xref: /illumos-gate/usr/src/uts/common/rpc/clnt_cots.c (revision c5bab7026b8e0ac44b25ee08507ea360f177d844)
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  */
26 
27 /*
28  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
29  * Use is subject to license terms.
30  */
31 
32 /*
33  * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T
34  *		All Rights Reserved
35  */
36 
37 /*
38  * Portions of this source code were derived from Berkeley 4.3 BSD
39  * under license from the Regents of the University of California.
40  */
41 
42 
43 /*
44  * Implements a kernel based, client side RPC over Connection Oriented
45  * Transports (COTS).
46  */
47 
48 /*
49  * Much of this file has been re-written to let NFS work better over slow
50  * transports. A description follows.
51  *
52  * One of the annoying things about kRPC/COTS is that it will temporarily
53  * create more than one connection between a client and server. This
54  * happens because when a connection is made, the end-points entry in the
55  * linked list of connections (headed by cm_hd), is removed so that other
56  * threads don't mess with it. Went ahead and bit the bullet by keeping
57  * the endpoint on the connection list and introducing state bits,
58  * condition variables etc. to the connection entry data structure (struct
59  * cm_xprt).
60  *
61  * Here is a summary of the changes to cm-xprt:
62  *
63  *	x_ctime is the timestamp of when the endpoint was last
64  *	connected or disconnected. If an end-point is ever disconnected
65  *	or re-connected, then any outstanding RPC request is presumed
66  *	lost, telling clnt_cots_kcallit that it needs to re-send the
67  *	request, not just wait for the original request's reply to
68  *	arrive.
69  *
70  *	x_thread flag which tells us if a thread is doing a connection attempt.
71  *
72  *	x_waitdis flag which tells us we are waiting a disconnect ACK.
73  *
74  *	x_needdis flag which tells us we need to send a T_DISCONN_REQ
75  *	to kill the connection.
76  *
77  *	x_needrel flag which tells us we need to send a T_ORDREL_REQ to
78  *	gracefully close the connection.
79  *
80  *	#defined bitmasks for the all the b_* bits so that more
81  *	efficient (and at times less clumsy) masks can be used to
82  *	manipulated state in cases where multiple bits have to
83  *	set/cleared/checked in the same critical section.
84  *
85  *	x_conn_cv and x_dis-_cv are new condition variables to let
86  *	threads knows when the connection attempt is done, and to let
87  *	the connecting thread know when the disconnect handshake is
88  *	done.
89  *
90  * Added the CONN_HOLD() macro so that all reference holds have the same
91  * look and feel.
92  *
93  * In the private (cku_private) portion of the client handle,
94  *
95  *	cku_flags replaces the cku_sent a boolean. cku_flags keeps
96  *	track of whether a request as been sent, and whether the
97  *	client's handles call record is on the dispatch list (so that
98  *	the reply can be matched by XID to the right client handle).
99  *	The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit()
100  *	and still have the response find the right client handle so
101  *	that the retry of CLNT_CALL() gets the result. Testing, found
102  *	situations where if the timeout was increased, performance
103  *	degraded. This was due to us hitting a window where the thread
104  *	was back in rfscall() (probably printing server not responding)
105  *	while the response came back but no place to put it.
106  *
107  *	cku_ctime is just a cache of x_ctime. If they match,
108  *	clnt_cots_kcallit() won't to send a retry (unless the maximum
109  *	receive count limit as been reached). If the don't match, then
110  *	we assume the request has been lost, and a retry of the request
111  *	is needed.
112  *
113  *	cku_recv_attempts counts the number of receive count attempts
114  *	after one try is sent on the wire.
115  *
116  * Added the clnt_delay() routine so that interruptible and
117  * noninterruptible delays are possible.
118  *
119  * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to
120  * control how long the client delays before returned after getting
121  * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash
122  * a server that may be booting and not yet started nfsd.
123  *
124  * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable)
125  * Why don't we just wait forever (receive an infinite # of times)?
126  * Because the server may have rebooted. More insidious is that some
127  * servers (ours) will drop NFS/TCP requests in some cases. This is bad,
128  * but it is a reality.
129  *
130  * The case of a server doing orderly release really messes up the
131  * client's recovery, especially if the server's TCP implementation is
132  * buggy.  It was found was that the kRPC/COTS client was breaking some
133  * TPI rules, such as not waiting for the acknowledgement of a
134  * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and
135  * T_DISCON_REQ in clnt_dispatch_notifyall()).
136  *
137  * One of things that we've seen is that a kRPC TCP endpoint goes into
138  * TIMEWAIT and a thus a reconnect takes a long time to satisfy because
139  * that the TIMEWAIT state takes a while to finish.  If a server sends a
140  * T_ORDREL_IND, there is little point in an RPC client doing a
141  * T_ORDREL_REQ, because the RPC request isn't going to make it (the
142  * server is saying that it won't accept any more data). So kRPC was
143  * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the
144  * connection skips the TIMEWAIT state and goes straight to a bound state
145  * that kRPC can quickly switch to connected.
146  *
147  * Code that issues TPI request must use waitforack() to wait for the
148  * corresponding ack (assuming there is one) in any future modifications.
149  * This works around problems that may be introduced by breaking TPI rules
150  * (by submitting new calls before earlier requests have been acked) in the
151  * case of a signal or other early return.  waitforack() depends on
152  * clnt_dispatch_notifyconn() to issue the wakeup when the ack
153  * arrives, so adding new TPI calls may require corresponding changes
154  * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on
155  * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure
156  * not to set it too low or TPI ACKS will be lost.
157  */
158 
159 #include <sys/param.h>
160 #include <sys/types.h>
161 #include <sys/user.h>
162 #include <sys/systm.h>
163 #include <sys/sysmacros.h>
164 #include <sys/proc.h>
165 #include <sys/socket.h>
166 #include <sys/file.h>
167 #include <sys/stream.h>
168 #include <sys/strsubr.h>
169 #include <sys/stropts.h>
170 #include <sys/strsun.h>
171 #include <sys/timod.h>
172 #include <sys/tiuser.h>
173 #include <sys/tihdr.h>
174 #include <sys/t_kuser.h>
175 #include <sys/fcntl.h>
176 #include <sys/errno.h>
177 #include <sys/kmem.h>
178 #include <sys/debug.h>
179 #include <sys/systm.h>
180 #include <sys/kstat.h>
181 #include <sys/t_lock.h>
182 #include <sys/ddi.h>
183 #include <sys/cmn_err.h>
184 #include <sys/time.h>
185 #include <sys/isa_defs.h>
186 #include <sys/callb.h>
187 #include <sys/sunddi.h>
188 #include <sys/atomic.h>
189 #include <sys/sdt.h>
190 
191 #include <netinet/in.h>
192 #include <netinet/tcp.h>
193 
194 #include <rpc/types.h>
195 #include <rpc/xdr.h>
196 #include <rpc/auth.h>
197 #include <rpc/clnt.h>
198 #include <rpc/rpc_msg.h>
199 
200 #define	COTS_DEFAULT_ALLOCSIZE	2048
201 
202 #define	WIRE_HDR_SIZE	20	/* serialized call header, sans proc number */
203 #define	MSG_OFFSET	128	/* offset of call into the mblk */
204 
205 const char *kinet_ntop6(uchar_t *, char *, size_t);
206 
207 static int	clnt_cots_ksettimers(CLIENT *, struct rpc_timers *,
208     struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t);
209 static enum clnt_stat	clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t,
210     caddr_t, xdrproc_t, caddr_t, struct timeval);
211 static void	clnt_cots_kabort(CLIENT *);
212 static void	clnt_cots_kerror(CLIENT *, struct rpc_err *);
213 static bool_t	clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t);
214 static void	clnt_cots_kdestroy(CLIENT *);
215 static bool_t	clnt_cots_kcontrol(CLIENT *, int, char *);
216 
217 
218 /* List of transports managed by the connection manager. */
219 struct cm_xprt {
220 	TIUSER		*x_tiptr;	/* transport handle */
221 	queue_t		*x_wq;		/* send queue */
222 	clock_t		x_time;		/* last time we handed this xprt out */
223 	clock_t		x_ctime;	/* time we went to CONNECTED */
224 	int		x_tidu_size;    /* TIDU size of this transport */
225 	union {
226 	    struct {
227 		unsigned int
228 #ifdef	_BIT_FIELDS_HTOL
229 		b_closing:	1,	/* we've sent a ord rel on this conn */
230 		b_dead:		1,	/* transport is closed or disconn */
231 		b_doomed:	1,	/* too many conns, let this go idle */
232 		b_connected:	1,	/* this connection is connected */
233 
234 		b_ordrel:	1,	/* do an orderly release? */
235 		b_thread:	1,	/* thread doing connect */
236 		b_waitdis:	1,	/* waiting for disconnect ACK */
237 		b_needdis:	1,	/* need T_DISCON_REQ */
238 
239 		b_needrel:	1,	/* need T_ORDREL_REQ */
240 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
241 					/* disconnect during connect */
242 
243 		b_pad:		22;
244 
245 #endif
246 
247 #ifdef	_BIT_FIELDS_LTOH
248 		b_pad:		22,
249 
250 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
251 					/* disconnect during connect */
252 		b_needrel:	1,	/* need T_ORDREL_REQ */
253 
254 		b_needdis:	1,	/* need T_DISCON_REQ */
255 		b_waitdis:	1,	/* waiting for disconnect ACK */
256 		b_thread:	1,	/* thread doing connect */
257 		b_ordrel:	1,	/* do an orderly release? */
258 
259 		b_connected:	1,	/* this connection is connected */
260 		b_doomed:	1,	/* too many conns, let this go idle */
261 		b_dead:		1,	/* transport is closed or disconn */
262 		b_closing:	1;	/* we've sent a ord rel on this conn */
263 #endif
264 	    } bit;	    unsigned int word;
265 
266 #define	x_closing	x_state.bit.b_closing
267 #define	x_dead		x_state.bit.b_dead
268 #define	x_doomed	x_state.bit.b_doomed
269 #define	x_connected	x_state.bit.b_connected
270 
271 #define	x_ordrel	x_state.bit.b_ordrel
272 #define	x_thread	x_state.bit.b_thread
273 #define	x_waitdis	x_state.bit.b_waitdis
274 #define	x_needdis	x_state.bit.b_needdis
275 
276 #define	x_needrel	x_state.bit.b_needrel
277 #define	x_early_disc    x_state.bit.b_early_disc
278 
279 #define	x_state_flags	x_state.word
280 
281 #define	X_CLOSING	0x80000000
282 #define	X_DEAD		0x40000000
283 #define	X_DOOMED	0x20000000
284 #define	X_CONNECTED	0x10000000
285 
286 #define	X_ORDREL	0x08000000
287 #define	X_THREAD	0x04000000
288 #define	X_WAITDIS	0x02000000
289 #define	X_NEEDDIS	0x01000000
290 
291 #define	X_NEEDREL	0x00800000
292 #define	X_EARLYDISC	0x00400000
293 
294 #define	X_BADSTATES	(X_CLOSING | X_DEAD | X_DOOMED)
295 
296 	}		x_state;
297 	int		x_ref;		/* number of users of this xprt */
298 	int		x_family;	/* address family of transport */
299 	dev_t		x_rdev;		/* device number of transport */
300 	struct cm_xprt	*x_next;
301 
302 	struct netbuf	x_server;	/* destination address */
303 	struct netbuf	x_src;		/* src address (for retries) */
304 	kmutex_t	x_lock;		/* lock on this entry */
305 	kcondvar_t	x_cv;		/* to signal when can be closed */
306 	kcondvar_t	x_conn_cv;	/* to signal when connection attempt */
307 					/* is complete */
308 	kstat_t		*x_ksp;
309 
310 	kcondvar_t	x_dis_cv;	/* to signal when disconnect attempt */
311 					/* is complete */
312 	zoneid_t	x_zoneid;	/* zone this xprt belongs to */
313 };
314 
315 typedef struct cm_kstat_xprt {
316 	kstat_named_t	x_wq;
317 	kstat_named_t	x_server;
318 	kstat_named_t	x_family;
319 	kstat_named_t	x_rdev;
320 	kstat_named_t	x_time;
321 	kstat_named_t	x_state;
322 	kstat_named_t	x_ref;
323 	kstat_named_t	x_port;
324 } cm_kstat_xprt_t;
325 
326 static cm_kstat_xprt_t cm_kstat_template = {
327 	{ "write_queue", KSTAT_DATA_UINT32 },
328 	{ "server",	KSTAT_DATA_STRING },
329 	{ "addr_family", KSTAT_DATA_UINT32 },
330 	{ "device",	KSTAT_DATA_UINT32 },
331 	{ "time_stamp",	KSTAT_DATA_UINT32 },
332 	{ "status",	KSTAT_DATA_UINT32 },
333 	{ "ref_count",	KSTAT_DATA_INT32 },
334 	{ "port",	KSTAT_DATA_UINT32 },
335 };
336 
337 /*
338  * The inverse of this is connmgr_release().
339  */
340 #define	CONN_HOLD(Cm_entry)	{\
341 	mutex_enter(&(Cm_entry)->x_lock);	\
342 	(Cm_entry)->x_ref++;	\
343 	mutex_exit(&(Cm_entry)->x_lock);	\
344 }
345 
346 
347 /*
348  * Private data per rpc handle.  This structure is allocated by
349  * clnt_cots_kcreate, and freed by clnt_cots_kdestroy.
350  */
351 typedef struct cku_private_s {
352 	CLIENT			cku_client;	/* client handle */
353 	calllist_t		cku_call;	/* for dispatching calls */
354 	struct rpc_err		cku_err;	/* error status */
355 
356 	struct netbuf		cku_srcaddr;	/* source address for retries */
357 	int			cku_addrfmly;  /* for binding port */
358 	struct netbuf		cku_addr;	/* remote address */
359 	dev_t			cku_device;	/* device to use */
360 	uint_t			cku_flags;
361 #define	CKU_ONQUEUE		0x1
362 #define	CKU_SENT		0x2
363 
364 	bool_t			cku_progress;	/* for CLSET_PROGRESS */
365 	uint32_t		cku_xid;	/* current XID */
366 	clock_t			cku_ctime;	/* time stamp of when */
367 						/* connection was created */
368 	uint_t			cku_recv_attempts;
369 	XDR			cku_outxdr;	/* xdr routine for output */
370 	XDR			cku_inxdr;	/* xdr routine for input */
371 	char			cku_rpchdr[WIRE_HDR_SIZE + 4];
372 						/* pre-serialized rpc header */
373 
374 	uint_t			cku_outbuflen;	/* default output mblk length */
375 	struct cred		*cku_cred;	/* credentials */
376 	bool_t			cku_nodelayonerr;
377 						/* for CLSET_NODELAYONERR */
378 	int			cku_useresvport; /* Use reserved port */
379 	struct rpc_cots_client	*cku_stats;	/* stats for zone */
380 } cku_private_t;
381 
382 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *,
383 	const struct timeval *, struct netbuf *, int, struct netbuf *,
384 	struct rpc_err *, bool_t, bool_t, cred_t *);
385 
386 static bool_t	connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *,
387 				int, calllist_t *, int *, bool_t reconnect,
388 				const struct timeval *, bool_t, cred_t *);
389 
390 static void	*connmgr_opt_getoff(mblk_t *mp, t_uscalar_t offset,
391 				t_uscalar_t length, uint_t align_size);
392 static bool_t	connmgr_setbufsz(calllist_t *e, queue_t *wq, cred_t *cr);
393 static bool_t	connmgr_getopt_int(queue_t *wq, int level, int name, int *val,
394 				calllist_t *e, cred_t *cr);
395 static bool_t	connmgr_setopt_int(queue_t *wq, int level, int name, int val,
396 				calllist_t *e, cred_t *cr);
397 static bool_t	connmgr_setopt(queue_t *, int, int, calllist_t *, cred_t *cr);
398 static void	connmgr_sndrel(struct cm_xprt *);
399 static void	connmgr_snddis(struct cm_xprt *);
400 static void	connmgr_close(struct cm_xprt *);
401 static void	connmgr_release(struct cm_xprt *);
402 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *,
403 	cku_private_t *);
404 
405 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *,
406 	struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t,
407 	bool_t, int, cred_t *);
408 
409 static void connmgr_cancelconn(struct cm_xprt *);
410 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *,
411 	bool_t);
412 static void connmgr_dis_and_wait(struct cm_xprt *);
413 
414 static int	clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t,
415 					uint_t);
416 
417 static int clnt_delay(clock_t, bool_t);
418 
419 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t);
420 
421 /*
422  * Operations vector for TCP/IP based RPC
423  */
424 static struct clnt_ops tcp_ops = {
425 	clnt_cots_kcallit,	/* do rpc call */
426 	clnt_cots_kabort,	/* abort call */
427 	clnt_cots_kerror,	/* return error status */
428 	clnt_cots_kfreeres,	/* free results */
429 	clnt_cots_kdestroy,	/* destroy rpc handle */
430 	clnt_cots_kcontrol,	/* the ioctl() of rpc */
431 	clnt_cots_ksettimers,	/* set retry timers */
432 };
433 
434 static int rpc_kstat_instance = 0;  /* keeps the current instance */
435 				/* number for the next kstat_create */
436 
437 static struct cm_xprt *cm_hd = NULL;
438 static kmutex_t connmgr_lock;	/* for connection mngr's list of transports */
439 
440 extern kmutex_t clnt_max_msg_lock;
441 
442 static calllist_t *clnt_pending = NULL;
443 extern kmutex_t clnt_pending_lock;
444 
445 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE;
446 
447 static call_table_t *cots_call_ht;
448 
449 static const struct rpc_cots_client {
450 	kstat_named_t	rccalls;
451 	kstat_named_t	rcbadcalls;
452 	kstat_named_t	rcbadxids;
453 	kstat_named_t	rctimeouts;
454 	kstat_named_t	rcnewcreds;
455 	kstat_named_t	rcbadverfs;
456 	kstat_named_t	rctimers;
457 	kstat_named_t	rccantconn;
458 	kstat_named_t	rcnomem;
459 	kstat_named_t	rcintrs;
460 } cots_rcstat_tmpl = {
461 	{ "calls",	KSTAT_DATA_UINT64 },
462 	{ "badcalls",	KSTAT_DATA_UINT64 },
463 	{ "badxids",	KSTAT_DATA_UINT64 },
464 	{ "timeouts",	KSTAT_DATA_UINT64 },
465 	{ "newcreds",	KSTAT_DATA_UINT64 },
466 	{ "badverfs",	KSTAT_DATA_UINT64 },
467 	{ "timers",	KSTAT_DATA_UINT64 },
468 	{ "cantconn",	KSTAT_DATA_UINT64 },
469 	{ "nomem",	KSTAT_DATA_UINT64 },
470 	{ "interrupts", KSTAT_DATA_UINT64 }
471 };
472 
473 #define	COTSRCSTAT_INCR(p, x)	\
474 	atomic_inc_64(&(p)->x.value.ui64)
475 
476 #define	CLNT_MAX_CONNS	1	/* concurrent connections between clnt/srvr */
477 int clnt_max_conns = CLNT_MAX_CONNS;
478 
479 #define	CLNT_MIN_TIMEOUT	10	/* seconds to wait after we get a */
480 					/* connection reset */
481 #define	CLNT_MIN_CONNTIMEOUT	5	/* seconds to wait for a connection */
482 
483 
484 int clnt_cots_min_tout = CLNT_MIN_TIMEOUT;
485 int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT;
486 
487 /*
488  * Limit the number of times we will attempt to receive a reply without
489  * re-sending a response.
490  */
491 #define	CLNT_MAXRECV_WITHOUT_RETRY	3
492 uint_t clnt_cots_maxrecv	= CLNT_MAXRECV_WITHOUT_RETRY;
493 
494 uint_t *clnt_max_msg_sizep;
495 void (*clnt_stop_idle)(queue_t *wq);
496 
497 #define	ptoh(p)		(&((p)->cku_client))
498 #define	htop(h)		((cku_private_t *)((h)->cl_private))
499 
500 /*
501  * Times to retry
502  */
503 #define	REFRESHES	2	/* authentication refreshes */
504 
505 /*
506  * The following is used to determine the global default behavior for
507  * COTS when binding to a local port.
508  *
509  * If the value is set to 1 the default will be to select a reserved
510  * (aka privileged) port, if the value is zero the default will be to
511  * use non-reserved ports.  Users of kRPC may override this by using
512  * CLNT_CONTROL() and CLSET_BINDRESVPORT.
513  */
514 int clnt_cots_do_bindresvport = 1;
515 
516 static zone_key_t zone_cots_key;
517 
518 /*
519  * Defaults TCP send and receive buffer size for RPC connections.
520  * These values can be tuned by /etc/system.
521  */
522 int rpc_send_bufsz = 1024*1024;
523 int rpc_recv_bufsz = 1024*1024;
524 /*
525  * To use system-wide default for TCP send and receive buffer size,
526  * use /etc/system to set rpc_default_tcp_bufsz to 1:
527  *
528  * set rpcmod:rpc_default_tcp_bufsz=1
529  */
530 int rpc_default_tcp_bufsz = 0;
531 
532 /*
533  * We need to do this after all kernel threads in the zone have exited.
534  */
535 /* ARGSUSED */
536 static void
537 clnt_zone_destroy(zoneid_t zoneid, void *unused)
538 {
539 	struct cm_xprt **cmp;
540 	struct cm_xprt *cm_entry;
541 	struct cm_xprt *freelist = NULL;
542 
543 	mutex_enter(&connmgr_lock);
544 	cmp = &cm_hd;
545 	while ((cm_entry = *cmp) != NULL) {
546 		if (cm_entry->x_zoneid == zoneid) {
547 			*cmp = cm_entry->x_next;
548 			cm_entry->x_next = freelist;
549 			freelist = cm_entry;
550 		} else {
551 			cmp = &cm_entry->x_next;
552 		}
553 	}
554 	mutex_exit(&connmgr_lock);
555 	while ((cm_entry = freelist) != NULL) {
556 		freelist = cm_entry->x_next;
557 		connmgr_close(cm_entry);
558 	}
559 }
560 
561 int
562 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog,
563     rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl)
564 {
565 	CLIENT *h;
566 	cku_private_t *p;
567 	struct rpc_msg call_msg;
568 	struct rpcstat *rpcstat;
569 
570 	RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog);
571 
572 	rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
573 	ASSERT(rpcstat != NULL);
574 
575 	/* Allocate and intialize the client handle. */
576 	p = kmem_zalloc(sizeof (*p), KM_SLEEP);
577 
578 	h = ptoh(p);
579 
580 	h->cl_private = (caddr_t)p;
581 	h->cl_auth = authkern_create();
582 	h->cl_ops = &tcp_ops;
583 
584 	cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL);
585 	mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL);
586 
587 	/*
588 	 * If the current sanity check size in rpcmod is smaller
589 	 * than the size needed, then increase the sanity check.
590 	 */
591 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
592 	    max_msgsize > *clnt_max_msg_sizep) {
593 		mutex_enter(&clnt_max_msg_lock);
594 		if (max_msgsize > *clnt_max_msg_sizep)
595 			*clnt_max_msg_sizep = max_msgsize;
596 		mutex_exit(&clnt_max_msg_lock);
597 	}
598 
599 	p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE;
600 
601 	/* Preserialize the call message header */
602 
603 	call_msg.rm_xid = 0;
604 	call_msg.rm_direction = CALL;
605 	call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION;
606 	call_msg.rm_call.cb_prog = prog;
607 	call_msg.rm_call.cb_vers = vers;
608 
609 	xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE);
610 
611 	if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) {
612 		XDR_DESTROY(&p->cku_outxdr);
613 		RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization "
614 		    "error\n");
615 		auth_destroy(h->cl_auth);
616 		kmem_free(p, sizeof (cku_private_t));
617 		RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n");
618 		return (EINVAL);		/* XXX */
619 	}
620 	XDR_DESTROY(&p->cku_outxdr);
621 
622 	/*
623 	 * The zalloc initialized the fields below.
624 	 * p->cku_xid = 0;
625 	 * p->cku_flags = 0;
626 	 * p->cku_srcaddr.len = 0;
627 	 * p->cku_srcaddr.maxlen = 0;
628 	 */
629 
630 	p->cku_cred = cred;
631 	p->cku_device = dev;
632 	p->cku_addrfmly = family;
633 	p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
634 	p->cku_addr.maxlen = addr->maxlen;
635 	p->cku_addr.len = addr->len;
636 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
637 	p->cku_stats = rpcstat->rpc_cots_client;
638 	p->cku_useresvport = -1; /* value is has not been set */
639 
640 	*ncl = h;
641 	return (0);
642 }
643 
644 /*ARGSUSED*/
645 static void
646 clnt_cots_kabort(CLIENT *h)
647 {
648 }
649 
650 /*
651  * Return error info on this handle.
652  */
653 static void
654 clnt_cots_kerror(CLIENT *h, struct rpc_err *err)
655 {
656 	/* LINTED pointer alignment */
657 	cku_private_t *p = htop(h);
658 
659 	*err = p->cku_err;
660 }
661 
662 /*ARGSUSED*/
663 static bool_t
664 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr)
665 {
666 	xdr_free(xdr_res, res_ptr);
667 
668 	return (TRUE);
669 }
670 
671 static bool_t
672 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg)
673 {
674 	cku_private_t *p = htop(h);
675 
676 	switch (cmd) {
677 	case CLSET_PROGRESS:
678 		p->cku_progress = TRUE;
679 		return (TRUE);
680 
681 	case CLSET_XID:
682 		if (arg == NULL)
683 			return (FALSE);
684 
685 		p->cku_xid = *((uint32_t *)arg);
686 		return (TRUE);
687 
688 	case CLGET_XID:
689 		if (arg == NULL)
690 			return (FALSE);
691 
692 		*((uint32_t *)arg) = p->cku_xid;
693 		return (TRUE);
694 
695 	case CLSET_NODELAYONERR:
696 		if (arg == NULL)
697 			return (FALSE);
698 
699 		if (*((bool_t *)arg) == TRUE) {
700 			p->cku_nodelayonerr = TRUE;
701 			return (TRUE);
702 		}
703 		if (*((bool_t *)arg) == FALSE) {
704 			p->cku_nodelayonerr = FALSE;
705 			return (TRUE);
706 		}
707 		return (FALSE);
708 
709 	case CLGET_NODELAYONERR:
710 		if (arg == NULL)
711 			return (FALSE);
712 
713 		*((bool_t *)arg) = p->cku_nodelayonerr;
714 		return (TRUE);
715 
716 	case CLSET_BINDRESVPORT:
717 		if (arg == NULL)
718 			return (FALSE);
719 
720 		if (*(int *)arg != 1 && *(int *)arg != 0)
721 			return (FALSE);
722 
723 		p->cku_useresvport = *(int *)arg;
724 
725 		return (TRUE);
726 
727 	case CLGET_BINDRESVPORT:
728 		if (arg == NULL)
729 			return (FALSE);
730 
731 		*(int *)arg = p->cku_useresvport;
732 
733 		return (TRUE);
734 
735 	default:
736 		return (FALSE);
737 	}
738 }
739 
740 /*
741  * Destroy rpc handle.  Frees the space used for output buffer,
742  * private data, and handle structure.
743  */
744 static void
745 clnt_cots_kdestroy(CLIENT *h)
746 {
747 	/* LINTED pointer alignment */
748 	cku_private_t *p = htop(h);
749 	calllist_t *call = &p->cku_call;
750 
751 	RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h);
752 	RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid);
753 
754 	if (p->cku_flags & CKU_ONQUEUE) {
755 		RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x "
756 		    "from dispatch list\n", p->cku_xid);
757 		call_table_remove(call);
758 	}
759 
760 	if (call->call_reply)
761 		freemsg(call->call_reply);
762 	cv_destroy(&call->call_cv);
763 	mutex_destroy(&call->call_lock);
764 
765 	kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen);
766 	kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
767 	kmem_free(p, sizeof (*p));
768 }
769 
770 static int clnt_cots_pulls;
771 #define	RM_HDR_SIZE	4	/* record mark header size */
772 
773 /*
774  * Call remote procedure.
775  */
776 static enum clnt_stat
777 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
778     caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait)
779 {
780 	/* LINTED pointer alignment */
781 	cku_private_t *p = htop(h);
782 	calllist_t *call = &p->cku_call;
783 	XDR *xdrs;
784 	struct rpc_msg reply_msg;
785 	mblk_t *mp;
786 #ifdef	RPCDEBUG
787 	clock_t time_sent;
788 #endif
789 	struct netbuf *retryaddr;
790 	struct cm_xprt *cm_entry = NULL;
791 	queue_t *wq;
792 	int len, waitsecs, max_waitsecs;
793 	int mpsize;
794 	int refreshes = REFRESHES;
795 	int interrupted;
796 	int tidu_size;
797 	enum clnt_stat status;
798 	struct timeval cwait;
799 	bool_t delay_first = FALSE;
800 	clock_t ticks, now;
801 
802 	RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum);
803 	COTSRCSTAT_INCR(p->cku_stats, rccalls);
804 
805 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec);
806 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec);
807 	/*
808 	 * Bug ID 1240234:
809 	 * Look out for zero length timeouts. We don't want to
810 	 * wait zero seconds for a connection to be established.
811 	 */
812 	if (wait.tv_sec < clnt_cots_min_conntout) {
813 		cwait.tv_sec = clnt_cots_min_conntout;
814 		cwait.tv_usec = 0;
815 		RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,",
816 		    wait.tv_sec);
817 		RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout);
818 	} else {
819 		cwait = wait;
820 	}
821 
822 call_again:
823 	if (cm_entry) {
824 		connmgr_release(cm_entry);
825 		cm_entry = NULL;
826 	}
827 
828 	mp = NULL;
829 
830 	/*
831 	 * If the call is not a retry, allocate a new xid and cache it
832 	 * for future retries.
833 	 * Bug ID 1246045:
834 	 * Treat call as a retry for purposes of binding the source
835 	 * port only if we actually attempted to send anything on
836 	 * the previous call.
837 	 */
838 	if (p->cku_xid == 0) {
839 		p->cku_xid = alloc_xid();
840 		call->call_zoneid = rpc_zoneid();
841 
842 		/*
843 		 * We need to ASSERT here that our xid != 0 because this
844 		 * determines whether or not our call record gets placed on
845 		 * the hash table or the linked list.  By design, we mandate
846 		 * that RPC calls over cots must have xid's != 0, so we can
847 		 * ensure proper management of the hash table.
848 		 */
849 		ASSERT(p->cku_xid != 0);
850 
851 		retryaddr = NULL;
852 		p->cku_flags &= ~CKU_SENT;
853 
854 		if (p->cku_flags & CKU_ONQUEUE) {
855 			RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old"
856 			    " one (%p)\n", (void *)call);
857 			call_table_remove(call);
858 			p->cku_flags &= ~CKU_ONQUEUE;
859 			RPCLOG(64, "clnt_cots_kcallit: removing call from "
860 			    "dispatch list because xid was zero (now 0x%x)\n",
861 			    p->cku_xid);
862 		}
863 
864 		if (call->call_reply != NULL) {
865 			freemsg(call->call_reply);
866 			call->call_reply = NULL;
867 		}
868 	} else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) {
869 		retryaddr = NULL;
870 
871 	} else if (p->cku_flags & CKU_SENT) {
872 		retryaddr = &p->cku_srcaddr;
873 
874 	} else {
875 		/*
876 		 * Bug ID 1246045: Nothing was sent, so set retryaddr to
877 		 * NULL and let connmgr_get() bind to any source port it
878 		 * can get.
879 		 */
880 		retryaddr = NULL;
881 	}
882 
883 	RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid);
884 	RPCLOG(64, " flags = 0x%x\n", p->cku_flags);
885 
886 	p->cku_err.re_status = RPC_TIMEDOUT;
887 	p->cku_err.re_errno = p->cku_err.re_terrno = 0;
888 
889 	cm_entry = connmgr_wrapget(retryaddr, &cwait, p);
890 
891 	if (cm_entry == NULL) {
892 		RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n",
893 		    clnt_sperrno(p->cku_err.re_status));
894 
895 		/*
896 		 * The reasons why we fail to create a connection are
897 		 * varied. In most cases we don't want the caller to
898 		 * immediately retry. This could have one or more
899 		 * bad effects. This includes flooding the net with
900 		 * connect requests to ports with no listener; a hard
901 		 * kernel loop due to all the "reserved" TCP ports being
902 		 * in use.
903 		 */
904 		delay_first = TRUE;
905 
906 		/*
907 		 * Even if we end up returning EINTR, we still count a
908 		 * a "can't connect", because the connection manager
909 		 * might have been committed to waiting for or timing out on
910 		 * a connection.
911 		 */
912 		COTSRCSTAT_INCR(p->cku_stats, rccantconn);
913 		switch (p->cku_err.re_status) {
914 		case RPC_INTR:
915 			p->cku_err.re_errno = EINTR;
916 
917 			/*
918 			 * No need to delay because a UNIX signal(2)
919 			 * interrupted us. The caller likely won't
920 			 * retry the CLNT_CALL() and even if it does,
921 			 * we assume the caller knows what it is doing.
922 			 */
923 			delay_first = FALSE;
924 			break;
925 
926 		case RPC_TIMEDOUT:
927 			p->cku_err.re_errno = ETIMEDOUT;
928 
929 			/*
930 			 * No need to delay because timed out already
931 			 * on the connection request and assume that the
932 			 * transport time out is longer than our minimum
933 			 * timeout, or least not too much smaller.
934 			 */
935 			delay_first = FALSE;
936 			break;
937 
938 		case RPC_SYSTEMERROR:
939 		case RPC_TLIERROR:
940 			/*
941 			 * We want to delay here because a transient
942 			 * system error has a better chance of going away
943 			 * if we delay a bit. If it's not transient, then
944 			 * we don't want end up in a hard kernel loop
945 			 * due to retries.
946 			 */
947 			ASSERT(p->cku_err.re_errno != 0);
948 			break;
949 
950 
951 		case RPC_CANTCONNECT:
952 			/*
953 			 * RPC_CANTCONNECT is set on T_ERROR_ACK which
954 			 * implies some error down in the TCP layer or
955 			 * below. If cku_nodelayonerror is set then we
956 			 * assume the caller knows not to try too hard.
957 			 */
958 			RPCLOG0(8, "clnt_cots_kcallit: connection failed,");
959 			RPCLOG0(8, " re_status=RPC_CANTCONNECT,");
960 			RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno);
961 			RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr);
962 			if (p->cku_nodelayonerr == TRUE)
963 				delay_first = FALSE;
964 
965 			p->cku_err.re_errno = EIO;
966 
967 			break;
968 
969 		case RPC_XPRTFAILED:
970 			/*
971 			 * We want to delay here because we likely
972 			 * got a refused connection.
973 			 */
974 			if (p->cku_err.re_errno == 0)
975 				p->cku_err.re_errno = EIO;
976 
977 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
978 			    p->cku_err.re_errno);
979 
980 			break;
981 
982 		default:
983 			/*
984 			 * We delay here because it is better to err
985 			 * on the side of caution. If we got here then
986 			 * status could have been RPC_SUCCESS, but we
987 			 * know that we did not get a connection, so
988 			 * force the rpc status to RPC_CANTCONNECT.
989 			 */
990 			p->cku_err.re_status = RPC_CANTCONNECT;
991 			p->cku_err.re_errno = EIO;
992 			break;
993 		}
994 		if (delay_first == TRUE)
995 			ticks = clnt_cots_min_tout * drv_usectohz(1000000);
996 		goto cots_done;
997 	}
998 
999 	/*
1000 	 * If we've never sent any request on this connection (send count
1001 	 * is zero, or the connection has been reset), cache the
1002 	 * the connection's create time and send a request (possibly a retry)
1003 	 */
1004 	if ((p->cku_flags & CKU_SENT) == 0 ||
1005 	    p->cku_ctime != cm_entry->x_ctime) {
1006 		p->cku_ctime = cm_entry->x_ctime;
1007 
1008 	} else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) &&
1009 	    (call->call_reply != NULL ||
1010 	    p->cku_recv_attempts < clnt_cots_maxrecv)) {
1011 
1012 		/*
1013 		 * If we've sent a request and our call is on the dispatch
1014 		 * queue and we haven't made too many receive attempts, then
1015 		 * don't re-send, just receive.
1016 		 */
1017 		p->cku_recv_attempts++;
1018 		goto read_again;
1019 	}
1020 
1021 	/*
1022 	 * Now we create the RPC request in a STREAMS message.  We have to do
1023 	 * this after the call to connmgr_get so that we have the correct
1024 	 * TIDU size for the transport.
1025 	 */
1026 	tidu_size = cm_entry->x_tidu_size;
1027 	len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE);
1028 
1029 	while ((mp = allocb(len, BPRI_MED)) == NULL) {
1030 		if (strwaitbuf(len, BPRI_MED)) {
1031 			p->cku_err.re_status = RPC_SYSTEMERROR;
1032 			p->cku_err.re_errno = ENOSR;
1033 			COTSRCSTAT_INCR(p->cku_stats, rcnomem);
1034 			goto cots_done;
1035 		}
1036 	}
1037 	xdrs = &p->cku_outxdr;
1038 	xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size);
1039 	mpsize = MBLKSIZE(mp);
1040 	ASSERT(mpsize >= len);
1041 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
1042 
1043 	/*
1044 	 * If the size of mblk is not appreciably larger than what we
1045 	 * asked, then resize the mblk to exactly len bytes. The reason for
1046 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
1047 	 * (from TCP over ethernet), and the arguments to the RPC require
1048 	 * 2800 bytes. Ideally we want the protocol to render two
1049 	 * ~1400 byte segments over the wire. However if allocb() gives us a 2k
1050 	 * mblk, and we allocate a second mblk for the remainder, the protocol
1051 	 * module may generate 3 segments over the wire:
1052 	 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and
1053 	 * 892 for the third. If we "waste" 448 bytes in the first mblk,
1054 	 * the XDR encoding will generate two ~1400 byte mblks, and the
1055 	 * protocol module is more likely to produce properly sized segments.
1056 	 */
1057 	if ((mpsize >> 1) <= len)
1058 		mp->b_rptr += (mpsize - len);
1059 
1060 	/*
1061 	 * Adjust b_rptr to reserve space for the non-data protocol headers
1062 	 * any downstream modules might like to add, and for the
1063 	 * record marking header.
1064 	 */
1065 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
1066 
1067 	if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) {
1068 		/* Copy in the preserialized RPC header information. */
1069 		bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE);
1070 
1071 		/* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */
1072 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base +
1073 		    WIRE_HDR_SIZE));
1074 
1075 		ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE);
1076 
1077 		/* Serialize the procedure number and the arguments. */
1078 		if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) ||
1079 		    (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) ||
1080 		    (!(*xdr_args)(xdrs, argsp))) {
1081 			XDR_DESTROY(xdrs);
1082 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1083 			p->cku_err.re_errno = EIO;
1084 			goto cots_done;
1085 		}
1086 
1087 		(*(uint32_t *)(mp->b_rptr)) = p->cku_xid;
1088 	} else {
1089 		uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE];
1090 		IXDR_PUT_U_INT32(uproc, procnum);
1091 
1092 		(*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid;
1093 
1094 		/* Use XDR_SETPOS() to set the b_wptr. */
1095 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
1096 
1097 		/* Serialize the procedure number and the arguments. */
1098 		if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4,
1099 		    xdrs, xdr_args, argsp)) {
1100 			XDR_DESTROY(xdrs);
1101 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1102 			p->cku_err.re_errno = EIO;
1103 			goto cots_done;
1104 		}
1105 	}
1106 
1107 	XDR_DESTROY(xdrs);
1108 
1109 	RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n",
1110 	    tidu_size);
1111 
1112 	wq = cm_entry->x_wq;
1113 	waitsecs = 0;
1114 
1115 dispatch_again:
1116 	status = clnt_dispatch_send(wq, mp, call, p->cku_xid,
1117 	    (p->cku_flags & CKU_ONQUEUE));
1118 
1119 	if ((status == RPC_CANTSEND) && (call->call_reason == ENOBUFS)) {
1120 		/*
1121 		 * QFULL condition, allow some time for queue to drain
1122 		 * and try again. Give up after waiting for all timeout
1123 		 * specified for the call, or zone is going away.
1124 		 */
1125 		max_waitsecs = wait.tv_sec ? wait.tv_sec : clnt_cots_min_tout;
1126 		if ((waitsecs++ < max_waitsecs) &&
1127 		    !(zone_status_get(curproc->p_zone) >=
1128 		    ZONE_IS_SHUTTING_DOWN)) {
1129 
1130 			/* wait 1 sec for queue to drain */
1131 			if (clnt_delay(drv_usectohz(1000000),
1132 			    h->cl_nosignal) == EINTR) {
1133 				p->cku_err.re_errno = EINTR;
1134 				p->cku_err.re_status = RPC_INTR;
1135 
1136 				goto cots_done;
1137 			}
1138 
1139 			/* and try again */
1140 			goto dispatch_again;
1141 		}
1142 		p->cku_err.re_status = status;
1143 		p->cku_err.re_errno = call->call_reason;
1144 		DTRACE_PROBE(krpc__e__clntcots__kcallit__cantsend);
1145 
1146 		goto cots_done;
1147 	}
1148 
1149 	if (waitsecs) {
1150 		/* adjust timeout to account for time wait to send */
1151 		wait.tv_sec -= waitsecs;
1152 		if (wait.tv_sec < 0) {
1153 			/* pick up reply on next retry */
1154 			wait.tv_sec = 0;
1155 		}
1156 		DTRACE_PROBE2(clnt_cots__sendwait, CLIENT *, h,
1157 		    int, waitsecs);
1158 	}
1159 
1160 	RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n",
1161 	    (uint_t)p->cku_xid);
1162 	p->cku_flags = (CKU_ONQUEUE|CKU_SENT);
1163 	p->cku_recv_attempts = 1;
1164 
1165 #ifdef	RPCDEBUG
1166 	time_sent = ddi_get_lbolt();
1167 #endif
1168 
1169 	/*
1170 	 * Wait for a reply or a timeout.  If there is no error or timeout,
1171 	 * (both indicated by call_status), call->call_reply will contain
1172 	 * the RPC reply message.
1173 	 */
1174 read_again:
1175 	mutex_enter(&call->call_lock);
1176 	interrupted = 0;
1177 	if (call->call_status == RPC_TIMEDOUT) {
1178 		/*
1179 		 * Indicate that the lwp is not to be stopped while waiting
1180 		 * for this network traffic.  This is to avoid deadlock while
1181 		 * debugging a process via /proc and also to avoid recursive
1182 		 * mutex_enter()s due to NFS page faults while stopping
1183 		 * (NFS holds locks when it calls here).
1184 		 */
1185 		clock_t cv_wait_ret;
1186 		clock_t timout;
1187 		clock_t oldlbolt;
1188 
1189 		klwp_t *lwp = ttolwp(curthread);
1190 
1191 		if (lwp != NULL)
1192 			lwp->lwp_nostop++;
1193 
1194 		oldlbolt = ddi_get_lbolt();
1195 		timout = wait.tv_sec * drv_usectohz(1000000) +
1196 		    drv_usectohz(wait.tv_usec) + oldlbolt;
1197 		/*
1198 		 * Iterate until the call_status is changed to something
1199 		 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns
1200 		 * something <=0 zero. The latter means that we timed
1201 		 * out.
1202 		 */
1203 		if (h->cl_nosignal)
1204 			while ((cv_wait_ret = cv_timedwait(&call->call_cv,
1205 			    &call->call_lock, timout)) > 0 &&
1206 			    call->call_status == RPC_TIMEDOUT)
1207 				;
1208 		else
1209 			while ((cv_wait_ret = cv_timedwait_sig(
1210 			    &call->call_cv,
1211 			    &call->call_lock, timout)) > 0 &&
1212 			    call->call_status == RPC_TIMEDOUT)
1213 				;
1214 
1215 		switch (cv_wait_ret) {
1216 		case 0:
1217 			/*
1218 			 * If we got out of the above loop with
1219 			 * cv_timedwait_sig() returning 0, then we were
1220 			 * interrupted regardless what call_status is.
1221 			 */
1222 			interrupted = 1;
1223 			break;
1224 		case -1:
1225 			/* cv_timedwait_sig() timed out */
1226 			break;
1227 		default:
1228 
1229 			/*
1230 			 * We were cv_signaled(). If we didn't
1231 			 * get a successful call_status and returned
1232 			 * before time expired, delay up to clnt_cots_min_tout
1233 			 * seconds so that the caller doesn't immediately
1234 			 * try to call us again and thus force the
1235 			 * same condition that got us here (such
1236 			 * as a RPC_XPRTFAILED due to the server not
1237 			 * listening on the end-point.
1238 			 */
1239 			if (call->call_status != RPC_SUCCESS) {
1240 				clock_t curlbolt;
1241 				clock_t diff;
1242 
1243 				curlbolt = ddi_get_lbolt();
1244 				ticks = clnt_cots_min_tout *
1245 				    drv_usectohz(1000000);
1246 				diff = curlbolt - oldlbolt;
1247 				if (diff < ticks) {
1248 					delay_first = TRUE;
1249 					if (diff > 0)
1250 						ticks -= diff;
1251 				}
1252 			}
1253 			break;
1254 		}
1255 
1256 		if (lwp != NULL)
1257 			lwp->lwp_nostop--;
1258 	}
1259 	/*
1260 	 * Get the reply message, if any.  This will be freed at the end
1261 	 * whether or not an error occurred.
1262 	 */
1263 	mp = call->call_reply;
1264 	call->call_reply = NULL;
1265 
1266 	/*
1267 	 * call_err is the error info when the call is on dispatch queue.
1268 	 * cku_err is the error info returned to the caller.
1269 	 * Sync cku_err with call_err for local message processing.
1270 	 */
1271 
1272 	status = call->call_status;
1273 	p->cku_err = call->call_err;
1274 	mutex_exit(&call->call_lock);
1275 
1276 	if (status != RPC_SUCCESS) {
1277 		switch (status) {
1278 		case RPC_TIMEDOUT:
1279 			now = ddi_get_lbolt();
1280 			if (interrupted) {
1281 				COTSRCSTAT_INCR(p->cku_stats, rcintrs);
1282 				p->cku_err.re_status = RPC_INTR;
1283 				p->cku_err.re_errno = EINTR;
1284 				RPCLOG(1, "clnt_cots_kcallit: xid 0x%x",
1285 				    p->cku_xid);
1286 				RPCLOG(1, "signal interrupted at %ld", now);
1287 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1288 			} else {
1289 				COTSRCSTAT_INCR(p->cku_stats, rctimeouts);
1290 				p->cku_err.re_errno = ETIMEDOUT;
1291 				RPCLOG(1, "clnt_cots_kcallit: timed out at %ld",
1292 				    now);
1293 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1294 			}
1295 			break;
1296 
1297 		case RPC_XPRTFAILED:
1298 			if (p->cku_err.re_errno == 0)
1299 				p->cku_err.re_errno = EIO;
1300 
1301 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
1302 			    p->cku_err.re_errno);
1303 			break;
1304 
1305 		case RPC_SYSTEMERROR:
1306 			ASSERT(p->cku_err.re_errno);
1307 			RPCLOG(1, "clnt_cots_kcallit: system error: %d\n",
1308 			    p->cku_err.re_errno);
1309 			break;
1310 
1311 		default:
1312 			p->cku_err.re_status = RPC_SYSTEMERROR;
1313 			p->cku_err.re_errno = EIO;
1314 			RPCLOG(1, "clnt_cots_kcallit: error: %s\n",
1315 			    clnt_sperrno(status));
1316 			break;
1317 		}
1318 		if (p->cku_err.re_status != RPC_TIMEDOUT) {
1319 
1320 			if (p->cku_flags & CKU_ONQUEUE) {
1321 				call_table_remove(call);
1322 				p->cku_flags &= ~CKU_ONQUEUE;
1323 			}
1324 
1325 			RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x "
1326 			    "taken off dispatch list\n", p->cku_xid);
1327 			if (call->call_reply) {
1328 				freemsg(call->call_reply);
1329 				call->call_reply = NULL;
1330 			}
1331 		} else if (wait.tv_sec != 0) {
1332 			/*
1333 			 * We've sent the request over TCP and so we have
1334 			 * every reason to believe it will get
1335 			 * delivered. In which case returning a timeout is not
1336 			 * appropriate.
1337 			 */
1338 			if (p->cku_progress == TRUE &&
1339 			    p->cku_recv_attempts < clnt_cots_maxrecv) {
1340 				p->cku_err.re_status = RPC_INPROGRESS;
1341 			}
1342 		}
1343 		goto cots_done;
1344 	}
1345 
1346 	xdrs = &p->cku_inxdr;
1347 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
1348 
1349 	reply_msg.rm_direction = REPLY;
1350 	reply_msg.rm_reply.rp_stat = MSG_ACCEPTED;
1351 	reply_msg.acpted_rply.ar_stat = SUCCESS;
1352 
1353 	reply_msg.acpted_rply.ar_verf = _null_auth;
1354 	/*
1355 	 *  xdr_results will be done in AUTH_UNWRAP.
1356 	 */
1357 	reply_msg.acpted_rply.ar_results.where = NULL;
1358 	reply_msg.acpted_rply.ar_results.proc = xdr_void;
1359 
1360 	if (xdr_replymsg(xdrs, &reply_msg)) {
1361 		enum clnt_stat re_status;
1362 
1363 		_seterr_reply(&reply_msg, &p->cku_err);
1364 
1365 		re_status = p->cku_err.re_status;
1366 		if (re_status == RPC_SUCCESS) {
1367 			/*
1368 			 * Reply is good, check auth.
1369 			 */
1370 			if (!AUTH_VALIDATE(h->cl_auth,
1371 			    &reply_msg.acpted_rply.ar_verf)) {
1372 				COTSRCSTAT_INCR(p->cku_stats, rcbadverfs);
1373 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1374 				    "failure\n");
1375 				freemsg(mp);
1376 				(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1377 				XDR_DESTROY(xdrs);
1378 				mutex_enter(&call->call_lock);
1379 				if (call->call_reply == NULL)
1380 					call->call_status = RPC_TIMEDOUT;
1381 				mutex_exit(&call->call_lock);
1382 				goto read_again;
1383 			} else if (!AUTH_UNWRAP(h->cl_auth, xdrs,
1384 			    xdr_results, resultsp)) {
1385 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1386 				    "failure (unwrap)\n");
1387 				p->cku_err.re_status = RPC_CANTDECODERES;
1388 				p->cku_err.re_errno = EIO;
1389 			}
1390 		} else {
1391 			/* set errno in case we can't recover */
1392 			if (re_status != RPC_VERSMISMATCH &&
1393 			    re_status != RPC_AUTHERROR &&
1394 			    re_status != RPC_PROGVERSMISMATCH)
1395 				p->cku_err.re_errno = EIO;
1396 
1397 			if (re_status == RPC_AUTHERROR) {
1398 				/*
1399 				 * Maybe our credential need to be refreshed
1400 				 */
1401 				if (cm_entry) {
1402 					/*
1403 					 * There is the potential that the
1404 					 * cm_entry has/will be marked dead,
1405 					 * so drop the connection altogether,
1406 					 * force REFRESH to establish new
1407 					 * connection.
1408 					 */
1409 					connmgr_cancelconn(cm_entry);
1410 					cm_entry = NULL;
1411 				}
1412 
1413 				(void) xdr_rpc_free_verifier(xdrs,
1414 				    &reply_msg);
1415 				XDR_DESTROY(xdrs);
1416 
1417 				if (p->cku_flags & CKU_ONQUEUE) {
1418 					call_table_remove(call);
1419 					p->cku_flags &= ~CKU_ONQUEUE;
1420 				}
1421 				RPCLOG(64,
1422 				    "clnt_cots_kcallit: AUTH_ERROR, xid"
1423 				    " 0x%x removed off dispatch list\n",
1424 				    p->cku_xid);
1425 				if (call->call_reply) {
1426 					freemsg(call->call_reply);
1427 					call->call_reply = NULL;
1428 				}
1429 
1430 				if ((refreshes > 0) &&
1431 				    AUTH_REFRESH(h->cl_auth, &reply_msg,
1432 				    p->cku_cred)) {
1433 					refreshes--;
1434 					freemsg(mp);
1435 					mp = NULL;
1436 
1437 					COTSRCSTAT_INCR(p->cku_stats,
1438 					    rcbadcalls);
1439 					COTSRCSTAT_INCR(p->cku_stats,
1440 					    rcnewcreds);
1441 					goto call_again;
1442 				}
1443 
1444 				/*
1445 				 * We have used the client handle to
1446 				 * do an AUTH_REFRESH and the RPC status may
1447 				 * be set to RPC_SUCCESS; Let's make sure to
1448 				 * set it to RPC_AUTHERROR.
1449 				 */
1450 				p->cku_err.re_status = RPC_AUTHERROR;
1451 
1452 				/*
1453 				 * Map recoverable and unrecoverable
1454 				 * authentication errors to appropriate errno
1455 				 */
1456 				switch (p->cku_err.re_why) {
1457 				case AUTH_TOOWEAK:
1458 					/*
1459 					 * This could be a failure where the
1460 					 * server requires use of a reserved
1461 					 * port,  check and optionally set the
1462 					 * client handle useresvport trying
1463 					 * one more time. Next go round we
1464 					 * fall out with the tooweak error.
1465 					 */
1466 					if (p->cku_useresvport != 1) {
1467 						p->cku_useresvport = 1;
1468 						p->cku_xid = 0;
1469 						freemsg(mp);
1470 						mp = NULL;
1471 						goto call_again;
1472 					}
1473 					/* FALLTHRU */
1474 				case AUTH_BADCRED:
1475 				case AUTH_BADVERF:
1476 				case AUTH_INVALIDRESP:
1477 				case AUTH_FAILED:
1478 				case RPCSEC_GSS_NOCRED:
1479 				case RPCSEC_GSS_FAILED:
1480 						p->cku_err.re_errno = EACCES;
1481 						break;
1482 				case AUTH_REJECTEDCRED:
1483 				case AUTH_REJECTEDVERF:
1484 				default:	p->cku_err.re_errno = EIO;
1485 						break;
1486 				}
1487 				RPCLOG(1, "clnt_cots_kcallit : authentication"
1488 				    " failed with RPC_AUTHERROR of type %d\n",
1489 				    (int)p->cku_err.re_why);
1490 				goto cots_done;
1491 			}
1492 		}
1493 	} else {
1494 		/* reply didn't decode properly. */
1495 		p->cku_err.re_status = RPC_CANTDECODERES;
1496 		p->cku_err.re_errno = EIO;
1497 		RPCLOG0(1, "clnt_cots_kcallit: decode failure\n");
1498 	}
1499 
1500 	(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1501 	XDR_DESTROY(xdrs);
1502 
1503 	if (p->cku_flags & CKU_ONQUEUE) {
1504 		call_table_remove(call);
1505 		p->cku_flags &= ~CKU_ONQUEUE;
1506 	}
1507 
1508 	RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list",
1509 	    p->cku_xid);
1510 	RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status));
1511 cots_done:
1512 	if (cm_entry)
1513 		connmgr_release(cm_entry);
1514 
1515 	if (mp != NULL)
1516 		freemsg(mp);
1517 	if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) {
1518 		freemsg(call->call_reply);
1519 		call->call_reply = NULL;
1520 	}
1521 	if (p->cku_err.re_status != RPC_SUCCESS) {
1522 		RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n");
1523 		COTSRCSTAT_INCR(p->cku_stats, rcbadcalls);
1524 	}
1525 
1526 	/*
1527 	 * No point in delaying if the zone is going away.
1528 	 */
1529 	if (delay_first == TRUE &&
1530 	    !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) {
1531 		if (clnt_delay(ticks, h->cl_nosignal) == EINTR) {
1532 			p->cku_err.re_errno = EINTR;
1533 			p->cku_err.re_status = RPC_INTR;
1534 		}
1535 	}
1536 	return (p->cku_err.re_status);
1537 }
1538 
1539 /*
1540  * Kinit routine for cots.  This sets up the correct operations in
1541  * the client handle, as the handle may have previously been a clts
1542  * handle, and clears the xid field so there is no way a new call
1543  * could be mistaken for a retry.  It also sets in the handle the
1544  * information that is passed at create/kinit time but needed at
1545  * call time, as cots creates the transport at call time - device,
1546  * address of the server, protocol family.
1547  */
1548 void
1549 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr,
1550     int max_msgsize, cred_t *cred)
1551 {
1552 	/* LINTED pointer alignment */
1553 	cku_private_t *p = htop(h);
1554 	calllist_t *call = &p->cku_call;
1555 
1556 	h->cl_ops = &tcp_ops;
1557 	if (p->cku_flags & CKU_ONQUEUE) {
1558 		call_table_remove(call);
1559 		p->cku_flags &= ~CKU_ONQUEUE;
1560 		RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from"
1561 		    " dispatch list\n", p->cku_xid);
1562 	}
1563 
1564 	if (call->call_reply != NULL) {
1565 		freemsg(call->call_reply);
1566 		call->call_reply = NULL;
1567 	}
1568 
1569 	call->call_bucket = NULL;
1570 	call->call_hash = 0;
1571 
1572 	/*
1573 	 * We don't clear cku_flags here, because clnt_cots_kcallit()
1574 	 * takes care of handling the cku_flags reset.
1575 	 */
1576 	p->cku_xid = 0;
1577 	p->cku_device = dev;
1578 	p->cku_addrfmly = family;
1579 	p->cku_cred = cred;
1580 
1581 	if (p->cku_addr.maxlen < addr->len) {
1582 		if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL)
1583 			kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
1584 		p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
1585 		p->cku_addr.maxlen = addr->maxlen;
1586 	}
1587 
1588 	p->cku_addr.len = addr->len;
1589 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
1590 
1591 	/*
1592 	 * If the current sanity check size in rpcmod is smaller
1593 	 * than the size needed, then increase the sanity check.
1594 	 */
1595 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
1596 	    max_msgsize > *clnt_max_msg_sizep) {
1597 		mutex_enter(&clnt_max_msg_lock);
1598 		if (max_msgsize > *clnt_max_msg_sizep)
1599 			*clnt_max_msg_sizep = max_msgsize;
1600 		mutex_exit(&clnt_max_msg_lock);
1601 	}
1602 }
1603 
1604 /*
1605  * ksettimers is a no-op for cots, with the exception of setting the xid.
1606  */
1607 /* ARGSUSED */
1608 static int
1609 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all,
1610     int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg, uint32_t xid)
1611 {
1612 	/* LINTED pointer alignment */
1613 	cku_private_t *p = htop(h);
1614 
1615 	if (xid)
1616 		p->cku_xid = xid;
1617 	COTSRCSTAT_INCR(p->cku_stats, rctimers);
1618 	return (0);
1619 }
1620 
1621 extern void rpc_poptimod(struct vnode *);
1622 extern int kstr_push(struct vnode *, char *);
1623 
1624 int
1625 conn_kstat_update(kstat_t *ksp, int rw)
1626 {
1627 	struct cm_xprt *cm_entry;
1628 	struct cm_kstat_xprt *cm_ksp_data;
1629 	uchar_t *b;
1630 	char *fbuf;
1631 
1632 	if (rw == KSTAT_WRITE)
1633 		return (EACCES);
1634 	if (ksp == NULL || ksp->ks_private == NULL)
1635 		return (EIO);
1636 	cm_entry  = (struct cm_xprt *)ksp->ks_private;
1637 	cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data;
1638 
1639 	cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq;
1640 	cm_ksp_data->x_family.value.ui32 = cm_entry->x_family;
1641 	cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev;
1642 	cm_ksp_data->x_time.value.ui32 = cm_entry->x_time;
1643 	cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref;
1644 	cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags;
1645 
1646 	if (cm_entry->x_server.buf) {
1647 		fbuf = cm_ksp_data->x_server.value.str.addr.ptr;
1648 		if (cm_entry->x_family == AF_INET &&
1649 		    cm_entry->x_server.len ==
1650 		    sizeof (struct sockaddr_in)) {
1651 			struct sockaddr_in  *sa;
1652 			sa = (struct sockaddr_in *)
1653 				cm_entry->x_server.buf;
1654 			b = (uchar_t *)&sa->sin_addr;
1655 			(void) sprintf(fbuf,
1656 			    "%d.%d.%d.%d", b[0] & 0xFF, b[1] & 0xFF,
1657 			    b[2] & 0xFF, b[3] & 0xFF);
1658 			cm_ksp_data->x_port.value.ui32 = ntohs(sa->sin_port);
1659 		} else if (cm_entry->x_family == AF_INET6 &&
1660 				cm_entry->x_server.len >=
1661 				sizeof (struct sockaddr_in6)) {
1662 			/* extract server IP address & port */
1663 			struct sockaddr_in6 *sin6;
1664 			sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf;
1665 			(void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf,
1666 				INET6_ADDRSTRLEN);
1667 			cm_ksp_data->x_port.value.ui32 = ntohs(sin6->sin6_port);
1668 		} else {
1669 			struct sockaddr_in  *sa;
1670 
1671 			sa = (struct sockaddr_in *)cm_entry->x_server.buf;
1672 			b = (uchar_t *)&sa->sin_addr;
1673 			(void) sprintf(fbuf,
1674 			    "%d.%d.%d.%d", b[0] & 0xFF, b[1] & 0xFF,
1675 			    b[2] & 0xFF, b[3] & 0xFF);
1676 		}
1677 		KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) =
1678 		    strlen(fbuf) + 1;
1679 	}
1680 
1681 	return (0);
1682 }
1683 
1684 
1685 /*
1686  * We want a version of delay which is interruptible by a UNIX signal
1687  * Return EINTR if an interrupt occured.
1688  */
1689 static int
1690 clnt_delay(clock_t ticks, bool_t nosignal)
1691 {
1692 	if (nosignal == TRUE) {
1693 		delay(ticks);
1694 		return (0);
1695 	}
1696 	return (delay_sig(ticks));
1697 }
1698 
1699 /*
1700  * Wait for a connection until a timeout, or until we are
1701  * signalled that there has been a connection state change.
1702  */
1703 static enum clnt_stat
1704 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp,
1705     bool_t nosignal)
1706 {
1707 	bool_t interrupted;
1708 	clock_t timout, cv_stat;
1709 	enum clnt_stat clstat;
1710 	unsigned int old_state;
1711 
1712 	ASSERT(MUTEX_HELD(&connmgr_lock));
1713 	/*
1714 	 * We wait for the transport connection to be made, or an
1715 	 * indication that it could not be made.
