xref: /titanic_41/usr/src/uts/common/rpc/clnt_cots.c (revision b509e89b2befbaa42939abad9da1d7f5a8c6aaae)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T
28  *		All Rights Reserved
29  */
30 
31 /*
32  * Portions of this source code were derived from Berkeley 4.3 BSD
33  * under license from the Regents of the University of California.
34  */
35 
36 
37 /*
38  * Implements a kernel based, client side RPC over Connection Oriented
39  * Transports (COTS).
40  */
41 
42 /*
43  * Much of this file has been re-written to let NFS work better over slow
44  * transports. A description follows.
45  *
46  * One of the annoying things about kRPC/COTS is that it will temporarily
47  * create more than one connection between a client and server. This
48  * happens because when a connection is made, the end-points entry in the
49  * linked list of connections (headed by cm_hd), is removed so that other
50  * threads don't mess with it. Went ahead and bit the bullet by keeping
51  * the endpoint on the connection list and introducing state bits,
52  * condition variables etc. to the connection entry data structure (struct
53  * cm_xprt).
54  *
55  * Here is a summary of the changes to cm-xprt:
56  *
57  *	x_ctime is the timestamp of when the endpoint was last
58  *	connected or disconnected. If an end-point is ever disconnected
59  *	or re-connected, then any outstanding RPC request is presumed
60  *	lost, telling clnt_cots_kcallit that it needs to re-send the
61  *	request, not just wait for the original request's reply to
62  *	arrive.
63  *
64  *	x_thread flag which tells us if a thread is doing a connection attempt.
65  *
66  *	x_waitdis flag which tells us we are waiting a disconnect ACK.
67  *
68  *	x_needdis flag which tells us we need to send a T_DISCONN_REQ
69  *	to kill the connection.
70  *
71  *	x_needrel flag which tells us we need to send a T_ORDREL_REQ to
72  *	gracefully close the connection.
73  *
74  *	#defined bitmasks for the all the b_* bits so that more
75  *	efficient (and at times less clumsy) masks can be used to
76  *	manipulated state in cases where multiple bits have to
77  *	set/cleared/checked in the same critical section.
78  *
79  *	x_conn_cv and x_dis-_cv are new condition variables to let
80  *	threads knows when the connection attempt is done, and to let
81  *	the connecting thread know when the disconnect handshake is
82  *	done.
83  *
84  * Added the CONN_HOLD() macro so that all reference holds have the same
85  * look and feel.
86  *
87  * In the private (cku_private) portion of the client handle,
88  *
89  *	cku_flags replaces the cku_sent a boolean. cku_flags keeps
90  *	track of whether a request as been sent, and whether the
91  *	client's handles call record is on the dispatch list (so that
92  *	the reply can be matched by XID to the right client handle).
93  *	The idea of CKU_ONQUEUE is that we can exit clnt_cots_kcallit()
94  *	and still have the response find the right client handle so
95  *	that the retry of CLNT_CALL() gets the result. Testing, found
96  *	situations where if the timeout was increased, performance
97  *	degraded. This was due to us hitting a window where the thread
98  *	was back in rfscall() (probably printing server not responding)
99  *	while the response came back but no place to put it.
100  *
101  *	cku_ctime is just a cache of x_ctime. If they match,
102  *	clnt_cots_kcallit() won't to send a retry (unless the maximum
103  *	receive count limit as been reached). If the don't match, then
104  *	we assume the request has been lost, and a retry of the request
105  *	is needed.
106  *
107  *	cku_recv_attempts counts the number of receive count attempts
108  *	after one try is sent on the wire.
109  *
110  * Added the clnt_delay() routine so that interruptible and
111  * noninterruptible delays are possible.
112  *
113  * CLNT_MIN_TIMEOUT has been bumped to 10 seconds from 3. This is used to
114  * control how long the client delays before returned after getting
115  * ECONNREFUSED. At 3 seconds, 8 client threads per mount really does bash
116  * a server that may be booting and not yet started nfsd.
117  *
118  * CLNT_MAXRECV_WITHOUT_RETRY is a new macro (value of 3) (with a tunable)
119  * Why don't we just wait forever (receive an infinite # of times)?
120  * Because the server may have rebooted. More insidious is that some
121  * servers (ours) will drop NFS/TCP requests in some cases. This is bad,
122  * but it is a reality.
123  *
124  * The case of a server doing orderly release really messes up the
125  * client's recovery, especially if the server's TCP implementation is
126  * buggy.  It was found was that the kRPC/COTS client was breaking some
127  * TPI rules, such as not waiting for the acknowledgement of a
128  * T_DISCON_REQ (hence the added case statements T_ERROR_ACK, T_OK_ACK and
129  * T_DISCON_REQ in clnt_dispatch_notifyall()).
130  *
131  * One of things that we've seen is that a kRPC TCP endpoint goes into
132  * TIMEWAIT and a thus a reconnect takes a long time to satisfy because
133  * that the TIMEWAIT state takes a while to finish.  If a server sends a
134  * T_ORDREL_IND, there is little point in an RPC client doing a
135  * T_ORDREL_REQ, because the RPC request isn't going to make it (the
136  * server is saying that it won't accept any more data). So kRPC was
137  * changed to send a T_DISCON_REQ when we get a T_ORDREL_IND. So now the
138  * connection skips the TIMEWAIT state and goes straight to a bound state
139  * that kRPC can quickly switch to connected.
140  *
141  * Code that issues TPI request must use waitforack() to wait for the
142  * corresponding ack (assuming there is one) in any future modifications.
143  * This works around problems that may be introduced by breaking TPI rules
144  * (by submitting new calls before earlier requests have been acked) in the
145  * case of a signal or other early return.  waitforack() depends on
146  * clnt_dispatch_notifyconn() to issue the wakeup when the ack
147  * arrives, so adding new TPI calls may require corresponding changes
148  * to clnt_dispatch_notifyconn(). Presently, the timeout period is based on
149  * CLNT_MIN_TIMEOUT which is 10 seconds. If you modify this value, be sure
150  * not to set it too low or TPI ACKS will be lost.
151  */
152 
153 #include <sys/param.h>
154 #include <sys/types.h>
155 #include <sys/user.h>
156 #include <sys/systm.h>
157 #include <sys/sysmacros.h>
158 #include <sys/proc.h>
159 #include <sys/socket.h>
160 #include <sys/file.h>
161 #include <sys/stream.h>
162 #include <sys/strsubr.h>
163 #include <sys/stropts.h>
164 #include <sys/strsun.h>
165 #include <sys/timod.h>
166 #include <sys/tiuser.h>
167 #include <sys/tihdr.h>
168 #include <sys/t_kuser.h>
169 #include <sys/fcntl.h>
170 #include <sys/errno.h>
171 #include <sys/kmem.h>
172 #include <sys/debug.h>
173 #include <sys/systm.h>
174 #include <sys/kstat.h>
175 #include <sys/t_lock.h>
176 #include <sys/ddi.h>
177 #include <sys/cmn_err.h>
178 #include <sys/time.h>
179 #include <sys/isa_defs.h>
180 #include <sys/callb.h>
181 #include <sys/sunddi.h>
182 #include <sys/atomic.h>
183 #include <sys/sdt.h>
184 
185 #include <netinet/in.h>
186 #include <netinet/tcp.h>
187 
188 #include <rpc/types.h>
189 #include <rpc/xdr.h>
190 #include <rpc/auth.h>
191 #include <rpc/clnt.h>
192 #include <rpc/rpc_msg.h>
193 
194 #define	COTS_DEFAULT_ALLOCSIZE	2048
195 
196 #define	WIRE_HDR_SIZE	20	/* serialized call header, sans proc number */
197 #define	MSG_OFFSET	128	/* offset of call into the mblk */
198 
199 const char *kinet_ntop6(uchar_t *, char *, size_t);
200 
201 static int	clnt_cots_ksettimers(CLIENT *, struct rpc_timers *,
202     struct rpc_timers *, int, void(*)(int, int, caddr_t), caddr_t, uint32_t);
203 static enum clnt_stat	clnt_cots_kcallit(CLIENT *, rpcproc_t, xdrproc_t,
204     caddr_t, xdrproc_t, caddr_t, struct timeval);
205 static void	clnt_cots_kabort(CLIENT *);
206 static void	clnt_cots_kerror(CLIENT *, struct rpc_err *);
207 static bool_t	clnt_cots_kfreeres(CLIENT *, xdrproc_t, caddr_t);
208 static void	clnt_cots_kdestroy(CLIENT *);
209 static bool_t	clnt_cots_kcontrol(CLIENT *, int, char *);
210 
211 
212 /* List of transports managed by the connection manager. */
213 struct cm_xprt {
214 	TIUSER		*x_tiptr;	/* transport handle */
215 	queue_t		*x_wq;		/* send queue */
216 	clock_t		x_time;		/* last time we handed this xprt out */
217 	clock_t		x_ctime;	/* time we went to CONNECTED */
218 	int		x_tidu_size;    /* TIDU size of this transport */
219 	union {
220 	    struct {
221 		unsigned int
222 #ifdef	_BIT_FIELDS_HTOL
223 		b_closing:	1,	/* we've sent a ord rel on this conn */
224 		b_dead:		1,	/* transport is closed or disconn */
225 		b_doomed:	1,	/* too many conns, let this go idle */
226 		b_connected:	1,	/* this connection is connected */
227 
228 		b_ordrel:	1,	/* do an orderly release? */
229 		b_thread:	1,	/* thread doing connect */
230 		b_waitdis:	1,	/* waiting for disconnect ACK */
231 		b_needdis:	1,	/* need T_DISCON_REQ */
232 
233 		b_needrel:	1,	/* need T_ORDREL_REQ */
234 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
235 					/* disconnect during connect */
236 
237 		b_pad:		22;
238 
239 #endif
240 
241 #ifdef	_BIT_FIELDS_LTOH
242 		b_pad:		22,
243 
244 		b_early_disc:	1,	/* got a T_ORDREL_IND or T_DISCON_IND */
245 					/* disconnect during connect */
246 		b_needrel:	1,	/* need T_ORDREL_REQ */
247 
248 		b_needdis:	1,	/* need T_DISCON_REQ */
249 		b_waitdis:	1,	/* waiting for disconnect ACK */
250 		b_thread:	1,	/* thread doing connect */
251 		b_ordrel:	1,	/* do an orderly release? */
252 
253 		b_connected:	1,	/* this connection is connected */
254 		b_doomed:	1,	/* too many conns, let this go idle */
255 		b_dead:		1,	/* transport is closed or disconn */
256 		b_closing:	1;	/* we've sent a ord rel on this conn */
257 #endif
258 	    } bit;	    unsigned int word;
259 
260 #define	x_closing	x_state.bit.b_closing
261 #define	x_dead		x_state.bit.b_dead
262 #define	x_doomed	x_state.bit.b_doomed
263 #define	x_connected	x_state.bit.b_connected
264 
265 #define	x_ordrel	x_state.bit.b_ordrel
266 #define	x_thread	x_state.bit.b_thread
267 #define	x_waitdis	x_state.bit.b_waitdis
268 #define	x_needdis	x_state.bit.b_needdis
269 
270 #define	x_needrel	x_state.bit.b_needrel
271 #define	x_early_disc    x_state.bit.b_early_disc
272 
273 #define	x_state_flags	x_state.word
274 
275 #define	X_CLOSING	0x80000000
276 #define	X_DEAD		0x40000000
277 #define	X_DOOMED	0x20000000
278 #define	X_CONNECTED	0x10000000
279 
280 #define	X_ORDREL	0x08000000
281 #define	X_THREAD	0x04000000
282 #define	X_WAITDIS	0x02000000
283 #define	X_NEEDDIS	0x01000000
284 
285 #define	X_NEEDREL	0x00800000
286 #define	X_EARLYDISC	0x00400000
287 
288 #define	X_BADSTATES	(X_CLOSING | X_DEAD | X_DOOMED)
289 
290 	}		x_state;
291 	int		x_ref;		/* number of users of this xprt */
292 	int		x_family;	/* address family of transport */
293 	dev_t		x_rdev;		/* device number of transport */
294 	struct cm_xprt	*x_next;
295 
296 	struct netbuf	x_server;	/* destination address */
297 	struct netbuf	x_src;		/* src address (for retries) */
298 	kmutex_t	x_lock;		/* lock on this entry */
299 	kcondvar_t	x_cv;		/* to signal when can be closed */
300 	kcondvar_t	x_conn_cv;	/* to signal when connection attempt */
301 					/* is complete */
302 	kstat_t		*x_ksp;
303 
304 	kcondvar_t	x_dis_cv;	/* to signal when disconnect attempt */
305 					/* is complete */
306 	zoneid_t	x_zoneid;	/* zone this xprt belongs to */
307 };
308 
309 typedef struct cm_kstat_xprt {
310 	kstat_named_t	x_wq;
311 	kstat_named_t	x_server;
312 	kstat_named_t	x_family;
313 	kstat_named_t	x_rdev;
314 	kstat_named_t	x_time;
315 	kstat_named_t	x_state;
316 	kstat_named_t	x_ref;
317 	kstat_named_t	x_port;
318 } cm_kstat_xprt_t;
319 
320 static cm_kstat_xprt_t cm_kstat_template = {
321 	{ "write_queue", KSTAT_DATA_UINT32 },
322 	{ "server",	KSTAT_DATA_STRING },
323 	{ "addr_family", KSTAT_DATA_UINT32 },
324 	{ "device",	KSTAT_DATA_UINT32 },
325 	{ "time_stamp",	KSTAT_DATA_UINT32 },
326 	{ "status",	KSTAT_DATA_UINT32 },
327 	{ "ref_count",	KSTAT_DATA_INT32 },
328 	{ "port",	KSTAT_DATA_UINT32 },
329 };
330 
331 /*
332  * The inverse of this is connmgr_release().
333  */
334 #define	CONN_HOLD(Cm_entry)	{\
335 	mutex_enter(&(Cm_entry)->x_lock);	\
336 	(Cm_entry)->x_ref++;	\
337 	mutex_exit(&(Cm_entry)->x_lock);	\
338 }
339 
340 
341 /*
342  * Private data per rpc handle.  This structure is allocated by
343  * clnt_cots_kcreate, and freed by clnt_cots_kdestroy.
344  */
345 typedef struct cku_private_s {
346 	CLIENT			cku_client;	/* client handle */
347 	calllist_t		cku_call;	/* for dispatching calls */
348 	struct rpc_err		cku_err;	/* error status */
349 
350 	struct netbuf		cku_srcaddr;	/* source address for retries */
351 	int			cku_addrfmly;  /* for binding port */
352 	struct netbuf		cku_addr;	/* remote address */
353 	dev_t			cku_device;	/* device to use */
354 	uint_t			cku_flags;
355 #define	CKU_ONQUEUE		0x1
356 #define	CKU_SENT		0x2
357 
358 	bool_t			cku_progress;	/* for CLSET_PROGRESS */
359 	uint32_t		cku_xid;	/* current XID */
360 	clock_t			cku_ctime;	/* time stamp of when */
361 						/* connection was created */
362 	uint_t			cku_recv_attempts;
363 	XDR			cku_outxdr;	/* xdr routine for output */
364 	XDR			cku_inxdr;	/* xdr routine for input */
365 	char			cku_rpchdr[WIRE_HDR_SIZE + 4];
366 						/* pre-serialized rpc header */
367 
368 	uint_t			cku_outbuflen;	/* default output mblk length */
369 	struct cred		*cku_cred;	/* credentials */
370 	bool_t			cku_nodelayonerr;
371 						/* for CLSET_NODELAYONERR */
372 	int			cku_useresvport; /* Use reserved port */
373 	struct rpc_cots_client	*cku_stats;	/* stats for zone */
374 } cku_private_t;
375 
376 static struct cm_xprt *connmgr_wrapconnect(struct cm_xprt *,
377 	const struct timeval *, struct netbuf *, int, struct netbuf *,
378 	struct rpc_err *, bool_t, bool_t, cred_t *);
379 
380 static bool_t	connmgr_connect(struct cm_xprt *, queue_t *, struct netbuf *,
381 				int, calllist_t *, int *, bool_t reconnect,
382 				const struct timeval *, bool_t, cred_t *);
383 
384 static bool_t	connmgr_setopt(queue_t *, int, int, calllist_t *, cred_t *cr);
385 static void	connmgr_sndrel(struct cm_xprt *);
386 static void	connmgr_snddis(struct cm_xprt *);
387 static void	connmgr_close(struct cm_xprt *);
388 static void	connmgr_release(struct cm_xprt *);
389 static struct cm_xprt *connmgr_wrapget(struct netbuf *, const struct timeval *,
390 	cku_private_t *);
391 
392 static struct cm_xprt *connmgr_get(struct netbuf *, const struct timeval *,
393 	struct netbuf *, int, struct netbuf *, struct rpc_err *, dev_t,
394 	bool_t, int, cred_t *);
395 
396 static void connmgr_cancelconn(struct cm_xprt *);
397 static enum clnt_stat connmgr_cwait(struct cm_xprt *, const struct timeval *,
398 	bool_t);
399 static void connmgr_dis_and_wait(struct cm_xprt *);
400 
401 static int	clnt_dispatch_send(queue_t *, mblk_t *, calllist_t *, uint_t,
402 					uint_t);
403 
404 static int clnt_delay(clock_t, bool_t);
405 
406 static int waitforack(calllist_t *, t_scalar_t, const struct timeval *, bool_t);
407 
408 /*
409  * Operations vector for TCP/IP based RPC
410  */
411 static struct clnt_ops tcp_ops = {
412 	clnt_cots_kcallit,	/* do rpc call */
413 	clnt_cots_kabort,	/* abort call */
414 	clnt_cots_kerror,	/* return error status */
415 	clnt_cots_kfreeres,	/* free results */
416 	clnt_cots_kdestroy,	/* destroy rpc handle */
417 	clnt_cots_kcontrol,	/* the ioctl() of rpc */
418 	clnt_cots_ksettimers,	/* set retry timers */
419 };
420 
421 static int rpc_kstat_instance = 0;  /* keeps the current instance */
422 				/* number for the next kstat_create */
423 
424 static struct cm_xprt *cm_hd = NULL;
425 static kmutex_t connmgr_lock;	/* for connection mngr's list of transports */
426 
427 extern kmutex_t clnt_max_msg_lock;
428 
429 static calllist_t *clnt_pending = NULL;
430 extern kmutex_t clnt_pending_lock;
431 
432 static int clnt_cots_hash_size = DEFAULT_HASH_SIZE;
433 
434 static call_table_t *cots_call_ht;
435 
436 static const struct rpc_cots_client {
437 	kstat_named_t	rccalls;
438 	kstat_named_t	rcbadcalls;
439 	kstat_named_t	rcbadxids;
440 	kstat_named_t	rctimeouts;
441 	kstat_named_t	rcnewcreds;
442 	kstat_named_t	rcbadverfs;
443 	kstat_named_t	rctimers;
444 	kstat_named_t	rccantconn;
445 	kstat_named_t	rcnomem;
446 	kstat_named_t	rcintrs;
447 } cots_rcstat_tmpl = {
448 	{ "calls",	KSTAT_DATA_UINT64 },
449 	{ "badcalls",	KSTAT_DATA_UINT64 },
450 	{ "badxids",	KSTAT_DATA_UINT64 },
451 	{ "timeouts",	KSTAT_DATA_UINT64 },
452 	{ "newcreds",	KSTAT_DATA_UINT64 },
453 	{ "badverfs",	KSTAT_DATA_UINT64 },
454 	{ "timers",	KSTAT_DATA_UINT64 },
455 	{ "cantconn",	KSTAT_DATA_UINT64 },
456 	{ "nomem",	KSTAT_DATA_UINT64 },
457 	{ "interrupts", KSTAT_DATA_UINT64 }
458 };
459 
460 #define	COTSRCSTAT_INCR(p, x)	\
461 	atomic_add_64(&(p)->x.value.ui64, 1)
462 
463 #define	CLNT_MAX_CONNS	1	/* concurrent connections between clnt/srvr */
464 int clnt_max_conns = CLNT_MAX_CONNS;
465 
466 #define	CLNT_MIN_TIMEOUT	10	/* seconds to wait after we get a */
467 					/* connection reset */
468 #define	CLNT_MIN_CONNTIMEOUT	5	/* seconds to wait for a connection */
469 
470 
471 int clnt_cots_min_tout = CLNT_MIN_TIMEOUT;
472 int clnt_cots_min_conntout = CLNT_MIN_CONNTIMEOUT;
473 
474 /*
475  * Limit the number of times we will attempt to receive a reply without
476  * re-sending a response.
477  */
478 #define	CLNT_MAXRECV_WITHOUT_RETRY	3
479 uint_t clnt_cots_maxrecv	= CLNT_MAXRECV_WITHOUT_RETRY;
480 
481 uint_t *clnt_max_msg_sizep;
482 void (*clnt_stop_idle)(queue_t *wq);
483 
484 #define	ptoh(p)		(&((p)->cku_client))
485 #define	htop(h)		((cku_private_t *)((h)->cl_private))
486 
487 /*
488  * Times to retry
489  */
490 #define	REFRESHES	2	/* authentication refreshes */
491 
492 /*
493  * The following is used to determine the global default behavior for
494  * COTS when binding to a local port.
495  *
496  * If the value is set to 1 the default will be to select a reserved
497  * (aka privileged) port, if the value is zero the default will be to
498  * use non-reserved ports.  Users of kRPC may override this by using
499  * CLNT_CONTROL() and CLSET_BINDRESVPORT.
500  */
501 int clnt_cots_do_bindresvport = 1;
502 
503 static zone_key_t zone_cots_key;
504 
505 /*
506  * We need to do this after all kernel threads in the zone have exited.
507  */
508 /* ARGSUSED */
509 static void
510 clnt_zone_destroy(zoneid_t zoneid, void *unused)
511 {
512 	struct cm_xprt **cmp;
513 	struct cm_xprt *cm_entry;
514 	struct cm_xprt *freelist = NULL;
515 
516 	mutex_enter(&connmgr_lock);
517 	cmp = &cm_hd;
518 	while ((cm_entry = *cmp) != NULL) {
519 		if (cm_entry->x_zoneid == zoneid) {
520 			*cmp = cm_entry->x_next;
521 			cm_entry->x_next = freelist;
522 			freelist = cm_entry;
523 		} else {
524 			cmp = &cm_entry->x_next;
525 		}
526 	}
527 	mutex_exit(&connmgr_lock);
528 	while ((cm_entry = freelist) != NULL) {
529 		freelist = cm_entry->x_next;
530 		connmgr_close(cm_entry);
531 	}
532 }
533 
534 int
535 clnt_cots_kcreate(dev_t dev, struct netbuf *addr, int family, rpcprog_t prog,
536 	rpcvers_t vers, uint_t max_msgsize, cred_t *cred, CLIENT **ncl)
537 {
538 	CLIENT *h;
539 	cku_private_t *p;
540 	struct rpc_msg call_msg;
541 	struct rpcstat *rpcstat;
542 
543 	RPCLOG(8, "clnt_cots_kcreate: prog %u\n", prog);
544 
545 	rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
546 	ASSERT(rpcstat != NULL);
547 
548 	/* Allocate and intialize the client handle. */
549 	p = kmem_zalloc(sizeof (*p), KM_SLEEP);
550 
551 	h = ptoh(p);
552 
553 	h->cl_private = (caddr_t)p;
554 	h->cl_auth = authkern_create();
555 	h->cl_ops = &tcp_ops;
556 
557 	cv_init(&p->cku_call.call_cv, NULL, CV_DEFAULT, NULL);
558 	mutex_init(&p->cku_call.call_lock, NULL, MUTEX_DEFAULT, NULL);
559 
560 	/*
561 	 * If the current sanity check size in rpcmod is smaller
562 	 * than the size needed, then increase the sanity check.
563 	 */
564 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
565 	    max_msgsize > *clnt_max_msg_sizep) {
566 		mutex_enter(&clnt_max_msg_lock);
567 		if (max_msgsize > *clnt_max_msg_sizep)
568 			*clnt_max_msg_sizep = max_msgsize;
569 		mutex_exit(&clnt_max_msg_lock);
570 	}
571 
572 	p->cku_outbuflen = COTS_DEFAULT_ALLOCSIZE;
573 
574 	/* Preserialize the call message header */
575 
576 	call_msg.rm_xid = 0;
577 	call_msg.rm_direction = CALL;
578 	call_msg.rm_call.cb_rpcvers = RPC_MSG_VERSION;
579 	call_msg.rm_call.cb_prog = prog;
580 	call_msg.rm_call.cb_vers = vers;
581 
582 	xdrmem_create(&p->cku_outxdr, p->cku_rpchdr, WIRE_HDR_SIZE, XDR_ENCODE);
583 
584 	if (!xdr_callhdr(&p->cku_outxdr, &call_msg)) {
585 		RPCLOG0(1, "clnt_cots_kcreate - Fatal header serialization "
586 		    "error\n");
587 		auth_destroy(h->cl_auth);
588 		kmem_free(p, sizeof (cku_private_t));
589 		RPCLOG0(1, "clnt_cots_kcreate: create failed error EINVAL\n");
590 		return (EINVAL);		/* XXX */
591 	}
592 
593 	/*
594 	 * The zalloc initialized the fields below.
595 	 * p->cku_xid = 0;
596 	 * p->cku_flags = 0;
597 	 * p->cku_srcaddr.len = 0;
598 	 * p->cku_srcaddr.maxlen = 0;
599 	 */
600 
601 	p->cku_cred = cred;
602 	p->cku_device = dev;
603 	p->cku_addrfmly = family;
604 	p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
605 	p->cku_addr.maxlen = addr->maxlen;
606 	p->cku_addr.len = addr->len;
607 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
608 	p->cku_stats = rpcstat->rpc_cots_client;
609 	p->cku_useresvport = -1; /* value is has not been set */
610 
611 	*ncl = h;
612 	return (0);
613 }
614 
615 /*ARGSUSED*/
616 static void
617 clnt_cots_kabort(CLIENT *h)
618 {
619 }
620 
621 /*
622  * Return error info on this handle.
623  */
624 static void
625 clnt_cots_kerror(CLIENT *h, struct rpc_err *err)
626 {
627 	/* LINTED pointer alignment */
628 	cku_private_t *p = htop(h);
629 
630 	*err = p->cku_err;
631 }
632 
633 static bool_t
634 clnt_cots_kfreeres(CLIENT *h, xdrproc_t xdr_res, caddr_t res_ptr)
635 {
636 	/* LINTED pointer alignment */
637 	cku_private_t *p = htop(h);
638 	XDR *xdrs;
639 
640 	xdrs = &(p->cku_outxdr);
641 	xdrs->x_op = XDR_FREE;
642 	return ((*xdr_res)(xdrs, res_ptr));
643 }
644 
645 static bool_t
646 clnt_cots_kcontrol(CLIENT *h, int cmd, char *arg)
647 {
648 	cku_private_t *p = htop(h);
649 
650 	switch (cmd) {
651 	case CLSET_PROGRESS:
652 		p->cku_progress = TRUE;
653 		return (TRUE);
654 
655 	case CLSET_XID:
656 		if (arg == NULL)
657 			return (FALSE);
658 
659 		p->cku_xid = *((uint32_t *)arg);
660 		return (TRUE);
661 
662 	case CLGET_XID:
663 		if (arg == NULL)
664 			return (FALSE);
665 
666 		*((uint32_t *)arg) = p->cku_xid;
667 		return (TRUE);
668 
669 	case CLSET_NODELAYONERR:
670 		if (arg == NULL)
671 			return (FALSE);
672 
673 		if (*((bool_t *)arg) == TRUE) {
674 			p->cku_nodelayonerr = TRUE;
675 			return (TRUE);
676 		}
677 		if (*((bool_t *)arg) == FALSE) {
678 			p->cku_nodelayonerr = FALSE;
679 			return (TRUE);
680 		}
681 		return (FALSE);
682 
683 	case CLGET_NODELAYONERR:
684 		if (arg == NULL)
685 			return (FALSE);
686 
687 		*((bool_t *)arg) = p->cku_nodelayonerr;
688 		return (TRUE);
689 
690 	case CLSET_BINDRESVPORT:
691 		if (arg == NULL)
692 			return (FALSE);
693 
694 		if (*(int *)arg != 1 && *(int *)arg != 0)
695 			return (FALSE);
696 
697 		p->cku_useresvport = *(int *)arg;
698 
699 		return (TRUE);
700 
701 	case CLGET_BINDRESVPORT:
702 		if (arg == NULL)
703 			return (FALSE);
704 
705 		*(int *)arg = p->cku_useresvport;
706 
707 		return (TRUE);
708 
709 	default:
710 		return (FALSE);
711 	}
712 }
713 
714 /*
715  * Destroy rpc handle.  Frees the space used for output buffer,
716  * private data, and handle structure.
717  */
718 static void
719 clnt_cots_kdestroy(CLIENT *h)
720 {
721 	/* LINTED pointer alignment */
722 	cku_private_t *p = htop(h);
723 	calllist_t *call = &p->cku_call;
724 
725 	RPCLOG(8, "clnt_cots_kdestroy h: %p\n", (void *)h);
726 	RPCLOG(8, "clnt_cots_kdestroy h: xid=0x%x\n", p->cku_xid);
727 
728 	if (p->cku_flags & CKU_ONQUEUE) {
729 		RPCLOG(64, "clnt_cots_kdestroy h: removing call for xid 0x%x "
730 		    "from dispatch list\n", p->cku_xid);
731 		call_table_remove(call);
732 	}
733 
734 	if (call->call_reply)
735 		freemsg(call->call_reply);
736 	cv_destroy(&call->call_cv);
737 	mutex_destroy(&call->call_lock);
738 
739 	kmem_free(p->cku_srcaddr.buf, p->cku_srcaddr.maxlen);
740 	kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
741 	kmem_free(p, sizeof (*p));
742 }
743 
744 static int clnt_cots_pulls;
745 #define	RM_HDR_SIZE	4	/* record mark header size */
746 
747 /*
748  * Call remote procedure.
749  */
750 static enum clnt_stat
751 clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
752     caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval wait)
753 {
754 	/* LINTED pointer alignment */
755 	cku_private_t *p = htop(h);
756 	calllist_t *call = &p->cku_call;
757 	XDR *xdrs;
758 	struct rpc_msg reply_msg;
759 	mblk_t *mp;
760 #ifdef	RPCDEBUG
761 	clock_t time_sent;
762 #endif
763 	struct netbuf *retryaddr;
764 	struct cm_xprt *cm_entry = NULL;
765 	queue_t *wq;
766 	int len, waitsecs, max_waitsecs;
767 	int mpsize;
768 	int refreshes = REFRESHES;
769 	int interrupted;
770 	int tidu_size;
771 	enum clnt_stat status;
772 	struct timeval cwait;
773 	bool_t delay_first = FALSE;
774 	clock_t ticks;
775 
776 	RPCLOG(2, "clnt_cots_kcallit, procnum %u\n", procnum);
777 	COTSRCSTAT_INCR(p->cku_stats, rccalls);
778 
779 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_sec: %ld\n", wait.tv_sec);
780 	RPCLOG(2, "clnt_cots_kcallit: wait.tv_usec: %ld\n", wait.tv_usec);
781 	/*
782 	 * Bug ID 1240234:
783 	 * Look out for zero length timeouts. We don't want to
784 	 * wait zero seconds for a connection to be established.
785 	 */
786 	if (wait.tv_sec < clnt_cots_min_conntout) {
787 		cwait.tv_sec = clnt_cots_min_conntout;
788 		cwait.tv_usec = 0;
789 		RPCLOG(8, "clnt_cots_kcallit: wait.tv_sec (%ld) too low,",
790 		    wait.tv_sec);
791 		RPCLOG(8, " setting to: %d\n", clnt_cots_min_conntout);
792 	} else {
793 		cwait = wait;
794 	}
795 
796 call_again:
797 	if (cm_entry) {
798 		connmgr_release(cm_entry);
799 		cm_entry = NULL;
800 	}
801 
802 	mp = NULL;
803 
804 	/*
805 	 * If the call is not a retry, allocate a new xid and cache it
806 	 * for future retries.
807 	 * Bug ID 1246045:
808 	 * Treat call as a retry for purposes of binding the source
809 	 * port only if we actually attempted to send anything on
810 	 * the previous call.
811 	 */
812 	if (p->cku_xid == 0) {
813 		p->cku_xid = alloc_xid();
814 		call->call_zoneid = rpc_zoneid();
815 
816 		/*
817 		 * We need to ASSERT here that our xid != 0 because this
818 		 * determines whether or not our call record gets placed on
819 		 * the hash table or the linked list.  By design, we mandate
820 		 * that RPC calls over cots must have xid's != 0, so we can
821 		 * ensure proper management of the hash table.
822 		 */
823 		ASSERT(p->cku_xid != 0);
824 
825 		retryaddr = NULL;
826 		p->cku_flags &= ~CKU_SENT;
827 
828 		if (p->cku_flags & CKU_ONQUEUE) {
829 			RPCLOG(8, "clnt_cots_kcallit: new call, dequeuing old"
830 			    " one (%p)\n", (void *)call);
831 			call_table_remove(call);
832 			p->cku_flags &= ~CKU_ONQUEUE;
833 			RPCLOG(64, "clnt_cots_kcallit: removing call from "
834 			    "dispatch list because xid was zero (now 0x%x)\n",
835 			    p->cku_xid);
836 		}
837 
838 		if (call->call_reply != NULL) {
839 			freemsg(call->call_reply);
840 			call->call_reply = NULL;
841 		}
842 	} else if (p->cku_srcaddr.buf == NULL || p->cku_srcaddr.len == 0) {
843 		retryaddr = NULL;
844 
845 	} else if (p->cku_flags & CKU_SENT) {
846 		retryaddr = &p->cku_srcaddr;
847 
848 	} else {
849 		/*
850 		 * Bug ID 1246045: Nothing was sent, so set retryaddr to
851 		 * NULL and let connmgr_get() bind to any source port it
852 		 * can get.
853 		 */
854 		retryaddr = NULL;
855 	}
856 
857 	RPCLOG(64, "clnt_cots_kcallit: xid = 0x%x", p->cku_xid);
858 	RPCLOG(64, " flags = 0x%x\n", p->cku_flags);
859 
860 	p->cku_err.re_status = RPC_TIMEDOUT;
861 	p->cku_err.re_errno = p->cku_err.re_terrno = 0;
862 
863 	cm_entry = connmgr_wrapget(retryaddr, &cwait, p);
864 
865 	if (cm_entry == NULL) {
866 		RPCLOG(1, "clnt_cots_kcallit: can't connect status %s\n",
867 		    clnt_sperrno(p->cku_err.re_status));
868 
869 		/*
870 		 * The reasons why we fail to create a connection are
871 		 * varied. In most cases we don't want the caller to
872 		 * immediately retry. This could have one or more
873 		 * bad effects. This includes flooding the net with
874 		 * connect requests to ports with no listener; a hard
875 		 * kernel loop due to all the "reserved" TCP ports being
876 		 * in use.
877 		 */
878 		delay_first = TRUE;
879 
880 		/*
881 		 * Even if we end up returning EINTR, we still count a
882 		 * a "can't connect", because the connection manager
883 		 * might have been committed to waiting for or timing out on
884 		 * a connection.
885 		 */
886 		COTSRCSTAT_INCR(p->cku_stats, rccantconn);
887 		switch (p->cku_err.re_status) {
888 		case RPC_INTR:
889 			p->cku_err.re_errno = EINTR;
890 
891 			/*
892 			 * No need to delay because a UNIX signal(2)
893 			 * interrupted us. The caller likely won't
894 			 * retry the CLNT_CALL() and even if it does,
895 			 * we assume the caller knows what it is doing.
896 			 */
897 			delay_first = FALSE;
898 			break;
899 
900 		case RPC_TIMEDOUT:
901 			p->cku_err.re_errno = ETIMEDOUT;
902 
903 			/*
904 			 * No need to delay because timed out already
905 			 * on the connection request and assume that the
906 			 * transport time out is longer than our minimum
907 			 * timeout, or least not too much smaller.
908 			 */
909 			delay_first = FALSE;
910 			break;
911 
912 		case RPC_SYSTEMERROR:
913 		case RPC_TLIERROR:
914 			/*
915 			 * We want to delay here because a transient
916 			 * system error has a better chance of going away
917 			 * if we delay a bit. If it's not transient, then
918 			 * we don't want end up in a hard kernel loop
919 			 * due to retries.
920 			 */
921 			ASSERT(p->cku_err.re_errno != 0);
922 			break;
923 
924 
925 		case RPC_CANTCONNECT:
926 			/*
927 			 * RPC_CANTCONNECT is set on T_ERROR_ACK which
928 			 * implies some error down in the TCP layer or
929 			 * below. If cku_nodelayonerror is set then we
930 			 * assume the caller knows not to try too hard.
931 			 */
932 			RPCLOG0(8, "clnt_cots_kcallit: connection failed,");
933 			RPCLOG0(8, " re_status=RPC_CANTCONNECT,");
934 			RPCLOG(8, " re_errno=%d,", p->cku_err.re_errno);
935 			RPCLOG(8, " cku_nodelayonerr=%d", p->cku_nodelayonerr);
936 			if (p->cku_nodelayonerr == TRUE)
937 				delay_first = FALSE;
938 
939 			p->cku_err.re_errno = EIO;
940 
941 			break;
942 
943 		case RPC_XPRTFAILED:
944 			/*
945 			 * We want to delay here because we likely
946 			 * got a refused connection.
947 			 */
948 			if (p->cku_err.re_errno == 0)
949 				p->cku_err.re_errno = EIO;
950 
951 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
952 			    p->cku_err.re_errno);
953 
954 			break;
955 
956 		default:
957 			/*
958 			 * We delay here because it is better to err
959 			 * on the side of caution. If we got here then
960 			 * status could have been RPC_SUCCESS, but we
961 			 * know that we did not get a connection, so
962 			 * force the rpc status to RPC_CANTCONNECT.
963 			 */
964 			p->cku_err.re_status = RPC_CANTCONNECT;
965 			p->cku_err.re_errno = EIO;
966 			break;
967 		}
968 		if (delay_first == TRUE)
969 			ticks = clnt_cots_min_tout * drv_usectohz(1000000);
970 		goto cots_done;
971 	}
972 
973 	/*
974 	 * If we've never sent any request on this connection (send count
975 	 * is zero, or the connection has been reset), cache the
976 	 * the connection's create time and send a request (possibly a retry)
977 	 */
978 	if ((p->cku_flags & CKU_SENT) == 0 ||
979 	    p->cku_ctime != cm_entry->x_ctime) {
980 		p->cku_ctime = cm_entry->x_ctime;
981 
982 	} else if ((p->cku_flags & CKU_SENT) && (p->cku_flags & CKU_ONQUEUE) &&
983 	    (call->call_reply != NULL ||
984 	    p->cku_recv_attempts < clnt_cots_maxrecv)) {
985 
986 		/*
987 		 * If we've sent a request and our call is on the dispatch
988 		 * queue and we haven't made too many receive attempts, then
989 		 * don't re-send, just receive.
990 		 */
991 		p->cku_recv_attempts++;
992 		goto read_again;
993 	}
994 
995 	/*
996 	 * Now we create the RPC request in a STREAMS message.  We have to do
997 	 * this after the call to connmgr_get so that we have the correct
998 	 * TIDU size for the transport.
999 	 */
1000 	tidu_size = cm_entry->x_tidu_size;
1001 	len = MSG_OFFSET + MAX(tidu_size, RM_HDR_SIZE + WIRE_HDR_SIZE);
1002 
1003 	while ((mp = allocb(len, BPRI_MED)) == NULL) {
1004 		if (strwaitbuf(len, BPRI_MED)) {
1005 			p->cku_err.re_status = RPC_SYSTEMERROR;
1006 			p->cku_err.re_errno = ENOSR;
1007 			COTSRCSTAT_INCR(p->cku_stats, rcnomem);
1008 			goto cots_done;
1009 		}
1010 	}
1011 	xdrs = &p->cku_outxdr;
1012 	xdrmblk_init(xdrs, mp, XDR_ENCODE, tidu_size);
1013 	mpsize = MBLKSIZE(mp);
1014 	ASSERT(mpsize >= len);
1015 	ASSERT(mp->b_rptr == mp->b_datap->db_base);
1016 
1017 	/*
1018 	 * If the size of mblk is not appreciably larger than what we
1019 	 * asked, then resize the mblk to exactly len bytes. The reason for
1020 	 * this: suppose len is 1600 bytes, the tidu is 1460 bytes
1021 	 * (from TCP over ethernet), and the arguments to the RPC require
1022 	 * 2800 bytes. Ideally we want the protocol to render two
1023 	 * ~1400 byte segments over the wire. However if allocb() gives us a 2k
1024 	 * mblk, and we allocate a second mblk for the remainder, the protocol
1025 	 * module may generate 3 segments over the wire:
1026 	 * 1460 bytes for the first, 448 (2048 - 1600) for the second, and
1027 	 * 892 for the third. If we "waste" 448 bytes in the first mblk,
1028 	 * the XDR encoding will generate two ~1400 byte mblks, and the
1029 	 * protocol module is more likely to produce properly sized segments.
1030 	 */
1031 	if ((mpsize >> 1) <= len)
1032 		mp->b_rptr += (mpsize - len);
1033 
1034 	/*
1035 	 * Adjust b_rptr to reserve space for the non-data protocol headers
1036 	 * any downstream modules might like to add, and for the
1037 	 * record marking header.
1038 	 */
1039 	mp->b_rptr += (MSG_OFFSET + RM_HDR_SIZE);
1040 
1041 	if (h->cl_auth->ah_cred.oa_flavor != RPCSEC_GSS) {
1042 		/* Copy in the preserialized RPC header information. */
1043 		bcopy(p->cku_rpchdr, mp->b_rptr, WIRE_HDR_SIZE);
1044 
1045 		/* Use XDR_SETPOS() to set the b_wptr to past the RPC header. */
1046 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base +
1047 		    WIRE_HDR_SIZE));
1048 
1049 		ASSERT((mp->b_wptr - mp->b_rptr) == WIRE_HDR_SIZE);
1050 
1051 		/* Serialize the procedure number and the arguments. */
1052 		if ((!XDR_PUTINT32(xdrs, (int32_t *)&procnum)) ||
1053 		    (!AUTH_MARSHALL(h->cl_auth, xdrs, p->cku_cred)) ||
1054 		    (!(*xdr_args)(xdrs, argsp))) {
1055 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1056 			p->cku_err.re_errno = EIO;
1057 			goto cots_done;
1058 		}
1059 
1060 		(*(uint32_t *)(mp->b_rptr)) = p->cku_xid;
1061 	} else {
1062 		uint32_t *uproc = (uint32_t *)&p->cku_rpchdr[WIRE_HDR_SIZE];
1063 		IXDR_PUT_U_INT32(uproc, procnum);
1064 
1065 		(*(uint32_t *)(&p->cku_rpchdr[0])) = p->cku_xid;
1066 
1067 		/* Use XDR_SETPOS() to set the b_wptr. */
1068 		XDR_SETPOS(xdrs, (uint_t)(mp->b_rptr - mp->b_datap->db_base));
1069 
1070 		/* Serialize the procedure number and the arguments. */
1071 		if (!AUTH_WRAP(h->cl_auth, p->cku_rpchdr, WIRE_HDR_SIZE+4,
1072 		    xdrs, xdr_args, argsp)) {
1073 			p->cku_err.re_status = RPC_CANTENCODEARGS;
1074 			p->cku_err.re_errno = EIO;
1075 			goto cots_done;
1076 		}
1077 	}
1078 
1079 	RPCLOG(2, "clnt_cots_kcallit: connected, sending call, tidu_size %d\n",
1080 	    tidu_size);
1081 
1082 	wq = cm_entry->x_wq;
1083 	waitsecs = 0;
1084 
1085 dispatch_again:
1086 	status = clnt_dispatch_send(wq, mp, call, p->cku_xid,
1087 	    (p->cku_flags & CKU_ONQUEUE));
1088 
1089 	if ((status == RPC_CANTSEND) && (call->call_reason == ENOBUFS)) {
1090 		/*
1091 		 * QFULL condition, allow some time for queue to drain
1092 		 * and try again. Give up after waiting for all timeout
1093 		 * specified for the call, or zone is going away.
1094 		 */
1095 		max_waitsecs = wait.tv_sec ? wait.tv_sec : clnt_cots_min_tout;
1096 		if ((waitsecs++ < max_waitsecs) &&
1097 		    !(zone_status_get(curproc->p_zone) >=
1098 		    ZONE_IS_SHUTTING_DOWN)) {
1099 
1100 			/* wait 1 sec for queue to drain */
1101 			if (clnt_delay(drv_usectohz(1000000),
1102 			    h->cl_nosignal) == EINTR) {
1103 				p->cku_err.re_errno = EINTR;
1104 				p->cku_err.re_status = RPC_INTR;
1105 
1106 				goto cots_done;
1107 			}
1108 
1109 			/* and try again */
1110 			goto dispatch_again;
1111 		}
1112 		p->cku_err.re_status = status;
1113 		p->cku_err.re_errno = call->call_reason;
1114 		DTRACE_PROBE(krpc__e__clntcots__kcallit__cantsend);
1115 
1116 		goto cots_done;
1117 	}
1118 
1119 	if (waitsecs) {
1120 		/* adjust timeout to account for time wait to send */
1121 		wait.tv_sec -= waitsecs;
1122 		if (wait.tv_sec < 0) {
1123 			/* pick up reply on next retry */
1124 			wait.tv_sec = 0;
1125 		}
1126 		DTRACE_PROBE2(clnt_cots__sendwait, CLIENT *, h,
1127 		    int, waitsecs);
1128 	}
1129 
1130 	RPCLOG(64, "clnt_cots_kcallit: sent call for xid 0x%x\n",
1131 	    (uint_t)p->cku_xid);
1132 	p->cku_flags = (CKU_ONQUEUE|CKU_SENT);
1133 	p->cku_recv_attempts = 1;
1134 
1135 #ifdef	RPCDEBUG
1136 	time_sent = lbolt;
1137 #endif
1138 
1139 	/*
1140 	 * Wait for a reply or a timeout.  If there is no error or timeout,
1141 	 * (both indicated by call_status), call->call_reply will contain
1142 	 * the RPC reply message.
1143 	 */
1144 read_again:
1145 	mutex_enter(&call->call_lock);
1146 	interrupted = 0;
1147 	if (call->call_status == RPC_TIMEDOUT) {
1148 		/*
1149 		 * Indicate that the lwp is not to be stopped while waiting
1150 		 * for this network traffic.  This is to avoid deadlock while
1151 		 * debugging a process via /proc and also to avoid recursive
1152 		 * mutex_enter()s due to NFS page faults while stopping
1153 		 * (NFS holds locks when it calls here).
1154 		 */
1155 		clock_t cv_wait_ret;
1156 		clock_t timout;
1157 		clock_t oldlbolt;
1158 
1159 		klwp_t *lwp = ttolwp(curthread);
1160 
1161 		if (lwp != NULL)
1162 			lwp->lwp_nostop++;
1163 
1164 		oldlbolt = lbolt;
1165 		timout = wait.tv_sec * drv_usectohz(1000000) +
1166 		    drv_usectohz(wait.tv_usec) + oldlbolt;
1167 		/*
1168 		 * Iterate until the call_status is changed to something
1169 		 * other that RPC_TIMEDOUT, or if cv_timedwait_sig() returns
1170 		 * something <=0 zero. The latter means that we timed
1171 		 * out.
1172 		 */
1173 		if (h->cl_nosignal)
1174 			while ((cv_wait_ret = cv_timedwait(&call->call_cv,
1175 			    &call->call_lock, timout)) > 0 &&
1176 			    call->call_status == RPC_TIMEDOUT)
1177 				;
1178 		else
1179 			while ((cv_wait_ret = cv_timedwait_sig(
1180 			    &call->call_cv,
1181 			    &call->call_lock, timout)) > 0 &&
1182 			    call->call_status == RPC_TIMEDOUT)
1183 				;
1184 
1185 		switch (cv_wait_ret) {
1186 		case 0:
1187 			/*
1188 			 * If we got out of the above loop with
1189 			 * cv_timedwait_sig() returning 0, then we were
1190 			 * interrupted regardless what call_status is.
1191 			 */
1192 			interrupted = 1;
1193 			break;
1194 		case -1:
1195 			/* cv_timedwait_sig() timed out */
1196 			break;
1197 		default:
1198 
1199 			/*
1200 			 * We were cv_signaled(). If we didn't
1201 			 * get a successful call_status and returned
1202 			 * before time expired, delay up to clnt_cots_min_tout
1203 			 * seconds so that the caller doesn't immediately
1204 			 * try to call us again and thus force the
1205 			 * same condition that got us here (such
1206 			 * as a RPC_XPRTFAILED due to the server not
1207 			 * listening on the end-point.
1208 			 */
1209 			if (call->call_status != RPC_SUCCESS) {
1210 				clock_t curlbolt;
1211 				clock_t diff;
1212 
1213 				curlbolt = ddi_get_lbolt();
1214 				ticks = clnt_cots_min_tout *
1215 				    drv_usectohz(1000000);
1216 				diff = curlbolt - oldlbolt;
1217 				if (diff < ticks) {
1218 					delay_first = TRUE;
1219 					if (diff > 0)
1220 						ticks -= diff;
1221 				}
1222 			}
1223 			break;
1224 		}
1225 
1226 		if (lwp != NULL)
1227 			lwp->lwp_nostop--;
1228 	}
1229 	/*
1230 	 * Get the reply message, if any.  This will be freed at the end
1231 	 * whether or not an error occurred.
1232 	 */
1233 	mp = call->call_reply;
1234 	call->call_reply = NULL;
1235 
1236 	/*
1237 	 * call_err is the error info when the call is on dispatch queue.
1238 	 * cku_err is the error info returned to the caller.
1239 	 * Sync cku_err with call_err for local message processing.
1240 	 */
1241 
1242 	status = call->call_status;
1243 	p->cku_err = call->call_err;
1244 	mutex_exit(&call->call_lock);
1245 
1246 	if (status != RPC_SUCCESS) {
1247 		switch (status) {
1248 		case RPC_TIMEDOUT:
1249 			if (interrupted) {
1250 				COTSRCSTAT_INCR(p->cku_stats, rcintrs);
1251 				p->cku_err.re_status = RPC_INTR;
1252 				p->cku_err.re_errno = EINTR;
1253 				RPCLOG(1, "clnt_cots_kcallit: xid 0x%x",
1254 				    p->cku_xid);
1255 				RPCLOG(1, "signal interrupted at %ld", lbolt);
1256 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1257 			} else {
1258 				COTSRCSTAT_INCR(p->cku_stats, rctimeouts);
1259 				p->cku_err.re_errno = ETIMEDOUT;
1260 				RPCLOG(1, "clnt_cots_kcallit: timed out at %ld",
1261 				    lbolt);
1262 				RPCLOG(1, ", was sent at %ld\n", time_sent);
1263 			}
1264 			break;
1265 
1266 		case RPC_XPRTFAILED:
1267 			if (p->cku_err.re_errno == 0)
1268 				p->cku_err.re_errno = EIO;
1269 
1270 			RPCLOG(1, "clnt_cots_kcallit: transport failed: %d\n",
1271 			    p->cku_err.re_errno);
1272 			break;
1273 
1274 		case RPC_SYSTEMERROR:
1275 			ASSERT(p->cku_err.re_errno);
1276 			RPCLOG(1, "clnt_cots_kcallit: system error: %d\n",
1277 			    p->cku_err.re_errno);
1278 			break;
1279 
1280 		default:
1281 			p->cku_err.re_status = RPC_SYSTEMERROR;
1282 			p->cku_err.re_errno = EIO;
1283 			RPCLOG(1, "clnt_cots_kcallit: error: %s\n",
1284 			    clnt_sperrno(status));
1285 			break;
1286 		}
1287 		if (p->cku_err.re_status != RPC_TIMEDOUT) {
1288 
1289 			if (p->cku_flags & CKU_ONQUEUE) {
1290 				call_table_remove(call);
1291 				p->cku_flags &= ~CKU_ONQUEUE;
1292 			}
1293 
1294 			RPCLOG(64, "clnt_cots_kcallit: non TIMEOUT so xid 0x%x "
1295 			    "taken off dispatch list\n", p->cku_xid);
1296 			if (call->call_reply) {
1297 				freemsg(call->call_reply);
1298 				call->call_reply = NULL;
1299 			}
1300 		} else if (wait.tv_sec != 0) {
1301 			/*
1302 			 * We've sent the request over TCP and so we have
1303 			 * every reason to believe it will get
1304 			 * delivered. In which case returning a timeout is not
1305 			 * appropriate.
1306 			 */
1307 			if (p->cku_progress == TRUE &&
1308 			    p->cku_recv_attempts < clnt_cots_maxrecv) {
1309 				p->cku_err.re_status = RPC_INPROGRESS;
1310 			}
1311 		}
1312 		goto cots_done;
1313 	}
1314 
1315 	xdrs = &p->cku_inxdr;
1316 	xdrmblk_init(xdrs, mp, XDR_DECODE, 0);
1317 
1318 	reply_msg.rm_direction = REPLY;
1319 	reply_msg.rm_reply.rp_stat = MSG_ACCEPTED;
1320 	reply_msg.acpted_rply.ar_stat = SUCCESS;
1321 
1322 	reply_msg.acpted_rply.ar_verf = _null_auth;
1323 	/*
1324 	 *  xdr_results will be done in AUTH_UNWRAP.
1325 	 */
1326 	reply_msg.acpted_rply.ar_results.where = NULL;
1327 	reply_msg.acpted_rply.ar_results.proc = xdr_void;
1328 
1329 	if (xdr_replymsg(xdrs, &reply_msg)) {
1330 		enum clnt_stat re_status;
1331 
1332 		_seterr_reply(&reply_msg, &p->cku_err);
1333 
1334 		re_status = p->cku_err.re_status;
1335 		if (re_status == RPC_SUCCESS) {
1336 			/*
1337 			 * Reply is good, check auth.
1338 			 */
1339 			if (!AUTH_VALIDATE(h->cl_auth,
1340 			    &reply_msg.acpted_rply.ar_verf)) {
1341 				COTSRCSTAT_INCR(p->cku_stats, rcbadverfs);
1342 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1343 				    "failure\n");
1344 				freemsg(mp);
1345 				(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1346 				mutex_enter(&call->call_lock);
1347 				if (call->call_reply == NULL)
1348 					call->call_status = RPC_TIMEDOUT;
1349 				mutex_exit(&call->call_lock);
1350 				goto read_again;
1351 			} else if (!AUTH_UNWRAP(h->cl_auth, xdrs,
1352 			    xdr_results, resultsp)) {
1353 				RPCLOG0(1, "clnt_cots_kcallit: validation "
1354 				    "failure (unwrap)\n");
1355 				p->cku_err.re_status = RPC_CANTDECODERES;
1356 				p->cku_err.re_errno = EIO;
1357 			}
1358 		} else {
1359 			/* set errno in case we can't recover */
1360 			if (re_status != RPC_VERSMISMATCH &&
1361 			    re_status != RPC_AUTHERROR &&
1362 			    re_status != RPC_PROGVERSMISMATCH)
1363 				p->cku_err.re_errno = EIO;
1364 
1365 			if (re_status == RPC_AUTHERROR) {
1366 				/*
1367 				 * Maybe our credential need to be refreshed
1368 				 */
1369 				if (cm_entry) {
1370 					/*
1371 					 * There is the potential that the
1372 					 * cm_entry has/will be marked dead,
1373 					 * so drop the connection altogether,
1374 					 * force REFRESH to establish new
1375 					 * connection.
1376 					 */
1377 					connmgr_cancelconn(cm_entry);
1378 					cm_entry = NULL;
1379 				}
1380 
1381 				if ((refreshes > 0) &&
1382 				    AUTH_REFRESH(h->cl_auth, &reply_msg,
1383 				    p->cku_cred)) {
1384 					refreshes--;
1385 					(void) xdr_rpc_free_verifier(xdrs,
1386 					    &reply_msg);
1387 					freemsg(mp);
1388 					mp = NULL;
1389 
1390 					if (p->cku_flags & CKU_ONQUEUE) {
1391 						call_table_remove(call);
1392 						p->cku_flags &= ~CKU_ONQUEUE;
1393 					}
1394 
1395 					RPCLOG(64,
1396 					    "clnt_cots_kcallit: AUTH_ERROR, xid"
1397 					    " 0x%x removed off dispatch list\n",
1398 					    p->cku_xid);
1399 					if (call->call_reply) {
1400 						freemsg(call->call_reply);
1401 						call->call_reply = NULL;
1402 					}
1403 
1404 					COTSRCSTAT_INCR(p->cku_stats,
1405 					    rcbadcalls);
1406 					COTSRCSTAT_INCR(p->cku_stats,
1407 					    rcnewcreds);
1408 					goto call_again;
1409 				}
1410 
1411 				/*
1412 				 * We have used the client handle to
1413 				 * do an AUTH_REFRESH and the RPC status may
1414 				 * be set to RPC_SUCCESS; Let's make sure to
1415 				 * set it to RPC_AUTHERROR.
1416 				 */
1417 				p->cku_err.re_status = RPC_AUTHERROR;
1418 
1419 				/*
1420 				 * Map recoverable and unrecoverable
1421 				 * authentication errors to appropriate errno
1422 				 */
1423 				switch (p->cku_err.re_why) {
1424 				case AUTH_TOOWEAK:
1425 					/*
1426 					 * This could be a failure where the
1427 					 * server requires use of a reserved
1428 					 * port,  check and optionally set the
1429 					 * client handle useresvport trying
1430 					 * one more time. Next go round we
1431 					 * fall out with the tooweak error.
1432 					 */
1433 					if (p->cku_useresvport != 1) {
1434 						p->cku_useresvport = 1;
1435 						p->cku_xid = 0;
1436 						(void) xdr_rpc_free_verifier
1437 						    (xdrs, &reply_msg);
1438 						freemsg(mp);
1439 						goto call_again;
1440 					}
1441 					/* FALLTHRU */
1442 				case AUTH_BADCRED:
1443 				case AUTH_BADVERF:
1444 				case AUTH_INVALIDRESP:
1445 				case AUTH_FAILED:
1446 				case RPCSEC_GSS_NOCRED:
1447 				case RPCSEC_GSS_FAILED:
1448 						p->cku_err.re_errno = EACCES;
1449 						break;
1450 				case AUTH_REJECTEDCRED:
1451 				case AUTH_REJECTEDVERF:
1452 				default:	p->cku_err.re_errno = EIO;
1453 						break;
1454 				}
1455 				RPCLOG(1, "clnt_cots_kcallit : authentication"
1456 				    " failed with RPC_AUTHERROR of type %d\n",
1457 				    (int)p->cku_err.re_why);
1458 			}
1459 		}
1460 	} else {
1461 		/* reply didn't decode properly. */
1462 		p->cku_err.re_status = RPC_CANTDECODERES;
1463 		p->cku_err.re_errno = EIO;
1464 		RPCLOG0(1, "clnt_cots_kcallit: decode failure\n");
1465 	}
1466 
1467 	(void) xdr_rpc_free_verifier(xdrs, &reply_msg);
1468 
1469 	if (p->cku_flags & CKU_ONQUEUE) {
1470 		call_table_remove(call);
1471 		p->cku_flags &= ~CKU_ONQUEUE;
1472 	}
1473 
1474 	RPCLOG(64, "clnt_cots_kcallit: xid 0x%x taken off dispatch list",
1475 	    p->cku_xid);
1476 	RPCLOG(64, " status is %s\n", clnt_sperrno(p->cku_err.re_status));
1477 cots_done:
1478 	if (cm_entry)
1479 		connmgr_release(cm_entry);
1480 
1481 	if (mp != NULL)
1482 		freemsg(mp);
1483 	if ((p->cku_flags & CKU_ONQUEUE) == 0 && call->call_reply) {
1484 		freemsg(call->call_reply);
1485 		call->call_reply = NULL;
1486 	}
1487 	if (p->cku_err.re_status != RPC_SUCCESS) {
1488 		RPCLOG0(1, "clnt_cots_kcallit: tail-end failure\n");
1489 		COTSRCSTAT_INCR(p->cku_stats, rcbadcalls);
1490 	}
1491 
1492 	/*
1493 	 * No point in delaying if the zone is going away.
1494 	 */
1495 	if (delay_first == TRUE &&
1496 	    !(zone_status_get(curproc->p_zone) >= ZONE_IS_SHUTTING_DOWN)) {
1497 		if (clnt_delay(ticks, h->cl_nosignal) == EINTR) {
1498 			p->cku_err.re_errno = EINTR;
1499 			p->cku_err.re_status = RPC_INTR;
1500 		}
1501 	}
1502 	return (p->cku_err.re_status);
1503 }
1504 
1505 /*
1506  * Kinit routine for cots.  This sets up the correct operations in
1507  * the client handle, as the handle may have previously been a clts
1508  * handle, and clears the xid field so there is no way a new call
1509  * could be mistaken for a retry.  It also sets in the handle the
1510  * information that is passed at create/kinit time but needed at
1511  * call time, as cots creates the transport at call time - device,
1512  * address of the server, protocol family.
1513  */
1514 void
1515 clnt_cots_kinit(CLIENT *h, dev_t dev, int family, struct netbuf *addr,
1516 	int max_msgsize, cred_t *cred)
1517 {
1518 	/* LINTED pointer alignment */
1519 	cku_private_t *p = htop(h);
1520 	calllist_t *call = &p->cku_call;
1521 
1522 	h->cl_ops = &tcp_ops;
1523 	if (p->cku_flags & CKU_ONQUEUE) {
1524 		call_table_remove(call);
1525 		p->cku_flags &= ~CKU_ONQUEUE;
1526 		RPCLOG(64, "clnt_cots_kinit: removing call for xid 0x%x from"
1527 		    " dispatch list\n", p->cku_xid);
1528 	}
1529 
1530 	if (call->call_reply != NULL) {
1531 		freemsg(call->call_reply);
1532 		call->call_reply = NULL;
1533 	}
1534 
1535 	call->call_bucket = NULL;
1536 	call->call_hash = 0;
1537 
1538 	/*
1539 	 * We don't clear cku_flags here, because clnt_cots_kcallit()
1540 	 * takes care of handling the cku_flags reset.
1541 	 */
1542 	p->cku_xid = 0;
1543 	p->cku_device = dev;
1544 	p->cku_addrfmly = family;
1545 	p->cku_cred = cred;
1546 
1547 	if (p->cku_addr.maxlen < addr->len) {
1548 		if (p->cku_addr.maxlen != 0 && p->cku_addr.buf != NULL)
1549 			kmem_free(p->cku_addr.buf, p->cku_addr.maxlen);
1550 		p->cku_addr.buf = kmem_zalloc(addr->maxlen, KM_SLEEP);
1551 		p->cku_addr.maxlen = addr->maxlen;
1552 	}
1553 
1554 	p->cku_addr.len = addr->len;
1555 	bcopy(addr->buf, p->cku_addr.buf, addr->len);
1556 
1557 	/*
1558 	 * If the current sanity check size in rpcmod is smaller
1559 	 * than the size needed, then increase the sanity check.
1560 	 */
1561 	if (max_msgsize != 0 && clnt_max_msg_sizep != NULL &&
1562 	    max_msgsize > *clnt_max_msg_sizep) {
1563 		mutex_enter(&clnt_max_msg_lock);
1564 		if (max_msgsize > *clnt_max_msg_sizep)
1565 			*clnt_max_msg_sizep = max_msgsize;
1566 		mutex_exit(&clnt_max_msg_lock);
1567 	}
1568 }
1569 
1570 /*
1571  * ksettimers is a no-op for cots, with the exception of setting the xid.
1572  */
1573 /* ARGSUSED */
1574 static int
1575 clnt_cots_ksettimers(CLIENT *h, struct rpc_timers *t, struct rpc_timers *all,
1576 	int minimum, void (*feedback)(int, int, caddr_t), caddr_t arg,
1577 	uint32_t xid)
1578 {
1579 	/* LINTED pointer alignment */
1580 	cku_private_t *p = htop(h);
1581 
1582 	if (xid)
1583 		p->cku_xid = xid;
1584 	COTSRCSTAT_INCR(p->cku_stats, rctimers);
1585 	return (0);
1586 }
1587 
1588 extern void rpc_poptimod(struct vnode *);
1589 extern int kstr_push(struct vnode *, char *);
1590 
1591 int
1592 conn_kstat_update(kstat_t *ksp, int rw)
1593 {
1594 	struct cm_xprt *cm_entry;
1595 	struct cm_kstat_xprt *cm_ksp_data;
1596 	uchar_t *b;
1597 	char *fbuf;
1598 
1599 	if (rw == KSTAT_WRITE)
1600 		return (EACCES);
1601 	if (ksp == NULL || ksp->ks_private == NULL)
1602 		return (EIO);
1603 	cm_entry  = (struct cm_xprt *)ksp->ks_private;
1604 	cm_ksp_data = (struct cm_kstat_xprt *)ksp->ks_data;
1605 
1606 	cm_ksp_data->x_wq.value.ui32 = (uint32_t)(uintptr_t)cm_entry->x_wq;
1607 	cm_ksp_data->x_family.value.ui32 = cm_entry->x_family;
1608 	cm_ksp_data->x_rdev.value.ui32 = (uint32_t)cm_entry->x_rdev;
1609 	cm_ksp_data->x_time.value.ui32 = cm_entry->x_time;
1610 	cm_ksp_data->x_ref.value.ui32 = cm_entry->x_ref;
1611 	cm_ksp_data->x_state.value.ui32 = cm_entry->x_state_flags;
1612 
1613 	if (cm_entry->x_server.buf) {
1614 		fbuf = cm_ksp_data->x_server.value.str.addr.ptr;
1615 		if (cm_entry->x_family == AF_INET &&
1616 		    cm_entry->x_server.len ==
1617 		    sizeof (struct sockaddr_in)) {
1618 			struct sockaddr_in  *sa;
1619 			sa = (struct sockaddr_in *)
1620 				cm_entry->x_server.buf;
1621 			b = (uchar_t *)&sa->sin_addr;
1622 			(void) sprintf(fbuf,
1623 			    "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF,
1624 			    b[2] & 0xFF, b[3] & 0xFF);
1625 			cm_ksp_data->x_port.value.ui32 =
1626 				(uint32_t)sa->sin_port;
1627 		} else if (cm_entry->x_family == AF_INET6 &&
1628 				cm_entry->x_server.len >=
1629 				sizeof (struct sockaddr_in6)) {
1630 			/* extract server IP address & port */
1631 			struct sockaddr_in6 *sin6;
1632 			sin6 = (struct sockaddr_in6 *)cm_entry->x_server.buf;
1633 			(void) kinet_ntop6((uchar_t *)&sin6->sin6_addr, fbuf,
1634 				INET6_ADDRSTRLEN);
1635 			cm_ksp_data->x_port.value.ui32 = sin6->sin6_port;
1636 		} else {
1637 			struct sockaddr_in  *sa;
1638 
1639 			sa = (struct sockaddr_in *)cm_entry->x_server.buf;
1640 			b = (uchar_t *)&sa->sin_addr;
1641 			(void) sprintf(fbuf,
1642 			    "%03d.%03d.%03d.%03d", b[0] & 0xFF, b[1] & 0xFF,
1643 			    b[2] & 0xFF, b[3] & 0xFF);
1644 		}
1645 		KSTAT_NAMED_STR_BUFLEN(&cm_ksp_data->x_server) =
1646 		    strlen(fbuf) + 1;
1647 	}
1648 
1649 	return (0);
1650 }
1651 
1652 
1653 /*
1654  * We want a version of delay which is interruptible by a UNIX signal
1655  * Return EINTR if an interrupt occured.
1656  */
1657 static int
1658 clnt_delay(clock_t ticks, bool_t nosignal)
1659 {
1660 	if (nosignal == TRUE) {
1661 		delay(ticks);
1662 		return (0);
1663 	}
1664 	return (delay_sig(ticks));
1665 }
1666 
1667 /*
1668  * Wait for a connection until a timeout, or until we are
1669  * signalled that there has been a connection state change.
1670  */
1671 static enum clnt_stat
1672 connmgr_cwait(struct cm_xprt *cm_entry, const struct timeval *waitp,
1673 	bool_t nosignal)
1674 {
1675 	bool_t interrupted;
1676 	clock_t timout, cv_stat;
1677 	enum clnt_stat clstat;
1678 	unsigned int old_state;
1679 
1680 	ASSERT(MUTEX_HELD(&connmgr_lock));
1681 	/*
1682 	 * We wait for the transport connection to be made, or an
1683 	 * indication that it could not be made.
1684 	 */
1685 	clstat = RPC_TIMEDOUT;
1686 	interrupted = FALSE;
1687 
1688 	old_state = cm_entry->x_state_flags;
1689 	/*
1690 	 * Now loop until cv_timedwait{_sig} returns because of
1691 	 * a signal(0) or timeout(-1) or cv_signal(>0). But it may be
1692 	 * cv_signalled for various other reasons too. So loop
1693 	 * until there is a state change on the connection.
1694 	 */
1695 
1696 	timout = waitp->tv_sec * drv_usectohz(1000000) +
1697 	    drv_usectohz(waitp->tv_usec) + lbolt;
1698 
1699 	if (nosignal) {
1700 		while ((cv_stat = cv_timedwait(&cm_entry->x_conn_cv,
1701 		    &connmgr_lock, timout)) > 0 &&
1702 		    cm_entry->x_state_flags == old_state)
1703 			;
1704 	} else {
1705 		while ((cv_stat = cv_timedwait_sig(&cm_entry->x_conn_cv,
1706 		    &connmgr_lock, timout)) > 0 &&
1707 		    cm_entry->x_state_flags == old_state)
1708 			;
1709 
1710 		if (cv_stat == 0) /* got intr signal? */
1711 			interrupted = TRUE;
1712 	}
1713 
1714 	if ((cm_entry->x_state_flags & (X_BADSTATES|X_CONNECTED)) ==
1715 	    X_CONNECTED) {
1716 		clstat = RPC_SUCCESS;
1717 	} else {
1718 		if (interrupted == TRUE)
1719 			clstat = RPC_INTR;
1720 		RPCLOG(1, "connmgr_cwait: can't connect, error: %s\n",
1721 		    clnt_sperrno(clstat));
1722 	}
1723 
1724 	return (clstat);
1725 }
1726 
1727 /*
1728  * Primary interface for how RPC grabs a connection.
1729  */
1730 static struct cm_xprt *
1731 connmgr_wrapget(
1732 	struct netbuf *retryaddr,
1733 	const struct timeval *waitp,
1734 	cku_private_t *p)
1735 {
1736 	struct cm_xprt *cm_entry;
1737 
1738 	cm_entry = connmgr_get(retryaddr, waitp, &p->cku_addr, p->cku_addrfmly,
1739 	    &p->cku_srcaddr, &p->cku_err, p->cku_device,
1740 	    p->cku_client.cl_nosignal, p->cku_useresvport, p->cku_cred);
1741 
1742 	if (cm_entry == NULL) {
1743 		/*
1744 		 * Re-map the call status to RPC_INTR if the err code is
1745 		 * EINTR. This can happen if calls status is RPC_TLIERROR.
1746 		 * However, don't re-map if signalling has been turned off.
1747 		 * XXX Really need to create a separate thread whenever
1748 		 * there isn't an existing connection.
1749 		 */
1750 		if (p->cku_err.re_errno == EINTR) {
1751 			if (p->cku_client.cl_nosignal == TRUE)
1752 				p->cku_err.re_errno = EIO;
1753 			else
1754 				p->cku_err.re_status = RPC_INTR;
1755 		}
1756 	}
1757 
1758 	return (cm_entry);
1759 }
1760 
1761 /*
1762  * Obtains a transport to the server specified in addr.  If a suitable transport
1763  * does not already exist in the list of cached transports, a new connection
1764  * is created, connected, and added to the list. The connection is for sending
1765  * only - the reply message may come back on another transport connection.
1766  *
1767  * To implement round-robin load balancing with multiple client connections,
1768  * the last entry on the list is always selected. Once the entry is selected
1769  * it's re-inserted to the head of the list.
1770  */
1771 static struct cm_xprt *
1772 connmgr_get(
1773 	struct netbuf	*retryaddr,
1774 	const struct timeval	*waitp,	/* changed to a ptr to converse stack */
1775 	struct netbuf	*destaddr,
1776 	int		addrfmly,
1777 	struct netbuf	*srcaddr,
1778 	struct rpc_err	*rpcerr,
1779 	dev_t		device,
1780 	bool_t		nosignal,
1781 	int		useresvport,
1782 	cred_t		*cr)
1783 {
1784 	struct cm_xprt *cm_entry;
1785 	struct cm_xprt *lru_entry;
1786 	struct cm_xprt **cmp, **prev;
1787 	queue_t *wq;
1788 	TIUSER *tiptr;
1789 	int i;
1790 	int retval;
1791 	int tidu_size;
1792 	bool_t	connected;
1793 	zoneid_t zoneid = rpc_zoneid();
1794 
1795 	/*
1796 	 * If the call is not a retry, look for a transport entry that
1797 	 * goes to the server of interest.
1798 	 */
1799 	mutex_enter(&connmgr_lock);
1800 
1801 	if (retryaddr == NULL) {
1802 use_new_conn:
1803 		i = 0;
1804 		cm_entry = lru_entry = NULL;
1805 
1806 		prev = cmp = &cm_hd;
1807 		while ((cm_entry = *cmp) != NULL) {
1808 			ASSERT(cm_entry != cm_entry->x_next);
1809 			/*
1810 			 * Garbage collect conections that are marked
1811 			 * for needs disconnect.
1812 			 */
1813 			if (cm_entry->x_needdis) {
1814 				CONN_HOLD(cm_entry);
1815 				connmgr_dis_and_wait(cm_entry);
1816 				connmgr_release(cm_entry);
1817 				/*
1818 				 * connmgr_lock could have been
1819 				 * dropped for the disconnect
1820 				 * processing so start over.
1821 				 */
1822 				goto use_new_conn;
1823 			}
1824 
1825 			/*
1826 			 * Garbage collect the dead connections that have
1827 			 * no threads working on them.
1828 			 */
1829 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1830 			    X_DEAD) {
1831 				mutex_enter(&cm_entry->x_lock);
1832 				if (cm_entry->x_ref != 0) {
1833 					/*
1834 					 * Currently in use.
1835 					 * Cleanup later.
1836 					 */
1837 					cmp = &cm_entry->x_next;
1838 					mutex_exit(&cm_entry->x_lock);
1839 					continue;
1840 				}
1841 				mutex_exit(&cm_entry->x_lock);
1842 				*cmp = cm_entry->x_next;
1843 				mutex_exit(&connmgr_lock);
1844 				connmgr_close(cm_entry);
1845 				mutex_enter(&connmgr_lock);
1846 				goto use_new_conn;
1847 			}
1848 
1849 
1850 			if ((cm_entry->x_state_flags & X_BADSTATES) == 0 &&
1851 			    cm_entry->x_zoneid == zoneid &&
1852 			    cm_entry->x_rdev == device &&
1853 			    destaddr->len == cm_entry->x_server.len &&
1854 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
1855 			    destaddr->len) == 0) {
1856 				/*
1857 				 * If the matching entry isn't connected,
1858 				 * attempt to reconnect it.
1859 				 */
1860 				if (cm_entry->x_connected == FALSE) {
1861 					/*
1862 					 * We don't go through trying
1863 					 * to find the least recently
1864 					 * used connected because
1865 					 * connmgr_reconnect() briefly
1866 					 * dropped the connmgr_lock,
1867 					 * allowing a window for our
1868 					 * accounting to be messed up.
1869 					 * In any case, a re-connected
1870 					 * connection is as good as
1871 					 * a LRU connection.
1872 					 */
1873 					return (connmgr_wrapconnect(cm_entry,
1874 					    waitp, destaddr, addrfmly, srcaddr,
1875 					    rpcerr, TRUE, nosignal, cr));
1876 				}
1877 				i++;
1878 
1879 				/* keep track of the last entry */
1880 				lru_entry = cm_entry;
1881 				prev = cmp;
1882 			}
1883 			cmp = &cm_entry->x_next;
1884 		}
1885 
1886 		if (i > clnt_max_conns) {
1887 			RPCLOG(8, "connmgr_get: too many conns, dooming entry"
1888 			    " %p\n", (void *)lru_entry->x_tiptr);
1889 			lru_entry->x_doomed = TRUE;
1890 			goto use_new_conn;
1891 		}
1892 
1893 		/*
1894 		 * If we are at the maximum number of connections to
1895 		 * the server, hand back the least recently used one.
1896 		 */
1897 		if (i == clnt_max_conns) {
1898 			/*
1899 			 * Copy into the handle the source address of
1900 			 * the connection, which we will use in case of
1901 			 * a later retry.
1902 			 */
1903 			if (srcaddr->len != lru_entry->x_src.len) {
1904 				if (srcaddr->len > 0)
1905 					kmem_free(srcaddr->buf,
1906 					    srcaddr->maxlen);
1907 				srcaddr->buf = kmem_zalloc(
1908 				    lru_entry->x_src.len, KM_SLEEP);
1909 				srcaddr->maxlen = srcaddr->len =
1910 				    lru_entry->x_src.len;
1911 			}
1912 			bcopy(lru_entry->x_src.buf, srcaddr->buf, srcaddr->len);
1913 			RPCLOG(2, "connmgr_get: call going out on %p\n",
1914 			    (void *)lru_entry);
1915 			lru_entry->x_time = lbolt;
1916 			CONN_HOLD(lru_entry);
1917 
1918 			if ((i > 1) && (prev != &cm_hd)) {
1919 				/*
1920 				 * remove and re-insert entry at head of list.
1921 				 */
1922 				*prev = lru_entry->x_next;
1923 				lru_entry->x_next = cm_hd;
1924 				cm_hd = lru_entry;
1925 			}
1926 
1927 			mutex_exit(&connmgr_lock);
1928 			return (lru_entry);
1929 		}
1930 
1931 	} else {
1932 		/*
1933 		 * This is the retry case (retryaddr != NULL).  Retries must
1934 		 * be sent on the same source port as the original call.
1935 		 */
1936 
1937 		/*
1938 		 * Walk the list looking for a connection with a source address
1939 		 * that matches the retry address.
1940 		 */
1941 start_retry_loop:
1942 		cmp = &cm_hd;
1943 		while ((cm_entry = *cmp) != NULL) {
1944 			ASSERT(cm_entry != cm_entry->x_next);
1945 
1946 			/*
1947 			 * determine if this connection matches the passed
1948 			 * in retry address.  If it does not match, advance
1949 			 * to the next element on the list.
1950 			 */
1951 			if (zoneid != cm_entry->x_zoneid ||
1952 			    device != cm_entry->x_rdev ||
1953 			    retryaddr->len != cm_entry->x_src.len ||
1954 			    bcmp(retryaddr->buf, cm_entry->x_src.buf,
1955 			    retryaddr->len) != 0) {
1956 				cmp = &cm_entry->x_next;
1957 				continue;
1958 			}
1959 			/*
1960 			 * Garbage collect conections that are marked
1961 			 * for needs disconnect.
1962 			 */
1963 			if (cm_entry->x_needdis) {
1964 				CONN_HOLD(cm_entry);
1965 				connmgr_dis_and_wait(cm_entry);
1966 				connmgr_release(cm_entry);
1967 				/*
1968 				 * connmgr_lock could have been
1969 				 * dropped for the disconnect
1970 				 * processing so start over.
1971 				 */
1972 				goto start_retry_loop;
1973 			}
1974 			/*
1975 			 * Garbage collect the dead connections that have
1976 			 * no threads working on them.
1977 			 */
1978 			if ((cm_entry->x_state_flags & (X_DEAD|X_THREAD)) ==
1979 			    X_DEAD) {
1980 				mutex_enter(&cm_entry->x_lock);
1981 				if (cm_entry->x_ref != 0) {
1982 					/*
1983 					 * Currently in use.
1984 					 * Cleanup later.
1985 					 */
1986 					cmp = &cm_entry->x_next;
1987 					mutex_exit(&cm_entry->x_lock);
1988 					continue;
1989 				}
1990 				mutex_exit(&cm_entry->x_lock);
1991 				*cmp = cm_entry->x_next;
1992 				mutex_exit(&connmgr_lock);
1993 				connmgr_close(cm_entry);
1994 				mutex_enter(&connmgr_lock);
1995 				goto start_retry_loop;
1996 			}
1997 
1998 			/*
1999 			 * Sanity check: if the connection with our source
2000 			 * port is going to some other server, something went
2001 			 * wrong, as we never delete connections (i.e. release
2002 			 * ports) unless they have been idle.  In this case,
2003 			 * it is probably better to send the call out using
2004 			 * a new source address than to fail it altogether,
2005 			 * since that port may never be released.
2006 			 */
2007 			if (destaddr->len != cm_entry->x_server.len ||
2008 			    bcmp(destaddr->buf, cm_entry->x_server.buf,
2009 			    destaddr->len) != 0) {
2010 				RPCLOG(1, "connmgr_get: tiptr %p"
2011 				    " is going to a different server"
2012 				    " with the port that belongs"
2013 				    " to us!\n", (void *)cm_entry->x_tiptr);
2014 				retryaddr = NULL;
2015 				goto use_new_conn;
2016 			}
2017 
2018 			/*
2019 			 * If the connection of interest is not connected and we
2020 			 * can't reconnect it, then the server is probably
2021 			 * still down.  Return NULL to the caller and let it
2022 			 * retry later if it wants to.  We have a delay so the
2023 			 * machine doesn't go into a tight retry loop.  If the
2024 			 * entry was already connected, or the reconnected was
2025 			 * successful, return this entry.
2026 			 */
2027 			if (cm_entry->x_connected == FALSE) {
2028 				return (connmgr_wrapconnect(cm_entry,
2029 				    waitp, destaddr, addrfmly, NULL,
2030 				    rpcerr, TRUE, nosignal, cr));
2031 			} else {
2032 				CONN_HOLD(cm_entry);
2033 
2034 				cm_entry->x_time = lbolt;
2035 				mutex_exit(&connmgr_lock);
2036 				RPCLOG(2, "connmgr_get: found old "
2037 				    "transport %p for retry\n",
2038 				    (void *)cm_entry);
2039 				return (cm_entry);
2040 			}
2041 		}
2042 
2043 		/*
2044 		 * We cannot find an entry in the list for this retry.
2045 		 * Either the entry has been removed temporarily to be
2046 		 * reconnected by another thread, or the original call
2047 		 * got a port but never got connected,
2048 		 * and hence the transport never got put in the
2049 		 * list.  Fall through to the "create new connection" code -
2050 		 * the former case will fail there trying to rebind the port,
2051 		 * and the later case (and any other pathological cases) will
2052 		 * rebind and reconnect and not hang the client machine.
2053 		 */
2054 		RPCLOG0(8, "connmgr_get: no entry in list for retry\n");
2055 	}
2056 	/*
2057 	 * Set up a transport entry in the connection manager's list.
2058 	 */
2059 	cm_entry = (struct cm_xprt *)
2060 	    kmem_zalloc(sizeof (struct cm_xprt), KM_SLEEP);
2061 
2062 	cm_entry->x_server.buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2063 	bcopy(destaddr->buf, cm_entry->x_server.buf, destaddr->len);
2064 	cm_entry->x_server.len = cm_entry->x_server.maxlen = destaddr->len;
2065 
2066 	cm_entry->x_state_flags = X_THREAD;
2067 	cm_entry->x_ref = 1;
2068 	cm_entry->x_family = addrfmly;
2069 	cm_entry->x_rdev = device;
2070 	cm_entry->x_zoneid = zoneid;
2071 	mutex_init(&cm_entry->x_lock, NULL, MUTEX_DEFAULT, NULL);
2072 	cv_init(&cm_entry->x_cv, NULL, CV_DEFAULT, NULL);
2073 	cv_init(&cm_entry->x_conn_cv, NULL, CV_DEFAULT, NULL);
2074 	cv_init(&cm_entry->x_dis_cv, NULL, CV_DEFAULT, NULL);
2075 
2076 	/*
2077 	 * Note that we add this partially initialized entry to the
2078 	 * connection list. This is so that we don't have connections to
2079 	 * the same server.
2080 	 *
2081 	 * Note that x_src is not initialized at this point. This is because
2082 	 * retryaddr might be NULL in which case x_src is whatever
2083 	 * t_kbind/bindresvport gives us. If another thread wants a
2084 	 * connection to the same server, seemingly we have an issue, but we
2085 	 * don't. If the other thread comes in with retryaddr == NULL, then it
2086 	 * will never look at x_src, and it will end up waiting in
2087 	 * connmgr_cwait() for the first thread to finish the connection
2088 	 * attempt. If the other thread comes in with retryaddr != NULL, then
2089 	 * that means there was a request sent on a connection, in which case
2090 	 * the the connection should already exist. Thus the first thread
2091 	 * never gets here ... it finds the connection it its server in the
2092 	 * connection list.
2093 	 *
2094 	 * But even if theory is wrong, in the retryaddr != NULL case, the 2nd
2095 	 * thread will skip us because x_src.len == 0.
2096 	 */
2097 	cm_entry->x_next = cm_hd;
2098 	cm_hd = cm_entry;
2099 	mutex_exit(&connmgr_lock);
2100 
2101 	/*
2102 	 * Either we didn't find an entry to the server of interest, or we
2103 	 * don't have the maximum number of connections to that server -
2104 	 * create a new connection.
2105 	 */
2106 	RPCLOG0(8, "connmgr_get: creating new connection\n");
2107 	rpcerr->re_status = RPC_TLIERROR;
2108 
2109 	i = t_kopen(NULL, device, FREAD|FWRITE|FNDELAY, &tiptr, zone_kcred());
2110 	if (i) {
2111 		RPCLOG(1, "connmgr_get: can't open cots device, error %d\n", i);
2112 		rpcerr->re_errno = i;
2113 		connmgr_cancelconn(cm_entry);
2114 		return (NULL);
2115 	}
2116 	rpc_poptimod(tiptr->fp->f_vnode);
2117 
2118 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"rpcmod", 0,
2119 	    K_TO_K, kcred, &retval)) {
2120 		RPCLOG(1, "connmgr_get: can't push cots module, %d\n", i);
2121 		(void) t_kclose(tiptr, 1);
2122 		rpcerr->re_errno = i;
2123 		connmgr_cancelconn(cm_entry);
2124 		return (NULL);
2125 	}
2126 
2127 	if (i = strioctl(tiptr->fp->f_vnode, RPC_CLIENT, 0, 0, K_TO_K,
2128 	    kcred, &retval)) {
2129 		RPCLOG(1, "connmgr_get: can't set client status with cots "
2130 		    "module, %d\n", i);
2131 		(void) t_kclose(tiptr, 1);
2132 		rpcerr->re_errno = i;
2133 		connmgr_cancelconn(cm_entry);
2134 		return (NULL);
2135 	}
2136 
2137 	mutex_enter(&connmgr_lock);
2138 
2139 	wq = tiptr->fp->f_vnode->v_stream->sd_wrq->q_next;
2140 	cm_entry->x_wq = wq;
2141 
2142 	mutex_exit(&connmgr_lock);
2143 
2144 	if (i = strioctl(tiptr->fp->f_vnode, I_PUSH, (intptr_t)"timod", 0,
2145 	    K_TO_K, kcred, &retval)) {
2146 		RPCLOG(1, "connmgr_get: can't push timod, %d\n", i);
2147 		(void) t_kclose(tiptr, 1);
2148 		rpcerr->re_errno = i;
2149 		connmgr_cancelconn(cm_entry);
2150 		return (NULL);
2151 	}
2152 
2153 	/*
2154 	 * If the caller has not specified reserved port usage then
2155 	 * take the system default.
2156 	 */
2157 	if (useresvport == -1)
2158 		useresvport = clnt_cots_do_bindresvport;
2159 
2160 	if ((useresvport || retryaddr != NULL) &&
2161 	    (addrfmly == AF_INET || addrfmly == AF_INET6)) {
2162 		bool_t alloc_src = FALSE;
2163 
2164 		if (srcaddr->len != destaddr->len) {
2165 			kmem_free(srcaddr->buf, srcaddr->maxlen);
2166 			srcaddr->buf = kmem_zalloc(destaddr->len, KM_SLEEP);
2167 			srcaddr->maxlen = destaddr->len;
2168 			srcaddr->len = destaddr->len;
2169 			alloc_src = TRUE;
2170 		}
2171 
2172 		if ((i = bindresvport(tiptr, retryaddr, srcaddr, TRUE)) != 0) {
2173 			(void) t_kclose(tiptr, 1);
2174 			RPCLOG(1, "connmgr_get: couldn't bind, retryaddr: "
2175 			    "%p\n", (void *)retryaddr);
2176 
2177 			/*
2178 			 * 1225408: If we allocated a source address, then it
2179 			 * is either garbage or all zeroes. In that case
2180 			 * we need to clear srcaddr.
2181 			 */
2182 			if (alloc_src == TRUE) {
2183 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2184 				srcaddr->maxlen = srcaddr->len = 0;
2185 				srcaddr->buf = NULL;
2186 			}
2187 			rpcerr->re_errno = i;
2188 			connmgr_cancelconn(cm_entry);
2189 			return (NULL);
2190 		}
2191 	} else {
2192 		if ((i = t_kbind(tiptr, NULL, NULL)) != 0) {
2193 			RPCLOG(1, "clnt_cots_kcreate: t_kbind: %d\n", i);
2194 			(void) t_kclose(tiptr, 1);
2195 			rpcerr->re_errno = i;
2196 			connmgr_cancelconn(cm_entry);
2197 			return (NULL);
2198 		}
2199 	}
2200 
2201 	{
2202 		/*
2203 		 * Keep the kernel stack lean. Don't move this call
2204 		 * declaration to the top of this function because a
2205 		 * call is declared in connmgr_wrapconnect()
2206 		 */
2207 		calllist_t call;
2208 
2209 		bzero(&call, sizeof (call));
2210 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2211 
2212 		/*
2213 		 * This is a bound end-point so don't close it's stream.
2214 		 */
2215 		connected = connmgr_connect(cm_entry, wq, destaddr, addrfmly,
2216 		    &call, &tidu_size, FALSE, waitp, nosignal, cr);
2217 		*rpcerr = call.call_err;
2218 		cv_destroy(&call.call_cv);
2219 
2220 	}
2221 
2222 	mutex_enter(&connmgr_lock);
2223 
2224 	/*
2225 	 * Set up a transport entry in the connection manager's list.
2226 	 */
2227 	cm_entry->x_src.buf = kmem_zalloc(srcaddr->len, KM_SLEEP);
2228 	bcopy(srcaddr->buf, cm_entry->x_src.buf, srcaddr->len);
2229 	cm_entry->x_src.len = cm_entry->x_src.maxlen = srcaddr->len;
2230 
2231 	cm_entry->x_tiptr = tiptr;
2232 	cm_entry->x_time = lbolt;
2233 
2234 	if (tiptr->tp_info.servtype == T_COTS_ORD)
2235 		cm_entry->x_ordrel = TRUE;
2236 	else
2237 		cm_entry->x_ordrel = FALSE;
2238 
2239 	cm_entry->x_tidu_size = tidu_size;
2240 
2241 	if (cm_entry->x_early_disc) {
2242 		/*
2243 		 * We need to check if a disconnect request has come
2244 		 * while we are connected, if so, then we need to
2245 		 * set rpcerr->re_status appropriately before returning
2246 		 * NULL to caller.
2247 		 */
2248 		if (rpcerr->re_status == RPC_SUCCESS)
2249 			rpcerr->re_status = RPC_XPRTFAILED;
2250 		cm_entry->x_connected = FALSE;
2251 	} else
2252 		cm_entry->x_connected = connected;
2253 
2254 	/*
2255 	 * There could be a discrepancy here such that
2256 	 * x_early_disc is TRUE yet connected is TRUE as well
2257 	 * and the connection is actually connected. In that case
2258 	 * lets be conservative and declare the connection as not
2259 	 * connected.
2260 	 */
2261 	cm_entry->x_early_disc = FALSE;
2262 	cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2263 	cm_entry->x_ctime = lbolt;
2264 
2265 	/*
2266 	 * Notify any threads waiting that the connection attempt is done.
2267 	 */
2268 	cm_entry->x_thread = FALSE;
2269 	cv_broadcast(&cm_entry->x_conn_cv);
2270 
2271 	if (cm_entry->x_connected == FALSE) {
2272 		mutex_exit(&connmgr_lock);
2273 		connmgr_release(cm_entry);
2274 		return (NULL);
2275 	}
2276 
2277 	mutex_exit(&connmgr_lock);
2278 
2279 	return (cm_entry);
2280 }
2281 
2282 /*
2283  * Keep the cm_xprt entry on the connecton list when making a connection. This
2284  * is to prevent multiple connections to a slow server from appearing.
2285  * We use the bit field x_thread to tell if a thread is doing a connection
2286  * which keeps other interested threads from messing with connection.
2287  * Those other threads just wait if x_thread is set.
2288  *
2289  * If x_thread is not set, then we do the actual work of connecting via
2290  * connmgr_connect().
2291  *
2292  * mutex convention: called with connmgr_lock held, returns with it released.
2293  */
2294 static struct cm_xprt *
2295 connmgr_wrapconnect(
2296 	struct cm_xprt	*cm_entry,
2297 	const struct timeval	*waitp,
2298 	struct netbuf	*destaddr,
2299 	int		addrfmly,
2300 	struct netbuf	*srcaddr,
2301 	struct rpc_err	*rpcerr,
2302 	bool_t		reconnect,
2303 	bool_t		nosignal,
2304 	cred_t		*cr)
2305 {
2306 	ASSERT(MUTEX_HELD(&connmgr_lock));
2307 	/*
2308 	 * Hold this entry as we are about to drop connmgr_lock.
2309 	 */
2310 	CONN_HOLD(cm_entry);
2311 
2312 	/*
2313 	 * If there is a thread already making a connection for us, then
2314 	 * wait for it to complete the connection.
2315 	 */
2316 	if (cm_entry->x_thread == TRUE) {
2317 		rpcerr->re_status = connmgr_cwait(cm_entry, waitp, nosignal);
2318 
2319 		if (rpcerr->re_status != RPC_SUCCESS) {
2320 			mutex_exit(&connmgr_lock);
2321 			connmgr_release(cm_entry);
2322 			return (NULL);
2323 		}
2324 	} else {
2325 		bool_t connected;
2326 		calllist_t call;
2327 
2328 		cm_entry->x_thread = TRUE;
2329 
2330 		while (cm_entry->x_needrel == TRUE) {
2331 			cm_entry->x_needrel = FALSE;
2332 
2333 			connmgr_sndrel(cm_entry);
2334 			delay(drv_usectohz(1000000));
2335 
2336 			mutex_enter(&connmgr_lock);
2337 		}
2338 
2339 		/*
2340 		 * If we need to send a T_DISCON_REQ, send one.
2341 		 */
2342 		connmgr_dis_and_wait(cm_entry);
2343 
2344 		mutex_exit(&connmgr_lock);
2345 
2346 		bzero(&call, sizeof (call));
2347 		cv_init(&call.call_cv, NULL, CV_DEFAULT, NULL);
2348 
2349 		connected = connmgr_connect(cm_entry, cm_entry->x_wq,
2350 		    destaddr, addrfmly, &call, &cm_entry->x_tidu_size,
2351 		    reconnect, waitp, nosignal, cr);
2352 
2353 		*rpcerr = call.call_err;
2354 		cv_destroy(&call.call_cv);
2355 
2356 		mutex_enter(&connmgr_lock);
2357 
2358 
2359 		if (cm_entry->x_early_disc) {
2360 			/*
2361 			 * We need to check if a disconnect request has come
2362 			 * while we are connected, if so, then we need to
2363 			 * set rpcerr->re_status appropriately before returning
2364 			 * NULL to caller.
2365 			 */
2366 			if (rpcerr->re_status == RPC_SUCCESS)
2367 				rpcerr->re_status = RPC_XPRTFAILED;
2368 			cm_entry->x_connected = FALSE;
2369 		} else
2370 			cm_entry->x_connected = connected;
2371 
2372 		/*
2373 		 * There could be a discrepancy here such that
2374 		 * x_early_disc is TRUE yet connected is TRUE as well
2375 		 * and the connection is actually connected. In that case
2376 		 * lets be conservative and declare the connection as not
2377 		 * connected.
2378 		 */
2379 
2380 		cm_entry->x_early_disc = FALSE;
2381 		cm_entry->x_needdis = (cm_entry->x_connected == FALSE);
2382 
2383 
2384 		/*
2385 		 * connmgr_connect() may have given up before the connection
2386 		 * actually timed out. So ensure that before the next
2387 		 * connection attempt we do a disconnect.
2388 		 */
2389 		cm_entry->x_ctime = lbolt;
2390 		cm_entry->x_thread = FALSE;
2391 
2392 		cv_broadcast(&cm_entry->x_conn_cv);
2393 
2394 		if (cm_entry->x_connected == FALSE) {
2395 			mutex_exit(&connmgr_lock);
2396 			connmgr_release(cm_entry);
2397 			return (NULL);
2398 		}
2399 	}
2400 
2401 	if (srcaddr != NULL) {
2402 		/*
2403 		 * Copy into the handle the
2404 		 * source address of the
2405 		 * connection, which we will use
2406 		 * in case of a later retry.
2407 		 */
2408 		if (srcaddr->len != cm_entry->x_src.len) {
2409 			if (srcaddr->maxlen > 0)
2410 				kmem_free(srcaddr->buf, srcaddr->maxlen);
2411 			srcaddr->buf = kmem_zalloc(cm_entry->x_src.len,
2412 			    KM_SLEEP);
2413 			srcaddr->maxlen = srcaddr->len =
2414 			    cm_entry->x_src.len;
2415 		}
2416 		bcopy(cm_entry->x_src.buf, srcaddr->buf, srcaddr->len);
2417 	}
2418 	cm_entry->x_time = lbolt;
2419 	mutex_exit(&connmgr_lock);
2420 	return (cm_entry);
2421 }
2422 
2423 /*
2424  * If we need to send a T_DISCON_REQ, send one.
2425  */
2426 static void
2427 connmgr_dis_and_wait(struct cm_xprt *cm_entry)
2428 {
2429 	ASSERT(MUTEX_HELD(&connmgr_lock));
2430 	for (;;) {
2431 		while (cm_entry->x_needdis == TRUE) {
2432 			RPCLOG(8, "connmgr_dis_and_wait: need "
2433 			    "T_DISCON_REQ for connection 0x%p\n",
2434 			    (void *)cm_entry);
2435 			cm_entry->x_needdis = FALSE;
2436 			cm_entry->x_waitdis = TRUE;
2437 
2438 			connmgr_snddis(cm_entry);
2439 
2440 			mutex_enter(&connmgr_lock);
2441 		}
2442 
2443 		if (cm_entry->x_waitdis == TRUE) {
2444 			clock_t curlbolt;
2445 			clock_t timout;
2446 
2447 			RPCLOG(8, "connmgr_dis_and_wait waiting for "
2448 			    "T_DISCON_REQ's ACK for connection %p\n",
2449 			    (void *)cm_entry);
2450 			curlbolt = ddi_get_lbolt();
2451 
2452 			timout = clnt_cots_min_conntout *
2453 			    drv_usectohz(1000000) + curlbolt;
2454 
2455 			/*
2456 			 * The TPI spec says that the T_DISCON_REQ
2457 			 * will get acknowledged, but in practice
2458 			 * the ACK may never get sent. So don't
2459 			 * block forever.
2460 			 */
2461 			(void) cv_timedwait(&cm_entry->x_dis_cv,
2462 			    &connmgr_lock, timout);
2463 		}
2464 		/*
2465 		 * If we got the ACK, break. If we didn't,
2466 		 * then send another T_DISCON_REQ.
2467 		 */
2468 		if (cm_entry->x_waitdis == FALSE) {
2469 			break;
2470 		} else {
2471 			RPCLOG(8, "connmgr_dis_and_wait: did"
2472 			    "not get T_DISCON_REQ's ACK for "
2473 			    "connection  %p\n", (void *)cm_entry);
2474 			cm_entry->x_needdis = TRUE;
2475 		}
2476 	}
2477 }
2478 
2479 static void
2480 connmgr_cancelconn(struct cm_xprt *cm_entry)
2481 {
2482 	/*
2483 	 * Mark the connection table entry as dead; the next thread that
2484 	 * goes through connmgr_release() will notice this and deal with it.
2485 	 */
2486 	mutex_enter(&connmgr_lock);
2487 	cm_entry->x_dead = TRUE;
2488 
2489 	/*
2490 	 * Notify any threads waiting for the connection that it isn't
2491 	 * going to happen.
2492 	 */
2493 	cm_entry->x_thread = FALSE;
2494 	cv_broadcast(&cm_entry->x_conn_cv);
2495 	mutex_exit(&connmgr_lock);
2496 
2497 	connmgr_release(cm_entry);
2498 }
2499 
2500 static void
2501 connmgr_close(struct cm_xprt *cm_entry)
2502 {
2503 	mutex_enter(&cm_entry->x_lock);
2504 	while (cm_entry->x_ref != 0) {
2505 		/*
2506 		 * Must be a noninterruptible wait.
2507 		 */
2508 		cv_wait(&cm_entry->x_cv, &cm_entry->x_lock);
2509 	}
2510 
2511 	if (cm_entry->x_tiptr != NULL)
2512 		(void) t_kclose(cm_entry->x_tiptr, 1);
2513 
2514 	mutex_exit(&cm_entry->x_lock);
2515 	if (cm_entry->x_ksp != NULL) {
2516 		mutex_enter(&connmgr_lock);
2517 		cm_entry->x_ksp->ks_private = NULL;
2518 		mutex_exit(&connmgr_lock);
2519 
2520 		/*
2521 		 * Must free the buffer we allocated for the
2522 		 * server address in the update function
2523 		 */
2524 		if (((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2525 		    x_server.value.str.addr.ptr != NULL)
2526 			kmem_free(((struct cm_kstat_xprt *)(cm_entry->x_ksp->
2527 			    ks_data))->x_server.value.str.addr.ptr,
2528 			    INET6_ADDRSTRLEN);
2529 		kmem_free(cm_entry->x_ksp->ks_data,
2530 		    cm_entry->x_ksp->ks_data_size);
2531 		kstat_delete(cm_entry->x_ksp);
2532 	}
2533 
2534 	mutex_destroy(&cm_entry->x_lock);
2535 	cv_destroy(&cm_entry->x_cv);
2536 	cv_destroy(&cm_entry->x_conn_cv);
2537 	cv_destroy(&cm_entry->x_dis_cv);
2538 
2539 	if (cm_entry->x_server.buf != NULL)
2540 		kmem_free(cm_entry->x_server.buf, cm_entry->x_server.maxlen);
2541 	if (cm_entry->x_src.buf != NULL)
2542 		kmem_free(cm_entry->x_src.buf, cm_entry->x_src.maxlen);
2543 	kmem_free(cm_entry, sizeof (struct cm_xprt));
2544 }
2545 
2546 /*
2547  * Called by KRPC after sending the call message to release the connection
2548  * it was using.
2549  */
2550 static void
2551 connmgr_release(struct cm_xprt *cm_entry)
2552 {
2553 	mutex_enter(&cm_entry->x_lock);
2554 	cm_entry->x_ref--;
2555 	if (cm_entry->x_ref == 0)
2556 		cv_signal(&cm_entry->x_cv);
2557 	mutex_exit(&cm_entry->x_lock);
2558 }
2559 
2560 /*
2561  * Given an open stream, connect to the remote.  Returns true if connected,
2562  * false otherwise.
2563  */
2564 static bool_t
2565 connmgr_connect(
2566 	struct cm_xprt		*cm_entry,
2567 	queue_t			*wq,
2568 	struct netbuf		*addr,
2569 	int			addrfmly,
2570 	calllist_t 		*e,
2571 	int 			*tidu_ptr,
2572 	bool_t 			reconnect,
2573 	const struct timeval 	*waitp,
2574 	bool_t 			nosignal,
2575 	cred_t			*cr)
2576 {
2577 	mblk_t *mp;
2578 	struct T_conn_req *tcr;
2579 	struct T_info_ack *tinfo;
2580 	int interrupted, error;
2581 	int tidu_size, kstat_instance;
2582 
2583 	/* if it's a reconnect, flush any lingering data messages */
2584 	if (reconnect)
2585 		(void) putctl1(wq, M_FLUSH, FLUSHRW);
2586 
2587 	/*
2588 	 * Note: if the receiver uses SCM_UCRED/getpeerucred the pid will
2589 	 * appear as -1.
2590 	 */
2591 	mp = allocb_cred(sizeof (*tcr) + addr->len, cr, NOPID);
2592 	if (mp == NULL) {
2593 		/*
2594 		 * This is unfortunate, but we need to look up the stats for
2595 		 * this zone to increment the "memory allocation failed"
2596 		 * counter.  curproc->p_zone is safe since we're initiating a
2597 		 * connection and not in some strange streams context.
2598 		 */
2599 		struct rpcstat *rpcstat;
2600 
2601 		rpcstat = zone_getspecific(rpcstat_zone_key, rpc_zone());
2602 		ASSERT(rpcstat != NULL);
2603 
2604 		RPCLOG0(1, "connmgr_connect: cannot alloc mp for "
2605 		    "sending conn request\n");
2606 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcnomem);
2607 		e->call_status = RPC_SYSTEMERROR;
2608 		e->call_reason = ENOSR;
2609 		return (FALSE);
2610 	}
2611 
2612 	mp->b_datap->db_type = M_PROTO;
2613 	tcr = (struct T_conn_req *)mp->b_rptr;
2614 	bzero(tcr, sizeof (*tcr));
2615 	tcr->PRIM_type = T_CONN_REQ;
2616 	tcr->DEST_length = addr->len;
2617 	tcr->DEST_offset = sizeof (struct T_conn_req);
2618 	mp->b_wptr = mp->b_rptr + sizeof (*tcr);
2619 
2620 	bcopy(addr->buf, mp->b_wptr, tcr->DEST_length);
2621 	mp->b_wptr += tcr->DEST_length;
2622 
2623 	RPCLOG(8, "connmgr_connect: sending conn request on queue "
2624 	    "%p", (void *)wq);
2625 	RPCLOG(8, " call %p\n", (void *)wq);
2626 	/*
2627 	 * We use the entry in the handle that is normally used for
2628 	 * waiting for RPC replies to wait for the connection accept.
2629 	 */
2630 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2631 		DTRACE_PROBE(krpc__e__connmgr__connect__cantsend);
2632 		freemsg(mp);
2633 		return (FALSE);
2634 	}
2635 
2636 	mutex_enter(&clnt_pending_lock);
2637 
2638 	/*
2639 	 * We wait for the transport connection to be made, or an
2640 	 * indication that it could not be made.
2641 	 */
2642 	interrupted = 0;
2643 
2644 	/*
2645 	 * waitforack should have been called with T_OK_ACK, but the
2646 	 * present implementation needs to be passed T_INFO_ACK to
2647 	 * work correctly.
2648 	 */
2649 	error = waitforack(e, T_INFO_ACK, waitp, nosignal);
2650 	if (error == EINTR)
2651 		interrupted = 1;
2652 	if (zone_status_get(curproc->p_zone) >= ZONE_IS_EMPTY) {
2653 		/*
2654 		 * No time to lose; we essentially have been signaled to
2655 		 * quit.
2656 		 */
2657 		interrupted = 1;
2658 	}
2659 #ifdef RPCDEBUG
2660 	if (error == ETIME)
2661 		RPCLOG0(8, "connmgr_connect: giving up "
2662 		    "on connection attempt; "
2663 		    "clnt_dispatch notifyconn "
2664 		    "diagnostic 'no one waiting for "
2665 		    "connection' should not be "
2666 		    "unexpected\n");
2667 #endif
2668 	if (e->call_prev)
2669 		e->call_prev->call_next = e->call_next;
2670 	else
2671 		clnt_pending = e->call_next;
2672 	if (e->call_next)
2673 		e->call_next->call_prev = e->call_prev;
2674 	mutex_exit(&clnt_pending_lock);
2675 
2676 	if (e->call_status != RPC_SUCCESS || error != 0) {
2677 		if (interrupted)
2678 			e->call_status = RPC_INTR;
2679 		else if (error == ETIME)
2680 			e->call_status = RPC_TIMEDOUT;
2681 		else if (error == EPROTO) {
2682 			e->call_status = RPC_SYSTEMERROR;
2683 			e->call_reason = EPROTO;
2684 		}
2685 
2686 		RPCLOG(8, "connmgr_connect: can't connect, status: "
2687 		    "%s\n", clnt_sperrno(e->call_status));
2688 
2689 		if (e->call_reply) {
2690 			freemsg(e->call_reply);
2691 			e->call_reply = NULL;
2692 		}
2693 
2694 		return (FALSE);
2695 	}
2696 	/*
2697 	 * The result of the "connection accept" is a T_info_ack
2698 	 * in the call_reply field.
2699 	 */
2700 	ASSERT(e->call_reply != NULL);
2701 	mp = e->call_reply;
2702 	e->call_reply = NULL;
2703 	tinfo = (struct T_info_ack *)mp->b_rptr;
2704 
2705 	tidu_size = tinfo->TIDU_size;
2706 	tidu_size -= (tidu_size % BYTES_PER_XDR_UNIT);
2707 	if (tidu_size > COTS_DEFAULT_ALLOCSIZE || (tidu_size <= 0))
2708 		tidu_size = COTS_DEFAULT_ALLOCSIZE;
2709 	*tidu_ptr = tidu_size;
2710 
2711 	freemsg(mp);
2712 
2713 	/*
2714 	 * Set up the pertinent options.  NODELAY is so the transport doesn't
2715 	 * buffer up RPC messages on either end.  This may not be valid for
2716 	 * all transports.  Failure to set this option is not cause to
2717 	 * bail out so we return success anyway.  Note that lack of NODELAY
2718 	 * or some other way to flush the message on both ends will cause
2719 	 * lots of retries and terrible performance.
2720 	 */
2721 	if (addrfmly == AF_INET || addrfmly == AF_INET6) {
2722 		(void) connmgr_setopt(wq, IPPROTO_TCP, TCP_NODELAY, e, cr);
2723 		if (e->call_status == RPC_XPRTFAILED)
2724 			return (FALSE);
2725 	}
2726 
2727 	/*
2728 	 * Since we have a connection, we now need to figure out if
2729 	 * we need to create a kstat. If x_ksp is not NULL then we
2730 	 * are reusing a connection and so we do not need to create
2731 	 * another kstat -- lets just return.
2732 	 */
2733 	if (cm_entry->x_ksp != NULL)
2734 		return (TRUE);
2735 
2736 	/*
2737 	 * We need to increment rpc_kstat_instance atomically to prevent
2738 	 * two kstats being created with the same instance.
2739 	 */
2740 	kstat_instance = atomic_add_32_nv((uint32_t *)&rpc_kstat_instance, 1);
2741 
2742 	if ((cm_entry->x_ksp = kstat_create_zone("unix", kstat_instance,
2743 	    "rpc_cots_connections", "rpc", KSTAT_TYPE_NAMED,
2744 	    (uint_t)(sizeof (cm_kstat_xprt_t) / sizeof (kstat_named_t)),
2745 	    KSTAT_FLAG_VIRTUAL, cm_entry->x_zoneid)) == NULL) {
2746 		return (TRUE);
2747 	}
2748 
2749 	cm_entry->x_ksp->ks_lock = &connmgr_lock;
2750 	cm_entry->x_ksp->ks_private = cm_entry;
2751 	cm_entry->x_ksp->ks_data_size = ((INET6_ADDRSTRLEN * sizeof (char))
2752 	    + sizeof (cm_kstat_template));
2753 	cm_entry->x_ksp->ks_data = kmem_alloc(cm_entry->x_ksp->ks_data_size,
2754 	    KM_SLEEP);
2755 	bcopy(&cm_kstat_template, cm_entry->x_ksp->ks_data,
2756 	    cm_entry->x_ksp->ks_data_size);
2757 	((struct cm_kstat_xprt *)(cm_entry->x_ksp->ks_data))->
2758 	    x_server.value.str.addr.ptr =
2759 	    kmem_alloc(INET6_ADDRSTRLEN, KM_SLEEP);
2760 
2761 	cm_entry->x_ksp->ks_update = conn_kstat_update;
2762 	kstat_install(cm_entry->x_ksp);
2763 	return (TRUE);
2764 }
2765 
2766 /*
2767  * Called by connmgr_connect to set an option on the new stream.
2768  */
2769 static bool_t
2770 connmgr_setopt(queue_t *wq, int level, int name, calllist_t *e, cred_t *cr)
2771 {
2772 	mblk_t *mp;
2773 	struct opthdr *opt;
2774 	struct T_optmgmt_req *tor;
2775 	struct timeval waitp;
2776 	int error;
2777 
2778 	mp = allocb_cred(sizeof (struct T_optmgmt_req) +
2779 	    sizeof (struct opthdr) + sizeof (int), cr, NOPID);
2780 	if (mp == NULL) {
2781 		RPCLOG0(1, "connmgr_setopt: cannot alloc mp for option "
2782 		    "request\n");
2783 		return (FALSE);
2784 	}
2785 
2786 	mp->b_datap->db_type = M_PROTO;
2787 	tor = (struct T_optmgmt_req *)(mp->b_rptr);
2788 	tor->PRIM_type = T_SVR4_OPTMGMT_REQ;
2789 	tor->MGMT_flags = T_NEGOTIATE;
2790 	tor->OPT_length = sizeof (struct opthdr) + sizeof (int);
2791 	tor->OPT_offset = sizeof (struct T_optmgmt_req);
2792 
2793 	opt = (struct opthdr *)(mp->b_rptr + sizeof (struct T_optmgmt_req));
2794 	opt->level = level;
2795 	opt->name = name;
2796 	opt->len = sizeof (int);
2797 	*(int *)((char *)opt + sizeof (*opt)) = 1;
2798 	mp->b_wptr += sizeof (struct T_optmgmt_req) + sizeof (struct opthdr) +
2799 	    sizeof (int);
2800 
2801 	/*
2802 	 * We will use this connection regardless
2803 	 * of whether or not the option is settable.
2804 	 */
2805 	if (clnt_dispatch_send(wq, mp, e, 0, 0) != RPC_SUCCESS) {
2806 		DTRACE_PROBE(krpc__e__connmgr__setopt__cantsend);
2807 		freemsg(mp);
2808 		return (FALSE);
2809 	}
2810 
2811 	mutex_enter(&clnt_pending_lock);
2812 
2813 	waitp.tv_sec = clnt_cots_min_conntout;
2814 	waitp.tv_usec = 0;
2815 	error = waitforack(e, T_OPTMGMT_ACK, &waitp, 1);
2816 
2817 	if (e->call_prev)
2818 		e->call_prev->call_next = e->call_next;
2819 	else
2820 		clnt_pending = e->call_next;
2821 	if (e->call_next)
2822 		e->call_next->call_prev = e->call_prev;
2823 	mutex_exit(&clnt_pending_lock);
2824 
2825 	if (e->call_reply != NULL) {
2826 		freemsg(e->call_reply);
2827 		e->call_reply = NULL;
2828 	}
2829 
2830 	if (e->call_status != RPC_SUCCESS || error != 0) {
2831 		RPCLOG(1, "connmgr_setopt: can't set option: %d\n", name);
2832 		return (FALSE);
2833 	}
2834 	RPCLOG(8, "connmgr_setopt: successfully set option: %d\n", name);
2835 	return (TRUE);
2836 }
2837 
2838 #ifdef	DEBUG
2839 
2840 /*
2841  * This is a knob to let us force code coverage in allocation failure
2842  * case.
2843  */
2844 static int	connmgr_failsnd;
2845 #define	CONN_SND_ALLOC(Size, Pri)	\
2846 	((connmgr_failsnd-- > 0) ? NULL : allocb(Size, Pri))
2847 
2848 #else
2849 
2850 #define	CONN_SND_ALLOC(Size, Pri)	allocb(Size, Pri)
2851 
2852 #endif
2853 
2854 /*
2855  * Sends an orderly release on the specified queue.
2856  * Entered with connmgr_lock. Exited without connmgr_lock
2857  */
2858 static void
2859 connmgr_sndrel(struct cm_xprt *cm_entry)
2860 {
2861 	struct T_ordrel_req *torr;
2862 	mblk_t *mp;
2863 	queue_t *q = cm_entry->x_wq;
2864 	ASSERT(MUTEX_HELD(&connmgr_lock));
2865 	mp = CONN_SND_ALLOC(sizeof (struct T_ordrel_req), BPRI_LO);
2866 	if (mp == NULL) {
2867 		cm_entry->x_needrel = TRUE;
2868 		mutex_exit(&connmgr_lock);
2869 		RPCLOG(1, "connmgr_sndrel: cannot alloc mp for sending ordrel "
2870 		    "to queue %p\n", (void *)q);
2871 		return;
2872 	}
2873 	mutex_exit(&connmgr_lock);
2874 
2875 	mp->b_datap->db_type = M_PROTO;
2876 	torr = (struct T_ordrel_req *)(mp->b_rptr);
2877 	torr->PRIM_type = T_ORDREL_REQ;
2878 	mp->b_wptr = mp->b_rptr + sizeof (struct T_ordrel_req);
2879 
2880 	RPCLOG(8, "connmgr_sndrel: sending ordrel to queue %p\n", (void *)q);
2881 	put(q, mp);
2882 }
2883 
2884 /*
2885  * Sends an disconnect on the specified queue.
2886  * Entered with connmgr_lock. Exited without connmgr_lock
2887  */
2888 static void
2889 connmgr_snddis(struct cm_xprt *cm_entry)
2890 {
2891 	struct T_discon_req *tdis;
2892 	mblk_t *mp;
2893 	queue_t *q = cm_entry->x_wq;
2894 
2895 	ASSERT(MUTEX_HELD(&connmgr_lock));
2896 	mp = CONN_SND_ALLOC(sizeof (*tdis), BPRI_LO);
2897 	if (mp == NULL) {
2898 		cm_entry->x_needdis = TRUE;
2899 		mutex_exit(&connmgr_lock);
2900 		RPCLOG(1, "connmgr_snddis: cannot alloc mp for sending discon "
2901 		    "to queue %p\n", (void *)q);
2902 		return;
2903 	}
2904 	mutex_exit(&connmgr_lock);
2905 
2906 	mp->b_datap->db_type = M_PROTO;
2907 	tdis = (struct T_discon_req *)mp->b_rptr;
2908 	tdis->PRIM_type = T_DISCON_REQ;
2909 	mp->b_wptr = mp->b_rptr + sizeof (*tdis);
2910 
2911 	RPCLOG(8, "connmgr_snddis: sending discon to queue %p\n", (void *)q);
2912 	put(q, mp);
2913 }
2914 
2915 /*
2916  * Sets up the entry for receiving replies, and calls rpcmod's write put proc
2917  * (through put) to send the call.
2918  */
2919 static int
2920 clnt_dispatch_send(queue_t *q, mblk_t *mp, calllist_t *e, uint_t xid,
2921 			uint_t queue_flag)
2922 {
2923 	ASSERT(e != NULL);
2924 
2925 	e->call_status = RPC_TIMEDOUT;	/* optimistic, eh? */
2926 	e->call_reason = 0;
2927 	e->call_wq = q;
2928 	e->call_xid = xid;
2929 	e->call_notified = FALSE;
2930 
2931 	if (!canput(q)) {
2932 		e->call_status = RPC_CANTSEND;
2933 		e->call_reason = ENOBUFS;
2934 		return (RPC_CANTSEND);
2935 	}
2936 
2937 	/*
2938 	 * If queue_flag is set then the calllist_t is already on the hash
2939 	 * queue.  In this case just send the message and return.
2940 	 */
2941 	if (queue_flag) {
2942 		put(q, mp);
2943 		return (RPC_SUCCESS);
2944 
2945 	}
2946 
2947 	/*
2948 	 * Set up calls for RPC requests (with XID != 0) on the hash
2949 	 * queue for fast lookups and place other calls (i.e.
2950 	 * connection management) on the linked list.
2951 	 */
2952 	if (xid != 0) {
2953 		RPCLOG(64, "clnt_dispatch_send: putting xid 0x%x on "
2954 		    "dispatch list\n", xid);
2955 		e->call_hash = call_hash(xid, clnt_cots_hash_size);
2956 		e->call_bucket = &cots_call_ht[e->call_hash];
2957 		call_table_enter(e);
2958 	} else {
2959 		mutex_enter(&clnt_pending_lock);
2960 		if (clnt_pending)
2961 			clnt_pending->call_prev = e;
2962 		e->call_next = clnt_pending;
2963 		e->call_prev = NULL;
2964 		clnt_pending = e;
2965 		mutex_exit(&clnt_pending_lock);
2966 	}
2967 
2968 	put(q, mp);
2969 	return (RPC_SUCCESS);
2970 }
2971 
2972 /*
2973  * Called by rpcmod to notify a client with a clnt_pending call that its reply
2974  * has arrived.  If we can't find a client waiting for this reply, we log
2975  * the error and return.
2976  */
2977 bool_t
2978 clnt_dispatch_notify(mblk_t *mp, zoneid_t zoneid)
2979 {
2980 	calllist_t *e = NULL;
2981 	call_table_t *chtp;
2982 	uint32_t xid;
2983 	uint_t hash;
2984 
2985 	if ((IS_P2ALIGNED(mp->b_rptr, sizeof (uint32_t))) &&
2986 	    (mp->b_wptr - mp->b_rptr) >= sizeof (xid))
2987 		xid = *((uint32_t *)mp->b_rptr);
2988 	else {
2989 		int i = 0;
2990 		unsigned char *p = (unsigned char *)&xid;
2991 		unsigned char *rptr;
2992 		mblk_t *tmp = mp;
2993 
2994 		/*
2995 		 * Copy the xid, byte-by-byte into xid.
2996 		 */
2997 		while (tmp) {
2998 			rptr = tmp->b_rptr;
2999 			while (rptr < tmp->b_wptr) {
3000 				*p++ = *rptr++;
3001 				if (++i >= sizeof (xid))
3002 					goto done_xid_copy;
3003 			}
3004 			tmp = tmp->b_cont;
3005 		}
3006 
3007 		/*
3008 		 * If we got here, we ran out of mblk space before the
3009 		 * xid could be copied.
3010 		 */
3011 		ASSERT(tmp == NULL && i < sizeof (xid));
3012 
3013 		RPCLOG0(1,
3014 		    "clnt_dispatch_notify: message less than size of xid\n");
3015 		return (FALSE);
3016 
3017 	}
3018 done_xid_copy:
3019 
3020 	hash = call_hash(xid, clnt_cots_hash_size);
3021 	chtp = &cots_call_ht[hash];
3022 	/* call_table_find returns with the hash bucket locked */
3023 	call_table_find(chtp, xid, e);
3024 
3025 	if (e != NULL) {
3026 		/*
3027 		 * Found thread waiting for this reply
3028 		 */
3029 		mutex_enter(&e->call_lock);
3030 
3031 		/*
3032 		 * verify that the reply is coming in on
3033 		 * the same zone that it was sent from.
3034 		 */
3035 		if (e->call_zoneid != zoneid) {
3036 			mutex_exit(&e->call_lock);
3037 			mutex_exit(&chtp->ct_lock);
3038 			RPCLOG0(1, "clnt_dispatch_notify: incorrect zoneid\n");
3039 			return (FALSE);
3040 		}
3041 
3042 		if (e->call_reply)
3043 			/*
3044 			 * This can happen under the following scenario:
3045 			 * clnt_cots_kcallit() times out on the response,
3046 			 * rfscall() repeats the CLNT_CALL() with
3047 			 * the same xid, clnt_cots_kcallit() sends the retry,
3048 			 * thereby putting the clnt handle on the pending list,
3049 			 * the first response arrives, signalling the thread
3050 			 * in clnt_cots_kcallit(). Before that thread is
3051 			 * dispatched, the second response arrives as well,
3052 			 * and clnt_dispatch_notify still finds the handle on
3053 			 * the pending list, with call_reply set. So free the
3054 			 * old reply now.
3055 			 *
3056 			 * It is also possible for a response intended for
3057 			 * an RPC call with a different xid to reside here.
3058 			 * This can happen if the thread that owned this
3059 			 * client handle prior to the current owner bailed
3060 			 * out and left its call record on the dispatch
3061 			 * queue.  A window exists where the response can
3062 			 * arrive before the current owner dispatches its
3063 			 * RPC call.
3064 			 *
3065 			 * In any case, this is the very last point where we
3066 			 * can safely check the call_reply field before
3067 			 * placing the new response there.
3068 			 */
3069 			freemsg(e->call_reply);
3070 		e->call_reply = mp;
3071 		e->call_status = RPC_SUCCESS;
3072 		e->call_notified = TRUE;
3073 		cv_signal(&e->call_cv);
3074 		mutex_exit(&e->call_lock);
3075 		mutex_exit(&chtp->ct_lock);
3076 		return (TRUE);
3077 	} else {
3078 		zone_t *zone;
3079 		struct rpcstat *rpcstat;
3080 
3081 		mutex_exit(&chtp->ct_lock);
3082 		RPCLOG(65, "clnt_dispatch_notify: no caller for reply 0x%x\n",
3083 		    xid);
3084 		/*
3085 		 * This is unfortunate, but we need to lookup the zone so we
3086 		 * can increment its "rcbadxids" counter.
3087 		 */
3088 		zone = zone_find_by_id(zoneid);
3089 		if (zone == NULL) {
3090 			/*
3091 			 * The zone went away...
3092 			 */
3093 			return (FALSE);
3094 		}
3095 		rpcstat = zone_getspecific(rpcstat_zone_key, zone);
3096 		if (zone_status_get(zone) >= ZONE_IS_SHUTTING_DOWN) {
3097 			/*
3098 			 * Not interested
3099 			 */
3100 			zone_rele(zone);
3101 			return (FALSE);
3102 		}
3103 		COTSRCSTAT_INCR(rpcstat->rpc_cots_client, rcbadxids);
3104 		zone_rele(zone);
3105 	}
3106 	return (FALSE);
3107 }
3108 
3109 /*
3110  * Called by rpcmod when a non-data indication arrives.  The ones in which we
3111  * are interested are connection indications and options acks.  We dispatch
3112  * based on the queue the indication came in on.  If we are not interested in
3113  * what came in, we return false to rpcmod, who will then pass it upstream.
3114  */
3115 bool_t
3116 clnt_dispatch_notifyconn(queue_t *q, mblk_t *mp)
3117 {
3118 	calllist_t *e;
3119 	int type;
3120 
3121 	ASSERT((q->q_flag & QREADR) == 0);
3122 
3123 	type = ((union T_primitives *)mp->b_rptr)->type;
3124 	RPCLOG(8, "clnt_dispatch_notifyconn: prim type: [%s]\n",
3125 	    rpc_tpiprim2name(type));
3126 	mutex_enter(&clnt_pending_lock);
3127 	for (e = clnt_pending; /* NO CONDITION */; e = e->call_next) {
3128 		if (e == NULL) {
3129 			mutex_exit(&clnt_pending_lock);
3130 			RPCLOG(1, "clnt_dispatch_notifyconn: no one waiting "
3131 			    "for connection on queue 0x%p\n", (void *)q);
3132 			return (FALSE);
3133 		}
3134 		if (e->call_wq == q)
3135 			break;
3136 	}
3137 
3138 	switch (type) {
3139 	case T_CONN_CON:
3140 		/*
3141 		 * The transport is now connected, send a T_INFO_REQ to get
3142 		 * the tidu size.
3143 		 */
3144 		mutex_exit(&clnt_pending_lock);
3145 		ASSERT(mp->b_datap->db_lim - mp->b_datap->db_base >=
3146 		    sizeof (struct T_info_req));
3147 		mp->b_rptr = mp->b_datap->db_base;
3148 		((union T_primitives *)mp->b_rptr)->type = T_INFO_REQ;
3149 		mp->b_wptr = mp->b_rptr + sizeof (struct T_info_req);
3150 		mp->b_datap->db_type = M_PCPROTO;
3151 		put(q, mp);
3152 		return (TRUE);
3153 	case T_INFO_ACK:
3154 	case T_OPTMGMT_ACK:
3155 		e->call_status = RPC_SUCCESS;
3156 		e->call_reply = mp;
3157 		e->call_notified = TRUE;
3158 		cv_signal(&e->call_cv);
3159 		break;
3160 	case T_ERROR_ACK:
3161 		e->call_status = RPC_CANTCONNECT;
3162 		e->call_reply = mp;
3163 		e->call_notified = TRUE;
3164 		cv_signal(&e->call_cv);
3165 		break;
3166 	case T_OK_ACK:
3167 		/*
3168 		 * Great, but we are really waiting for a T_CONN_CON
3169 		 */
3170 		freemsg(mp);
3171 		break;
3172 	default:
3173 		mutex_exit(&clnt_pending_lock);
3174 		RPCLOG(1, "clnt_dispatch_notifyconn: bad type %d\n", type);
3175 		return (FALSE);
3176 	}
3177 
3178 	mutex_exit(&clnt_pending_lock);
3179 	return (TRUE);
3180 }
3181 
3182 /*
3183  * Called by rpcmod when the transport is (or should be) going away.  Informs
3184  * all callers waiting for replies and marks the entry in the connection
3185  * manager's list as unconnected, and either closing (close handshake in
3186  * progress) or dead.
3187  */
3188 void
3189 clnt_dispatch_notifyall(queue_t *q, int32_t msg_type, int32_t reason)
3190 {
3191 	calllist_t *e;
3192 	call_table_t *ctp;
3193 	struct cm_xprt *cm_entry;
3194 	int have_connmgr_lock;
3195 	int i;
3196 
3197 	ASSERT((q->q_flag & QREADR) == 0);
3198 
3199 	RPCLOG(1, "clnt_dispatch_notifyall on queue %p", (void *)q);
3200 	RPCLOG(1, " received a notifcation prim type [%s]",
3201 	    rpc_tpiprim2name(msg_type));
3202 	RPCLOG(1, " and reason %d\n", reason);
3203 
3204 	/*
3205 	 * Find the transport entry in the connection manager's list, close
3206 	 * the transport and delete the entry.  In the case where rpcmod's
3207 	 * idle timer goes off, it sends us a T_ORDREL_REQ, indicating we
3208 	 * should gracefully close the connection.
3209 	 */
3210 	have_connmgr_lock = 1;
3211 	mutex_enter(&connmgr_lock);
3212 	for (cm_entry = cm_hd; cm_entry; cm_entry = cm_entry->x_next) {
3213 		ASSERT(cm_entry != cm_entry->x_next);
3214 		if (cm_entry->x_wq == q) {
3215 			ASSERT(MUTEX_HELD(&connmgr_lock));
3216 			ASSERT(have_connmgr_lock == 1);
3217 			switch (msg_type) {
3218 			case T_ORDREL_REQ:
3219 
3220 				if (cm_entry->x_dead) {
3221 					RPCLOG(1, "idle timeout on dead "
3222 					    "connection: %p\n",
3223 					    (void *)cm_entry);
3224 					if (clnt_stop_idle != NULL)
3225 						(*clnt_stop_idle)(q);
3226 					break;
3227 				}
3228 
3229 				/*
3230 				 * Only mark the connection as dead if it is
3231 				 * connected and idle.
3232 				 * An unconnected connection has probably
3233 				 * gone idle because the server is down,
3234 				 * and when it comes back up there will be
3235 				 * retries that need to use that connection.
3236 				 */
3237 				if (cm_entry->x_connected ||
3238 				    cm_entry->x_doomed) {
3239 					if (cm_entry->x_ordrel) {
3240 						if (cm_entry->x_closing ==
3241 						    TRUE) {
3242 							/*
3243 							 * The connection is
3244 							 * obviously wedged due
3245 							 * to a bug or problem
3246 							 * with the transport.
3247 							 * Mark it as dead.
3248 							 * Otherwise we can
3249 							 * leak connections.
3250 							 */
3251 							cm_entry->x_dead = TRUE;
3252 							mutex_exit(
3253 							    &connmgr_lock);
3254 							have_connmgr_lock = 0;
3255 							if (clnt_stop_idle !=
3256 							    NULL)
3257 							(*clnt_stop_idle)(q);
3258 							break;
3259 						}
3260 						cm_entry->x_closing = TRUE;
3261 						connmgr_sndrel(cm_entry);
3262 						have_connmgr_lock = 0;
3263 					} else {
3264 						cm_entry->x_dead = TRUE;
3265 						mutex_exit(&connmgr_lock);
3266 						have_connmgr_lock = 0;
3267 						if (clnt_stop_idle != NULL)
3268 							(*clnt_stop_idle)(q);
3269 					}
3270 				} else {
3271 					/*
3272 					 * We don't mark the connection
3273 					 * as dead, but we turn off the
3274 					 * idle timer.
3275 					 */
3276 					mutex_exit(&connmgr_lock);
3277 					have_connmgr_lock = 0;
3278 					if (clnt_stop_idle != NULL)
3279 						(*clnt_stop_idle)(q);
3280 					RPCLOG(1, "clnt_dispatch_notifyall:"
3281 					    " ignoring timeout from rpcmod"
3282 					    " (q %p) because we are not "
3283 					    " connected\n", (void *)q);
3284 				}
3285 				break;
3286 			case T_ORDREL_IND:
3287 				/*
3288 				 * If this entry is marked closing, then we are
3289 				 * completing a close handshake, and the
3290 				 * connection is dead.  Otherwise, the server is
3291 				 * trying to close. Since the server will not
3292 				 * be sending any more RPC replies, we abort
3293 				 * the connection, including flushing
3294 				 * any RPC requests that are in-transit.
3295 				 * In either case, mark the entry as dead so
3296 				 * that it can be closed by the connection
3297 				 * manager's garbage collector.
3298 				 */
3299 				cm_entry->x_dead = TRUE;
3300 				if (cm_entry->x_closing) {
3301 					mutex_exit(&connmgr_lock);
3302 					have_connmgr_lock = 0;
3303 					if (clnt_stop_idle != NULL)
3304 						(*clnt_stop_idle)(q);
3305 				} else {
3306 					/*
3307 					 * if we're getting a disconnect
3308 					 * before we've finished our
3309 					 * connect attempt, mark it for
3310 					 * later processing
3311 					 */
3312 					if (cm_entry->x_thread)
3313 						cm_entry->x_early_disc = TRUE;
3314 					else
3315 						cm_entry->x_connected = FALSE;
3316 					cm_entry->x_waitdis = TRUE;
3317 					connmgr_snddis(cm_entry);
3318 					have_connmgr_lock = 0;
3319 				}
3320 				break;
3321 
3322 			case T_ERROR_ACK:
3323 			case T_OK_ACK:
3324 				cm_entry->x_waitdis = FALSE;
3325 				cv_signal(&cm_entry->x_dis_cv);
3326 				mutex_exit(&connmgr_lock);
3327 				return;
3328 
3329 			case T_DISCON_REQ:
3330 				if (cm_entry->x_thread)
3331 					cm_entry->x_early_disc = TRUE;
3332 				else
3333 					cm_entry->x_connected = FALSE;
3334 				cm_entry->x_waitdis = TRUE;
3335 
3336 				connmgr_snddis(cm_entry);
3337 				have_connmgr_lock = 0;
3338 				break;
3339 
3340 			case T_DISCON_IND:
3341 			default:
3342 				/*
3343 				 * if we're getting a disconnect before
3344 				 * we've finished our connect attempt,
3345 				 * mark it for later processing
3346 				 */
3347 				if (cm_entry->x_closing) {
3348 					cm_entry->x_dead = TRUE;
3349 					mutex_exit(&connmgr_lock);
3350 					have_connmgr_lock = 0;
3351 					if (clnt_stop_idle != NULL)
3352 						(*clnt_stop_idle)(q);
3353 				} else {
3354 					if (cm_entry->x_thread) {
3355 						cm_entry->x_early_disc = TRUE;
3356 					} else {
3357 						cm_entry->x_dead = TRUE;
3358 						cm_entry->x_connected = FALSE;
3359 					}
3360 				}
3361 				break;
3362 			}
3363 			break;
3364 		}
3365 	}
3366 
3367 	if (have_connmgr_lock)
3368 		mutex_exit(&connmgr_lock);
3369 
3370 	if (msg_type == T_ERROR_ACK || msg_type == T_OK_ACK) {
3371 		RPCLOG(1, "clnt_dispatch_notifyall: (wq %p) could not find "
3372 		    "connmgr entry for discon ack\n", (void *)q);
3373 		return;
3374 	}
3375 
3376 	/*
3377 	 * Then kick all the clnt_pending calls out of their wait.  There
3378 	 * should be no clnt_pending calls in the case of rpcmod's idle
3379 	 * timer firing.
3380 	 */
3381 	for (i = 0; i < clnt_cots_hash_size; i++) {
3382 		ctp = &cots_call_ht[i];
3383 		mutex_enter(&ctp->ct_lock);
3384 		for (e = ctp->ct_call_next;
3385 		    e != (calllist_t *)ctp;
3386 		    e = e->call_next) {
3387 			if (e->call_wq == q && e->call_notified == FALSE) {
3388 				RPCLOG(1,
3389 				    "clnt_dispatch_notifyall for queue %p ",
3390 				    (void *)q);
3391 				RPCLOG(1, "aborting clnt_pending call %p\n",
3392 				    (void *)e);
3393 
3394 				if (msg_type == T_DISCON_IND)
3395 					e->call_reason = reason;
3396 				e->call_notified = TRUE;
3397 				e->call_status = RPC_XPRTFAILED;
3398 				cv_signal(&e->call_cv);
3399 			}
3400 		}
3401 		mutex_exit(&ctp->ct_lock);
3402 	}
3403 
3404 	mutex_enter(&clnt_pending_lock);
3405 	for (e = clnt_pending; e; e = e->call_next) {
3406 		/*
3407 		 * Only signal those RPC handles that haven't been
3408 		 * signalled yet. Otherwise we can get a bogus call_reason.
3409 		 * This can happen if thread A is making a call over a
3410 		 * connection. If the server is killed, it will cause
3411 		 * reset, and reason will default to EIO as a result of
3412 		 * a T_ORDREL_IND. Thread B then attempts to recreate
3413 		 * the connection but gets a T_DISCON_IND. If we set the
3414 		 * call_reason code for all threads, then if thread A
3415 		 * hasn't been dispatched yet, it will get the wrong
3416 		 * reason. The bogus call_reason can make it harder to
3417 		 * discriminate between calls that fail because the
3418 		 * connection attempt failed versus those where the call
3419 		 * may have been executed on the server.
3420 		 */
3421 		if (e->call_wq == q && e->call_notified == FALSE) {
3422 			RPCLOG(1, "clnt_dispatch_notifyall for queue %p ",
3423 			    (void *)q);
3424 			RPCLOG(1, " aborting clnt_pending call %p\n",
3425 			    (void *)e);
3426 
3427 			if (msg_type == T_DISCON_IND)
3428 				e->call_reason = reason;
3429 			e->call_notified = TRUE;
3430 			/*
3431 			 * Let the caller timeout, else he will retry
3432 			 * immediately.
3433 			 */
3434 			e->call_status = RPC_XPRTFAILED;
3435 
3436 			/*
3437 			 * We used to just signal those threads
3438 			 * waiting for a connection, (call_xid = 0).
3439 			 * That meant that threads waiting for a response
3440 			 * waited till their timeout expired. This
3441 			 * could be a long time if they've specified a
3442 			 * maximum timeout. (2^31 - 1). So we
3443 			 * Signal all threads now.
3444 			 */
3445 			cv_signal(&e->call_cv);
3446 		}
3447 	}
3448 	mutex_exit(&clnt_pending_lock);
3449 }
3450 
3451 
3452 /*ARGSUSED*/
3453 /*
3454  * after resuming a system that's been suspended for longer than the
3455  * NFS server's idle timeout (svc_idle_timeout for Solaris 2), rfscall()
3456  * generates "NFS server X not responding" and "NFS server X ok" messages;
3457  * here we reset inet connections to cause a re-connect and avoid those
3458  * NFS messages.  see 4045054
3459  */
3460 boolean_t
3461 connmgr_cpr_reset(void *arg, int code)
3462 {
3463 	struct cm_xprt *cxp;
3464 
3465 	if (code == CB_CODE_CPR_CHKPT)
3466 		return (B_TRUE);
3467 
3468 	if (mutex_tryenter(&connmgr_lock) == 0)
3469 		return (B_FALSE);
3470 	for (cxp = cm_hd; cxp; cxp = cxp->x_next) {
3471 		if ((cxp->x_family == AF_INET || cxp->x_family == AF_INET6) &&
3472 		    cxp->x_connected == TRUE) {
3473 			if (cxp->x_thread)
3474 				cxp->x_early_disc = TRUE;
3475 			else
3476 				cxp->x_connected = FALSE;
3477 			cxp->x_needdis = TRUE;
3478 		}
3479 	}
3480 	mutex_exit(&connmgr_lock);
3481 	return (B_TRUE);
3482 }
3483 
3484 void
3485 clnt_cots_stats_init(zoneid_t zoneid, struct rpc_cots_client **statsp)
3486 {
3487 
3488 	*statsp = (struct rpc_cots_client *)rpcstat_zone_init_common(zoneid,
3489 	    "unix", "rpc_cots_client", (const kstat_named_t *)&cots_rcstat_tmpl,
3490 	    sizeof (cots_rcstat_tmpl));
3491 }
3492 
3493 void
3494 clnt_cots_stats_fini(zoneid_t zoneid, struct rpc_cots_client **statsp)
3495 {
3496 	rpcstat_zone_fini_common(zoneid, "unix", "rpc_cots_client");
3497 	kmem_free(*statsp, sizeof (cots_rcstat_tmpl));
3498 }
3499 
3500 void
3501 clnt_cots_init(void)
3502 {
3503 	mutex_init(&connmgr_lock, NULL, MUTEX_DEFAULT, NULL);
3504 	mutex_init(&clnt_pending_lock, NULL, MUTEX_DEFAULT, NULL);
3505 
3506 	if (clnt_cots_hash_size < DEFAULT_MIN_HASH_SIZE)
3507 		clnt_cots_hash_size = DEFAULT_MIN_HASH_SIZE;
3508 
3509 	cots_call_ht = call_table_init(clnt_cots_hash_size);
3510 	zone_key_create(&zone_cots_key, NULL, NULL, clnt_zone_destroy);
3511 }
3512 
3513 void
3514 clnt_cots_fini(void)
3515 {
3516 	(void) zone_key_delete(zone_cots_key);
3517 }
3518 
3519 /*
3520  * Wait for TPI ack, returns success only if expected ack is received
3521  * within timeout period.
3522  */
3523 
3524 static int
3525 waitforack(calllist_t *e, t_scalar_t ack_prim, const struct timeval *waitp,
3526     bool_t nosignal)
3527 {
3528 	union T_primitives *tpr;
3529 	clock_t timout;
3530 	int cv_stat = 1;
3531 
3532 	ASSERT(MUTEX_HELD(&clnt_pending_lock));
3533 	while (e->call_reply == NULL) {
3534 		if (waitp != NULL) {
3535 			timout = waitp->tv_sec * drv_usectohz(MICROSEC) +
3536 			    drv_usectohz(waitp->tv_usec) + lbolt;
3537 			if (nosignal)
3538 				cv_stat = cv_timedwait(&e->call_cv,
3539 				    &clnt_pending_lock, timout);
3540 			else
3541 				cv_stat = cv_timedwait_sig(&e->call_cv,
3542 				    &clnt_pending_lock, timout);
3543 		} else {
3544 			if (nosignal)
3545 				cv_wait(&e->call_cv, &clnt_pending_lock);
3546 			else
3547 				cv_stat = cv_wait_sig(&e->call_cv,
3548 				    &clnt_pending_lock);
3549 		}
3550 		if (cv_stat == -1)
3551 			return (ETIME);
3552 		if (cv_stat == 0)
3553 			return (EINTR);
3554 		/*
3555 		 * if we received an error from the server and we know a reply
3556 		 * is not going to be sent, do not wait for the full timeout,
3557 		 * return now.
3558 		 */
3559 		if (e->call_status == RPC_XPRTFAILED)
3560 			return (e->call_reason);
3561 	}
3562 	tpr = (union T_primitives *)e->call_reply->b_rptr;
3563 	if (tpr->type == ack_prim)
3564 		return (0); /* Success */
3565 
3566 	if (tpr->type == T_ERROR_ACK) {
3567 		if (tpr->error_ack.TLI_error == TSYSERR)
3568 			return (tpr->error_ack.UNIX_error);
3569 		else
3570 			return (t_tlitosyserr(tpr->error_ack.TLI_error));
3571 	}
3572 
3573 	return (EPROTO); /* unknown or unexpected primitive */
3574 }
3575