mutex_enter(&stp->sd_lock);
3297 3296 stp->sd_flag |= STZCNOTIFY;
3298 3297 cv_broadcast(&stp->sd_zcopy_wait);
3299 3298 mutex_exit(&stp->sd_lock);
3300 3299 }
3301 3300
3302 3301 /*
3303 3302 * Update the TCP connection according to change of LSO capability.
3304 3303 */
3305 3304 static void
3306 3305 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3307 3306 {
3308 3307 /*
3309 3308 * We check against IPv4 header length to preserve the old behavior
3310 3309 * of only enabling LSO when there are no IP options.
3311 3310 * But this restriction might not be necessary at all. Before removing
3312 3311 * it, need to verify how LSO is handled for source routing case, with
3313 3312 * which IP does software checksum.
3314 3313 *
3315 3314 * For IPv6, whenever any extension header is needed, LSO is supressed.
3316 3315 */
3317 3316 if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3318 3317 IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3319 3318 return;
3320 3319
3321 3320 /*
3322 3321 * Either the LSO capability newly became usable, or it has changed.
3323 3322 */
3324 3323 if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3325 3324 ill_lso_capab_t *lsoc = &ixa->ixa_lso_capab;
3326 3325
3327 3326 ASSERT(lsoc->ill_lso_max > 0);
3328 3327 tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lsoc->ill_lso_max);
3329 3328
3330 3329 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3331 3330 boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3332 3331
3333 3332 /*
3334 3333 * If LSO to be enabled, notify the STREAM header with larger
3335 3334 * data block.
3336 3335 */
3337 3336 if (!tcp->tcp_lso)
3338 3337 tcp->tcp_maxpsz_multiplier = 0;
3339 3338
3340 3339 tcp->tcp_lso = B_TRUE;
3341 3340 TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3342 3341 } else { /* LSO capability is not usable any more. */
3343 3342 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3344 3343 boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
3345 3344
3346 3345 /*
3347 3346 * If LSO to be disabled, notify the STREAM header with smaller
3348 3347 * data block. And need to restore fragsize to PMTU.
3349 3348 */
3350 3349 if (tcp->tcp_lso) {
3351 3350 tcp->tcp_maxpsz_multiplier =
3352 3351 tcp->tcp_tcps->tcps_maxpsz_multiplier;
3353 3352 ixa->ixa_fragsize = ixa->ixa_pmtu;
3354 3353 tcp->tcp_lso = B_FALSE;
3355 3354 TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
3356 3355 }
3357 3356 }
3358 3357
3359 3358 (void) tcp_maxpsz_set(tcp, B_TRUE);
3360 3359 }
3361 3360
3362 3361 /*
3363 3362 * Update the TCP connection according to change of ZEROCOPY capability.
3364 3363 */
3365 3364 static void
3366 3365 tcp_update_zcopy(tcp_t *tcp)
3367 3366 {
3368 3367 conn_t *connp = tcp->tcp_connp;
3369 3368 tcp_stack_t *tcps = tcp->tcp_tcps;
3370 3369
3371 3370 if (tcp->tcp_snd_zcopy_on) {
3372 3371 tcp->tcp_snd_zcopy_on = B_FALSE;
3373 3372 if (!TCP_IS_DETACHED(tcp)) {
3374 3373 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3375 3374 ZCVMUNSAFE);
3376 3375 TCP_STAT(tcps, tcp_zcopy_off);
3377 3376 }
3378 3377 } else {
3379 3378 tcp->tcp_snd_zcopy_on = B_TRUE;
3380 3379 if (!TCP_IS_DETACHED(tcp)) {
3381 3380 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3382 3381 ZCVMSAFE);
3383 3382 TCP_STAT(tcps, tcp_zcopy_on);
3384 3383 }
3385 3384 }
3386 3385 }
3387 3386
3388 3387 /*
3389 3388 * Notify function registered with ip_xmit_attr_t. It's called in the squeue
3390 3389 * so it's safe to update the TCP connection.
3391 3390 */
3392 3391 /* ARGSUSED1 */
3393 3392 static void
3394 3393 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
3395 3394 ixa_notify_arg_t narg)
3396 3395 {
3397 3396 tcp_t *tcp = (tcp_t *)arg;
3398 3397 conn_t *connp = tcp->tcp_connp;
3399 3398
3400 3399 switch (ntype) {
3401 3400 case IXAN_LSO:
3402 3401 tcp_update_lso(tcp, connp->conn_ixa);
3403 3402 break;
3404 3403 case IXAN_PMTU:
3405 3404 tcp_update_pmtu(tcp, B_FALSE);
3406 3405 break;
3407 3406 case IXAN_ZCOPY:
3408 3407 tcp_update_zcopy(tcp);
3409 3408 break;
3410 3409 default:
3411 3410 break;
3412 3411 }
3413 3412 }
3414 3413
3415 3414 /*
3416 3415 * The TCP write service routine should never be called...
3417 3416 */
3418 3417 /* ARGSUSED */
3419 3418 static void
3420 3419 tcp_wsrv(queue_t *q)
3421 3420 {
3422 3421 tcp_stack_t *tcps = Q_TO_TCP(q)->tcp_tcps;
3423 3422
3424 3423 TCP_STAT(tcps, tcp_wsrv_called);
3425 3424 }
3426 3425
3427 3426 /*
3428 3427 * Hash list lookup routine for tcp_t structures.
3429 3428 * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
3430 3429 */
3431 3430 tcp_t *
3432 3431 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
3433 3432 {
3434 3433 tf_t *tf;
3435 3434 tcp_t *tcp;
3436 3435
3437 3436 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3438 3437 mutex_enter(&tf->tf_lock);
3439 3438 for (tcp = tf->tf_tcp; tcp != NULL;
3440 3439 tcp = tcp->tcp_acceptor_hash) {
3441 3440 if (tcp->tcp_acceptor_id == id) {
3442 3441 CONN_INC_REF(tcp->tcp_connp);
3443 3442 mutex_exit(&tf->tf_lock);
3444 3443 return (tcp);
3445 3444 }
3446 3445 }
3447 3446 mutex_exit(&tf->tf_lock);
3448 3447 return (NULL);
3449 3448 }
3450 3449
3451 3450 /*
3452 3451 * Hash list insertion routine for tcp_t structures.
