Print this page
916 False-sharing in TCP hash buckets, plus size increase
(originally Joyent OS-299 tcp_bindi() hash lock contention affects
TCP client performance)
| 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
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
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
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 + * Copyright (c) 2011, Joyent Inc. All rights reserved.
24 25 */
25 26 /* Copyright (c) 1990 Mentat Inc. */
26 27
27 28 #include <sys/types.h>
28 29 #include <sys/stream.h>
29 30 #include <sys/strsun.h>
30 31 #include <sys/strsubr.h>
31 32 #include <sys/stropts.h>
32 33 #include <sys/strlog.h>
33 34 #define _SUN_TPI_VERSION 2
34 35 #include <sys/tihdr.h>
35 36 #include <sys/timod.h>
36 37 #include <sys/ddi.h>
37 38 #include <sys/sunddi.h>
38 39 #include <sys/suntpi.h>
39 40 #include <sys/xti_inet.h>
40 41 #include <sys/cmn_err.h>
41 42 #include <sys/debug.h>
42 43 #include <sys/sdt.h>
43 44 #include <sys/vtrace.h>
44 45 #include <sys/kmem.h>
45 46 #include <sys/ethernet.h>
46 47 #include <sys/cpuvar.h>
47 48 #include <sys/dlpi.h>
48 49 #include <sys/pattr.h>
49 50 #include <sys/policy.h>
50 51 #include <sys/priv.h>
51 52 #include <sys/zone.h>
52 53 #include <sys/sunldi.h>
53 54
54 55 #include <sys/errno.h>
55 56 #include <sys/signal.h>
56 57 #include <sys/socket.h>
57 58 #include <sys/socketvar.h>
58 59 #include <sys/sockio.h>
59 60 #include <sys/isa_defs.h>
60 61 #include <sys/md5.h>
61 62 #include <sys/random.h>
62 63 #include <sys/uio.h>
63 64 #include <sys/systm.h>
64 65 #include <netinet/in.h>
65 66 #include <netinet/tcp.h>
66 67 #include <netinet/ip6.h>
67 68 #include <netinet/icmp6.h>
68 69 #include <net/if.h>
69 70 #include <net/route.h>
70 71 #include <inet/ipsec_impl.h>
71 72
72 73 #include <inet/common.h>
73 74 #include <inet/ip.h>
74 75 #include <inet/ip_impl.h>
75 76 #include <inet/ip6.h>
76 77 #include <inet/ip_ndp.h>
77 78 #include <inet/proto_set.h>
78 79 #include <inet/mib2.h>
79 80 #include <inet/optcom.h>
80 81 #include <inet/snmpcom.h>
81 82 #include <inet/kstatcom.h>
82 83 #include <inet/tcp.h>
83 84 #include <inet/tcp_impl.h>
84 85 #include <inet/tcp_cluster.h>
85 86 #include <inet/udp_impl.h>
86 87 #include <net/pfkeyv2.h>
87 88 #include <inet/ipdrop.h>
88 89
89 90 #include <inet/ipclassifier.h>
90 91 #include <inet/ip_ire.h>
91 92 #include <inet/ip_ftable.h>
92 93 #include <inet/ip_if.h>
93 94 #include <inet/ipp_common.h>
94 95 #include <inet/ip_rts.h>
95 96 #include <inet/ip_netinfo.h>
96 97 #include <sys/squeue_impl.h>
97 98 #include <sys/squeue.h>
98 99 #include <sys/tsol/label.h>
99 100 #include <sys/tsol/tnet.h>
100 101 #include <rpc/pmap_prot.h>
101 102 #include <sys/callo.h>
102 103
103 104 /*
104 105 * TCP Notes: aka FireEngine Phase I (PSARC 2002/433)
105 106 *
106 107 * (Read the detailed design doc in PSARC case directory)
107 108 *
108 109 * The entire tcp state is contained in tcp_t and conn_t structure
109 110 * which are allocated in tandem using ipcl_conn_create() and passing
110 111 * IPCL_TCPCONN as a flag. We use 'conn_ref' and 'conn_lock' to protect
111 112 * the references on the tcp_t. The tcp_t structure is never compressed
112 113 * and packets always land on the correct TCP perimeter from the time
113 114 * eager is created till the time tcp_t dies (as such the old mentat
114 115 * TCP global queue is not used for detached state and no IPSEC checking
115 116 * is required). The global queue is still allocated to send out resets
116 117 * for connection which have no listeners and IP directly calls
117 118 * tcp_xmit_listeners_reset() which does any policy check.
118 119 *
119 120 * Protection and Synchronisation mechanism:
120 121 *
121 122 * The tcp data structure does not use any kind of lock for protecting
122 123 * its state but instead uses 'squeues' for mutual exclusion from various
123 124 * read and write side threads. To access a tcp member, the thread should
124 125 * always be behind squeue (via squeue_enter with flags as SQ_FILL, SQ_PROCESS,
125 126 * or SQ_NODRAIN). Since the squeues allow a direct function call, caller
126 127 * can pass any tcp function having prototype of edesc_t as argument
127 128 * (different from traditional STREAMs model where packets come in only
128 129 * designated entry points). The list of functions that can be directly
129 130 * called via squeue are listed before the usual function prototype.
130 131 *
131 132 * Referencing:
132 133 *
133 134 * TCP is MT-Hot and we use a reference based scheme to make sure that the
134 135 * tcp structure doesn't disappear when its needed. When the application
135 136 * creates an outgoing connection or accepts an incoming connection, we
136 137 * start out with 2 references on 'conn_ref'. One for TCP and one for IP.
137 138 * The IP reference is just a symbolic reference since ip_tcpclose()
138 139 * looks at tcp structure after tcp_close_output() returns which could
139 140 * have dropped the last TCP reference. So as long as the connection is
140 141 * in attached state i.e. !TCP_IS_DETACHED, we have 2 references on the
141 142 * conn_t. The classifier puts its own reference when the connection is
142 143 * inserted in listen or connected hash. Anytime a thread needs to enter
143 144 * the tcp connection perimeter, it retrieves the conn/tcp from q->ptr
144 145 * on write side or by doing a classify on read side and then puts a
145 146 * reference on the conn before doing squeue_enter/tryenter/fill. For
146 147 * read side, the classifier itself puts the reference under fanout lock
147 148 * to make sure that tcp can't disappear before it gets processed. The
148 149 * squeue will drop this reference automatically so the called function
149 150 * doesn't have to do a DEC_REF.
150 151 *
151 152 * Opening a new connection:
152 153 *
153 154 * The outgoing connection open is pretty simple. tcp_open() does the
154 155 * work in creating the conn/tcp structure and initializing it. The
155 156 * squeue assignment is done based on the CPU the application
156 157 * is running on. So for outbound connections, processing is always done
157 158 * on application CPU which might be different from the incoming CPU
158 159 * being interrupted by the NIC. An optimal way would be to figure out
159 160 * the NIC <-> CPU binding at listen time, and assign the outgoing
160 161 * connection to the squeue attached to the CPU that will be interrupted
161 162 * for incoming packets (we know the NIC based on the bind IP address).
162 163 * This might seem like a problem if more data is going out but the
163 164 * fact is that in most cases the transmit is ACK driven transmit where
164 165 * the outgoing data normally sits on TCP's xmit queue waiting to be
165 166 * transmitted.
166 167 *
167 168 * Accepting a connection:
168 169 *
169 170 * This is a more interesting case because of various races involved in
170 171 * establishing a eager in its own perimeter. Read the meta comment on
171 172 * top of tcp_input_listener(). But briefly, the squeue is picked by
172 173 * ip_fanout based on the ring or the sender (if loopback).
173 174 *
174 175 * Closing a connection:
175 176 *
176 177 * The close is fairly straight forward. tcp_close() calls tcp_close_output()
177 178 * via squeue to do the close and mark the tcp as detached if the connection
178 179 * was in state TCPS_ESTABLISHED or greater. In the later case, TCP keep its
179 180 * reference but tcp_close() drop IP's reference always. So if tcp was
180 181 * not killed, it is sitting in time_wait list with 2 reference - 1 for TCP
181 182 * and 1 because it is in classifier's connected hash. This is the condition
182 183 * we use to determine that its OK to clean up the tcp outside of squeue
183 184 * when time wait expires (check the ref under fanout and conn_lock and
184 185 * if it is 2, remove it from fanout hash and kill it).
185 186 *
186 187 * Although close just drops the necessary references and marks the
187 188 * tcp_detached state, tcp_close needs to know the tcp_detached has been
188 189 * set (under squeue) before letting the STREAM go away (because a
189 190 * inbound packet might attempt to go up the STREAM while the close
190 191 * has happened and tcp_detached is not set). So a special lock and
191 192 * flag is used along with a condition variable (tcp_closelock, tcp_closed,
192 193 * and tcp_closecv) to signal tcp_close that tcp_close_out() has marked
193 194 * tcp_detached.
194 195 *
195 196 * Special provisions and fast paths:
196 197 *
197 198 * We make special provisions for sockfs by marking tcp_issocket
198 199 * whenever we have only sockfs on top of TCP. This allows us to skip
199 200 * putting the tcp in acceptor hash since a sockfs listener can never
200 201 * become acceptor and also avoid allocating a tcp_t for acceptor STREAM
201 202 * since eager has already been allocated and the accept now happens
202 203 * on acceptor STREAM. There is a big blob of comment on top of
203 204 * tcp_input_listener explaining the new accept. When socket is POP'd,
204 205 * sockfs sends us an ioctl to mark the fact and we go back to old
205 206 * behaviour. Once tcp_issocket is unset, its never set for the
206 207 * life of that connection.
207 208 *
208 209 * IPsec notes :
209 210 *
210 211 * Since a packet is always executed on the correct TCP perimeter
211 212 * all IPsec processing is defered to IP including checking new
212 213 * connections and setting IPSEC policies for new connection. The
213 214 * only exception is tcp_xmit_listeners_reset() which is called
214 215 * directly from IP and needs to policy check to see if TH_RST
215 216 * can be sent out.
216 217 */
217 218
218 219 /*
219 220 * Values for squeue switch:
220 221 * 1: SQ_NODRAIN
221 222 * 2: SQ_PROCESS
222 223 * 3: SQ_FILL
223 224 */
224 225 int tcp_squeue_wput = 2; /* /etc/systems */
225 226 int tcp_squeue_flag;
226 227
227 228 /*
228 229 * To prevent memory hog, limit the number of entries in tcp_free_list
229 230 * to 1% of available memory / number of cpus
230 231 */
231 232 uint_t tcp_free_list_max_cnt = 0;
232 233
|
↓ open down ↓ |
199 lines elided |
↑ open up ↑ |
233 234 #define TCP_XMIT_LOWATER 4096
234 235 #define TCP_XMIT_HIWATER 49152
235 236 #define TCP_RECV_LOWATER 2048
236 237 #define TCP_RECV_HIWATER 128000
237 238
238 239 #define TIDUSZ 4096 /* transport interface data unit size */
239 240
240 241 /*
241 242 * Size of acceptor hash list. It has to be a power of 2 for hashing.
242 243 */
243 -#define TCP_ACCEPTOR_FANOUT_SIZE 256
244 +#define TCP_ACCEPTOR_FANOUT_SIZE 512
244 245
245 246 #ifdef _ILP32
246 247 #define TCP_ACCEPTOR_HASH(accid) \
247 248 (((uint_t)(accid) >> 8) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
248 249 #else
249 250 #define TCP_ACCEPTOR_HASH(accid) \
250 251 ((uint_t)(accid) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
251 252 #endif /* _ILP32 */
252 253
253 254 /*
254 255 * Minimum number of connections which can be created per listener. Used
255 256 * when the listener connection count is in effect.
256 257 */
257 258 static uint32_t tcp_min_conn_listener = 2;
258 259
259 260 uint32_t tcp_early_abort = 30;
260 261
261 262 /* TCP Timer control structure */
262 263 typedef struct tcpt_s {
263 264 pfv_t tcpt_pfv; /* The routine we are to call */
264 265 tcp_t *tcpt_tcp; /* The parameter we are to pass in */
265 266 } tcpt_t;
266 267
267 268 /*
268 269 * Functions called directly via squeue having a prototype of edesc_t.
269 270 */
270 271 void tcp_input_listener(void *arg, mblk_t *mp, void *arg2,
271 272 ip_recv_attr_t *ira);
272 273 void tcp_input_data(void *arg, mblk_t *mp, void *arg2,
273 274 ip_recv_attr_t *ira);
274 275 static void tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2,
275 276 ip_recv_attr_t *dummy);
276 277
277 278
278 279 /* Prototype for TCP functions */
279 280 static void tcp_random_init(void);
280 281 int tcp_random(void);
281 282 static int tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp,
282 283 in_port_t dstport, uint_t srcid);
283 284 static int tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp,
284 285 in_port_t dstport, uint32_t flowinfo,
285 286 uint_t srcid, uint32_t scope_id);
286 287 static void tcp_iss_init(tcp_t *tcp);
287 288 static void tcp_reinit(tcp_t *tcp);
288 289 static void tcp_reinit_values(tcp_t *tcp);
289 290
290 291 static void tcp_wsrv(queue_t *q);
291 292 static void tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa);
292 293 static void tcp_update_zcopy(tcp_t *tcp);
293 294 static void tcp_notify(void *, ip_xmit_attr_t *, ixa_notify_type_t,
294 295 ixa_notify_arg_t);
295 296 static void *tcp_stack_init(netstackid_t stackid, netstack_t *ns);
296 297 static void tcp_stack_fini(netstackid_t stackid, void *arg);
297 298
298 299 static int tcp_squeue_switch(int);
299 300
300 301 static int tcp_open(queue_t *, dev_t *, int, int, cred_t *, boolean_t);
301 302 static int tcp_openv4(queue_t *, dev_t *, int, int, cred_t *);
302 303 static int tcp_openv6(queue_t *, dev_t *, int, int, cred_t *);
303 304
304 305 static void tcp_squeue_add(squeue_t *);
305 306
306 307 struct module_info tcp_rinfo = {
307 308 TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, TCP_RECV_HIWATER, TCP_RECV_LOWATER
308 309 };
309 310
310 311 static struct module_info tcp_winfo = {
311 312 TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, 127, 16
312 313 };
313 314
314 315 /*
315 316 * Entry points for TCP as a device. The normal case which supports
316 317 * the TCP functionality.
317 318 * We have separate open functions for the /dev/tcp and /dev/tcp6 devices.
318 319 */
319 320 struct qinit tcp_rinitv4 = {
320 321 NULL, (pfi_t)tcp_rsrv, tcp_openv4, tcp_tpi_close, NULL, &tcp_rinfo
321 322 };
322 323
323 324 struct qinit tcp_rinitv6 = {
324 325 NULL, (pfi_t)tcp_rsrv, tcp_openv6, tcp_tpi_close, NULL, &tcp_rinfo
325 326 };
326 327
327 328 struct qinit tcp_winit = {
328 329 (pfi_t)tcp_wput, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
329 330 };
330 331
331 332 /* Initial entry point for TCP in socket mode. */
332 333 struct qinit tcp_sock_winit = {
333 334 (pfi_t)tcp_wput_sock, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
334 335 };
335 336
336 337 /* TCP entry point during fallback */
337 338 struct qinit tcp_fallback_sock_winit = {
338 339 (pfi_t)tcp_wput_fallback, NULL, NULL, NULL, NULL, &tcp_winfo
339 340 };
340 341
341 342 /*
342 343 * Entry points for TCP as a acceptor STREAM opened by sockfs when doing
343 344 * an accept. Avoid allocating data structures since eager has already
344 345 * been created.
345 346 */
346 347 struct qinit tcp_acceptor_rinit = {
347 348 NULL, (pfi_t)tcp_rsrv, NULL, tcp_tpi_close_accept, NULL, &tcp_winfo
348 349 };
349 350
350 351 struct qinit tcp_acceptor_winit = {
351 352 (pfi_t)tcp_tpi_accept, NULL, NULL, NULL, NULL, &tcp_winfo
352 353 };
353 354
354 355 /* For AF_INET aka /dev/tcp */
355 356 struct streamtab tcpinfov4 = {
356 357 &tcp_rinitv4, &tcp_winit
357 358 };
358 359
359 360 /* For AF_INET6 aka /dev/tcp6 */
360 361 struct streamtab tcpinfov6 = {
361 362 &tcp_rinitv6, &tcp_winit
362 363 };
363 364
364 365 /*
365 366 * Following assumes TPI alignment requirements stay along 32 bit
366 367 * boundaries
367 368 */
368 369 #define ROUNDUP32(x) \
369 370 (((x) + (sizeof (int32_t) - 1)) & ~(sizeof (int32_t) - 1))
370 371
371 372 /* Template for response to info request. */
372 373 struct T_info_ack tcp_g_t_info_ack = {
373 374 T_INFO_ACK, /* PRIM_type */
374 375 0, /* TSDU_size */
375 376 T_INFINITE, /* ETSDU_size */
376 377 T_INVALID, /* CDATA_size */
377 378 T_INVALID, /* DDATA_size */
378 379 sizeof (sin_t), /* ADDR_size */
379 380 0, /* OPT_size - not initialized here */
380 381 TIDUSZ, /* TIDU_size */
381 382 T_COTS_ORD, /* SERV_type */
382 383 TCPS_IDLE, /* CURRENT_state */
383 384 (XPG4_1|EXPINLINE) /* PROVIDER_flag */
384 385 };
385 386
386 387 struct T_info_ack tcp_g_t_info_ack_v6 = {
387 388 T_INFO_ACK, /* PRIM_type */
388 389 0, /* TSDU_size */
389 390 T_INFINITE, /* ETSDU_size */
390 391 T_INVALID, /* CDATA_size */
391 392 T_INVALID, /* DDATA_size */
392 393 sizeof (sin6_t), /* ADDR_size */
393 394 0, /* OPT_size - not initialized here */
394 395 TIDUSZ, /* TIDU_size */
395 396 T_COTS_ORD, /* SERV_type */
396 397 TCPS_IDLE, /* CURRENT_state */
397 398 (XPG4_1|EXPINLINE) /* PROVIDER_flag */
398 399 };
399 400
400 401 /*
401 402 * TCP tunables related declarations. Definitions are in tcp_tunables.c
402 403 */
403 404 extern mod_prop_info_t tcp_propinfo_tbl[];
404 405 extern int tcp_propinfo_count;
405 406
406 407 #define IS_VMLOANED_MBLK(mp) \
407 408 (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0)
408 409
409 410 uint32_t do_tcpzcopy = 1; /* 0: disable, 1: enable, 2: force */
410 411
411 412 /*
412 413 * Forces all connections to obey the value of the tcps_maxpsz_multiplier
413 414 * tunable settable via NDD. Otherwise, the per-connection behavior is
414 415 * determined dynamically during tcp_set_destination(), which is the default.
415 416 */
416 417 boolean_t tcp_static_maxpsz = B_FALSE;
417 418
418 419 /*
419 420 * If the receive buffer size is changed, this function is called to update
420 421 * the upper socket layer on the new delayed receive wake up threshold.
421 422 */
422 423 static void
423 424 tcp_set_recv_threshold(tcp_t *tcp, uint32_t new_rcvthresh)
424 425 {
425 426 uint32_t default_threshold = SOCKET_RECVHIWATER >> 3;
426 427
427 428 if (IPCL_IS_NONSTR(tcp->tcp_connp)) {
428 429 conn_t *connp = tcp->tcp_connp;
429 430 struct sock_proto_props sopp;
430 431
431 432 /*
432 433 * only increase rcvthresh upto default_threshold
433 434 */
434 435 if (new_rcvthresh > default_threshold)
435 436 new_rcvthresh = default_threshold;
436 437
437 438 sopp.sopp_flags = SOCKOPT_RCVTHRESH;
438 439 sopp.sopp_rcvthresh = new_rcvthresh;
439 440
440 441 (*connp->conn_upcalls->su_set_proto_props)
441 442 (connp->conn_upper_handle, &sopp);
442 443 }
443 444 }
444 445
445 446 /*
446 447 * Figure out the value of window scale opton. Note that the rwnd is
447 448 * ASSUMED to be rounded up to the nearest MSS before the calculation.
448 449 * We cannot find the scale value and then do a round up of tcp_rwnd
449 450 * because the scale value may not be correct after that.
450 451 *
451 452 * Set the compiler flag to make this function inline.
452 453 */
453 454 void
454 455 tcp_set_ws_value(tcp_t *tcp)
455 456 {
456 457 int i;
457 458 uint32_t rwnd = tcp->tcp_rwnd;
458 459
459 460 for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT;
460 461 i++, rwnd >>= 1)
461 462 ;
462 463 tcp->tcp_rcv_ws = i;
463 464 }
464 465
465 466 /*
466 467 * Remove cached/latched IPsec references.
467 468 */
468 469 void
469 470 tcp_ipsec_cleanup(tcp_t *tcp)
470 471 {
471 472 conn_t *connp = tcp->tcp_connp;
472 473
473 474 ASSERT(connp->conn_flags & IPCL_TCPCONN);
474 475
475 476 if (connp->conn_latch != NULL) {
476 477 IPLATCH_REFRELE(connp->conn_latch);
477 478 connp->conn_latch = NULL;
478 479 }
479 480 if (connp->conn_latch_in_policy != NULL) {
480 481 IPPOL_REFRELE(connp->conn_latch_in_policy);
481 482 connp->conn_latch_in_policy = NULL;
482 483 }
483 484 if (connp->conn_latch_in_action != NULL) {
484 485 IPACT_REFRELE(connp->conn_latch_in_action);
485 486 connp->conn_latch_in_action = NULL;
486 487 }
487 488 if (connp->conn_policy != NULL) {
488 489 IPPH_REFRELE(connp->conn_policy, connp->conn_netstack);
489 490 connp->conn_policy = NULL;
490 491 }
491 492 }
492 493
493 494 /*
494 495 * Cleaup before placing on free list.
495 496 * Disassociate from the netstack/tcp_stack_t since the freelist
496 497 * is per squeue and not per netstack.
497 498 */
498 499 void
499 500 tcp_cleanup(tcp_t *tcp)
500 501 {
501 502 mblk_t *mp;
502 503 conn_t *connp = tcp->tcp_connp;
503 504 tcp_stack_t *tcps = tcp->tcp_tcps;
504 505 netstack_t *ns = tcps->tcps_netstack;
505 506 mblk_t *tcp_rsrv_mp;
506 507
507 508 tcp_bind_hash_remove(tcp);
508 509
509 510 /* Cleanup that which needs the netstack first */
510 511 tcp_ipsec_cleanup(tcp);
511 512 ixa_cleanup(connp->conn_ixa);
512 513
513 514 if (connp->conn_ht_iphc != NULL) {
514 515 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
515 516 connp->conn_ht_iphc = NULL;
516 517 connp->conn_ht_iphc_allocated = 0;
517 518 connp->conn_ht_iphc_len = 0;
518 519 connp->conn_ht_ulp = NULL;
519 520 connp->conn_ht_ulp_len = 0;
520 521 tcp->tcp_ipha = NULL;
521 522 tcp->tcp_ip6h = NULL;
522 523 tcp->tcp_tcpha = NULL;
523 524 }
524 525
525 526 /* We clear any IP_OPTIONS and extension headers */
526 527 ip_pkt_free(&connp->conn_xmit_ipp);
527 528
528 529 tcp_free(tcp);
529 530
530 531 /*
531 532 * Since we will bzero the entire structure, we need to
532 533 * remove it and reinsert it in global hash list. We
533 534 * know the walkers can't get to this conn because we
534 535 * had set CONDEMNED flag earlier and checked reference
535 536 * under conn_lock so walker won't pick it and when we
536 537 * go the ipcl_globalhash_remove() below, no walker
537 538 * can get to it.
538 539 */
539 540 ipcl_globalhash_remove(connp);
540 541
541 542 /* Save some state */
542 543 mp = tcp->tcp_timercache;
543 544
544 545 tcp_rsrv_mp = tcp->tcp_rsrv_mp;
545 546
546 547 if (connp->conn_cred != NULL) {
547 548 crfree(connp->conn_cred);
548 549 connp->conn_cred = NULL;
549 550 }
550 551 ipcl_conn_cleanup(connp);
551 552 connp->conn_flags = IPCL_TCPCONN;
552 553
553 554 /*
554 555 * Now it is safe to decrement the reference counts.
555 556 * This might be the last reference on the netstack
556 557 * in which case it will cause the freeing of the IP Instance.
557 558 */
558 559 connp->conn_netstack = NULL;
559 560 connp->conn_ixa->ixa_ipst = NULL;
560 561 netstack_rele(ns);
561 562 ASSERT(tcps != NULL);
562 563 tcp->tcp_tcps = NULL;
563 564
564 565 bzero(tcp, sizeof (tcp_t));
565 566
566 567 /* restore the state */
567 568 tcp->tcp_timercache = mp;
568 569
569 570 tcp->tcp_rsrv_mp = tcp_rsrv_mp;
570 571
571 572 tcp->tcp_connp = connp;
572 573
573 574 ASSERT(connp->conn_tcp == tcp);
574 575 ASSERT(connp->conn_flags & IPCL_TCPCONN);
575 576 connp->conn_state_flags = CONN_INCIPIENT;
576 577 ASSERT(connp->conn_proto == IPPROTO_TCP);
577 578 ASSERT(connp->conn_ref == 1);
578 579 }
579 580
580 581 /*
581 582 * Adapt to the information, such as rtt and rtt_sd, provided from the
582 583 * DCE and IRE maintained by IP.
583 584 *
584 585 * Checks for multicast and broadcast destination address.
585 586 * Returns zero if ok; an errno on failure.
586 587 *
587 588 * Note that the MSS calculation here is based on the info given in
588 589 * the DCE and IRE. We do not do any calculation based on TCP options. They
589 590 * will be handled in tcp_input_data() when TCP knows which options to use.
590 591 *
591 592 * Note on how TCP gets its parameters for a connection.
592 593 *
593 594 * When a tcp_t structure is allocated, it gets all the default parameters.
594 595 * In tcp_set_destination(), it gets those metric parameters, like rtt, rtt_sd,
595 596 * spipe, rpipe, ... from the route metrics. Route metric overrides the
596 597 * default.
597 598 *
598 599 * An incoming SYN with a multicast or broadcast destination address is dropped
599 600 * in ip_fanout_v4/v6.
600 601 *
601 602 * An incoming SYN with a multicast or broadcast source address is always
602 603 * dropped in tcp_set_destination, since IPDF_ALLOW_MCBC is not set in
603 604 * conn_connect.
604 605 * The same logic in tcp_set_destination also serves to
605 606 * reject an attempt to connect to a broadcast or multicast (destination)
606 607 * address.
607 608 */
608 609 int
609 610 tcp_set_destination(tcp_t *tcp)
610 611 {
611 612 uint32_t mss_max;
612 613 uint32_t mss;
613 614 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
614 615 conn_t *connp = tcp->tcp_connp;
615 616 tcp_stack_t *tcps = tcp->tcp_tcps;
616 617 iulp_t uinfo;
617 618 int error;
618 619 uint32_t flags;
619 620
620 621 flags = IPDF_LSO | IPDF_ZCOPY;
621 622 /*
622 623 * Make sure we have a dce for the destination to avoid dce_ident
623 624 * contention for connected sockets.
624 625 */
625 626 flags |= IPDF_UNIQUE_DCE;
626 627
627 628 if (!tcps->tcps_ignore_path_mtu)
628 629 connp->conn_ixa->ixa_flags |= IXAF_PMTU_DISCOVERY;
629 630
630 631 /* Use conn_lock to satify ASSERT; tcp is already serialized */
631 632 mutex_enter(&connp->conn_lock);
632 633 error = conn_connect(connp, &uinfo, flags);
633 634 mutex_exit(&connp->conn_lock);
634 635 if (error != 0)
635 636 return (error);
636 637
637 638 error = tcp_build_hdrs(tcp);
638 639 if (error != 0)
639 640 return (error);
640 641
641 642 tcp->tcp_localnet = uinfo.iulp_localnet;
642 643
643 644 if (uinfo.iulp_rtt != 0) {
644 645 clock_t rto;
645 646
646 647 tcp->tcp_rtt_sa = uinfo.iulp_rtt;
647 648 tcp->tcp_rtt_sd = uinfo.iulp_rtt_sd;
648 649 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
649 650 tcps->tcps_rexmit_interval_extra +
650 651 (tcp->tcp_rtt_sa >> 5);
651 652
652 653 TCP_SET_RTO(tcp, rto);
653 654 }
654 655 if (uinfo.iulp_ssthresh != 0)
655 656 tcp->tcp_cwnd_ssthresh = uinfo.iulp_ssthresh;
656 657 else
657 658 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
658 659 if (uinfo.iulp_spipe > 0) {
659 660 connp->conn_sndbuf = MIN(uinfo.iulp_spipe,
660 661 tcps->tcps_max_buf);
661 662 if (tcps->tcps_snd_lowat_fraction != 0) {
662 663 connp->conn_sndlowat = connp->conn_sndbuf /
663 664 tcps->tcps_snd_lowat_fraction;
664 665 }
665 666 (void) tcp_maxpsz_set(tcp, B_TRUE);
666 667 }
667 668 /*
668 669 * Note that up till now, acceptor always inherits receive
669 670 * window from the listener. But if there is a metrics
670 671 * associated with a host, we should use that instead of
671 672 * inheriting it from listener. Thus we need to pass this
672 673 * info back to the caller.
673 674 */
674 675 if (uinfo.iulp_rpipe > 0) {
675 676 tcp->tcp_rwnd = MIN(uinfo.iulp_rpipe,
676 677 tcps->tcps_max_buf);
677 678 }
678 679
679 680 if (uinfo.iulp_rtomax > 0) {
680 681 tcp->tcp_second_timer_threshold =
681 682 uinfo.iulp_rtomax;
682 683 }
683 684
684 685 /*
685 686 * Use the metric option settings, iulp_tstamp_ok and
686 687 * iulp_wscale_ok, only for active open. What this means
687 688 * is that if the other side uses timestamp or window
688 689 * scale option, TCP will also use those options. That
689 690 * is for passive open. If the application sets a
690 691 * large window, window scale is enabled regardless of
691 692 * the value in iulp_wscale_ok. This is the behavior
692 693 * since 2.6. So we keep it.
693 694 * The only case left in passive open processing is the
694 695 * check for SACK.
695 696 * For ECN, it should probably be like SACK. But the
696 697 * current value is binary, so we treat it like the other
697 698 * cases. The metric only controls active open.For passive
698 699 * open, the ndd param, tcp_ecn_permitted, controls the
699 700 * behavior.
700 701 */
701 702 if (!tcp_detached) {
702 703 /*
703 704 * The if check means that the following can only
704 705 * be turned on by the metrics only IRE, but not off.
705 706 */
706 707 if (uinfo.iulp_tstamp_ok)
707 708 tcp->tcp_snd_ts_ok = B_TRUE;
708 709 if (uinfo.iulp_wscale_ok)
709 710 tcp->tcp_snd_ws_ok = B_TRUE;
710 711 if (uinfo.iulp_sack == 2)
711 712 tcp->tcp_snd_sack_ok = B_TRUE;
712 713 if (uinfo.iulp_ecn_ok)
713 714 tcp->tcp_ecn_ok = B_TRUE;
714 715 } else {
715 716 /*
716 717 * Passive open.
717 718 *
718 719 * As above, the if check means that SACK can only be
719 720 * turned on by the metric only IRE.
720 721 */
721 722 if (uinfo.iulp_sack > 0) {
722 723 tcp->tcp_snd_sack_ok = B_TRUE;
723 724 }
724 725 }
725 726
726 727 /*
727 728 * XXX Note that currently, iulp_mtu can be as small as 68
728 729 * because of PMTUd. So tcp_mss may go to negative if combined
729 730 * length of all those options exceeds 28 bytes. But because
730 731 * of the tcp_mss_min check below, we may not have a problem if
731 732 * tcp_mss_min is of a reasonable value. The default is 1 so
732 733 * the negative problem still exists. And the check defeats PMTUd.
733 734 * In fact, if PMTUd finds that the MSS should be smaller than
734 735 * tcp_mss_min, TCP should turn off PMUTd and use the tcp_mss_min
735 736 * value.
736 737 *
737 738 * We do not deal with that now. All those problems related to
738 739 * PMTUd will be fixed later.
739 740 */
740 741 ASSERT(uinfo.iulp_mtu != 0);
741 742 mss = tcp->tcp_initial_pmtu = uinfo.iulp_mtu;
742 743
743 744 /* Sanity check for MSS value. */
744 745 if (connp->conn_ipversion == IPV4_VERSION)
745 746 mss_max = tcps->tcps_mss_max_ipv4;
746 747 else
747 748 mss_max = tcps->tcps_mss_max_ipv6;
748 749
749 750 if (tcp->tcp_ipsec_overhead == 0)
750 751 tcp->tcp_ipsec_overhead = conn_ipsec_length(connp);
751 752
752 753 mss -= tcp->tcp_ipsec_overhead;
753 754
754 755 if (mss < tcps->tcps_mss_min)
755 756 mss = tcps->tcps_mss_min;
756 757 if (mss > mss_max)
757 758 mss = mss_max;
758 759
759 760 /* Note that this is the maximum MSS, excluding all options. */
760 761 tcp->tcp_mss = mss;
761 762
762 763 /*
763 764 * Update the tcp connection with LSO capability.
764 765 */
765 766 tcp_update_lso(tcp, connp->conn_ixa);
766 767
767 768 /*
768 769 * Initialize the ISS here now that we have the full connection ID.
769 770 * The RFC 1948 method of initial sequence number generation requires
770 771 * knowledge of the full connection ID before setting the ISS.
771 772 */
772 773 tcp_iss_init(tcp);
773 774
774 775 tcp->tcp_loopback = (uinfo.iulp_loopback | uinfo.iulp_local);
775 776
776 777 /*
777 778 * Make sure that conn is not marked incipient
778 779 * for incoming connections. A blind
779 780 * removal of incipient flag is cheaper than
780 781 * check and removal.
781 782 */
782 783 mutex_enter(&connp->conn_lock);
783 784 connp->conn_state_flags &= ~CONN_INCIPIENT;
784 785 mutex_exit(&connp->conn_lock);
785 786 return (0);
786 787 }
787 788
788 789 /*
789 790 * tcp_clean_death / tcp_close_detached must not be called more than once
790 791 * on a tcp. Thus every function that potentially calls tcp_clean_death
791 792 * must check for the tcp state before calling tcp_clean_death.
792 793 * Eg. tcp_input_data, tcp_eager_kill, tcp_clean_death_wrapper,
793 794 * tcp_timer_handler, all check for the tcp state.
794 795 */
795 796 /* ARGSUSED */
796 797 void
797 798 tcp_clean_death_wrapper(void *arg, mblk_t *mp, void *arg2,
798 799 ip_recv_attr_t *dummy)
799 800 {
800 801 tcp_t *tcp = ((conn_t *)arg)->conn_tcp;
801 802
802 803 freemsg(mp);
803 804 if (tcp->tcp_state > TCPS_BOUND)
804 805 (void) tcp_clean_death(((conn_t *)arg)->conn_tcp, ETIMEDOUT);
805 806 }
806 807
807 808 /*
808 809 * We are dying for some reason. Try to do it gracefully. (May be called
809 810 * as writer.)
810 811 *
811 812 * Return -1 if the structure was not cleaned up (if the cleanup had to be
812 813 * done by a service procedure).
813 814 * TBD - Should the return value distinguish between the tcp_t being
814 815 * freed and it being reinitialized?
815 816 */
816 817 int
817 818 tcp_clean_death(tcp_t *tcp, int err)
818 819 {
819 820 mblk_t *mp;
820 821 queue_t *q;
821 822 conn_t *connp = tcp->tcp_connp;
822 823 tcp_stack_t *tcps = tcp->tcp_tcps;
823 824
824 825 if (tcp->tcp_fused)
825 826 tcp_unfuse(tcp);
826 827
827 828 if (tcp->tcp_linger_tid != 0 &&
828 829 TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
829 830 tcp_stop_lingering(tcp);
830 831 }
831 832
832 833 ASSERT(tcp != NULL);
833 834 ASSERT((connp->conn_family == AF_INET &&
834 835 connp->conn_ipversion == IPV4_VERSION) ||
835 836 (connp->conn_family == AF_INET6 &&
836 837 (connp->conn_ipversion == IPV4_VERSION ||
837 838 connp->conn_ipversion == IPV6_VERSION)));
838 839
839 840 if (TCP_IS_DETACHED(tcp)) {
840 841 if (tcp->tcp_hard_binding) {
841 842 /*
842 843 * Its an eager that we are dealing with. We close the
843 844 * eager but in case a conn_ind has already gone to the
844 845 * listener, let tcp_accept_finish() send a discon_ind
845 846 * to the listener and drop the last reference. If the
846 847 * listener doesn't even know about the eager i.e. the
847 848 * conn_ind hasn't gone up, blow away the eager and drop
848 849 * the last reference as well. If the conn_ind has gone
849 850 * up, state should be BOUND. tcp_accept_finish
850 851 * will figure out that the connection has received a
851 852 * RST and will send a DISCON_IND to the application.
852 853 */
853 854 tcp_closei_local(tcp);
854 855 if (!tcp->tcp_tconnind_started) {
855 856 CONN_DEC_REF(connp);
856 857 } else {
857 858 tcp->tcp_state = TCPS_BOUND;
858 859 DTRACE_TCP6(state__change, void, NULL,
859 860 ip_xmit_attr_t *, connp->conn_ixa,
860 861 void, NULL, tcp_t *, tcp, void, NULL,
861 862 int32_t, TCPS_CLOSED);
862 863 }
863 864 } else {
864 865 tcp_close_detached(tcp);
865 866 }
866 867 return (0);
867 868 }
868 869
869 870 TCP_STAT(tcps, tcp_clean_death_nondetached);
870 871
871 872 /*
872 873 * The connection is dead. Decrement listener connection counter if
873 874 * necessary.
874 875 */
875 876 if (tcp->tcp_listen_cnt != NULL)
876 877 TCP_DECR_LISTEN_CNT(tcp);
877 878
878 879 /*
879 880 * When a connection is moved to TIME_WAIT state, the connection
880 881 * counter is already decremented. So no need to decrement here
881 882 * again. See SET_TIME_WAIT() macro.
882 883 */
883 884 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
884 885 tcp->tcp_state < TCPS_TIME_WAIT) {
885 886 TCPS_CONN_DEC(tcps);
886 887 }
887 888
888 889 q = connp->conn_rq;
889 890
890 891 /* Trash all inbound data */
891 892 if (!IPCL_IS_NONSTR(connp)) {
892 893 ASSERT(q != NULL);
893 894 flushq(q, FLUSHALL);
894 895 }
895 896
896 897 /*
897 898 * If we are at least part way open and there is error
898 899 * (err==0 implies no error)
899 900 * notify our client by a T_DISCON_IND.
900 901 */
901 902 if ((tcp->tcp_state >= TCPS_SYN_SENT) && err) {
902 903 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
903 904 !TCP_IS_SOCKET(tcp)) {
904 905 /*
905 906 * Send M_FLUSH according to TPI. Because sockets will
906 907 * (and must) ignore FLUSHR we do that only for TPI
907 908 * endpoints and sockets in STREAMS mode.
908 909 */
909 910 (void) putnextctl1(q, M_FLUSH, FLUSHR);
910 911 }
911 912 if (connp->conn_debug) {
912 913 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE|SL_ERROR,
913 914 "tcp_clean_death: discon err %d", err);
914 915 }
915 916 if (IPCL_IS_NONSTR(connp)) {
916 917 /* Direct socket, use upcall */
917 918 (*connp->conn_upcalls->su_disconnected)(
918 919 connp->conn_upper_handle, tcp->tcp_connid, err);
919 920 } else {
920 921 mp = mi_tpi_discon_ind(NULL, err, 0);
921 922 if (mp != NULL) {
922 923 putnext(q, mp);
923 924 } else {
924 925 if (connp->conn_debug) {
925 926 (void) strlog(TCP_MOD_ID, 0, 1,
926 927 SL_ERROR|SL_TRACE,
927 928 "tcp_clean_death, sending M_ERROR");
928 929 }
929 930 (void) putnextctl1(q, M_ERROR, EPROTO);
930 931 }
931 932 }
932 933 if (tcp->tcp_state <= TCPS_SYN_RCVD) {
933 934 /* SYN_SENT or SYN_RCVD */
934 935 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
935 936 } else if (tcp->tcp_state <= TCPS_CLOSE_WAIT) {
936 937 /* ESTABLISHED or CLOSE_WAIT */
937 938 TCPS_BUMP_MIB(tcps, tcpEstabResets);
938 939 }
939 940 }
940 941
941 942 /*
942 943 * ESTABLISHED non-STREAMS eagers are not 'detached' because
943 944 * an upper handle is obtained when the SYN-ACK comes in. So it
944 945 * should receive the 'disconnected' upcall, but tcp_reinit should
945 946 * not be called since this is an eager.
946 947 */
947 948 if (tcp->tcp_listener != NULL && IPCL_IS_NONSTR(connp)) {
948 949 tcp_closei_local(tcp);
949 950 tcp->tcp_state = TCPS_BOUND;
950 951 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
951 952 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
952 953 int32_t, TCPS_CLOSED);
953 954 return (0);
954 955 }
955 956
956 957 tcp_reinit(tcp);
957 958 if (IPCL_IS_NONSTR(connp))
958 959 (void) tcp_do_unbind(connp);
959 960
960 961 return (-1);
961 962 }
962 963
963 964 /*
964 965 * In case tcp is in the "lingering state" and waits for the SO_LINGER timeout
965 966 * to expire, stop the wait and finish the close.
966 967 */
967 968 void
968 969 tcp_stop_lingering(tcp_t *tcp)
969 970 {
970 971 clock_t delta = 0;
971 972 tcp_stack_t *tcps = tcp->tcp_tcps;
972 973 conn_t *connp = tcp->tcp_connp;
973 974
974 975 tcp->tcp_linger_tid = 0;
975 976 if (tcp->tcp_state > TCPS_LISTEN) {
976 977 tcp_acceptor_hash_remove(tcp);
977 978 mutex_enter(&tcp->tcp_non_sq_lock);
978 979 if (tcp->tcp_flow_stopped) {
979 980 tcp_clrqfull(tcp);
980 981 }
981 982 mutex_exit(&tcp->tcp_non_sq_lock);
982 983
983 984 if (tcp->tcp_timer_tid != 0) {
984 985 delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
985 986 tcp->tcp_timer_tid = 0;
986 987 }
987 988 /*
988 989 * Need to cancel those timers which will not be used when
989 990 * TCP is detached. This has to be done before the conn_wq
990 991 * is cleared.
991 992 */
992 993 tcp_timers_stop(tcp);
993 994
994 995 tcp->tcp_detached = B_TRUE;
995 996 connp->conn_rq = NULL;
996 997 connp->conn_wq = NULL;
997 998
998 999 if (tcp->tcp_state == TCPS_TIME_WAIT) {
999 1000 tcp_time_wait_append(tcp);
1000 1001 TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1001 1002 goto finish;
1002 1003 }
1003 1004
1004 1005 /*
1005 1006 * If delta is zero the timer event wasn't executed and was
1006 1007 * successfully canceled. In this case we need to restart it
1007 1008 * with the minimal delta possible.
1008 1009 */
1009 1010 if (delta >= 0) {
1010 1011 tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1011 1012 delta ? delta : 1);
1012 1013 }
1013 1014 } else {
1014 1015 tcp_closei_local(tcp);
1015 1016 CONN_DEC_REF(connp);
1016 1017 }
1017 1018 finish:
1018 1019 tcp->tcp_detached = B_TRUE;
1019 1020 connp->conn_rq = NULL;
1020 1021 connp->conn_wq = NULL;
1021 1022
1022 1023 /* Signal closing thread that it can complete close */
1023 1024 mutex_enter(&tcp->tcp_closelock);
1024 1025 tcp->tcp_closed = 1;
1025 1026 cv_signal(&tcp->tcp_closecv);
1026 1027 mutex_exit(&tcp->tcp_closelock);
1027 1028
1028 1029 /* If we have an upper handle (socket), release it */
1029 1030 if (IPCL_IS_NONSTR(connp)) {
1030 1031 ASSERT(connp->conn_upper_handle != NULL);
1031 1032 (*connp->conn_upcalls->su_closed)(connp->conn_upper_handle);
1032 1033 connp->conn_upper_handle = NULL;
1033 1034 connp->conn_upcalls = NULL;
1034 1035 }
1035 1036 }
1036 1037
1037 1038 void
1038 1039 tcp_close_common(conn_t *connp, int flags)
1039 1040 {
1040 1041 tcp_t *tcp = connp->conn_tcp;
1041 1042 mblk_t *mp = &tcp->tcp_closemp;
1042 1043 boolean_t conn_ioctl_cleanup_reqd = B_FALSE;
1043 1044 mblk_t *bp;
1044 1045
1045 1046 ASSERT(connp->conn_ref >= 2);
1046 1047
1047 1048 /*
1048 1049 * Mark the conn as closing. ipsq_pending_mp_add will not
1049 1050 * add any mp to the pending mp list, after this conn has
1050 1051 * started closing.
1051 1052 */
1052 1053 mutex_enter(&connp->conn_lock);
1053 1054 connp->conn_state_flags |= CONN_CLOSING;
1054 1055 if (connp->conn_oper_pending_ill != NULL)
1055 1056 conn_ioctl_cleanup_reqd = B_TRUE;
1056 1057 CONN_INC_REF_LOCKED(connp);
1057 1058 mutex_exit(&connp->conn_lock);
1058 1059 tcp->tcp_closeflags = (uint8_t)flags;
1059 1060 ASSERT(connp->conn_ref >= 3);
1060 1061
1061 1062 /*
1062 1063 * tcp_closemp_used is used below without any protection of a lock
1063 1064 * as we don't expect any one else to use it concurrently at this
1064 1065 * point otherwise it would be a major defect.
1065 1066 */
1066 1067
1067 1068 if (mp->b_prev == NULL)
1068 1069 tcp->tcp_closemp_used = B_TRUE;
1069 1070 else
1070 1071 cmn_err(CE_PANIC, "tcp_close: concurrent use of tcp_closemp: "
1071 1072 "connp %p tcp %p\n", (void *)connp, (void *)tcp);
1072 1073
1073 1074 TCP_DEBUG_GETPCSTACK(tcp->tcmp_stk, 15);
1074 1075
1075 1076 /*
1076 1077 * Cleanup any queued ioctls here. This must be done before the wq/rq
1077 1078 * are re-written by tcp_close_output().
1078 1079 */
1079 1080 if (conn_ioctl_cleanup_reqd)
1080 1081 conn_ioctl_cleanup(connp);
1081 1082
1082 1083 /*
1083 1084 * As CONN_CLOSING is set, no further ioctls should be passed down to
1084 1085 * IP for this conn (see the guards in tcp_ioctl, tcp_wput_ioctl and
1085 1086 * tcp_wput_iocdata). If the ioctl was queued on an ipsq,
1086 1087 * conn_ioctl_cleanup should have found it and removed it. If the ioctl
1087 1088 * was still in flight at the time, we wait for it here. See comments
1088 1089 * for CONN_INC_IOCTLREF in ip.h for details.
1089 1090 */
1090 1091 mutex_enter(&connp->conn_lock);
1091 1092 while (connp->conn_ioctlref > 0)
1092 1093 cv_wait(&connp->conn_cv, &connp->conn_lock);
1093 1094 ASSERT(connp->conn_ioctlref == 0);
1094 1095 ASSERT(connp->conn_oper_pending_ill == NULL);
1095 1096 mutex_exit(&connp->conn_lock);
1096 1097
1097 1098 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_close_output, connp,
1098 1099 NULL, tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1099 1100
1100 1101 /*
1101 1102 * For non-STREAMS sockets, the normal case is that the conn makes
1102 1103 * an upcall when it's finally closed, so there is no need to wait
1103 1104 * in the protocol. But in case of SO_LINGER the thread sleeps here
1104 1105 * so it can properly deal with the thread being interrupted.
1105 1106 */
1106 1107 if (IPCL_IS_NONSTR(connp) && connp->conn_linger == 0)
1107 1108 goto nowait;
1108 1109
1109 1110 mutex_enter(&tcp->tcp_closelock);
1110 1111 while (!tcp->tcp_closed) {
1111 1112 if (!cv_wait_sig(&tcp->tcp_closecv, &tcp->tcp_closelock)) {
1112 1113 /*
1113 1114 * The cv_wait_sig() was interrupted. We now do the
1114 1115 * following:
1115 1116 *
1116 1117 * 1) If the endpoint was lingering, we allow this
1117 1118 * to be interrupted by cancelling the linger timeout
1118 1119 * and closing normally.
1119 1120 *
1120 1121 * 2) Revert to calling cv_wait()
1121 1122 *
1122 1123 * We revert to using cv_wait() to avoid an
1123 1124 * infinite loop which can occur if the calling
1124 1125 * thread is higher priority than the squeue worker
1125 1126 * thread and is bound to the same cpu.
1126 1127 */
1127 1128 if (connp->conn_linger && connp->conn_lingertime > 0) {
1128 1129 mutex_exit(&tcp->tcp_closelock);
1129 1130 /* Entering squeue, bump ref count. */
1130 1131 CONN_INC_REF(connp);
1131 1132 bp = allocb_wait(0, BPRI_HI, STR_NOSIG, NULL);
1132 1133 SQUEUE_ENTER_ONE(connp->conn_sqp, bp,
1133 1134 tcp_linger_interrupted, connp, NULL,
1134 1135 tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1135 1136 mutex_enter(&tcp->tcp_closelock);
1136 1137 }
1137 1138 break;
1138 1139 }
1139 1140 }
1140 1141 while (!tcp->tcp_closed)
1141 1142 cv_wait(&tcp->tcp_closecv, &tcp->tcp_closelock);
1142 1143 mutex_exit(&tcp->tcp_closelock);
1143 1144
1144 1145 /*
1145 1146 * In the case of listener streams that have eagers in the q or q0
1146 1147 * we wait for the eagers to drop their reference to us. conn_rq and
1147 1148 * conn_wq of the eagers point to our queues. By waiting for the
1148 1149 * refcnt to drop to 1, we are sure that the eagers have cleaned
1149 1150 * up their queue pointers and also dropped their references to us.
1150 1151 *
1151 1152 * For non-STREAMS sockets we do not have to wait here; the
1152 1153 * listener will instead make a su_closed upcall when the last
1153 1154 * reference is dropped.
1154 1155 */
1155 1156 if (tcp->tcp_wait_for_eagers && !IPCL_IS_NONSTR(connp)) {
1156 1157 mutex_enter(&connp->conn_lock);
1157 1158 while (connp->conn_ref != 1) {
1158 1159 cv_wait(&connp->conn_cv, &connp->conn_lock);
1159 1160 }
1160 1161 mutex_exit(&connp->conn_lock);
1161 1162 }
1162 1163
1163 1164 nowait:
1164 1165 connp->conn_cpid = NOPID;
1165 1166 }
1166 1167
1167 1168 /*
1168 1169 * Called by tcp_close() routine via squeue when lingering is
1169 1170 * interrupted by a signal.
1170 1171 */
1171 1172
1172 1173 /* ARGSUSED */
1173 1174 static void
1174 1175 tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1175 1176 {
1176 1177 conn_t *connp = (conn_t *)arg;
1177 1178 tcp_t *tcp = connp->conn_tcp;
1178 1179
1179 1180 freeb(mp);
1180 1181 if (tcp->tcp_linger_tid != 0 &&
1181 1182 TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
1182 1183 tcp_stop_lingering(tcp);
1183 1184 tcp->tcp_client_errno = EINTR;
1184 1185 }
1185 1186 }
1186 1187
1187 1188 /*
1188 1189 * Clean up the b_next and b_prev fields of every mblk pointed at by *mpp.
1189 1190 * Some stream heads get upset if they see these later on as anything but NULL.
1190 1191 */
1191 1192 void
1192 1193 tcp_close_mpp(mblk_t **mpp)
1193 1194 {
1194 1195 mblk_t *mp;
1195 1196
1196 1197 if ((mp = *mpp) != NULL) {
1197 1198 do {
1198 1199 mp->b_next = NULL;
1199 1200 mp->b_prev = NULL;
1200 1201 } while ((mp = mp->b_cont) != NULL);
1201 1202
1202 1203 mp = *mpp;
1203 1204 *mpp = NULL;
1204 1205 freemsg(mp);
1205 1206 }
1206 1207 }
1207 1208
1208 1209 /* Do detached close. */
1209 1210 void
1210 1211 tcp_close_detached(tcp_t *tcp)
1211 1212 {
1212 1213 if (tcp->tcp_fused)
1213 1214 tcp_unfuse(tcp);
1214 1215
1215 1216 /*
1216 1217 * Clustering code serializes TCP disconnect callbacks and
1217 1218 * cluster tcp list walks by blocking a TCP disconnect callback
1218 1219 * if a cluster tcp list walk is in progress. This ensures
1219 1220 * accurate accounting of TCPs in the cluster code even though
1220 1221 * the TCP list walk itself is not atomic.
1221 1222 */
1222 1223 tcp_closei_local(tcp);
1223 1224 CONN_DEC_REF(tcp->tcp_connp);
1224 1225 }
1225 1226
1226 1227 /*
1227 1228 * The tcp_t is going away. Remove it from all lists and set it
1228 1229 * to TCPS_CLOSED. The freeing up of memory is deferred until
1229 1230 * tcp_inactive. This is needed since a thread in tcp_rput might have
1230 1231 * done a CONN_INC_REF on this structure before it was removed from the
1231 1232 * hashes.
1232 1233 */
1233 1234 void
1234 1235 tcp_closei_local(tcp_t *tcp)
1235 1236 {
1236 1237 conn_t *connp = tcp->tcp_connp;
1237 1238 tcp_stack_t *tcps = tcp->tcp_tcps;
1238 1239 int32_t oldstate;
1239 1240
1240 1241 if (!TCP_IS_SOCKET(tcp))
1241 1242 tcp_acceptor_hash_remove(tcp);
1242 1243
1243 1244 TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1244 1245 tcp->tcp_ibsegs = 0;
1245 1246 TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1246 1247 tcp->tcp_obsegs = 0;
1247 1248
1248 1249 /*
1249 1250 * This can be called via tcp_time_wait_processing() if TCP gets a
1250 1251 * SYN with sequence number outside the TIME-WAIT connection's
1251 1252 * window. So we need to check for TIME-WAIT state here as the
1252 1253 * connection counter is already decremented. See SET_TIME_WAIT()
1253 1254 * macro
1254 1255 */
1255 1256 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1256 1257 tcp->tcp_state < TCPS_TIME_WAIT) {
1257 1258 TCPS_CONN_DEC(tcps);
1258 1259 }
1259 1260
1260 1261 /*
1261 1262 * If we are an eager connection hanging off a listener that
1262 1263 * hasn't formally accepted the connection yet, get off his
1263 1264 * list and blow off any data that we have accumulated.
1264 1265 */
1265 1266 if (tcp->tcp_listener != NULL) {
1266 1267 tcp_t *listener = tcp->tcp_listener;
1267 1268 mutex_enter(&listener->tcp_eager_lock);
1268 1269 /*
1269 1270 * tcp_tconnind_started == B_TRUE means that the
1270 1271 * conn_ind has already gone to listener. At
1271 1272 * this point, eager will be closed but we
1272 1273 * leave it in listeners eager list so that
1273 1274 * if listener decides to close without doing
1274 1275 * accept, we can clean this up. In tcp_tli_accept
1275 1276 * we take care of the case of accept on closed
1276 1277 * eager.
1277 1278 */
1278 1279 if (!tcp->tcp_tconnind_started) {
1279 1280 tcp_eager_unlink(tcp);
1280 1281 mutex_exit(&listener->tcp_eager_lock);
1281 1282 /*
1282 1283 * We don't want to have any pointers to the
1283 1284 * listener queue, after we have released our
1284 1285 * reference on the listener
1285 1286 */
1286 1287 ASSERT(tcp->tcp_detached);
1287 1288 connp->conn_rq = NULL;
1288 1289 connp->conn_wq = NULL;
1289 1290 CONN_DEC_REF(listener->tcp_connp);
1290 1291 } else {
1291 1292 mutex_exit(&listener->tcp_eager_lock);
1292 1293 }
1293 1294 }
1294 1295
1295 1296 /* Stop all the timers */
1296 1297 tcp_timers_stop(tcp);
1297 1298
1298 1299 if (tcp->tcp_state == TCPS_LISTEN) {
1299 1300 if (tcp->tcp_ip_addr_cache) {
1300 1301 kmem_free((void *)tcp->tcp_ip_addr_cache,
1301 1302 IP_ADDR_CACHE_SIZE * sizeof (ipaddr_t));
1302 1303 tcp->tcp_ip_addr_cache = NULL;
1303 1304 }
1304 1305 }
1305 1306
1306 1307 /* Decrement listerner connection counter if necessary. */
1307 1308 if (tcp->tcp_listen_cnt != NULL)
1308 1309 TCP_DECR_LISTEN_CNT(tcp);
1309 1310
1310 1311 mutex_enter(&tcp->tcp_non_sq_lock);
1311 1312 if (tcp->tcp_flow_stopped)
1312 1313 tcp_clrqfull(tcp);
1313 1314 mutex_exit(&tcp->tcp_non_sq_lock);
1314 1315
1315 1316 tcp_bind_hash_remove(tcp);
1316 1317 /*
1317 1318 * If the tcp_time_wait_collector (which runs outside the squeue)
1318 1319 * is trying to remove this tcp from the time wait list, we will
1319 1320 * block in tcp_time_wait_remove while trying to acquire the
1320 1321 * tcp_time_wait_lock. The logic in tcp_time_wait_collector also
1321 1322 * requires the ipcl_hash_remove to be ordered after the
1322 1323 * tcp_time_wait_remove for the refcnt checks to work correctly.
1323 1324 */
1324 1325 if (tcp->tcp_state == TCPS_TIME_WAIT)
1325 1326 (void) tcp_time_wait_remove(tcp, NULL);
1326 1327 CL_INET_DISCONNECT(connp);
1327 1328 ipcl_hash_remove(connp);
1328 1329 oldstate = tcp->tcp_state;
1329 1330 tcp->tcp_state = TCPS_CLOSED;
1330 1331 /* Need to probe before ixa_cleanup() is called */
1331 1332 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1332 1333 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
1333 1334 int32_t, oldstate);
1334 1335 ixa_cleanup(connp->conn_ixa);
1335 1336
1336 1337 /*
1337 1338 * Mark the conn as CONDEMNED
1338 1339 */
1339 1340 mutex_enter(&connp->conn_lock);
1340 1341 connp->conn_state_flags |= CONN_CONDEMNED;
1341 1342 mutex_exit(&connp->conn_lock);
1342 1343
1343 1344 ASSERT(tcp->tcp_time_wait_next == NULL);
1344 1345 ASSERT(tcp->tcp_time_wait_prev == NULL);
1345 1346 ASSERT(tcp->tcp_time_wait_expire == 0);
1346 1347
1347 1348 tcp_ipsec_cleanup(tcp);
1348 1349 }
1349 1350
1350 1351 /*
1351 1352 * tcp is dying (called from ipcl_conn_destroy and error cases).
1352 1353 * Free the tcp_t in either case.
1353 1354 */
1354 1355 void
1355 1356 tcp_free(tcp_t *tcp)
1356 1357 {
1357 1358 mblk_t *mp;
1358 1359 conn_t *connp = tcp->tcp_connp;
1359 1360
1360 1361 ASSERT(tcp != NULL);
1361 1362 ASSERT(tcp->tcp_ptpahn == NULL && tcp->tcp_acceptor_hash == NULL);
1362 1363
1363 1364 connp->conn_rq = NULL;
1364 1365 connp->conn_wq = NULL;
1365 1366
1366 1367 tcp_close_mpp(&tcp->tcp_xmit_head);
1367 1368 tcp_close_mpp(&tcp->tcp_reass_head);
1368 1369 if (tcp->tcp_rcv_list != NULL) {
1369 1370 /* Free b_next chain */
1370 1371 tcp_close_mpp(&tcp->tcp_rcv_list);
1371 1372 }
1372 1373 if ((mp = tcp->tcp_urp_mp) != NULL) {
1373 1374 freemsg(mp);
1374 1375 }
1375 1376 if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1376 1377 freemsg(mp);
1377 1378 }
1378 1379
1379 1380 if (tcp->tcp_fused_sigurg_mp != NULL) {
1380 1381 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1381 1382 freeb(tcp->tcp_fused_sigurg_mp);
1382 1383 tcp->tcp_fused_sigurg_mp = NULL;
1383 1384 }
1384 1385
1385 1386 if (tcp->tcp_ordrel_mp != NULL) {
1386 1387 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1387 1388 freeb(tcp->tcp_ordrel_mp);
1388 1389 tcp->tcp_ordrel_mp = NULL;
1389 1390 }
1390 1391
1391 1392 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
1392 1393 bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
1393 1394
1394 1395 if (tcp->tcp_hopopts != NULL) {
1395 1396 mi_free(tcp->tcp_hopopts);
1396 1397 tcp->tcp_hopopts = NULL;
1397 1398 tcp->tcp_hopoptslen = 0;
1398 1399 }
1399 1400 ASSERT(tcp->tcp_hopoptslen == 0);
1400 1401 if (tcp->tcp_dstopts != NULL) {
1401 1402 mi_free(tcp->tcp_dstopts);
1402 1403 tcp->tcp_dstopts = NULL;
1403 1404 tcp->tcp_dstoptslen = 0;
1404 1405 }
1405 1406 ASSERT(tcp->tcp_dstoptslen == 0);
1406 1407 if (tcp->tcp_rthdrdstopts != NULL) {
1407 1408 mi_free(tcp->tcp_rthdrdstopts);
1408 1409 tcp->tcp_rthdrdstopts = NULL;
1409 1410 tcp->tcp_rthdrdstoptslen = 0;
1410 1411 }
1411 1412 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
1412 1413 if (tcp->tcp_rthdr != NULL) {
1413 1414 mi_free(tcp->tcp_rthdr);
1414 1415 tcp->tcp_rthdr = NULL;
1415 1416 tcp->tcp_rthdrlen = 0;
1416 1417 }
1417 1418 ASSERT(tcp->tcp_rthdrlen == 0);
1418 1419
1419 1420 /*
1420 1421 * Following is really a blowing away a union.
1421 1422 * It happens to have exactly two members of identical size
1422 1423 * the following code is enough.
1423 1424 */
1424 1425 tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1425 1426
1426 1427 /*
1427 1428 * If this is a non-STREAM socket still holding on to an upper
1428 1429 * handle, release it. As a result of fallback we might also see
1429 1430 * STREAMS based conns with upper handles, in which case there is
1430 1431 * nothing to do other than clearing the field.
1431 1432 */
1432 1433 if (connp->conn_upper_handle != NULL) {
1433 1434 if (IPCL_IS_NONSTR(connp)) {
1434 1435 (*connp->conn_upcalls->su_closed)(
1435 1436 connp->conn_upper_handle);
1436 1437 tcp->tcp_detached = B_TRUE;
1437 1438 }
1438 1439 connp->conn_upper_handle = NULL;
1439 1440 connp->conn_upcalls = NULL;
1440 1441 }
1441 1442 }
1442 1443
1443 1444 /*
1444 1445 * tcp_get_conn/tcp_free_conn
1445 1446 *
1446 1447 * tcp_get_conn is used to get a clean tcp connection structure.
1447 1448 * It tries to reuse the connections put on the freelist by the
1448 1449 * time_wait_collector failing which it goes to kmem_cache. This
1449 1450 * way has two benefits compared to just allocating from and
1450 1451 * freeing to kmem_cache.
1451 1452 * 1) The time_wait_collector can free (which includes the cleanup)
1452 1453 * outside the squeue. So when the interrupt comes, we have a clean
1453 1454 * connection sitting in the freelist. Obviously, this buys us
1454 1455 * performance.
1455 1456 *
1456 1457 * 2) Defence against DOS attack. Allocating a tcp/conn in tcp_input_listener
1457 1458 * has multiple disadvantages - tying up the squeue during alloc.
1458 1459 * But allocating the conn/tcp in IP land is also not the best since
1459 1460 * we can't check the 'q' and 'q0' which are protected by squeue and
1460 1461 * blindly allocate memory which might have to be freed here if we are
1461 1462 * not allowed to accept the connection. By using the freelist and
1462 1463 * putting the conn/tcp back in freelist, we don't pay a penalty for
1463 1464 * allocating memory without checking 'q/q0' and freeing it if we can't
1464 1465 * accept the connection.
1465 1466 *
1466 1467 * Care should be taken to put the conn back in the same squeue's freelist
1467 1468 * from which it was allocated. Best results are obtained if conn is
1468 1469 * allocated from listener's squeue and freed to the same. Time wait
1469 1470 * collector will free up the freelist is the connection ends up sitting
1470 1471 * there for too long.
1471 1472 */
1472 1473 void *
1473 1474 tcp_get_conn(void *arg, tcp_stack_t *tcps)
1474 1475 {
1475 1476 tcp_t *tcp = NULL;
1476 1477 conn_t *connp = NULL;
1477 1478 squeue_t *sqp = (squeue_t *)arg;
1478 1479 tcp_squeue_priv_t *tcp_time_wait;
1479 1480 netstack_t *ns;
1480 1481 mblk_t *tcp_rsrv_mp = NULL;
1481 1482
1482 1483 tcp_time_wait =
1483 1484 *((tcp_squeue_priv_t **)squeue_getprivate(sqp, SQPRIVATE_TCP));
1484 1485
1485 1486 mutex_enter(&tcp_time_wait->tcp_time_wait_lock);
1486 1487 tcp = tcp_time_wait->tcp_free_list;
1487 1488 ASSERT((tcp != NULL) ^ (tcp_time_wait->tcp_free_list_cnt == 0));
1488 1489 if (tcp != NULL) {
1489 1490 tcp_time_wait->tcp_free_list = tcp->tcp_time_wait_next;
1490 1491 tcp_time_wait->tcp_free_list_cnt--;
1491 1492 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1492 1493 tcp->tcp_time_wait_next = NULL;
1493 1494 connp = tcp->tcp_connp;
1494 1495 connp->conn_flags |= IPCL_REUSED;
1495 1496
1496 1497 ASSERT(tcp->tcp_tcps == NULL);
1497 1498 ASSERT(connp->conn_netstack == NULL);
1498 1499 ASSERT(tcp->tcp_rsrv_mp != NULL);
1499 1500 ns = tcps->tcps_netstack;
1500 1501 netstack_hold(ns);
1501 1502 connp->conn_netstack = ns;
1502 1503 connp->conn_ixa->ixa_ipst = ns->netstack_ip;
1503 1504 tcp->tcp_tcps = tcps;
1504 1505 ipcl_globalhash_insert(connp);
1505 1506
1506 1507 connp->conn_ixa->ixa_notify_cookie = tcp;
1507 1508 ASSERT(connp->conn_ixa->ixa_notify == tcp_notify);
1508 1509 connp->conn_recv = tcp_input_data;
1509 1510 ASSERT(connp->conn_recvicmp == tcp_icmp_input);
1510 1511 ASSERT(connp->conn_verifyicmp == tcp_verifyicmp);
1511 1512 return ((void *)connp);
1512 1513 }
1513 1514 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1514 1515 /*
1515 1516 * Pre-allocate the tcp_rsrv_mp. This mblk will not be freed until
1516 1517 * this conn_t/tcp_t is freed at ipcl_conn_destroy().
1517 1518 */
1518 1519 tcp_rsrv_mp = allocb(0, BPRI_HI);
1519 1520 if (tcp_rsrv_mp == NULL)
1520 1521 return (NULL);
1521 1522
1522 1523 if ((connp = ipcl_conn_create(IPCL_TCPCONN, KM_NOSLEEP,
1523 1524 tcps->tcps_netstack)) == NULL) {
1524 1525 freeb(tcp_rsrv_mp);
1525 1526 return (NULL);
1526 1527 }
1527 1528
1528 1529 tcp = connp->conn_tcp;
1529 1530 tcp->tcp_rsrv_mp = tcp_rsrv_mp;
1530 1531 mutex_init(&tcp->tcp_rsrv_mp_lock, NULL, MUTEX_DEFAULT, NULL);
1531 1532
1532 1533 tcp->tcp_tcps = tcps;
1533 1534
1534 1535 connp->conn_recv = tcp_input_data;
1535 1536 connp->conn_recvicmp = tcp_icmp_input;
1536 1537 connp->conn_verifyicmp = tcp_verifyicmp;
1537 1538
1538 1539 /*
1539 1540 * Register tcp_notify to listen to capability changes detected by IP.
1540 1541 * This upcall is made in the context of the call to conn_ip_output
1541 1542 * thus it is inside the squeue.
1542 1543 */
1543 1544 connp->conn_ixa->ixa_notify = tcp_notify;
1544 1545 connp->conn_ixa->ixa_notify_cookie = tcp;
1545 1546
1546 1547 return ((void *)connp);
1547 1548 }
1548 1549
1549 1550 /*
1550 1551 * Handle connect to IPv4 destinations, including connections for AF_INET6
1551 1552 * sockets connecting to IPv4 mapped IPv6 destinations.
1552 1553 * Returns zero if OK, a positive errno, or a negative TLI error.
1553 1554 */
1554 1555 static int
1555 1556 tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp, in_port_t dstport,
1556 1557 uint_t srcid)
1557 1558 {
1558 1559 ipaddr_t dstaddr = *dstaddrp;
1559 1560 uint16_t lport;
1560 1561 conn_t *connp = tcp->tcp_connp;
1561 1562 tcp_stack_t *tcps = tcp->tcp_tcps;
1562 1563 int error;
1563 1564
1564 1565 ASSERT(connp->conn_ipversion == IPV4_VERSION);
1565 1566
1566 1567 /* Check for attempt to connect to INADDR_ANY */
1567 1568 if (dstaddr == INADDR_ANY) {
1568 1569 /*
1569 1570 * SunOS 4.x and 4.3 BSD allow an application
1570 1571 * to connect a TCP socket to INADDR_ANY.
1571 1572 * When they do this, the kernel picks the
1572 1573 * address of one interface and uses it
1573 1574 * instead. The kernel usually ends up
1574 1575 * picking the address of the loopback
1575 1576 * interface. This is an undocumented feature.
1576 1577 * However, we provide the same thing here
1577 1578 * in order to have source and binary
1578 1579 * compatibility with SunOS 4.x.
1579 1580 * Update the T_CONN_REQ (sin/sin6) since it is used to
1580 1581 * generate the T_CONN_CON.
1581 1582 */
1582 1583 dstaddr = htonl(INADDR_LOOPBACK);
1583 1584 *dstaddrp = dstaddr;
1584 1585 }
1585 1586
1586 1587 /* Handle __sin6_src_id if socket not bound to an IP address */
1587 1588 if (srcid != 0 && connp->conn_laddr_v4 == INADDR_ANY) {
1588 1589 ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1589 1590 IPCL_ZONEID(connp), tcps->tcps_netstack);
1590 1591 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1591 1592 }
1592 1593
1593 1594 IN6_IPADDR_TO_V4MAPPED(dstaddr, &connp->conn_faddr_v6);
1594 1595 connp->conn_fport = dstport;
1595 1596
1596 1597 /*
1597 1598 * At this point the remote destination address and remote port fields
1598 1599 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1599 1600 * have to see which state tcp was in so we can take appropriate action.
1600 1601 */
1601 1602 if (tcp->tcp_state == TCPS_IDLE) {
1602 1603 /*
1603 1604 * We support a quick connect capability here, allowing
1604 1605 * clients to transition directly from IDLE to SYN_SENT
1605 1606 * tcp_bindi will pick an unused port, insert the connection
1606 1607 * in the bind hash and transition to BOUND state.
1607 1608 */
1608 1609 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1609 1610 tcp, B_TRUE);
1610 1611 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1611 1612 B_FALSE, B_FALSE);
1612 1613 if (lport == 0)
1613 1614 return (-TNOADDR);
1614 1615 }
1615 1616
1616 1617 /*
1617 1618 * Lookup the route to determine a source address and the uinfo.
1618 1619 * Setup TCP parameters based on the metrics/DCE.
1619 1620 */
1620 1621 error = tcp_set_destination(tcp);
1621 1622 if (error != 0)
1622 1623 return (error);
1623 1624
1624 1625 /*
1625 1626 * Don't let an endpoint connect to itself.
