1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* 27 * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T 28 * All Rights Reserved 29 */ 30 31 /* 32 * Portions of this source code were derived from Berkeley 4.3 BSD 33 * under license from the Regents of the University of California. 34 */ 35 36 #pragma ident "%Z%%M% %I% %E% SMI" 37 38 /* 39 * Implements a kernel based, client side RPC over Connection Oriented 40 * Transports (COTS). 41 */ 42 43 /* 44 * Much of this file has been re-written to let NFS work better over slow 45 * transports. A description follows. 46 * 47 * One of the annoying things about kRPC/COTS is that it will temporarily 48 * create more than one connection between a client and server. This 49 * happens because when a connection is made, the end-points entry in the 50 * linked list of connections (headed by cm_hd), is removed so that other 51 * threads don't mess with it. Went ahead and bit the bullet by keeping 52 * the endpoint on the connection list and introducing state bits, 53 * condition variables etc. to the connection entry data structure (struct 54 * cm_xprt). 55 * 56 * Here is a summary of the changes to cm-xprt: 57 * 58 * x_ctime is the timestamp of when the endpoint was last 59 * connected or disconnected. If an end-point is ever disconnected 60 * or re-connected, then any outstanding RPC request is presumed 61 * lost, telling clnt_cots_kcallit that it needs to re-send the 62 * request, not just wait for the original request's reply to 63 * arrive. 64 * 65 * x_thread flag which tells us if a thread is doing a connection attempt. 66 * 67 * x_waitdis flag which tells us we are waiting a disconnect ACK. 68 * 69 * x_needdis flag which tells us we need to send a T_DISCONN_REQ 70 * to kill the connection. 71 * 72 * x_needrel flag which tells us we need to send a T_ORDREL_REQ to 73 * gracefully close the connection. 74 * 75 * #defined bitmasks for the all the b_* bits so that more 76 * efficient (and at times less clumsy) masks can be used to 77 * manipulated state in cases where multiple bits have to 78 * set/cleared/checked in the same critical section. 79 * 80 * x_conn_cv and x_dis-_cv are new condition variables to let 81 * threads knows when the connection attempt is done, and to let 82 * the connecting thread know when the disconnect handshake is 83 * done. 84 * 85 * Added the CONN_HOLD() macro so that all reference holds have the same 86 * look and feel. 87 * 88 * In the private (cku_private) portion of the client handle, 89 * 90 * cku_flags replaces the cku_sent a boolean. cku_flags keeps 91 * track of whether a request as been sent, and whether the 92 * client's handles call record is on the dispatch list (so that 93 * the reply can be matched by XID to the right client handle). 94 * The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit() 95 * and still have the response find the right client handle so 96 * that the retry of CLNT_CALL() gets the result. Testing, found 97 * situations where if the timeout was increased, performance 98 * degraded. This was due to us hitting a window where the thread 99 * was back in rfscall() (probably printing server not responding) 100 * while the response came back but no place to put it. 101 * 102 * cku_ctime is just a cache of x_ctime. If they match, 103 * clnt_cots_kcallit() won't to send a retry (unless the maximum 104 * receive count limit as been reached). If the don't match, then 105 * we assume the request has been lost, and a retry of the request 106 * is needed. 107 * 108 * cku_recv_attempts counts the number of receive count attempts 109 * after one try is sent on the wire. 110 * 111 * Added the clnt_delay() routine so that interruptible and 112 * noninterruptible delays are possible. 113 * 114 * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to 115 * control how long the client delays before returned after getting 116 * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash 117 * a server that may be booting and not yet started nfsd. 118 * 119 * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable) 120 * Why don't we just wait forever (receive an infinite # of times)? 121 * Because the server may have rebooted. More insidious is that some 122 * servers (ours) will drop NFS/TCP requests in some cases. This is bad, 123 * but it is a reality. 124 * 125 * The case of a server doing orderly release really messes up the 126 * client's recovery, especially if the server's TCP implementation is 127 * buggy. It was found was that the kRPC/COTS client was breaking some 128 * TPI rules, such as not waiting for the acknowledgement of a 129 * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and 130 * T_DISCON_REQ in clnt_dispatch_notifyall()). 131 * 132 * One of things that we've seen is that a kRPC TCP endpoint goes into 133 * TIMEWAIT and a thus a reconnect takes a long time to satisfy because 134 * that the TIMEWAIT state takes a while to finish. If a server sends a 135 * T_ORDREL_IND, there is little point in an RPC client doing a 136 * T_ORDREL_REQ, because the RPC request isn't going to make it (the 137 * server is saying that it won't accept any more data). So kRPC was 138 * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the 139 * connection skips the TIMEWAIT state and goes straight to a bound state 140 * that kRPC can quickly switch to connected. 141 * 142 * Code that issues TPI request must use waitforack() to wait for the 143 * corresponding ack (assuming there is one) in any future modifications. 144 * This works around problems that may be introduced by breaking TPI rules 145 * (by submitting new calls before earlier requests have been acked) in the 146 * case of a signal or other early return. waitforack() depends on 147 * clnt_dispatch_notifyconn() to issue the wakeup when the ack 148 * arrives, so adding new TPI calls may require corresponding changes 149 * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on 150 * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure 151 * not to set it too low or TPI ACKS will be lost. 152 */ 153 154 #include <sys/param.h> 155 #include <sys/types.h> 156 #include <sys/user.h> 157 #include <sys/systm.h> 158 #include <sys/sysmacros.h> 159 #include <sys/proc.h> 160 #include <sys/socket.h> 161 #include <sys/file.h> 162 #include <sys/stream.h> 163 #include <sys/strsubr.h> 164 #include <sys/stropts.h> 165 #include <sys/strsun.h> 166 #include <sys/timod.h> 167 #include <sys/tiuser.h> 168 #include <sys/tihdr.h> 169 #include <sys/t_kuser.h> 170 #include <sys/fcntl.h> 171 #include <sys/errno.h> 172 #include <sys/kmem.h> 173 #include <sys/debug.h> 174 #include <sys/systm.h> 175 #include <sys/kstat.h> 176 #include <sys/t_lock.h> 177 #include <sys/ddi.h> 178 #include <sys/cmn_err.h> 179 #include <sys/time.h> 180 #include <sys/isa_defs.h> 181 #include <sys/callb.h> 182 #include <sys/sunddi.h> 183 #include <sys/atomic.h> 184 185 #include <netinet/in.h> 186 #include <netinet/tcp.h> 187 188 #include <rpc/types.h> 189 #include <rpc/xdr.h> 190 #include <rpc/auth.h> 191 #include <rpc/clnt.h> 192 #include <rpc/rpc_msg.h> 193 194 #define COTS_DEFAULT_ALLOCSIZE 2048 195 196 #define WIRE_HDR_SIZE 20 /* serialized call header, sans proc number */ 197 #define MSG_OFFSET 128 /* offset of call into the mblk */ 198 199 const char *kinet_ntop6(uchar_t *, char *, size_t); 200 201 static int clnt_cots_ksettimers(CLIENT *, struct rpc_timers *, 202 struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t); 203 static enum clnt_stat clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t, 204 caddr_t, xdrproc_t, caddr_t, struct timeval); 205 static void clnt_cots_kabort(CLIENT *); 206 static void clnt_cots_kerror(CLIENT *, struct rpc_err *); 207 static bool_t clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t); 208 static void clnt_cots_kdestroy(CLIENT *); 209 static bool_t clnt_cots_kcontrol(CLIENT *, int, char *); 210 211 212 /* List of transports managed by the connection manager. */ 213 struct cm_xprt { 214 TIUSER *x_tiptr; /* transport handle */ 215 queue_t *x_wq; /* send queue */ 216 clock_t x_time; /* last time we handed this xprt out */ 217 clock_t x_ctime; /* time we went to CONNECTED */ 218 int x_tidu_size; /* TIDU size of this transport */ 219 union { 220 struct { 221 unsigned int 222 #ifdef _BIT_FIELDS_HTOL 223 b_closing: 1, /* we've sent a ord rel on this conn */ 224 b_dead: 1, /* transport is closed or disconn */ 225 b_doomed: 1, /* too many conns, let this go idle */ 226 b_connected: 1, /* this connection is connected */ 227 228 b_ordrel: 1, /* do an orderly release? */ 229 b_thread: 1, /* thread doing connect */ 230 b_waitdis: 1, /* waiting for disconnect ACK */ 231 b_needdis: 1, /* need T_DISCON_REQ */ 232 233 b_needrel: 1, /* need T_ORDREL_REQ */ 234 b_early_disc: 1, /* got a T_ORDREL_IND or T_DISCON_IND */ 235 /* disconnect during connect */ 236 237 b_pad: 22; 238 239 #endif 240 241 #ifdef _BIT_FIELDS_LTOH 242 b_pad: 22, 243 244 b_early_disc: 1, /* got a T_ORDREL_IND or T_DISCON_IND */ 245 /* disconnect during connect */ 246 b_needrel: 1, /* need T_ORDREL_REQ */ 247 248 b_needdis: 1, /* need T_DISCON_REQ */ 249 b_waitdis: 1, /* waiting for disconnect ACK */ 250 b_thread: 1, /* thread doing connect */ 251 b_ordrel: 1, /* do an orderly release? */ 252 253 b_connected: 1, /* this connection is connected */ 254 b_doomed: 1, /* too many conns, let this go idle */ 255 b_dead: 1, /* transport is closed or disconn */ 256 b_closing: 1; /* we've sent a ord rel on this conn */ 257 #endif 258 } bit; unsigned int word; 259 260 #define x_closing x_state.bit.b_closing 261 #define x_dead x_state.bit.b_dead 262 #define x_doomed x_state.bit.b_doomed 263 #define x_connected x_state.bit.b_connected 264 265 #define x_ordrel x_state.bit.b_ordrel 266 #define x_thread x_state.bit.b_thread 267 #define x_waitdis x_state.bit.b_waitdis 268 #define x_needdis x_state.bit.b_needdis 269 270 #define x_needrel x_state.bit.b_needrel 271 #define x_early_disc x_state.bit.b_early_disc 272 273 #define x_state_flags x_state.word 274 275 #define X_CLOSING 0x80000000 276 #define X_DEAD 0x40000000 277 #define X_DOOMED 0x20000000 278 #define X_CONNECTED 0x10000000 279 280 #define X_ORDREL 0x08000000 281 #define X_THREAD 0x04000000 282 #define X_WAITDIS 0x02000000 283 #define X_NEEDDIS 0x01000000 284 285 #define X_NEEDREL 0x00800000 286 #define X_EARLYDISC 0x00400000 287 288 #define X_BADSTATES (X_CLOSING | X_DEAD | X_DOOMED) 289 290 } x_state; 291 int x_ref; /* number of users of this xprt */ 292 int x_family; /* address family of transport */ 293 dev_t x_rdev; /* device number of transport */ 294 struct cm_xprt *x_next; 295 296 struct netbuf x_server; /* destination address */ 297 struct netbuf x_src; /* src address (for retries) */ 298 kmutex_t x_lock; /* lock on this entry */ 299 kcondvar_t x_cv; /* to signal when can be closed */ 300 kcondvar_t x_conn_cv; /* to signal when connection attempt */ 301 /* is complete */ 302 kstat_t *x_ksp; 303 304 kcondvar_t x_dis_cv; /* to signal when disconnect attempt */ 305 /* is complete */ 306 zoneid_t x_zoneid; /* zone this xprt belongs to */ 307 }; 308 309 typedef struct cm_kstat_xprt { 310 kstat_named_t x_wq; 311 kstat_named_t x_server; 312 kstat_named_t x_family; 313 kstat_named_t x_rdev; 314 kstat_named_t x_time; 315 kstat_named_t x_state; 316 kstat_named_t x_ref; 317 kstat_named_t x_port; 318 } cm_kstat_xprt_t; 319 320 static cm_kstat_xprt_t cm_kstat_template = { 321 { "write_queue", KSTAT_DATA_UINT32 }, 322 { "server", KSTAT_DATA_STRING }, 323 { "addr_family", KSTAT_DATA_UINT32 }, 324 { "device", KSTAT_DATA_UINT32 }, 325 { "time_stamp", KSTAT_DATA_UINT32 }, 326 { "status", KSTAT_DATA_UINT32 }, 327 { "ref_count", KSTAT_DATA_INT32 }, 328 { "port", KSTAT_DATA_UINT32 }, 329 }; 330 331 /* 332 * The inverse of this is connmgr_release(). 333 */ 334 #define CONN_HOLD(Cm_entry) {\ 335 mutex_enter(&(Cm_entry)->x_lock); \ 336 (Cm_entry)->x_ref++; \ 337 mutex_exit(&(Cm_entry)->x_lock); \ 338 } 339 340 341 /* 342 * Private data per rpc handle. This structure is allocated by 343 * clnt_cots_kcreate, and freed by clnt_cots_kdestroy. 