3453 3452 */
3454 3453 void
3455 3454 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
3456 3455 {
3457 3456 tf_t *tf;
3458 3457 tcp_t **tcpp;
3459 3458 tcp_t *tcpnext;
3460 3459 tcp_stack_t *tcps = tcp->tcp_tcps;
3461 3460
3462 3461 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3463 3462
3464 3463 if (tcp->tcp_ptpahn != NULL)
3465 3464 tcp_acceptor_hash_remove(tcp);
3466 3465 tcpp = &tf->tf_tcp;
3467 3466 mutex_enter(&tf->tf_lock);
3468 3467 tcpnext = tcpp[0];
3469 3468 if (tcpnext)
3470 3469 tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
3471 3470 tcp->tcp_acceptor_hash = tcpnext;
3472 3471 tcp->tcp_ptpahn = tcpp;
3473 3472 tcpp[0] = tcp;
3474 3473 tcp->tcp_acceptor_lockp = &tf->tf_lock; /* For tcp_*_hash_remove */
3475 3474 mutex_exit(&tf->tf_lock);
3476 3475 }
3477 3476
3478 3477 /*
3479 3478 * Hash list removal routine for tcp_t structures.
3480 3479 */
3481 3480 void
3482 3481 tcp_acceptor_hash_remove(tcp_t *tcp)
3483 3482 {
3484 3483 tcp_t *tcpnext;
3485 3484 kmutex_t *lockp;
3486 3485
3487 3486 /*
3488 3487 * Extract the lock pointer in case there are concurrent
3489 3488 * hash_remove's for this instance.
3490 3489 */
3491 3490 lockp = tcp->tcp_acceptor_lockp;
3492 3491
3493 3492 if (tcp->tcp_ptpahn == NULL)
3494 3493 return;
3495 3494
3496 3495 ASSERT(lockp != NULL);
3497 3496 mutex_enter(lockp);
3498 3497 if (tcp->tcp_ptpahn) {
3499 3498 tcpnext = tcp->tcp_acceptor_hash;
3500 3499 if (tcpnext) {
3501 3500 tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
3502 3501 tcp->tcp_acceptor_hash = NULL;
3503 3502 }
3504 3503 *tcp->tcp_ptpahn = tcpnext;
3505 3504 tcp->tcp_ptpahn = NULL;
3506 3505 }
3507 3506 mutex_exit(lockp);
3508 3507 tcp->tcp_acceptor_lockp = NULL;
3509 3508 }
3510 3509
3511 3510 /*
3512 3511 * Type three generator adapted from the random() function in 4.4 BSD:
3513 3512 */
3514 3513
3515 3514 /*
3516 3515 * Copyright (c) 1983, 1993
3517 3516 * The Regents of the University of California. All rights reserved.
3518 3517 *
3519 3518 * Redistribution and use in source and binary forms, with or without
3520 3519 * modification, are permitted provided that the following conditions
3521 3520 * are met:
3522 3521 * 1. Redistributions of source code must retain the above copyright
3523 3522 * notice, this list of conditions and the following disclaimer.
3524 3523 * 2. Redistributions in binary form must reproduce the above copyright
3525 3524 * notice, this list of conditions and the following disclaimer in the
3526 3525 * documentation and/or other materials provided with the distribution.
3527 3526 * 3. All advertising materials mentioning features or use of this software
3528 3527 * must display the following acknowledgement:
3529 3528 * This product includes software developed by the University of
3530 3529 * California, Berkeley and its contributors.
3531 3530 * 4. Neither the name of the University nor the names of its contributors
3532 3531 * may be used to endorse or promote products derived from this software
3533 3532 * without specific prior written permission.
3534 3533 *
3535 3534 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
3536 3535 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
3537 3536 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
3538 3537 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
3539 3538 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
3540 3539 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
3541 3540 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
3542 3541 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
3543 3542 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
3544 3543 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
3545 3544 * SUCH DAMAGE.
3546 3545 */
3547 3546
3548 3547 /* Type 3 -- x**31 + x**3 + 1 */
3549 3548 #define DEG_3 31
3550 3549 #define SEP_3 3
3551 3550
3552 3551
3553 3552 /* Protected by tcp_random_lock */
3554 3553 static int tcp_randtbl[DEG_3 + 1];
3555 3554
3556 3555 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
3557 3556 static int *tcp_random_rptr = &tcp_randtbl[1];
3558 3557
3559 3558 static int *tcp_random_state = &tcp_randtbl[1];
3560 3559 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
3561 3560
3562 3561 kmutex_t tcp_random_lock;
3563 3562
3564 3563 void
3565 3564 tcp_random_init(void)
3566 3565 {
3567 3566 int i;
3568 3567 hrtime_t hrt;
3569 3568 time_t wallclock;
3570 3569 uint64_t result;
3571 3570
3572 3571 /*
3573 3572 * Use high-res timer and current time for seed. Gethrtime() returns
3574 3573 * a longlong, which may contain resolution down to nanoseconds.
3575 3574 * The current time will either be a 32-bit or a 64-bit quantity.
3576 3575 * XOR the two together in a 64-bit result variable.
3577 3576 * Convert the result to a 32-bit value by multiplying the high-order
3578 3577 * 32-bits by the low-order 32-bits.
3579 3578 */
3580 3579
3581 3580 hrt = gethrtime();
3582 3581 (void) drv_getparm(TIME, &wallclock);
3583 3582 result = (uint64_t)wallclock ^ (uint64_t)hrt;
3584 3583 mutex_enter(&tcp_random_lock);
3585 3584 tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
3586 3585 (result & 0xffffffff);
3587 3586
3588 3587 for (i = 1; i < DEG_3; i++)
3589 3588 tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
3590 3589 + 12345;
3591 3590 tcp_random_fptr = &tcp_random_state[SEP_3];
3592 3591 tcp_random_rptr = &tcp_random_state[0];
3593 3592 mutex_exit(&tcp_random_lock);
3594 3593 for (i = 0; i < 10 * DEG_3; i++)
3595 3594 (void) tcp_random();
3596 3595 }
3597 3596
3598 3597 /*
3599 3598 * tcp_random: Return a random number in the range [1 - (128K + 1)].
3600 3599 * This range is selected to be approximately centered on TCP_ISS / 2,
3601 3600 * and easy to compute. We get this value by generating a 32-bit random
3602 3601 * number, selecting out the high-order 17 bits, and then adding one so
3603 3602 * that we never return zero.
3604 3603 */
3605 3604 int
3606 3605 tcp_random(void)
3607 3606 {
3608 3607 int i;
3609 3608
3610 3609 mutex_enter(&tcp_random_lock);
3611 3610 *tcp_random_fptr += *tcp_random_rptr;
3612 3611
3613 3612 /*
3614 3613 * The high-order bits are more random than the low-order bits,
3615 3614 * so we select out the high-order 17 bits and add one so that
3616 3615 * we never return zero.
3617 3616 */
3618 3617 i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
3619 3618 if (++tcp_random_fptr >= tcp_random_end_ptr) {
3620 3619 tcp_random_fptr = tcp_random_state;
3621 3620 ++tcp_random_rptr;
3622 3621 } else if (++tcp_random_rptr >= tcp_random_end_ptr)
3623 3622 tcp_random_rptr = tcp_random_state;
3624 3623
3625 3624 mutex_exit(&tcp_random_lock);
3626 3625 return (i);
3627 3626 }
3628 3627
3629 3628 /*
3630 3629 * Split this function out so that if the secret changes, I'm okay.
3631 3630 *
3632 3631 * Initialize the tcp_iss_cookie and tcp_iss_key.
3633 3632 */
3634 3633
3635 3634 #define PASSWD_SIZE 16 /* MUST be multiple of 4 */
3636 3635
3637 3636 void
3638 3637 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
3639 3638 {
3640 3639 struct {
3641 3640 int32_t current_time;
3642 3641 uint32_t randnum;
3643 3642 uint16_t pad;
3644 3643 uint8_t ether[6];
3645 3644 uint8_t passwd[PASSWD_SIZE];
3646 3645 } tcp_iss_cookie;
3647 3646 time_t t;
3648 3647
3649 3648 /*
3650 3649 * Start with the current absolute time.
3651 3650 */
3652 3651 (void) drv_getparm(TIME, &t);
3653 3652 tcp_iss_cookie.current_time = t;
3654 3653
3655 3654 /*
3656 3655 * XXX - Need a more random number per RFC 1750, not this crap.
3657 3656 * OTOH, if what follows is pretty random, then I'm in better shape.
3658 3657 */
3659 3658 tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
3660 3659 tcp_iss_cookie.pad = 0x365c; /* Picked from HMAC pad values. */
3661 3660
3662 3661 /*
3663 3662 * The cpu_type_info is pretty non-random. Ugggh. It does serve
3664 3663 * as a good template.
3665 3664 */
3666 3665 bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
3667 3666 min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
3668 3667
3669 3668 /*
3670 3669 * The pass-phrase. Normally this is supplied by user-called NDD.
3671 3670 */
3672 3671 bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
3673 3672
3674 3673 /*
3675 3674 * See 4010593 if this section becomes a problem again,
3676 3675 * but the local ethernet address is useful here.
3677 3676 */
3678 3677 (void) localetheraddr(NULL,
3679 3678 (struct ether_addr *)&tcp_iss_cookie.ether);
3680 3679
3681 3680 /*
3682 3681 * Hash 'em all together. The MD5Final is called per-connection.
3683 3682 */
3684 3683 mutex_enter(&tcps->tcps_iss_key_lock);
3685 3684 MD5Init(&tcps->tcps_iss_key);
3686 3685 MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
3687 3686 sizeof (tcp_iss_cookie));
3688 3687 mutex_exit(&tcps->tcps_iss_key_lock);
3689 3688 }
3690 3689
3691 3690 /*
3692 3691 * Called by IP when IP is loaded into the kernel
3693 3692 */
3694 3693 void
3695 3694 tcp_ddi_g_init(void)
3696 3695 {
3697 3696 tcp_timercache = kmem_cache_create("tcp_timercache",
3698 3697 sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
3699 3698 NULL, NULL, NULL, NULL, NULL, 0);
3700 3699
3701 3700 tcp_notsack_blk_cache = kmem_cache_create("tcp_notsack_blk_cache",
3702 3701 sizeof (notsack_blk_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3703 3702
3704 3703 mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
3705 3704
3706 3705 /* Initialize the random number generator */
3707 3706 tcp_random_init();
3708 3707
3709 3708 /* A single callback independently of how many netstacks we have */
3710 3709 ip_squeue_init(tcp_squeue_add);
3711 3710
3712 3711 tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
3713 3712
3714 3713 tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
3715 3714
3716 3715 /*
3717 3716 * We want to be informed each time a stack is created or
3718 3717 * destroyed in the kernel, so we can maintain the
3719 3718 * set of tcp_stack_t's.
3720 3719 */
3721 3720 netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
3722 3721 }
3723 3722
3724 3723
3725 3724 #define INET_NAME "ip"
3726 3725
3727 3726 /*
3728 3727 * Initialize the TCP stack instance.
3729 3728 */
3730 3729 static void *
3731 3730 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
3732 3731 {
3733 3732 tcp_stack_t *tcps;
3734 3733 int i;
3735 3734 int error = 0;
3736 3735 major_t major;
3737 3736 size_t arrsz;
3738 3737
3739 3738 tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
3740 3739 tcps->tcps_netstack = ns;
3741 3740
3742 3741 /* Initialize locks */
3743 3742 mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
3744 3743 mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
3745 3744
3746 3745 tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
3747 3746 tcps->tcps_g_epriv_ports[0] = ULP_DEF_EPRIV_PORT1;
3748 3747 tcps->tcps_g_epriv_ports[1] = ULP_DEF_EPRIV_PORT2;
3749 3748 tcps->tcps_min_anonpriv_port = 512;
3750 3749
3751 3750 tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
3752 3751 TCP_BIND_FANOUT_SIZE, KM_SLEEP);
3753 3752 tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
3754 3753 TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
3755 3754
3756 3755 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3757 3756 mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
3758 3757 MUTEX_DEFAULT, NULL);
3759 3758 }
3760 3759
3761 3760 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3762 3761 mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
3763 3762 MUTEX_DEFAULT, NULL);
3764 3763 }
3765 3764
3766 3765 /* TCP's IPsec code calls the packet dropper. */
3767 3766 ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
3768 3767
3769 3768 arrsz = tcp_propinfo_count * sizeof (mod_prop_info_t);
3770 3769 tcps->tcps_propinfo_tbl = (mod_prop_info_t *)kmem_alloc(arrsz,
3771 3770 KM_SLEEP);
3772 3771 bcopy(tcp_propinfo_tbl, tcps->tcps_propinfo_tbl, arrsz);
3773 3772
3774 3773 /*
3775 3774 * Note: To really walk the device tree you need the devinfo
3776 3775 * pointer to your device which is only available after probe/attach.