1716 	 */
1717 	clstat = RPC_TIMEDOUT;
1718 	interrupted = FALSE;
1719 
1720 	old_state = cm_entry->x_state_flags;
1721 	/*
1722 	 * Now loop until cv_timedwait{_sig} returns because of
1723 	 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be
1724 	 * cv_signalled for various other reasons too. So loop
1725 	 * until there is a state change on the connection.
1726 	 */
1727 
1728 	timout = waitp->tv_sec * drv_usectohz(1000000) +
1729 	    drv_usectohz(waitp->tv_usec) + ddi_get_lbolt();
1730 
1731 	if (nosignal) {
1732 		while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv,
1733 		    &connmgr_lock, timout)) > 0 &&
1734 		    cm_entry->x_state_flags == old_state)
1735 			;
1736 	} else {
1737 		while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv,
1738 		    &connmgr_lock, timout)) > 0 &&
1739 		    cm_entry->x_state_flags == old_state)
1740 			;
1741 
1742 		if (cv_stat == 0) /* got intr signal? */
1743 			interrupted = TRUE;
1744 	}
1745 
1746 	if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) ==
1747 	    X_CONNECTED) {
1748 		clstat = RPC_SUCCESS;
1749 	} else {
1750 		if (interrupted == TRUE)
1751 			clstat = RPC_INTR;
1752 		RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n",
1753 		    clnt_sperrno(clstat));
1754 	}
1755 
1756 	return (clstat);
1757 }
1758 
1759 /*
1760  * Primary interface for how RPC grabs a connection.
1761  */
1762 static struct cm_xprt *
1763 connmgr_wrapget(
1764 	struct netbuf *retryaddr,
1765 	const struct timeval *waitp,
1766 	cku_private_t *p)
1767 {
1768 	struct cm_xprt *cm_entry;
1769 
1770 	cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly,
1771 	    &p->cku_srcaddr, &p->cku_err, p->cku_device,
1772 	    p->cku_client.cl_nosignal, p->cku_useresvport, p->cku_cred);
1773 
1774 	if (cm_entry == NULL) {
1775 		/*
1776 		 * Re-map the call status to RPC_INTR if the err code is
1777 		 * EINTR. This can happen if calls status is RPC_TLIERROR.
1778 		 * However, don't re-map if signalling has been turned off.
1779 		 * XXX Really need to create a separate thread whenever
1780 		 * there isn't an existing connection.
1781 		 */
1782 		if (p->cku_err.re_errno == EINTR) {
1783 			if (p->cku_client.cl_nosignal == TRUE)
1784 				p->cku_err.re_errno = EIO;
1785 			else
1786 				p->cku_err.re_status = RPC_INTR;
1787 		}
1788 	}
1789 
1790 	return (cm_entry);
1791 }
1792 
1793 /*
1794  * Obtains a transport to the server specified in addr.  If a suitable transport
1795  * does not already exist in the list of cached transports, a new connection
1796  * is created, connected, and added to the list. The connection is for sending
1797  * only - the reply message may come back on another transport connection.
1798  *
1799  * To implement round-robin load balancing with multiple client connections,
1800  * the last entry on the list is always selected. Once the entry is selected
1801  * it's re-inserted to the head of the list.
1802  */
1803 static struct cm_xprt *
1804 connmgr_get(
1805 	struct netbuf	*retryaddr,
1806 	const struct timeval	*waitp,	/* changed to a ptr to converse stack */
1807 	struct netbuf	*destaddr,
1808 	int		addrfmly,
1809 	struct netbuf	*srcaddr,
1810 	struct rpc_err	*rpcerr,
1811 	dev_t		device,
1812 	bool_t		nosignal,
1813 	int		useresvport,
1814 	cred_t		*cr)
1815 {
1816 	struct cm_xprt *cm_entry;
1817 	struct cm_xprt *lru_entry;
1818 	struct cm_xprt **cmp, **prev;
1819 	queue_t *wq;
1820 	TIUSER *tiptr;
1821 	int i;
1822 	int retval;
1823 	int tidu_size;
1824 	bool_t	connected;
1825 	zoneid_t zoneid = rpc_zoneid();
1826 
1827 	/*
1828 	 * If the call is not a retry, look for a transport entry that
1829 	 * goes to the server of interest.
1830 	 */
1831 	mutex_enter(&connmgr_lock);
1832 
1833 	if (retryaddr == NULL) {
1834 use_new_conn:
1835 		i = 0;
1836 		cm_entry = lru_entry = NULL;
1837 
1838 		prev = cmp = &cm_hd;
1839 		while ((cm_entry = *cmp) != NULL) {
1840 			ASSERT(cm_entry != cm_entry->x_next);
1841 			/*
1842 			 * Garbage collect conections that are marked
1843 			 * for needs disconnect.
1844 			 */
1845 			if (cm_entry->x_needdis) {
1846 				CONN_HOLD(cm_entry);
1847 				connmgr_dis_and_wait(cm_entry);
1848 				connmgr_release(cm_entry);
1849 				/*
1850 				 * connmgr_lock could have been
1851 				 * dropped for the disconnect
1852 				 * processing so start over.
1853 				 */
1854 				goto use_new_conn;
1855 			}
1856 
1857 			/*
1858 			 * Garbage collect the dead connections that have
1859 			 * no threads working on them.
1860 			 */
1861 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1862 			    X_DEAD) {
1863 				mutex_enter(&cm_entry->x_lock);
1864 				if (cm_entry->x_ref != 0) {
1865 					/*
1866 					 * Currently in use.
1867 					 * Cleanup later.
1868 					 */
1869 					cmp = &cm_entry->x_next;
1870 					mutex_exit(&cm_entry->x_lock);
1871 					continue;
1872 				}
1873 				mutex_exit(&cm_entry->x_lock);
1874 				*cmp = cm_entry->x_next;
1875 				mutex_exit(&connmgr_lock);
1876 				connmgr_close(cm_entry);
1877 				mutex_enter(&connmgr_lock);
1878 				goto use_new_conn;
1879 			}
1880 
1881 
1882 			if ((cm_entry->x_state_flags & X_BADSTATES) == 0 &&
1883 			    cm_entry->x_zoneid == zoneid &&
1884 			    cm_entry->x_rdev == device &&
1885 			    destaddr->len == cm_entry->x_server.len &&
1886 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
1887 			    destaddr->len) == 0) {
1888 				/*
1889 				 * If the matching entry isn't connected,
1890 				 * attempt to reconnect it.
1891 				 */
1892 				if (cm_entry->x_connected == FALSE) {
1893 					/*
1894 					 * We don't go through trying
1895 					 * to find the least recently
1896 					 * used connected because
1897 					 * connmgr_reconnect() briefly
1898 					 * dropped the connmgr_lock,
1899 					 * allowing a window for our
1900 					 * accounting to be messed up.
1901 					 * In any case, a re-connected
1902 					 * connection is as good as
1903 					 * a LRU connection.
1904 					 */
1905 					return (connmgr_wrapconnect(cm_entry,
1906 					    waitp, destaddr, addrfmly, srcaddr,
1907 					    rpcerr, TRUE, nosignal, cr));
1908 				}
1909 				i++;
1910 
1911 				/* keep track of the last entry */
1912 				lru_entry = cm_entry;
1913 				prev = cmp;
1914 			}
1915 			cmp = &cm_entry->x_next;
1916 		}
1917 
1918 		if (i > clnt_max_conns) {
1919 			RPCLOG(8, "connmgr_get: too many conns, dooming entry"
1920 			    " %p\n", (void *)lru_entry->x_tiptr);
1921 			lru_entry->x_doomed = TRUE;
1922 			goto use_new_conn;
1923 		}
1924 
1925 		/*
1926 		 * If we are at the maximum number of connections to
1927 		 * the server, hand back the least recently used one.
1928 		 */
1929 		if (i == clnt_max_conns) {
1930 			/*
1931 			 * Copy into the handle the source address of
1932 			 * the connection, which we will use in case of
1933 			 * a later retry.
1934 			 */
1935 			if (srcaddr->len != lru_entry->x_src.len) {
1936 				if (srcaddr->len > 0)
1937 					kmem_free(srcaddr->buf,
1938 					    srcaddr->maxlen);
1939 				srcaddr->buf = kmem_zalloc(
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_zalloc(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 	cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP);
2260 	bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len);
2261 	cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len;
2262 
2263 	cm_entry->x_tiptr = tiptr;
2264 	cm_entry->x_time = ddi_get_lbolt();
2265 
2266 	if (tiptr->tp_info.servtype == T_COTS_ORD)
2267 		cm_entry->x_ordrel = TRUE;
2268 	else
2269 		cm_entry->x_ordrel = FALSE;
2270 
2271 	cm_entry->x_tidu_size = tidu_size;
2272 
2273 	if (cm_entry->x_early_disc) {
2274 		/*
2275 		 * We need to check if a disconnect request has come
2276 		 * while we are connected, if so, then we need to
2277 		 * set rpcerr->re_status appropriately before returning
2278 		 * NULL to caller.
2279 		 */
2280 		if (rpcerr->re_status == RPC_SUCCESS)
2281 			rpcerr->re_status = RPC_XPRTFAILED;
2282 		cm_entry->x_connected = FALSE;
2283 	} else
2284 		cm_entry->x_connected = connected;
2285 
2286 	/*
2287 	 * There could be a discrepancy here such that
2288 	 * x_early_disc is TRUE yet connected is TRUE as well
2289 	 * and the connection is actually connected. In that case
2290 	 * lets be conservative and declare the connection as not
2291 	 * connected.
2292 	 */
2293 	cm_entry->x_early_disc = FALSE;
2294 	cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2295 	cm_entry->x_ctime = ddi_get_lbolt();
2296 
2297 	/*
2298 	 * Notify any threads waiting that the connection attempt is done.
2299 	 */
2300 	cm_entry->x_thread = FALSE;
2301 	cv_broadcast(&cm_entry->x_conn_cv);
2302 
2303 	if (cm_entry->x_connected == FALSE) {
2304 		mutex_exit(&connmgr_lock);
2305 		connmgr_release(cm_entry);
2306 		return (NULL);
2307 	}
2308 
2309 	mutex_exit(&connmgr_lock);
2310 
2311 	return (cm_entry);
2312 }
2313 
2314 /*
2315  * Keep the cm_xprt entry on the connecton list when making a connection. This
2316  * is to prevent multiple connections to a slow server from appearing.
2317  * We use the bit field x_thread to tell if a thread is doing a connection
2318  * which keeps other interested threads from messing with connection.
2319  * Those other threads just wait if x_thread is set.
2320  *
2321  * If x_thread is not set, then we do the actual work of connecting via
2322  * connmgr_connect().
2323  *
2324  * mutex convention: called with connmgr_lock held, returns with it released.
2325  */
2326 static struct cm_xprt *
2327 connmgr_wrapconnect(
2328 	struct cm_xprt	*cm_entry,
2329 	const struct timeval	*waitp,
2330 	struct netbuf	*destaddr,
2331 	int		addrfmly,
2332 	struct netbuf	*srcaddr,
2333 	struct rpc_err	*rpcerr,
2334 	bool_t		reconnect,
2335 	bool_t		nosignal,
2336 	cred_t		*cr)
2337 {
2338 	ASSERT(MUTEX_HELD(&connmgr_lock));
2339 	/*
2340 	 * Hold this entry as we are about to drop connmgr_lock.
2341 	 */
2342 	CONN_HOLD(cm_entry);
2343 
2344 	/*
2345 	 * If there is a thread already making a connection for us, then
2346 	 * wait for it to complete the connection.
2347 	 */
2348 	if (cm_entry->x_thread == TRUE) {
2349 		rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal);
2350 
2351 		if (rpcerr->re_status != RPC_SUCCESS) {
2352 			mutex_exit(&connmgr_lock);
2353 			connmgr_release(cm_entry);
2354 			return (NULL);
2355 		}
2356 	} else {
2357 		bool_t connected;
2358 		calllist_t call;
2359 
2360 		cm_entry->x_thread = TRUE;
2361 
2362 		while (cm_entry->x_needrel == TRUE) {
2363 			cm_entry->x_needrel = FALSE;
2364 
2365 			connmgr_sndrel(cm_entry);
2366 			delay(drv_usectohz(1000000));
2367 
2368 			mutex_enter(&connmgr_lock);
2369 		}
2370 
2371 		/*
2372 		 * If we need to send a T_DISCON_REQ, send one.
2373 		 */
2374 		connmgr_dis_and_wait(cm_entry);
2375 
2376 		mutex_exit(&connmgr_lock);
2377 
2378 		bzero(&call, sizeof (call));
2379 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2380 
2381 		connected = connmgr_connect(cm_entry, cm_entry->x_wq,
2382 		    destaddr, addrfmly, &call, &cm_entry->x_tidu_size,
2383 		    reconnect, waitp, nosignal, cr);
2384 
2385 		*rpcerr = call.call_err;
2386 		cv_destroy(&call.call_cv);
2387 
2388 		mutex_enter(&connmgr_lock);
2389 
2390 
2391 		if (cm_entry->x_early_disc) {
2392 			/*
2393 			 * We need to check if a disconnect request has come
2394 			 * while we are connected, if so, then we need to
2395 			 * set rpcerr->re_status appropriately before returning
2396 			 * NULL to caller.
2397 			 */
2398 			if (rpcerr->re_status == RPC_SUCCESS)
2399 				rpcerr->re_status = RPC_XPRTFAILED;
2400 			cm_entry->x_connected = FALSE;
2401 		} else
2402 			cm_entry->x_connected = connected;
2403 
2404 		/*
2405 		 * There could be a discrepancy here such that
2406 		 * x_early_disc is TRUE yet connected is TRUE as well
2407 		 * and the connection is actually connected. In that case
2408 		 * lets be conservative and declare the connection as not
2409 		 * connected.
2410 		 */
2411 
2412 		cm_entry->x_early_disc = FALSE;
2413 		cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2414 
2415 
2416 		/*
2417 		 * connmgr_connect() may have given up before the connection
2418 		 * actually timed out. So ensure that before the next
2419 		 * connection attempt we do a disconnect.
2420 		 */
2421 		cm_entry->x_ctime = ddi_get_lbolt();
2422 		cm_entry->x_thread = FALSE;
2423 
2424 		cv_broadcast(&cm_entry->x_conn_cv);
2425 
2426 		if (cm_entry->x_connected == FALSE) {
2427 			mutex_exit(&connmgr_lock);
2428 			connmgr_release(cm_entry);
2429 			return (NULL);
2430 		}
2431 	}
2432 
2433 	if (srcaddr != NULL) {
2434 		/*
2435 		 * Copy into the handle the
2436 		 * source address of the
2437 		 * connection, which we will use
2438 		 * in case of a later retry.
2439 		 */
2440 		if (srcaddr->len != cm_entry->x_src.len) {
2441 			if (srcaddr->maxlen > 0)
2442 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2443 			srcaddr->buf = kmem_zalloc(cm_entry->x_src.len,
2444 			    KM_SLEEP);
2445 			srcaddr->maxlen = srcaddr->len =
2446 			    cm_entry->x_src.len;
2447 		}
2448 		bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len);
2449 	}
2450 	cm_entry->x_time = ddi_get_lbolt();
2451 	mutex_exit(&connmgr_lock);
2452 	return (cm_entry);
2453 }
2454 
2455 /*
2456  * If we need to send a T_DISCON_REQ, send one.
2457  */
2458 static void
2459 connmgr_dis_and_wait(struct cm_xprt *cm_entry)
2460 {
2461 	ASSERT(MUTEX_HELD(&connmgr_lock));
2462 	for (;;) {
2463 		while (cm_entry->x_needdis == TRUE) {
2464 			RPCLOG(8, "connmgr_dis_and_wait: need "
2465 			    "T_DISCON_REQ for connection 0x%p\n",
2466 			    (void *)cm_entry);
2467 			cm_entry->x_needdis = FALSE;
2468 			cm_entry->x_waitdis = TRUE;
2469 
2470 			connmgr_snddis(cm_entry);
2471 
2472 			mutex_enter(&connmgr_lock);
2473 		}
2474 
2475 		if (cm_entry->x_waitdis == TRUE) {
2476 			clock_t timout;
2477 
2478 			RPCLOG(8, "connmgr_dis_and_wait waiting for "
2479 			    "T_DISCON_REQ's ACK for connection %p\n",
2480 			    (void *)cm_entry);
2481 
2482 			timout = clnt_cots_min_conntout * drv_usectohz(1000000);
2483 
2484 			/*
2485 			 * The TPI spec says that the T_DISCON_REQ
2486 			 * will get acknowledged, but in practice
2487 			 * the ACK may never get sent. So don't
2488 			 * block forever.
2489 			 */
2490 			(void) cv_reltimedwait(&cm_entry->x_dis_cv,
2491 			    &connmgr_lock, timout, TR_CLOCK_TICK);
2492 		}
2493 		/*
2494 		 * If we got the ACK, break. If we didn't,
2495 		 * then send another T_DISCON_REQ.
2496 		 */
2497 		if (cm_entry->x_waitdis == FALSE) {
2498 			break;
2499 		} else {
2500 			RPCLOG(8, "connmgr_dis_and_wait: did"
2501 			    "not get T_DISCON_REQ's ACK for "
2502 			    "connection  %p\n", (void *)cm_entry);
2503 			cm_entry->x_needdis = TRUE;
2504 		}
2505 	}
2506 }
2507 
2508 static void
2509 connmgr_cancelconn(struct cm_xprt *cm_entry)
2510 {
2511 	/*
2512 	 * Mark the connection table entry as dead; the next thread that
2513 	 * goes through connmgr_release() will notice this and deal with it.
2514 	 */
2515 	mutex_enter(&connmgr_lock);
2516 	cm_entry->x_dead = TRUE;
2517 
2518 	/*
2519 	 * Notify any threads waiting for the connection that it isn't
2520 	 * going to happen.
2521 	 */
2522 	cm_entry->x_thread = FALSE;
2523 	cv_broadcast(&cm_entry->x_conn_cv);
2524 	mutex_exit(&connmgr_lock);
2525 
2526 	connmgr_release(cm_entry);
2527 }
2528 
2529 static void
2530 connmgr_close(struct cm_xprt *cm_entry)
2531 {
2532 	mutex_enter(&cm_entry->x_lock);
2533 	while (cm_entry->x_ref != 0) {
2534 		/*
2535 		 * Must be a noninterruptible wait.
2536 		 */
2537 		cv_wait(&cm_entry->x_cv, &cm_entry->x_lock);
2538 	}
2539 
2540 	if (cm_entry->x_tiptr != NULL)
2541 		(void) t_kclose(cm_entry->x_tiptr, 1);
2542 
2543 	mutex_exit(&cm_entry->x_lock);
2544 	if (cm_entry->x_ksp != NULL) {
2545 		mutex_enter(&connmgr_lock);
2546 		cm_entry->x_ksp->ks_private = NULL;
2547 		mutex_exit(&connmgr_lock);
2548 
2549 		/*
2550 		 * Must free the buffer we allocated for the
2551 		 * server address in the update function
2552 		 */
2553 		if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2554 		    x_server.value.str.addr.ptr != NULL)
2555 			kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp->
2556 			    ks_data))->x_server.value.str.addr.ptr,
2557 			    INET6_ADDRSTRLEN);
2558 		kmem_free(cm_entry->x_ksp->ks_data,
2559 		    cm_entry->x_ksp->ks_data_size);
2560 		kstat_delete(cm_entry->x_ksp);
2561 	}
2562 
2563 	mutex_destroy(&cm_entry->x_lock);
2564 	cv_destroy(&cm_entry->x_cv);
2565 	cv_destroy(&cm_entry->x_conn_cv);
2566 	cv_destroy(&cm_entry->x_dis_cv);
2567 
2568 	if (cm_entry->x_server.buf != NULL)
2569 		kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen);
2570 	if (cm_entry->x_src.buf != NULL)
2571 		kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen);
2572 	kmem_free(cm_entry, sizeof (struct cm_xprt));
2573 }
2574 
2575 /*
2576  * Called by KRPC after sending the call message to release the connection
2577  * it was using.
2578  */
2579 static void
2580 connmgr_release(struct cm_xprt *cm_entry)
2581 {
2582 	mutex_enter(&cm_entry->x_lock);
2583 	cm_entry->x_ref--;
2584 	if (cm_entry->x_ref == 0)
2585 		cv_signal(&cm_entry->x_cv);
2586 	mutex_exit(&cm_entry->x_lock);
2587 }
2588 
2589 /*
2590  * Set TCP receive and xmit buffer size for RPC connections.
2591  */
2592 static bool_t
2593 connmgr_setbufsz(calllist_t *e, queue_t *wq, cred_t *cr)
2594 {
2595 	int ok = FALSE;
2596 	int val;
2597 
2598 	if (rpc_default_tcp_bufsz)
2599 		return (FALSE);
2600 
2601 	/*
2602 	 * Only set new buffer size if it's larger than the system
2603 	 * default buffer size. If smaller buffer size is needed
2604 	 * then use /etc/system to set rpc_default_tcp_bufsz to 1.
2605 	 */
2606 	ok = connmgr_getopt_int(wq, SOL_SOCKET, SO_RCVBUF, &val, e, cr);
2607 	if ((ok == TRUE) && (val < rpc_send_bufsz)) {
2608 		ok = connmgr_setopt_int(wq, SOL_SOCKET, SO_RCVBUF,
2609 		    rpc_send_bufsz, e, cr);
2610 		DTRACE_PROBE2(krpc__i__connmgr_rcvbufsz,
2611 		    int, ok, calllist_t *, e);
2612 	}
2613 
2614 	ok = connmgr_getopt_int(wq, SOL_SOCKET, SO_SNDBUF, &val, e, cr);
2615 	if ((ok == TRUE) && (val < rpc_recv_bufsz)) {
2616 		ok = connmgr_setopt_int(wq, SOL_SOCKET, SO_SNDBUF,
2617 		    rpc_recv_bufsz, e, cr);
2618 		DTRACE_PROBE2(krpc__i__connmgr_sndbufsz,
2619 		    int, ok, calllist_t *, e);
2620 	}
2621 	return (TRUE);
2622 }
2623 
2624 /*
2625  * Given an open stream, connect to the remote.  Returns true if connected,
2626  * false otherwise.
2627  */
2628 static bool_t
2629 connmgr_connect(
2630 	struct cm_xprt		*cm_entry,
2631 	queue_t			*wq,
2632 	struct netbuf		*addr,
2633 	int			addrfmly,
2634 	calllist_t 		*e,
2635 	int 			*tidu_ptr,
2636 	bool_t 			reconnect,
2637 	const struct timeval 	*waitp,
2638 	bool_t 			nosignal,
2639 	cred_t			*cr)
2640 {
2641 	mblk_t *mp;
2642 	struct T_conn_req *tcr;
2643 	struct T_info_ack *tinfo;
2644 	int interrupted, error;
2645 	int tidu_size, kstat_instance;
2646 
2647 	/* if it's a reconnect, flush any lingering data messages */
2648 	if (reconnect)
2649 		(void) putctl1(wq, M_FLUSH, FLUSHRW);
2650 
2651 	/*
2652 	 * Note: if the receiver uses SCM_UCRED/getpeerucred the pid will
2653 	 * appear as -1.
2654 	 */
2655 	mp = allocb_cred(sizeof (*tcr) + addr->len, cr, NOPID);
2656 	if (mp == NULL) {
2657 		/*
2658 		 * This is unfortunate, but we need to look up the stats for
2659 		 * this zone to increment the "memory allocation failed"
2660 		 * counter.  curproc->p_zone is safe since we're initiating a
2661 		 * connection and not in some strange streams context.
2662 		 */
2663 		struct rpcstat *rpcstat;
2664 
2665 		rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
2666 		ASSERT(rpcstat != NULL);
2667 
2668 		RPCLOG0(1, "connmgr_connect: cannot alloc mp for "
2669 		    "sending conn request\n");
2670 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem);
2671 		e->call_status = RPC_SYSTEMERROR;
2672 		e->call_reason = ENOSR;
2673 		return (FALSE);
2674 	}
2675 
2676 	/* Set TCP buffer size for RPC connections if needed */
2677 	if (addrfmly == AF_INET || addrfmly == AF_INET6)
2678 		(void) connmgr_setbufsz(e, wq, cr);
2679 
2680 	mp->b_datap->db_type = M_PROTO;
2681 	tcr = (struct T_conn_req *)mp->b_rptr;
2682 	bzero(tcr, sizeof (*tcr));
2683 	tcr->PRIM_type = T_CONN_REQ;
2684 	tcr->DEST_length = addr->len;
2685 	tcr->DEST_offset = sizeof (struct T_conn_req);
2686 	mp->b_wptr = mp->b_rptr + sizeof (*tcr);
2687 
2688 	bcopy(addr->buf, mp->b_wptr, tcr->DEST_length);
2689 	mp->b_wptr += tcr->DEST_length;
2690 
2691 	RPCLOG(8, "connmgr_connect: sending conn request on queue "
2692 	    "%p", (void *)wq);
2693 	RPCLOG(8, " call %p\n", (void *)wq);
2694 	/*
2695 	 * We use the entry in the handle that is normally used for
2696 	 * waiting for RPC replies to wait for the connection accept.
2697 	 */
2698 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2699 		DTRACE_PROBE(krpc__e__connmgr__connect__cantsend);
2700 		freemsg(mp);
2701 		return (FALSE);
2702 	}
2703 
2704 	mutex_enter(&clnt_pending_lock);
2705 
2706 	/*
2707 	 * We wait for the transport connection to be made, or an
2708 	 * indication that it could not be made.
2709 	 */
2710 	interrupted = 0;
2711 
2712 	/*
2713 	 * waitforack should have been called with T_OK_ACK, but the
2714 	 * present implementation needs to be passed T_INFO_ACK to
2715 	 * work correctly.
2716 	 */
2717 	error = waitforack(e, T_INFO_ACK, waitp, nosignal);
2718 	if (error == EINTR)
2719 		interrupted = 1;
2720 	if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) {
2721 		/*
2722 		 * No time to lose; we essentially have been signaled to
2723 		 * quit.
2724 		 */
2725 		interrupted = 1;
2726 	}
2727 #ifdef RPCDEBUG
2728 	if (error == ETIME)
2729 		RPCLOG0(8, "connmgr_connect: giving up "
2730 		    "on connection attempt; "
2731 		    "clnt_dispatch notifyconn "
2732 		    "diagnostic 'no one waiting for "
2733 		    "connection' should not be "
2734 		    "unexpected\n");
2735 #endif
2736 	if (e->call_prev)
2737 		e->call_prev->call_next = e->call_next;
2738 	else
2739 		clnt_pending = e->call_next;
2740 	if (e->call_next)
2741 		e->call_next->call_prev = e->call_prev;
2742 	mutex_exit(&clnt_pending_lock);
2743 
2744 	if (e->call_status != RPC_SUCCESS || error != 0) {
2745 		if (interrupted)
2746 			e->call_status = RPC_INTR;
2747 		else if (error == ETIME)
2748 			e->call_status = RPC_TIMEDOUT;
2749 		else if (error == EPROTO) {
2750 			e->call_status = RPC_SYSTEMERROR;
2751 			e->call_reason = EPROTO;
2752 		}
2753 
2754 		RPCLOG(8, "connmgr_connect: can't connect, status: "
2755 		    "%s\n", clnt_sperrno(e->call_status));
2756 
2757 		if (e->call_reply) {
2758 			freemsg(e->call_reply);
2759 			e->call_reply = NULL;
2760 		}
2761 
2762 		return (FALSE);
2763 	}
2764 	/*
2765 	 * The result of the "connection accept" is a T_info_ack
2766 	 * in the call_reply field.
2767 	 */
2768 	ASSERT(e->call_reply != NULL);
2769 	mp = e->call_reply;
2770 	e->call_reply = NULL;
2771 	tinfo = (struct T_info_ack *)mp->b_rptr;
2772 
2773 	tidu_size = tinfo->TIDU_size;
2774 	tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT);
2775 	if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0))
2776 		tidu_size = COTS_DEFAULT_ALLOCSIZE;
2777 	*tidu_ptr = tidu_size;
2778 
2779 	freemsg(mp);
2780 
2781 	/*
2782 	 * Set up the pertinent options.  NODELAY is so the transport doesn't
2783 	 * buffer up RPC messages on either end.  This may not be valid for
2784 	 * all transports.  Failure to set this option is not cause to
2785 	 * bail out so we return success anyway.  Note that lack of NODELAY
2786 	 * or some other way to flush the message on both ends will cause
2787 	 * lots of retries and terrible performance.
2788 	 */
2789 	if (addrfmly == AF_INET || addrfmly == AF_INET6) {
2790 		(void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e, cr);
2791 		if (e->call_status == RPC_XPRTFAILED)
2792 			return (FALSE);
2793 	}
2794 
2795 	/*
2796 	 * Since we have a connection, we now need to figure out if
2797 	 * we need to create a kstat. If x_ksp is not NULL then we
2798 	 * are reusing a connection and so we do not need to create
2799 	 * another kstat -- lets just return.
2800 	 */
2801 	if (cm_entry->x_ksp != NULL)
2802 		return (TRUE);
2803 
2804 	/*
2805 	 * We need to increment rpc_kstat_instance atomically to prevent
2806 	 * two kstats being created with the same instance.
2807 	 */
2808 	kstat_instance = atomic_inc_32_nv((uint32_t *)&rpc_kstat_instance);
2809 
2810 	if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance,
2811 	    "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED,
2812 	    (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)),
2813 	    KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) {
2814 		return (TRUE);
2815 	}
2816 
2817 	cm_entry->x_ksp->ks_lock = &connmgr_lock;
2818 	cm_entry->x_ksp->ks_private = cm_entry;
2819 	cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char))
2820 	    + sizeof (cm_kstat_template));
2821 	cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size,
2822 	    KM_SLEEP);
2823 	bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data,
2824 	    cm_entry->x_ksp->ks_data_size);
2825 	((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2826 	    x_server.value.str.addr.ptr =
2827 	    kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP);
2828 
2829 	cm_entry->x_ksp->ks_update = conn_kstat_update;
2830 	kstat_install(cm_entry->x_ksp);
2831 	return (TRUE);
2832 }
2833 
2834 /*
2835  * Verify that the specified offset falls within the mblk and
2836  * that the resulting pointer is aligned.
2837  * Returns NULL if not.
2838  *
2839  * code from fs/sockfs/socksubr.c
2840  */
2841 static void *
2842 connmgr_opt_getoff(mblk_t *mp, t_uscalar_t offset,
2843     t_uscalar_t length, uint_t align_size)
2844 {
2845 	uintptr_t ptr1, ptr2;
2846 
2847 	ASSERT(mp && mp->b_wptr >= mp->b_rptr);
2848 	ptr1 = (uintptr_t)mp->b_rptr + offset;
2849 	ptr2 = (uintptr_t)ptr1 + length;
2850 	if (ptr1 < (uintptr_t)mp->b_rptr || ptr2 > (uintptr_t)mp->b_wptr) {
2851 		return (NULL);
2852 	}
2853 	if ((ptr1 & (align_size - 1)) != 0) {
2854 		return (NULL);
2855 	}
2856 	return ((void *)ptr1);
2857 }
2858 
2859 static bool_t
2860 connmgr_getopt_int(queue_t *wq, int level, int name, int *val,
2861     calllist_t *e, cred_t *cr)
2862 {
2863 	mblk_t *mp;
2864 	struct opthdr *opt, *opt_res;
2865 	struct T_optmgmt_req *tor;
2866 	struct T_optmgmt_ack *opt_ack;
2867 	struct timeval waitp;
2868 	int error;
2869 
2870 	mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2871 	    sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2872 	if (mp == NULL)
2873 		return (FALSE);
2874 
2875 	mp->b_datap->db_type = M_PROTO;
2876 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2877 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2878 	tor->MGMT_flags = T_CURRENT;
2879 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2880 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2881 
2882 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2883 	opt->level = level;
2884 	opt->name = name;
2885 	opt->len = sizeof (int);
2886 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2887 	    sizeof (int);
2888 
2889 	/*
2890 	 * We will use this connection regardless
2891 	 * of whether or not the option is readable.
2892 	 */
2893 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2894 		DTRACE_PROBE(krpc__e__connmgr__getopt__cantsend);
2895 		freemsg(mp);
2896 		return (FALSE);
2897 	}
2898 
2899 	mutex_enter(&clnt_pending_lock);
2900 
2901 	waitp.tv_sec = clnt_cots_min_conntout;
2902 	waitp.tv_usec = 0;
2903 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2904 
2905 	if (e->call_prev)
2906 		e->call_prev->call_next = e->call_next;
2907 	else
2908 		clnt_pending = e->call_next;
2909 	if (e->call_next)
2910 		e->call_next->call_prev = e->call_prev;
2911 	mutex_exit(&clnt_pending_lock);
2912 
2913 	/* get reply message */
2914 	mp = e->call_reply;
2915 	e->call_reply = NULL;
2916 
2917 	if ((!mp) || (e->call_status != RPC_SUCCESS) || (error != 0)) {
2918 
2919 		DTRACE_PROBE4(krpc__e__connmgr_getopt, int, name,
2920 		    int, e->call_status, int, error, mblk_t *, mp);
2921 
2922 		if (mp)
2923 			freemsg(mp);
2924 		return (FALSE);
2925 	}
2926 
2927 	opt_ack = (struct T_optmgmt_ack *)mp->b_rptr;
2928 	opt_res = (struct opthdr *)connmgr_opt_getoff(mp, opt_ack->OPT_offset,
2929 	    opt_ack->OPT_length, __TPI_ALIGN_SIZE);
2930 
2931 	if (!opt_res) {
2932 		DTRACE_PROBE4(krpc__e__connmgr_optres, mblk_t *, mp, int, name,
2933 		    int, opt_ack->OPT_offset, int, opt_ack->OPT_length);
2934 		freemsg(mp);
2935 		return (FALSE);
2936 	}
2937 	*val = *(int *)&opt_res[1];
2938 
2939 	DTRACE_PROBE2(connmgr_getopt__ok, int, name, int, *val);
2940 
2941 	freemsg(mp);
2942 	return (TRUE);
2943 }
2944 
2945 /*
2946  * Called by connmgr_connect to set an option on the new stream.
2947  */
2948 static bool_t
2949 connmgr_setopt_int(queue_t *wq, int level, int name, int val,
2950     calllist_t *e, cred_t *cr)
2951 {
2952 	mblk_t *mp;
2953 	struct opthdr *opt;
2954 	struct T_optmgmt_req *tor;
2955 	struct timeval waitp;
2956 	int error;
2957 
2958 	mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2959 	    sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2960 	if (mp == NULL) {
2961 		RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option "
2962 		    "request\n");
2963 		return (FALSE);
2964 	}
2965 
2966 	mp->b_datap->db_type = M_PROTO;
2967 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2968 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2969 	tor->MGMT_flags = T_NEGOTIATE;
2970 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2971 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2972 
2973 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2974 	opt->level = level;
2975 	opt->name = name;
2976 	opt->len = sizeof (int);
2977 	*(int *)((char *)opt + sizeof (*opt)) = val;
2978 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2979 	    sizeof (int);
2980 
2981 	/*
2982 	 * We will use this connection regardless
2983 	 * of whether or not the option is settable.
2984 	 */
2985 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2986 		DTRACE_PROBE(krpc__e__connmgr__setopt__cantsend);
2987 		freemsg(mp);
2988 		return (FALSE);
2989 	}
2990 
2991 	mutex_enter(&clnt_pending_lock);
2992 
2993 	waitp.tv_sec = clnt_cots_min_conntout;
2994 	waitp.tv_usec = 0;
2995 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2996 
2997 	if (e->call_prev)
2998 		e->call_prev->call_next = e->call_next;
2999 	else
3000 		clnt_pending = e->call_next;
3001 	if (e->call_next)
3002 		e->call_next->call_prev = e->call_prev;
3003 	mutex_exit(&clnt_pending_lock);
3004 
3005 	if (e->call_reply != NULL) {
3006 		freemsg(e->call_reply);
3007 		e->call_reply = NULL;
3008 	}
3009 
3010 	if (e->call_status != RPC_SUCCESS || error != 0) {
3011 		RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name);
3012 		return (FALSE);
3013 	}
3014 	RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name);
3015 	return (TRUE);
3016 }
3017 
3018 static bool_t
3019 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e, cred_t *cr)
3020 {
3021 	return (connmgr_setopt_int(wq, level, name, 1, e, cr));
3022 }
3023 
3024 #ifdef	DEBUG
3025 
3026 /*
3027  * This is a knob to let us force code coverage in allocation failure
3028  * case.
3029  */
3030 static int	connmgr_failsnd;
3031 #define	CONN_SND_ALLOC(Size, Pri)	\
3032 	((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri))
3033 
3034 #else
3035 
3036 #define	CONN_SND_ALLOC(Size, Pri)	allocb(Size, Pri)
3037 
3038 #endif
3039 
3040 /*
3041  * Sends an orderly release on the specified queue.
3042  * Entered with connmgr_lock. Exited without connmgr_lock
3043  */
3044 static void
3045 connmgr_sndrel(struct cm_xprt *cm_entry)
3046 {
3047 	struct T_ordrel_req *torr;
3048 	mblk_t *mp;
3049 	queue_t *q = cm_entry->x_wq;
3050 	ASSERT(MUTEX_HELD(&connmgr_lock));
3051 	mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO);
3052 	if (mp == NULL) {
3053 		cm_entry->x_needrel = TRUE;
3054 		mutex_exit(&connmgr_lock);
3055 		RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel "
3056 		    "to queue %p\n", (void *)q);
3057 		return;
3058 	}
3059 	mutex_exit(&connmgr_lock);
3060 
3061 	mp->b_datap->db_type = M_PROTO;
3062 	torr = (struct T_ordrel_req *)(mp->b_rptr);
3063 	torr->PRIM_type = T_ORDREL_REQ;
3064 	mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req);
3065 
3066 	RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q);
3067 	put(q, mp);
3068 }
3069 
3070 /*
3071  * Sends an disconnect on the specified queue.
3072  * Entered with connmgr_lock. Exited without connmgr_lock
3073  */
3074 static void
3075 connmgr_snddis(struct cm_xprt *cm_entry)
3076 {
3077 	struct T_discon_req *tdis;
3078 	mblk_t *mp;
3079 	queue_t *q = cm_entry->x_wq;
3080 
3081 	ASSERT(MUTEX_HELD(&connmgr_lock));
3082 	mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO);
3083 	if (mp == NULL) {
3084 		cm_entry->x_needdis = TRUE;
3085 		mutex_exit(&connmgr_lock);
3086 		RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon "
3087 		    "to queue %p\n", (void *)q);
3088 		return;
3089 	}
3090 	mutex_exit(&connmgr_lock);
3091 
3092 	mp->b_datap->db_type = M_PROTO;
3093 	tdis = (struct T_discon_req *)mp->b_rptr;
3094 	tdis->PRIM_type = T_DISCON_REQ;
3095 	mp->b_wptr = mp->b_rptr + sizeof (*tdis);
3096 
3097 	RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q);
3098 	put(q, mp);
3099 }
3100 
3101 /*
3102  * Sets up the entry for receiving replies, and calls rpcmod's write put proc
3103  * (through put) to send the call.
3104  */
3105 static int
3106 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid,
3107     uint_t queue_flag)
3108 {
3109 	ASSERT(e != NULL);
3110 
3111 	e->call_status = RPC_TIMEDOUT;	/* optimistic, eh? */
3112 	e->call_reason = 0;
3113 	e->call_wq = q;
3114 	e->call_xid = xid;
3115 	e->call_notified = FALSE;
3116 
3117 	if (!canput(q)) {
3118 		e->call_status = RPC_CANTSEND;
3119 		e->call_reason = ENOBUFS;
3120 		return (RPC_CANTSEND);
3121 	}
3122 
3123 	/*
3124 	 * If queue_flag is set then the calllist_t is already on the hash
3125 	 * queue.  In this case just send the message and return.
3126 	 */
3127 	if (queue_flag) {
3128 		put(q, mp);
3129 		return (RPC_SUCCESS);
3130 
3131 	}
3132 
3133 	/*
3134 	 * Set up calls for RPC requests (with XID != 0) on the hash
3135 	 * queue for fast lookups and place other calls (i.e.
3136 	 * connection management) on the linked list.
3137 	 */
3138 	if (xid != 0) {
3139 		RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on "
3140 		    "dispatch list\n", xid);
3141 		e->call_hash = call_hash(xid, clnt_cots_hash_size);
3142 		e->call_bucket = &cots_call_ht[e->call_hash];
3143 		call_table_enter(e);
3144 	} else {
3145 		mutex_enter(&clnt_pending_lock);
3146 		if (clnt_pending)
3147 			clnt_pending->call_prev = e;
3148 		e->call_next = clnt_pending;
3149 		e->call_prev = NULL;
3150 		clnt_pending = e;
3151 		mutex_exit(&clnt_pending_lock);
3152 	}
3153 
3154 	put(q, mp);
3155 	return (RPC_SUCCESS);
3156 }
3157 
3158 /*
3159  * Called by rpcmod to notify a client with a clnt_pending call that its reply
3160  * has arrived.  If we can't find a client waiting for this reply, we log
3161  * the error and return.
3162  */
3163 bool_t
3164 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid)
3165 {
3166 	calllist_t *e = NULL;
3167 	call_table_t *chtp;
3168 	uint32_t xid;
3169 	uint_t hash;
3170 
3171 	if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) &&
3172 	    (mp->b_wptr - mp->b_rptr) >= sizeof (xid))
3173 		xid = *((uint32_t *)mp->b_rptr);
3174 	else {
3175 		int i = 0;
3176 		unsigned char *p = (unsigned char *)&xid;
3177 		unsigned char *rptr;
3178 		mblk_t *tmp = mp;
3179 
3180 		/*
3181 		 * Copy the xid, byte-by-byte into xid.
3182 		 */
3183 		while (tmp) {
3184 			rptr = tmp->b_rptr;
3185 			while (rptr < tmp->b_wptr) {
3186 				*p++ = *rptr++;
3187 				if (++i >= sizeof (xid))
3188 					goto done_xid_copy;
3189 			}
3190 			tmp = tmp->b_cont;
3191 		}
3192 
3193 		/*
3194 		 * If we got here, we ran out of mblk space before the
3195 		 * xid could be copied.
3196 		 */
3197 		ASSERT(tmp == NULL && i < sizeof (xid));
3198 
3199 		RPCLOG0(1,
3200 		    "clnt_dispatch_notify: message less than size of xid\n");
3201 		return (FALSE);
3202 
3203 	}
3204 done_xid_copy:
3205 
3206 	hash = call_hash(xid, clnt_cots_hash_size);
3207 	chtp = &cots_call_ht[hash];
3208 	/* call_table_find returns with the hash bucket locked */
3209 	call_table_find(chtp, xid, e);
3210 
3211 	if (e != NULL) {
3212 		/*
3213 		 * Found thread waiting for this reply
3214 		 */
3215 		mutex_enter(&e->call_lock);
3216 
3217 		/*
3218 		 * verify that the reply is coming in on
3219 		 * the same zone that it was sent from.
3220 		 */
3221 		if (e->call_zoneid != zoneid) {
3222 			mutex_exit(&e->call_lock);
3223 			mutex_exit(&chtp->ct_lock);
3224 			RPCLOG0(1, "clnt_dispatch_notify: incorrect zoneid\n");
3225 			return (FALSE);
3226 		}
3227 
3228 		if (e->call_reply)
3229 			/*
3230 			 * This can happen under the following scenario:
3231 			 * clnt_cots_kcallit() times out on the response,
3232 			 * rfscall() repeats the CLNT_CALL() with
3233 			 * the same xid, clnt_cots_kcallit() sends the retry,
3234 			 * thereby putting the clnt handle on the pending list,
3235 			 * the first response arrives, signalling the thread
3236 			 * in clnt_cots_kcallit(). Before that thread is
3237 			 * dispatched, the second response arrives as well,
3238 			 * and clnt_dispatch_notify still finds the handle on
3239 			 * the pending list, with call_reply set. So free the
3240 			 * old reply now.
3241 			 *
3242 			 * It is also possible for a response intended for
3243 			 * an RPC call with a different xid to reside here.
3244 			 * This can happen if the thread that owned this
3245 			 * client handle prior to the current owner bailed
3246 			 * out and left its call record on the dispatch
3247 			 * queue.  A window exists where the response can
3248 			 * arrive before the current owner dispatches its
3249 			 * RPC call.
3250 			 *
3251 			 * In any case, this is the very last point where we
3252 			 * can safely check the call_reply field before
3253 			 * placing the new response there.
3254 			 */
3255 			freemsg(e->call_reply);
3256 		e->call_reply = mp;
3257 		e->call_status = RPC_SUCCESS;
3258 		e->call_notified = TRUE;
3259 		cv_signal(&e->call_cv);
3260 		mutex_exit(&e->call_lock);
3261 		mutex_exit(&chtp->ct_lock);
3262 		return (TRUE);
3263 	} else {
3264 		zone_t *zone;
3265 		struct rpcstat *rpcstat;
3266 
3267 		mutex_exit(&chtp->ct_lock);
3268 		RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n",
3269 		    xid);
3270 		/*
3271 		 * This is unfortunate, but we need to lookup the zone so we
3272 		 * can increment its "rcbadxids" counter.
3273 		 */
3274 		zone = zone_find_by_id(zoneid);
3275 		if (zone == NULL) {
3276 			/*
3277 			 * The zone went away...
3278 			 */
3279 			return (FALSE);
3280 		}
3281 		rpcstat = zone_getspecific(rpcstat_zone_key, zone);
3282 		if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) {
3283 			/*
3284 			 * Not interested
3285 			 */
3286 			zone_rele(zone);
3287 			return (FALSE);
3288 		}
3289 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids);
3290 		zone_rele(zone);
3291 	}
3292 	return (FALSE);
3293 }
3294 
3295 /*
3296  * Called by rpcmod when a non-data indication arrives.  The ones in which we
3297  * are interested are connection indications and options acks.  We dispatch
3298  * based on the queue the indication came in on.  If we are not interested in
3299  * what came in, we return false to rpcmod, who will then pass it upstream.
3300  */
3301 bool_t
3302 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp)
3303 {
3304 	calllist_t *e;
3305 	int type;
3306 
3307 	ASSERT((q->q_flag & QREADR) == 0);
3308 
3309 	type = ((union T_primitives *)mp->b_rptr)->type;
3310 	RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n",
3311 	    rpc_tpiprim2name(type));
3312 	mutex_enter(&clnt_pending_lock);
3313 	for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) {
3314 		if (e == NULL) {
3315 			mutex_exit(&clnt_pending_lock);
3316 			RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting "
3317 			    "for connection on queue 0x%p\n", (void *)q);
3318 			return (FALSE);
3319 		}
3320 		if (e->call_wq == q)
3321 			break;
3322 	}
3323 
3324 	switch (type) {
3325 	case T_CONN_CON:
3326 		/*
3327 		 * The transport is now connected, send a T_INFO_REQ to get
3328 		 * the tidu size.
3329 		 */
3330 		mutex_exit(&clnt_pending_lock);
3331 		ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >=
3332 		    sizeof (struct T_info_req));
3333 		mp->b_rptr = mp->b_datap->db_base;
3334 		((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ;
3335 		mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req);
3336 		mp->b_datap->db_type = M_PCPROTO;
3337 		put(q, mp);
3338 		return (TRUE);
3339 	case T_INFO_ACK:
3340 	case T_OPTMGMT_ACK:
3341 		e->call_status = RPC_SUCCESS;
3342 		e->call_reply = mp;
3343 		e->call_notified = TRUE;
3344 		cv_signal(&e->call_cv);
3345 		break;
3346 	case T_ERROR_ACK:
3347 		e->call_status = RPC_CANTCONNECT;
3348 		e->call_reply = mp;
3349 		e->call_notified = TRUE;
3350 		cv_signal(&e->call_cv);
3351 		break;
3352 	case T_OK_ACK:
3353 		/*
3354 		 * Great, but we are really waiting for a T_CONN_CON
3355 		 */
3356 		freemsg(mp);
3357 		break;
3358 	default:
3359 		mutex_exit(&clnt_pending_lock);
3360 		RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type);
3361 		return (FALSE);
3362 	}
3363 
3364 	mutex_exit(&clnt_pending_lock);
3365 	return (TRUE);
3366 }
3367 
3368 /*
3369  * Called by rpcmod when the transport is (or should be) going away.  Informs
3370  * all callers waiting for replies and marks the entry in the connection
3371  * manager's list as unconnected, and either closing (close handshake in
3372  * progress) or dead.
3373  */
3374 void
3375 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason)
3376 {
3377 	calllist_t *e;
3378 	call_table_t *ctp;
3379 	struct cm_xprt *cm_entry;
3380 	int have_connmgr_lock;
3381 	int i;
3382 
3383 	ASSERT((q->q_flag & QREADR) == 0);
3384 
3385 	RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q);
3386 	RPCLOG(1, " received a notifcation prim type [%s]",
3387 	    rpc_tpiprim2name(msg_type));
3388 	RPCLOG(1, " and reason %d\n", reason);
3389 
3390 	/*
3391 	 * Find the transport entry in the connection manager's list, close
3392 	 * the transport and delete the entry.  In the case where rpcmod's
3393 	 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we
3394 	 * should gracefully close the connection.
3395 	 */
3396 	have_connmgr_lock = 1;
3397 	mutex_enter(&connmgr_lock);
3398 	for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) {
3399 		ASSERT(cm_entry != cm_entry->x_next);
3400 		if (cm_entry->x_wq == q) {
3401 			ASSERT(MUTEX_HELD(&connmgr_lock));
3402 			ASSERT(have_connmgr_lock == 1);
3403 			switch (msg_type) {
3404 			case T_ORDREL_REQ:
3405 
3406 				if (cm_entry->x_dead) {
3407 					RPCLOG(1, "idle timeout on dead "
3408 					    "connection: %p\n",
3409 					    (void *)cm_entry);
3410 					if (clnt_stop_idle != NULL)
3411 						(*clnt_stop_idle)(q);
3412 					break;
3413 				}
3414 
3415 				/*
3416 				 * Only mark the connection as dead if it is
3417 				 * connected and idle.
3418 				 * An unconnected connection has probably
3419 				 * gone idle because the server is down,
3420 				 * and when it comes back up there will be
3421 				 * retries that need to use that connection.
3422 				 */
3423 				if (cm_entry->x_connected ||
3424 				    cm_entry->x_doomed) {
3425 					if (cm_entry->x_ordrel) {
3426 						if (cm_entry->x_closing ==
3427 						    TRUE) {
3428 							/*
3429 							 * The connection is
3430 							 * obviously wedged due
3431 							 * to a bug or problem
3432 							 * with the transport.
3433 							 * Mark it as dead.
3434 							 * Otherwise we can
3435 							 * leak connections.
3436 							 */
3437 							cm_entry->x_dead = TRUE;
3438 							mutex_exit(
3439 							    &connmgr_lock);
3440 							have_connmgr_lock = 0;
3441 							if (clnt_stop_idle !=
3442 							    NULL)
3443 							(*clnt_stop_idle)(q);
3444 							break;
3445 						}
3446 						cm_entry->x_closing = TRUE;
3447 						connmgr_sndrel(cm_entry);
3448 						have_connmgr_lock = 0;
3449 					} else {
3450 						cm_entry->x_dead = TRUE;
3451 						mutex_exit(&connmgr_lock);
3452 						have_connmgr_lock = 0;
3453 						if (clnt_stop_idle != NULL)
3454 							(*clnt_stop_idle)(q);
3455 					}
3456 				} else {
3457 					/*
3458 					 * We don't mark the connection
3459 					 * as dead, but we turn off the
3460 					 * idle timer.
3461 					 */
3462 					mutex_exit(&connmgr_lock);
3463 					have_connmgr_lock = 0;
3464 					if (clnt_stop_idle != NULL)
3465 						(*clnt_stop_idle)(q);
3466 					RPCLOG(1, "clnt_dispatch_notifyall:"
3467 					    " ignoring timeout from rpcmod"
3468 					    " (q %p) because we are not "
3469 					    " connected\n", (void *)q);
3470 				}
3471 				break;
3472 			case T_ORDREL_IND:
3473 				/*
3474 				 * If this entry is marked closing, then we are
3475 				 * completing a close handshake, and the
3476 				 * connection is dead.  Otherwise, the server is
3477 				 * trying to close. Since the server will not
3478 				 * be sending any more RPC replies, we abort
3479 				 * the connection, including flushing
3480 				 * any RPC requests that are in-transit.
3481 				 * In either case, mark the entry as dead so
3482 				 * that it can be closed by the connection
3483 				 * manager's garbage collector.
3484 				 */
3485 				cm_entry->x_dead = TRUE;
3486 				if (cm_entry->x_closing) {
3487 					mutex_exit(&connmgr_lock);
3488 					have_connmgr_lock = 0;
3489 					if (clnt_stop_idle != NULL)
3490 						(*clnt_stop_idle)(q);
3491 				} else {
3492 					/*
3493 					 * if we're getting a disconnect
3494 					 * before we've finished our
3495 					 * connect attempt, mark it for
3496 					 * later processing
3497 					 */
3498 					if (cm_entry->x_thread)
3499 						cm_entry->x_early_disc = TRUE;
3500 					else
3501 						cm_entry->x_connected = FALSE;
3502 					cm_entry->x_waitdis = TRUE;
3503 					connmgr_snddis(cm_entry);
3504 					have_connmgr_lock = 0;
3505 				}
3506 				break;
3507 
3508 			case T_ERROR_ACK:
3509 			case T_OK_ACK:
3510 				cm_entry->x_waitdis = FALSE;
3511 				cv_signal(&cm_entry->x_dis_cv);
3512 				mutex_exit(&connmgr_lock);
3513 				return;
3514 
3515 			case T_DISCON_REQ:
3516 				if (cm_entry->x_thread)
3517 					cm_entry->x_early_disc = TRUE;
3518 				else
3519 					cm_entry->x_connected = FALSE;
3520 				cm_entry->x_waitdis = TRUE;
3521 
3522 				connmgr_snddis(cm_entry);
3523 				have_connmgr_lock = 0;
3524 				break;
3525 
3526 			case T_DISCON_IND:
3527 			default:
3528 				/*
3529 				 * if we're getting a disconnect before
3530 				 * we've finished our connect attempt,
3531 				 * mark it for later processing
3532 				 */
3533 				if (cm_entry->x_closing) {
3534 					cm_entry->x_dead = TRUE;
3535 					mutex_exit(&connmgr_lock);
3536 					have_connmgr_lock = 0;
3537 					if (clnt_stop_idle != NULL)
3538 						(*clnt_stop_idle)(q);
3539 				} else {
3540 					if (cm_entry->x_thread) {
3541 						cm_entry->x_early_disc = TRUE;
3542 					} else {
3543 						cm_entry->x_dead = TRUE;
3544 						cm_entry->x_connected = FALSE;
3545 					}
3546 				}
3547 				break;
3548 			}
3549 			break;
3550 		}
3551 	}
3552 
3553 	if (have_connmgr_lock)
3554 		mutex_exit(&connmgr_lock);
3555 
3556 	if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) {
3557 		RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find "
3558 		    "connmgr entry for discon ack\n", (void *)q);
3559 		return;
3560 	}
3561 
3562 	/*
3563 	 * Then kick all the clnt_pending calls out of their wait.  There
3564 	 * should be no clnt_pending calls in the case of rpcmod's idle
3565 	 * timer firing.
3566 	 */
3567 	for (i = 0; i < clnt_cots_hash_size; i++) {
3568 		ctp = &cots_call_ht[i];
3569 		mutex_enter(&ctp->ct_lock);
3570 		for (e = ctp->ct_call_next;
3571 		    e != (calllist_t *)ctp;
3572 		    e = e->call_next) {
3573 			if (e->call_wq == q && e->call_notified == FALSE) {
3574 				RPCLOG(1,
3575 				    "clnt_dispatch_notifyall for queue %p ",
3576 				    (void *)q);
3577 				RPCLOG(1, "aborting clnt_pending call %p\n",
3578 				    (void *)e);
3579 
3580 				if (msg_type == T_DISCON_IND)
3581 					e->call_reason = reason;
3582 				e->call_notified = TRUE;
3583 				e->call_status = RPC_XPRTFAILED;
3584 				cv_signal(&e->call_cv);
3585 			}
3586 		}
3587 		mutex_exit(&ctp->ct_lock);
3588 	}
3589 
3590 	mutex_enter(&clnt_pending_lock);
3591 	for (e = clnt_pending; e; e = e->call_next) {
3592 		/*
3593 		 * Only signal those RPC handles that haven't been
3594 		 * signalled yet. Otherwise we can get a bogus call_reason.
3595 		 * This can happen if thread A is making a call over a
3596 		 * connection. If the server is killed, it will cause
3597 		 * reset, and reason will default to EIO as a result of
3598 		 * a T_ORDREL_IND. Thread B then attempts to recreate
3599 		 * the connection but gets a T_DISCON_IND. If we set the
3600 		 * call_reason code for all threads, then if thread A
3601 		 * hasn't been dispatched yet, it will get the wrong
3602 		 * reason. The bogus call_reason can make it harder to
3603 		 * discriminate between calls that fail because the
3604 		 * connection attempt failed versus those where the call
3605 		 * may have been executed on the server.
3606 		 */
3607 		if (e->call_wq == q && e->call_notified == FALSE) {
3608 			RPCLOG(1, "clnt_dispatch_notifyall for queue %p ",
3609 			    (void *)q);
3610 			RPCLOG(1, " aborting clnt_pending call %p\n",
3611 			    (void *)e);
3612 
3613 			if (msg_type == T_DISCON_IND)
3614 				e->call_reason = reason;
3615 			e->call_notified = TRUE;
3616 			/*
3617 			 * Let the caller timeout, else it will retry
3618 			 * immediately.
3619 			 */
3620 			e->call_status = RPC_XPRTFAILED;
3621 
3622 			/*
3623 			 * We used to just signal those threads
3624 			 * waiting for a connection, (call_xid = 0).
3625 			 * That meant that threads waiting for a response
3626 			 * waited till their timeout expired. This
3627 			 * could be a long time if they've specified a
3628 			 * maximum timeout. (2^31 - 1). So we
3629 			 * Signal all threads now.
3630 			 */
3631 			cv_signal(&e->call_cv);
3632 		}
3633 	}
3634 	mutex_exit(&clnt_pending_lock);
3635 }
3636 
3637 
3638 /*ARGSUSED*/
3639 /*
3640  * after resuming a system that's been suspended for longer than the
3641  * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall()
3642  * generates "NFS server X not responding" and "NFS server X ok" messages;
3643  * here we reset inet connections to cause a re-connect and avoid those
3644  * NFS messages.  see 4045054
3645  */
3646 boolean_t
3647 connmgr_cpr_reset(void *arg, int code)
3648 {
3649 	struct cm_xprt *cxp;
3650 
3651 	if (code == CB_CODE_CPR_CHKPT)
3652 		return (B_TRUE);
3653 
3654 	if (mutex_tryenter(&connmgr_lock) == 0)
3655 		return (B_FALSE);
3656 	for (cxp = cm_hd; cxp; cxp = cxp->x_next) {
3657 		if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) &&
3658 		    cxp->x_connected == TRUE) {
3659 			if (cxp->x_thread)
3660 				cxp->x_early_disc = TRUE;
3661 			else
3662 				cxp->x_connected = FALSE;
3663 			cxp->x_needdis = TRUE;
3664 		}
3665 	}
3666 	mutex_exit(&connmgr_lock);
3667 	return (B_TRUE);
3668 }
3669 
3670 void
3671 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp)
3672 {
3673 
3674 	*statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid,
3675 	    "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl,
3676 	    sizeof (cots_rcstat_tmpl));
3677 }
3678 
3679 void
3680 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp)
3681 {
3682 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client");
3683 	kmem_free(*statsp, sizeof (cots_rcstat_tmpl));
3684 }
3685 
3686 void
3687 clnt_cots_init(void)
3688 {
3689 	mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL);
3690 	mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL);
3691 
3692 	if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE)
3693 		clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE;
3694 
3695 	cots_call_ht = call_table_init(clnt_cots_hash_size);
3696 	zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy);
3697 }
3698 
3699 void
3700 clnt_cots_fini(void)
3701 {
3702 	(void) zone_key_delete(zone_cots_key);
3703 }
3704 
3705 /*
3706  * Wait for TPI ack, returns success only if expected ack is received
3707  * within timeout period.
3708  */
3709 
3710 static int
3711 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp,
3712     bool_t nosignal)
3713 {
3714 	union T_primitives *tpr;
3715 	clock_t timout;
3716 	int cv_stat = 1;
3717 
3718 	ASSERT(MUTEX_HELD(&clnt_pending_lock));
3719 	while (e->call_reply == NULL) {
3720 		if (waitp != NULL) {
3721 			timout = waitp->tv_sec * drv_usectohz(MICROSEC) +
3722 			    drv_usectohz(waitp->tv_usec);
3723 			if (nosignal)
3724 				cv_stat = cv_reltimedwait(&e->call_cv,
3725 				    &clnt_pending_lock, timout, TR_CLOCK_TICK);
3726 			else
3727 				cv_stat = cv_reltimedwait_sig(&e->call_cv,
3728 				    &clnt_pending_lock, timout, TR_CLOCK_TICK);
3729 		} else {
3730 			if (nosignal)
3731 				cv_wait(&e->call_cv, &clnt_pending_lock);
3732 			else
3733 				cv_stat = cv_wait_sig(&e->call_cv,
3734 				    &clnt_pending_lock);
3735 		}
3736 		if (cv_stat == -1)
3737 			return (ETIME);
3738 		if (cv_stat == 0)
3739 			return (EINTR);
3740 		/*
3741 		 * if we received an error from the server and we know a reply
3742 		 * is not going to be sent, do not wait for the full timeout,
3743 		 * return now.
3744 		 */
3745 		if (e->call_status == RPC_XPRTFAILED)
3746 			return (e->call_reason);
3747 	}
3748 	tpr = (union T_primitives *)e->call_reply->b_rptr;
3749 	if (tpr->type == ack_prim)
3750 		return (0); /* Success */
3751 
3752 	if (tpr->type == T_ERROR_ACK) {
3753 		if (tpr->error_ack.TLI_error == TSYSERR)
3754 			return (tpr->error_ack.UNIX_error);
3755 		else
3756 			return (t_tlitosyserr(tpr->error_ack.TLI_error));
3757 	}
3758 
3759 	return (EPROTO); /* unknown or unexpected primitive */
3760 }
3761