xref: /titanic_41/usr/src/uts/common/rpc/clnt_cots.c (revision 1a1a84a324206b6b1f5f704ab166c4ebf78aed76)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T
28  *		All Rights Reserved
29  */
30 
31 /*
32  * Portions of this source code were derived from Berkeley 4.3 BSD
33  * under license from the Regents of the University of California.
34  */
35 
36 
37 /*
38  * Implements a kernel based, client side RPC over Connection Oriented
39  * Transports (COTS).
40  */
41 
42 /*
43  * Much of this file has been re-written to let NFS work better over slow
44  * transports. A description follows.
45  *
46  * One of the annoying things about kRPC/COTS is that it will temporarily
47  * create more than one connection between a client and server. This
48  * happens because when a connection is made, the end-points entry in the
49  * linked list of connections (headed by cm_hd), is removed so that other
50  * threads don't mess with it. Went ahead and bit the bullet by keeping
51  * the endpoint on the connection list and introducing state bits,
52  * condition variables etc. to the connection entry data structure (struct
53  * cm_xprt).
54  *
55  * Here is a summary of the changes to cm-xprt:
56  *
57  *	x_ctime is the timestamp of when the endpoint was last
58  *	connected or disconnected. If an end-point is ever disconnected
59  *	or re-connected, then any outstanding RPC request is presumed
60  *	lost, telling clnt_cots_kcallit that it needs to re-send the
61  *	request, not just wait for the original request's reply to
62  *	arrive.
63  *
64  *	x_thread flag which tells us if a thread is doing a connection attempt.
65  *
66  *	x_waitdis flag which tells us we are waiting a disconnect ACK.
67  *
68  *	x_needdis flag which tells us we need to send a T_DISCONN_REQ
69  *	to kill the connection.
70  *
71  *	x_needrel flag which tells us we need to send a T_ORDREL_REQ to
72  *	gracefully close the connection.
73  *
74  *	#defined bitmasks for the all the b_* bits so that more
75  *	efficient (and at times less clumsy) masks can be used to
76  *	manipulated state in cases where multiple bits have to
77  *	set/cleared/checked in the same critical section.
78  *
79  *	x_conn_cv and x_dis-_cv are new condition variables to let
80  *	threads knows when the connection attempt is done, and to let
81  *	the connecting thread know when the disconnect handshake is
82  *	done.
83  *
84  * Added the CONN_HOLD() macro so that all reference holds have the same
85  * look and feel.
86  *
87  * In the private (cku_private) portion of the client handle,
88  *
89  *	cku_flags replaces the cku_sent a boolean. cku_flags keeps
90  *	track of whether a request as been sent, and whether the
91  *	client's handles call record is on the dispatch list (so that
92  *	the reply can be matched by XID to the right client handle).
93  *	The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit()
94  *	and still have the response find the right client handle so
95  *	that the retry of CLNT_CALL() gets the result. Testing, found
96  *	situations where if the timeout was increased, performance
97  *	degraded. This was due to us hitting a window where the thread
98  *	was back in rfscall() (probably printing server not responding)
99  *	while the response came back but no place to put it.
100  *
101  *	cku_ctime is just a cache of x_ctime. If they match,
102  *	clnt_cots_kcallit() won't to send a retry (unless the maximum
103  *	receive count limit as been reached). If the don't match, then
104  *	we assume the request has been lost, and a retry of the request
105  *	is needed.
106  *
107  *	cku_recv_attempts counts the number of receive count attempts
108  *	after one try is sent on the wire.
109  *
110  * Added the clnt_delay() routine so that interruptible and
111  * noninterruptible delays are possible.
112  *
113  * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to
114  * control how long the client delays before returned after getting
115  * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash
116  * a server that may be booting and not yet started nfsd.
117  *
118  * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable)
119  * Why don't we just wait forever (receive an infinite # of times)?
120  * Because the server may have rebooted. More insidious is that some
121  * servers (ours) will drop NFS/TCP requests in some cases. This is bad,
122  * but it is a reality.
123  *
124  * The case of a server doing orderly release really messes up the
125  * client's recovery, especially if the server's TCP implementation is
126  * buggy.  It was found was that the kRPC/COTS client was breaking some
127  * TPI rules, such as not waiting for the acknowledgement of a
128  * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and
129  * T_DISCON_REQ in clnt_dispatch_notifyall()).
130  *
131  * One of things that we've seen is that a kRPC TCP endpoint goes into
132  * TIMEWAIT and a thus a reconnect takes a long time to satisfy because
133  * that the TIMEWAIT state takes a while to finish.  If a server sends a
134  * T_ORDREL_IND, there is little point in an RPC client doing a
135  * T_ORDREL_REQ, because the RPC request isn't going to make it (the
136  * server is saying that it won't accept any more data). So kRPC was
137  * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the
138  * connection skips the TIMEWAIT state and goes straight to a bound state
139  * that kRPC can quickly switch to connected.
140  *
141  * Code that issues TPI request must use waitforack() to wait for the
142  * corresponding ack (assuming there is one) in any future modifications.
143  * This works around problems that may be introduced by breaking TPI rules
144  * (by submitting new calls before earlier requests have been acked) in the
145  * case of a signal or other early return.  waitforack() depends on
146  * clnt_dispatch_notifyconn() to issue the wakeup when the ack
147  * arrives, so adding new TPI calls may require corresponding changes
148  * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on
149  * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure
150  * not to set it too low or TPI ACKS will be lost.
151  */
152 
153 #include <sys/param.h>
154 #include <sys/types.h>
155 #include <sys/user.h>
156 #include <sys/systm.h>
157 #include <sys/sysmacros.h>
158 #include <sys/proc.h>
159 #include <sys/socket.h>
160 #include <sys/file.h>
161 #include <sys/stream.h>
162 #include <sys/strsubr.h>
163 #include <sys/stropts.h>
164 #include <sys/strsun.h>
165 #include <sys/timod.h>
166 #include <sys/tiuser.h>
167 #include <sys/tihdr.h>
168 #include <sys/t_kuser.h>
169 #include <sys/fcntl.h>
170 #include <sys/errno.h>
171 #include <sys/kmem.h>
172 #include <sys/debug.h>
173 #include <sys/systm.h>
174 #include <sys/kstat.h>
175 #include <sys/t_lock.h>
176 #include <sys/ddi.h>
177 #include <sys/cmn_err.h>
178 #include <sys/time.h>
179 #include <sys/isa_defs.h>
180 #include <sys/callb.h>
181 #include <sys/sunddi.h>
182 #include <sys/atomic.h>
183 #include <sys/sdt.h>
184 
185 #include <netinet/in.h>
186 #include <netinet/tcp.h>
187 
188 #include <rpc/types.h>
189 #include <rpc/xdr.h>
190 #include <rpc/auth.h>
191 #include <rpc/clnt.h>
192 #include <rpc/rpc_msg.h>
193 
194 #define	COTS_DEFAULT_ALLOCSIZE	2048
195 
196 #define	WIRE_HDR_SIZE	20	/* serialized call header, sans proc number */
197 #define	MSG_OFFSET	128	/* offset of call into the mblk */
198 
199 const char *kinet_ntop6(uchar_t *, char *, size_t);
200 
201 static int	clnt_cots_ksettimers(CLIENT *, struct rpc_timers *,
202     struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t);
203 static enum clnt_stat	clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t,
204     caddr_t, xdrproc_t, caddr_t, struct timeval);
205 static void	clnt_cots_kabort(CLIENT *);
206 static void	clnt_cots_kerror(CLIENT *, struct rpc_err *);
207 static bool_t	clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t);
208 static void	clnt_cots_kdestroy(CLIENT *);
209 static bool_t	clnt_cots_kcontrol(CLIENT *, int, char *);
210 
211 
212 /* List of transports managed by the connection manager. */
213 struct cm_xprt {
214 	TIUSER		*x_tiptr;	/* transport handle */
215 	queue_t		*x_wq;		/* send queue */
216 	clock_t		x_time;		/* last time we handed this xprt out */
217 	clock_t		x_ctime;	/* time we went to CONNECTED */
218 	int		x_tidu_size;    /* TIDU size of this transport */
219 	union {
220 	    struct {
221 		unsigned int
222 #ifdef	_BIT_FIELDS_HTOL
223 		b_closing:	1,	/* we've sent a ord rel on this conn */
224 		b_dead:		1,	/* transport is closed or disconn */
225 		b_doomed:	1,	/* too many conns, let this go idle */
226 		b_connected:	1,	/* this connection is connected */
227 
228 		b_ordrel:	1,	/* do an orderly release? */
229 		b_thread:	1,	/* thread doing connect */
230 		b_waitdis:	1,	/* waiting for disconnect ACK */
231 		b_needdis:	1,	/* need T_DISCON_REQ */
232 
233 		b_needrel:	1,	/* need T_ORDREL_REQ */
234 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
235 					/* disconnect during connect */
236 
237 		b_pad:		22;
238 
239 #endif
240 
241 #ifdef	_BIT_FIELDS_LTOH
242 		b_pad:		22,
243 
244 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
245 					/* disconnect during connect */
246 		b_needrel:	1,	/* need T_ORDREL_REQ */
247 
248 		b_needdis:	1,	/* need T_DISCON_REQ */
249 		b_waitdis:	1,	/* waiting for disconnect ACK */
250 		b_thread:	1,	/* thread doing connect */
251 		b_ordrel:	1,	/* do an orderly release? */
252 
253 		b_connected:	1,	/* this connection is connected */
254 		b_doomed:	1,	/* too many conns, let this go idle */
255 		b_dead:		1,	/* transport is closed or disconn */
256 		b_closing:	1;	/* we've sent a ord rel on this conn */
257 #endif
258 	    } bit;	    unsigned int word;
259 
260 #define	x_closing	x_state.bit.b_closing
261 #define	x_dead		x_state.bit.b_dead
262 #define	x_doomed	x_state.bit.b_doomed
263 #define	x_connected	x_state.bit.b_connected
264 
265 #define	x_ordrel	x_state.bit.b_ordrel
266 #define	x_thread	x_state.bit.b_thread
267 #define	x_waitdis	x_state.bit.b_waitdis
268 #define	x_needdis	x_state.bit.b_needdis
269 
270 #define	x_needrel	x_state.bit.b_needrel
271 #define	x_early_disc    x_state.bit.b_early_disc
272 
273 #define	x_state_flags	x_state.word
274 
275 #define	X_CLOSING	0x80000000
276 #define	X_DEAD		0x40000000
277 #define	X_DOOMED	0x20000000
278 #define	X_CONNECTED	0x10000000
279 
280 #define	X_ORDREL	0x08000000
281 #define	X_THREAD	0x04000000
282 #define	X_WAITDIS	0x02000000
283 #define	X_NEEDDIS	0x01000000
284 
285 #define	X_NEEDREL	0x00800000
286 #define	X_EARLYDISC	0x00400000
287 
288 #define	X_BADSTATES	(X_CLOSING | X_DEAD | X_DOOMED)
289 
290 	}		x_state;
291 	int		x_ref;		/* number of users of this xprt */
292 	int		x_family;	/* address family of transport */
293 	dev_t		x_rdev;		/* device number of transport */
294 	struct cm_xprt	*x_next;
295 
296 	struct netbuf	x_server;	/* destination address */
297 	struct netbuf	x_src;		/* src address (for retries) */
298 	kmutex_t	x_lock;		/* lock on this entry */
299 	kcondvar_t	x_cv;		/* to signal when can be closed */
300 	kcondvar_t	x_conn_cv;	/* to signal when connection attempt */
301 					/* is complete */
302 	kstat_t		*x_ksp;
303 
304 	kcondvar_t	x_dis_cv;	/* to signal when disconnect attempt */
305 					/* is complete */
306 	zoneid_t	x_zoneid;	/* zone this xprt belongs to */
307 };
308 
309 typedef struct cm_kstat_xprt {
310 	kstat_named_t	x_wq;
311 	kstat_named_t	x_server;
312 	kstat_named_t	x_family;
313 	kstat_named_t	x_rdev;
314 	kstat_named_t	x_time;
315 	kstat_named_t	x_state;
316 	kstat_named_t	x_ref;
317 	kstat_named_t	x_port;
318 } cm_kstat_xprt_t;
319 
320 static cm_kstat_xprt_t cm_kstat_template = {
321 	{ "write_queue", KSTAT_DATA_UINT32 },
322 	{ "server",	KSTAT_DATA_STRING },
323 	{ "addr_family", KSTAT_DATA_UINT32 },
324 	{ "device",	KSTAT_DATA_UINT32 },
325 	{ "time_stamp",	KSTAT_DATA_UINT32 },
326 	{ "status",	KSTAT_DATA_UINT32 },
327 	{ "ref_count",	KSTAT_DATA_INT32 },
328 	{ "port",	KSTAT_DATA_UINT32 },
329 };
330 
331 /*
332  * The inverse of this is connmgr_release().
333  */
334 #define	CONN_HOLD(Cm_entry)	{\
335 	mutex_enter(&(Cm_entry)->x_lock);	\
336 	(Cm_entry)->x_ref++;	\
337 	mutex_exit(&(Cm_entry)->x_lock);	\
338 }
339 
340 
341 /*
342  * Private data per rpc handle.  This structure is allocated by
343  * clnt_cots_kcreate, and freed by clnt_cots_kdestroy.
344  */
345 typedef struct cku_private_s {
346 	CLIENT			cku_client;	/* client handle */
347 	calllist_t		cku_call;	/* for dispatching calls */
348 	struct rpc_err		cku_err;	/* error status */
349 
350 	struct netbuf		cku_srcaddr;	/* source address for retries */
351 	int			cku_addrfmly;  /* for binding port */
352 	struct netbuf		cku_addr;	/* remote address */
353 	dev_t			cku_device;	/* device to use */
354 	uint_t			cku_flags;
355 #define	CKU_ONQUEUE		0x1
356 #define	CKU_SENT		0x2
357 
358 	bool_t			cku_progress;	/* for CLSET_PROGRESS */
359 	uint32_t		cku_xid;	/* current XID */
360 	clock_t			cku_ctime;	/* time stamp of when */
361 						/* connection was created */
362 	uint_t			cku_recv_attempts;
363 	XDR			cku_outxdr;	/* xdr routine for output */
364 	XDR			cku_inxdr;	/* xdr routine for input */
365 	char			cku_rpchdr[WIRE_HDR_SIZE + 4];
366 						/* pre-serialized rpc header */
367 
368 	uint_t			cku_outbuflen;	/* default output mblk length */
369 	struct cred		*cku_cred;	/* credentials */
370 	bool_t			cku_nodelayonerr;
371 						/* for CLSET_NODELAYONERR */
372 	int			cku_useresvport; /* Use reserved port */
373 	struct rpc_cots_client	*cku_stats;	/* stats for zone */
374 } cku_private_t;
375 
376 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *,
377 	const struct timeval *, struct netbuf *, int, struct netbuf *,
378 	struct rpc_err *, bool_t, bool_t);
379 
380 static bool_t	connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *,
381 				int, calllist_t *, int *, bool_t reconnect,
382 				const struct timeval *, bool_t);
383 
384 static bool_t	connmgr_setopt(queue_t *, int, int, calllist_t *);
385 static void	connmgr_sndrel(struct cm_xprt *);
386 static void	connmgr_snddis(struct cm_xprt *);
387 static void	connmgr_close(struct cm_xprt *);
388 static void	connmgr_release(struct cm_xprt *);
389 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *,
390 	cku_private_t *);
391 
392 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *,
393 	struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t,
394 	bool_t, int);
395 
396 static void connmgr_cancelconn(struct cm_xprt *);
397 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *,
398 	bool_t);
399 static void connmgr_dis_and_wait(struct cm_xprt *);
400 
401 static int	clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t,
402 					uint_t);
403 
404 static int clnt_delay(clock_t, bool_t);
405 
406 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t);
407 
408 /*
409  * Operations vector for TCP/IP based RPC
410  */
411 static struct clnt_ops tcp_ops = {
412 	clnt_cots_kcallit,	/* do rpc call */
413 	clnt_cots_kabort,	/* abort call */
414 	clnt_cots_kerror,	/* return error status */
415 	clnt_cots_kfreeres,	/* free results */
416 	clnt_cots_kdestroy,	/* destroy rpc handle */
417 	clnt_cots_kcontrol,	/* the ioctl() of rpc */
418 	clnt_cots_ksettimers,	/* set retry timers */
419 };
420 
421 static int rpc_kstat_instance = 0;  /* keeps the current instance */
422 				/* number for the next kstat_create */
423 
424 static struct cm_xprt *cm_hd = NULL;
425 static kmutex_t connmgr_lock;	/* for connection mngr's list of transports */
426 
427 extern kmutex_t clnt_max_msg_lock;
428 
429 static calllist_t *clnt_pending = NULL;
430 extern kmutex_t clnt_pending_lock;
431 
432 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE;
433 
434 static call_table_t *cots_call_ht;
435 
436 static const struct rpc_cots_client {
437 	kstat_named_t	rccalls;
438 	kstat_named_t	rcbadcalls;
439 	kstat_named_t	rcbadxids;
440 	kstat_named_t	rctimeouts;
441 	kstat_named_t	rcnewcreds;
442 	kstat_named_t	rcbadverfs;
443 	kstat_named_t	rctimers;
444 	kstat_named_t	rccantconn;
445 	kstat_named_t	rcnomem;
446 	kstat_named_t	rcintrs;
447 } cots_rcstat_tmpl = {
448 	{ "calls",	KSTAT_DATA_UINT64 },
449 	{ "badcalls",	KSTAT_DATA_UINT64 },
450 	{ "badxids",	KSTAT_DATA_UINT64 },
451 	{ "timeouts",	KSTAT_DATA_UINT64 },
452 	{ "newcreds",	KSTAT_DATA_UINT64 },
453 	{ "badverfs",	KSTAT_DATA_UINT64 },
454 	{ "timers",	KSTAT_DATA_UINT64 },
455 	{ "cantconn",	KSTAT_DATA_UINT64 },
456 	{ "nomem",	KSTAT_DATA_UINT64 },
457 	{ "interrupts", KSTAT_DATA_UINT64 }
458 };
459 
460 #define	COTSRCSTAT_INCR(p, x)	\
461 	atomic_add_64(&(p)->x.value.ui64, 1)
462 
463 #define	CLNT_MAX_CONNS	1	/* concurrent connections between clnt/srvr */
464 static int clnt_max_conns = CLNT_MAX_CONNS;
465 
466 #define	CLNT_MIN_TIMEOUT	10	/* seconds to wait after we get a */
467 					/* connection reset */
468 #define	CLNT_MIN_CONNTIMEOUT	5	/* seconds to wait for a connection */
469 
470 
471 static int clnt_cots_min_tout = CLNT_MIN_TIMEOUT;
472 static int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT;
473 
474 /*
475  * Limit the number of times we will attempt to receive a reply without
476  * re-sending a response.
477  */
478 #define	CLNT_MAXRECV_WITHOUT_RETRY	3
479 static uint_t clnt_cots_maxrecv	= CLNT_MAXRECV_WITHOUT_RETRY;
480 
481 uint_t *clnt_max_msg_sizep;
482 void (*clnt_stop_idle)(queue_t *wq);
483 
484 #define	ptoh(p)		(&((p)->cku_client))
485 #define	htop(h)		((cku_private_t *)((h)->cl_private))
486 
487 /*
488  * Times to retry
489  */
490 #define	REFRESHES	2	/* authentication refreshes */
491 
492 /*
493  * The following is used to determine the global default behavior for
494  * COTS when binding to a local port.
495  *
496  * If the value is set to 1 the default will be to select a reserved
497  * (aka privileged) port, if the value is zero the default will be to
498  * use non-reserved ports.  Users of kRPC may override this by using
499  * CLNT_CONTROL() and CLSET_BINDRESVPORT.
500  */
501 static int clnt_cots_do_bindresvport = 1;
502 
503 static zone_key_t zone_cots_key;
504 
505 /*
506  * We need to do this after all kernel threads in the zone have exited.
507  */
508 /* ARGSUSED */
509 static void
510 clnt_zone_destroy(zoneid_t zoneid, void *unused)
511 {
512 	struct cm_xprt **cmp;
513 	struct cm_xprt *cm_entry;
514 	struct cm_xprt *freelist = NULL;
515 
516 	mutex_enter(&connmgr_lock);
517 	cmp = &cm_hd;
518 	while ((cm_entry = *cmp) != NULL) {
519 		if (cm_entry->x_zoneid == zoneid) {
520 			*cmp = cm_entry->x_next;
521 			cm_entry->x_next = freelist;
522 			freelist = cm_entry;
523 		} else {
524 			cmp = &cm_entry->x_next;
525 		}
526 	}
527 	mutex_exit(&connmgr_lock);
528 	while ((cm_entry = freelist) != NULL) {
529 		freelist = cm_entry->x_next;
530 		connmgr_close(cm_entry);
531 	}
532 }
533 
534 int
535 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog,
536 	rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl)
537 {
538 	CLIENT *h;
539 	cku_private_t *p;
540 	struct rpc_msg call_msg;
541 	struct rpcstat *rpcstat;
542 
543 	RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog);
544 
545 	rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
546 	ASSERT(rpcstat != NULL);
547 
548 	/* Allocate and intialize the client handle. */
549 	p = kmem_zalloc(sizeof (*p), KM_SLEEP);
550 
551 	h = ptoh(p);
552 
553 	h->cl_private = (caddr_t)p;
554 	h->cl_auth = authkern_create();
555 	h->cl_ops = &tcp_ops;
556 
557 	cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL);
558 	mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL);
559 
560 	/*
561 	 * If the current sanity check size in rpcmod is smaller
562 	 * than the size needed, then increase the sanity check.
563 	 */
564 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
565 	    max_msgsize > *clnt_max_msg_sizep) {
566 		mutex_enter(&clnt_max_msg_lock);
567 		if (max_msgsize > *clnt_max_msg_sizep)
568 			*clnt_max_msg_sizep = max_msgsize;
569 		mutex_exit(&clnt_max_msg_lock);
570 	}
571 
572 	p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE;
573 
574 	/* Preserialize the call message header */
575 
576 	call_msg.rm_xid = 0;
577 	call_msg.rm_direction = CALL;
578 	call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION;
579 	call_msg.rm_call.cb_prog = prog;
580 	call_msg.rm_call.cb_vers = vers;
581 
582 	xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE);
583 
584 	if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) {
585 		RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization "
586 		    "error\n");
587 		auth_destroy(h->cl_auth);
588 		kmem_free(p, sizeof (cku_private_t));
589 		RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n");
590 		return (EINVAL);		/* XXX */
591 	}
592 
593 	/*
594 	 * The zalloc initialized the fields below.
595 	 * p->cku_xid = 0;
596 	 * p->cku_flags = 0;
597 	 * p->cku_srcaddr.len = 0;
598 	 * p->cku_srcaddr.maxlen = 0;
599 	 */
600 
601 	p->cku_cred = cred;
602 	p->cku_device = dev;
603 	p->cku_addrfmly = family;
604 	p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
605 	p->cku_addr.maxlen = addr->maxlen;
606 	p->cku_addr.len = addr->len;
607 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
608 	p->cku_stats = rpcstat->rpc_cots_client;
609 	p->cku_useresvport = -1; /* value is has not been set */
610 
611 	*ncl = h;
612 	return (0);
613 }
614 
615 /*ARGSUSED*/
616 static void
617 clnt_cots_kabort(CLIENT *h)
618 {
619 }
620 
621 /*
622  * Return error info on this handle.
623  */
624 static void
625 clnt_cots_kerror(CLIENT *h, struct rpc_err *err)
626 {
627 	/* LINTED pointer alignment */
628 	cku_private_t *p = htop(h);
629 
630 	*err = p->cku_err;
631 }
632 
633 static bool_t
634 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr)
635 {
636 	/* LINTED pointer alignment */
637 	cku_private_t *p = htop(h);
638 	XDR *xdrs;
639 
640 	xdrs = &(p->cku_outxdr);
641 	xdrs->x_op = XDR_FREE;
642 	return ((*xdr_res)(xdrs, res_ptr));
643 }
644 
645 static bool_t
646 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg)
647 {
648 	cku_private_t *p = htop(h);
649 
650 	switch (cmd) {
651 	case CLSET_PROGRESS:
652 		p->cku_progress = TRUE;
653 		return (TRUE);
654 
655 	case CLSET_XID:
656 		if (arg == NULL)
657 			return (FALSE);
658 
659 		p->cku_xid = *((uint32_t *)arg);
660 		return (TRUE);
661 
662 	case CLGET_XID:
663 		if (arg == NULL)
664 			return (FALSE);
665 
666 		*((uint32_t *)arg) = p->cku_xid;
667 		return (TRUE);
668 
669 	case CLSET_NODELAYONERR:
670 		if (arg == NULL)
671 			return (FALSE);
672 
673 		if (*((bool_t *)arg) == TRUE) {
674 			p->cku_nodelayonerr = TRUE;
675 			return (TRUE);
676 		}
677 		if (*((bool_t *)arg) == FALSE) {
678 			p->cku_nodelayonerr = FALSE;
679 			return (TRUE);
680 		}
681 		return (FALSE);
682 
683 	case CLGET_NODELAYONERR:
684 		if (arg == NULL)
685 			return (FALSE);
686 
687 		*((bool_t *)arg) = p->cku_nodelayonerr;
688 		return (TRUE);
689 
690 	case CLSET_BINDRESVPORT:
691 		if (arg == NULL)
692 			return (FALSE);
693 
694 		if (*(int *)arg != 1 && *(int *)arg != 0)
695 			return (FALSE);
696 
697 		p->cku_useresvport = *(int *)arg;
698 
699 		return (TRUE);
700 
701 	case CLGET_BINDRESVPORT:
702 		if (arg == NULL)
703 			return (FALSE);
704 
705 		*(int *)arg = p->cku_useresvport;
706 
707 		return (TRUE);
708 
709 	default:
710 		return (FALSE);
711 	}
712 }
713 
714 /*
715  * Destroy rpc handle.  Frees the space used for output buffer,
716  * private data, and handle structure.
717  */
718 static void
719 clnt_cots_kdestroy(CLIENT *h)
720 {
721 	/* LINTED pointer alignment */
722 	cku_private_t *p = htop(h);
723 	calllist_t *call = &p->cku_call;
724 
725 	RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h);
726 	RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid);
727 
728 	if (p->cku_flags & CKU_ONQUEUE) {
729 		RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x "
730 		    "from dispatch list\n", p->cku_xid);
731 		call_table_remove(call);
732 	}
733 
734 	if (call->call_reply)
735 		freemsg(call->call_reply);
736 	cv_destroy(&call->call_cv);
737 	mutex_destroy(&call->call_lock);
738 
739 	kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen);
740 	kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
741 	kmem_free(p, sizeof (*p));
742 }
743 
744 static int clnt_cots_pulls;
745 #define	RM_HDR_SIZE	4	/* record mark header size */
746 
747 /*
748  * Call remote procedure.
749  */
750 static enum clnt_stat
751 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
752     caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait)
753 {
754 	/* LINTED pointer alignment */
755 	cku_private_t *p = htop(h);
756 	calllist_t *call = &p->cku_call;
757 	XDR *xdrs;
758 	struct rpc_msg reply_msg;
759 	mblk_t *mp;
760 #ifdef	RPCDEBUG
761 	clock_t time_sent;
762 #endif
763 	struct netbuf *retryaddr;
764 	struct cm_xprt *cm_entry = NULL;
765 	queue_t *wq;
766 	int len;
767 	int mpsize;
768 	int refreshes = REFRESHES;
769 	int interrupted;
770 	int tidu_size;
771 	enum clnt_stat status;
772 	struct timeval cwait;
773 	bool_t delay_first = FALSE;
774 	clock_t ticks;
775 
776 	RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum);
777 	COTSRCSTAT_INCR(p->cku_stats, rccalls);
778 
779 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec);
780 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec);
781 
782 	/*
783 	 * Bug ID 1240234:
784 	 * Look out for zero length timeouts. We don't want to
785 	 * wait zero seconds for a connection to be established.
786 	 */
787 	if (wait.tv_sec < clnt_cots_min_conntout) {
788 		cwait.tv_sec = clnt_cots_min_conntout;
789 		cwait.tv_usec = 0;
790 		RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,",
791 		    wait.tv_sec);
792 		RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout);
793 	} else {
794 		cwait = wait;
795 	}
796 
797 call_again:
798 	if (cm_entry) {
799 		connmgr_release(cm_entry);
800 		cm_entry = NULL;
801 	}
802 
803 	mp = NULL;
804 
805 	/*
806 	 * If the call is not a retry, allocate a new xid and cache it
807 	 * for future retries.
808 	 * Bug ID 1246045:
809 	 * Treat call as a retry for purposes of binding the source
810 	 * port only if we actually attempted to send anything on
811 	 * the previous call.
812 	 */
813 	if (p->cku_xid == 0) {
814 		p->cku_xid = alloc_xid();
815 		call->call_zoneid = rpc_zoneid();
816 
817 		/*
818 		 * We need to ASSERT here that our xid != 0 because this
819 		 * determines whether or not our call record gets placed on
820 		 * the hash table or the linked list.  By design, we mandate
821 		 * that RPC calls over cots must have xid's != 0, so we can
822 		 * ensure proper management of the hash table.
823 		 */
824 		ASSERT(p->cku_xid != 0);
825 
826 		retryaddr = NULL;
827 		p->cku_flags &= ~CKU_SENT;
828 
829 		if (p->cku_flags & CKU_ONQUEUE) {
830 			RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old"
831 			    " one (%p)\n", (void *)call);
832 			call_table_remove(call);
833 			p->cku_flags &= ~CKU_ONQUEUE;
834 			RPCLOG(64, "clnt_cots_kcallit: removing call from "
835 			    "dispatch list because xid was zero (now 0x%x)\n",
836 			    p->cku_xid);
837 		}
838 
839 		if (call->call_reply != NULL) {
840 			freemsg(call->call_reply);
841 			call->call_reply = NULL;
842 		}
843 	} else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) {
844 		retryaddr = NULL;
845 
846 	} else if (p->cku_flags & CKU_SENT) {
847 		retryaddr = &p->cku_srcaddr;
848 
849 	} else {
850 		/*
851 		 * Bug ID 1246045: Nothing was sent, so set retryaddr to
852 		 * NULL and let connmgr_get() bind to any source port it
853 		 * can get.
854 		 */
855 		retryaddr = NULL;
856 	}
857 
858 	RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid);
859 	RPCLOG(64, " flags = 0x%x\n", p->cku_flags);
860 
861 	p->cku_err.re_status = RPC_TIMEDOUT;
862 	p->cku_err.re_errno = p->cku_err.re_terrno = 0;
863 
864 	cm_entry = connmgr_wrapget(retryaddr, &cwait, p);
865 
866 	if (cm_entry == NULL) {
867 		RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n",
868 		    clnt_sperrno(p->cku_err.re_status));
869 
870 		/*
871 		 * The reasons why we fail to create a connection are
872 		 * varied. In most cases we don't want the caller to
873 		 * immediately retry. This could have one or more
874 		 * bad effects. This includes flooding the net with
875 		 * connect requests to ports with no listener; a hard
876 		 * kernel loop due to all the "reserved" TCP ports being
877 		 * in use.
878 		 */
879 		delay_first = TRUE;
880 
881 		/*
882 		 * Even if we end up returning EINTR, we still count a
883 		 * a "can't connect", because the connection manager
884 		 * might have been committed to waiting for or timing out on
885 		 * a connection.
886 		 */
887 		COTSRCSTAT_INCR(p->cku_stats, rccantconn);
888 		switch (p->cku_err.re_status) {
889 		case RPC_INTR:
890 			p->cku_err.re_errno = EINTR;
891 
892 			/*
893 			 * No need to delay because a UNIX signal(2)
894 			 * interrupted us. The caller likely won't
895 			 * retry the CLNT_CALL() and even if it does,
896 			 * we assume the caller knows what it is doing.
897 			 */
898 			delay_first = FALSE;
899 			break;
900 
901 		case RPC_TIMEDOUT:
902 			p->cku_err.re_errno = ETIMEDOUT;
903 
904 			/*
905 			 * No need to delay because timed out already
906 			 * on the connection request and assume that the
907 			 * transport time out is longer than our minimum
908 			 * timeout, or least not too much smaller.
909 			 */
910 			delay_first = FALSE;
911 			break;
912 
913 		case RPC_SYSTEMERROR:
914 		case RPC_TLIERROR:
915 			/*
916 			 * We want to delay here because a transient
917 			 * system error has a better chance of going away
918 			 * if we delay a bit. If it's not transient, then
919 			 * we don't want end up in a hard kernel loop
920 			 * due to retries.
921 			 */
922 			ASSERT(p->cku_err.re_errno != 0);
923 			break;
924 
925 
926 		case RPC_CANTCONNECT:
927 			/*
928 			 * RPC_CANTCONNECT is set on T_ERROR_ACK which
929 			 * implies some error down in the TCP layer or
930 			 * below. If cku_nodelayonerror is set then we
931 			 * assume the caller knows not to try too hard.
932 			 */
933 			RPCLOG0(8, "clnt_cots_kcallit: connection failed,");
934 			RPCLOG0(8, " re_status=RPC_CANTCONNECT,");
935 			RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno);
936 			RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr);
937 			if (p->cku_nodelayonerr == TRUE)
938 				delay_first = FALSE;
939 
940 			p->cku_err.re_errno = EIO;
941 
942 			break;
943 
944 		case RPC_XPRTFAILED:
945 			/*
946 			 * We want to delay here because we likely
947 			 * got a refused connection.
948 			 */
949 			if (p->cku_err.re_errno == 0)
950 				p->cku_err.re_errno = EIO;
951 
952 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
953 			    p->cku_err.re_errno);
954 
955 			break;
956 
957 		default:
958 			/*
959 			 * We delay here because it is better to err
960 			 * on the side of caution. If we got here then
961 			 * status could have been RPC_SUCCESS, but we
962 			 * know that we did not get a connection, so
963 			 * force the rpc status to RPC_CANTCONNECT.
964 			 */
965 			p->cku_err.re_status = RPC_CANTCONNECT;
966 			p->cku_err.re_errno = EIO;
967 			break;
968 		}
969 		if (delay_first == TRUE)
970 			ticks = clnt_cots_min_tout * drv_usectohz(1000000);
971 		goto cots_done;
972 	}
973 
974 	/*
975 	 * If we've never sent any request on this connection (send count
976 	 * is zero, or the connection has been reset), cache the
977 	 * the connection's create time and send a request (possibly a retry)
978 	 */
979 	if ((p->cku_flags & CKU_SENT) == 0 ||
980 	    p->cku_ctime != cm_entry->x_ctime) {
981 		p->cku_ctime = cm_entry->x_ctime;
982 
983 	} else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) &&
984 	    (call->call_reply != NULL ||
985 	    p->cku_recv_attempts < clnt_cots_maxrecv)) {
986 
987 		/*
988 		 * If we've sent a request and our call is on the dispatch
989 		 * queue and we haven't made too many receive attempts, then
990 		 * don't re-send, just receive.
991 		 */
992 		p->cku_recv_attempts++;
993 		goto read_again;
994 	}
995 
996 	/*
997 	 * Now we create the RPC request in a STREAMS message.  We have to do
998 	 * this after the call to connmgr_get so that we have the correct
999 	 * TIDU size for the transport.
1000 	 */
1001 	tidu_size = cm_entry->x_tidu_size;
1002 	len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE);
1003 
1004 	while ((mp = allocb(len, BPRI_MED)) == NULL) {
1005 		if (strwaitbuf(len, BPRI_MED)) {
1006 			p->cku_err.re_status = RPC_SYSTEMERROR;
1007 			p->cku_err.re_errno = ENOSR;
1008 			COTSRCSTAT_INCR(p->cku_stats, rcnomem);
1009 			goto cots_done;
1010 		}
1011 	}
1012 	xdrs = &p->cku_outxdr;
1013 	xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size);
1014 	mpsize = MBLKSIZE(mp);
1015 	ASSERT(mpsize >= len);
1016 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
1017 
1018 	/*
1019 	 * If the size of mblk is not appreciably larger than what we
1020 	 * asked, then resize the mblk to exactly len bytes. The reason for
1021 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
1022 	 * (from TCP over ethernet), and the arguments to the RPC require
1023 	 * 2800 bytes. Ideally we want the protocol to render two
1024 	 * ~1400 byte segments over the wire. However if allocb() gives us a 2k
1025 	 * mblk, and we allocate a second mblk for the remainder, the protocol
1026 	 * module may generate 3 segments over the wire:
1027 	 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and
1028 	 * 892 for the third. If we "waste" 448 bytes in the first mblk,
1029 	 * the XDR encoding will generate two ~1400 byte mblks, and the
1030 	 * protocol module is more likely to produce properly sized segments.
1031 	 */
1032 	if ((mpsize >> 1) <= len)
1033 		mp->b_rptr += (mpsize - len);
1034 
1035 	/*
1036 	 * Adjust b_rptr to reserve space for the non-data protocol headers
1037 	 * any downstream modules might like to add, and for the
1038 	 * record marking header.
1039 	 */
1040 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
1041 
1042 	if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) {
1043 		/* Copy in the preserialized RPC header information. */
1044 		bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE);
1045 
1046 		/* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */
1047 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base +
1048 		    WIRE_HDR_SIZE));
1049 
1050 		ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE);
1051 
1052 		/* Serialize the procedure number and the arguments. */
1053 		if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) ||
1054 		    (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) ||
1055 		    (!(*xdr_args)(xdrs, argsp))) {
1056 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1057 			p->cku_err.re_errno = EIO;
1058 			goto cots_done;
1059 		}
1060 
1061 		(*(uint32_t *)(mp->b_rptr)) = p->cku_xid;
1062 	} else {
1063 		uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE];
1064 		IXDR_PUT_U_INT32(uproc, procnum);
1065 
1066 		(*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid;
1067 
1068 		/* Use XDR_SETPOS() to set the b_wptr. */
1069 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
1070 
1071 		/* Serialize the procedure number and the arguments. */
1072 		if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4,
1073 		    xdrs, xdr_args, argsp)) {
1074 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1075 			p->cku_err.re_errno = EIO;
1076 			goto cots_done;
1077 		}
1078 	}
1079 
1080 	RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n",
1081 	    tidu_size);
1082 
1083 	wq = cm_entry->x_wq;
1084 	status = clnt_dispatch_send(wq, mp, call, p->cku_xid,
1085 	    (p->cku_flags & CKU_ONQUEUE));
1086 
1087 	if (status == RPC_CANTSEND) {
1088 		p->cku_err.re_status = status;
1089 		p->cku_err.re_errno = EIO;
1090 		DTRACE_PROBE(krpc__e__clntcots__kcallit__cantsend);
1091 
1092 		/*
1093 		 * Allow for processing of the QFULL queue.
1094 		 */
1095 		delay_first = TRUE;
1096 		ticks = clnt_cots_min_tout * drv_usectohz(1000000);
1097 
1098 		goto cots_done;
1099 	}
1100 
1101 	RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n",
1102 	    (uint_t)p->cku_xid);
1103 	p->cku_flags = (CKU_ONQUEUE|CKU_SENT);
1104 	p->cku_recv_attempts = 1;
1105 
1106 #ifdef	RPCDEBUG
1107 	time_sent = lbolt;
1108 #endif
1109 
1110 	/*
1111 	 * Wait for a reply or a timeout.  If there is no error or timeout,
1112 	 * (both indicated by call_status), call->call_reply will contain
1113 	 * the RPC reply message.
1114 	 */
1115 read_again:
1116 	mutex_enter(&call->call_lock);
1117 	interrupted = 0;
1118 	if (call->call_status == RPC_TIMEDOUT) {
1119 		/*
1120 		 * Indicate that the lwp is not to be stopped while waiting
1121 		 * for this network traffic.  This is to avoid deadlock while
1122 		 * debugging a process via /proc and also to avoid recursive
1123 		 * mutex_enter()s due to NFS page faults while stopping
1124 		 * (NFS holds locks when it calls here).
1125 		 */
1126 		clock_t cv_wait_ret;
1127 		clock_t timout;
1128 		clock_t oldlbolt;
1129 
1130 		klwp_t *lwp = ttolwp(curthread);
1131 
1132 		if (lwp != NULL)
1133 			lwp->lwp_nostop++;
1134 
1135 		oldlbolt = lbolt;
1136 		timout = wait.tv_sec * drv_usectohz(1000000) +
1137 		    drv_usectohz(wait.tv_usec) + oldlbolt;
1138 		/*
1139 		 * Iterate until the call_status is changed to something
1140 		 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns
1141 		 * something <=0 zero. The latter means that we timed
1142 		 * out.
1143 		 */
1144 		if (h->cl_nosignal)
1145 			while ((cv_wait_ret = cv_timedwait(&call->call_cv,
1146 			    &call->call_lock, timout)) > 0 &&
1147 			    call->call_status == RPC_TIMEDOUT)
1148 				;
1149 		else
1150 			while ((cv_wait_ret = cv_timedwait_sig(
1151 			    &call->call_cv,
1152 			    &call->call_lock, timout)) > 0 &&
1153 			    call->call_status == RPC_TIMEDOUT)
1154 				;
1155 
1156 		switch (cv_wait_ret) {
1157 		case 0:
1158 			/*
1159 			 * If we got out of the above loop with
1160 			 * cv_timedwait_sig() returning 0, then we were
1161 			 * interrupted regardless what call_status is.
1162 			 */
1163 			interrupted = 1;
1164 			break;
1165 		case -1:
1166 			/* cv_timedwait_sig() timed out */
1167 			break;
1168 		default:
1169 
1170 			/*
1171 			 * We were cv_signaled(). If we didn't
1172 			 * get a successful call_status and returned
1173 			 * before time expired, delay up to clnt_cots_min_tout
1174 			 * seconds so that the caller doesn't immediately
1175 			 * try to call us again and thus force the
1176 			 * same condition that got us here (such
1177 			 * as a RPC_XPRTFAILED due to the server not
1178 			 * listening on the end-point.
1179 			 */
1180 			if (call->call_status != RPC_SUCCESS) {
1181 				clock_t curlbolt;
1182 				clock_t diff;
1183 
1184 				curlbolt = ddi_get_lbolt();
1185 				ticks = clnt_cots_min_tout *
1186 				    drv_usectohz(1000000);
1187 				diff = curlbolt - oldlbolt;
1188 				if (diff < ticks) {
1189 					delay_first = TRUE;
1190 					if (diff > 0)
1191 						ticks -= diff;
1192 				}
1193 			}
1194 			break;
1195 		}
1196 
1197 		if (lwp != NULL)
1198 			lwp->lwp_nostop--;
1199 	}
1200 	/*
1201 	 * Get the reply message, if any.  This will be freed at the end
1202 	 * whether or not an error occurred.
1203 	 */
1204 	mp = call->call_reply;
1205 	call->call_reply = NULL;
1206 
1207 	/*
1208 	 * call_err is the error info when the call is on dispatch queue.
1209 	 * cku_err is the error info returned to the caller.
1210 	 * Sync cku_err with call_err for local message processing.
1211 	 */
1212 
1213 	status = call->call_status;
1214 	p->cku_err = call->call_err;
1215 	mutex_exit(&call->call_lock);
1216 
1217 	if (status != RPC_SUCCESS) {
1218 		switch (status) {
1219 		case RPC_TIMEDOUT:
1220 			if (interrupted) {
1221 				COTSRCSTAT_INCR(p->cku_stats, rcintrs);
1222 				p->cku_err.re_status = RPC_INTR;
1223 				p->cku_err.re_errno = EINTR;
1224 				RPCLOG(1, "clnt_cots_kcallit: xid 0x%x",
1225 				    p->cku_xid);
1226 				RPCLOG(1, "signal interrupted at %ld", lbolt);
1227 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1228 			} else {
1229 				COTSRCSTAT_INCR(p->cku_stats, rctimeouts);
1230 				p->cku_err.re_errno = ETIMEDOUT;
1231 				RPCLOG(1, "clnt_cots_kcallit: timed out at %ld",
1232 				    lbolt);
1233 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1234 			}
1235 			break;
1236 
1237 		case RPC_XPRTFAILED:
1238 			if (p->cku_err.re_errno == 0)
1239 				p->cku_err.re_errno = EIO;
1240 
1241 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
1242 			    p->cku_err.re_errno);
1243 			break;
1244 
1245 		case RPC_SYSTEMERROR:
1246 			ASSERT(p->cku_err.re_errno);
1247 			RPCLOG(1, "clnt_cots_kcallit: system error: %d\n",
1248 			    p->cku_err.re_errno);
1249 			break;
1250 
1251 		default:
1252 			p->cku_err.re_status = RPC_SYSTEMERROR;
1253 			p->cku_err.re_errno = EIO;
1254 			RPCLOG(1, "clnt_cots_kcallit: error: %s\n",
1255 			    clnt_sperrno(status));
1256 			break;
1257 		}
1258 		if (p->cku_err.re_status != RPC_TIMEDOUT) {
1259 
1260 			if (p->cku_flags & CKU_ONQUEUE) {
1261 				call_table_remove(call);
1262 				p->cku_flags &= ~CKU_ONQUEUE;
1263 			}
1264 
1265 			RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x "
1266 			    "taken off dispatch list\n", p->cku_xid);
1267 			if (call->call_reply) {
1268 				freemsg(call->call_reply);
1269 				call->call_reply = NULL;
1270 			}
1271 		} else if (wait.tv_sec != 0) {
1272 			/*
1273 			 * We've sent the request over TCP and so we have
1274 			 * every reason to believe it will get
1275 			 * delivered. In which case returning a timeout is not
1276 			 * appropriate.
1277 			 */
1278 			if (p->cku_progress == TRUE &&
1279 			    p->cku_recv_attempts < clnt_cots_maxrecv) {
1280 				p->cku_err.re_status = RPC_INPROGRESS;
1281 			}
1282 		}
1283 		goto cots_done;
1284 	}
1285 
1286 	xdrs = &p->cku_inxdr;
1287 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
1288 
1289 	reply_msg.rm_direction = REPLY;
1290 	reply_msg.rm_reply.rp_stat = MSG_ACCEPTED;
1291 	reply_msg.acpted_rply.ar_stat = SUCCESS;
1292 
1293 	reply_msg.acpted_rply.ar_verf = _null_auth;
1294 	/*
1295 	 *  xdr_results will be done in AUTH_UNWRAP.
1296 	 */
1297 	reply_msg.acpted_rply.ar_results.where = NULL;
1298 	reply_msg.acpted_rply.ar_results.proc = xdr_void;
1299 
1300 	if (xdr_replymsg(xdrs, &reply_msg)) {
1301 		enum clnt_stat re_status;
1302 
1303 		_seterr_reply(&reply_msg, &p->cku_err);
1304 
1305 		re_status = p->cku_err.re_status;
1306 		if (re_status == RPC_SUCCESS) {
1307 			/*
1308 			 * Reply is good, check auth.
1309 			 */
1310 			if (!AUTH_VALIDATE(h->cl_auth,
1311 			    &reply_msg.acpted_rply.ar_verf)) {
1312 				COTSRCSTAT_INCR(p->cku_stats, rcbadverfs);
1313 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1314 				    "failure\n");
1315 				freemsg(mp);
1316 				(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1317 				mutex_enter(&call->call_lock);
1318 				if (call->call_reply == NULL)
1319 					call->call_status = RPC_TIMEDOUT;
1320 				mutex_exit(&call->call_lock);
1321 				goto read_again;
1322 			} else if (!AUTH_UNWRAP(h->cl_auth, xdrs,
1323 			    xdr_results, resultsp)) {
1324 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1325 				    "failure (unwrap)\n");
1326 				p->cku_err.re_status = RPC_CANTDECODERES;
1327 				p->cku_err.re_errno = EIO;
1328 			}
1329 		} else {
1330 			/* set errno in case we can't recover */
1331 			if (re_status != RPC_VERSMISMATCH &&
1332 			    re_status != RPC_AUTHERROR &&
1333 			    re_status != RPC_PROGVERSMISMATCH)
1334 				p->cku_err.re_errno = EIO;
1335 
1336 			if (re_status == RPC_AUTHERROR) {
1337 				/*
1338 				 * Maybe our credential need to be refreshed
1339 				 */
1340 				if (cm_entry) {
1341 					/*
1342 					 * There is the potential that the
1343 					 * cm_entry has/will be marked dead,
1344 					 * so drop the connection altogether,
1345 					 * force REFRESH to establish new
1346 					 * connection.
1347 					 */
1348 					connmgr_cancelconn(cm_entry);
1349 					cm_entry = NULL;
1350 				}
1351 
1352 				if ((refreshes > 0) &&
1353 				    AUTH_REFRESH(h->cl_auth, &reply_msg,
1354 				    p->cku_cred)) {
1355 					refreshes--;
1356 					(void) xdr_rpc_free_verifier(xdrs,
1357 					    &reply_msg);
1358 					freemsg(mp);
1359 					mp = NULL;
1360 
1361 					if (p->cku_flags & CKU_ONQUEUE) {
1362 						call_table_remove(call);
1363 						p->cku_flags &= ~CKU_ONQUEUE;
1364 					}
1365 
1366 					RPCLOG(64,
1367 					    "clnt_cots_kcallit: AUTH_ERROR, xid"
1368 					    " 0x%x removed off dispatch list\n",
1369 					    p->cku_xid);
1370 					if (call->call_reply) {
1371 						freemsg(call->call_reply);
1372 						call->call_reply = NULL;
1373 					}
1374 
1375 					COTSRCSTAT_INCR(p->cku_stats,
1376 					    rcbadcalls);
1377 					COTSRCSTAT_INCR(p->cku_stats,
1378 					    rcnewcreds);
1379 					goto call_again;
1380 				}
1381 
1382 				/*
1383 				 * We have used the client handle to
1384 				 * do an AUTH_REFRESH and the RPC status may
1385 				 * be set to RPC_SUCCESS; Let's make sure to
1386 				 * set it to RPC_AUTHERROR.
1387 				 */
1388 				p->cku_err.re_status = RPC_AUTHERROR;
1389 
1390 				/*
1391 				 * Map recoverable and unrecoverable
1392 				 * authentication errors to appropriate errno
1393 				 */
1394 				switch (p->cku_err.re_why) {
1395 				case AUTH_TOOWEAK:
1396 					/*
1397 					 * This could be a failure where the
1398 					 * server requires use of a reserved
1399 					 * port,  check and optionally set the
1400 					 * client handle useresvport trying
1401 					 * one more time. Next go round we
1402 					 * fall out with the tooweak error.
1403 					 */
1404 					if (p->cku_useresvport != 1) {
1405 						p->cku_useresvport = 1;
1406 						p->cku_xid = 0;
1407 						(void) xdr_rpc_free_verifier
1408 						    (xdrs, &reply_msg);
1409 						freemsg(mp);
1410 						goto call_again;
1411 					}
1412 					/* FALLTHRU */
1413 				case AUTH_BADCRED:
1414 				case AUTH_BADVERF:
1415 				case AUTH_INVALIDRESP:
1416 				case AUTH_FAILED:
1417 				case RPCSEC_GSS_NOCRED:
1418 				case RPCSEC_GSS_FAILED:
1419 						p->cku_err.re_errno = EACCES;
1420 						break;
1421 				case AUTH_REJECTEDCRED:
1422 				case AUTH_REJECTEDVERF:
1423 				default:	p->cku_err.re_errno = EIO;
1424 						break;
1425 				}
1426 				RPCLOG(1, "clnt_cots_kcallit : authentication"
1427 				    " failed with RPC_AUTHERROR of type %d\n",
1428 				    (int)p->cku_err.re_why);
1429 			}
1430 		}
1431 	} else {
1432 		/* reply didn't decode properly. */
1433 		p->cku_err.re_status = RPC_CANTDECODERES;
1434 		p->cku_err.re_errno = EIO;
1435 		RPCLOG0(1, "clnt_cots_kcallit: decode failure\n");
1436 	}
1437 
1438 	(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1439 
1440 	if (p->cku_flags & CKU_ONQUEUE) {
1441 		call_table_remove(call);
1442 		p->cku_flags &= ~CKU_ONQUEUE;
1443 	}
1444 
1445 	RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list",
1446 	    p->cku_xid);
1447 	RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status));
1448 cots_done:
1449 	if (cm_entry)
1450 		connmgr_release(cm_entry);
1451 
1452 	if (mp != NULL)
1453 		freemsg(mp);
1454 	if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) {
1455 		freemsg(call->call_reply);
1456 		call->call_reply = NULL;
1457 	}
1458 	if (p->cku_err.re_status != RPC_SUCCESS) {
1459 		RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n");
1460 		COTSRCSTAT_INCR(p->cku_stats, rcbadcalls);
1461 	}
1462 
1463 	/*
1464 	 * No point in delaying if the zone is going away.
1465 	 */
1466 	if (delay_first == TRUE &&
1467 	    !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) {
1468 		if (clnt_delay(ticks, h->cl_nosignal) == EINTR) {
1469 			p->cku_err.re_errno = EINTR;
1470 			p->cku_err.re_status = RPC_INTR;
1471 		}
1472 	}
1473 	return (p->cku_err.re_status);
1474 }
1475 
1476 /*
1477  * Kinit routine for cots.  This sets up the correct operations in
1478  * the client handle, as the handle may have previously been a clts
1479  * handle, and clears the xid field so there is no way a new call
1480  * could be mistaken for a retry.  It also sets in the handle the
1481  * information that is passed at create/kinit time but needed at
1482  * call time, as cots creates the transport at call time - device,
1483  * address of the server, protocol family.
1484  */
1485 void
1486 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr,
1487 	int max_msgsize, cred_t *cred)
1488 {
1489 	/* LINTED pointer alignment */
1490 	cku_private_t *p = htop(h);
1491 	calllist_t *call = &p->cku_call;
1492 
1493 	h->cl_ops = &tcp_ops;
1494 	if (p->cku_flags & CKU_ONQUEUE) {
1495 		call_table_remove(call);
1496 		p->cku_flags &= ~CKU_ONQUEUE;
1497 		RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from"
1498 		    " dispatch list\n", p->cku_xid);
1499 	}
1500 
1501 	if (call->call_reply != NULL) {
1502 		freemsg(call->call_reply);
1503 		call->call_reply = NULL;
1504 	}
1505 
1506 	call->call_bucket = NULL;
1507 	call->call_hash = 0;
1508 
1509 	/*
1510 	 * We don't clear cku_flags here, because clnt_cots_kcallit()
1511 	 * takes care of handling the cku_flags reset.
1512 	 */
1513 	p->cku_xid = 0;
1514 	p->cku_device = dev;
1515 	p->cku_addrfmly = family;
1516 	p->cku_cred = cred;
1517 
1518 	if (p->cku_addr.maxlen < addr->len) {
1519 		if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL)
1520 			kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
1521 		p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
1522 		p->cku_addr.maxlen = addr->maxlen;
1523 	}
1524 
1525 	p->cku_addr.len = addr->len;
1526 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
1527 
1528 	/*
1529 	 * If the current sanity check size in rpcmod is smaller
1530 	 * than the size needed, then increase the sanity check.
1531 	 */
1532 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
1533 	    max_msgsize > *clnt_max_msg_sizep) {
1534 		mutex_enter(&clnt_max_msg_lock);
1535 		if (max_msgsize > *clnt_max_msg_sizep)
1536 			*clnt_max_msg_sizep = max_msgsize;
1537 		mutex_exit(&clnt_max_msg_lock);
1538 	}
1539 }
1540 
1541 /*
1542  * ksettimers is a no-op for cots, with the exception of setting the xid.
1543  */
1544 /* ARGSUSED */
1545 static int
1546 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all,
1547 	int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg,
1548 	uint32_t xid)
1549 {
1550 	/* LINTED pointer alignment */
1551 	cku_private_t *p = htop(h);
1552 
1553 	if (xid)
1554 		p->cku_xid = xid;
1555 	COTSRCSTAT_INCR(p->cku_stats, rctimers);
1556 	return (0);
1557 }
1558 
1559 extern void rpc_poptimod(struct vnode *);
1560 extern int kstr_push(struct vnode *, char *);
1561 
1562 int
1563 conn_kstat_update(kstat_t *ksp, int rw)
1564 {
1565 	struct cm_xprt *cm_entry;
1566 	struct cm_kstat_xprt *cm_ksp_data;
1567 	uchar_t *b;
1568 	char *fbuf;
1569 
1570 	if (rw == KSTAT_WRITE)
1571 		return (EACCES);
1572 	if (ksp == NULL || ksp->ks_private == NULL)
1573 		return (EIO);
1574 	cm_entry  = (struct cm_xprt *)ksp->ks_private;
1575 	cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data;
1576 
1577 	cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq;
1578 	cm_ksp_data->x_family.value.ui32 = cm_entry->x_family;
1579 	cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev;
1580 	cm_ksp_data->x_time.value.ui32 = cm_entry->x_time;
1581 	cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref;
1582 	cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags;
1583 
1584 	if (cm_entry->x_server.buf) {
1585 		fbuf = cm_ksp_data->x_server.value.str.addr.ptr;
1586 		if (cm_entry->x_family == AF_INET &&
1587 		    cm_entry->x_server.len ==
1588 		    sizeof (struct sockaddr_in)) {
1589 			struct sockaddr_in  *sa;
1590 			sa = (struct sockaddr_in *)
1591 				cm_entry->x_server.buf;
1592 			b = (uchar_t *)&sa->sin_addr;
1593 			(void) sprintf(fbuf,
1594 			    "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF,
1595 			    b[2] & 0xFF, b[3] & 0xFF);
1596 			cm_ksp_data->x_port.value.ui32 =
1597 				(uint32_t)sa->sin_port;
1598 		} else if (cm_entry->x_family == AF_INET6 &&
1599 				cm_entry->x_server.len >=
1600 				sizeof (struct sockaddr_in6)) {
1601 			/* extract server IP address & port */
1602 			struct sockaddr_in6 *sin6;
1603 			sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf;
1604 			(void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf,
1605 				INET6_ADDRSTRLEN);
1606 			cm_ksp_data->x_port.value.ui32 = sin6->sin6_port;
1607 		} else {
1608 			struct sockaddr_in  *sa;
1609 
1610 			sa = (struct sockaddr_in *)cm_entry->x_server.buf;
1611 			b = (uchar_t *)&sa->sin_addr;
1612 			(void) sprintf(fbuf,
1613 			    "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF,
1614 			    b[2] & 0xFF, b[3] & 0xFF);
1615 		}
1616 		KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) =
1617 		    strlen(fbuf) + 1;
1618 	}
1619 
1620 	return (0);
1621 }
1622 
1623 
1624 /*
1625  * We want a version of delay which is interruptible by a UNIX signal
1626  * Return EINTR if an interrupt occured.
1627  */
1628 static int
1629 clnt_delay(clock_t ticks, bool_t nosignal)
1630 {
1631 	if (nosignal == TRUE) {
1632 		delay(ticks);
1633 		return (0);
1634 	}
1635 	return (delay_sig(ticks));
1636 }
1637 
1638 /*
1639  * Wait for a connection until a timeout, or until we are
1640  * signalled that there has been a connection state change.
1641  */
1642 static enum clnt_stat
1643 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp,
1644 	bool_t nosignal)
1645 {
1646 	bool_t interrupted;
1647 	clock_t timout, cv_stat;
1648 	enum clnt_stat clstat;
1649 	unsigned int old_state;
1650 
1651 	ASSERT(MUTEX_HELD(&connmgr_lock));
1652 	/*
1653 	 * We wait for the transport connection to be made, or an
1654 	 * indication that it could not be made.
1655 	 */
1656 	clstat = RPC_TIMEDOUT;
1657 	interrupted = FALSE;
1658 
1659 	old_state = cm_entry->x_state_flags;
1660 	/*
1661 	 * Now loop until cv_timedwait{_sig} returns because of
1662 	 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be
1663 	 * cv_signalled for various other reasons too. So loop
1664 	 * until there is a state change on the connection.
1665 	 */
1666 
1667 	timout = waitp->tv_sec * drv_usectohz(1000000) +
1668 	    drv_usectohz(waitp->tv_usec) + lbolt;
1669 
1670 	if (nosignal) {
1671 		while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv,
1672 		    &connmgr_lock, timout)) > 0 &&
1673 		    cm_entry->x_state_flags == old_state)
1674 			;
1675 	} else {
1676 		while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv,
1677 		    &connmgr_lock, timout)) > 0 &&
1678 		    cm_entry->x_state_flags == old_state)
1679 			;
1680 
1681 		if (cv_stat == 0) /* got intr signal? */
1682 			interrupted = TRUE;
1683 	}
1684 
1685 	if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) ==
1686 	    X_CONNECTED) {
1687 		clstat = RPC_SUCCESS;
1688 	} else {
1689 		if (interrupted == TRUE)
1690 			clstat = RPC_INTR;
1691 		RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n",
1692 		    clnt_sperrno(clstat));
1693 	}
1694 
1695 	return (clstat);
1696 }
1697 
1698 /*
1699  * Primary interface for how RPC grabs a connection.
1700  */
1701 static struct cm_xprt *
1702 connmgr_wrapget(
1703 	struct netbuf *retryaddr,
1704 	const struct timeval *waitp,
1705 	cku_private_t *p)
1706 {
1707 	struct cm_xprt *cm_entry;
1708 
1709 	cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly,
1710 	    &p->cku_srcaddr, &p->cku_err, p->cku_device,
1711 	    p->cku_client.cl_nosignal, p->cku_useresvport);
1712 
1713 	if (cm_entry == NULL) {
1714 		/*
1715 		 * Re-map the call status to RPC_INTR if the err code is
1716 		 * EINTR. This can happen if calls status is RPC_TLIERROR.
1717 		 * However, don't re-map if signalling has been turned off.
1718 		 * XXX Really need to create a separate thread whenever
1719 		 * there isn't an existing connection.
1720 		 */
1721 		if (p->cku_err.re_errno == EINTR) {
1722 			if (p->cku_client.cl_nosignal == TRUE)
1723 				p->cku_err.re_errno = EIO;
1724 			else
1725 				p->cku_err.re_status = RPC_INTR;
1726 		}
1727 	}
1728 
1729 	return (cm_entry);
1730 }
1731 
1732 /*
1733  * Obtains a transport to the server specified in addr.  If a suitable transport
1734  * does not already exist in the list of cached transports, a new connection
1735  * is created, connected, and added to the list. The connection is for sending
1736  * only - the reply message may come back on another transport connection.
1737  */
1738 static struct cm_xprt *
1739 connmgr_get(
1740 	struct netbuf	*retryaddr,
1741 	const struct timeval	*waitp,	/* changed to a ptr to converse stack */
1742 	struct netbuf	*destaddr,
1743 	int		addrfmly,
1744 	struct netbuf	*srcaddr,
1745 	struct rpc_err	*rpcerr,
1746 	dev_t		device,
1747 	bool_t		nosignal,
1748 	int		useresvport)
1749 {
1750 	struct cm_xprt *cm_entry;
1751 	struct cm_xprt *lru_entry;
1752 	struct cm_xprt **cmp;
1753 	queue_t *wq;
1754 	TIUSER *tiptr;
1755 	int i;
1756 	int retval;
1757 	clock_t prev_time;
1758 	int tidu_size;
1759 	bool_t	connected;
1760 	zoneid_t zoneid = rpc_zoneid();
1761 
1762 	/*
1763 	 * If the call is not a retry, look for a transport entry that
1764 	 * goes to the server of interest.
1765 	 */
1766 	mutex_enter(&connmgr_lock);
1767 
1768 	if (retryaddr == NULL) {
1769 use_new_conn:
1770 		i = 0;
1771 		cm_entry = lru_entry = NULL;
1772 		prev_time = lbolt;
1773 
1774 		cmp = &cm_hd;
1775 		while ((cm_entry = *cmp) != NULL) {
1776 			ASSERT(cm_entry != cm_entry->x_next);
1777 			/*
1778 			 * Garbage collect conections that are marked
1779 			 * for needs disconnect.
1780 			 */
1781 			if (cm_entry->x_needdis) {
1782 				CONN_HOLD(cm_entry);
1783 				connmgr_dis_and_wait(cm_entry);
1784 				connmgr_release(cm_entry);
1785 				/*
1786 				 * connmgr_lock could have been
1787 				 * dropped for the disconnect
1788 				 * processing so start over.
1789 				 */
1790 				goto use_new_conn;
1791 			}
1792 
1793 			/*
1794 			 * Garbage collect the dead connections that have
1795 			 * no threads working on them.
1796 			 */
1797 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1798 			    X_DEAD) {
1799 				mutex_enter(&cm_entry->x_lock);
1800 				if (cm_entry->x_ref != 0) {
1801 					/*
1802 					 * Currently in use.
1803 					 * Cleanup later.
1804 					 */
1805 					cmp = &cm_entry->x_next;
1806 					mutex_exit(&cm_entry->x_lock);
1807 					continue;
1808 				}
1809 				mutex_exit(&cm_entry->x_lock);
1810 				*cmp = cm_entry->x_next;
1811 				mutex_exit(&connmgr_lock);
1812 				connmgr_close(cm_entry);
1813 				mutex_enter(&connmgr_lock);
1814 				goto use_new_conn;
1815 			}
1816 
1817 
1818 			if ((cm_entry->x_state_flags & X_BADSTATES) == 0 &&
1819 			    cm_entry->x_zoneid == zoneid &&
1820 			    cm_entry->x_rdev == device &&
1821 			    destaddr->len == cm_entry->x_server.len &&
1822 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
1823 			    destaddr->len) == 0) {
1824 				/*
1825 				 * If the matching entry isn't connected,
1826 				 * attempt to reconnect it.
1827 				 */
1828 				if (cm_entry->x_connected == FALSE) {
1829 					/*
1830 					 * We don't go through trying
1831 					 * to find the least recently
1832 					 * used connected because
1833 					 * connmgr_reconnect() briefly
1834 					 * dropped the connmgr_lock,
1835 					 * allowing a window for our
1836 					 * accounting to be messed up.
1837 					 * In any case, a re-connected
1838 					 * connection is as good as
1839 					 * a LRU connection.
1840 					 */
1841 					return (connmgr_wrapconnect(cm_entry,
1842 					    waitp, destaddr, addrfmly, srcaddr,
1843 					    rpcerr, TRUE, nosignal));
1844 				}
1845 				i++;
1846 				if (cm_entry->x_time - prev_time <= 0 ||
1847 				    lru_entry == NULL) {
1848 					prev_time = cm_entry->x_time;
1849 					lru_entry = cm_entry;
1850 				}
1851 			}
1852 			cmp = &cm_entry->x_next;
1853 		}
1854 
1855 		if (i > clnt_max_conns) {
1856 			RPCLOG(8, "connmgr_get: too many conns, dooming entry"
1857 			    " %p\n", (void *)lru_entry->x_tiptr);
1858 			lru_entry->x_doomed = TRUE;
1859 			goto use_new_conn;
1860 		}
1861 
1862 		/*
1863 		 * If we are at the maximum number of connections to
1864 		 * the server, hand back the least recently used one.
1865 		 */
1866 		if (i == clnt_max_conns) {
1867 			/*
1868 			 * Copy into the handle the source address of
1869 			 * the connection, which we will use in case of
1870 			 * a later retry.
1871 			 */
1872 			if (srcaddr->len != lru_entry->x_src.len) {
1873 				if (srcaddr->len > 0)
1874 					kmem_free(srcaddr->buf,
1875 					    srcaddr->maxlen);
1876 				srcaddr->buf = kmem_zalloc(
1877 				    lru_entry->x_src.len, KM_SLEEP);
1878 				srcaddr->maxlen = srcaddr->len =
1879 				    lru_entry->x_src.len;
1880 			}
1881 			bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len);
1882 			RPCLOG(2, "connmgr_get: call going out on %p\n",
1883 			    (void *)lru_entry);
1884 			lru_entry->x_time = lbolt;
1885 			CONN_HOLD(lru_entry);
1886 			mutex_exit(&connmgr_lock);
1887 			return (lru_entry);
1888 		}
1889 
1890 	} else {
1891 		/*
1892 		 * This is the retry case (retryaddr != NULL).  Retries must
1893 		 * be sent on the same source port as the original call.
1894 		 */
1895 
1896 		/*
1897 		 * Walk the list looking for a connection with a source address
1898 		 * that matches the retry address.
1899 		 */
1900 		cmp = &cm_hd;
1901 		while ((cm_entry = *cmp) != NULL) {
1902 			ASSERT(cm_entry != cm_entry->x_next);
1903 			if (zoneid != cm_entry->x_zoneid ||
1904 			    device != cm_entry->x_rdev ||
1905 			    retryaddr->len != cm_entry->x_src.len ||
1906 			    bcmp(retryaddr->buf, cm_entry->x_src.buf,
1907 			    retryaddr->len) != 0) {
1908 				cmp = &cm_entry->x_next;
1909 				continue;
1910 			}
1911 
1912 			/*
1913 			 * Sanity check: if the connection with our source
1914 			 * port is going to some other server, something went
1915 			 * wrong, as we never delete connections (i.e. release
1916 			 * ports) unless they have been idle.  In this case,
1917 			 * it is probably better to send the call out using
1918 			 * a new source address than to fail it altogether,
1919 			 * since that port may never be released.
1920 			 */
1921 			if (destaddr->len != cm_entry->x_server.len ||
1922 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
1923 			    destaddr->len) != 0) {
1924 				RPCLOG(1, "connmgr_get: tiptr %p"
1925 				    " is going to a different server"
1926 				    " with the port that belongs"
1927 				    " to us!\n", (void *)cm_entry->x_tiptr);
1928 				retryaddr = NULL;
1929 				goto use_new_conn;
1930 			}
1931 
1932 			/*
1933 			 * If the connection of interest is not connected and we
1934 			 * can't reconnect it, then the server is probably
1935 			 * still down.  Return NULL to the caller and let it
1936 			 * retry later if it wants to.  We have a delay so the
1937 			 * machine doesn't go into a tight retry loop.  If the
1938 			 * entry was already connected, or the reconnected was
1939 			 * successful, return this entry.
1940 			 */
1941 			if (cm_entry->x_connected == FALSE) {
1942 				return (connmgr_wrapconnect(cm_entry,
1943 				    waitp, destaddr, addrfmly, NULL,
1944 				    rpcerr, TRUE, nosignal));
1945 			} else {
1946 				CONN_HOLD(cm_entry);
1947 
1948 				cm_entry->x_time = lbolt;
1949 				mutex_exit(&connmgr_lock);
1950 				RPCLOG(2, "connmgr_get: found old "
1951 				    "transport %p for retry\n",
1952 				    (void *)cm_entry);
1953 				return (cm_entry);
1954 			}
1955 		}
1956 
1957 		/*
1958 		 * We cannot find an entry in the list for this retry.
1959 		 * Either the entry has been removed temporarily to be
1960 		 * reconnected by another thread, or the original call
1961 		 * got a port but never got connected,
1962 		 * and hence the transport never got put in the
1963 		 * list.  Fall through to the "create new connection" code -
1964 		 * the former case will fail there trying to rebind the port,
1965 		 * and the later case (and any other pathological cases) will
1966 		 * rebind and reconnect and not hang the client machine.
1967 		 */
1968 		RPCLOG0(8, "connmgr_get: no entry in list for retry\n");
1969 	}
1970 	/*
1971 	 * Set up a transport entry in the connection manager's list.
1972 	 */
1973 	cm_entry = (struct cm_xprt *)
1974 	    kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP);
1975 
1976 	cm_entry->x_server.buf = kmem_zalloc(destaddr->len, KM_SLEEP);
1977 	bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len);
1978 	cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len;
1979 
1980 	cm_entry->x_state_flags = X_THREAD;
1981 	cm_entry->x_ref = 1;
1982 	cm_entry->x_family = addrfmly;
1983 	cm_entry->x_rdev = device;
1984 	cm_entry->x_zoneid = zoneid;
1985 	mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL);
1986 	cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL);
1987 	cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL);
1988 	cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL);
1989 
1990 	/*
1991 	 * Note that we add this partially initialized entry to the
1992 	 * connection list. This is so that we don't have connections to
1993 	 * the same server.
1994 	 *
1995 	 * Note that x_src is not initialized at this point. This is because
1996 	 * retryaddr might be NULL in which case x_src is whatever
1997 	 * t_kbind/bindresvport gives us. If another thread wants a
1998 	 * connection to the same server, seemingly we have an issue, but we
1999 	 * don't. If the other thread comes in with retryaddr == NULL, then it
2000 	 * will never look at x_src, and it will end up waiting in
2001 	 * connmgr_cwait() for the first thread to finish the connection
2002 	 * attempt. If the other thread comes in with retryaddr != NULL, then
2003 	 * that means there was a request sent on a connection, in which case
2004 	 * the the connection should already exist. Thus the first thread
2005 	 * never gets here ... it finds the connection it its server in the
2006 	 * connection list.
2007 	 *
2008 	 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd
2009 	 * thread will skip us because x_src.len == 0.
2010 	 */
2011 	cm_entry->x_next = cm_hd;
2012 	cm_hd = cm_entry;
2013 	mutex_exit(&connmgr_lock);
2014 
2015 	/*
2016 	 * Either we didn't find an entry to the server of interest, or we
2017 	 * don't have the maximum number of connections to that server -
2018 	 * create a new connection.
2019 	 */
2020 	RPCLOG0(8, "connmgr_get: creating new connection\n");
2021 	rpcerr->re_status = RPC_TLIERROR;
2022 
2023 	i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, zone_kcred());
2024 	if (i) {
2025 		RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i);
2026 		rpcerr->re_errno = i;
2027 		connmgr_cancelconn(cm_entry);
2028 		return (NULL);
2029 	}
2030 	rpc_poptimod(tiptr->fp->f_vnode);
2031 
2032 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0,
2033 	    K_TO_K, kcred, &retval)) {
2034 		RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i);
2035 		(void) t_kclose(tiptr, 1);
2036 		rpcerr->re_errno = i;
2037 		connmgr_cancelconn(cm_entry);
2038 		return (NULL);
2039 	}
2040 
2041 	if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K,
2042 	    kcred, &retval)) {
2043 		RPCLOG(1, "connmgr_get: can't set client status with cots "
2044 		    "module, %d\n", i);
2045 		(void) t_kclose(tiptr, 1);
2046 		rpcerr->re_errno = i;
2047 		connmgr_cancelconn(cm_entry);
2048 		return (NULL);
2049 	}
2050 
2051 	mutex_enter(&connmgr_lock);
2052 
2053 	wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next;
2054 	cm_entry->x_wq = wq;
2055 
2056 	mutex_exit(&connmgr_lock);
2057 
2058 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0,
2059 	    K_TO_K, kcred, &retval)) {
2060 		RPCLOG(1, "connmgr_get: can't push timod, %d\n", i);
2061 		(void) t_kclose(tiptr, 1);
2062 		rpcerr->re_errno = i;
2063 		connmgr_cancelconn(cm_entry);
2064 		return (NULL);
2065 	}
2066 
2067 	/*
2068 	 * If the caller has not specified reserved port usage then
2069 	 * take the system default.
2070 	 */
2071 	if (useresvport == -1)
2072 		useresvport = clnt_cots_do_bindresvport;
2073 
2074 	if ((useresvport || retryaddr != NULL) &&
2075 	    (addrfmly == AF_INET || addrfmly == AF_INET6)) {
2076 		bool_t alloc_src = FALSE;
2077 
2078 		if (srcaddr->len != destaddr->len) {
2079 			kmem_free(srcaddr->buf, srcaddr->maxlen);
2080 			srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2081 			srcaddr->maxlen = destaddr->len;
2082 			srcaddr->len = destaddr->len;
2083 			alloc_src = TRUE;
2084 		}
2085 
2086 		if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) {
2087 			(void) t_kclose(tiptr, 1);
2088 			RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: "
2089 			    "%p\n", (void *)retryaddr);
2090 
2091 			/*
2092 			 * 1225408: If we allocated a source address, then it
2093 			 * is either garbage or all zeroes. In that case
2094 			 * we need to clear srcaddr.
2095 			 */
2096 			if (alloc_src == TRUE) {
2097 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2098 				srcaddr->maxlen = srcaddr->len = 0;
2099 				srcaddr->buf = NULL;
2100 			}
2101 			rpcerr->re_errno = i;
2102 			connmgr_cancelconn(cm_entry);
2103 			return (NULL);
2104 		}
2105 	} else {
2106 		if ((i = t_kbind(tiptr, NULL, NULL)) != 0) {
2107 			RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i);
2108 			(void) t_kclose(tiptr, 1);
2109 			rpcerr->re_errno = i;
2110 			connmgr_cancelconn(cm_entry);
2111 			return (NULL);
2112 		}
2113 	}
2114 
2115 	{
2116 		/*
2117 		 * Keep the kernel stack lean. Don't move this call
2118 		 * declaration to the top of this function because a
2119 		 * call is declared in connmgr_wrapconnect()
2120 		 */
2121 		calllist_t call;
2122 
2123 		bzero(&call, sizeof (call));
2124 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2125 
2126 		/*
2127 		 * This is a bound end-point so don't close it's stream.
2128 		 */
2129 		connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly,
2130 		    &call, &tidu_size, FALSE, waitp, nosignal);
2131 		*rpcerr = call.call_err;
2132 		cv_destroy(&call.call_cv);
2133 
2134 	}
2135 
2136 	mutex_enter(&connmgr_lock);
2137 
2138 	/*
2139 	 * Set up a transport entry in the connection manager's list.
2140 	 */
2141 	cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP);
2142 	bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len);
2143 	cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len;
2144 
2145 	cm_entry->x_tiptr = tiptr;
2146 	cm_entry->x_time = lbolt;
2147 
2148 	if (tiptr->tp_info.servtype == T_COTS_ORD)
2149 		cm_entry->x_ordrel = TRUE;
2150 	else
2151 		cm_entry->x_ordrel = FALSE;
2152 
2153 	cm_entry->x_tidu_size = tidu_size;
2154 
2155 	if (cm_entry->x_early_disc) {
2156 		/*
2157 		 * We need to check if a disconnect request has come
2158 		 * while we are connected, if so, then we need to
2159 		 * set rpcerr->re_status appropriately before returning
2160 		 * NULL to caller.
2161 		 */
2162 		if (rpcerr->re_status == RPC_SUCCESS)
2163 			rpcerr->re_status = RPC_XPRTFAILED;
2164 		cm_entry->x_connected = FALSE;
2165 	} else
2166 		cm_entry->x_connected = connected;
2167 
2168 	/*
2169 	 * There could be a discrepancy here such that
2170 	 * x_early_disc is TRUE yet connected is TRUE as well
2171 	 * and the connection is actually connected. In that case
2172 	 * lets be conservative and declare the connection as not
2173 	 * connected.
2174 	 */
2175 	cm_entry->x_early_disc = FALSE;
2176 	cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2177 	cm_entry->x_ctime = lbolt;
2178 
2179 	/*
2180 	 * Notify any threads waiting that the connection attempt is done.
2181 	 */
2182 	cm_entry->x_thread = FALSE;
2183 	cv_broadcast(&cm_entry->x_conn_cv);
2184 
2185 	if (cm_entry->x_connected == FALSE) {
2186 		mutex_exit(&connmgr_lock);
2187 		connmgr_release(cm_entry);
2188 		return (NULL);
2189 	}
2190 
2191 	mutex_exit(&connmgr_lock);
2192 
2193 	return (cm_entry);
2194 }
2195 
2196 /*
2197  * Keep the cm_xprt entry on the connecton list when making a connection. This
2198  * is to prevent multiple connections to a slow server from appearing.
2199  * We use the bit field x_thread to tell if a thread is doing a connection
2200  * which keeps other interested threads from messing with connection.
2201  * Those other threads just wait if x_thread is set.
2202  *
2203  * If x_thread is not set, then we do the actual work of connecting via
2204  * connmgr_connect().
2205  *
2206  * mutex convention: called with connmgr_lock held, returns with it released.
2207  */
2208 static struct cm_xprt *
2209 connmgr_wrapconnect(
2210 	struct cm_xprt	*cm_entry,
2211 	const struct timeval	*waitp,
2212 	struct netbuf	*destaddr,
2213 	int		addrfmly,
2214 	struct netbuf	*srcaddr,
2215 	struct rpc_err	*rpcerr,
2216 	bool_t		reconnect,
2217 	bool_t		nosignal)
2218 {
2219 	ASSERT(MUTEX_HELD(&connmgr_lock));
2220 	/*
2221 	 * Hold this entry as we are about to drop connmgr_lock.
2222 	 */
2223 	CONN_HOLD(cm_entry);
2224 
2225 	/*
2226 	 * If there is a thread already making a connection for us, then
2227 	 * wait for it to complete the connection.
2228 	 */
2229 	if (cm_entry->x_thread == TRUE) {
2230 		rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal);
2231 
2232 		if (rpcerr->re_status != RPC_SUCCESS) {
2233 			mutex_exit(&connmgr_lock);
2234 			connmgr_release(cm_entry);
2235 			return (NULL);
2236 		}
2237 	} else {
2238 		bool_t connected;
2239 		calllist_t call;
2240 
2241 		cm_entry->x_thread = TRUE;
2242 
2243 		while (cm_entry->x_needrel == TRUE) {
2244 			cm_entry->x_needrel = FALSE;
2245 
2246 			connmgr_sndrel(cm_entry);
2247 			delay(drv_usectohz(1000000));
2248 
2249 			mutex_enter(&connmgr_lock);
2250 		}
2251 
2252 		/*
2253 		 * If we need to send a T_DISCON_REQ, send one.
2254 		 */
2255 		connmgr_dis_and_wait(cm_entry);
2256 
2257 		mutex_exit(&connmgr_lock);
2258 
2259 		bzero(&call, sizeof (call));
2260 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2261 
2262 		connected = connmgr_connect(cm_entry, cm_entry->x_wq,
2263 		    destaddr, addrfmly, &call, &cm_entry->x_tidu_size,
2264 		    reconnect, waitp, nosignal);
2265 
2266 		*rpcerr = call.call_err;
2267 		cv_destroy(&call.call_cv);
2268 
2269 		mutex_enter(&connmgr_lock);
2270 
2271 
2272 		if (cm_entry->x_early_disc) {
2273 			/*
2274 			 * We need to check if a disconnect request has come
2275 			 * while we are connected, if so, then we need to
2276 			 * set rpcerr->re_status appropriately before returning
2277 			 * NULL to caller.
2278 			 */
2279 			if (rpcerr->re_status == RPC_SUCCESS)
2280 				rpcerr->re_status = RPC_XPRTFAILED;
2281 			cm_entry->x_connected = FALSE;
2282 		} else
2283 			cm_entry->x_connected = connected;
2284 
2285 		/*
2286 		 * There could be a discrepancy here such that
2287 		 * x_early_disc is TRUE yet connected is TRUE as well
2288 		 * and the connection is actually connected. In that case
2289 		 * lets be conservative and declare the connection as not
2290 		 * connected.
2291 		 */
2292 
2293 		cm_entry->x_early_disc = FALSE;
2294 		cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2295 
2296 
2297 		/*
2298 		 * connmgr_connect() may have given up before the connection
2299 		 * actually timed out. So ensure that before the next
2300 		 * connection attempt we do a disconnect.
2301 		 */
2302 		cm_entry->x_ctime = lbolt;
2303 		cm_entry->x_thread = FALSE;
2304 
2305 		cv_broadcast(&cm_entry->x_conn_cv);
2306 
2307 		if (cm_entry->x_connected == FALSE) {
2308 			mutex_exit(&connmgr_lock);
2309 			connmgr_release(cm_entry);
2310 			return (NULL);
2311 		}
2312 	}
2313 
2314 	if (srcaddr != NULL) {
2315 		/*
2316 		 * Copy into the handle the
2317 		 * source address of the
2318 		 * connection, which we will use
2319 		 * in case of a later retry.
2320 		 */
2321 		if (srcaddr->len != cm_entry->x_src.len) {
2322 			if (srcaddr->maxlen > 0)
2323 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2324 			srcaddr->buf = kmem_zalloc(cm_entry->x_src.len,
2325 			    KM_SLEEP);
2326 			srcaddr->maxlen = srcaddr->len =
2327 			    cm_entry->x_src.len;
2328 		}
2329 		bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len);
2330 	}
2331 	cm_entry->x_time = lbolt;
2332 	mutex_exit(&connmgr_lock);
2333 	return (cm_entry);
2334 }
2335 
2336 /*
2337  * If we need to send a T_DISCON_REQ, send one.
2338  */
2339 static void
2340 connmgr_dis_and_wait(struct cm_xprt *cm_entry)
2341 {
2342 	ASSERT(MUTEX_HELD(&connmgr_lock));
2343 	for (;;) {
2344 		while (cm_entry->x_needdis == TRUE) {
2345 			RPCLOG(8, "connmgr_dis_and_wait: need "
2346 			    "T_DISCON_REQ for connection 0x%p\n",
2347 			    (void *)cm_entry);
2348 			cm_entry->x_needdis = FALSE;
2349 			cm_entry->x_waitdis = TRUE;
2350 
2351 			connmgr_snddis(cm_entry);
2352 
2353 			mutex_enter(&connmgr_lock);
2354 		}
2355 
2356 		if (cm_entry->x_waitdis == TRUE) {
2357 			clock_t curlbolt;
2358 			clock_t timout;
2359 
2360 			RPCLOG(8, "connmgr_dis_and_wait waiting for "
2361 			    "T_DISCON_REQ's ACK for connection %p\n",
2362 			    (void *)cm_entry);
2363 			curlbolt = ddi_get_lbolt();
2364 
2365 			timout = clnt_cots_min_conntout *
2366 			    drv_usectohz(1000000) + curlbolt;
2367 
2368 			/*
2369 			 * The TPI spec says that the T_DISCON_REQ
2370 			 * will get acknowledged, but in practice
2371 			 * the ACK may never get sent. So don't
2372 			 * block forever.
2373 			 */
2374 			(void) cv_timedwait(&cm_entry->x_dis_cv,
2375 			    &connmgr_lock, timout);
2376 		}
2377 		/*
2378 		 * If we got the ACK, break. If we didn't,
2379 		 * then send another T_DISCON_REQ.
2380 		 */
2381 		if (cm_entry->x_waitdis == FALSE) {
2382 			break;
2383 		} else {
2384 			RPCLOG(8, "connmgr_dis_and_wait: did"
2385 			    "not get T_DISCON_REQ's ACK for "
2386 			    "connection  %p\n", (void *)cm_entry);
2387 			cm_entry->x_needdis = TRUE;
2388 		}
2389 	}
2390 }
2391 
2392 static void
2393 connmgr_cancelconn(struct cm_xprt *cm_entry)
2394 {
2395 	/*
2396 	 * Mark the connection table entry as dead; the next thread that
2397 	 * goes through connmgr_release() will notice this and deal with it.
2398 	 */
2399 	mutex_enter(&connmgr_lock);
2400 	cm_entry->x_dead = TRUE;
2401 
2402 	/*
2403 	 * Notify any threads waiting for the connection that it isn't
2404 	 * going to happen.
2405 	 */
2406 	cm_entry->x_thread = FALSE;
2407 	cv_broadcast(&cm_entry->x_conn_cv);
2408 	mutex_exit(&connmgr_lock);
2409 
2410 	connmgr_release(cm_entry);
2411 }
2412 
2413 static void
2414 connmgr_close(struct cm_xprt *cm_entry)
2415 {
2416 	mutex_enter(&cm_entry->x_lock);
2417 	while (cm_entry->x_ref != 0) {
2418 		/*
2419 		 * Must be a noninterruptible wait.
2420 		 */
2421 		cv_wait(&cm_entry->x_cv, &cm_entry->x_lock);
2422 	}
2423 
2424 	if (cm_entry->x_tiptr != NULL)
2425 		(void) t_kclose(cm_entry->x_tiptr, 1);
2426 
2427 	mutex_exit(&cm_entry->x_lock);
2428 	if (cm_entry->x_ksp != NULL) {
2429 		mutex_enter(&connmgr_lock);
2430 		cm_entry->x_ksp->ks_private = NULL;
2431 		mutex_exit(&connmgr_lock);
2432 
2433 		/*
2434 		 * Must free the buffer we allocated for the
2435 		 * server address in the update function
2436 		 */
2437 		if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2438 		    x_server.value.str.addr.ptr != NULL)
2439 			kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp->
2440 			    ks_data))->x_server.value.str.addr.ptr,
2441 			    INET6_ADDRSTRLEN);
2442 		kmem_free(cm_entry->x_ksp->ks_data,
2443 		    cm_entry->x_ksp->ks_data_size);
2444 		kstat_delete(cm_entry->x_ksp);
2445 	}
2446 
2447 	mutex_destroy(&cm_entry->x_lock);
2448 	cv_destroy(&cm_entry->x_cv);
2449 	cv_destroy(&cm_entry->x_conn_cv);
2450 	cv_destroy(&cm_entry->x_dis_cv);
2451 
2452 	if (cm_entry->x_server.buf != NULL)
2453 		kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen);
2454 	if (cm_entry->x_src.buf != NULL)
2455 		kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen);
2456 	kmem_free(cm_entry, sizeof (struct cm_xprt));
2457 }
2458 
2459 /*
2460  * Called by KRPC after sending the call message to release the connection
2461  * it was using.
2462  */
2463 static void
2464 connmgr_release(struct cm_xprt *cm_entry)
2465 {
2466 	mutex_enter(&cm_entry->x_lock);
2467 	cm_entry->x_ref--;
2468 	if (cm_entry->x_ref == 0)
2469 		cv_signal(&cm_entry->x_cv);
2470 	mutex_exit(&cm_entry->x_lock);
2471 }
2472 
2473 /*
2474  * Given an open stream, connect to the remote.  Returns true if connected,
2475  * false otherwise.
2476  */
2477 static bool_t
2478 connmgr_connect(
2479 	struct cm_xprt		*cm_entry,
2480 	queue_t			*wq,
2481 	struct netbuf		*addr,
2482 	int			addrfmly,
2483 	calllist_t 		*e,
2484 	int 			*tidu_ptr,
2485 	bool_t 			reconnect,
2486 	const struct timeval 	*waitp,
2487 	bool_t 			nosignal)
2488 {
2489 	mblk_t *mp;
2490 	struct T_conn_req *tcr;
2491 	struct T_info_ack *tinfo;
2492 	int interrupted, error;
2493 	int tidu_size, kstat_instance;
2494 
2495 	/* if it's a reconnect, flush any lingering data messages */
2496 	if (reconnect)
2497 		(void) putctl1(wq, M_FLUSH, FLUSHRW);
2498 
2499 	mp = allocb(sizeof (*tcr) + addr->len, BPRI_LO);
2500 	if (mp == NULL) {
2501 		/*
2502 		 * This is unfortunate, but we need to look up the stats for
2503 		 * this zone to increment the "memory allocation failed"
2504 		 * counter.  curproc->p_zone is safe since we're initiating a
2505 		 * connection and not in some strange streams context.
2506 		 */
2507 		struct rpcstat *rpcstat;
2508 
2509 		rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
2510 		ASSERT(rpcstat != NULL);
2511 
2512 		RPCLOG0(1, "connmgr_connect: cannot alloc mp for "
2513 		    "sending conn request\n");
2514 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem);
2515 		e->call_status = RPC_SYSTEMERROR;
2516 		e->call_reason = ENOSR;
2517 		return (FALSE);
2518 	}
2519 
2520 	mp->b_datap->db_type = M_PROTO;
2521 	tcr = (struct T_conn_req *)mp->b_rptr;
2522 	bzero(tcr, sizeof (*tcr));
2523 	tcr->PRIM_type = T_CONN_REQ;
2524 	tcr->DEST_length = addr->len;
2525 	tcr->DEST_offset = sizeof (struct T_conn_req);
2526 	mp->b_wptr = mp->b_rptr + sizeof (*tcr);
2527 
2528 	bcopy(addr->buf, mp->b_wptr, tcr->DEST_length);
2529 	mp->b_wptr += tcr->DEST_length;
2530 
2531 	RPCLOG(8, "connmgr_connect: sending conn request on queue "
2532 	    "%p", (void *)wq);
2533 	RPCLOG(8, " call %p\n", (void *)wq);
2534 	/*
2535 	 * We use the entry in the handle that is normally used for
2536 	 * waiting for RPC replies to wait for the connection accept.
2537 	 */
2538 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2539 		DTRACE_PROBE(krpc__e__connmgr__connect__cantsend);
2540 		freemsg(mp);
2541 		return (FALSE);
2542 	}
2543 
2544 	mutex_enter(&clnt_pending_lock);
2545 
2546 	/*
2547 	 * We wait for the transport connection to be made, or an
2548 	 * indication that it could not be made.
2549 	 */
2550 	interrupted = 0;
2551 
2552 	/*
2553 	 * waitforack should have been called with T_OK_ACK, but the
2554 	 * present implementation needs to be passed T_INFO_ACK to
2555 	 * work correctly.
2556 	 */
2557 	error = waitforack(e, T_INFO_ACK, waitp, nosignal);
2558 	if (error == EINTR)
2559 		interrupted = 1;
2560 	if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) {
2561 		/*
2562 		 * No time to lose; we essentially have been signaled to
2563 		 * quit.
2564 		 */
2565 		interrupted = 1;
2566 	}
2567 #ifdef RPCDEBUG
2568 	if (error == ETIME)
2569 		RPCLOG0(8, "connmgr_connect: giving up "
2570 		    "on connection attempt; "
2571 		    "clnt_dispatch notifyconn "
2572 		    "diagnostic 'no one waiting for "
2573 		    "connection' should not be "
2574 		    "unexpected\n");
2575 #endif
2576 	if (e->call_prev)
2577 		e->call_prev->call_next = e->call_next;
2578 	else
2579 		clnt_pending = e->call_next;
2580 	if (e->call_next)
2581 		e->call_next->call_prev = e->call_prev;
2582 	mutex_exit(&clnt_pending_lock);
2583 
2584 	if (e->call_status != RPC_SUCCESS || error != 0) {
2585 		if (interrupted)
2586 			e->call_status = RPC_INTR;
2587 		else if (error == ETIME)
2588 			e->call_status = RPC_TIMEDOUT;
2589 		else if (error == EPROTO)
2590 			e->call_status = RPC_SYSTEMERROR;
2591 
2592 		RPCLOG(8, "connmgr_connect: can't connect, status: "
2593 		    "%s\n", clnt_sperrno(e->call_status));
2594 
2595 		if (e->call_reply) {
2596 			freemsg(e->call_reply);
2597 			e->call_reply = NULL;
2598 		}
2599 
2600 		return (FALSE);
2601 	}
2602 	/*
2603 	 * The result of the "connection accept" is a T_info_ack
2604 	 * in the call_reply field.
2605 	 */
2606 	ASSERT(e->call_reply != NULL);
2607 	mp = e->call_reply;
2608 	e->call_reply = NULL;
2609 	tinfo = (struct T_info_ack *)mp->b_rptr;
2610 
2611 	tidu_size = tinfo->TIDU_size;
2612 	tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT);
2613 	if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0))
2614 		tidu_size = COTS_DEFAULT_ALLOCSIZE;
2615 	*tidu_ptr = tidu_size;
2616 
2617 	freemsg(mp);
2618 
2619 	/*
2620 	 * Set up the pertinent options.  NODELAY is so the transport doesn't
2621 	 * buffer up RPC messages on either end.  This may not be valid for
2622 	 * all transports.  Failure to set this option is not cause to
2623 	 * bail out so we return success anyway.  Note that lack of NODELAY
2624 	 * or some other way to flush the message on both ends will cause
2625 	 * lots of retries and terrible performance.
2626 	 */
2627 	if (addrfmly == AF_INET || addrfmly == AF_INET6) {
2628 		(void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e);
2629 		if (e->call_status == RPC_XPRTFAILED)
2630 			return (FALSE);
2631 	}
2632 
2633 	/*
2634 	 * Since we have a connection, we now need to figure out if
2635 	 * we need to create a kstat. If x_ksp is not NULL then we
2636 	 * are reusing a connection and so we do not need to create
2637 	 * another kstat -- lets just return.
2638 	 */
2639 	if (cm_entry->x_ksp != NULL)
2640 		return (TRUE);
2641 
2642 	/*
2643 	 * We need to increment rpc_kstat_instance atomically to prevent
2644 	 * two kstats being created with the same instance.
2645 	 */
2646 	kstat_instance = atomic_add_32_nv((uint32_t *)&rpc_kstat_instance, 1);
2647 
2648 	if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance,
2649 	    "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED,
2650 	    (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)),
2651 	    KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) {
2652 		return (TRUE);
2653 	}
2654 
2655 	cm_entry->x_ksp->ks_lock = &connmgr_lock;
2656 	cm_entry->x_ksp->ks_private = cm_entry;
2657 	cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char))
2658 	    + sizeof (cm_kstat_template));
2659 	cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size,
2660 	    KM_SLEEP);
2661 	bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data,
2662 	    cm_entry->x_ksp->ks_data_size);
2663 	((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2664 	    x_server.value.str.addr.ptr =
2665 	    kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP);
2666 
2667 	cm_entry->x_ksp->ks_update = conn_kstat_update;
2668 	kstat_install(cm_entry->x_ksp);
2669 	return (TRUE);
2670 }
2671 
2672 /*
2673  * Called by connmgr_connect to set an option on the new stream.
2674  */
2675 static bool_t
2676 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e)
2677 {
2678 	mblk_t *mp;
2679 	struct opthdr *opt;
2680 	struct T_optmgmt_req *tor;
2681 	struct timeval waitp;
2682 	int error;
2683 
2684 	mp = allocb(sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2685 	    sizeof (int), BPRI_LO);
2686 	if (mp == NULL) {
2687 		RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option "
2688 		    "request\n");
2689 		return (FALSE);
2690 	}
2691 
2692 	mp->b_datap->db_type = M_PROTO;
2693 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2694 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2695 	tor->MGMT_flags = T_NEGOTIATE;
2696 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2697 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2698 
2699 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2700 	opt->level = level;
2701 	opt->name = name;
2702 	opt->len = sizeof (int);
2703 	*(int *)((char *)opt + sizeof (*opt)) = 1;
2704 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2705 	    sizeof (int);
2706 
2707 	/*
2708 	 * We will use this connection regardless
2709 	 * of whether or not the option is settable.
2710 	 */
2711 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2712 		DTRACE_PROBE(krpc__e__connmgr__setopt__cantsend);
2713 		freemsg(mp);
2714 		return (FALSE);
2715 	}
2716 
2717 	mutex_enter(&clnt_pending_lock);
2718 
2719 	waitp.tv_sec = clnt_cots_min_conntout;
2720 	waitp.tv_usec = 0;
2721 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2722 
2723 	if (e->call_prev)
2724 		e->call_prev->call_next = e->call_next;
2725 	else
2726 		clnt_pending = e->call_next;
2727 	if (e->call_next)
2728 		e->call_next->call_prev = e->call_prev;
2729 	mutex_exit(&clnt_pending_lock);
2730 
2731 	if (e->call_reply != NULL) {
2732 		freemsg(e->call_reply);
2733 		e->call_reply = NULL;
2734 	}
2735 
2736 	if (e->call_status != RPC_SUCCESS || error != 0) {
2737 		RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name);
2738 		return (FALSE);
2739 	}
2740 	RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name);
2741 	return (TRUE);
2742 }
2743 
2744 #ifdef	DEBUG
2745 
2746 /*
2747  * This is a knob to let us force code coverage in allocation failure
2748  * case.
2749  */
2750 static int	connmgr_failsnd;
2751 #define	CONN_SND_ALLOC(Size, Pri)	\
2752 	((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri))
2753 
2754 #else
2755 
2756 #define	CONN_SND_ALLOC(Size, Pri)	allocb(Size, Pri)
2757 
2758 #endif
2759 
2760 /*
2761  * Sends an orderly release on the specified queue.
2762  * Entered with connmgr_lock. Exited without connmgr_lock
2763  */
2764 static void
2765 connmgr_sndrel(struct cm_xprt *cm_entry)
2766 {
2767 	struct T_ordrel_req *torr;
2768 	mblk_t *mp;
2769 	queue_t *q = cm_entry->x_wq;
2770 	ASSERT(MUTEX_HELD(&connmgr_lock));
2771 	mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO);
2772 	if (mp == NULL) {
2773 		cm_entry->x_needrel = TRUE;
2774 		mutex_exit(&connmgr_lock);
2775 		RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel "
2776 		    "to queue %p\n", (void *)q);
2777 		return;
2778 	}
2779 	mutex_exit(&connmgr_lock);
2780 
2781 	mp->b_datap->db_type = M_PROTO;
2782 	torr = (struct T_ordrel_req *)(mp->b_rptr);
2783 	torr->PRIM_type = T_ORDREL_REQ;
2784 	mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req);
2785 
2786 	RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q);
2787 	put(q, mp);
2788 }
2789 
2790 /*
2791  * Sends an disconnect on the specified queue.
2792  * Entered with connmgr_lock. Exited without connmgr_lock
2793  */
2794 static void
2795 connmgr_snddis(struct cm_xprt *cm_entry)
2796 {
2797 	struct T_discon_req *tdis;
2798 	mblk_t *mp;
2799 	queue_t *q = cm_entry->x_wq;
2800 
2801 	ASSERT(MUTEX_HELD(&connmgr_lock));
2802 	mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO);
2803 	if (mp == NULL) {
2804 		cm_entry->x_needdis = TRUE;
2805 		mutex_exit(&connmgr_lock);
2806 		RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon "
2807 		    "to queue %p\n", (void *)q);
2808 		return;
2809 	}
2810 	mutex_exit(&connmgr_lock);
2811 
2812 	mp->b_datap->db_type = M_PROTO;
2813 	tdis = (struct T_discon_req *)mp->b_rptr;
2814 	tdis->PRIM_type = T_DISCON_REQ;
2815 	mp->b_wptr = mp->b_rptr + sizeof (*tdis);
2816 
2817 	RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q);
2818 	put(q, mp);
2819 }
2820 
2821 /*
2822  * Sets up the entry for receiving replies, and calls rpcmod's write put proc
2823  * (through put) to send the call.
2824  */
2825 static int
2826 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid,
2827 			uint_t queue_flag)
2828 {
2829 	ASSERT(e != NULL);
2830 
2831 	e->call_status = RPC_TIMEDOUT;	/* optimistic, eh? */
2832 	e->call_reason = 0;
2833 	e->call_wq = q;
2834 	e->call_xid = xid;
2835 	e->call_notified = FALSE;
2836 
2837 	if (!canput(q)) {
2838 		e->call_status = RPC_CANTSEND;
2839 		e->call_reason = EIO;
2840 		return (RPC_CANTSEND);
2841 	}
2842 
2843 	/*
2844 	 * If queue_flag is set then the calllist_t is already on the hash
2845 	 * queue.  In this case just send the message and return.
2846 	 */
2847 	if (queue_flag) {
2848 		put(q, mp);
2849 		return (RPC_SUCCESS);
2850 
2851 	}
2852 
2853 	/*
2854 	 * Set up calls for RPC requests (with XID != 0) on the hash
2855 	 * queue for fast lookups and place other calls (i.e.
2856 	 * connection management) on the linked list.
2857 	 */
2858 	if (xid != 0) {
2859 		RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on "
2860 		    "dispatch list\n", xid);
2861 		e->call_hash = call_hash(xid, clnt_cots_hash_size);
2862 		e->call_bucket = &cots_call_ht[e->call_hash];
2863 		call_table_enter(e);
2864 	} else {
2865 		mutex_enter(&clnt_pending_lock);
2866 		if (clnt_pending)
2867 			clnt_pending->call_prev = e;
2868 		e->call_next = clnt_pending;
2869 		e->call_prev = NULL;
2870 		clnt_pending = e;
2871 		mutex_exit(&clnt_pending_lock);
2872 	}
2873 
2874 	put(q, mp);
2875 	return (RPC_SUCCESS);
2876 }
2877 
2878 /*
2879  * Called by rpcmod to notify a client with a clnt_pending call that its reply
2880  * has arrived.  If we can't find a client waiting for this reply, we log
2881  * the error and return.
2882  */
2883 bool_t
2884 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid)
2885 {
2886 	calllist_t *e = NULL;
2887 	call_table_t *chtp;
2888 	uint32_t xid;
2889 	uint_t hash;
2890 
2891 	if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) &&
2892 	    (mp->b_wptr - mp->b_rptr) >= sizeof (xid))
2893 		xid = *((uint32_t *)mp->b_rptr);
2894 	else {
2895 		int i = 0;
2896 		unsigned char *p = (unsigned char *)&xid;
2897 		unsigned char *rptr;
2898 		mblk_t *tmp = mp;
2899 
2900 		/*
2901 		 * Copy the xid, byte-by-byte into xid.
2902 		 */
2903 		while (tmp) {
2904 			rptr = tmp->b_rptr;
2905 			while (rptr < tmp->b_wptr) {
2906 				*p++ = *rptr++;
2907 				if (++i >= sizeof (xid))
2908 					goto done_xid_copy;
2909 			}
2910 			tmp = tmp->b_cont;
2911 		}
2912 
2913 		/*
2914 		 * If we got here, we ran out of mblk space before the
2915 		 * xid could be copied.
2916 		 */
2917 		ASSERT(tmp == NULL && i < sizeof (xid));
2918 
2919 		RPCLOG0(1,
2920 		    "clnt_dispatch_notify: message less than size of xid\n");
2921 		return (FALSE);
2922 
2923 	}
2924 done_xid_copy:
2925 
2926 	hash = call_hash(xid, clnt_cots_hash_size);
2927 	chtp = &cots_call_ht[hash];
2928 	/* call_table_find returns with the hash bucket locked */
2929 	call_table_find(chtp, xid, e);
2930 
2931 	if (e != NULL) {
2932 		/*
2933 		 * Found thread waiting for this reply
2934 		 */
2935 		mutex_enter(&e->call_lock);
2936 
2937 		/*
2938 		 * verify that the reply is coming in on
2939 		 * the same zone that it was sent from.
2940 		 */
2941 		if (e->call_zoneid != zoneid) {
2942 			mutex_exit(&e->call_lock);
2943 			mutex_exit(&chtp->ct_lock);
2944 			return (FALSE);
2945 		}
2946 
2947 		if (e->call_reply)
2948 			/*
2949 			 * This can happen under the following scenario:
2950 			 * clnt_cots_kcallit() times out on the response,
2951 			 * rfscall() repeats the CLNT_CALL() with
2952 			 * the same xid, clnt_cots_kcallit() sends the retry,
2953 			 * thereby putting the clnt handle on the pending list,
2954 			 * the first response arrives, signalling the thread
2955 			 * in clnt_cots_kcallit(). Before that thread is
2956 			 * dispatched, the second response arrives as well,
2957 			 * and clnt_dispatch_notify still finds the handle on
2958 			 * the pending list, with call_reply set. So free the
2959 			 * old reply now.
2960 			 *
2961 			 * It is also possible for a response intended for
2962 			 * an RPC call with a different xid to reside here.
2963 			 * This can happen if the thread that owned this
2964 			 * client handle prior to the current owner bailed
2965 			 * out and left its call record on the dispatch
2966 			 * queue.  A window exists where the response can
2967 			 * arrive before the current owner dispatches its
2968 			 * RPC call.
2969 			 *
2970 			 * In any case, this is the very last point where we
2971 			 * can safely check the call_reply field before
2972 			 * placing the new response there.
2973 			 */
2974 			freemsg(e->call_reply);
2975 		e->call_reply = mp;
2976 		e->call_status = RPC_SUCCESS;
2977 		e->call_notified = TRUE;
2978 		cv_signal(&e->call_cv);
2979 		mutex_exit(&e->call_lock);
2980 		mutex_exit(&chtp->ct_lock);
2981 		return (TRUE);
2982 	} else {
2983 		zone_t *zone;
2984 		struct rpcstat *rpcstat;
2985 
2986 		mutex_exit(&chtp->ct_lock);
2987 		RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n",
2988 		    xid);
2989 		/*
2990 		 * This is unfortunate, but we need to lookup the zone so we
2991 		 * can increment its "rcbadxids" counter.
2992 		 */
2993 		zone = zone_find_by_id(zoneid);
2994 		if (zone == NULL) {
2995 			/*
2996 			 * The zone went away...
2997 			 */
2998 			return (FALSE);
2999 		}
3000 		rpcstat = zone_getspecific(rpcstat_zone_key, zone);
3001 		if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) {
3002 			/*
3003 			 * Not interested
3004 			 */
3005 			zone_rele(zone);
3006 			return (FALSE);
3007 		}
3008 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids);
3009 		zone_rele(zone);
3010 	}
3011 	return (FALSE);
3012 }
3013 
3014 /*
3015  * Called by rpcmod when a non-data indication arrives.  The ones in which we
3016  * are interested are connection indications and options acks.  We dispatch
3017  * based on the queue the indication came in on.  If we are not interested in
3018  * what came in, we return false to rpcmod, who will then pass it upstream.
3019  */
3020 bool_t
3021 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp)
3022 {
3023 	calllist_t *e;
3024 	int type;
3025 
3026 	ASSERT((q->q_flag & QREADR) == 0);
3027 
3028 	type = ((union T_primitives *)mp->b_rptr)->type;
3029 	RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n",
3030 	    rpc_tpiprim2name(type));
3031 	mutex_enter(&clnt_pending_lock);
3032 	for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) {
3033 		if (e == NULL) {
3034 			mutex_exit(&clnt_pending_lock);
3035 			RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting "
3036 			    "for connection on queue 0x%p\n", (void *)q);
3037 			return (FALSE);
3038 		}
3039 		if (e->call_wq == q)
3040 			break;
3041 	}
3042 
3043 	switch (type) {
3044 	case T_CONN_CON:
3045 		/*
3046 		 * The transport is now connected, send a T_INFO_REQ to get
3047 		 * the tidu size.
3048 		 */
3049 		mutex_exit(&clnt_pending_lock);
3050 		ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >=
3051 		    sizeof (struct T_info_req));
3052 		mp->b_rptr = mp->b_datap->db_base;
3053 		((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ;
3054 		mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req);
3055 		mp->b_datap->db_type = M_PCPROTO;
3056 		put(q, mp);
3057 		return (TRUE);
3058 	case T_INFO_ACK:
3059 	case T_OPTMGMT_ACK:
3060 		e->call_status = RPC_SUCCESS;
3061 		e->call_reply = mp;
3062 		e->call_notified = TRUE;
3063 		cv_signal(&e->call_cv);
3064 		break;
3065 	case T_ERROR_ACK:
3066 		e->call_status = RPC_CANTCONNECT;
3067 		e->call_reply = mp;
3068 		e->call_notified = TRUE;
3069 		cv_signal(&e->call_cv);
3070 		break;
3071 	case T_OK_ACK:
3072 		/*
3073 		 * Great, but we are really waiting for a T_CONN_CON
3074 		 */
3075 		freemsg(mp);
3076 		break;
3077 	default:
3078 		mutex_exit(&clnt_pending_lock);
3079 		RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type);
3080 		return (FALSE);
3081 	}
3082 
3083 	mutex_exit(&clnt_pending_lock);
3084 	return (TRUE);
3085 }
3086 
3087 /*
3088  * Called by rpcmod when the transport is (or should be) going away.  Informs
3089  * all callers waiting for replies and marks the entry in the connection
3090  * manager's list as unconnected, and either closing (close handshake in
3091  * progress) or dead.
3092  */
3093 void
3094 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason)
3095 {
3096 	calllist_t *e;
3097 	call_table_t *ctp;
3098 	struct cm_xprt *cm_entry;
3099 	int have_connmgr_lock;
3100 	int i;
3101 
3102 	ASSERT((q->q_flag & QREADR) == 0);
3103 
3104 	RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q);
3105 	RPCLOG(1, " received a notifcation prim type [%s]",
3106 	    rpc_tpiprim2name(msg_type));
3107 	RPCLOG(1, " and reason %d\n", reason);
3108 
3109 	/*
3110 	 * Find the transport entry in the connection manager's list, close
3111 	 * the transport and delete the entry.  In the case where rpcmod's
3112 	 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we
3113 	 * should gracefully close the connection.
3114 	 */
3115 	have_connmgr_lock = 1;
3116 	mutex_enter(&connmgr_lock);
3117 	for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) {
3118 		ASSERT(cm_entry != cm_entry->x_next);
3119 		if (cm_entry->x_wq == q) {
3120 			ASSERT(MUTEX_HELD(&connmgr_lock));
3121 			ASSERT(have_connmgr_lock == 1);
3122 			switch (msg_type) {
3123 			case T_ORDREL_REQ:
3124 
3125 				if (cm_entry->x_dead) {
3126 					RPCLOG(1, "idle timeout on dead "
3127 					    "connection: %p\n",
3128 					    (void *)cm_entry);
3129 					if (clnt_stop_idle != NULL)
3130 						(*clnt_stop_idle)(q);
3131 					break;
3132 				}
3133 
3134 				/*
3135 				 * Only mark the connection as dead if it is
3136 				 * connected and idle.
3137 				 * An unconnected connection has probably
3138 				 * gone idle because the server is down,
3139 				 * and when it comes back up there will be
3140 				 * retries that need to use that connection.
3141 				 */
3142 				if (cm_entry->x_connected ||
3143 				    cm_entry->x_doomed) {
3144 					if (cm_entry->x_ordrel) {
3145 						if (cm_entry->x_closing ==
3146 						    TRUE) {
3147 							/*
3148 							 * The connection is
3149 							 * obviously wedged due
3150 							 * to a bug or problem
3151 							 * with the transport.
3152 							 * Mark it as dead.
3153 							 * Otherwise we can
3154 							 * leak connections.
3155 							 */
3156 							cm_entry->x_dead = TRUE;
3157 							mutex_exit(
3158 							    &connmgr_lock);
3159 							have_connmgr_lock = 0;
3160 							if (clnt_stop_idle !=
3161 							    NULL)
3162 							(*clnt_stop_idle)(q);
3163 							break;
3164 						}
3165 						cm_entry->x_closing = TRUE;
3166 						connmgr_sndrel(cm_entry);
3167 						have_connmgr_lock = 0;
3168 					} else {
3169 						cm_entry->x_dead = TRUE;
3170 						mutex_exit(&connmgr_lock);
3171 						have_connmgr_lock = 0;
3172 						if (clnt_stop_idle != NULL)
3173 							(*clnt_stop_idle)(q);
3174 					}
3175 				} else {
3176 					/*
3177 					 * We don't mark the connection
3178 					 * as dead, but we turn off the
3179 					 * idle timer.
3180 					 */
3181 					mutex_exit(&connmgr_lock);
3182 					have_connmgr_lock = 0;
3183 					if (clnt_stop_idle != NULL)
3184 						(*clnt_stop_idle)(q);
3185 					RPCLOG(1, "clnt_dispatch_notifyall:"
3186 					    " ignoring timeout from rpcmod"
3187 					    " (q %p) because we are not "
3188 					    " connected\n", (void *)q);
3189 				}
3190 				break;
3191 			case T_ORDREL_IND:
3192 				/*
3193 				 * If this entry is marked closing, then we are
3194 				 * completing a close handshake, and the
3195 				 * connection is dead.  Otherwise, the server is
3196 				 * trying to close. Since the server will not
3197 				 * be sending any more RPC replies, we abort
3198 				 * the connection, including flushing
3199 				 * any RPC requests that are in-transit.
3200 				 */
3201 				if (cm_entry->x_closing) {
3202 					cm_entry->x_dead = TRUE;
3203 					mutex_exit(&connmgr_lock);
3204 					have_connmgr_lock = 0;
3205 					if (clnt_stop_idle != NULL)
3206 						(*clnt_stop_idle)(q);
3207 				} else {
3208 					/*
3209 					 * if we're getting a disconnect
3210 					 * before we've finished our
3211 					 * connect attempt, mark it for
3212 					 * later processing
3213 					 */
3214 					if (cm_entry->x_thread)
3215 						cm_entry->x_early_disc = TRUE;
3216 					else
3217 						cm_entry->x_connected = FALSE;
3218 					cm_entry->x_waitdis = TRUE;
3219 					connmgr_snddis(cm_entry);
3220 					have_connmgr_lock = 0;
3221 				}
3222 				break;
3223 
3224 			case T_ERROR_ACK:
3225 			case T_OK_ACK:
3226 				cm_entry->x_waitdis = FALSE;
3227 				cv_signal(&cm_entry->x_dis_cv);
3228 				mutex_exit(&connmgr_lock);
3229 				return;
3230 
3231 			case T_DISCON_REQ:
3232 				if (cm_entry->x_thread)
3233 					cm_entry->x_early_disc = TRUE;
3234 				else
3235 					cm_entry->x_connected = FALSE;
3236 				cm_entry->x_waitdis = TRUE;
3237 
3238 				connmgr_snddis(cm_entry);
3239 				have_connmgr_lock = 0;
3240 				break;
3241 
3242 			case T_DISCON_IND:
3243 			default:
3244 				/*
3245 				 * if we're getting a disconnect before
3246 				 * we've finished our connect attempt,
3247 				 * mark it for later processing
3248 				 */
3249 				if (cm_entry->x_closing) {
3250 					cm_entry->x_dead = TRUE;
3251 					mutex_exit(&connmgr_lock);
3252 					have_connmgr_lock = 0;
3253 					if (clnt_stop_idle != NULL)
3254 						(*clnt_stop_idle)(q);
3255 				} else {
3256 					if (cm_entry->x_thread) {
3257 						cm_entry->x_early_disc = TRUE;
3258 					} else {
3259 						cm_entry->x_dead = TRUE;
3260 						cm_entry->x_connected = FALSE;
3261 					}
3262 				}
3263 				break;
3264 			}
3265 			break;
3266 		}
3267 	}
3268 
3269 	if (have_connmgr_lock)
3270 		mutex_exit(&connmgr_lock);
3271 
3272 	if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) {
3273 		RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find "
3274 		    "connmgr entry for discon ack\n", (void *)q);
3275 		return;
3276 	}
3277 
3278 	/*
3279 	 * Then kick all the clnt_pending calls out of their wait.  There
3280 	 * should be no clnt_pending calls in the case of rpcmod's idle
3281 	 * timer firing.
3282 	 */
3283 	for (i = 0; i < clnt_cots_hash_size; i++) {
3284 		ctp = &cots_call_ht[i];
3285 		mutex_enter(&ctp->ct_lock);
3286 		for (e = ctp->ct_call_next;
3287 		    e != (calllist_t *)ctp;
3288 		    e = e->call_next) {
3289 			if (e->call_wq == q && e->call_notified == FALSE) {
3290 				RPCLOG(1,
3291 				    "clnt_dispatch_notifyall for queue %p ",
3292 				    (void *)q);
3293 				RPCLOG(1, "aborting clnt_pending call %p\n",
3294 				    (void *)e);
3295 
3296 				if (msg_type == T_DISCON_IND)
3297 					e->call_reason = reason;
3298 				e->call_notified = TRUE;
3299 				e->call_status = RPC_XPRTFAILED;
3300 				cv_signal(&e->call_cv);
3301 			}
3302 		}
3303 		mutex_exit(&ctp->ct_lock);
3304 	}
3305 
3306 	mutex_enter(&clnt_pending_lock);
3307 	for (e = clnt_pending; e; e = e->call_next) {
3308 		/*
3309 		 * Only signal those RPC handles that haven't been
3310 		 * signalled yet. Otherwise we can get a bogus call_reason.
3311 		 * This can happen if thread A is making a call over a
3312 		 * connection. If the server is killed, it will cause
3313 		 * reset, and reason will default to EIO as a result of
3314 		 * a T_ORDREL_IND. Thread B then attempts to recreate
3315 		 * the connection but gets a T_DISCON_IND. If we set the
3316 		 * call_reason code for all threads, then if thread A
3317 		 * hasn't been dispatched yet, it will get the wrong
3318 		 * reason. The bogus call_reason can make it harder to
3319 		 * discriminate between calls that fail because the
3320 		 * connection attempt failed versus those where the call
3321 		 * may have been executed on the server.
3322 		 */
3323 		if (e->call_wq == q && e->call_notified == FALSE) {
3324 			RPCLOG(1, "clnt_dispatch_notifyall for queue %p ",
3325 			    (void *)q);
3326 			RPCLOG(1, " aborting clnt_pending call %p\n",
3327 			    (void *)e);
3328 
3329 			if (msg_type == T_DISCON_IND)
3330 				e->call_reason = reason;
3331 			e->call_notified = TRUE;
3332 			/*
3333 			 * Let the caller timeout, else he will retry
3334 			 * immediately.
3335 			 */
3336 			e->call_status = RPC_XPRTFAILED;
3337 
3338 			/*
3339 			 * We used to just signal those threads
3340 			 * waiting for a connection, (call_xid = 0).
3341 			 * That meant that threads waiting for a response
3342 			 * waited till their timeout expired. This
3343 			 * could be a long time if they've specified a
3344 			 * maximum timeout. (2^31 - 1). So we
3345 			 * Signal all threads now.
3346 			 */
3347 			cv_signal(&e->call_cv);
3348 		}
3349 	}
3350 	mutex_exit(&clnt_pending_lock);
3351 }
3352 
3353 
3354 /*ARGSUSED*/
3355 /*
3356  * after resuming a system that's been suspended for longer than the
3357  * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall()
3358  * generates "NFS server X not responding" and "NFS server X ok" messages;
3359  * here we reset inet connections to cause a re-connect and avoid those
3360  * NFS messages.  see 4045054
3361  */
3362 boolean_t
3363 connmgr_cpr_reset(void *arg, int code)
3364 {
3365 	struct cm_xprt *cxp;
3366 
3367 	if (code == CB_CODE_CPR_CHKPT)
3368 		return (B_TRUE);
3369 
3370 	if (mutex_tryenter(&connmgr_lock) == 0)
3371 		return (B_FALSE);
3372 	for (cxp = cm_hd; cxp; cxp = cxp->x_next) {
3373 		if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) &&
3374 		    cxp->x_connected == TRUE) {
3375 			if (cxp->x_thread)
3376 				cxp->x_early_disc = TRUE;
3377 			else
3378 				cxp->x_connected = FALSE;
3379 			cxp->x_needdis = TRUE;
3380 		}
3381 	}
3382 	mutex_exit(&connmgr_lock);
3383 	return (B_TRUE);
3384 }
3385 
3386 void
3387 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp)
3388 {
3389 
3390 	*statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid,
3391 	    "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl,
3392 	    sizeof (cots_rcstat_tmpl));
3393 }
3394 
3395 void
3396 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp)
3397 {
3398 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client");
3399 	kmem_free(*statsp, sizeof (cots_rcstat_tmpl));
3400 }
3401 
3402 void
3403 clnt_cots_init(void)
3404 {
3405 	mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL);
3406 	mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL);
3407 
3408 	if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE)
3409 		clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE;
3410 
3411 	cots_call_ht = call_table_init(clnt_cots_hash_size);
3412 	zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy);
3413 }
3414 
3415 void
3416 clnt_cots_fini(void)
3417 {
3418 	(void) zone_key_delete(zone_cots_key);
3419 }
3420 
3421 /*
3422  * Wait for TPI ack, returns success only if expected ack is received
3423  * within timeout period.
3424  */
3425 
3426 static int
3427 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp,
3428     bool_t nosignal)
3429 {
3430 	union T_primitives *tpr;
3431 	clock_t timout;
3432 	int cv_stat = 1;
3433 
3434 	ASSERT(MUTEX_HELD(&clnt_pending_lock));
3435 	while (e->call_reply == NULL) {
3436 		if (waitp != NULL) {
3437 			timout = waitp->tv_sec * drv_usectohz(MICROSEC) +
3438 			    drv_usectohz(waitp->tv_usec) + lbolt;
3439 			if (nosignal)
3440 				cv_stat = cv_timedwait(&e->call_cv,
3441 				    &clnt_pending_lock, timout);
3442 			else
3443 				cv_stat = cv_timedwait_sig(&e->call_cv,
3444 				    &clnt_pending_lock, timout);
3445 		} else {
3446 			if (nosignal)
3447 				cv_wait(&e->call_cv, &clnt_pending_lock);
3448 			else
3449 				cv_stat = cv_wait_sig(&e->call_cv,
3450 				    &clnt_pending_lock);
3451 		}
3452 		if (cv_stat == -1)
3453 			return (ETIME);
3454 		if (cv_stat == 0)
3455 			return (EINTR);
3456 		/*
3457 		 * if we received an error from the server and we know a reply
3458 		 * is not going to be sent, do not wait for the full timeout,
3459 		 * return now.
3460 		 */
3461 		if (e->call_status == RPC_XPRTFAILED)
3462 			return (e->call_reason);
3463 	}
3464 	tpr = (union T_primitives *)e->call_reply->b_rptr;
3465 	if (tpr->type == ack_prim)
3466 		return (0); /* Success */
3467 
3468 	if (tpr->type == T_ERROR_ACK) {
3469 		if (tpr->error_ack.TLI_error == TSYSERR)
3470 			return (tpr->error_ack.UNIX_error);
3471 		else
3472 			return (t_tlitosyserr(tpr->error_ack.TLI_error));
3473 	}
3474 
3475 	return (EPROTO); /* unknown or unexpected primitive */
3476 }
3477