3777 3776 * The following is safe only because it uses ddi_root_node()
3778 3777 */
3779 3778 tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
3780 3779 tcp_opt_obj.odb_opt_arr_cnt);
3781 3780
3782 3781 /*
3783 3782 * Initialize RFC 1948 secret values. This will probably be reset once
3784 3783 * by the boot scripts.
3785 3784 *
3786 3785 * Use NULL name, as the name is caught by the new lockstats.
3787 3786 *
3788 3787 * Initialize with some random, non-guessable string, like the global
3789 3788 * T_INFO_ACK.
3790 3789 */
3791 3790
3792 3791 tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
3793 3792 sizeof (tcp_g_t_info_ack), tcps);
3794 3793
3795 3794 tcps->tcps_kstat = tcp_kstat2_init(stackid);
3796 3795 tcps->tcps_mibkp = tcp_kstat_init(stackid);
3797 3796
3798 3797 major = mod_name_to_major(INET_NAME);
3799 3798 error = ldi_ident_from_major(major, &tcps->tcps_ldi_ident);
3800 3799 ASSERT(error == 0);
3801 3800 tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
3802 3801 ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
3803 3802 cv_init(&tcps->tcps_ixa_cleanup_ready_cv, NULL, CV_DEFAULT, NULL);
3804 3803 cv_init(&tcps->tcps_ixa_cleanup_done_cv, NULL, CV_DEFAULT, NULL);
3805 3804 mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
3806 3805
3807 3806 mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
3808 3807 tcps->tcps_reclaim = B_FALSE;
3809 3808 tcps->tcps_reclaim_tid = 0;
3810 3809 tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
3811 3810
3812 3811 /*
3813 3812 * ncpus is the current number of CPUs, which can be bigger than
3814 3813 * boot_ncpus. But we don't want to use ncpus to allocate all the
3815 3814 * tcp_stats_cpu_t at system boot up time since it will be 1. While
3816 3815 * we handle adding CPU in tcp_cpu_update(), it will be slow if
3817 3816 * there are many CPUs as we will be adding them 1 by 1.
3818 3817 *
3819 3818 * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
3820 3819 * are not freed until the stack is going away. So there is no need
3821 3820 * to grab a lock to access the per CPU tcps_sc[x] pointer.
3822 3821 */
3823 3822 mutex_enter(&cpu_lock);
3824 3823 tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
3825 3824 mutex_exit(&cpu_lock);
3826 3825 tcps->tcps_sc = kmem_zalloc(max_ncpus * sizeof (tcp_stats_cpu_t *),
3827 3826 KM_SLEEP);
3828 3827 for (i = 0; i < tcps->tcps_sc_cnt; i++) {
3829 3828 tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
3830 3829 KM_SLEEP);
3831 3830 }
3832 3831
3833 3832 mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
3834 3833 list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
3835 3834 offsetof(tcp_listener_t, tl_link));
3836 3835
3837 3836 return (tcps);
3838 3837 }
3839 3838
3840 3839 /*
3841 3840 * Called when the IP module is about to be unloaded.
3842 3841 */
3843 3842 void
3844 3843 tcp_ddi_g_destroy(void)
3845 3844 {
3846 3845 tcp_g_kstat_fini(tcp_g_kstat);
3847 3846 tcp_g_kstat = NULL;
3848 3847 bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
3849 3848
3850 3849 mutex_destroy(&tcp_random_lock);
3851 3850
3852 3851 kmem_cache_destroy(tcp_timercache);
3853 3852 kmem_cache_destroy(tcp_notsack_blk_cache);
3854 3853
3855 3854 netstack_unregister(NS_TCP);
3856 3855 }
3857 3856
3858 3857 /*
3859 3858 * Free the TCP stack instance.
3860 3859 */
3861 3860 static void
3862 3861 tcp_stack_fini(netstackid_t stackid, void *arg)
3863 3862 {
3864 3863 tcp_stack_t *tcps = (tcp_stack_t *)arg;
3865 3864 int i;
3866 3865
3867 3866 freeb(tcps->tcps_ixa_cleanup_mp);
3868 3867 tcps->tcps_ixa_cleanup_mp = NULL;
3869 3868 cv_destroy(&tcps->tcps_ixa_cleanup_ready_cv);
3870 3869 cv_destroy(&tcps->tcps_ixa_cleanup_done_cv);
3871 3870 mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
3872 3871
3873 3872 /*
3874 3873 * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
3875 3874 * the timer.