344 */ 345 typedef struct cku_private_s { 346 CLIENT cku_client; /* client handle */ 347 calllist_t cku_call; /* for dispatching calls */ 348 struct rpc_err cku_err; /* error status */ 349 350 struct netbuf cku_srcaddr; /* source address for retries */ 351 int cku_addrfmly; /* for binding port */ 352 struct netbuf cku_addr; /* remote address */ 353 dev_t cku_device; /* device to use */ 354 uint_t cku_flags; 355 #define CKU_ONQUEUE 0x1 356 #define CKU_SENT 0x2 357 358 bool_t cku_progress; /* for CLSET_PROGRESS */ 359 uint32_t cku_xid; /* current XID */ 360 clock_t cku_ctime; /* time stamp of when */ 361 /* connection was created */ 362 uint_t cku_recv_attempts; 363 XDR cku_outxdr; /* xdr routine for output */ 364 XDR cku_inxdr; /* xdr routine for input */ 365 char cku_rpchdr[WIRE_HDR_SIZE + 4]; 366 /* pre-serialized rpc header */ 367 368 uint_t cku_outbuflen; /* default output mblk length */ 369 struct cred *cku_cred; /* credentials */ 370 bool_t cku_nodelayonerr; 371 /* for CLSET_NODELAYONERR */ 372 int cku_useresvport; /* Use reserved port */ 373 struct rpc_cots_client *cku_stats; /* stats for zone */ 374 } cku_private_t; 375 376 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *, 377 const struct timeval *, struct netbuf *, int, struct netbuf *, 378 struct rpc_err *, bool_t, bool_t); 379 380 static bool_t connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *, 381 int, calllist_t *, int *, bool_t reconnect, 382 const struct timeval *, bool_t); 383 384 static bool_t connmgr_setopt(queue_t *, int, int, calllist_t *); 385 static void connmgr_sndrel(struct cm_xprt *); 386 static void connmgr_snddis(struct cm_xprt *); 387 static void connmgr_close(struct cm_xprt *); 388 static void connmgr_release(struct cm_xprt *); 389 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *, 390 cku_private_t *); 391 392 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *, 393 struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t, 394 bool_t, int); 395 396 static void connmgr_cancelconn(struct cm_xprt *); 397 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *, 398 bool_t); 399 static void connmgr_dis_and_wait(struct cm_xprt *); 400 401 static void clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t, 402 uint_t); 403 404 static int clnt_delay(clock_t, bool_t); 405 406 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t); 407 408 /* 409 * Operations vector for TCP/IP based RPC 410 */ 411 static struct clnt_ops tcp_ops = { 412 clnt_cots_kcallit, /* do rpc call */ 413 clnt_cots_kabort, /* abort call */ 414 clnt_cots_kerror, /* return error status */ 415 clnt_cots_kfreeres, /* free results */ 416 clnt_cots_kdestroy, /* destroy rpc handle */ 417 clnt_cots_kcontrol, /* the ioctl() of rpc */ 418 clnt_cots_ksettimers, /* set retry timers */ 419 }; 420 421 static int rpc_kstat_instance = 0; /* keeps the current instance */ 422 /* number for the next kstat_create */ 423 424 static struct cm_xprt *cm_hd = NULL; 425 static kmutex_t connmgr_lock; /* for connection mngr's list of transports */ 426 427 extern kmutex_t clnt_max_msg_lock; 428 429 static calllist_t *clnt_pending = NULL; 430 extern kmutex_t clnt_pending_lock; 431 432 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE; 433 434 static call_table_t *cots_call_ht; 435 436 static const struct rpc_cots_client { 437 kstat_named_t rccalls; 438 kstat_named_t rcbadcalls; 439 kstat_named_t rcbadxids; 440 kstat_named_t rctimeouts; 441 kstat_named_t rcnewcreds; 442 kstat_named_t rcbadverfs; 443 kstat_named_t rctimers; 444 kstat_named_t rccantconn; 445 kstat_named_t rcnomem; 446 kstat_named_t rcintrs; 447 } cots_rcstat_tmpl = { 448 { "calls", KSTAT_DATA_UINT64 }, 449 { "badcalls", KSTAT_DATA_UINT64 }, 450 { "badxids", KSTAT_DATA_UINT64 }, 451 { "timeouts", KSTAT_DATA_UINT64 }, 452 { "newcreds", KSTAT_DATA_UINT64 }, 453 { "badverfs", KSTAT_DATA_UINT64 }, 454 { "timers", KSTAT_DATA_UINT64 }, 455 { "cantconn", KSTAT_DATA_UINT64 }, 456 { "nomem", KSTAT_DATA_UINT64 }, 457 { "interrupts", KSTAT_DATA_UINT64 } 458 }; 459 460 #define COTSRCSTAT_INCR(p, x) \ 461 atomic_add_64(&(p)->x.value.ui64, 1) 462 463 #define CLNT_MAX_CONNS 1 /* concurrent connections between clnt/srvr */ 464 static int clnt_max_conns = CLNT_MAX_CONNS; 465 466 #define CLNT_MIN_TIMEOUT 10 /* seconds to wait after we get a */ 467 /* connection reset */ 468 #define CLNT_MIN_CONNTIMEOUT 5 /* seconds to wait for a connection */ 469 470 471 static int clnt_cots_min_tout = CLNT_MIN_TIMEOUT; 472 static int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT; 473 474 /* 475 * Limit the number of times we will attempt to receive a reply without 476 * re-sending a response. 477 */ 478 #define CLNT_MAXRECV_WITHOUT_RETRY 3 479 static uint_t clnt_cots_maxrecv = CLNT_MAXRECV_WITHOUT_RETRY; 480 481 uint_t *clnt_max_msg_sizep; 482 void (*clnt_stop_idle)(queue_t *wq); 483 484 #define ptoh(p) (&((p)->cku_client)) 485 #define htop(h) ((cku_private_t *)((h)->cl_private)) 486 487 /* 488 * Times to retry 489 */ 490 #define REFRESHES 2 /* authentication refreshes */ 491 492 /* 493 * The following is used to determine the global default behavior for 494 * COTS when binding to a local port. 495 * 496 * If the value is set to 1 the default will be to select a reserved 497 * (aka privileged) port, if the value is zero the default will be to 498 * use non-reserved ports. Users of kRPC may override this by using 499 * CLNT_CONTROL() and CLSET_BINDRESVPORT. 500 */ 501 static int clnt_cots_do_bindresvport = 1; 502 503 static zone_key_t zone_cots_key; 504 505 /* 506 * We need to do this after all kernel threads in the zone have exited. 507 */ 508 /* ARGSUSED */ 509 static void 510 clnt_zone_destroy(zoneid_t zoneid, void *unused) 511 { 512 struct cm_xprt **cmp; 513 struct cm_xprt *cm_entry; 514 struct cm_xprt *freelist = NULL; 515 516 mutex_enter(&connmgr_lock); 517 cmp = &cm_hd; 518 while ((cm_entry = *cmp) != NULL) { 519 if (cm_entry->x_zoneid == zoneid) { 520 *cmp = cm_entry->x_next; 521 cm_entry->x_next = freelist; 522 freelist = cm_entry; 523 } else { 524 cmp = &cm_entry->x_next; 525 } 526 } 527 mutex_exit(&connmgr_lock); 528 while ((cm_entry = freelist) != NULL) { 529 freelist = cm_entry->x_next; 530 connmgr_close(cm_entry); 531 } 532 } 533 534 int 535 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog, 536 rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl) 537 { 538 CLIENT *h; 539 cku_private_t *p; 540 struct rpc_msg call_msg; 541 struct rpcstat *rpcstat; 542 543 RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog); 544 545 rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone()); 546 ASSERT(rpcstat != NULL); 547 548 /* Allocate and intialize the client handle. */ 549 p = kmem_zalloc(sizeof (*p), KM_SLEEP); 550 551 h = ptoh(p); 552 553 h->cl_private = (caddr_t)p; 554 h->cl_auth = authkern_create(); 555 h->cl_ops = &tcp_ops; 556 557 cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL); 558 mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL); 559 560 /* 561 * If the current sanity check size in rpcmod is smaller 562 * than the size needed, then increase the sanity check. 563 */ 564 if (max_msgsize != 0 && clnt_max_msg_sizep != NULL && 565 max_msgsize > *clnt_max_msg_sizep) { 566 mutex_enter(&clnt_max_msg_lock); 567 if (max_msgsize > *clnt_max_msg_sizep) 568 *clnt_max_msg_sizep = max_msgsize; 569 mutex_exit(&clnt_max_msg_lock); 570 } 571 572 p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE; 573 574 /* Preserialize the call message header */ 575 576 call_msg.rm_xid = 0; 577 call_msg.rm_direction = CALL; 578 call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION; 579 call_msg.rm_call.cb_prog = prog; 580 call_msg.rm_call.cb_vers = vers; 581 582 xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE); 583 584 if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) { 585 RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization " 586 "error\n"); 587 auth_destroy(h->cl_auth); 588 kmem_free(p, sizeof (cku_private_t)); 589 RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n"); 590 return (EINVAL); /* XXX */ 591 } 592 593 /* 594 * The zalloc initialized the fields below. 595 * p->cku_xid = 0; 596 * p->cku_flags = 0; 597 * p->cku_srcaddr.len = 0; 598 * p->cku_srcaddr.maxlen = 0; 599 */ 600 601 p->cku_cred = cred; 602 p->cku_device = dev; 603 p->cku_addrfmly = family; 604 p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP); 605 p->cku_addr.maxlen = addr->maxlen; 606 p->cku_addr.len = addr->len; 607 bcopy(addr->buf, p->cku_addr.buf, addr->len); 608 p->cku_stats = rpcstat->rpc_cots_client; 609 p->cku_useresvport = -1; /* value is has not been set */ 610 611 *ncl = h; 612 return (0); 613 } 614 615 /*ARGSUSED*/ 616 static void 617 clnt_cots_kabort(CLIENT *h) 618 { 619 } 620 621 /* 622 * Return error info on this handle. 623 */ 624 static void 625 clnt_cots_kerror(CLIENT *h, struct rpc_err *err) 626 { 627 /* LINTED pointer alignment */ 628 cku_private_t *p = htop(h); 629 630 *err = p->cku_err; 631 } 632 633 static bool_t 634 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr) 635 { 636 /* LINTED pointer alignment */ 637 cku_private_t *p = htop(h); 638 XDR *xdrs; 639 640 xdrs = &(p->cku_outxdr); 641 xdrs->x_op = XDR_FREE; 642 return ((*xdr_res)(xdrs, res_ptr)); 643 } 644 645 static bool_t 646 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg) 647 { 648 cku_private_t *p = htop(h); 649 650 switch (cmd) { 651 case CLSET_PROGRESS: 652 p->cku_progress = TRUE; 653 return (TRUE); 654 655 case CLSET_XID: 656 if (arg == NULL) 657 return (FALSE); 658 659 p->cku_xid = *((uint32_t *)arg); 660 return (TRUE); 661 662 case CLGET_XID: 663 if (arg == NULL) 664 return (FALSE); 665 666 *((uint32_t *)arg) = p->cku_xid; 667 return (TRUE); 668 669 case CLSET_NODELAYONERR: 670 if (arg == NULL) 671 return (FALSE); 672 673 if (*((bool_t *)arg) == TRUE) { 674 p->cku_nodelayonerr = TRUE; 675 return (TRUE); 676 } 677 if (*((bool_t *)arg) == FALSE) { 678 p->cku_nodelayonerr = FALSE; 679 return (TRUE); 680 } 681 return (FALSE); 682 683 case CLGET_NODELAYONERR: 684 if (arg == NULL) 685 return (FALSE); 686 687 *((bool_t *)arg) = p->cku_nodelayonerr; 688 return (TRUE); 689 690 case CLSET_BINDRESVPORT: 691 if (arg == NULL) 692 return (FALSE); 693 694 if (*(int *)arg != 1 && *(int *)arg != 0) 695 return (FALSE); 696 697 p->cku_useresvport = *(int *)arg; 698 699 return (TRUE); 700 701 case CLGET_BINDRESVPORT: 702 if (arg == NULL) 703 return (FALSE); 704 705 *(int *)arg = p->cku_useresvport; 706 707 return (TRUE); 708 709 default: 710 return (FALSE); 711 } 712 } 713 714 /* 715 * Destroy rpc handle. Frees the space used for output buffer, 716 * private data, and handle structure. 717 */ 718 static void 719 clnt_cots_kdestroy(CLIENT *h) 720 { 721 /* LINTED pointer alignment */ 722 cku_private_t *p = htop(h); 723 calllist_t *call = &p->cku_call; 724 725 RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h); 726 RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid); 727 728 if (p->cku_flags & CKU_ONQUEUE) { 729 RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x " 730 "from dispatch list\n", p->cku_xid); 731 call_table_remove(call); 732 } 733 734 if (call->call_reply) 735 freemsg(call->call_reply); 736 cv_destroy(&call->call_cv); 737 mutex_destroy(&call->call_lock); 738 739 kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen); 740 kmem_free(p->cku_addr.buf, p->cku_addr.maxlen); 741 kmem_free(p, sizeof (*p)); 742 } 743 744 static int clnt_cots_pulls; 745 #define RM_HDR_SIZE 4 /* record mark header size */ 746 747 /* 748 * Call remote procedure. 749 */ 750 static enum clnt_stat 751 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args, 752 caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait) 753 { 754 /* LINTED pointer alignment */ 755 cku_private_t *p = htop(h); 756 calllist_t *call = &p->cku_call; 757 XDR *xdrs; 758 struct rpc_msg reply_msg; 759 mblk_t *mp; 760 #ifdef RPCDEBUG 761 clock_t time_sent; 762 #endif 763 struct netbuf *retryaddr; 764 struct cm_xprt *cm_entry = NULL; 765 queue_t *wq; 766 int len; 767 int mpsize; 768 int refreshes = REFRESHES; 769 int interrupted; 770 int tidu_size; 771 enum clnt_stat status; 772 struct timeval cwait; 773 bool_t delay_first = FALSE; 774 clock_t ticks; 775 776 RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum); 777 COTSRCSTAT_INCR(p->cku_stats, rccalls); 778 779 RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec); 780 RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec); 781 782 /* 783 * Bug ID 1240234: 784 * Look out for zero length timeouts. We don't want to 785 * wait zero seconds for a connection to be established. 786 */ 787 if (wait.tv_sec < clnt_cots_min_conntout) { 788 cwait.tv_sec = clnt_cots_min_conntout; 789 cwait.tv_usec = 0; 790 RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,", 791 wait.tv_sec); 792 RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout); 793 } else { 794 cwait = wait; 795 } 796 797 call_again: 798 if (cm_entry) { 799 connmgr_release(cm_entry); 800 cm_entry = NULL; 801 } 802 803 mp = NULL; 804 805 /* 806 * If the call is not a retry, allocate a new xid and cache it 807 * for future retries. 808 * Bug ID 1246045: 809 * Treat call as a retry for purposes of binding the source 810 * port only if we actually attempted to send anything on 811 * the previous call. 812 */ 813 if (p->cku_xid == 0) { 814 p->cku_xid = alloc_xid(); 815 call->call_zoneid = rpc_zoneid(); 816 817 /* 818 * We need to ASSERT here that our xid != 0 because this 819 * determines whether or not our call record gets placed on 820 * the hash table or the linked list. By design, we mandate 821 * that RPC calls over cots must have xid's != 0, so we can 822 * ensure proper management of the hash table. 