xref: /titanic_52/usr/src/lib/libnsl/rpc/svc_run.c (revision ea8dc4b6d2251b437950c0056bc626b311c73c27)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 
23 /*
24  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
25  * Use is subject to license terms.
26  */
27 
28 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
29 /* All Rights Reserved */
30 /*
31  * Portions of this source code were derived from Berkeley
32  * 4.3 BSD under license from the Regents of the University of
33  * California.
34  */
35 
36 #pragma ident	"%Z%%M%	%I%	%E% SMI"
37 
38 /*
39  * This is the rpc server side idle loop
40  * Wait for input, call server program.
41  */
42 #include "mt.h"
43 #include "rpc_mt.h"
44 #include <stdlib.h>
45 #include <unistd.h>
46 #include <signal.h>
47 #include <rpc/rpc.h>
48 #include <errno.h>
49 #include <sys/poll.h>
50 #include <sys/types.h>
51 #include <syslog.h>
52 #include <thread.h>
53 #include <assert.h>
54 #include <libintl.h>
55 #include <values.h>
56 
57 extern const char __nsl_dom[];
58 
59 extern int __rpc_compress_pollfd(int, pollfd_t *, pollfd_t *);
60 extern bool_t __is_a_userfd(int P_fd);
61 extern void   __destroy_userfd();
62 extern void clear_pollfd(int);
63 extern void set_pollfd(int /* fd */, short /* events */);
64 extern void svc_getreq_poll();
65 extern void (*__proc_cleanup_cb)();
66 
67 static void start_threads();
68 static void create_pipe();
69 static void clear_pipe();
70 static int select_next_pollfd();
71 static SVCXPRT *make_xprt_copy();
72 static void _svc_run_mt();
73 static void _svc_run();
74 
75 int _svc_prog_dispatch();
76 static void _svc_done_private();
77 
78 extern rwlock_t svc_fd_lock;
79 extern mutex_t	svc_door_mutex;
80 extern cond_t	svc_door_waitcv;
81 extern int	svc_ndoorfds;
82 extern void	__svc_cleanup_door_xprts();
83 extern void	__svc_free_xprtlist();
84 extern void	__svc_getreq_user(struct pollfd *);
85 
86 /*
87  * Maximum fragment size allowed for connection oriented connections.
88  * Zero means that no maximum size limit has been requested.
89  */
90 int __rpc_connmaxrec = 0;
91 
92 /* Inter-Record Timeout in secs for non-blocked connection RPC */
93 int __rpc_irtimeout = 35;
94 
95 /*
96  * Request exclusive access to tcp and udp non-priv ports bound with a
97  * wildcard addr.
98  */
99 bool_t __rpc_tp_exclbind = FALSE;
100 
101 /*
102  * XXX - eventually, all mutexes and their initializations static
103  */
104 
105 /*
106  * Variables used for MT
107  */
108 int svc_mt_mode;		/* multi-threading mode */
109 
110 int svc_pipe[2];	/* pipe for breaking out of poll: read(0), write(1) */
111 
112 /* BEGIN PROTECTED BY svc_mutex */
113 
114 static int svc_thr_max = 16;	/* default maximum number of threads allowed */
115 
116 static int svc_thr_total;	/* current number of threads */
117 
118 static int svc_thr_active;	/* current number of threads active */
119 
120 /* circular array of file descriptors with pending data */
121 
122 #define	CIRCULAR_BUFSIZE	1024
123 
124 static int svc_pending_fds[CIRCULAR_BUFSIZE+1];	/* fds with pending data */
125 
126 static int svc_next_pending;			/* next one to be processed */
127 
128 static int svc_last_pending;			/* last one in list */
129 
130 static int svc_total_pending;			/* total in list */
131 
132 static int svc_thr_total_creates;	/* total created - stats */
133 
134 static int svc_thr_total_create_errors;	/* total create errors - stats */
135 
136 static int svc_waiters;		/* number of waiting threads */
137 
138 /* END PROTECTED BY svc_mutex */
139 
140 /* BEGIN PROTECTED BY svc_fd_lock: */
141 
142 int svc_nfds;		/* total number of active file descriptors */
143 
144 int svc_nfds_set;	/* total number of fd bits set in svc_fdset */
145 
146 int svc_max_fd = 0;	/* largest active file descriptor */
147 
148 int svc_npollfds;	/* total number of active pollfds */
149 
150 int svc_npollfds_set;	/* total number of pollfd set in svc_pollfd */
151 
152 int svc_max_pollfd;	/* largest active pollfd so far */
153 
154 int svc_pollfd_allocd;  /* number of pollfd structures allocated */
155 
156 /* END PROTECTED BY svc_fd_lock: */
157 
158 /* BEGIN PROTECTED BY svc_thr_mutex */
159 
160 #define	POLLSET_EXTEND	256
161 static int svc_pollset_allocd;
162 static struct pollfd *svc_pollset;
163 				/*
164 				 * array of file descriptors currently active
165 				 */
166 static int svc_polled;		/* no of fds polled in last poll() - input */
167 
168 static int svc_pollfds;		/* no of active fds in last poll() - output */
169 
170 static int svc_next_pollfd;	/* next fd  to processin svc_pollset */
171 
172 bool_t svc_polling;		/* true if a thread is polling */
173 
174 /* END PROTECTED BY svc_thr_mutex */
175 
176 /* BEGIN PROTECTED BY svc_exit_mutex */
177 
178 static bool_t svc_exit_done = TRUE;
179 
180 /* END PROTECTED BY svc_exit_mutex */
181 
182 /*
183  * Warlock section
184  */
185 
186 /* VARIABLES PROTECTED BY svc_mutex:
187 	svc_thr_total, svc_thr_active, svc_pending_fds, svc_next_pending,
188 	svc_last_pending, svc_total_pending, svc_thr_total_creates,
189 	svc_thr_total_create_errors,
190 	svcxprt_list_t::next, svcxprt_ext_t::my_xlist,
191 	svc_thr_max, svc_waiters
192  */
193 
194 /* VARIABLES PROTECTED BY svc_fd_lock:
195 	svc_xports, svc_fdset, svc_nfds, svc_nfds_set, svc_max_fd,
196 	svc_pollfd, svc_npollfds, svc_npollfds_set, svc_max_pollfd
197  */
198 
199 /* VARIABLES PROTECTED BY svc_thr_mutex:
200 	svc_pollset, svc_pollfds, svc_next_pollfd, svc_polling
201 	svc_pollset_allocd, svc_polled
202  */
203 
204 /* VARIABLES PROTECTED BY svc_exit_mutex:
205 	svc_exit_done
206  */
207 
208 /* VARIABLES READABLE WITHOUT LOCK:
209 	svc_thr_total, svc_thr_active, svc_thr_total_creates,
210 	svc_thr_total_create_errors,
211 	svc_xports, svc_nfds, svc_nfds_set, svc_max_fd,
212 	svc_npollfds, svc_npollfds_set, svc_max_pollfd,
213 	svc_pollfds, svc_next_pollfd, svc_exit_done, svc_polling,
214 	svc_thr_max, svc_waiters
215  */
216 
217 /* VARIABLES PROTECTED BY "program_logic":
218 	rpc_msg::, svc_req::, svcxprt_ext_t::flags, svc_mt_mode,
219 	svcxprt_ext_t::parent
220  */
221 
222 /* LOCK ORDER:
223 	svc_exit_mutex, svc_thr_mutex, svc_mutex, svc_fd_lock
224  */
225 
226 
227 void
228 svc_run(void)
229 {
230 	/* NO OTHER THREADS ARE RUNNING */
231 
232 	svc_exit_done = FALSE;
233 
234 	while ((svc_npollfds > 0 || svc_ndoorfds > 0) && !svc_exit_done) {
235 		if (svc_npollfds > 0) {
236 			switch (svc_mt_mode) {
237 			case RPC_SVC_MT_NONE:
238 				_svc_run();
239 				break;
240 			default:
241 				_svc_run_mt();
242 				break;
243 			}
244 			continue;
245 		}
246 
247 		(void) mutex_lock(&svc_door_mutex);
248 		if (svc_ndoorfds > 0)
249 			(void) cond_wait(&svc_door_waitcv, &svc_door_mutex);
250 		(void) mutex_unlock(&svc_door_mutex);
251 	}
252 }
253 
254 
255 /*
256  *	This function causes svc_run() to exit by destroying all
257  *	service handles.
258  */
259 void
260 svc_exit(void)
261 {
262 	SVCXPRT	*xprt;
263 	int fd;
264 	char dummy;
265 
266 	/* NO LOCKS HELD */
267 
268 	(void) mutex_lock(&svc_exit_mutex);
269 	if (svc_exit_done) {
270 		(void) mutex_unlock(&svc_exit_mutex);
271 		return;
272 	}
273 	svc_exit_done = TRUE;
274 	for (fd = 0; fd < svc_max_pollfd; fd++) {
275 		xprt = svc_xports[fd];
276 		if (xprt) {
277 			SVC_DESTROY(xprt);
278 		}
279 	}
280 	__svc_free_xprtlist();
281 	__svc_cleanup_door_xprts();
282 	(void) mutex_unlock(&svc_exit_mutex);
283 
284 	if (svc_mt_mode != RPC_SVC_MT_NONE) {
285 		(void) mutex_lock(&svc_mutex);
286 		(void) cond_broadcast(&svc_thr_fdwait);
287 		(void) mutex_unlock(&svc_mutex);
288 
289 		(void) write(svc_pipe[1], &dummy, sizeof (dummy));
290 	}
291 
292 	(void) mutex_lock(&svc_door_mutex);
293 	(void) cond_signal(&svc_door_waitcv);	/* wake up door dispatching */
294 	(void) mutex_unlock(&svc_door_mutex);
295 
296 	/* destroy reactor information if any */
297 	__destroy_userfd();
298 }
299 
300 
301 /*
302  * this funtion is called with svc_fd_lock and svc_thr_mutex
303  */
304 
305 static int
306 alloc_pollset(int npollfds)
307 {
308 	if (npollfds > svc_pollset_allocd) {
309 		pollfd_t *tmp;
310 		do {
311 			svc_pollset_allocd += POLLSET_EXTEND;
312 		} while (npollfds > svc_pollset_allocd);
313 		tmp = realloc(svc_pollset,
314 				sizeof (pollfd_t) * svc_pollset_allocd);
315 		if (tmp == NULL) {
316 			syslog(LOG_ERR, "alloc_pollset: out of memory");
317 			return (-1);
318 		}
319 		svc_pollset = tmp;
320 	}
321 	return (0);
322 }
323 
324 static void
325 _svc_run(void)
326 {
327 	sigset_t set, oldset;
328 	int npollfds;
329 	int i;
330 
331 	/*
332 	 * Block SIGALRM while doing work.  Unblock it while doing poll().
333 	 * This is so that services like rpc.rstatd can cause the poll()
334 	 * to be interrupted due to alarm() but that we don't end up in
335 	 * an MT-unsafe signal handler at an inopportune time.
336 	 */
337 	(void) sigemptyset(&set);
338 	(void) sigaddset(&set, SIGALRM);
339 	(void) sigprocmask(SIG_BLOCK, &set, &oldset);
340 	while (!svc_exit_done) {
341 		/*
342 		 * Check whether there is any server fd on which we may want
343 		 * to wait.
344 		 */
345 		(void) rw_rdlock(&svc_fd_lock);
346 		if (alloc_pollset(svc_npollfds) == -1)
347 			break;
348 		npollfds = __rpc_compress_pollfd(svc_max_pollfd,
349 			svc_pollfd, svc_pollset);
350 		(void) rw_unlock(&svc_fd_lock);
351 		if (npollfds == 0)
352 			break;	/* None waiting, hence return */
353 
354 		(void) sigprocmask(SIG_SETMASK, &oldset, NULL);
355 		i = poll(svc_pollset, npollfds, -1);
356 		(void) sigprocmask(SIG_BLOCK, &set, &oldset);
357 		switch (i) {
358 		case -1:
359 			/*
360 			 * We ignore all errors, continuing with the assumption
361 			 * that it was set by the signal handlers (or any
362 			 * other outside event) and not caused by poll().
363 			 */
364 		case 0:
365 			continue;
366 		default:
367 			svc_getreq_poll(svc_pollset, i);
368 		}
369 	}
370 	(void) sigprocmask(SIG_SETMASK, &oldset, NULL);
371 }
372 
373 /*
374  * In _svc_run_mt, myfd is linked with mypollfd
375  * svc_pollset[mypollfd].fd == myfd
376  * However, in some cases, the link can not be made, thus we define the
377  * following values for these special cases
378  */
379 enum {
380 	INVALID_POLLFD	= -200,
381 	FD_FROM_PENDING
382 };
383 
384 static void
385 _svc_run_mt(void)
386 {
387 	int npollfds;
388 	int n_polled, dispatch;
389 
390 	static bool_t first_time = TRUE;
391 	bool_t main_thread = FALSE;
392 	int n_new;
393 	int myfd, mypollfd;
394 	SVCXPRT *parent_xprt, *xprt;
395 
396 	/*
397 	 * Server is multi-threaded.  Do "first time" initializations.
398 	 * Since only one thread exists in the beginning, there's no
399 	 * need for mutex protection for first time initializations.
400 	 */
401 	if (first_time) {
402 		first_time = FALSE;
403 		main_thread = TRUE;
404 		svc_thr_total = 1;	/* this thread */
405 		svc_next_pending = svc_last_pending = 0;
406 
407 		/*
408 		 * Create a pipe for waking up the poll, if new
409 		 * descriptors have been added to svc_fdset.
410 		 */
411 		create_pipe();
412 	}
413 
414 	/* OTHER THREADS ARE RUNNING */
415 
416 	if (svc_exit_done)
417 		return;
418 
419 	for (;;) {
420 		/*
421 		 * svc_thr_mutex prevents more than one thread from
422 		 * trying to select a descriptor to process further.
423 		 * svc_thr_mutex is unlocked after a thread selects
424 		 * a descriptor on which to receive data.  If there are
425 		 * no such descriptors, the thread will poll with
426 		 * svc_thr_mutex locked, after unlocking all other
427 		 * locks.  This prevents more than one thread from
428 		 * trying to poll at the same time.
429 		 */
430 		(void) mutex_lock(&svc_thr_mutex);
431 		(void) mutex_lock(&svc_mutex);
432 continue_with_locks:
433 		myfd = -1;
434 		mypollfd = INVALID_POLLFD;
435 
436 		/*
437 		 * Check if there are any descriptors with data pending.
438 		 */
439 		if (svc_total_pending > 0) {
440 			myfd = svc_pending_fds[svc_next_pending++];
441 			mypollfd = FD_FROM_PENDING;
442 			if (svc_next_pending > CIRCULAR_BUFSIZE)
443 				svc_next_pending = 0;
444 			svc_total_pending--;
445 		}
446 
447 		/*
448 		 * Get the next active file descriptor to process.
449 		 */
450 		if (myfd == -1 && svc_pollfds == 0) {
451 			/*
452 			 * svc_pollset is empty; do polling
453 			 */
454 			svc_polling = TRUE;
455 
456 			/*
457 			 * if there are no file descriptors, return
458 			 */
459 			(void) rw_rdlock(&svc_fd_lock);
460 			if (svc_npollfds == 0 ||
461 					alloc_pollset(svc_npollfds + 1) == -1) {
462 				(void) rw_unlock(&svc_fd_lock);
463 				svc_polling = FALSE;
464 				svc_thr_total--;
465 				(void) mutex_unlock(&svc_mutex);
466 				(void) mutex_unlock(&svc_thr_mutex);
467 				if (!main_thread) {
468 					thr_exit(NULL);
469 					/* NOTREACHED */
470 				}
471 				break;
472 			}
473 
474 			npollfds = __rpc_compress_pollfd(svc_max_pollfd,
475 					svc_pollfd, svc_pollset);
476 			(void) rw_unlock(&svc_fd_lock);
477 
478 			if (npollfds == 0) {
479 				/*
480 				 * There are file descriptors, but none of them
481 				 * are available for polling.  If this is the
482 				 * main thread, or if no thread is waiting,
483 				 * wait on condition variable, otherwise exit.
484 				 */
485 				svc_polling = FALSE;
486 				(void) mutex_unlock(&svc_thr_mutex);
487 				if ((!main_thread) && svc_waiters > 0) {
488 					svc_thr_total--;
489 					(void) mutex_unlock(&svc_mutex);
490 					thr_exit(NULL);
491 					/* NOTREACHED */
492 				}
493 
494 				while (svc_npollfds_set == 0 &&
495 					svc_pollfds == 0 &&
496 					svc_total_pending == 0 &&
497 							!svc_exit_done) {
498 					svc_waiters++;
499 					(void) cond_wait(&svc_thr_fdwait,
500 								&svc_mutex);
501 					svc_waiters--;
502 				}
503 
504 				/*
505 				 * Check exit flag.  If this is not the main
506 				 * thread, exit.
507 				 */
508 				if (svc_exit_done) {
509 					svc_thr_total--;
510 					(void) mutex_unlock(&svc_mutex);
511 					if (!main_thread)
512 						thr_exit(NULL);
513 					break;
514 				}
515 
516 				(void) mutex_unlock(&svc_mutex);
517 				continue;
518 			}
519 
520 			/*
521 			 * We're ready to poll.  Always set svc_pipe[0]
522 			 * as the last one, since the poll will occasionally
523 			 * need to be interrupted.  Release svc_mutex for
524 			 * the duration of the poll, but hold on to
525 			 * svc_thr_mutex, as we don't want any other thread
526 			 * to do the same.
527 			 */
528 			svc_pollset[npollfds].fd = svc_pipe[0];
529 			svc_pollset[npollfds].events = MASKVAL;
530 
531 			do {
532 				int i, j;
533 
534 				(void) mutex_unlock(&svc_mutex);
535 				n_polled = poll(svc_pollset, npollfds + 1, -1);
536 				(void) mutex_lock(&svc_mutex);
537 				if (n_polled <= 0)
538 					continue;
539 
540 				/*
541 				 * Check if information returned indicates one
542 				 * or more closed fd's; find and remove any such
543 				 * information
544 				 */
545 				for (i = 0; i <= npollfds; i++) {
546 					if (svc_pollset[i].revents & POLLNVAL) {
547 						/* Overwrite svc_pollset[i] */
548 						for (j = i; j < npollfds; j++)
549 							svc_pollset[j] =
550 							    svc_pollset[j + 1];
551 						(void) memset(&svc_pollset[j],
552 						    0, sizeof (struct pollfd));
553 						npollfds--;
554 						n_polled--;
555 						i--;
556 					}
557 				}
558 			} while (n_polled <= 0);
559 			svc_polling = FALSE;
560 
561 			/*
562 			 * If there's data in the pipe, clear it.
563 			 */
564 			if (svc_pollset[npollfds].revents) {
565 				clear_pipe();
566 				n_polled--;
567 				svc_pollset[npollfds].revents = 0;
568 			}
569 			svc_polled = npollfds;
570 			svc_pollfds = n_polled;
571 			svc_next_pollfd = 0;
572 
573 			/*
574 			 * Check exit flag.
575 			 */
576 			if (svc_exit_done) {
577 				svc_thr_total--;
578 				(void) mutex_unlock(&svc_mutex);
579 				(void) mutex_unlock(&svc_thr_mutex);
580 				if (!main_thread) {
581 					thr_exit(NULL);
582 					/* NOTREACHED */
583 				}
584 				break;
585 			}
586 
587 			/*
588 			 * If no descriptor is active, continue.
589 			 */
590 			if (svc_pollfds == 0)
591 				goto continue_with_locks;
592 		}
593 
594 		/*
595 		 * If a file descriptor has already not been selected,
596 		 * choose a file descriptor.
597 		 * svc_pollfds and svc_next_pollfd are updated.
598 		 */
599 		if (myfd == -1) {
600 			if (select_next_pollfd(&myfd, &mypollfd) == -1)
601 				goto continue_with_locks;
602 		}
603 
604 		/*
605 		 * Check to see if new threads need to be started.
606 		 * Count of threads that could be gainfully employed is
607 		 * obtained as follows:
608 		 *	- count 1 for poller
609 		 *	- count 1 for this request
610 		 *	- count active file descriptors (svc_pollfds)
611 		 *	- count pending file descriptors
612 		 *
613 		 * (svc_thr_total - svc_thr_active) are already available.
614 		 * This thread is one of the available threads.
615 		 *
616 		 * Number of new threads should not exceed
617 		 *	(svc_thr_max - svc_thr_total).
618 		 */
619 		if (svc_thr_total < svc_thr_max &&
620 			    svc_mt_mode == RPC_SVC_MT_AUTO && !svc_exit_done) {
621 			n_new = 1 + 1 + svc_pollfds + svc_total_pending -
622 					(svc_thr_total - svc_thr_active);
623 			if (n_new > (svc_thr_max - svc_thr_total))
624 				n_new = svc_thr_max - svc_thr_total;
625 			if (n_new > 0)
626 				start_threads(n_new);
627 		}
628 
629 		/*
630 		 * Get parent xprt.  It is possible for the parent service
631 		 * handle to be destroyed by now, due to a race condition.
632 		 * Check for this, and if so, log a warning and go on.
633 		 */
634 		parent_xprt = svc_xports[myfd];
635 		if (parent_xprt == NULL) {
636 			/* Check if it is not a user FD */
637 			if (__is_a_userfd(myfd) == TRUE)
638 				__svc_getreq_user(&(svc_pollset[mypollfd]));
639 			goto continue_with_locks;
640 		}
641 /* LINTED pointer alignment */
642 		if (svc_defunct(parent_xprt) || svc_failed(parent_xprt))
643 			goto continue_with_locks;
644 
645 		/*
646 		 * Make a copy of parent xprt, update svc_fdset.
647 		 */
648 		if ((xprt = make_xprt_copy(parent_xprt)) == NULL)
649 			goto continue_with_locks;
650 
651 		/*
652 		 * Keep track of active threads in automatic mode.
653 		 */
654 		if (svc_mt_mode == RPC_SVC_MT_AUTO)
655 			svc_thr_active++;
656 
657 		/*
658 		 * Release mutexes so other threads can get going.
659 		 */
660 		(void) mutex_unlock(&svc_mutex);
661 		(void) mutex_unlock(&svc_thr_mutex);
662 
663 		/*
664 		 * Process request.
665 		 */
666 		{
667 			struct rpc_msg *msg;
668 			struct svc_req *r;
669 			char *cred_area;
670 
671 /* LINTED pointer alignment */
672 			msg = SVCEXT(xprt)->msg;
673 /* LINTED pointer alignment */
674 			r = SVCEXT(xprt)->req;
675 /* LINTED pointer alignment */
676 			cred_area = SVCEXT(xprt)->cred_area;
677 
678 
679 			msg->rm_call.cb_cred.oa_base = cred_area;
680 			msg->rm_call.cb_verf.oa_base =
681 						&(cred_area[MAX_AUTH_BYTES]);
682 			r->rq_clntcred = &(cred_area[2 * MAX_AUTH_BYTES]);
683 
684 			/*
685 			 * receive RPC message
686 			 */
687 			if ((dispatch = SVC_RECV(xprt, msg))) {
688 				if (svc_mt_mode != RPC_SVC_MT_NONE)
689 /* LINTED pointer alignment */
690 					svc_flags(xprt) |= SVC_ARGS_CHECK;
691 				dispatch = _svc_prog_dispatch(xprt, msg, r);
692 
693 				/*
694 				 * Call cleanup procedure if set.
695 				 */
696 				if (__proc_cleanup_cb != NULL)
697 					(*__proc_cleanup_cb)(xprt);
698 			} else
699 				svc_args_done(xprt);
700 
701 			/*
702 			 * Finish up, if automatic mode, or not dispatched.
703 			 */
704 			if (svc_mt_mode == RPC_SVC_MT_AUTO || !dispatch) {
705 /* LINTED pointer alignment */
706 				if (svc_flags(xprt) & SVC_ARGS_CHECK)
707 					svc_args_done(xprt);
708 				(void) mutex_lock(&svc_mutex);
709 				_svc_done_private(xprt);
710 				if (svc_mt_mode == RPC_SVC_MT_AUTO) {
711 					/*
712 					 * not active any more
713 					 */
714 					svc_thr_active--;
715 
716 					/*
717 					 * If not main thread, exit unless
718 					 * there's some immediate work.
719 					 */
720 					if (!main_thread &&
721 						    svc_pollfds <= 0 &&
722 						    svc_total_pending <= 0 &&
723 						    (svc_polling ||
724 							svc_waiters > 0)) {
725 						svc_thr_total--;
726 						if (svc_thr_total ==
727 						    svc_waiters) {
728 							(void) cond_broadcast(
729 							    &svc_thr_fdwait);
730 						}
731 						(void) mutex_unlock(&svc_mutex);
732 						thr_exit(NULL);
733 						/* NOTREACHED */
734 					}
735 				}
736 				(void) mutex_unlock(&svc_mutex);
737 			}
738 		}
739 
740 	}
741 }
742 
743 
744 /*
745  * start_threads() - Start specified number of threads.
746  */
747 static void
748 start_threads(int num_threads)
749 {
750 	int		i;
751 
752 	assert(MUTEX_HELD(&svc_mutex));
753 
754 	for (i = 0; i < num_threads; i++) {
755 		if (thr_create(NULL, 0, (void *(*)(void *))_svc_run_mt, NULL,
756 		    THR_DETACHED, NULL) == 0) {
757 			svc_thr_total++;
758 			svc_thr_total_creates++;
759 		} else {
760 			svc_thr_total_create_errors++;
761 		}
762 	}
763 }
764 
765 
766 /*
767  * create_pipe() - create pipe for breaking out of poll.
768  */
769 static void
770 create_pipe(void)
771 {
772 	if (pipe(svc_pipe) == -1) {
773 		syslog(LOG_ERR, dgettext(__nsl_dom,
774 				"RPC: svc could not create pipe - exiting"));
775 		exit(1);
776 	}
777 	if (fcntl(svc_pipe[0], F_SETFL, O_NONBLOCK) == -1) {
778 		syslog(LOG_ERR, dgettext(__nsl_dom,
779 					"RPC: svc pipe error - exiting"));
780 		exit(1);
781 	}
782 	if (fcntl(svc_pipe[1], F_SETFL, O_NONBLOCK) == -1) {
783 		syslog(LOG_ERR, dgettext(__nsl_dom,
784 					"RPC: svc pipe error - exiting"));
785 		exit(1);
786 	}
787 }
788 
789 
790 /*
791  * clear_pipe() - Empty data in pipe.
792  */
793 static void
794 clear_pipe(void)
795 {
796 	char	buf[16];
797 	int	i;
798 
799 	do {
800 		i = read(svc_pipe[0], buf, sizeof (buf));
801 	} while (i == sizeof (buf));
802 }
803 
804 
805 /*
806  * select_next_pollfd() - Select the next active fd in svc_pollset.
807  */
808 static int
809 select_next_pollfd(int *fd, int *pollfdIndex)
810 {
811 	int i;
812 
813 	assert(MUTEX_HELD(&svc_thr_mutex));
814 	assert(MUTEX_HELD(&svc_mutex));
815 
816 	for (i = svc_next_pollfd; svc_pollfds > 0 && i < svc_polled;
817 							i++) {
818 		if (svc_pollset[i].revents) {
819 			svc_pollfds--;
820 			/*
821 			 * No more special case for POLLNVAL, because it may
822 			 * be linked with a user file descriptot callback
823 			 */
824 			svc_next_pollfd = i + 1;
825 
826 			*fd = svc_pollset[i].fd;
827 			*pollfdIndex = i;
828 
829 			return (0);
830 		}
831 	}
832 	svc_next_pollfd = svc_pollfds = 0;
833 	*fd = -1;
834 	*pollfdIndex = INVALID_POLLFD;
835 	return (-1);
836 }
837 
838 
839 /*
840  * make_xprt_copy() - make a copy of the parent xprt.
841  * Clear fd bit in svc_fdset.
842  */
843 static SVCXPRT *
844 make_xprt_copy(SVCXPRT *parent)
845 {
846 /* LINTED pointer alignment */
847 	SVCXPRT_LIST	*xlist = SVCEXT(parent)->my_xlist;
848 	SVCXPRT_LIST	*xret;
849 	SVCXPRT		*xprt;
850 	int		fd = parent->xp_fd;
851 
852 	assert(MUTEX_HELD(&svc_mutex));
853 
854 	xret = xlist->next;
855 	if (xret) {
856 		xlist->next = xret->next;
857 		xret->next = NULL;
858 		xprt = xret->xprt;
859 /* LINTED pointer alignment */
860 		svc_flags(xprt) = svc_flags(parent);
861 	} else
862 		xprt = svc_copy(parent);
863 
864 	if (xprt) {
865 /* LINTED pointer alignment */
866 		SVCEXT(parent)->refcnt++;
867 		(void) rw_wrlock(&svc_fd_lock);
868 		clear_pollfd(fd);
869 		(void) rw_unlock(&svc_fd_lock);
870 	}
871 	return (xprt);
872 }
873 
874 /*
875  * _svc_done_private() - return copies to library.
876  */
877 static void
878 _svc_done_private(SVCXPRT *xprt)
879 {
880 	SVCXPRT		*parent;
881 	SVCXPRT_LIST	*xhead, *xlist;
882 
883 	assert(MUTEX_HELD(&svc_mutex));
884 
885 /* LINTED pointer alignment */
886 	if ((parent = SVCEXT(xprt)->parent) == NULL)
887 		return;
888 
889 /* LINTED pointer alignment */
890 	xhead = SVCEXT(parent)->my_xlist;
891 /* LINTED pointer alignment */
892 	xlist = SVCEXT(xprt)->my_xlist;
893 	xlist->next = xhead->next;
894 	xhead->next = xlist;
895 
896 /* LINTED pointer alignment */
897 	SVCEXT(parent)->refcnt--;
898 
899 	/*
900 	 * Propagate any error flags.  This is done in both directions to
901 	 * ensure that if one child gets an error, everyone will see it
902 	 * (even if there are multiple outstanding children) and the
903 	 * transport will get closed.
904 	 */
905 /* LINTED pointer alignment */
906 	svc_flags(xprt) |= svc_flags(parent);
907 /* LINTED pointer alignment */
908 	if (svc_failed(xprt) || svc_defunct(xprt)) {
909 /* LINTED pointer alignment */
910 		svc_flags(parent) |= (svc_flags(xprt) &
911 				(SVC_FAILED | SVC_DEFUNCT));
912 /* LINTED pointer alignment */
913 		if (SVCEXT(parent)->refcnt == 0)
914 			_svc_destroy_private(xprt);
915 	}
916 }
917 
918 void
919 svc_done(SVCXPRT *xprt)
920 {
921 	if (svc_mt_mode != RPC_SVC_MT_USER)
922 		return;
923 
924 	/*
925 	 * Make sure file descriptor is released in user mode.
926 	 * If the xprt is a door, do nothing: this work is performed by
927 	 * svc_door.c's return_xprt_copy() routine, which is basically a
928 	 * door-specific copy of _svc_done_private().
929 	 */
930 /* LINTED pointer alignment */
931 	if (svc_type(xprt) == SVC_DOOR)
932 		return;
933 
934 /* LINTED pointer alignment */
935 	if (svc_flags(xprt) & SVC_ARGS_CHECK)
936 		svc_args_done(xprt);
937 
938 	(void) mutex_lock(&svc_mutex);
939 	_svc_done_private(xprt);
940 	(void) mutex_unlock(&svc_mutex);
941 }
942 
943 
944 /*
945  * Mark argument completion.  Release file descriptor.
946  */
947 void
948 svc_args_done(SVCXPRT *xprt)
949 {
950 	char	dummy;
951 /* LINTED pointer alignment */
952 	SVCXPRT	*parent = SVCEXT(xprt)->parent;
953 	bool_t	wake_up_poller;
954 	enum	xprt_stat stat;
955 
956 /* LINTED pointer alignment */
957 	svc_flags(xprt) |= svc_flags(parent);
958 /* LINTED pointer alignment */
959 	svc_flags(xprt) &= ~SVC_ARGS_CHECK;
960 /* LINTED pointer alignment */
961 	if (svc_failed(xprt) || svc_defunct(parent))
962 		return;
963 
964 /* LINTED pointer alignment */
965 	if (svc_type(xprt) == SVC_CONNECTION &&
966 				(stat = SVC_STAT(xprt)) != XPRT_IDLE) {
967 		if (stat == XPRT_MOREREQS) {
968 			(void) mutex_lock(&svc_mutex);
969 			svc_pending_fds[svc_last_pending++] = xprt->xp_fd;
970 			if (svc_last_pending > CIRCULAR_BUFSIZE)
971 				svc_last_pending = 0;
972 			svc_total_pending++;
973 			(void) mutex_unlock(&svc_mutex);
974 			wake_up_poller = FALSE;
975 		} else {
976 			/*
977 			 * connection failed
978 			 */
979 			return;
980 		}
981 	} else {
982 		(void) rw_wrlock(&svc_fd_lock);
983 		set_pollfd(xprt->xp_fd, MASKVAL);
984 		(void) rw_unlock(&svc_fd_lock);
985 		wake_up_poller = TRUE;
986 	}
987 
988 	if (!wake_up_poller || !svc_polling) {
989 		/*
990 		 * Wake up any waiting threads.
991 		 */
992 		(void) mutex_lock(&svc_mutex);
993 		if (svc_waiters > 0) {
994 			(void) cond_broadcast(&svc_thr_fdwait);
995 			(void) mutex_unlock(&svc_mutex);
996 			return;
997 		}
998 		(void) mutex_unlock(&svc_mutex);
999 	}
1000 
1001 	/*
1002 	 * Wake up any polling thread.
1003 	 */
1004 	if (svc_polling)
1005 		(void) write(svc_pipe[1], &dummy, sizeof (dummy));
1006 }
1007 
1008 
1009 int
1010 __rpc_legal_connmaxrec(int suggested) {
1011 	if (suggested == -1) {
1012 		/* Supply default */
1013 		return (RPC_MAXDATASIZE + 2*sizeof (uint32_t));
1014 	} else if (suggested < 0) {
1015 		return (-1);
1016 	} else if (suggested > 0) {
1017 		/* Round down to multiple of BYTES_PER_XDR_UNIT */
1018 		suggested -= suggested % BYTES_PER_XDR_UNIT;
1019 		/* If possible, allow for two fragment headers */
1020 		if (suggested < MAXINT-(2*sizeof (uint32_t))) {
1021 			/* Allow for two fragment headers */
1022 			suggested += 2 * sizeof (uint32_t);
1023 		} else {
1024 			suggested = MAXINT;
1025 		}
1026 		if (suggested < sizeof (struct rpc_msg)) {
1027 			return (-1);
1028 		}
1029 	}
1030 	return (suggested);
1031 }
1032 
1033 
1034 bool_t
1035 rpc_control(int op, void *info)
1036 {
1037 	int		tmp;
1038 	extern int	__rpc_minfd;
1039 
1040 	switch (op) {
1041 	case RPC_SVC_MTMODE_SET:
1042 		tmp = *((int *)info);
1043 		if (tmp != RPC_SVC_MT_NONE && tmp != RPC_SVC_MT_AUTO &&
1044 						tmp != RPC_SVC_MT_USER)
1045 			return (FALSE);
1046 		if (svc_mt_mode != RPC_SVC_MT_NONE && svc_mt_mode != tmp)
1047 			return (FALSE);
1048 		svc_mt_mode = tmp;
1049 		return (TRUE);
1050 	case RPC_SVC_MTMODE_GET:
1051 		*((int *)info) = svc_mt_mode;
1052 		return (TRUE);
1053 	case RPC_SVC_THRMAX_SET:
1054 		if ((tmp = *((int *)info)) < 1)
1055 			return (FALSE);
1056 		(void) mutex_lock(&svc_mutex);
1057 		svc_thr_max = tmp;
1058 		(void) mutex_unlock(&svc_mutex);
1059 		return (TRUE);
1060 	case RPC_SVC_THRMAX_GET:
1061 		*((int *)info) = svc_thr_max;
1062 		return (TRUE);
1063 	case RPC_SVC_THRTOTAL_GET:
1064 		*((int *)info) = svc_thr_total;
1065 		return (TRUE);
1066 	case RPC_SVC_THRCREATES_GET:
1067 		*((int *)info) = svc_thr_total_creates;
1068 		return (TRUE);
1069 	case RPC_SVC_THRERRORS_GET:
1070 		*((int *)info) = svc_thr_total_create_errors;
1071 		return (TRUE);
1072 	case RPC_SVC_USE_POLLFD:
1073 		if (*((int *)info) && !__rpc_use_pollfd_done) {
1074 			__rpc_use_pollfd_done = 1;
1075 			return (TRUE);
1076 		}
1077 		return (FALSE);
1078 	case __RPC_CLNT_MINFD_SET:
1079 		tmp = *((int *)info);
1080 		if (tmp < 0)
1081 			return (FALSE);
1082 		__rpc_minfd = tmp;
1083 		return (TRUE);
1084 	case __RPC_CLNT_MINFD_GET:
1085 		*((int *)info) = __rpc_minfd;
1086 		return (TRUE);
1087 	case RPC_SVC_CONNMAXREC_SET:
1088 		tmp = __rpc_legal_connmaxrec(*(int *)info);
1089 		if (tmp >= 0) {
1090 			__rpc_connmaxrec = tmp;
1091 			return (TRUE);
1092 		} else {
1093 			return (FALSE);
1094 		}
1095 	case RPC_SVC_CONNMAXREC_GET:
1096 		*((int *)info) = __rpc_connmaxrec;
1097 		return (TRUE);
1098 	case RPC_SVC_IRTIMEOUT_SET:
1099 		tmp = *((int *)info);
1100 		if (tmp >= 0) {
1101 			__rpc_irtimeout = tmp;
1102 			return (TRUE);
1103 		} else {
1104 			return (FALSE);
1105 		}
1106 	/*
1107 	 * No mutex necessary as _EXCLBIND_SET will/should only
1108 	 * be used before an RPC daemon goes mt-hot.
1109 	 */
1110 	case __RPC_SVC_EXCLBIND_SET:
1111 		if (info) {
1112 			__rpc_tp_exclbind = *((bool_t *)info);
1113 			return (TRUE);
1114 		}
1115 		return (FALSE);
1116 	case __RPC_SVC_EXCLBIND_GET:
1117 		if (info) {
1118 			*((bool_t *)info) = __rpc_tp_exclbind;
1119 			return (TRUE);
1120 		}
1121 		return (FALSE);
1122 
1123 	default:
1124 		return (FALSE);
1125 	}
1126 }
1127