xref: /freebsd/sys/kern/vfs_aio.c (revision eacee0ff7ec955b32e09515246bd97b6edcd2b0f)
1 /*
2  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. John S. Dyson's name may not be used to endorse or promote products
10  *    derived from this software without specific prior written permission.
11  *
12  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
13  * bad that happens because of using this software isn't the responsibility
14  * of the author.  This software is distributed AS-IS.
15  *
16  * $FreeBSD$
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/param.h>
24 #include <sys/systm.h>
25 #include <sys/bio.h>
26 #include <sys/buf.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/kthread.h>
31 #include <sys/fcntl.h>
32 #include <sys/file.h>
33 #include <sys/lock.h>
34 #include <sys/mutex.h>
35 #include <sys/unistd.h>
36 #include <sys/proc.h>
37 #include <sys/resourcevar.h>
38 #include <sys/signalvar.h>
39 #include <sys/protosw.h>
40 #include <sys/socketvar.h>
41 #include <sys/syscall.h>
42 #include <sys/sysent.h>
43 #include <sys/sysctl.h>
44 #include <sys/vnode.h>
45 #include <sys/conf.h>
46 #include <sys/event.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_extern.h>
50 #include <vm/pmap.h>
51 #include <vm/vm_map.h>
52 #include <vm/vm_zone.h>
53 #include <sys/aio.h>
54 
55 #include <machine/limits.h>
56 
57 #include "opt_vfs_aio.h"
58 
59 static	long jobrefid;
60 
61 #define JOBST_NULL		0x0
62 #define JOBST_JOBQGLOBAL	0x2
63 #define JOBST_JOBRUNNING	0x3
64 #define JOBST_JOBFINISHED	0x4
65 #define	JOBST_JOBQBUF		0x5
66 #define	JOBST_JOBBFINISHED	0x6
67 
68 #ifndef MAX_AIO_PER_PROC
69 #define MAX_AIO_PER_PROC	32
70 #endif
71 
72 #ifndef MAX_AIO_QUEUE_PER_PROC
73 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
74 #endif
75 
76 #ifndef MAX_AIO_PROCS
77 #define MAX_AIO_PROCS		32
78 #endif
79 
80 #ifndef MAX_AIO_QUEUE
81 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
82 #endif
83 
84 #ifndef TARGET_AIO_PROCS
85 #define TARGET_AIO_PROCS	4
86 #endif
87 
88 #ifndef MAX_BUF_AIO
89 #define MAX_BUF_AIO		16
90 #endif
91 
92 #ifndef AIOD_TIMEOUT_DEFAULT
93 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
94 #endif
95 
96 #ifndef AIOD_LIFETIME_DEFAULT
97 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
98 #endif
99 
100 static int max_aio_procs = MAX_AIO_PROCS;
101 static int num_aio_procs = 0;
102 static int target_aio_procs = TARGET_AIO_PROCS;
103 static int max_queue_count = MAX_AIO_QUEUE;
104 static int num_queue_count = 0;
105 static int num_buf_aio = 0;
106 static int num_aio_resv_start = 0;
107 static int aiod_timeout;
108 static int aiod_lifetime;
109 static int unloadable = 0;
110 
111 static int max_aio_per_proc = MAX_AIO_PER_PROC;
112 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
113 static int max_buf_aio = MAX_BUF_AIO;
114 
115 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "AIO mgmt");
116 
117 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc,
118 	CTLFLAG_RW, &max_aio_per_proc, 0, "");
119 
120 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc,
121 	CTLFLAG_RW, &max_aio_queue_per_proc, 0, "");
122 
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
124 	CTLFLAG_RW, &max_aio_procs, 0, "");
125 
126 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
127 	CTLFLAG_RD, &num_aio_procs, 0, "");
128 
129 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count,
130 	CTLFLAG_RD, &num_queue_count, 0, "");
131 
132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue,
133 	CTLFLAG_RW, &max_queue_count, 0, "");
134 
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs,
136 	CTLFLAG_RW, &target_aio_procs, 0, "");
137 
138 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio,
139 	CTLFLAG_RW, &max_buf_aio, 0, "");
140 
141 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio,
142 	CTLFLAG_RD, &num_buf_aio, 0, "");
143 
144 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime,
145 	CTLFLAG_RW, &aiod_lifetime, 0, "");
146 
147 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout,
148 	CTLFLAG_RW, &aiod_timeout, 0, "");
149 
150 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
151     "Allow unload of aio (not recommended)");
152 
153 struct aiocblist {
154         TAILQ_ENTRY(aiocblist) list;	/* List of jobs */
155         TAILQ_ENTRY(aiocblist) plist;	/* List of jobs for proc */
156         int	jobflags;
157         int	jobstate;
158 	int	inputcharge;
159 	int	outputcharge;
160 	struct	callout_handle timeouthandle;
161         struct	buf *bp;		/* Buffer pointer */
162         struct	proc *userproc;		/* User process */ /* Not td! */
163         struct	file *fd_file;		/* Pointer to file structure */
164 	struct	aiothreadlist *jobaiothread;  /* AIO process descriptor */
165         struct	aio_liojob *lio;	/* Optional lio job */
166         struct	aiocb *uuaiocb;		/* Pointer in userspace of aiocb */
167 	struct	klist klist;		/* list of knotes */
168         struct	aiocb uaiocb;		/* Kernel I/O control block */
169 };
170 
171 /* jobflags */
172 #define AIOCBLIST_RUNDOWN       0x4
173 #define AIOCBLIST_ASYNCFREE     0x8
174 #define AIOCBLIST_DONE          0x10
175 
176 /*
177  * AIO process info
178  */
179 #define AIOP_FREE	0x1			/* proc on free queue */
180 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
181 
182 struct aiothreadlist {
183 	int aiothreadflags;			/* AIO proc flags */
184 	TAILQ_ENTRY(aiothreadlist) list;	/* List of processes */
185 	struct thread *aiothread;		/* The AIO thread */
186 };
187 
188 /*
189  * data-structure for lio signal management
190  */
191 struct aio_liojob {
192 	int	lioj_flags;
193 	int	lioj_buffer_count;
194 	int	lioj_buffer_finished_count;
195 	int	lioj_queue_count;
196 	int	lioj_queue_finished_count;
197 	struct	sigevent lioj_signal;	/* signal on all I/O done */
198 	TAILQ_ENTRY(aio_liojob) lioj_list;
199 	struct	kaioinfo *lioj_ki;
200 };
201 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
202 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
203 
204 /*
205  * per process aio data structure
206  */
207 struct kaioinfo {
208 	int	kaio_flags;		/* per process kaio flags */
209 	int	kaio_maxactive_count;	/* maximum number of AIOs */
210 	int	kaio_active_count;	/* number of currently used AIOs */
211 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
212 	int	kaio_queue_count;	/* size of AIO queue */
213 	int	kaio_ballowed_count;	/* maximum number of buffers */
214 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
215 	int	kaio_buffer_count;	/* number of physio buffers */
216 	int	kaio_buffer_finished_count; /* count of I/O done */
217 	struct 	proc *kaio_p;		/* process that uses this kaio block */
218 	TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
219 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* job queue for process */
220 	TAILQ_HEAD(,aiocblist) kaio_jobdone;	/* done queue for process */
221 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* buffer job queue for process */
222 	TAILQ_HEAD(,aiocblist) kaio_bufdone;	/* buffer done queue for process */
223 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;	/* queue for aios waiting on sockets */
224 };
225 
226 #define KAIO_RUNDOWN	0x1	/* process is being run down */
227 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
228 
229 static TAILQ_HEAD(,aiothreadlist) aio_freeproc, aio_activeproc;
230 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
231 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
232 
233 static void	aio_init_aioinfo(struct proc *p);
234 static void	aio_onceonly(void);
235 static int	aio_free_entry(struct aiocblist *aiocbe);
236 static void	aio_process(struct aiocblist *aiocbe);
237 static int	aio_newproc(void);
238 static int	aio_aqueue(struct thread *td, struct aiocb *job, int type);
239 static void	aio_physwakeup(struct buf *bp);
240 static void	aio_proc_rundown(struct proc *p);
241 static int	aio_fphysio(struct aiocblist *aiocbe);
242 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
243 static void	aio_daemon(void *uproc);
244 static void	aio_swake_cb(struct socket *, struct sockbuf *);
245 static int	aio_unload(void);
246 static void	process_signal(void *aioj);
247 static int	filt_aioattach(struct knote *kn);
248 static void	filt_aiodetach(struct knote *kn);
249 static int	filt_aio(struct knote *kn, long hint);
250 
251 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone;
252 static vm_zone_t aiolio_zone;
253 
254 static struct filterops aio_filtops =
255 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
256 
257 static int
258 aio_modload(struct module *module, int cmd, void *arg)
259 {
260 	int error = 0;
261 
262 	switch (cmd) {
263 	case MOD_LOAD:
264 		aio_onceonly();
265 		break;
266 	case MOD_UNLOAD:
267 		error = aio_unload();
268 		break;
269 	case MOD_SHUTDOWN:
270 		break;
271 	default:
272 		error = EINVAL;
273 		break;
274 	}
275 	return (error);
276 }
277 
278 static moduledata_t aio_mod = {
279 	"aio",
280 	&aio_modload,
281 	NULL
282 };
283 
284 SYSCALL_MODULE_HELPER(aio_return);
285 SYSCALL_MODULE_HELPER(aio_suspend);
286 SYSCALL_MODULE_HELPER(aio_cancel);
287 SYSCALL_MODULE_HELPER(aio_error);
288 SYSCALL_MODULE_HELPER(aio_read);
289 SYSCALL_MODULE_HELPER(aio_write);
290 SYSCALL_MODULE_HELPER(aio_waitcomplete);
291 SYSCALL_MODULE_HELPER(lio_listio);
292 
293 DECLARE_MODULE(aio, aio_mod,
294 	SI_SUB_VFS, SI_ORDER_ANY);
295 MODULE_VERSION(aio, 1);
296 
297 /*
298  * Startup initialization
299  */
300 static void
301 aio_onceonly(void)
302 {
303 
304 	/* XXX: should probably just use so->callback */
305 	aio_swake = &aio_swake_cb;
306 	at_exit(aio_proc_rundown);
307 	at_exec(aio_proc_rundown);
308 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
309 	TAILQ_INIT(&aio_freeproc);
310 	TAILQ_INIT(&aio_activeproc);
311 	TAILQ_INIT(&aio_jobs);
312 	TAILQ_INIT(&aio_bufjobs);
313 	kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
314 	aiop_zone = zinit("AIOP", sizeof(struct aiothreadlist), 0, 0, 1);
315 	aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
316 	aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
317 	aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
318 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
319 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
320 	jobrefid = 1;
321 }
322 
323 static int
324 aio_unload(void)
325 {
326 
327 	/*
328 	 * XXX: no unloads by default, it's too dangerous.
329 	 * perhaps we could do it if locked out callers and then
330 	 * did an aio_proc_rundown() on each process.
331 	 */
332 	if (!unloadable)
333 		return (EOPNOTSUPP);
334 
335 	aio_swake = NULL;
336 	rm_at_exit(aio_proc_rundown);
337 	rm_at_exec(aio_proc_rundown);
338 	kqueue_del_filteropts(EVFILT_AIO);
339 	return (0);
340 }
341 
342 /*
343  * Init the per-process aioinfo structure.  The aioinfo limits are set
344  * per-process for user limit (resource) management.
345  */
346 static void
347 aio_init_aioinfo(struct proc *p)
348 {
349 	struct kaioinfo *ki;
350 	if (p->p_aioinfo == NULL) {
351 		ki = zalloc(kaio_zone);
352 		p->p_aioinfo = ki;
353 		ki->kaio_flags = 0;
354 		ki->kaio_maxactive_count = max_aio_per_proc;
355 		ki->kaio_active_count = 0;
356 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
357 		ki->kaio_queue_count = 0;
358 		ki->kaio_ballowed_count = max_buf_aio;
359 		ki->kaio_buffer_count = 0;
360 		ki->kaio_buffer_finished_count = 0;
361 		ki->kaio_p = p;
362 		TAILQ_INIT(&ki->kaio_jobdone);
363 		TAILQ_INIT(&ki->kaio_jobqueue);
364 		TAILQ_INIT(&ki->kaio_bufdone);
365 		TAILQ_INIT(&ki->kaio_bufqueue);
366 		TAILQ_INIT(&ki->kaio_liojoblist);
367 		TAILQ_INIT(&ki->kaio_sockqueue);
368 	}
369 
370 	while (num_aio_procs < target_aio_procs)
371 		aio_newproc();
372 }
373 
374 /*
375  * Free a job entry.  Wait for completion if it is currently active, but don't
376  * delay forever.  If we delay, we return a flag that says that we have to
377  * restart the queue scan.
378  */
379 static int
380 aio_free_entry(struct aiocblist *aiocbe)
381 {
382 	struct kaioinfo *ki;
383 	struct aio_liojob *lj;
384 	struct proc *p;
385 	int error;
386 	int s;
387 
388 	if (aiocbe->jobstate == JOBST_NULL)
389 		panic("aio_free_entry: freeing already free job");
390 
391 	p = aiocbe->userproc;
392 	ki = p->p_aioinfo;
393 	lj = aiocbe->lio;
394 	if (ki == NULL)
395 		panic("aio_free_entry: missing p->p_aioinfo");
396 
397 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
398 		if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE)
399 			return 0;
400 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
401 		tsleep(aiocbe, PRIBIO, "jobwai", 0);
402 	}
403 	aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
404 
405 	if (aiocbe->bp == NULL) {
406 		if (ki->kaio_queue_count <= 0)
407 			panic("aio_free_entry: process queue size <= 0");
408 		if (num_queue_count <= 0)
409 			panic("aio_free_entry: system wide queue size <= 0");
410 
411 		if (lj) {
412 			lj->lioj_queue_count--;
413 			if (aiocbe->jobflags & AIOCBLIST_DONE)
414 				lj->lioj_queue_finished_count--;
415 		}
416 		ki->kaio_queue_count--;
417 		if (aiocbe->jobflags & AIOCBLIST_DONE)
418 			ki->kaio_queue_finished_count--;
419 		num_queue_count--;
420 	} else {
421 		if (lj) {
422 			lj->lioj_buffer_count--;
423 			if (aiocbe->jobflags & AIOCBLIST_DONE)
424 				lj->lioj_buffer_finished_count--;
425 		}
426 		if (aiocbe->jobflags & AIOCBLIST_DONE)
427 			ki->kaio_buffer_finished_count--;
428 		ki->kaio_buffer_count--;
429 		num_buf_aio--;
430 	}
431 
432 	/* aiocbe is going away, we need to destroy any knotes */
433 	/* XXXKSE Note the thread here is used to eventually find the
434 	 * owning process again, but it is also used to do a fo_close
435 	 * and that requires the thread. (but does it require the
436 	 * OWNING thread? (or maybe the running thread?)
437 	 * There is a semantic problem here...
438 	 */
439 	knote_remove(FIRST_THREAD_IN_PROC(p), &aiocbe->klist); /* XXXKSE */
440 
441 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
442 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
443 		ki->kaio_flags &= ~KAIO_WAKEUP;
444 		wakeup(p);
445 	}
446 
447 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
448 		if ((error = aio_fphysio(aiocbe)) != 0)
449 			return error;
450 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
451 			panic("aio_free_entry: invalid physio finish-up state");
452 		s = splbio();
453 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
454 		splx(s);
455 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
456 		s = splnet();
457 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
458 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
459 		splx(s);
460 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
461 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
462 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
463 		s = splbio();
464 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
465 		splx(s);
466 		if (aiocbe->bp) {
467 			vunmapbuf(aiocbe->bp);
468 			relpbuf(aiocbe->bp, NULL);
469 			aiocbe->bp = NULL;
470 		}
471 	}
472 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
473 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
474 		zfree(aiolio_zone, lj);
475 	}
476 	aiocbe->jobstate = JOBST_NULL;
477 	untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
478 	zfree(aiocb_zone, aiocbe);
479 	return 0;
480 }
481 
482 /*
483  * Rundown the jobs for a given process.
484  */
485 static void
486 aio_proc_rundown(struct proc *p)
487 {
488 	int s;
489 	struct kaioinfo *ki;
490 	struct aio_liojob *lj, *ljn;
491 	struct aiocblist *aiocbe, *aiocbn;
492 	struct file *fp;
493 	struct filedesc *fdp;
494 	struct socket *so;
495 
496 	ki = p->p_aioinfo;
497 	if (ki == NULL)
498 		return;
499 
500 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
501 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
502 	    ki->kaio_buffer_finished_count)) {
503 		ki->kaio_flags |= KAIO_RUNDOWN;
504 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
505 			break;
506 	}
507 
508 	/*
509 	 * Move any aio ops that are waiting on socket I/O to the normal job
510 	 * queues so they are cleaned up with any others.
511 	 */
512 	fdp = p->p_fd;
513 
514 	s = splnet();
515 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
516 	    aiocbn) {
517 		aiocbn = TAILQ_NEXT(aiocbe, plist);
518 		fp = fdp->fd_ofiles[aiocbe->uaiocb.aio_fildes];
519 
520 		/*
521 		 * Under some circumstances, the aio_fildes and the file
522 		 * structure don't match.  This would leave aiocbe's in the
523 		 * TAILQ associated with the socket and cause a panic later.
524 		 *
525 		 * Detect and fix.
526 		 */
527 		if ((fp == NULL) || (fp != aiocbe->fd_file))
528 			fp = aiocbe->fd_file;
529 		if (fp) {
530 			so = (struct socket *)fp->f_data;
531 			TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
532 			if (TAILQ_EMPTY(&so->so_aiojobq)) {
533 				so->so_snd.sb_flags &= ~SB_AIO;
534 				so->so_rcv.sb_flags &= ~SB_AIO;
535 			}
536 		}
537 		TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
538 		TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
539 		TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
540 	}
541 	splx(s);
542 
543 restart1:
544 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
545 		aiocbn = TAILQ_NEXT(aiocbe, plist);
546 		if (aio_free_entry(aiocbe))
547 			goto restart1;
548 	}
549 
550 restart2:
551 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
552 	    aiocbn) {
553 		aiocbn = TAILQ_NEXT(aiocbe, plist);
554 		if (aio_free_entry(aiocbe))
555 			goto restart2;
556 	}
557 
558 /*
559  * Note the use of lots of splbio here, trying to avoid splbio for long chains
560  * of I/O.  Probably unnecessary.
561  */
562 restart3:
563 	s = splbio();
564 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
565 		ki->kaio_flags |= KAIO_WAKEUP;
566 		tsleep(p, PRIBIO, "aioprn", 0);
567 		splx(s);
568 		goto restart3;
569 	}
570 	splx(s);
571 
572 restart4:
573 	s = splbio();
574 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
575 		aiocbn = TAILQ_NEXT(aiocbe, plist);
576 		if (aio_free_entry(aiocbe)) {
577 			splx(s);
578 			goto restart4;
579 		}
580 	}
581 	splx(s);
582 
583         /*
584          * If we've slept, jobs might have moved from one queue to another.
585          * Retry rundown if we didn't manage to empty the queues.
586          */
587         if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
588 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
589 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
590 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
591 		goto restart1;
592 
593 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
594 		ljn = TAILQ_NEXT(lj, lioj_list);
595 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
596 		    0)) {
597 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
598 			zfree(aiolio_zone, lj);
599 		} else {
600 #ifdef DIAGNOSTIC
601 			printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
602 			    "QF:%d\n", lj->lioj_buffer_count,
603 			    lj->lioj_buffer_finished_count,
604 			    lj->lioj_queue_count,
605 			    lj->lioj_queue_finished_count);
606 #endif
607 		}
608 	}
609 
610 	zfree(kaio_zone, ki);
611 	p->p_aioinfo = NULL;
612 }
613 
614 /*
615  * Select a job to run (called by an AIO daemon).
616  */
617 static struct aiocblist *
618 aio_selectjob(struct aiothreadlist *aiop)
619 {
620 	int s;
621 	struct aiocblist *aiocbe;
622 	struct kaioinfo *ki;
623 	struct proc *userp;
624 
625 	s = splnet();
626 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
627 	    TAILQ_NEXT(aiocbe, list)) {
628 		userp = aiocbe->userproc;
629 		ki = userp->p_aioinfo;
630 
631 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
632 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
633 			splx(s);
634 			return aiocbe;
635 		}
636 	}
637 	splx(s);
638 
639 	return NULL;
640 }
641 
642 /*
643  * The AIO processing activity.  This is the code that does the I/O request for
644  * the non-physio version of the operations.  The normal vn operations are used,
645  * and this code should work in all instances for every type of file, including
646  * pipes, sockets, fifos, and regular files.
647  */
648 static void
649 aio_process(struct aiocblist *aiocbe)
650 {
651 	struct filedesc *fdp;
652 	struct thread *td;
653 	struct proc *userp;
654 	struct proc *mycp;
655 	struct aiocb *cb;
656 	struct file *fp;
657 	struct uio auio;
658 	struct iovec aiov;
659 	unsigned int fd;
660 	int cnt;
661 	int error;
662 	off_t offset;
663 	int oublock_st, oublock_end;
664 	int inblock_st, inblock_end;
665 
666 	userp = aiocbe->userproc;
667 	td = curthread;
668 	mycp = td->td_proc;
669 	cb = &aiocbe->uaiocb;
670 
671 	fdp = mycp->p_fd;
672 	fd = cb->aio_fildes;
673 	fp = fdp->fd_ofiles[fd];
674 
675 	if ((fp == NULL) || (fp != aiocbe->fd_file)) {
676 		cb->_aiocb_private.error = EBADF;
677 		cb->_aiocb_private.status = -1;
678 		return;
679 	}
680 
681 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
682 	aiov.iov_len = cb->aio_nbytes;
683 
684 	auio.uio_iov = &aiov;
685 	auio.uio_iovcnt = 1;
686 	auio.uio_offset = offset = cb->aio_offset;
687 	auio.uio_resid = cb->aio_nbytes;
688 	cnt = cb->aio_nbytes;
689 	auio.uio_segflg = UIO_USERSPACE;
690 	auio.uio_td = td;
691 
692 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
693 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
694 	/*
695 	 * Temporarily bump the ref count while reading to avoid the
696 	 * descriptor being ripped out from under us.
697 	 */
698 	fhold(fp);
699 	if (cb->aio_lio_opcode == LIO_READ) {
700 		auio.uio_rw = UIO_READ;
701 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
702 	} else {
703 		auio.uio_rw = UIO_WRITE;
704 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
705 	}
706 	fdrop(fp, td);
707 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
708 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
709 
710 	aiocbe->inputcharge = inblock_end - inblock_st;
711 	aiocbe->outputcharge = oublock_end - oublock_st;
712 
713 	if ((error) && (auio.uio_resid != cnt)) {
714 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
715 			error = 0;
716 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
717 			PROC_LOCK(userp);
718 			psignal(userp, SIGPIPE);
719 			PROC_UNLOCK(userp);
720 		}
721 	}
722 
723 	cnt -= auio.uio_resid;
724 	cb->_aiocb_private.error = error;
725 	cb->_aiocb_private.status = cnt;
726 }
727 
728 /*
729  * The AIO daemon, most of the actual work is done in aio_process,
730  * but the setup (and address space mgmt) is done in this routine.
731  */
732 static void
733 aio_daemon(void *uproc)
734 {
735 	int s;
736 	struct aio_liojob *lj;
737 	struct aiocb *cb;
738 	struct aiocblist *aiocbe;
739 	struct aiothreadlist *aiop;
740 	struct kaioinfo *ki;
741 	struct proc *curcp, *mycp, *userp;
742 	struct vmspace *myvm, *tmpvm;
743 	struct thread *td = curthread;
744 
745 	mtx_lock(&Giant);
746 	/*
747 	 * Local copies of curproc (cp) and vmspace (myvm)
748 	 */
749 	mycp = td->td_proc;
750 	myvm = mycp->p_vmspace;
751 
752 	if (mycp->p_textvp) {
753 		vrele(mycp->p_textvp);
754 		mycp->p_textvp = NULL;
755 	}
756 
757 	/*
758 	 * Allocate and ready the aio control info.  There is one aiop structure
759 	 * per daemon.
760 	 */
761 	aiop = zalloc(aiop_zone);
762 	aiop->aiothread = td;
763 	aiop->aiothreadflags |= AIOP_FREE;
764 
765 	s = splnet();
766 
767 	/*
768 	 * Place thread (lightweight process) onto the AIO free thread list.
769 	 */
770 	if (TAILQ_EMPTY(&aio_freeproc))
771 		wakeup(&aio_freeproc);
772 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
773 
774 	splx(s);
775 
776 	/*
777 	 * Get rid of our current filedescriptors.  AIOD's don't need any
778 	 * filedescriptors, except as temporarily inherited from the client.
779 	 */
780 	fdfree(td);
781 	mycp->p_fd = NULL;
782 
783 	/* The daemon resides in its own pgrp. */
784 	enterpgrp(mycp, mycp->p_pid, 1);
785 
786 	/* Mark special process type. */
787 	mycp->p_flag |= P_SYSTEM;
788 
789 	/*
790 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
791 	 * and creating too many daemons.)
792 	 */
793 	wakeup(mycp);
794 
795 	for (;;) {
796 		/*
797 		 * curcp is the current daemon process context.
798 		 * userp is the current user process context.
799 		 */
800 		curcp = mycp;
801 
802 		/*
803 		 * Take daemon off of free queue
804 		 */
805 		if (aiop->aiothreadflags & AIOP_FREE) {
806 			s = splnet();
807 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
808 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
809 			aiop->aiothreadflags &= ~AIOP_FREE;
810 			splx(s);
811 		}
812 		aiop->aiothreadflags &= ~AIOP_SCHED;
813 
814 		/*
815 		 * Check for jobs.
816 		 */
817 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
818 			cb = &aiocbe->uaiocb;
819 			userp = aiocbe->userproc;
820 
821 			aiocbe->jobstate = JOBST_JOBRUNNING;
822 
823 			/*
824 			 * Connect to process address space for user program.
825 			 */
826 			if (userp != curcp) {
827 				/*
828 				 * Save the current address space that we are
829 				 * connected to.
830 				 */
831 				tmpvm = mycp->p_vmspace;
832 
833 				/*
834 				 * Point to the new user address space, and
835 				 * refer to it.
836 				 */
837 				mycp->p_vmspace = userp->p_vmspace;
838 				mycp->p_vmspace->vm_refcnt++;
839 
840 				/* Activate the new mapping. */
841 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
842 
843 				/*
844 				 * If the old address space wasn't the daemons
845 				 * own address space, then we need to remove the
846 				 * daemon's reference from the other process
847 				 * that it was acting on behalf of.
848 				 */
849 				if (tmpvm != myvm) {
850 					vmspace_free(tmpvm);
851 				}
852 
853 				/*
854 				 * Disassociate from previous clients file
855 				 * descriptors, and associate to the new clients
856 				 * descriptors.  Note that the daemon doesn't
857 				 * need to worry about its orginal descriptors,
858 				 * because they were originally freed.
859 				 */
860 				if (mycp->p_fd)
861 					fdfree(td);
862 				mycp->p_fd = fdshare(userp);
863 				curcp = userp;
864 			}
865 
866 			ki = userp->p_aioinfo;
867 			lj = aiocbe->lio;
868 
869 			/* Account for currently active jobs. */
870 			ki->kaio_active_count++;
871 
872 			/* Do the I/O function. */
873 			aiocbe->jobaiothread = aiop;
874 			aio_process(aiocbe);
875 
876 			/* Decrement the active job count. */
877 			ki->kaio_active_count--;
878 
879 			/*
880 			 * Increment the completion count for wakeup/signal
881 			 * comparisons.
882 			 */
883 			aiocbe->jobflags |= AIOCBLIST_DONE;
884 			ki->kaio_queue_finished_count++;
885 			if (lj)
886 				lj->lioj_queue_finished_count++;
887 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
888 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
889 				ki->kaio_flags &= ~KAIO_WAKEUP;
890 				wakeup(userp);
891 			}
892 
893 			s = splbio();
894 			if (lj && (lj->lioj_flags &
895 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
896 				if ((lj->lioj_queue_finished_count ==
897 				    lj->lioj_queue_count) &&
898 				    (lj->lioj_buffer_finished_count ==
899 				    lj->lioj_buffer_count)) {
900 					PROC_LOCK(userp);
901 					psignal(userp,
902 					    lj->lioj_signal.sigev_signo);
903 					PROC_UNLOCK(userp);
904 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
905 				}
906 			}
907 			splx(s);
908 
909 			aiocbe->jobstate = JOBST_JOBFINISHED;
910 
911 			/*
912 			 * If the I/O request should be automatically rundown,
913 			 * do the needed cleanup.  Otherwise, place the queue
914 			 * entry for the just finished I/O request into the done
915 			 * queue for the associated client.
916 			 */
917 			s = splnet();
918 			if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE) {
919 				aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
920 				zfree(aiocb_zone, aiocbe);
921 			} else {
922 				TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
923 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe,
924 				    plist);
925 			}
926 			splx(s);
927 			KNOTE(&aiocbe->klist, 0);
928 
929 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
930 				wakeup(aiocbe);
931 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
932 			}
933 
934 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
935 				PROC_LOCK(userp);
936 				psignal(userp, cb->aio_sigevent.sigev_signo);
937 				PROC_UNLOCK(userp);
938 			}
939 		}
940 
941 		/*
942 		 * Disconnect from user address space.
943 		 */
944 		if (curcp != mycp) {
945 			/* Get the user address space to disconnect from. */
946 			tmpvm = mycp->p_vmspace;
947 
948 			/* Get original address space for daemon. */
949 			mycp->p_vmspace = myvm;
950 
951 			/* Activate the daemon's address space. */
952 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
953 #ifdef DIAGNOSTIC
954 			if (tmpvm == myvm) {
955 				printf("AIOD: vmspace problem -- %d\n",
956 				    mycp->p_pid);
957 			}
958 #endif
959 			/* Remove our vmspace reference. */
960 			vmspace_free(tmpvm);
961 
962 			/*
963 			 * Disassociate from the user process's file
964 			 * descriptors.
965 			 */
966 			if (mycp->p_fd)
967 				fdfree(td);
968 			mycp->p_fd = NULL;
969 			curcp = mycp;
970 		}
971 
972 		/*
973 		 * If we are the first to be put onto the free queue, wakeup
974 		 * anyone waiting for a daemon.
975 		 */
976 		s = splnet();
977 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
978 		if (TAILQ_EMPTY(&aio_freeproc))
979 			wakeup(&aio_freeproc);
980 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
981 		aiop->aiothreadflags |= AIOP_FREE;
982 		splx(s);
983 
984 		/*
985 		 * If daemon is inactive for a long time, allow it to exit,
986 		 * thereby freeing resources.
987 		 */
988 		if ((aiop->aiothreadflags & AIOP_SCHED) == 0 &&
989 		    tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) {
990 			s = splnet();
991 			if (TAILQ_EMPTY(&aio_jobs)) {
992 				if ((aiop->aiothreadflags & AIOP_FREE) &&
993 				    (num_aio_procs > target_aio_procs)) {
994 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
995 					splx(s);
996 					zfree(aiop_zone, aiop);
997 					num_aio_procs--;
998 #ifdef DIAGNOSTIC
999 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1000 						printf("AIOD: bad vm refcnt for"
1001 						    " exiting daemon: %d\n",
1002 						    mycp->p_vmspace->vm_refcnt);
1003 					}
1004 #endif
1005 					kthread_exit(0);
1006 				}
1007 			}
1008 			splx(s);
1009 		}
1010 	}
1011 }
1012 
1013 /*
1014  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
1015  * AIO daemon modifies its environment itself.
1016  */
1017 static int
1018 aio_newproc()
1019 {
1020 	int error;
1021 	struct proc *p;
1022 
1023 	error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, "aiod%d",
1024 			       num_aio_procs);
1025 	if (error)
1026 		return error;
1027 
1028 	/*
1029 	 * Wait until daemon is started, but continue on just in case to
1030 	 * handle error conditions.
1031 	 */
1032 	error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1033 
1034 	num_aio_procs++;
1035 
1036 	return error;
1037 }
1038 
1039 /*
1040  * Try the high-performance, low-overhead physio method for eligible
1041  * VCHR devices.  This method doesn't use an aio helper thread, and
1042  * thus has very low overhead.
1043  *
1044  * Assumes that the caller, _aio_aqueue(), has incremented the file
1045  * structure's reference count, preventing its deallocation for the
1046  * duration of this call.
1047  */
1048 static int
1049 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1050 {
1051 	int error;
1052 	struct aiocb *cb;
1053 	struct file *fp;
1054 	struct buf *bp;
1055 	struct vnode *vp;
1056 	struct kaioinfo *ki;
1057 	struct aio_liojob *lj;
1058 	int s;
1059 	int notify;
1060 
1061 	cb = &aiocbe->uaiocb;
1062 	fp = aiocbe->fd_file;
1063 
1064 	if (fp->f_type != DTYPE_VNODE)
1065 		return (-1);
1066 
1067 	vp = (struct vnode *)fp->f_data;
1068 
1069 	/*
1070 	 * If its not a disk, we don't want to return a positive error.
1071 	 * It causes the aio code to not fall through to try the thread
1072 	 * way when you're talking to a regular file.
1073 	 */
1074 	if (!vn_isdisk(vp, &error)) {
1075 		if (error == ENOTBLK)
1076 			return (-1);
1077 		else
1078 			return (error);
1079 	}
1080 
1081  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
1082 		return (-1);
1083 
1084 	if (cb->aio_nbytes >
1085 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1086 		return (-1);
1087 
1088 	ki = p->p_aioinfo;
1089 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1090 		return (-1);
1091 
1092 	ki->kaio_buffer_count++;
1093 
1094 	lj = aiocbe->lio;
1095 	if (lj)
1096 		lj->lioj_buffer_count++;
1097 
1098 	/* Create and build a buffer header for a transfer. */
1099 	bp = (struct buf *)getpbuf(NULL);
1100 	BUF_KERNPROC(bp);
1101 
1102 	/*
1103 	 * Get a copy of the kva from the physical buffer.
1104 	 */
1105 	bp->b_caller1 = p;
1106 	bp->b_dev = vp->v_rdev;
1107 	error = bp->b_error = 0;
1108 
1109 	bp->b_bcount = cb->aio_nbytes;
1110 	bp->b_bufsize = cb->aio_nbytes;
1111 	bp->b_flags = B_PHYS;
1112 	bp->b_iodone = aio_physwakeup;
1113 	bp->b_saveaddr = bp->b_data;
1114 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1115 	bp->b_blkno = btodb(cb->aio_offset);
1116 
1117 	if (cb->aio_lio_opcode == LIO_WRITE) {
1118 		bp->b_iocmd = BIO_WRITE;
1119 		if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_READ)) {
1120 			error = EFAULT;
1121 			goto doerror;
1122 		}
1123 	} else {
1124 		bp->b_iocmd = BIO_READ;
1125 		if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_WRITE)) {
1126 			error = EFAULT;
1127 			goto doerror;
1128 		}
1129 	}
1130 
1131 	/* Bring buffer into kernel space. */
1132 	vmapbuf(bp);
1133 
1134 	s = splbio();
1135 	aiocbe->bp = bp;
1136 	bp->b_spc = (void *)aiocbe;
1137 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1138 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1139 	aiocbe->jobstate = JOBST_JOBQBUF;
1140 	cb->_aiocb_private.status = cb->aio_nbytes;
1141 	num_buf_aio++;
1142 	bp->b_error = 0;
1143 
1144 	splx(s);
1145 
1146 	/* Perform transfer. */
1147 	DEV_STRATEGY(bp, 0);
1148 
1149 	notify = 0;
1150 	s = splbio();
1151 
1152 	/*
1153 	 * If we had an error invoking the request, or an error in processing
1154 	 * the request before we have returned, we process it as an error in
1155 	 * transfer.  Note that such an I/O error is not indicated immediately,
1156 	 * but is returned using the aio_error mechanism.  In this case,
1157 	 * aio_suspend will return immediately.
1158 	 */
1159 	if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1160 		struct aiocb *job = aiocbe->uuaiocb;
1161 
1162 		aiocbe->uaiocb._aiocb_private.status = 0;
1163 		suword(&job->_aiocb_private.status, 0);
1164 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1165 		suword(&job->_aiocb_private.error, bp->b_error);
1166 
1167 		ki->kaio_buffer_finished_count++;
1168 
1169 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1170 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1171 			aiocbe->jobflags |= AIOCBLIST_DONE;
1172 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1173 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1174 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1175 			notify = 1;
1176 		}
1177 	}
1178 	splx(s);
1179 	if (notify)
1180 		KNOTE(&aiocbe->klist, 0);
1181 	return 0;
1182 
1183 doerror:
1184 	ki->kaio_buffer_count--;
1185 	if (lj)
1186 		lj->lioj_buffer_count--;
1187 	aiocbe->bp = NULL;
1188 	relpbuf(bp, NULL);
1189 	return error;
1190 }
1191 
1192 /*
1193  * This waits/tests physio completion.
1194  */
1195 static int
1196 aio_fphysio(struct aiocblist *iocb)
1197 {
1198 	int s;
1199 	struct buf *bp;
1200 	int error;
1201 
1202 	bp = iocb->bp;
1203 
1204 	s = splbio();
1205 	while ((bp->b_flags & B_DONE) == 0) {
1206 		if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1207 			if ((bp->b_flags & B_DONE) == 0) {
1208 				splx(s);
1209 				return EINPROGRESS;
1210 			} else
1211 				break;
1212 		}
1213 	}
1214 	splx(s);
1215 
1216 	/* Release mapping into kernel space. */
1217 	vunmapbuf(bp);
1218 	iocb->bp = 0;
1219 
1220 	error = 0;
1221 
1222 	/* Check for an error. */
1223 	if (bp->b_ioflags & BIO_ERROR)
1224 		error = bp->b_error;
1225 
1226 	relpbuf(bp, NULL);
1227 	return (error);
1228 }
1229 
1230 /*
1231  * Wake up aio requests that may be serviceable now.
1232  */
1233 static void
1234 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1235 {
1236 	struct aiocblist *cb,*cbn;
1237 	struct proc *p;
1238 	struct kaioinfo *ki = NULL;
1239 	int opcode, wakecount = 0;
1240 	struct aiothreadlist *aiop;
1241 
1242 	if (sb == &so->so_snd) {
1243 		opcode = LIO_WRITE;
1244 		so->so_snd.sb_flags &= ~SB_AIO;
1245 	} else {
1246 		opcode = LIO_READ;
1247 		so->so_rcv.sb_flags &= ~SB_AIO;
1248 	}
1249 
1250 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1251 		cbn = TAILQ_NEXT(cb, list);
1252 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1253 			p = cb->userproc;
1254 			ki = p->p_aioinfo;
1255 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1256 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1257 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1258 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1259 			wakecount++;
1260 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1261 				panic("invalid queue value");
1262 		}
1263 	}
1264 
1265 	while (wakecount--) {
1266 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1267 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1268 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1269 			aiop->aiothreadflags &= ~AIOP_FREE;
1270 			wakeup(aiop->aiothread);
1271 		}
1272 	}
1273 }
1274 
1275 /*
1276  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1277  * technique is done in this code.
1278  */
1279 static int
1280 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1281 {
1282 	struct proc *p = td->td_proc;
1283 	struct filedesc *fdp;
1284 	struct file *fp;
1285 	unsigned int fd;
1286 	struct socket *so;
1287 	int s;
1288 	int error;
1289 	int opcode;
1290 	struct aiocblist *aiocbe;
1291 	struct aiothreadlist *aiop;
1292 	struct kaioinfo *ki;
1293 	struct kevent kev;
1294 	struct kqueue *kq;
1295 	struct file *kq_fp;
1296 
1297 	aiocbe = zalloc(aiocb_zone);
1298 	aiocbe->inputcharge = 0;
1299 	aiocbe->outputcharge = 0;
1300 	callout_handle_init(&aiocbe->timeouthandle);
1301 	SLIST_INIT(&aiocbe->klist);
1302 
1303 	suword(&job->_aiocb_private.status, -1);
1304 	suword(&job->_aiocb_private.error, 0);
1305 	suword(&job->_aiocb_private.kernelinfo, -1);
1306 
1307 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1308 	if (error) {
1309 		suword(&job->_aiocb_private.error, error);
1310 		zfree(aiocb_zone, aiocbe);
1311 		return error;
1312 	}
1313 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1314 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1315 		zfree(aiocb_zone, aiocbe);
1316 		return EINVAL;
1317 	}
1318 
1319 	/* Save userspace address of the job info. */
1320 	aiocbe->uuaiocb = job;
1321 
1322 	/* Get the opcode. */
1323 	if (type != LIO_NOP)
1324 		aiocbe->uaiocb.aio_lio_opcode = type;
1325 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1326 
1327 	/* Get the fd info for process. */
1328 	fdp = p->p_fd;
1329 
1330 	/*
1331 	 * Range check file descriptor.
1332 	 */
1333 	fd = aiocbe->uaiocb.aio_fildes;
1334 	if (fd >= fdp->fd_nfiles) {
1335 		zfree(aiocb_zone, aiocbe);
1336 		if (type == 0)
1337 			suword(&job->_aiocb_private.error, EBADF);
1338 		return EBADF;
1339 	}
1340 
1341 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1342 	if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1343 	    0))) {
1344 		zfree(aiocb_zone, aiocbe);
1345 		if (type == 0)
1346 			suword(&job->_aiocb_private.error, EBADF);
1347 		return EBADF;
1348 	}
1349 
1350 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1351 		zfree(aiocb_zone, aiocbe);
1352 		if (type == 0)
1353 			suword(&job->_aiocb_private.error, EINVAL);
1354 		return EINVAL;
1355 	}
1356 
1357 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1358 	if (error) {
1359 		zfree(aiocb_zone, aiocbe);
1360 		if (type == 0)
1361 			suword(&job->_aiocb_private.error, EINVAL);
1362 		return error;
1363 	}
1364 
1365 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1366 	if (jobrefid == LONG_MAX)
1367 		jobrefid = 1;
1368 	else
1369 		jobrefid++;
1370 
1371 	if (opcode == LIO_NOP) {
1372 		zfree(aiocb_zone, aiocbe);
1373 		if (type == 0) {
1374 			suword(&job->_aiocb_private.error, 0);
1375 			suword(&job->_aiocb_private.status, 0);
1376 			suword(&job->_aiocb_private.kernelinfo, 0);
1377 		}
1378 		return 0;
1379 	}
1380 
1381 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1382 		zfree(aiocb_zone, aiocbe);
1383 		if (type == 0) {
1384 			suword(&job->_aiocb_private.status, 0);
1385 			suword(&job->_aiocb_private.error, EINVAL);
1386 		}
1387 		return EINVAL;
1388 	}
1389 
1390 	fhold(fp);
1391 
1392 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1393 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1394 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1395 	}
1396 	else {
1397 		/*
1398 		 * This method for requesting kevent-based notification won't
1399 		 * work on the alpha, since we're passing in a pointer
1400 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1401 		 * based method instead.
1402 		 */
1403 		struct kevent *kevp;
1404 
1405 		kevp = (struct kevent *)(uintptr_t)job->aio_lio_opcode;
1406 		if (kevp == NULL)
1407 			goto no_kqueue;
1408 
1409 		error = copyin(kevp, &kev, sizeof(kev));
1410 		if (error)
1411 			goto aqueue_fail;
1412 	}
1413 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1414 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1415 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1416 		error = EBADF;
1417 		goto aqueue_fail;
1418 	}
1419 	kq = (struct kqueue *)kq_fp->f_data;
1420 	kev.ident = (uintptr_t)aiocbe;
1421 	kev.filter = EVFILT_AIO;
1422 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1423 	error = kqueue_register(kq, &kev, td);
1424 aqueue_fail:
1425 	if (error) {
1426 		zfree(aiocb_zone, aiocbe);
1427 		if (type == 0)
1428 			suword(&job->_aiocb_private.error, error);
1429 		goto done;
1430 	}
1431 no_kqueue:
1432 
1433 	suword(&job->_aiocb_private.error, EINPROGRESS);
1434 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1435 	aiocbe->userproc = p;
1436 	aiocbe->jobflags = 0;
1437 	aiocbe->lio = lj;
1438 	ki = p->p_aioinfo;
1439 
1440 	if (fp->f_type == DTYPE_SOCKET) {
1441 		/*
1442 		 * Alternate queueing for socket ops: Reach down into the
1443 		 * descriptor to get the socket data.  Then check to see if the
1444 		 * socket is ready to be read or written (based on the requested
1445 		 * operation).
1446 		 *
1447 		 * If it is not ready for io, then queue the aiocbe on the
1448 		 * socket, and set the flags so we get a call when sbnotify()
1449 		 * happens.
1450 		 */
1451 		so = (struct socket *)fp->f_data;
1452 		s = splnet();
1453 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1454 		    LIO_WRITE) && (!sowriteable(so)))) {
1455 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1456 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1457 			if (opcode == LIO_READ)
1458 				so->so_rcv.sb_flags |= SB_AIO;
1459 			else
1460 				so->so_snd.sb_flags |= SB_AIO;
1461 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1462 			ki->kaio_queue_count++;
1463 			num_queue_count++;
1464 			splx(s);
1465 			error = 0;
1466 			goto done;
1467 		}
1468 		splx(s);
1469 	}
1470 
1471 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1472 		goto done;
1473 	if (error > 0) {
1474 		suword(&job->_aiocb_private.status, 0);
1475 		aiocbe->uaiocb._aiocb_private.error = error;
1476 		suword(&job->_aiocb_private.error, error);
1477 		goto done;
1478 	}
1479 
1480 	/* No buffer for daemon I/O. */
1481 	aiocbe->bp = NULL;
1482 
1483 	ki->kaio_queue_count++;
1484 	if (lj)
1485 		lj->lioj_queue_count++;
1486 	s = splnet();
1487 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1488 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1489 	splx(s);
1490 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1491 
1492 	num_queue_count++;
1493 	error = 0;
1494 
1495 	/*
1496 	 * If we don't have a free AIO process, and we are below our quota, then
1497 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1498 	 * pick-up this job.  If we don't sucessfully create the new process
1499 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1500 	 * which is likely not the correct thing to do.
1501 	 */
1502 	s = splnet();
1503 retryproc:
1504 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1505 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1506 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1507 		aiop->aiothreadflags &= ~AIOP_FREE;
1508 		wakeup(aiop->aiothread);
1509 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1510 	    ((ki->kaio_active_count + num_aio_resv_start) <
1511 	    ki->kaio_maxactive_count)) {
1512 		num_aio_resv_start++;
1513 		if ((error = aio_newproc()) == 0) {
1514 			num_aio_resv_start--;
1515 			goto retryproc;
1516 		}
1517 		num_aio_resv_start--;
1518 	}
1519 	splx(s);
1520 done:
1521 	fdrop(fp, td);
1522 	return error;
1523 }
1524 
1525 /*
1526  * This routine queues an AIO request, checking for quotas.
1527  */
1528 static int
1529 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1530 {
1531 	struct proc *p = td->td_proc;
1532 	struct kaioinfo *ki;
1533 
1534 	if (p->p_aioinfo == NULL)
1535 		aio_init_aioinfo(p);
1536 
1537 	if (num_queue_count >= max_queue_count)
1538 		return EAGAIN;
1539 
1540 	ki = p->p_aioinfo;
1541 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1542 		return EAGAIN;
1543 
1544 	return _aio_aqueue(td, job, NULL, type);
1545 }
1546 
1547 /*
1548  * Support the aio_return system call, as a side-effect, kernel resources are
1549  * released.
1550  */
1551 int
1552 aio_return(struct thread *td, struct aio_return_args *uap)
1553 {
1554 	struct proc *p = td->td_proc;
1555 	int s;
1556 	int jobref;
1557 	struct aiocblist *cb, *ncb;
1558 	struct aiocb *ujob;
1559 	struct kaioinfo *ki;
1560 
1561 	ki = p->p_aioinfo;
1562 	if (ki == NULL)
1563 		return EINVAL;
1564 
1565 	ujob = uap->aiocbp;
1566 
1567 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1568 	if (jobref == -1 || jobref == 0)
1569 		return EINVAL;
1570 
1571 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1572 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1573 		    jobref) {
1574 			if (ujob == cb->uuaiocb) {
1575 				td->td_retval[0] =
1576 				    cb->uaiocb._aiocb_private.status;
1577 			} else
1578 				td->td_retval[0] = EFAULT;
1579 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1580 				p->p_stats->p_ru.ru_oublock +=
1581 				    cb->outputcharge;
1582 				cb->outputcharge = 0;
1583 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1584 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1585 				cb->inputcharge = 0;
1586 			}
1587 			aio_free_entry(cb);
1588 			return 0;
1589 		}
1590 	}
1591 	s = splbio();
1592 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1593 		ncb = TAILQ_NEXT(cb, plist);
1594 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1595 		    == jobref) {
1596 			splx(s);
1597 			if (ujob == cb->uuaiocb) {
1598 				td->td_retval[0] =
1599 				    cb->uaiocb._aiocb_private.status;
1600 			} else
1601 				td->td_retval[0] = EFAULT;
1602 			aio_free_entry(cb);
1603 			return 0;
1604 		}
1605 	}
1606 	splx(s);
1607 
1608 	return (EINVAL);
1609 }
1610 
1611 /*
1612  * Allow a process to wakeup when any of the I/O requests are completed.
1613  */
1614 int
1615 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1616 {
1617 	struct proc *p = td->td_proc;
1618 	struct timeval atv;
1619 	struct timespec ts;
1620 	struct aiocb *const *cbptr, *cbp;
1621 	struct kaioinfo *ki;
1622 	struct aiocblist *cb;
1623 	int i;
1624 	int njoblist;
1625 	int error, s, timo;
1626 	int *ijoblist;
1627 	struct aiocb **ujoblist;
1628 
1629 	if (uap->nent > AIO_LISTIO_MAX)
1630 		return EINVAL;
1631 
1632 	timo = 0;
1633 	if (uap->timeout) {
1634 		/* Get timespec struct. */
1635 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1636 			return error;
1637 
1638 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1639 			return (EINVAL);
1640 
1641 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1642 		if (itimerfix(&atv))
1643 			return (EINVAL);
1644 		timo = tvtohz(&atv);
1645 	}
1646 
1647 	ki = p->p_aioinfo;
1648 	if (ki == NULL)
1649 		return EAGAIN;
1650 
1651 	njoblist = 0;
1652 	ijoblist = zalloc(aiol_zone);
1653 	ujoblist = zalloc(aiol_zone);
1654 	cbptr = uap->aiocbp;
1655 
1656 	for (i = 0; i < uap->nent; i++) {
1657 		cbp = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
1658 		if (cbp == 0)
1659 			continue;
1660 		ujoblist[njoblist] = cbp;
1661 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1662 		njoblist++;
1663 	}
1664 
1665 	if (njoblist == 0) {
1666 		zfree(aiol_zone, ijoblist);
1667 		zfree(aiol_zone, ujoblist);
1668 		return 0;
1669 	}
1670 
1671 	error = 0;
1672 	for (;;) {
1673 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1674 			for (i = 0; i < njoblist; i++) {
1675 				if (((intptr_t)
1676 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1677 				    ijoblist[i]) {
1678 					if (ujoblist[i] != cb->uuaiocb)
1679 						error = EINVAL;
1680 					zfree(aiol_zone, ijoblist);
1681 					zfree(aiol_zone, ujoblist);
1682 					return error;
1683 				}
1684 			}
1685 		}
1686 
1687 		s = splbio();
1688 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1689 		    TAILQ_NEXT(cb, plist)) {
1690 			for (i = 0; i < njoblist; i++) {
1691 				if (((intptr_t)
1692 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1693 				    ijoblist[i]) {
1694 					splx(s);
1695 					if (ujoblist[i] != cb->uuaiocb)
1696 						error = EINVAL;
1697 					zfree(aiol_zone, ijoblist);
1698 					zfree(aiol_zone, ujoblist);
1699 					return error;
1700 				}
1701 			}
1702 		}
1703 
1704 		ki->kaio_flags |= KAIO_WAKEUP;
1705 		error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1706 		splx(s);
1707 
1708 		if (error == ERESTART || error == EINTR) {
1709 			zfree(aiol_zone, ijoblist);
1710 			zfree(aiol_zone, ujoblist);
1711 			return EINTR;
1712 		} else if (error == EWOULDBLOCK) {
1713 			zfree(aiol_zone, ijoblist);
1714 			zfree(aiol_zone, ujoblist);
1715 			return EAGAIN;
1716 		}
1717 	}
1718 
1719 /* NOTREACHED */
1720 	return EINVAL;
1721 }
1722 
1723 /*
1724  * aio_cancel cancels any non-physio aio operations not currently in
1725  * progress.
1726  */
1727 int
1728 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1729 {
1730 	struct proc *p = td->td_proc;
1731 	struct kaioinfo *ki;
1732 	struct aiocblist *cbe, *cbn;
1733 	struct file *fp;
1734 	struct filedesc *fdp;
1735 	struct socket *so;
1736 	struct proc *po;
1737 	int s,error;
1738 	int cancelled=0;
1739 	int notcancelled=0;
1740 	struct vnode *vp;
1741 
1742 	fdp = p->p_fd;
1743 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1744 	    (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1745 		return (EBADF);
1746 
1747         if (fp->f_type == DTYPE_VNODE) {
1748 		vp = (struct vnode *)fp->f_data;
1749 
1750 		if (vn_isdisk(vp,&error)) {
1751 			td->td_retval[0] = AIO_NOTCANCELED;
1752         	        return 0;
1753 		}
1754 	} else if (fp->f_type == DTYPE_SOCKET) {
1755 		so = (struct socket *)fp->f_data;
1756 
1757 		s = splnet();
1758 
1759 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1760 			cbn = TAILQ_NEXT(cbe, list);
1761 			if ((uap->aiocbp == NULL) ||
1762 				(uap->aiocbp == cbe->uuaiocb) ) {
1763 				po = cbe->userproc;
1764 				ki = po->p_aioinfo;
1765 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1766 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1767 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1768 				if (ki->kaio_flags & KAIO_WAKEUP) {
1769 					wakeup(po);
1770 				}
1771 				cbe->jobstate = JOBST_JOBFINISHED;
1772 				cbe->uaiocb._aiocb_private.status=-1;
1773 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1774 				cancelled++;
1775 /* XXX cancelled, knote? */
1776 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1777 				    SIGEV_SIGNAL) {
1778 					PROC_LOCK(cbe->userproc);
1779 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1780 					PROC_UNLOCK(cbe->userproc);
1781 				}
1782 				if (uap->aiocbp)
1783 					break;
1784 			}
1785 		}
1786 		splx(s);
1787 
1788 		if ((cancelled) && (uap->aiocbp)) {
1789 			td->td_retval[0] = AIO_CANCELED;
1790 			return 0;
1791 		}
1792 	}
1793 	ki=p->p_aioinfo;
1794 	s = splnet();
1795 
1796 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1797 		cbn = TAILQ_NEXT(cbe, plist);
1798 
1799 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1800 		    ((uap->aiocbp == NULL ) ||
1801 		     (uap->aiocbp == cbe->uuaiocb))) {
1802 
1803 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1804 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1805                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1806                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1807                                     plist);
1808 				cancelled++;
1809 				ki->kaio_queue_finished_count++;
1810 				cbe->jobstate = JOBST_JOBFINISHED;
1811 				cbe->uaiocb._aiocb_private.status = -1;
1812 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1813 /* XXX cancelled, knote? */
1814 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1815 				    SIGEV_SIGNAL) {
1816 					PROC_LOCK(cbe->userproc);
1817 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1818 					PROC_UNLOCK(cbe->userproc);
1819 				}
1820 			} else {
1821 				notcancelled++;
1822 			}
1823 		}
1824 	}
1825 	splx(s);
1826 
1827 	if (notcancelled) {
1828 		td->td_retval[0] = AIO_NOTCANCELED;
1829 		return 0;
1830 	}
1831 	if (cancelled) {
1832 		td->td_retval[0] = AIO_CANCELED;
1833 		return 0;
1834 	}
1835 	td->td_retval[0] = AIO_ALLDONE;
1836 
1837 	return 0;
1838 }
1839 
1840 /*
1841  * aio_error is implemented in the kernel level for compatibility purposes only.
1842  * For a user mode async implementation, it would be best to do it in a userland
1843  * subroutine.
1844  */
1845 int
1846 aio_error(struct thread *td, struct aio_error_args *uap)
1847 {
1848 	struct proc *p = td->td_proc;
1849 	int s;
1850 	struct aiocblist *cb;
1851 	struct kaioinfo *ki;
1852 	int jobref;
1853 
1854 	ki = p->p_aioinfo;
1855 	if (ki == NULL)
1856 		return EINVAL;
1857 
1858 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1859 	if ((jobref == -1) || (jobref == 0))
1860 		return EINVAL;
1861 
1862 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1863 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1864 		    jobref) {
1865 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1866 			return 0;
1867 		}
1868 	}
1869 
1870 	s = splnet();
1871 
1872 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1873 	    plist)) {
1874 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1875 		    jobref) {
1876 			td->td_retval[0] = EINPROGRESS;
1877 			splx(s);
1878 			return 0;
1879 		}
1880 	}
1881 
1882 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1883 	    plist)) {
1884 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1885 		    jobref) {
1886 			td->td_retval[0] = EINPROGRESS;
1887 			splx(s);
1888 			return 0;
1889 		}
1890 	}
1891 	splx(s);
1892 
1893 	s = splbio();
1894 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1895 	    plist)) {
1896 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1897 		    jobref) {
1898 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1899 			splx(s);
1900 			return 0;
1901 		}
1902 	}
1903 
1904 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1905 	    plist)) {
1906 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1907 		    jobref) {
1908 			td->td_retval[0] = EINPROGRESS;
1909 			splx(s);
1910 			return 0;
1911 		}
1912 	}
1913 	splx(s);
1914 
1915 #if (0)
1916 	/*
1917 	 * Hack for lio.
1918 	 */
1919 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1920 	if (status == -1)
1921 		return fuword(&uap->aiocbp->_aiocb_private.error);
1922 #endif
1923 	return EINVAL;
1924 }
1925 
1926 int
1927 aio_read(struct thread *td, struct aio_read_args *uap)
1928 {
1929 
1930 	return aio_aqueue(td, uap->aiocbp, LIO_READ);
1931 }
1932 
1933 int
1934 aio_write(struct thread *td, struct aio_write_args *uap)
1935 {
1936 
1937 	return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1938 }
1939 
1940 int
1941 lio_listio(struct thread *td, struct lio_listio_args *uap)
1942 {
1943 	struct proc *p = td->td_proc;
1944 	int nent, nentqueued;
1945 	struct aiocb *iocb, * const *cbptr;
1946 	struct aiocblist *cb;
1947 	struct kaioinfo *ki;
1948 	struct aio_liojob *lj;
1949 	int error, runningcode;
1950 	int nerror;
1951 	int i;
1952 	int s;
1953 
1954 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1955 		return EINVAL;
1956 
1957 	nent = uap->nent;
1958 	if (nent > AIO_LISTIO_MAX)
1959 		return EINVAL;
1960 
1961 	if (p->p_aioinfo == NULL)
1962 		aio_init_aioinfo(p);
1963 
1964 	if ((nent + num_queue_count) > max_queue_count)
1965 		return EAGAIN;
1966 
1967 	ki = p->p_aioinfo;
1968 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1969 		return EAGAIN;
1970 
1971 	lj = zalloc(aiolio_zone);
1972 	if (!lj)
1973 		return EAGAIN;
1974 
1975 	lj->lioj_flags = 0;
1976 	lj->lioj_buffer_count = 0;
1977 	lj->lioj_buffer_finished_count = 0;
1978 	lj->lioj_queue_count = 0;
1979 	lj->lioj_queue_finished_count = 0;
1980 	lj->lioj_ki = ki;
1981 
1982 	/*
1983 	 * Setup signal.
1984 	 */
1985 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1986 		error = copyin(uap->sig, &lj->lioj_signal,
1987 			       sizeof(lj->lioj_signal));
1988 		if (error) {
1989 			zfree(aiolio_zone, lj);
1990 			return error;
1991 		}
1992 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1993 			zfree(aiolio_zone, lj);
1994 			return EINVAL;
1995 		}
1996 		lj->lioj_flags |= LIOJ_SIGNAL;
1997 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1998 	} else
1999 		lj->lioj_flags &= ~LIOJ_SIGNAL;
2000 
2001 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2002 	/*
2003 	 * Get pointers to the list of I/O requests.
2004 	 */
2005 	nerror = 0;
2006 	nentqueued = 0;
2007 	cbptr = uap->acb_list;
2008 	for (i = 0; i < uap->nent; i++) {
2009 		iocb = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
2010 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != NULL)) {
2011 			error = _aio_aqueue(td, iocb, lj, 0);
2012 			if (error == 0)
2013 				nentqueued++;
2014 			else
2015 				nerror++;
2016 		}
2017 	}
2018 
2019 	/*
2020 	 * If we haven't queued any, then just return error.
2021 	 */
2022 	if (nentqueued == 0)
2023 		return 0;
2024 
2025 	/*
2026 	 * Calculate the appropriate error return.
2027 	 */
2028 	runningcode = 0;
2029 	if (nerror)
2030 		runningcode = EIO;
2031 
2032 	if (uap->mode == LIO_WAIT) {
2033 		int command, found, jobref;
2034 
2035 		for (;;) {
2036 			found = 0;
2037 			for (i = 0; i < uap->nent; i++) {
2038 				/*
2039 				 * Fetch address of the control buf pointer in
2040 				 * user space.
2041 				 */
2042 				iocb = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
2043 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2044 				    == 0))
2045 					continue;
2046 
2047 				/*
2048 				 * Fetch the associated command from user space.
2049 				 */
2050 				command = fuword(&iocb->aio_lio_opcode);
2051 				if (command == LIO_NOP) {
2052 					found++;
2053 					continue;
2054 				}
2055 
2056 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
2057 
2058 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2059 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2060 					    == jobref) {
2061 						if (cb->uaiocb.aio_lio_opcode
2062 						    == LIO_WRITE) {
2063 							p->p_stats->p_ru.ru_oublock
2064 							    +=
2065 							    cb->outputcharge;
2066 							cb->outputcharge = 0;
2067 						} else if (cb->uaiocb.aio_lio_opcode
2068 						    == LIO_READ) {
2069 							p->p_stats->p_ru.ru_inblock
2070 							    += cb->inputcharge;
2071 							cb->inputcharge = 0;
2072 						}
2073 						found++;
2074 						break;
2075 					}
2076 				}
2077 
2078 				s = splbio();
2079 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2080 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2081 					    == jobref) {
2082 						found++;
2083 						break;
2084 					}
2085 				}
2086 				splx(s);
2087 			}
2088 
2089 			/*
2090 			 * If all I/Os have been disposed of, then we can
2091 			 * return.
2092 			 */
2093 			if (found == nentqueued)
2094 				return runningcode;
2095 
2096 			ki->kaio_flags |= KAIO_WAKEUP;
2097 			error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2098 
2099 			if (error == EINTR)
2100 				return EINTR;
2101 			else if (error == EWOULDBLOCK)
2102 				return EAGAIN;
2103 		}
2104 	}
2105 
2106 	return runningcode;
2107 }
2108 
2109 /*
2110  * This is a weird hack so that we can post a signal.  It is safe to do so from
2111  * a timeout routine, but *not* from an interrupt routine.
2112  */
2113 static void
2114 process_signal(void *aioj)
2115 {
2116 	struct aiocblist *aiocbe = aioj;
2117 	struct aio_liojob *lj = aiocbe->lio;
2118 	struct aiocb *cb = &aiocbe->uaiocb;
2119 
2120 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2121 		(lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2122 		PROC_LOCK(lj->lioj_ki->kaio_p);
2123 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2124 		PROC_UNLOCK(lj->lioj_ki->kaio_p);
2125 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2126 	}
2127 
2128 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2129 		PROC_LOCK(aiocbe->userproc);
2130 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2131 		PROC_UNLOCK(aiocbe->userproc);
2132 	}
2133 }
2134 
2135 /*
2136  * Interrupt handler for physio, performs the necessary process wakeups, and
2137  * signals.
2138  */
2139 static void
2140 aio_physwakeup(struct buf *bp)
2141 {
2142 	struct aiocblist *aiocbe;
2143 	struct proc *p;
2144 	struct kaioinfo *ki;
2145 	struct aio_liojob *lj;
2146 
2147 	wakeup(bp);
2148 
2149 	aiocbe = (struct aiocblist *)bp->b_spc;
2150 	if (aiocbe) {
2151 		p = bp->b_caller1;
2152 
2153 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2154 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2155 		aiocbe->uaiocb._aiocb_private.error = 0;
2156 		aiocbe->jobflags |= AIOCBLIST_DONE;
2157 
2158 		if (bp->b_ioflags & BIO_ERROR)
2159 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2160 
2161 		lj = aiocbe->lio;
2162 		if (lj) {
2163 			lj->lioj_buffer_finished_count++;
2164 
2165 			/*
2166 			 * wakeup/signal if all of the interrupt jobs are done.
2167 			 */
2168 			if (lj->lioj_buffer_finished_count ==
2169 			    lj->lioj_buffer_count) {
2170 				/*
2171 				 * Post a signal if it is called for.
2172 				 */
2173 				if ((lj->lioj_flags &
2174 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2175 				    LIOJ_SIGNAL) {
2176 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2177 					aiocbe->timeouthandle =
2178 						timeout(process_signal,
2179 							aiocbe, 0);
2180 				}
2181 			}
2182 		}
2183 
2184 		ki = p->p_aioinfo;
2185 		if (ki) {
2186 			ki->kaio_buffer_finished_count++;
2187 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2188 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2189 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2190 
2191 			KNOTE(&aiocbe->klist, 0);
2192 			/* Do the wakeup. */
2193 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2194 				ki->kaio_flags &= ~KAIO_WAKEUP;
2195 				wakeup(p);
2196 			}
2197 		}
2198 
2199 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2200 			aiocbe->timeouthandle =
2201 				timeout(process_signal, aiocbe, 0);
2202 	}
2203 }
2204 
2205 int
2206 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2207 {
2208 	struct proc *p = td->td_proc;
2209 	struct timeval atv;
2210 	struct timespec ts;
2211 	struct aiocb **cbptr;
2212 	struct kaioinfo *ki;
2213 	struct aiocblist *cb = NULL;
2214 	int error, s, timo;
2215 
2216 	suword(uap->aiocbp, (int)NULL);
2217 
2218 	timo = 0;
2219 	if (uap->timeout) {
2220 		/* Get timespec struct. */
2221 		error = copyin(uap->timeout, &ts, sizeof(ts));
2222 		if (error)
2223 			return error;
2224 
2225 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2226 			return (EINVAL);
2227 
2228 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2229 		if (itimerfix(&atv))
2230 			return (EINVAL);
2231 		timo = tvtohz(&atv);
2232 	}
2233 
2234 	ki = p->p_aioinfo;
2235 	if (ki == NULL)
2236 		return EAGAIN;
2237 
2238 	cbptr = uap->aiocbp;
2239 
2240 	for (;;) {
2241 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2242 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2243 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2244 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2245 				p->p_stats->p_ru.ru_oublock +=
2246 				    cb->outputcharge;
2247 				cb->outputcharge = 0;
2248 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2249 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2250 				cb->inputcharge = 0;
2251 			}
2252 			aio_free_entry(cb);
2253 			return cb->uaiocb._aiocb_private.error;
2254 		}
2255 
2256 		s = splbio();
2257  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2258 			splx(s);
2259 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2260 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2261 			aio_free_entry(cb);
2262 			return cb->uaiocb._aiocb_private.error;
2263 		}
2264 
2265 		ki->kaio_flags |= KAIO_WAKEUP;
2266 		error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2267 		splx(s);
2268 
2269 		if (error == ERESTART)
2270 			return EINTR;
2271 		else if (error < 0)
2272 			return error;
2273 		else if (error == EINTR)
2274 			return EINTR;
2275 		else if (error == EWOULDBLOCK)
2276 			return EAGAIN;
2277 	}
2278 }
2279 
2280 static int
2281 filt_aioattach(struct knote *kn)
2282 {
2283 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2284 
2285 	/*
2286 	 * The aiocbe pointer must be validated before using it, so
2287 	 * registration is restricted to the kernel; the user cannot
2288 	 * set EV_FLAG1.
2289 	 */
2290 	if ((kn->kn_flags & EV_FLAG1) == 0)
2291 		return (EPERM);
2292 	kn->kn_flags &= ~EV_FLAG1;
2293 
2294 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2295 
2296 	return (0);
2297 }
2298 
2299 static void
2300 filt_aiodetach(struct knote *kn)
2301 {
2302 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2303 
2304 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2305 }
2306 
2307 /*ARGSUSED*/
2308 static int
2309 filt_aio(struct knote *kn, long hint)
2310 {
2311 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2312 
2313 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2314 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2315 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2316 		return (0);
2317 	kn->kn_flags |= EV_EOF;
2318 	return (1);
2319 }
2320