3876 3875 */
3877 3876 mutex_enter(&tcps->tcps_reclaim_lock);
3878 3877 tcps->tcps_reclaim = B_FALSE;
3879 3878 mutex_exit(&tcps->tcps_reclaim_lock);
3880 3879 if (tcps->tcps_reclaim_tid != 0)
3881 3880 (void) untimeout(tcps->tcps_reclaim_tid);
3882 3881 mutex_destroy(&tcps->tcps_reclaim_lock);
3883 3882
3884 3883 tcp_listener_conf_cleanup(tcps);
3885 3884
3886 3885 for (i = 0; i < tcps->tcps_sc_cnt; i++)
3887 3886 kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
3888 3887 kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
3889 3888
3890 3889 kmem_free(tcps->tcps_propinfo_tbl,
3891 3890 tcp_propinfo_count * sizeof (mod_prop_info_t));
3892 3891 tcps->tcps_propinfo_tbl = NULL;
3893 3892
3894 3893 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3895 3894 ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
3896 3895 mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
3897 3896 }
3898 3897
3899 3898 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3900 3899 ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
3901 3900 mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
3902 3901 }
3903 3902
3904 3903 kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
3905 3904 tcps->tcps_bind_fanout = NULL;
3906 3905
3907 3906 kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
3908 3907 TCP_ACCEPTOR_FANOUT_SIZE);
3909 3908 tcps->tcps_acceptor_fanout = NULL;
3910 3909
3911 3910 mutex_destroy(&tcps->tcps_iss_key_lock);
3912 3911 mutex_destroy(&tcps->tcps_epriv_port_lock);
3913 3912
3914 3913 ip_drop_unregister(&tcps->tcps_dropper);
3915 3914
3916 3915 tcp_kstat2_fini(stackid, tcps->tcps_kstat);
3917 3916 tcps->tcps_kstat = NULL;
3918 3917
3919 3918 tcp_kstat_fini(stackid, tcps->tcps_mibkp);
3920 3919 tcps->tcps_mibkp = NULL;
3921 3920
3922 3921 ldi_ident_release(tcps->tcps_ldi_ident);
3923 3922 kmem_free(tcps, sizeof (*tcps));
3924 3923 }
3925 3924
3926 3925 /*
3927 3926 * Generate ISS, taking into account NDD changes may happen halfway through.
3928 3927 * (If the iss is not zero, set it.)
3929 3928 */
3930 3929
3931 3930 static void
3932 3931 tcp_iss_init(tcp_t *tcp)
3933 3932 {
3934 3933 MD5_CTX context;
3935 3934 struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
3936 3935 uint32_t answer[4];
3937 3936 tcp_stack_t *tcps = tcp->tcp_tcps;
3938 3937 conn_t *connp = tcp->tcp_connp;
3939 3938
3940 3939 tcps->tcps_iss_incr_extra += (tcps->tcps_iss_incr >> 1);
3941 3940 tcp->tcp_iss = tcps->tcps_iss_incr_extra;
3942 3941 switch (tcps->tcps_strong_iss) {
3943 3942 case 2:
3944 3943 mutex_enter(&tcps->tcps_iss_key_lock);
3945 3944 context = tcps->tcps_iss_key;
3946 3945 mutex_exit(&tcps->tcps_iss_key_lock);
3947 3946 arg.ports = connp->conn_ports;
3948 3947 arg.src = connp->conn_laddr_v6;
3949 3948 arg.dst = connp->conn_faddr_v6;
3950 3949 MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
3951 3950 MD5Final((uchar_t *)answer, &context);
3952 3951 tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
3953 3952 /*
3954 3953 * Now that we've hashed into a unique per-connection sequence
3955 3954 * space, add a random increment per strong_iss == 1. So I
3956 3955 * guess we'll have to...
3957 3956 */
3958 3957 /* FALLTHRU */
3959 3958 case 1:
3960 3959 tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
3961 3960 break;
3962 3961 default:
3963 3962 tcp->tcp_iss += (uint32_t)gethrestime_sec() *
3964 3963 tcps->tcps_iss_incr;
3965 3964 break;
3966 3965 }
3967 3966 tcp->tcp_valid_bits = TCP_ISS_VALID;
3968 3967 tcp->tcp_fss = tcp->tcp_iss - 1;
3969 3968 tcp->tcp_suna = tcp->tcp_iss;
3970 3969 tcp->tcp_snxt = tcp->tcp_iss + 1;
3971 3970 tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3972 3971 tcp->tcp_csuna = tcp->tcp_snxt;
3973 3972 }
3974 3973
3975 3974 /*
3976 3975 * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
3977 3976 * on the specified backing STREAMS q. Note, the caller may make the
3978 3977 * decision to call based on the tcp_t.tcp_flow_stopped value which
3979 3978 * when check outside the q's lock is only an advisory check ...
3980 3979 */
3981 3980 void
3982 3981 tcp_setqfull(tcp_t *tcp)
3983 3982 {
3984 3983 tcp_stack_t *tcps = tcp->tcp_tcps;
3985 3984 conn_t *connp = tcp->tcp_connp;
3986 3985
3987 3986 if (tcp->tcp_closed)
3988 3987 return;
3989 3988
3990 3989 conn_setqfull(connp, &tcp->tcp_flow_stopped);
3991 3990 if (tcp->tcp_flow_stopped)
3992 3991 TCP_STAT(tcps, tcp_flwctl_on);
3993 3992 }
3994 3993
3995 3994 void
3996 3995 tcp_clrqfull(tcp_t *tcp)
3997 3996 {
3998 3997 conn_t *connp = tcp->tcp_connp;
3999 3998
4000 3999 if (tcp->tcp_closed)
4001 4000 return;
4002 4001 conn_clrqfull(connp, &tcp->tcp_flow_stopped);
4003 4002 }
4004 4003
4005 4004 static int
4006 4005 tcp_squeue_switch(int val)
4007 4006 {
4008 4007 int rval = SQ_FILL;
4009 4008
4010 4009 switch (val) {
4011 4010 case 1:
4012 4011 rval = SQ_NODRAIN;
4013 4012 break;
4014 4013 case 2:
4015 4014 rval = SQ_PROCESS;
4016 4015 break;
4017 4016 default:
4018 4017 break;
4019 4018 }
4020 4019 return (rval);
4021 4020 }
4022 4021
4023 4022 /*
4024 4023 * This is called once for each squeue - globally for all stack
4025 4024 * instances.
4026 4025 */
4027 4026 static void
4028 4027 tcp_squeue_add(squeue_t *sqp)
4029 4028 {
4030 4029 tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4031 4030 sizeof (tcp_squeue_priv_t), KM_SLEEP);
4032 4031
4033 4032 *squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4034 4033 if (tcp_free_list_max_cnt == 0) {
4035 4034 int tcp_ncpus = ((boot_max_ncpus == -1) ?
4036 4035 max_ncpus : boot_max_ncpus);
4037 4036
4038 4037 /*
4039 4038 * Limit number of entries to 1% of availble memory / tcp_ncpus
4040 4039 */
4041 4040 tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4042 4041 (tcp_ncpus * sizeof (tcp_t) * 100);
4043 4042 }
4044 4043 tcp_time_wait->tcp_free_list_cnt = 0;
4045 4044 }
4046 4045 /*
4047 4046 * Return unix error is tli error is TSYSERR, otherwise return a negative
4048 4047 * tli error.
