xref: /illumos-gate/usr/src/uts/common/rpc/clnt_cots.c (revision 9b9d39d2a32ff806d2431dbcc50968ef1e6d46b2)
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 /*
23  * Copyright 2016 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 2016 by Delphix. All rights reserved.
25  * Copyright 2019 Joyent, Inc.
26  */
27 
28 /*
29  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
30  * Use is subject to license terms.
31  */
32 
33 /*
34  * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T
35  *		All Rights Reserved
36  */
37 
38 /*
39  * Portions of this source code were derived from Berkeley 4.3 BSD
40  * under license from the Regents of the University of California.
41  */
42 
43 
44 /*
45  * Implements a kernel based, client side RPC over Connection Oriented
46  * Transports (COTS).
47  */
48 
49 /*
50  * Much of this file has been re-written to let NFS work better over slow
51  * transports. A description follows.
52  *
53  * One of the annoying things about kRPC/COTS is that it will temporarily
54  * create more than one connection between a client and server. This
55  * happens because when a connection is made, the end-points entry in the
56  * linked list of connections (headed by cm_hd), is removed so that other
57  * threads don't mess with it. Went ahead and bit the bullet by keeping
58  * the endpoint on the connection list and introducing state bits,
59  * condition variables etc. to the connection entry data structure (struct
60  * cm_xprt).
61  *
62  * Here is a summary of the changes to cm-xprt:
63  *
64  *	x_ctime is the timestamp of when the endpoint was last
65  *	connected or disconnected. If an end-point is ever disconnected
66  *	or re-connected, then any outstanding RPC request is presumed
67  *	lost, telling clnt_cots_kcallit that it needs to re-send the
68  *	request, not just wait for the original request's reply to
69  *	arrive.
70  *
71  *	x_thread flag which tells us if a thread is doing a connection attempt.
72  *
73  *	x_waitdis flag which tells us we are waiting a disconnect ACK.
74  *
75  *	x_needdis flag which tells us we need to send a T_DISCONN_REQ
76  *	to kill the connection.
77  *
78  *	x_needrel flag which tells us we need to send a T_ORDREL_REQ to
79  *	gracefully close the connection.
80  *
81  *	#defined bitmasks for the all the b_* bits so that more
82  *	efficient (and at times less clumsy) masks can be used to
83  *	manipulated state in cases where multiple bits have to
84  *	set/cleared/checked in the same critical section.
85  *
86  *	x_conn_cv and x_dis-_cv are new condition variables to let
87  *	threads knows when the connection attempt is done, and to let
88  *	the connecting thread know when the disconnect handshake is
89  *	done.
90  *
91  * Added the CONN_HOLD() macro so that all reference holds have the same
92  * look and feel.
93  *
94  * In the private (cku_private) portion of the client handle,
95  *
96  *	cku_flags replaces the cku_sent a boolean. cku_flags keeps
97  *	track of whether a request as been sent, and whether the
98  *	client's handles call record is on the dispatch list (so that
99  *	the reply can be matched by XID to the right client handle).
100  *	The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit()
101  *	and still have the response find the right client handle so
102  *	that the retry of CLNT_CALL() gets the result. Testing, found
103  *	situations where if the timeout was increased, performance
104  *	degraded. This was due to us hitting a window where the thread
105  *	was back in rfscall() (probably printing server not responding)
106  *	while the response came back but no place to put it.
107  *
108  *	cku_ctime is just a cache of x_ctime. If they match,
109  *	clnt_cots_kcallit() won't to send a retry (unless the maximum
110  *	receive count limit as been reached). If the don't match, then
111  *	we assume the request has been lost, and a retry of the request
112  *	is needed.
113  *
114  *	cku_recv_attempts counts the number of receive count attempts
115  *	after one try is sent on the wire.
116  *
117  * Added the clnt_delay() routine so that interruptible and
118  * noninterruptible delays are possible.
119  *
120  * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to
121  * control how long the client delays before returned after getting
122  * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash
123  * a server that may be booting and not yet started nfsd.
124  *
125  * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable)
126  * Why don't we just wait forever (receive an infinite # of times)?
127  * Because the server may have rebooted. More insidious is that some
128  * servers (ours) will drop NFS/TCP requests in some cases. This is bad,
129  * but it is a reality.
130  *
131  * The case of a server doing orderly release really messes up the
132  * client's recovery, especially if the server's TCP implementation is
133  * buggy.  It was found was that the kRPC/COTS client was breaking some
134  * TPI rules, such as not waiting for the acknowledgement of a
135  * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and
136  * T_DISCON_REQ in clnt_dispatch_notifyall()).
137  *
138  * One of things that we've seen is that a kRPC TCP endpoint goes into
139  * TIMEWAIT and a thus a reconnect takes a long time to satisfy because
140  * that the TIMEWAIT state takes a while to finish.  If a server sends a
141  * T_ORDREL_IND, there is little point in an RPC client doing a
142  * T_ORDREL_REQ, because the RPC request isn't going to make it (the
143  * server is saying that it won't accept any more data). So kRPC was
144  * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the
145  * connection skips the TIMEWAIT state and goes straight to a bound state
146  * that kRPC can quickly switch to connected.
147  *
148  * Code that issues TPI request must use waitforack() to wait for the
149  * corresponding ack (assuming there is one) in any future modifications.
150  * This works around problems that may be introduced by breaking TPI rules
151  * (by submitting new calls before earlier requests have been acked) in the
152  * case of a signal or other early return.  waitforack() depends on
153  * clnt_dispatch_notifyconn() to issue the wakeup when the ack
154  * arrives, so adding new TPI calls may require corresponding changes
155  * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on
156  * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure
157  * not to set it too low or TPI ACKS will be lost.
158  */
159 
160 #include <sys/param.h>
161 #include <sys/types.h>
162 #include <sys/user.h>
163 #include <sys/systm.h>
164 #include <sys/sysmacros.h>
165 #include <sys/proc.h>
166 #include <sys/socket.h>
167 #include <sys/file.h>
168 #include <sys/stream.h>
169 #include <sys/strsubr.h>
170 #include <sys/stropts.h>
171 #include <sys/strsun.h>
172 #include <sys/timod.h>
173 #include <sys/tiuser.h>
174 #include <sys/tihdr.h>
175 #include <sys/t_kuser.h>
176 #include <sys/fcntl.h>
177 #include <sys/errno.h>
178 #include <sys/kmem.h>
179 #include <sys/debug.h>
180 #include <sys/systm.h>
181 #include <sys/kstat.h>
182 #include <sys/t_lock.h>
183 #include <sys/ddi.h>
184 #include <sys/cmn_err.h>
185 #include <sys/time.h>
186 #include <sys/isa_defs.h>
187 #include <sys/callb.h>
188 #include <sys/sunddi.h>
189 #include <sys/atomic.h>
190 #include <sys/sdt.h>
191 
192 #include <netinet/in.h>
193 #include <netinet/tcp.h>
194 
195 #include <rpc/types.h>
196 #include <rpc/xdr.h>
197 #include <rpc/auth.h>
198 #include <rpc/clnt.h>
199 #include <rpc/rpc_msg.h>
200 
201 #define	COTS_DEFAULT_ALLOCSIZE	2048
202 
203 #define	WIRE_HDR_SIZE	20	/* serialized call header, sans proc number */
204 #define	MSG_OFFSET	128	/* offset of call into the mblk */
205 
206 const char *kinet_ntop6(uchar_t *, char *, size_t);
207 
208 static int	clnt_cots_ksettimers(CLIENT *, struct rpc_timers *,
209     struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t);
210 static enum clnt_stat	clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t,
211     caddr_t, xdrproc_t, caddr_t, struct timeval);
212 static void	clnt_cots_kabort(CLIENT *);
213 static void	clnt_cots_kerror(CLIENT *, struct rpc_err *);
214 static bool_t	clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t);
215 static void	clnt_cots_kdestroy(CLIENT *);
216 static bool_t	clnt_cots_kcontrol(CLIENT *, int, char *);
217 
218 
219 /* List of transports managed by the connection manager. */
220 struct cm_xprt {
221 	TIUSER		*x_tiptr;	/* transport handle */
222 	queue_t		*x_wq;		/* send queue */
223 	clock_t		x_time;		/* last time we handed this xprt out */
224 	clock_t		x_ctime;	/* time we went to CONNECTED */
225 	int		x_tidu_size;    /* TIDU size of this transport */
226 	union {
227 	    struct {
228 		unsigned int
229 #ifdef	_BIT_FIELDS_HTOL
230 		b_closing:	1,	/* we've sent a ord rel on this conn */
231 		b_dead:		1,	/* transport is closed or disconn */
232 		b_doomed:	1,	/* too many conns, let this go idle */
233 		b_connected:	1,	/* this connection is connected */
234 
235 		b_ordrel:	1,	/* do an orderly release? */
236 		b_thread:	1,	/* thread doing connect */
237 		b_waitdis:	1,	/* waiting for disconnect ACK */
238 		b_needdis:	1,	/* need T_DISCON_REQ */
239 
240 		b_needrel:	1,	/* need T_ORDREL_REQ */
241 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
242 					/* disconnect during connect */
243 
244 		b_pad:		22;
245 
246 #endif
247 
248 #ifdef	_BIT_FIELDS_LTOH
249 		b_pad:		22,
250 
251 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
252 					/* disconnect during connect */
253 		b_needrel:	1,	/* need T_ORDREL_REQ */
254 
255 		b_needdis:	1,	/* need T_DISCON_REQ */
256 		b_waitdis:	1,	/* waiting for disconnect ACK */
257 		b_thread:	1,	/* thread doing connect */
258 		b_ordrel:	1,	/* do an orderly release? */
259 
260 		b_connected:	1,	/* this connection is connected */
261 		b_doomed:	1,	/* too many conns, let this go idle */
262 		b_dead:		1,	/* transport is closed or disconn */
263 		b_closing:	1;	/* we've sent a ord rel on this conn */
264 #endif
265 	    } bit;	    unsigned int word;
266 
267 #define	x_closing	x_state.bit.b_closing
268 #define	x_dead		x_state.bit.b_dead
269 #define	x_doomed	x_state.bit.b_doomed
270 #define	x_connected	x_state.bit.b_connected
271 
272 #define	x_ordrel	x_state.bit.b_ordrel
273 #define	x_thread	x_state.bit.b_thread
274 #define	x_waitdis	x_state.bit.b_waitdis
275 #define	x_needdis	x_state.bit.b_needdis
276 
277 #define	x_needrel	x_state.bit.b_needrel
278 #define	x_early_disc    x_state.bit.b_early_disc
279 
280 #define	x_state_flags	x_state.word
281 
282 #define	X_CLOSING	0x80000000
283 #define	X_DEAD		0x40000000
284 #define	X_DOOMED	0x20000000
285 #define	X_CONNECTED	0x10000000
286 
287 #define	X_ORDREL	0x08000000
288 #define	X_THREAD	0x04000000
289 #define	X_WAITDIS	0x02000000
290 #define	X_NEEDDIS	0x01000000
291 
292 #define	X_NEEDREL	0x00800000
293 #define	X_EARLYDISC	0x00400000
294 
295 #define	X_BADSTATES	(X_CLOSING | X_DEAD | X_DOOMED)
296 
297 	}		x_state;
298 	int		x_ref;		/* number of users of this xprt */
299 	int		x_family;	/* address family of transport */
300 	dev_t		x_rdev;		/* device number of transport */
301 	struct cm_xprt	*x_next;
302 
303 	struct netbuf	x_server;	/* destination address */
304 	struct netbuf	x_src;		/* src address (for retries) */
305 	kmutex_t	x_lock;		/* lock on this entry */
306 	kcondvar_t	x_cv;		/* to signal when can be closed */
307 	kcondvar_t	x_conn_cv;	/* to signal when connection attempt */
308 					/* is complete */
309 	kstat_t		*x_ksp;
310 
311 	kcondvar_t	x_dis_cv;	/* to signal when disconnect attempt */
312 					/* is complete */
313 	zoneid_t	x_zoneid;	/* zone this xprt belongs to */
314 };
315 
316 typedef struct cm_kstat_xprt {
317 	kstat_named_t	x_wq;
318 	kstat_named_t	x_server;
319 	kstat_named_t	x_family;
320 	kstat_named_t	x_rdev;
321 	kstat_named_t	x_time;
322 	kstat_named_t	x_state;
323 	kstat_named_t	x_ref;
324 	kstat_named_t	x_port;
325 } cm_kstat_xprt_t;
326 
327 static cm_kstat_xprt_t cm_kstat_template = {
328 	{ "write_queue", KSTAT_DATA_UINT32 },
329 	{ "server",	KSTAT_DATA_STRING },
330 	{ "addr_family", KSTAT_DATA_UINT32 },
331 	{ "device",	KSTAT_DATA_UINT32 },
332 	{ "time_stamp",	KSTAT_DATA_UINT32 },
333 	{ "status",	KSTAT_DATA_UINT32 },
334 	{ "ref_count",	KSTAT_DATA_INT32 },
335 	{ "port",	KSTAT_DATA_UINT32 },
336 };
337 
338 /*
339  * The inverse of this is connmgr_release().
340  */
341 #define	CONN_HOLD(Cm_entry)	{\
342 	mutex_enter(&(Cm_entry)->x_lock);	\
343 	(Cm_entry)->x_ref++;	\
344 	mutex_exit(&(Cm_entry)->x_lock);	\
345 }
346 
347 
348 /*
349  * Private data per rpc handle.  This structure is allocated by
350  * clnt_cots_kcreate, and freed by clnt_cots_kdestroy.
351  */
352 typedef struct cku_private_s {
353 	CLIENT			cku_client;	/* client handle */
354 	calllist_t		cku_call;	/* for dispatching calls */
355 	struct rpc_err		cku_err;	/* error status */
356 
357 	struct netbuf		cku_srcaddr;	/* source address for retries */
358 	int			cku_addrfmly;  /* for binding port */
359 	struct netbuf		cku_addr;	/* remote address */
360 	dev_t			cku_device;	/* device to use */
361 	uint_t			cku_flags;
362 #define	CKU_ONQUEUE		0x1
363 #define	CKU_SENT		0x2
364 
365 	bool_t			cku_progress;	/* for CLSET_PROGRESS */
366 	uint32_t		cku_xid;	/* current XID */
367 	clock_t			cku_ctime;	/* time stamp of when */
368 						/* connection was created */
369 	uint_t			cku_recv_attempts;
370 	XDR			cku_outxdr;	/* xdr routine for output */
371 	XDR			cku_inxdr;	/* xdr routine for input */
372 	char			cku_rpchdr[WIRE_HDR_SIZE + 4];
373 						/* pre-serialized rpc header */
374 
375 	uint_t			cku_outbuflen;	/* default output mblk length */
376 	struct cred		*cku_cred;	/* credentials */
377 	bool_t			cku_nodelayonerr;
378 						/* for CLSET_NODELAYONERR */
379 	int			cku_useresvport; /* Use reserved port */
380 	struct rpc_cots_client	*cku_stats;	/* stats for zone */
381 } cku_private_t;
382 
383 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *,
384 	const struct timeval *, struct netbuf *, int, struct netbuf *,
385 	struct rpc_err *, bool_t, bool_t, cred_t *);
386 
387 static bool_t	connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *,
388 				int, calllist_t *, int *, bool_t reconnect,
389 				const struct timeval *, bool_t, cred_t *);
390 
391 static void	*connmgr_opt_getoff(mblk_t *mp, t_uscalar_t offset,
392 				t_uscalar_t length, uint_t align_size);
393 static bool_t	connmgr_setbufsz(calllist_t *e, queue_t *wq, cred_t *cr);
394 static bool_t	connmgr_getopt_int(queue_t *wq, int level, int name, int *val,
395 				calllist_t *e, cred_t *cr);
396 static bool_t	connmgr_setopt_int(queue_t *wq, int level, int name, int val,
397 				calllist_t *e, cred_t *cr);
398 static bool_t	connmgr_setopt(queue_t *, int, int, calllist_t *, cred_t *cr);
399 static void	connmgr_sndrel(struct cm_xprt *);
400 static void	connmgr_snddis(struct cm_xprt *);
401 static void	connmgr_close(struct cm_xprt *);
402 static void	connmgr_release(struct cm_xprt *);
403 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *,
404 	cku_private_t *);
405 
406 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *,
407 	struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t,
408 	bool_t, int, cred_t *);
409 
410 static void connmgr_cancelconn(struct cm_xprt *);
411 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *,
412 	bool_t);
413 static void connmgr_dis_and_wait(struct cm_xprt *);
414 
415 static int	clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t,
416 					uint_t);
417 
418 static int clnt_delay(clock_t, bool_t);
419 
420 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t);
421 
422 /*
423  * Operations vector for TCP/IP based RPC
424  */
425 static struct clnt_ops tcp_ops = {
426 	clnt_cots_kcallit,	/* do rpc call */
427 	clnt_cots_kabort,	/* abort call */
428 	clnt_cots_kerror,	/* return error status */
429 	clnt_cots_kfreeres,	/* free results */
430 	clnt_cots_kdestroy,	/* destroy rpc handle */
431 	clnt_cots_kcontrol,	/* the ioctl() of rpc */
432 	clnt_cots_ksettimers,	/* set retry timers */
433 };
434 
435 static int rpc_kstat_instance = 0;  /* keeps the current instance */
436 				/* number for the next kstat_create */
437 
438 static struct cm_xprt *cm_hd = NULL;
439 static kmutex_t connmgr_lock;	/* for connection mngr's list of transports */
440 
441 extern kmutex_t clnt_max_msg_lock;
442 
443 static calllist_t *clnt_pending = NULL;
444 extern kmutex_t clnt_pending_lock;
445 
446 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE;
447 
448 static call_table_t *cots_call_ht;
449 
450 static const struct rpc_cots_client {
451 	kstat_named_t	rccalls;
452 	kstat_named_t	rcbadcalls;
453 	kstat_named_t	rcbadxids;
454 	kstat_named_t	rctimeouts;
455 	kstat_named_t	rcnewcreds;
456 	kstat_named_t	rcbadverfs;
457 	kstat_named_t	rctimers;
458 	kstat_named_t	rccantconn;
459 	kstat_named_t	rcnomem;
460 	kstat_named_t	rcintrs;
461 } cots_rcstat_tmpl = {
462 	{ "calls",	KSTAT_DATA_UINT64 },
463 	{ "badcalls",	KSTAT_DATA_UINT64 },
464 	{ "badxids",	KSTAT_DATA_UINT64 },
465 	{ "timeouts",	KSTAT_DATA_UINT64 },
466 	{ "newcreds",	KSTAT_DATA_UINT64 },
467 	{ "badverfs",	KSTAT_DATA_UINT64 },
468 	{ "timers",	KSTAT_DATA_UINT64 },
469 	{ "cantconn",	KSTAT_DATA_UINT64 },
470 	{ "nomem",	KSTAT_DATA_UINT64 },
471 	{ "interrupts", KSTAT_DATA_UINT64 }
472 };
473 
474 #define	COTSRCSTAT_INCR(p, x)	\
475 	atomic_inc_64(&(p)->x.value.ui64)
476 
477 #define	CLNT_MAX_CONNS	1	/* concurrent connections between clnt/srvr */
478 int clnt_max_conns = CLNT_MAX_CONNS;
479 
480 #define	CLNT_MIN_TIMEOUT	10	/* seconds to wait after we get a */
481 					/* connection reset */
482 #define	CLNT_MIN_CONNTIMEOUT	5	/* seconds to wait for a connection */
483 
484 
485 int clnt_cots_min_tout = CLNT_MIN_TIMEOUT;
486 int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT;
487 
488 /*
489  * Limit the number of times we will attempt to receive a reply without
490  * re-sending a response.
491  */
492 #define	CLNT_MAXRECV_WITHOUT_RETRY	3
493 uint_t clnt_cots_maxrecv	= CLNT_MAXRECV_WITHOUT_RETRY;
494 
495 uint_t *clnt_max_msg_sizep;
496 void (*clnt_stop_idle)(queue_t *wq);
497 
498 #define	ptoh(p)		(&((p)->cku_client))
499 #define	htop(h)		((cku_private_t *)((h)->cl_private))
500 
501 /*
502  * Times to retry
503  */
504 #define	REFRESHES	2	/* authentication refreshes */
505 
506 /*
507  * The following is used to determine the global default behavior for
508  * COTS when binding to a local port.
509  *
510  * If the value is set to 1 the default will be to select a reserved
511  * (aka privileged) port, if the value is zero the default will be to
512  * use non-reserved ports.  Users of kRPC may override this by using
513  * CLNT_CONTROL() and CLSET_BINDRESVPORT.
514  */
515 int clnt_cots_do_bindresvport = 1;
516 
517 static zone_key_t zone_cots_key;
518 
519 /*
520  * Defaults TCP send and receive buffer size for RPC connections.
521  * These values can be tuned by /etc/system.
522  */
523 int rpc_send_bufsz = 1024*1024;
524 int rpc_recv_bufsz = 1024*1024;
525 /*
526  * To use system-wide default for TCP send and receive buffer size,
527  * use /etc/system to set rpc_default_tcp_bufsz to 1:
528  *
529  * set rpcmod:rpc_default_tcp_bufsz=1
530  */
531 int rpc_default_tcp_bufsz = 0;
532 
533 /*
534  * We need to do this after all kernel threads in the zone have exited.
535  */
536 /* ARGSUSED */
537 static void
538 clnt_zone_destroy(zoneid_t zoneid, void *unused)
539 {
540 	struct cm_xprt **cmp;
541 	struct cm_xprt *cm_entry;
542 	struct cm_xprt *freelist = NULL;
543 
544 	mutex_enter(&connmgr_lock);
545 	cmp = &cm_hd;
546 	while ((cm_entry = *cmp) != NULL) {
547 		if (cm_entry->x_zoneid == zoneid) {
548 			*cmp = cm_entry->x_next;
549 			cm_entry->x_next = freelist;
550 			freelist = cm_entry;
551 		} else {
552 			cmp = &cm_entry->x_next;
553 		}
554 	}
555 	mutex_exit(&connmgr_lock);
556 	while ((cm_entry = freelist) != NULL) {
557 		freelist = cm_entry->x_next;
558 		connmgr_close(cm_entry);
559 	}
560 }
561 
562 int
563 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog,
564     rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl)
565 {
566 	CLIENT *h;
567 	cku_private_t *p;
568 	struct rpc_msg call_msg;
569 	struct rpcstat *rpcstat;
570 
571 	RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog);
572 
573 	rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
574 	ASSERT(rpcstat != NULL);
575 
576 	/* Allocate and intialize the client handle. */
577 	p = kmem_zalloc(sizeof (*p), KM_SLEEP);
578 
579 	h = ptoh(p);
580 
581 	h->cl_private = (caddr_t)p;
582 	h->cl_auth = authkern_create();
583 	h->cl_ops = &tcp_ops;
584 
585 	cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL);
586 	mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL);
587 
588 	/*
589 	 * If the current sanity check size in rpcmod is smaller
590 	 * than the size needed, then increase the sanity check.
591 	 */
592 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
593 	    max_msgsize > *clnt_max_msg_sizep) {
594 		mutex_enter(&clnt_max_msg_lock);
595 		if (max_msgsize > *clnt_max_msg_sizep)
596 			*clnt_max_msg_sizep = max_msgsize;
597 		mutex_exit(&clnt_max_msg_lock);
598 	}
599 
600 	p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE;
601 
602 	/* Preserialize the call message header */
603 
604 	call_msg.rm_xid = 0;
605 	call_msg.rm_direction = CALL;
606 	call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION;
607 	call_msg.rm_call.cb_prog = prog;
608 	call_msg.rm_call.cb_vers = vers;
609 
610 	xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE);
611 
612 	if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) {
613 		XDR_DESTROY(&p->cku_outxdr);
614 		RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization "
615 		    "error\n");
616 		auth_destroy(h->cl_auth);
617 		kmem_free(p, sizeof (cku_private_t));
618 		RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n");
619 		return (EINVAL);		/* XXX */
620 	}
621 	XDR_DESTROY(&p->cku_outxdr);
622 
623 	/*
624 	 * The zalloc initialized the fields below.
625 	 * p->cku_xid = 0;
626 	 * p->cku_flags = 0;
627 	 * p->cku_srcaddr.buf = NULL;
628 	 * p->cku_srcaddr.len = 0;
629 	 * p->cku_srcaddr.maxlen = 0;
630 	 */
631 
632 	p->cku_cred = cred;
633 	p->cku_device = dev;
634 	p->cku_addrfmly = family;
635 	p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
636 	p->cku_addr.maxlen = addr->maxlen;
637 	p->cku_addr.len = addr->len;
638 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
639 	p->cku_stats = rpcstat->rpc_cots_client;
640 	p->cku_useresvport = -1; /* value is has not been set */
641 
642 	*ncl = h;
643 	return (0);
644 }
645 
646 /*ARGSUSED*/
647 static void
648 clnt_cots_kabort(CLIENT *h)
649 {
650 }
651 
652 /*
653  * Return error info on this handle.
654  */
655 static void
656 clnt_cots_kerror(CLIENT *h, struct rpc_err *err)
657 {
658 	/* LINTED pointer alignment */
659 	cku_private_t *p = htop(h);
660 
661 	*err = p->cku_err;
662 }
663 
664 /*ARGSUSED*/
665 static bool_t
666 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr)
667 {
668 	xdr_free(xdr_res, res_ptr);
669 
670 	return (TRUE);
671 }
672 
673 static bool_t
674 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg)
675 {
676 	cku_private_t *p = htop(h);
677 
678 	switch (cmd) {
679 	case CLSET_PROGRESS:
680 		p->cku_progress = TRUE;
681 		return (TRUE);
682 
683 	case CLSET_XID:
684 		if (arg == NULL)
685 			return (FALSE);
686 
687 		p->cku_xid = *((uint32_t *)arg);
688 		return (TRUE);
689 
690 	case CLGET_XID:
691 		if (arg == NULL)
692 			return (FALSE);
693 
694 		*((uint32_t *)arg) = p->cku_xid;
695 		return (TRUE);
696 
697 	case CLSET_NODELAYONERR:
698 		if (arg == NULL)
699 			return (FALSE);
700 
701 		if (*((bool_t *)arg) == TRUE) {
702 			p->cku_nodelayonerr = TRUE;
703 			return (TRUE);
704 		}
705 		if (*((bool_t *)arg) == FALSE) {
706 			p->cku_nodelayonerr = FALSE;
707 			return (TRUE);
708 		}
709 		return (FALSE);
710 
711 	case CLGET_NODELAYONERR:
712 		if (arg == NULL)
713 			return (FALSE);
714 
715 		*((bool_t *)arg) = p->cku_nodelayonerr;
716 		return (TRUE);
717 
718 	case CLSET_BINDRESVPORT:
719 		if (arg == NULL)
720 			return (FALSE);
721 
722 		if (*(int *)arg != 1 && *(int *)arg != 0)
723 			return (FALSE);
724 
725 		p->cku_useresvport = *(int *)arg;
726 
727 		return (TRUE);
728 
729 	case CLGET_BINDRESVPORT:
730 		if (arg == NULL)
731 			return (FALSE);
732 
733 		*(int *)arg = p->cku_useresvport;
734 
735 		return (TRUE);
736 
737 	default:
738 		return (FALSE);
739 	}
740 }
741 
742 /*
743  * Destroy rpc handle.  Frees the space used for output buffer,
744  * private data, and handle structure.
745  */
746 static void
747 clnt_cots_kdestroy(CLIENT *h)
748 {
749 	/* LINTED pointer alignment */
750 	cku_private_t *p = htop(h);
751 	calllist_t *call = &p->cku_call;
752 
753 	RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h);
754 	RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid);
755 
756 	if (p->cku_flags & CKU_ONQUEUE) {
757 		RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x "
758 		    "from dispatch list\n", p->cku_xid);
759 		call_table_remove(call);
760 	}
761 
762 	if (call->call_reply)
763 		freemsg(call->call_reply);
764 	cv_destroy(&call->call_cv);
765 	mutex_destroy(&call->call_lock);
766 
767 	kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen);
768 	kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
769 	kmem_free(p, sizeof (*p));
770 }
771 
772 static int clnt_cots_pulls;
773 #define	RM_HDR_SIZE	4	/* record mark header size */
774 
775 /*
776  * Call remote procedure.
777  */
778 static enum clnt_stat
779 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
780     caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait)
781 {
782 	/* LINTED pointer alignment */
783 	cku_private_t *p = htop(h);
784 	calllist_t *call = &p->cku_call;
785 	XDR *xdrs;
786 	struct rpc_msg reply_msg;
787 	mblk_t *mp;
788 #ifdef	RPCDEBUG
789 	clock_t time_sent;
790 #endif
791 	struct netbuf *retryaddr;
792 	struct cm_xprt *cm_entry = NULL;
793 	queue_t *wq;
794 	int len, waitsecs, max_waitsecs;
795 	int mpsize;
796 	int refreshes = REFRESHES;
797 	int interrupted;
798 	int tidu_size;
799 	enum clnt_stat status;
800 	struct timeval cwait;
801 	bool_t delay_first = FALSE;
802 	clock_t ticks, now;
803 
804 	RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum);
805 	COTSRCSTAT_INCR(p->cku_stats, rccalls);
806 
807 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec);
808 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec);
809 	/*
810 	 * Bug ID 1240234:
811 	 * Look out for zero length timeouts. We don't want to
812 	 * wait zero seconds for a connection to be established.
813 	 */
814 	if (wait.tv_sec < clnt_cots_min_conntout) {
815 		cwait.tv_sec = clnt_cots_min_conntout;
816 		cwait.tv_usec = 0;
817 		RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,",
818 		    wait.tv_sec);
819 		RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout);
820 	} else {
821 		cwait = wait;
822 	}
823 
824 call_again:
825 	if (cm_entry) {
826 		connmgr_release(cm_entry);
827 		cm_entry = NULL;
828 	}
829 
830 	mp = NULL;
831 
832 	/*
833 	 * If the call is not a retry, allocate a new xid and cache it
834 	 * for future retries.
835 	 * Bug ID 1246045:
836 	 * Treat call as a retry for purposes of binding the source
837 	 * port only if we actually attempted to send anything on
838 	 * the previous call.
839 	 */
840 	if (p->cku_xid == 0) {
841 		p->cku_xid = alloc_xid();
842 		call->call_zoneid = rpc_zoneid();
843 
844 		/*
845 		 * We need to ASSERT here that our xid != 0 because this
846 		 * determines whether or not our call record gets placed on
847 		 * the hash table or the linked list.  By design, we mandate
848 		 * that RPC calls over cots must have xid's != 0, so we can
849 		 * ensure proper management of the hash table.
850 		 */
851 		ASSERT(p->cku_xid != 0);
852 
853 		retryaddr = NULL;
854 		p->cku_flags &= ~CKU_SENT;
855 
856 		if (p->cku_flags & CKU_ONQUEUE) {
857 			RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old"
858 			    " one (%p)\n", (void *)call);
859 			call_table_remove(call);
860 			p->cku_flags &= ~CKU_ONQUEUE;
861 			RPCLOG(64, "clnt_cots_kcallit: removing call from "
862 			    "dispatch list because xid was zero (now 0x%x)\n",
863 			    p->cku_xid);
864 		}
865 
866 		if (call->call_reply != NULL) {
867 			freemsg(call->call_reply);
868 			call->call_reply = NULL;
869 		}
870 	} else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) {
871 		retryaddr = NULL;
872 
873 	} else if (p->cku_flags & CKU_SENT) {
874 		retryaddr = &p->cku_srcaddr;
875 
876 	} else {
877 		/*
878 		 * Bug ID 1246045: Nothing was sent, so set retryaddr to
879 		 * NULL and let connmgr_get() bind to any source port it
880 		 * can get.
881 		 */
882 		retryaddr = NULL;
883 	}
884 
885 	RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid);
886 	RPCLOG(64, " flags = 0x%x\n", p->cku_flags);
887 
888 	p->cku_err.re_status = RPC_TIMEDOUT;
889 	p->cku_err.re_errno = p->cku_err.re_terrno = 0;
890 
891 	cm_entry = connmgr_wrapget(retryaddr, &cwait, p);
892 
893 	if (cm_entry == NULL) {
894 		RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n",
895 		    clnt_sperrno(p->cku_err.re_status));
896 
897 		/*
898 		 * The reasons why we fail to create a connection are
899 		 * varied. In most cases we don't want the caller to
900 		 * immediately retry. This could have one or more
901 		 * bad effects. This includes flooding the net with
902 		 * connect requests to ports with no listener; a hard
903 		 * kernel loop due to all the "reserved" TCP ports being
904 		 * in use.
905 		 */
906 		delay_first = TRUE;
907 
908 		/*
909 		 * Even if we end up returning EINTR, we still count a
910 		 * a "can't connect", because the connection manager
911 		 * might have been committed to waiting for or timing out on
912 		 * a connection.
913 		 */
914 		COTSRCSTAT_INCR(p->cku_stats, rccantconn);
915 		switch (p->cku_err.re_status) {
916 		case RPC_INTR:
917 			p->cku_err.re_errno = EINTR;
918 
919 			/*
920 			 * No need to delay because a UNIX signal(2)
921 			 * interrupted us. The caller likely won't
922 			 * retry the CLNT_CALL() and even if it does,
923 			 * we assume the caller knows what it is doing.
924 			 */
925 			delay_first = FALSE;
926 			break;
927 
928 		case RPC_TIMEDOUT:
929 			p->cku_err.re_errno = ETIMEDOUT;
930 
931 			/*
932 			 * No need to delay because timed out already
933 			 * on the connection request and assume that the
934 			 * transport time out is longer than our minimum
935 			 * timeout, or least not too much smaller.
936 			 */
937 			delay_first = FALSE;
938 			break;
939 
940 		case RPC_SYSTEMERROR:
941 		case RPC_TLIERROR:
942 			/*
943 			 * We want to delay here because a transient
944 			 * system error has a better chance of going away
945 			 * if we delay a bit. If it's not transient, then
946 			 * we don't want end up in a hard kernel loop
947 			 * due to retries.
948 			 */
949 			ASSERT(p->cku_err.re_errno != 0);
950 			break;
951 
952 
953 		case RPC_CANTCONNECT:
954 			/*
955 			 * RPC_CANTCONNECT is set on T_ERROR_ACK which
956 			 * implies some error down in the TCP layer or
957 			 * below. If cku_nodelayonerror is set then we
958 			 * assume the caller knows not to try too hard.
959 			 */
960 			RPCLOG0(8, "clnt_cots_kcallit: connection failed,");
961 			RPCLOG0(8, " re_status=RPC_CANTCONNECT,");
962 			RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno);
963 			RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr);
964 			if (p->cku_nodelayonerr == TRUE)
965 				delay_first = FALSE;
966 
967 			p->cku_err.re_errno = EIO;
968 
969 			break;
970 
971 		case RPC_XPRTFAILED:
972 			/*
973 			 * We want to delay here because we likely
974 			 * got a refused connection.
975 			 */
976 			if (p->cku_err.re_errno == 0)
977 				p->cku_err.re_errno = EIO;
978 
979 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
980 			    p->cku_err.re_errno);
981 
982 			break;
983 
984 		default:
985 			/*
986 			 * We delay here because it is better to err
987 			 * on the side of caution. If we got here then
988 			 * status could have been RPC_SUCCESS, but we
989 			 * know that we did not get a connection, so
990 			 * force the rpc status to RPC_CANTCONNECT.
991 			 */
992 			p->cku_err.re_status = RPC_CANTCONNECT;
993 			p->cku_err.re_errno = EIO;
994 			break;
995 		}
996 		if (delay_first == TRUE)
997 			ticks = clnt_cots_min_tout * drv_usectohz(1000000);
998 		goto cots_done;
999 	}
1000 
1001 	/*
1002 	 * If we've never sent any request on this connection (send count
1003 	 * is zero, or the connection has been reset), cache the
1004 	 * the connection's create time and send a request (possibly a retry)
1005 	 */
1006 	if ((p->cku_flags & CKU_SENT) == 0 ||
1007 	    p->cku_ctime != cm_entry->x_ctime) {
1008 		p->cku_ctime = cm_entry->x_ctime;
1009 
1010 	} else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) &&
1011 	    (call->call_reply != NULL ||
1012 	    p->cku_recv_attempts < clnt_cots_maxrecv)) {
1013 
1014 		/*
1015 		 * If we've sent a request and our call is on the dispatch
1016 		 * queue and we haven't made too many receive attempts, then
1017 		 * don't re-send, just receive.
1018 		 */
1019 		p->cku_recv_attempts++;
1020 		goto read_again;
1021 	}
1022 
1023 	/*
1024 	 * Now we create the RPC request in a STREAMS message.  We have to do
1025 	 * this after the call to connmgr_get so that we have the correct
1026 	 * TIDU size for the transport.
1027 	 */
1028 	tidu_size = cm_entry->x_tidu_size;
1029 	len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE);
1030 
1031 	while ((mp = allocb(len, BPRI_MED)) == NULL) {
1032 		if (strwaitbuf(len, BPRI_MED)) {
1033 			p->cku_err.re_status = RPC_SYSTEMERROR;
1034 			p->cku_err.re_errno = ENOSR;
1035 			COTSRCSTAT_INCR(p->cku_stats, rcnomem);
1036 			goto cots_done;
1037 		}
1038 	}
1039 	xdrs = &p->cku_outxdr;
1040 	xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size);
1041 	mpsize = MBLKSIZE(mp);
1042 	ASSERT(mpsize >= len);
1043 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
1044 
1045 	/*
1046 	 * If the size of mblk is not appreciably larger than what we
1047 	 * asked, then resize the mblk to exactly len bytes. The reason for
1048 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
1049 	 * (from TCP over ethernet), and the arguments to the RPC require
1050 	 * 2800 bytes. Ideally we want the protocol to render two
1051 	 * ~1400 byte segments over the wire. However if allocb() gives us a 2k
1052 	 * mblk, and we allocate a second mblk for the remainder, the protocol
1053 	 * module may generate 3 segments over the wire:
1054 	 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and
1055 	 * 892 for the third. If we "waste" 448 bytes in the first mblk,
1056 	 * the XDR encoding will generate two ~1400 byte mblks, and the
1057 	 * protocol module is more likely to produce properly sized segments.
1058 	 */
1059 	if ((mpsize >> 1) <= len)
1060 		mp->b_rptr += (mpsize - len);
1061 
1062 	/*
1063 	 * Adjust b_rptr to reserve space for the non-data protocol headers
1064 	 * any downstream modules might like to add, and for the
1065 	 * record marking header.
1066 	 */
1067 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
1068 
1069 	if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) {
1070 		/* Copy in the preserialized RPC header information. */
1071 		bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE);
1072 
1073 		/* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */
1074 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base +
1075 		    WIRE_HDR_SIZE));
1076 
1077 		ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE);
1078 
1079 		/* Serialize the procedure number and the arguments. */
1080 		if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) ||
1081 		    (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) ||
1082 		    (!(*xdr_args)(xdrs, argsp))) {
1083 			XDR_DESTROY(xdrs);
1084 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1085 			p->cku_err.re_errno = EIO;
1086 			goto cots_done;
1087 		}
1088 
1089 		(*(uint32_t *)(mp->b_rptr)) = p->cku_xid;
1090 	} else {
1091 		uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE];
1092 		IXDR_PUT_U_INT32(uproc, procnum);
1093 
1094 		(*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid;
1095 
1096 		/* Use XDR_SETPOS() to set the b_wptr. */
1097 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
1098 
1099 		/* Serialize the procedure number and the arguments. */
1100 		if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4,
1101 		    xdrs, xdr_args, argsp)) {
1102 			XDR_DESTROY(xdrs);
1103 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1104 			p->cku_err.re_errno = EIO;
1105 			goto cots_done;
1106 		}
1107 	}
1108 
1109 	XDR_DESTROY(xdrs);
1110 
1111 	RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n",
1112 	    tidu_size);
1113 
1114 	wq = cm_entry->x_wq;
1115 	waitsecs = 0;
1116 
1117 dispatch_again:
1118 	status = clnt_dispatch_send(wq, mp, call, p->cku_xid,
1119 	    (p->cku_flags & CKU_ONQUEUE));
1120 
1121 	if ((status == RPC_CANTSEND) && (call->call_reason == ENOBUFS)) {
1122 		/*
1123 		 * QFULL condition, allow some time for queue to drain
1124 		 * and try again. Give up after waiting for all timeout
1125 		 * specified for the call, or zone is going away.
1126 		 */
1127 		max_waitsecs = wait.tv_sec ? wait.tv_sec : clnt_cots_min_tout;
1128 		if ((waitsecs++ < max_waitsecs) &&
1129 		    !(zone_status_get(curproc->p_zone) >=
1130 		    ZONE_IS_SHUTTING_DOWN)) {
1131 
1132 			/* wait 1 sec for queue to drain */
1133 			if (clnt_delay(drv_usectohz(1000000),
1134 			    h->cl_nosignal) == EINTR) {
1135 				p->cku_err.re_errno = EINTR;
1136 				p->cku_err.re_status = RPC_INTR;
1137 
1138 				goto cots_done;
1139 			}
1140 
1141 			/* and try again */
1142 			goto dispatch_again;
1143 		}
1144 		p->cku_err.re_status = status;
1145 		p->cku_err.re_errno = call->call_reason;
1146 		DTRACE_PROBE(krpc__e__clntcots__kcallit__cantsend);
1147 
1148 		goto cots_done;
1149 	}
1150 
1151 	if (waitsecs) {
1152 		/* adjust timeout to account for time wait to send */
1153 		wait.tv_sec -= waitsecs;
1154 		if (wait.tv_sec < 0) {
1155 			/* pick up reply on next retry */
1156 			wait.tv_sec = 0;
1157 		}
1158 		DTRACE_PROBE2(clnt_cots__sendwait, CLIENT *, h,
1159 		    int, waitsecs);
1160 	}
1161 
1162 	RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n",
1163 	    (uint_t)p->cku_xid);
1164 	p->cku_flags = (CKU_ONQUEUE|CKU_SENT);
1165 	p->cku_recv_attempts = 1;
1166 
1167 #ifdef	RPCDEBUG
1168 	time_sent = ddi_get_lbolt();
1169 #endif
1170 
1171 	/*
1172 	 * Wait for a reply or a timeout.  If there is no error or timeout,
1173 	 * (both indicated by call_status), call->call_reply will contain
1174 	 * the RPC reply message.
1175 	 */
1176 read_again:
1177 	mutex_enter(&call->call_lock);
1178 	interrupted = 0;
1179 	if (call->call_status == RPC_TIMEDOUT) {
1180 		/*
1181 		 * Indicate that the lwp is not to be stopped while waiting
1182 		 * for this network traffic.  This is to avoid deadlock while
1183 		 * debugging a process via /proc and also to avoid recursive
1184 		 * mutex_enter()s due to NFS page faults while stopping
1185 		 * (NFS holds locks when it calls here).
1186 		 */
1187 		clock_t cv_wait_ret;
1188 		clock_t timout;
1189 		clock_t oldlbolt;
1190 
1191 		klwp_t *lwp = ttolwp(curthread);
1192 
1193 		if (lwp != NULL)
1194 			lwp->lwp_nostop++;
1195 
1196 		oldlbolt = ddi_get_lbolt();
1197 		timout = wait.tv_sec * drv_usectohz(1000000) +
1198 		    drv_usectohz(wait.tv_usec) + oldlbolt;
1199 		/*
1200 		 * Iterate until the call_status is changed to something
1201 		 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns
1202 		 * something <=0 zero. The latter means that we timed
1203 		 * out.
1204 		 */
1205 		if (h->cl_nosignal)
1206 			while ((cv_wait_ret = cv_timedwait(&call->call_cv,
1207 			    &call->call_lock, timout)) > 0 &&
1208 			    call->call_status == RPC_TIMEDOUT)
1209 				;
1210 		else
1211 			while ((cv_wait_ret = cv_timedwait_sig(
1212 			    &call->call_cv,
1213 			    &call->call_lock, timout)) > 0 &&
1214 			    call->call_status == RPC_TIMEDOUT)
1215 				;
1216 
1217 		switch (cv_wait_ret) {
1218 		case 0:
1219 			/*
1220 			 * If we got out of the above loop with
1221 			 * cv_timedwait_sig() returning 0, then we were
1222 			 * interrupted regardless what call_status is.
1223 			 */
1224 			interrupted = 1;
1225 			break;
1226 		case -1:
1227 			/* cv_timedwait_sig() timed out */
1228 			break;
1229 		default:
1230 
1231 			/*
1232 			 * We were cv_signaled(). If we didn't
1233 			 * get a successful call_status and returned
1234 			 * before time expired, delay up to clnt_cots_min_tout
1235 			 * seconds so that the caller doesn't immediately
1236 			 * try to call us again and thus force the
1237 			 * same condition that got us here (such
1238 			 * as a RPC_XPRTFAILED due to the server not
1239 			 * listening on the end-point.
1240 			 */
1241 			if (call->call_status != RPC_SUCCESS) {
1242 				clock_t curlbolt;
1243 				clock_t diff;
1244 
1245 				curlbolt = ddi_get_lbolt();
1246 				ticks = clnt_cots_min_tout *
1247 				    drv_usectohz(1000000);
1248 				diff = curlbolt - oldlbolt;
1249 				if (diff < ticks) {
1250 					delay_first = TRUE;
1251 					if (diff > 0)
1252 						ticks -= diff;
1253 				}
1254 			}
1255 			break;
1256 		}
1257 
1258 		if (lwp != NULL)
1259 			lwp->lwp_nostop--;
1260 	}
1261 	/*
1262 	 * Get the reply message, if any.  This will be freed at the end
1263 	 * whether or not an error occurred.
1264 	 */
1265 	mp = call->call_reply;
1266 	call->call_reply = NULL;
1267 
1268 	/*
1269 	 * call_err is the error info when the call is on dispatch queue.
1270 	 * cku_err is the error info returned to the caller.
1271 	 * Sync cku_err with call_err for local message processing.
1272 	 */
1273 
1274 	status = call->call_status;
1275 	p->cku_err = call->call_err;
1276 	mutex_exit(&call->call_lock);
1277 
1278 	if (status != RPC_SUCCESS) {
1279 		switch (status) {
1280 		case RPC_TIMEDOUT:
1281 			now = ddi_get_lbolt();
1282 			if (interrupted) {
1283 				COTSRCSTAT_INCR(p->cku_stats, rcintrs);
1284 				p->cku_err.re_status = RPC_INTR;
1285 				p->cku_err.re_errno = EINTR;
1286 				RPCLOG(1, "clnt_cots_kcallit: xid 0x%x",
1287 				    p->cku_xid);
1288 				RPCLOG(1, "signal interrupted at %ld", now);
1289 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1290 			} else {
1291 				COTSRCSTAT_INCR(p->cku_stats, rctimeouts);
1292 				p->cku_err.re_errno = ETIMEDOUT;
1293 				RPCLOG(1, "clnt_cots_kcallit: timed out at %ld",
1294 				    now);
1295 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1296 			}
1297 			break;
1298 
1299 		case RPC_XPRTFAILED:
1300 			if (p->cku_err.re_errno == 0)
1301 				p->cku_err.re_errno = EIO;
1302 
1303 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
1304 			    p->cku_err.re_errno);
1305 			break;
1306 
1307 		case RPC_SYSTEMERROR:
1308 			ASSERT(p->cku_err.re_errno);
1309 			RPCLOG(1, "clnt_cots_kcallit: system error: %d\n",
1310 			    p->cku_err.re_errno);
1311 			break;
1312 
1313 		default:
1314 			p->cku_err.re_status = RPC_SYSTEMERROR;
1315 			p->cku_err.re_errno = EIO;
1316 			RPCLOG(1, "clnt_cots_kcallit: error: %s\n",
1317 			    clnt_sperrno(status));
1318 			break;
1319 		}
1320 		if (p->cku_err.re_status != RPC_TIMEDOUT) {
1321 
1322 			if (p->cku_flags & CKU_ONQUEUE) {
1323 				call_table_remove(call);
1324 				p->cku_flags &= ~CKU_ONQUEUE;
1325 			}
1326 
1327 			RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x "
1328 			    "taken off dispatch list\n", p->cku_xid);
1329 			if (call->call_reply) {
1330 				freemsg(call->call_reply);
1331 				call->call_reply = NULL;
1332 			}
1333 		} else if (wait.tv_sec != 0) {
1334 			/*
1335 			 * We've sent the request over TCP and so we have
1336 			 * every reason to believe it will get
1337 			 * delivered. In which case returning a timeout is not
1338 			 * appropriate.
1339 			 */
1340 			if (p->cku_progress == TRUE &&
1341 			    p->cku_recv_attempts < clnt_cots_maxrecv) {
1342 				p->cku_err.re_status = RPC_INPROGRESS;
1343 			}
1344 		}
1345 		goto cots_done;
1346 	}
1347 
1348 	xdrs = &p->cku_inxdr;
1349 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
1350 
1351 	reply_msg.rm_direction = REPLY;
1352 	reply_msg.rm_reply.rp_stat = MSG_ACCEPTED;
1353 	reply_msg.acpted_rply.ar_stat = SUCCESS;
1354 
1355 	reply_msg.acpted_rply.ar_verf = _null_auth;
1356 	/*
1357 	 *  xdr_results will be done in AUTH_UNWRAP.
1358 	 */
1359 	reply_msg.acpted_rply.ar_results.where = NULL;
1360 	reply_msg.acpted_rply.ar_results.proc = xdr_void;
1361 
1362 	if (xdr_replymsg(xdrs, &reply_msg)) {
1363 		enum clnt_stat re_status;
1364 
1365 		_seterr_reply(&reply_msg, &p->cku_err);
1366 
1367 		re_status = p->cku_err.re_status;
1368 		if (re_status == RPC_SUCCESS) {
1369 			/*
1370 			 * Reply is good, check auth.
1371 			 */
1372 			if (!AUTH_VALIDATE(h->cl_auth,
1373 			    &reply_msg.acpted_rply.ar_verf)) {
1374 				COTSRCSTAT_INCR(p->cku_stats, rcbadverfs);
1375 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1376 				    "failure\n");
1377 				freemsg(mp);
1378 				(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1379 				XDR_DESTROY(xdrs);
1380 				mutex_enter(&call->call_lock);
1381 				if (call->call_reply == NULL)
1382 					call->call_status = RPC_TIMEDOUT;
1383 				mutex_exit(&call->call_lock);
1384 				goto read_again;
1385 			} else if (!AUTH_UNWRAP(h->cl_auth, xdrs,
1386 			    xdr_results, resultsp)) {
1387 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1388 				    "failure (unwrap)\n");
1389 				p->cku_err.re_status = RPC_CANTDECODERES;
1390 				p->cku_err.re_errno = EIO;
1391 			}
1392 		} else {
1393 			/* set errno in case we can't recover */
1394 			if (re_status != RPC_VERSMISMATCH &&
1395 			    re_status != RPC_AUTHERROR &&
1396 			    re_status != RPC_PROGVERSMISMATCH)
1397 				p->cku_err.re_errno = EIO;
1398 
1399 			if (re_status == RPC_AUTHERROR) {
1400 				/*
1401 				 * Maybe our credential need to be refreshed
1402 				 */
1403 				if (cm_entry) {
1404 					/*
1405 					 * There is the potential that the
1406 					 * cm_entry has/will be marked dead,
1407 					 * so drop the connection altogether,
1408 					 * force REFRESH to establish new
1409 					 * connection.
1410 					 */
1411 					connmgr_cancelconn(cm_entry);
1412 					cm_entry = NULL;
1413 				}
1414 
1415 				(void) xdr_rpc_free_verifier(xdrs,
1416 				    &reply_msg);
1417 				XDR_DESTROY(xdrs);
1418 
1419 				if (p->cku_flags & CKU_ONQUEUE) {
1420 					call_table_remove(call);
1421 					p->cku_flags &= ~CKU_ONQUEUE;
1422 				}
1423 				RPCLOG(64,
1424 				    "clnt_cots_kcallit: AUTH_ERROR, xid"
1425 				    " 0x%x removed off dispatch list\n",
1426 				    p->cku_xid);
1427 				if (call->call_reply) {
1428 					freemsg(call->call_reply);
1429 					call->call_reply = NULL;
1430 				}
1431 
1432 				if ((refreshes > 0) &&
1433 				    AUTH_REFRESH(h->cl_auth, &reply_msg,
1434 				    p->cku_cred)) {
1435 					refreshes--;
1436 					freemsg(mp);
1437 					mp = NULL;
1438 
1439 					COTSRCSTAT_INCR(p->cku_stats,
1440 					    rcbadcalls);
1441 					COTSRCSTAT_INCR(p->cku_stats,
1442 					    rcnewcreds);
1443 					goto call_again;
1444 				}
1445 
1446 				/*
1447 				 * We have used the client handle to
1448 				 * do an AUTH_REFRESH and the RPC status may
1449 				 * be set to RPC_SUCCESS; Let's make sure to
1450 				 * set it to RPC_AUTHERROR.
1451 				 */
1452 				p->cku_err.re_status = RPC_AUTHERROR;
1453 
1454 				/*
1455 				 * Map recoverable and unrecoverable
1456 				 * authentication errors to appropriate errno
1457 				 */
1458 				switch (p->cku_err.re_why) {
1459 				case AUTH_TOOWEAK:
1460 					/*
1461 					 * This could be a failure where the
1462 					 * server requires use of a reserved
1463 					 * port,  check and optionally set the
1464 					 * client handle useresvport trying
1465 					 * one more time. Next go round we
1466 					 * fall out with the tooweak error.
1467 					 */
1468 					if (p->cku_useresvport != 1) {
1469 						p->cku_useresvport = 1;
1470 						p->cku_xid = 0;
1471 						freemsg(mp);
1472 						mp = NULL;
1473 						goto call_again;
1474 					}
1475 					/* FALLTHRU */
1476 				case AUTH_BADCRED:
1477 				case AUTH_BADVERF:
1478 				case AUTH_INVALIDRESP:
1479 				case AUTH_FAILED:
1480 				case RPCSEC_GSS_NOCRED:
1481 				case RPCSEC_GSS_FAILED:
1482 						p->cku_err.re_errno = EACCES;
1483 						break;
1484 				case AUTH_REJECTEDCRED:
1485 				case AUTH_REJECTEDVERF:
1486 				default:	p->cku_err.re_errno = EIO;
1487 						break;
1488 				}
1489 				RPCLOG(1, "clnt_cots_kcallit : authentication"
1490 				    " failed with RPC_AUTHERROR of type %d\n",
1491 				    (int)p->cku_err.re_why);
1492 				goto cots_done;
1493 			}
1494 		}
1495 	} else {
1496 		/* reply didn't decode properly. */
1497 		p->cku_err.re_status = RPC_CANTDECODERES;
1498 		p->cku_err.re_errno = EIO;
1499 		RPCLOG0(1, "clnt_cots_kcallit: decode failure\n");
1500 	}
1501 
1502 	(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1503 	XDR_DESTROY(xdrs);
1504 
1505 	if (p->cku_flags & CKU_ONQUEUE) {
1506 		call_table_remove(call);
1507 		p->cku_flags &= ~CKU_ONQUEUE;
1508 	}
1509 
1510 	RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list",
1511 	    p->cku_xid);
1512 	RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status));
1513 cots_done:
1514 	if (cm_entry)
1515 		connmgr_release(cm_entry);
1516 
1517 	if (mp != NULL)
1518 		freemsg(mp);
1519 	if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) {
1520 		freemsg(call->call_reply);
1521 		call->call_reply = NULL;
1522 	}
1523 	if (p->cku_err.re_status != RPC_SUCCESS) {
1524 		RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n");
1525 		COTSRCSTAT_INCR(p->cku_stats, rcbadcalls);
1526 	}
1527 
1528 	/*
1529 	 * No point in delaying if the zone is going away.
1530 	 */
1531 	if (delay_first == TRUE &&
1532 	    !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) {
1533 		if (clnt_delay(ticks, h->cl_nosignal) == EINTR) {
1534 			p->cku_err.re_errno = EINTR;
1535 			p->cku_err.re_status = RPC_INTR;
1536 		}
1537 	}
1538 	return (p->cku_err.re_status);
1539 }
1540 
1541 /*
1542  * Kinit routine for cots.  This sets up the correct operations in
1543  * the client handle, as the handle may have previously been a clts
1544  * handle, and clears the xid field so there is no way a new call
1545  * could be mistaken for a retry.  It also sets in the handle the
1546  * information that is passed at create/kinit time but needed at
1547  * call time, as cots creates the transport at call time - device,
1548  * address of the server, protocol family.
1549  */
1550 void
1551 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr,
1552     int max_msgsize, cred_t *cred)
1553 {
1554 	/* LINTED pointer alignment */
1555 	cku_private_t *p = htop(h);
1556 	calllist_t *call = &p->cku_call;
1557 
1558 	h->cl_ops = &tcp_ops;
1559 	if (p->cku_flags & CKU_ONQUEUE) {
1560 		call_table_remove(call);
1561 		p->cku_flags &= ~CKU_ONQUEUE;
1562 		RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from"
1563 		    " dispatch list\n", p->cku_xid);
1564 	}
1565 
1566 	if (call->call_reply != NULL) {
1567 		freemsg(call->call_reply);
1568 		call->call_reply = NULL;
1569 	}
1570 
1571 	call->call_bucket = NULL;
1572 	call->call_hash = 0;
1573 
1574 	/*
1575 	 * We don't clear cku_flags here, because clnt_cots_kcallit()
1576 	 * takes care of handling the cku_flags reset.
1577 	 */
1578 	p->cku_xid = 0;
1579 	p->cku_device = dev;
1580 	p->cku_addrfmly = family;
1581 	p->cku_cred = cred;
1582 
1583 	if (p->cku_addr.maxlen < addr->len) {
1584 		kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
1585 		p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
1586 		p->cku_addr.maxlen = addr->maxlen;
1587 	}
1588 
1589 	p->cku_addr.len = addr->len;
1590 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
1591 
1592 	/*
1593 	 * If the current sanity check size in rpcmod is smaller
1594 	 * than the size needed, then increase the sanity check.
1595 	 */
1596 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
1597 	    max_msgsize > *clnt_max_msg_sizep) {
1598 		mutex_enter(&clnt_max_msg_lock);
1599 		if (max_msgsize > *clnt_max_msg_sizep)
1600 			*clnt_max_msg_sizep = max_msgsize;
1601 		mutex_exit(&clnt_max_msg_lock);
1602 	}
1603 }
1604 
1605 /*
1606  * ksettimers is a no-op for cots, with the exception of setting the xid.
1607  */
1608 /* ARGSUSED */
1609 static int
1610 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all,
1611     int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg, uint32_t xid)
1612 {
1613 	/* LINTED pointer alignment */
1614 	cku_private_t *p = htop(h);
1615 
1616 	if (xid)
1617 		p->cku_xid = xid;
1618 	COTSRCSTAT_INCR(p->cku_stats, rctimers);
1619 	return (0);
1620 }
1621 
1622 extern void rpc_poptimod(struct vnode *);
1623 extern int kstr_push(struct vnode *, char *);
1624 
1625 int
1626 conn_kstat_update(kstat_t *ksp, int rw)
1627 {
1628 	struct cm_xprt *cm_entry;
1629 	struct cm_kstat_xprt *cm_ksp_data;
1630 	uchar_t *b;
1631 	char *fbuf;
1632 
1633 	if (rw == KSTAT_WRITE)
1634 		return (EACCES);
1635 	if (ksp == NULL || ksp->ks_private == NULL)
1636 		return (EIO);
1637 	cm_entry  = (struct cm_xprt *)ksp->ks_private;
1638 	cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data;
1639 
1640 	cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq;
1641 	cm_ksp_data->x_family.value.ui32 = cm_entry->x_family;
1642 	cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev;
1643 	cm_ksp_data->x_time.value.ui32 = cm_entry->x_time;
1644 	cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref;
1645 	cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags;
1646 
1647 	if (cm_entry->x_server.buf) {
1648 		fbuf = cm_ksp_data->x_server.value.str.addr.ptr;
1649 		if (cm_entry->x_family == AF_INET &&
1650 		    cm_entry->x_server.len ==
1651 		    sizeof (struct sockaddr_in)) {
1652 			struct sockaddr_in  *sa;
1653 			sa = (struct sockaddr_in *)
1654 				cm_entry->x_server.buf;
1655 			b = (uchar_t *)&sa->sin_addr;
1656 			(void) sprintf(fbuf,
1657 			    "%d.%d.%d.%d", b[0] & 0xFF, b[1] & 0xFF,
1658 			    b[2] & 0xFF, b[3] & 0xFF);
1659 			cm_ksp_data->x_port.value.ui32 = ntohs(sa->sin_port);
1660 		} else if (cm_entry->x_family == AF_INET6 &&
1661 				cm_entry->x_server.len >=
1662 				sizeof (struct sockaddr_in6)) {
1663 			/* extract server IP address & port */
1664 			struct sockaddr_in6 *sin6;
1665 			sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf;
1666 			(void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf,
1667 				INET6_ADDRSTRLEN);
1668 			cm_ksp_data->x_port.value.ui32 = ntohs(sin6->sin6_port);
1669 		} else {
1670 			struct sockaddr_in  *sa;
1671 
1672 			sa = (struct sockaddr_in *)cm_entry->x_server.buf;
1673 			b = (uchar_t *)&sa->sin_addr;
1674 			(void) sprintf(fbuf,
1675 			    "%d.%d.%d.%d", b[0] & 0xFF, b[1] & 0xFF,
1676 			    b[2] & 0xFF, b[3] & 0xFF);
1677 		}
1678 		KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) =
1679 		    strlen(fbuf) + 1;
1680 	}
1681 
1682 	return (0);
1683 }
1684 
1685 
1686 /*
1687  * We want a version of delay which is interruptible by a UNIX signal
1688  * Return EINTR if an interrupt occured.
1689  */
1690 static int
1691 clnt_delay(clock_t ticks, bool_t nosignal)
1692 {
1693 	if (nosignal == TRUE) {
1694 		delay(ticks);
1695 		return (0);
1696 	}
1697 	return (delay_sig(ticks));
1698 }
1699 
1700 /*
1701  * Wait for a connection until a timeout, or until we are
1702  * signalled that there has been a connection state change.
1703  */
1704 static enum clnt_stat
1705 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp,
1706     bool_t nosignal)
1707 {
1708 	bool_t interrupted;
1709 	clock_t timout, cv_stat;
1710 	enum clnt_stat clstat;
1711 	unsigned int old_state;
1712 
1713 	ASSERT(MUTEX_HELD(&connmgr_lock));
1714 	/*
1715 	 * We wait for the transport connection to be made, or an
1716 	 * indication that it could not be made.
1717 	 */
1718 	clstat = RPC_TIMEDOUT;
1719 	interrupted = FALSE;
1720 
1721 	old_state = cm_entry->x_state_flags;
1722 	/*
1723 	 * Now loop until cv_timedwait{_sig} returns because of
1724 	 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be
1725 	 * cv_signalled for various other reasons too. So loop
1726 	 * until there is a state change on the connection.
1727 	 */
1728 
1729 	timout = waitp->tv_sec * drv_usectohz(1000000) +
1730 	    drv_usectohz(waitp->tv_usec) + ddi_get_lbolt();
1731 
1732 	if (nosignal) {
1733 		while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv,
1734 		    &connmgr_lock, timout)) > 0 &&
1735 		    cm_entry->x_state_flags == old_state)
1736 			;
1737 	} else {
1738 		while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv,
1739 		    &connmgr_lock, timout)) > 0 &&
1740 		    cm_entry->x_state_flags == old_state)
1741 			;
1742 
1743 		if (cv_stat == 0) /* got intr signal? */
1744 			interrupted = TRUE;
1745 	}
1746 
1747 	if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) ==
1748 	    X_CONNECTED) {
1749 		clstat = RPC_SUCCESS;
1750 	} else {
1751 		if (interrupted == TRUE)
1752 			clstat = RPC_INTR;
1753 		RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n",
1754 		    clnt_sperrno(clstat));
1755 	}
1756 
1757 	return (clstat);
1758 }
1759 
1760 /*
1761  * Primary interface for how RPC grabs a connection.
1762  */
1763 static struct cm_xprt *
1764 connmgr_wrapget(
1765 	struct netbuf *retryaddr,
1766 	const struct timeval *waitp,
1767 	cku_private_t *p)
1768 {
1769 	struct cm_xprt *cm_entry;
1770 
1771 	cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly,
1772 	    &p->cku_srcaddr, &p->cku_err, p->cku_device,
1773 	    p->cku_client.cl_nosignal, p->cku_useresvport, p->cku_cred);
1774 
1775 	if (cm_entry == NULL) {
1776 		/*
1777 		 * Re-map the call status to RPC_INTR if the err code is
1778 		 * EINTR. This can happen if calls status is RPC_TLIERROR.
1779 		 * However, don't re-map if signalling has been turned off.
1780 		 * XXX Really need to create a separate thread whenever
1781 		 * there isn't an existing connection.
1782 		 */
1783 		if (p->cku_err.re_errno == EINTR) {
1784 			if (p->cku_client.cl_nosignal == TRUE)
1785 				p->cku_err.re_errno = EIO;
1786 			else
1787 				p->cku_err.re_status = RPC_INTR;
1788 		}
1789 	}
1790 
1791 	return (cm_entry);
1792 }
1793 
1794 /*
1795  * Obtains a transport to the server specified in addr.  If a suitable transport
1796  * does not already exist in the list of cached transports, a new connection
1797  * is created, connected, and added to the list. The connection is for sending
1798  * only - the reply message may come back on another transport connection.
1799  *
1800  * To implement round-robin load balancing with multiple client connections,
1801  * the last entry on the list is always selected. Once the entry is selected
1802  * it's re-inserted to the head of the list.
1803  */
1804 static struct cm_xprt *
1805 connmgr_get(
1806 	struct netbuf	*retryaddr,
1807 	const struct timeval	*waitp,	/* changed to a ptr to converse stack */
1808 	struct netbuf	*destaddr,
1809 	int		addrfmly,
1810 	struct netbuf	*srcaddr,
1811 	struct rpc_err	*rpcerr,
1812 	dev_t		device,
1813 	bool_t		nosignal,
1814 	int		useresvport,
1815 	cred_t		*cr)
1816 {
1817 	struct cm_xprt *cm_entry;
1818 	struct cm_xprt *lru_entry;
1819 	struct cm_xprt **cmp, **prev;
1820 	queue_t *wq;
1821 	TIUSER *tiptr;
1822 	int i;
1823 	int retval;
1824 	int tidu_size;
1825 	bool_t	connected;
1826 	zoneid_t zoneid = rpc_zoneid();
1827 
1828 	/*
1829 	 * If the call is not a retry, look for a transport entry that
1830 	 * goes to the server of interest.
1831 	 */
1832 	mutex_enter(&connmgr_lock);
1833 
1834 	if (retryaddr == NULL) {
1835 use_new_conn:
1836 		i = 0;
1837 		cm_entry = lru_entry = NULL;
1838 
1839 		prev = cmp = &cm_hd;
1840 		while ((cm_entry = *cmp) != NULL) {
1841 			ASSERT(cm_entry != cm_entry->x_next);
1842 			/*
1843 			 * Garbage collect conections that are marked
1844 			 * for needs disconnect.
1845 			 */
1846 			if (cm_entry->x_needdis) {
1847 				CONN_HOLD(cm_entry);
1848 				connmgr_dis_and_wait(cm_entry);
1849 				connmgr_release(cm_entry);
1850 				/*
1851 				 * connmgr_lock could have been
1852 				 * dropped for the disconnect
1853 				 * processing so start over.
1854 				 */
1855 				goto use_new_conn;
1856 			}
1857 
1858 			/*
1859 			 * Garbage collect the dead connections that have
1860 			 * no threads working on them.
1861 			 */
1862 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1863 			    X_DEAD) {
1864 				mutex_enter(&cm_entry->x_lock);
1865 				if (cm_entry->x_ref != 0) {
1866 					/*
1867 					 * Currently in use.
1868 					 * Cleanup later.
1869 					 */
1870 					cmp = &cm_entry->x_next;
1871 					mutex_exit(&cm_entry->x_lock);
1872 					continue;
1873 				}
1874 				mutex_exit(&cm_entry->x_lock);
1875 				*cmp = cm_entry->x_next;
1876 				mutex_exit(&connmgr_lock);
1877 				connmgr_close(cm_entry);
1878 				mutex_enter(&connmgr_lock);
1879 				goto use_new_conn;
1880 			}
1881 
1882 
1883 			if ((cm_entry->x_state_flags & X_BADSTATES) == 0 &&
1884 			    cm_entry->x_zoneid == zoneid &&
1885 			    cm_entry->x_rdev == device &&
1886 			    destaddr->len == cm_entry->x_server.len &&
1887 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
1888 			    destaddr->len) == 0) {
1889 				/*
1890 				 * If the matching entry isn't connected,
1891 				 * attempt to reconnect it.
1892 				 */
1893 				if (cm_entry->x_connected == FALSE) {
1894 					/*
1895 					 * We don't go through trying
1896 					 * to find the least recently
1897 					 * used connected because
1898 					 * connmgr_reconnect() briefly
1899 					 * dropped the connmgr_lock,
1900 					 * allowing a window for our
1901 					 * accounting to be messed up.
1902 					 * In any case, a re-connected
1903 					 * connection is as good as
1904 					 * a LRU connection.
1905 					 */
1906 					return (connmgr_wrapconnect(cm_entry,
1907 					    waitp, destaddr, addrfmly, srcaddr,
1908 					    rpcerr, TRUE, nosignal, cr));
1909 				}
1910 				i++;
1911 
1912 				/* keep track of the last entry */
1913 				lru_entry = cm_entry;
1914 				prev = cmp;
1915 			}
1916 			cmp = &cm_entry->x_next;
1917 		}
1918 
1919 		if (i > clnt_max_conns) {
1920 			RPCLOG(8, "connmgr_get: too many conns, dooming entry"
1921 			    " %p\n", (void *)lru_entry->x_tiptr);
1922 			lru_entry->x_doomed = TRUE;
1923 			goto use_new_conn;
1924 		}
1925 
1926 		/*
1927 		 * If we are at the maximum number of connections to
1928 		 * the server, hand back the least recently used one.
1929 		 */
1930 		if (i == clnt_max_conns) {
1931 			/*
1932 			 * Copy into the handle the source address of
1933 			 * the connection, which we will use in case of
1934 			 * a later retry.
1935 			 */
1936 			if (srcaddr->len != lru_entry->x_src.len) {
1937 				kmem_free(srcaddr->buf, srcaddr->maxlen);
1938 				ASSERT(lru_entry->x_src.len != 0);
1939 				srcaddr->buf = kmem_alloc(
1940 				    lru_entry->x_src.len, KM_SLEEP);
1941 				srcaddr->maxlen = srcaddr->len =
1942 				    lru_entry->x_src.len;
1943 			}
1944 			bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len);
1945 			RPCLOG(2, "connmgr_get: call going out on %p\n",
1946 			    (void *)lru_entry);
1947 			lru_entry->x_time = ddi_get_lbolt();
1948 			CONN_HOLD(lru_entry);
1949 
1950 			if ((i > 1) && (prev != &cm_hd)) {
1951 				/*
1952 				 * remove and re-insert entry at head of list.
1953 				 */
1954 				*prev = lru_entry->x_next;
1955 				lru_entry->x_next = cm_hd;
1956 				cm_hd = lru_entry;
1957 			}
1958 
1959 			mutex_exit(&connmgr_lock);
1960 			return (lru_entry);
1961 		}
1962 
1963 	} else {
1964 		/*
1965 		 * This is the retry case (retryaddr != NULL).  Retries must
1966 		 * be sent on the same source port as the original call.
1967 		 */
1968 
1969 		/*
1970 		 * Walk the list looking for a connection with a source address
1971 		 * that matches the retry address.
1972 		 */
1973 start_retry_loop:
1974 		cmp = &cm_hd;
1975 		while ((cm_entry = *cmp) != NULL) {
1976 			ASSERT(cm_entry != cm_entry->x_next);
1977 
1978 			/*
1979 			 * determine if this connection matches the passed
1980 			 * in retry address.  If it does not match, advance
1981 			 * to the next element on the list.
1982 			 */
1983 			if (zoneid != cm_entry->x_zoneid ||
1984 			    device != cm_entry->x_rdev ||
1985 			    retryaddr->len != cm_entry->x_src.len ||
1986 			    bcmp(retryaddr->buf, cm_entry->x_src.buf,
1987 			    retryaddr->len) != 0) {
1988 				cmp = &cm_entry->x_next;
1989 				continue;
1990 			}
1991 			/*
1992 			 * Garbage collect conections that are marked
1993 			 * for needs disconnect.
1994 			 */
1995 			if (cm_entry->x_needdis) {
1996 				CONN_HOLD(cm_entry);
1997 				connmgr_dis_and_wait(cm_entry);
1998 				connmgr_release(cm_entry);
1999 				/*
2000 				 * connmgr_lock could have been
2001 				 * dropped for the disconnect
2002 				 * processing so start over.
2003 				 */
2004 				goto start_retry_loop;
2005 			}
2006 			/*
2007 			 * Garbage collect the dead connections that have
2008 			 * no threads working on them.
2009 			 */
2010 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
2011 			    X_DEAD) {
2012 				mutex_enter(&cm_entry->x_lock);
2013 				if (cm_entry->x_ref != 0) {
2014 					/*
2015 					 * Currently in use.
2016 					 * Cleanup later.
2017 					 */
2018 					cmp = &cm_entry->x_next;
2019 					mutex_exit(&cm_entry->x_lock);
2020 					continue;
2021 				}
2022 				mutex_exit(&cm_entry->x_lock);
2023 				*cmp = cm_entry->x_next;
2024 				mutex_exit(&connmgr_lock);
2025 				connmgr_close(cm_entry);
2026 				mutex_enter(&connmgr_lock);
2027 				goto start_retry_loop;
2028 			}
2029 
2030 			/*
2031 			 * Sanity check: if the connection with our source
2032 			 * port is going to some other server, something went
2033 			 * wrong, as we never delete connections (i.e. release
2034 			 * ports) unless they have been idle.  In this case,
2035 			 * it is probably better to send the call out using
2036 			 * a new source address than to fail it altogether,
2037 			 * since that port may never be released.
2038 			 */
2039 			if (destaddr->len != cm_entry->x_server.len ||
2040 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
2041 			    destaddr->len) != 0) {
2042 				RPCLOG(1, "connmgr_get: tiptr %p"
2043 				    " is going to a different server"
2044 				    " with the port that belongs"
2045 				    " to us!\n", (void *)cm_entry->x_tiptr);
2046 				retryaddr = NULL;
2047 				goto use_new_conn;
2048 			}
2049 
2050 			/*
2051 			 * If the connection of interest is not connected and we
2052 			 * can't reconnect it, then the server is probably
2053 			 * still down.  Return NULL to the caller and let it
2054 			 * retry later if it wants to.  We have a delay so the
2055 			 * machine doesn't go into a tight retry loop.  If the
2056 			 * entry was already connected, or the reconnected was
2057 			 * successful, return this entry.
2058 			 */
2059 			if (cm_entry->x_connected == FALSE) {
2060 				return (connmgr_wrapconnect(cm_entry,
2061 				    waitp, destaddr, addrfmly, NULL,
2062 				    rpcerr, TRUE, nosignal, cr));
2063 			} else {
2064 				CONN_HOLD(cm_entry);
2065 
2066 				cm_entry->x_time = ddi_get_lbolt();
2067 				mutex_exit(&connmgr_lock);
2068 				RPCLOG(2, "connmgr_get: found old "
2069 				    "transport %p for retry\n",
2070 				    (void *)cm_entry);
2071 				return (cm_entry);
2072 			}
2073 		}
2074 
2075 		/*
2076 		 * We cannot find an entry in the list for this retry.
2077 		 * Either the entry has been removed temporarily to be
2078 		 * reconnected by another thread, or the original call
2079 		 * got a port but never got connected,
2080 		 * and hence the transport never got put in the
2081 		 * list.  Fall through to the "create new connection" code -
2082 		 * the former case will fail there trying to rebind the port,
2083 		 * and the later case (and any other pathological cases) will
2084 		 * rebind and reconnect and not hang the client machine.
2085 		 */
2086 		RPCLOG0(8, "connmgr_get: no entry in list for retry\n");
2087 	}
2088 	/*
2089 	 * Set up a transport entry in the connection manager's list.
2090 	 */
2091 	cm_entry = (struct cm_xprt *)
2092 	    kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP);
2093 
2094 	cm_entry->x_server.buf = kmem_alloc(destaddr->len, KM_SLEEP);
2095 	bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len);
2096 	cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len;
2097 
2098 	cm_entry->x_state_flags = X_THREAD;
2099 	cm_entry->x_ref = 1;
2100 	cm_entry->x_family = addrfmly;
2101 	cm_entry->x_rdev = device;
2102 	cm_entry->x_zoneid = zoneid;
2103 	mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL);
2104 	cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL);
2105 	cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL);
2106 	cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL);
2107 
2108 	/*
2109 	 * Note that we add this partially initialized entry to the
2110 	 * connection list. This is so that we don't have connections to
2111 	 * the same server.
2112 	 *
2113 	 * Note that x_src is not initialized at this point. This is because
2114 	 * retryaddr might be NULL in which case x_src is whatever
2115 	 * t_kbind/bindresvport gives us. If another thread wants a
2116 	 * connection to the same server, seemingly we have an issue, but we
2117 	 * don't. If the other thread comes in with retryaddr == NULL, then it
2118 	 * will never look at x_src, and it will end up waiting in
2119 	 * connmgr_cwait() for the first thread to finish the connection
2120 	 * attempt. If the other thread comes in with retryaddr != NULL, then
2121 	 * that means there was a request sent on a connection, in which case
2122 	 * the the connection should already exist. Thus the first thread
2123 	 * never gets here ... it finds the connection it its server in the
2124 	 * connection list.
2125 	 *
2126 	 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd
2127 	 * thread will skip us because x_src.len == 0.
2128 	 */
2129 	cm_entry->x_next = cm_hd;
2130 	cm_hd = cm_entry;
2131 	mutex_exit(&connmgr_lock);
2132 
2133 	/*
2134 	 * Either we didn't find an entry to the server of interest, or we
2135 	 * don't have the maximum number of connections to that server -
2136 	 * create a new connection.
2137 	 */
2138 	RPCLOG0(8, "connmgr_get: creating new connection\n");
2139 	rpcerr->re_status = RPC_TLIERROR;
2140 
2141 	i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, zone_kcred());
2142 	if (i) {
2143 		RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i);
2144 		rpcerr->re_errno = i;
2145 		connmgr_cancelconn(cm_entry);
2146 		return (NULL);
2147 	}
2148 	rpc_poptimod(tiptr->fp->f_vnode);
2149 
2150 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0,
2151 	    K_TO_K, kcred, &retval)) {
2152 		RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i);
2153 		(void) t_kclose(tiptr, 1);
2154 		rpcerr->re_errno = i;
2155 		connmgr_cancelconn(cm_entry);
2156 		return (NULL);
2157 	}
2158 
2159 	if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K,
2160 	    kcred, &retval)) {
2161 		RPCLOG(1, "connmgr_get: can't set client status with cots "
2162 		    "module, %d\n", i);
2163 		(void) t_kclose(tiptr, 1);
2164 		rpcerr->re_errno = i;
2165 		connmgr_cancelconn(cm_entry);
2166 		return (NULL);
2167 	}
2168 
2169 	mutex_enter(&connmgr_lock);
2170 
2171 	wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next;
2172 	cm_entry->x_wq = wq;
2173 
2174 	mutex_exit(&connmgr_lock);
2175 
2176 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0,
2177 	    K_TO_K, kcred, &retval)) {
2178 		RPCLOG(1, "connmgr_get: can't push timod, %d\n", i);
2179 		(void) t_kclose(tiptr, 1);
2180 		rpcerr->re_errno = i;
2181 		connmgr_cancelconn(cm_entry);
2182 		return (NULL);
2183 	}
2184 
2185 	/*
2186 	 * If the caller has not specified reserved port usage then
2187 	 * take the system default.
2188 	 */
2189 	if (useresvport == -1)
2190 		useresvport = clnt_cots_do_bindresvport;
2191 
2192 	if ((useresvport || retryaddr != NULL) &&
2193 	    (addrfmly == AF_INET || addrfmly == AF_INET6)) {
2194 		bool_t alloc_src = FALSE;
2195 
2196 		if (srcaddr->len != destaddr->len) {
2197 			kmem_free(srcaddr->buf, srcaddr->maxlen);
2198 			srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2199 			srcaddr->maxlen = destaddr->len;
2200 			srcaddr->len = destaddr->len;
2201 			alloc_src = TRUE;
2202 		}
2203 
2204 		if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) {
2205 			(void) t_kclose(tiptr, 1);
2206 			RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: "
2207 			    "%p\n", (void *)retryaddr);
2208 
2209 			/*
2210 			 * 1225408: If we allocated a source address, then it
2211 			 * is either garbage or all zeroes. In that case
2212 			 * we need to clear srcaddr.
2213 			 */
2214 			if (alloc_src == TRUE) {
2215 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2216 				srcaddr->maxlen = srcaddr->len = 0;
2217 				srcaddr->buf = NULL;
2218 			}
2219 			rpcerr->re_errno = i;
2220 			connmgr_cancelconn(cm_entry);
2221 			return (NULL);
2222 		}
2223 	} else {
2224 		if ((i = t_kbind(tiptr, NULL, NULL)) != 0) {
2225 			RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i);
2226 			(void) t_kclose(tiptr, 1);
2227 			rpcerr->re_errno = i;
2228 			connmgr_cancelconn(cm_entry);
2229 			return (NULL);
2230 		}
2231 	}
2232 
2233 	{
2234 		/*
2235 		 * Keep the kernel stack lean. Don't move this call
2236 		 * declaration to the top of this function because a
2237 		 * call is declared in connmgr_wrapconnect()
2238 		 */
2239 		calllist_t call;
2240 
2241 		bzero(&call, sizeof (call));
2242 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2243 
2244 		/*
2245 		 * This is a bound end-point so don't close it's stream.
2246 		 */
2247 		connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly,
2248 		    &call, &tidu_size, FALSE, waitp, nosignal, cr);
2249 		*rpcerr = call.call_err;
2250 		cv_destroy(&call.call_cv);
2251 
2252 	}
2253 
2254 	mutex_enter(&connmgr_lock);
2255 
2256 	/*
2257 	 * Set up a transport entry in the connection manager's list.
2258 	 */
2259 	if (srcaddr->len > 0) {
2260 		cm_entry->x_src.buf = kmem_alloc(srcaddr->len, KM_SLEEP);
2261 		bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len);
2262 		cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len;
2263 	} /* Else kmem_zalloc() of cm_entry already sets its x_src to NULL. */
2264 
2265 	cm_entry->x_tiptr = tiptr;
2266 	cm_entry->x_time = ddi_get_lbolt();
2267 
2268 	if (tiptr->tp_info.servtype == T_COTS_ORD)
2269 		cm_entry->x_ordrel = TRUE;
2270 	else
2271 		cm_entry->x_ordrel = FALSE;
2272 
2273 	cm_entry->x_tidu_size = tidu_size;
2274 
2275 	if (cm_entry->x_early_disc) {
2276 		/*
2277 		 * We need to check if a disconnect request has come
2278 		 * while we are connected, if so, then we need to
2279 		 * set rpcerr->re_status appropriately before returning
2280 		 * NULL to caller.
2281 		 */
2282 		if (rpcerr->re_status == RPC_SUCCESS)
2283 			rpcerr->re_status = RPC_XPRTFAILED;
2284 		cm_entry->x_connected = FALSE;
2285 	} else
2286 		cm_entry->x_connected = connected;
2287 
2288 	/*
2289 	 * There could be a discrepancy here such that
2290 	 * x_early_disc is TRUE yet connected is TRUE as well
2291 	 * and the connection is actually connected. In that case
2292 	 * lets be conservative and declare the connection as not
2293 	 * connected.
2294 	 */
2295 	cm_entry->x_early_disc = FALSE;
2296 	cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2297 	cm_entry->x_ctime = ddi_get_lbolt();
2298 
2299 	/*
2300 	 * Notify any threads waiting that the connection attempt is done.
2301 	 */
2302 	cm_entry->x_thread = FALSE;
2303 	cv_broadcast(&cm_entry->x_conn_cv);
2304 
2305 	if (cm_entry->x_connected == FALSE) {
2306 		mutex_exit(&connmgr_lock);
2307 		connmgr_release(cm_entry);
2308 		return (NULL);
2309 	}
2310 
2311 	mutex_exit(&connmgr_lock);
2312 
2313 	return (cm_entry);
2314 }
2315 
2316 /*
2317  * Keep the cm_xprt entry on the connecton list when making a connection. This
2318  * is to prevent multiple connections to a slow server from appearing.
2319  * We use the bit field x_thread to tell if a thread is doing a connection
2320  * which keeps other interested threads from messing with connection.
2321  * Those other threads just wait if x_thread is set.
2322  *
2323  * If x_thread is not set, then we do the actual work of connecting via
2324  * connmgr_connect().
2325  *
2326  * mutex convention: called with connmgr_lock held, returns with it released.
2327  */
2328 static struct cm_xprt *
2329 connmgr_wrapconnect(
2330 	struct cm_xprt	*cm_entry,
2331 	const struct timeval	*waitp,
2332 	struct netbuf	*destaddr,
2333 	int		addrfmly,
2334 	struct netbuf	*srcaddr,
2335 	struct rpc_err	*rpcerr,
2336 	bool_t		reconnect,
2337 	bool_t		nosignal,
2338 	cred_t		*cr)
2339 {
2340 	ASSERT(MUTEX_HELD(&connmgr_lock));
2341 	/*
2342 	 * Hold this entry as we are about to drop connmgr_lock.
2343 	 */
2344 	CONN_HOLD(cm_entry);
2345 
2346 	/*
2347 	 * If there is a thread already making a connection for us, then
2348 	 * wait for it to complete the connection.
2349 	 */
2350 	if (cm_entry->x_thread == TRUE) {
2351 		rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal);
2352 
2353 		if (rpcerr->re_status != RPC_SUCCESS) {
2354 			mutex_exit(&connmgr_lock);
2355 			connmgr_release(cm_entry);
2356 			return (NULL);
2357 		}
2358 	} else {
2359 		bool_t connected;
2360 		calllist_t call;
2361 
2362 		cm_entry->x_thread = TRUE;
2363 
2364 		while (cm_entry->x_needrel == TRUE) {
2365 			cm_entry->x_needrel = FALSE;
2366 
2367 			connmgr_sndrel(cm_entry);
2368 			delay(drv_usectohz(1000000));
2369 
2370 			mutex_enter(&connmgr_lock);
2371 		}
2372 
2373 		/*
2374 		 * If we need to send a T_DISCON_REQ, send one.
2375 		 */
2376 		connmgr_dis_and_wait(cm_entry);
2377 
2378 		mutex_exit(&connmgr_lock);
2379 
2380 		bzero(&call, sizeof (call));
2381 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2382 
2383 		connected = connmgr_connect(cm_entry, cm_entry->x_wq,
2384 		    destaddr, addrfmly, &call, &cm_entry->x_tidu_size,
2385 		    reconnect, waitp, nosignal, cr);
2386 
2387 		*rpcerr = call.call_err;
2388 		cv_destroy(&call.call_cv);
2389 
2390 		mutex_enter(&connmgr_lock);
2391 
2392 
2393 		if (cm_entry->x_early_disc) {
2394 			/*
2395 			 * We need to check if a disconnect request has come
2396 			 * while we are connected, if so, then we need to
2397 			 * set rpcerr->re_status appropriately before returning
2398 			 * NULL to caller.
2399 			 */
2400 			if (rpcerr->re_status == RPC_SUCCESS)
2401 				rpcerr->re_status = RPC_XPRTFAILED;
2402 			cm_entry->x_connected = FALSE;
2403 		} else
2404 			cm_entry->x_connected = connected;
2405 
2406 		/*
2407 		 * There could be a discrepancy here such that
2408 		 * x_early_disc is TRUE yet connected is TRUE as well
2409 		 * and the connection is actually connected. In that case
2410 		 * lets be conservative and declare the connection as not
2411 		 * connected.
2412 		 */
2413 
2414 		cm_entry->x_early_disc = FALSE;
2415 		cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2416 
2417 
2418 		/*
2419 		 * connmgr_connect() may have given up before the connection
2420 		 * actually timed out. So ensure that before the next
2421 		 * connection attempt we do a disconnect.
2422 		 */
2423 		cm_entry->x_ctime = ddi_get_lbolt();
2424 		cm_entry->x_thread = FALSE;
2425 
2426 		cv_broadcast(&cm_entry->x_conn_cv);
2427 
2428 		if (cm_entry->x_connected == FALSE) {
2429 			mutex_exit(&connmgr_lock);
2430 			connmgr_release(cm_entry);
2431 			return (NULL);
2432 		}
2433 	}
2434 
2435 	if (srcaddr != NULL) {
2436 		/*
2437 		 * Copy into the handle the
2438 		 * source address of the
2439 		 * connection, which we will use
2440 		 * in case of a later retry.
2441 		 */
2442 		if (srcaddr->len != cm_entry->x_src.len) {
2443 			kmem_free(srcaddr->buf, srcaddr->maxlen);
2444 			ASSERT(cm_entry->x_src.len != 0);
2445 			srcaddr->buf = kmem_alloc(cm_entry->x_src.len,
2446 			    KM_SLEEP);
2447 			srcaddr->maxlen = srcaddr->len = cm_entry->x_src.len;
2448 		}
2449 		bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len);
2450 	}
2451 	cm_entry->x_time = ddi_get_lbolt();
2452 	mutex_exit(&connmgr_lock);
2453 	return (cm_entry);
2454 }
2455 
2456 /*
2457  * If we need to send a T_DISCON_REQ, send one.
2458  */
2459 static void
2460 connmgr_dis_and_wait(struct cm_xprt *cm_entry)
2461 {
2462 	ASSERT(MUTEX_HELD(&connmgr_lock));
2463 	for (;;) {
2464 		while (cm_entry->x_needdis == TRUE) {
2465 			RPCLOG(8, "connmgr_dis_and_wait: need "
2466 			    "T_DISCON_REQ for connection 0x%p\n",
2467 			    (void *)cm_entry);
2468 			cm_entry->x_needdis = FALSE;
2469 			cm_entry->x_waitdis = TRUE;
2470 
2471 			connmgr_snddis(cm_entry);
2472 
2473 			mutex_enter(&connmgr_lock);
2474 		}
2475 
2476 		if (cm_entry->x_waitdis == TRUE) {
2477 			clock_t timout;
2478 
2479 			RPCLOG(8, "connmgr_dis_and_wait waiting for "
2480 			    "T_DISCON_REQ's ACK for connection %p\n",
2481 			    (void *)cm_entry);
2482 
2483 			timout = clnt_cots_min_conntout * drv_usectohz(1000000);
2484 
2485 			/*
2486 			 * The TPI spec says that the T_DISCON_REQ
2487 			 * will get acknowledged, but in practice
2488 			 * the ACK may never get sent. So don't
2489 			 * block forever.
2490 			 */
2491 			(void) cv_reltimedwait(&cm_entry->x_dis_cv,
2492 			    &connmgr_lock, timout, TR_CLOCK_TICK);
2493 		}
2494 		/*
2495 		 * If we got the ACK, break. If we didn't,
2496 		 * then send another T_DISCON_REQ.
2497 		 */
2498 		if (cm_entry->x_waitdis == FALSE) {
2499 			break;
2500 		} else {
2501 			RPCLOG(8, "connmgr_dis_and_wait: did"
2502 			    "not get T_DISCON_REQ's ACK for "
2503 			    "connection  %p\n", (void *)cm_entry);
2504 			cm_entry->x_needdis = TRUE;
2505 		}
2506 	}
2507 }
2508 
2509 static void
2510 connmgr_cancelconn(struct cm_xprt *cm_entry)
2511 {
2512 	/*
2513 	 * Mark the connection table entry as dead; the next thread that
2514 	 * goes through connmgr_release() will notice this and deal with it.
2515 	 */
2516 	mutex_enter(&connmgr_lock);
2517 	cm_entry->x_dead = TRUE;
2518 
2519 	/*
2520 	 * Notify any threads waiting for the connection that it isn't
2521 	 * going to happen.
2522 	 */
2523 	cm_entry->x_thread = FALSE;
2524 	cv_broadcast(&cm_entry->x_conn_cv);
2525 	mutex_exit(&connmgr_lock);
2526 
2527 	connmgr_release(cm_entry);
2528 }
2529 
2530 static void
2531 connmgr_close(struct cm_xprt *cm_entry)
2532 {
2533 	mutex_enter(&cm_entry->x_lock);
2534 	while (cm_entry->x_ref != 0) {
2535 		/*
2536 		 * Must be a noninterruptible wait.
2537 		 */
2538 		cv_wait(&cm_entry->x_cv, &cm_entry->x_lock);
2539 	}
2540 
2541 	if (cm_entry->x_tiptr != NULL)
2542 		(void) t_kclose(cm_entry->x_tiptr, 1);
2543 
2544 	mutex_exit(&cm_entry->x_lock);
2545 	if (cm_entry->x_ksp != NULL) {
2546 		mutex_enter(cm_entry->x_ksp->ks_lock);
2547 		cm_entry->x_ksp->ks_private = NULL;
2548 		mutex_exit(cm_entry->x_ksp->ks_lock);
2549 
2550 		/*
2551 		 * Must free the buffer we allocated for the
2552 		 * server address in the update function
2553 		 */
2554 		if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2555 		    x_server.value.str.addr.ptr != NULL)
2556 			kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp->
2557 			    ks_data))->x_server.value.str.addr.ptr,
2558 			    INET6_ADDRSTRLEN);
2559 		kmem_free(cm_entry->x_ksp->ks_data,
2560 		    cm_entry->x_ksp->ks_data_size);
2561 		kstat_delete(cm_entry->x_ksp);
2562 	}
2563 
2564 	mutex_destroy(&cm_entry->x_lock);
2565 	cv_destroy(&cm_entry->x_cv);
2566 	cv_destroy(&cm_entry->x_conn_cv);
2567 	cv_destroy(&cm_entry->x_dis_cv);
2568 
2569 	kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen);
2570 	kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen);
2571 	kmem_free(cm_entry, sizeof (struct cm_xprt));
2572 }
2573 
2574 /*
2575  * Called by KRPC after sending the call message to release the connection
2576  * it was using.
2577  */
2578 static void
2579 connmgr_release(struct cm_xprt *cm_entry)
2580 {
2581 	mutex_enter(&cm_entry->x_lock);
2582 	cm_entry->x_ref--;
2583 	if (cm_entry->x_ref == 0)
2584 		cv_signal(&cm_entry->x_cv);
2585 	mutex_exit(&cm_entry->x_lock);
2586 }
2587 
2588 /*
2589  * Set TCP receive and xmit buffer size for RPC connections.
2590  */
2591 static bool_t
2592 connmgr_setbufsz(calllist_t *e, queue_t *wq, cred_t *cr)
2593 {
2594 	int ok = FALSE;
2595 	int val;
2596 
2597 	if (rpc_default_tcp_bufsz)
2598 		return (FALSE);
2599 
2600 	/*
2601 	 * Only set new buffer size if it's larger than the system
2602 	 * default buffer size. If smaller buffer size is needed
2603 	 * then use /etc/system to set rpc_default_tcp_bufsz to 1.
2604 	 */
2605 	ok = connmgr_getopt_int(wq, SOL_SOCKET, SO_RCVBUF, &val, e, cr);
2606 	if ((ok == TRUE) && (val < rpc_send_bufsz)) {
2607 		ok = connmgr_setopt_int(wq, SOL_SOCKET, SO_RCVBUF,
2608 		    rpc_send_bufsz, e, cr);
2609 		DTRACE_PROBE2(krpc__i__connmgr_rcvbufsz,
2610 		    int, ok, calllist_t *, e);
2611 	}
2612 
2613 	ok = connmgr_getopt_int(wq, SOL_SOCKET, SO_SNDBUF, &val, e, cr);
2614 	if ((ok == TRUE) && (val < rpc_recv_bufsz)) {
2615 		ok = connmgr_setopt_int(wq, SOL_SOCKET, SO_SNDBUF,
2616 		    rpc_recv_bufsz, e, cr);
2617 		DTRACE_PROBE2(krpc__i__connmgr_sndbufsz,
2618 		    int, ok, calllist_t *, e);
2619 	}
2620 	return (TRUE);
2621 }
2622 
2623 /*
2624  * Given an open stream, connect to the remote.  Returns true if connected,
2625  * false otherwise.
2626  */
2627 static bool_t
2628 connmgr_connect(
2629 	struct cm_xprt		*cm_entry,
2630 	queue_t			*wq,
2631 	struct netbuf		*addr,
2632 	int			addrfmly,
2633 	calllist_t		*e,
2634 	int			*tidu_ptr,
2635 	bool_t			reconnect,
2636 	const struct timeval	*waitp,
2637 	bool_t			nosignal,
2638 	cred_t			*cr)
2639 {
2640 	mblk_t *mp;
2641 	struct T_conn_req *tcr;
2642 	struct T_info_ack *tinfo;
2643 	int interrupted, error;
2644 	int tidu_size, kstat_instance;
2645 
2646 	/* if it's a reconnect, flush any lingering data messages */
2647 	if (reconnect)
2648 		(void) putctl1(wq, M_FLUSH, FLUSHRW);
2649 
2650 	/*
2651 	 * Note: if the receiver uses SCM_UCRED/getpeerucred the pid will
2652 	 * appear as -1.
2653 	 */
2654 	mp = allocb_cred(sizeof (*tcr) + addr->len, cr, NOPID);
2655 	if (mp == NULL) {
2656 		/*
2657 		 * This is unfortunate, but we need to look up the stats for
2658 		 * this zone to increment the "memory allocation failed"
2659 		 * counter.  curproc->p_zone is safe since we're initiating a
2660 		 * connection and not in some strange streams context.
2661 		 */
2662 		struct rpcstat *rpcstat;
2663 
2664 		rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
2665 		ASSERT(rpcstat != NULL);
2666 
2667 		RPCLOG0(1, "connmgr_connect: cannot alloc mp for "
2668 		    "sending conn request\n");
2669 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem);
2670 		e->call_status = RPC_SYSTEMERROR;
2671 		e->call_reason = ENOSR;
2672 		return (FALSE);
2673 	}
2674 
2675 	/* Set TCP buffer size for RPC connections if needed */
2676 	if (addrfmly == AF_INET || addrfmly == AF_INET6)
2677 		(void) connmgr_setbufsz(e, wq, cr);
2678 
2679 	mp->b_datap->db_type = M_PROTO;
2680 	tcr = (struct T_conn_req *)mp->b_rptr;
2681 	bzero(tcr, sizeof (*tcr));
2682 	tcr->PRIM_type = T_CONN_REQ;
2683 	tcr->DEST_length = addr->len;
2684 	tcr->DEST_offset = sizeof (struct T_conn_req);
2685 	mp->b_wptr = mp->b_rptr + sizeof (*tcr);
2686 
2687 	bcopy(addr->buf, mp->b_wptr, tcr->DEST_length);
2688 	mp->b_wptr += tcr->DEST_length;
2689 
2690 	RPCLOG(8, "connmgr_connect: sending conn request on queue "
2691 	    "%p", (void *)wq);
2692 	RPCLOG(8, " call %p\n", (void *)wq);
2693 	/*
2694 	 * We use the entry in the handle that is normally used for
2695 	 * waiting for RPC replies to wait for the connection accept.
2696 	 */
2697 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2698 		DTRACE_PROBE(krpc__e__connmgr__connect__cantsend);
2699 		freemsg(mp);
2700 		return (FALSE);
2701 	}
2702 
2703 	mutex_enter(&clnt_pending_lock);
2704 
2705 	/*
2706 	 * We wait for the transport connection to be made, or an
2707 	 * indication that it could not be made.
2708 	 */
2709 	interrupted = 0;
2710 
2711 	/*
2712 	 * waitforack should have been called with T_OK_ACK, but the
2713 	 * present implementation needs to be passed T_INFO_ACK to
2714 	 * work correctly.
2715 	 */
2716 	error = waitforack(e, T_INFO_ACK, waitp, nosignal);
2717 	if (error == EINTR)
2718 		interrupted = 1;
2719 	if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) {
2720 		/*
2721 		 * No time to lose; we essentially have been signaled to
2722 		 * quit.
2723 		 */
2724 		interrupted = 1;
2725 	}
2726 #ifdef RPCDEBUG
2727 	if (error == ETIME)
2728 		RPCLOG0(8, "connmgr_connect: giving up "
2729 		    "on connection attempt; "
2730 		    "clnt_dispatch notifyconn "
2731 		    "diagnostic 'no one waiting for "
2732 		    "connection' should not be "
2733 		    "unexpected\n");
2734 #endif
2735 	if (e->call_prev)
2736 		e->call_prev->call_next = e->call_next;
2737 	else
2738 		clnt_pending = e->call_next;
2739 	if (e->call_next)
2740 		e->call_next->call_prev = e->call_prev;
2741 	mutex_exit(&clnt_pending_lock);
2742 
2743 	if (e->call_status != RPC_SUCCESS || error != 0) {
2744 		if (interrupted)
2745 			e->call_status = RPC_INTR;
2746 		else if (error == ETIME)
2747 			e->call_status = RPC_TIMEDOUT;
2748 		else if (error == EPROTO) {
2749 			e->call_status = RPC_SYSTEMERROR;
2750 			e->call_reason = EPROTO;
2751 		}
2752 
2753 		RPCLOG(8, "connmgr_connect: can't connect, status: "
2754 		    "%s\n", clnt_sperrno(e->call_status));
2755 
2756 		if (e->call_reply) {
2757 			freemsg(e->call_reply);
2758 			e->call_reply = NULL;
2759 		}
2760 
2761 		return (FALSE);
2762 	}
2763 	/*
2764 	 * The result of the "connection accept" is a T_info_ack
2765 	 * in the call_reply field.
2766 	 */
2767 	ASSERT(e->call_reply != NULL);
2768 	mp = e->call_reply;
2769 	e->call_reply = NULL;
2770 	tinfo = (struct T_info_ack *)mp->b_rptr;
2771 
2772 	tidu_size = tinfo->TIDU_size;
2773 	tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT);
2774 	if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0))
2775 		tidu_size = COTS_DEFAULT_ALLOCSIZE;
2776 	*tidu_ptr = tidu_size;
2777 
2778 	freemsg(mp);
2779 
2780 	/*
2781 	 * Set up the pertinent options.  NODELAY is so the transport doesn't
2782 	 * buffer up RPC messages on either end.  This may not be valid for
2783 	 * all transports.  Failure to set this option is not cause to
2784 	 * bail out so we return success anyway.  Note that lack of NODELAY
2785 	 * or some other way to flush the message on both ends will cause
2786 	 * lots of retries and terrible performance.
2787 	 */
2788 	if (addrfmly == AF_INET || addrfmly == AF_INET6) {
2789 		(void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e, cr);
2790 		if (e->call_status == RPC_XPRTFAILED)
2791 			return (FALSE);
2792 	}
2793 
2794 	/*
2795 	 * Since we have a connection, we now need to figure out if
2796 	 * we need to create a kstat. If x_ksp is not NULL then we
2797 	 * are reusing a connection and so we do not need to create
2798 	 * another kstat -- lets just return.
2799 	 */
2800 	if (cm_entry->x_ksp != NULL)
2801 		return (TRUE);
2802 
2803 	/*
2804 	 * We need to increment rpc_kstat_instance atomically to prevent
2805 	 * two kstats being created with the same instance.
2806 	 */
2807 	kstat_instance = atomic_inc_32_nv((uint32_t *)&rpc_kstat_instance);
2808 
2809 	if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance,
2810 	    "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED,
2811 	    (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)),
2812 	    KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) {
2813 		return (TRUE);
2814 	}
2815 
2816 	cm_entry->x_ksp->ks_lock = &connmgr_lock;
2817 	cm_entry->x_ksp->ks_private = cm_entry;
2818 	cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char))
2819 	    + sizeof (cm_kstat_template));
2820 	cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size,
2821 	    KM_SLEEP);
2822 	bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data,
2823 	    cm_entry->x_ksp->ks_data_size);
2824 	((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2825 	    x_server.value.str.addr.ptr =
2826 	    kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP);
2827 
2828 	cm_entry->x_ksp->ks_update = conn_kstat_update;
2829 	kstat_install(cm_entry->x_ksp);
2830 	return (TRUE);
2831 }
2832 
2833 /*
2834  * Verify that the specified offset falls within the mblk and
2835  * that the resulting pointer is aligned.
2836  * Returns NULL if not.
2837  *
2838  * code from fs/sockfs/socksubr.c
2839  */
2840 static void *
2841 connmgr_opt_getoff(mblk_t *mp, t_uscalar_t offset,
2842     t_uscalar_t length, uint_t align_size)
2843 {
2844 	uintptr_t ptr1, ptr2;
2845 
2846 	ASSERT(mp && mp->b_wptr >= mp->b_rptr);
2847 	ptr1 = (uintptr_t)mp->b_rptr + offset;
2848 	ptr2 = (uintptr_t)ptr1 + length;
2849 	if (ptr1 < (uintptr_t)mp->b_rptr || ptr2 > (uintptr_t)mp->b_wptr) {
2850 		return (NULL);
2851 	}
2852 	if ((ptr1 & (align_size - 1)) != 0) {
2853 		return (NULL);
2854 	}
2855 	return ((void *)ptr1);
2856 }
2857 
2858 static bool_t
2859 connmgr_getopt_int(queue_t *wq, int level, int name, int *val,
2860     calllist_t *e, cred_t *cr)
2861 {
2862 	mblk_t *mp;
2863 	struct opthdr *opt, *opt_res;
2864 	struct T_optmgmt_req *tor;
2865 	struct T_optmgmt_ack *opt_ack;
2866 	struct timeval waitp;
2867 	int error;
2868 
2869 	mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2870 	    sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2871 	if (mp == NULL)
2872 		return (FALSE);
2873 
2874 	mp->b_datap->db_type = M_PROTO;
2875 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2876 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2877 	tor->MGMT_flags = T_CURRENT;
2878 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2879 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2880 
2881 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2882 	opt->level = level;
2883 	opt->name = name;
2884 	opt->len = sizeof (int);
2885 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2886 	    sizeof (int);
2887 
2888 	/*
2889 	 * We will use this connection regardless
2890 	 * of whether or not the option is readable.
2891 	 */
2892 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2893 		DTRACE_PROBE(krpc__e__connmgr__getopt__cantsend);
2894 		freemsg(mp);
2895 		return (FALSE);
2896 	}
2897 
2898 	mutex_enter(&clnt_pending_lock);
2899 
2900 	waitp.tv_sec = clnt_cots_min_conntout;
2901 	waitp.tv_usec = 0;
2902 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2903 
2904 	if (e->call_prev)
2905 		e->call_prev->call_next = e->call_next;
2906 	else
2907 		clnt_pending = e->call_next;
2908 	if (e->call_next)
2909 		e->call_next->call_prev = e->call_prev;
2910 	mutex_exit(&clnt_pending_lock);
2911 
2912 	/* get reply message */
2913 	mp = e->call_reply;
2914 	e->call_reply = NULL;
2915 
2916 	if ((!mp) || (e->call_status != RPC_SUCCESS) || (error != 0)) {
2917 
2918 		DTRACE_PROBE4(krpc__e__connmgr_getopt, int, name,
2919 		    int, e->call_status, int, error, mblk_t *, mp);
2920 
2921 		if (mp)
2922 			freemsg(mp);
2923 		return (FALSE);
2924 	}
2925 
2926 	opt_ack = (struct T_optmgmt_ack *)mp->b_rptr;
2927 	opt_res = (struct opthdr *)connmgr_opt_getoff(mp, opt_ack->OPT_offset,
2928 	    opt_ack->OPT_length, __TPI_ALIGN_SIZE);
2929 
2930 	if (!opt_res) {
2931 		DTRACE_PROBE4(krpc__e__connmgr_optres, mblk_t *, mp, int, name,
2932 		    int, opt_ack->OPT_offset, int, opt_ack->OPT_length);
2933 		freemsg(mp);
2934 		return (FALSE);
2935 	}
2936 	*val = *(int *)&opt_res[1];
2937 
2938 	DTRACE_PROBE2(connmgr_getopt__ok, int, name, int, *val);
2939 
2940 	freemsg(mp);
2941 	return (TRUE);
2942 }
2943 
2944 /*
2945  * Called by connmgr_connect to set an option on the new stream.
2946  */
2947 static bool_t
2948 connmgr_setopt_int(queue_t *wq, int level, int name, int val,
2949     calllist_t *e, cred_t *cr)
2950 {
2951 	mblk_t *mp;
2952 	struct opthdr *opt;
2953 	struct T_optmgmt_req *tor;
2954 	struct timeval waitp;
2955 	int error;
2956 
2957 	mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2958 	    sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2959 	if (mp == NULL) {
2960 		RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option "
2961 		    "request\n");
2962 		return (FALSE);
2963 	}
2964 
2965 	mp->b_datap->db_type = M_PROTO;
2966 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2967 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2968 	tor->MGMT_flags = T_NEGOTIATE;
2969 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2970 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2971 
2972 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2973 	opt->level = level;
2974 	opt->name = name;
2975 	opt->len = sizeof (int);
2976 	*(int *)((char *)opt + sizeof (*opt)) = val;
2977 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2978 	    sizeof (int);
2979 
2980 	/*
2981 	 * We will use this connection regardless
2982 	 * of whether or not the option is settable.
2983 	 */
2984 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2985 		DTRACE_PROBE(krpc__e__connmgr__setopt__cantsend);
2986 		freemsg(mp);
2987 		return (FALSE);
2988 	}
2989 
2990 	mutex_enter(&clnt_pending_lock);
2991 
2992 	waitp.tv_sec = clnt_cots_min_conntout;
2993 	waitp.tv_usec = 0;
2994 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2995 
2996 	if (e->call_prev)
2997 		e->call_prev->call_next = e->call_next;
2998 	else
2999 		clnt_pending = e->call_next;
3000 	if (e->call_next)
3001 		e->call_next->call_prev = e->call_prev;
3002 	mutex_exit(&clnt_pending_lock);
3003 
3004 	if (e->call_reply != NULL) {
3005 		freemsg(e->call_reply);
3006 		e->call_reply = NULL;
3007 	}
3008 
3009 	if (e->call_status != RPC_SUCCESS || error != 0) {
3010 		RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name);
3011 		return (FALSE);
3012 	}
3013 	RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name);
3014 	return (TRUE);
3015 }
3016 
3017 static bool_t
3018 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e, cred_t *cr)
3019 {
3020 	return (connmgr_setopt_int(wq, level, name, 1, e, cr));
3021 }
3022 
3023 #ifdef	DEBUG
3024 
3025 /*
3026  * This is a knob to let us force code coverage in allocation failure
3027  * case.
3028  */
3029 static int	connmgr_failsnd;
3030 #define	CONN_SND_ALLOC(Size, Pri)	\
3031 	((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri))
3032 
3033 #else
3034 
3035 #define	CONN_SND_ALLOC(Size, Pri)	allocb(Size, Pri)
3036 
3037 #endif
3038 
3039 /*
3040  * Sends an orderly release on the specified queue.
3041  * Entered with connmgr_lock. Exited without connmgr_lock
3042  */
3043 static void
3044 connmgr_sndrel(struct cm_xprt *cm_entry)
3045 {
3046 	struct T_ordrel_req *torr;
3047 	mblk_t *mp;
3048 	queue_t *q = cm_entry->x_wq;
3049 	ASSERT(MUTEX_HELD(&connmgr_lock));
3050 	mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO);
3051 	if (mp == NULL) {
3052 		cm_entry->x_needrel = TRUE;
3053 		mutex_exit(&connmgr_lock);
3054 		RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel "
3055 		    "to queue %p\n", (void *)q);
3056 		return;
3057 	}
3058 	mutex_exit(&connmgr_lock);
3059 
3060 	mp->b_datap->db_type = M_PROTO;
3061 	torr = (struct T_ordrel_req *)(mp->b_rptr);
3062 	torr->PRIM_type = T_ORDREL_REQ;
3063 	mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req);
3064 
3065 	RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q);
3066 	put(q, mp);
3067 }
3068 
3069 /*
3070  * Sends an disconnect on the specified queue.
3071  * Entered with connmgr_lock. Exited without connmgr_lock
3072  */
3073 static void
3074 connmgr_snddis(struct cm_xprt *cm_entry)
3075 {
3076 	struct T_discon_req *tdis;
3077 	mblk_t *mp;
3078 	queue_t *q = cm_entry->x_wq;
3079 
3080 	ASSERT(MUTEX_HELD(&connmgr_lock));
3081 	mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO);
3082 	if (mp == NULL) {
3083 		cm_entry->x_needdis = TRUE;
3084 		mutex_exit(&connmgr_lock);
3085 		RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon "
3086 		    "to queue %p\n", (void *)q);
3087 		return;
3088 	}
3089 	mutex_exit(&connmgr_lock);
3090 
3091 	mp->b_datap->db_type = M_PROTO;
3092 	tdis = (struct T_discon_req *)mp->b_rptr;
3093 	tdis->PRIM_type = T_DISCON_REQ;
3094 	mp->b_wptr = mp->b_rptr + sizeof (*tdis);
3095 
3096 	RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q);
3097 	put(q, mp);
3098 }
3099 
3100 /*
3101  * Sets up the entry for receiving replies, and calls rpcmod's write put proc
3102  * (through put) to send the call.
3103  */
3104 static int
3105 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid,
3106     uint_t queue_flag)
3107 {
3108 	ASSERT(e != NULL);
3109 
3110 	e->call_status = RPC_TIMEDOUT;	/* optimistic, eh? */
3111 	e->call_reason = 0;
3112 	e->call_wq = q;
3113 	e->call_xid = xid;
3114 	e->call_notified = FALSE;
3115 
3116 	if (!canput(q)) {
3117 		e->call_status = RPC_CANTSEND;
3118 		e->call_reason = ENOBUFS;
3119 		return (RPC_CANTSEND);
3120 	}
3121 
3122 	/*
3123 	 * If queue_flag is set then the calllist_t is already on the hash
3124 	 * queue.  In this case just send the message and return.
3125 	 */
3126 	if (queue_flag) {
3127 		put(q, mp);
3128 		return (RPC_SUCCESS);
3129 
3130 	}
3131 
3132 	/*
3133 	 * Set up calls for RPC requests (with XID != 0) on the hash
3134 	 * queue for fast lookups and place other calls (i.e.
3135 	 * connection management) on the linked list.
3136 	 */
3137 	if (xid != 0) {
3138 		RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on "
3139 		    "dispatch list\n", xid);
3140 		e->call_hash = call_hash(xid, clnt_cots_hash_size);
3141 		e->call_bucket = &cots_call_ht[e->call_hash];
3142 		call_table_enter(e);
3143 	} else {
3144 		mutex_enter(&clnt_pending_lock);
3145 		if (clnt_pending)
3146 			clnt_pending->call_prev = e;
3147 		e->call_next = clnt_pending;
3148 		e->call_prev = NULL;
3149 		clnt_pending = e;
3150 		mutex_exit(&clnt_pending_lock);
3151 	}
3152 
3153 	put(q, mp);
3154 	return (RPC_SUCCESS);
3155 }
3156 
3157 /*
3158  * Called by rpcmod to notify a client with a clnt_pending call that its reply
3159  * has arrived.  If we can't find a client waiting for this reply, we log
3160  * the error and return.
3161  */
3162 bool_t
3163 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid)
3164 {
3165 	calllist_t *e = NULL;
3166 	call_table_t *chtp;
3167 	uint32_t xid;
3168 	uint_t hash;
3169 
3170 	if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) &&
3171 	    (mp->b_wptr - mp->b_rptr) >= sizeof (xid))
3172 		xid = *((uint32_t *)mp->b_rptr);
3173 	else {
3174 		int i = 0;
3175 		unsigned char *p = (unsigned char *)&xid;
3176 		unsigned char *rptr;
3177 		mblk_t *tmp = mp;
3178 
3179 		/*
3180 		 * Copy the xid, byte-by-byte into xid.
3181 		 */
3182 		while (tmp) {
3183 			rptr = tmp->b_rptr;
3184 			while (rptr < tmp->b_wptr) {
3185 				*p++ = *rptr++;
3186 				if (++i >= sizeof (xid))
3187 					goto done_xid_copy;
3188 			}
3189 			tmp = tmp->b_cont;
3190 		}
3191 
3192 		/*
3193 		 * If we got here, we ran out of mblk space before the
3194 		 * xid could be copied.
3195 		 */
3196 		ASSERT(tmp == NULL && i < sizeof (xid));
3197 
3198 		RPCLOG0(1,
3199 		    "clnt_dispatch_notify: message less than size of xid\n");
3200 		return (FALSE);
3201 
3202 	}
3203 done_xid_copy:
3204 
3205 	hash = call_hash(xid, clnt_cots_hash_size);
3206 	chtp = &cots_call_ht[hash];
3207 	/* call_table_find returns with the hash bucket locked */
3208 	call_table_find(chtp, xid, e);
3209 
3210 	if (e != NULL) {
3211 		/*
3212 		 * Found thread waiting for this reply
3213 		 */
3214 		mutex_enter(&e->call_lock);
3215 
3216 		/*
3217 		 * verify that the reply is coming in on
3218 		 * the same zone that it was sent from.
3219 		 */
3220 		if (e->call_zoneid != zoneid) {
3221 			mutex_exit(&e->call_lock);
3222 			mutex_exit(&chtp->ct_lock);
3223 			RPCLOG0(1, "clnt_dispatch_notify: incorrect zoneid\n");
3224 			return (FALSE);
3225 		}
3226 
3227 		if (e->call_reply)
3228 			/*
3229 			 * This can happen under the following scenario:
3230 			 * clnt_cots_kcallit() times out on the response,
3231 			 * rfscall() repeats the CLNT_CALL() with
3232 			 * the same xid, clnt_cots_kcallit() sends the retry,
3233 			 * thereby putting the clnt handle on the pending list,
3234 			 * the first response arrives, signalling the thread
3235 			 * in clnt_cots_kcallit(). Before that thread is
3236 			 * dispatched, the second response arrives as well,
3237 			 * and clnt_dispatch_notify still finds the handle on
3238 			 * the pending list, with call_reply set. So free the
3239 			 * old reply now.
3240 			 *
3241 			 * It is also possible for a response intended for
3242 			 * an RPC call with a different xid to reside here.
3243 			 * This can happen if the thread that owned this
3244 			 * client handle prior to the current owner bailed
3245 			 * out and left its call record on the dispatch
3246 			 * queue.  A window exists where the response can
3247 			 * arrive before the current owner dispatches its
3248 			 * RPC call.
3249 			 *
3250 			 * In any case, this is the very last point where we
3251 			 * can safely check the call_reply field before
3252 			 * placing the new response there.
3253 			 */
3254 			freemsg(e->call_reply);
3255 		e->call_reply = mp;
3256 		e->call_status = RPC_SUCCESS;
3257 		e->call_notified = TRUE;
3258 		cv_signal(&e->call_cv);
3259 		mutex_exit(&e->call_lock);
3260 		mutex_exit(&chtp->ct_lock);
3261 		return (TRUE);
3262 	} else {
3263 		zone_t *zone;
3264 		struct rpcstat *rpcstat;
3265 
3266 		mutex_exit(&chtp->ct_lock);
3267 		RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n",
3268 		    xid);
3269 		/*
3270 		 * This is unfortunate, but we need to lookup the zone so we
3271 		 * can increment its "rcbadxids" counter.
3272 		 */
3273 		zone = zone_find_by_id(zoneid);
3274 		if (zone == NULL) {
3275 			/*
3276 			 * The zone went away...
3277 			 */
3278 			return (FALSE);
3279 		}
3280 		rpcstat = zone_getspecific(rpcstat_zone_key, zone);
3281 		if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) {
3282 			/*
3283 			 * Not interested
3284 			 */
3285 			zone_rele(zone);
3286 			return (FALSE);
3287 		}
3288 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids);
3289 		zone_rele(zone);
3290 	}
3291 	return (FALSE);
3292 }
3293 
3294 /*
3295  * Called by rpcmod when a non-data indication arrives.  The ones in which we
3296  * are interested are connection indications and options acks.  We dispatch
3297  * based on the queue the indication came in on.  If we are not interested in
3298  * what came in, we return false to rpcmod, who will then pass it upstream.
3299  */
3300 bool_t
3301 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp)
3302 {
3303 	calllist_t *e;
3304 	int type;
3305 
3306 	ASSERT((q->q_flag & QREADR) == 0);
3307 
3308 	type = ((union T_primitives *)mp->b_rptr)->type;
3309 	RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n",
3310 	    rpc_tpiprim2name(type));
3311 	mutex_enter(&clnt_pending_lock);
3312 	for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) {
3313 		if (e == NULL) {
3314 			mutex_exit(&clnt_pending_lock);
3315 			RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting "
3316 			    "for connection on queue 0x%p\n", (void *)q);
3317 			return (FALSE);
3318 		}
3319 		if (e->call_wq == q)
3320 			break;
3321 	}
3322 
3323 	switch (type) {
3324 	case T_CONN_CON:
3325 		/*
3326 		 * The transport is now connected, send a T_INFO_REQ to get
3327 		 * the tidu size.
3328 		 */
3329 		mutex_exit(&clnt_pending_lock);
3330 		ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >=
3331 		    sizeof (struct T_info_req));
3332 		mp->b_rptr = mp->b_datap->db_base;
3333 		((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ;
3334 		mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req);
3335 		mp->b_datap->db_type = M_PCPROTO;
3336 		put(q, mp);
3337 		return (TRUE);
3338 	case T_INFO_ACK:
3339 	case T_OPTMGMT_ACK:
3340 		e->call_status = RPC_SUCCESS;
3341 		e->call_reply = mp;
3342 		e->call_notified = TRUE;
3343 		cv_signal(&e->call_cv);
3344 		break;
3345 	case T_ERROR_ACK:
3346 		e->call_status = RPC_CANTCONNECT;
3347 		e->call_reply = mp;
3348 		e->call_notified = TRUE;
3349 		cv_signal(&e->call_cv);
3350 		break;
3351 	case T_OK_ACK:
3352 		/*
3353 		 * Great, but we are really waiting for a T_CONN_CON
3354 		 */
3355 		freemsg(mp);
3356 		break;
3357 	default:
3358 		mutex_exit(&clnt_pending_lock);
3359 		RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type);
3360 		return (FALSE);
3361 	}
3362 
3363 	mutex_exit(&clnt_pending_lock);
3364 	return (TRUE);
3365 }
3366 
3367 /*
3368  * Called by rpcmod when the transport is (or should be) going away.  Informs
3369  * all callers waiting for replies and marks the entry in the connection
3370  * manager's list as unconnected, and either closing (close handshake in
3371  * progress) or dead.
3372  */
3373 void
3374 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason)
3375 {
3376 	calllist_t *e;
3377 	call_table_t *ctp;
3378 	struct cm_xprt *cm_entry;
3379 	int have_connmgr_lock;
3380 	int i;
3381 
3382 	ASSERT((q->q_flag & QREADR) == 0);
3383 
3384 	RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q);
3385 	RPCLOG(1, " received a notifcation prim type [%s]",
3386 	    rpc_tpiprim2name(msg_type));
3387 	RPCLOG(1, " and reason %d\n", reason);
3388 
3389 	/*
3390 	 * Find the transport entry in the connection manager's list, close
3391 	 * the transport and delete the entry.  In the case where rpcmod's
3392 	 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we
3393 	 * should gracefully close the connection.
3394 	 */
3395 	have_connmgr_lock = 1;
3396 	mutex_enter(&connmgr_lock);
3397 	for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) {
3398 		ASSERT(cm_entry != cm_entry->x_next);
3399 		if (cm_entry->x_wq == q) {
3400 			ASSERT(MUTEX_HELD(&connmgr_lock));
3401 			ASSERT(have_connmgr_lock == 1);
3402 			switch (msg_type) {
3403 			case T_ORDREL_REQ:
3404 
3405 				if (cm_entry->x_dead) {
3406 					RPCLOG(1, "idle timeout on dead "
3407 					    "connection: %p\n",
3408 					    (void *)cm_entry);
3409 					if (clnt_stop_idle != NULL)
3410 						(*clnt_stop_idle)(q);
3411 					break;
3412 				}
3413 
3414 				/*
3415 				 * Only mark the connection as dead if it is
3416 				 * connected and idle.
3417 				 * An unconnected connection has probably
3418 				 * gone idle because the server is down,
3419 				 * and when it comes back up there will be
3420 				 * retries that need to use that connection.
3421 				 */
3422 				if (cm_entry->x_connected ||
3423 				    cm_entry->x_doomed) {
3424 					if (cm_entry->x_ordrel) {
3425 						if (cm_entry->x_closing ==
3426 						    TRUE) {
3427 							/*
3428 							 * The connection is
3429 							 * obviously wedged due
3430 							 * to a bug or problem
3431 							 * with the transport.
3432 							 * Mark it as dead.
3433 							 * Otherwise we can
3434 							 * leak connections.
3435 							 */
3436 							cm_entry->x_dead = TRUE;
3437 							mutex_exit(
3438 							    &connmgr_lock);
3439 							have_connmgr_lock = 0;
3440 							if (clnt_stop_idle !=
3441 							    NULL)
3442 							(*clnt_stop_idle)(q);
3443 							break;
3444 						}
3445 						cm_entry->x_closing = TRUE;
3446 						connmgr_sndrel(cm_entry);
3447 						have_connmgr_lock = 0;
3448 					} else {
3449 						cm_entry->x_dead = TRUE;
3450 						mutex_exit(&connmgr_lock);
3451 						have_connmgr_lock = 0;
3452 						if (clnt_stop_idle != NULL)
3453 							(*clnt_stop_idle)(q);
3454 					}
3455 				} else {
3456 					/*
3457 					 * We don't mark the connection
3458 					 * as dead, but we turn off the
3459 					 * idle timer.
3460 					 */
3461 					mutex_exit(&connmgr_lock);
3462 					have_connmgr_lock = 0;
3463 					if (clnt_stop_idle != NULL)
3464 						(*clnt_stop_idle)(q);
3465 					RPCLOG(1, "clnt_dispatch_notifyall:"
3466 					    " ignoring timeout from rpcmod"
3467 					    " (q %p) because we are not "
3468 					    " connected\n", (void *)q);
3469 				}
3470 				break;
3471 			case T_ORDREL_IND:
3472 				/*
3473 				 * If this entry is marked closing, then we are
3474 				 * completing a close handshake, and the
3475 				 * connection is dead.  Otherwise, the server is
3476 				 * trying to close. Since the server will not
3477 				 * be sending any more RPC replies, we abort
3478 				 * the connection, including flushing
3479 				 * any RPC requests that are in-transit.
3480 				 * In either case, mark the entry as dead so
3481 				 * that it can be closed by the connection
3482 				 * manager's garbage collector.
3483 				 */
3484 				cm_entry->x_dead = TRUE;
3485 				if (cm_entry->x_closing) {
3486 					mutex_exit(&connmgr_lock);
3487 					have_connmgr_lock = 0;
3488 					if (clnt_stop_idle != NULL)
3489 						(*clnt_stop_idle)(q);
3490 				} else {
3491 					/*
3492 					 * if we're getting a disconnect
3493 					 * before we've finished our
3494 					 * connect attempt, mark it for
3495 					 * later processing
3496 					 */
3497 					if (cm_entry->x_thread)
3498 						cm_entry->x_early_disc = TRUE;
3499 					else
3500 						cm_entry->x_connected = FALSE;
3501 					cm_entry->x_waitdis = TRUE;
3502 					connmgr_snddis(cm_entry);
3503 					have_connmgr_lock = 0;
3504 				}
3505 				break;
3506 
3507 			case T_ERROR_ACK:
3508 			case T_OK_ACK:
3509 				cm_entry->x_waitdis = FALSE;
3510 				cv_signal(&cm_entry->x_dis_cv);
3511 				mutex_exit(&connmgr_lock);
3512 				return;
3513 
3514 			case T_DISCON_REQ:
3515 				if (cm_entry->x_thread)
3516 					cm_entry->x_early_disc = TRUE;
3517 				else
3518 					cm_entry->x_connected = FALSE;
3519 				cm_entry->x_waitdis = TRUE;
3520 
3521 				connmgr_snddis(cm_entry);
3522 				have_connmgr_lock = 0;
3523 				break;
3524 
3525 			case T_DISCON_IND:
3526 			default:
3527 				/*
3528 				 * if we're getting a disconnect before
3529 				 * we've finished our connect attempt,
3530 				 * mark it for later processing
3531 				 */
3532 				if (cm_entry->x_closing) {
3533 					cm_entry->x_dead = TRUE;
3534 					mutex_exit(&connmgr_lock);
3535 					have_connmgr_lock = 0;
3536 					if (clnt_stop_idle != NULL)
3537 						(*clnt_stop_idle)(q);
3538 				} else {
3539 					if (cm_entry->x_thread) {
3540 						cm_entry->x_early_disc = TRUE;
3541 					} else {
3542 						cm_entry->x_dead = TRUE;
3543 						cm_entry->x_connected = FALSE;
3544 					}
3545 				}
3546 				break;
3547 			}
3548 			break;
3549 		}
3550 	}
3551 
3552 	if (have_connmgr_lock)
3553 		mutex_exit(&connmgr_lock);
3554 
3555 	if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) {
3556 		RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find "
3557 		    "connmgr entry for discon ack\n", (void *)q);
3558 		return;
3559 	}
3560 
3561 	/*
3562 	 * Then kick all the clnt_pending calls out of their wait.  There
3563 	 * should be no clnt_pending calls in the case of rpcmod's idle
3564 	 * timer firing.
3565 	 */
3566 	for (i = 0; i < clnt_cots_hash_size; i++) {
3567 		ctp = &cots_call_ht[i];
3568 		mutex_enter(&ctp->ct_lock);
3569 		for (e = ctp->ct_call_next;
3570 		    e != (calllist_t *)ctp;
3571 		    e = e->call_next) {
3572 			if (e->call_wq == q && e->call_notified == FALSE) {
3573 				RPCLOG(1,
3574 				    "clnt_dispatch_notifyall for queue %p ",
3575 				    (void *)q);
3576 				RPCLOG(1, "aborting clnt_pending call %p\n",
3577 				    (void *)e);
3578 
3579 				if (msg_type == T_DISCON_IND)
3580 					e->call_reason = reason;
3581 				e->call_notified = TRUE;
3582 				e->call_status = RPC_XPRTFAILED;
3583 				cv_signal(&e->call_cv);
3584 			}
3585 		}
3586 		mutex_exit(&ctp->ct_lock);
3587 	}
3588 
3589 	mutex_enter(&clnt_pending_lock);
3590 	for (e = clnt_pending; e; e = e->call_next) {
3591 		/*
3592 		 * Only signal those RPC handles that haven't been
3593 		 * signalled yet. Otherwise we can get a bogus call_reason.
3594 		 * This can happen if thread A is making a call over a
3595 		 * connection. If the server is killed, it will cause
3596 		 * reset, and reason will default to EIO as a result of
3597 		 * a T_ORDREL_IND. Thread B then attempts to recreate
3598 		 * the connection but gets a T_DISCON_IND. If we set the
3599 		 * call_reason code for all threads, then if thread A
3600 		 * hasn't been dispatched yet, it will get the wrong
3601 		 * reason. The bogus call_reason can make it harder to
3602 		 * discriminate between calls that fail because the
3603 		 * connection attempt failed versus those where the call
3604 		 * may have been executed on the server.
3605 		 */
3606 		if (e->call_wq == q && e->call_notified == FALSE) {
3607 			RPCLOG(1, "clnt_dispatch_notifyall for queue %p ",
3608 			    (void *)q);
3609 			RPCLOG(1, " aborting clnt_pending call %p\n",
3610 			    (void *)e);
3611 
3612 			if (msg_type == T_DISCON_IND)
3613 				e->call_reason = reason;
3614 			e->call_notified = TRUE;
3615 			/*
3616 			 * Let the caller timeout, else it will retry
3617 			 * immediately.
3618 			 */
3619 			e->call_status = RPC_XPRTFAILED;
3620 
3621 			/*
3622 			 * We used to just signal those threads
3623 			 * waiting for a connection, (call_xid = 0).
3624 			 * That meant that threads waiting for a response
3625 			 * waited till their timeout expired. This
3626 			 * could be a long time if they've specified a
3627 			 * maximum timeout. (2^31 - 1). So we
3628 			 * Signal all threads now.
3629 			 */
3630 			cv_signal(&e->call_cv);
3631 		}
3632 	}
3633 	mutex_exit(&clnt_pending_lock);
3634 }
3635 
3636 
3637 /*ARGSUSED*/
3638 /*
3639  * after resuming a system that's been suspended for longer than the
3640  * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall()
3641  * generates "NFS server X not responding" and "NFS server X ok" messages;
3642  * here we reset inet connections to cause a re-connect and avoid those
3643  * NFS messages.  see 4045054
3644  */
3645 boolean_t
3646 connmgr_cpr_reset(void *arg, int code)
3647 {
3648 	struct cm_xprt *cxp;
3649 
3650 	if (code == CB_CODE_CPR_CHKPT)
3651 		return (B_TRUE);
3652 
3653 	if (mutex_tryenter(&connmgr_lock) == 0)
3654 		return (B_FALSE);
3655 	for (cxp = cm_hd; cxp; cxp = cxp->x_next) {
3656 		if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) &&
3657 		    cxp->x_connected == TRUE) {
3658 			if (cxp->x_thread)
3659 				cxp->x_early_disc = TRUE;
3660 			else
3661 				cxp->x_connected = FALSE;
3662 			cxp->x_needdis = TRUE;
3663 		}
3664 	}
3665 	mutex_exit(&connmgr_lock);
3666 	return (B_TRUE);
3667 }
3668 
3669 void
3670 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp)
3671 {
3672 
3673 	*statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid,
3674 	    "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl,
3675 	    sizeof (cots_rcstat_tmpl));
3676 }
3677 
3678 void
3679 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp)
3680 {
3681 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client");
3682 	kmem_free(*statsp, sizeof (cots_rcstat_tmpl));
3683 }
3684 
3685 void
3686 clnt_cots_init(void)
3687 {
3688 	mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL);
3689 	mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL);
3690 
3691 	if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE)
3692 		clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE;
3693 
3694 	cots_call_ht = call_table_init(clnt_cots_hash_size);
3695 	zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy);
3696 }
3697 
3698 void
3699 clnt_cots_fini(void)
3700 {
3701 	(void) zone_key_delete(zone_cots_key);
3702 }
3703 
3704 /*
3705  * Wait for TPI ack, returns success only if expected ack is received
3706  * within timeout period.
3707  */
3708 
3709 static int
3710 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp,
3711     bool_t nosignal)
3712 {
3713 	union T_primitives *tpr;
3714 	clock_t timout;
3715 	int cv_stat = 1;
3716 
3717 	ASSERT(MUTEX_HELD(&clnt_pending_lock));
3718 	while (e->call_reply == NULL) {
3719 		if (waitp != NULL) {
3720 			timout = waitp->tv_sec * drv_usectohz(MICROSEC) +
3721 			    drv_usectohz(waitp->tv_usec);
3722 			if (nosignal)
3723 				cv_stat = cv_reltimedwait(&e->call_cv,
3724 				    &clnt_pending_lock, timout, TR_CLOCK_TICK);
3725 			else
3726 				cv_stat = cv_reltimedwait_sig(&e->call_cv,
3727 				    &clnt_pending_lock, timout, TR_CLOCK_TICK);
3728 		} else {
3729 			if (nosignal)
3730 				cv_wait(&e->call_cv, &clnt_pending_lock);
3731 			else
3732 				cv_stat = cv_wait_sig(&e->call_cv,
3733 				    &clnt_pending_lock);
3734 		}
3735 		if (cv_stat == -1)
3736 			return (ETIME);
3737 		if (cv_stat == 0)
3738 			return (EINTR);
3739 		/*
3740 		 * if we received an error from the server and we know a reply
3741 		 * is not going to be sent, do not wait for the full timeout,
3742 		 * return now.
3743 		 */
3744 		if (e->call_status == RPC_XPRTFAILED)
3745 			return (e->call_reason);
3746 	}
3747 	tpr = (union T_primitives *)e->call_reply->b_rptr;
3748 	if (tpr->type == ack_prim)
3749 		return (0); /* Success */
3750 
3751 	if (tpr->type == T_ERROR_ACK) {
3752 		if (tpr->error_ack.TLI_error == TSYSERR)
3753 			return (tpr->error_ack.UNIX_error);
3754 		else
3755 			return (t_tlitosyserr(tpr->error_ack.TLI_error));
3756 	}
3757 
3758 	return (EPROTO); /* unknown or unexpected primitive */
3759 }
3760