xref: /titanic_51/usr/src/uts/common/rpc/clnt_cots.c (revision 5d54f3d8999eac1762fe0a8c7177d20f1f201fae)
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 /*
494  * The following is used to determine the global default behavior for
495  * COTS when binding to a local port.
496  *
497  * If the value is set to 1 the default will be to select a reserved
498  * (aka privileged) port, if the value is zero the default will be to
499  * use non-reserved ports.  Users of kRPC may override this by using
500  * CLNT_CONTROL() and CLSET_BINDRESVPORT.
501  */
502 static int clnt_cots_do_bindresvport = 1;
503 
504 static zone_key_t zone_cots_key;
505 
506 /*
507  * We need to do this after all kernel threads in the zone have exited.
508  */
509 /* ARGSUSED */
510 static void
511 clnt_zone_destroy(zoneid_t zoneid, void *unused)
512 {
513 	struct cm_xprt **cmp;
514 	struct cm_xprt *cm_entry;
515 	struct cm_xprt *freelist = NULL;
516 
517 	mutex_enter(&connmgr_lock);
518 	cmp = &cm_hd;
519 	while ((cm_entry = *cmp) != NULL) {
520 		if (cm_entry->x_zoneid == zoneid) {
521 			*cmp = cm_entry->x_next;
522 			cm_entry->x_next = freelist;
523 			freelist = cm_entry;
524 		} else {
525 			cmp = &cm_entry->x_next;
526 		}
527 	}
528 	mutex_exit(&connmgr_lock);
529 	while ((cm_entry = freelist) != NULL) {
530 		freelist = cm_entry->x_next;
531 		connmgr_close(cm_entry);
532 	}
533 }
534 
535 int
536 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog,
537 	rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl)
538 {
539 	CLIENT *h;
540 	cku_private_t *p;
541 	struct rpc_msg call_msg;
542 	struct rpcstat *rpcstat;
543 
544 	RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog);
545 
546 	rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone);
547 	ASSERT(rpcstat != NULL);
548 
549 	/* Allocate and intialize the client handle. */
550 	p = kmem_zalloc(sizeof (*p), KM_SLEEP);
551 
552 	h = ptoh(p);
553 
554 	h->cl_private = (caddr_t)p;
555 	h->cl_auth = authkern_create();
556 	h->cl_ops = &tcp_ops;
557 
558 	cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL);
559 	mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL);
560 
561 	/*
562 	 * If the current sanity check size in rpcmod is smaller
563 	 * than the size needed, then increase the sanity check.
564 	 */
565 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
566 	    max_msgsize > *clnt_max_msg_sizep) {
567 		mutex_enter(&clnt_max_msg_lock);
568 		if (max_msgsize > *clnt_max_msg_sizep)
569 			*clnt_max_msg_sizep = max_msgsize;
570 		mutex_exit(&clnt_max_msg_lock);
571 	}
572 
573 	p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE;
574 
575 	/* Preserialize the call message header */
576 
577 	call_msg.rm_xid = 0;
578 	call_msg.rm_direction = CALL;
579 	call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION;
580 	call_msg.rm_call.cb_prog = prog;
581 	call_msg.rm_call.cb_vers = vers;
582 
583 	xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE);
584 
585 	if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) {
586 		RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization "
587 		    "error\n");
588 		auth_destroy(h->cl_auth);
589 		kmem_free(p, sizeof (cku_private_t));
590 		RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n");
591 		return (EINVAL);		/* XXX */
592 	}
593 
594 	/*
595 	 * The zalloc initialized the fields below.
596 	 * p->cku_xid = 0;
597 	 * p->cku_flags = 0;
598 	 * p->cku_srcaddr.len = 0;
599 	 * p->cku_srcaddr.maxlen = 0;
600 	 */
601 
602 	p->cku_cred = cred;
603 	p->cku_device = dev;
604 	p->cku_addrfmly = family;
605 	p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
606 	p->cku_addr.maxlen = addr->maxlen;
607 	p->cku_addr.len = addr->len;
608 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
609 	p->cku_stats = rpcstat->rpc_cots_client;
610 	p->cku_useresvport = -1; /* value is has not been set */
611 
612 	*ncl = h;
613 	return (0);
614 }
615 
616 /*ARGSUSED*/
617 static void
618 clnt_cots_kabort(CLIENT *h)
619 {
620 }
621 
622 /*
623  * Return error info on this handle.
624  */
625 static void
626 clnt_cots_kerror(CLIENT *h, struct rpc_err *err)
627 {
628 	/* LINTED pointer alignment */
629 	cku_private_t *p = htop(h);
630 
631 	*err = p->cku_err;
632 }
633 
634 static bool_t
635 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr)
636 {
637 	/* LINTED pointer alignment */
638 	cku_private_t *p = htop(h);
639 	XDR *xdrs;
640 
641 	xdrs = &(p->cku_outxdr);
642 	xdrs->x_op = XDR_FREE;
643 	return ((*xdr_res)(xdrs, res_ptr));
644 }
645 
646 static bool_t
647 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg)
648 {
649 	cku_private_t *p = htop(h);
650 
651 	switch (cmd) {
652 	case CLSET_PROGRESS:
653 		p->cku_progress = TRUE;
654 		return (TRUE);
655 
656 	case CLSET_XID:
657 		if (arg == NULL)
658 			return (FALSE);
659 
660 		p->cku_xid = *((uint32_t *)arg);
661 		return (TRUE);
662 
663 	case CLGET_XID:
664 		if (arg == NULL)
665 			return (FALSE);
666 
667 		*((uint32_t *)arg) = p->cku_xid;
668 		return (TRUE);
669 
670 	case CLSET_NODELAYONERR:
671 		if (arg == NULL)
672 			return (FALSE);
673 
674 		if (*((bool_t *)arg) == TRUE) {
675 			p->cku_nodelayonerr = TRUE;
676 			return (TRUE);
677 		}
678 		if (*((bool_t *)arg) == FALSE) {
679 			p->cku_nodelayonerr = FALSE;
680 			return (TRUE);
681 		}
682 		return (FALSE);
683 
684 	case CLGET_NODELAYONERR:
685 		if (arg == NULL)
686 			return (FALSE);
687 
688 		*((bool_t *)arg) = p->cku_nodelayonerr;
689 		return (TRUE);
690 
691 	case CLSET_BINDRESVPORT:
692 		if (arg == NULL)
693 			return (FALSE);
694 
695 		if (*(int *)arg != 1 && *(int *)arg != 0)
696 			return (FALSE);
697 
698 		p->cku_useresvport = *(int *)arg;
699 
700 		return (TRUE);
701 
702 	case CLGET_BINDRESVPORT:
703 		if (arg == NULL)
704 			return (FALSE);
705 
706 		*(int *)arg = p->cku_useresvport;
707 
708 		return (TRUE);
709 
710 	default:
711 		return (FALSE);
712 	}
713 }
714 
715 /*
716  * Destroy rpc handle.  Frees the space used for output buffer,
717  * private data, and handle structure.
718  */
719 static void
720 clnt_cots_kdestroy(CLIENT *h)
721 {
722 	/* LINTED pointer alignment */
723 	cku_private_t *p = htop(h);
724 	calllist_t *call = &p->cku_call;
725 
726 	RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h);
727 	RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid);
728 
729 	if (p->cku_flags & CKU_ONQUEUE) {
730 		RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x "
731 		    "from dispatch list\n", p->cku_xid);
732 		call_table_remove(call);
733 	}
734 
735 	if (call->call_reply)
736 		freemsg(call->call_reply);
737 	cv_destroy(&call->call_cv);
738 	mutex_destroy(&call->call_lock);
739 
740 	kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen);
741 	kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
742 	kmem_free(p, sizeof (*p));
743 }
744 
745 static int clnt_cots_pulls;
746 #define	RM_HDR_SIZE	4	/* record mark header size */
747 
748 /*
749  * Call remote procedure.
750  */
751 static enum clnt_stat
752 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
753     caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait)
754 {
755 	/* LINTED pointer alignment */
756 	cku_private_t *p = htop(h);
757 	calllist_t *call = &p->cku_call;
758 	XDR *xdrs;
759 	struct rpc_msg reply_msg;
760 	mblk_t *mp;
761 #ifdef	RPCDEBUG
762 	clock_t time_sent;
763 #endif
764 	struct netbuf *retryaddr;
765 	struct cm_xprt *cm_entry = NULL;
766 	queue_t *wq;
767 	int len;
768 	int mpsize;
769 	int refreshes = REFRESHES;
770 	int interrupted;
771 	int tidu_size;
772 	enum clnt_stat status;
773 	struct timeval cwait;
774 	bool_t delay_first = FALSE;
775 	clock_t ticks;
776 
777 	RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum);
778 	COTSRCSTAT_INCR(p->cku_stats, rccalls);
779 
780 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec);
781 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec);
782 
783 	/*
784 	 * Bug ID 1240234:
785 	 * Look out for zero length timeouts. We don't want to
786 	 * wait zero seconds for a connection to be established.
787 	 */
788 	if (wait.tv_sec < clnt_cots_min_conntout) {
789 		cwait.tv_sec = clnt_cots_min_conntout;
790 		cwait.tv_usec = 0;
791 		RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,",
792 		    wait.tv_sec);
793 		RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout);
794 	} else {
795 		cwait = wait;
796 	}
797 
798 call_again:
799 	if (cm_entry) {
800 		connmgr_release(cm_entry);
801 		cm_entry = NULL;
802 	}
803 
804 	mp = NULL;
805 
806 	/*
807 	 * If the call is not a retry, allocate a new xid and cache it
808 	 * for future retries.
809 	 * Bug ID 1246045:
810 	 * Treat call as a retry for purposes of binding the source
811 	 * port only if we actually attempted to send anything on
812 	 * the previous call.
813 	 */
814 	if (p->cku_xid == 0) {
815 		p->cku_xid = alloc_xid();
816 		/*
817 		 * We need to ASSERT here that our xid != 0 because this
818 		 * determines whether or not our call record gets placed on
819 		 * the hash table or the linked list.  By design, we mandate
820 		 * that RPC calls over cots must have xid's != 0, so we can
821 		 * ensure proper management of the hash table.
822 		 */
823 		ASSERT(p->cku_xid != 0);
824 
825 		retryaddr = NULL;
826 		p->cku_flags &= ~CKU_SENT;
827 
828 		if (p->cku_flags & CKU_ONQUEUE) {
829 			RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old"
830 			    " one (%p)\n", (void *)call);
831 			call_table_remove(call);
832 			p->cku_flags &= ~CKU_ONQUEUE;
833 			RPCLOG(64, "clnt_cots_kcallit: removing call from "
834 			    "dispatch list because xid was zero (now 0x%x)\n",
835 			    p->cku_xid);
836 		}
837 
838 		if (call->call_reply != NULL) {
839 			freemsg(call->call_reply);
840 			call->call_reply = NULL;
841 		}
842 	} else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) {
843 		retryaddr = NULL;
844 
845 	} else if (p->cku_flags & CKU_SENT) {
846 		retryaddr = &p->cku_srcaddr;
847 
848 	} else {
849 		/*
850 		 * Bug ID 1246045: Nothing was sent, so set retryaddr to
851 		 * NULL and let connmgr_get() bind to any source port it
852 		 * can get.
853 		 */
854 		retryaddr = NULL;
855 	}
856 
857 	RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid);
858 	RPCLOG(64, " flags = 0x%x\n", p->cku_flags);
859 
860 	p->cku_err.re_status = RPC_TIMEDOUT;
861 	p->cku_err.re_errno = p->cku_err.re_terrno = 0;
862 
863 	cm_entry = connmgr_wrapget(retryaddr, &cwait, p);
864 
865 	if (cm_entry == NULL) {
866 		RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n",
867 		    clnt_sperrno(p->cku_err.re_status));
868 
869 		/*
870 		 * The reasons why we fail to create a connection are
871 		 * varied. In most cases we don't want the caller to
872 		 * immediately retry. This could have one or more
873 		 * bad effects. This includes flooding the net with
874 		 * connect requests to ports with no listener; a hard
875 		 * kernel loop due to all the "reserved" TCP ports being
876 		 * in use.
877 		 */
878 		delay_first = TRUE;
879 
880 		/*
881 		 * Even if we end up returning EINTR, we still count a
882 		 * a "can't connect", because the connection manager
883 		 * might have been committed to waiting for or timing out on
884 		 * a connection.
885 		 */
886 		COTSRCSTAT_INCR(p->cku_stats, rccantconn);
887 		switch (p->cku_err.re_status) {
888 		case RPC_INTR:
889 			p->cku_err.re_errno = EINTR;
890 
891 			/*
892 			 * No need to delay because a UNIX signal(2)
893 			 * interrupted us. The caller likely won't
894 			 * retry the CLNT_CALL() and even if it does,
895 			 * we assume the caller knows what it is doing.
896 			 */
897 			delay_first = FALSE;
898 			break;
899 
900 		case RPC_TIMEDOUT:
901 			p->cku_err.re_errno = ETIMEDOUT;
902 
903 			/*
904 			 * No need to delay because timed out already
905 			 * on the connection request and assume that the
906 			 * transport time out is longer than our minimum
907 			 * timeout, or least not too much smaller.
908 			 */
909 			delay_first = FALSE;
910 			break;
911 
912 		case RPC_SYSTEMERROR:
913 		case RPC_TLIERROR:
914 			/*
915 			 * We want to delay here because a transient
916 			 * system error has a better chance of going away
917 			 * if we delay a bit. If it's not transient, then
918 			 * we don't want end up in a hard kernel loop
919 			 * due to retries.
920 			 */
921 			ASSERT(p->cku_err.re_errno != 0);
922 			break;
923 
924 
925 		case RPC_CANTCONNECT:
926 			/*
927 			 * RPC_CANTCONNECT is set on T_ERROR_ACK which
928 			 * implies some error down in the TCP layer or
929 			 * below. If cku_nodelayonerror is set then we
930 			 * assume the caller knows not to try too hard.
931 			 */
932 			RPCLOG0(8, "clnt_cots_kcallit: connection failed,");
933 			RPCLOG0(8, " re_status=RPC_CANTCONNECT,");
934 			RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno);
935 			RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr);
936 			if (p->cku_nodelayonerr == TRUE)
937 				delay_first = FALSE;
938 
939 			p->cku_err.re_errno = EIO;
940 
941 			break;
942 
943 		case RPC_XPRTFAILED:
944 			/*
945 			 * We want to delay here because we likely
946 			 * got a refused connection.
947 			 */
948 			if (p->cku_err.re_errno != 0)
949 				break;
950 
951 			/* fall thru */
952 
953 		default:
954 			/*
955 			 * We delay here because it is better to err
956 			 * on the side of caution. If we got here then
957 			 * status could have been RPC_SUCCESS, but we
958 			 * know that we did not get a connection, so
959 			 * force the rpc status to RPC_CANTCONNECT.
960 			 */
961 			p->cku_err.re_status = RPC_CANTCONNECT;
962 			p->cku_err.re_errno = EIO;
963 			break;
964 		}
965 		if (delay_first == TRUE)
966 			ticks = clnt_cots_min_tout * drv_usectohz(1000000);
967 		goto cots_done;
968 	}
969 
970 	/*
971 	 * If we've never sent any request on this connection (send count
972 	 * is zero, or the connection has been reset), cache the
973 	 * the connection's create time and send a request (possibly a retry)
974 	 */
975 	if ((p->cku_flags & CKU_SENT) == 0 ||
976 	    p->cku_ctime != cm_entry->x_ctime) {
977 		p->cku_ctime = cm_entry->x_ctime;
978 
979 	} else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) &&
980 	    (call->call_reply != NULL ||
981 	    p->cku_recv_attempts < clnt_cots_maxrecv)) {
982 
983 		/*
984 		 * If we've sent a request and our call is on the dispatch
985 		 * queue and we haven't made too many receive attempts, then
986 		 * don't re-send, just receive.
987 		 */
988 		p->cku_recv_attempts++;
989 		goto read_again;
990 	}
991 
992 	/*
993 	 * Now we create the RPC request in a STREAMS message.  We have to do
994 	 * this after the call to connmgr_get so that we have the correct
995 	 * TIDU size for the transport.
996 	 */
997 	tidu_size = cm_entry->x_tidu_size;
998 	len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE);
999 
1000 	while ((mp = allocb(len, BPRI_MED)) == NULL) {
1001 		if (strwaitbuf(len, BPRI_MED)) {
1002 			p->cku_err.re_status = RPC_SYSTEMERROR;
1003 			p->cku_err.re_errno = ENOSR;
1004 			COTSRCSTAT_INCR(p->cku_stats, rcnomem);
1005 			goto cots_done;
1006 		}
1007 	}
1008 	xdrs = &p->cku_outxdr;
1009 	xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size);
1010 	mpsize = MBLKSIZE(mp);
1011 	ASSERT(mpsize >= len);
1012 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
1013 
1014 	/*
1015 	 * If the size of mblk is not appreciably larger than what we
1016 	 * asked, then resize the mblk to exactly len bytes. The reason for
1017 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
1018 	 * (from TCP over ethernet), and the arguments to the RPC require
1019 	 * 2800 bytes. Ideally we want the protocol to render two
1020 	 * ~1400 byte segments over the wire. However if allocb() gives us a 2k
1021 	 * mblk, and we allocate a second mblk for the remainder, the protocol
1022 	 * module may generate 3 segments over the wire:
1023 	 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and
1024 	 * 892 for the third. If we "waste" 448 bytes in the first mblk,
1025 	 * the XDR encoding will generate two ~1400 byte mblks, and the
1026 	 * protocol module is more likely to produce properly sized segments.
1027 	 */
1028 	if ((mpsize >> 1) <= len)
1029 		mp->b_rptr += (mpsize - len);
1030 
1031 	/*
1032 	 * Adjust b_rptr to reserve space for the non-data protocol headers
1033 	 * any downstream modules might like to add, and for the
1034 	 * record marking header.
1035 	 */
1036 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
1037 
1038 	if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) {
1039 		/* Copy in the preserialized RPC header information. */
1040 		bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE);
1041 
1042 		/* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */
1043 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base +
1044 		    WIRE_HDR_SIZE));
1045 
1046 		ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE);
1047 
1048 		/* Serialize the procedure number and the arguments. */
1049 		if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) ||
1050 		    (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) ||
1051 		    (!(*xdr_args)(xdrs, argsp))) {
1052 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1053 			p->cku_err.re_errno = EIO;
1054 			goto cots_done;
1055 		}
1056 
1057 		(*(uint32_t *)(mp->b_rptr)) = p->cku_xid;
1058 	} else {
1059 		uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE];
1060 		IXDR_PUT_U_INT32(uproc, procnum);
1061 
1062 		(*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid;
1063 
1064 		/* Use XDR_SETPOS() to set the b_wptr. */
1065 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
1066 
1067 		/* Serialize the procedure number and the arguments. */
1068 		if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4,
1069 		    xdrs, xdr_args, argsp)) {
1070 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1071 			p->cku_err.re_errno = EIO;
1072 			goto cots_done;
1073 		}
1074 	}
1075 
1076 	RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n",
1077 	    tidu_size);
1078 
1079 	wq = cm_entry->x_wq;
1080 	clnt_dispatch_send(wq, mp, call, p->cku_xid,
1081 				(p->cku_flags & CKU_ONQUEUE));
1082 
1083 	RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n",
1084 		(uint_t)p->cku_xid);
1085 	p->cku_flags = (CKU_ONQUEUE|CKU_SENT);
1086 	p->cku_recv_attempts = 1;
1087 
1088 #ifdef	RPCDEBUG
1089 	time_sent = lbolt;
1090 #endif
1091 
1092 	/*
1093 	 * Wait for a reply or a timeout.  If there is no error or timeout,
1094 	 * (both indicated by call_status), call->call_reply will contain
1095 	 * the RPC reply message.
1096 	 */
1097 read_again:
1098 	mutex_enter(&call->call_lock);
1099 	interrupted = 0;
1100 	if (call->call_status == RPC_TIMEDOUT) {
1101 		/*
1102 		 * Indicate that the lwp is not to be stopped while waiting
1103 		 * for this network traffic.  This is to avoid deadlock while
1104 		 * debugging a process via /proc and also to avoid recursive
1105 		 * mutex_enter()s due to NFS page faults while stopping
1106 		 * (NFS holds locks when it calls here).
1107 		 */
1108 		clock_t cv_wait_ret;
1109 		clock_t timout;
1110 		clock_t oldlbolt;
1111 
1112 		klwp_t *lwp = ttolwp(curthread);
1113 
1114 		if (lwp != NULL)
1115 			lwp->lwp_nostop++;
1116 
1117 		oldlbolt = lbolt;
1118 		timout = wait.tv_sec * drv_usectohz(1000000) +
1119 		    drv_usectohz(wait.tv_usec) + oldlbolt;
1120 		/*
1121 		 * Iterate until the call_status is changed to something
1122 		 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns
1123 		 * something <=0 zero. The latter means that we timed
1124 		 * out.
1125 		 */
1126 		if (h->cl_nosignal)
1127 			while ((cv_wait_ret = cv_timedwait(&call->call_cv,
1128 			    &call->call_lock, timout)) > 0 &&
1129 			    call->call_status == RPC_TIMEDOUT);
1130 		else
1131 			while ((cv_wait_ret = cv_timedwait_sig(
1132 			    &call->call_cv,
1133 			    &call->call_lock, timout)) > 0 &&
1134 			    call->call_status == RPC_TIMEDOUT);
1135 
1136 		switch (cv_wait_ret) {
1137 		case 0:
1138 			/*
1139 			 * If we got out of the above loop with
1140 			 * cv_timedwait_sig() returning 0, then we were
1141 			 * interrupted regardless what call_status is.
1142 			 */
1143 			interrupted = 1;
1144 			break;
1145 		case -1:
1146 			/* cv_timedwait_sig() timed out */
1147 			break;
1148 		default:
1149 
1150 			/*
1151 			 * We were cv_signaled(). If we didn't
1152 			 * get a successful call_status and returned
1153 			 * before time expired, delay up to clnt_cots_min_tout
1154 			 * seconds so that the caller doesn't immediately
1155 			 * try to call us again and thus force the
1156 			 * same condition that got us here (such
1157 			 * as a RPC_XPRTFAILED due to the server not
1158 			 * listening on the end-point.
1159 			 */
1160 			if (call->call_status != RPC_SUCCESS) {
1161 				clock_t curlbolt;
1162 				clock_t diff;
1163 
1164 				curlbolt = ddi_get_lbolt();
1165 				ticks = clnt_cots_min_tout *
1166 				    drv_usectohz(1000000);
1167 				diff = curlbolt - oldlbolt;
1168 				if (diff < ticks) {
1169 					delay_first = TRUE;
1170 					if (diff > 0)
1171 						ticks -= diff;
1172 				}
1173 			}
1174 			break;
1175 		}
1176 
1177 		if (lwp != NULL)
1178 			lwp->lwp_nostop--;
1179 	}
1180 	/*
1181 	 * Get the reply message, if any.  This will be freed at the end
1182 	 * whether or not an error occurred.
1183 	 */
1184 	mp = call->call_reply;
1185 	call->call_reply = NULL;
1186 
1187 	/*
1188 	 * call_err is the error info when the call is on dispatch queue.
1189 	 * cku_err is the error info returned to the caller.
1190 	 * Sync cku_err with call_err for local message processing.
1191 	 */
1192 
1193 	status = call->call_status;
1194 	p->cku_err = call->call_err;
1195 	mutex_exit(&call->call_lock);
1196 
1197 	if (status != RPC_SUCCESS) {
1198 		switch (status) {
1199 		case RPC_TIMEDOUT:
1200 			if (interrupted) {
1201 				COTSRCSTAT_INCR(p->cku_stats, rcintrs);
1202 				p->cku_err.re_status = RPC_INTR;
1203 				p->cku_err.re_errno = EINTR;
1204 				RPCLOG(1, "clnt_cots_kcallit: xid 0x%x",
1205 				    p->cku_xid);
1206 				RPCLOG(1, "signal interrupted at %ld", lbolt);
1207 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1208 			} else {
1209 				COTSRCSTAT_INCR(p->cku_stats, rctimeouts);
1210 				p->cku_err.re_errno = ETIMEDOUT;
1211 				RPCLOG(1, "clnt_cots_kcallit: timed out at %ld",
1212 				    lbolt);
1213 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1214 			}
1215 			break;
1216 
1217 		case RPC_XPRTFAILED:
1218 			if (p->cku_err.re_errno == 0)
1219 				p->cku_err.re_errno = EIO;
1220 
1221 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
1222 			    p->cku_err.re_errno);
1223 			break;
1224 
1225 		case RPC_SYSTEMERROR:
1226 			ASSERT(p->cku_err.re_errno);
1227 			RPCLOG(1, "clnt_cots_kcallit: system error: %d\n",
1228 			    p->cku_err.re_errno);
1229 			break;
1230 
1231 		default:
1232 			p->cku_err.re_status = RPC_SYSTEMERROR;
1233 			p->cku_err.re_errno = EIO;
1234 			RPCLOG(1, "clnt_cots_kcallit: error: %s\n",
1235 			    clnt_sperrno(status));
1236 			break;
1237 		}
1238 		if (p->cku_err.re_status != RPC_TIMEDOUT) {
1239 
1240 			if (p->cku_flags & CKU_ONQUEUE) {
1241 				call_table_remove(call);
1242 				p->cku_flags &= ~CKU_ONQUEUE;
1243 			}
1244 
1245 			RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x "
1246 			    "taken off dispatch list\n", p->cku_xid);
1247 			if (call->call_reply) {
1248 				freemsg(call->call_reply);
1249 				call->call_reply = NULL;
1250 			}
1251 		} else if (wait.tv_sec != 0) {
1252 			/*
1253 			 * We've sent the request over TCP and so we have
1254 			 * every reason to believe it will get
1255 			 * delivered. In which case returning a timeout is not
1256 			 * appropriate.
1257 			 */
1258 			if (p->cku_progress == TRUE &&
1259 			    p->cku_recv_attempts < clnt_cots_maxrecv) {
1260 				p->cku_err.re_status = RPC_INPROGRESS;
1261 			}
1262 		}
1263 		goto cots_done;
1264 	}
1265 
1266 	xdrs = &p->cku_inxdr;
1267 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
1268 
1269 	reply_msg.rm_direction = REPLY;
1270 	reply_msg.rm_reply.rp_stat = MSG_ACCEPTED;
1271 	reply_msg.acpted_rply.ar_stat = SUCCESS;
1272 
1273 	reply_msg.acpted_rply.ar_verf = _null_auth;
1274 	/*
1275 	 *  xdr_results will be done in AUTH_UNWRAP.
1276 	 */
1277 	reply_msg.acpted_rply.ar_results.where = NULL;
1278 	reply_msg.acpted_rply.ar_results.proc = xdr_void;
1279 
1280 	if (xdr_replymsg(xdrs, &reply_msg)) {
1281 		enum clnt_stat re_status;
1282 
1283 		_seterr_reply(&reply_msg, &p->cku_err);
1284 
1285 		re_status = p->cku_err.re_status;
1286 		if (re_status == RPC_SUCCESS) {
1287 			/*
1288 			 * Reply is good, check auth.
1289 			 */
1290 			if (!AUTH_VALIDATE(h->cl_auth,
1291 					&reply_msg.acpted_rply.ar_verf)) {
1292 				COTSRCSTAT_INCR(p->cku_stats, rcbadverfs);
1293 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1294 					"failure\n");
1295 				freemsg(mp);
1296 				(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1297 				mutex_enter(&call->call_lock);
1298 				if (call->call_reply == NULL)
1299 					call->call_status = RPC_TIMEDOUT;
1300 				mutex_exit(&call->call_lock);
1301 				goto read_again;
1302 			} else if (!AUTH_UNWRAP(h->cl_auth, xdrs,
1303 						xdr_results, resultsp)) {
1304 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1305 					"failure (unwrap)\n");
1306 				p->cku_err.re_status = RPC_CANTDECODERES;
1307 				p->cku_err.re_errno = EIO;
1308 			}
1309 		} else {
1310 			/* set errno in case we can't recover */
1311 			if (re_status != RPC_VERSMISMATCH &&
1312 			    re_status != RPC_AUTHERROR &&
1313 			    re_status != RPC_PROGVERSMISMATCH)
1314 				p->cku_err.re_errno = EIO;
1315 
1316 			if (re_status == RPC_AUTHERROR) {
1317 				/*
1318 				 * Maybe our credential need to be refreshed
1319 				 */
1320 				if (cm_entry) {
1321 					/*
1322 					 * There is the potential that the
1323 					 * cm_entry has/will be marked dead,
1324 					 * so drop the connection altogether,
1325 					 * force REFRESH to establish new
1326 					 * connection.
1327 					 */
1328 					connmgr_cancelconn(cm_entry);
1329 					cm_entry = NULL;
1330 				}
1331 
1332 				if ((refreshes > 0) &&
1333 				    AUTH_REFRESH(h->cl_auth, &reply_msg,
1334 						p->cku_cred)) {
1335 					refreshes--;
1336 					(void) xdr_rpc_free_verifier(xdrs,
1337 								&reply_msg);
1338 					freemsg(mp);
1339 					mp = NULL;
1340 
1341 					if (p->cku_flags & CKU_ONQUEUE) {
1342 						call_table_remove(call);
1343 						p->cku_flags &= ~CKU_ONQUEUE;
1344 					}
1345 
1346 					RPCLOG(64,
1347 					    "clnt_cots_kcallit: AUTH_ERROR, xid"
1348 					    " 0x%x removed off dispatch list\n",
1349 					    p->cku_xid);
1350 					if (call->call_reply) {
1351 						freemsg(call->call_reply);
1352 						call->call_reply = NULL;
1353 					}
1354 
1355 					COTSRCSTAT_INCR(p->cku_stats,
1356 							rcbadcalls);
1357 					COTSRCSTAT_INCR(p->cku_stats,
1358 							rcnewcreds);
1359 					goto call_again;
1360 				}
1361 
1362 				/*
1363 				 * We have used the client handle to
1364 				 * do an AUTH_REFRESH and the RPC status may
1365 				 * be set to RPC_SUCCESS; Let's make sure to
1366 				 * set it to RPC_AUTHERROR.
1367 				 */
1368 				p->cku_err.re_status = RPC_AUTHERROR;
1369 
1370 				/*
1371 				 * Map recoverable and unrecoverable
1372 				 * authentication errors to appropriate errno
1373 				 */
1374 				switch (p->cku_err.re_why) {
1375 				case AUTH_TOOWEAK:
1376 					/*
1377 					 * This could be a failure where the
1378 					 * server requires use of a reserved
1379 					 * port,  check and optionally set the
1380 					 * client handle useresvport trying
1381 					 * one more time. Next go round we
1382 					 * fall out with the tooweak error.
1383 					 */
1384 					if (p->cku_useresvport != 1) {
1385 						p->cku_useresvport = 1;
1386 						p->cku_xid = 0;
1387 						(void) xdr_rpc_free_verifier
1388 							    (xdrs, &reply_msg);
1389 						freemsg(mp);
1390 						goto call_again;
1391 					}
1392 					/* FALLTHRU */
1393 				case AUTH_BADCRED:
1394 				case AUTH_BADVERF:
1395 				case AUTH_INVALIDRESP:
1396 				case AUTH_FAILED:
1397 				case RPCSEC_GSS_NOCRED:
1398 				case RPCSEC_GSS_FAILED:
1399 						p->cku_err.re_errno = EACCES;
1400 						break;
1401 				case AUTH_REJECTEDCRED:
1402 				case AUTH_REJECTEDVERF:
1403 				default:	p->cku_err.re_errno = EIO;
1404 						break;
1405 				}
1406 				RPCLOG(1, "clnt_cots_kcallit : authentication"
1407 				    " failed with RPC_AUTHERROR of type %d\n",
1408 				    (int)p->cku_err.re_why);
1409 			}
1410 		}
1411 	} else {
1412 		/* reply didn't decode properly. */
1413 		p->cku_err.re_status = RPC_CANTDECODERES;
1414 		p->cku_err.re_errno = EIO;
1415 		RPCLOG0(1, "clnt_cots_kcallit: decode failure\n");
1416 	}
1417 
1418 	(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1419 
1420 	if (p->cku_flags & CKU_ONQUEUE) {
1421 		call_table_remove(call);
1422 		p->cku_flags &= ~CKU_ONQUEUE;
1423 	}
1424 
1425 	RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list",
1426 	    p->cku_xid);
1427 	RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status));
1428 cots_done:
1429 	if (cm_entry)
1430 		connmgr_release(cm_entry);
1431 
1432 	if (mp != NULL)
1433 		freemsg(mp);
1434 	if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) {
1435 		freemsg(call->call_reply);
1436 		call->call_reply = NULL;
1437 	}
1438 	if (p->cku_err.re_status != RPC_SUCCESS) {
1439 		RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n");
1440 		COTSRCSTAT_INCR(p->cku_stats, rcbadcalls);
1441 	}
1442 
1443 	/*
1444 	 * No point in delaying if the zone is going away.
1445 	 */
1446 	if (delay_first == TRUE &&
1447 	    !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) {
1448 		if (clnt_delay(ticks, h->cl_nosignal) == EINTR) {
1449 			p->cku_err.re_errno = EINTR;
1450 			p->cku_err.re_status = RPC_INTR;
1451 		}
1452 	}
1453 	return (p->cku_err.re_status);
1454 }
1455 
1456 /*
1457  * Kinit routine for cots.  This sets up the correct operations in
1458  * the client handle, as the handle may have previously been a clts
1459  * handle, and clears the xid field so there is no way a new call
1460  * could be mistaken for a retry.  It also sets in the handle the
1461  * information that is passed at create/kinit time but needed at
1462  * call time, as cots creates the transport at call time - device,
1463  * address of the server, protocol family.
1464  */
1465 void
1466 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr,
1467 	int max_msgsize, cred_t *cred)
1468 {
1469 	/* LINTED pointer alignment */
1470 	cku_private_t *p = htop(h);
1471 	calllist_t *call = &p->cku_call;
1472 
1473 	h->cl_ops = &tcp_ops;
1474 	if (p->cku_flags & CKU_ONQUEUE) {
1475 		call_table_remove(call);
1476 		p->cku_flags &= ~CKU_ONQUEUE;
1477 		RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from"
1478 		    " dispatch list\n", p->cku_xid);
1479 	}
1480 
1481 	if (call->call_reply != NULL) {
1482 		freemsg(call->call_reply);
1483 		call->call_reply = NULL;
1484 	}
1485 
1486 	call->call_bucket = NULL;
1487 	call->call_hash = 0;
1488 
1489 	/*
1490 	 * We don't clear cku_flags here, because clnt_cots_kcallit()
1491 	 * takes care of handling the cku_flags reset.
1492 	 */
1493 	p->cku_xid = 0;
1494 	p->cku_device = dev;
1495 	p->cku_addrfmly = family;
1496 	p->cku_cred = cred;
1497 
1498 	if (p->cku_addr.maxlen < addr->len) {
1499 		if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL)
1500 			kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
1501 		p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
1502 		p->cku_addr.maxlen = addr->maxlen;
1503 	}
1504 
1505 	p->cku_addr.len = addr->len;
1506 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
1507 
1508 	/*
1509 	 * If the current sanity check size in rpcmod is smaller
1510 	 * than the size needed, then increase the sanity check.
1511 	 */
1512 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
1513 	    max_msgsize > *clnt_max_msg_sizep) {
1514 		mutex_enter(&clnt_max_msg_lock);
1515 		if (max_msgsize > *clnt_max_msg_sizep)
1516 			*clnt_max_msg_sizep = max_msgsize;
1517 		mutex_exit(&clnt_max_msg_lock);
1518 	}
1519 }
1520 
1521 /*
1522  * ksettimers is a no-op for cots, with the exception of setting the xid.
1523  */
1524 /* ARGSUSED */
1525 static int
1526 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all,
1527 	int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg,
1528 	uint32_t xid)
1529 {
1530 	/* LINTED pointer alignment */
1531 	cku_private_t *p = htop(h);
1532 
1533 	if (xid)
1534 		p->cku_xid = xid;
1535 	COTSRCSTAT_INCR(p->cku_stats, rctimers);
1536 	return (0);
1537 }
1538 
1539 extern void rpc_poptimod(struct vnode *);
1540 extern int kstr_push(struct vnode *, char *);
1541 
1542 int
1543 conn_kstat_update(kstat_t *ksp, int rw)
1544 {
1545 	struct cm_xprt *cm_entry;
1546 	struct cm_kstat_xprt *cm_ksp_data;
1547 	uchar_t *b;
1548 	char *fbuf;
1549 
1550 	if (rw == KSTAT_WRITE)
1551 		return (EACCES);
1552 	if (ksp == NULL || ksp->ks_private == NULL)
1553 		return (EIO);
1554 	cm_entry  = (struct cm_xprt *)ksp->ks_private;
1555 	cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data;
1556 
1557 	cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq;
1558 	cm_ksp_data->x_family.value.ui32 = cm_entry->x_family;
1559 	cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev;
1560 	cm_ksp_data->x_time.value.ui32 = cm_entry->x_time;
1561 	cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref;
1562 	cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags;
1563 
1564 	if (cm_entry->x_server.buf) {
1565 		fbuf = cm_ksp_data->x_server.value.str.addr.ptr;
1566 		if (cm_entry->x_family == AF_INET &&
1567 		    cm_entry->x_server.len ==
1568 		    sizeof (struct sockaddr_in)) {
1569 			struct sockaddr_in  *sa;
1570 			sa = (struct sockaddr_in *)
1571 				cm_entry->x_server.buf;
1572 			b = (uchar_t *)&sa->sin_addr;
1573 			(void) sprintf(fbuf,
1574 			    "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF,
1575 			    b[2] & 0xFF, b[3] & 0xFF);
1576 			cm_ksp_data->x_port.value.ui32 =
1577 				(uint32_t)sa->sin_port;
1578 		} else if (cm_entry->x_family == AF_INET6 &&
1579 				cm_entry->x_server.len >=
1580 				sizeof (struct sockaddr_in6)) {
1581 			/* extract server IP address & port */
1582 			struct sockaddr_in6 *sin6;
1583 			sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf;
1584 			(void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf,
1585 				INET6_ADDRSTRLEN);
1586 			cm_ksp_data->x_port.value.ui32 = sin6->sin6_port;
1587 		} else {
1588 			struct sockaddr_in  *sa;
1589 
1590 			sa = (struct sockaddr_in *)cm_entry->x_server.buf;
1591 			b = (uchar_t *)&sa->sin_addr;
1592 			(void) sprintf(fbuf,
1593 			    "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF,
1594 			    b[2] & 0xFF, b[3] & 0xFF);
1595 		}
1596 		KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) =
1597 			strlen(fbuf) + 1;
1598 	}
1599 
1600 	return (0);
1601 }
1602 
1603 
1604 /*
1605  * We want a version of delay which is interruptible by a UNIX signal
1606  * Return EINTR if an interrupt occured.
1607  */
1608 static int
1609 clnt_delay(clock_t ticks, bool_t nosignal)
1610 {
1611 	if (nosignal == TRUE) {
1612 		delay(ticks);
1613 		return (0);
1614 	}
1615 	return (delay_sig(ticks));
1616 }
1617 
1618 /*
1619  * Wait for a connection until a timeout, or until we are
1620  * signalled that there has been a connection state change.
1621  */
1622 static enum clnt_stat
1623 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp,
1624 	bool_t nosignal)
1625 {
1626 	bool_t interrupted;
1627 	clock_t timout, cv_stat;
1628 	enum clnt_stat clstat;
1629 	unsigned int old_state;
1630 
1631 	ASSERT(MUTEX_HELD(&connmgr_lock));
1632 	/*
1633 	 * We wait for the transport connection to be made, or an
1634 	 * indication that it could not be made.
1635 	 */
1636 	clstat = RPC_TIMEDOUT;
1637 	interrupted = FALSE;
1638 
1639 	old_state = cm_entry->x_state_flags;
1640 	/*
1641 	 * Now loop until cv_timedwait{_sig} returns because of
1642 	 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be
1643 	 * cv_signalled for various other reasons too. So loop
1644 	 * until there is a state change on the connection.
1645 	 */
1646 
1647 	timout = waitp->tv_sec * drv_usectohz(1000000) +
1648 	    drv_usectohz(waitp->tv_usec) + lbolt;
1649 
1650 	if (nosignal) {
1651 		while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv,
1652 		    &connmgr_lock, timout)) > 0 &&
1653 		    cm_entry->x_state_flags == old_state)
1654 			;
1655 	} else {
1656 		while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv,
1657 		    &connmgr_lock, timout)) > 0 &&
1658 		    cm_entry->x_state_flags == old_state)
1659 			;
1660 
1661 		if (cv_stat == 0) /* got intr signal? */
1662 			interrupted = TRUE;
1663 	}
1664 
1665 	if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) ==
1666 	    X_CONNECTED) {
1667 		clstat = RPC_SUCCESS;
1668 	} else {
1669 		if (interrupted == TRUE)
1670 			clstat = RPC_INTR;
1671 		RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n",
1672 		    clnt_sperrno(clstat));
1673 	}
1674 
1675 	return (clstat);
1676 }
1677 
1678 /*
1679  * Primary interface for how RPC grabs a connection.
1680  */
1681 static struct cm_xprt *
1682 connmgr_wrapget(
1683 	struct netbuf *retryaddr,
1684 	const struct timeval *waitp,
1685 	cku_private_t *p)
1686 {
1687 	struct cm_xprt *cm_entry;
1688 
1689 	cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly,
1690 	    &p->cku_srcaddr, &p->cku_err, p->cku_device,
1691 	    p->cku_client.cl_nosignal, p->cku_useresvport);
1692 
1693 	if (cm_entry == NULL) {
1694 		/*
1695 		 * Re-map the call status to RPC_INTR if the err code is
1696 		 * EINTR. This can happen if calls status is RPC_TLIERROR.
1697 		 * However, don't re-map if signalling has been turned off.
1698 		 * XXX Really need to create a separate thread whenever
1699 		 * there isn't an existing connection.
1700 		 */
1701 		if (p->cku_err.re_errno == EINTR) {
1702 			if (p->cku_client.cl_nosignal == TRUE)
1703 				p->cku_err.re_errno = EIO;
1704 			else
1705 				p->cku_err.re_status = RPC_INTR;
1706 		}
1707 	}
1708 
1709 	return (cm_entry);
1710 }
1711 
1712 /*
1713  * Obtains a transport to the server specified in addr.  If a suitable transport
1714  * does not already exist in the list of cached transports, a new connection
1715  * is created, connected, and added to the list. The connection is for sending
1716  * only - the reply message may come back on another transport connection.
1717  */
1718 static struct cm_xprt *
1719 connmgr_get(
1720 	struct netbuf	*retryaddr,
1721 	const struct timeval	*waitp,	/* changed to a ptr to converse stack */
1722 	struct netbuf	*destaddr,
1723 	int		addrfmly,
1724 	struct netbuf	*srcaddr,
1725 	struct rpc_err	*rpcerr,
1726 	dev_t		device,
1727 	bool_t		nosignal,
1728 	int		useresvport)
1729 {
1730 	struct cm_xprt *cm_entry;
1731 	struct cm_xprt *lru_entry;
1732 	struct cm_xprt **cmp;
1733 	queue_t *wq;
1734 	TIUSER *tiptr;
1735 	int i;
1736 	int retval;
1737 	clock_t prev_time;
1738 	int tidu_size;
1739 	bool_t	connected;
1740 	zoneid_t zoneid = getzoneid();
1741 
1742 	/*
1743 	 * If the call is not a retry, look for a transport entry that
1744 	 * goes to the server of interest.
1745 	 */
1746 	mutex_enter(&connmgr_lock);
1747 
1748 	if (retryaddr == NULL) {
1749 use_new_conn:
1750 		i = 0;
1751 		cm_entry = lru_entry = NULL;
1752 		prev_time = lbolt;
1753 
1754 		cmp = &cm_hd;
1755 		while ((cm_entry = *cmp) != NULL) {
1756 			ASSERT(cm_entry != cm_entry->x_next);
1757 			/*
1758 			 * Garbage collect conections that are marked
1759 			 * for needs disconnect.
1760 			 */
1761 			if (cm_entry->x_needdis) {
1762 				CONN_HOLD(cm_entry);
1763 				connmgr_dis_and_wait(cm_entry);
1764 				connmgr_release(cm_entry);
1765 				/*
1766 				 * connmgr_lock could have been
1767 				 * dropped for the disconnect
1768 				 * processing so start over.
1769 				 */
1770 				goto use_new_conn;
1771 			}
1772 
1773 			/*
1774 			 * Garbage collect the dead connections that have
1775 			 * no threads working on them.
1776 			 */
1777 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1778 			    X_DEAD) {
1779 				*cmp = cm_entry->x_next;
1780 				mutex_exit(&connmgr_lock);
1781 				connmgr_close(cm_entry);
1782 				mutex_enter(&connmgr_lock);
1783 				goto use_new_conn;
1784 			}
1785 
1786 
1787 			if ((cm_entry->x_state_flags & X_BADSTATES) == 0 &&
1788 			    cm_entry->x_zoneid == zoneid &&
1789 			    cm_entry->x_rdev == device &&
1790 			    destaddr->len == cm_entry->x_server.len &&
1791 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
1792 			    destaddr->len) == 0) {
1793 				/*
1794 				 * If the matching entry isn't connected,
1795 				 * attempt to reconnect it.
1796 				 */
1797 				if (cm_entry->x_connected == FALSE) {
1798 					/*
1799 					 * We don't go through trying
1800 					 * to find the least recently
1801 					 * used connected because
1802 					 * connmgr_reconnect() briefly
1803 					 * dropped the connmgr_lock,
1804 					 * allowing a window for our
1805 					 * accounting to be messed up.
1806 					 * In any case, a re-connected
1807 					 * connection is as good as
1808 					 * a LRU connection.
1809 					 */
1810 					return (connmgr_wrapconnect(cm_entry,
1811 					    waitp, destaddr, addrfmly, srcaddr,
1812 					    rpcerr, TRUE, nosignal));
1813 				}
1814 				i++;
1815 				if (cm_entry->x_time - prev_time <= 0 ||
1816 				    lru_entry == NULL) {
1817 					prev_time = cm_entry->x_time;
1818 					lru_entry = cm_entry;
1819 				}
1820 			}
1821 			cmp = &cm_entry->x_next;
1822 		}
1823 
1824 		if (i > clnt_max_conns) {
1825 			RPCLOG(8, "connmgr_get: too many conns, dooming entry"
1826 			    " %p\n", (void *)lru_entry->x_tiptr);
1827 			lru_entry->x_doomed = TRUE;
1828 			goto use_new_conn;
1829 		}
1830 
1831 		/*
1832 		 * If we are at the maximum number of connections to
1833 		 * the server, hand back the least recently used one.
1834 		 */
1835 		if (i == clnt_max_conns) {
1836 			/*
1837 			 * Copy into the handle the source address of
1838 			 * the connection, which we will use in case of
1839 			 * a later retry.
1840 			 */
1841 			if (srcaddr->len != lru_entry->x_src.len) {
1842 				if (srcaddr->len > 0)
1843 					kmem_free(srcaddr->buf,
1844 					    srcaddr->maxlen);
1845 				srcaddr->buf = kmem_zalloc(
1846 				    lru_entry->x_src.len, KM_SLEEP);
1847 				srcaddr->maxlen = srcaddr->len =
1848 				    lru_entry->x_src.len;
1849 			}
1850 			bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len);
1851 			RPCLOG(2, "connmgr_get: call going out on %p\n",
1852 			    (void *)lru_entry);
1853 			lru_entry->x_time = lbolt;
1854 			CONN_HOLD(lru_entry);
1855 			mutex_exit(&connmgr_lock);
1856 			return (lru_entry);
1857 		}
1858 
1859 	} else {
1860 		/*
1861 		 * This is the retry case (retryaddr != NULL).  Retries must
1862 		 * be sent on the same source port as the original call.
1863 		 */
1864 
1865 		/*
1866 		 * Walk the list looking for a connection with a source address
1867 		 * that matches the retry address.
1868 		 */
1869 		cmp = &cm_hd;
1870 		while ((cm_entry = *cmp) != NULL) {
1871 			ASSERT(cm_entry != cm_entry->x_next);
1872 			if (zoneid != cm_entry->x_zoneid ||
1873 			    device != cm_entry->x_rdev ||
1874 			    retryaddr->len != cm_entry->x_src.len ||
1875 			    bcmp(retryaddr->buf, cm_entry->x_src.buf,
1876 				    retryaddr->len) != 0) {
1877 				cmp = &cm_entry->x_next;
1878 				continue;
1879 			}
1880 
1881 			/*
1882 			 * Sanity check: if the connection with our source
1883 			 * port is going to some other server, something went
1884 			 * wrong, as we never delete connections (i.e. release
1885 			 * ports) unless they have been idle.  In this case,
1886 			 * it is probably better to send the call out using
1887 			 * a new source address than to fail it altogether,
1888 			 * since that port may never be released.
1889 			 */
1890 			if (destaddr->len != cm_entry->x_server.len ||
1891 				bcmp(destaddr->buf, cm_entry->x_server.buf,
1892 					destaddr->len) != 0) {
1893 				RPCLOG(1, "connmgr_get: tiptr %p"
1894 				    " is going to a different server"
1895 				    " with the port that belongs"
1896 				    " to us!\n", (void *)cm_entry->x_tiptr);
1897 				retryaddr = NULL;
1898 				goto use_new_conn;
1899 			}
1900 
1901 			/*
1902 			 * If the connection of interest is not connected and we
1903 			 * can't reconnect it, then the server is probably
1904 			 * still down.  Return NULL to the caller and let it
1905 			 * retry later if it wants to.  We have a delay so the
1906 			 * machine doesn't go into a tight retry loop.  If the
1907 			 * entry was already connected, or the reconnected was
1908 			 * successful, return this entry.
1909 			 */
1910 			if (cm_entry->x_connected == FALSE) {
1911 				return (connmgr_wrapconnect(cm_entry,
1912 				    waitp, destaddr, addrfmly, NULL,
1913 				    rpcerr, TRUE, nosignal));
1914 			} else {
1915 				CONN_HOLD(cm_entry);
1916 
1917 				cm_entry->x_time = lbolt;
1918 				mutex_exit(&connmgr_lock);
1919 				RPCLOG(2, "connmgr_get: found old "
1920 				    "transport %p for retry\n",
1921 				    (void *)cm_entry);
1922 				return (cm_entry);
1923 			}
1924 		}
1925 
1926 		/*
1927 		 * We cannot find an entry in the list for this retry.
1928 		 * Either the entry has been removed temporarily to be
1929 		 * reconnected by another thread, or the original call
1930 		 * got a port but never got connected,
1931 		 * and hence the transport never got put in the
1932 		 * list.  Fall through to the "create new connection" code -
1933 		 * the former case will fail there trying to rebind the port,
1934 		 * and the later case (and any other pathological cases) will
1935 		 * rebind and reconnect and not hang the client machine.
1936 		 */
1937 		RPCLOG0(8, "connmgr_get: no entry in list for retry\n");
1938 	}
1939 	/*
1940 	 * Set up a transport entry in the connection manager's list.
1941 	 */
1942 	cm_entry = (struct cm_xprt *)
1943 	    kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP);
1944 
1945 	cm_entry->x_server.buf = kmem_zalloc(destaddr->len, KM_SLEEP);
1946 	bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len);
1947 	cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len;
1948 
1949 	cm_entry->x_state_flags = X_THREAD;
1950 	cm_entry->x_ref = 1;
1951 	cm_entry->x_family = addrfmly;
1952 	cm_entry->x_rdev = device;
1953 	cm_entry->x_zoneid = zoneid;
1954 	mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL);
1955 	cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL);
1956 	cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL);
1957 	cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL);
1958 
1959 	/*
1960 	 * Note that we add this partially initialized entry to the
1961 	 * connection list. This is so that we don't have connections to
1962 	 * the same server.
1963 	 *
1964 	 * Note that x_src is not initialized at this point. This is because
1965 	 * retryaddr might be NULL in which case x_src is whatever
1966 	 * t_kbind/bindresvport gives us. If another thread wants a
1967 	 * connection to the same server, seemingly we have an issue, but we
1968 	 * don't. If the other thread comes in with retryaddr == NULL, then it
1969 	 * will never look at x_src, and it will end up waiting in
1970 	 * connmgr_cwait() for the first thread to finish the connection
1971 	 * attempt. If the other thread comes in with retryaddr != NULL, then
1972 	 * that means there was a request sent on a connection, in which case
1973 	 * the the connection should already exist. Thus the first thread
1974 	 * never gets here ... it finds the connection it its server in the
1975 	 * connection list.
1976 	 *
1977 	 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd
1978 	 * thread will skip us because x_src.len == 0.
1979 	 */
1980 	cm_entry->x_next = cm_hd;
1981 	cm_hd = cm_entry;
1982 	mutex_exit(&connmgr_lock);
1983 
1984 	/*
1985 	 * Either we didn't find an entry to the server of interest, or we
1986 	 * don't have the maximum number of connections to that server -
1987 	 * create a new connection.
1988 	 */
1989 	RPCLOG0(8, "connmgr_get: creating new connection\n");
1990 	rpcerr->re_status = RPC_TLIERROR;
1991 
1992 	i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, kcred);
1993 	if (i) {
1994 		RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i);
1995 		rpcerr->re_errno = i;
1996 		connmgr_cancelconn(cm_entry);
1997 		return (NULL);
1998 	}
1999 	rpc_poptimod(tiptr->fp->f_vnode);
2000 
2001 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0,
2002 			K_TO_K, kcred, &retval)) {
2003 		RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i);
2004 		(void) t_kclose(tiptr, 1);
2005 		rpcerr->re_errno = i;
2006 		connmgr_cancelconn(cm_entry);
2007 		return (NULL);
2008 	}
2009 
2010 	if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K,
2011 		kcred, &retval)) {
2012 		RPCLOG(1, "connmgr_get: can't set client status with cots "
2013 		    "module, %d\n", i);
2014 		(void) t_kclose(tiptr, 1);
2015 		rpcerr->re_errno = i;
2016 		connmgr_cancelconn(cm_entry);
2017 		return (NULL);
2018 	}
2019 
2020 	mutex_enter(&connmgr_lock);
2021 
2022 	wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next;
2023 	cm_entry->x_wq = wq;
2024 
2025 	mutex_exit(&connmgr_lock);
2026 
2027 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0,
2028 			K_TO_K, kcred, &retval)) {
2029 		RPCLOG(1, "connmgr_get: can't push timod, %d\n", i);
2030 		(void) t_kclose(tiptr, 1);
2031 		rpcerr->re_errno = i;
2032 		connmgr_cancelconn(cm_entry);
2033 		return (NULL);
2034 	}
2035 
2036 	/*
2037 	 * If the caller has not specified reserved port usage then
2038 	 * take the system default.
2039 	 */
2040 	if (useresvport == -1)
2041 		useresvport = clnt_cots_do_bindresvport;
2042 
2043 	if ((useresvport || retryaddr != NULL) &&
2044 	    (addrfmly == AF_INET || addrfmly == AF_INET6)) {
2045 		bool_t alloc_src = FALSE;
2046 
2047 		if (srcaddr->len != destaddr->len) {
2048 			kmem_free(srcaddr->buf, srcaddr->maxlen);
2049 			srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2050 			srcaddr->maxlen = destaddr->len;
2051 			srcaddr->len = destaddr->len;
2052 			alloc_src = TRUE;
2053 		}
2054 
2055 		if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) {
2056 			(void) t_kclose(tiptr, 1);
2057 			RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: "
2058 				"%p\n", (void *)retryaddr);
2059 
2060 			/*
2061 			 * 1225408: If we allocated a source address, then it
2062 			 * is either garbage or all zeroes. In that case
2063 			 * we need to clear srcaddr.
2064 			 */
2065 			if (alloc_src == TRUE) {
2066 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2067 				srcaddr->maxlen = srcaddr->len = 0;
2068 				srcaddr->buf = NULL;
2069 			}
2070 			rpcerr->re_errno = i;
2071 			connmgr_cancelconn(cm_entry);
2072 			return (NULL);
2073 		}
2074 	} else {
2075 		if ((i = t_kbind(tiptr, NULL, NULL)) != 0) {
2076 			RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i);
2077 			(void) t_kclose(tiptr, 1);
2078 			rpcerr->re_errno = i;
2079 			connmgr_cancelconn(cm_entry);
2080 			return (NULL);
2081 		}
2082 	}
2083 
2084 	{
2085 		/*
2086 		 * Keep the kernel stack lean. Don't move this call
2087 		 * declaration to the top of this function because a
2088 		 * call is declared in connmgr_wrapconnect()
2089 		 */
2090 		calllist_t call;
2091 
2092 		bzero(&call, sizeof (call));
2093 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2094 
2095 		/*
2096 		 * This is a bound end-point so don't close it's stream.
2097 		 */
2098 		connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly,
2099 						&call, &tidu_size, FALSE, waitp,
2100 						nosignal);
2101 		*rpcerr = call.call_err;
2102 		cv_destroy(&call.call_cv);
2103 
2104 	}
2105 
2106 	mutex_enter(&connmgr_lock);
2107 
2108 	/*
2109 	 * Set up a transport entry in the connection manager's list.
2110 	 */
2111 	cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP);
2112 	bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len);
2113 	cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len;
2114 
2115 	cm_entry->x_tiptr = tiptr;
2116 	cm_entry->x_time = lbolt;
2117 
2118 	if (tiptr->tp_info.servtype == T_COTS_ORD)
2119 		cm_entry->x_ordrel = TRUE;
2120 	else
2121 		cm_entry->x_ordrel = FALSE;
2122 
2123 	cm_entry->x_tidu_size = tidu_size;
2124 
2125 	if (cm_entry->x_early_disc)
2126 		cm_entry->x_connected = FALSE;
2127 	else
2128 		cm_entry->x_connected = connected;
2129 
2130 	/*
2131 	 * There could be a discrepancy here such that
2132 	 * x_early_disc is TRUE yet connected is TRUE as well
2133 	 * and the connection is actually connected. In that case
2134 	 * lets be conservative and declare the connection as not
2135 	 * connected.
2136 	 */
2137 	cm_entry->x_early_disc = FALSE;
2138 	cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2139 	cm_entry->x_ctime = lbolt;
2140 
2141 	/*
2142 	 * Notify any threads waiting that the connection attempt is done.
2143 	 */
2144 	cm_entry->x_thread = FALSE;
2145 	cv_broadcast(&cm_entry->x_conn_cv);
2146 
2147 	mutex_exit(&connmgr_lock);
2148 
2149 	if (cm_entry->x_connected == FALSE) {
2150 		connmgr_release(cm_entry);
2151 		return (NULL);
2152 	}
2153 	return (cm_entry);
2154 }
2155 
2156 /*
2157  * Keep the cm_xprt entry on the connecton list when making a connection. This
2158  * is to prevent multiple connections to a slow server from appearing.
2159  * We use the bit field x_thread to tell if a thread is doing a connection
2160  * which keeps other interested threads from messing with connection.
2161  * Those other threads just wait if x_thread is set.
2162  *
2163  * If x_thread is not set, then we do the actual work of connecting via
2164  * connmgr_connect().
2165  *
2166  * mutex convention: called with connmgr_lock held, returns with it released.
2167  */
2168 static struct cm_xprt *
2169 connmgr_wrapconnect(
2170 	struct cm_xprt	*cm_entry,
2171 	const struct timeval	*waitp,
2172 	struct netbuf	*destaddr,
2173 	int		addrfmly,
2174 	struct netbuf	*srcaddr,
2175 	struct rpc_err	*rpcerr,
2176 	bool_t		reconnect,
2177 	bool_t		nosignal)
2178 {
2179 	ASSERT(MUTEX_HELD(&connmgr_lock));
2180 	/*
2181 	 * Hold this entry as we are about to drop connmgr_lock.
2182 	 */
2183 	CONN_HOLD(cm_entry);
2184 
2185 	/*
2186 	 * If there is a thread already making a connection for us, then
2187 	 * wait for it to complete the connection.
2188 	 */
2189 	if (cm_entry->x_thread == TRUE) {
2190 		rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal);
2191 
2192 		if (rpcerr->re_status != RPC_SUCCESS) {
2193 			mutex_exit(&connmgr_lock);
2194 			connmgr_release(cm_entry);
2195 			return (NULL);
2196 		}
2197 	} else {
2198 		bool_t connected;
2199 		calllist_t call;
2200 
2201 		cm_entry->x_thread = TRUE;
2202 
2203 		while (cm_entry->x_needrel == TRUE) {
2204 			cm_entry->x_needrel = FALSE;
2205 
2206 			connmgr_sndrel(cm_entry);
2207 			delay(drv_usectohz(1000000));
2208 
2209 			mutex_enter(&connmgr_lock);
2210 		}
2211 
2212 		/*
2213 		 * If we need to send a T_DISCON_REQ, send one.
2214 		 */
2215 		connmgr_dis_and_wait(cm_entry);
2216 
2217 		mutex_exit(&connmgr_lock);
2218 
2219 		bzero(&call, sizeof (call));
2220 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2221 
2222 		connected = connmgr_connect(cm_entry, cm_entry->x_wq,
2223 					    destaddr, addrfmly, &call,
2224 					    &cm_entry->x_tidu_size,
2225 					    reconnect, waitp, nosignal);
2226 
2227 		*rpcerr = call.call_err;
2228 		cv_destroy(&call.call_cv);
2229 
2230 		mutex_enter(&connmgr_lock);
2231 
2232 
2233 		if (cm_entry->x_early_disc)
2234 			cm_entry->x_connected = FALSE;
2235 		else
2236 			cm_entry->x_connected = connected;
2237 
2238 		/*
2239 		 * There could be a discrepancy here such that
2240 		 * x_early_disc is TRUE yet connected is TRUE as well
2241 		 * and the connection is actually connected. In that case
2242 		 * lets be conservative and declare the connection as not
2243 		 * connected.
2244 		 */
2245 
2246 		cm_entry->x_early_disc = FALSE;
2247 		cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2248 
2249 
2250 		/*
2251 		 * connmgr_connect() may have given up before the connection
2252 		 * actually timed out. So ensure that before the next
2253 		 * connection attempt we do a disconnect.
2254 		 */
2255 		cm_entry->x_ctime = lbolt;
2256 		cm_entry->x_thread = FALSE;
2257 
2258 		cv_broadcast(&cm_entry->x_conn_cv);
2259 
2260 		if (cm_entry->x_connected == FALSE) {
2261 			mutex_exit(&connmgr_lock);
2262 			connmgr_release(cm_entry);
2263 			return (NULL);
2264 		}
2265 	}
2266 
2267 	if (srcaddr != NULL) {
2268 		/*
2269 		 * Copy into the handle the
2270 		 * source address of the
2271 		 * connection, which we will use
2272 		 * in case of a later retry.
2273 		 */
2274 		if (srcaddr->len != cm_entry->x_src.len) {
2275 			if (srcaddr->maxlen > 0)
2276 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2277 			srcaddr->buf = kmem_zalloc(cm_entry->x_src.len,
2278 			    KM_SLEEP);
2279 			srcaddr->maxlen = srcaddr->len =
2280 			    cm_entry->x_src.len;
2281 		}
2282 		bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len);
2283 	}
2284 	cm_entry->x_time = lbolt;
2285 	mutex_exit(&connmgr_lock);
2286 	return (cm_entry);
2287 }
2288 
2289 /*
2290  * If we need to send a T_DISCON_REQ, send one.
2291  */
2292 static void
2293 connmgr_dis_and_wait(struct cm_xprt *cm_entry)
2294 {
2295 	ASSERT(MUTEX_HELD(&connmgr_lock));
2296 	for (;;) {
2297 		while (cm_entry->x_needdis == TRUE) {
2298 			RPCLOG(8, "connmgr_dis_and_wait: need "
2299 				"T_DISCON_REQ for connection 0x%p\n",
2300 				(void *)cm_entry);
2301 			cm_entry->x_needdis = FALSE;
2302 			cm_entry->x_waitdis = TRUE;
2303 
2304 			connmgr_snddis(cm_entry);
2305 
2306 			mutex_enter(&connmgr_lock);
2307 		}
2308 
2309 		if (cm_entry->x_waitdis == TRUE) {
2310 			clock_t curlbolt;
2311 			clock_t timout;
2312 
2313 			RPCLOG(8, "connmgr_dis_and_wait waiting for "
2314 				"T_DISCON_REQ's ACK for connection %p\n",
2315 				(void *)cm_entry);
2316 			curlbolt = ddi_get_lbolt();
2317 
2318 			timout = clnt_cots_min_conntout *
2319 				drv_usectohz(1000000) + curlbolt;
2320 
2321 			/*
2322 			 * The TPI spec says that the T_DISCON_REQ
2323 			 * will get acknowledged, but in practice
2324 			 * the ACK may never get sent. So don't
2325 			 * block forever.
2326 			 */
2327 			(void) cv_timedwait(&cm_entry->x_dis_cv,
2328 					    &connmgr_lock, timout);
2329 		}
2330 		/*
2331 		 * If we got the ACK, break. If we didn't,
2332 		 * then send another T_DISCON_REQ.
2333 		 */
2334 		if (cm_entry->x_waitdis == FALSE) {
2335 			break;
2336 		} else {
2337 			RPCLOG(8, "connmgr_dis_and_wait: did"
2338 				"not get T_DISCON_REQ's ACK for "
2339 				"connection  %p\n", (void *)cm_entry);
2340 			cm_entry->x_needdis = TRUE;
2341 		}
2342 	}
2343 }
2344 
2345 static void
2346 connmgr_cancelconn(struct cm_xprt *cm_entry)
2347 {
2348 	/*
2349 	 * Mark the connection table entry as dead; the next thread that
2350 	 * goes through connmgr_release() will notice this and deal with it.
2351 	 */
2352 	mutex_enter(&connmgr_lock);
2353 	cm_entry->x_dead = TRUE;
2354 
2355 	/*
2356 	 * Notify any threads waiting for the connection that it isn't
2357 	 * going to happen.
2358 	 */
2359 	cm_entry->x_thread = FALSE;
2360 	cv_broadcast(&cm_entry->x_conn_cv);
2361 	mutex_exit(&connmgr_lock);
2362 
2363 	connmgr_release(cm_entry);
2364 }
2365 
2366 static void
2367 connmgr_close(struct cm_xprt *cm_entry)
2368 {
2369 	mutex_enter(&cm_entry->x_lock);
2370 	while (cm_entry->x_ref != 0) {
2371 		/*
2372 		 * Must be a noninterruptible wait.
2373 		 */
2374 		cv_wait(&cm_entry->x_cv, &cm_entry->x_lock);
2375 	}
2376 
2377 	if (cm_entry->x_tiptr != NULL)
2378 		(void) t_kclose(cm_entry->x_tiptr, 1);
2379 
2380 	mutex_exit(&cm_entry->x_lock);
2381 	if (cm_entry->x_ksp != NULL) {
2382 		mutex_enter(&connmgr_lock);
2383 		cm_entry->x_ksp->ks_private = NULL;
2384 		mutex_exit(&connmgr_lock);
2385 
2386 		/*
2387 		 * Must free the buffer we allocated for the
2388 		 * server address in the update function
2389 		 */
2390 		if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2391 		    x_server.value.str.addr.ptr != NULL)
2392 			kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp->
2393 			    ks_data))->x_server.value.str.addr.ptr,
2394 				    INET6_ADDRSTRLEN);
2395 		kmem_free(cm_entry->x_ksp->ks_data,
2396 			    cm_entry->x_ksp->ks_data_size);
2397 		kstat_delete(cm_entry->x_ksp);
2398 	}
2399 
2400 	mutex_destroy(&cm_entry->x_lock);
2401 	cv_destroy(&cm_entry->x_cv);
2402 	cv_destroy(&cm_entry->x_conn_cv);
2403 	cv_destroy(&cm_entry->x_dis_cv);
2404 
2405 	if (cm_entry->x_server.buf != NULL)
2406 		kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen);
2407 	if (cm_entry->x_src.buf != NULL)
2408 		kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen);
2409 	kmem_free(cm_entry, sizeof (struct cm_xprt));
2410 }
2411 
2412 /*
2413  * Called by KRPC after sending the call message to release the connection
2414  * it was using.
2415  */
2416 static void
2417 connmgr_release(struct cm_xprt *cm_entry)
2418 {
2419 	mutex_enter(&cm_entry->x_lock);
2420 	cm_entry->x_ref--;
2421 	if (cm_entry->x_ref == 0)
2422 		cv_signal(&cm_entry->x_cv);
2423 	mutex_exit(&cm_entry->x_lock);
2424 }
2425 
2426 /*
2427  * Given an open stream, connect to the remote.  Returns true if connected,
2428  * false otherwise.
2429  */
2430 static bool_t
2431 connmgr_connect(
2432 	struct cm_xprt		*cm_entry,
2433 	queue_t			*wq,
2434 	struct netbuf		*addr,
2435 	int			addrfmly,
2436 	calllist_t 		*e,
2437 	int 			*tidu_ptr,
2438 	bool_t 			reconnect,
2439 	const struct timeval 	*waitp,
2440 	bool_t 			nosignal)
2441 {
2442 	mblk_t *mp;
2443 	struct T_conn_req *tcr;
2444 	struct T_info_ack *tinfo;
2445 	int interrupted, error;
2446 	int tidu_size, kstat_instance;
2447 
2448 	/* if it's a reconnect, flush any lingering data messages */
2449 	if (reconnect)
2450 		(void) putctl1(wq, M_FLUSH, FLUSHRW);
2451 
2452 	mp = allocb(sizeof (*tcr) + addr->len, BPRI_LO);
2453 	if (mp == NULL) {
2454 		/*
2455 		 * This is unfortunate, but we need to look up the stats for
2456 		 * this zone to increment the "memory allocation failed"
2457 		 * counter.  curproc->p_zone is safe since we're initiating a
2458 		 * connection and not in some strange streams context.
2459 		 */
2460 		struct rpcstat *rpcstat;
2461 
2462 		rpcstat = zone_getspecific(rpcstat_zone_key, curproc->p_zone);
2463 		ASSERT(rpcstat != NULL);
2464 
2465 		RPCLOG0(1, "connmgr_connect: cannot alloc mp for "
2466 		    "sending conn request\n");
2467 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem);
2468 		e->call_status = RPC_SYSTEMERROR;
2469 		e->call_reason = ENOSR;
2470 		return (FALSE);
2471 	}
2472 
2473 	mp->b_datap->db_type = M_PROTO;
2474 	tcr = (struct T_conn_req *)mp->b_rptr;
2475 	bzero(tcr, sizeof (*tcr));
2476 	tcr->PRIM_type = T_CONN_REQ;
2477 	tcr->DEST_length = addr->len;
2478 	tcr->DEST_offset = sizeof (struct T_conn_req);
2479 	mp->b_wptr = mp->b_rptr + sizeof (*tcr);
2480 
2481 	bcopy(addr->buf, mp->b_wptr, tcr->DEST_length);
2482 	mp->b_wptr += tcr->DEST_length;
2483 
2484 	RPCLOG(8, "connmgr_connect: sending conn request on queue "
2485 	    "%p", (void *)wq);
2486 	RPCLOG(8, " call %p\n", (void *)wq);
2487 	/*
2488 	 * We use the entry in the handle that is normally used for
2489 	 * waiting for RPC replies to wait for the connection accept.
2490 	 */
2491 	clnt_dispatch_send(wq, mp, e, 0, 0);
2492 
2493 	mutex_enter(&clnt_pending_lock);
2494 
2495 	/*
2496 	 * We wait for the transport connection to be made, or an
2497 	 * indication that it could not be made.
2498 	 */
2499 	interrupted = 0;
2500 
2501 	/*
2502 	 * waitforack should have been called with T_OK_ACK, but the
2503 	 * present implementation needs to be passed T_INFO_ACK to
2504 	 * work correctly.
2505 	 */
2506 	error = waitforack(e, T_INFO_ACK, waitp, nosignal);
2507 	if (error == EINTR)
2508 		interrupted = 1;
2509 	if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) {
2510 		/*
2511 		 * No time to lose; we essentially have been signaled to
2512 		 * quit.
2513 		 */
2514 		interrupted = 1;
2515 	}
2516 #ifdef RPCDEBUG
2517 	if (error == ETIME)
2518 		RPCLOG0(8, "connmgr_connect: giving up "
2519 		    "on connection attempt; "
2520 		    "clnt_dispatch notifyconn "
2521 		    "diagnostic 'no one waiting for "
2522 		    "connection' should not be "
2523 		    "unexpected\n");
2524 #endif
2525 	if (e->call_prev)
2526 		e->call_prev->call_next = e->call_next;
2527 	else
2528 		clnt_pending = e->call_next;
2529 	if (e->call_next)
2530 		e->call_next->call_prev = e->call_prev;
2531 	mutex_exit(&clnt_pending_lock);
2532 
2533 	if (e->call_status != RPC_SUCCESS || error != 0) {
2534 		if (interrupted)
2535 			e->call_status = RPC_INTR;
2536 		else if (error == ETIME)
2537 			e->call_status = RPC_TIMEDOUT;
2538 		else if (error == EPROTO)
2539 			e->call_status = RPC_SYSTEMERROR;
2540 
2541 		RPCLOG(8, "connmgr_connect: can't connect, status: "
2542 		    "%s\n", clnt_sperrno(e->call_status));
2543 
2544 		if (e->call_reply) {
2545 			freemsg(e->call_reply);
2546 			e->call_reply = NULL;
2547 		}
2548 
2549 		return (FALSE);
2550 	}
2551 	/*
2552 	 * The result of the "connection accept" is a T_info_ack
2553 	 * in the call_reply field.
2554 	 */
2555 	ASSERT(e->call_reply != NULL);
2556 	mp = e->call_reply;
2557 	e->call_reply = NULL;
2558 	tinfo = (struct T_info_ack *)mp->b_rptr;
2559 
2560 	tidu_size = tinfo->TIDU_size;
2561 	tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT);
2562 	if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0))
2563 		tidu_size = COTS_DEFAULT_ALLOCSIZE;
2564 	*tidu_ptr = tidu_size;
2565 
2566 	freemsg(mp);
2567 
2568 	/*
2569 	 * Set up the pertinent options.  NODELAY is so the transport doesn't
2570 	 * buffer up RPC messages on either end.  This may not be valid for
2571 	 * all transports.  Failure to set this option is not cause to
2572 	 * bail out so we return success anyway.  Note that lack of NODELAY
2573 	 * or some other way to flush the message on both ends will cause
2574 	 * lots of retries and terrible performance.
2575 	 */
2576 	if (addrfmly == AF_INET || addrfmly == AF_INET6) {
2577 		(void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e);
2578 		if (e->call_status == RPC_XPRTFAILED)
2579 			return (FALSE);
2580 	}
2581 
2582 	/*
2583 	 * Since we have a connection, we now need to figure out if
2584 	 * we need to create a kstat. If x_ksp is not NULL then we
2585 	 * are reusing a connection and so we do not need to create
2586 	 * another kstat -- lets just return.
2587 	 */
2588 	if (cm_entry->x_ksp != NULL)
2589 		return (TRUE);
2590 
2591 	/*
2592 	 * We need to increment rpc_kstat_instance atomically to prevent
2593 	 * two kstats being created with the same instance.
2594 	 */
2595 	kstat_instance = atomic_add_32_nv((uint32_t *)&rpc_kstat_instance, 1);
2596 
2597 	if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance,
2598 	    "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED,
2599 	    (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)),
2600 	    KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) {
2601 		return (TRUE);
2602 	    }
2603 
2604 	cm_entry->x_ksp->ks_lock = &connmgr_lock;
2605 	cm_entry->x_ksp->ks_private = cm_entry;
2606 	cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char))
2607 					    + sizeof (cm_kstat_template));
2608 	cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size,
2609 					    KM_SLEEP);
2610 	bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data,
2611 	    cm_entry->x_ksp->ks_data_size);
2612 	((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2613 		    x_server.value.str.addr.ptr =
2614 		    kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP);
2615 
2616 	cm_entry->x_ksp->ks_update = conn_kstat_update;
2617 	kstat_install(cm_entry->x_ksp);
2618 	return (TRUE);
2619 }
2620 
2621 /*
2622  * Called by connmgr_connect to set an option on the new stream.
2623  */
2624 static bool_t
2625 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e)
2626 {
2627 	mblk_t *mp;
2628 	struct opthdr *opt;
2629 	struct T_optmgmt_req *tor;
2630 	struct timeval waitp;
2631 	int error;
2632 
2633 	mp = allocb(sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2634 	    sizeof (int), BPRI_LO);
2635 	if (mp == NULL) {
2636 		RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option "
2637 		    "request\n");
2638 		return (FALSE);
2639 	}
2640 
2641 	mp->b_datap->db_type = M_PROTO;
2642 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2643 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2644 	tor->MGMT_flags = T_NEGOTIATE;
2645 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2646 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2647 
2648 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2649 	opt->level = level;
2650 	opt->name = name;
2651 	opt->len = sizeof (int);
2652 	*(int *)((char *)opt + sizeof (*opt)) = 1;
2653 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2654 	    sizeof (int);
2655 
2656 	/*
2657 	 * We will use this connection regardless
2658 	 * of whether or not the option is settable.
2659 	 */
2660 	clnt_dispatch_send(wq, mp, e, 0, 0);
2661 	mutex_enter(&clnt_pending_lock);
2662 
2663 	waitp.tv_sec = clnt_cots_min_conntout;
2664 	waitp.tv_usec = 0;
2665 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2666 
2667 	if (e->call_prev)
2668 		e->call_prev->call_next = e->call_next;
2669 	else
2670 		clnt_pending = e->call_next;
2671 	if (e->call_next)
2672 		e->call_next->call_prev = e->call_prev;
2673 	mutex_exit(&clnt_pending_lock);
2674 
2675 	if (e->call_reply != NULL) {
2676 		freemsg(e->call_reply);
2677 		e->call_reply = NULL;
2678 	}
2679 
2680 	if (e->call_status != RPC_SUCCESS || error != 0) {
2681 		RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name);
2682 		return (FALSE);
2683 	}
2684 	RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name);
2685 	return (TRUE);
2686 }
2687 
2688 #ifdef	DEBUG
2689 
2690 /*
2691  * This is a knob to let us force code coverage in allocation failure
2692  * case.
2693  */
2694 static int	connmgr_failsnd;
2695 #define	CONN_SND_ALLOC(Size, Pri)	\
2696 	((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri))
2697 
2698 #else
2699 
2700 #define	CONN_SND_ALLOC(Size, Pri)	allocb(Size, Pri)
2701 
2702 #endif
2703 
2704 /*
2705  * Sends an orderly release on the specified queue.
2706  * Entered with connmgr_lock. Exited without connmgr_lock
2707  */
2708 static void
2709 connmgr_sndrel(struct cm_xprt *cm_entry)
2710 {
2711 	struct T_ordrel_req *torr;
2712 	mblk_t *mp;
2713 	queue_t *q = cm_entry->x_wq;
2714 	ASSERT(MUTEX_HELD(&connmgr_lock));
2715 	mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO);
2716 	if (mp == NULL) {
2717 		cm_entry->x_needrel = TRUE;
2718 		mutex_exit(&connmgr_lock);
2719 		RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel "
2720 			"to queue %p\n", (void *)q);
2721 		return;
2722 	}
2723 	mutex_exit(&connmgr_lock);
2724 
2725 	mp->b_datap->db_type = M_PROTO;
2726 	torr = (struct T_ordrel_req *)(mp->b_rptr);
2727 	torr->PRIM_type = T_ORDREL_REQ;
2728 	mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req);
2729 
2730 	RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q);
2731 	put(q, mp);
2732 }
2733 
2734 /*
2735  * Sends an disconnect on the specified queue.
2736  * Entered with connmgr_lock. Exited without connmgr_lock
2737  */
2738 static void
2739 connmgr_snddis(struct cm_xprt *cm_entry)
2740 {
2741 	struct T_discon_req *tdis;
2742 	mblk_t *mp;
2743 	queue_t *q = cm_entry->x_wq;
2744 
2745 	ASSERT(MUTEX_HELD(&connmgr_lock));
2746 	mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO);
2747 	if (mp == NULL) {
2748 		cm_entry->x_needdis = TRUE;
2749 		mutex_exit(&connmgr_lock);
2750 		RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon "
2751 		    "to queue %p\n", (void *)q);
2752 		return;
2753 	}
2754 	mutex_exit(&connmgr_lock);
2755 
2756 	mp->b_datap->db_type = M_PROTO;
2757 	tdis = (struct T_discon_req *)mp->b_rptr;
2758 	tdis->PRIM_type = T_DISCON_REQ;
2759 	mp->b_wptr = mp->b_rptr + sizeof (*tdis);
2760 
2761 	RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q);
2762 	put(q, mp);
2763 }
2764 
2765 /*
2766  * Sets up the entry for receiving replies, and calls rpcmod's write put proc
2767  * (through put) to send the call.
2768  */
2769 static void
2770 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid,
2771 			uint_t queue_flag)
2772 {
2773 	ASSERT(e != NULL);
2774 
2775 	e->call_status = RPC_TIMEDOUT;	/* optimistic, eh? */
2776 	e->call_reason = 0;
2777 	e->call_wq = q;
2778 	e->call_xid = xid;
2779 	e->call_notified = FALSE;
2780 
2781 	/*
2782 	 * If queue_flag is set then the calllist_t is already on the hash
2783 	 * queue.  In this case just send the message and return.
2784 	 */
2785 	if (queue_flag) {
2786 		put(q, mp);
2787 		return;
2788 	}
2789 
2790 	/*
2791 	 * Set up calls for RPC requests (with XID != 0) on the hash
2792 	 * queue for fast lookups and place other calls (i.e.
2793 	 * connection management) on the linked list.
2794 	 */
2795 	if (xid != 0) {
2796 		RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on "
2797 			"dispatch list\n", xid);
2798 		e->call_hash = call_hash(xid, clnt_cots_hash_size);
2799 		e->call_bucket = &cots_call_ht[e->call_hash];
2800 		call_table_enter(e);
2801 	} else {
2802 		mutex_enter(&clnt_pending_lock);
2803 		if (clnt_pending)
2804 			clnt_pending->call_prev = e;
2805 		e->call_next = clnt_pending;
2806 		e->call_prev = NULL;
2807 		clnt_pending = e;
2808 		mutex_exit(&clnt_pending_lock);
2809 	}
2810 
2811 	put(q, mp);
2812 }
2813 
2814 /*
2815  * Called by rpcmod to notify a client with a clnt_pending call that its reply
2816  * has arrived.  If we can't find a client waiting for this reply, we log
2817  * the error and return.
2818  */
2819 bool_t
2820 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid)
2821 {
2822 	calllist_t *e = NULL;
2823 	call_table_t *chtp;
2824 	uint32_t xid;
2825 	uint_t hash;
2826 
2827 	if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) &&
2828 	    (mp->b_wptr - mp->b_rptr) >= sizeof (xid))
2829 		xid = *((uint32_t *)mp->b_rptr);
2830 	else {
2831 		int i = 0;
2832 		unsigned char *p = (unsigned char *)&xid;
2833 		unsigned char *rptr;
2834 		mblk_t *tmp = mp;
2835 
2836 		/*
2837 		 * Copy the xid, byte-by-byte into xid.
2838 		 */
2839 		while (tmp) {
2840 			rptr = tmp->b_rptr;
2841 			while (rptr < tmp->b_wptr) {
2842 				*p++ = *rptr++;
2843 				if (++i >= sizeof (xid))
2844 					goto done_xid_copy;
2845 			}
2846 			tmp = tmp->b_cont;
2847 		}
2848 
2849 		/*
2850 		 * If we got here, we ran out of mblk space before the
2851 		 * xid could be copied.
2852 		 */
2853 		ASSERT(tmp == NULL && i < sizeof (xid));
2854 
2855 		RPCLOG0(1,
2856 		    "clnt_dispatch_notify: message less than size of xid\n");
2857 		return (FALSE);
2858 
2859 	}
2860 done_xid_copy:
2861 
2862 	hash = call_hash(xid, clnt_cots_hash_size);
2863 	chtp = &cots_call_ht[hash];
2864 	/* call_table_find returns with the hash bucket locked */
2865 	call_table_find(chtp, xid, e);
2866 
2867 	if (e != NULL) {
2868 		/*
2869 		 * Found thread waiting for this reply
2870 		 */
2871 		mutex_enter(&e->call_lock);
2872 		if (e->call_reply)
2873 			/*
2874 			 * This can happen under the following scenario:
2875 			 * clnt_cots_kcallit() times out on the response,
2876 			 * rfscall() repeats the CLNT_CALL() with
2877 			 * the same xid, clnt_cots_kcallit() sends the retry,
2878 			 * thereby putting the clnt handle on the pending list,
2879 			 * the first response arrives, signalling the thread
2880 			 * in clnt_cots_kcallit(). Before that thread is
2881 			 * dispatched, the second response arrives as well,
2882 			 * and clnt_dispatch_notify still finds the handle on
2883 			 * the pending list, with call_reply set. So free the
2884 			 * old reply now.
2885 			 *
2886 			 * It is also possible for a response intended for
2887 			 * an RPC call with a different xid to reside here.
2888 			 * This can happen if the thread that owned this
2889 			 * client handle prior to the current owner bailed
2890 			 * out and left its call record on the dispatch
2891 			 * queue.  A window exists where the response can
2892 			 * arrive before the current owner dispatches its
2893 			 * RPC call.
2894 			 *
2895 			 * In any case, this is the very last point where we
2896 			 * can safely check the call_reply field before
2897 			 * placing the new response there.
2898 			 */
2899 			freemsg(e->call_reply);
2900 		e->call_reply = mp;
2901 		e->call_status = RPC_SUCCESS;
2902 		e->call_notified = TRUE;
2903 		cv_signal(&e->call_cv);
2904 		mutex_exit(&e->call_lock);
2905 		mutex_exit(&chtp->ct_lock);
2906 		return (TRUE);
2907 	} else {
2908 		zone_t *zone;
2909 		struct rpcstat *rpcstat;
2910 
2911 		mutex_exit(&chtp->ct_lock);
2912 		RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n",
2913 		    xid);
2914 		/*
2915 		 * This is unfortunate, but we need to lookup the zone so we
2916 		 * can increment its "rcbadxids" counter.
2917 		 */
2918 		zone = zone_find_by_id(zoneid);
2919 		if (zone == NULL) {
2920 			/*
2921 			 * The zone went away...
2922 			 */
2923 			return (FALSE);
2924 		}
2925 		rpcstat = zone_getspecific(rpcstat_zone_key, zone);
2926 		if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) {
2927 			/*
2928 			 * Not interested
2929 			 */
2930 			zone_rele(zone);
2931 			return (FALSE);
2932 		}
2933 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids);
2934 		zone_rele(zone);
2935 	}
2936 	return (FALSE);
2937 }
2938 
2939 /*
2940  * Called by rpcmod when a non-data indication arrives.  The ones in which we
2941  * are interested are connection indications and options acks.  We dispatch
2942  * based on the queue the indication came in on.  If we are not interested in
2943  * what came in, we return false to rpcmod, who will then pass it upstream.
2944  */
2945 bool_t
2946 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp)
2947 {
2948 	calllist_t *e;
2949 	int type;
2950 
2951 	ASSERT((q->q_flag & QREADR) == 0);
2952 
2953 	type = ((union T_primitives *)mp->b_rptr)->type;
2954 	RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n",
2955 	    rpc_tpiprim2name(type));
2956 	mutex_enter(&clnt_pending_lock);
2957 	for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) {
2958 		if (e == NULL) {
2959 			mutex_exit(&clnt_pending_lock);
2960 			RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting "
2961 			    "for connection on queue 0x%p\n", (void *)q);
2962 			return (FALSE);
2963 		}
2964 		if (e->call_wq == q)
2965 			break;
2966 	}
2967 
2968 	switch (type) {
2969 	case T_CONN_CON:
2970 		/*
2971 		 * The transport is now connected, send a T_INFO_REQ to get
2972 		 * the tidu size.
2973 		 */
2974 		mutex_exit(&clnt_pending_lock);
2975 		ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >=
2976 			sizeof (struct T_info_req));
2977 		mp->b_rptr = mp->b_datap->db_base;
2978 		((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ;
2979 		mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req);
2980 		mp->b_datap->db_type = M_PCPROTO;
2981 		put(q, mp);
2982 		return (TRUE);
2983 	case T_INFO_ACK:
2984 	case T_OPTMGMT_ACK:
2985 		e->call_status = RPC_SUCCESS;
2986 		e->call_reply = mp;
2987 		e->call_notified = TRUE;
2988 		cv_signal(&e->call_cv);
2989 		break;
2990 	case T_ERROR_ACK:
2991 		e->call_status = RPC_CANTCONNECT;
2992 		e->call_reply = mp;
2993 		e->call_notified = TRUE;
2994 		cv_signal(&e->call_cv);
2995 		break;
2996 	case T_OK_ACK:
2997 		/*
2998 		 * Great, but we are really waiting for a T_CONN_CON
2999 		 */
3000 		freemsg(mp);
3001 		break;
3002 	default:
3003 		mutex_exit(&clnt_pending_lock);
3004 		RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type);
3005 		return (FALSE);
3006 	}
3007 
3008 	mutex_exit(&clnt_pending_lock);
3009 	return (TRUE);
3010 }
3011 
3012 /*
3013  * Called by rpcmod when the transport is (or should be) going away.  Informs
3014  * all callers waiting for replies and marks the entry in the connection
3015  * manager's list as unconnected, and either closing (close handshake in
3016  * progress) or dead.
3017  */
3018 void
3019 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason)
3020 {
3021 	calllist_t *e;
3022 	call_table_t *ctp;
3023 	struct cm_xprt *cm_entry;
3024 	int have_connmgr_lock;
3025 	int i;
3026 
3027 	ASSERT((q->q_flag & QREADR) == 0);
3028 
3029 	RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q);
3030 	RPCLOG(1, " received a notifcation prim type [%s]",
3031 	    rpc_tpiprim2name(msg_type));
3032 	RPCLOG(1, " and reason %d\n", reason);
3033 
3034 	/*
3035 	 * Find the transport entry in the connection manager's list, close
3036 	 * the transport and delete the entry.  In the case where rpcmod's
3037 	 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we
3038 	 * should gracefully close the connection.
3039 	 */
3040 	have_connmgr_lock = 1;
3041 	mutex_enter(&connmgr_lock);
3042 	for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) {
3043 		ASSERT(cm_entry != cm_entry->x_next);
3044 		if (cm_entry->x_wq == q) {
3045 			ASSERT(MUTEX_HELD(&connmgr_lock));
3046 			ASSERT(have_connmgr_lock == 1);
3047 			switch (msg_type) {
3048 			case T_ORDREL_REQ:
3049 
3050 				if (cm_entry->x_dead) {
3051 					RPCLOG(1, "idle timeout on dead "
3052 					    "connection: %p\n",
3053 					    (void *)cm_entry);
3054 					if (clnt_stop_idle != NULL)
3055 						(*clnt_stop_idle)(q);
3056 					break;
3057 				}
3058 
3059 				/*
3060 				 * Only mark the connection as dead if it is
3061 				 * connected and idle.
3062 				 * An unconnected connection has probably
3063 				 * gone idle because the server is down,
3064 				 * and when it comes back up there will be
3065 				 * retries that need to use that connection.
3066 				 */
3067 				if (cm_entry->x_connected ||
3068 				    cm_entry->x_doomed) {
3069 				    if (cm_entry->x_ordrel) {
3070 					if (cm_entry->x_closing == TRUE) {
3071 					/*
3072 					 * The connection is obviously
3073 					 * wedged due to a bug or problem
3074 					 * with the transport. Mark it
3075 					 * as dead. Otherwise we can leak
3076 					 * connections.
3077 					 */
3078 					    cm_entry->x_dead = TRUE;
3079 					    mutex_exit(&connmgr_lock);
3080 					    have_connmgr_lock = 0;
3081 					    if (clnt_stop_idle != NULL)
3082 						(*clnt_stop_idle)(q);
3083 					    break;
3084 					}
3085 					cm_entry->x_closing = TRUE;
3086 					connmgr_sndrel(cm_entry);
3087 					have_connmgr_lock = 0;
3088 				    } else {
3089 					cm_entry->x_dead = TRUE;
3090 					mutex_exit(&connmgr_lock);
3091 					have_connmgr_lock = 0;
3092 					if (clnt_stop_idle != NULL)
3093 						(*clnt_stop_idle)(q);
3094 				    }
3095 				} else {
3096 					/*
3097 					 * We don't mark the connection
3098 					 * as dead, but we turn off the
3099 					 * idle timer.
3100 					 */
3101 					mutex_exit(&connmgr_lock);
3102 					have_connmgr_lock = 0;
3103 					if (clnt_stop_idle != NULL)
3104 						(*clnt_stop_idle)(q);
3105 					RPCLOG(1, "clnt_dispatch_notifyall:"
3106 					    " ignoring timeout from rpcmod"
3107 					    " (q %p) because we are not "
3108 					    " connected\n", (void *)q);
3109 				}
3110 				break;
3111 			case T_ORDREL_IND:
3112 				/*
3113 				 * If this entry is marked closing, then we are
3114 				 * completing a close handshake, and the
3115 				 * connection is dead.  Otherwise, the server is
3116 				 * trying to close. Since the server will not
3117 				 * be sending any more RPC replies, we abort
3118 				 * the connection, including flushing
3119 				 * any RPC requests that are in-transit.
3120 				 */
3121 				if (cm_entry->x_closing) {
3122 					cm_entry->x_dead = TRUE;
3123 					mutex_exit(&connmgr_lock);
3124 					have_connmgr_lock = 0;
3125 					if (clnt_stop_idle != NULL)
3126 						(*clnt_stop_idle)(q);
3127 				} else {
3128 					/*
3129 					 * if we're getting a disconnect
3130 					 * before we've finished our
3131 					 * connect attempt, mark it for
3132 					 * later processing
3133 					 */
3134 					if (cm_entry->x_thread)
3135 						cm_entry->x_early_disc = TRUE;
3136 					else
3137 						cm_entry->x_connected = FALSE;
3138 					cm_entry->x_waitdis = TRUE;
3139 					connmgr_snddis(cm_entry);
3140 					have_connmgr_lock = 0;
3141 				}
3142 				break;
3143 
3144 			case T_ERROR_ACK:
3145 			case T_OK_ACK:
3146 				cm_entry->x_waitdis = FALSE;
3147 				cv_signal(&cm_entry->x_dis_cv);
3148 				mutex_exit(&connmgr_lock);
3149 				return;
3150 
3151 			case T_DISCON_REQ:
3152 				if (cm_entry->x_thread)
3153 					cm_entry->x_early_disc = TRUE;
3154 				else
3155 					cm_entry->x_connected = FALSE;
3156 				cm_entry->x_waitdis = TRUE;
3157 
3158 				connmgr_snddis(cm_entry);
3159 				have_connmgr_lock = 0;
3160 				break;
3161 
3162 			case T_DISCON_IND:
3163 			default:
3164 				/*
3165 				 * if we're getting a disconnect before
3166 				 * we've finished our connect attempt,
3167 				 * mark it for later processing
3168 				 */
3169 				if (cm_entry->x_closing) {
3170 					cm_entry->x_dead = TRUE;
3171 					mutex_exit(&connmgr_lock);
3172 					have_connmgr_lock = 0;
3173 					if (clnt_stop_idle != NULL)
3174 						(*clnt_stop_idle)(q);
3175 				} else {
3176 					if (cm_entry->x_thread) {
3177 						cm_entry->x_early_disc = TRUE;
3178 					} else {
3179 						cm_entry->x_dead = TRUE;
3180 						cm_entry->x_connected = FALSE;
3181 					}
3182 				}
3183 				break;
3184 			}
3185 			break;
3186 		}
3187 	}
3188 
3189 	if (have_connmgr_lock)
3190 		mutex_exit(&connmgr_lock);
3191 
3192 	if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) {
3193 		RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find "
3194 		    "connmgr entry for discon ack\n", (void *)q);
3195 		return;
3196 	}
3197 
3198 	/*
3199 	 * Then kick all the clnt_pending calls out of their wait.  There
3200 	 * should be no clnt_pending calls in the case of rpcmod's idle
3201 	 * timer firing.
3202 	 */
3203 	for (i = 0; i < clnt_cots_hash_size; i++) {
3204 		ctp = &cots_call_ht[i];
3205 		mutex_enter(&ctp->ct_lock);
3206 		for (e = ctp->ct_call_next;
3207 			e != (calllist_t *)ctp;
3208 			e = e->call_next) {
3209 			if (e->call_wq == q && e->call_notified == FALSE) {
3210 				RPCLOG(1,
3211 				"clnt_dispatch_notifyall for queue %p ",
3212 					(void *)q);
3213 				RPCLOG(1, "aborting clnt_pending call %p\n",
3214 					(void *)e);
3215 
3216 				if (msg_type == T_DISCON_IND)
3217 					e->call_reason = reason;
3218 				e->call_notified = TRUE;
3219 				e->call_status = RPC_XPRTFAILED;
3220 				cv_signal(&e->call_cv);
3221 			}
3222 		}
3223 		mutex_exit(&ctp->ct_lock);
3224 	}
3225 
3226 	mutex_enter(&clnt_pending_lock);
3227 	for (e = clnt_pending; e; e = e->call_next) {
3228 		/*
3229 		 * Only signal those RPC handles that haven't been
3230 		 * signalled yet. Otherwise we can get a bogus call_reason.
3231 		 * This can happen if thread A is making a call over a
3232 		 * connection. If the server is killed, it will cause
3233 		 * reset, and reason will default to EIO as a result of
3234 		 * a T_ORDREL_IND. Thread B then attempts to recreate
3235 		 * the connection but gets a T_DISCON_IND. If we set the
3236 		 * call_reason code for all threads, then if thread A
3237 		 * hasn't been dispatched yet, it will get the wrong
3238 		 * reason. The bogus call_reason can make it harder to
3239 		 * discriminate between calls that fail because the
3240 		 * connection attempt failed versus those where the call
3241 		 * may have been executed on the server.
3242 		 */
3243 		if (e->call_wq == q && e->call_notified == FALSE) {
3244 			RPCLOG(1, "clnt_dispatch_notifyall for queue %p ",
3245 			    (void *)q);
3246 			RPCLOG(1, " aborting clnt_pending call %p\n",
3247 			    (void *)e);
3248 
3249 			if (msg_type == T_DISCON_IND)
3250 				e->call_reason = reason;
3251 			e->call_notified = TRUE;
3252 			/*
3253 			 * Let the caller timeout, else he will retry
3254 			 * immediately.
3255 			 */
3256 			e->call_status = RPC_XPRTFAILED;
3257 
3258 			/*
3259 			 * We used to just signal those threads
3260 			 * waiting for a connection, (call_xid = 0).
3261 			 * That meant that threads waiting for a response
3262 			 * waited till their timeout expired. This
3263 			 * could be a long time if they've specified a
3264 			 * maximum timeout. (2^31 - 1). So we
3265 			 * Signal all threads now.
3266 			 */
3267 			cv_signal(&e->call_cv);
3268 		}
3269 	}
3270 	mutex_exit(&clnt_pending_lock);
3271 }
3272 
3273 
3274 /*ARGSUSED*/
3275 /*
3276  * after resuming a system that's been suspended for longer than the
3277  * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall()
3278  * generates "NFS server X not responding" and "NFS server X ok" messages;
3279  * here we reset inet connections to cause a re-connect and avoid those
3280  * NFS messages.  see 4045054
3281  */
3282 boolean_t
3283 connmgr_cpr_reset(void *arg, int code)
3284 {
3285 	struct cm_xprt *cxp;
3286 
3287 	if (code == CB_CODE_CPR_CHKPT)
3288 		return (B_TRUE);
3289 
3290 	if (mutex_tryenter(&connmgr_lock) == 0)
3291 		return (B_FALSE);
3292 	for (cxp = cm_hd; cxp; cxp = cxp->x_next) {
3293 		if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) &&
3294 			cxp->x_connected == TRUE) {
3295 			if (cxp->x_thread)
3296 				cxp->x_early_disc = TRUE;
3297 			else
3298 				cxp->x_connected = FALSE;
3299 			cxp->x_needdis = TRUE;
3300 		}
3301 	}
3302 	mutex_exit(&connmgr_lock);
3303 	return (B_TRUE);
3304 }
3305 
3306 void
3307 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp)
3308 {
3309 
3310 	*statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid,
3311 	    "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl,
3312 	    sizeof (cots_rcstat_tmpl));
3313 }
3314 
3315 void
3316 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp)
3317 {
3318 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client");
3319 	kmem_free(*statsp, sizeof (cots_rcstat_tmpl));
3320 }
3321 
3322 void
3323 clnt_cots_init(void)
3324 {
3325 	mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL);
3326 	mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL);
3327 
3328 	if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE)
3329 		clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE;
3330 
3331 	cots_call_ht = call_table_init(clnt_cots_hash_size);
3332 	zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy);
3333 }
3334 
3335 void
3336 clnt_cots_fini(void)
3337 {
3338 	(void) zone_key_delete(zone_cots_key);
3339 }
3340 
3341 /*
3342  * Wait for TPI ack, returns success only if expected ack is received
3343  * within timeout period.
3344  */
3345 
3346 static int
3347 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp,
3348     bool_t nosignal)
3349 {
3350 	union T_primitives *tpr;
3351 	clock_t timout;
3352 	int cv_stat = 1;
3353 
3354 	ASSERT(MUTEX_HELD(&clnt_pending_lock));
3355 	while (e->call_reply == NULL) {
3356 		if (waitp != NULL) {
3357 			timout = waitp->tv_sec * drv_usectohz(MICROSEC) +
3358 			    drv_usectohz(waitp->tv_usec) + lbolt;
3359 			if (nosignal)
3360 				cv_stat = cv_timedwait(&e->call_cv,
3361 				    &clnt_pending_lock, timout);
3362 			else
3363 				cv_stat = cv_timedwait_sig(&e->call_cv,
3364 				    &clnt_pending_lock, timout);
3365 		} else {
3366 			if (nosignal)
3367 				cv_wait(&e->call_cv, &clnt_pending_lock);
3368 			else
3369 				cv_stat = cv_wait_sig(&e->call_cv,
3370 				    &clnt_pending_lock);
3371 		}
3372 		if (cv_stat == -1)
3373 			return (ETIME);
3374 		if (cv_stat == 0)
3375 			return (EINTR);
3376 	}
3377 	tpr = (union T_primitives *)e->call_reply->b_rptr;
3378 	if (tpr->type == ack_prim)
3379 		return (0); /* Success */
3380 
3381 	if (tpr->type == T_ERROR_ACK) {
3382 		if (tpr->error_ack.TLI_error == TSYSERR)
3383 			return (tpr->error_ack.UNIX_error);
3384 		else
3385 			return (t_tlitosyserr(tpr->error_ack.TLI_error));
3386 	}
3387 
3388 	return (EPROTO); /* unknown or unexpected primitive */
3389 }
3390