4049 4048 */
4050 4049 int
4051 4050 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4052 4051 boolean_t bind_to_req_port_only)
4053 4052 {
4054 4053 int error;
4055 4054 tcp_t *tcp = connp->conn_tcp;
4056 4055
4057 4056 if (tcp->tcp_state >= TCPS_BOUND) {
4058 4057 if (connp->conn_debug) {
4059 4058 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4060 4059 "tcp_bind: bad state, %d", tcp->tcp_state);
4061 4060 }
4062 4061 return (-TOUTSTATE);
4063 4062 }
4064 4063
4065 4064 error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4066 4065 if (error != 0)
4067 4066 return (error);
4068 4067
4069 4068 ASSERT(tcp->tcp_state == TCPS_BOUND);
4070 4069 tcp->tcp_conn_req_max = 0;
4071 4070 return (0);
4072 4071 }
4073 4072
4074 4073 /*
4075 4074 * If the return value from this function is positive, it's a UNIX error.
4076 4075 * Otherwise, if it's negative, then the absolute value is a TLI error.
4077 4076 * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4078 4077 */
4079 4078 int
4080 4079 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4081 4080 cred_t *cr, pid_t pid)
4082 4081 {
4083 4082 tcp_t *tcp = connp->conn_tcp;
4084 4083 sin_t *sin = (sin_t *)sa;
4085 4084 sin6_t *sin6 = (sin6_t *)sa;
4086 4085 ipaddr_t *dstaddrp;
4087 4086 in_port_t dstport;
4088 4087 uint_t srcid;
4089 4088 int error;
4090 4089 uint32_t mss;
4091 4090 mblk_t *syn_mp;
4092 4091 tcp_stack_t *tcps = tcp->tcp_tcps;
4093 4092 int32_t oldstate;
4094 4093 ip_xmit_attr_t *ixa = connp->conn_ixa;
4095 4094
4096 4095 oldstate = tcp->tcp_state;
4097 4096
4098 4097 switch (len) {
4099 4098 default:
4100 4099 /*
4101 4100 * Should never happen
4102 4101 */
4103 4102 return (EINVAL);
4104 4103
4105 4104 case sizeof (sin_t):
4106 4105 sin = (sin_t *)sa;
4107 4106 if (sin->sin_port == 0) {
4108 4107 return (-TBADADDR);
4109 4108 }
4110 4109 if (connp->conn_ipv6_v6only) {
4111 4110 return (EAFNOSUPPORT);
4112 4111 }
4113 4112 break;
4114 4113
4115 4114 case sizeof (sin6_t):
4116 4115 sin6 = (sin6_t *)sa;
4117 4116 if (sin6->sin6_port == 0) {
4118 4117 return (-TBADADDR);
4119 4118 }
4120 4119 break;
4121 4120 }
4122 4121 /*
4123 4122 * If we're connecting to an IPv4-mapped IPv6 address, we need to
4124 4123 * make sure that the conn_ipversion is IPV4_VERSION. We
4125 4124 * need to this before we call tcp_bindi() so that the port lookup
4126 4125 * code will look for ports in the correct port space (IPv4 and
4127 4126 * IPv6 have separate port spaces).
4128 4127 */
4129 4128 if (connp->conn_family == AF_INET6 &&
4130 4129 connp->conn_ipversion == IPV6_VERSION &&
4131 4130 IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4132 4131 if (connp->conn_ipv6_v6only)
4133 4132 return (EADDRNOTAVAIL);
4134 4133
4135 4134 connp->conn_ipversion = IPV4_VERSION;
4136 4135 }
4137 4136
4138 4137 switch (tcp->tcp_state) {
4139 4138 case TCPS_LISTEN:
4140 4139 /*
4141 4140 * Listening sockets are not allowed to issue connect().
4142 4141 */
4143 4142 if (IPCL_IS_NONSTR(connp))
4144 4143 return (EOPNOTSUPP);
4145 4144 /* FALLTHRU */
4146 4145 case TCPS_IDLE:
4147 4146 /*
4148 4147 * We support quick connect, refer to comments in
4149 4148 * tcp_connect_*()
4150 4149 */
4151 4150 /* FALLTHRU */
4152 4151 case TCPS_BOUND:
4153 4152 break;
4154 4153 default:
4155 4154 return (-TOUTSTATE);
4156 4155 }
4157 4156
4158 4157 /*
4159 4158 * We update our cred/cpid based on the caller of connect
4160 4159 */
4161 4160 if (connp->conn_cred != cr) {
4162 4161 crhold(cr);
4163 4162 crfree(connp->conn_cred);
4164 4163 connp->conn_cred = cr;
4165 4164 }
4166 4165 connp->conn_cpid = pid;
4167 4166
4168 4167 /* Cache things in the ixa without any refhold */
4169 4168 ASSERT(!(ixa->ixa_free_flags & IXA_FREE_CRED));
4170 4169 ixa->ixa_cred = cr;
4171 4170 ixa->ixa_cpid = pid;
4172 4171 if (is_system_labeled()) {
4173 4172 /* We need to restart with a label based on the cred */
4174 4173 ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4175 4174 }
4176 4175
4177 4176 if (connp->conn_family == AF_INET6) {
4178 4177 if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4179 4178 error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4180 4179 sin6->sin6_port, sin6->sin6_flowinfo,
4181 4180 sin6->__sin6_src_id, sin6->sin6_scope_id);
4182 4181 } else {
4183 4182 /*
4184 4183 * Destination adress is mapped IPv6 address.
4185 4184 * Source bound address should be unspecified or
4186 4185 * IPv6 mapped address as well.
4187 4186 */
4188 4187 if (!IN6_IS_ADDR_UNSPECIFIED(
4189 4188 &connp->conn_bound_addr_v6) &&
4190 4189 !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4191 4190 return (EADDRNOTAVAIL);
4192 4191 }
4193 4192 dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4194 4193 dstport = sin6->sin6_port;
4195 4194 srcid = sin6->__sin6_src_id;
4196 4195 error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4197 4196 srcid);
4198 4197 }
4199 4198 } else {
4200 4199 dstaddrp = &sin->sin_addr.s_addr;
4201 4200 dstport = sin->sin_port;
4202 4201 srcid = 0;
4203 4202 error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4204 4203 }
4205 4204
4206 4205 if (error != 0)
4207 4206 goto connect_failed;
4208 4207
4209 4208 CL_INET_CONNECT(connp, B_TRUE, error);
4210 4209 if (error != 0)
4211 4210 goto connect_failed;
4212 4211
4213 4212 /* connect succeeded */
4214 4213 TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4215 4214 tcp->tcp_active_open = 1;
4216 4215
4217 4216 /*
4218 4217 * tcp_set_destination() does not adjust for TCP/IP header length.
4219 4218 */
4220 4219 mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4221 4220
4222 4221 /*
4223 4222 * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4224 4223 * to the nearest MSS.
4225 4224 *
4226 4225 * We do the round up here because we need to get the interface MTU
4227 4226 * first before we can do the round up.
4228 4227 */
4229 4228 tcp->tcp_rwnd = connp->conn_rcvbuf;
4230 4229 tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4231 4230 tcps->tcps_recv_hiwat_minmss * mss);
4232 4231 connp->conn_rcvbuf = tcp->tcp_rwnd;
4233 4232 tcp_set_ws_value(tcp);
4234 4233 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4235 4234 if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4236 4235 tcp->tcp_snd_ws_ok = B_TRUE;
4237 4236
4238 4237 /*
4239 4238 * Set tcp_snd_ts_ok to true
4240 4239 * so that tcp_xmit_mp will
4241 4240 * include the timestamp
4242 4241 * option in the SYN segment.
4243 4242 */
4244 4243 if (tcps->tcps_tstamp_always ||
4245 4244 (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4246 4245 tcp->tcp_snd_ts_ok = B_TRUE;
4247 4246 }
4248 4247
4249 4248 /*
4250 4249 * Note that tcp_snd_sack_ok can be set in tcp_set_destination() if
4251 4250 * the SACK metric is set. So here we just check the per stack SACK
4252 4251 * permitted param.
4253 4252 */
4254 4253 if (tcps->tcps_sack_permitted == 2) {
4255 4254 ASSERT(tcp->tcp_num_sack_blk == 0);
4256 4255 ASSERT(tcp->tcp_notsack_list == NULL);
4257 4256 tcp->tcp_snd_sack_ok = B_TRUE;
4258 4257 }
4259 4258
4260 4259 /*
4261 4260 * Should we use ECN? Note that the current
4262 4261 * default value (SunOS 5.9) of tcp_ecn_permitted
4263 4262 * is 1. The reason for doing this is that there
4264 4263 * are equipments out there that will drop ECN
4265 4264 * enabled IP packets. Setting it to 1 avoids
4266 4265 * compatibility problems.
4267 4266 */
4268 4267 if (tcps->tcps_ecn_permitted == 2)
4269 4268 tcp->tcp_ecn_ok = B_TRUE;
4270 4269
4271 4270 /* Trace change from BOUND -> SYN_SENT here */
4272 4271 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4273 4272 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4274 4273 int32_t, TCPS_BOUND);
4275 4274
4276 4275 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4277 4276 syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4278 4277 tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4279 4278 if (syn_mp != NULL) {
4280 4279 /*
4281 4280 * We must bump the generation before sending the syn
4282 4281 * to ensure that we use the right generation in case
4283 4282 * this thread issues a "connected" up call.
4284 4283 */
4285 4284 SOCK_CONNID_BUMP(tcp->tcp_connid);
4286 4285 /*
4287 4286 * DTrace sending the first SYN as a
4288 4287 * tcp:::connect-request event.
4289 4288 */
4290 4289 DTRACE_TCP5(connect__request, mblk_t *, NULL,
4291 4290 ip_xmit_attr_t *, connp->conn_ixa,
4292 4291 void_ip_t *, syn_mp->b_rptr, tcp_t *, tcp,
4293 4292 tcph_t *,
4294 4293 &syn_mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
4295 4294 tcp_send_data(tcp, syn_mp);
4296 4295 }
4297 4296
4298 4297 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4299 4298 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4300 4299 return (0);
4301 4300
4302 4301 connect_failed:
4303 4302 connp->conn_faddr_v6 = ipv6_all_zeros;
4304 4303 connp->conn_fport = 0;
4305 4304 tcp->tcp_state = oldstate;
4306 4305 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4307 4306 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4308 4307 return (error);
4309 4308 }
4310 4309
4311 4310 int
4312 4311 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4313 4312 int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4314 4313 {
4315 4314 tcp_t *tcp = connp->conn_tcp;
4316 4315 int error = 0;
4317 4316 tcp_stack_t *tcps = tcp->tcp_tcps;
4318 4317 int32_t oldstate;
4319 4318
4320 4319 /* All Solaris components should pass a cred for this operation. */
4321 4320 ASSERT(cr != NULL);
4322 4321
4323 4322 if (tcp->tcp_state >= TCPS_BOUND) {
4324 4323 if ((tcp->tcp_state == TCPS_BOUND ||
4325 4324 tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
4326 4325 /*
4327 4326 * Handle listen() increasing backlog.
4328 4327 * This is more "liberal" then what the TPI spec
4329 4328 * requires but is needed to avoid a t_unbind
4330 4329 * when handling listen() since the port number
4331 4330 * might be "stolen" between the unbind and bind.
4332 4331 */
4333 4332 goto do_listen;
4334 4333 }
4335 4334 if (connp->conn_debug) {
4336 4335 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4337 4336 "tcp_listen: bad state, %d", tcp->tcp_state);
4338 4337 }
4339 4338 return (-TOUTSTATE);
4340 4339 } else {
4341 4340 if (sa == NULL) {
4342 4341 sin6_t addr;
4343 4342 sin_t *sin;
4344 4343 sin6_t *sin6;
4345 4344
4346 4345 ASSERT(IPCL_IS_NONSTR(connp));
4347 4346 /* Do an implicit bind: Request for a generic port. */
4348 4347 if (connp->conn_family == AF_INET) {
4349 4348 len = sizeof (sin_t);
4350 4349 sin = (sin_t *)&addr;
4351 4350 *sin = sin_null;
4352 4351 sin->sin_family = AF_INET;
4353 4352 } else {
4354 4353 ASSERT(connp->conn_family == AF_INET6);
4355 4354 len = sizeof (sin6_t);
4356 4355 sin6 = (sin6_t *)&addr;
4357 4356 *sin6 = sin6_null;
4358 4357 sin6->sin6_family = AF_INET6;
4359 4358 }
4360 4359 sa = (struct sockaddr *)&addr;
4361 4360 }
4362 4361
4363 4362 error = tcp_bind_check(connp, sa, len, cr,
4364 4363 bind_to_req_port_only);
4365 4364 if (error)
4366 4365 return (error);
4367 4366 /* Fall through and do the fanout insertion */
4368 4367 }
4369 4368
4370 4369 do_listen:
4371 4370 ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
4372 4371 tcp->tcp_conn_req_max = backlog;
4373 4372 if (tcp->tcp_conn_req_max) {
4374 4373 if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
4375 4374 tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
4376 4375 if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
4377 4376 tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
4378 4377 /*
4379 4378 * If this is a listener, do not reset the eager list
4380 4379 * and other stuffs. Note that we don't check if the
4381 4380 * existing eager list meets the new tcp_conn_req_max
4382 4381 * requirement.
4383 4382 */
4384 4383 if (tcp->tcp_state != TCPS_LISTEN) {
4385 4384 tcp->tcp_state = TCPS_LISTEN;
4386 4385 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4387 4386 connp->conn_ixa, void, NULL, tcp_t *, tcp,
4388 4387 void, NULL, int32_t, TCPS_BOUND);
4389 4388 /* Initialize the chain. Don't need the eager_lock */
4390 4389 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
4391 4390 tcp->tcp_eager_next_drop_q0 = tcp;
4392 4391 tcp->tcp_eager_prev_drop_q0 = tcp;
4393 4392 tcp->tcp_second_ctimer_threshold =
4394 4393 tcps->tcps_ip_abort_linterval;
4395 4394 }
4396 4395 }
4397 4396
4398 4397 /*
4399 4398 * We need to make sure that the conn_recv is set to a non-null
4400 4399 * value before we insert the conn into the classifier table.
4401 4400 * This is to avoid a race with an incoming packet which does an
4402 4401 * ipcl_classify().
4403 4402 * We initially set it to tcp_input_listener_unbound to try to
4404 4403 * pick a good squeue for the listener when the first SYN arrives.
4405 4404 * tcp_input_listener_unbound sets it to tcp_input_listener on that
4406 4405 * first SYN.
4407 4406 */
4408 4407 connp->conn_recv = tcp_input_listener_unbound;
4409 4408
4410 4409 /* Insert the listener in the classifier table */
4411 4410 error = ip_laddr_fanout_insert(connp);
4412 4411 if (error != 0) {
4413 4412 /* Undo the bind - release the port number */
4414 4413 oldstate = tcp->tcp_state;
4415 4414 tcp->tcp_state = TCPS_IDLE;
4416 4415 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4417 4416 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4418 4417 int32_t, oldstate);
4419 4418 connp->conn_bound_addr_v6 = ipv6_all_zeros;
4420 4419
4421 4420 connp->conn_laddr_v6 = ipv6_all_zeros;
4422 4421 connp->conn_saddr_v6 = ipv6_all_zeros;
4423 4422 connp->conn_ports = 0;
4424 4423
4425 4424 if (connp->conn_anon_port) {
4426 4425 zone_t *zone;
4427 4426
4428 4427 zone = crgetzone(cr);
4429 4428 connp->conn_anon_port = B_FALSE;
4430 4429 (void) tsol_mlp_anon(zone, connp->conn_mlp_type,
4431 4430 connp->conn_proto, connp->conn_lport, B_FALSE);
4432 4431 }
4433 4432 connp->conn_mlp_type = mlptSingle;
4434 4433
4435 4434 tcp_bind_hash_remove(tcp);
4436 4435 return (error);
4437 4436 } else {
4438 4437 /*
4439 4438 * If there is a connection limit, allocate and initialize
4440 4439 * the counter struct. Note that since listen can be called
4441 4440 * multiple times, the struct may have been allready allocated.
4442 4441 */
4443 4442 if (!list_is_empty(&tcps->tcps_listener_conf) &&
4444 4443 tcp->tcp_listen_cnt == NULL) {
4445 4444 tcp_listen_cnt_t *tlc;
4446 4445 uint32_t ratio;
4447 4446
4448 4447 ratio = tcp_find_listener_conf(tcps,
4449 4448 ntohs(connp->conn_lport));
4450 4449 if (ratio != 0) {
4451 4450 uint32_t mem_ratio, tot_buf;
4452 4451
4453 4452 tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
4454 4453 KM_SLEEP);
4455 4454 /*
4456 4455 * Calculate the connection limit based on
4457 4456 * the configured ratio and maxusers. Maxusers
4458 4457 * are calculated based on memory size,
4459 4458 * ~ 1 user per MB. Note that the conn_rcvbuf
4460 4459 * and conn_sndbuf may change after a
4461 4460 * connection is accepted. So what we have
4462 4461 * is only an approximation.
4463 4462 */
4464 4463 if ((tot_buf = connp->conn_rcvbuf +
4465 4464 connp->conn_sndbuf) < MB) {
4466 4465 mem_ratio = MB / tot_buf;
4467 4466 tlc->tlc_max = maxusers / ratio *
4468 4467 mem_ratio;
4469 4468 } else {
4470 4469 mem_ratio = tot_buf / MB;
4471 4470 tlc->tlc_max = maxusers / ratio /
4472 4471 mem_ratio;
4473 4472 }
4474 4473 /* At least we should allow two connections! */
4475 4474 if (tlc->tlc_max <= tcp_min_conn_listener)
4476 4475 tlc->tlc_max = tcp_min_conn_listener;
4477 4476 tlc->tlc_cnt = 1;
4478 4477 tlc->tlc_drop = 0;
4479 4478 tcp->tcp_listen_cnt = tlc;
4480 4479 }
4481 4480 }
4482 4481 }
4483 4482 return (error);
4484 4483 }
|
↓ open down ↓ |
2055 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX