xref: /freebsd/sys/kern/vfs_aio.c (revision 2ad872c5794e4c26fdf6ed219ad3f09ca0d5304a)
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  * $Id: vfs_aio.c,v 1.35 1998/11/27 01:14:21 tegge Exp $
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/sysproto.h>
26 #include <sys/filedesc.h>
27 #include <sys/kernel.h>
28 #include <sys/fcntl.h>
29 #include <sys/file.h>
30 #include <sys/lock.h>
31 #include <sys/unistd.h>
32 #include <sys/proc.h>
33 #include <sys/resourcevar.h>
34 #include <sys/signalvar.h>
35 #include <sys/sysctl.h>
36 #include <sys/vnode.h>
37 #include <sys/conf.h>
38 #include <miscfs/specfs/specdev.h>
39 
40 #include <vm/vm.h>
41 #include <vm/vm_param.h>
42 #include <vm/vm_extern.h>
43 #include <vm/pmap.h>
44 #include <vm/vm_map.h>
45 #include <vm/vm_zone.h>
46 #include <sys/aio.h>
47 #include <sys/shm.h>
48 
49 #include <machine/cpu.h>
50 #include <machine/limits.h>
51 
52 static	long jobrefid;
53 
54 #define JOBST_NULL			0x0
55 #define	JOBST_JOBQPROC		0x1
56 #define JOBST_JOBQGLOBAL	0x2
57 #define JOBST_JOBRUNNING	0x3
58 #define JOBST_JOBFINISHED	0x4
59 #define	JOBST_JOBQBUF		0x5
60 #define	JOBST_JOBBFINISHED	0x6
61 
62 #ifndef MAX_AIO_PER_PROC
63 #define MAX_AIO_PER_PROC	32
64 #endif
65 
66 #ifndef MAX_AIO_QUEUE_PER_PROC
67 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
68 #endif
69 
70 #ifndef MAX_AIO_PROCS
71 #define MAX_AIO_PROCS		32
72 #endif
73 
74 #ifndef MAX_AIO_QUEUE
75 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
76 #endif
77 
78 #ifndef TARGET_AIO_PROCS
79 #define TARGET_AIO_PROCS	0
80 #endif
81 
82 #ifndef MAX_BUF_AIO
83 #define MAX_BUF_AIO 16
84 #endif
85 
86 #ifndef AIOD_TIMEOUT_DEFAULT
87 #define	AIOD_TIMEOUT_DEFAULT (10 * hz)
88 #endif
89 
90 #ifndef AIOD_LIFETIME_DEFAULT
91 #define AIOD_LIFETIME_DEFAULT (30 * hz)
92 #endif
93 
94 static int max_aio_procs = MAX_AIO_PROCS;
95 static int num_aio_procs = 0;
96 static int target_aio_procs = TARGET_AIO_PROCS;
97 static int max_queue_count = MAX_AIO_QUEUE;
98 static int num_queue_count = 0;
99 static int num_buf_aio = 0;
100 static int num_aio_resv_start = 0;
101 static int aiod_timeout;
102 static int aiod_lifetime;
103 
104 static int max_aio_per_proc = MAX_AIO_PER_PROC,
105 	max_aio_queue_per_proc=MAX_AIO_QUEUE_PER_PROC;
106 
107 static int max_buf_aio = MAX_BUF_AIO;
108 
109 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "AIO mgmt");
110 
111 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc,
112 	CTLFLAG_RW, &max_aio_per_proc, 0, "");
113 
114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc,
115 	CTLFLAG_RW, &max_aio_queue_per_proc, 0, "");
116 
117 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
118 	CTLFLAG_RW, &max_aio_procs, 0, "");
119 
120 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
121 	CTLFLAG_RD, &num_aio_procs, 0, "");
122 
123 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count,
124 	CTLFLAG_RD, &num_queue_count, 0, "");
125 
126 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue,
127 	CTLFLAG_RW, &max_queue_count, 0, "");
128 
129 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs,
130 	CTLFLAG_RW, &target_aio_procs, 0, "");
131 
132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio,
133 	CTLFLAG_RW, &max_buf_aio, 0, "");
134 
135 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio,
136 	CTLFLAG_RD, &num_buf_aio, 0, "");
137 
138 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime,
139 	CTLFLAG_RW, &aiod_lifetime, 0, "");
140 
141 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout,
142 	CTLFLAG_RW, &aiod_timeout, 0, "");
143 
144 
145 /*
146  * Job queue item
147  */
148 
149 #define AIOCBLIST_CANCELLED	0x1
150 #define AIOCBLIST_RUNDOWN	0x4
151 #define AIOCBLIST_ASYNCFREE	0x8
152 #define AIOCBLIST_DONE		0x10
153 
154 struct aiocblist {
155 	TAILQ_ENTRY (aiocblist) list;		/* List of jobs */
156 	TAILQ_ENTRY (aiocblist) plist;		/* List of jobs for proc */
157 	int	jobflags;
158 	int	jobstate;
159 	int inputcharge, outputcharge;
160 	struct	buf *bp;				/* buffer pointer */
161 	struct	proc *userproc;			/* User process */
162 	struct	aioproclist	*jobaioproc;	/* AIO process descriptor */
163 	struct	aio_liojob	*lio;		/* optional lio job */
164 	struct	aiocb *uuaiocb;			/* pointer in userspace of aiocb */
165 	struct	aiocb uaiocb;			/* Kernel I/O control block */
166 };
167 
168 
169 /*
170  * AIO process info
171  */
172 #define AIOP_FREE	0x1			/* proc on free queue */
173 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
174 
175 struct aioproclist {
176 	int aioprocflags;			/* AIO proc flags */
177 	TAILQ_ENTRY(aioproclist) list;		/* List of processes */
178 	struct proc *aioproc;			/* The AIO thread */
179 	TAILQ_HEAD (,aiocblist) jobtorun;	/* suggested job to run */
180 };
181 
182 /*
183  * data-structure for lio signal management
184  */
185 struct aio_liojob {
186 	int lioj_flags;
187 	int	lioj_buffer_count;
188 	int	lioj_buffer_finished_count;
189 	int	lioj_queue_count;
190 	int	lioj_queue_finished_count;
191 	struct sigevent lioj_signal;	/* signal on all I/O done */
192 	TAILQ_ENTRY (aio_liojob) lioj_list;
193 	struct kaioinfo *lioj_ki;
194 };
195 #define	LIOJ_SIGNAL			0x1 /* signal on all done (lio) */
196 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
197 
198 /*
199  * per process aio data structure
200  */
201 struct kaioinfo {
202 	int	kaio_flags;			/* per process kaio flags */
203 	int	kaio_maxactive_count;	/* maximum number of AIOs */
204 	int	kaio_active_count;	/* number of currently used AIOs */
205 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
206 	int	kaio_queue_count;	/* size of AIO queue */
207 	int	kaio_ballowed_count;	/* maximum number of buffers */
208 	int	kaio_queue_finished_count;	/* number of daemon jobs finished */
209 	int	kaio_buffer_count;	/* number of physio buffers */
210 	int	kaio_buffer_finished_count;	/* count of I/O done */
211 	struct proc *kaio_p;			/* process that uses this kaio block */
212 	TAILQ_HEAD (,aio_liojob) kaio_liojoblist;	/* list of lio jobs */
213 	TAILQ_HEAD (,aiocblist)	kaio_jobqueue;	/* job queue for process */
214 	TAILQ_HEAD (,aiocblist)	kaio_jobdone;	/* done queue for process */
215 	TAILQ_HEAD (,aiocblist)	kaio_bufqueue;	/* buffer job queue for process */
216 	TAILQ_HEAD (,aiocblist)	kaio_bufdone;	/* buffer done queue for process */
217 };
218 
219 #define KAIO_RUNDOWN 0x1		/* process is being run down */
220 #define KAIO_WAKEUP 0x2			/* wakeup process when there is a significant
221 								   event */
222 
223 
224 static TAILQ_HEAD (,aioproclist) aio_freeproc, aio_activeproc;
225 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
226 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
227 static TAILQ_HEAD(,aiocblist) aio_freejobs;		/* Pool of free jobs */
228 
229 static void aio_init_aioinfo(struct proc *p) ;
230 static void aio_onceonly(void *) ;
231 static int aio_free_entry(struct aiocblist *aiocbe);
232 static void aio_process(struct aiocblist *aiocbe);
233 static int aio_newproc(void) ;
234 static int aio_aqueue(struct proc *p, struct aiocb *job, int type) ;
235 static void aio_physwakeup(struct buf *bp);
236 static int aio_fphysio(struct proc *p, struct aiocblist *aiocbe, int type);
237 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
238 static void aio_daemon(void *uproc);
239 
240 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
241 
242 static vm_zone_t kaio_zone=0, aiop_zone=0,
243 	aiocb_zone=0, aiol_zone=0, aiolio_zone=0;
244 
245 /*
246  * Single AIOD vmspace shared amongst all of them
247  */
248 struct vmspace *aiovmspace = NULL;
249 
250 /*
251  * Startup initialization
252  */
253 void
254 aio_onceonly(void *na)
255 {
256 	TAILQ_INIT(&aio_freeproc);
257 	TAILQ_INIT(&aio_activeproc);
258 	TAILQ_INIT(&aio_jobs);
259 	TAILQ_INIT(&aio_bufjobs);
260 	TAILQ_INIT(&aio_freejobs);
261 	kaio_zone = zinit("AIO", sizeof (struct kaioinfo), 0, 0, 1);
262 	aiop_zone = zinit("AIOP", sizeof (struct aioproclist), 0, 0, 1);
263 	aiocb_zone = zinit("AIOCB", sizeof (struct aiocblist), 0, 0, 1);
264 	aiol_zone = zinit("AIOL", AIO_LISTIO_MAX * sizeof (int), 0, 0, 1);
265 	aiolio_zone = zinit("AIOLIO",
266 		AIO_LISTIO_MAX * sizeof (struct aio_liojob), 0, 0, 1);
267 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
268 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
269 	jobrefid = 1;
270 }
271 
272 /*
273  * Init the per-process aioinfo structure.
274  * The aioinfo limits are set per-process for user limit (resource) management.
275  */
276 void
277 aio_init_aioinfo(struct proc *p)
278 {
279 	struct kaioinfo *ki;
280 	if (p->p_aioinfo == NULL) {
281 		ki = zalloc(kaio_zone);
282 		p->p_aioinfo = ki;
283 		ki->kaio_flags = 0;
284 		ki->kaio_maxactive_count = max_aio_per_proc;
285 		ki->kaio_active_count = 0;
286 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
287 		ki->kaio_queue_count = 0;
288 		ki->kaio_ballowed_count = max_buf_aio;
289 		ki->kaio_buffer_count = 0;
290 		ki->kaio_buffer_finished_count = 0;
291 		ki->kaio_p = p;
292 		TAILQ_INIT(&ki->kaio_jobdone);
293 		TAILQ_INIT(&ki->kaio_jobqueue);
294 		TAILQ_INIT(&ki->kaio_bufdone);
295 		TAILQ_INIT(&ki->kaio_bufqueue);
296 		TAILQ_INIT(&ki->kaio_liojoblist);
297 	}
298 }
299 
300 /*
301  * Free a job entry.  Wait for completion if it is currently
302  * active, but don't delay forever.  If we delay, we return
303  * a flag that says that we have to restart the queue scan.
304  */
305 int
306 aio_free_entry(struct aiocblist *aiocbe)
307 {
308 	struct kaioinfo *ki;
309 	struct aioproclist *aiop;
310 	struct aio_liojob *lj;
311 	struct proc *p;
312 	int error;
313 	int s;
314 
315 	if (aiocbe->jobstate == JOBST_NULL)
316 		panic("aio_free_entry: freeing already free job");
317 
318 	p = aiocbe->userproc;
319 	ki = p->p_aioinfo;
320 	lj = aiocbe->lio;
321 	if (ki == NULL)
322 		panic("aio_free_entry: missing p->p_aioinfo");
323 
324 	if (aiocbe->jobstate == JOBST_JOBRUNNING) {
325 		if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE)
326 			return 0;
327 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
328 		tsleep(aiocbe, PRIBIO|PCATCH, "jobwai", 0);
329 	}
330 	aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
331 
332 	if (aiocbe->bp == NULL) {
333 		if (ki->kaio_queue_count <= 0)
334 			panic("aio_free_entry: process queue size <= 0");
335 		if (num_queue_count <= 0)
336 			panic("aio_free_entry: system wide queue size <= 0");
337 
338 		if(lj) {
339 			lj->lioj_queue_count--;
340 			if (aiocbe->jobflags & AIOCBLIST_DONE)
341 				lj->lioj_queue_finished_count--;
342 		}
343 		ki->kaio_queue_count--;
344 		if (aiocbe->jobflags & AIOCBLIST_DONE)
345 			ki->kaio_queue_finished_count--;
346 		num_queue_count--;
347 
348 	} else {
349 		if(lj) {
350 			lj->lioj_buffer_count--;
351 			if (aiocbe->jobflags & AIOCBLIST_DONE)
352 				lj->lioj_buffer_finished_count--;
353 		}
354 		if (aiocbe->jobflags & AIOCBLIST_DONE)
355 			ki->kaio_buffer_finished_count--;
356 		ki->kaio_buffer_count--;
357 		num_buf_aio--;
358 
359 	}
360 
361 	if ((ki->kaio_flags & KAIO_WAKEUP) ||
362 		(ki->kaio_flags & KAIO_RUNDOWN) &&
363 		((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0))) {
364 		ki->kaio_flags &= ~KAIO_WAKEUP;
365 		wakeup(p);
366 	}
367 
368 	if ( aiocbe->jobstate == JOBST_JOBQBUF) {
369 		if ((error = aio_fphysio(p, aiocbe, 1)) != 0)
370 			return error;
371 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
372 			panic("aio_free_entry: invalid physio finish-up state");
373 		s = splbio();
374 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
375 		splx(s);
376 	} else if ( aiocbe->jobstate == JOBST_JOBQPROC) {
377 		aiop = aiocbe->jobaioproc;
378 		TAILQ_REMOVE(&aiop->jobtorun, aiocbe, list);
379 	} else if ( aiocbe->jobstate == JOBST_JOBQGLOBAL) {
380 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
381 	} else if ( aiocbe->jobstate == JOBST_JOBFINISHED) {
382 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
383 	} else if ( aiocbe->jobstate == JOBST_JOBBFINISHED) {
384 		s = splbio();
385 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
386 		splx(s);
387 		if (aiocbe->bp) {
388 			vunmapbuf(aiocbe->bp);
389 			relpbuf(aiocbe->bp);
390 			aiocbe->bp = NULL;
391 		}
392 	}
393 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
394 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
395 		zfree(aiolio_zone, lj);
396 	}
397 	TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
398 	aiocbe->jobstate = JOBST_NULL;
399 	return 0;
400 }
401 
402 /*
403  * Rundown the jobs for a given process.
404  */
405 void
406 aio_proc_rundown(struct proc *p)
407 {
408 	int s;
409 	struct kaioinfo *ki;
410 	struct aio_liojob *lj, *ljn;
411 	struct aiocblist *aiocbe, *aiocbn;
412 
413 	ki = p->p_aioinfo;
414 	if (ki == NULL)
415 		return;
416 
417 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
418 	while ((ki->kaio_active_count > 0) ||
419 		(ki->kaio_buffer_count > ki->kaio_buffer_finished_count)) {
420 		ki->kaio_flags |= KAIO_RUNDOWN;
421 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
422 			break;
423 	}
424 
425 restart1:
426 	for ( aiocbe = TAILQ_FIRST(&ki->kaio_jobdone);
427 		aiocbe;
428 		aiocbe = aiocbn) {
429 		aiocbn = TAILQ_NEXT(aiocbe, plist);
430 		if (aio_free_entry(aiocbe))
431 			goto restart1;
432 	}
433 
434 restart2:
435 	for ( aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue);
436 		aiocbe;
437 		aiocbe = aiocbn) {
438 		aiocbn = TAILQ_NEXT(aiocbe, plist);
439 		if (aio_free_entry(aiocbe))
440 			goto restart2;
441 	}
442 
443 /*
444  * Note the use of lots of splbio here, trying to avoid
445  * splbio for long chains of I/O.  Probably unnecessary.
446  */
447 
448 restart3:
449 	s = splbio();
450 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
451 		ki->kaio_flags |= KAIO_WAKEUP;
452 		tsleep (p, PRIBIO, "aioprn", 0);
453 		splx(s);
454 		goto restart3;
455 	}
456 	splx(s);
457 
458 restart4:
459 	s = splbio();
460 	for ( aiocbe = TAILQ_FIRST(&ki->kaio_bufdone);
461 		aiocbe;
462 		aiocbe = aiocbn) {
463 		aiocbn = TAILQ_NEXT(aiocbe, plist);
464 		if (aio_free_entry(aiocbe)) {
465 			splx(s);
466 			goto restart4;
467 		}
468 	}
469 	splx(s);
470 
471 	for ( lj = TAILQ_FIRST(&ki->kaio_liojoblist);
472 		  lj;
473 		  lj = ljn) {
474 			ljn = TAILQ_NEXT(lj, lioj_list);
475 			if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
476 				TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
477 				zfree(aiolio_zone, lj);
478 			} else {
479 #if defined(DIAGNOSTIC)
480 				printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, QF:%d\n",
481 					lj->lioj_buffer_count, lj->lioj_buffer_finished_count,
482 					lj->lioj_queue_count, lj->lioj_queue_finished_count);
483 #endif
484 			}
485 	}
486 
487 	zfree(kaio_zone, ki);
488 	p->p_aioinfo = NULL;
489 }
490 
491 /*
492  * Select a job to run (called by an AIO daemon)
493  */
494 static struct aiocblist *
495 aio_selectjob(struct aioproclist *aiop)
496 {
497 
498 	struct aiocblist *aiocbe;
499 
500 	aiocbe = TAILQ_FIRST(&aiop->jobtorun);
501 	if (aiocbe) {
502 		TAILQ_REMOVE(&aiop->jobtorun, aiocbe, list);
503 		return aiocbe;
504 	}
505 
506 	for (aiocbe = TAILQ_FIRST(&aio_jobs);
507 		aiocbe;
508 		aiocbe = TAILQ_NEXT(aiocbe, list)) {
509 		struct kaioinfo *ki;
510 		struct proc *userp;
511 
512 		userp = aiocbe->userproc;
513 		ki = userp->p_aioinfo;
514 
515 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
516 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
517 			return aiocbe;
518 		}
519 	}
520 
521 	return NULL;
522 }
523 
524 /*
525  * The AIO processing activity.  This is the code that does the
526  * I/O request for the non-physio version of the operations.  The
527  * normal vn operations are used, and this code should work in
528  * all instances for every type of file, including pipes, sockets,
529  * fifos, and regular files.
530  */
531 void
532 aio_process(struct aiocblist *aiocbe)
533 {
534 	struct filedesc *fdp;
535 	struct proc *userp, *mycp;
536 	struct aiocb *cb;
537 	struct file *fp;
538 	struct uio auio;
539 	struct iovec aiov;
540 	unsigned int fd;
541 	int cnt;
542 	int error;
543 	off_t offset;
544 	int oublock_st, oublock_end;
545 	int inblock_st, inblock_end;
546 
547 	userp = aiocbe->userproc;
548 	cb = &aiocbe->uaiocb;
549 
550 	mycp = curproc;
551 
552 	fdp = mycp->p_fd;
553 	fd = cb->aio_fildes;
554 	fp = fdp->fd_ofiles[fd];
555 
556 	aiov.iov_base = (void *) cb->aio_buf;
557 	aiov.iov_len = cb->aio_nbytes;
558 
559 	auio.uio_iov = &aiov;
560 	auio.uio_iovcnt = 1;
561 	auio.uio_offset = offset = cb->aio_offset;
562 	auio.uio_resid = cb->aio_nbytes;
563 	cnt = cb->aio_nbytes;
564 	auio.uio_segflg = UIO_USERSPACE;
565 	auio.uio_procp = mycp;
566 
567 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
568 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
569 	if (cb->aio_lio_opcode == LIO_READ) {
570 		auio.uio_rw = UIO_READ;
571 		error = (*fp->f_ops->fo_read)(fp, &auio, fp->f_cred);
572 	} else {
573 		auio.uio_rw = UIO_WRITE;
574 		error = (*fp->f_ops->fo_write)(fp, &auio, fp->f_cred);
575 	}
576 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
577 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
578 
579 	aiocbe->inputcharge = inblock_end - inblock_st;
580 	aiocbe->outputcharge = oublock_end - oublock_st;
581 
582 	if (error) {
583 		if (auio.uio_resid != cnt) {
584 			if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
585 				error = 0;
586 			if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
587 				psignal(userp, SIGPIPE);
588 		}
589 	}
590 
591 	cnt -= auio.uio_resid;
592 	cb->_aiocb_private.error = error;
593 	cb->_aiocb_private.status = cnt;
594 
595 	return;
596 
597 }
598 
599 /*
600  * The AIO daemon, most of the actual work is done in aio_process,
601  * but the setup (and address space mgmt) is done in this routine.
