xref: /titanic_41/usr/src/lib/libc/port/aio/aio.c (revision fb49a0d14105507530a42288b4a7bec770c7b016)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include "synonyms.h"
30 #include "thr_uberdata.h"
31 #include "asyncio.h"
32 #include <atomic.h>
33 #include <sys/param.h>
34 #include <sys/file.h>
35 #include <sys/port.h>
36 
37 static int _aio_hash_insert(aio_result_t *, aio_req_t *);
38 static aio_req_t *_aio_req_get(aio_worker_t *);
39 static void _aio_req_add(aio_req_t *, aio_worker_t **, int);
40 static void _aio_req_del(aio_worker_t *, aio_req_t *, int);
41 static void _aio_work_done(aio_worker_t *);
42 static void _aio_enq_doneq(aio_req_t *);
43 
44 extern void _aio_lio_free(aio_lio_t *);
45 
46 extern int __fdsync(int, int);
47 extern int __fcntl(int, int, ...);
48 extern int _port_dispatch(int, int, int, int, uintptr_t, void *);
49 
50 static int _aio_fsync_del(aio_worker_t *, aio_req_t *);
51 static void _aiodone(aio_req_t *, ssize_t, int);
52 static void _aio_cancel_work(aio_worker_t *, int, int *, int *);
53 static void _aio_finish_request(aio_worker_t *, ssize_t, int);
54 
55 /*
56  * switch for kernel async I/O
57  */
58 int _kaio_ok = 0;		/* 0 = disabled, 1 = on, -1 = error */
59 
60 /*
61  * Key for thread-specific data
62  */
63 pthread_key_t _aio_key;
64 
65 /*
66  * Array for determining whether or not a file supports kaio.
67  * Initialized in _kaio_init().
68  */
69 uint32_t *_kaio_supported = NULL;
70 
71 /*
72  *  workers for read/write requests
73  * (__aio_mutex lock protects circular linked list of workers)
74  */
75 aio_worker_t *__workers_rw;	/* circular list of AIO workers */
76 aio_worker_t *__nextworker_rw;	/* next worker in list of workers */
77 int __rw_workerscnt;		/* number of read/write workers */
78 
79 /*
80  * worker for notification requests.
81  */
82 aio_worker_t *__workers_no;	/* circular list of AIO workers */
83 aio_worker_t *__nextworker_no;	/* next worker in list of workers */
84 int __no_workerscnt;		/* number of write workers */
85 
86 aio_req_t *_aio_done_tail;		/* list of done requests */
87 aio_req_t *_aio_done_head;
88 
89 mutex_t __aio_initlock = DEFAULTMUTEX;	/* makes aio initialization atomic */
90 cond_t __aio_initcv = DEFAULTCV;
91 int __aio_initbusy = 0;
92 
93 mutex_t __aio_mutex = DEFAULTMUTEX;	/* protects counts, and linked lists */
94 cond_t _aio_iowait_cv = DEFAULTCV;	/* wait for userland I/Os */
95 
96 pid_t __pid = (pid_t)-1;		/* initialize as invalid pid */
97 int _sigio_enabled = 0;			/* when set, send SIGIO signal */
98 
99 aio_hash_t *_aio_hash;
100 
101 aio_req_t *_aio_doneq;			/* double linked done queue list */
102 
103 int _aio_donecnt = 0;
104 int _aio_waitncnt = 0;			/* # of requests for aio_waitn */
105 int _aio_doneq_cnt = 0;
106 int _aio_outstand_cnt = 0;		/* # of outstanding requests */
107 int _kaio_outstand_cnt = 0;		/* # of outstanding kaio requests */
108 int _aio_req_done_cnt = 0;		/* req. done but not in "done queue" */
109 int _aio_kernel_suspend = 0;		/* active kernel kaio calls */
110 int _aio_suscv_cnt = 0;			/* aio_suspend calls waiting on cv's */
111 
112 int _max_workers = 256;			/* max number of workers permitted */
113 int _min_workers = 4;			/* min number of workers */
114 int _minworkload = 2;			/* min number of request in q */
115 int _aio_worker_cnt = 0;		/* number of workers to do requests */
116 int __uaio_ok = 0;			/* AIO has been enabled */
117 sigset_t _worker_set;			/* worker's signal mask */
118 
119 int _aiowait_flag = 0;			/* when set, aiowait() is inprogress */
120 int _aio_flags = 0;			/* see asyncio.h defines for */
121 
122 aio_worker_t *_kaiowp = NULL;		/* points to kaio cleanup thread */
123 
124 int hz;					/* clock ticks per second */
125 
126 static int
127 _kaio_supported_init(void)
128 {
129 	void *ptr;
130 	size_t size;
131 
132 	if (_kaio_supported != NULL)	/* already initialized */
133 		return (0);
134 
135 	size = MAX_KAIO_FDARRAY_SIZE * sizeof (uint32_t);
136 	ptr = mmap(NULL, size, PROT_READ | PROT_WRITE,
137 	    MAP_PRIVATE | MAP_ANON, -1, (off_t)0);
138 	if (ptr == MAP_FAILED)
139 		return (-1);
140 	_kaio_supported = ptr;
141 	return (0);
142 }
143 
144 /*
145  * The aio subsystem is initialized when an AIO request is made.
146  * Constants are initialized like the max number of workers that
147  * the subsystem can create, and the minimum number of workers
148  * permitted before imposing some restrictions.  Also, some
149  * workers are created.
150  */
151 int
152 __uaio_init(void)
153 {
154 	int ret = -1;
155 	int i;
156 	int cancel_state;
157 
158 	lmutex_lock(&__aio_initlock);
159 	(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cancel_state);
160 	while (__aio_initbusy)
161 		(void) cond_wait(&__aio_initcv, &__aio_initlock);
162 	(void) pthread_setcancelstate(cancel_state, NULL);
163 	if (__uaio_ok) {	/* already initialized */
164 		lmutex_unlock(&__aio_initlock);
165 		return (0);
166 	}
167 	__aio_initbusy = 1;
168 	lmutex_unlock(&__aio_initlock);
169 
170 	hz = (int)sysconf(_SC_CLK_TCK);
171 	__pid = getpid();
172 
173 	setup_cancelsig(SIGAIOCANCEL);
174 
175 	if (_kaio_supported_init() != 0)
176 		goto out;
177 
178 	/*
179 	 * Allocate and initialize the hash table.
180 	 * Do this only once, even if __uaio_init() is called twice.
181 	 */
182 	if (_aio_hash == NULL) {
183 		/* LINTED pointer cast */
184 		_aio_hash = (aio_hash_t *)mmap(NULL,
185 		    HASHSZ * sizeof (aio_hash_t), PROT_READ | PROT_WRITE,
186 		    MAP_PRIVATE | MAP_ANON, -1, (off_t)0);
187 		if ((void *)_aio_hash == MAP_FAILED) {
188 			_aio_hash = NULL;
189 			goto out;
190 		}
191 		for (i = 0; i < HASHSZ; i++)
192 			(void) mutex_init(&_aio_hash[i].hash_lock,
193 			    USYNC_THREAD, NULL);
194 	}
195 
196 	/*
197 	 * Initialize worker's signal mask to only catch SIGAIOCANCEL.
198 	 */
199 	(void) sigfillset(&_worker_set);
200 	(void) sigdelset(&_worker_set, SIGAIOCANCEL);
201 
202 	/*
203 	 * Create one worker to send asynchronous notifications.
204 	 * Do this only once, even if __uaio_init() is called twice.
205 	 */
206 	if (__no_workerscnt == 0 &&
207 	    (_aio_create_worker(NULL, AIONOTIFY) != 0)) {
208 		errno = EAGAIN;
209 		goto out;
210 	}
211 
212 	/*
213 	 * Create the minimum number of read/write workers.
214 	 * And later check whether atleast one worker is created;
215 	 * lwp_create() calls could fail because of segkp exhaustion.
216 	 */
217 	for (i = 0; i < _min_workers; i++)
218 		(void) _aio_create_worker(NULL, AIOREAD);
219 	if (__rw_workerscnt == 0) {
220 		errno = EAGAIN;
221 		goto out;
222 	}
223 
224 	ret = 0;
225 out:
226 	lmutex_lock(&__aio_initlock);
227 	if (ret == 0)
228 		__uaio_ok = 1;
229 	__aio_initbusy = 0;
230 	(void) cond_broadcast(&__aio_initcv);
231 	lmutex_unlock(&__aio_initlock);
232 	return (ret);
233 }
234 
235 /*
236  * Called from close() before actually performing the real _close().
237  */
238 void
239 _aio_close(int fd)
240 {
241 	if (fd < 0)	/* avoid cancelling everything */
242 		return;
243 	/*
244 	 * Cancel all outstanding aio requests for this file descriptor.
245 	 */
246 	if (__uaio_ok)
247 		(void) aiocancel_all(fd);
248 	/*
249 	 * If we have allocated the bit array, clear the bit for this file.
250 	 * The next open may re-use this file descriptor and the new file
251 	 * may have different kaio() behaviour.
252 	 */
253 	if (_kaio_supported != NULL)
254 		CLEAR_KAIO_SUPPORTED(fd);
255 }
256 
257 /*
258  * special kaio cleanup thread sits in a loop in the
259  * kernel waiting for pending kaio requests to complete.
260  */
261 void *
262 _kaio_cleanup_thread(void *arg)
263 {
264 	if (pthread_setspecific(_aio_key, arg) != 0)
265 		aio_panic("_kaio_cleanup_thread, pthread_setspecific()");
266 	(void) _kaio(AIOSTART);
267 	return (arg);
268 }
269 
270 /*
271  * initialize kaio.
272  */
273 void
274 _kaio_init()
275 {
276 	int error;
277 	sigset_t oset;
278 	int cancel_state;
279 
280 	lmutex_lock(&__aio_initlock);
281 	(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cancel_state);
282 	while (__aio_initbusy)
283 		(void) cond_wait(&__aio_initcv, &__aio_initlock);
284 	(void) pthread_setcancelstate(cancel_state, NULL);
285 	if (_kaio_ok) {		/* already initialized */
286 		lmutex_unlock(&__aio_initlock);
287 		return;
288 	}
289 	__aio_initbusy = 1;
290 	lmutex_unlock(&__aio_initlock);
291 
292 	if (_kaio_supported_init() != 0)
293 		error = ENOMEM;
294 	else if ((_kaiowp = _aio_worker_alloc()) == NULL)
295 		error = ENOMEM;
296 	else if ((error = (int)_kaio(AIOINIT)) == 0) {
297 		(void) pthread_sigmask(SIG_SETMASK, &maskset, &oset);
298 		error = thr_create(NULL, AIOSTKSIZE, _kaio_cleanup_thread,
299 		    _kaiowp, THR_DAEMON, &_kaiowp->work_tid);
300 		(void) pthread_sigmask(SIG_SETMASK, &oset, NULL);
301 	}
302 	if (error && _kaiowp != NULL) {
303 		_aio_worker_free(_kaiowp);
304 		_kaiowp = NULL;
305 	}
306 
307 	lmutex_lock(&__aio_initlock);
308 	if (error)
309 		_kaio_ok = -1;
310 	else
311 		_kaio_ok = 1;
312 	__aio_initbusy = 0;
313 	(void) cond_broadcast(&__aio_initcv);
314 	lmutex_unlock(&__aio_initlock);
315 }
316 
317 int
318 aioread(int fd, caddr_t buf, int bufsz, off_t offset, int whence,
319     aio_result_t *resultp)
320 {
321 	return (_aiorw(fd, buf, bufsz, offset, whence, resultp, AIOREAD));
322 }
323 
324 int
325 aiowrite(int fd, caddr_t buf, int bufsz, off_t offset, int whence,
326     aio_result_t *resultp)
327 {
328 	return (_aiorw(fd, buf, bufsz, offset, whence, resultp, AIOWRITE));
329 }
330 
331 #if !defined(_LP64)
332 int
333 aioread64(int fd, caddr_t buf, int bufsz, off64_t offset, int whence,
334     aio_result_t *resultp)
335 {
336 	return (_aiorw(fd, buf, bufsz, offset, whence, resultp, AIOAREAD64));
337 }
338 
339 int
340 aiowrite64(int fd, caddr_t buf, int bufsz, off64_t offset, int whence,
341     aio_result_t *resultp)
342 {
343 	return (_aiorw(fd, buf, bufsz, offset, whence, resultp, AIOAWRITE64));
344 }
345 #endif	/* !defined(_LP64) */
346 
347 int
348 _aiorw(int fd, caddr_t buf, int bufsz, offset_t offset, int whence,
349     aio_result_t *resultp, int mode)
350 {
351 	aio_req_t *reqp;
352 	aio_args_t *ap;
353 	offset_t loffset;
354 	struct stat64 stat64;
355 	int error = 0;
356 	int kerr;
357 	int umode;
358 
359 	switch (whence) {
360 
361 	case SEEK_SET:
362 		loffset = offset;
363 		break;
364 	case SEEK_CUR:
365 		if ((loffset = llseek(fd, 0, SEEK_CUR)) == -1)
366 			error = -1;
367 		else
368 			loffset += offset;
369 		break;
370 	case SEEK_END:
371 		if (fstat64(fd, &stat64) == -1)
372 			error = -1;
373 		else
374 			loffset = offset + stat64.st_size;
375 		break;
376 	default:
377 		errno = EINVAL;
378 		error = -1;
379 	}
380 
381 	if (error)
382 		return (error);
383 
384 	/* initialize kaio */
385 	if (!_kaio_ok)
386 		_kaio_init();
387 
388 	/*
389 	 * _aio_do_request() needs the original request code (mode) to be able
390 	 * to choose the appropiate 32/64 bit function.  All other functions
391 	 * only require the difference between READ and WRITE (umode).
392 	 */
393 	if (mode == AIOAREAD64 || mode == AIOAWRITE64)
394 		umode = mode - AIOAREAD64;
395 	else
396 		umode = mode;
397 
398 	/*
399 	 * Try kernel aio first.
400 	 * If errno is ENOTSUP/EBADFD, fall back to the thread implementation.
401 	 */
402 	if (_kaio_ok > 0 && KAIO_SUPPORTED(fd)) {
403 		resultp->aio_errno = 0;
404 		sig_mutex_lock(&__aio_mutex);
405 		_kaio_outstand_cnt++;
406 		sig_mutex_unlock(&__aio_mutex);
407 		kerr = (int)_kaio(((resultp->aio_return == AIO_INPROGRESS) ?
408 		    (umode | AIO_POLL_BIT) : umode),
409 		    fd, buf, bufsz, loffset, resultp);
410 		if (kerr == 0) {
411 			return (0);
412 		}
413 		sig_mutex_lock(&__aio_mutex);
414 		_kaio_outstand_cnt--;
415 		sig_mutex_unlock(&__aio_mutex);
416 		if (errno != ENOTSUP && errno != EBADFD)
417 			return (-1);
418 		if (errno == EBADFD)
419 			SET_KAIO_NOT_SUPPORTED(fd);
420 	}
421 
422 	if (!__uaio_ok && __uaio_init() == -1)
423 		return (-1);
424 
425 	if ((reqp = _aio_req_alloc()) == NULL) {
426 		errno = EAGAIN;
427 		return (-1);
428 	}
429 
430 	/*
431 	 * _aio_do_request() checks reqp->req_op to differentiate
432 	 * between 32 and 64 bit access.
433 	 */
434 	reqp->req_op = mode;
435 	reqp->req_resultp = resultp;
436 	ap = &reqp->req_args;
437 	ap->fd = fd;
438 	ap->buf = buf;
439 	ap->bufsz = bufsz;
440 	ap->offset = loffset;
441 
442 	if (_aio_hash_insert(resultp, reqp) != 0) {
443 		_aio_req_free(reqp);
444 		errno = EINVAL;
445 		return (-1);
446 	}
447 	/*
448 	 * _aio_req_add() only needs the difference between READ and
449 	 * WRITE to choose the right worker queue.
450 	 */
451 	_aio_req_add(reqp, &__nextworker_rw, umode);
452 	return (0);
453 }
454 
455 int
456 aiocancel(aio_result_t *resultp)
457 {
458 	aio_req_t *reqp;
459 	aio_worker_t *aiowp;
460 	int ret;
461 	int done = 0;
462 	int canceled = 0;
463 
464 	if (!__uaio_ok) {
465 		errno = EINVAL;
466 		return (-1);
467 	}
468 
469 	sig_mutex_lock(&__aio_mutex);
470 	reqp = _aio_hash_find(resultp);
471 	if (reqp == NULL) {
472 		if (_aio_outstand_cnt == _aio_req_done_cnt)
473 			errno = EINVAL;
474 		else
475 			errno = EACCES;
476 		ret = -1;
477 	} else {
478 		aiowp = reqp->req_worker;
479 		sig_mutex_lock(&aiowp->work_qlock1);
480 		(void) _aio_cancel_req(aiowp, reqp, &canceled, &done);
481 		sig_mutex_unlock(&aiowp->work_qlock1);
482 
483 		if (canceled) {
484 			ret = 0;
485 		} else {
486 			if (_aio_outstand_cnt == 0 ||
487 			    _aio_outstand_cnt == _aio_req_done_cnt)
488 				errno = EINVAL;
489 			else
490 				errno = EACCES;
491 			ret = -1;
492 		}
493 	}
494 	sig_mutex_unlock(&__aio_mutex);
495 	return (ret);
496 }
497 
498 /* ARGSUSED */
499 static void
500 _aiowait_cleanup(void *arg)
501 {
502 	sig_mutex_lock(&__aio_mutex);
503 	_aiowait_flag--;
504 	sig_mutex_unlock(&__aio_mutex);
505 }
506 
507 /*
508  * This must be asynch safe and cancel safe
509  */
510 aio_result_t *
511 aiowait(struct timeval *uwait)
512 {
513 	aio_result_t *uresultp;
514 	aio_result_t *kresultp;
515 	aio_result_t *resultp;
516 	int dontblock;
517 	int timedwait = 0;
518 	int kaio_errno = 0;
519 	struct timeval twait;
520 	struct timeval *wait = NULL;
521 	hrtime_t hrtend;
522 	hrtime_t hres;
523 
524 	if (uwait) {
525 		/*
526 		 * Check for a valid specified wait time.
527 		 * If it is invalid, fail the call right away.
528 		 */
529 		if (uwait->tv_sec < 0 || uwait->tv_usec < 0 ||
530 		    uwait->tv_usec >= MICROSEC) {
531 			errno = EINVAL;
532 			return ((aio_result_t *)-1);
533 		}
534 
535 		if (uwait->tv_sec > 0 || uwait->tv_usec > 0) {
536 			hrtend = gethrtime() +
537 			    (hrtime_t)uwait->tv_sec * NANOSEC +
538 			    (hrtime_t)uwait->tv_usec * (NANOSEC / MICROSEC);
539 			twait = *uwait;
540 			wait = &twait;
541 			timedwait++;
542 		} else {
543 			/* polling */
544 			sig_mutex_lock(&__aio_mutex);
545 			if (_kaio_outstand_cnt == 0) {
546 				kresultp = (aio_result_t *)-1;
547 			} else {
548 				kresultp = (aio_result_t *)_kaio(AIOWAIT,
549 				    (struct timeval *)-1, 1);
550 				if (kresultp != (aio_result_t *)-1 &&
551 				    kresultp != NULL &&
552 				    kresultp != (aio_result_t *)1) {
553 					_kaio_outstand_cnt--;
554 					sig_mutex_unlock(&__aio_mutex);
555 					return (kresultp);
556 				}
557 			}
558 			uresultp = _aio_req_done();
559 			sig_mutex_unlock(&__aio_mutex);
560 			if (uresultp != NULL &&
561 			    uresultp != (aio_result_t *)-1) {
562 				return (uresultp);
563 			}
564 			if (uresultp == (aio_result_t *)-1 &&
565 			    kresultp == (aio_result_t *)-1) {
566 				errno = EINVAL;
567 				return ((aio_result_t *)-1);
568 			} else {
569 				return (NULL);
570 			}
571 		}
572 	}
573 
574 	for (;;) {
575 		sig_mutex_lock(&__aio_mutex);
576 		uresultp = _aio_req_done();
577 		if (uresultp != NULL && uresultp != (aio_result_t *)-1) {
578 			sig_mutex_unlock(&__aio_mutex);
579 			resultp = uresultp;
580 			break;
581 		}
582 		_aiowait_flag++;
583 		dontblock = (uresultp == (aio_result_t *)-1);
584 		if (dontblock && _kaio_outstand_cnt == 0) {
585 			kresultp = (aio_result_t *)-1;
586 			kaio_errno = EINVAL;
587 		} else {
588 			sig_mutex_unlock(&__aio_mutex);
589 			pthread_cleanup_push(_aiowait_cleanup, NULL);
590 			_cancel_prologue();
591 			kresultp = (aio_result_t *)_kaio(AIOWAIT,
592 			    wait, dontblock);
593 			_cancel_epilogue();
594 			pthread_cleanup_pop(0);
595 			sig_mutex_lock(&__aio_mutex);
596 			kaio_errno = errno;
597 		}
598 		_aiowait_flag--;
599 		sig_mutex_unlock(&__aio_mutex);
600 		if (kresultp == (aio_result_t *)1) {
601 			/* aiowait() awakened by an aionotify() */
602 			continue;
603 		} else if (kresultp != NULL &&
604 		    kresultp != (aio_result_t *)-1) {
605 			resultp = kresultp;
606 			sig_mutex_lock(&__aio_mutex);
607 			_kaio_outstand_cnt--;
608 			sig_mutex_unlock(&__aio_mutex);
609 			break;
610 		} else if (kresultp == (aio_result_t *)-1 &&
611 		    kaio_errno == EINVAL &&
612 		    uresultp == (aio_result_t *)-1) {
613 			errno = kaio_errno;
614 			resultp = (aio_result_t *)-1;
615 			break;
616 		} else if (kresultp == (aio_result_t *)-1 &&
617 		    kaio_errno == EINTR) {
618 			errno = kaio_errno;
619 			resultp = (aio_result_t *)-1;
620 			break;
621 		} else if (timedwait) {
622 			hres = hrtend - gethrtime();
623 			if (hres <= 0) {
624 				/* time is up; return */
625 				resultp = NULL;
626 				break;
627 			} else {
628 				/*
629 				 * Some time left.  Round up the remaining time
630 				 * in nanoseconds to microsec.  Retry the call.
631 				 */
632 				hres += (NANOSEC / MICROSEC) - 1;
633 				wait->tv_sec = hres / NANOSEC;
634 				wait->tv_usec =
635 				    (hres % NANOSEC) / (NANOSEC / MICROSEC);
636 			}
637 		} else {
638 			ASSERT(kresultp == NULL && uresultp == NULL);
639 			resultp = NULL;
640 			continue;
641 		}
642 	}
643 	return (resultp);
644 }
645 
646 /*
647  * _aio_get_timedelta calculates the remaining time and stores the result
648  * into timespec_t *wait.
649  */
650 
651 int
652 _aio_get_timedelta(timespec_t *end, timespec_t *wait)
653 {
654 	int	ret = 0;
655 	struct	timeval cur;
656 	timespec_t curtime;
657 
658 	(void) gettimeofday(&cur, NULL);
659 	curtime.tv_sec = cur.tv_sec;
660 	curtime.tv_nsec = cur.tv_usec * 1000;   /* convert us to ns */
661 
662 	if (end->tv_sec >= curtime.tv_sec) {
663 		wait->tv_sec = end->tv_sec - curtime.tv_sec;
664 		if (end->tv_nsec >= curtime.tv_nsec) {
665 			wait->tv_nsec = end->tv_nsec - curtime.tv_nsec;
666 			if (wait->tv_sec == 0 && wait->tv_nsec == 0)
667 				ret = -1;	/* timer expired */
668 		} else {
669 			if (end->tv_sec > curtime.tv_sec) {
670 				wait->tv_sec -= 1;
671 				wait->tv_nsec = NANOSEC -
672 				    (curtime.tv_nsec - end->tv_nsec);
673 			} else {
674 				ret = -1;	/* timer expired */
675 			}
676 		}
677 	} else {
678 		ret = -1;
679 	}
680 	return (ret);
681 }
682 
683 /*
684  * If closing by file descriptor: we will simply cancel all the outstanding
685  * aio`s and return.  Those aio's in question will have either noticed the
686  * cancellation notice before, during, or after initiating io.
687  */
688 int
689 aiocancel_all(int fd)
690 {
691 	aio_req_t *reqp;
692 	aio_req_t **reqpp;
693 	aio_worker_t *first;
694 	aio_worker_t *next;
695 	int canceled = 0;
696 	int done = 0;
697 	int cancelall = 0;
698 
699 	sig_mutex_lock(&__aio_mutex);
700 
701 	if (_aio_outstand_cnt == 0) {
702 		sig_mutex_unlock(&__aio_mutex);
703 		return (AIO_ALLDONE);
704 	}
705 
706 	/*
707 	 * Cancel requests from the read/write workers' queues.
708 	 */
709 	first = __nextworker_rw;
710 	next = first;
711 	do {
712 		_aio_cancel_work(next, fd, &canceled, &done);
713 	} while ((next = next->work_forw) != first);
714 
715 	/*
716 	 * finally, check if there are requests on the done queue that
717 	 * should be canceled.
718 	 */
719 	if (fd < 0)
720 		cancelall = 1;
721 	reqpp = &_aio_done_tail;
722 	while ((reqp = *reqpp) != NULL) {
723 		if (cancelall || reqp->req_args.fd == fd) {
724 			*reqpp = reqp->req_next;
725 			_aio_donecnt--;
726 			(void) _aio_hash_del(reqp->req_resultp);
727 			_aio_req_free(reqp);
728 		} else
729 			reqpp = &reqp->req_next;
730 	}
731 	if (cancelall) {
732 		ASSERT(_aio_donecnt == 0);
733 		_aio_done_head = NULL;
734 	}
735 	sig_mutex_unlock(&__aio_mutex);
736 
737 	if (canceled && done == 0)
738 		return (AIO_CANCELED);
739 	else if (done && canceled == 0)
740 		return (AIO_ALLDONE);
741 	else if ((canceled + done == 0) && KAIO_SUPPORTED(fd))
742 		return ((int)_kaio(AIOCANCEL, fd, NULL));
743 	return (AIO_NOTCANCELED);
744 }
745 
746 /*
747  * Cancel requests from a given work queue.  If the file descriptor
748  * parameter, fd, is non-negative, then only cancel those requests
749  * in this queue that are to this file descriptor.  If the fd
750  * parameter is -1, then cancel all requests.
751  */
752 static void
753 _aio_cancel_work(aio_worker_t *aiowp, int fd, int *canceled, int *done)
754 {
755 	aio_req_t *reqp;
756 
757 	sig_mutex_lock(&aiowp->work_qlock1);
758 	/*
759 	 * cancel queued requests first.
760 	 */
761 	reqp = aiowp->work_tail1;
762 	while (reqp != NULL) {
763 		if (fd < 0 || reqp->req_args.fd == fd) {
764 			if (_aio_cancel_req(aiowp, reqp, canceled, done)) {
765 				/*
766 				 * Callers locks were dropped.
767 				 * reqp is invalid; start traversing
768 				 * the list from the beginning again.
769 				 */
770 				reqp = aiowp->work_tail1;
771 				continue;
772 			}
773 		}
774 		reqp = reqp->req_next;
775 	}
776 	/*
777 	 * Since the queued requests have been canceled, there can
778 	 * only be one inprogress request that should be canceled.
779 	 */
780 	if ((reqp = aiowp->work_req) != NULL &&
781 	    (fd < 0 || reqp->req_args.fd == fd))
782 		(void) _aio_cancel_req(aiowp, reqp, canceled, done);
783 	sig_mutex_unlock(&aiowp->work_qlock1);
784 }
785 
786 /*
787  * Cancel a request.  Return 1 if the callers locks were temporarily
788  * dropped, otherwise return 0.
789  */
790 int
791 _aio_cancel_req(aio_worker_t *aiowp, aio_req_t *reqp, int *canceled, int *done)
792 {
793 	int ostate = reqp->req_state;
794 
795 	ASSERT(MUTEX_HELD(&__aio_mutex));
796 	ASSERT(MUTEX_HELD(&aiowp->work_qlock1));
797 	if (ostate == AIO_REQ_CANCELED)
798 		return (0);
799 	if (ostate == AIO_REQ_DONE || ostate == AIO_REQ_DONEQ) {
800 		(*done)++;
801 		return (0);
802 	}
803 	if (reqp->req_op == AIOFSYNC && reqp != aiowp->work_req) {
804 		ASSERT(POSIX_AIO(reqp));
805 		/* Cancel the queued aio_fsync() request */
806 		if (!reqp->req_head->lio_canned) {
807 			reqp->req_head->lio_canned = 1;
808 			_aio_outstand_cnt--;
809 			(*canceled)++;
810 		}
811 		return (0);
812 	}
813 	reqp->req_state = AIO_REQ_CANCELED;
814 	_aio_req_del(aiowp, reqp, ostate);
815 	(void) _aio_hash_del(reqp->req_resultp);
816 	(*canceled)++;
817 	if (reqp == aiowp->work_req) {
818 		ASSERT(ostate == AIO_REQ_INPROGRESS);
819 		/*
820 		 * Set the result values now, before _aiodone() is called.
821 		 * We do this because the application can expect aio_return
822 		 * and aio_errno to be set to -1 and ECANCELED, respectively,
823 		 * immediately after a successful return from aiocancel()
824 		 * or aio_cancel().
825 		 */
826 		_aio_set_result(reqp, -1, ECANCELED);
827 		(void) thr_kill(aiowp->work_tid, SIGAIOCANCEL);
828 		return (0);
829 	}
830 	if (!POSIX_AIO(reqp)) {
831 		_aio_outstand_cnt--;
832 		_aio_set_result(reqp, -1, ECANCELED);
833 		return (0);
834 	}
835 	sig_mutex_unlock(&aiowp->work_qlock1);
836 	sig_mutex_unlock(&__aio_mutex);
837 	_aiodone(reqp, -1, ECANCELED);
838 	sig_mutex_lock(&__aio_mutex);
839 	sig_mutex_lock(&aiowp->work_qlock1);
840 	return (1);
841 }
842 
843 int
844 _aio_create_worker(aio_req_t *reqp, int mode)
845 {
846 	aio_worker_t *aiowp, **workers, **nextworker;
847 	int *aio_workerscnt;
848 	void *(*func)(void *);
849 	sigset_t oset;
850 	int error;
851 
852 	/*
853 	 * Put the new worker thread in the right queue.
854 	 */
855 	switch (mode) {
856 	case AIOREAD:
857 	case AIOWRITE:
858 	case AIOAREAD:
859 	case AIOAWRITE:
860 #if !defined(_LP64)
861 	case AIOAREAD64:
862 	case AIOAWRITE64:
863 #endif
864 		workers = &__workers_rw;
865 		nextworker = &__nextworker_rw;
866 		aio_workerscnt = &__rw_workerscnt;
867 		func = _aio_do_request;
868 		break;
869 	case AIONOTIFY:
870 		workers = &__workers_no;
871 		nextworker = &__nextworker_no;
872 		func = _aio_do_notify;
873 		aio_workerscnt = &__no_workerscnt;
874 		break;
875 	default:
876 		aio_panic("_aio_create_worker: invalid mode");
877 		break;
878 	}
879 
880 	if ((aiowp = _aio_worker_alloc()) == NULL)
881 		return (-1);
882 
883 	if (reqp) {
884 		reqp->req_state = AIO_REQ_QUEUED;
885 		reqp->req_worker = aiowp;
886 		aiowp->work_head1 = reqp;
887 		aiowp->work_tail1 = reqp;
888 		aiowp->work_next1 = reqp;
889 		aiowp->work_count1 = 1;
890 		aiowp->work_minload1 = 1;
891 	}
892 
893 	(void) pthread_sigmask(SIG_SETMASK, &maskset, &oset);
894 	error = thr_create(NULL, AIOSTKSIZE, func, aiowp,
895 	    THR_DAEMON | THR_SUSPENDED, &aiowp->work_tid);
896 	(void) pthread_sigmask(SIG_SETMASK, &oset, NULL);
897 	if (error) {
898 		if (reqp) {
899 			reqp->req_state = 0;
900 			reqp->req_worker = NULL;
901 		}
902 		_aio_worker_free(aiowp);
903 		return (-1);
904 	}
905 
906 	lmutex_lock(&__aio_mutex);
907 	(*aio_workerscnt)++;
908 	if (*workers == NULL) {
909 		aiowp->work_forw = aiowp;
910 		aiowp->work_backw = aiowp;
911 		*nextworker = aiowp;
912 		*workers = aiowp;
913 	} else {
914 		aiowp->work_backw = (*workers)->work_backw;
915 		aiowp->work_forw = (*workers);
916 		(*workers)->work_backw->work_forw = aiowp;
917 		(*workers)->work_backw = aiowp;
918 	}
919 	_aio_worker_cnt++;
920 	lmutex_unlock(&__aio_mutex);
921 
922 	(void) thr_continue(aiowp->work_tid);
923 
924 	return (0);
925 }
926 
927 /*
928  * This is the worker's main routine.
929  * The task of this function is to execute all queued requests;
930  * once the last pending request is executed this function will block
931  * in _aio_idle().  A new incoming request must wakeup this thread to
932  * restart the work.
933  * Every worker has an own work queue.  The queue lock is required
934  * to synchronize the addition of new requests for this worker or
935  * cancellation of pending/running requests.
936  *
937  * Cancellation scenarios:
938  * The cancellation of a request is being done asynchronously using
939  * _aio_cancel_req() from another thread context.
940  * A queued request can be cancelled in different manners :
941  * a) request is queued but not "in progress" or "done" (AIO_REQ_QUEUED):
942  *	- lock the queue -> remove the request -> unlock the queue
943  *	- this function/thread does not detect this cancellation process
944  * b) request is in progress (AIO_REQ_INPROGRESS) :
945  *	- this function first allow the cancellation of the running
946  *	  request with the flag "work_cancel_flg=1"
947  * 		see _aio_req_get() -> _aio_cancel_on()
948  *	  During this phase, it is allowed to interrupt the worker
949  *	  thread running the request (this thread) using the SIGAIOCANCEL
950  *	  signal.
951  *	  Once this thread returns from the kernel (because the request
952  *	  is just done), then it must disable a possible cancellation
953  *	  and proceed to finish the request.  To disable the cancellation
954  *	  this thread must use _aio_cancel_off() to set "work_cancel_flg=0".
955  * c) request is already done (AIO_REQ_DONE || AIO_REQ_DONEQ):
956  *	  same procedure as in a)
957  *
958  * To b)
959  *	This thread uses sigsetjmp() to define the position in the code, where
960  *	it wish to continue working in the case that a SIGAIOCANCEL signal
961  *	is detected.
962  *	Normally this thread should get the cancellation signal during the
963  *	kernel phase (reading or writing).  In that case the signal handler
964  *	aiosigcancelhndlr() is activated using the worker thread context,
965  *	which again will use the siglongjmp() function to break the standard
966  *	code flow and jump to the "sigsetjmp" position, provided that
967  *	"work_cancel_flg" is set to "1".
968  *	Because the "work_cancel_flg" is only manipulated by this worker
969  *	thread and it can only run on one CPU at a given time, it is not
970  *	necessary to protect that flag with the queue lock.
971  *	Returning from the kernel (read or write system call) we must
972  *	first disable the use of the SIGAIOCANCEL signal and accordingly
973  *	the use of the siglongjmp() function to prevent a possible deadlock:
974  *	- It can happens that this worker thread returns from the kernel and
975  *	  blocks in "work_qlock1",
976  *	- then a second thread cancels the apparently "in progress" request
977  *	  and sends the SIGAIOCANCEL signal to the worker thread,
978  *	- the worker thread gets assigned the "work_qlock1" and will returns
979  *	  from the kernel,
980  *	- the kernel detects the pending signal and activates the signal
981  *	  handler instead,
982  *	- if the "work_cancel_flg" is still set then the signal handler
983  *	  should use siglongjmp() to cancel the "in progress" request and
984  *	  it would try to acquire the same work_qlock1 in _aio_req_get()
985  *	  for a second time => deadlock.
986  *	To avoid that situation we disable the cancellation of the request
987  *	in progress BEFORE we try to acquire the work_qlock1.
988  *	In that case the signal handler will not call siglongjmp() and the
989  *	worker thread will continue running the standard code flow.
990  *	Then this thread must check the AIO_REQ_CANCELED flag to emulate
991  *	an eventually required siglongjmp() freeing the work_qlock1 and
992  *	avoiding a deadlock.
993  */
994 void *
995 _aio_do_request(void *arglist)
996 {
997 	aio_worker_t *aiowp = (aio_worker_t *)arglist;
998 	ulwp_t *self = curthread;
999 	struct aio_args *arg;
1000 	aio_req_t *reqp;		/* current AIO request */
1001 	ssize_t retval;
1002 	int append;
1003 	int error;
1004 
1005 	if (pthread_setspecific(_aio_key, aiowp) != 0)
1006 		aio_panic("_aio_do_request, pthread_setspecific()");
1007 	(void) pthread_sigmask(SIG_SETMASK, &_worker_set, NULL);
1008 	ASSERT(aiowp->work_req == NULL);
1009 
1010 	/*
1011 	 * We resume here when an operation is cancelled.
1012 	 * On first entry, aiowp->work_req == NULL, so all
1013 	 * we do is block SIGAIOCANCEL.
1014 	 */
1015 	(void) sigsetjmp(aiowp->work_jmp_buf, 0);
1016 	ASSERT(self->ul_sigdefer == 0);
1017 
1018 	sigoff(self);	/* block SIGAIOCANCEL */
1019 	if (aiowp->work_req != NULL)
1020 		_aio_finish_request(aiowp, -1, ECANCELED);
1021 
1022 	for (;;) {
1023 		/*
1024 		 * Put completed requests on aio_done_list.  This has
1025 		 * to be done as part of the main loop to ensure that
1026 		 * we don't artificially starve any aiowait'ers.
1027 		 */
1028 		if (aiowp->work_done1)
1029 			_aio_work_done(aiowp);
1030 
1031 top:
1032 		/* consume any deferred SIGAIOCANCEL signal here */
1033 		sigon(self);
1034 		sigoff(self);
1035 
1036 		while ((reqp = _aio_req_get(aiowp)) == NULL) {
1037 			if (_aio_idle(aiowp) != 0)
1038 				goto top;
1039 		}
1040 		arg = &reqp->req_args;
1041 		ASSERT(reqp->req_state == AIO_REQ_INPROGRESS ||
1042 		    reqp->req_state == AIO_REQ_CANCELED);
1043 		error = 0;
1044 
1045 		switch (reqp->req_op) {
1046 		case AIOREAD:
1047 		case AIOAREAD:
1048 			sigon(self);	/* unblock SIGAIOCANCEL */
1049 			retval = pread(arg->fd, arg->buf,
1050 			    arg->bufsz, arg->offset);
1051 			if (retval == -1) {
1052 				if (errno == ESPIPE) {
1053 					retval = read(arg->fd,
1054 					    arg->buf, arg->bufsz);
1055 					if (retval == -1)
1056 						error = errno;
1057 				} else {
1058 					error = errno;
1059 				}
1060 			}
1061 			sigoff(self);	/* block SIGAIOCANCEL */
1062 			break;
1063 		case AIOWRITE:
1064 		case AIOAWRITE:
1065 			/*
1066 			 * The SUSv3 POSIX spec for aio_write() states:
1067 			 *	If O_APPEND is set for the file descriptor,
1068 			 *	write operations append to the file in the
1069 			 *	same order as the calls were made.
1070 			 * but, somewhat inconsistently, it requires pwrite()
1071 			 * to ignore the O_APPEND setting.  So we have to use
1072 			 * fcntl() to get the open modes and call write() for
1073 			 * the O_APPEND case.
1074 			 */
1075 			append = (__fcntl(arg->fd, F_GETFL) & O_APPEND);
1076 			sigon(self);	/* unblock SIGAIOCANCEL */
1077 			retval = append?
1078 			    write(arg->fd, arg->buf, arg->bufsz) :
1079 			    pwrite(arg->fd, arg->buf, arg->bufsz,
1080 			    arg->offset);
1081 			if (retval == -1) {
1082 				if (errno == ESPIPE) {
1083 					retval = write(arg->fd,
1084 					    arg->buf, arg->bufsz);
1085 					if (retval == -1)
1086 						error = errno;
1087 				} else {
1088 					error = errno;
1089 				}
1090 			}
1091 			sigoff(self);	/* block SIGAIOCANCEL */
1092 			break;
1093 #if !defined(_LP64)
1094 		case AIOAREAD64:
1095 			sigon(self);	/* unblock SIGAIOCANCEL */
1096 			retval = pread64(arg->fd, arg->buf,
1097 			    arg->bufsz, arg->offset);
1098 			if (retval == -1) {
1099 				if (errno == ESPIPE) {
1100 					retval = read(arg->fd,
1101 					    arg->buf, arg->bufsz);
1102 					if (retval == -1)
1103 						error = errno;
1104 				} else {
1105 					error = errno;
1106 				}
1107 			}
1108 			sigoff(self);	/* block SIGAIOCANCEL */
1109 			break;
1110 		case AIOAWRITE64:
1111 			/*
1112 			 * The SUSv3 POSIX spec for aio_write() states:
1113 			 *	If O_APPEND is set for the file descriptor,
1114 			 *	write operations append to the file in the
1115 			 *	same order as the calls were made.
1116 			 * but, somewhat inconsistently, it requires pwrite()
1117 			 * to ignore the O_APPEND setting.  So we have to use
1118 			 * fcntl() to get the open modes and call write() for
1119 			 * the O_APPEND case.
1120 			 */
1121 			append = (__fcntl(arg->fd, F_GETFL) & O_APPEND);
1122 			sigon(self);	/* unblock SIGAIOCANCEL */
1123 			retval = append?
1124 			    write(arg->fd, arg->buf, arg->bufsz) :
1125 			    pwrite64(arg->fd, arg->buf, arg->bufsz,
1126 			    arg->offset);
1127 			if (retval == -1) {
1128 				if (errno == ESPIPE) {
1129 					retval = write(arg->fd,
1130 					    arg->buf, arg->bufsz);
1131 					if (retval == -1)
1132 						error = errno;
1133 				} else {
1134 					error = errno;
1135 				}
1136 			}
1137 			sigoff(self);	/* block SIGAIOCANCEL */
1138 			break;
1139 #endif	/* !defined(_LP64) */
1140 		case AIOFSYNC:
1141 			if (_aio_fsync_del(aiowp, reqp))
1142 				goto top;
1143 			ASSERT(reqp->req_head == NULL);
1144 			/*
1145 			 * All writes for this fsync request are now
1146 			 * acknowledged.  Now make these writes visible
1147 			 * and put the final request into the hash table.
1148 			 */
1149 			if (reqp->req_state == AIO_REQ_CANCELED) {
1150 				/* EMPTY */;
1151 			} else if (arg->offset == O_SYNC) {
1152 				if ((retval = __fdsync(arg->fd, FSYNC)) == -1)
1153 					error = errno;
1154 			} else {
1155 				if ((retval = __fdsync(arg->fd, FDSYNC)) == -1)
1156 					error = errno;
1157 			}
1158 			if (_aio_hash_insert(reqp->req_resultp, reqp) != 0)
1159 				aio_panic("_aio_do_request(): AIOFSYNC: "
1160 				    "request already in hash table");
1161 			break;
1162 		default:
1163 			aio_panic("_aio_do_request, bad op");
1164 		}
1165 
1166 		_aio_finish_request(aiowp, retval, error);
1167 	}
1168 	/* NOTREACHED */
1169 	return (NULL);
1170 }
1171 
1172 /*
1173  * Perform the tail processing for _aio_do_request().
1174  * The in-progress request may or may not have been cancelled.
1175  */
1176 static void
1177 _aio_finish_request(aio_worker_t *aiowp, ssize_t retval, int error)
1178 {
1179 	aio_req_t *reqp;
1180 
1181 	sig_mutex_lock(&aiowp->work_qlock1);
1182 	if ((reqp = aiowp->work_req) == NULL)
1183 		sig_mutex_unlock(&aiowp->work_qlock1);
1184 	else {
1185 		aiowp->work_req = NULL;
1186 		if (reqp->req_state == AIO_REQ_CANCELED) {
1187 			retval = -1;
1188 			error = ECANCELED;
1189 		}
1190 		if (!POSIX_AIO(reqp)) {
1191 			int notify;
1192 			sig_mutex_unlock(&aiowp->work_qlock1);
1193 			sig_mutex_lock(&__aio_mutex);
1194 			if (reqp->req_state == AIO_REQ_INPROGRESS)
1195 				reqp->req_state = AIO_REQ_DONE;
1196 			/*
1197 			 * If it was canceled, this request will not be
1198 			 * added to done list. Just free it.
1199 			 */
1200 			if (error == ECANCELED) {
1201 				_aio_outstand_cnt--;
1202 				_aio_req_free(reqp);
1203 			} else {
1204 				_aio_set_result(reqp, retval, error);
1205 				_aio_req_done_cnt++;
1206 			}
1207 			/*
1208 			 * Notify any thread that may have blocked
1209 			 * because it saw an outstanding request.
1210 			 */
1211 			notify = 0;
1212 			if (_aio_outstand_cnt == 0 && _aiowait_flag) {
1213 				notify = 1;
1214 			}
1215 			sig_mutex_unlock(&__aio_mutex);
1216 			if (notify) {
1217 				(void) _kaio(AIONOTIFY);
1218 			}
1219 		} else {
1220 			if (reqp->req_state == AIO_REQ_INPROGRESS)
1221 				reqp->req_state = AIO_REQ_DONE;
1222 			sig_mutex_unlock(&aiowp->work_qlock1);
1223 			_aiodone(reqp, retval, error);
1224 		}
1225 	}
1226 }
1227 
1228 void
1229 _aio_req_mark_done(aio_req_t *reqp)
1230 {
1231 #if !defined(_LP64)
1232 	if (reqp->req_largefile)
1233 		((aiocb64_t *)reqp->req_aiocbp)->aio_state = USERAIO_DONE;
1234 	else
1235 #endif
1236 		((aiocb_t *)reqp->req_aiocbp)->aio_state = USERAIO_DONE;
1237 }
1238 
1239 /*
1240  * Sleep for 'ticks' clock ticks to give somebody else a chance to run,
1241  * hopefully to consume one of our queued signals.
1242  */
1243 static void
1244 _aio_delay(int ticks)
1245 {
1246 	(void) usleep(ticks * (MICROSEC / hz));
1247 }
1248 
1249 /*
1250  * Actually send the notifications.
1251  * We could block indefinitely here if the application
1252  * is not listening for the signal or port notifications.
1253  */
1254 static void
1255 send_notification(notif_param_t *npp)
1256 {
1257 	extern int __sigqueue(pid_t pid, int signo,
1258 	    /* const union sigval */ void *value, int si_code, int block);
1259 
1260 	if (npp->np_signo)
1261 		(void) __sigqueue(__pid, npp->np_signo, npp->np_user,
1262 		    SI_ASYNCIO, 1);
1263 	else if (npp->np_port >= 0)
1264 		(void) _port_dispatch(npp->np_port, 0, PORT_SOURCE_AIO,
1265 		    npp->np_event, npp->np_object, npp->np_user);
1266 
1267 	if (npp->np_lio_signo)
1268 		(void) __sigqueue(__pid, npp->np_lio_signo, npp->np_lio_user,
1269 		    SI_ASYNCIO, 1);
1270 	else if (npp->np_lio_port >= 0)
1271 		(void) _port_dispatch(npp->np_lio_port, 0, PORT_SOURCE_AIO,
1272 		    npp->np_lio_event, npp->np_lio_object, npp->np_lio_user);
1273 }
1274 
1275 /*
1276  * Asynchronous notification worker.
1277  */
1278 void *
1279 _aio_do_notify(void *arg)
1280 {
1281 	aio_worker_t *aiowp = (aio_worker_t *)arg;
1282 	aio_req_t *reqp;
1283 
1284 	/*
1285 	 * This isn't really necessary.  All signals are blocked.
1286 	 */
1287 	if (pthread_setspecific(_aio_key, aiowp) != 0)
1288 		aio_panic("_aio_do_notify, pthread_setspecific()");
1289 
1290 	/*
1291 	 * Notifications are never cancelled.
1292 	 * All signals remain blocked, forever.
1293 	 */
1294 	for (;;) {
1295 		while ((reqp = _aio_req_get(aiowp)) == NULL) {
1296 			if (_aio_idle(aiowp) != 0)
1297 				aio_panic("_aio_do_notify: _aio_idle() failed");
1298 		}
1299 		send_notification(&reqp->req_notify);
1300 		_aio_req_free(reqp);
1301 	}
1302 
1303 	/* NOTREACHED */
1304 	return (NULL);
1305 }
1306 
1307 /*
1308  * Do the completion semantics for a request that was either canceled
1309  * by _aio_cancel_req() or was completed by _aio_do_request().
1310  */
1311 static void
1312 _aiodone(aio_req_t *reqp, ssize_t retval, int error)
1313 {
1314 	aio_result_t *resultp = reqp->req_resultp;
1315 	int notify = 0;
1316 	aio_lio_t *head;
1317 	int sigev_none;
1318 	int sigev_signal;
1319 	int sigev_thread;
1320 	int sigev_port;
1321 	notif_param_t np;
1322 
1323 	/*
1324 	 * We call _aiodone() only for Posix I/O.
1325 	 */
1326 	ASSERT(POSIX_AIO(reqp));
1327 
1328 	sigev_none = 0;
1329 	sigev_signal = 0;
1330 	sigev_thread = 0;
1331 	sigev_port = 0;
1332 	np.np_signo = 0;
1333 	np.np_port = -1;
1334 	np.np_lio_signo = 0;
1335 	np.np_lio_port = -1;
1336 
1337 	switch (reqp->req_sigevent.sigev_notify) {
1338 	case SIGEV_NONE:
1339 		sigev_none = 1;
1340 		break;
1341 	case SIGEV_SIGNAL:
1342 		sigev_signal = 1;
1343 		break;
1344 	case SIGEV_THREAD:
1345 		sigev_thread = 1;
1346 		break;
1347 	case SIGEV_PORT:
1348 		sigev_port = 1;
1349 		break;
1350 	default:
1351 		aio_panic("_aiodone: improper sigev_notify");
1352 		break;
1353 	}
1354 
1355 	/*
1356 	 * Figure out the notification parameters while holding __aio_mutex.
1357 	 * Actually perform the notifications after dropping __aio_mutex.
1358 	 * This allows us to sleep for a long time (if the notifications
1359 	 * incur delays) without impeding other async I/O operations.
1360 	 */
1361 
1362 	sig_mutex_lock(&__aio_mutex);
1363 
1364 	if (sigev_signal) {
1365 		if ((np.np_signo = reqp->req_sigevent.sigev_signo) != 0)
1366 			notify = 1;
1367 		np.np_user = reqp->req_sigevent.sigev_value.sival_ptr;
1368 	} else if (sigev_thread | sigev_port) {
1369 		if ((np.np_port = reqp->req_sigevent.sigev_signo) >= 0)
1370 			notify = 1;
1371 		np.np_event = reqp->req_op;
1372 		if (np.np_event == AIOFSYNC && reqp->req_largefile)
1373 			np.np_event = AIOFSYNC64;
1374 		np.np_object = (uintptr_t)reqp->req_aiocbp;
1375 		np.np_user = reqp->req_sigevent.sigev_value.sival_ptr;
1376 	}
1377 
1378 	if (resultp->aio_errno == EINPROGRESS)
1379 		_aio_set_result(reqp, retval, error);
1380 
1381 	_aio_outstand_cnt--;
1382 
1383 	head = reqp->req_head;
1384 	reqp->req_head = NULL;
1385 
1386 	if (sigev_none) {
1387 		_aio_enq_doneq(reqp);
1388 		reqp = NULL;
1389 	} else {
1390 		(void) _aio_hash_del(resultp);
1391 		_aio_req_mark_done(reqp);
1392 	}
1393 
1394 	_aio_waitn_wakeup();
1395 
1396 	/*
1397 	 * __aio_waitn() sets AIO_WAIT_INPROGRESS and
1398 	 * __aio_suspend() increments "_aio_kernel_suspend"
1399 	 * when they are waiting in the kernel for completed I/Os.
1400 	 *
1401 	 * _kaio(AIONOTIFY) awakes the corresponding function
1402 	 * in the kernel; then the corresponding __aio_waitn() or
1403 	 * __aio_suspend() function could reap the recently
1404 	 * completed I/Os (_aiodone()).
1405 	 */
1406 	if ((_aio_flags & AIO_WAIT_INPROGRESS) || _aio_kernel_suspend > 0)
1407 		(void) _kaio(AIONOTIFY);
1408 
1409 	sig_mutex_unlock(&__aio_mutex);
1410 
1411 	if (head != NULL) {
1412 		/*
1413 		 * If all the lio requests have completed,
1414 		 * prepare to notify the waiting thread.
1415 		 */
1416 		sig_mutex_lock(&head->lio_mutex);
1417 		ASSERT(head->lio_refcnt == head->lio_nent);
1418 		if (head->lio_refcnt == 1) {
1419 			int waiting = 0;
1420 			if (head->lio_mode == LIO_WAIT) {
1421 				if ((waiting = head->lio_waiting) != 0)
1422 					(void) cond_signal(&head->lio_cond_cv);
1423 			} else if (head->lio_port < 0) { /* none or signal */
1424 				if ((np.np_lio_signo = head->lio_signo) != 0)
1425 					notify = 1;
1426 				np.np_lio_user = head->lio_sigval.sival_ptr;
1427 			} else {			/* thread or port */
1428 				notify = 1;
1429 				np.np_lio_port = head->lio_port;
1430 				np.np_lio_event = head->lio_event;
1431 				np.np_lio_object =
1432 				    (uintptr_t)head->lio_sigevent;
1433 				np.np_lio_user = head->lio_sigval.sival_ptr;
1434 			}
1435 			head->lio_nent = head->lio_refcnt = 0;
1436 			sig_mutex_unlock(&head->lio_mutex);
1437 			if (waiting == 0)
1438 				_aio_lio_free(head);
1439 		} else {
1440 			head->lio_nent--;
1441 			head->lio_refcnt--;
1442 			sig_mutex_unlock(&head->lio_mutex);
1443 		}
1444 	}
1445 
1446 	/*
1447 	 * The request is completed; now perform the notifications.
1448 	 */
1449 	if (notify) {
1450 		if (reqp != NULL) {
1451 			/*
1452 			 * We usually put the request on the notification
1453 			 * queue because we don't want to block and delay
1454 			 * other operations behind us in the work queue.
1455 			 * Also we must never block on a cancel notification
1456 			 * because we are being called from an application
1457 			 * thread in this case and that could lead to deadlock
1458 			 * if no other thread is receiving notificatins.
1459 			 */
1460 			reqp->req_notify = np;
1461 			reqp->req_op = AIONOTIFY;
1462 			_aio_req_add(reqp, &__workers_no, AIONOTIFY);
1463 			reqp = NULL;
1464 		} else {
1465 			/*
1466 			 * We already put the request on the done queue,
1467 			 * so we can't queue it to the notification queue.
1468 			 * Just do the notification directly.
1469 			 */
1470 			send_notification(&np);
1471 		}
1472 	}
1473 
1474 	if (reqp != NULL)
1475 		_aio_req_free(reqp);
1476 }
1477 
1478 /*
1479  * Delete fsync requests from list head until there is
1480  * only one left.  Return 0 when there is only one,
1481  * otherwise return a non-zero value.
1482  */
1483 static int
1484 _aio_fsync_del(aio_worker_t *aiowp, aio_req_t *reqp)
1485 {
1486 	aio_lio_t *head = reqp->req_head;
1487 	int rval = 0;
1488 
1489 	ASSERT(reqp == aiowp->work_req);
1490 	sig_mutex_lock(&aiowp->work_qlock1);
1491 	sig_mutex_lock(&head->lio_mutex);
1492 	if (head->lio_refcnt > 1) {
1493 		head->lio_refcnt--;
1494 		head->lio_nent--;
1495 		aiowp->work_req = NULL;
1496 		sig_mutex_unlock(&head->lio_mutex);
1497 		sig_mutex_unlock(&aiowp->work_qlock1);
1498 		sig_mutex_lock(&__aio_mutex);
1499 		_aio_outstand_cnt--;
1500 		_aio_waitn_wakeup();
1501 		sig_mutex_unlock(&__aio_mutex);
1502 		_aio_req_free(reqp);
1503 		return (1);
1504 	}
1505 	ASSERT(head->lio_nent == 1 && head->lio_refcnt == 1);
1506 	reqp->req_head = NULL;
1507 	if (head->lio_canned)
1508 		reqp->req_state = AIO_REQ_CANCELED;
1509 	if (head->lio_mode == LIO_DESTROY) {
1510 		aiowp->work_req = NULL;
1511 		rval = 1;
1512 	}
1513 	sig_mutex_unlock(&head->lio_mutex);
1514 	sig_mutex_unlock(&aiowp->work_qlock1);
1515 	head->lio_refcnt--;
1516 	head->lio_nent--;
1517 	_aio_lio_free(head);
1518 	if (rval != 0)
1519 		_aio_req_free(reqp);
1520 	return (rval);
1521 }
1522 
1523 /*
1524  * A worker is set idle when its work queue is empty.
1525  * The worker checks again that it has no more work
1526  * and then goes to sleep waiting for more work.
1527  */
1528 int
1529 _aio_idle(aio_worker_t *aiowp)
1530 {
1531 	int error = 0;
1532 
1533 	sig_mutex_lock(&aiowp->work_qlock1);
1534 	if (aiowp->work_count1 == 0) {
1535 		ASSERT(aiowp->work_minload1 == 0);
1536 		aiowp->work_idleflg = 1;
1537 		/*
1538 		 * A cancellation handler is not needed here.
1539 		 * aio worker threads are never cancelled via pthread_cancel().
1540 		 */
1541 		error = sig_cond_wait(&aiowp->work_idle_cv,
1542 		    &aiowp->work_qlock1);
1543 		/*
1544 		 * The idle flag is normally cleared before worker is awakened
1545 		 * by aio_req_add().  On error (EINTR), we clear it ourself.
1546 		 */
1547 		if (error)
1548 			aiowp->work_idleflg = 0;
1549 	}
1550 	sig_mutex_unlock(&aiowp->work_qlock1);
1551 	return (error);
1552 }
1553 
1554 /*
1555  * A worker's completed AIO requests are placed onto a global
1556  * done queue.  The application is only sent a SIGIO signal if
1557  * the process has a handler enabled and it is not waiting via
1558  * aiowait().
1559  */
1560 static void
1561 _aio_work_done(aio_worker_t *aiowp)
1562 {
1563 	aio_req_t *reqp;
1564 
1565 	sig_mutex_lock(&aiowp->work_qlock1);
1566 	reqp = aiowp->work_prev1;
1567 	reqp->req_next = NULL;
1568 	aiowp->work_done1 = 0;
1569 	aiowp->work_tail1 = aiowp->work_next1;
1570 	if (aiowp->work_tail1 == NULL)
1571 		aiowp->work_head1 = NULL;
1572 	aiowp->work_prev1 = NULL;
1573 	sig_mutex_unlock(&aiowp->work_qlock1);
1574 	sig_mutex_lock(&__aio_mutex);
1575 	_aio_donecnt++;
1576 	_aio_outstand_cnt--;
1577 	_aio_req_done_cnt--;
1578 	ASSERT(_aio_donecnt > 0 &&
1579 	    _aio_outstand_cnt >= 0 &&
1580 	    _aio_req_done_cnt >= 0);
1581 	ASSERT(reqp != NULL);
1582 
1583 	if (_aio_done_tail == NULL) {
1584 		_aio_done_head = _aio_done_tail = reqp;
1585 	} else {
1586 		_aio_done_head->req_next = reqp;
1587 		_aio_done_head = reqp;
1588 	}
1589 
1590 	if (_aiowait_flag) {
1591 		sig_mutex_unlock(&__aio_mutex);
1592 		(void) _kaio(AIONOTIFY);
1593 	} else {
1594 		sig_mutex_unlock(&__aio_mutex);
1595 		if (_sigio_enabled)
1596 			(void) kill(__pid, SIGIO);
1597 	}
1598 }
1599 
1600 /*
1601  * The done queue consists of AIO requests that are in either the
1602  * AIO_REQ_DONE or AIO_REQ_CANCELED state.  Requests that were cancelled
1603  * are discarded.  If the done queue is empty then NULL is returned.
1604  * Otherwise the address of a done aio_result_t is returned.
1605  */
1606 aio_result_t *
1607 _aio_req_done(void)
1608 {
1609 	aio_req_t *reqp;
1610 	aio_result_t *resultp;
1611 
1612 	ASSERT(MUTEX_HELD(&__aio_mutex));
1613 
1614 	if ((reqp = _aio_done_tail) != NULL) {
1615 		if ((_aio_done_tail = reqp->req_next) == NULL)
1616 			_aio_done_head = NULL;
1617 		ASSERT(_aio_donecnt > 0);
1618 		_aio_donecnt--;
1619 		(void) _aio_hash_del(reqp->req_resultp);
1620 		resultp = reqp->req_resultp;
1621 		ASSERT(reqp->req_state == AIO_REQ_DONE);
1622 		_aio_req_free(reqp);
1623 		return (resultp);
1624 	}
1625 	/* is queue empty? */
1626 	if (reqp == NULL && _aio_outstand_cnt == 0) {
1627 		return ((aio_result_t *)-1);
1628 	}
1629 	return (NULL);
1630 }
1631 
1632 /*
1633  * Set the return and errno values for the application's use.
1634  *
1635  * For the Posix interfaces, we must set the return value first followed
1636  * by the errno value because the Posix interfaces allow for a change
1637  * in the errno value from EINPROGRESS to something else to signal
1638  * the completion of the asynchronous request.
1639  *
1640  * The opposite is true for the Solaris interfaces.  These allow for
1641  * a change in the return value from AIO_INPROGRESS to something else
1642  * to signal the completion of the asynchronous request.
1643  */
1644 void
1645 _aio_set_result(aio_req_t *reqp, ssize_t retval, int error)
1646 {
1647 	aio_result_t *resultp = reqp->req_resultp;
1648 
1649 	if (POSIX_AIO(reqp)) {
1650 		resultp->aio_return = retval;
1651 		membar_producer();
1652 		resultp->aio_errno = error;
1653 	} else {
1654 		resultp->aio_errno = error;
1655 		membar_producer();
1656 		resultp->aio_return = retval;
1657 	}
1658 }
1659 
1660 /*
1661  * Add an AIO request onto the next work queue.
1662  * A circular list of workers is used to choose the next worker.
1663  */
1664 void
1665 _aio_req_add(aio_req_t *reqp, aio_worker_t **nextworker, int mode)
1666 {
1667 	ulwp_t *self = curthread;
1668 	aio_worker_t *aiowp;
1669 	aio_worker_t *first;
1670 	int load_bal_flg = 1;
1671 	int found;
1672 
1673 	ASSERT(reqp->req_state != AIO_REQ_DONEQ);
1674 	reqp->req_next = NULL;
1675 	/*
1676 	 * Try to acquire the next worker's work queue.  If it is locked,
1677 	 * then search the list of workers until a queue is found unlocked,
1678 	 * or until the list is completely traversed at which point another
1679 	 * worker will be created.
1680 	 */
1681 	sigoff(self);		/* defer SIGIO */
1682 	sig_mutex_lock(&__aio_mutex);
1683 	first = aiowp = *nextworker;
1684 	if (mode != AIONOTIFY)
1685 		_aio_outstand_cnt++;
1686 	sig_mutex_unlock(&__aio_mutex);
1687 
1688 	switch (mode) {
1689 	case AIOREAD:
1690 	case AIOWRITE:
1691 	case AIOAREAD:
1692 	case AIOAWRITE:
1693 #if !defined(_LP64)
1694 	case AIOAREAD64:
1695 	case AIOAWRITE64:
1696 #endif
1697 		/* try to find an idle worker */
1698 		found = 0;
1699 		do {
1700 			if (sig_mutex_trylock(&aiowp->work_qlock1) == 0) {
1701 				if (aiowp->work_idleflg) {
1702 					found = 1;
1703 					break;
1704 				}
1705 				sig_mutex_unlock(&aiowp->work_qlock1);
1706 			}
1707 		} while ((aiowp = aiowp->work_forw) != first);
1708 
1709 		if (found) {
1710 			aiowp->work_minload1++;
1711 			break;
1712 		}
1713 
1714 		/* try to acquire some worker's queue lock */
1715 		do {
1716 			if (sig_mutex_trylock(&aiowp->work_qlock1) == 0) {
1717 				found = 1;
1718 				break;
1719 			}
1720 		} while ((aiowp = aiowp->work_forw) != first);
1721 
1722 		/*
1723 		 * Create more workers when the workers appear overloaded.
1724 		 * Either all the workers are busy draining their queues
1725 		 * or no worker's queue lock could be acquired.
1726 		 */
1727 		if (!found) {
1728 			if (_aio_worker_cnt < _max_workers) {
1729 				if (_aio_create_worker(reqp, mode))
1730 					aio_panic("_aio_req_add: add worker");
1731 				sigon(self);	/* reenable SIGIO */
1732 				return;
1733 			}
1734 
1735 			/*
1736 			 * No worker available and we have created
1737 			 * _max_workers, keep going through the
1738 			 * list slowly until we get a lock
1739 			 */
1740 			while (sig_mutex_trylock(&aiowp->work_qlock1) != 0) {
1741 				/*
1742 				 * give someone else a chance
1743 				 */
1744 				_aio_delay(1);
1745 				aiowp = aiowp->work_forw;
1746 			}
1747 		}
1748 
1749 		ASSERT(MUTEX_HELD(&aiowp->work_qlock1));
1750 		if (_aio_worker_cnt < _max_workers &&
1751 		    aiowp->work_minload1 >= _minworkload) {
1752 			sig_mutex_unlock(&aiowp->work_qlock1);
1753 			sig_mutex_lock(&__aio_mutex);
1754 			*nextworker = aiowp->work_forw;
1755 			sig_mutex_unlock(&__aio_mutex);
1756 			if (_aio_create_worker(reqp, mode))
1757 				aio_panic("aio_req_add: add worker");
1758 			sigon(self);	/* reenable SIGIO */
1759 			return;
1760 		}
1761 		aiowp->work_minload1++;
1762 		break;
1763 	case AIOFSYNC:
1764 	case AIONOTIFY:
1765 		load_bal_flg = 0;
1766 		sig_mutex_lock(&aiowp->work_qlock1);
1767 		break;
1768 	default:
1769 		aio_panic("_aio_req_add: invalid mode");
1770 		break;
1771 	}
1772 	/*
1773 	 * Put request onto worker's work queue.
1774 	 */
1775 	if (aiowp->work_tail1 == NULL) {
1776 		ASSERT(aiowp->work_count1 == 0);
1777 		aiowp->work_tail1 = reqp;
1778 		aiowp->work_next1 = reqp;
1779 	} else {
1780 		aiowp->work_head1->req_next = reqp;
1781 		if (aiowp->work_next1 == NULL)
1782 			aiowp->work_next1 = reqp;
1783 	}
1784 	reqp->req_state = AIO_REQ_QUEUED;
1785 	reqp->req_worker = aiowp;
1786 	aiowp->work_head1 = reqp;
1787 	/*
1788 	 * Awaken worker if it is not currently active.
1789 	 */
1790 	if (aiowp->work_count1++ == 0 && aiowp->work_idleflg) {
1791 		aiowp->work_idleflg = 0;
1792 		(void) cond_signal(&aiowp->work_idle_cv);
1793 	}
1794 	sig_mutex_unlock(&aiowp->work_qlock1);
1795 
1796 	if (load_bal_flg) {
1797 		sig_mutex_lock(&__aio_mutex);
1798 		*nextworker = aiowp->work_forw;
1799 		sig_mutex_unlock(&__aio_mutex);
1800 	}
1801 	sigon(self);	/* reenable SIGIO */
1802 }
1803 
1804 /*
1805  * Get an AIO request for a specified worker.
1806  * If the work queue is empty, return NULL.
1807  */
1808 aio_req_t *
1809 _aio_req_get(aio_worker_t *aiowp)
1810 {
1811 	aio_req_t *reqp;
1812 
1813 	sig_mutex_lock(&aiowp->work_qlock1);
1814 	if ((reqp = aiowp->work_next1) != NULL) {
1815 		/*
1816 		 * Remove a POSIX request from the queue; the
1817 		 * request queue is a singularly linked list
1818 		 * with a previous pointer.  The request is
1819 		 * removed by updating the previous pointer.
1820 		 *
1821 		 * Non-posix requests are left on the queue
1822 		 * to eventually be placed on the done queue.
1823 		 */
1824 
1825 		if (POSIX_AIO(reqp)) {
1826 			if (aiowp->work_prev1 == NULL) {
1827 				aiowp->work_tail1 = reqp->req_next;
1828 				if (aiowp->work_tail1 == NULL)
1829 					aiowp->work_head1 = NULL;
1830 			} else {
1831 				aiowp->work_prev1->req_next = reqp->req_next;
1832 				if (aiowp->work_head1 == reqp)
1833 					aiowp->work_head1 = reqp->req_next;
1834 			}
1835 
1836 		} else {
1837 			aiowp->work_prev1 = reqp;
1838 			ASSERT(aiowp->work_done1 >= 0);
1839 			aiowp->work_done1++;
1840 		}
1841 		ASSERT(reqp != reqp->req_next);
1842 		aiowp->work_next1 = reqp->req_next;
1843 		ASSERT(aiowp->work_count1 >= 1);
1844 		aiowp->work_count1--;
1845 		switch (reqp->req_op) {
1846 		case AIOREAD:
1847 		case AIOWRITE:
1848 		case AIOAREAD:
1849 		case AIOAWRITE:
1850 #if !defined(_LP64)
1851 		case AIOAREAD64:
1852 		case AIOAWRITE64:
1853 #endif
1854 			ASSERT(aiowp->work_minload1 > 0);
1855 			aiowp->work_minload1--;
1856 			break;
1857 		}
1858 		reqp->req_state = AIO_REQ_INPROGRESS;
1859 	}
1860 	aiowp->work_req = reqp;
1861 	ASSERT(reqp != NULL || aiowp->work_count1 == 0);
1862 	sig_mutex_unlock(&aiowp->work_qlock1);
1863 	return (reqp);
1864 }
1865 
1866 static void
1867 _aio_req_del(aio_worker_t *aiowp, aio_req_t *reqp, int ostate)
1868 {
1869 	aio_req_t **last;
1870 	aio_req_t *lastrp;
1871 	aio_req_t *next;
1872 
1873 	ASSERT(aiowp != NULL);
1874 	ASSERT(MUTEX_HELD(&aiowp->work_qlock1));
1875 	if (POSIX_AIO(reqp)) {
1876 		if (ostate != AIO_REQ_QUEUED)
1877 			return;
1878 	}
1879 	last = &aiowp->work_tail1;
1880 	lastrp = aiowp->work_tail1;
1881 	ASSERT(ostate == AIO_REQ_QUEUED || ostate == AIO_REQ_INPROGRESS);
1882 	while ((next = *last) != NULL) {
1883 		if (next == reqp) {
1884 			*last = next->req_next;
1885 			if (aiowp->work_next1 == next)
1886 				aiowp->work_next1 = next->req_next;
1887 
1888 			if ((next->req_next != NULL) ||
1889 			    (aiowp->work_done1 == 0)) {
1890 				if (aiowp->work_head1 == next)
1891 					aiowp->work_head1 = next->req_next;
1892 				if (aiowp->work_prev1 == next)
1893 					aiowp->work_prev1 = next->req_next;
1894 			} else {
1895 				if (aiowp->work_head1 == next)
1896 					aiowp->work_head1 = lastrp;
1897 				if (aiowp->work_prev1 == next)
1898 					aiowp->work_prev1 = lastrp;
1899 			}
1900 
1901 			if (ostate == AIO_REQ_QUEUED) {
1902 				ASSERT(aiowp->work_count1 >= 1);
1903 				aiowp->work_count1--;
1904 				ASSERT(aiowp->work_minload1 >= 1);
1905 				aiowp->work_minload1--;
1906 			} else {
1907 				ASSERT(ostate == AIO_REQ_INPROGRESS &&
1908 				    !POSIX_AIO(reqp));
1909 				aiowp->work_done1--;
1910 			}
1911 			return;
1912 		}
1913 		last = &next->req_next;
1914 		lastrp = next;
1915 	}
1916 	/* NOTREACHED */
1917 }
1918 
1919 static void
1920 _aio_enq_doneq(aio_req_t *reqp)
1921 {
1922 	if (_aio_doneq == NULL) {
1923 		_aio_doneq = reqp;
1924 		reqp->req_next = reqp->req_prev = reqp;
1925 	} else {
1926 		reqp->req_next = _aio_doneq;
1927 		reqp->req_prev = _aio_doneq->req_prev;
1928 		_aio_doneq->req_prev->req_next = reqp;
1929 		_aio_doneq->req_prev = reqp;
1930 	}
1931 	reqp->req_state = AIO_REQ_DONEQ;
1932 	_aio_doneq_cnt++;
1933 }
1934 
1935 /*
1936  * caller owns the _aio_mutex
1937  */
1938 aio_req_t *
1939 _aio_req_remove(aio_req_t *reqp)
1940 {
1941 	if (reqp && reqp->req_state != AIO_REQ_DONEQ)
1942 		return (NULL);
1943 
1944 	if (reqp) {
1945 		/* request in done queue */
1946 		if (_aio_doneq == reqp)
1947 			_aio_doneq = reqp->req_next;
1948 		if (_aio_doneq == reqp) {
1949 			/* only one request on queue */
1950 			_aio_doneq = NULL;
1951 		} else {
1952 			aio_req_t *tmp = reqp->req_next;
1953 			reqp->req_prev->req_next = tmp;
1954 			tmp->req_prev = reqp->req_prev;
1955 		}
1956 	} else if ((reqp = _aio_doneq) != NULL) {
1957 		if (reqp == reqp->req_next) {
1958 			/* only one request on queue */
1959 			_aio_doneq = NULL;
1960 		} else {
1961 			reqp->req_prev->req_next = _aio_doneq = reqp->req_next;
1962 			_aio_doneq->req_prev = reqp->req_prev;
1963 		}
1964 	}
1965 	if (reqp) {
1966 		_aio_doneq_cnt--;
1967 		reqp->req_next = reqp->req_prev = reqp;
1968 		reqp->req_state = AIO_REQ_DONE;
1969 	}
1970 	return (reqp);
1971 }
1972 
1973 /*
1974  * An AIO request is identified by an aio_result_t pointer.  The library
1975  * maps this aio_result_t pointer to its internal representation using a
1976  * hash table.  This function adds an aio_result_t pointer to the hash table.
1977  */
1978 static int
1979 _aio_hash_insert(aio_result_t *resultp, aio_req_t *reqp)
1980 {
1981 	aio_hash_t *hashp;
1982 	aio_req_t **prev;
1983 	aio_req_t *next;
1984 
1985 	hashp = _aio_hash + AIOHASH(resultp);
1986 	lmutex_lock(&hashp->hash_lock);
1987 	prev = &hashp->hash_ptr;
1988 	while ((next = *prev) != NULL) {
1989 		if (resultp == next->req_resultp) {
1990 			lmutex_unlock(&hashp->hash_lock);
1991 			return (-1);
1992 		}
1993 		prev = &next->req_link;
1994 	}
1995 	*prev = reqp;
1996 	ASSERT(reqp->req_link == NULL);
1997 	lmutex_unlock(&hashp->hash_lock);
1998 	return (0);
1999 }
2000 
2001 /*
2002  * Remove an entry from the hash table.
2003  */
2004 aio_req_t *
2005 _aio_hash_del(aio_result_t *resultp)
2006 {
2007 	aio_hash_t *hashp;
2008 	aio_req_t **prev;
2009 	aio_req_t *next = NULL;
2010 
2011 	if (_aio_hash != NULL) {
2012 		hashp = _aio_hash + AIOHASH(resultp);
2013 		lmutex_lock(&hashp->hash_lock);
2014 		prev = &hashp->hash_ptr;
2015 		while ((next = *prev) != NULL) {
2016 			if (resultp == next->req_resultp) {
2017 				*prev = next->req_link;
2018 				next->req_link = NULL;
2019 				break;
2020 			}
2021 			prev = &next->req_link;
2022 		}
2023 		lmutex_unlock(&hashp->hash_lock);
2024 	}
2025 	return (next);
2026 }
2027 
2028 /*
2029  *  find an entry in the hash table
2030  */
2031 aio_req_t *
2032 _aio_hash_find(aio_result_t *resultp)
2033 {
2034 	aio_hash_t *hashp;
2035 	aio_req_t **prev;
2036 	aio_req_t *next = NULL;
2037 
2038 	if (_aio_hash != NULL) {
2039 		hashp = _aio_hash + AIOHASH(resultp);
2040 		lmutex_lock(&hashp->hash_lock);
2041 		prev = &hashp->hash_ptr;
2042 		while ((next = *prev) != NULL) {
2043 			if (resultp == next->req_resultp)
2044 				break;
2045 			prev = &next->req_link;
2046 		}
2047 		lmutex_unlock(&hashp->hash_lock);
2048 	}
2049 	return (next);
2050 }
2051 
2052 /*
2053  * AIO interface for POSIX
2054  */
2055 int
2056 _aio_rw(aiocb_t *aiocbp, aio_lio_t *lio_head, aio_worker_t **nextworker,
2057     int mode, int flg)
2058 {
2059 	aio_req_t *reqp;
2060 	aio_args_t *ap;
2061 	int kerr;
2062 
2063 	if (aiocbp == NULL) {
2064 		errno = EINVAL;
2065 		return (-1);
2066 	}
2067 
2068 	/* initialize kaio */
2069 	if (!_kaio_ok)
2070 		_kaio_init();
2071 
2072 	aiocbp->aio_state = NOCHECK;
2073 
2074 	/*
2075 	 * If we have been called because a list I/O
2076 	 * kaio() failed, we dont want to repeat the
2077 	 * system call
2078 	 */
2079 
2080 	if (flg & AIO_KAIO) {
2081 		/*
2082 		 * Try kernel aio first.
2083 		 * If errno is ENOTSUP/EBADFD,
2084 		 * fall back to the thread implementation.
2085 		 */
2086 		if (_kaio_ok > 0 && KAIO_SUPPORTED(aiocbp->aio_fildes)) {
2087 			aiocbp->aio_resultp.aio_errno = EINPROGRESS;
2088 			aiocbp->aio_state = CHECK;
2089 			kerr = (int)_kaio(mode, aiocbp);
2090 			if (kerr == 0)
2091 				return (0);
2092 			if (errno != ENOTSUP && errno != EBADFD) {
2093 				aiocbp->aio_resultp.aio_errno = errno;
2094 				aiocbp->aio_resultp.aio_return = -1;
2095 				aiocbp->aio_state = NOCHECK;
2096 				return (-1);
2097 			}
2098 			if (errno == EBADFD)
2099 				SET_KAIO_NOT_SUPPORTED(aiocbp->aio_fildes);
2100 		}
2101 	}
2102 
2103 	aiocbp->aio_resultp.aio_errno = EINPROGRESS;
2104 	aiocbp->aio_state = USERAIO;
2105 
2106 	if (!__uaio_ok && __uaio_init() == -1)
2107 		return (-1);
2108 
2109 	if ((reqp = _aio_req_alloc()) == NULL) {
2110 		errno = EAGAIN;
2111 		return (-1);
2112 	}
2113 
2114 	/*
2115 	 * If an LIO request, add the list head to the aio request
2116 	 */
2117 	reqp->req_head = lio_head;
2118 	reqp->req_type = AIO_POSIX_REQ;
2119 	reqp->req_op = mode;
2120 	reqp->req_largefile = 0;
2121 
2122 	if (aiocbp->aio_sigevent.sigev_notify == SIGEV_NONE) {
2123 		reqp->req_sigevent.sigev_notify = SIGEV_NONE;
2124 	} else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2125 		reqp->req_sigevent.sigev_notify = SIGEV_SIGNAL;
2126 		reqp->req_sigevent.sigev_signo =
2127 		    aiocbp->aio_sigevent.sigev_signo;
2128 		reqp->req_sigevent.sigev_value.sival_ptr =
2129 		    aiocbp->aio_sigevent.sigev_value.sival_ptr;
2130 	} else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_PORT) {
2131 		port_notify_t *pn = aiocbp->aio_sigevent.sigev_value.sival_ptr;
2132 		reqp->req_sigevent.sigev_notify = SIGEV_PORT;
2133 		/*
2134 		 * Reuse the sigevent structure to contain the port number
2135 		 * and the user value.  Same for SIGEV_THREAD, below.
2136 		 */
2137 		reqp->req_sigevent.sigev_signo =
2138 		    pn->portnfy_port;
2139 		reqp->req_sigevent.sigev_value.sival_ptr =
2140 		    pn->portnfy_user;
2141 	} else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_THREAD) {
2142 		reqp->req_sigevent.sigev_notify = SIGEV_THREAD;
2143 		/*
2144 		 * The sigevent structure contains the port number
2145 		 * and the user value.  Same for SIGEV_PORT, above.
2146 		 */
2147 		reqp->req_sigevent.sigev_signo =
2148 		    aiocbp->aio_sigevent.sigev_signo;
2149 		reqp->req_sigevent.sigev_value.sival_ptr =
2150 		    aiocbp->aio_sigevent.sigev_value.sival_ptr;
2151 	}
2152 
2153 	reqp->req_resultp = &aiocbp->aio_resultp;
2154 	reqp->req_aiocbp = aiocbp;
2155 	ap = &reqp->req_args;
2156 	ap->fd = aiocbp->aio_fildes;
2157 	ap->buf = (caddr_t)aiocbp->aio_buf;
2158 	ap->bufsz = aiocbp->aio_nbytes;
2159 	ap->offset = aiocbp->aio_offset;
2160 
2161 	if ((flg & AIO_NO_DUPS) &&
2162 	    _aio_hash_insert(&aiocbp->aio_resultp, reqp) != 0) {
2163 		aio_panic("_aio_rw(): request already in hash table");
2164 		_aio_req_free(reqp);
2165 		errno = EINVAL;
2166 		return (-1);
2167 	}
2168 	_aio_req_add(reqp, nextworker, mode);
2169 	return (0);
2170 }
2171 
2172 #if !defined(_LP64)
2173 /*
2174  * 64-bit AIO interface for POSIX
2175  */
2176 int
2177 _aio_rw64(aiocb64_t *aiocbp, aio_lio_t *lio_head, aio_worker_t **nextworker,
2178     int mode, int flg)
2179 {
2180 	aio_req_t *reqp;
2181 	aio_args_t *ap;
2182 	int kerr;
2183 
2184 	if (aiocbp == NULL) {
2185 		errno = EINVAL;
2186 		return (-1);
2187 	}
2188 
2189 	/* initialize kaio */
2190 	if (!_kaio_ok)
2191 		_kaio_init();
2192 
2193 	aiocbp->aio_state = NOCHECK;
2194 
2195 	/*
2196 	 * If we have been called because a list I/O
2197 	 * kaio() failed, we dont want to repeat the
2198 	 * system call
2199 	 */
2200 
2201 	if (flg & AIO_KAIO) {
2202 		/*
2203 		 * Try kernel aio first.
2204 		 * If errno is ENOTSUP/EBADFD,
2205 		 * fall back to the thread implementation.
2206 		 */
2207 		if (_kaio_ok > 0 && KAIO_SUPPORTED(aiocbp->aio_fildes)) {
2208 			aiocbp->aio_resultp.aio_errno = EINPROGRESS;
2209 			aiocbp->aio_state = CHECK;
2210 			kerr = (int)_kaio(mode, aiocbp);
2211 			if (kerr == 0)
2212 				return (0);
2213 			if (errno != ENOTSUP && errno != EBADFD) {
2214 				aiocbp->aio_resultp.aio_errno = errno;
2215 				aiocbp->aio_resultp.aio_return = -1;
2216 				aiocbp->aio_state = NOCHECK;
2217 				return (-1);
2218 			}
2219 			if (errno == EBADFD)
2220 				SET_KAIO_NOT_SUPPORTED(aiocbp->aio_fildes);
2221 		}
2222 	}
2223 
2224 	aiocbp->aio_resultp.aio_errno = EINPROGRESS;
2225 	aiocbp->aio_state = USERAIO;
2226 
2227 	if (!__uaio_ok && __uaio_init() == -1)
2228 		return (-1);
2229 
2230 	if ((reqp = _aio_req_alloc()) == NULL) {
2231 		errno = EAGAIN;
2232 		return (-1);
2233 	}
2234 
2235 	/*
2236 	 * If an LIO request, add the list head to the aio request
2237 	 */
2238 	reqp->req_head = lio_head;
2239 	reqp->req_type = AIO_POSIX_REQ;
2240 	reqp->req_op = mode;
2241 	reqp->req_largefile = 1;
2242 
2243 	if (aiocbp->aio_sigevent.sigev_notify == SIGEV_NONE) {
2244 		reqp->req_sigevent.sigev_notify = SIGEV_NONE;
2245 	} else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2246 		reqp->req_sigevent.sigev_notify = SIGEV_SIGNAL;
2247 		reqp->req_sigevent.sigev_signo =
2248 		    aiocbp->aio_sigevent.sigev_signo;
2249 		reqp->req_sigevent.sigev_value.sival_ptr =
2250 		    aiocbp->aio_sigevent.sigev_value.sival_ptr;
2251 	} else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_PORT) {
2252 		port_notify_t *pn = aiocbp->aio_sigevent.sigev_value.sival_ptr;
2253 		reqp->req_sigevent.sigev_notify = SIGEV_PORT;
2254 		reqp->req_sigevent.sigev_signo =
2255 		    pn->portnfy_port;
2256 		reqp->req_sigevent.sigev_value.sival_ptr =
2257 		    pn->portnfy_user;
2258 	} else if (aiocbp->aio_sigevent.sigev_notify == SIGEV_THREAD) {
2259 		reqp->req_sigevent.sigev_notify = SIGEV_THREAD;
2260 		reqp->req_sigevent.sigev_signo =
2261 		    aiocbp->aio_sigevent.sigev_signo;
2262 		reqp->req_sigevent.sigev_value.sival_ptr =
2263 		    aiocbp->aio_sigevent.sigev_value.sival_ptr;
2264 	}
2265 
2266 	reqp->req_resultp = &aiocbp->aio_resultp;
2267 	reqp->req_aiocbp = aiocbp;
2268 	ap = &reqp->req_args;
2269 	ap->fd = aiocbp->aio_fildes;
2270 	ap->buf = (caddr_t)aiocbp->aio_buf;
2271 	ap->bufsz = aiocbp->aio_nbytes;
2272 	ap->offset = aiocbp->aio_offset;
2273 
2274 	if ((flg & AIO_NO_DUPS) &&
2275 	    _aio_hash_insert(&aiocbp->aio_resultp, reqp) != 0) {
2276 		aio_panic("_aio_rw64(): request already in hash table");
2277 		_aio_req_free(reqp);
2278 		errno = EINVAL;
2279 		return (-1);
2280 	}
2281 	_aio_req_add(reqp, nextworker, mode);
2282 	return (0);
2283 }
2284 #endif	/* !defined(_LP64) */
2285