xref: /freebsd/contrib/ntp/libntp/work_thread.c (revision 61e21613)
1 /*
2  * work_thread.c - threads implementation for blocking worker child.
3  */
4 #include <config.h>
5 #include "ntp_workimpl.h"
6 
7 #ifdef WORK_THREAD
8 
9 #include <stdio.h>
10 #include <ctype.h>
11 #include <signal.h>
12 #ifndef SYS_WINNT
13 #include <pthread.h>
14 #endif
15 
16 #include "ntp_stdlib.h"
17 #include "ntp_malloc.h"
18 #include "ntp_syslog.h"
19 #include "ntpd.h"
20 #include "ntp_io.h"
21 #include "ntp_assert.h"
22 #include "ntp_unixtime.h"
23 #include "timespecops.h"
24 #include "ntp_worker.h"
25 
26 #define CHILD_EXIT_REQ	((blocking_pipe_header *)(intptr_t)-1)
27 #define CHILD_GONE_RESP	CHILD_EXIT_REQ
28 /* Queue size increments:
29  * The request queue grows a bit faster than the response queue -- the
30  * daemon can push requests and pull results faster on avarage than the
31  * worker can process requests and push results...  If this really pays
32  * off is debatable.
33  */
34 #define WORKITEMS_ALLOC_INC	16
35 #define RESPONSES_ALLOC_INC	4
36 
37 /* Fiddle with min/max stack sizes. 64kB minimum seems to work, so we
38  * set the maximum to 256kB. If the minimum goes below the
39  * system-defined minimum stack size, we have to adjust accordingly.
40  */
41 #ifndef THREAD_MINSTACKSIZE
42 # define THREAD_MINSTACKSIZE	(64U * 1024)
43 #endif
44 
45 #ifndef THREAD_MAXSTACKSIZE
46 # define THREAD_MAXSTACKSIZE	(256U * 1024)
47 #endif
48 
49 /* need a good integer to store a pointer... */
50 #ifndef UINTPTR_T
51 # if defined(UINTPTR_MAX)
52 #  define UINTPTR_T uintptr_t
53 # elif defined(UINT_PTR)
54 #  define UINTPTR_T UINT_PTR
55 # else
56 #  define UINTPTR_T size_t
57 # endif
58 #endif
59 
60 
61 #ifdef SYS_WINNT
62 
63 # define thread_exit(c)	_endthreadex(c)
64 # define tickle_sem(sh) ReleaseSemaphore((sh->shnd), 1, NULL)
65 u_int	WINAPI	blocking_thread(void *);
66 static BOOL	same_os_sema(const sem_ref obj, void * osobj);
67 
68 #else
69 
70 # define thread_exit(c)	pthread_exit((void*)(UINTPTR_T)(c))
71 # define tickle_sem	sem_post
72 void *		blocking_thread(void *);
73 static	void	block_thread_signals(sigset_t *);
74 
75 #endif
76 
77 #ifdef WORK_PIPE
78 addremove_io_fd_func		addremove_io_fd;
79 #else
80 addremove_io_semaphore_func	addremove_io_semaphore;
81 #endif
82 
83 static	void	start_blocking_thread(blocking_child *);
84 static	void	start_blocking_thread_internal(blocking_child *);
85 static	void	prepare_child_sems(blocking_child *);
86 static	int	wait_for_sem(sem_ref, struct timespec *);
87 static	int	ensure_workitems_empty_slot(blocking_child *);
88 static	int	ensure_workresp_empty_slot(blocking_child *);
89 static	int	queue_req_pointer(blocking_child *, blocking_pipe_header *);
90 static	void	cleanup_after_child(blocking_child *);
91 
92 static sema_type worker_mmutex;
93 static sem_ref   worker_memlock;
94 
95 /* --------------------------------------------------------------------
96  * locking the global worker state table (and other global stuff)
97  */
98 void
99 worker_global_lock(
100 	int inOrOut)
101 {
102 	if (worker_memlock) {
103 		if (inOrOut)
104 			wait_for_sem(worker_memlock, NULL);
105 		else
106 			tickle_sem(worker_memlock);
107 	}
108 }
109 
110 /* --------------------------------------------------------------------
111  * implementation isolation wrapper
112  */
113 void
114 exit_worker(
115 	int	exitcode
116 	)
117 {
118 	thread_exit(exitcode);	/* see #define thread_exit */
119 }
120 
121 /* --------------------------------------------------------------------
122  * sleep for a given time or until the wakup semaphore is tickled.
123  */
124 int
125 worker_sleep(
126 	blocking_child *	c,
127 	time_t			seconds
128 	)
129 {
130 	struct timespec	until;
131 	int		rc;
132 
133 # ifdef HAVE_CLOCK_GETTIME
134 	if (0 != clock_gettime(CLOCK_REALTIME, &until)) {
135 		msyslog(LOG_ERR, "worker_sleep: clock_gettime() failed: %m");
136 		return -1;
137 	}
138 # else
139 	if (0 != getclock(TIMEOFDAY, &until)) {
140 		msyslog(LOG_ERR, "worker_sleep: getclock() failed: %m");
141 		return -1;
142 	}
143 # endif
144 	until.tv_sec += seconds;
145 	rc = wait_for_sem(c->wake_scheduled_sleep, &until);
146 	if (0 == rc)
147 		return -1;
148 	if (-1 == rc && ETIMEDOUT == errno)
149 		return 0;
150 	msyslog(LOG_ERR, "worker_sleep: sem_timedwait: %m");
151 	return -1;
152 }
153 
154 
155 /* --------------------------------------------------------------------
156  * Wake up a worker that takes a nap.
157  */
158 void
159 interrupt_worker_sleep(void)
160 {
161 	u_int			idx;
162 	blocking_child *	c;
163 
164 	for (idx = 0; idx < blocking_children_alloc; idx++) {
165 		c = blocking_children[idx];
166 		if (NULL == c || NULL == c->wake_scheduled_sleep)
167 			continue;
168 		tickle_sem(c->wake_scheduled_sleep);
169 	}
170 }
171 
172 /* --------------------------------------------------------------------
173  * Make sure there is an empty slot at the head of the request
174  * queue. Tell if the queue is currently empty.
175  */
176 static int
177 ensure_workitems_empty_slot(
178 	blocking_child *c
179 	)
180 {
181 	/*
182 	** !!! PRECONDITION: caller holds access lock!
183 	**
184 	** This simply tries to increase the size of the buffer if it
185 	** becomes full. The resize operation does *not* maintain the
186 	** order of requests, but that should be irrelevant since the
187 	** processing is considered asynchronous anyway.
188 	**
189 	** Return if the buffer is currently empty.
190 	*/
191 
192 	static const size_t each =
193 	    sizeof(blocking_children[0]->workitems[0]);
194 
195 	size_t	new_alloc;
196 	size_t  slots_used;
197 	size_t	sidx;
198 
199 	slots_used = c->head_workitem - c->tail_workitem;
200 	if (slots_used >= c->workitems_alloc) {
201 		new_alloc  = c->workitems_alloc + WORKITEMS_ALLOC_INC;
202 		c->workitems = erealloc(c->workitems, new_alloc * each);
203 		for (sidx = c->workitems_alloc; sidx < new_alloc; ++sidx)
204 		    c->workitems[sidx] = NULL;
205 		c->tail_workitem   = 0;
206 		c->head_workitem   = c->workitems_alloc;
207 		c->workitems_alloc = new_alloc;
208 	}
209 	INSIST(NULL == c->workitems[c->head_workitem % c->workitems_alloc]);
210 	return (0 == slots_used);
211 }
212 
213 /* --------------------------------------------------------------------
214  * Make sure there is an empty slot at the head of the response
215  * queue. Tell if the queue is currently empty.
216  */
217 static int
218 ensure_workresp_empty_slot(
219 	blocking_child *c
220 	)
221 {
222 	/*
223 	** !!! PRECONDITION: caller holds access lock!
224 	**
225 	** Works like the companion function above.
226 	*/
227 
228 	static const size_t each =
229 	    sizeof(blocking_children[0]->responses[0]);
230 
231 	size_t	new_alloc;
232 	size_t  slots_used;
233 	size_t	sidx;
234 
235 	slots_used = c->head_response - c->tail_response;
236 	if (slots_used >= c->responses_alloc) {
237 		new_alloc  = c->responses_alloc + RESPONSES_ALLOC_INC;
238 		c->responses = erealloc(c->responses, new_alloc * each);
239 		for (sidx = c->responses_alloc; sidx < new_alloc; ++sidx)
240 		    c->responses[sidx] = NULL;
241 		c->tail_response   = 0;
242 		c->head_response   = c->responses_alloc;
243 		c->responses_alloc = new_alloc;
244 	}
245 	INSIST(NULL == c->responses[c->head_response % c->responses_alloc]);
246 	return (0 == slots_used);
247 }
248 
249 
250 /* --------------------------------------------------------------------
251  * queue_req_pointer() - append a work item or idle exit request to
252  *			 blocking_workitems[]. Employ proper locking.
253  */
254 static int
255 queue_req_pointer(
256 	blocking_child	*	c,
257 	blocking_pipe_header *	hdr
258 	)
259 {
260 	size_t qhead;
261 
262 	/* >>>> ACCESS LOCKING STARTS >>>> */
263 	wait_for_sem(c->accesslock, NULL);
264 	ensure_workitems_empty_slot(c);
265 	qhead = c->head_workitem;
266 	c->workitems[qhead % c->workitems_alloc] = hdr;
267 	c->head_workitem = 1 + qhead;
268 	tickle_sem(c->accesslock);
269 	/* <<<< ACCESS LOCKING ENDS <<<< */
270 
271 	/* queue consumer wake-up notification */
272 	tickle_sem(c->workitems_pending);
273 
274 	return 0;
275 }
276 
277 /* --------------------------------------------------------------------
278  * API function to make sure a worker is running, a proper private copy
279  * of the data is made, the data eneterd into the queue and the worker
280  * is signalled.
281  */
282 int
283 send_blocking_req_internal(
284 	blocking_child *	c,
285 	blocking_pipe_header *	hdr,
286 	void *			data
287 	)
288 {
289 	blocking_pipe_header *	threadcopy;
290 	size_t			payload_octets;
291 
292 	REQUIRE(hdr != NULL);
293 	REQUIRE(data != NULL);
294 	DEBUG_REQUIRE(BLOCKING_REQ_MAGIC == hdr->magic_sig);
295 
296 	if (hdr->octets <= sizeof(*hdr))
297 		return 1;	/* failure */
298 	payload_octets = hdr->octets - sizeof(*hdr);
299 
300 	if (NULL == c->thread_ref)
301 		start_blocking_thread(c);
302 	threadcopy = emalloc(hdr->octets);
303 	memcpy(threadcopy, hdr, sizeof(*hdr));
304 	memcpy((char *)threadcopy + sizeof(*hdr), data, payload_octets);
305 
306 	return queue_req_pointer(c, threadcopy);
307 }
308 
309 /* --------------------------------------------------------------------
310  * Wait for the 'incoming queue no longer empty' signal, lock the shared
311  * structure and dequeue an item.
312  */
313 blocking_pipe_header *
314 receive_blocking_req_internal(
315 	blocking_child *	c
316 	)
317 {
318 	blocking_pipe_header *	req;
319 	size_t			qhead, qtail;
320 
321 	req = NULL;
322 	do {
323 		/* wait for tickle from the producer side */
324 		wait_for_sem(c->workitems_pending, NULL);
325 
326 		/* >>>> ACCESS LOCKING STARTS >>>> */
327 		wait_for_sem(c->accesslock, NULL);
328 		qhead = c->head_workitem;
329 		do {
330 			qtail = c->tail_workitem;
331 			if (qhead == qtail)
332 				break;
333 			c->tail_workitem = qtail + 1;
334 			qtail %= c->workitems_alloc;
335 			req = c->workitems[qtail];
336 			c->workitems[qtail] = NULL;
337 		} while (NULL == req);
338 		tickle_sem(c->accesslock);
339 		/* <<<< ACCESS LOCKING ENDS <<<< */
340 
341 	} while (NULL == req);
342 
343 	INSIST(NULL != req);
344 	if (CHILD_EXIT_REQ == req) {	/* idled out */
345 		send_blocking_resp_internal(c, CHILD_GONE_RESP);
346 		req = NULL;
347 	}
348 
349 	return req;
350 }
351 
352 /* --------------------------------------------------------------------
353  * Push a response into the return queue and eventually tickle the
354  * receiver.
355  */
356 int
357 send_blocking_resp_internal(
358 	blocking_child *	c,
359 	blocking_pipe_header *	resp
360 	)
361 {
362 	size_t	qhead;
363 	int	empty;
364 
365 	/* >>>> ACCESS LOCKING STARTS >>>> */
366 	wait_for_sem(c->accesslock, NULL);
367 	empty = ensure_workresp_empty_slot(c);
368 	qhead = c->head_response;
369 	c->responses[qhead % c->responses_alloc] = resp;
370 	c->head_response = 1 + qhead;
371 	tickle_sem(c->accesslock);
372 	/* <<<< ACCESS LOCKING ENDS <<<< */
373 
374 	/* queue consumer wake-up notification */
375 	if (empty)
376 	{
377 #	    ifdef WORK_PIPE
378 		if (1 != write(c->resp_write_pipe, "", 1))
379 			msyslog(LOG_WARNING, "async resolver: %s",
380 				"failed to notify main thread!");
381 #	    else
382 		tickle_sem(c->responses_pending);
383 #	    endif
384 	}
385 	return 0;
386 }
387 
388 
389 #ifndef WORK_PIPE
390 
391 /* --------------------------------------------------------------------
392  * Check if a (Windows-)handle to a semaphore is actually the same we
393  * are using inside the sema wrapper.
394  */
395 static BOOL
396 same_os_sema(
397 	const sem_ref	obj,
398 	void*		osh
399 	)
400 {
401 	return obj && osh && (obj->shnd == (HANDLE)osh);
402 }
403 
404 /* --------------------------------------------------------------------
405  * Find the shared context that associates to an OS handle and make sure
406  * the data is dequeued and processed.
407  */
408 void
409 handle_blocking_resp_sem(
410 	void *	context
411 	)
412 {
413 	blocking_child *	c;
414 	u_int			idx;
415 
416 	c = NULL;
417 	for (idx = 0; idx < blocking_children_alloc; idx++) {
418 		c = blocking_children[idx];
419 		if (c != NULL &&
420 			c->thread_ref != NULL &&
421 			same_os_sema(c->responses_pending, context))
422 			break;
423 	}
424 	if (idx < blocking_children_alloc)
425 		process_blocking_resp(c);
426 }
427 #endif	/* !WORK_PIPE */
428 
429 /* --------------------------------------------------------------------
430  * Fetch the next response from the return queue. In case of signalling
431  * via pipe, make sure the pipe is flushed, too.
432  */
433 blocking_pipe_header *
434 receive_blocking_resp_internal(
435 	blocking_child *	c
436 	)
437 {
438 	blocking_pipe_header *	removed;
439 	size_t			qhead, qtail, slot;
440 
441 #ifdef WORK_PIPE
442 	int			rc;
443 	char			scratch[32];
444 
445 	do
446 		rc = read(c->resp_read_pipe, scratch, sizeof(scratch));
447 	while (-1 == rc && EINTR == errno);
448 #endif
449 
450 	/* >>>> ACCESS LOCKING STARTS >>>> */
451 	wait_for_sem(c->accesslock, NULL);
452 	qhead = c->head_response;
453 	qtail = c->tail_response;
454 	for (removed = NULL; !removed && (qhead != qtail); ++qtail) {
455 		slot = qtail % c->responses_alloc;
456 		removed = c->responses[slot];
457 		c->responses[slot] = NULL;
458 	}
459 	c->tail_response = qtail;
460 	tickle_sem(c->accesslock);
461 	/* <<<< ACCESS LOCKING ENDS <<<< */
462 
463 	if (NULL != removed) {
464 		DEBUG_ENSURE(CHILD_GONE_RESP == removed ||
465 			     BLOCKING_RESP_MAGIC == removed->magic_sig);
466 	}
467 	if (CHILD_GONE_RESP == removed) {
468 		cleanup_after_child(c);
469 		removed = NULL;
470 	}
471 
472 	return removed;
473 }
474 
475 /* --------------------------------------------------------------------
476  * Light up a new worker.
477  */
478 static void
479 start_blocking_thread(
480 	blocking_child *	c
481 	)
482 {
483 
484 	DEBUG_INSIST(!c->reusable);
485 
486 	prepare_child_sems(c);
487 	start_blocking_thread_internal(c);
488 }
489 
490 /* --------------------------------------------------------------------
491  * Create a worker thread. There are several differences between POSIX
492  * and Windows, of course -- most notably the Windows thread is no
493  * detached thread, and we keep the handle around until we want to get
494  * rid of the thread. The notification scheme also differs: Windows
495  * makes use of semaphores in both directions, POSIX uses a pipe for
496  * integration with 'select()' or alike.
497  */
498 static void
499 start_blocking_thread_internal(
500 	blocking_child *	c
501 	)
502 #ifdef SYS_WINNT
503 {
504 	BOOL	resumed;
505 
506 	c->thread_ref = NULL;
507 	(*addremove_io_semaphore)(c->responses_pending->shnd, FALSE);
508 	c->thr_table[0].thnd =
509 		(HANDLE)_beginthreadex(
510 			NULL,
511 			0,
512 			&blocking_thread,
513 			c,
514 			CREATE_SUSPENDED,
515 			NULL);
516 
517 	if (NULL == c->thr_table[0].thnd) {
518 		msyslog(LOG_ERR, "start blocking thread failed: %m");
519 		exit(-1);
520 	}
521 	/* remember the thread priority is only within the process class */
522 	if (!SetThreadPriority(c->thr_table[0].thnd,
523 			       THREAD_PRIORITY_BELOW_NORMAL))
524 		msyslog(LOG_ERR, "Error lowering blocking thread priority: %m");
525 
526 	resumed = ResumeThread(c->thr_table[0].thnd);
527 	DEBUG_INSIST(resumed);
528 	c->thread_ref = &c->thr_table[0];
529 }
530 #else	/* pthreads start_blocking_thread_internal() follows */
531 {
532 # ifdef NEED_PTHREAD_INIT
533 	static int	pthread_init_called;
534 # endif
535 	pthread_attr_t	thr_attr;
536 	int		rc;
537 	int		pipe_ends[2];	/* read then write */
538 	int		is_pipe;
539 	int		flags;
540 	size_t		ostacksize;
541 	size_t		nstacksize;
542 	sigset_t	saved_sig_mask;
543 
544 	c->thread_ref = NULL;
545 
546 # ifdef NEED_PTHREAD_INIT
547 	/*
548 	 * from lib/isc/unix/app.c:
549 	 * BSDI 3.1 seg faults in pthread_sigmask() if we don't do this.
550 	 */
551 	if (!pthread_init_called) {
552 		pthread_init();
553 		pthread_init_called = TRUE;
554 	}
555 # endif
556 
557 	rc = pipe_socketpair(&pipe_ends[0], &is_pipe);
558 	if (0 != rc) {
559 		msyslog(LOG_ERR, "start_blocking_thread: pipe_socketpair() %m");
560 		exit(1);
561 	}
562 	c->resp_read_pipe = move_fd(pipe_ends[0]);
563 	c->resp_write_pipe = move_fd(pipe_ends[1]);
564 	c->ispipe = is_pipe;
565 	flags = fcntl(c->resp_read_pipe, F_GETFL, 0);
566 	if (-1 == flags) {
567 		msyslog(LOG_ERR, "start_blocking_thread: fcntl(F_GETFL) %m");
568 		exit(1);
569 	}
570 	rc = fcntl(c->resp_read_pipe, F_SETFL, O_NONBLOCK | flags);
571 	if (-1 == rc) {
572 		msyslog(LOG_ERR,
573 			"start_blocking_thread: fcntl(F_SETFL, O_NONBLOCK) %m");
574 		exit(1);
575 	}
576 	(*addremove_io_fd)(c->resp_read_pipe, c->ispipe, FALSE);
577 	pthread_attr_init(&thr_attr);
578 	pthread_attr_setdetachstate(&thr_attr, PTHREAD_CREATE_DETACHED);
579 #if defined(HAVE_PTHREAD_ATTR_GETSTACKSIZE) && \
580     defined(HAVE_PTHREAD_ATTR_SETSTACKSIZE)
581 	rc = pthread_attr_getstacksize(&thr_attr, &ostacksize);
582 	if (0 != rc) {
583 		msyslog(LOG_ERR,
584 			"start_blocking_thread: pthread_attr_getstacksize() -> %s",
585 			strerror(rc));
586 	} else {
587 		nstacksize = ostacksize;
588 		/* order is important here: first clamp on upper limit,
589 		 * and the PTHREAD min stack size is ultimate override!
590 		 */
591 		if (nstacksize > THREAD_MAXSTACKSIZE)
592 			nstacksize = THREAD_MAXSTACKSIZE;
593 #            ifdef PTHREAD_STACK_MAX
594 		if (nstacksize > PTHREAD_STACK_MAX)
595 			nstacksize = PTHREAD_STACK_MAX;
596 #            endif
597 
598 		/* now clamp on lower stack limit. */
599 		if (nstacksize < THREAD_MINSTACKSIZE)
600 			nstacksize = THREAD_MINSTACKSIZE;
601 #            ifdef PTHREAD_STACK_MIN
602 		if (nstacksize < PTHREAD_STACK_MIN)
603 			nstacksize = PTHREAD_STACK_MIN;
604 #            endif
605 
606 		if (nstacksize != ostacksize)
607 			rc = pthread_attr_setstacksize(&thr_attr, nstacksize);
608 		if (0 != rc)
609 			msyslog(LOG_ERR,
610 				"start_blocking_thread: pthread_attr_setstacksize(0x%lx -> 0x%lx) -> %s",
611 				(u_long)ostacksize, (u_long)nstacksize,
612 				strerror(rc));
613 	}
614 #else
615 	UNUSED_ARG(nstacksize);
616 	UNUSED_ARG(ostacksize);
617 #endif
618 #if defined(PTHREAD_SCOPE_SYSTEM) && defined(NEED_PTHREAD_SCOPE_SYSTEM)
619 	pthread_attr_setscope(&thr_attr, PTHREAD_SCOPE_SYSTEM);
620 #endif
621 	c->thread_ref = emalloc_zero(sizeof(*c->thread_ref));
622 	block_thread_signals(&saved_sig_mask);
623 	rc = pthread_create(&c->thr_table[0], &thr_attr,
624 			    &blocking_thread, c);
625 	pthread_sigmask(SIG_SETMASK, &saved_sig_mask, NULL);
626 	pthread_attr_destroy(&thr_attr);
627 	if (0 != rc) {
628 		msyslog(LOG_ERR, "start_blocking_thread: pthread_create() -> %s",
629 			strerror(rc));
630 		exit(1);
631 	}
632 	c->thread_ref = &c->thr_table[0];
633 }
634 #endif
635 
636 /* --------------------------------------------------------------------
637  * block_thread_signals()
638  *
639  * Temporarily block signals used by ntpd main thread, so that signal
640  * mask inherited by child threads leaves them blocked.  Returns prior
641  * active signal mask via pmask, to be restored by the main thread
642  * after pthread_create().
643  */
644 #ifndef SYS_WINNT
645 void
646 block_thread_signals(
647 	sigset_t *	pmask
648 	)
649 {
650 	sigset_t	block;
651 
652 	sigemptyset(&block);
653 # ifdef HAVE_SIGNALED_IO
654 #  ifdef SIGIO
655 	sigaddset(&block, SIGIO);
656 #  endif
657 #  ifdef SIGPOLL
658 	sigaddset(&block, SIGPOLL);
659 #  endif
660 # endif	/* HAVE_SIGNALED_IO */
661 	sigaddset(&block, SIGALRM);
662 	sigaddset(&block, MOREDEBUGSIG);
663 	sigaddset(&block, LESSDEBUGSIG);
664 # ifdef SIGDIE1
665 	sigaddset(&block, SIGDIE1);
666 # endif
667 # ifdef SIGDIE2
668 	sigaddset(&block, SIGDIE2);
669 # endif
670 # ifdef SIGDIE3
671 	sigaddset(&block, SIGDIE3);
672 # endif
673 # ifdef SIGDIE4
674 	sigaddset(&block, SIGDIE4);
675 # endif
676 # ifdef SIGBUS
677 	sigaddset(&block, SIGBUS);
678 # endif
679 	sigemptyset(pmask);
680 	pthread_sigmask(SIG_BLOCK, &block, pmask);
681 }
682 #endif	/* !SYS_WINNT */
683 
684 
685 /* --------------------------------------------------------------------
686  * Create & destroy semaphores. This is sufficiently different between
687  * POSIX and Windows to warrant wrapper functions and close enough to
688  * use the concept of synchronization via semaphore for all platforms.
689  */
690 static sem_ref
691 create_sema(
692 	sema_type*	semptr,
693 	u_int		inival,
694 	u_int		maxval)
695 {
696 #ifdef SYS_WINNT
697 
698 	long svini, svmax;
699 	if (NULL != semptr) {
700 		svini = (inival < LONG_MAX)
701 		    ? (long)inival : LONG_MAX;
702 		svmax = (maxval < LONG_MAX && maxval > 0)
703 		    ? (long)maxval : LONG_MAX;
704 		semptr->shnd = CreateSemaphore(NULL, svini, svmax, NULL);
705 		if (NULL == semptr->shnd)
706 			semptr = NULL;
707 	}
708 
709 #else
710 
711 	(void)maxval;
712 	if (semptr && sem_init(semptr, FALSE, inival))
713 		semptr = NULL;
714 
715 #endif
716 
717 	return semptr;
718 }
719 
720 /* ------------------------------------------------------------------ */
721 static sem_ref
722 delete_sema(
723 	sem_ref obj)
724 {
725 
726 #   ifdef SYS_WINNT
727 
728 	if (obj) {
729 		if (obj->shnd)
730 			CloseHandle(obj->shnd);
731 		obj->shnd = NULL;
732 	}
733 
734 #   else
735 
736 	if (obj)
737 		sem_destroy(obj);
738 
739 #   endif
740 
741 	return NULL;
742 }
743 
744 /* --------------------------------------------------------------------
745  * prepare_child_sems()
746  *
747  * create sync & access semaphores
748  *
749  * All semaphores are cleared, only the access semaphore has 1 unit.
750  * Childs wait on 'workitems_pending', then grabs 'sema_access'
751  * and dequeues jobs. When done, 'sema_access' is given one unit back.
752  *
753  * The producer grabs 'sema_access', manages the queue, restores
754  * 'sema_access' and puts one unit into 'workitems_pending'.
755  *
756  * The story goes the same for the response queue.
757  */
758 static void
759 prepare_child_sems(
760 	blocking_child *c
761 	)
762 {
763 	if (NULL == worker_memlock)
764 		worker_memlock = create_sema(&worker_mmutex, 1, 1);
765 
766 	c->accesslock           = create_sema(&c->sem_table[0], 1, 1);
767 	c->workitems_pending    = create_sema(&c->sem_table[1], 0, 0);
768 	c->wake_scheduled_sleep = create_sema(&c->sem_table[2], 0, 1);
769 #   ifndef WORK_PIPE
770 	c->responses_pending    = create_sema(&c->sem_table[3], 0, 0);
771 #   endif
772 }
773 
774 /* --------------------------------------------------------------------
775  * wait for semaphore. Where the wait can be interrupted, it will
776  * internally resume -- When this function returns, there is either no
777  * semaphore at all, a timeout occurred, or the caller could
778  * successfully take a token from the semaphore.
779  *
780  * For untimed wait, not checking the result of this function at all is
781  * definitely an option.
782  */
783 static int
784 wait_for_sem(
785 	sem_ref			sem,
786 	struct timespec *	timeout		/* wall-clock */
787 	)
788 #ifdef SYS_WINNT
789 {
790 	struct timespec now;
791 	struct timespec delta;
792 	DWORD		msec;
793 	DWORD		rc;
794 
795 	if (!(sem && sem->shnd)) {
796 		errno = EINVAL;
797 		return -1;
798 	}
799 
800 	if (NULL == timeout) {
801 		msec = INFINITE;
802 	} else {
803 		getclock(TIMEOFDAY, &now);
804 		delta = sub_tspec(*timeout, now);
805 		if (delta.tv_sec < 0) {
806 			msec = 0;
807 		} else if ((delta.tv_sec + 1) >= (MAXDWORD / 1000)) {
808 			msec = INFINITE;
809 		} else {
810 			msec = 1000 * (DWORD)delta.tv_sec;
811 			msec += delta.tv_nsec / (1000 * 1000);
812 		}
813 	}
814 	rc = WaitForSingleObject(sem->shnd, msec);
815 	if (WAIT_OBJECT_0 == rc)
816 		return 0;
817 	if (WAIT_TIMEOUT == rc) {
818 		errno = ETIMEDOUT;
819 		return -1;
820 	}
821 	msyslog(LOG_ERR, "WaitForSingleObject unexpected 0x%x", rc);
822 	errno = EFAULT;
823 	return -1;
824 }
825 #else	/* pthreads wait_for_sem() follows */
826 {
827 	int rc = -1;
828 
829 	if (sem) do {
830 			if (NULL == timeout)
831 				rc = sem_wait(sem);
832 			else
833 				rc = sem_timedwait(sem, timeout);
834 		} while (rc == -1 && errno == EINTR);
835 	else
836 		errno = EINVAL;
837 
838 	return rc;
839 }
840 #endif
841 
842 /* --------------------------------------------------------------------
843  * blocking_thread - thread functions have WINAPI (aka 'stdcall')
844  * calling conventions under Windows and POSIX-defined signature
845  * otherwise.
846  */
847 #ifdef SYS_WINNT
848 u_int WINAPI
849 #else
850 void *
851 #endif
852 blocking_thread(
853 	void *	ThreadArg
854 	)
855 {
856 	blocking_child *c;
857 
858 	c = ThreadArg;
859 	exit_worker(blocking_child_common(c));
860 
861 	/* NOTREACHED */
862 	return 0;
863 }
864 
865 /* --------------------------------------------------------------------
866  * req_child_exit() runs in the parent.
867  *
868  * This function is called from from the idle timer, too, and possibly
869  * without a thread being there any longer. Since we have folded up our
870  * tent in that case and all the semaphores are already gone, we simply
871  * ignore this request in this case.
872  *
873  * Since the existence of the semaphores is controlled exclusively by
874  * the parent, there's no risk of data race here.
875  */
876 int
877 req_child_exit(
878 	blocking_child *c
879 	)
880 {
881 	return (c->accesslock)
882 	    ? queue_req_pointer(c, CHILD_EXIT_REQ)
883 	    : 0;
884 }
885 
886 /* --------------------------------------------------------------------
887  * cleanup_after_child() runs in parent.
888  */
889 static void
890 cleanup_after_child(
891 	blocking_child *	c
892 	)
893 {
894 	DEBUG_INSIST(!c->reusable);
895 
896 #   ifdef SYS_WINNT
897 	/* The thread was not created in detached state, so we better
898 	 * clean up.
899 	 */
900 	if (c->thread_ref && c->thread_ref->thnd) {
901 		WaitForSingleObject(c->thread_ref->thnd, INFINITE);
902 		INSIST(CloseHandle(c->thread_ref->thnd));
903 		c->thread_ref->thnd = NULL;
904 	}
905 #   endif
906 	c->thread_ref = NULL;
907 
908 	/* remove semaphores and (if signalling vi IO) pipes */
909 
910 	c->accesslock           = delete_sema(c->accesslock);
911 	c->workitems_pending    = delete_sema(c->workitems_pending);
912 	c->wake_scheduled_sleep = delete_sema(c->wake_scheduled_sleep);
913 
914 #   ifdef WORK_PIPE
915 	DEBUG_INSIST(-1 != c->resp_read_pipe);
916 	DEBUG_INSIST(-1 != c->resp_write_pipe);
917 	(*addremove_io_fd)(c->resp_read_pipe, c->ispipe, TRUE);
918 	close(c->resp_write_pipe);
919 	close(c->resp_read_pipe);
920 	c->resp_write_pipe = -1;
921 	c->resp_read_pipe = -1;
922 #   else
923 	DEBUG_INSIST(NULL != c->responses_pending);
924 	(*addremove_io_semaphore)(c->responses_pending->shnd, TRUE);
925 	c->responses_pending = delete_sema(c->responses_pending);
926 #   endif
927 
928 	/* Is it necessary to check if there are pending requests and
929 	 * responses? If so, and if there are, what to do with them?
930 	 */
931 
932 	/* re-init buffer index sequencers */
933 	c->head_workitem = 0;
934 	c->tail_workitem = 0;
935 	c->head_response = 0;
936 	c->tail_response = 0;
937 
938 	c->reusable = TRUE;
939 }
940 
941 
942 #else	/* !WORK_THREAD follows */
943 char work_thread_nonempty_compilation_unit;
944 #endif
945