823 */ 824 ASSERT(p->cku_xid != 0); 825 826 retryaddr = NULL; 827 p->cku_flags &= ~CKU_SENT; 828 829 if (p->cku_flags & CKU_ONQUEUE) { 830 RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old" 831 " one (%p)\n", (void *)call); 832 call_table_remove(call); 833 p->cku_flags &= ~CKU_ONQUEUE; 834 RPCLOG(64, "clnt_cots_kcallit: removing call from " 835 "dispatch list because xid was zero (now 0x%x)\n", 836 p->cku_xid); 837 } 838 839 if (call->call_reply != NULL) { 840 freemsg(call->call_reply); 841 call->call_reply = NULL; 842 } 843 } else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) { 844 retryaddr = NULL; 845 846 } else if (p->cku_flags & CKU_SENT) { 847 retryaddr = &p->cku_srcaddr; 848 849 } else { 850 /* 851 * Bug ID 1246045: Nothing was sent, so set retryaddr to 852 * NULL and let connmgr_get() bind to any source port it 853 * can get. 854 */ 855 retryaddr = NULL; 856 } 857 858 RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid); 859 RPCLOG(64, " flags = 0x%x\n", p->cku_flags); 860 861 p->cku_err.re_status = RPC_TIMEDOUT; 862 p->cku_err.re_errno = p->cku_err.re_terrno = 0; 863 864 cm_entry = connmgr_wrapget(retryaddr, &cwait, p); 865 866 if (cm_entry == NULL) { 867 RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n", 868 clnt_sperrno(p->cku_err.re_status)); 869 870 /* 871 * The reasons why we fail to create a connection are 872 * varied. In most cases we don't want the caller to 873 * immediately retry. This could have one or more 874 * bad effects. This includes flooding the net with 875 * connect requests to ports with no listener; a hard 876 * kernel loop due to all the "reserved" TCP ports being 877 * in use. 878 */ 879 delay_first = TRUE; 880 881 /* 882 * Even if we end up returning EINTR, we still count a 883 * a "can't connect", because the connection manager 884 * might have been committed to waiting for or timing out on 885 * a connection. 886 */ 887 COTSRCSTAT_INCR(p->cku_stats, rccantconn); 888 switch (p->cku_err.re_status) { 889 case RPC_INTR: 890 p->cku_err.re_errno = EINTR; 891 892 /* 893 * No need to delay because a UNIX signal(2) 894 * interrupted us. The caller likely won't 895 * retry the CLNT_CALL() and even if it does, 896 * we assume the caller knows what it is doing. 897 */ 898 delay_first = FALSE; 899 break; 900 901 case RPC_TIMEDOUT: 902 p->cku_err.re_errno = ETIMEDOUT; 903 904 /* 905 * No need to delay because timed out already 906 * on the connection request and assume that the 907 * transport time out is longer than our minimum 908 * timeout, or least not too much smaller. 909 */ 910 delay_first = FALSE; 911 break; 912 913 case RPC_SYSTEMERROR: 914 case RPC_TLIERROR: 915 /* 916 * We want to delay here because a transient 917 * system error has a better chance of going away 918 * if we delay a bit. If it's not transient, then 919 * we don't want end up in a hard kernel loop 920 * due to retries. 921 */ 922 ASSERT(p->cku_err.re_errno != 0); 923 break; 924 925 926 case RPC_CANTCONNECT: 927 /* 928 * RPC_CANTCONNECT is set on T_ERROR_ACK which 929 * implies some error down in the TCP layer or 930 * below. If cku_nodelayonerror is set then we 931 * assume the caller knows not to try too hard. 932 */ 933 RPCLOG0(8, "clnt_cots_kcallit: connection failed,"); 934 RPCLOG0(8, " re_status=RPC_CANTCONNECT,"); 935 RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno); 936 RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr); 937 if (p->cku_nodelayonerr == TRUE) 938 delay_first = FALSE; 939 940 p->cku_err.re_errno = EIO; 941 942 break; 943 944 case RPC_XPRTFAILED: 945 /* 946 * We want to delay here because we likely 947 * got a refused connection. 948 */ 949 if (p->cku_err.re_errno == 0) 950 p->cku_err.re_errno = EIO; 951 952 RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n", 953 p->cku_err.re_errno); 954 955 break; 956 957 default: 958 /* 959 * We delay here because it is better to err 960 * on the side of caution. If we got here then 961 * status could have been RPC_SUCCESS, but we 962 * know that we did not get a connection, so 963 * force the rpc status to RPC_CANTCONNECT. 964 */ 965 p->cku_err.re_status = RPC_CANTCONNECT; 966 p->cku_err.re_errno = EIO; 967 break; 968 } 969 if (delay_first == TRUE) 970 ticks = clnt_cots_min_tout * drv_usectohz(1000000); 971 goto cots_done; 972 } 973 974 /* 975 * If we've never sent any request on this connection (send count 976 * is zero, or the connection has been reset), cache the 977 * the connection's create time and send a request (possibly a retry) 978 */ 979 if ((p->cku_flags & CKU_SENT) == 0 || 980 p->cku_ctime != cm_entry->x_ctime) { 981 p->cku_ctime = cm_entry->x_ctime; 982 983 } else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) && 984 (call->call_reply != NULL || 985 p->cku_recv_attempts < clnt_cots_maxrecv)) { 986 987 /* 988 * If we've sent a request and our call is on the dispatch 989 * queue and we haven't made too many receive attempts, then 990 * don't re-send, just receive. 991 */ 992 p->cku_recv_attempts++; 993 goto read_again; 994 } 995 996 /* 997 * Now we create the RPC request in a STREAMS message. We have to do 998 * this after the call to connmgr_get so that we have the correct 999 * TIDU size for the transport. 1000 */ 1001 tidu_size = cm_entry->x_tidu_size; 1002 len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE); 1003 1004 while ((mp = allocb(len, BPRI_MED)) == NULL) { 1005 if (strwaitbuf(len, BPRI_MED)) { 1006 p->cku_err.re_status = RPC_SYSTEMERROR; 1007 p->cku_err.re_errno = ENOSR; 1008 COTSRCSTAT_INCR(p->cku_stats, rcnomem); 1009 goto cots_done; 1010 } 1011 } 1012 xdrs = &p->cku_outxdr; 1013 xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size); 1014 mpsize = MBLKSIZE(mp); 1015 ASSERT(mpsize >= len); 1016 ASSERT(mp->b_rptr == mp->b_datap->db_base); 1017 1018 /* 1019 * If the size of mblk is not appreciably larger than what we 1020 * asked, then resize the mblk to exactly len bytes. The reason for 1021 * this: suppose len is 1600 bytes, the tidu is 1460 bytes 1022 * (from TCP over ethernet), and the arguments to the RPC require 1023 * 2800 bytes. Ideally we want the protocol to render two 1024 * ~1400 byte segments over the wire. However if allocb() gives us a 2k 1025 * mblk, and we allocate a second mblk for the remainder, the protocol 1026 * module may generate 3 segments over the wire: 1027 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and 1028 * 892 for the third. If we "waste" 448 bytes in the first mblk, 1029 * the XDR encoding will generate two ~1400 byte mblks, and the 1030 * protocol module is more likely to produce properly sized segments. 1031 */ 1032 if ((mpsize >> 1) <= len) 1033 mp->b_rptr += (mpsize - len); 1034 1035 /* 1036 * Adjust b_rptr to reserve space for the non-data protocol headers 1037 * any downstream modules might like to add, and for the 1038 * record marking header. 1039 */ 1040 mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE); 1041 1042 if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) { 1043 /* Copy in the preserialized RPC header information. */ 1044 bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE); 1045 1046 /* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */ 1047 XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base + 1048 WIRE_HDR_SIZE)); 1049 1050 ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE); 1051 1052 /* Serialize the procedure number and the arguments. */ 1053 if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) || 1054 (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) || 1055 (!(*xdr_args)(xdrs, argsp))) { 1056 p->cku_err.re_status = RPC_CANTENCODEARGS; 1057 p->cku_err.re_errno = EIO; 1058 goto cots_done; 1059 } 1060 1061 (*(uint32_t *)(mp->b_rptr)) = p->cku_xid; 1062 } else { 1063 uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE]; 1064 IXDR_PUT_U_INT32(uproc, procnum); 1065 1066 (*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid; 1067 1068 /* Use XDR_SETPOS() to set the b_wptr. */ 1069 XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base)); 1070 1071 /* Serialize the procedure number and the arguments. */ 1072 if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4, 1073 xdrs, xdr_args, argsp)) { 1074 p->cku_err.re_status = RPC_CANTENCODEARGS; 1075 p->cku_err.re_errno = EIO; 1076 goto cots_done; 1077 } 1078 } 1079 1080 RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n", 1081 tidu_size); 1082 1083 wq = cm_entry->x_wq; 1084 clnt_dispatch_send(wq, mp, call, p->cku_xid, 1085 (p->cku_flags & CKU_ONQUEUE)); 1086 1087 RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n", 1088 (uint_t)p->cku_xid); 1089 p->cku_flags = (CKU_ONQUEUE|CKU_SENT); 1090 p->cku_recv_attempts = 1; 1091 1092 #ifdef RPCDEBUG 1093 time_sent = lbolt; 1094 #endif 1095 1096 /* 1097 * Wait for a reply or a timeout. If there is no error or timeout, 1098 * (both indicated by call_status), call->call_reply will contain 1099 * the RPC reply message. 1100 */ 1101 read_again: 1102 mutex_enter(&call->call_lock); 1103 interrupted = 0; 1104 if (call->call_status == RPC_TIMEDOUT) { 1105 /* 1106 * Indicate that the lwp is not to be stopped while waiting 1107 * for this network traffic. This is to avoid deadlock while 1108 * debugging a process via /proc and also to avoid recursive 1109 * mutex_enter()s due to NFS page faults while stopping 1110 * (NFS holds locks when it calls here). 1111 */ 1112 clock_t cv_wait_ret; 1113 clock_t timout; 1114 clock_t oldlbolt; 1115 1116 klwp_t *lwp = ttolwp(curthread); 1117 1118 if (lwp != NULL) 1119 lwp->lwp_nostop++; 1120 1121 oldlbolt = lbolt; 1122 timout = wait.tv_sec * drv_usectohz(1000000) + 1123 drv_usectohz(wait.tv_usec) + oldlbolt; 1124 /* 1125 * Iterate until the call_status is changed to something 1126 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns 1127 * something <=0 zero. The latter means that we timed 1128 * out. 1129 */ 1130 if (h->cl_nosignal) 1131 while ((cv_wait_ret = cv_timedwait(&call->call_cv, 1132 &call->call_lock, timout)) > 0 && 1133 call->call_status == RPC_TIMEDOUT) 1134 ; 1135 else 1136 while ((cv_wait_ret = cv_timedwait_sig( 1137 &call->call_cv, 1138 &call->call_lock, timout)) > 0 && 1139 call->call_status == RPC_TIMEDOUT) 1140 ; 1141 1142 switch (cv_wait_ret) { 1143 case 0: 1144 /* 1145 * If we got out of the above loop with 1146 * cv_timedwait_sig() returning 0, then we were 1147 * interrupted regardless what call_status is. 1148 */ 1149 interrupted = 1; 1150 break; 1151 case -1: 1152 /* cv_timedwait_sig() timed out */ 1153 break; 1154 default: 1155 1156 /* 1157 * We were cv_signaled(). If we didn't 1158 * get a successful call_status and returned 1159 * before time expired, delay up to clnt_cots_min_tout 1160 * seconds so that the caller doesn't immediately 1161 * try to call us again and thus force the 1162 * same condition that got us here (such 1163 * as a RPC_XPRTFAILED due to the server not 1164 * listening on the end-point. 1165 */ 1166 if (call->call_status != RPC_SUCCESS) { 1167 clock_t curlbolt; 1168 clock_t diff; 1169 1170 curlbolt = ddi_get_lbolt(); 1171 ticks = clnt_cots_min_tout * 1172 drv_usectohz(1000000); 1173 diff = curlbolt - oldlbolt; 1174 if (diff < ticks) { 1175 delay_first = TRUE; 1176 if (diff > 0) 1177 ticks -= diff; 1178 } 1179 } 1180 break; 1181 } 1182 1183 if (lwp != NULL) 1184 lwp->lwp_nostop--; 1185 } 1186 /* 1187 * Get the reply message, if any. This will be freed at the end 1188 * whether or not an error occurred. 1189 */ 1190 mp = call->call_reply; 1191 call->call_reply = NULL; 1192 1193 /* 1194 * call_err is the error info when the call is on dispatch queue. 1195 * cku_err is the error info returned to the caller. 1196 * Sync cku_err with call_err for local message processing. 1197 */ 1198 1199 status = call->call_status; 1200 p->cku_err = call->call_err; 1201 mutex_exit(&call->call_lock); 1202 1203 if (status != RPC_SUCCESS) { 1204 switch (status) { 1205 case RPC_TIMEDOUT: 1206 if (interrupted) { 1207 COTSRCSTAT_INCR(p->cku_stats, rcintrs); 1208 p->cku_err.re_status = RPC_INTR; 1209 p->cku_err.re_errno = EINTR; 1210 RPCLOG(1, "clnt_cots_kcallit: xid 0x%x", 1211 p->cku_xid); 1212 RPCLOG(1, "signal interrupted at %ld", lbolt); 1213 RPCLOG(1, ", was sent at %ld\n", time_sent); 1214 } else { 1215 COTSRCSTAT_INCR(p->cku_stats, rctimeouts); 1216 p->cku_err.re_errno = ETIMEDOUT; 1217 RPCLOG(1, "clnt_cots_kcallit: timed out at %ld", 1218 lbolt); 1219 RPCLOG(1, ", was sent at %ld\n", time_sent); 1220 } 1221 break; 1222 1223 case RPC_XPRTFAILED: 1224 if (p->cku_err.re_errno == 0) 1225 p->cku_err.re_errno = EIO; 1226 1227 RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n", 1228 p->cku_err.re_errno); 1229 break; 1230 1231 case RPC_SYSTEMERROR: 1232 ASSERT(p->cku_err.re_errno); 1233 RPCLOG(1, "clnt_cots_kcallit: system error: %d\n", 1234 p->cku_err.re_errno); 1235 break; 1236 1237 default: 1238 p->cku_err.re_status = RPC_SYSTEMERROR; 1239 p->cku_err.re_errno = EIO; 1240 RPCLOG(1, "clnt_cots_kcallit: error: %s\n", 1241 clnt_sperrno(status)); 1242 break; 1243 } 1244 if (p->cku_err.re_status != RPC_TIMEDOUT) { 1245 1246 if (p->cku_flags & CKU_ONQUEUE) { 1247 call_table_remove(call); 1248 p->cku_flags &= ~CKU_ONQUEUE; 1249 } 1250 1251 RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x " 1252 "taken off dispatch list\n", p->cku_xid); 1253 if (call->call_reply) { 1254 freemsg(call->call_reply); 1255 call->call_reply = NULL; 1256 } 1257 } else if (wait.tv_sec != 0) { 1258 /* 1259 * We've sent the request over TCP and so we have 1260 * every reason to believe it will get 1261 * delivered. In which case returning a timeout is not 1262 * appropriate. 1263 */ 1264 if (p->cku_progress == TRUE && 1265 p->cku_recv_attempts < clnt_cots_maxrecv) { 1266 p->cku_err.re_status = RPC_INPROGRESS; 1267 } 1268 } 1269 goto cots_done; 1270 } 1271 1272 xdrs = &p->cku_inxdr; 1273 xdrmblk_init(xdrs, mp, XDR_DECODE, 0); 1274 1275 reply_msg.rm_direction = REPLY; 1276 reply_msg.rm_reply.rp_stat = MSG_ACCEPTED; 1277 reply_msg.acpted_rply.ar_stat = SUCCESS; 1278 1279 reply_msg.acpted_rply.ar_verf = _null_auth; 1280 /* 1281 * xdr_results will be done in AUTH_UNWRAP. 1282 */ 1283 reply_msg.acpted_rply.ar_results.where = NULL; 1284 reply_msg.acpted_rply.ar_results.proc = xdr_void; 1285 1286 if (xdr_replymsg(xdrs, &reply_msg)) { 1287 enum clnt_stat re_status; 1288 1289 _seterr_reply(&reply_msg, &p->cku_err); 1290 1291 re_status = p->cku_err.re_status; 1292 if (re_status == RPC_SUCCESS) { 1293 /* 1294 * Reply is good, check auth. 1295 */ 1296 if (!AUTH_VALIDATE(h->cl_auth, 1297 &reply_msg.acpted_rply.ar_verf)) { 1298 COTSRCSTAT_INCR(p->cku_stats, rcbadverfs); 1299 RPCLOG0(1, "clnt_cots_kcallit: validation " 1300 "failure\n"); 1301 freemsg(mp); 1302 (void) xdr_rpc_free_verifier(xdrs, &reply_msg); 1303 mutex_enter(&call->call_lock); 1304 if (call->call_reply == NULL) 1305 call->call_status = RPC_TIMEDOUT; 1306 mutex_exit(&call->call_lock); 1307 goto read_again; 1308 } else if (!AUTH_UNWRAP(h->cl_auth, xdrs, 1309 xdr_results, resultsp)) { 1310 RPCLOG0(1, "clnt_cots_kcallit: validation " 1311 "failure (unwrap)\n"); 1312 p->cku_err.re_status = RPC_CANTDECODERES; 1313 p->cku_err.re_errno = EIO; 1314 } 1315 } else { 1316 /* set errno in case we can't recover */ 1317 if (re_status != RPC_VERSMISMATCH && 1318 re_status != RPC_AUTHERROR && 1319 re_status != RPC_PROGVERSMISMATCH) 1320 p->cku_err.re_errno = EIO; 1321 1322 if (re_status == RPC_AUTHERROR) { 1323 /* 1324 * Maybe our credential need to be refreshed 1325 */ 1326 if (cm_entry) { 1327 /* 1328 * There is the potential that the 1329 * cm_entry has/will be marked dead, 1330 * so drop the connection altogether, 1331 * force REFRESH to establish new 1332 * connection. 1333 */ 1334 connmgr_cancelconn(cm_entry); 1335 cm_entry = NULL; 1336 } 1337 1338 if ((refreshes > 0) && 1339 AUTH_REFRESH(h->cl_auth, &reply_msg, 1340 p->cku_cred)) { 1341 refreshes--; 1342 (void) xdr_rpc_free_verifier(xdrs, 1343 &reply_msg); 1344 freemsg(mp); 1345 mp = NULL; 1346 1347 if (p->cku_flags & CKU_ONQUEUE) { 1348 call_table_remove(call); 1349 p->cku_flags &= ~CKU_ONQUEUE; 1350 } 1351 1352 RPCLOG(64, 1353 "clnt_cots_kcallit: AUTH_ERROR, xid" 1354 " 0x%x removed off dispatch list\n", 1355 p->cku_xid); 1356 if (call->call_reply) { 1357 freemsg(call->call_reply); 1358 call->call_reply = NULL; 1359 } 1360 1361 COTSRCSTAT_INCR(p->cku_stats, 1362 rcbadcalls); 1363 COTSRCSTAT_INCR(p->cku_stats, 1364 rcnewcreds); 1365 goto call_again; 1366 } 1367 1368 /* 1369 * We have used the client handle to 1370 * do an AUTH_REFRESH and the RPC status may 1371 * be set to RPC_SUCCESS; Let's make sure to 1372 * set it to RPC_AUTHERROR. 1373 */ 1374 p->cku_err.re_status = RPC_AUTHERROR; 1375 1376 /* 1377 * Map recoverable and unrecoverable 1378 * authentication errors to appropriate errno 1379 */ 1380 switch (p->cku_err.re_why) { 1381 case AUTH_TOOWEAK: 1382 /* 1383 * This could be a failure where the 1384 * server requires use of a reserved 1385 * port, check and optionally set the 1386 * client handle useresvport trying 1387 * one more time. Next go round we 1388 * fall out with the tooweak error. 1389 */ 1390 if (p->cku_useresvport != 1) { 1391 p->cku_useresvport = 1; 1392 p->cku_xid = 0; 1393 (void) xdr_rpc_free_verifier 1394 (xdrs, &reply_msg); 1395 freemsg(mp); 1396 goto call_again; 1397 } 1398 /* FALLTHRU */ 1399 case AUTH_BADCRED: 1400 case AUTH_BADVERF: 1401 case AUTH_INVALIDRESP: 1402 case AUTH_FAILED: 1403 case RPCSEC_GSS_NOCRED: 1404 case RPCSEC_GSS_FAILED: 1405 p->cku_err.re_errno = EACCES; 1406 break; 1407 case AUTH_REJECTEDCRED: 1408 case AUTH_REJECTEDVERF: 1409 default: p->cku_err.re_errno = EIO; 1410 break; 1411 } 1412 RPCLOG(1, "clnt_cots_kcallit : authentication" 1413 " failed with RPC_AUTHERROR of type %d\n", 1414 (int)p->cku_err.re_why); 1415 } 1416 } 1417 } else { 1418 /* reply didn't decode properly. */ 1419 p->cku_err.re_status = RPC_CANTDECODERES; 1420 p->cku_err.re_errno = EIO; 1421 RPCLOG0(1, "clnt_cots_kcallit: decode failure\n"); 1422 } 1423 1424 (void) xdr_rpc_free_verifier(xdrs, &reply_msg); 1425 1426 if (p->cku_flags & CKU_ONQUEUE) { 1427 call_table_remove(call); 1428 p->cku_flags &= ~CKU_ONQUEUE; 1429 } 1430 1431 RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list", 1432 p->cku_xid); 1433 RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status)); 1434 cots_done: 1435 if (cm_entry) 1436 connmgr_release(cm_entry); 1437 1438 if (mp != NULL) 1439 freemsg(mp); 1440 if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) { 1441 freemsg(call->call_reply); 1442 call->call_reply = NULL; 1443 } 1444 if (p->cku_err.re_status != RPC_SUCCESS) { 1445 RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n"); 1446 COTSRCSTAT_INCR(p->cku_stats, rcbadcalls); 1447 } 1448 1449 /* 1450 * No point in delaying if the zone is going away. 1451 */ 1452 if (delay_first == TRUE && 1453 !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) { 1454 if (clnt_delay(ticks, h->cl_nosignal) == EINTR) { 1455 p->cku_err.re_errno = EINTR; 1456 p->cku_err.re_status = RPC_INTR; 1457 } 1458 } 1459 return (p->cku_err.re_status); 1460 } 1461 1462 /* 1463 * Kinit routine for cots. This sets up the correct operations in 1464 * the client handle, as the handle may have previously been a clts 1465 * handle, and clears the xid field so there is no way a new call 1466 * could be mistaken for a retry. It also sets in the handle the 1467 * information that is passed at create/kinit time but needed at 1468 * call time, as cots creates the transport at call time - device, 1469 * address of the server, protocol family. 1470 */ 1471 void 1472 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr, 1473 int max_msgsize, cred_t *cred) 1474 { 1475 /* LINTED pointer alignment */ 1476 cku_private_t *p = htop(h); 1477 calllist_t *call = &p->cku_call; 1478 1479 h->cl_ops = &tcp_ops; 1480 if (p->cku_flags & CKU_ONQUEUE) { 1481 call_table_remove(call); 1482 p->cku_flags &= ~CKU_ONQUEUE; 1483 RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from" 1484 " dispatch list\n", p->cku_xid); 1485 } 1486 1487 if (call->call_reply != NULL) { 1488 freemsg(call->call_reply); 1489 call->call_reply = NULL; 1490 } 1491 1492 call->call_bucket = NULL; 1493 call->call_hash = 0; 1494 1495 /* 1496 * We don't clear cku_flags here, because clnt_cots_kcallit() 1497 * takes care of handling the cku_flags reset. 1498 */ 1499 p->cku_xid = 0; 1500 p->cku_device = dev; 1501 p->cku_addrfmly = family; 1502 p->cku_cred = cred; 1503 1504 if (p->cku_addr.maxlen < addr->len) { 1505 if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL) 1506 kmem_free(p->cku_addr.buf, p->cku_addr.maxlen); 1507 p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP); 1508 p->cku_addr.maxlen = addr->maxlen; 1509 } 1510 1511 p->cku_addr.len = addr->len; 1512 bcopy(addr->buf, p->cku_addr.buf, addr->len); 1513 1514 /* 1515 * If the current sanity check size in rpcmod is smaller 1516 * than the size needed, then increase the sanity check. 1517 */ 1518 if (max_msgsize != 0 && clnt_max_msg_sizep != NULL && 1519 max_msgsize > *clnt_max_msg_sizep) { 1520 mutex_enter(&clnt_max_msg_lock); 1521 if (max_msgsize > *clnt_max_msg_sizep) 1522 *clnt_max_msg_sizep = max_msgsize; 1523 mutex_exit(&clnt_max_msg_lock); 1524 } 1525 } 1526 1527 /* 1528 * ksettimers is a no-op for cots, with the exception of setting the xid. 1529 */ 1530 /* ARGSUSED */ 1531 static int 1532 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all, 1533 int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg, 1534 uint32_t xid) 1535 { 1536 /* LINTED pointer alignment */ 1537 cku_private_t *p = htop(h); 1538 1539 if (xid) 1540 p->cku_xid = xid; 1541 COTSRCSTAT_INCR(p->cku_stats, rctimers); 1542 return (0); 1543 } 1544 1545 extern void rpc_poptimod(struct vnode *); 1546 extern int kstr_push(struct vnode *, char *); 1547 1548 int 1549 conn_kstat_update(kstat_t *ksp, int rw) 1550 { 1551 struct cm_xprt *cm_entry; 1552 struct cm_kstat_xprt *cm_ksp_data; 1553 uchar_t *b; 1554 char *fbuf; 1555 1556 if (rw == KSTAT_WRITE) 1557 return (EACCES); 1558 if (ksp == NULL || ksp->ks_private == NULL) 1559 return (EIO); 1560 cm_entry = (struct cm_xprt *)ksp->ks_private; 1561 cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data; 1562 1563 cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq; 1564 cm_ksp_data->x_family.value.ui32 = cm_entry->x_family; 1565 cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev; 1566 cm_ksp_data->x_time.value.ui32 = cm_entry->x_time; 1567 cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref; 1568 cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags; 1569 1570 if (cm_entry->x_server.buf) { 1571 fbuf = cm_ksp_data->x_server.value.str.addr.ptr; 1572 if (cm_entry->x_family == AF_INET && 1573 cm_entry->x_server.len == 1574 sizeof (struct sockaddr_in)) { 1575 struct sockaddr_in *sa; 1576 sa = (struct sockaddr_in *) 1577 cm_entry->x_server.buf; 1578 b = (uchar_t *)&sa->sin_addr; 1579 (void) sprintf(fbuf, 1580 "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF, 1581 b[2] & 0xFF, b[3] & 0xFF); 1582 cm_ksp_data->x_port.value.ui32 = 1583 (uint32_t)sa->sin_port; 1584 } else if (cm_entry->x_family == AF_INET6 && 1585 cm_entry->x_server.len >= 1586 sizeof (struct sockaddr_in6)) { 1587 /* extract server IP address & port */ 1588 struct sockaddr_in6 *sin6; 1589 sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf; 1590 (void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf, 1591 INET6_ADDRSTRLEN); 1592 cm_ksp_data->x_port.value.ui32 = sin6->sin6_port; 1593 } else { 1594 struct sockaddr_in *sa; 1595 1596 sa = (struct sockaddr_in *)cm_entry->x_server.buf; 1597 b = (uchar_t *)&sa->sin_addr; 1598 (void) sprintf(fbuf, 1599 "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF, 1600 b[2] & 0xFF, b[3] & 0xFF); 1601 } 1602 KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) = 1603 strlen(fbuf) + 1; 1604 } 1605 1606 return (0); 1607 } 1608 1609 1610 /* 1611 * We want a version of delay which is interruptible by a UNIX signal 1612 * Return EINTR if an interrupt occured. 1613 */ 1614 static int 1615 clnt_delay(clock_t ticks, bool_t nosignal) 1616 { 1617 if (nosignal == TRUE) { 1618 delay(ticks); 1619 return (0); 1620 } 1621 return (delay_sig(ticks)); 1622 } 1623 1624 /* 1625 * Wait for a connection until a timeout, or until we are 1626 * signalled that there has been a connection state change. 1627 */ 1628 static enum clnt_stat 1629 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp, 1630 bool_t nosignal) 1631 { 1632 bool_t interrupted; 1633 clock_t timout, cv_stat; 1634 enum clnt_stat clstat; 1635 unsigned int old_state; 1636 1637 ASSERT(MUTEX_HELD(&connmgr_lock)); 1638 /* 1639 * We wait for the transport connection to be made, or an 1640 * indication that it could not be made. 1641 */ 1642 clstat = RPC_TIMEDOUT; 1643 interrupted = FALSE; 1644 1645 old_state = cm_entry->x_state_flags; 1646 /* 1647 * Now loop until cv_timedwait{_sig} returns because of 1648 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be 1649 * cv_signalled for various other reasons too. So loop 1650 * until there is a state change on the connection. 1651 */ 1652 1653 timout = waitp->tv_sec * drv_usectohz(1000000) + 1654 drv_usectohz(waitp->tv_usec) + lbolt; 1655 1656 if (nosignal) { 1657 while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv, 1658 &connmgr_lock, timout)) > 0 && 1659 cm_entry->x_state_flags == old_state) 1660 ; 1661 } else { 1662 while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv, 1663 &connmgr_lock, timout)) > 0 && 1664 cm_entry->x_state_flags == old_state) 1665 ; 1666 1667 if (cv_stat == 0) /* got intr signal? */ 1668 interrupted = TRUE; 1669 } 1670 1671 if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) == 1672 X_CONNECTED) { 1673 clstat = RPC_SUCCESS; 1674 } else { 1675 if (interrupted == TRUE) 1676 clstat = RPC_INTR; 1677 RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n", 1678 clnt_sperrno(clstat)); 1679 } 1680 1681 return (clstat); 1682 } 1683 1684 /* 1685 * Primary interface for how RPC grabs a connection. 1686 */ 1687 static struct cm_xprt * 1688 connmgr_wrapget( 1689 struct netbuf *retryaddr, 1690 const struct timeval *waitp, 1691 cku_private_t *p) 1692 { 1693 struct cm_xprt *cm_entry; 1694 1695 cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly, 1696 &p->cku_srcaddr, &p->cku_err, p->cku_device, 1697 p->cku_client.cl_nosignal, p->cku_useresvport); 1698 1699 if (cm_entry == NULL) { 1700 /* 1701 * Re-map the call status to RPC_INTR if the err code is 1702 * EINTR. This can happen if calls status is RPC_TLIERROR. 1703 * However, don't re-map if signalling has been turned off. 1704 * XXX Really need to create a separate thread whenever 1705 * there isn't an existing connection. 1706 */ 1707 if (p->cku_err.re_errno == EINTR) { 1708 if (p->cku_client.cl_nosignal == TRUE) 1709 p->cku_err.re_errno = EIO; 1710 else 1711 p->cku_err.re_status = RPC_INTR; 1712 } 1713 } 1714 1715 return (cm_entry); 1716 } 1717 1718 /* 1719 * Obtains a transport to the server specified in addr. If a suitable transport 1720 * does not already exist in the list of cached transports, a new connection 1721 * is created, connected, and added to the list. The connection is for sending 1722 * only - the reply message may come back on another transport connection. 1723 */ 1724 static struct cm_xprt * 1725 connmgr_get( 1726 struct netbuf *retryaddr, 1727 const struct timeval *waitp, /* changed to a ptr to converse stack */ 1728 struct netbuf *destaddr, 1729 int addrfmly, 1730 struct netbuf *srcaddr, 1731 struct rpc_err *rpcerr, 1732 dev_t device, 1733 bool_t nosignal, 1734 int useresvport) 1735 { 1736 struct cm_xprt *cm_entry; 1737 struct cm_xprt *lru_entry; 1738 struct cm_xprt **cmp; 1739 queue_t *wq; 1740 TIUSER *tiptr; 1741 int i; 1742 int retval; 1743 clock_t prev_time; 1744 int tidu_size; 1745 bool_t connected; 1746 zoneid_t zoneid = rpc_zoneid(); 1747 1748 /* 1749 * If the call is not a retry, look for a transport entry that 1750 * goes to the server of interest. 1751 */ 1752 mutex_enter(&connmgr_lock); 1753 1754 if (retryaddr == NULL) { 1755 use_new_conn: 1756 i = 0; 1757 cm_entry = lru_entry = NULL; 1758 prev_time = lbolt; 1759 1760 cmp = &cm_hd; 1761 while ((cm_entry = *cmp) != NULL) { 1762 ASSERT(cm_entry != cm_entry->x_next); 1763 /* 1764 * Garbage collect conections that are marked 1765 * for needs disconnect. 1766 */ 1767 if (cm_entry->x_needdis) { 1768 CONN_HOLD(cm_entry); 1769 connmgr_dis_and_wait(cm_entry); 1770 connmgr_release(cm_entry); 1771 /* 1772 * connmgr_lock could have been 1773 * dropped for the disconnect 1774 * processing so start over. 1775 */ 1776 goto use_new_conn; 1777 } 1778 1779 /* 1780 * Garbage collect the dead connections that have 1781 * no threads working on them. 1782 */ 1783 if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) == 1784 X_DEAD) { 1785 mutex_enter(&cm_entry->x_lock); 1786 if (cm_entry->x_ref != 0) { 1787 /* 1788 * Currently in use. 1789 * Cleanup later. 1790 */ 1791 cmp = &cm_entry->x_next; 1792 mutex_exit(&cm_entry->x_lock); 1793 continue; 1794 } 1795 mutex_exit(&cm_entry->x_lock); 1796 *cmp = cm_entry->x_next; 1797 mutex_exit(&connmgr_lock); 1798 connmgr_close(cm_entry); 1799 mutex_enter(&connmgr_lock); 1800 goto use_new_conn; 1801 } 1802 1803 1804 if ((cm_entry->x_state_flags & X_BADSTATES) == 0 && 1805 cm_entry->x_zoneid == zoneid && 1806 cm_entry->x_rdev == device && 1807 destaddr->len == cm_entry->x_server.len && 1808 bcmp(destaddr->buf, cm_entry->x_server.buf, 1809 destaddr->len) == 0) { 1810 /* 1811 * If the matching entry isn't connected, 1812 * attempt to reconnect it. 1813 */ 1814 if (cm_entry->x_connected == FALSE) { 1815 /* 1816 * We don't go through trying 1817 * to find the least recently 1818 * used connected because 1819 * connmgr_reconnect() briefly 1820 * dropped the connmgr_lock, 1821 * allowing a window for our 1822 * accounting to be messed up. 1823 * In any case, a re-connected 1824 * connection is as good as 1825 * a LRU connection. 1826 */ 1827 return (connmgr_wrapconnect(cm_entry, 1828 waitp, destaddr, addrfmly, srcaddr, 1829 rpcerr, TRUE, nosignal)); 1830 } 1831 i++; 1832 if (cm_entry->x_time - prev_time <= 0 || 1833 lru_entry == NULL) { 1834 prev_time = cm_entry->x_time; 1835 lru_entry = cm_entry; 1836 } 1837 } 1838 cmp = &cm_entry->x_next; 1839 } 1840 1841 if (i > clnt_max_conns) { 1842 RPCLOG(8, "connmgr_get: too many conns, dooming entry" 1843 " %p\n", (void *)lru_entry->x_tiptr); 1844 lru_entry->x_doomed = TRUE; 1845 goto use_new_conn; 1846 } 1847 1848 /* 1849 * If we are at the maximum number of connections to 1850 * the server, hand back the least recently used one. 1851 */ 1852 if (i == clnt_max_conns) { 1853 /* 1854 * Copy into the handle the source address of 1855 * the connection, which we will use in case of 1856 * a later retry. 1857 */ 1858 if (srcaddr->len != lru_entry->x_src.len) { 1859 if (srcaddr->len > 0) 1860 kmem_free(srcaddr->buf, 1861 srcaddr->maxlen); 1862 srcaddr->buf = kmem_zalloc( 1863 lru_entry->x_src.len, KM_SLEEP); 1864 srcaddr->maxlen = srcaddr->len = 1865 lru_entry->x_src.len; 1866 } 1867 bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len); 1868 RPCLOG(2, "connmgr_get: call going out on %p\n", 1869 (void *)lru_entry); 1870 lru_entry->x_time = lbolt; 1871 CONN_HOLD(lru_entry); 1872 mutex_exit(&connmgr_lock); 1873 return (lru_entry); 1874 } 1875 1876 } else { 1877 /* 1878 * This is the retry case (retryaddr != NULL). Retries must 1879 * be sent on the same source port as the original call. 1880 */ 1881 1882 /* 1883 * Walk the list looking for a connection with a source address 1884 * that matches the retry address. 1885 */ 1886 cmp = &cm_hd; 1887 while ((cm_entry = *cmp) != NULL) { 1888 ASSERT(cm_entry != cm_entry->x_next); 1889 if (zoneid != cm_entry->x_zoneid || 1890 device != cm_entry->x_rdev || 1891 retryaddr->len != cm_entry->x_src.len || 1892 bcmp(retryaddr->buf, cm_entry->x_src.buf, 1893 retryaddr->len) != 0) { 1894 cmp = &cm_entry->x_next; 1895 continue; 1896 } 1897 1898 /* 1899 * Sanity check: if the connection with our source 1900 * port is going to some other server, something went 1901 * wrong, as we never delete connections (i.e. release 1902 * ports) unless they have been idle. In this case, 1903 * it is probably better to send the call out using 1904 * a new source address than to fail it altogether, 1905 * since that port may never be released. 1906 */ 1907 if (destaddr->len != cm_entry->x_server.len || 1908 bcmp(destaddr->buf, cm_entry->x_server.buf, 1909 destaddr->len) != 0) { 1910 RPCLOG(1, "connmgr_get: tiptr %p" 1911 " is going to a different server" 1912 " with the port that belongs" 1913 " to us!\n", (void *)cm_entry->x_tiptr); 1914 retryaddr = NULL; 1915 goto use_new_conn; 1916 } 1917 1918 /* 1919 * If the connection of interest is not connected and we 1920 * can't reconnect it, then the server is probably 1921 * still down. Return NULL to the caller and let it 1922 * retry later if it wants to. We have a delay so the 1923 * machine doesn't go into a tight retry loop. If the 1924 * entry was already connected, or the reconnected was 1925 * successful, return this entry. 1926 */ 1927 if (cm_entry->x_connected == FALSE) { 1928 return (connmgr_wrapconnect(cm_entry, 1929 waitp, destaddr, addrfmly, NULL, 1930 rpcerr, TRUE, nosignal)); 1931 } else { 1932 CONN_HOLD(cm_entry); 1933 1934 cm_entry->x_time = lbolt; 1935 mutex_exit(&connmgr_lock); 1936 RPCLOG(2, "connmgr_get: found old " 1937 "transport %p for retry\n", 1938 (void *)cm_entry); 1939 return (cm_entry); 1940 } 1941 } 1942 1943 /* 1944 * We cannot find an entry in the list for this retry. 1945 * Either the entry has been removed temporarily to be 1946 * reconnected by another thread, or the original call 1947 * got a port but never got connected, 1948 * and hence the transport never got put in the 1949 * list. Fall through to the "create new connection" code - 1950 * the former case will fail there trying to rebind the port, 1951 * and the later case (and any other pathological cases) will 1952 * rebind and reconnect and not hang the client machine. 1953 */ 1954 RPCLOG0(8, "connmgr_get: no entry in list for retry\n"); 1955 } 1956 /* 1957 * Set up a transport entry in the connection manager's list. 1958 */ 1959 cm_entry = (struct cm_xprt *) 1960 kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP); 1961 1962 cm_entry->x_server.buf = kmem_zalloc(destaddr->len, KM_SLEEP); 1963 bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len); 1964 cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len; 1965 1966 cm_entry->x_state_flags = X_THREAD; 1967 cm_entry->x_ref = 1; 1968 cm_entry->x_family = addrfmly; 1969 cm_entry->x_rdev = device; 1970 cm_entry->x_zoneid = zoneid; 1971 mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL); 1972 cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL); 1973 cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL); 1974 cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL); 1975 1976 /* 1977 * Note that we add this partially initialized entry to the 1978 * connection list. This is so that we don't have connections to 1979 * the same server. 1980 * 1981 * Note that x_src is not initialized at this point. This is because 1982 * retryaddr might be NULL in which case x_src is whatever 1983 * t_kbind/bindresvport gives us. If another thread wants a 1984 * connection to the same server, seemingly we have an issue, but we 1985 * don't. If the other thread comes in with retryaddr == NULL, then it 1986 * will never look at x_src, and it will end up waiting in 1987 * connmgr_cwait() for the first thread to finish the connection 1988 * attempt. If the other thread comes in with retryaddr != NULL, then 1989 * that means there was a request sent on a connection, in which case 1990 * the the connection should already exist. Thus the first thread 1991 * never gets here ... it finds the connection it its server in the 1992 * connection list. 1993 * 1994 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd 1995 * thread will skip us because x_src.len == 0. 1996 */ 1997 cm_entry->x_next = cm_hd; 1998 cm_hd = cm_entry; 1999 mutex_exit(&connmgr_lock); 2000 2001 /* 2002 * Either we didn't find an entry to the server of interest, or we 2003 * don't have the maximum number of connections to that server - 2004 * create a new connection. 2005 */ 2006 RPCLOG0(8, "connmgr_get: creating new connection\n"); 2007 rpcerr->re_status = RPC_TLIERROR; 2008 2009 i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, zone_kcred()); 2010 if (i) { 2011 RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i); 2012 rpcerr->re_errno = i; 2013 connmgr_cancelconn(cm_entry); 2014 return (NULL); 2015 } 2016 rpc_poptimod(tiptr->fp->f_vnode); 2017 2018 if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0, 2019 K_TO_K, kcred, &retval)) { 2020 RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i); 2021 (void) t_kclose(tiptr, 1); 2022 rpcerr->re_errno = i; 2023 connmgr_cancelconn(cm_entry); 2024 return (NULL); 2025 } 2026 2027 if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K, 2028 kcred, &retval)) { 2029 RPCLOG(1, "connmgr_get: can't set client status with cots " 2030 "module, %d\n", i); 2031 (void) t_kclose(tiptr, 1); 2032 rpcerr->re_errno = i; 2033 connmgr_cancelconn(cm_entry); 2034 return (NULL); 2035 } 2036 2037 mutex_enter(&connmgr_lock); 2038 2039 wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next; 2040 cm_entry->x_wq = wq; 2041 2042 mutex_exit(&connmgr_lock); 2043 2044 if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0, 2045 K_TO_K, kcred, &retval)) { 2046 RPCLOG(1, "connmgr_get: can't push timod, %d\n", i); 2047 (void) t_kclose(tiptr, 1); 2048 rpcerr->re_errno = i; 2049 connmgr_cancelconn(cm_entry); 2050 return (NULL); 2051 } 2052 2053 /* 2054 * If the caller has not specified reserved port usage then 2055 * take the system default. 2056 */ 2057 if (useresvport == -1) 2058 useresvport = clnt_cots_do_bindresvport; 2059 2060 if ((useresvport || retryaddr != NULL) && 2061 (addrfmly == AF_INET || addrfmly == AF_INET6)) { 2062 bool_t alloc_src = FALSE; 2063 2064 if (srcaddr->len != destaddr->len) { 2065 kmem_free(srcaddr->buf, srcaddr->maxlen); 2066 srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP); 2067 srcaddr->maxlen = destaddr->len; 2068 srcaddr->len = destaddr->len; 2069 alloc_src = TRUE; 2070 } 2071 2072 if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) { 2073 (void) t_kclose(tiptr, 1); 2074 RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: " 2075 "%p\n", (void *)retryaddr); 2076 2077 /* 2078 * 1225408: If we allocated a source address, then it 2079 * is either garbage or all zeroes. In that case 2080 * we need to clear srcaddr. 2081 */ 2082 if (alloc_src == TRUE) { 2083 kmem_free(srcaddr->buf, srcaddr->maxlen); 2084 srcaddr->maxlen = srcaddr->len = 0; 2085 srcaddr->buf = NULL; 2086 } 2087 rpcerr->re_errno = i; 2088 connmgr_cancelconn(cm_entry); 2089 return (NULL); 2090 } 2091 } else { 2092 if ((i = t_kbind(tiptr, NULL, NULL)) != 0) { 2093 RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i); 2094 (void) t_kclose(tiptr, 1); 2095 rpcerr->re_errno = i; 2096 connmgr_cancelconn(cm_entry); 2097 return (NULL); 2098 } 2099 } 2100 2101 { 2102 /* 2103 * Keep the kernel stack lean. Don't move this call 2104 * declaration to the top of this function because a 2105 * call is declared in connmgr_wrapconnect() 2106 */ 2107 calllist_t call; 2108 2109 bzero(&call, sizeof (call)); 2110 cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL); 2111 2112 /* 2113 * This is a bound end-point so don't close it's stream. 2114 */ 2115 connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly, 2116 &call, &tidu_size, FALSE, waitp, nosignal); 2117 *rpcerr = call.call_err; 2118 cv_destroy(&call.call_cv); 2119 2120 } 2121 2122 mutex_enter(&connmgr_lock); 2123 2124 /* 2125 * Set up a transport entry in the connection manager's list. 2126 */ 2127 cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP); 2128 bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len); 2129 cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len; 2130 2131 cm_entry->x_tiptr = tiptr; 2132 cm_entry->x_time = lbolt; 2133 2134 if (tiptr->tp_info.servtype == T_COTS_ORD) 2135 cm_entry->x_ordrel = TRUE; 2136 else 2137 cm_entry->x_ordrel = FALSE; 2138 2139 cm_entry->x_tidu_size = tidu_size; 2140 2141 if (cm_entry->x_early_disc) { 2142 /* 2143 * We need to check if a disconnect request has come 2144 * while we are connected, if so, then we need to 2145 * set rpcerr->re_status appropriately before returning 2146 * NULL to caller. 2147 */ 2148 if (rpcerr->re_status == RPC_SUCCESS) 2149 rpcerr->re_status = RPC_XPRTFAILED; 2150 cm_entry->x_connected = FALSE; 2151 } else 2152 cm_entry->x_connected = connected; 2153 2154 /* 2155 * There could be a discrepancy here such that 2156 * x_early_disc is TRUE yet connected is TRUE as well 2157 * and the connection is actually connected. In that case 2158 * lets be conservative and declare the connection as not 2159 * connected. 2160 */ 2161 cm_entry->x_early_disc = FALSE; 2162 cm_entry->x_needdis = (cm_entry->x_connected == FALSE); 2163 cm_entry->x_ctime = lbolt; 2164 2165 /* 2166 * Notify any threads waiting that the connection attempt is done. 2167 */ 2168 cm_entry->x_thread = FALSE; 2169 cv_broadcast(&cm_entry->x_conn_cv); 2170 2171 if (cm_entry->x_connected == FALSE) { 2172 mutex_exit(&connmgr_lock); 2173 connmgr_release(cm_entry); 2174 return (NULL); 2175 } 2176 2177 mutex_exit(&connmgr_lock); 2178 2179 return (cm_entry); 2180 } 2181 2182 /* 2183 * Keep the cm_xprt entry on the connecton list when making a connection. This 2184 * is to prevent multiple connections to a slow server from appearing. 2185 * We use the bit field x_thread to tell if a thread is doing a connection 2186 * which keeps other interested threads from messing with connection. 2187 * Those other threads just wait if x_thread is set. 2188 * 2189 * If x_thread is not set, then we do the actual work of connecting via 2190 * connmgr_connect(). 2191 * 2192 * mutex convention: called with connmgr_lock held, returns with it released. 2193 */ 2194 static struct cm_xprt * 2195 connmgr_wrapconnect( 2196 struct cm_xprt *cm_entry, 2197 const struct timeval *waitp, 2198 struct netbuf *destaddr, 2199 int addrfmly, 2200 struct netbuf *srcaddr, 2201 struct rpc_err *rpcerr, 2202 bool_t reconnect, 2203 bool_t nosignal) 2204 { 2205 ASSERT(MUTEX_HELD(&connmgr_lock)); 2206 /* 2207 * Hold this entry as we are about to drop connmgr_lock. 2208 */ 2209 CONN_HOLD(cm_entry); 2210 2211 /* 2212 * If there is a thread already making a connection for us, then 2213 * wait for it to complete the connection. 2214 */ 2215 if (cm_entry->x_thread == TRUE) { 2216 rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal); 2217 2218 if (rpcerr->re_status != RPC_SUCCESS) { 2219 mutex_exit(&connmgr_lock); 2220 connmgr_release(cm_entry); 2221 return (NULL); 2222 } 2223 } else { 2224 bool_t connected; 2225 calllist_t call; 2226 2227 cm_entry->x_thread = TRUE; 2228 2229 while (cm_entry->x_needrel == TRUE) { 2230 cm_entry->x_needrel = FALSE; 2231 2232 connmgr_sndrel(cm_entry); 2233 delay(drv_usectohz(1000000)); 2234 2235 mutex_enter(&connmgr_lock); 2236 } 2237 2238 /* 2239 * If we need to send a T_DISCON_REQ, send one. 2240 */ 2241 connmgr_dis_and_wait(cm_entry); 2242 2243 mutex_exit(&connmgr_lock); 2244 2245 bzero(&call, sizeof (call)); 2246 cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL); 2247 2248 connected = connmgr_connect(cm_entry, cm_entry->x_wq, 2249 destaddr, addrfmly, &call, &cm_entry->x_tidu_size, 2250 reconnect, waitp, nosignal); 2251 2252 *rpcerr = call.call_err; 2253 cv_destroy(&call.call_cv); 2254 2255 mutex_enter(&connmgr_lock); 2256 2257 2258 if (cm_entry->x_early_disc) { 2259 /* 2260 * We need to check if a disconnect request has come 2261 * while we are connected, if so, then we need to 2262 * set rpcerr->re_status appropriately before returning 2263 * NULL to caller. 2264 */ 2265 if (rpcerr->re_status == RPC_SUCCESS) 2266 rpcerr->re_status = RPC_XPRTFAILED; 2267 cm_entry->x_connected = FALSE; 2268 } else 2269 cm_entry->x_connected = connected; 2270 2271 /* 2272 * There could be a discrepancy here such that 2273 * x_early_disc is TRUE yet connected is TRUE as well 2274 * and the connection is actually connected. In that case 2275 * lets be conservative and declare the connection as not 2276 * connected. 2277 */ 2278 2279 cm_entry->x_early_disc = FALSE; 2280 cm_entry->x_needdis = (cm_entry->x_connected == FALSE); 2281 2282 2283 /* 2284 * connmgr_connect() may have given up before the connection 2285 * actually timed out. So ensure that before the next 2286 * connection attempt we do a disconnect. 2287 */ 2288 cm_entry->x_ctime = lbolt; 2289 cm_entry->x_thread = FALSE; 2290 2291 cv_broadcast(&cm_entry->x_conn_cv); 2292 2293 if (cm_entry->x_connected == FALSE) { 2294 mutex_exit(&connmgr_lock); 2295 connmgr_release(cm_entry); 2296 return (NULL); 2297 } 2298 } 2299 2300 if (srcaddr != NULL) { 2301 /* 2302 * Copy into the handle the 2303 * source address of the 2304 * connection, which we will use 2305 * in case of a later retry. 2306 */ 2307 if (srcaddr->len != cm_entry->x_src.len) { 2308 if (srcaddr->maxlen > 0) 2309 kmem_free(srcaddr->buf, srcaddr->maxlen); 2310 srcaddr->buf = kmem_zalloc(cm_entry->x_src.len, 2311 KM_SLEEP); 2312 srcaddr->maxlen = srcaddr->len = 2313 cm_entry->x_src.len; 2314 } 2315 bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len); 2316 } 2317 cm_entry->x_time = lbolt; 2318 mutex_exit(&connmgr_lock); 2319 return (cm_entry); 2320 } 2321 2322 /* 2323 * If we need to send a T_DISCON_REQ, send one. 2324 */ 2325 static void 2326 connmgr_dis_and_wait(struct cm_xprt *cm_entry) 2327 { 2328 ASSERT(MUTEX_HELD(&connmgr_lock)); 2329 for (;;) { 2330 while (cm_entry->x_needdis == TRUE) { 2331 RPCLOG(8, "connmgr_dis_and_wait: need " 2332 "T_DISCON_REQ for connection 0x%p\n", 2333 (void *)cm_entry); 2334 cm_entry->x_needdis = FALSE; 2335 cm_entry->x_waitdis = TRUE; 2336 2337 connmgr_snddis(cm_entry); 2338 2339 mutex_enter(&connmgr_lock); 2340 } 2341 2342 if (cm_entry->x_waitdis == TRUE) { 2343 clock_t curlbolt; 2344 clock_t timout; 2345 2346 RPCLOG(8, "connmgr_dis_and_wait waiting for " 2347 "T_DISCON_REQ's ACK for connection %p\n", 2348 (void *)cm_entry); 2349 curlbolt = ddi_get_lbolt(); 2350 2351 timout = clnt_cots_min_conntout * 2352 drv_usectohz(1000000) + curlbolt; 2353 2354 /* 2355 * The TPI spec says that the T_DISCON_REQ 2356 * will get acknowledged, but in practice 2357 * the ACK may never get sent. So don't 2358 * block forever. 2359 */ 2360 (void) cv_timedwait(&cm_entry->x_dis_cv, 2361 &connmgr_lock, timout); 2362 } 2363 /* 2364 * If we got the ACK, break. If we didn't, 2365 * then send another T_DISCON_REQ. 2366 */ 2367 if (cm_entry->x_waitdis == FALSE) { 2368 break; 2369 } else { 2370 RPCLOG(8, "connmgr_dis_and_wait: did" 2371 "not get T_DISCON_REQ's ACK for " 2372 "connection %p\n", (void *)cm_entry); 2373 cm_entry->x_needdis = TRUE; 2374 } 2375 } 2376 } 2377 2378 static void 2379 connmgr_cancelconn(struct cm_xprt *cm_entry) 2380 { 2381 /* 2382 * Mark the connection table entry as dead; the next thread that 2383 * goes through connmgr_release() will notice this and deal with it. 2384 */ 2385 mutex_enter(&connmgr_lock); 2386 cm_entry->x_dead = TRUE; 2387 2388 /* 2389 * Notify any threads waiting for the connection that it isn't 2390 * going to happen. 2391 */ 2392 cm_entry->x_thread = FALSE; 2393 cv_broadcast(&cm_entry->x_conn_cv); 2394 mutex_exit(&connmgr_lock); 2395 2396 connmgr_release(cm_entry); 2397 } 2398 2399 static void 2400 connmgr_close(struct cm_xprt *cm_entry) 2401 { 2402 mutex_enter(&cm_entry->x_lock); 2403 while (cm_entry->x_ref != 0) { 2404 /* 2405 * Must be a noninterruptible wait. 2406 */ 2407 cv_wait(&cm_entry->x_cv, &cm_entry->x_lock); 2408 } 2409 2410 if (cm_entry->x_tiptr != NULL) 2411 (void) t_kclose(cm_entry->x_tiptr, 1); 2412 2413 mutex_exit(&cm_entry->x_lock); 2414 if (cm_entry->x_ksp != NULL) { 2415 mutex_enter(&connmgr_lock); 2416 cm_entry->x_ksp->ks_private = NULL; 2417 mutex_exit(&connmgr_lock); 2418 2419 /* 2420 * Must free the buffer we allocated for the 2421 * server address in the update function 2422 */ 2423 if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))-> 2424 x_server.value.str.addr.ptr != NULL) 2425 kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp-> 2426 ks_data))->x_server.value.str.addr.ptr, 2427 INET6_ADDRSTRLEN); 2428 kmem_free(cm_entry->x_ksp->ks_data, 2429 cm_entry->x_ksp->ks_data_size); 2430 kstat_delete(cm_entry->x_ksp); 2431 } 2432 2433 mutex_destroy(&cm_entry->x_lock); 2434 cv_destroy(&cm_entry->x_cv); 2435 cv_destroy(&cm_entry->x_conn_cv); 2436 cv_destroy(&cm_entry->x_dis_cv); 2437 2438 if (cm_entry->x_server.buf != NULL) 2439 kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen); 2440 if (cm_entry->x_src.buf != NULL) 2441 kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen); 2442 kmem_free(cm_entry, sizeof (struct cm_xprt)); 2443 } 2444 2445 /* 2446 * Called by KRPC after sending the call message to release the connection 2447 * it was using. 2448 */ 2449 static void 2450 connmgr_release(struct cm_xprt *cm_entry) 2451 { 2452 mutex_enter(&cm_entry->x_lock); 2453 cm_entry->x_ref--; 2454 if (cm_entry->x_ref == 0) 2455 cv_signal(&cm_entry->x_cv); 2456 mutex_exit(&cm_entry->x_lock); 2457 } 2458 2459 /* 2460 * Given an open stream, connect to the remote. Returns true if connected, 2461 * false otherwise. 2462 */ 2463 static bool_t 2464 connmgr_connect( 2465 struct cm_xprt *cm_entry, 2466 queue_t *wq, 2467 struct netbuf *addr, 2468 int addrfmly, 2469 calllist_t *e, 2470 int *tidu_ptr, 2471 bool_t reconnect, 2472 const struct timeval *waitp, 2473 bool_t nosignal) 2474 { 2475 mblk_t *mp; 2476 struct T_conn_req *tcr; 2477 struct T_info_ack *tinfo; 2478 int interrupted, error; 2479 int tidu_size, kstat_instance; 2480 2481 /* if it's a reconnect, flush any lingering data messages */ 2482 if (reconnect) 2483 (void) putctl1(wq, M_FLUSH, FLUSHRW); 2484 2485 mp = allocb(sizeof (*tcr) + addr->len, BPRI_LO); 2486 if (mp == NULL) { 2487 /* 2488 * This is unfortunate, but we need to look up the stats for 2489 * this zone to increment the "memory allocation failed" 2490 * counter. curproc->p_zone is safe since we're initiating a 2491 * connection and not in some strange streams context. 2492 */ 2493 struct rpcstat *rpcstat; 2494 2495 rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone()); 2496 ASSERT(rpcstat != NULL); 2497 2498 RPCLOG0(1, "connmgr_connect: cannot alloc mp for " 2499 "sending conn request\n"); 2500 COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem); 2501 e->call_status = RPC_SYSTEMERROR; 2502 e->call_reason = ENOSR; 2503 return (FALSE); 2504 } 2505 2506 mp->b_datap->db_type = M_PROTO; 2507 tcr = (struct T_conn_req *)mp->b_rptr; 2508 bzero(tcr, sizeof (*tcr)); 2509 tcr->PRIM_type = T_CONN_REQ; 2510 tcr->DEST_length = addr->len; 2511 tcr->DEST_offset = sizeof (struct T_conn_req); 2512 mp->b_wptr = mp->b_rptr + sizeof (*tcr); 2513 2514 bcopy(addr->buf, mp->b_wptr, tcr->DEST_length); 2515 mp->b_wptr += tcr->DEST_length; 2516 2517 RPCLOG(8, "connmgr_connect: sending conn request on queue " 2518 "%p", (void *)wq); 2519 RPCLOG(8, " call %p\n", (void *)wq); 2520 /* 2521 * We use the entry in the handle that is normally used for 2522 * waiting for RPC replies to wait for the connection accept. 2523 */ 2524 clnt_dispatch_send(wq, mp, e, 0, 0); 2525 2526 mutex_enter(&clnt_pending_lock); 2527 2528 /* 2529 * We wait for the transport connection to be made, or an 2530 * indication that it could not be made. 2531 */ 2532 interrupted = 0; 2533 2534 /* 2535 * waitforack should have been called with T_OK_ACK, but the 2536 * present implementation needs to be passed T_INFO_ACK to 2537 * work correctly. 2538 */ 2539 error = waitforack(e, T_INFO_ACK, waitp, nosignal); 2540 if (error == EINTR) 2541 interrupted = 1; 2542 if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) { 2543 /* 2544 * No time to lose; we essentially have been signaled to 2545 * quit. 2546 */ 2547 interrupted = 1; 2548 } 2549 #ifdef RPCDEBUG 2550 if (error == ETIME) 2551 RPCLOG0(8, "connmgr_connect: giving up " 2552 "on connection attempt; " 2553 "clnt_dispatch notifyconn " 2554 "diagnostic 'no one waiting for " 2555 "connection' should not be " 2556 "unexpected\n"); 2557 #endif 2558 if (e->call_prev) 2559 e->call_prev->call_next = e->call_next; 2560 else 2561 clnt_pending = e->call_next; 2562 if (e->call_next) 2563 e->call_next->call_prev = e->call_prev; 2564 mutex_exit(&clnt_pending_lock); 2565 2566 if (e->call_status != RPC_SUCCESS || error != 0) { 2567 if (interrupted) 2568 e->call_status = RPC_INTR; 2569 else if (error == ETIME) 2570 e->call_status = RPC_TIMEDOUT; 2571 else if (error == EPROTO) 2572 e->call_status = RPC_SYSTEMERROR; 2573 2574 RPCLOG(8, "connmgr_connect: can't connect, status: " 2575 "%s\n", clnt_sperrno(e->call_status)); 2576 2577 if (e->call_reply) { 2578 freemsg(e->call_reply); 2579 e->call_reply = NULL; 2580 } 2581 2582 return (FALSE); 2583 } 2584 /* 2585 * The result of the "connection accept" is a T_info_ack 2586 * in the call_reply field. 2587 */ 2588 ASSERT(e->call_reply != NULL); 2589 mp = e->call_reply; 2590 e->call_reply = NULL; 2591 tinfo = (struct T_info_ack *)mp->b_rptr; 2592 2593 tidu_size = tinfo->TIDU_size; 2594 tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT); 2595 if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0)) 2596 tidu_size = COTS_DEFAULT_ALLOCSIZE; 2597 *tidu_ptr = tidu_size; 2598 2599 freemsg(mp); 2600 2601 /* 2602 * Set up the pertinent options. NODELAY is so the transport doesn't 2603 * buffer up RPC messages on either end. This may not be valid for 2604 * all transports. Failure to set this option is not cause to 2605 * bail out so we return success anyway. Note that lack of NODELAY 2606 * or some other way to flush the message on both ends will cause 2607 * lots of retries and terrible performance. 2608 */ 2609 if (addrfmly == AF_INET || addrfmly == AF_INET6) { 2610 (void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e); 2611 if (e->call_status == RPC_XPRTFAILED) 2612 return (FALSE); 2613 } 2614 2615 /* 2616 * Since we have a connection, we now need to figure out if 2617 * we need to create a kstat. If x_ksp is not NULL then we 2618 * are reusing a connection and so we do not need to create 2619 * another kstat -- lets just return. 2620 */ 2621 if (cm_entry->x_ksp != NULL) 2622 return (TRUE); 2623 2624 /* 2625 * We need to increment rpc_kstat_instance atomically to prevent 2626 * two kstats being created with the same instance. 2627 */ 2628 kstat_instance = atomic_add_32_nv((uint32_t *)&rpc_kstat_instance, 1); 2629 2630 if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance, 2631 "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED, 2632 (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)), 2633 KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) { 2634 return (TRUE); 2635 } 2636 2637 cm_entry->x_ksp->ks_lock = &connmgr_lock; 2638 cm_entry->x_ksp->ks_private = cm_entry; 2639 cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char)) 2640 + sizeof (cm_kstat_template)); 2641 cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size, 2642 KM_SLEEP); 2643 bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data, 2644 cm_entry->x_ksp->ks_data_size); 2645 ((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))-> 2646 x_server.value.str.addr.ptr = 2647 kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP); 2648 2649 cm_entry->x_ksp->ks_update = conn_kstat_update; 2650 kstat_install(cm_entry->x_ksp); 2651 return (TRUE); 2652 } 2653 2654 /* 2655 * Called by connmgr_connect to set an option on the new stream. 2656 */ 2657 static bool_t 2658 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e) 2659 { 2660 mblk_t *mp; 2661 struct opthdr *opt; 2662 struct T_optmgmt_req *tor; 2663 struct timeval waitp; 2664 int error; 2665 2666 mp = allocb(sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) + 2667 sizeof (int), BPRI_LO); 2668 if (mp == NULL) { 2669 RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option " 2670 "request\n"); 2671 return (FALSE); 2672 } 2673 2674 mp->b_datap->db_type = M_PROTO; 2675 tor = (struct T_optmgmt_req *)(mp->b_rptr); 2676 tor->PRIM_type = T_SVR4_OPTMGMT_REQ; 2677 tor->MGMT_flags = T_NEGOTIATE; 2678 tor->OPT_length = sizeof (struct opthdr) + sizeof (int); 2679 tor->OPT_offset = sizeof (struct T_optmgmt_req); 2680 2681 opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req)); 2682 opt->level = level; 2683 opt->name = name; 2684 opt->len = sizeof (int); 2685 *(int *)((char *)opt + sizeof (*opt)) = 1; 2686 mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) + 2687 sizeof (int); 2688 2689 /* 2690 * We will use this connection regardless 2691 * of whether or not the option is settable. 2692 */ 2693 clnt_dispatch_send(wq, mp, e, 0, 0); 2694 mutex_enter(&clnt_pending_lock); 2695 2696 waitp.tv_sec = clnt_cots_min_conntout; 2697 waitp.tv_usec = 0; 2698 error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1); 2699 2700 if (e->call_prev) 2701 e->call_prev->call_next = e->call_next; 2702 else 2703 clnt_pending = e->call_next; 2704 if (e->call_next) 2705 e->call_next->call_prev = e->call_prev; 2706 mutex_exit(&clnt_pending_lock); 2707 2708 if (e->call_reply != NULL) { 2709 freemsg(e->call_reply); 2710 e->call_reply = NULL; 2711 } 2712 2713 if (e->call_status != RPC_SUCCESS || error != 0) { 2714 RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name); 2715 return (FALSE); 2716 } 2717 RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name); 2718 return (TRUE); 2719 } 2720 2721 #ifdef DEBUG 2722 2723 /* 2724 * This is a knob to let us force code coverage in allocation failure 2725 * case. 2726 */ 2727 static int connmgr_failsnd; 2728 #define CONN_SND_ALLOC(Size, Pri) \ 2729 ((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri)) 2730 2731 #else 2732 2733 #define CONN_SND_ALLOC(Size, Pri) allocb(Size, Pri) 2734 2735 #endif 2736 2737 /* 2738 * Sends an orderly release on the specified queue. 2739 * Entered with connmgr_lock. Exited without connmgr_lock 2740 */ 2741 static void 2742 connmgr_sndrel(struct cm_xprt *cm_entry) 2743 { 2744 struct T_ordrel_req *torr; 2745 mblk_t *mp; 2746 queue_t *q = cm_entry->x_wq; 2747 ASSERT(MUTEX_HELD(&connmgr_lock)); 2748 mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO); 2749 if (mp == NULL) { 2750 cm_entry->x_needrel = TRUE; 2751 mutex_exit(&connmgr_lock); 2752 RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel " 2753 "to queue %p\n", (void *)q); 2754 return; 2755 } 2756 mutex_exit(&connmgr_lock); 2757 2758 mp->b_datap->db_type = M_PROTO; 2759 torr = (struct T_ordrel_req *)(mp->b_rptr); 2760 torr->PRIM_type = T_ORDREL_REQ; 2761 mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req); 2762 2763 RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q); 2764 put(q, mp); 2765 } 2766 2767 /* 2768 * Sends an disconnect on the specified queue. 2769 * Entered with connmgr_lock. Exited without connmgr_lock 2770 */ 2771 static void 2772 connmgr_snddis(struct cm_xprt *cm_entry) 2773 { 2774 struct T_discon_req *tdis; 2775 mblk_t *mp; 2776 queue_t *q = cm_entry->x_wq; 2777 2778 ASSERT(MUTEX_HELD(&connmgr_lock)); 2779 mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO); 2780 if (mp == NULL) { 2781 cm_entry->x_needdis = TRUE; 2782 mutex_exit(&connmgr_lock); 2783 RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon " 2784 "to queue %p\n", (void *)q); 2785 return; 2786 } 2787 mutex_exit(&connmgr_lock); 2788 2789 mp->b_datap->db_type = M_PROTO; 2790 tdis = (struct T_discon_req *)mp->b_rptr; 2791 tdis->PRIM_type = T_DISCON_REQ; 2792 mp->b_wptr = mp->b_rptr + sizeof (*tdis); 2793 2794 RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q); 2795 put(q, mp); 2796 } 2797 2798 /* 2799 * Sets up the entry for receiving replies, and calls rpcmod's write put proc 2800 * (through put) to send the call. 2801 */ 2802 static void 2803 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid, 2804 uint_t queue_flag) 2805 { 2806 ASSERT(e != NULL); 2807 2808 e->call_status = RPC_TIMEDOUT; /* optimistic, eh? */ 2809 e->call_reason = 0; 2810 e->call_wq = q; 2811 e->call_xid = xid; 2812 e->call_notified = FALSE; 2813 2814 /* 2815 * If queue_flag is set then the calllist_t is already on the hash 2816 * queue. In this case just send the message and return. 2817 */ 2818 if (queue_flag) { 2819 put(q, mp); 2820 return; 2821 } 2822 2823 /* 2824 * Set up calls for RPC requests (with XID != 0) on the hash 2825 * queue for fast lookups and place other calls (i.e. 2826 * connection management) on the linked list. 2827 */ 2828 if (xid != 0) { 2829 RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on " 2830 "dispatch list\n", xid); 2831 e->call_hash = call_hash(xid, clnt_cots_hash_size); 2832 e->call_bucket = &cots_call_ht[e->call_hash]; 2833 call_table_enter(e); 2834 } else { 2835 mutex_enter(&clnt_pending_lock); 2836 if (clnt_pending) 2837 clnt_pending->call_prev = e; 2838 e->call_next = clnt_pending; 2839 e->call_prev = NULL; 2840 clnt_pending = e; 2841 mutex_exit(&clnt_pending_lock); 2842 } 2843 2844 put(q, mp); 2845 } 2846 2847 /* 2848 * Called by rpcmod to notify a client with a clnt_pending call that its reply 2849 * has arrived. If we can't find a client waiting for this reply, we log 2850 * the error and return. 2851 */ 2852 bool_t 2853 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid) 2854 { 2855 calllist_t *e = NULL; 2856 call_table_t *chtp; 2857 uint32_t xid; 2858 uint_t hash; 2859 2860 if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) && 2861 (mp->b_wptr - mp->b_rptr) >= sizeof (xid)) 2862 xid = *((uint32_t *)mp->b_rptr); 2863 else { 2864 int i = 0; 2865 unsigned char *p = (unsigned char *)&xid; 2866 unsigned char *rptr; 2867 mblk_t *tmp = mp; 2868 2869 /* 2870 * Copy the xid, byte-by-byte into xid. 2871 */ 2872 while (tmp) { 2873 rptr = tmp->b_rptr; 2874 while (rptr < tmp->b_wptr) { 2875 *p++ = *rptr++; 2876 if (++i >= sizeof (xid)) 2877 goto done_xid_copy; 2878 } 2879 tmp = tmp->b_cont; 2880 } 2881 2882 /* 2883 * If we got here, we ran out of mblk space before the 2884 * xid could be copied. 2885 */ 2886 ASSERT(tmp == NULL && i < sizeof (xid)); 2887 2888 RPCLOG0(1, 2889 "clnt_dispatch_notify: message less than size of xid\n"); 2890 return (FALSE); 2891 2892 } 2893 done_xid_copy: 2894 2895 hash = call_hash(xid, clnt_cots_hash_size); 2896 chtp = &cots_call_ht[hash]; 2897 /* call_table_find returns with the hash bucket locked */ 2898 call_table_find(chtp, xid, e); 2899 2900 if (e != NULL) { 2901 /* 2902 * Found thread waiting for this reply 2903 */ 2904 mutex_enter(&e->call_lock); 2905 2906 /* 2907 * verify that the reply is coming in on 2908 * the same zone that it was sent from. 2909 */ 2910 if (e->call_zoneid != zoneid) { 2911 mutex_exit(&e->call_lock); 2912 mutex_exit(&chtp->ct_lock); 2913 return (FALSE); 2914 } 2915 2916 if (e->call_reply) 2917 /* 2918 * This can happen under the following scenario: 2919 * clnt_cots_kcallit() times out on the response, 2920 * rfscall() repeats the CLNT_CALL() with 2921 * the same xid, clnt_cots_kcallit() sends the retry, 2922 * thereby putting the clnt handle on the pending list, 2923 * the first response arrives, signalling the thread 2924 * in clnt_cots_kcallit(). Before that thread is 2925 * dispatched, the second response arrives as well, 2926 * and clnt_dispatch_notify still finds the handle on 2927 * the pending list, with call_reply set. So free the 2928 * old reply now. 2929 * 2930 * It is also possible for a response intended for 2931 * an RPC call with a different xid to reside here. 2932 * This can happen if the thread that owned this 2933 * client handle prior to the current owner bailed 2934 * out and left its call record on the dispatch 2935 * queue. A window exists where the response can 2936 * arrive before the current owner dispatches its 2937 * RPC call. 2938 * 2939 * In any case, this is the very last point where we 2940 * can safely check the call_reply field before 2941 * placing the new response there. 2942 */ 2943 freemsg(e->call_reply); 2944 e->call_reply = mp; 2945 e->call_status = RPC_SUCCESS; 2946 e->call_notified = TRUE; 2947 cv_signal(&e->call_cv); 2948 mutex_exit(&e->call_lock); 2949 mutex_exit(&chtp->ct_lock); 2950 return (TRUE); 2951 } else { 2952 zone_t *zone; 2953 struct rpcstat *rpcstat; 2954 2955 mutex_exit(&chtp->ct_lock); 2956 RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n", 2957 xid); 2958 /* 2959 * This is unfortunate, but we need to lookup the zone so we 2960 * can increment its "rcbadxids" counter. 2961 */ 2962 zone = zone_find_by_id(zoneid); 2963 if (zone == NULL) { 2964 /* 2965 * The zone went away... 2966 */ 2967 return (FALSE); 2968 } 2969 rpcstat = zone_getspecific(rpcstat_zone_key, zone); 2970 if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) { 2971 /* 2972 * Not interested 2973 */ 2974 zone_rele(zone); 2975 return (FALSE); 2976 } 2977 COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids); 2978 zone_rele(zone); 2979 } 2980 return (FALSE); 2981 } 2982 2983 /* 2984 * Called by rpcmod when a non-data indication arrives. The ones in which we 2985 * are interested are connection indications and options acks. We dispatch 2986 * based on the queue the indication came in on. If we are not interested in 2987 * what came in, we return false to rpcmod, who will then pass it upstream. 2988 */ 2989 bool_t 2990 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp) 2991 { 2992 calllist_t *e; 2993 int type; 2994 2995 ASSERT((q->q_flag & QREADR) == 0); 2996 2997 type = ((union T_primitives *)mp->b_rptr)->type; 2998 RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n", 2999 rpc_tpiprim2name(type)); 3000 mutex_enter(&clnt_pending_lock); 3001 for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) { 3002 if (e == NULL) { 3003 mutex_exit(&clnt_pending_lock); 3004 RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting " 3005 "for connection on queue 0x%p\n", (void *)q); 3006 return (FALSE); 3007 } 3008 if (e->call_wq == q) 3009 break; 3010 } 3011 3012 switch (type) { 3013 case T_CONN_CON: 3014 /* 3015 * The transport is now connected, send a T_INFO_REQ to get 3016 * the tidu size. 3017 */ 3018 mutex_exit(&clnt_pending_lock); 3019 ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >= 3020 sizeof (struct T_info_req)); 3021 mp->b_rptr = mp->b_datap->db_base; 3022 ((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ; 3023 mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req); 3024 mp->b_datap->db_type = M_PCPROTO; 3025 put(q, mp); 3026 return (TRUE); 3027 case T_INFO_ACK: 3028 case T_OPTMGMT_ACK: 3029 e->call_status = RPC_SUCCESS; 3030 e->call_reply = mp; 3031 e->call_notified = TRUE; 3032 cv_signal(&e->call_cv); 3033 break; 3034 case T_ERROR_ACK: 3035 e->call_status = RPC_CANTCONNECT; 3036 e->call_reply = mp; 3037 e->call_notified = TRUE; 3038 cv_signal(&e->call_cv); 3039 break; 3040 case T_OK_ACK: 3041 /* 3042 * Great, but we are really waiting for a T_CONN_CON 3043 */ 3044 freemsg(mp); 3045 break; 3046 default: 3047 mutex_exit(&clnt_pending_lock); 3048 RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type); 3049 return (FALSE); 3050 } 3051 3052 mutex_exit(&clnt_pending_lock); 3053 return (TRUE); 3054 } 3055 3056 /* 3057 * Called by rpcmod when the transport is (or should be) going away. Informs 3058 * all callers waiting for replies and marks the entry in the connection 3059 * manager's list as unconnected, and either closing (close handshake in 3060 * progress) or dead. 3061 */ 3062 void 3063 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason) 3064 { 3065 calllist_t *e; 3066 call_table_t *ctp; 3067 struct cm_xprt *cm_entry; 3068 int have_connmgr_lock; 3069 int i; 3070 3071 ASSERT((q->q_flag & QREADR) == 0); 3072 3073 RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q); 3074 RPCLOG(1, " received a notifcation prim type [%s]", 3075 rpc_tpiprim2name(msg_type)); 3076 RPCLOG(1, " and reason %d\n", reason); 3077 3078 /* 3079 * Find the transport entry in the connection manager's list, close 3080 * the transport and delete the entry. In the case where rpcmod's 3081 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we 3082 * should gracefully close the connection. 3083 */ 3084 have_connmgr_lock = 1; 3085 mutex_enter(&connmgr_lock); 3086 for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) { 3087 ASSERT(cm_entry != cm_entry->x_next); 3088 if (cm_entry->x_wq == q) { 3089 ASSERT(MUTEX_HELD(&connmgr_lock)); 3090 ASSERT(have_connmgr_lock == 1); 3091 switch (msg_type) { 3092 case T_ORDREL_REQ: 3093 3094 if (cm_entry->x_dead) { 3095 RPCLOG(1, "idle timeout on dead " 3096 "connection: %p\n", 3097 (void *)cm_entry); 3098 if (clnt_stop_idle != NULL) 3099 (*clnt_stop_idle)(q); 3100 break; 3101 } 3102 3103 /* 3104 * Only mark the connection as dead if it is 3105 * connected and idle. 3106 * An unconnected connection has probably 3107 * gone idle because the server is down, 3108 * and when it comes back up there will be 3109 * retries that need to use that connection. 3110 */ 3111 if (cm_entry->x_connected || 3112 cm_entry->x_doomed) { 3113 if (cm_entry->x_ordrel) { 3114 if (cm_entry->x_closing == 3115 TRUE) { 3116 /* 3117 * The connection is 3118 * obviously wedged due 3119 * to a bug or problem 3120 * with the transport. 3121 * Mark it as dead. 3122 * Otherwise we can 3123 * leak connections. 3124 */ 3125 cm_entry->x_dead = TRUE; 3126 mutex_exit( 3127 &connmgr_lock); 3128 have_connmgr_lock = 0; 3129 if (clnt_stop_idle != 3130 NULL) 3131 (*clnt_stop_idle)(q); 3132 break; 3133 } 3134 cm_entry->x_closing = TRUE; 3135 connmgr_sndrel(cm_entry); 3136 have_connmgr_lock = 0; 3137 } else { 3138 cm_entry->x_dead = TRUE; 3139 mutex_exit(&connmgr_lock); 3140 have_connmgr_lock = 0; 3141 if (clnt_stop_idle != NULL) 3142 (*clnt_stop_idle)(q); 3143 } 3144 } else { 3145 /* 3146 * We don't mark the connection 3147 * as dead, but we turn off the 3148 * idle timer. 3149 */ 3150 mutex_exit(&connmgr_lock); 3151 have_connmgr_lock = 0; 3152 if (clnt_stop_idle != NULL) 3153 (*clnt_stop_idle)(q); 3154 RPCLOG(1, "clnt_dispatch_notifyall:" 3155 " ignoring timeout from rpcmod" 3156 " (q %p) because we are not " 3157 " connected\n", (void *)q); 3158 } 3159 break; 3160 case T_ORDREL_IND: 3161 /* 3162 * If this entry is marked closing, then we are 3163 * completing a close handshake, and the 3164 * connection is dead. Otherwise, the server is 3165 * trying to close. Since the server will not 3166 * be sending any more RPC replies, we abort 3167 * the connection, including flushing 3168 * any RPC requests that are in-transit. 3169 */ 3170 if (cm_entry->x_closing) { 3171 cm_entry->x_dead = TRUE; 3172 mutex_exit(&connmgr_lock); 3173 have_connmgr_lock = 0; 3174 if (clnt_stop_idle != NULL) 3175 (*clnt_stop_idle)(q); 3176 } else { 3177 /* 3178 * if we're getting a disconnect 3179 * before we've finished our 3180 * connect attempt, mark it for 3181 * later processing 3182 */ 3183 if (cm_entry->x_thread) 3184 cm_entry->x_early_disc = TRUE; 3185 else 3186 cm_entry->x_connected = FALSE; 3187 cm_entry->x_waitdis = TRUE; 3188 connmgr_snddis(cm_entry); 3189 have_connmgr_lock = 0; 3190 } 3191 break; 3192 3193 case T_ERROR_ACK: 3194 case T_OK_ACK: 3195 cm_entry->x_waitdis = FALSE; 3196 cv_signal(&cm_entry->x_dis_cv); 3197 mutex_exit(&connmgr_lock); 3198 return; 3199 3200 case T_DISCON_REQ: 3201 if (cm_entry->x_thread) 3202 cm_entry->x_early_disc = TRUE; 3203 else 3204 cm_entry->x_connected = FALSE; 3205 cm_entry->x_waitdis = TRUE; 3206 3207 connmgr_snddis(cm_entry); 3208 have_connmgr_lock = 0; 3209 break; 3210 3211 case T_DISCON_IND: 3212 default: 3213 /* 3214 * if we're getting a disconnect before 3215 * we've finished our connect attempt, 3216 * mark it for later processing 3217 */ 3218 if (cm_entry->x_closing) { 3219 cm_entry->x_dead = TRUE; 3220 mutex_exit(&connmgr_lock); 3221 have_connmgr_lock = 0; 3222 if (clnt_stop_idle != NULL) 3223 (*clnt_stop_idle)(q); 3224 } else { 3225 if (cm_entry->x_thread) { 3226 cm_entry->x_early_disc = TRUE; 3227 } else { 3228 cm_entry->x_dead = TRUE; 3229 cm_entry->x_connected = FALSE; 3230 } 3231 } 3232 break; 3233 } 3234 break; 3235 } 3236 } 3237 3238 if (have_connmgr_lock) 3239 mutex_exit(&connmgr_lock); 3240 3241 if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) { 3242 RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find " 3243 "connmgr entry for discon ack\n", (void *)q); 3244 return; 3245 } 3246 3247 /* 3248 * Then kick all the clnt_pending calls out of their wait. There 3249 * should be no clnt_pending calls in the case of rpcmod's idle 3250 * timer firing. 3251 */ 3252 for (i = 0; i < clnt_cots_hash_size; i++) { 3253 ctp = &cots_call_ht[i]; 3254 mutex_enter(&ctp->ct_lock); 3255 for (e = ctp->ct_call_next; 3256 e != (calllist_t *)ctp; 3257 e = e->call_next) { 3258 if (e->call_wq == q && e->call_notified == FALSE) { 3259 RPCLOG(1, 3260 "clnt_dispatch_notifyall for queue %p ", 3261 (void *)q); 3262 RPCLOG(1, "aborting clnt_pending call %p\n", 3263 (void *)e); 3264 3265 if (msg_type == T_DISCON_IND) 3266 e->call_reason = reason; 3267 e->call_notified = TRUE; 3268 e->call_status = RPC_XPRTFAILED; 3269 cv_signal(&e->call_cv); 3270 } 3271 } 3272 mutex_exit(&ctp->ct_lock); 3273 } 3274 3275 mutex_enter(&clnt_pending_lock); 3276 for (e = clnt_pending; e; e = e->call_next) { 3277 /* 3278 * Only signal those RPC handles that haven't been 3279 * signalled yet. Otherwise we can get a bogus call_reason. 3280 * This can happen if thread A is making a call over a 3281 * connection. If the server is killed, it will cause 3282 * reset, and reason will default to EIO as a result of 3283 * a T_ORDREL_IND. Thread B then attempts to recreate 3284 * the connection but gets a T_DISCON_IND. If we set the 3285 * call_reason code for all threads, then if thread A 3286 * hasn't been dispatched yet, it will get the wrong 3287 * reason. The bogus call_reason can make it harder to 3288 * discriminate between calls that fail because the 3289 * connection attempt failed versus those where the call 3290 * may have been executed on the server. 3291 */ 3292 if (e->call_wq == q && e->call_notified == FALSE) { 3293 RPCLOG(1, "clnt_dispatch_notifyall for queue %p ", 3294 (void *)q); 3295 RPCLOG(1, " aborting clnt_pending call %p\n", 3296 (void *)e); 3297 3298 if (msg_type == T_DISCON_IND) 3299 e->call_reason = reason; 3300 e->call_notified = TRUE; 3301 /* 3302 * Let the caller timeout, else he will retry 3303 * immediately. 3304 */ 3305 e->call_status = RPC_XPRTFAILED; 3306 3307 /* 3308 * We used to just signal those threads 3309 * waiting for a connection, (call_xid = 0). 3310 * That meant that threads waiting for a response 3311 * waited till their timeout expired. This 3312 * could be a long time if they've specified a 3313 * maximum timeout. (2^31 - 1). So we 3314 * Signal all threads now. 3315 */ 3316 cv_signal(&e->call_cv); 3317 } 3318 } 3319 mutex_exit(&clnt_pending_lock); 3320 } 3321 3322 3323 /*ARGSUSED*/ 3324 /* 3325 * after resuming a system that's been suspended for longer than the 3326 * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall() 3327 * generates "NFS server X not responding" and "NFS server X ok" messages; 3328 * here we reset inet connections to cause a re-connect and avoid those 3329 * NFS messages. see 4045054 3330 */ 3331 boolean_t 3332 connmgr_cpr_reset(void *arg, int code) 3333 { 3334 struct cm_xprt *cxp; 3335 3336 if (code == CB_CODE_CPR_CHKPT) 3337 return (B_TRUE); 3338 3339 if (mutex_tryenter(&connmgr_lock) == 0) 3340 return (B_FALSE); 3341 for (cxp = cm_hd; cxp; cxp = cxp->x_next) { 3342 if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) && 3343 cxp->x_connected == TRUE) { 3344 if (cxp->x_thread) 3345 cxp->x_early_disc = TRUE; 3346 else 3347 cxp->x_connected = FALSE; 3348 cxp->x_needdis = TRUE; 3349 } 3350 } 3351 mutex_exit(&connmgr_lock); 3352 return (B_TRUE); 3353 } 3354 3355 void 3356 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp) 3357 { 3358 3359 *statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid, 3360 "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl, 3361 sizeof (cots_rcstat_tmpl)); 3362 } 3363 3364 void 3365 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp) 3366 { 3367 rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client"); 3368 kmem_free(*statsp, sizeof (cots_rcstat_tmpl)); 3369 } 3370 3371 void 3372 clnt_cots_init(void) 3373 { 3374 mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL); 3375 mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL); 3376 3377 if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE) 3378 clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE; 3379 3380 cots_call_ht = call_table_init(clnt_cots_hash_size); 3381 zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy); 3382 } 3383 3384 void 3385 clnt_cots_fini(void) 3386 { 3387 (void) zone_key_delete(zone_cots_key); 3388 } 3389 3390 /* 3391 * Wait for TPI ack, returns success only if expected ack is received 3392 * within timeout period. 3393 */ 3394 3395 static int 3396 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp, 3397 bool_t nosignal) 3398 { 3399 union T_primitives *tpr; 3400 clock_t timout; 3401 int cv_stat = 1; 3402 3403 ASSERT(MUTEX_HELD(&clnt_pending_lock)); 3404 while (e->call_reply == NULL) { 3405 if (waitp != NULL) { 3406 timout = waitp->tv_sec * drv_usectohz(MICROSEC) + 3407 drv_usectohz(waitp->tv_usec) + lbolt; 3408 if (nosignal) 3409 cv_stat = cv_timedwait(&e->call_cv, 3410 &clnt_pending_lock, timout); 3411 else 3412 cv_stat = cv_timedwait_sig(&e->call_cv, 3413 &clnt_pending_lock, timout); 3414 } else { 3415 if (nosignal) 3416 cv_wait(&e->call_cv, &clnt_pending_lock); 3417 else 3418 cv_stat = cv_wait_sig(&e->call_cv, 3419 &clnt_pending_lock); 3420 } 3421 if (cv_stat == -1) 3422 return (ETIME); 3423 if (cv_stat == 0) 3424 return (EINTR); 3425 /* 3426 * if we received an error from the server and we know a reply 3427 * is not going to be sent, do not wait for the full timeout, 3428 * return now. 3429 */ 3430 if (e->call_status == RPC_XPRTFAILED) 3431 return (e->call_reason); 3432 } 3433 tpr = (union T_primitives *)e->call_reply->b_rptr; 3434 if (tpr->type == ack_prim) 3435 return (0); /* Success */ 3436 3437 if (tpr->type == T_ERROR_ACK) { 3438 if (tpr->error_ack.TLI_error == TSYSERR) 3439 return (tpr->error_ack.UNIX_error); 3440 else 3441 return (t_tlitosyserr(tpr->error_ack.TLI_error)); 3442 } 3443 3444 return (EPROTO); /* unknown or unexpected primitive */ 3445 } 3446