1626 1627 */
1627 1628 if (connp->conn_faddr_v4 == connp->conn_laddr_v4 &&
1628 1629 connp->conn_fport == connp->conn_lport)
1629 1630 return (-TBADADDR);
1630 1631
1631 1632 tcp->tcp_state = TCPS_SYN_SENT;
1632 1633
1633 1634 return (ipcl_conn_insert_v4(connp));
1634 1635 }
1635 1636
1636 1637 /*
1637 1638 * Handle connect to IPv6 destinations.
1638 1639 * Returns zero if OK, a positive errno, or a negative TLI error.
1639 1640 */
1640 1641 static int
1641 1642 tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp, in_port_t dstport,
1642 1643 uint32_t flowinfo, uint_t srcid, uint32_t scope_id)
1643 1644 {
1644 1645 uint16_t lport;
1645 1646 conn_t *connp = tcp->tcp_connp;
1646 1647 tcp_stack_t *tcps = tcp->tcp_tcps;
1647 1648 int error;
1648 1649
1649 1650 ASSERT(connp->conn_family == AF_INET6);
1650 1651
1651 1652 /*
1652 1653 * If we're here, it means that the destination address is a native
1653 1654 * IPv6 address. Return an error if conn_ipversion is not IPv6. A
1654 1655 * reason why it might not be IPv6 is if the socket was bound to an
1655 1656 * IPv4-mapped IPv6 address.
1656 1657 */
1657 1658 if (connp->conn_ipversion != IPV6_VERSION)
1658 1659 return (-TBADADDR);
1659 1660
1660 1661 /*
1661 1662 * Interpret a zero destination to mean loopback.
1662 1663 * Update the T_CONN_REQ (sin/sin6) since it is used to
1663 1664 * generate the T_CONN_CON.
1664 1665 */
1665 1666 if (IN6_IS_ADDR_UNSPECIFIED(dstaddrp))
1666 1667 *dstaddrp = ipv6_loopback;
1667 1668
1668 1669 /* Handle __sin6_src_id if socket not bound to an IP address */
1669 1670 if (srcid != 0 && IN6_IS_ADDR_UNSPECIFIED(&connp->conn_laddr_v6)) {
1670 1671 ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1671 1672 IPCL_ZONEID(connp), tcps->tcps_netstack);
1672 1673 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1673 1674 }
1674 1675
1675 1676 /*
1676 1677 * Take care of the scope_id now.
1677 1678 */
1678 1679 if (scope_id != 0 && IN6_IS_ADDR_LINKSCOPE(dstaddrp)) {
1679 1680 connp->conn_ixa->ixa_flags |= IXAF_SCOPEID_SET;
1680 1681 connp->conn_ixa->ixa_scopeid = scope_id;
1681 1682 } else {
1682 1683 connp->conn_ixa->ixa_flags &= ~IXAF_SCOPEID_SET;
1683 1684 }
1684 1685
1685 1686 connp->conn_flowinfo = flowinfo;
1686 1687 connp->conn_faddr_v6 = *dstaddrp;
1687 1688 connp->conn_fport = dstport;
1688 1689
1689 1690 /*
1690 1691 * At this point the remote destination address and remote port fields
1691 1692 * in the tcp-four-tuple have been filled in the tcp structure. Now we
1692 1693 * have to see which state tcp was in so we can take appropriate action.
1693 1694 */
1694 1695 if (tcp->tcp_state == TCPS_IDLE) {
1695 1696 /*
1696 1697 * We support a quick connect capability here, allowing
1697 1698 * clients to transition directly from IDLE to SYN_SENT
1698 1699 * tcp_bindi will pick an unused port, insert the connection
1699 1700 * in the bind hash and transition to BOUND state.
1700 1701 */
1701 1702 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1702 1703 tcp, B_TRUE);
1703 1704 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1704 1705 B_FALSE, B_FALSE);
1705 1706 if (lport == 0)
1706 1707 return (-TNOADDR);
1707 1708 }
1708 1709
1709 1710 /*
1710 1711 * Lookup the route to determine a source address and the uinfo.
1711 1712 * Setup TCP parameters based on the metrics/DCE.
1712 1713 */
1713 1714 error = tcp_set_destination(tcp);
1714 1715 if (error != 0)
1715 1716 return (error);
1716 1717
1717 1718 /*
1718 1719 * Don't let an endpoint connect to itself.
1719 1720 */
1720 1721 if (IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6, &connp->conn_laddr_v6) &&
1721 1722 connp->conn_fport == connp->conn_lport)
1722 1723 return (-TBADADDR);
1723 1724
1724 1725 tcp->tcp_state = TCPS_SYN_SENT;
1725 1726
1726 1727 return (ipcl_conn_insert_v6(connp));
1727 1728 }
1728 1729
1729 1730 /*
1730 1731 * Disconnect
1731 1732 * Note that unlike other functions this returns a positive tli error
1732 1733 * when it fails; it never returns an errno.
1733 1734 */
1734 1735 static int
1735 1736 tcp_disconnect_common(tcp_t *tcp, t_scalar_t seqnum)
1736 1737 {
1737 1738 conn_t *lconnp;
1738 1739 tcp_stack_t *tcps = tcp->tcp_tcps;
1739 1740 conn_t *connp = tcp->tcp_connp;
1740 1741
1741 1742 /*
1742 1743 * Right now, upper modules pass down a T_DISCON_REQ to TCP,
1743 1744 * when the stream is in BOUND state. Do not send a reset,
1744 1745 * since the destination IP address is not valid, and it can
1745 1746 * be the initialized value of all zeros (broadcast address).
1746 1747 */
1747 1748 if (tcp->tcp_state <= TCPS_BOUND) {
1748 1749 if (connp->conn_debug) {
1749 1750 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1750 1751 "tcp_disconnect: bad state, %d", tcp->tcp_state);
1751 1752 }
1752 1753 return (TOUTSTATE);
1753 1754 } else if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1754 1755 TCPS_CONN_DEC(tcps);
1755 1756 }
1756 1757
1757 1758 if (seqnum == -1 || tcp->tcp_conn_req_max == 0) {
1758 1759
1759 1760 /*
1760 1761 * According to TPI, for non-listeners, ignore seqnum
1761 1762 * and disconnect.
1762 1763 * Following interpretation of -1 seqnum is historical
1763 1764 * and implied TPI ? (TPI only states that for T_CONN_IND,
1764 1765 * a valid seqnum should not be -1).
1765 1766 *
1766 1767 * -1 means disconnect everything
1767 1768 * regardless even on a listener.
1768 1769 */
1769 1770
1770 1771 int old_state = tcp->tcp_state;
1771 1772 ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
1772 1773
1773 1774 /*
1774 1775 * The connection can't be on the tcp_time_wait_head list
1775 1776 * since it is not detached.
1776 1777 */
1777 1778 ASSERT(tcp->tcp_time_wait_next == NULL);
1778 1779 ASSERT(tcp->tcp_time_wait_prev == NULL);
1779 1780 ASSERT(tcp->tcp_time_wait_expire == 0);
1780 1781 /*
1781 1782 * If it used to be a listener, check to make sure no one else
1782 1783 * has taken the port before switching back to LISTEN state.
1783 1784 */
1784 1785 if (connp->conn_ipversion == IPV4_VERSION) {
1785 1786 lconnp = ipcl_lookup_listener_v4(connp->conn_lport,
1786 1787 connp->conn_laddr_v4, IPCL_ZONEID(connp), ipst);
1787 1788 } else {
1788 1789 uint_t ifindex = 0;
1789 1790
1790 1791 if (connp->conn_ixa->ixa_flags & IXAF_SCOPEID_SET)
1791 1792 ifindex = connp->conn_ixa->ixa_scopeid;
1792 1793
1793 1794 /* Allow conn_bound_if listeners? */
1794 1795 lconnp = ipcl_lookup_listener_v6(connp->conn_lport,
1795 1796 &connp->conn_laddr_v6, ifindex, IPCL_ZONEID(connp),
1796 1797 ipst);
1797 1798 }
1798 1799 if (tcp->tcp_conn_req_max && lconnp == NULL) {
1799 1800 tcp->tcp_state = TCPS_LISTEN;
1800 1801 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1801 1802 connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1802 1803 NULL, int32_t, old_state);
1803 1804 } else if (old_state > TCPS_BOUND) {
1804 1805 tcp->tcp_conn_req_max = 0;
1805 1806 tcp->tcp_state = TCPS_BOUND;
1806 1807 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1807 1808 connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1808 1809 NULL, int32_t, old_state);
1809 1810
1810 1811 /*
1811 1812 * If this end point is not going to become a listener,
1812 1813 * decrement the listener connection count if
1813 1814 * necessary. Note that we do not do this if it is
1814 1815 * going to be a listner (the above if case) since
1815 1816 * then it may remove the counter struct.
1816 1817 */
1817 1818 if (tcp->tcp_listen_cnt != NULL)
1818 1819 TCP_DECR_LISTEN_CNT(tcp);
1819 1820 }
1820 1821 if (lconnp != NULL)
1821 1822 CONN_DEC_REF(lconnp);
1822 1823 switch (old_state) {
1823 1824 case TCPS_SYN_SENT:
1824 1825 case TCPS_SYN_RCVD:
1825 1826 TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1826 1827 break;
1827 1828 case TCPS_ESTABLISHED:
1828 1829 case TCPS_CLOSE_WAIT:
1829 1830 TCPS_BUMP_MIB(tcps, tcpEstabResets);
1830 1831 break;
1831 1832 }
1832 1833
1833 1834 if (tcp->tcp_fused)
1834 1835 tcp_unfuse(tcp);
1835 1836
1836 1837 mutex_enter(&tcp->tcp_eager_lock);
1837 1838 if ((tcp->tcp_conn_req_cnt_q0 != 0) ||
1838 1839 (tcp->tcp_conn_req_cnt_q != 0)) {
1839 1840 tcp_eager_cleanup(tcp, 0);
1840 1841 }
1841 1842 mutex_exit(&tcp->tcp_eager_lock);
1842 1843
1843 1844 tcp_xmit_ctl("tcp_disconnect", tcp, tcp->tcp_snxt,
1844 1845 tcp->tcp_rnxt, TH_RST | TH_ACK);
1845 1846
1846 1847 tcp_reinit(tcp);
1847 1848
1848 1849 return (0);
1849 1850 } else if (!tcp_eager_blowoff(tcp, seqnum)) {
1850 1851 return (TBADSEQ);
1851 1852 }
1852 1853 return (0);
1853 1854 }
1854 1855
1855 1856 /*
1856 1857 * Our client hereby directs us to reject the connection request
1857 1858 * that tcp_input_listener() marked with 'seqnum'. Rejection consists
1858 1859 * of sending the appropriate RST, not an ICMP error.
1859 1860 */
1860 1861 void
1861 1862 tcp_disconnect(tcp_t *tcp, mblk_t *mp)
1862 1863 {
1863 1864 t_scalar_t seqnum;
1864 1865 int error;
1865 1866 conn_t *connp = tcp->tcp_connp;
1866 1867
1867 1868 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1868 1869 if ((mp->b_wptr - mp->b_rptr) < sizeof (struct T_discon_req)) {
1869 1870 tcp_err_ack(tcp, mp, TPROTO, 0);
1870 1871 return;
1871 1872 }
1872 1873 seqnum = ((struct T_discon_req *)mp->b_rptr)->SEQ_number;
1873 1874 error = tcp_disconnect_common(tcp, seqnum);
1874 1875 if (error != 0)
1875 1876 tcp_err_ack(tcp, mp, error, 0);
1876 1877 else {
1877 1878 if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1878 1879 /* Send M_FLUSH according to TPI */
1879 1880 (void) putnextctl1(connp->conn_rq, M_FLUSH, FLUSHRW);
1880 1881 }
1881 1882 mp = mi_tpi_ok_ack_alloc(mp);
1882 1883 if (mp != NULL)
1883 1884 putnext(connp->conn_rq, mp);
1884 1885 }
1885 1886 }
1886 1887
1887 1888 /*
1888 1889 * Handle reinitialization of a tcp structure.
1889 1890 * Maintain "binding state" resetting the state to BOUND, LISTEN, or IDLE.
1890 1891 */
1891 1892 static void
1892 1893 tcp_reinit(tcp_t *tcp)
1893 1894 {
1894 1895 mblk_t *mp;
1895 1896 tcp_stack_t *tcps = tcp->tcp_tcps;
1896 1897 conn_t *connp = tcp->tcp_connp;
1897 1898 int32_t oldstate;
1898 1899
1899 1900 /* tcp_reinit should never be called for detached tcp_t's */
1900 1901 ASSERT(tcp->tcp_listener == NULL);
1901 1902 ASSERT((connp->conn_family == AF_INET &&
1902 1903 connp->conn_ipversion == IPV4_VERSION) ||
1903 1904 (connp->conn_family == AF_INET6 &&
1904 1905 (connp->conn_ipversion == IPV4_VERSION ||
1905 1906 connp->conn_ipversion == IPV6_VERSION)));
1906 1907
1907 1908 /* Cancel outstanding timers */
1908 1909 tcp_timers_stop(tcp);
1909 1910
1910 1911 /*
1911 1912 * Reset everything in the state vector, after updating global
1912 1913 * MIB data from instance counters.
1913 1914 */
1914 1915 TCPS_UPDATE_MIB(tcps, tcpHCInSegs, tcp->tcp_ibsegs);
1915 1916 tcp->tcp_ibsegs = 0;
1916 1917 TCPS_UPDATE_MIB(tcps, tcpHCOutSegs, tcp->tcp_obsegs);
1917 1918 tcp->tcp_obsegs = 0;
1918 1919
1919 1920 tcp_close_mpp(&tcp->tcp_xmit_head);
1920 1921 if (tcp->tcp_snd_zcopy_aware)
1921 1922 tcp_zcopy_notify(tcp);
1922 1923 tcp->tcp_xmit_last = tcp->tcp_xmit_tail = NULL;
1923 1924 tcp->tcp_unsent = tcp->tcp_xmit_tail_unsent = 0;
1924 1925 mutex_enter(&tcp->tcp_non_sq_lock);
1925 1926 if (tcp->tcp_flow_stopped &&
1926 1927 TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1927 1928 tcp_clrqfull(tcp);
1928 1929 }
1929 1930 mutex_exit(&tcp->tcp_non_sq_lock);
1930 1931 tcp_close_mpp(&tcp->tcp_reass_head);
1931 1932 tcp->tcp_reass_tail = NULL;
1932 1933 if (tcp->tcp_rcv_list != NULL) {
1933 1934 /* Free b_next chain */
1934 1935 tcp_close_mpp(&tcp->tcp_rcv_list);
1935 1936 tcp->tcp_rcv_last_head = NULL;
1936 1937 tcp->tcp_rcv_last_tail = NULL;
1937 1938 tcp->tcp_rcv_cnt = 0;
1938 1939 }
1939 1940 tcp->tcp_rcv_last_tail = NULL;
1940 1941
1941 1942 if ((mp = tcp->tcp_urp_mp) != NULL) {
1942 1943 freemsg(mp);
1943 1944 tcp->tcp_urp_mp = NULL;
1944 1945 }
1945 1946 if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1946 1947 freemsg(mp);
1947 1948 tcp->tcp_urp_mark_mp = NULL;
1948 1949 }
1949 1950 if (tcp->tcp_fused_sigurg_mp != NULL) {
1950 1951 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1951 1952 freeb(tcp->tcp_fused_sigurg_mp);
1952 1953 tcp->tcp_fused_sigurg_mp = NULL;
1953 1954 }
1954 1955 if (tcp->tcp_ordrel_mp != NULL) {
1955 1956 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1956 1957 freeb(tcp->tcp_ordrel_mp);
1957 1958 tcp->tcp_ordrel_mp = NULL;
1958 1959 }
1959 1960
1960 1961 /*
1961 1962 * Following is a union with two members which are
1962 1963 * identical types and size so the following cleanup
1963 1964 * is enough.
1964 1965 */
1965 1966 tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1966 1967
1967 1968 CL_INET_DISCONNECT(connp);
1968 1969
1969 1970 /*
1970 1971 * The connection can't be on the tcp_time_wait_head list
1971 1972 * since it is not detached.
1972 1973 */
1973 1974 ASSERT(tcp->tcp_time_wait_next == NULL);
1974 1975 ASSERT(tcp->tcp_time_wait_prev == NULL);
1975 1976 ASSERT(tcp->tcp_time_wait_expire == 0);
1976 1977
1977 1978 /*
1978 1979 * Reset/preserve other values
1979 1980 */
1980 1981 tcp_reinit_values(tcp);
1981 1982 ipcl_hash_remove(connp);
1982 1983 /* Note that ixa_cred gets cleared in ixa_cleanup */
1983 1984 ixa_cleanup(connp->conn_ixa);
1984 1985 tcp_ipsec_cleanup(tcp);
1985 1986
1986 1987 connp->conn_laddr_v6 = connp->conn_bound_addr_v6;
1987 1988 connp->conn_saddr_v6 = connp->conn_bound_addr_v6;
1988 1989 oldstate = tcp->tcp_state;
1989 1990
1990 1991 if (tcp->tcp_conn_req_max != 0) {
1991 1992 /*
1992 1993 * This is the case when a TLI program uses the same
1993 1994 * transport end point to accept a connection. This
1994 1995 * makes the TCP both a listener and acceptor. When
1995 1996 * this connection is closed, we need to set the state
1996 1997 * back to TCPS_LISTEN. Make sure that the eager list
1997 1998 * is reinitialized.
1998 1999 *
1999 2000 * Note that this stream is still bound to the four
2000 2001 * tuples of the previous connection in IP. If a new
2001 2002 * SYN with different foreign address comes in, IP will
2002 2003 * not find it and will send it to the global queue. In
2003 2004 * the global queue, TCP will do a tcp_lookup_listener()
2004 2005 * to find this stream. This works because this stream
2005 2006 * is only removed from connected hash.
2006 2007 *
2007 2008 */
2008 2009 tcp->tcp_state = TCPS_LISTEN;
2009 2010 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
2010 2011 tcp->tcp_eager_next_drop_q0 = tcp;
2011 2012 tcp->tcp_eager_prev_drop_q0 = tcp;
2012 2013 /*
2013 2014 * Initially set conn_recv to tcp_input_listener_unbound to try
2014 2015 * to pick a good squeue for the listener when the first SYN
2015 2016 * arrives. tcp_input_listener_unbound sets it to
2016 2017 * tcp_input_listener on that first SYN.
2017 2018 */
2018 2019 connp->conn_recv = tcp_input_listener_unbound;
2019 2020
2020 2021 connp->conn_proto = IPPROTO_TCP;
2021 2022 connp->conn_faddr_v6 = ipv6_all_zeros;
2022 2023 connp->conn_fport = 0;
2023 2024
2024 2025 (void) ipcl_bind_insert(connp);
2025 2026 } else {
2026 2027 tcp->tcp_state = TCPS_BOUND;
2027 2028 }
2028 2029
2029 2030 /*
2030 2031 * Initialize to default values
2031 2032 */
2032 2033 tcp_init_values(tcp, NULL);
2033 2034
2034 2035 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
2035 2036 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
2036 2037 int32_t, oldstate);
2037 2038
2038 2039 ASSERT(tcp->tcp_ptpbhn != NULL);
2039 2040 tcp->tcp_rwnd = connp->conn_rcvbuf;
2040 2041 tcp->tcp_mss = connp->conn_ipversion != IPV4_VERSION ?
2041 2042 tcps->tcps_mss_def_ipv6 : tcps->tcps_mss_def_ipv4;
2042 2043 }
2043 2044
2044 2045 /*
2045 2046 * Force values to zero that need be zero.
2046 2047 * Do not touch values asociated with the BOUND or LISTEN state
2047 2048 * since the connection will end up in that state after the reinit.
2048 2049 * NOTE: tcp_reinit_values MUST have a line for each field in the tcp_t
2049 2050 * structure!
2050 2051 */
2051 2052 static void
2052 2053 tcp_reinit_values(tcp)
2053 2054 tcp_t *tcp;
2054 2055 {
2055 2056 tcp_stack_t *tcps = tcp->tcp_tcps;
2056 2057 conn_t *connp = tcp->tcp_connp;
2057 2058
2058 2059 #ifndef lint
2059 2060 #define DONTCARE(x)
2060 2061 #define PRESERVE(x)
2061 2062 #else
2062 2063 #define DONTCARE(x) ((x) = (x))
2063 2064 #define PRESERVE(x) ((x) = (x))
2064 2065 #endif /* lint */
2065 2066
2066 2067 PRESERVE(tcp->tcp_bind_hash_port);
2067 2068 PRESERVE(tcp->tcp_bind_hash);
2068 2069 PRESERVE(tcp->tcp_ptpbhn);
2069 2070 PRESERVE(tcp->tcp_acceptor_hash);
2070 2071 PRESERVE(tcp->tcp_ptpahn);
2071 2072
2072 2073 /* Should be ASSERT NULL on these with new code! */
2073 2074 ASSERT(tcp->tcp_time_wait_next == NULL);
2074 2075 ASSERT(tcp->tcp_time_wait_prev == NULL);
2075 2076 ASSERT(tcp->tcp_time_wait_expire == 0);
2076 2077 PRESERVE(tcp->tcp_state);
2077 2078 PRESERVE(connp->conn_rq);
2078 2079 PRESERVE(connp->conn_wq);
2079 2080
2080 2081 ASSERT(tcp->tcp_xmit_head == NULL);
2081 2082 ASSERT(tcp->tcp_xmit_last == NULL);
2082 2083 ASSERT(tcp->tcp_unsent == 0);
2083 2084 ASSERT(tcp->tcp_xmit_tail == NULL);
2084 2085 ASSERT(tcp->tcp_xmit_tail_unsent == 0);
2085 2086
2086 2087 tcp->tcp_snxt = 0; /* Displayed in mib */
2087 2088 tcp->tcp_suna = 0; /* Displayed in mib */
2088 2089 tcp->tcp_swnd = 0;
2089 2090 DONTCARE(tcp->tcp_cwnd); /* Init in tcp_process_options */
2090 2091
2091 2092 ASSERT(tcp->tcp_ibsegs == 0);
2092 2093 ASSERT(tcp->tcp_obsegs == 0);
2093 2094
2094 2095 if (connp->conn_ht_iphc != NULL) {
2095 2096 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
2096 2097 connp->conn_ht_iphc = NULL;
2097 2098 connp->conn_ht_iphc_allocated = 0;
2098 2099 connp->conn_ht_iphc_len = 0;
2099 2100 connp->conn_ht_ulp = NULL;
2100 2101 connp->conn_ht_ulp_len = 0;
2101 2102 tcp->tcp_ipha = NULL;
2102 2103 tcp->tcp_ip6h = NULL;
2103 2104 tcp->tcp_tcpha = NULL;
2104 2105 }
2105 2106
2106 2107 /* We clear any IP_OPTIONS and extension headers */
2107 2108 ip_pkt_free(&connp->conn_xmit_ipp);
2108 2109
2109 2110 DONTCARE(tcp->tcp_naglim); /* Init in tcp_init_values */
2110 2111 DONTCARE(tcp->tcp_ipha);
2111 2112 DONTCARE(tcp->tcp_ip6h);
2112 2113 DONTCARE(tcp->tcp_tcpha);
2113 2114 tcp->tcp_valid_bits = 0;
2114 2115
2115 2116 DONTCARE(tcp->tcp_timer_backoff); /* Init in tcp_init_values */
2116 2117 DONTCARE(tcp->tcp_last_recv_time); /* Init in tcp_init_values */
2117 2118 tcp->tcp_last_rcv_lbolt = 0;
2118 2119
2119 2120 tcp->tcp_init_cwnd = 0;
2120 2121
2121 2122 tcp->tcp_urp_last_valid = 0;
2122 2123 tcp->tcp_hard_binding = 0;
2123 2124
2124 2125 tcp->tcp_fin_acked = 0;
2125 2126 tcp->tcp_fin_rcvd = 0;
2126 2127 tcp->tcp_fin_sent = 0;
2127 2128 tcp->tcp_ordrel_done = 0;
2128 2129
2129 2130 tcp->tcp_detached = 0;
2130 2131
2131 2132 tcp->tcp_snd_ws_ok = B_FALSE;
2132 2133 tcp->tcp_snd_ts_ok = B_FALSE;
2133 2134 tcp->tcp_zero_win_probe = 0;
2134 2135
2135 2136 tcp->tcp_loopback = 0;
2136 2137 tcp->tcp_localnet = 0;
2137 2138 tcp->tcp_syn_defense = 0;
2138 2139 tcp->tcp_set_timer = 0;
2139 2140
2140 2141 tcp->tcp_active_open = 0;
2141 2142 tcp->tcp_rexmit = B_FALSE;
2142 2143 tcp->tcp_xmit_zc_clean = B_FALSE;
2143 2144
2144 2145 tcp->tcp_snd_sack_ok = B_FALSE;
2145 2146 tcp->tcp_hwcksum = B_FALSE;
2146 2147
2147 2148 DONTCARE(tcp->tcp_maxpsz_multiplier); /* Init in tcp_init_values */
2148 2149
2149 2150 tcp->tcp_conn_def_q0 = 0;
2150 2151 tcp->tcp_ip_forward_progress = B_FALSE;
2151 2152 tcp->tcp_ecn_ok = B_FALSE;
2152 2153
2153 2154 tcp->tcp_cwr = B_FALSE;
2154 2155 tcp->tcp_ecn_echo_on = B_FALSE;
2155 2156 tcp->tcp_is_wnd_shrnk = B_FALSE;
2156 2157
2157 2158 TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
2158 2159 bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
2159 2160
2160 2161 tcp->tcp_rcv_ws = 0;
2161 2162 tcp->tcp_snd_ws = 0;
2162 2163 tcp->tcp_ts_recent = 0;
2163 2164 tcp->tcp_rnxt = 0; /* Displayed in mib */
2164 2165 DONTCARE(tcp->tcp_rwnd); /* Set in tcp_reinit() */
2165 2166 tcp->tcp_initial_pmtu = 0;
2166 2167
2167 2168 ASSERT(tcp->tcp_reass_head == NULL);
2168 2169 ASSERT(tcp->tcp_reass_tail == NULL);
2169 2170
2170 2171 tcp->tcp_cwnd_cnt = 0;
2171 2172
2172 2173 ASSERT(tcp->tcp_rcv_list == NULL);
2173 2174 ASSERT(tcp->tcp_rcv_last_head == NULL);
2174 2175 ASSERT(tcp->tcp_rcv_last_tail == NULL);
2175 2176 ASSERT(tcp->tcp_rcv_cnt == 0);
2176 2177
2177 2178 DONTCARE(tcp->tcp_cwnd_ssthresh); /* Init in tcp_set_destination */
2178 2179 DONTCARE(tcp->tcp_cwnd_max); /* Init in tcp_init_values */
2179 2180 tcp->tcp_csuna = 0;
2180 2181
2181 2182 tcp->tcp_rto = 0; /* Displayed in MIB */
2182 2183 DONTCARE(tcp->tcp_rtt_sa); /* Init in tcp_init_values */
2183 2184 DONTCARE(tcp->tcp_rtt_sd); /* Init in tcp_init_values */
2184 2185 tcp->tcp_rtt_update = 0;
2185 2186
2186 2187 DONTCARE(tcp->tcp_swl1); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2187 2188 DONTCARE(tcp->tcp_swl2); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2188 2189
2189 2190 tcp->tcp_rack = 0; /* Displayed in mib */
2190 2191 tcp->tcp_rack_cnt = 0;
2191 2192 tcp->tcp_rack_cur_max = 0;
2192 2193 tcp->tcp_rack_abs_max = 0;
2193 2194
2194 2195 tcp->tcp_max_swnd = 0;
2195 2196
2196 2197 ASSERT(tcp->tcp_listener == NULL);
2197 2198
2198 2199 DONTCARE(tcp->tcp_irs); /* tcp_valid_bits cleared */
2199 2200 DONTCARE(tcp->tcp_iss); /* tcp_valid_bits cleared */
2200 2201 DONTCARE(tcp->tcp_fss); /* tcp_valid_bits cleared */
2201 2202 DONTCARE(tcp->tcp_urg); /* tcp_valid_bits cleared */
2202 2203
2203 2204 ASSERT(tcp->tcp_conn_req_cnt_q == 0);
2204 2205 ASSERT(tcp->tcp_conn_req_cnt_q0 == 0);
2205 2206 PRESERVE(tcp->tcp_conn_req_max);
2206 2207 PRESERVE(tcp->tcp_conn_req_seqnum);
2207 2208
2208 2209 DONTCARE(tcp->tcp_first_timer_threshold); /* Init in tcp_init_values */
2209 2210 DONTCARE(tcp->tcp_second_timer_threshold); /* Init in tcp_init_values */
2210 2211 DONTCARE(tcp->tcp_first_ctimer_threshold); /* Init in tcp_init_values */
2211 2212 DONTCARE(tcp->tcp_second_ctimer_threshold); /* in tcp_init_values */
2212 2213
2213 2214 DONTCARE(tcp->tcp_urp_last); /* tcp_urp_last_valid is cleared */
2214 2215 ASSERT(tcp->tcp_urp_mp == NULL);
2215 2216 ASSERT(tcp->tcp_urp_mark_mp == NULL);
2216 2217 ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
2217 2218
2218 2219 ASSERT(tcp->tcp_eager_next_q == NULL);
2219 2220 ASSERT(tcp->tcp_eager_last_q == NULL);
2220 2221 ASSERT((tcp->tcp_eager_next_q0 == NULL &&
2221 2222 tcp->tcp_eager_prev_q0 == NULL) ||
2222 2223 tcp->tcp_eager_next_q0 == tcp->tcp_eager_prev_q0);
2223 2224 ASSERT(tcp->tcp_conn.tcp_eager_conn_ind == NULL);
2224 2225
2225 2226 ASSERT((tcp->tcp_eager_next_drop_q0 == NULL &&
2226 2227 tcp->tcp_eager_prev_drop_q0 == NULL) ||
2227 2228 tcp->tcp_eager_next_drop_q0 == tcp->tcp_eager_prev_drop_q0);
2228 2229
2229 2230 tcp->tcp_client_errno = 0;
2230 2231
2231 2232 DONTCARE(connp->conn_sum); /* Init in tcp_init_values */
2232 2233
2233 2234 connp->conn_faddr_v6 = ipv6_all_zeros; /* Displayed in MIB */
2234 2235
2235 2236 PRESERVE(connp->conn_bound_addr_v6);
2236 2237 tcp->tcp_last_sent_len = 0;
2237 2238 tcp->tcp_dupack_cnt = 0;
2238 2239
2239 2240 connp->conn_fport = 0; /* Displayed in MIB */
2240 2241 PRESERVE(connp->conn_lport);
2241 2242
2242 2243 PRESERVE(tcp->tcp_acceptor_lockp);
2243 2244
2244 2245 ASSERT(tcp->tcp_ordrel_mp == NULL);
2245 2246 PRESERVE(tcp->tcp_acceptor_id);
2246 2247 DONTCARE(tcp->tcp_ipsec_overhead);
2247 2248
2248 2249 PRESERVE(connp->conn_family);
2249 2250 /* Remove any remnants of mapped address binding */
2250 2251 if (connp->conn_family == AF_INET6) {
2251 2252 connp->conn_ipversion = IPV6_VERSION;
2252 2253 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2253 2254 } else {
2254 2255 connp->conn_ipversion = IPV4_VERSION;
2255 2256 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2256 2257 }
2257 2258
2258 2259 connp->conn_bound_if = 0;
2259 2260 connp->conn_recv_ancillary.crb_all = 0;
2260 2261 tcp->tcp_recvifindex = 0;
2261 2262 tcp->tcp_recvhops = 0;
2262 2263 tcp->tcp_closed = 0;
2263 2264 if (tcp->tcp_hopopts != NULL) {
2264 2265 mi_free(tcp->tcp_hopopts);
2265 2266 tcp->tcp_hopopts = NULL;
2266 2267 tcp->tcp_hopoptslen = 0;
2267 2268 }
2268 2269 ASSERT(tcp->tcp_hopoptslen == 0);
2269 2270 if (tcp->tcp_dstopts != NULL) {
2270 2271 mi_free(tcp->tcp_dstopts);
2271 2272 tcp->tcp_dstopts = NULL;
2272 2273 tcp->tcp_dstoptslen = 0;
2273 2274 }
2274 2275 ASSERT(tcp->tcp_dstoptslen == 0);
2275 2276 if (tcp->tcp_rthdrdstopts != NULL) {
2276 2277 mi_free(tcp->tcp_rthdrdstopts);
2277 2278 tcp->tcp_rthdrdstopts = NULL;
2278 2279 tcp->tcp_rthdrdstoptslen = 0;
2279 2280 }
2280 2281 ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2281 2282 if (tcp->tcp_rthdr != NULL) {
2282 2283 mi_free(tcp->tcp_rthdr);
2283 2284 tcp->tcp_rthdr = NULL;
2284 2285 tcp->tcp_rthdrlen = 0;
2285 2286 }
2286 2287 ASSERT(tcp->tcp_rthdrlen == 0);
2287 2288
2288 2289 /* Reset fusion-related fields */
2289 2290 tcp->tcp_fused = B_FALSE;
2290 2291 tcp->tcp_unfusable = B_FALSE;
2291 2292 tcp->tcp_fused_sigurg = B_FALSE;
2292 2293 tcp->tcp_loopback_peer = NULL;
2293 2294
2294 2295 tcp->tcp_lso = B_FALSE;
2295 2296
2296 2297 tcp->tcp_in_ack_unsent = 0;
2297 2298 tcp->tcp_cork = B_FALSE;
2298 2299 tcp->tcp_tconnind_started = B_FALSE;
2299 2300
2300 2301 PRESERVE(tcp->tcp_squeue_bytes);
2301 2302
2302 2303 tcp->tcp_closemp_used = B_FALSE;
2303 2304
2304 2305 PRESERVE(tcp->tcp_rsrv_mp);
2305 2306 PRESERVE(tcp->tcp_rsrv_mp_lock);
2306 2307
2307 2308 #ifdef DEBUG
2308 2309 DONTCARE(tcp->tcmp_stk[0]);
2309 2310 #endif
2310 2311
2311 2312 PRESERVE(tcp->tcp_connid);
2312 2313
2313 2314 ASSERT(tcp->tcp_listen_cnt == NULL);
2314 2315 ASSERT(tcp->tcp_reass_tid == 0);
2315 2316
2316 2317 #undef DONTCARE
2317 2318 #undef PRESERVE
2318 2319 }
2319 2320
2320 2321 /*
2321 2322 * Initialize the various fields in tcp_t. If parent (the listener) is non
2322 2323 * NULL, certain values will be inheritted from it.
2323 2324 */
2324 2325 void
2325 2326 tcp_init_values(tcp_t *tcp, tcp_t *parent)
2326 2327 {
2327 2328 tcp_stack_t *tcps = tcp->tcp_tcps;
2328 2329 conn_t *connp = tcp->tcp_connp;
2329 2330 clock_t rto;
2330 2331
2331 2332 ASSERT((connp->conn_family == AF_INET &&
2332 2333 connp->conn_ipversion == IPV4_VERSION) ||
2333 2334 (connp->conn_family == AF_INET6 &&
2334 2335 (connp->conn_ipversion == IPV4_VERSION ||
2335 2336 connp->conn_ipversion == IPV6_VERSION)));
2336 2337
2337 2338 if (parent == NULL) {
2338 2339 tcp->tcp_naglim = tcps->tcps_naglim_def;
2339 2340
2340 2341 tcp->tcp_rto_initial = tcps->tcps_rexmit_interval_initial;
2341 2342 tcp->tcp_rto_min = tcps->tcps_rexmit_interval_min;
2342 2343 tcp->tcp_rto_max = tcps->tcps_rexmit_interval_max;
2343 2344
2344 2345 tcp->tcp_first_ctimer_threshold =
2345 2346 tcps->tcps_ip_notify_cinterval;
2346 2347 tcp->tcp_second_ctimer_threshold =
2347 2348 tcps->tcps_ip_abort_cinterval;
2348 2349 tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2349 2350 tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2350 2351
2351 2352 tcp->tcp_fin_wait_2_flush_interval =
2352 2353 tcps->tcps_fin_wait_2_flush_interval;
2353 2354
2354 2355 tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2355 2356 tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2356 2357
2357 2358 /*
2358 2359 * Default value of tcp_init_cwnd is 0, so no need to set here
2359 2360 * if parent is NULL. But we need to inherit it from parent.
2360 2361 */
2361 2362 } else {
2362 2363 /* Inherit various TCP parameters from the parent. */
2363 2364 tcp->tcp_naglim = parent->tcp_naglim;
2364 2365
2365 2366 tcp->tcp_rto_initial = parent->tcp_rto_initial;
2366 2367 tcp->tcp_rto_min = parent->tcp_rto_min;
2367 2368 tcp->tcp_rto_max = parent->tcp_rto_max;
2368 2369
2369 2370 tcp->tcp_first_ctimer_threshold =
2370 2371 parent->tcp_first_ctimer_threshold;
2371 2372 tcp->tcp_second_ctimer_threshold =
2372 2373 parent->tcp_second_ctimer_threshold;
2373 2374 tcp->tcp_first_timer_threshold =
2374 2375 parent->tcp_first_timer_threshold;
2375 2376 tcp->tcp_second_timer_threshold =
2376 2377 parent->tcp_second_timer_threshold;
2377 2378
2378 2379 tcp->tcp_fin_wait_2_flush_interval =
2379 2380 parent->tcp_fin_wait_2_flush_interval;
2380 2381
2381 2382 tcp->tcp_ka_interval = parent->tcp_ka_interval;
2382 2383 tcp->tcp_ka_abort_thres = parent->tcp_ka_abort_thres;
2383 2384
2384 2385 tcp->tcp_init_cwnd = parent->tcp_init_cwnd;
2385 2386 }
2386 2387
2387 2388 /*
2388 2389 * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2389 2390 * will be close to tcp_rexmit_interval_initial. By doing this, we
2390 2391 * allow the algorithm to adjust slowly to large fluctuations of RTT
2391 2392 * during first few transmissions of a connection as seen in slow
2392 2393 * links.
2393 2394 */
2394 2395 tcp->tcp_rtt_sa = tcp->tcp_rto_initial << 2;
2395 2396 tcp->tcp_rtt_sd = tcp->tcp_rto_initial >> 1;
2396 2397 rto = (tcp->tcp_rtt_sa >> 3) + tcp->tcp_rtt_sd +
2397 2398 tcps->tcps_rexmit_interval_extra + (tcp->tcp_rtt_sa >> 5) +
2398 2399 tcps->tcps_conn_grace_period;
2399 2400 TCP_SET_RTO(tcp, rto);
2400 2401
2401 2402 tcp->tcp_timer_backoff = 0;
2402 2403 tcp->tcp_ms_we_have_waited = 0;
2403 2404 tcp->tcp_last_recv_time = ddi_get_lbolt();
2404 2405 tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2405 2406 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2406 2407 tcp->tcp_snd_burst = TCP_CWND_INFINITE;
2407 2408
2408 2409 tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2409 2410
2410 2411 /* NOTE: ISS is now set in tcp_set_destination(). */
2411 2412
2412 2413 /* Reset fusion-related fields */
2413 2414 tcp->tcp_fused = B_FALSE;
2414 2415 tcp->tcp_unfusable = B_FALSE;
2415 2416 tcp->tcp_fused_sigurg = B_FALSE;
2416 2417 tcp->tcp_loopback_peer = NULL;
2417 2418
2418 2419 /* We rebuild the header template on the next connect/conn_request */
2419 2420
2420 2421 connp->conn_mlp_type = mlptSingle;
2421 2422
2422 2423 /*
2423 2424 * Init the window scale to the max so tcp_rwnd_set() won't pare
2424 2425 * down tcp_rwnd. tcp_set_destination() will set the right value later.
2425 2426 */
2426 2427 tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2427 2428 tcp->tcp_rwnd = connp->conn_rcvbuf;
2428 2429
2429 2430 tcp->tcp_cork = B_FALSE;
2430 2431 /*
2431 2432 * Init the tcp_debug option if it wasn't already set. This value
2432 2433 * determines whether TCP
2433 2434 * calls strlog() to print out debug messages. Doing this
2434 2435 * initialization here means that this value is not inherited thru
2435 2436 * tcp_reinit().
2436 2437 */
2437 2438 if (!connp->conn_debug)
2438 2439 connp->conn_debug = tcps->tcps_dbg;
2439 2440 }
2440 2441
2441 2442 /*
2442 2443 * Update the TCP connection according to change of PMTU.
2443 2444 *
2444 2445 * Path MTU might have changed by either increase or decrease, so need to
2445 2446 * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2446 2447 * or negative MSS, since tcp_mss_set() will do it.
2447 2448 */
2448 2449 void
2449 2450 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2450 2451 {
2451 2452 uint32_t pmtu;
2452 2453 int32_t mss;
2453 2454 conn_t *connp = tcp->tcp_connp;
2454 2455 ip_xmit_attr_t *ixa = connp->conn_ixa;
2455 2456 iaflags_t ixaflags;
2456 2457
2457 2458 if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2458 2459 return;
2459 2460
2460 2461 if (tcp->tcp_state < TCPS_ESTABLISHED)
2461 2462 return;
2462 2463
2463 2464 /*
2464 2465 * Always call ip_get_pmtu() to make sure that IP has updated
2465 2466 * ixa_flags properly.
2466 2467 */
2467 2468 pmtu = ip_get_pmtu(ixa);
2468 2469 ixaflags = ixa->ixa_flags;
2469 2470
2470 2471 /*
2471 2472 * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2472 2473 * IPsec overhead if applied. Make sure to use the most recent
2473 2474 * IPsec information.
2474 2475 */
2475 2476 mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2476 2477
2477 2478 /*
2478 2479 * Nothing to change, so just return.
2479 2480 */
2480 2481 if (mss == tcp->tcp_mss)
2481 2482 return;
2482 2483
2483 2484 /*
2484 2485 * Currently, for ICMP errors, only PMTU decrease is handled.
2485 2486 */
2486 2487 if (mss > tcp->tcp_mss && decrease_only)
2487 2488 return;
2488 2489
2489 2490 DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2490 2491
2491 2492 /*
2492 2493 * Update ixa_fragsize and ixa_pmtu.
2493 2494 */
2494 2495 ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2495 2496
2496 2497 /*
2497 2498 * Adjust MSS and all relevant variables.
2498 2499 */
2499 2500 tcp_mss_set(tcp, mss);
2500 2501
2501 2502 /*
2502 2503 * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2503 2504 * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2504 2505 * has a (potentially different) min size we do the same. Make sure to
2505 2506 * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2506 2507 * fragment the packet.
2507 2508 *
2508 2509 * LSO over IPv6 can not be fragmented. So need to disable LSO
2509 2510 * when IPv6 fragmentation is needed.
2510 2511 */
2511 2512 if (mss < tcp->tcp_tcps->tcps_mss_min)
2512 2513 ixaflags |= IXAF_PMTU_TOO_SMALL;
2513 2514
2514 2515 if (ixaflags & IXAF_PMTU_TOO_SMALL)
2515 2516 ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2516 2517
2517 2518 if ((connp->conn_ipversion == IPV4_VERSION) &&
2518 2519 !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2519 2520 tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2520 2521 }
2521 2522 ixa->ixa_flags = ixaflags;
2522 2523 }
2523 2524
2524 2525 int
2525 2526 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2526 2527 {
2527 2528 conn_t *connp = tcp->tcp_connp;
2528 2529 queue_t *q = connp->conn_rq;
2529 2530 int32_t mss = tcp->tcp_mss;
2530 2531 int maxpsz;
2531 2532
2532 2533 if (TCP_IS_DETACHED(tcp))
2533 2534 return (mss);
2534 2535 if (tcp->tcp_fused) {
2535 2536 maxpsz = tcp_fuse_maxpsz(tcp);
2536 2537 mss = INFPSZ;
2537 2538 } else if (tcp->tcp_maxpsz_multiplier == 0) {
2538 2539 /*
2539 2540 * Set the sd_qn_maxpsz according to the socket send buffer
2540 2541 * size, and sd_maxblk to INFPSZ (-1). This will essentially
2541 2542 * instruct the stream head to copyin user data into contiguous
2542 2543 * kernel-allocated buffers without breaking it up into smaller
2543 2544 * chunks. We round up the buffer size to the nearest SMSS.
2544 2545 */
2545 2546 maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2546 2547 mss = INFPSZ;
2547 2548 } else {
2548 2549 /*
2549 2550 * Set sd_qn_maxpsz to approx half the (receivers) buffer
2550 2551 * (and a multiple of the mss). This instructs the stream
2551 2552 * head to break down larger than SMSS writes into SMSS-
2552 2553 * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2553 2554 */
2554 2555 maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2555 2556 if (maxpsz > connp->conn_sndbuf / 2) {
2556 2557 maxpsz = connp->conn_sndbuf / 2;
2557 2558 /* Round up to nearest mss */
2558 2559 maxpsz = MSS_ROUNDUP(maxpsz, mss);
2559 2560 }
2560 2561 }
2561 2562
2562 2563 (void) proto_set_maxpsz(q, connp, maxpsz);
2563 2564 if (!(IPCL_IS_NONSTR(connp)))
2564 2565 connp->conn_wq->q_maxpsz = maxpsz;
2565 2566 if (set_maxblk)
2566 2567 (void) proto_set_tx_maxblk(q, connp, mss);
2567 2568 return (mss);
2568 2569 }
2569 2570
2570 2571 /* For /dev/tcp aka AF_INET open */
2571 2572 static int
2572 2573 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2573 2574 {
2574 2575 return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2575 2576 }
2576 2577
2577 2578 /* For /dev/tcp6 aka AF_INET6 open */
2578 2579 static int
2579 2580 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2580 2581 {
2581 2582 return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2582 2583 }
2583 2584
2584 2585 conn_t *
2585 2586 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2586 2587 int *errorp)
2587 2588 {
2588 2589 tcp_t *tcp = NULL;
2589 2590 conn_t *connp;
2590 2591 zoneid_t zoneid;
2591 2592 tcp_stack_t *tcps;
2592 2593 squeue_t *sqp;
2593 2594
2594 2595 ASSERT(errorp != NULL);
2595 2596 /*
2596 2597 * Find the proper zoneid and netstack.
2597 2598 */
2598 2599 /*
2599 2600 * Special case for install: miniroot needs to be able to
2600 2601 * access files via NFS as though it were always in the
2601 2602 * global zone.
2602 2603 */
2603 2604 if (credp == kcred && nfs_global_client_only != 0) {
2604 2605 zoneid = GLOBAL_ZONEID;
2605 2606 tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2606 2607 netstack_tcp;
2607 2608 ASSERT(tcps != NULL);
2608 2609 } else {
2609 2610 netstack_t *ns;
2610 2611 int err;
2611 2612
2612 2613 if ((err = secpolicy_basic_net_access(credp)) != 0) {
2613 2614 *errorp = err;
2614 2615 return (NULL);
2615 2616 }
2616 2617
2617 2618 ns = netstack_find_by_cred(credp);
2618 2619 ASSERT(ns != NULL);
2619 2620 tcps = ns->netstack_tcp;
2620 2621 ASSERT(tcps != NULL);
2621 2622
2622 2623 /*
2623 2624 * For exclusive stacks we set the zoneid to zero
2624 2625 * to make TCP operate as if in the global zone.
2625 2626 */
2626 2627 if (tcps->tcps_netstack->netstack_stackid !=
2627 2628 GLOBAL_NETSTACKID)
2628 2629 zoneid = GLOBAL_ZONEID;
2629 2630 else
2630 2631 zoneid = crgetzoneid(credp);
2631 2632 }
2632 2633
2633 2634 sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2634 2635 connp = (conn_t *)tcp_get_conn(sqp, tcps);
2635 2636 /*
2636 2637 * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2637 2638 * so we drop it by one.
2638 2639 */
2639 2640 netstack_rele(tcps->tcps_netstack);
2640 2641 if (connp == NULL) {
2641 2642 *errorp = ENOSR;
2642 2643 return (NULL);
2643 2644 }
2644 2645 ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2645 2646
2646 2647 connp->conn_sqp = sqp;
2647 2648 connp->conn_initial_sqp = connp->conn_sqp;
2648 2649 connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2649 2650 tcp = connp->conn_tcp;
2650 2651
2651 2652 /*
2652 2653 * Besides asking IP to set the checksum for us, have conn_ip_output
2653 2654 * to do the following checks when necessary:
2654 2655 *
2655 2656 * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2656 2657 * IXAF_VERIFY_PMTU: verify PMTU changes
2657 2658 * IXAF_VERIFY_LSO: verify LSO capability changes
2658 2659 */
2659 2660 connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2660 2661 IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2661 2662
2662 2663 if (!tcps->tcps_dev_flow_ctl)
2663 2664 connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2664 2665
2665 2666 if (isv6) {
2666 2667 connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2667 2668 connp->conn_ipversion = IPV6_VERSION;
2668 2669 connp->conn_family = AF_INET6;
2669 2670 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2670 2671 connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2671 2672 } else {
2672 2673 connp->conn_ipversion = IPV4_VERSION;
2673 2674 connp->conn_family = AF_INET;
2674 2675 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2675 2676 connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2676 2677 }
2677 2678 connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2678 2679
2679 2680 crhold(credp);
2680 2681 connp->conn_cred = credp;
2681 2682 connp->conn_cpid = curproc->p_pid;
2682 2683 connp->conn_open_time = ddi_get_lbolt64();
2683 2684
2684 2685 /* Cache things in the ixa without any refhold */
2685 2686 ASSERT(!(connp->conn_ixa->ixa_free_flags & IXA_FREE_CRED));
2686 2687 connp->conn_ixa->ixa_cred = credp;
2687 2688 connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2688 2689
2689 2690 connp->conn_zoneid = zoneid;
2690 2691 /* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2691 2692 connp->conn_ixa->ixa_zoneid = zoneid;
2692 2693 connp->conn_mlp_type = mlptSingle;
2693 2694 ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2694 2695 ASSERT(tcp->tcp_tcps == tcps);
2695 2696
2696 2697 /*
2697 2698 * If the caller has the process-wide flag set, then default to MAC
2698 2699 * exempt mode. This allows read-down to unlabeled hosts.
2699 2700 */
2700 2701 if (getpflags(NET_MAC_AWARE, credp) != 0)
2701 2702 connp->conn_mac_mode = CONN_MAC_AWARE;
2702 2703
2703 2704 connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2704 2705
2705 2706 if (issocket) {
2706 2707 tcp->tcp_issocket = 1;
2707 2708 }
2708 2709
2709 2710 connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2710 2711 connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2711 2712 connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2712 2713 connp->conn_so_type = SOCK_STREAM;
2713 2714 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2714 2715 tcps->tcps_wroff_xtra;
2715 2716
2716 2717 SOCK_CONNID_INIT(tcp->tcp_connid);
2717 2718 /* DTrace ignores this - it isn't a tcp:::state-change */
2718 2719 tcp->tcp_state = TCPS_IDLE;
2719 2720 tcp_init_values(tcp, NULL);
2720 2721 return (connp);
2721 2722 }
2722 2723
2723 2724 static int
2724 2725 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2725 2726 boolean_t isv6)
2726 2727 {
2727 2728 tcp_t *tcp = NULL;
2728 2729 conn_t *connp = NULL;
2729 2730 int err;
2730 2731 vmem_t *minor_arena = NULL;
2731 2732 dev_t conn_dev;
2732 2733 boolean_t issocket;
2733 2734
2734 2735 if (q->q_ptr != NULL)
2735 2736 return (0);
2736 2737
2737 2738 if (sflag == MODOPEN)
2738 2739 return (EINVAL);
2739 2740
2740 2741 if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2741 2742 ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2742 2743 minor_arena = ip_minor_arena_la;
2743 2744 } else {
2744 2745 /*
2745 2746 * Either minor numbers in the large arena were exhausted
2746 2747 * or a non socket application is doing the open.
2747 2748 * Try to allocate from the small arena.
2748 2749 */
2749 2750 if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2750 2751 return (EBUSY);
2751 2752 }
2752 2753 minor_arena = ip_minor_arena_sa;
2753 2754 }
2754 2755
2755 2756 ASSERT(minor_arena != NULL);
2756 2757
2757 2758 *devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2758 2759
2759 2760 if (flag & SO_FALLBACK) {
2760 2761 /*
2761 2762 * Non streams socket needs a stream to fallback to
2762 2763 */
2763 2764 RD(q)->q_ptr = (void *)conn_dev;
2764 2765 WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2765 2766 WR(q)->q_ptr = (void *)minor_arena;
2766 2767 qprocson(q);
2767 2768 return (0);
2768 2769 } else if (flag & SO_ACCEPTOR) {
2769 2770 q->q_qinfo = &tcp_acceptor_rinit;
2770 2771 /*
2771 2772 * the conn_dev and minor_arena will be subsequently used by
2772 2773 * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2773 2774 * the minor device number for this connection from the q_ptr.
2774 2775 */
2775 2776 RD(q)->q_ptr = (void *)conn_dev;
2776 2777 WR(q)->q_qinfo = &tcp_acceptor_winit;
2777 2778 WR(q)->q_ptr = (void *)minor_arena;
2778 2779 qprocson(q);
2779 2780 return (0);
2780 2781 }
2781 2782
2782 2783 issocket = flag & SO_SOCKSTR;
2783 2784 connp = tcp_create_common(credp, isv6, issocket, &err);
2784 2785
2785 2786 if (connp == NULL) {
2786 2787 inet_minor_free(minor_arena, conn_dev);
2787 2788 q->q_ptr = WR(q)->q_ptr = NULL;
2788 2789 return (err);
2789 2790 }
2790 2791
2791 2792 connp->conn_rq = q;
2792 2793 connp->conn_wq = WR(q);
2793 2794 q->q_ptr = WR(q)->q_ptr = connp;
2794 2795
2795 2796 connp->conn_dev = conn_dev;
2796 2797 connp->conn_minor_arena = minor_arena;
2797 2798
2798 2799 ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2799 2800 ASSERT(WR(q)->q_qinfo == &tcp_winit);
2800 2801
2801 2802 tcp = connp->conn_tcp;
2802 2803
2803 2804 if (issocket) {
2804 2805 WR(q)->q_qinfo = &tcp_sock_winit;
2805 2806 } else {
2806 2807 #ifdef _ILP32
2807 2808 tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2808 2809 #else
2809 2810 tcp->tcp_acceptor_id = conn_dev;
2810 2811 #endif /* _ILP32 */
2811 2812 tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2812 2813 }
2813 2814
2814 2815 /*
2815 2816 * Put the ref for TCP. Ref for IP was already put
2816 2817 * by ipcl_conn_create. Also Make the conn_t globally
2817 2818 * visible to walkers
2818 2819 */
2819 2820 mutex_enter(&connp->conn_lock);
2820 2821 CONN_INC_REF_LOCKED(connp);
2821 2822 ASSERT(connp->conn_ref == 2);
2822 2823 connp->conn_state_flags &= ~CONN_INCIPIENT;
2823 2824 mutex_exit(&connp->conn_lock);
2824 2825
2825 2826 qprocson(q);
2826 2827 return (0);
2827 2828 }
2828 2829
2829 2830 /*
2830 2831 * Build/update the tcp header template (in conn_ht_iphc) based on
2831 2832 * conn_xmit_ipp. The headers include ip6_t, any extension
2832 2833 * headers, and the maximum size tcp header (to avoid reallocation
2833 2834 * on the fly for additional tcp options).
2834 2835 *
2835 2836 * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2836 2837 * Returns failure if can't allocate memory.
2837 2838 */
2838 2839 int
2839 2840 tcp_build_hdrs(tcp_t *tcp)
2840 2841 {
2841 2842 tcp_stack_t *tcps = tcp->tcp_tcps;
2842 2843 conn_t *connp = tcp->tcp_connp;
2843 2844 char buf[TCP_MAX_HDR_LENGTH];
2844 2845 uint_t buflen;
2845 2846 uint_t ulplen = TCP_MIN_HEADER_LENGTH;
2846 2847 uint_t extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
2847 2848 tcpha_t *tcpha;
2848 2849 uint32_t cksum;
2849 2850 int error;
2850 2851
2851 2852 /*
2852 2853 * We might be called after the connection is set up, and we might
2853 2854 * have TS options already in the TCP header. Thus we save any
2854 2855 * existing tcp header.
2855 2856 */
2856 2857 buflen = connp->conn_ht_ulp_len;
2857 2858 if (buflen != 0) {
2858 2859 bcopy(connp->conn_ht_ulp, buf, buflen);
2859 2860 extralen -= buflen - ulplen;
2860 2861 ulplen = buflen;
2861 2862 }
2862 2863
2863 2864 /* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
2864 2865 mutex_enter(&connp->conn_lock);
2865 2866 error = conn_build_hdr_template(connp, ulplen, extralen,
2866 2867 &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
2867 2868 mutex_exit(&connp->conn_lock);
2868 2869 if (error != 0)
2869 2870 return (error);
2870 2871
2871 2872 /*
2872 2873 * Any routing header/option has been massaged. The checksum difference
2873 2874 * is stored in conn_sum for later use.
2874 2875 */
2875 2876 tcpha = (tcpha_t *)connp->conn_ht_ulp;
2876 2877 tcp->tcp_tcpha = tcpha;
2877 2878
2878 2879 /* restore any old tcp header */
2879 2880 if (buflen != 0) {
2880 2881 bcopy(buf, connp->conn_ht_ulp, buflen);
2881 2882 } else {
2882 2883 tcpha->tha_sum = 0;
2883 2884 tcpha->tha_urp = 0;
2884 2885 tcpha->tha_ack = 0;
2885 2886 tcpha->tha_offset_and_reserved = (5 << 4);
2886 2887 tcpha->tha_lport = connp->conn_lport;
2887 2888 tcpha->tha_fport = connp->conn_fport;
2888 2889 }
2889 2890
2890 2891 /*
2891 2892 * IP wants our header length in the checksum field to
2892 2893 * allow it to perform a single pseudo-header+checksum
2893 2894 * calculation on behalf of TCP.
2894 2895 * Include the adjustment for a source route once IP_OPTIONS is set.
2895 2896 */
2896 2897 cksum = sizeof (tcpha_t) + connp->conn_sum;
2897 2898 cksum = (cksum >> 16) + (cksum & 0xFFFF);
2898 2899 ASSERT(cksum < 0x10000);
2899 2900 tcpha->tha_sum = htons(cksum);
2900 2901
2901 2902 if (connp->conn_ipversion == IPV4_VERSION)
2902 2903 tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
2903 2904 else
2904 2905 tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
2905 2906
2906 2907 if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
2907 2908 connp->conn_wroff) {
2908 2909 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2909 2910 tcps->tcps_wroff_xtra;
2910 2911 (void) proto_set_tx_wroff(connp->conn_rq, connp,
2911 2912 connp->conn_wroff);
2912 2913 }
2913 2914 return (0);
2914 2915 }
2915 2916
2916 2917 /*
2917 2918 * tcp_rwnd_set() is called to adjust the receive window to a desired value.
2918 2919 * We do not allow the receive window to shrink. After setting rwnd,
2919 2920 * set the flow control hiwat of the stream.
2920 2921 *
2921 2922 * This function is called in 2 cases:
2922 2923 *
2923 2924 * 1) Before data transfer begins, in tcp_input_listener() for accepting a
2924 2925 * connection (passive open) and in tcp_input_data() for active connect.
2925 2926 * This is called after tcp_mss_set() when the desired MSS value is known.
2926 2927 * This makes sure that our window size is a mutiple of the other side's
2927 2928 * MSS.
2928 2929 * 2) Handling SO_RCVBUF option.
2929 2930 *
2930 2931 * It is ASSUMED that the requested size is a multiple of the current MSS.
2931 2932 *
2932 2933 * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
2933 2934 * user requests so.
2934 2935 */
2935 2936 int
2936 2937 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
2937 2938 {
2938 2939 uint32_t mss = tcp->tcp_mss;
2939 2940 uint32_t old_max_rwnd;
2940 2941 uint32_t max_transmittable_rwnd;
2941 2942 boolean_t tcp_detached = TCP_IS_DETACHED(tcp);
2942 2943 tcp_stack_t *tcps = tcp->tcp_tcps;
2943 2944 conn_t *connp = tcp->tcp_connp;
2944 2945
2945 2946 /*
2946 2947 * Insist on a receive window that is at least
2947 2948 * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
2948 2949 * funny TCP interactions of Nagle algorithm, SWS avoidance
2949 2950 * and delayed acknowledgement.
2950 2951 */
2951 2952 rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
2952 2953
2953 2954 if (tcp->tcp_fused) {
2954 2955 size_t sth_hiwat;
2955 2956 tcp_t *peer_tcp = tcp->tcp_loopback_peer;
2956 2957
2957 2958 ASSERT(peer_tcp != NULL);
2958 2959 sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
2959 2960 if (!tcp_detached) {
2960 2961 (void) proto_set_rx_hiwat(connp->conn_rq, connp,
2961 2962 sth_hiwat);
2962 2963 tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
2963 2964 }
2964 2965
2965 2966 /* Caller could have changed tcp_rwnd; update tha_win */
2966 2967 if (tcp->tcp_tcpha != NULL) {
2967 2968 tcp->tcp_tcpha->tha_win =
2968 2969 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2969 2970 }
2970 2971 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
2971 2972 tcp->tcp_cwnd_max = rwnd;
2972 2973
2973 2974 /*
2974 2975 * In the fusion case, the maxpsz stream head value of
2975 2976 * our peer is set according to its send buffer size
2976 2977 * and our receive buffer size; since the latter may
2977 2978 * have changed we need to update the peer's maxpsz.
2978 2979 */
2979 2980 (void) tcp_maxpsz_set(peer_tcp, B_TRUE);
2980 2981 return (sth_hiwat);
2981 2982 }
2982 2983
2983 2984 if (tcp_detached)
2984 2985 old_max_rwnd = tcp->tcp_rwnd;
2985 2986 else
2986 2987 old_max_rwnd = connp->conn_rcvbuf;
2987 2988
2988 2989
2989 2990 /*
2990 2991 * If window size info has already been exchanged, TCP should not
2991 2992 * shrink the window. Shrinking window is doable if done carefully.
2992 2993 * We may add that support later. But so far there is not a real
2993 2994 * need to do that.
2994 2995 */
2995 2996 if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
2996 2997 /* MSS may have changed, do a round up again. */
2997 2998 rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
2998 2999 }
2999 3000
3000 3001 /*
3001 3002 * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3002 3003 * can be applied even before the window scale option is decided.
3003 3004 */
3004 3005 max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3005 3006 if (rwnd > max_transmittable_rwnd) {
3006 3007 rwnd = max_transmittable_rwnd -
3007 3008 (max_transmittable_rwnd % mss);
3008 3009 if (rwnd < mss)
3009 3010 rwnd = max_transmittable_rwnd;
3010 3011 /*
3011 3012 * If we're over the limit we may have to back down tcp_rwnd.
3012 3013 * The increment below won't work for us. So we set all three
3013 3014 * here and the increment below will have no effect.
3014 3015 */
3015 3016 tcp->tcp_rwnd = old_max_rwnd = rwnd;
3016 3017 }
3017 3018 if (tcp->tcp_localnet) {
3018 3019 tcp->tcp_rack_abs_max =
3019 3020 MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3020 3021 } else {
3021 3022 /*
3022 3023 * For a remote host on a different subnet (through a router),
3023 3024 * we ack every other packet to be conforming to RFC1122.
3024 3025 * tcp_deferred_acks_max is default to 2.
3025 3026 */
3026 3027 tcp->tcp_rack_abs_max =
3027 3028 MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3028 3029 }
3029 3030 if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3030 3031 tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3031 3032 else
3032 3033 tcp->tcp_rack_cur_max = 0;
3033 3034 /*
3034 3035 * Increment the current rwnd by the amount the maximum grew (we
3035 3036 * can not overwrite it since we might be in the middle of a
3036 3037 * connection.)
3037 3038 */
3038 3039 tcp->tcp_rwnd += rwnd - old_max_rwnd;
3039 3040 connp->conn_rcvbuf = rwnd;
3040 3041
3041 3042 /* Are we already connected? */
3042 3043 if (tcp->tcp_tcpha != NULL) {
3043 3044 tcp->tcp_tcpha->tha_win =
3044 3045 htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3045 3046 }
3046 3047
3047 3048 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3048 3049 tcp->tcp_cwnd_max = rwnd;
3049 3050
3050 3051 if (tcp_detached)
3051 3052 return (rwnd);
3052 3053
3053 3054 tcp_set_recv_threshold(tcp, rwnd >> 3);
3054 3055
3055 3056 (void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3056 3057 return (rwnd);
3057 3058 }
3058 3059
3059 3060 int
3060 3061 tcp_do_unbind(conn_t *connp)
3061 3062 {
3062 3063 tcp_t *tcp = connp->conn_tcp;
3063 3064 int32_t oldstate;
3064 3065
3065 3066 switch (tcp->tcp_state) {
3066 3067 case TCPS_BOUND:
3067 3068 case TCPS_LISTEN:
3068 3069 break;
3069 3070 default:
3070 3071 return (-TOUTSTATE);
3071 3072 }
3072 3073
3073 3074 /*
3074 3075 * Need to clean up all the eagers since after the unbind, segments
3075 3076 * will no longer be delivered to this listener stream.
3076 3077 */
3077 3078 mutex_enter(&tcp->tcp_eager_lock);
3078 3079 if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3079 3080 tcp_eager_cleanup(tcp, 0);
3080 3081 }
3081 3082 mutex_exit(&tcp->tcp_eager_lock);
3082 3083
3083 3084 /* Clean up the listener connection counter if necessary. */
3084 3085 if (tcp->tcp_listen_cnt != NULL)
3085 3086 TCP_DECR_LISTEN_CNT(tcp);
3086 3087 connp->conn_laddr_v6 = ipv6_all_zeros;
3087 3088 connp->conn_saddr_v6 = ipv6_all_zeros;
3088 3089 tcp_bind_hash_remove(tcp);
3089 3090 oldstate = tcp->tcp_state;
3090 3091 tcp->tcp_state = TCPS_IDLE;
3091 3092 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
3092 3093 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
3093 3094 int32_t, oldstate);
3094 3095
3095 3096 ip_unbind(connp);
3096 3097 bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3097 3098
3098 3099 return (0);
3099 3100 }
3100 3101
3101 3102 /*
3102 3103 * Collect protocol properties to send to the upper handle.
3103 3104 */
3104 3105 void
3105 3106 tcp_get_proto_props(tcp_t *tcp, struct sock_proto_props *sopp)
3106 3107 {
3107 3108 conn_t *connp = tcp->tcp_connp;
3108 3109
3109 3110 sopp->sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3110 3111 sopp->sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3111 3112
3112 3113 sopp->sopp_rxhiwat = tcp->tcp_fused ?
3113 3114 tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3114 3115 connp->conn_rcvbuf;
3115 3116 /*
3116 3117 * Determine what write offset value to use depending on SACK and
3117 3118 * whether the endpoint is fused or not.
3118 3119 */
3119 3120 if (tcp->tcp_fused) {
3120 3121 ASSERT(tcp->tcp_loopback);
3121 3122 ASSERT(tcp->tcp_loopback_peer != NULL);
3122 3123 /*
3123 3124 * For fused tcp loopback, set the stream head's write
3124 3125 * offset value to zero since we won't be needing any room
3125 3126 * for TCP/IP headers. This would also improve performance
3126 3127 * since it would reduce the amount of work done by kmem.
3127 3128 * Non-fused tcp loopback case is handled separately below.
3128 3129 */
3129 3130 sopp->sopp_wroff = 0;
3130 3131 /*
3131 3132 * Update the peer's transmit parameters according to
3132 3133 * our recently calculated high water mark value.
3133 3134 */
3134 3135 (void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3135 3136 } else if (tcp->tcp_snd_sack_ok) {
3136 3137 sopp->sopp_wroff = connp->conn_ht_iphc_allocated +
3137 3138 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3138 3139 } else {
3139 3140 sopp->sopp_wroff = connp->conn_ht_iphc_len +
3140 3141 (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3141 3142 }
3142 3143
3143 3144 if (tcp->tcp_loopback) {
3144 3145 sopp->sopp_flags |= SOCKOPT_LOOPBACK;
3145 3146 sopp->sopp_loopback = B_TRUE;
3146 3147 }
3147 3148 }
3148 3149
3149 3150 /*
3150 3151 * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3151 3152 */
3152 3153 boolean_t
3153 3154 tcp_zcopy_check(tcp_t *tcp)
3154 3155 {
3155 3156 conn_t *connp = tcp->tcp_connp;
3156 3157 ip_xmit_attr_t *ixa = connp->conn_ixa;
3157 3158 boolean_t zc_enabled = B_FALSE;
3158 3159 tcp_stack_t *tcps = tcp->tcp_tcps;
3159 3160
3160 3161 if (do_tcpzcopy == 2)
3161 3162 zc_enabled = B_TRUE;
3162 3163 else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3163 3164 zc_enabled = B_TRUE;
3164 3165
3165 3166 tcp->tcp_snd_zcopy_on = zc_enabled;
3166 3167 if (!TCP_IS_DETACHED(tcp)) {
3167 3168 if (zc_enabled) {
3168 3169 ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3169 3170 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3170 3171 ZCVMSAFE);
3171 3172 TCP_STAT(tcps, tcp_zcopy_on);
3172 3173 } else {
3173 3174 ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3174 3175 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3175 3176 ZCVMUNSAFE);
3176 3177 TCP_STAT(tcps, tcp_zcopy_off);
3177 3178 }
3178 3179 }
3179 3180 return (zc_enabled);
3180 3181 }
3181 3182
3182 3183 /*
3183 3184 * Backoff from a zero-copy message by copying data to a new allocated
3184 3185 * message and freeing the original desballoca'ed segmapped message.
3185 3186 *
3186 3187 * This function is called by following two callers:
3187 3188 * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3188 3189 * the origial desballoca'ed message and notify sockfs. This is in re-
3189 3190 * transmit state.
3190 3191 * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3191 3192 * to be copied to new message.
3192 3193 */
3193 3194 mblk_t *
3194 3195 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3195 3196 {
3196 3197 mblk_t *nbp;
3197 3198 mblk_t *head = NULL;
3198 3199 mblk_t *tail = NULL;
3199 3200 tcp_stack_t *tcps = tcp->tcp_tcps;
3200 3201
3201 3202 ASSERT(bp != NULL);
3202 3203 while (bp != NULL) {
3203 3204 if (IS_VMLOANED_MBLK(bp)) {
3204 3205 TCP_STAT(tcps, tcp_zcopy_backoff);
3205 3206 if ((nbp = copyb(bp)) == NULL) {
3206 3207 tcp->tcp_xmit_zc_clean = B_FALSE;
3207 3208 if (tail != NULL)
3208 3209 tail->b_cont = bp;
3209 3210 return ((head == NULL) ? bp : head);
3210 3211 }
3211 3212
3212 3213 if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3213 3214 if (fix_xmitlist)
3214 3215 tcp_zcopy_notify(tcp);
3215 3216 else
3216 3217 nbp->b_datap->db_struioflag |=
3217 3218 STRUIO_ZCNOTIFY;
3218 3219 }
3219 3220 nbp->b_cont = bp->b_cont;
3220 3221
3221 3222 /*
3222 3223 * Copy saved information and adjust tcp_xmit_tail
3223 3224 * if needed.
3224 3225 */
3225 3226 if (fix_xmitlist) {
3226 3227 nbp->b_prev = bp->b_prev;
3227 3228 nbp->b_next = bp->b_next;
3228 3229
3229 3230 if (tcp->tcp_xmit_tail == bp)
3230 3231 tcp->tcp_xmit_tail = nbp;
3231 3232 }
3232 3233
3233 3234 /* Free the original message. */
3234 3235 bp->b_prev = NULL;
3235 3236 bp->b_next = NULL;
3236 3237 freeb(bp);
3237 3238
3238 3239 bp = nbp;
3239 3240 }
3240 3241
3241 3242 if (head == NULL) {
3242 3243 head = bp;
3243 3244 }
3244 3245 if (tail == NULL) {
3245 3246 tail = bp;
3246 3247 } else {
3247 3248 tail->b_cont = bp;
3248 3249 tail = bp;
3249 3250 }
3250 3251
3251 3252 /* Move forward. */
3252 3253 bp = bp->b_cont;
3253 3254 }
3254 3255
3255 3256 if (fix_xmitlist) {
3256 3257 tcp->tcp_xmit_last = tail;
3257 3258 tcp->tcp_xmit_zc_clean = B_TRUE;
3258 3259 }
3259 3260
3260 3261 return (head);
3261 3262 }
3262 3263
3263 3264 void
3264 3265 tcp_zcopy_notify(tcp_t *tcp)
3265 3266 {
3266 3267 struct stdata *stp;
3267 3268 conn_t *connp;
3268 3269
3269 3270 if (tcp->tcp_detached)
3270 3271 return;
3271 3272 connp = tcp->tcp_connp;
3272 3273 if (IPCL_IS_NONSTR(connp)) {
3273 3274 (*connp->conn_upcalls->su_zcopy_notify)
3274 3275 (connp->conn_upper_handle);
3275 3276 return;
3276 3277 }
3277 3278 stp = STREAM(connp->conn_rq);
3278 3279 mutex_enter(&stp->sd_lock);
3279 3280 stp->sd_flag |= STZCNOTIFY;
3280 3281 cv_broadcast(&stp->sd_zcopy_wait);
3281 3282 mutex_exit(&stp->sd_lock);
3282 3283 }
3283 3284
3284 3285 /*
3285 3286 * Update the TCP connection according to change of LSO capability.
3286 3287 */
3287 3288 static void
3288 3289 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3289 3290 {
3290 3291 /*
3291 3292 * We check against IPv4 header length to preserve the old behavior
3292 3293 * of only enabling LSO when there are no IP options.
3293 3294 * But this restriction might not be necessary at all. Before removing
3294 3295 * it, need to verify how LSO is handled for source routing case, with
3295 3296 * which IP does software checksum.
3296 3297 *
3297 3298 * For IPv6, whenever any extension header is needed, LSO is supressed.
3298 3299 */
3299 3300 if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3300 3301 IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3301 3302 return;
3302 3303
3303 3304 /*
3304 3305 * Either the LSO capability newly became usable, or it has changed.
3305 3306 */
3306 3307 if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3307 3308 ill_lso_capab_t *lsoc = &ixa->ixa_lso_capab;
3308 3309
3309 3310 ASSERT(lsoc->ill_lso_max > 0);
3310 3311 tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lsoc->ill_lso_max);
3311 3312
3312 3313 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3313 3314 boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3314 3315
3315 3316 /*
3316 3317 * If LSO to be enabled, notify the STREAM header with larger
3317 3318 * data block.
3318 3319 */
3319 3320 if (!tcp->tcp_lso)
3320 3321 tcp->tcp_maxpsz_multiplier = 0;
3321 3322
3322 3323 tcp->tcp_lso = B_TRUE;
3323 3324 TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3324 3325 } else { /* LSO capability is not usable any more. */
3325 3326 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3326 3327 boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
3327 3328
3328 3329 /*
3329 3330 * If LSO to be disabled, notify the STREAM header with smaller
3330 3331 * data block. And need to restore fragsize to PMTU.
3331 3332 */
3332 3333 if (tcp->tcp_lso) {
3333 3334 tcp->tcp_maxpsz_multiplier =
3334 3335 tcp->tcp_tcps->tcps_maxpsz_multiplier;
3335 3336 ixa->ixa_fragsize = ixa->ixa_pmtu;
3336 3337 tcp->tcp_lso = B_FALSE;
3337 3338 TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
3338 3339 }
3339 3340 }
3340 3341
3341 3342 (void) tcp_maxpsz_set(tcp, B_TRUE);
3342 3343 }
3343 3344
3344 3345 /*
3345 3346 * Update the TCP connection according to change of ZEROCOPY capability.
3346 3347 */
3347 3348 static void
3348 3349 tcp_update_zcopy(tcp_t *tcp)
3349 3350 {
3350 3351 conn_t *connp = tcp->tcp_connp;
3351 3352 tcp_stack_t *tcps = tcp->tcp_tcps;
3352 3353
3353 3354 if (tcp->tcp_snd_zcopy_on) {
3354 3355 tcp->tcp_snd_zcopy_on = B_FALSE;
3355 3356 if (!TCP_IS_DETACHED(tcp)) {
3356 3357 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3357 3358 ZCVMUNSAFE);
3358 3359 TCP_STAT(tcps, tcp_zcopy_off);
3359 3360 }
3360 3361 } else {
3361 3362 tcp->tcp_snd_zcopy_on = B_TRUE;
3362 3363 if (!TCP_IS_DETACHED(tcp)) {
3363 3364 (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3364 3365 ZCVMSAFE);
3365 3366 TCP_STAT(tcps, tcp_zcopy_on);
3366 3367 }
3367 3368 }
3368 3369 }
3369 3370
3370 3371 /*
3371 3372 * Notify function registered with ip_xmit_attr_t. It's called in the squeue
3372 3373 * so it's safe to update the TCP connection.
3373 3374 */
3374 3375 /* ARGSUSED1 */
3375 3376 static void
3376 3377 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
3377 3378 ixa_notify_arg_t narg)
3378 3379 {
3379 3380 tcp_t *tcp = (tcp_t *)arg;
3380 3381 conn_t *connp = tcp->tcp_connp;
3381 3382
3382 3383 switch (ntype) {
3383 3384 case IXAN_LSO:
3384 3385 tcp_update_lso(tcp, connp->conn_ixa);
3385 3386 break;
3386 3387 case IXAN_PMTU:
3387 3388 tcp_update_pmtu(tcp, B_FALSE);
3388 3389 break;
3389 3390 case IXAN_ZCOPY:
3390 3391 tcp_update_zcopy(tcp);
3391 3392 break;
3392 3393 default:
3393 3394 break;
3394 3395 }
3395 3396 }
3396 3397
3397 3398 /*
3398 3399 * The TCP write service routine should never be called...
3399 3400 */
3400 3401 /* ARGSUSED */
3401 3402 static void
3402 3403 tcp_wsrv(queue_t *q)
3403 3404 {
3404 3405 tcp_stack_t *tcps = Q_TO_TCP(q)->tcp_tcps;
3405 3406
3406 3407 TCP_STAT(tcps, tcp_wsrv_called);
3407 3408 }
3408 3409
3409 3410 /*
3410 3411 * Hash list lookup routine for tcp_t structures.
3411 3412 * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
3412 3413 */
3413 3414 tcp_t *
3414 3415 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
3415 3416 {
3416 3417 tf_t *tf;
3417 3418 tcp_t *tcp;
3418 3419
3419 3420 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3420 3421 mutex_enter(&tf->tf_lock);
3421 3422 for (tcp = tf->tf_tcp; tcp != NULL;
3422 3423 tcp = tcp->tcp_acceptor_hash) {
3423 3424 if (tcp->tcp_acceptor_id == id) {
3424 3425 CONN_INC_REF(tcp->tcp_connp);
3425 3426 mutex_exit(&tf->tf_lock);
3426 3427 return (tcp);
3427 3428 }
3428 3429 }
3429 3430 mutex_exit(&tf->tf_lock);
3430 3431 return (NULL);
3431 3432 }
3432 3433
3433 3434 /*
3434 3435 * Hash list insertion routine for tcp_t structures.
3435 3436 */
3436 3437 void
3437 3438 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
3438 3439 {
3439 3440 tf_t *tf;
3440 3441 tcp_t **tcpp;
3441 3442 tcp_t *tcpnext;
3442 3443 tcp_stack_t *tcps = tcp->tcp_tcps;
3443 3444
3444 3445 tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3445 3446
3446 3447 if (tcp->tcp_ptpahn != NULL)
3447 3448 tcp_acceptor_hash_remove(tcp);
3448 3449 tcpp = &tf->tf_tcp;
3449 3450 mutex_enter(&tf->tf_lock);
3450 3451 tcpnext = tcpp[0];
3451 3452 if (tcpnext)
3452 3453 tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
3453 3454 tcp->tcp_acceptor_hash = tcpnext;
3454 3455 tcp->tcp_ptpahn = tcpp;
3455 3456 tcpp[0] = tcp;
3456 3457 tcp->tcp_acceptor_lockp = &tf->tf_lock; /* For tcp_*_hash_remove */
3457 3458 mutex_exit(&tf->tf_lock);
3458 3459 }
3459 3460
3460 3461 /*
3461 3462 * Hash list removal routine for tcp_t structures.
3462 3463 */
3463 3464 void
3464 3465 tcp_acceptor_hash_remove(tcp_t *tcp)
3465 3466 {
3466 3467 tcp_t *tcpnext;
3467 3468 kmutex_t *lockp;
3468 3469
3469 3470 /*
3470 3471 * Extract the lock pointer in case there are concurrent
3471 3472 * hash_remove's for this instance.
3472 3473 */
3473 3474 lockp = tcp->tcp_acceptor_lockp;
3474 3475
3475 3476 if (tcp->tcp_ptpahn == NULL)
3476 3477 return;
3477 3478
3478 3479 ASSERT(lockp != NULL);
3479 3480 mutex_enter(lockp);
3480 3481 if (tcp->tcp_ptpahn) {
3481 3482 tcpnext = tcp->tcp_acceptor_hash;
3482 3483 if (tcpnext) {
3483 3484 tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
3484 3485 tcp->tcp_acceptor_hash = NULL;
3485 3486 }
3486 3487 *tcp->tcp_ptpahn = tcpnext;
3487 3488 tcp->tcp_ptpahn = NULL;
3488 3489 }
3489 3490 mutex_exit(lockp);
3490 3491 tcp->tcp_acceptor_lockp = NULL;
3491 3492 }
3492 3493
3493 3494 /*
3494 3495 * Type three generator adapted from the random() function in 4.4 BSD:
3495 3496 */
3496 3497
3497 3498 /*
3498 3499 * Copyright (c) 1983, 1993
3499 3500 * The Regents of the University of California. All rights reserved.
3500 3501 *
3501 3502 * Redistribution and use in source and binary forms, with or without
3502 3503 * modification, are permitted provided that the following conditions
3503 3504 * are met:
3504 3505 * 1. Redistributions of source code must retain the above copyright
3505 3506 * notice, this list of conditions and the following disclaimer.
3506 3507 * 2. Redistributions in binary form must reproduce the above copyright
3507 3508 * notice, this list of conditions and the following disclaimer in the
3508 3509 * documentation and/or other materials provided with the distribution.
3509 3510 * 3. All advertising materials mentioning features or use of this software
3510 3511 * must display the following acknowledgement:
3511 3512 * This product includes software developed by the University of
3512 3513 * California, Berkeley and its contributors.
3513 3514 * 4. Neither the name of the University nor the names of its contributors
3514 3515 * may be used to endorse or promote products derived from this software
3515 3516 * without specific prior written permission.
3516 3517 *
3517 3518 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
3518 3519 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
3519 3520 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
3520 3521 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
3521 3522 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
3522 3523 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
3523 3524 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
3524 3525 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
3525 3526 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
3526 3527 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
3527 3528 * SUCH DAMAGE.
3528 3529 */
3529 3530
3530 3531 /* Type 3 -- x**31 + x**3 + 1 */
3531 3532 #define DEG_3 31
3532 3533 #define SEP_3 3
3533 3534
3534 3535
3535 3536 /* Protected by tcp_random_lock */
3536 3537 static int tcp_randtbl[DEG_3 + 1];
3537 3538
3538 3539 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
3539 3540 static int *tcp_random_rptr = &tcp_randtbl[1];
3540 3541
3541 3542 static int *tcp_random_state = &tcp_randtbl[1];
3542 3543 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
3543 3544
3544 3545 kmutex_t tcp_random_lock;
3545 3546
3546 3547 void
3547 3548 tcp_random_init(void)
3548 3549 {
3549 3550 int i;
3550 3551 hrtime_t hrt;
3551 3552 time_t wallclock;
3552 3553 uint64_t result;
3553 3554
3554 3555 /*
3555 3556 * Use high-res timer and current time for seed. Gethrtime() returns
3556 3557 * a longlong, which may contain resolution down to nanoseconds.
3557 3558 * The current time will either be a 32-bit or a 64-bit quantity.
3558 3559 * XOR the two together in a 64-bit result variable.
3559 3560 * Convert the result to a 32-bit value by multiplying the high-order
3560 3561 * 32-bits by the low-order 32-bits.
3561 3562 */
3562 3563
3563 3564 hrt = gethrtime();
3564 3565 (void) drv_getparm(TIME, &wallclock);
3565 3566 result = (uint64_t)wallclock ^ (uint64_t)hrt;
3566 3567 mutex_enter(&tcp_random_lock);
3567 3568 tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
3568 3569 (result & 0xffffffff);
3569 3570
3570 3571 for (i = 1; i < DEG_3; i++)
3571 3572 tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
3572 3573 + 12345;
3573 3574 tcp_random_fptr = &tcp_random_state[SEP_3];
3574 3575 tcp_random_rptr = &tcp_random_state[0];
3575 3576 mutex_exit(&tcp_random_lock);
3576 3577 for (i = 0; i < 10 * DEG_3; i++)
3577 3578 (void) tcp_random();
3578 3579 }
3579 3580
3580 3581 /*
3581 3582 * tcp_random: Return a random number in the range [1 - (128K + 1)].
3582 3583 * This range is selected to be approximately centered on TCP_ISS / 2,
3583 3584 * and easy to compute. We get this value by generating a 32-bit random
3584 3585 * number, selecting out the high-order 17 bits, and then adding one so
3585 3586 * that we never return zero.
3586 3587 */
3587 3588 int
3588 3589 tcp_random(void)
3589 3590 {
3590 3591 int i;
3591 3592
3592 3593 mutex_enter(&tcp_random_lock);
3593 3594 *tcp_random_fptr += *tcp_random_rptr;
3594 3595
3595 3596 /*
3596 3597 * The high-order bits are more random than the low-order bits,
3597 3598 * so we select out the high-order 17 bits and add one so that
3598 3599 * we never return zero.
3599 3600 */
3600 3601 i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
3601 3602 if (++tcp_random_fptr >= tcp_random_end_ptr) {
3602 3603 tcp_random_fptr = tcp_random_state;
3603 3604 ++tcp_random_rptr;
3604 3605 } else if (++tcp_random_rptr >= tcp_random_end_ptr)
3605 3606 tcp_random_rptr = tcp_random_state;
3606 3607
3607 3608 mutex_exit(&tcp_random_lock);
3608 3609 return (i);
3609 3610 }
3610 3611
3611 3612 /*
3612 3613 * Split this function out so that if the secret changes, I'm okay.
3613 3614 *
3614 3615 * Initialize the tcp_iss_cookie and tcp_iss_key.
3615 3616 */
3616 3617
3617 3618 #define PASSWD_SIZE 16 /* MUST be multiple of 4 */
3618 3619
3619 3620 void
3620 3621 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
3621 3622 {
3622 3623 struct {
3623 3624 int32_t current_time;
3624 3625 uint32_t randnum;
3625 3626 uint16_t pad;
3626 3627 uint8_t ether[6];
3627 3628 uint8_t passwd[PASSWD_SIZE];
3628 3629 } tcp_iss_cookie;
3629 3630 time_t t;
3630 3631
3631 3632 /*
3632 3633 * Start with the current absolute time.
3633 3634 */
3634 3635 (void) drv_getparm(TIME, &t);
3635 3636 tcp_iss_cookie.current_time = t;
3636 3637
3637 3638 /*
3638 3639 * XXX - Need a more random number per RFC 1750, not this crap.
3639 3640 * OTOH, if what follows is pretty random, then I'm in better shape.
3640 3641 */
3641 3642 tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
3642 3643 tcp_iss_cookie.pad = 0x365c; /* Picked from HMAC pad values. */
3643 3644
3644 3645 /*
3645 3646 * The cpu_type_info is pretty non-random. Ugggh. It does serve
3646 3647 * as a good template.
3647 3648 */
3648 3649 bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
3649 3650 min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
3650 3651
3651 3652 /*
3652 3653 * The pass-phrase. Normally this is supplied by user-called NDD.
3653 3654 */
3654 3655 bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
3655 3656
3656 3657 /*
3657 3658 * See 4010593 if this section becomes a problem again,
3658 3659 * but the local ethernet address is useful here.
3659 3660 */
3660 3661 (void) localetheraddr(NULL,
3661 3662 (struct ether_addr *)&tcp_iss_cookie.ether);
3662 3663
3663 3664 /*
3664 3665 * Hash 'em all together. The MD5Final is called per-connection.
3665 3666 */
3666 3667 mutex_enter(&tcps->tcps_iss_key_lock);
3667 3668 MD5Init(&tcps->tcps_iss_key);
3668 3669 MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
3669 3670 sizeof (tcp_iss_cookie));
3670 3671 mutex_exit(&tcps->tcps_iss_key_lock);
3671 3672 }
3672 3673
3673 3674 /*
3674 3675 * Called by IP when IP is loaded into the kernel
3675 3676 */
3676 3677 void
3677 3678 tcp_ddi_g_init(void)
3678 3679 {
3679 3680 tcp_timercache = kmem_cache_create("tcp_timercache",
3680 3681 sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
3681 3682 NULL, NULL, NULL, NULL, NULL, 0);
3682 3683
3683 3684 tcp_notsack_blk_cache = kmem_cache_create("tcp_notsack_blk_cache",
3684 3685 sizeof (notsack_blk_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3685 3686
3686 3687 mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
3687 3688
3688 3689 /* Initialize the random number generator */
3689 3690 tcp_random_init();
3690 3691
3691 3692 /* A single callback independently of how many netstacks we have */
3692 3693 ip_squeue_init(tcp_squeue_add);
3693 3694
3694 3695 tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
3695 3696
3696 3697 tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
3697 3698
3698 3699 /*
3699 3700 * We want to be informed each time a stack is created or
3700 3701 * destroyed in the kernel, so we can maintain the
3701 3702 * set of tcp_stack_t's.
3702 3703 */
3703 3704 netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
3704 3705 }
3705 3706
3706 3707
3707 3708 #define INET_NAME "ip"
3708 3709
3709 3710 /*
3710 3711 * Initialize the TCP stack instance.
3711 3712 */
3712 3713 static void *
3713 3714 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
3714 3715 {
3715 3716 tcp_stack_t *tcps;
3716 3717 int i;
3717 3718 int error = 0;
3718 3719 major_t major;
3719 3720 size_t arrsz;
3720 3721
3721 3722 tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
3722 3723 tcps->tcps_netstack = ns;
3723 3724
3724 3725 /* Initialize locks */
3725 3726 mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
3726 3727 mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
3727 3728
3728 3729 tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
3729 3730 tcps->tcps_g_epriv_ports[0] = ULP_DEF_EPRIV_PORT1;
3730 3731 tcps->tcps_g_epriv_ports[1] = ULP_DEF_EPRIV_PORT2;
3731 3732 tcps->tcps_min_anonpriv_port = 512;
3732 3733
3733 3734 tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
3734 3735 TCP_BIND_FANOUT_SIZE, KM_SLEEP);
3735 3736 tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
3736 3737 TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
3737 3738
3738 3739 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3739 3740 mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
3740 3741 MUTEX_DEFAULT, NULL);
3741 3742 }
3742 3743
3743 3744 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3744 3745 mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
3745 3746 MUTEX_DEFAULT, NULL);
3746 3747 }
3747 3748
3748 3749 /* TCP's IPsec code calls the packet dropper. */
3749 3750 ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
3750 3751
3751 3752 arrsz = tcp_propinfo_count * sizeof (mod_prop_info_t);
3752 3753 tcps->tcps_propinfo_tbl = (mod_prop_info_t *)kmem_alloc(arrsz,
3753 3754 KM_SLEEP);
3754 3755 bcopy(tcp_propinfo_tbl, tcps->tcps_propinfo_tbl, arrsz);
3755 3756
3756 3757 /*
3757 3758 * Note: To really walk the device tree you need the devinfo
3758 3759 * pointer to your device which is only available after probe/attach.
3759 3760 * The following is safe only because it uses ddi_root_node()
3760 3761 */
3761 3762 tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
3762 3763 tcp_opt_obj.odb_opt_arr_cnt);
3763 3764
3764 3765 /*
3765 3766 * Initialize RFC 1948 secret values. This will probably be reset once
3766 3767 * by the boot scripts.
3767 3768 *
3768 3769 * Use NULL name, as the name is caught by the new lockstats.
3769 3770 *
3770 3771 * Initialize with some random, non-guessable string, like the global
3771 3772 * T_INFO_ACK.
3772 3773 */
3773 3774
3774 3775 tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
3775 3776 sizeof (tcp_g_t_info_ack), tcps);
3776 3777
3777 3778 tcps->tcps_kstat = tcp_kstat2_init(stackid);
3778 3779 tcps->tcps_mibkp = tcp_kstat_init(stackid);
3779 3780
3780 3781 major = mod_name_to_major(INET_NAME);
3781 3782 error = ldi_ident_from_major(major, &tcps->tcps_ldi_ident);
3782 3783 ASSERT(error == 0);
3783 3784 tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
3784 3785 ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
3785 3786 cv_init(&tcps->tcps_ixa_cleanup_cv, NULL, CV_DEFAULT, NULL);
3786 3787 mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
3787 3788
3788 3789 mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
3789 3790 tcps->tcps_reclaim = B_FALSE;
3790 3791 tcps->tcps_reclaim_tid = 0;
3791 3792 tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
3792 3793
3793 3794 /*
3794 3795 * ncpus is the current number of CPUs, which can be bigger than
3795 3796 * boot_ncpus. But we don't want to use ncpus to allocate all the
3796 3797 * tcp_stats_cpu_t at system boot up time since it will be 1. While
3797 3798 * we handle adding CPU in tcp_cpu_update(), it will be slow if
3798 3799 * there are many CPUs as we will be adding them 1 by 1.
3799 3800 *
3800 3801 * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
3801 3802 * are not freed until the stack is going away. So there is no need
3802 3803 * to grab a lock to access the per CPU tcps_sc[x] pointer.
3803 3804 */
3804 3805 mutex_enter(&cpu_lock);
3805 3806 tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
3806 3807 mutex_exit(&cpu_lock);
3807 3808 tcps->tcps_sc = kmem_zalloc(max_ncpus * sizeof (tcp_stats_cpu_t *),
3808 3809 KM_SLEEP);
3809 3810 for (i = 0; i < tcps->tcps_sc_cnt; i++) {
3810 3811 tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
3811 3812 KM_SLEEP);
3812 3813 }
3813 3814
3814 3815 mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
3815 3816 list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
3816 3817 offsetof(tcp_listener_t, tl_link));
3817 3818
3818 3819 return (tcps);
3819 3820 }
3820 3821
3821 3822 /*
3822 3823 * Called when the IP module is about to be unloaded.
3823 3824 */
3824 3825 void
3825 3826 tcp_ddi_g_destroy(void)
3826 3827 {
3827 3828 tcp_g_kstat_fini(tcp_g_kstat);
3828 3829 tcp_g_kstat = NULL;
3829 3830 bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
3830 3831
3831 3832 mutex_destroy(&tcp_random_lock);
3832 3833
3833 3834 kmem_cache_destroy(tcp_timercache);
3834 3835 kmem_cache_destroy(tcp_notsack_blk_cache);
3835 3836
3836 3837 netstack_unregister(NS_TCP);
3837 3838 }
3838 3839
3839 3840 /*
3840 3841 * Free the TCP stack instance.
3841 3842 */
3842 3843 static void
3843 3844 tcp_stack_fini(netstackid_t stackid, void *arg)
3844 3845 {
3845 3846 tcp_stack_t *tcps = (tcp_stack_t *)arg;
3846 3847 int i;
3847 3848
3848 3849 freeb(tcps->tcps_ixa_cleanup_mp);
3849 3850 tcps->tcps_ixa_cleanup_mp = NULL;
3850 3851 cv_destroy(&tcps->tcps_ixa_cleanup_cv);
3851 3852 mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
3852 3853
3853 3854 /*
3854 3855 * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
3855 3856 * the timer.
3856 3857 */
3857 3858 mutex_enter(&tcps->tcps_reclaim_lock);
3858 3859 tcps->tcps_reclaim = B_FALSE;
3859 3860 mutex_exit(&tcps->tcps_reclaim_lock);
3860 3861 if (tcps->tcps_reclaim_tid != 0)
3861 3862 (void) untimeout(tcps->tcps_reclaim_tid);
3862 3863 mutex_destroy(&tcps->tcps_reclaim_lock);
3863 3864
3864 3865 tcp_listener_conf_cleanup(tcps);
3865 3866
3866 3867 for (i = 0; i < tcps->tcps_sc_cnt; i++)
3867 3868 kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
3868 3869 kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
3869 3870
3870 3871 kmem_free(tcps->tcps_propinfo_tbl,
3871 3872 tcp_propinfo_count * sizeof (mod_prop_info_t));
3872 3873 tcps->tcps_propinfo_tbl = NULL;
3873 3874
3874 3875 for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3875 3876 ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
3876 3877 mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
3877 3878 }
3878 3879
3879 3880 for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3880 3881 ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
3881 3882 mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
3882 3883 }
3883 3884
3884 3885 kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
3885 3886 tcps->tcps_bind_fanout = NULL;
3886 3887
3887 3888 kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
3888 3889 TCP_ACCEPTOR_FANOUT_SIZE);
3889 3890 tcps->tcps_acceptor_fanout = NULL;
3890 3891
3891 3892 mutex_destroy(&tcps->tcps_iss_key_lock);
3892 3893 mutex_destroy(&tcps->tcps_epriv_port_lock);
3893 3894
3894 3895 ip_drop_unregister(&tcps->tcps_dropper);
3895 3896
3896 3897 tcp_kstat2_fini(stackid, tcps->tcps_kstat);
3897 3898 tcps->tcps_kstat = NULL;
3898 3899
3899 3900 tcp_kstat_fini(stackid, tcps->tcps_mibkp);
3900 3901 tcps->tcps_mibkp = NULL;
3901 3902
3902 3903 ldi_ident_release(tcps->tcps_ldi_ident);
3903 3904 kmem_free(tcps, sizeof (*tcps));
3904 3905 }
3905 3906
3906 3907 /*
3907 3908 * Generate ISS, taking into account NDD changes may happen halfway through.
3908 3909 * (If the iss is not zero, set it.)
3909 3910 */
3910 3911
3911 3912 static void
3912 3913 tcp_iss_init(tcp_t *tcp)
3913 3914 {
3914 3915 MD5_CTX context;
3915 3916 struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
3916 3917 uint32_t answer[4];
3917 3918 tcp_stack_t *tcps = tcp->tcp_tcps;
3918 3919 conn_t *connp = tcp->tcp_connp;
3919 3920
3920 3921 tcps->tcps_iss_incr_extra += (ISS_INCR >> 1);
3921 3922 tcp->tcp_iss = tcps->tcps_iss_incr_extra;
3922 3923 switch (tcps->tcps_strong_iss) {
3923 3924 case 2:
3924 3925 mutex_enter(&tcps->tcps_iss_key_lock);
3925 3926 context = tcps->tcps_iss_key;
3926 3927 mutex_exit(&tcps->tcps_iss_key_lock);
3927 3928 arg.ports = connp->conn_ports;
3928 3929 arg.src = connp->conn_laddr_v6;
3929 3930 arg.dst = connp->conn_faddr_v6;
3930 3931 MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
3931 3932 MD5Final((uchar_t *)answer, &context);
3932 3933 tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
3933 3934 /*
3934 3935 * Now that we've hashed into a unique per-connection sequence
3935 3936 * space, add a random increment per strong_iss == 1. So I
3936 3937 * guess we'll have to...
3937 3938 */
3938 3939 /* FALLTHRU */
3939 3940 case 1:
3940 3941 tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
3941 3942 break;
3942 3943 default:
3943 3944 tcp->tcp_iss += (uint32_t)gethrestime_sec() * ISS_INCR;
3944 3945 break;
3945 3946 }
3946 3947 tcp->tcp_valid_bits = TCP_ISS_VALID;
3947 3948 tcp->tcp_fss = tcp->tcp_iss - 1;
3948 3949 tcp->tcp_suna = tcp->tcp_iss;
3949 3950 tcp->tcp_snxt = tcp->tcp_iss + 1;
3950 3951 tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
3951 3952 tcp->tcp_csuna = tcp->tcp_snxt;
3952 3953 }
3953 3954
3954 3955 /*
3955 3956 * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
3956 3957 * on the specified backing STREAMS q. Note, the caller may make the
3957 3958 * decision to call based on the tcp_t.tcp_flow_stopped value which
3958 3959 * when check outside the q's lock is only an advisory check ...
3959 3960 */
3960 3961 void
3961 3962 tcp_setqfull(tcp_t *tcp)
3962 3963 {
3963 3964 tcp_stack_t *tcps = tcp->tcp_tcps;
3964 3965 conn_t *connp = tcp->tcp_connp;
3965 3966
3966 3967 if (tcp->tcp_closed)
3967 3968 return;
3968 3969
3969 3970 conn_setqfull(connp, &tcp->tcp_flow_stopped);
3970 3971 if (tcp->tcp_flow_stopped)
3971 3972 TCP_STAT(tcps, tcp_flwctl_on);
3972 3973 }
3973 3974
3974 3975 void
3975 3976 tcp_clrqfull(tcp_t *tcp)
3976 3977 {
3977 3978 conn_t *connp = tcp->tcp_connp;
3978 3979
3979 3980 if (tcp->tcp_closed)
3980 3981 return;
3981 3982 conn_clrqfull(connp, &tcp->tcp_flow_stopped);
3982 3983 }
3983 3984
3984 3985 static int
3985 3986 tcp_squeue_switch(int val)
3986 3987 {
3987 3988 int rval = SQ_FILL;
3988 3989
3989 3990 switch (val) {
3990 3991 case 1:
3991 3992 rval = SQ_NODRAIN;
3992 3993 break;
3993 3994 case 2:
3994 3995 rval = SQ_PROCESS;
3995 3996 break;
3996 3997 default:
3997 3998 break;
3998 3999 }
3999 4000 return (rval);
4000 4001 }
4001 4002
4002 4003 /*
4003 4004 * This is called once for each squeue - globally for all stack
4004 4005 * instances.
4005 4006 */
4006 4007 static void
4007 4008 tcp_squeue_add(squeue_t *sqp)
4008 4009 {
4009 4010 tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4010 4011 sizeof (tcp_squeue_priv_t), KM_SLEEP);
4011 4012
4012 4013 *squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4013 4014 if (tcp_free_list_max_cnt == 0) {
4014 4015 int tcp_ncpus = ((boot_max_ncpus == -1) ?
4015 4016 max_ncpus : boot_max_ncpus);
4016 4017
4017 4018 /*
4018 4019 * Limit number of entries to 1% of availble memory / tcp_ncpus
4019 4020 */
4020 4021 tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4021 4022 (tcp_ncpus * sizeof (tcp_t) * 100);
4022 4023 }
4023 4024 tcp_time_wait->tcp_free_list_cnt = 0;
4024 4025 }
4025 4026 /*
4026 4027 * Return unix error is tli error is TSYSERR, otherwise return a negative
4027 4028 * tli error.
4028 4029 */
4029 4030 int
4030 4031 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4031 4032 boolean_t bind_to_req_port_only)
4032 4033 {
4033 4034 int error;
4034 4035 tcp_t *tcp = connp->conn_tcp;
4035 4036
4036 4037 if (tcp->tcp_state >= TCPS_BOUND) {
4037 4038 if (connp->conn_debug) {
4038 4039 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4039 4040 "tcp_bind: bad state, %d", tcp->tcp_state);
4040 4041 }
4041 4042 return (-TOUTSTATE);
4042 4043 }
4043 4044
4044 4045 error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4045 4046 if (error != 0)
4046 4047 return (error);
4047 4048
4048 4049 ASSERT(tcp->tcp_state == TCPS_BOUND);
4049 4050 tcp->tcp_conn_req_max = 0;
4050 4051 return (0);
4051 4052 }
4052 4053
4053 4054 /*
4054 4055 * If the return value from this function is positive, it's a UNIX error.
4055 4056 * Otherwise, if it's negative, then the absolute value is a TLI error.
4056 4057 * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4057 4058 */
4058 4059 int
4059 4060 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4060 4061 cred_t *cr, pid_t pid)
4061 4062 {
4062 4063 tcp_t *tcp = connp->conn_tcp;
4063 4064 sin_t *sin = (sin_t *)sa;
4064 4065 sin6_t *sin6 = (sin6_t *)sa;
4065 4066 ipaddr_t *dstaddrp;
4066 4067 in_port_t dstport;
4067 4068 uint_t srcid;
4068 4069 int error;
4069 4070 uint32_t mss;
4070 4071 mblk_t *syn_mp;
4071 4072 tcp_stack_t *tcps = tcp->tcp_tcps;
4072 4073 int32_t oldstate;
4073 4074 ip_xmit_attr_t *ixa = connp->conn_ixa;
4074 4075
4075 4076 oldstate = tcp->tcp_state;
4076 4077
4077 4078 switch (len) {
4078 4079 default:
4079 4080 /*
4080 4081 * Should never happen
4081 4082 */
4082 4083 return (EINVAL);
4083 4084
4084 4085 case sizeof (sin_t):
4085 4086 sin = (sin_t *)sa;
4086 4087 if (sin->sin_port == 0) {
4087 4088 return (-TBADADDR);
4088 4089 }
4089 4090 if (connp->conn_ipv6_v6only) {
4090 4091 return (EAFNOSUPPORT);
4091 4092 }
4092 4093 break;
4093 4094
4094 4095 case sizeof (sin6_t):
4095 4096 sin6 = (sin6_t *)sa;
4096 4097 if (sin6->sin6_port == 0) {
4097 4098 return (-TBADADDR);
4098 4099 }
4099 4100 break;
4100 4101 }
4101 4102 /*
4102 4103 * If we're connecting to an IPv4-mapped IPv6 address, we need to
4103 4104 * make sure that the conn_ipversion is IPV4_VERSION. We
4104 4105 * need to this before we call tcp_bindi() so that the port lookup
4105 4106 * code will look for ports in the correct port space (IPv4 and
4106 4107 * IPv6 have separate port spaces).
4107 4108 */
4108 4109 if (connp->conn_family == AF_INET6 &&
4109 4110 connp->conn_ipversion == IPV6_VERSION &&
4110 4111 IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4111 4112 if (connp->conn_ipv6_v6only)
4112 4113 return (EADDRNOTAVAIL);
4113 4114
4114 4115 connp->conn_ipversion = IPV4_VERSION;
4115 4116 }
4116 4117
4117 4118 switch (tcp->tcp_state) {
4118 4119 case TCPS_LISTEN:
4119 4120 /*
4120 4121 * Listening sockets are not allowed to issue connect().
4121 4122 */
4122 4123 if (IPCL_IS_NONSTR(connp))
4123 4124 return (EOPNOTSUPP);
4124 4125 /* FALLTHRU */
4125 4126 case TCPS_IDLE:
4126 4127 /*
4127 4128 * We support quick connect, refer to comments in
4128 4129 * tcp_connect_*()
4129 4130 */
4130 4131 /* FALLTHRU */
4131 4132 case TCPS_BOUND:
4132 4133 break;
4133 4134 default:
4134 4135 return (-TOUTSTATE);
4135 4136 }
4136 4137
4137 4138 /*
4138 4139 * We update our cred/cpid based on the caller of connect
4139 4140 */
4140 4141 if (connp->conn_cred != cr) {
4141 4142 crhold(cr);
4142 4143 crfree(connp->conn_cred);
4143 4144 connp->conn_cred = cr;
4144 4145 }
4145 4146 connp->conn_cpid = pid;
4146 4147
4147 4148 /* Cache things in the ixa without any refhold */
4148 4149 ASSERT(!(ixa->ixa_free_flags & IXA_FREE_CRED));
4149 4150 ixa->ixa_cred = cr;
4150 4151 ixa->ixa_cpid = pid;
4151 4152 if (is_system_labeled()) {
4152 4153 /* We need to restart with a label based on the cred */
4153 4154 ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4154 4155 }
4155 4156
4156 4157 if (connp->conn_family == AF_INET6) {
4157 4158 if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4158 4159 error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4159 4160 sin6->sin6_port, sin6->sin6_flowinfo,
4160 4161 sin6->__sin6_src_id, sin6->sin6_scope_id);
4161 4162 } else {
4162 4163 /*
4163 4164 * Destination adress is mapped IPv6 address.
4164 4165 * Source bound address should be unspecified or
4165 4166 * IPv6 mapped address as well.
4166 4167 */
4167 4168 if (!IN6_IS_ADDR_UNSPECIFIED(
4168 4169 &connp->conn_bound_addr_v6) &&
4169 4170 !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4170 4171 return (EADDRNOTAVAIL);
4171 4172 }
4172 4173 dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4173 4174 dstport = sin6->sin6_port;
4174 4175 srcid = sin6->__sin6_src_id;
4175 4176 error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4176 4177 srcid);
4177 4178 }
4178 4179 } else {
4179 4180 dstaddrp = &sin->sin_addr.s_addr;
4180 4181 dstport = sin->sin_port;
4181 4182 srcid = 0;
4182 4183 error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4183 4184 }
4184 4185
4185 4186 if (error != 0)
4186 4187 goto connect_failed;
4187 4188
4188 4189 CL_INET_CONNECT(connp, B_TRUE, error);
4189 4190 if (error != 0)
4190 4191 goto connect_failed;
4191 4192
4192 4193 /* connect succeeded */
4193 4194 TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4194 4195 tcp->tcp_active_open = 1;
4195 4196
4196 4197 /*
4197 4198 * tcp_set_destination() does not adjust for TCP/IP header length.
4198 4199 */
4199 4200 mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4200 4201
4201 4202 /*
4202 4203 * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4203 4204 * to the nearest MSS.
4204 4205 *
4205 4206 * We do the round up here because we need to get the interface MTU
4206 4207 * first before we can do the round up.
4207 4208 */
4208 4209 tcp->tcp_rwnd = connp->conn_rcvbuf;
4209 4210 tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4210 4211 tcps->tcps_recv_hiwat_minmss * mss);
4211 4212 connp->conn_rcvbuf = tcp->tcp_rwnd;
4212 4213 tcp_set_ws_value(tcp);
4213 4214 tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4214 4215 if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4215 4216 tcp->tcp_snd_ws_ok = B_TRUE;
4216 4217
4217 4218 /*
4218 4219 * Set tcp_snd_ts_ok to true
4219 4220 * so that tcp_xmit_mp will
4220 4221 * include the timestamp
4221 4222 * option in the SYN segment.
4222 4223 */
4223 4224 if (tcps->tcps_tstamp_always ||
4224 4225 (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4225 4226 tcp->tcp_snd_ts_ok = B_TRUE;
4226 4227 }
4227 4228
4228 4229 /*
4229 4230 * Note that tcp_snd_sack_ok can be set in tcp_set_destination() if
4230 4231 * the SACK metric is set. So here we just check the per stack SACK
4231 4232 * permitted param.
4232 4233 */
4233 4234 if (tcps->tcps_sack_permitted == 2) {
4234 4235 ASSERT(tcp->tcp_num_sack_blk == 0);
4235 4236 ASSERT(tcp->tcp_notsack_list == NULL);
4236 4237 tcp->tcp_snd_sack_ok = B_TRUE;
4237 4238 }
4238 4239
4239 4240 /*
4240 4241 * Should we use ECN? Note that the current
4241 4242 * default value (SunOS 5.9) of tcp_ecn_permitted
4242 4243 * is 1. The reason for doing this is that there
4243 4244 * are equipments out there that will drop ECN
4244 4245 * enabled IP packets. Setting it to 1 avoids
4245 4246 * compatibility problems.
4246 4247 */
4247 4248 if (tcps->tcps_ecn_permitted == 2)
4248 4249 tcp->tcp_ecn_ok = B_TRUE;
4249 4250
4250 4251 /* Trace change from BOUND -> SYN_SENT here */
4251 4252 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4252 4253 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4253 4254 int32_t, TCPS_BOUND);
4254 4255
4255 4256 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4256 4257 syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4257 4258 tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4258 4259 if (syn_mp != NULL) {
4259 4260 /*
4260 4261 * We must bump the generation before sending the syn
4261 4262 * to ensure that we use the right generation in case
4262 4263 * this thread issues a "connected" up call.
4263 4264 */
4264 4265 SOCK_CONNID_BUMP(tcp->tcp_connid);
4265 4266 /*
4266 4267 * DTrace sending the first SYN as a
4267 4268 * tcp:::connect-request event.
4268 4269 */
4269 4270 DTRACE_TCP5(connect__request, mblk_t *, NULL,
4270 4271 ip_xmit_attr_t *, connp->conn_ixa,
4271 4272 void_ip_t *, syn_mp->b_rptr, tcp_t *, tcp,
4272 4273 tcph_t *,
4273 4274 &syn_mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
4274 4275 tcp_send_data(tcp, syn_mp);
4275 4276 }
4276 4277
4277 4278 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4278 4279 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4279 4280 return (0);
4280 4281
4281 4282 connect_failed:
4282 4283 connp->conn_faddr_v6 = ipv6_all_zeros;
4283 4284 connp->conn_fport = 0;
4284 4285 tcp->tcp_state = oldstate;
4285 4286 if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4286 4287 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4287 4288 return (error);
4288 4289 }
4289 4290
4290 4291 int
4291 4292 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4292 4293 int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4293 4294 {
4294 4295 tcp_t *tcp = connp->conn_tcp;
4295 4296 int error = 0;
4296 4297 tcp_stack_t *tcps = tcp->tcp_tcps;
4297 4298 int32_t oldstate;
4298 4299
4299 4300 /* All Solaris components should pass a cred for this operation. */
4300 4301 ASSERT(cr != NULL);
4301 4302
4302 4303 if (tcp->tcp_state >= TCPS_BOUND) {
4303 4304 if ((tcp->tcp_state == TCPS_BOUND ||
4304 4305 tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
4305 4306 /*
4306 4307 * Handle listen() increasing backlog.
4307 4308 * This is more "liberal" then what the TPI spec
4308 4309 * requires but is needed to avoid a t_unbind
4309 4310 * when handling listen() since the port number
4310 4311 * might be "stolen" between the unbind and bind.
4311 4312 */
4312 4313 goto do_listen;
4313 4314 }
4314 4315 if (connp->conn_debug) {
4315 4316 (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4316 4317 "tcp_listen: bad state, %d", tcp->tcp_state);
4317 4318 }
4318 4319 return (-TOUTSTATE);
4319 4320 } else {
4320 4321 if (sa == NULL) {
4321 4322 sin6_t addr;
4322 4323 sin_t *sin;
4323 4324 sin6_t *sin6;
4324 4325
4325 4326 ASSERT(IPCL_IS_NONSTR(connp));
4326 4327 /* Do an implicit bind: Request for a generic port. */
4327 4328 if (connp->conn_family == AF_INET) {
4328 4329 len = sizeof (sin_t);
4329 4330 sin = (sin_t *)&addr;
4330 4331 *sin = sin_null;
4331 4332 sin->sin_family = AF_INET;
4332 4333 } else {
4333 4334 ASSERT(connp->conn_family == AF_INET6);
4334 4335 len = sizeof (sin6_t);
4335 4336 sin6 = (sin6_t *)&addr;
4336 4337 *sin6 = sin6_null;
4337 4338 sin6->sin6_family = AF_INET6;
4338 4339 }
4339 4340 sa = (struct sockaddr *)&addr;
4340 4341 }
4341 4342
4342 4343 error = tcp_bind_check(connp, sa, len, cr,
4343 4344 bind_to_req_port_only);
4344 4345 if (error)
4345 4346 return (error);
4346 4347 /* Fall through and do the fanout insertion */
4347 4348 }
4348 4349
4349 4350 do_listen:
4350 4351 ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
4351 4352 tcp->tcp_conn_req_max = backlog;
4352 4353 if (tcp->tcp_conn_req_max) {
4353 4354 if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
4354 4355 tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
4355 4356 if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
4356 4357 tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
4357 4358 /*
4358 4359 * If this is a listener, do not reset the eager list
4359 4360 * and other stuffs. Note that we don't check if the
4360 4361 * existing eager list meets the new tcp_conn_req_max
4361 4362 * requirement.
4362 4363 */
4363 4364 if (tcp->tcp_state != TCPS_LISTEN) {
4364 4365 tcp->tcp_state = TCPS_LISTEN;
4365 4366 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4366 4367 connp->conn_ixa, void, NULL, tcp_t *, tcp,
4367 4368 void, NULL, int32_t, TCPS_BOUND);
4368 4369 /* Initialize the chain. Don't need the eager_lock */
4369 4370 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
4370 4371 tcp->tcp_eager_next_drop_q0 = tcp;
4371 4372 tcp->tcp_eager_prev_drop_q0 = tcp;
4372 4373 tcp->tcp_second_ctimer_threshold =
4373 4374 tcps->tcps_ip_abort_linterval;
4374 4375 }
4375 4376 }
4376 4377
4377 4378 /*
4378 4379 * We need to make sure that the conn_recv is set to a non-null
4379 4380 * value before we insert the conn into the classifier table.
4380 4381 * This is to avoid a race with an incoming packet which does an
4381 4382 * ipcl_classify().
4382 4383 * We initially set it to tcp_input_listener_unbound to try to
4383 4384 * pick a good squeue for the listener when the first SYN arrives.
4384 4385 * tcp_input_listener_unbound sets it to tcp_input_listener on that
4385 4386 * first SYN.
4386 4387 */
4387 4388 connp->conn_recv = tcp_input_listener_unbound;
4388 4389
4389 4390 /* Insert the listener in the classifier table */
4390 4391 error = ip_laddr_fanout_insert(connp);
4391 4392 if (error != 0) {
4392 4393 /* Undo the bind - release the port number */
4393 4394 oldstate = tcp->tcp_state;
4394 4395 tcp->tcp_state = TCPS_IDLE;
4395 4396 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4396 4397 connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4397 4398 int32_t, oldstate);
4398 4399 connp->conn_bound_addr_v6 = ipv6_all_zeros;
4399 4400
4400 4401 connp->conn_laddr_v6 = ipv6_all_zeros;
4401 4402 connp->conn_saddr_v6 = ipv6_all_zeros;
4402 4403 connp->conn_ports = 0;
4403 4404
4404 4405 if (connp->conn_anon_port) {
4405 4406 zone_t *zone;
4406 4407
4407 4408 zone = crgetzone(cr);
4408 4409 connp->conn_anon_port = B_FALSE;
4409 4410 (void) tsol_mlp_anon(zone, connp->conn_mlp_type,
4410 4411 connp->conn_proto, connp->conn_lport, B_FALSE);
4411 4412 }
4412 4413 connp->conn_mlp_type = mlptSingle;
4413 4414
4414 4415 tcp_bind_hash_remove(tcp);
4415 4416 return (error);
4416 4417 } else {
4417 4418 /*
4418 4419 * If there is a connection limit, allocate and initialize
4419 4420 * the counter struct. Note that since listen can be called
4420 4421 * multiple times, the struct may have been allready allocated.
4421 4422 */
4422 4423 if (!list_is_empty(&tcps->tcps_listener_conf) &&
4423 4424 tcp->tcp_listen_cnt == NULL) {
4424 4425 tcp_listen_cnt_t *tlc;
4425 4426 uint32_t ratio;
4426 4427
4427 4428 ratio = tcp_find_listener_conf(tcps,
4428 4429 ntohs(connp->conn_lport));
4429 4430 if (ratio != 0) {
4430 4431 uint32_t mem_ratio, tot_buf;
4431 4432
4432 4433 tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
4433 4434 KM_SLEEP);
4434 4435 /*
4435 4436 * Calculate the connection limit based on
4436 4437 * the configured ratio and maxusers. Maxusers
4437 4438 * are calculated based on memory size,
4438 4439 * ~ 1 user per MB. Note that the conn_rcvbuf
4439 4440 * and conn_sndbuf may change after a
4440 4441 * connection is accepted. So what we have
4441 4442 * is only an approximation.
4442 4443 */
4443 4444 if ((tot_buf = connp->conn_rcvbuf +
4444 4445 connp->conn_sndbuf) < MB) {
4445 4446 mem_ratio = MB / tot_buf;
4446 4447 tlc->tlc_max = maxusers / ratio *
4447 4448 mem_ratio;
4448 4449 } else {
4449 4450 mem_ratio = tot_buf / MB;
4450 4451 tlc->tlc_max = maxusers / ratio /
4451 4452 mem_ratio;
4452 4453 }
4453 4454 /* At least we should allow two connections! */
4454 4455 if (tlc->tlc_max <= tcp_min_conn_listener)
4455 4456 tlc->tlc_max = tcp_min_conn_listener;
4456 4457 tlc->tlc_cnt = 1;
4457 4458 tlc->tlc_drop = 0;
4458 4459 tcp->tcp_listen_cnt = tlc;
4459 4460 }
4460 4461 }
4461 4462 }
4462 4463 return (error);
4463 4464 }
|
↓ open down ↓ |
4210 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX