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, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 /* 28 * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T 29 * All Rights Reserved 30 */ 31 32 /* 33 * Portions of this source code were derived from Berkeley 4.3 BSD 34 * under license from the Regents of the University of California. 35 */ 36 37 #pragma ident "%Z%%M% %I% %E% SMI" 38 39 /* 40 * Implements a kernel based, client side RPC over Connection Oriented 41 * Transports (COTS). 42 */ 43 44 /* 45 * Much of this file has been re-written to let NFS work better over slow 46 * transports. A description follows. 47 * 48 * One of the annoying things about kRPC/COTS is that it will temporarily 49 * create more than one connection between a client and server. This 50 * happens because when a connection is made, the end-points entry in the 51 * linked list of connections (headed by cm_hd), is removed so that other 52 * threads don't mess with it. Went ahead and bit the bullet by keeping 53 * the endpoint on the connection list and introducing state bits, 54 * condition variables etc. to the connection entry data structure (struct 55 * cm_xprt). 56 * 57 * Here is a summary of the changes to cm-xprt: 58 * 59 * x_ctime is the timestamp of when the endpoint was last 60 * connected or disconnected. If an end-point is ever disconnected 61 * or re-connected, then any outstanding RPC request is presumed 62 * lost, telling clnt_cots_kcallit that it needs to re-send the 63 * request, not just wait for the original request's reply to 64 * arrive. 65 * 66 * x_thread flag which tells us if a thread is doing a connection attempt. 67 * 68 * x_waitdis flag which tells us we are waiting a disconnect ACK. 69 * 70 * x_needdis flag which tells us we need to send a T_DISCONN_REQ 71 * to kill the connection. 72 * 73 * x_needrel flag which tells us we need to send a T_ORDREL_REQ to 74 * gracefully close the connection. 75 * 76 * #defined bitmasks for the all the b_* bits so that more 77 * efficient (and at times less clumsy) masks can be used to 78 * manipulated state in cases where multiple bits have to 79 * set/cleared/checked in the same critical section. 80 * 81 * x_conn_cv and x_dis-_cv are new condition variables to let 82 * threads knows when the connection attempt is done, and to let 83 * the connecting thread know when the disconnect handshake is 84 * done. 85 * 86 * Added the CONN_HOLD() macro so that all reference holds have the same 87 * look and feel. 88 * 89 * In the private (cku_private) portion of the client handle, 90 * 91 * cku_flags replaces the cku_sent a boolean. cku_flags keeps 92 * track of whether a request as been sent, and whether the 93 * client's handles call record is on the dispatch list (so that 94 * the reply can be matched by XID to the right client handle). 95 * The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit() 96 * and still have the response find the right client handle so 97 * that the retry of CLNT_CALL() gets the result. Testing, found 98 * situations where if the timeout was increased, performance 99 * degraded. This was due to us hitting a window where the thread 100 * was back in rfscall() (probably printing server not responding) 101 * while the response came back but no place to put it. 102 * 103 * cku_ctime is just a cache of x_ctime. If they match, 104 * clnt_cots_kcallit() won't to send a retry (unless the maximum 105 * receive count limit as been reached). If the don't match, then 106 * we assume the request has been lost, and a retry of the request 107 * is needed. 108 * 109 * cku_recv_attempts counts the number of receive count attempts 110 * after one try is sent on the wire. 111 * 112 * Added the clnt_delay() routine so that interruptible and 113 * noninterruptible delays are possible. 114 * 115 * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to 116 * control how long the client delays before returned after getting 117 * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash 118 * a server that may be booting and not yet started nfsd. 119 * 120 * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable) 121 * Why don't we just wait forever (receive an infinite # of times)? 122 * Because the server may have rebooted. More insidious is that some 123 * servers (ours) will drop NFS/TCP requests in some cases. This is bad, 124 * but it is a reality. 125 * 126 * The case of a server doing orderly release really messes up the 127 * client's recovery, especially if the server's TCP implementation is 128 * buggy. It was found was that the kRPC/COTS client was breaking some 129 * TPI rules, such as not waiting for the acknowledgement of a 130 * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and 131 * T_DISCON_REQ in clnt_dispatch_notifyall()). 132 * 133 * One of things that we've seen is that a kRPC TCP endpoint goes into 134 * TIMEWAIT and a thus a reconnect takes a long time to satisfy because 135 * that the TIMEWAIT state takes a while to finish. If a server sends a 136 * T_ORDREL_IND, there is little point in an RPC client doing a 137 * T_ORDREL_REQ, because the RPC request isn't going to make it (the 138 * server is saying that it won't accept any more data). So kRPC was 139 * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the 140 * connection skips the TIMEWAIT state and goes straight to a bound state 141 * that kRPC can quickly switch to connected. 142 * 143 * Code that issues TPI request must use waitforack() to wait for the 144 * corresponding ack (assuming there is one) in any future modifications. 145 * This works around problems that may be introduced by breaking TPI rules 146 * (by submitting new calls before earlier requests have been acked) in the 147 * case of a signal or other early return. waitforack() depends on 148 * clnt_dispatch_notifyconn() to issue the wakeup when the ack 149 * arrives, so adding new TPI calls may require corresponding changes 150 * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on 151 * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure 152 * not to set it too low or TPI ACKS will be lost. 153 */ 154 155 #include <sys/param.h> 156 #include <sys/types.h> 157 #include <sys/user.h> 158 #include <sys/systm.h> 159 #include <sys/sysmacros.h> 160 #include <sys/proc.h> 161 #include <sys/socket.h> 162 #include <sys/file.h> 163 #include <sys/stream.h> 164 #include <sys/strsubr.h> 165 #include <sys/stropts.h> 166 #include <sys/strsun.h> 167 #include <sys/timod.h> 168 #include <sys/tiuser.h> 169 #include <sys/tihdr.h> 170 #include <sys/t_kuser.h> 171 #include <sys/fcntl.h> 172 #include <sys/errno.h> 173 #include <sys/kmem.h> 174 #include <sys/debug.h> 175 #include <sys/systm.h> 176 #include <sys/kstat.h> 177 #include <sys/t_lock.h> 178 #include <sys/ddi.h> 179 #include <sys/cmn_err.h> 180 #include <sys/time.h> 181 #include <sys/isa_defs.h> 182 #include <sys/callb.h> 183 #include <sys/sunddi.h> 184 #include <sys/atomic.h> 185 186 #include <netinet/in.h> 187 #include <netinet/tcp.h> 188 189 #include <rpc/types.h> 190 #include <rpc/xdr.h> 191 #include <rpc/auth.h> 192 #include <rpc/clnt.h> 193 #include <rpc/rpc_msg.h> 194 195 #define COTS_DEFAULT_ALLOCSIZE 2048 196 197 #define WIRE_HDR_SIZE 20 /* serialized call header, sans proc number */ 198 #define MSG_OFFSET 128 /* offset of call into the mblk */ 199 200 const char *kinet_ntop6(uchar_t *, char *, size_t); 201 202 static int clnt_cots_ksettimers(CLIENT *, struct rpc_timers *, 203 struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t); 204 static enum clnt_stat clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t, 205 caddr_t, xdrproc_t, caddr_t, struct timeval); 206 static void clnt_cots_kabort(CLIENT *); 207 static void clnt_cots_kerror(CLIENT *, struct rpc_err *); 208 static bool_t clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t); 209 static void clnt_cots_kdestroy(CLIENT *); 210 static bool_t clnt_cots_kcontrol(CLIENT *, int, char *); 211 212 213 /* List of transports managed by the connection manager. */ 214 struct cm_xprt { 215 TIUSER *x_tiptr; /* transport handle */ 216 queue_t *x_wq; /* send queue */ 217 clock_t x_time; /* last time we handed this xprt out */ 218 clock_t x_ctime; /* time we went to CONNECTED */ 219 int x_tidu_size; /* TIDU size of this transport */ 220 union { 221 struct { 222 unsigned int 223 #ifdef _BIT_FIELDS_HTOL 224 b_closing: 1, /* we've sent a ord rel on this conn */ 225 b_dead: 1, /* transport is closed or disconn */ 226 b_doomed: 1, /* too many conns, let this go idle */ 227 b_connected: 1, /* this connection is connected */ 228 229 b_ordrel: 1, /* do an orderly release? */ 230 b_thread: 1, /* thread doing connect */ 231 b_waitdis: 1, /* waiting for disconnect ACK */ 232 b_needdis: 1, /* need T_DISCON_REQ */ 233 234 b_needrel: 1, /* need T_ORDREL_REQ */ 235 b_early_disc: 1, /* got a T_ORDREL_IND or T_DISCON_IND */ 236 /* disconnect during connect */ 237 238 b_pad: 22; 239 240 #endif 241 242 #ifdef _BIT_FIELDS_LTOH 243 b_pad: 22, 244 245 b_early_disc: 1, /* got a T_ORDREL_IND or T_DISCON_IND */ 246 /* disconnect during connect */ 247 b_needrel: 1, /* need T_ORDREL_REQ */ 248 249 b_needdis: 1, /* need T_DISCON_REQ */ 250 b_waitdis: 1, /* waiting for disconnect ACK */ 251 b_thread: 1, /* thread doing connect */ 252 b_ordrel: 1, /* do an orderly release? */ 253 254 b_connected: 1, /* this connection is connected */ 255 b_doomed: 1, /* too many conns, let this go idle */ 256 b_dead: 1, /* transport is closed or disconn */ 257 b_closing: 1; /* we've sent a ord rel on this conn */ 258 #endif 259 } bit; unsigned int word; 260 261 #define x_closing x_state.bit.b_closing 262 #define x_dead x_state.bit.b_dead 263 #define x_doomed x_state.bit.b_doomed 264 #define x_connected x_state.bit.b_connected 265 266 #define x_ordrel x_state.bit.b_ordrel 267 #define x_thread x_state.bit.b_thread 268 #define x_waitdis x_state.bit.b_waitdis 269 #define x_needdis x_state.bit.b_needdis 270 271 #define x_needrel x_state.bit.b_needrel 272 #define x_early_disc x_state.bit.b_early_disc 273 274 #define x_state_flags x_state.word 275 276 #define X_CLOSING 0x80000000 277 #define X_DEAD 0x40000000 278 #define X_DOOMED 0x20000000 279 #define X_CONNECTED 0x10000000 280 281 #define X_ORDREL 0x08000000 282 #define X_THREAD 0x04000000 283 #define X_WAITDIS 0x02000000 284 #define X_NEEDDIS 0x01000000 285 286 #define X_NEEDREL 0x00800000 287 #define X_EARLYDISC 0x00400000 288 289 #define X_BADSTATES (X_CLOSING | X_DEAD | X_DOOMED) 290 291 } x_state; 292 int x_ref; /* number of users of this xprt */ 293 int x_family; /* address family of transport */ 294 dev_t x_rdev; /* device number of transport */ 295 struct cm_xprt *x_next; 296 297 struct netbuf x_server; /* destination address */ 298 struct netbuf x_src; /* src address (for retries) */ 299 kmutex_t x_lock; /* lock on this entry */ 300 kcondvar_t x_cv; /* to signal when can be closed */ 301 kcondvar_t x_conn_cv; /* to signal when connection attempt */ 302 /* is complete */ 303 kstat_t *x_ksp; 304 305 kcondvar_t x_dis_cv; /* to signal when disconnect attempt */ 306 /* is complete */ 307 zoneid_t x_zoneid; /* zone this xprt belongs to */ 308 }; 309 310 typedef struct cm_kstat_xprt { 311 kstat_named_t x_wq; 312 kstat_named_t x_server; 313 kstat_named_t x_family; 314 kstat_named_t x_rdev; 315 kstat_named_t x_time; 316 kstat_named_t x_state; 317 kstat_named_t x_ref; 318 kstat_named_t x_port; 319 } cm_kstat_xprt_t; 320 321 static cm_kstat_xprt_t cm_kstat_template = { 322 { "write_queue", KSTAT_DATA_UINT32 }, 323 { "server", KSTAT_DATA_STRING }, 324 { "addr_family", KSTAT_DATA_UINT32 }, 325 { "device", KSTAT_DATA_UINT32 }, 326 { "time_stamp", KSTAT_DATA_UINT32 }, 327 { "status", KSTAT_DATA_UINT32 }, 328 { "ref_count", KSTAT_DATA_INT32 }, 329 { "port", KSTAT_DATA_UINT32 }, 330 }; 331 332 /* 333 * The inverse of this is connmgr_release(). 334 */ 335 #define CONN_HOLD(Cm_entry) {\ 336 mutex_enter(&(Cm_entry)->x_lock); \ 337 (Cm_entry)->x_ref++; \ 338 mutex_exit(&(Cm_entry)->x_lock); \ 339 } 340 341 342 /* 343 * Private data per rpc handle. This structure is allocated by 344 * clnt_cots_kcreate, and freed by clnt_cots_kdestroy. 345 */ 346 typedef struct cku_private_s { 347 CLIENT cku_client; /* client handle */ 348 calllist_t cku_call; /* for dispatching calls */ 349 struct rpc_err cku_err; /* error status */ 350 351 struct netbuf cku_srcaddr; /* source address for retries */ 352 int cku_addrfmly; /* for binding port */ 353 struct netbuf cku_addr; /* remote address */ 354 dev_t cku_device; /* device to use */ 355 uint_t cku_flags; 356 #define CKU_ONQUEUE 0x1 357 #define CKU_SENT 0x2 358 359 bool_t cku_progress; /* for CLSET_PROGRESS */ 360 uint32_t cku_xid; /* current XID */ 361 clock_t cku_ctime; /* time stamp of when */ 362 /* connection was created */ 363 uint_t cku_recv_attempts; 364 XDR cku_outxdr; /* xdr routine for output */ 365 XDR cku_inxdr; /* xdr routine for input */ 366 char cku_rpchdr[WIRE_HDR_SIZE + 4]; 367 /* pre-serialized rpc header */ 368 369 uint_t cku_outbuflen; /* default output mblk length */ 370 struct cred *cku_cred; /* credentials */ 371 bool_t cku_nodelayonerr; 372 /* for CLSET_NODELAYONERR */ 373 int cku_useresvport; /* Use reserved port */ 374 struct rpc_cots_client *cku_stats; /* stats for zone */ 375 } cku_private_t; 376 377 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *, 378 const struct timeval *, struct netbuf *, int, struct netbuf *, 379 struct rpc_err *, bool_t, bool_t); 380 381 static bool_t connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *, 382 int, calllist_t *, int *, bool_t reconnect, 383 const struct timeval *, bool_t); 384 385 static bool_t connmgr_setopt(queue_t *, int, int, calllist_t *); 386 static void connmgr_sndrel(struct cm_xprt *); 387 static void connmgr_snddis(struct cm_xprt *); 388 static void connmgr_close(struct cm_xprt *); 389 static void connmgr_release(struct cm_xprt *); 390 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *, 391 cku_private_t *); 392 393 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *, 394 struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t, 395 bool_t, int); 396 397 static void connmgr_cancelconn(struct cm_xprt *); 398 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *, 399 bool_t); 400 static void connmgr_dis_and_wait(struct cm_xprt *); 401 402 static void clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t, 403 uint_t); 404 405 static int clnt_delay(clock_t, bool_t); 406 407 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t); 408 409 /* 410 * Operations vector for TCP/IP based RPC 411 */ 412 static struct clnt_ops tcp_ops = { 413 clnt_cots_kcallit, /* do rpc call */ 414 clnt_cots_kabort, /* abort call */ 415 clnt_cots_kerror, /* return error status */ 416 clnt_cots_kfreeres, /* free results */ 417 clnt_cots_kdestroy, /* destroy rpc handle */ 418 clnt_cots_kcontrol, /* the ioctl() of rpc */ 419 clnt_cots_ksettimers, /* set retry timers */ 420 }; 421 422 static int rpc_kstat_instance = 0; /* keeps the current instance */ 423 /* number for the next kstat_create */ 424 425 static struct cm_xprt *cm_hd = NULL; 426 static kmutex_t connmgr_lock; /* for connection mngr's list of transports */ 427 428 extern kmutex_t clnt_max_msg_lock; 429 430 static calllist_t *clnt_pending = NULL; 431 extern kmutex_t clnt_pending_lock; 432 433 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE; 434 435 static call_table_t *cots_call_ht; 436 437 static const struct rpc_cots_client { 438 kstat_named_t rccalls; 439 kstat_named_t rcbadcalls; 440 kstat_named_t rcbadxids; 441 kstat_named_t rctimeouts; 442 kstat_named_t rcnewcreds; 443 kstat_named_t rcbadverfs; 444 kstat_named_t rctimers; 445 kstat_named_t rccantconn; 446 kstat_named_t rcnomem; 447 kstat_named_t rcintrs; 448 } cots_rcstat_tmpl = { 449 { "calls", KSTAT_DATA_UINT64 }, 450 { "badcalls", KSTAT_DATA_UINT64 }, 451 { "badxids", KSTAT_DATA_UINT64 }, 452 { "timeouts", KSTAT_DATA_UINT64 }, 453 { "newcreds", KSTAT_DATA_UINT64 }, 454 { "badverfs", KSTAT_DATA_UINT64 }, 455 { "timers", KSTAT_DATA_UINT64 }, 456 { "cantconn", KSTAT_DATA_UINT64 }, 457 { "nomem", KSTAT_DATA_UINT64 }, 458 { "interrupts", KSTAT_DATA_UINT64 } 459 }; 460 461 #define COTSRCSTAT_INCR(p, x) \ 462 atomic_add_64(&(p)->x.value.ui64, 1) 463 464 #define CLNT_MAX_CONNS 1 /* concurrent connections between clnt/srvr */ 465 static int clnt_max_conns = CLNT_MAX_CONNS; 466 467 #define CLNT_MIN_TIMEOUT 10 /* seconds to wait after we get a */ 468 /* connection reset */ 469 #define CLNT_MIN_CONNTIMEOUT 5 /* seconds to wait for a connection */ 470 471 472 static int clnt_cots_min_tout = CLNT_MIN_TIMEOUT; 473 static int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT; 474 475 /* 476 * Limit the number of times we will attempt to receive a reply without 477 * re-sending a response. 478 */ 479 #define CLNT_MAXRECV_WITHOUT_RETRY 3 480 static uint_t clnt_cots_maxrecv = CLNT_MAXRECV_WITHOUT_RETRY; 481 482 uint_t *clnt_max_msg_sizep; 483 void (*clnt_stop_idle)(queue_t *wq); 484 485 #define ptoh(p) (&((p)->cku_client)) 486 #define htop(h) ((cku_private_t *)((h)->cl_private)) 487 488 /* 489 * Times to retry 490 */ 491 #define REFRESHES 2 /* authentication refreshes */ 492 493 static int clnt_cots_do_bindresvport = 1; /* bind to reserved port */ 494 495 static zone_key_t zone_cots_key; 496 497 /* 498 * We need to do this after all kernel threads in the zone have exited. 499 */ 500 /* ARGSUSED */ 501 static void 502 clnt_zone_destroy(zoneid_t zoneid, void *unused) 503 { 504 struct cm_xprt **cmp; 505 struct cm_xprt *cm_entry; 506 struct cm_xprt *freelist = NULL; 507 508 mutex_enter(&connmgr_lock); 509 cmp = &cm_hd; 510 while ((cm_entry = *cmp) != NULL) { 511 if (cm_entry->x_zoneid == zoneid) { 512 *cmp = cm_entry->x_next; 513 cm_entry->x_next = freelist; 514 freelist = cm_entry; 515 } else { 516 cmp = &cm_entry->x_next; 517 } 518 } 519 mutex_exit(&connmgr_lock); 520 while ((cm_entry = freelist) != NULL) { 521 freelist = cm_entry->x_next; 522 connmgr_close(cm_entry); 523 } 524 } 525 526 int 527 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog, 528 rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl) 529 { 530 CLIENT *h; 531 cku_private_t *p; 532 struct rpc_msg call_msg; 533 struct rpcstat *rpcstat; 534 535 RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog); 536 537 rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone); 538 ASSERT(rpcstat != NULL); 539 540 /* Allocate and intialize the client handle. */ 541 p = kmem_zalloc(sizeof (*p), KM_SLEEP); 542 543 h = ptoh(p); 544 545 h->cl_private = (caddr_t)p; 546 h->cl_auth = authkern_create(); 547 h->cl_ops = &tcp_ops; 548 549 cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL); 550 mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL); 551 552 /* 553 * If the current sanity check size in rpcmod is smaller 554 * than the size needed, then increase the sanity check. 555 */ 556 if (max_msgsize != 0 && clnt_max_msg_sizep != NULL && 557 max_msgsize > *clnt_max_msg_sizep) { 558 mutex_enter(&clnt_max_msg_lock); 559 if (max_msgsize > *clnt_max_msg_sizep) 560 *clnt_max_msg_sizep = max_msgsize; 561 mutex_exit(&clnt_max_msg_lock); 562 } 563 564 p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE; 565 566 /* Preserialize the call message header */ 567 568 call_msg.rm_xid = 0; 569 call_msg.rm_direction = CALL; 570 call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION; 571 call_msg.rm_call.cb_prog = prog; 572 call_msg.rm_call.cb_vers = vers; 573 574 xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE); 575 576 if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) { 577 RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization " 578 "error\n"); 579 auth_destroy(h->cl_auth); 580 kmem_free(p, sizeof (cku_private_t)); 581 RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n"); 582 return (EINVAL); /* XXX */ 583 } 584 585 /* 586 * The zalloc initialized the fields below. 587 * p->cku_xid = 0; 588 * p->cku_flags = 0; 589 * p->cku_srcaddr.len = 0; 590 * p->cku_srcaddr.maxlen = 0; 591 */ 592 593 p->cku_cred = cred; 594 p->cku_device = dev; 595 p->cku_addrfmly = family; 596 p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP); 597 p->cku_addr.maxlen = addr->maxlen; 598 p->cku_addr.len = addr->len; 599 bcopy(addr->buf, p->cku_addr.buf, addr->len); 600 p->cku_stats = rpcstat->rpc_cots_client; 601 p->cku_useresvport = -1; /* value is has not been set */ 602 603 *ncl = h; 604 return (0); 605 } 606 607 /*ARGSUSED*/ 608 static void 609 clnt_cots_kabort(CLIENT *h) 610 { 611 } 612 613 /* 614 * Return error info on this handle. 615 */ 616 static void 617 clnt_cots_kerror(CLIENT *h, struct rpc_err *err) 618 { 619 /* LINTED pointer alignment */ 620 cku_private_t *p = htop(h); 621 622 *err = p->cku_err; 623 } 624 625 static bool_t 626 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr) 627 { 628 /* LINTED pointer alignment */ 629 cku_private_t *p = htop(h); 630 XDR *xdrs; 631 632 xdrs = &(p->cku_outxdr); 633 xdrs->x_op = XDR_FREE; 634 return ((*xdr_res)(xdrs, res_ptr)); 635 } 636 637 static bool_t 638 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg) 639 { 640 cku_private_t *p = htop(h); 641 642 switch (cmd) { 643 case CLSET_PROGRESS: 644 p->cku_progress = TRUE; 645 return (TRUE); 646 647 case CLSET_XID: 648 if (arg == NULL) 649 return (FALSE); 650 651 p->cku_xid = *((uint32_t *)arg); 652 return (TRUE); 653 654 case CLGET_XID: 655 if (arg == NULL) 656 return (FALSE); 657 658 *((uint32_t *)arg) = p->cku_xid; 659 return (TRUE); 660 661 case CLSET_NODELAYONERR: 662 if (arg == NULL) 663 return (FALSE); 664 665 if (*((bool_t *)arg) == TRUE) { 666 p->cku_nodelayonerr = TRUE; 667 return (TRUE); 668 } 669 if (*((bool_t *)arg) == FALSE) { 670 p->cku_nodelayonerr = FALSE; 671 return (TRUE); 672 } 673 return (FALSE); 674 675 case CLGET_NODELAYONERR: 676 if (arg == NULL) 677 return (FALSE); 678 679 *((bool_t *)arg) = p->cku_nodelayonerr; 680 return (TRUE); 681 682 case CLSET_BINDRESVPORT: 683 if (arg == NULL) 684 return (FALSE); 685 686 if (*(int *)arg != 1 && *(int *)arg != 0) 687 return (FALSE); 688 689 p->cku_useresvport = *(int *)arg; 690 691 return (TRUE); 692 693 case CLGET_BINDRESVPORT: 694 if (arg == NULL) 695 return (FALSE); 696 697 *(int *)arg = p->cku_useresvport; 698 699 return (TRUE); 700 701 default: 702 return (FALSE); 703 } 704 } 705 706 /* 707 * Destroy rpc handle. Frees the space used for output buffer, 708 * private data, and handle structure. 709 */ 710 static void 711 clnt_cots_kdestroy(CLIENT *h) 712 { 713 /* LINTED pointer alignment */ 714 cku_private_t *p = htop(h); 715 calllist_t *call = &p->cku_call; 716 717 RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h); 718 RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid); 719 720 if (p->cku_flags & CKU_ONQUEUE) { 721 RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x " 722 "from dispatch list\n", p->cku_xid); 723 call_table_remove(call); 724 } 725 726 if (call->call_reply) 727 freemsg(call->call_reply); 728 cv_destroy(&call->call_cv); 729 mutex_destroy(&call->call_lock); 730 731 kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen); 732 kmem_free(p->cku_addr.buf, p->cku_addr.maxlen); 733 kmem_free(p, sizeof (*p)); 734 } 735 736 static int clnt_cots_pulls; 737 #define RM_HDR_SIZE 4 /* record mark header size */ 738 739 /* 740 * Call remote procedure. 741 */ 742 static enum clnt_stat 743 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args, 744 caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait) 745 { 746 /* LINTED pointer alignment */ 747 cku_private_t *p = htop(h); 748 calllist_t *call = &p->cku_call; 749 XDR *xdrs; 750 struct rpc_msg reply_msg; 751 mblk_t *mp; 752 #ifdef RPCDEBUG 753 clock_t time_sent; 754 #endif 755 struct netbuf *retryaddr; 756 struct cm_xprt *cm_entry = NULL; 757 queue_t *wq; 758 int len; 759 int mpsize; 760 int refreshes = REFRESHES; 761 int interrupted; 762 int tidu_size; 763 enum clnt_stat status; 764 struct timeval cwait; 765 bool_t delay_first = FALSE; 766 clock_t ticks; 767 768 RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum); 769 COTSRCSTAT_INCR(p->cku_stats, rccalls); 770 771 RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec); 772 RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec); 773 774 /* 775 * Bug ID 1240234: 776 * Look out for zero length timeouts. We don't want to 777 * wait zero seconds for a connection to be established. 778 */ 779 if (wait.tv_sec < clnt_cots_min_conntout) { 780 cwait.tv_sec = clnt_cots_min_conntout; 781 cwait.tv_usec = 0; 782 RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,", 783 wait.tv_sec); 784 RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout); 785 } else { 786 cwait = wait; 787 } 788 789 call_again: 790 if (cm_entry) { 791 connmgr_release(cm_entry); 792 cm_entry = NULL; 793 } 794 795 mp = NULL; 796 797 /* 798 * If the call is not a retry, allocate a new xid and cache it 799 * for future retries. 800 * Bug ID 1246045: 801 * Treat call as a retry for purposes of binding the source 802 * port only if we actually attempted to send anything on 803 * the previous call. 804 */ 805 if (p->cku_xid == 0) { 806 p->cku_xid = alloc_xid(); 807 /* 808 * We need to ASSERT here that our xid != 0 because this 809 * determines whether or not our call record gets placed on 810 * the hash table or the linked list. By design, we mandate 811 * that RPC calls over cots must have xid's != 0, so we can 812 * ensure proper management of the hash table. 813 */ 814 ASSERT(p->cku_xid != 0); 815 816 retryaddr = NULL; 817 p->cku_flags &= ~CKU_SENT; 818 819 if (p->cku_flags & CKU_ONQUEUE) { 820 RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old" 821 " one (%p)\n", (void *)call); 822 call_table_remove(call); 823 p->cku_flags &= ~CKU_ONQUEUE; 824 RPCLOG(64, "clnt_cots_kcallit: removing call from " 825 "dispatch list because xid was zero (now 0x%x)\n", 826 p->cku_xid); 827 } 828 829 if (call->call_reply != NULL) { 830 freemsg(call->call_reply); 831 call->call_reply = NULL; 832 } 833 } else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) { 834 retryaddr = NULL; 835 836 } else if (p->cku_flags & CKU_SENT) { 837 retryaddr = &p->cku_srcaddr; 838 839 } else { 840 /* 841 * Bug ID 1246045: Nothing was sent, so set retryaddr to 842 * NULL and let connmgr_get() bind to any source port it 843 * can get. 844 */ 845 retryaddr = NULL; 846 } 847 848 RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid); 849 RPCLOG(64, " flags = 0x%x\n", p->cku_flags); 850 851 p->cku_err.re_status = RPC_TIMEDOUT; 852 p->cku_err.re_errno = p->cku_err.re_terrno = 0; 853 854 cm_entry = connmgr_wrapget(retryaddr, &cwait, p); 855 856 if (cm_entry == NULL) { 857 RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n", 858 clnt_sperrno(p->cku_err.re_status)); 859 860 /* 861 * The reasons why we fail to create a connection are 862 * varied. In most cases we don't want the caller to 863 * immediately retry. This could have one or more 864 * bad effects. This includes flooding the net with 865 * connect requests to ports with no listener; a hard 866 * kernel loop due to all the "reserved" TCP ports being 867 * in use. 868 */ 869 delay_first = TRUE; 870 871 /* 872 * Even if we end up returning EINTR, we still count a 873 * a "can't connect", because the connection manager 874 * might have been committed to waiting for or timing out on 875 * a connection. 876 */ 877 COTSRCSTAT_INCR(p->cku_stats, rccantconn); 878 switch (p->cku_err.re_status) { 879 case RPC_INTR: 880 p->cku_err.re_errno = EINTR; 881 882 /* 883 * No need to delay because a UNIX signal(2) 884 * interrupted us. The caller likely won't 885 * retry the CLNT_CALL() and even if it does, 886 * we assume the caller knows what it is doing. 887 */ 888 delay_first = FALSE; 889 break; 890 891 case RPC_TIMEDOUT: 892 p->cku_err.re_errno = ETIMEDOUT; 893 894 /* 895 * No need to delay because timed out already 896 * on the connection request and assume that the 897 * transport time out is longer than our minimum 898 * timeout, or least not too much smaller. 899 */ 900 delay_first = FALSE; 901 break; 902 903 case RPC_SYSTEMERROR: 904 case RPC_TLIERROR: 905 /* 906 * We want to delay here because a transient 907 * system error has a better chance of going away 908 * if we delay a bit. If it's not transient, then 909 * we don't want end up in a hard kernel loop 910 * due to retries. 911 */ 912 ASSERT(p->cku_err.re_errno != 0); 913 break; 914 915 916 case RPC_CANTCONNECT: 917 /* 918 * RPC_CANTCONNECT is set on T_ERROR_ACK which 919 * implies some error down in the TCP layer or 920 * below. If cku_nodelayonerror is set then we 921 * assume the caller knows not to try too hard. 922 */ 923 RPCLOG0(8, "clnt_cots_kcallit: connection failed,"); 924 RPCLOG0(8, " re_status=RPC_CANTCONNECT,"); 925 RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno); 926 RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr); 927 if (p->cku_nodelayonerr == TRUE) 928 delay_first = FALSE; 929 930 p->cku_err.re_errno = EIO; 931 932 break; 933 934 case RPC_XPRTFAILED: 935 /* 936 * We want to delay here because we likely 937 * got a refused connection. 938 */ 939 if (p->cku_err.re_errno != 0) 940 break; 941 942 /* fall thru */ 943 944 default: 945 /* 946 * We delay here because it is better to err 947 * on the side of caution. If we got here then 948 * status could have been RPC_SUCCESS, but we 949 * know that we did not get a connection, so 950 * force the rpc status to RPC_CANTCONNECT. 951 */ 952 p->cku_err.re_status = RPC_CANTCONNECT; 953 p->cku_err.re_errno = EIO; 954 break; 955 } 956 if (delay_first == TRUE) 957 ticks = clnt_cots_min_tout * drv_usectohz(1000000); 958 goto cots_done; 959 } 960 961 /* 962 * If we've never sent any request on this connection (send count 963 * is zero, or the connection has been reset), cache the 964 * the connection's create time and send a request (possibly a retry) 965 */ 966 if ((p->cku_flags & CKU_SENT) == 0 || 967 p->cku_ctime != cm_entry->x_ctime) { 968 p->cku_ctime = cm_entry->x_ctime; 969 970 } else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) && 971 (call->call_reply != NULL || 972 p->cku_recv_attempts < clnt_cots_maxrecv)) { 973 974 /* 975 * If we've sent a request and our call is on the dispatch 976 * queue and we haven't made too many receive attempts, then 977 * don't re-send, just receive. 978 */ 979 p->cku_recv_attempts++; 980 goto read_again; 981 } 982 983 /* 984 * Now we create the RPC request in a STREAMS message. We have to do 985 * this after the call to connmgr_get so that we have the correct 986 * TIDU size for the transport. 987 */ 988 tidu_size = cm_entry->x_tidu_size; 989 len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE); 990 991 while ((mp = allocb(len, BPRI_MED)) == NULL) { 992 if (strwaitbuf(len, BPRI_MED)) { 993 p->cku_err.re_status = RPC_SYSTEMERROR; 994 p->cku_err.re_errno = ENOSR; 995 COTSRCSTAT_INCR(p->cku_stats, rcnomem); 996 goto cots_done; 997 } 998 } 999 xdrs = &p->cku_outxdr; 1000 xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size); 1001 mpsize = MBLKSIZE(mp); 1002 ASSERT(mpsize >= len); 1003 ASSERT(mp->b_rptr == mp->b_datap->db_base); 1004 1005 /* 1006 * If the size of mblk is not appreciably larger than what we 1007 * asked, then resize the mblk to exactly len bytes. The reason for 1008 * this: suppose len is 1600 bytes, the tidu is 1460 bytes 1009 * (from TCP over ethernet), and the arguments to the RPC require 1010 * 2800 bytes. Ideally we want the protocol to render two 1011 * ~1400 byte segments over the wire. However if allocb() gives us a 2k 1012 * mblk, and we allocate a second mblk for the remainder, the protocol 1013 * module may generate 3 segments over the wire: 1014 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and 1015 * 892 for the third. If we "waste" 448 bytes in the first mblk, 1016 * the XDR encoding will generate two ~1400 byte mblks, and the 1017 * protocol module is more likely to produce properly sized segments. 1018 */ 1019 if ((mpsize >> 1) <= len) 1020 mp->b_rptr += (mpsize - len); 1021 1022 /* 1023 * Adjust b_rptr to reserve space for the non-data protocol headers 1024 * any downstream modules might like to add, and for the 1025 * record marking header. 1026 */ 1027 mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE); 1028 1029 if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) { 1030 /* Copy in the preserialized RPC header information. */ 1031 bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE); 1032 1033 /* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */ 1034 XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base + 1035 WIRE_HDR_SIZE)); 1036 1037 ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE); 1038 1039 /* Serialize the procedure number and the arguments. */ 1040 if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) || 1041 (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) || 1042 (!(*xdr_args)(xdrs, argsp))) { 1043 p->cku_err.re_status = RPC_CANTENCODEARGS; 1044 p->cku_err.re_errno = EIO; 1045 goto cots_done; 1046 } 1047 1048 (*(uint32_t *)(mp->b_rptr)) = p->cku_xid; 1049 } else { 1050 uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE]; 1051 IXDR_PUT_U_INT32(uproc, procnum); 1052 1053 (*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid; 1054 1055 /* Use XDR_SETPOS() to set the b_wptr. */ 1056 XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base)); 1057 1058 /* Serialize the procedure number and the arguments. */ 1059 if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4, 1060 xdrs, xdr_args, argsp)) { 1061 p->cku_err.re_status = RPC_CANTENCODEARGS; 1062 p->cku_err.re_errno = EIO; 1063 goto cots_done; 1064 } 1065 } 1066 1067 RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n", 1068 tidu_size); 1069 1070 wq = cm_entry->x_wq; 1071 clnt_dispatch_send(wq, mp, call, p->cku_xid, 1072 (p->cku_flags & CKU_ONQUEUE)); 1073 1074 RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n", 1075 (uint_t)p->cku_xid); 1076 p->cku_flags = (CKU_ONQUEUE|CKU_SENT); 1077 p->cku_recv_attempts = 1; 1078 1079 #ifdef RPCDEBUG 1080 time_sent = lbolt; 1081 #endif 1082 1083 /* 1084 * Wait for a reply or a timeout. If there is no error or timeout, 1085 * (both indicated by call_status), call->call_reply will contain 1086 * the RPC reply message. 1087 */ 1088 read_again: 1089 mutex_enter(&call->call_lock); 1090 interrupted = 0; 1091 if (call->call_status == RPC_TIMEDOUT) { 1092 /* 1093 * Indicate that the lwp is not to be stopped while waiting 1094 * for this network traffic. This is to avoid deadlock while 1095 * debugging a process via /proc and also to avoid recursive 1096 * mutex_enter()s due to NFS page faults while stopping 1097 * (NFS holds locks when it calls here). 1098 */ 1099 clock_t cv_wait_ret; 1100 clock_t timout; 1101 clock_t oldlbolt; 1102 1103 klwp_t *lwp = ttolwp(curthread); 1104 1105 if (lwp != NULL) 1106 lwp->lwp_nostop++; 1107 1108 oldlbolt = lbolt; 1109 timout = wait.tv_sec * drv_usectohz(1000000) + 1110 drv_usectohz(wait.tv_usec) + oldlbolt; 1111 /* 1112 * Iterate until the call_status is changed to something 1113 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns 1114 * something <=0 zero. The latter means that we timed 1115 * out. 1116 */ 1117 if (h->cl_nosignal) 1118 while ((cv_wait_ret = cv_timedwait(&call->call_cv, 1119 &call->call_lock, timout)) > 0 && 1120 call->call_status == RPC_TIMEDOUT); 1121 else 1122 while ((cv_wait_ret = cv_timedwait_sig( 1123 &call->call_cv, 1124 &call->call_lock, timout)) > 0 && 1125 call->call_status == RPC_TIMEDOUT); 1126 1127 switch (cv_wait_ret) { 1128 case 0: 1129 /* 1130 * If we got out of the above loop with 1131 * cv_timedwait_sig() returning 0, then we were 1132 * interrupted regardless what call_status is. 1133 */ 1134 interrupted = 1; 1135 break; 1136 case -1: 1137 /* cv_timedwait_sig() timed out */ 1138 break; 1139 default: 1140 1141 /* 1142 * We were cv_signaled(). If we didn't 1143 * get a successful call_status and returned 1144 * before time expired, delay up to clnt_cots_min_tout 1145 * seconds so that the caller doesn't immediately 1146 * try to call us again and thus force the 1147 * same condition that got us here (such 1148 * as a RPC_XPRTFAILED due to the server not 1149 * listening on the end-point. 1150 */ 1151 if (call->call_status != RPC_SUCCESS) { 1152 clock_t curlbolt; 1153 clock_t diff; 1154 1155 curlbolt = ddi_get_lbolt(); 1156 ticks = clnt_cots_min_tout * 1157 drv_usectohz(1000000); 1158 diff = curlbolt - oldlbolt; 1159 if (diff < ticks) { 1160 delay_first = TRUE; 1161 if (diff > 0) 1162 ticks -= diff; 1163 } 1164 } 1165 break; 1166 } 1167 1168 if (lwp != NULL) 1169 lwp->lwp_nostop--; 1170 } 1171 /* 1172 * Get the reply message, if any. This will be freed at the end 1173 * whether or not an error occurred. 1174 */ 1175 mp = call->call_reply; 1176 call->call_reply = NULL; 1177 1178 /* 1179 * call_err is the error info when the call is on dispatch queue. 1180 * cku_err is the error info returned to the caller. 1181 * Sync cku_err with call_err for local message processing. 1182 */ 1183 1184 status = call->call_status; 1185 p->cku_err = call->call_err; 1186 mutex_exit(&call->call_lock); 1187 1188 if (status != RPC_SUCCESS) { 1189 switch (status) { 1190 case RPC_TIMEDOUT: 1191 if (interrupted) { 1192 COTSRCSTAT_INCR(p->cku_stats, rcintrs); 1193 p->cku_err.re_status = RPC_INTR; 1194 p->cku_err.re_errno = EINTR; 1195 RPCLOG(1, "clnt_cots_kcallit: xid 0x%x", 1196 p->cku_xid); 1197 RPCLOG(1, "signal interrupted at %ld", lbolt); 1198 RPCLOG(1, ", was sent at %ld\n", time_sent); 1199 } else { 1200 COTSRCSTAT_INCR(p->cku_stats, rctimeouts); 1201 p->cku_err.re_errno = ETIMEDOUT; 1202 RPCLOG(1, "clnt_cots_kcallit: timed out at %ld", 1203 lbolt); 1204 RPCLOG(1, ", was sent at %ld\n", time_sent); 1205 } 1206 break; 1207 1208 case RPC_XPRTFAILED: 1209 if (p->cku_err.re_errno == 0) 1210 p->cku_err.re_errno = EIO; 1211 1212 RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n", 1213 p->cku_err.re_errno); 1214 break; 1215 1216 case RPC_SYSTEMERROR: 1217 ASSERT(p->cku_err.re_errno); 1218 RPCLOG(1, "clnt_cots_kcallit: system error: %d\n", 1219 p->cku_err.re_errno); 1220 break; 1221 1222 default: 1223 p->cku_err.re_status = RPC_SYSTEMERROR; 1224 p->cku_err.re_errno = EIO; 1225 RPCLOG(1, "clnt_cots_kcallit: error: %s\n", 1226 clnt_sperrno(status)); 1227 break; 1228 } 1229 if (p->cku_err.re_status != RPC_TIMEDOUT) { 1230 1231 if (p->cku_flags & CKU_ONQUEUE) { 1232 call_table_remove(call); 1233 p->cku_flags &= ~CKU_ONQUEUE; 1234 } 1235 1236 RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x " 1237 "taken off dispatch list\n", p->cku_xid); 1238 if (call->call_reply) { 1239 freemsg(call->call_reply); 1240 call->call_reply = NULL; 1241 } 1242 } else if (wait.tv_sec != 0) { 1243 /* 1244 * We've sent the request over TCP and so we have 1245 * every reason to believe it will get 1246 * delivered. In which case returning a timeout is not 1247 * appropriate. 1248 */ 1249 if (p->cku_progress == TRUE && 1250 p->cku_recv_attempts < clnt_cots_maxrecv) { 1251 p->cku_err.re_status = RPC_INPROGRESS; 1252 } 1253 } 1254 goto cots_done; 1255 } 1256 1257 xdrs = &p->cku_inxdr; 1258 xdrmblk_init(xdrs, mp, XDR_DECODE, 0); 1259 1260 reply_msg.rm_direction = REPLY; 1261 reply_msg.rm_reply.rp_stat = MSG_ACCEPTED; 1262 reply_msg.acpted_rply.ar_stat = SUCCESS; 1263 1264 reply_msg.acpted_rply.ar_verf = _null_auth; 1265 /* 1266 * xdr_results will be done in AUTH_UNWRAP. 1267 */ 1268 reply_msg.acpted_rply.ar_results.where = NULL; 1269 reply_msg.acpted_rply.ar_results.proc = xdr_void; 1270 1271 if (xdr_replymsg(xdrs, &reply_msg)) { 1272 enum clnt_stat re_status; 1273 1274 _seterr_reply(&reply_msg, &p->cku_err); 1275 1276 re_status = p->cku_err.re_status; 1277 if (re_status == RPC_SUCCESS) { 1278 /* 1279 * Reply is good, check auth. 1280 */ 1281 if (!AUTH_VALIDATE(h->cl_auth, 1282 &reply_msg.acpted_rply.ar_verf)) { 1283 COTSRCSTAT_INCR(p->cku_stats, rcbadverfs); 1284 RPCLOG0(1, "clnt_cots_kcallit: validation " 1285 "failure\n"); 1286 freemsg(mp); 1287 (void) xdr_rpc_free_verifier(xdrs, &reply_msg); 1288 mutex_enter(&call->call_lock); 1289 if (call->call_reply == NULL) 1290 call->call_status = RPC_TIMEDOUT; 1291 mutex_exit(&call->call_lock); 1292 goto read_again; 1293 } else if (!AUTH_UNWRAP(h->cl_auth, xdrs, 1294 xdr_results, resultsp)) { 1295 RPCLOG0(1, "clnt_cots_kcallit: validation " 1296 "failure (unwrap)\n"); 1297 p->cku_err.re_status = RPC_CANTDECODERES; 1298 p->cku_err.re_errno = EIO; 1299 } 1300 } else { 1301 /* set errno in case we can't recover */ 1302 if (re_status != RPC_VERSMISMATCH && 1303 re_status != RPC_AUTHERROR && 1304 re_status != RPC_PROGVERSMISMATCH) 1305 p->cku_err.re_errno = EIO; 1306 1307 if (re_status == RPC_AUTHERROR) { 1308 /* 1309 * Maybe our credential need to be 1310 * refreshed 1311 */ 1312 if ((refreshes > 0) && 1313 AUTH_REFRESH(h->cl_auth, &reply_msg, 1314 p->cku_cred)) { 1315 refreshes--; 1316 (void) xdr_rpc_free_verifier(xdrs, &reply_msg); 1317 freemsg(mp); 1318 mp = NULL; 1319 1320 if (p->cku_flags & CKU_ONQUEUE) { 1321 call_table_remove(call); 1322 p->cku_flags &= ~CKU_ONQUEUE; 1323 } 1324 1325 RPCLOG(64, "clnt_cots_kcallit: AUTH_ERROR, so " 1326 "xid 0x%x taken off dispatch list\n", 1327 p->cku_xid); 1328 if (call->call_reply) { 1329 freemsg(call->call_reply); 1330 call->call_reply = NULL; 1331 } 1332 COTSRCSTAT_INCR(p->cku_stats, rcbadcalls); 1333 COTSRCSTAT_INCR(p->cku_stats, rcnewcreds); 1334 goto call_again; 1335 } else { 1336 /* 1337 * We have used the client handle to 1338 * do an AUTH_REFRESH and the RPC status may 1339 * be set to RPC_SUCCESS; Let's make sure to 1340 * set it to RPC_AUTHERROR. 1341 */ 1342 p->cku_err.re_status = RPC_AUTHERROR; 1343 /* 1344 * Map recoverable and unrecoverable 1345 * authentication errors to appropriate errno 1346 */ 1347 switch (p->cku_err.re_why) { 1348 case AUTH_BADCRED: 1349 case AUTH_BADVERF: 1350 case AUTH_INVALIDRESP: 1351 case AUTH_TOOWEAK: 1352 case AUTH_FAILED: 1353 case RPCSEC_GSS_NOCRED: 1354 case RPCSEC_GSS_FAILED: 1355 p->cku_err.re_errno = EACCES; 1356 break; 1357 case AUTH_REJECTEDCRED: 1358 case AUTH_REJECTEDVERF: 1359 default: p->cku_err.re_errno = EIO; 1360 break; 1361 } 1362 RPCLOG(1, "clnt_cots_kcallit : authentication" 1363 " failed with RPC_AUTHERROR of type %d\n", 1364 (int)p->cku_err.re_why); 1365 } 1366 } 1367 } 1368 } else { 1369 /* reply didn't decode properly. */ 1370 p->cku_err.re_status = RPC_CANTDECODERES; 1371 p->cku_err.re_errno = EIO; 1372 RPCLOG0(1, "clnt_cots_kcallit: decode failure\n"); 1373 } 1374 1375 (void) xdr_rpc_free_verifier(xdrs, &reply_msg); 1376 1377 if (p->cku_flags & CKU_ONQUEUE) { 1378 call_table_remove(call); 1379 p->cku_flags &= ~CKU_ONQUEUE; 1380 } 1381 1382 RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list", 1383 p->cku_xid); 1384 RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status)); 1385 cots_done: 1386 if (cm_entry) 1387 connmgr_release(cm_entry); 1388 1389 if (mp != NULL) 1390 freemsg(mp); 1391 if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) { 1392 freemsg(call->call_reply); 1393 call->call_reply = NULL; 1394 } 1395 if (p->cku_err.re_status != RPC_SUCCESS) { 1396 RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n"); 1397 COTSRCSTAT_INCR(p->cku_stats, rcbadcalls); 1398 } 1399 1400 /* 1401 * No point in delaying if the zone is going away. 1402 */ 1403 if (delay_first == TRUE && 1404 !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) { 1405 if (clnt_delay(ticks, h->cl_nosignal) == EINTR) { 1406 p->cku_err.re_errno = EINTR; 1407 p->cku_err.re_status = RPC_INTR; 1408 } 1409 } 1410 return (p->cku_err.re_status); 1411 } 1412 1413 /* 1414 * Kinit routine for cots. This sets up the correct operations in 1415 * the client handle, as the handle may have previously been a clts 1416 * handle, and clears the xid field so there is no way a new call 1417 * could be mistaken for a retry. It also sets in the handle the 1418 * information that is passed at create/kinit time but needed at 1419 * call time, as cots creates the transport at call time - device, 1420 * address of the server, protocol family. 1421 */ 1422 void 1423 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr, 1424 int max_msgsize, cred_t *cred) 1425 { 1426 /* LINTED pointer alignment */ 1427 cku_private_t *p = htop(h); 1428 calllist_t *call = &p->cku_call; 1429 1430 h->cl_ops = &tcp_ops; 1431 if (p->cku_flags & CKU_ONQUEUE) { 1432 call_table_remove(call); 1433 p->cku_flags &= ~CKU_ONQUEUE; 1434 RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from" 1435 " dispatch list\n", p->cku_xid); 1436 } 1437 1438 if (call->call_reply != NULL) { 1439 freemsg(call->call_reply); 1440 call->call_reply = NULL; 1441 } 1442 1443 call->call_bucket = NULL; 1444 call->call_hash = 0; 1445 1446 /* 1447 * We don't clear cku_flags here, because clnt_cots_kcallit() 1448 * takes care of handling the cku_flags reset. 1449 */ 1450 p->cku_xid = 0; 1451 p->cku_device = dev; 1452 p->cku_addrfmly = family; 1453 p->cku_cred = cred; 1454 1455 if (p->cku_addr.maxlen < addr->len) { 1456 if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL) 1457 kmem_free(p->cku_addr.buf, p->cku_addr.maxlen); 1458 p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP); 1459 p->cku_addr.maxlen = addr->maxlen; 1460 } 1461 1462 p->cku_addr.len = addr->len; 1463 bcopy(addr->buf, p->cku_addr.buf, addr->len); 1464 1465 /* 1466 * If the current sanity check size in rpcmod is smaller 1467 * than the size needed, then increase the sanity check. 1468 */ 1469 if (max_msgsize != 0 && clnt_max_msg_sizep != NULL && 1470 max_msgsize > *clnt_max_msg_sizep) { 1471 mutex_enter(&clnt_max_msg_lock); 1472 if (max_msgsize > *clnt_max_msg_sizep) 1473 *clnt_max_msg_sizep = max_msgsize; 1474 mutex_exit(&clnt_max_msg_lock); 1475 } 1476 } 1477 1478 /* 1479 * ksettimers is a no-op for cots, with the exception of setting the xid. 1480 */ 1481 /* ARGSUSED */ 1482 static int 1483 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all, 1484 int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg, 1485 uint32_t xid) 1486 { 1487 /* LINTED pointer alignment */ 1488 cku_private_t *p = htop(h); 1489 1490 if (xid) 1491 p->cku_xid = xid; 1492 COTSRCSTAT_INCR(p->cku_stats, rctimers); 1493 return (0); 1494 } 1495 1496 extern void rpc_poptimod(struct vnode *); 1497 extern int kstr_push(struct vnode *, char *); 1498 1499 int 1500 conn_kstat_update(kstat_t *ksp, int rw) 1501 { 1502 struct cm_xprt *cm_entry; 1503 struct cm_kstat_xprt *cm_ksp_data; 1504 uchar_t *b; 1505 char *fbuf; 1506 1507 if (rw == KSTAT_WRITE) 1508 return (EACCES); 1509 if (ksp == NULL || ksp->ks_private == NULL) 1510 return (EIO); 1511 cm_entry = (struct cm_xprt *)ksp->ks_private; 1512 cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data; 1513 1514 cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq; 1515 cm_ksp_data->x_family.value.ui32 = cm_entry->x_family; 1516 cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev; 1517 cm_ksp_data->x_time.value.ui32 = cm_entry->x_time; 1518 cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref; 1519 cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags; 1520 1521 if (cm_entry->x_server.buf) { 1522 fbuf = cm_ksp_data->x_server.value.string.addr.ptr; 1523 if (cm_entry->x_family == AF_INET && 1524 cm_entry->x_server.len == 1525 sizeof (struct sockaddr_in)) { 1526 struct sockaddr_in *sa; 1527 sa = (struct sockaddr_in *) 1528 cm_entry->x_server.buf; 1529 b = (uchar_t *)&sa->sin_addr; 1530 (void) sprintf(fbuf, 1531 "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF, 1532 b[2] & 0xFF, b[3] & 0xFF); 1533 cm_ksp_data->x_port.value.ui32 = 1534 (uint32_t)sa->sin_port; 1535 } else if (cm_entry->x_family == AF_INET6 && 1536 cm_entry->x_server.len >= 1537 sizeof (struct sockaddr_in6)) { 1538 /* extract server IP address & port */ 1539 struct sockaddr_in6 *sin6; 1540 sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf; 1541 (void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf, 1542 INET6_ADDRSTRLEN); 1543 cm_ksp_data->x_port.value.ui32 = sin6->sin6_port; 1544 } else { 1545 struct sockaddr_in *sa; 1546 1547 sa = (struct sockaddr_in *)cm_entry->x_server.buf; 1548 b = (uchar_t *)&sa->sin_addr; 1549 (void) sprintf(fbuf, 1550 "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF, 1551 b[2] & 0xFF, b[3] & 0xFF); 1552 } 1553 KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) = 1554 strlen(fbuf) + 1; 1555 } 1556 1557 return (0); 1558 } 1559 1560 1561 /* 1562 * We want a version of delay which is interruptible by a UNIX signal 1563 * Return EINTR if an interrupt occured. 1564 */ 1565 static int 1566 clnt_delay(clock_t ticks, bool_t nosignal) 1567 { 1568 if (nosignal == TRUE) { 1569 delay(ticks); 1570 return (0); 1571 } 1572 return (delay_sig(ticks)); 1573 } 1574 1575 /* 1576 * Wait for a connection until a timeout, or until we are 1577 * signalled that there has been a connection state change. 1578 */ 1579 static enum clnt_stat 1580 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp, 1581 bool_t nosignal) 1582 { 1583 bool_t interrupted; 1584 clock_t timout, cv_stat; 1585 enum clnt_stat clstat; 1586 unsigned int old_state; 1587 1588 ASSERT(MUTEX_HELD(&connmgr_lock)); 1589 /* 1590 * We wait for the transport connection to be made, or an 1591 * indication that it could not be made. 1592 */ 1593 clstat = RPC_TIMEDOUT; 1594 interrupted = FALSE; 1595 1596 old_state = cm_entry->x_state_flags; 1597 /* 1598 * Now loop until cv_timedwait{_sig} returns because of 1599 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be 1600 * cv_signalled for various other reasons too. So loop 1601 * until there is a state change on the connection. 1602 */ 1603 1604 timout = waitp->tv_sec * drv_usectohz(1000000) + 1605 drv_usectohz(waitp->tv_usec) + lbolt; 1606 1607 if (nosignal) { 1608 while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv, 1609 &connmgr_lock, timout)) > 0 && 1610 cm_entry->x_state_flags == old_state) 1611 ; 1612 } else { 1613 while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv, 1614 &connmgr_lock, timout)) > 0 && 1615 cm_entry->x_state_flags == old_state) 1616 ; 1617 1618 if (cv_stat == 0) /* got intr signal? */ 1619 interrupted = TRUE; 1620 } 1621 1622 if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) == 1623 X_CONNECTED) { 1624 clstat = RPC_SUCCESS; 1625 } else { 1626 if (interrupted == TRUE) 1627 clstat = RPC_INTR; 1628 RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n", 1629 clnt_sperrno(clstat)); 1630 } 1631 1632 return (clstat); 1633 } 1634 1635 /* 1636 * Primary interface for how RPC grabs a connection. 1637 */ 1638 static struct cm_xprt * 1639 connmgr_wrapget( 1640 struct netbuf *retryaddr, 1641 const struct timeval *waitp, 1642 cku_private_t *p) 1643 { 1644 struct cm_xprt *cm_entry; 1645 1646 cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly, 1647 &p->cku_srcaddr, &p->cku_err, p->cku_device, 1648 p->cku_client.cl_nosignal, p->cku_useresvport); 1649 1650 if (cm_entry == NULL) { 1651 /* 1652 * Re-map the call status to RPC_INTR if the err code is 1653 * EINTR. This can happen if calls status is RPC_TLIERROR. 1654 * However, don't re-map if signalling has been turned off. 1655 * XXX Really need to create a separate thread whenever 1656 * there isn't an existing connection. 1657 */ 1658 if (p->cku_err.re_errno == EINTR) { 1659 if (p->cku_client.cl_nosignal == TRUE) 1660 p->cku_err.re_errno = EIO; 1661 else 1662 p->cku_err.re_status = RPC_INTR; 1663 } 1664 } 1665 1666 return (cm_entry); 1667 } 1668 1669 /* 1670 * Obtains a transport to the server specified in addr. If a suitable transport 1671 * does not already exist in the list of cached transports, a new connection 1672 * is created, connected, and added to the list. The connection is for sending 1673 * only - the reply message may come back on another transport connection. 1674 */ 1675 static struct cm_xprt * 1676 connmgr_get( 1677 struct netbuf *retryaddr, 1678 const struct timeval *waitp, /* changed to a ptr to converse stack */ 1679 struct netbuf *destaddr, 1680 int addrfmly, 1681 struct netbuf *srcaddr, 1682 struct rpc_err *rpcerr, 1683 dev_t device, 1684 bool_t nosignal, 1685 int useresvport) 1686 { 1687 struct cm_xprt *cm_entry; 1688 struct cm_xprt *lru_entry; 1689 struct cm_xprt **cmp; 1690 queue_t *wq; 1691 TIUSER *tiptr; 1692 int i; 1693 int retval; 1694 clock_t prev_time; 1695 int tidu_size; 1696 bool_t connected; 1697 zoneid_t zoneid = getzoneid(); 1698 1699 /* 1700 * If the call is not a retry, look for a transport entry that 1701 * goes to the server of interest. 1702 */ 1703 mutex_enter(&connmgr_lock); 1704 1705 if (retryaddr == NULL) { 1706 use_new_conn: 1707 i = 0; 1708 cm_entry = lru_entry = NULL; 1709 prev_time = lbolt; 1710 1711 cmp = &cm_hd; 1712 while ((cm_entry = *cmp) != NULL) { 1713 ASSERT(cm_entry != cm_entry->x_next); 1714 /* 1715 * Garbage collect conections that are marked 1716 * for needs disconnect. 1717 */ 1718 if (cm_entry->x_needdis) { 1719 connmgr_dis_and_wait(cm_entry); 1720 /* 1721 * connmgr_lock could have been 1722 * dropped for the disconnect 1723 * processing so start over. 1724 */ 1725 goto use_new_conn; 1726 } 1727 1728 /* 1729 * Garbage collect the dead connections that have 1730 * no threads working on them. 1731 */ 1732 if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) == 1733 X_DEAD) { 1734 *cmp = cm_entry->x_next; 1735 mutex_exit(&connmgr_lock); 1736 connmgr_close(cm_entry); 1737 mutex_enter(&connmgr_lock); 1738 goto use_new_conn; 1739 } 1740 1741 1742 if ((cm_entry->x_state_flags & X_BADSTATES) == 0 && 1743 cm_entry->x_zoneid == zoneid && 1744 cm_entry->x_rdev == device && 1745 destaddr->len == cm_entry->x_server.len && 1746 bcmp(destaddr->buf, cm_entry->x_server.buf, 1747 destaddr->len) == 0) { 1748 /* 1749 * If the matching entry isn't connected, 1750 * attempt to reconnect it. 1751 */ 1752 if (cm_entry->x_connected == FALSE) { 1753 /* 1754 * We don't go through trying 1755 * to find the least recently 1756 * used connected because 1757 * connmgr_reconnect() briefly 1758 * dropped the connmgr_lock, 1759 * allowing a window for our 1760 * accounting to be messed up. 1761 * In any case, a re-connected 1762 * connection is as good as 1763 * a LRU connection. 1764 */ 1765 return (connmgr_wrapconnect(cm_entry, 1766 waitp, destaddr, addrfmly, srcaddr, 1767 rpcerr, TRUE, nosignal)); 1768 } 1769 i++; 1770 if (cm_entry->x_time - prev_time <= 0 || 1771 lru_entry == NULL) { 1772 prev_time = cm_entry->x_time; 1773 lru_entry = cm_entry; 1774 } 1775 } 1776 cmp = &cm_entry->x_next; 1777 } 1778 1779 if (i > clnt_max_conns) { 1780 RPCLOG(8, "connmgr_get: too many conns, dooming entry" 1781 " %p\n", (void *)lru_entry->x_tiptr); 1782 lru_entry->x_doomed = TRUE; 1783 goto use_new_conn; 1784 } 1785 1786 /* 1787 * If we are at the maximum number of connections to 1788 * the server, hand back the least recently used one. 1789 */ 1790 if (i == clnt_max_conns) { 1791 /* 1792 * Copy into the handle the source address of 1793 * the connection, which we will use in case of 1794 * a later retry. 1795 */ 1796 if (srcaddr->len != lru_entry->x_src.len) { 1797 if (srcaddr->len > 0) 1798 kmem_free(srcaddr->buf, 1799 srcaddr->maxlen); 1800 srcaddr->buf = kmem_zalloc( 1801 lru_entry->x_src.len, KM_SLEEP); 1802 srcaddr->maxlen = srcaddr->len = 1803 lru_entry->x_src.len; 1804 } 1805 bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len); 1806 RPCLOG(2, "connmgr_get: call going out on %p\n", 1807 (void *)lru_entry); 1808 lru_entry->x_time = lbolt; 1809 CONN_HOLD(lru_entry); 1810 mutex_exit(&connmgr_lock); 1811 return (lru_entry); 1812 } 1813 1814 } else { 1815 /* 1816 * This is the retry case (retryaddr != NULL). Retries must 1817 * be sent on the same source port as the original call. 1818 */ 1819 1820 /* 1821 * Walk the list looking for a connection with a source address 1822 * that matches the retry address. 1823 */ 1824 cmp = &cm_hd; 1825 while ((cm_entry = *cmp) != NULL) { 1826 ASSERT(cm_entry != cm_entry->x_next); 1827 if (zoneid != cm_entry->x_zoneid || 1828 device != cm_entry->x_rdev || 1829 retryaddr->len != cm_entry->x_src.len || 1830 bcmp(retryaddr->buf, cm_entry->x_src.buf, 1831 retryaddr->len) != 0) { 1832 cmp = &cm_entry->x_next; 1833 continue; 1834 } 1835 1836 /* 1837 * Sanity check: if the connection with our source 1838 * port is going to some other server, something went 1839 * wrong, as we never delete connections (i.e. release 1840 * ports) unless they have been idle. In this case, 1841 * it is probably better to send the call out using 1842 * a new source address than to fail it altogether, 1843 * since that port may never be released. 1844 */ 1845 if (destaddr->len != cm_entry->x_server.len || 1846 bcmp(destaddr->buf, cm_entry->x_server.buf, 1847 destaddr->len) != 0) { 1848 RPCLOG(1, "connmgr_get: tiptr %p" 1849 " is going to a different server" 1850 " with the port that belongs" 1851 " to us!\n", (void *)cm_entry->x_tiptr); 1852 retryaddr = NULL; 1853 goto use_new_conn; 1854 } 1855 1856 /* 1857 * If the connection of interest is not connected and we 1858 * can't reconnect it, then the server is probably 1859 * still down. Return NULL to the caller and let it 1860 * retry later if it wants to. We have a delay so the 1861 * machine doesn't go into a tight retry loop. If the 1862 * entry was already connected, or the reconnected was 1863 * successful, return this entry. 1864 */ 1865 if (cm_entry->x_connected == FALSE) { 1866 return (connmgr_wrapconnect(cm_entry, 1867 waitp, destaddr, addrfmly, NULL, 1868 rpcerr, TRUE, nosignal)); 1869 } else { 1870 CONN_HOLD(cm_entry); 1871 1872 cm_entry->x_time = lbolt; 1873 mutex_exit(&connmgr_lock); 1874 RPCLOG(2, "connmgr_get: found old " 1875 "transport %p for retry\n", 1876 (void *)cm_entry); 1877 return (cm_entry); 1878 } 1879 } 1880 1881 /* 1882 * We cannot find an entry in the list for this retry. 1883 * Either the entry has been removed temporarily to be 1884 * reconnected by another thread, or the original call 1885 * got a port but never got connected, 1886 * and hence the transport never got put in the 1887 * list. Fall through to the "create new connection" code - 1888 * the former case will fail there trying to rebind the port, 1889 * and the later case (and any other pathological cases) will 1890 * rebind and reconnect and not hang the client machine. 1891 */ 1892 RPCLOG0(8, "connmgr_get: no entry in list for retry\n"); 1893 } 1894 /* 1895 * Set up a transport entry in the connection manager's list. 1896 */ 1897 cm_entry = (struct cm_xprt *) 1898 kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP); 1899 1900 cm_entry->x_server.buf = kmem_zalloc(destaddr->len, KM_SLEEP); 1901 bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len); 1902 cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len; 1903 1904 cm_entry->x_state_flags = X_THREAD; 1905 cm_entry->x_ref = 1; 1906 cm_entry->x_family = addrfmly; 1907 cm_entry->x_rdev = device; 1908 cm_entry->x_zoneid = zoneid; 1909 mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL); 1910 cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL); 1911 cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL); 1912 cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL); 1913 1914 /* 1915 * Note that we add this partially initialized entry to the 1916 * connection list. This is so that we don't have connections to 1917 * the same server. 1918 * 1919 * Note that x_src is not initialized at this point. This is because 1920 * retryaddr might be NULL in which case x_src is whatever 1921 * t_kbind/bindresvport gives us. If another thread wants a 1922 * connection to the same server, seemingly we have an issue, but we 1923 * don't. If the other thread comes in with retryaddr == NULL, then it 1924 * will never look at x_src, and it will end up waiting in 1925 * connmgr_cwait() for the first thread to finish the connection 1926 * attempt. If the other thread comes in with retryaddr != NULL, then 1927 * that means there was a request sent on a connection, in which case 1928 * the the connection should already exist. Thus the first thread 1929 * never gets here ... it finds the connection it its server in the 1930 * connection list. 1931 * 1932 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd 1933 * thread will skip us because x_src.len == 0. 1934 */ 1935 cm_entry->x_next = cm_hd; 1936 cm_hd = cm_entry; 1937 mutex_exit(&connmgr_lock); 1938 1939 /* 1940 * Either we didn't find an entry to the server of interest, or we 1941 * don't have the maximum number of connections to that server - 1942 * create a new connection. 1943 */ 1944 RPCLOG0(8, "connmgr_get: creating new connection\n"); 1945 rpcerr->re_status = RPC_TLIERROR; 1946 1947 i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, kcred); 1948 if (i) { 1949 RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i); 1950 rpcerr->re_errno = i; 1951 connmgr_cancelconn(cm_entry); 1952 return (NULL); 1953 } 1954 rpc_poptimod(tiptr->fp->f_vnode); 1955 1956 if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0, 1957 K_TO_K, kcred, &retval)) { 1958 RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i); 1959 (void) t_kclose(tiptr, 1); 1960 rpcerr->re_errno = i; 1961 connmgr_cancelconn(cm_entry); 1962 return (NULL); 1963 } 1964 1965 if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K, 1966 kcred, &retval)) { 1967 RPCLOG(1, "connmgr_get: can't set client status with cots " 1968 "module, %d\n", i); 1969 (void) t_kclose(tiptr, 1); 1970 rpcerr->re_errno = i; 1971 connmgr_cancelconn(cm_entry); 1972 return (NULL); 1973 } 1974 1975 mutex_enter(&connmgr_lock); 1976 1977 wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next; 1978 cm_entry->x_wq = wq; 1979 1980 mutex_exit(&connmgr_lock); 1981 1982 if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0, 1983 K_TO_K, kcred, &retval)) { 1984 RPCLOG(1, "connmgr_get: can't push timod, %d\n", i); 1985 (void) t_kclose(tiptr, 1); 1986 rpcerr->re_errno = i; 1987 connmgr_cancelconn(cm_entry); 1988 return (NULL); 1989 } 1990 1991 /* 1992 * If the caller has not specified reserved port usage then 1993 * take the system default. 1994 */ 1995 if (useresvport == -1) 1996 useresvport = clnt_cots_do_bindresvport; 1997 1998 if ((useresvport || retryaddr != NULL) && 1999 (addrfmly == AF_INET || addrfmly == AF_INET6)) { 2000 bool_t alloc_src = FALSE; 2001 2002 if (srcaddr->len != destaddr->len) { 2003 kmem_free(srcaddr->buf, srcaddr->maxlen); 2004 srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP); 2005 srcaddr->maxlen = destaddr->len; 2006 srcaddr->len = destaddr->len; 2007 alloc_src = TRUE; 2008 } 2009 2010 if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) { 2011 (void) t_kclose(tiptr, 1); 2012 RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: " 2013 "%p\n", (void *)retryaddr); 2014 2015 /* 2016 * 1225408: If we allocated a source address, then it 2017 * is either garbage or all zeroes. In that case 2018 * we need to clear srcaddr. 2019 */ 2020 if (alloc_src == TRUE) { 2021 kmem_free(srcaddr->buf, srcaddr->maxlen); 2022 srcaddr->maxlen = srcaddr->len = 0; 2023 srcaddr->buf = NULL; 2024 } 2025 rpcerr->re_errno = i; 2026 connmgr_cancelconn(cm_entry); 2027 return (NULL); 2028 } 2029 } else { 2030 if ((i = t_kbind(tiptr, NULL, NULL)) != 0) { 2031 RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i); 2032 (void) t_kclose(tiptr, 1); 2033 rpcerr->re_errno = i; 2034 connmgr_cancelconn(cm_entry); 2035 return (NULL); 2036 } 2037 } 2038 2039 { 2040 /* 2041 * Keep the kernel stack lean. Don't move this call 2042 * declaration to the top of this function because a 2043 * call is declared in connmgr_wrapconnect() 2044 */ 2045 calllist_t call; 2046 2047 bzero(&call, sizeof (call)); 2048 cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL); 2049 2050 /* 2051 * This is a bound end-point so don't close it's stream. 2052 */ 2053 connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly, 2054 &call, &tidu_size, FALSE, waitp, 2055 nosignal); 2056 *rpcerr = call.call_err; 2057 cv_destroy(&call.call_cv); 2058 2059 } 2060 2061 mutex_enter(&connmgr_lock); 2062 2063 /* 2064 * Set up a transport entry in the connection manager's list. 2065 */ 2066 cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP); 2067 bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len); 2068 cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len; 2069 2070 cm_entry->x_tiptr = tiptr; 2071 cm_entry->x_time = lbolt; 2072 2073 if (tiptr->tp_info.servtype == T_COTS_ORD) 2074 cm_entry->x_ordrel = TRUE; 2075 else 2076 cm_entry->x_ordrel = FALSE; 2077 2078 cm_entry->x_tidu_size = tidu_size; 2079 2080 if (cm_entry->x_early_disc) 2081 cm_entry->x_connected = FALSE; 2082 else 2083 cm_entry->x_connected = connected; 2084 2085 /* 2086 * There could be a discrepancy here such that 2087 * x_early_disc is TRUE yet connected is TRUE as well 2088 * and the connection is actually connected. In that case 2089 * lets be conservative and declare the connection as not 2090 * connected. 2091 */ 2092 cm_entry->x_early_disc = FALSE; 2093 cm_entry->x_needdis = (cm_entry->x_connected == FALSE); 2094 cm_entry->x_ctime = lbolt; 2095 2096 /* 2097 * Notify any threads waiting that the connection attempt is done. 2098 */ 2099 cm_entry->x_thread = FALSE; 2100 cv_broadcast(&cm_entry->x_conn_cv); 2101 2102 mutex_exit(&connmgr_lock); 2103 2104 if (cm_entry->x_connected == FALSE) { 2105 connmgr_release(cm_entry); 2106 return (NULL); 2107 } 2108 return (cm_entry); 2109 } 2110 2111 /* 2112 * Keep the cm_xprt entry on the connecton list when making a connection. This 2113 * is to prevent multiple connections to a slow server from appearing. 2114 * We use the bit field x_thread to tell if a thread is doing a connection 2115 * which keeps other interested threads from messing with connection. 2116 * Those other threads just wait if x_thread is set. 2117 * 2118 * If x_thread is not set, then we do the actual work of connecting via 2119 * connmgr_connect(). 2120 * 2121 * mutex convention: called with connmgr_lock held, returns with it released. 2122 */ 2123 static struct cm_xprt * 2124 connmgr_wrapconnect( 2125 struct cm_xprt *cm_entry, 2126 const struct timeval *waitp, 2127 struct netbuf *destaddr, 2128 int addrfmly, 2129 struct netbuf *srcaddr, 2130 struct rpc_err *rpcerr, 2131 bool_t reconnect, 2132 bool_t nosignal) 2133 { 2134 ASSERT(MUTEX_HELD(&connmgr_lock)); 2135 /* 2136 * Hold this entry as we are about to drop connmgr_lock. 2137 */ 2138 CONN_HOLD(cm_entry); 2139 2140 /* 2141 * If there is a thread already making a connection for us, then 2142 * wait for it to complete the connection. 2143 */ 2144 if (cm_entry->x_thread == TRUE) { 2145 rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal); 2146 2147 if (rpcerr->re_status != RPC_SUCCESS) { 2148 mutex_exit(&connmgr_lock); 2149 connmgr_release(cm_entry); 2150 return (NULL); 2151 } 2152 } else { 2153 bool_t connected; 2154 calllist_t call; 2155 2156 cm_entry->x_thread = TRUE; 2157 2158 while (cm_entry->x_needrel == TRUE) { 2159 cm_entry->x_needrel = FALSE; 2160 2161 connmgr_sndrel(cm_entry); 2162 delay(drv_usectohz(1000000)); 2163 2164 mutex_enter(&connmgr_lock); 2165 } 2166 2167 /* 2168 * If we need to send a T_DISCON_REQ, send one. 2169 */ 2170 connmgr_dis_and_wait(cm_entry); 2171 2172 mutex_exit(&connmgr_lock); 2173 2174 bzero(&call, sizeof (call)); 2175 cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL); 2176 2177 connected = connmgr_connect(cm_entry, cm_entry->x_wq, 2178 destaddr, addrfmly, &call, 2179 &cm_entry->x_tidu_size, 2180 reconnect, waitp, nosignal); 2181 2182 *rpcerr = call.call_err; 2183 cv_destroy(&call.call_cv); 2184 2185 mutex_enter(&connmgr_lock); 2186 2187 2188 if (cm_entry->x_early_disc) 2189 cm_entry->x_connected = FALSE; 2190 else 2191 cm_entry->x_connected = connected; 2192 2193 /* 2194 * There could be a discrepancy here such that 2195 * x_early_disc is TRUE yet connected is TRUE as well 2196 * and the connection is actually connected. In that case 2197 * lets be conservative and declare the connection as not 2198 * connected. 2199 */ 2200 2201 cm_entry->x_early_disc = FALSE; 2202 cm_entry->x_needdis = (cm_entry->x_connected == FALSE); 2203 2204 2205 /* 2206 * connmgr_connect() may have given up before the connection 2207 * actually timed out. So ensure that before the next 2208 * connection attempt we do a disconnect. 2209 */ 2210 cm_entry->x_ctime = lbolt; 2211 cm_entry->x_thread = FALSE; 2212 2213 cv_broadcast(&cm_entry->x_conn_cv); 2214 2215 if (cm_entry->x_connected == FALSE) { 2216 mutex_exit(&connmgr_lock); 2217 connmgr_release(cm_entry); 2218 return (NULL); 2219 } 2220 } 2221 2222 if (srcaddr != NULL) { 2223 /* 2224 * Copy into the handle the 2225 * source address of the 2226 * connection, which we will use 2227 * in case of a later retry. 2228 */ 2229 if (srcaddr->len != cm_entry->x_src.len) { 2230 if (srcaddr->maxlen > 0) 2231 kmem_free(srcaddr->buf, srcaddr->maxlen); 2232 srcaddr->buf = kmem_zalloc(cm_entry->x_src.len, 2233 KM_SLEEP); 2234 srcaddr->maxlen = srcaddr->len = 2235 cm_entry->x_src.len; 2236 } 2237 bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len); 2238 } 2239 cm_entry->x_time = lbolt; 2240 mutex_exit(&connmgr_lock); 2241 return (cm_entry); 2242 } 2243 2244 /* 2245 * If we need to send a T_DISCON_REQ, send one. 2246 */ 2247 static void 2248 connmgr_dis_and_wait(struct cm_xprt *cm_entry) 2249 { 2250 ASSERT(MUTEX_HELD(&connmgr_lock)); 2251 for (;;) { 2252 while (cm_entry->x_needdis == TRUE) { 2253 RPCLOG(8, "connmgr_dis_and_wait: need " 2254 "T_DISCON_REQ for connection 0x%p\n", 2255 (void *)cm_entry); 2256 cm_entry->x_needdis = FALSE; 2257 cm_entry->x_waitdis = TRUE; 2258 2259 connmgr_snddis(cm_entry); 2260 2261 mutex_enter(&connmgr_lock); 2262 } 2263 2264 if (cm_entry->x_waitdis == TRUE) { 2265 clock_t curlbolt; 2266 clock_t timout; 2267 2268 RPCLOG(8, "connmgr_dis_and_wait waiting for " 2269 "T_DISCON_REQ's ACK for connection %p\n", 2270 (void *)cm_entry); 2271 curlbolt = ddi_get_lbolt(); 2272 2273 timout = clnt_cots_min_conntout * 2274 drv_usectohz(1000000) + curlbolt; 2275 2276 /* 2277 * The TPI spec says that the T_DISCON_REQ 2278 * will get acknowledged, but in practice 2279 * the ACK may never get sent. So don't 2280 * block forever. 2281 */ 2282 (void) cv_timedwait(&cm_entry->x_dis_cv, 2283 &connmgr_lock, timout); 2284 } 2285 /* 2286 * If we got the ACK, break. If we didn't, 2287 * then send another T_DISCON_REQ. 2288 */ 2289 if (cm_entry->x_waitdis == FALSE) { 2290 break; 2291 } else { 2292 RPCLOG(8, "connmgr_dis_and_wait: did" 2293 "not get T_DISCON_REQ's ACK for " 2294 "connection %p\n", (void *)cm_entry); 2295 cm_entry->x_needdis = TRUE; 2296 } 2297 } 2298 } 2299 2300 static void 2301 connmgr_cancelconn(struct cm_xprt *cm_entry) 2302 { 2303 /* 2304 * Mark the connection table entry as dead; the next thread that 2305 * goes through connmgr_release() will notice this and deal with it. 2306 */ 2307 mutex_enter(&connmgr_lock); 2308 cm_entry->x_dead = TRUE; 2309 2310 /* 2311 * Notify any threads waiting for the connection that it isn't 2312 * going to happen. 2313 */ 2314 cm_entry->x_thread = FALSE; 2315 cv_broadcast(&cm_entry->x_conn_cv); 2316 mutex_exit(&connmgr_lock); 2317 2318 connmgr_release(cm_entry); 2319 } 2320 2321 static void 2322 connmgr_close(struct cm_xprt *cm_entry) 2323 { 2324 mutex_enter(&cm_entry->x_lock); 2325 while (cm_entry->x_ref != 0) { 2326 /* 2327 * Must be a noninterruptible wait. 2328 */ 2329 cv_wait(&cm_entry->x_cv, &cm_entry->x_lock); 2330 } 2331 2332 if (cm_entry->x_tiptr != NULL) 2333 (void) t_kclose(cm_entry->x_tiptr, 1); 2334 2335 mutex_exit(&cm_entry->x_lock); 2336 if (cm_entry->x_ksp != NULL) { 2337 mutex_enter(&connmgr_lock); 2338 cm_entry->x_ksp->ks_private = NULL; 2339 mutex_exit(&connmgr_lock); 2340 2341 /* 2342 * Must free the buffer we allocated for the 2343 * server address in the update function 2344 */ 2345 if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))-> 2346 x_server.value.string.addr.ptr != NULL) 2347 kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp-> 2348 ks_data))->x_server.value.string.addr.ptr, 2349 INET6_ADDRSTRLEN); 2350 kmem_free(cm_entry->x_ksp->ks_data, 2351 cm_entry->x_ksp->ks_data_size); 2352 kstat_delete(cm_entry->x_ksp); 2353 } 2354 2355 mutex_destroy(&cm_entry->x_lock); 2356 cv_destroy(&cm_entry->x_cv); 2357 cv_destroy(&cm_entry->x_conn_cv); 2358 cv_destroy(&cm_entry->x_dis_cv); 2359 2360 if (cm_entry->x_server.buf != NULL) 2361 kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen); 2362 if (cm_entry->x_src.buf != NULL) 2363 kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen); 2364 kmem_free(cm_entry, sizeof (struct cm_xprt)); 2365 } 2366 2367 /* 2368 * Called by KRPC after sending the call message to release the connection 2369 * it was using. 2370 */ 2371 static void 2372 connmgr_release(struct cm_xprt *cm_entry) 2373 { 2374 mutex_enter(&cm_entry->x_lock); 2375 cm_entry->x_ref--; 2376 if (cm_entry->x_ref == 0) 2377 cv_signal(&cm_entry->x_cv); 2378 mutex_exit(&cm_entry->x_lock); 2379 } 2380 2381 /* 2382 * Given an open stream, connect to the remote. Returns true if connected, 2383 * false otherwise. 2384 */ 2385 static bool_t 2386 connmgr_connect( 2387 struct cm_xprt *cm_entry, 2388 queue_t *wq, 2389 struct netbuf *addr, 2390 int addrfmly, 2391 calllist_t *e, 2392 int *tidu_ptr, 2393 bool_t reconnect, 2394 const struct timeval *waitp, 2395 bool_t nosignal) 2396 { 2397 mblk_t *mp; 2398 struct T_conn_req *tcr; 2399 struct T_info_ack *tinfo; 2400 int interrupted, error; 2401 int tidu_size, kstat_instance; 2402 2403 /* if it's a reconnect, flush any lingering data messages */ 2404 if (reconnect) 2405 (void) putctl1(wq, M_FLUSH, FLUSHRW); 2406 2407 mp = allocb(sizeof (*tcr) + addr->len, BPRI_LO); 2408 if (mp == NULL) { 2409 /* 2410 * This is unfortunate, but we need to look up the stats for 2411 * this zone to increment the "memory allocation failed" 2412 * counter. curproc->p_zone is safe since we're initiating a 2413 * connection and not in some strange streams context. 2414 */ 2415 struct rpcstat *rpcstat; 2416 2417 rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone); 2418 ASSERT(rpcstat != NULL); 2419 2420 RPCLOG0(1, "connmgr_connect: cannot alloc mp for " 2421 "sending conn request\n"); 2422 COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem); 2423 e->call_status = RPC_SYSTEMERROR; 2424 e->call_reason = ENOSR; 2425 return (FALSE); 2426 } 2427 2428 mp->b_datap->db_type = M_PROTO; 2429 tcr = (struct T_conn_req *)mp->b_rptr; 2430 bzero(tcr, sizeof (*tcr)); 2431 tcr->PRIM_type = T_CONN_REQ; 2432 tcr->DEST_length = addr->len; 2433 tcr->DEST_offset = sizeof (struct T_conn_req); 2434 mp->b_wptr = mp->b_rptr + sizeof (*tcr); 2435 2436 bcopy(addr->buf, mp->b_wptr, tcr->DEST_length); 2437 mp->b_wptr += tcr->DEST_length; 2438 2439 RPCLOG(8, "connmgr_connect: sending conn request on queue " 2440 "%p", (void *)wq); 2441 RPCLOG(8, " call %p\n", (void *)wq); 2442 /* 2443 * We use the entry in the handle that is normally used for 2444 * waiting for RPC replies to wait for the connection accept. 2445 */ 2446 clnt_dispatch_send(wq, mp, e, 0, 0); 2447 2448 mutex_enter(&clnt_pending_lock); 2449 2450 /* 2451 * We wait for the transport connection to be made, or an 2452 * indication that it could not be made. 2453 */ 2454 interrupted = 0; 2455 2456 /* 2457 * waitforack should have been called with T_OK_ACK, but the 2458 * present implementation needs to be passed T_INFO_ACK to 2459 * work correctly. 2460 */ 2461 error = waitforack(e, T_INFO_ACK, waitp, nosignal); 2462 if (error == EINTR) 2463 interrupted = 1; 2464 if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) { 2465 /* 2466 * No time to lose; we essentially have been signaled to 2467 * quit. 2468 */ 2469 interrupted = 1; 2470 } 2471 #ifdef RPCDEBUG 2472 if (error == ETIME) 2473 RPCLOG0(8, "connmgr_connect: giving up " 2474 "on connection attempt; " 2475 "clnt_dispatch notifyconn " 2476 "diagnostic 'no one waiting for " 2477 "connection' should not be " 2478 "unexpected\n"); 2479 #endif 2480 if (e->call_prev) 2481 e->call_prev->call_next = e->call_next; 2482 else 2483 clnt_pending = e->call_next; 2484 if (e->call_next) 2485 e->call_next->call_prev = e->call_prev; 2486 mutex_exit(&clnt_pending_lock); 2487 2488 if (e->call_status != RPC_SUCCESS || error != 0) { 2489 if (interrupted) 2490 e->call_status = RPC_INTR; 2491 else if (error == ETIME) 2492 e->call_status = RPC_TIMEDOUT; 2493 else if (error == EPROTO) 2494 e->call_status = RPC_SYSTEMERROR; 2495 2496 RPCLOG(8, "connmgr_connect: can't connect, status: " 2497 "%s\n", clnt_sperrno(e->call_status)); 2498 2499 if (e->call_reply) { 2500 freemsg(e->call_reply); 2501 e->call_reply = NULL; 2502 } 2503 2504 return (FALSE); 2505 } 2506 /* 2507 * The result of the "connection accept" is a T_info_ack 2508 * in the call_reply field. 2509 */ 2510 ASSERT(e->call_reply != NULL); 2511 mp = e->call_reply; 2512 e->call_reply = NULL; 2513 tinfo = (struct T_info_ack *)mp->b_rptr; 2514 2515 tidu_size = tinfo->TIDU_size; 2516 tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT); 2517 if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0)) 2518 tidu_size = COTS_DEFAULT_ALLOCSIZE; 2519 *tidu_ptr = tidu_size; 2520 2521 freemsg(mp); 2522 2523 /* 2524 * Set up the pertinent options. NODELAY is so the transport doesn't 2525 * buffer up RPC messages on either end. This may not be valid for 2526 * all transports. Failure to set this option is not cause to 2527 * bail out so we return success anyway. Note that lack of NODELAY 2528 * or some other way to flush the message on both ends will cause 2529 * lots of retries and terrible performance. 2530 */ 2531 if (addrfmly == AF_INET || addrfmly == AF_INET6) { 2532 (void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e); 2533 if (e->call_status == RPC_XPRTFAILED) 2534 return (FALSE); 2535 } 2536 2537 /* 2538 * Since we have a connection, we now need to figure out if 2539 * we need to create a kstat. If x_ksp is not NULL then we 2540 * are reusing a connection and so we do not need to create 2541 * another kstat -- lets just return. 2542 */ 2543 if (cm_entry->x_ksp != NULL) 2544 return (TRUE); 2545 2546 /* 2547 * We need to increment rpc_kstat_instance atomically to prevent 2548 * two kstats being created with the same instance. 2549 */ 2550 kstat_instance = atomic_add_32_nv((uint32_t *)&rpc_kstat_instance, 1); 2551 2552 if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance, 2553 "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED, 2554 (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)), 2555 KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) { 2556 return (TRUE); 2557 } 2558 2559 cm_entry->x_ksp->ks_lock = &connmgr_lock; 2560 cm_entry->x_ksp->ks_private = cm_entry; 2561 cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char)) 2562 + sizeof (cm_kstat_template)); 2563 cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size, 2564 KM_SLEEP); 2565 bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data, 2566 cm_entry->x_ksp->ks_data_size); 2567 ((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))-> 2568 x_server.value.string.addr.ptr = 2569 kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP); 2570 2571 cm_entry->x_ksp->ks_update = conn_kstat_update; 2572 kstat_install(cm_entry->x_ksp); 2573 return (TRUE); 2574 } 2575 2576 /* 2577 * Called by connmgr_connect to set an option on the new stream. 2578 */ 2579 static bool_t 2580 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e) 2581 { 2582 mblk_t *mp; 2583 struct opthdr *opt; 2584 struct T_optmgmt_req *tor; 2585 struct timeval waitp; 2586 int error; 2587 2588 mp = allocb(sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) + 2589 sizeof (int), BPRI_LO); 2590 if (mp == NULL) { 2591 RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option " 2592 "request\n"); 2593 return (FALSE); 2594 } 2595 2596 mp->b_datap->db_type = M_PROTO; 2597 tor = (struct T_optmgmt_req *)(mp->b_rptr); 2598 tor->PRIM_type = T_SVR4_OPTMGMT_REQ; 2599 tor->MGMT_flags = T_NEGOTIATE; 2600 tor->OPT_length = sizeof (struct opthdr) + sizeof (int); 2601 tor->OPT_offset = sizeof (struct T_optmgmt_req); 2602 2603 opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req)); 2604 opt->level = level; 2605 opt->name = name; 2606 opt->len = sizeof (int); 2607 *(int *)((char *)opt + sizeof (*opt)) = 1; 2608 mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) + 2609 sizeof (int); 2610 2611 /* 2612 * We will use this connection regardless 2613 * of whether or not the option is settable. 2614 */ 2615 clnt_dispatch_send(wq, mp, e, 0, 0); 2616 mutex_enter(&clnt_pending_lock); 2617 2618 waitp.tv_sec = clnt_cots_min_conntout; 2619 waitp.tv_usec = 0; 2620 error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1); 2621 2622 if (e->call_prev) 2623 e->call_prev->call_next = e->call_next; 2624 else 2625 clnt_pending = e->call_next; 2626 if (e->call_next) 2627 e->call_next->call_prev = e->call_prev; 2628 mutex_exit(&clnt_pending_lock); 2629 2630 if (e->call_reply != NULL) { 2631 freemsg(e->call_reply); 2632 e->call_reply = NULL; 2633 } 2634 2635 if (e->call_status != RPC_SUCCESS || error != 0) { 2636 RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name); 2637 return (FALSE); 2638 } 2639 RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name); 2640 return (TRUE); 2641 } 2642 2643 #ifdef DEBUG 2644 2645 /* 2646 * This is a knob to let us force code coverage in allocation failure 2647 * case. 2648 */ 2649 static int connmgr_failsnd; 2650 #define CONN_SND_ALLOC(Size, Pri) \ 2651 ((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri)) 2652 2653 #else 2654 2655 #define CONN_SND_ALLOC(Size, Pri) allocb(Size, Pri) 2656 2657 #endif 2658 2659 /* 2660 * Sends an orderly release on the specified queue. 2661 * Entered with connmgr_lock. Exited without connmgr_lock 2662 */ 2663 static void 2664 connmgr_sndrel(struct cm_xprt *cm_entry) 2665 { 2666 struct T_ordrel_req *torr; 2667 mblk_t *mp; 2668 queue_t *q = cm_entry->x_wq; 2669 ASSERT(MUTEX_HELD(&connmgr_lock)); 2670 mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO); 2671 if (mp == NULL) { 2672 cm_entry->x_needrel = TRUE; 2673 mutex_exit(&connmgr_lock); 2674 RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel " 2675 "to queue %p\n", (void *)q); 2676 return; 2677 } 2678 mutex_exit(&connmgr_lock); 2679 2680 mp->b_datap->db_type = M_PROTO; 2681 torr = (struct T_ordrel_req *)(mp->b_rptr); 2682 torr->PRIM_type = T_ORDREL_REQ; 2683 mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req); 2684 2685 RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q); 2686 put(q, mp); 2687 } 2688 2689 /* 2690 * Sends an disconnect on the specified queue. 2691 * Entered with connmgr_lock. Exited without connmgr_lock 2692 */ 2693 static void 2694 connmgr_snddis(struct cm_xprt *cm_entry) 2695 { 2696 struct T_discon_req *tdis; 2697 mblk_t *mp; 2698 queue_t *q = cm_entry->x_wq; 2699 2700 ASSERT(MUTEX_HELD(&connmgr_lock)); 2701 mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO); 2702 if (mp == NULL) { 2703 cm_entry->x_needdis = TRUE; 2704 mutex_exit(&connmgr_lock); 2705 RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon " 2706 "to queue %p\n", (void *)q); 2707 return; 2708 } 2709 mutex_exit(&connmgr_lock); 2710 2711 mp->b_datap->db_type = M_PROTO; 2712 tdis = (struct T_discon_req *)mp->b_rptr; 2713 tdis->PRIM_type = T_DISCON_REQ; 2714 mp->b_wptr = mp->b_rptr + sizeof (*tdis); 2715 2716 RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q); 2717 put(q, mp); 2718 } 2719 2720 /* 2721 * Sets up the entry for receiving replies, and calls rpcmod's write put proc 2722 * (through put) to send the call. 2723 */ 2724 static void 2725 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid, 2726 uint_t queue_flag) 2727 { 2728 ASSERT(e != NULL); 2729 2730 e->call_status = RPC_TIMEDOUT; /* optimistic, eh? */ 2731 e->call_reason = 0; 2732 e->call_wq = q; 2733 e->call_xid = xid; 2734 e->call_notified = FALSE; 2735 2736 /* 2737 * If queue_flag is set then the calllist_t is already on the hash 2738 * queue. In this case just send the message and return. 2739 */ 2740 if (queue_flag) { 2741 put(q, mp); 2742 return; 2743 } 2744 2745 /* 2746 * Set up calls for RPC requests (with XID != 0) on the hash 2747 * queue for fast lookups and place other calls (i.e. 2748 * connection management) on the linked list. 2749 */ 2750 if (xid != 0) { 2751 RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on " 2752 "dispatch list\n", xid); 2753 e->call_hash = call_hash(xid, clnt_cots_hash_size); 2754 e->call_bucket = &cots_call_ht[e->call_hash]; 2755 call_table_enter(e); 2756 } else { 2757 mutex_enter(&clnt_pending_lock); 2758 if (clnt_pending) 2759 clnt_pending->call_prev = e; 2760 e->call_next = clnt_pending; 2761 e->call_prev = NULL; 2762 clnt_pending = e; 2763 mutex_exit(&clnt_pending_lock); 2764 } 2765 2766 put(q, mp); 2767 } 2768 2769 /* 2770 * Called by rpcmod to notify a client with a clnt_pending call that its reply 2771 * has arrived. If we can't find a client waiting for this reply, we log 2772 * the error and return. 2773 */ 2774 bool_t 2775 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid) 2776 { 2777 calllist_t *e = NULL; 2778 call_table_t *chtp; 2779 uint32_t xid; 2780 uint_t hash; 2781 2782 if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) && 2783 (mp->b_wptr - mp->b_rptr) >= sizeof (xid)) 2784 xid = *((uint32_t *)mp->b_rptr); 2785 else { 2786 int i = 0; 2787 unsigned char *p = (unsigned char *)&xid; 2788 unsigned char *rptr; 2789 mblk_t *tmp = mp; 2790 2791 /* 2792 * Copy the xid, byte-by-byte into xid. 2793 */ 2794 while (tmp) { 2795 rptr = tmp->b_rptr; 2796 while (rptr < tmp->b_wptr) { 2797 *p++ = *rptr++; 2798 if (++i >= sizeof (xid)) 2799 goto done_xid_copy; 2800 } 2801 tmp = tmp->b_cont; 2802 } 2803 2804 /* 2805 * If we got here, we ran out of mblk space before the 2806 * xid could be copied. 2807 */ 2808 ASSERT(tmp == NULL && i < sizeof (xid)); 2809 2810 RPCLOG0(1, 2811 "clnt_dispatch_notify: message less than size of xid\n"); 2812 return (FALSE); 2813 2814 } 2815 done_xid_copy: 2816 2817 hash = call_hash(xid, clnt_cots_hash_size); 2818 chtp = &cots_call_ht[hash]; 2819 /* call_table_find returns with the hash bucket locked */ 2820 call_table_find(chtp, xid, e); 2821 2822 if (e != NULL) { 2823 /* 2824 * Found thread waiting for this reply 2825 */ 2826 mutex_enter(&e->call_lock); 2827 if (e->call_reply) 2828 /* 2829 * This can happen under the following scenario: 2830 * clnt_cots_kcallit() times out on the response, 2831 * rfscall() repeats the CLNT_CALL() with 2832 * the same xid, clnt_cots_kcallit() sends the retry, 2833 * thereby putting the clnt handle on the pending list, 2834 * the first response arrives, signalling the thread 2835 * in clnt_cots_kcallit(). Before that thread is 2836 * dispatched, the second response arrives as well, 2837 * and clnt_dispatch_notify still finds the handle on 2838 * the pending list, with call_reply set. So free the 2839 * old reply now. 2840 * 2841 * It is also possible for a response intended for 2842 * an RPC call with a different xid to reside here. 2843 * This can happen if the thread that owned this 2844 * client handle prior to the current owner bailed 2845 * out and left its call record on the dispatch 2846 * queue. A window exists where the response can 2847 * arrive before the current owner dispatches its 2848 * RPC call. 2849 * 2850 * In any case, this is the very last point where we 2851 * can safely check the call_reply field before 2852 * placing the new response there. 2853 */ 2854 freemsg(e->call_reply); 2855 e->call_reply = mp; 2856 e->call_status = RPC_SUCCESS; 2857 e->call_notified = TRUE; 2858 cv_signal(&e->call_cv); 2859 mutex_exit(&e->call_lock); 2860 mutex_exit(&chtp->ct_lock); 2861 return (TRUE); 2862 } else { 2863 zone_t *zone; 2864 struct rpcstat *rpcstat; 2865 2866 mutex_exit(&chtp->ct_lock); 2867 RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n", 2868 xid); 2869 /* 2870 * This is unfortunate, but we need to lookup the zone so we 2871 * can increment its "rcbadxids" counter. 2872 */ 2873 zone = zone_find_by_id(zoneid); 2874 if (zone == NULL) { 2875 /* 2876 * The zone went away... 2877 */ 2878 return (FALSE); 2879 } 2880 rpcstat = zone_getspecific(rpcstat_zone_key, zone); 2881 if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) { 2882 /* 2883 * Not interested 2884 */ 2885 zone_rele(zone); 2886 return (FALSE); 2887 } 2888 COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids); 2889 zone_rele(zone); 2890 } 2891 return (FALSE); 2892 } 2893 2894 /* 2895 * Called by rpcmod when a non-data indication arrives. The ones in which we 2896 * are interested are connection indications and options acks. We dispatch 2897 * based on the queue the indication came in on. If we are not interested in 2898 * what came in, we return false to rpcmod, who will then pass it upstream. 2899 */ 2900 bool_t 2901 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp) 2902 { 2903 calllist_t *e; 2904 int type; 2905 2906 ASSERT((q->q_flag & QREADR) == 0); 2907 2908 type = ((union T_primitives *)mp->b_rptr)->type; 2909 RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n", 2910 rpc_tpiprim2name(type)); 2911 mutex_enter(&clnt_pending_lock); 2912 for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) { 2913 if (e == NULL) { 2914 mutex_exit(&clnt_pending_lock); 2915 RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting " 2916 "for connection on queue 0x%p\n", (void *)q); 2917 return (FALSE); 2918 } 2919 if (e->call_wq == q) 2920 break; 2921 } 2922 2923 switch (type) { 2924 case T_CONN_CON: 2925 /* 2926 * The transport is now connected, send a T_INFO_REQ to get 2927 * the tidu size. 2928 */ 2929 mutex_exit(&clnt_pending_lock); 2930 ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >= 2931 sizeof (struct T_info_req)); 2932 mp->b_rptr = mp->b_datap->db_base; 2933 ((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ; 2934 mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req); 2935 mp->b_datap->db_type = M_PCPROTO; 2936 put(q, mp); 2937 return (TRUE); 2938 case T_INFO_ACK: 2939 case T_OPTMGMT_ACK: 2940 e->call_status = RPC_SUCCESS; 2941 e->call_reply = mp; 2942 e->call_notified = TRUE; 2943 cv_signal(&e->call_cv); 2944 break; 2945 case T_ERROR_ACK: 2946 e->call_status = RPC_CANTCONNECT; 2947 e->call_reply = mp; 2948 e->call_notified = TRUE; 2949 cv_signal(&e->call_cv); 2950 break; 2951 case T_OK_ACK: 2952 /* 2953 * Great, but we are really waiting for a T_CONN_CON 2954 */ 2955 freemsg(mp); 2956 break; 2957 default: 2958 mutex_exit(&clnt_pending_lock); 2959 RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type); 2960 return (FALSE); 2961 } 2962 2963 mutex_exit(&clnt_pending_lock); 2964 return (TRUE); 2965 } 2966 2967 /* 2968 * Called by rpcmod when the transport is (or should be) going away. Informs 2969 * all callers waiting for replies and marks the entry in the connection 2970 * manager's list as unconnected, and either closing (close handshake in 2971 * progress) or dead. 2972 */ 2973 void 2974 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason) 2975 { 2976 calllist_t *e; 2977 call_table_t *ctp; 2978 struct cm_xprt *cm_entry; 2979 int have_connmgr_lock; 2980 int i; 2981 2982 ASSERT((q->q_flag & QREADR) == 0); 2983 2984 RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q); 2985 RPCLOG(1, " received a notifcation prim type [%s]", 2986 rpc_tpiprim2name(msg_type)); 2987 RPCLOG(1, " and reason %d\n", reason); 2988 2989 /* 2990 * Find the transport entry in the connection manager's list, close 2991 * the transport and delete the entry. In the case where rpcmod's 2992 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we 2993 * should gracefully close the connection. 2994 */ 2995 have_connmgr_lock = 1; 2996 mutex_enter(&connmgr_lock); 2997 for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) { 2998 ASSERT(cm_entry != cm_entry->x_next); 2999 if (cm_entry->x_wq == q) { 3000 ASSERT(MUTEX_HELD(&connmgr_lock)); 3001 ASSERT(have_connmgr_lock == 1); 3002 switch (msg_type) { 3003 case T_ORDREL_REQ: 3004 3005 if (cm_entry->x_dead) { 3006 RPCLOG(1, "idle timeout on dead " 3007 "connection: %p\n", 3008 (void *)cm_entry); 3009 if (clnt_stop_idle != NULL) 3010 (*clnt_stop_idle)(q); 3011 break; 3012 } 3013 3014 /* 3015 * Only mark the connection as dead if it is 3016 * connected and idle. 3017 * An unconnected connection has probably 3018 * gone idle because the server is down, 3019 * and when it comes back up there will be 3020 * retries that need to use that connection. 3021 */ 3022 if (cm_entry->x_connected || 3023 cm_entry->x_doomed) { 3024 if (cm_entry->x_ordrel) { 3025 if (cm_entry->x_closing == TRUE) { 3026 /* 3027 * The connection is obviously 3028 * wedged due to a bug or problem 3029 * with the transport. Mark it 3030 * as dead. Otherwise we can leak 3031 * connections. 3032 */ 3033 cm_entry->x_dead = TRUE; 3034 mutex_exit(&connmgr_lock); 3035 have_connmgr_lock = 0; 3036 if (clnt_stop_idle != NULL) 3037 (*clnt_stop_idle)(q); 3038 break; 3039 } 3040 cm_entry->x_closing = TRUE; 3041 connmgr_sndrel(cm_entry); 3042 have_connmgr_lock = 0; 3043 } else { 3044 cm_entry->x_dead = TRUE; 3045 mutex_exit(&connmgr_lock); 3046 have_connmgr_lock = 0; 3047 if (clnt_stop_idle != NULL) 3048 (*clnt_stop_idle)(q); 3049 } 3050 } else { 3051 /* 3052 * We don't mark the connection 3053 * as dead, but we turn off the 3054 * idle timer. 3055 */ 3056 mutex_exit(&connmgr_lock); 3057 have_connmgr_lock = 0; 3058 if (clnt_stop_idle != NULL) 3059 (*clnt_stop_idle)(q); 3060 RPCLOG(1, "clnt_dispatch_notifyall:" 3061 " ignoring timeout from rpcmod" 3062 " (q %p) because we are not " 3063 " connected\n", (void *)q); 3064 } 3065 break; 3066 case T_ORDREL_IND: 3067 /* 3068 * If this entry is marked closing, then we are 3069 * completing a close handshake, and the 3070 * connection is dead. Otherwise, the server is 3071 * trying to close. Since the server will not 3072 * be sending any more RPC replies, we abort 3073 * the connection, including flushing 3074 * any RPC requests that are in-transit. 3075 */ 3076 if (cm_entry->x_closing) { 3077 cm_entry->x_dead = TRUE; 3078 mutex_exit(&connmgr_lock); 3079 have_connmgr_lock = 0; 3080 if (clnt_stop_idle != NULL) 3081 (*clnt_stop_idle)(q); 3082 } else { 3083 /* 3084 * if we're getting a disconnect 3085 * before we've finished our 3086 * connect attempt, mark it for 3087 * later processing 3088 */ 3089 if (cm_entry->x_thread) 3090 cm_entry->x_early_disc = TRUE; 3091 else 3092 cm_entry->x_connected = FALSE; 3093 cm_entry->x_waitdis = TRUE; 3094 connmgr_snddis(cm_entry); 3095 have_connmgr_lock = 0; 3096 } 3097 break; 3098 3099 case T_ERROR_ACK: 3100 case T_OK_ACK: 3101 cm_entry->x_waitdis = FALSE; 3102 cv_signal(&cm_entry->x_dis_cv); 3103 mutex_exit(&connmgr_lock); 3104 return; 3105 3106 case T_DISCON_REQ: 3107 if (cm_entry->x_thread) 3108 cm_entry->x_early_disc = TRUE; 3109 else 3110 cm_entry->x_connected = FALSE; 3111 cm_entry->x_waitdis = TRUE; 3112 3113 connmgr_snddis(cm_entry); 3114 have_connmgr_lock = 0; 3115 break; 3116 3117 case T_DISCON_IND: 3118 default: 3119 /* 3120 * if we're getting a disconnect before 3121 * we've finished our connect attempt, 3122 * mark it for later processing 3123 */ 3124 if (cm_entry->x_closing) { 3125 cm_entry->x_dead = TRUE; 3126 mutex_exit(&connmgr_lock); 3127 have_connmgr_lock = 0; 3128 if (clnt_stop_idle != NULL) 3129 (*clnt_stop_idle)(q); 3130 } else { 3131 if (cm_entry->x_thread) { 3132 cm_entry->x_early_disc = TRUE; 3133 } else { 3134 cm_entry->x_dead = TRUE; 3135 cm_entry->x_connected = FALSE; 3136 } 3137 } 3138 break; 3139 } 3140 break; 3141 } 3142 } 3143 3144 if (have_connmgr_lock) 3145 mutex_exit(&connmgr_lock); 3146 3147 if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) { 3148 RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find " 3149 "connmgr entry for discon ack\n", (void *)q); 3150 return; 3151 } 3152 3153 /* 3154 * Then kick all the clnt_pending calls out of their wait. There 3155 * should be no clnt_pending calls in the case of rpcmod's idle 3156 * timer firing. 3157 */ 3158 for (i = 0; i < clnt_cots_hash_size; i++) { 3159 ctp = &cots_call_ht[i]; 3160 mutex_enter(&ctp->ct_lock); 3161 for (e = ctp->ct_call_next; 3162 e != (calllist_t *)ctp; 3163 e = e->call_next) { 3164 if (e->call_wq == q && e->call_notified == FALSE) { 3165 RPCLOG(1, 3166 "clnt_dispatch_notifyall for queue %p ", 3167 (void *)q); 3168 RPCLOG(1, "aborting clnt_pending call %p\n", 3169 (void *)e); 3170 3171 if (msg_type == T_DISCON_IND) 3172 e->call_reason = reason; 3173 e->call_notified = TRUE; 3174 e->call_status = RPC_XPRTFAILED; 3175 cv_signal(&e->call_cv); 3176 } 3177 } 3178 mutex_exit(&ctp->ct_lock); 3179 } 3180 3181 mutex_enter(&clnt_pending_lock); 3182 for (e = clnt_pending; e; e = e->call_next) { 3183 /* 3184 * Only signal those RPC handles that haven't been 3185 * signalled yet. Otherwise we can get a bogus call_reason. 3186 * This can happen if thread A is making a call over a 3187 * connection. If the server is killed, it will cause 3188 * reset, and reason will default to EIO as a result of 3189 * a T_ORDREL_IND. Thread B then attempts to recreate 3190 * the connection but gets a T_DISCON_IND. If we set the 3191 * call_reason code for all threads, then if thread A 3192 * hasn't been dispatched yet, it will get the wrong 3193 * reason. The bogus call_reason can make it harder to 3194 * discriminate between calls that fail because the 3195 * connection attempt failed versus those where the call 3196 * may have been executed on the server. 3197 */ 3198 if (e->call_wq == q && e->call_notified == FALSE) { 3199 RPCLOG(1, "clnt_dispatch_notifyall for queue %p ", 3200 (void *)q); 3201 RPCLOG(1, " aborting clnt_pending call %p\n", 3202 (void *)e); 3203 3204 if (msg_type == T_DISCON_IND) 3205 e->call_reason = reason; 3206 e->call_notified = TRUE; 3207 /* 3208 * Let the caller timeout, else he will retry 3209 * immediately. 3210 */ 3211 e->call_status = RPC_XPRTFAILED; 3212 3213 /* 3214 * We used to just signal those threads 3215 * waiting for a connection, (call_xid = 0). 3216 * That meant that threads waiting for a response 3217 * waited till their timeout expired. This 3218 * could be a long time if they've specified a 3219 * maximum timeout. (2^31 - 1). So we 3220 * Signal all threads now. 3221 */ 3222 cv_signal(&e->call_cv); 3223 } 3224 } 3225 mutex_exit(&clnt_pending_lock); 3226 } 3227 3228 3229 /*ARGSUSED*/ 3230 /* 3231 * after resuming a system that's been suspended for longer than the 3232 * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall() 3233 * generates "NFS server X not responding" and "NFS server X ok" messages; 3234 * here we reset inet connections to cause a re-connect and avoid those 3235 * NFS messages. see 4045054 3236 */ 3237 boolean_t 3238 connmgr_cpr_reset(void *arg, int code) 3239 { 3240 struct cm_xprt *cxp; 3241 3242 if (code == CB_CODE_CPR_CHKPT) 3243 return (B_TRUE); 3244 3245 if (mutex_tryenter(&connmgr_lock) == 0) 3246 return (B_FALSE); 3247 for (cxp = cm_hd; cxp; cxp = cxp->x_next) { 3248 if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) && 3249 cxp->x_connected == TRUE) { 3250 if (cxp->x_thread) 3251 cxp->x_early_disc = TRUE; 3252 else 3253 cxp->x_connected = FALSE; 3254 cxp->x_needdis = TRUE; 3255 } 3256 } 3257 mutex_exit(&connmgr_lock); 3258 return (B_TRUE); 3259 } 3260 3261 void 3262 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp) 3263 { 3264 3265 *statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid, 3266 "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl, 3267 sizeof (cots_rcstat_tmpl)); 3268 } 3269 3270 void 3271 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp) 3272 { 3273 rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client"); 3274 kmem_free(*statsp, sizeof (cots_rcstat_tmpl)); 3275 } 3276 3277 void 3278 clnt_cots_init(void) 3279 { 3280 mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL); 3281 mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL); 3282 3283 if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE) 3284 clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE; 3285 3286 cots_call_ht = call_table_init(clnt_cots_hash_size); 3287 zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy); 3288 } 3289 3290 void 3291 clnt_cots_fini(void) 3292 { 3293 (void) zone_key_delete(zone_cots_key); 3294 } 3295 3296 /* 3297 * Wait for TPI ack, returns success only if expected ack is received 3298 * within timeout period. 3299 */ 3300 3301 static int 3302 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp, 3303 bool_t nosignal) 3304 { 3305 union T_primitives *tpr; 3306 clock_t timout; 3307 int cv_stat = 1; 3308 3309 ASSERT(MUTEX_HELD(&clnt_pending_lock)); 3310 while (e->call_reply == NULL) { 3311 if (waitp != NULL) { 3312 timout = waitp->tv_sec * drv_usectohz(MICROSEC) + 3313 drv_usectohz(waitp->tv_usec) + lbolt; 3314 if (nosignal) 3315 cv_stat = cv_timedwait(&e->call_cv, 3316 &clnt_pending_lock, timout); 3317 else 3318 cv_stat = cv_timedwait_sig(&e->call_cv, 3319 &clnt_pending_lock, timout); 3320 } else { 3321 if (nosignal) 3322 cv_wait(&e->call_cv, &clnt_pending_lock); 3323 else 3324 cv_stat = cv_wait_sig(&e->call_cv, 3325 &clnt_pending_lock); 3326 } 3327 if (cv_stat == -1) 3328 return (ETIME); 3329 if (cv_stat == 0) 3330 return (EINTR); 3331 } 3332 tpr = (union T_primitives *)e->call_reply->b_rptr; 3333 if (tpr->type == ack_prim) 3334 return (0); /* Success */ 3335 3336 if (tpr->type == T_ERROR_ACK) { 3337 if (tpr->error_ack.TLI_error == TSYSERR) 3338 return (tpr->error_ack.UNIX_error); 3339 else 3340 return (t_tlitosyserr(tpr->error_ack.TLI_error)); 3341 } 3342 3343 return (EPROTO); /* unknown or unexpected primitive */ 3344 } 3345