602  */
603 static void
604 aio_daemon(void *uproc)
605 {
606 	int s;
607 	struct aioproclist *aiop;
608 	struct vmspace *myvm, *aiovm;
609 	struct proc *mycp;
610 
611 	/*
612 	 * Local copies of curproc (cp) and vmspace (myvm)
613 	 */
614 	mycp = curproc;
615 	myvm = mycp->p_vmspace;
616 
617 	/*
618 	 * We manage to create only one VM space for all AIOD processes.
619 	 * The VM space for the first AIOD created becomes the shared VM
620 	 * space for all of them.  We add an additional reference count,
621 	 * even for the first AIOD, so the address space does not go away,
622 	 * and we continue to use that original VM space even if the first
623 	 * AIOD exits.
624 	 */
625 	if ((aiovm = aiovmspace) == NULL) {
626 		aiovmspace = myvm;
627 		myvm->vm_refcnt++;
628 		/*
629 		 * Remove userland cruft from address space.
630 		 */
631 		if (myvm->vm_shm)
632 			shmexit(mycp);
633 		pmap_remove_pages(&myvm->vm_pmap, 0, USRSTACK);
634 		vm_map_remove(&myvm->vm_map, 0, USRSTACK);
635 		myvm->vm_tsize = 0;
636 		myvm->vm_dsize = 0;
637 		myvm->vm_ssize = 0;
638 	} else {
639 		aiovm->vm_refcnt++;
640 		mycp->p_vmspace = aiovm;
641 		pmap_activate(mycp);
642 		vmspace_free(myvm);
643 		myvm = aiovm;
644 	}
645 
646 	if (mycp->p_textvp) {
647 		vrele(mycp->p_textvp);
648 		mycp->p_textvp = NULL;
649 	}
650 
651 	/*
652 	 * Allocate and ready the aio control info.  There is one
653 	 * aiop structure per daemon.
654 	 */
655 	aiop = zalloc(aiop_zone);
656 	aiop->aioproc = mycp;
657 	aiop->aioprocflags |= AIOP_FREE;
658 	TAILQ_INIT(&aiop->jobtorun);
659 
660 	/*
661 	 * Place thread (lightweight process) onto the AIO free thread list
662 	 */
663 	if (TAILQ_EMPTY(&aio_freeproc))
664 		wakeup(&aio_freeproc);
665 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
666 
667 	/*
668 	 * Make up a name for the daemon
669 	 */
670 	strcpy(mycp->p_comm, "aiod");
671 
672 	/*
673 	 * Get rid of our current filedescriptors.  AIOD's don't need any
674 	 * filedescriptors, except as temporarily inherited from the client.
675 	 * Credentials are also cloned, and made equivalent to "root."
676 	 */
677 	fdfree(mycp);
678 	mycp->p_fd = NULL;
679 	mycp->p_ucred = crcopy(mycp->p_ucred);
680 	mycp->p_ucred->cr_uid = 0;
681 	mycp->p_ucred->cr_ngroups = 1;
682 	mycp->p_ucred->cr_groups[0] = 1;
683 
684 	/*
685 	 * The daemon resides in its own pgrp.
686 	 */
687 	enterpgrp(mycp, mycp->p_pid, 1);
688 
689 	/*
690 	 * Mark special process type
691 	 */
692 	mycp->p_flag |= P_SYSTEM|P_KTHREADP;
693 
694 	/*
695 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
696 	 * creating to many daemons.)
697 	 */
698 	wakeup(mycp);
699 
700 	while(1) {
701 		struct proc *curcp;
702 		struct	aiocblist *aiocbe;
703 
704 		/*
705 		 * curcp is the current daemon process context.
706 		 * userp is the current user process context.
707 		 */
708 		curcp = mycp;
709 
710 		/*
711 		 * Take daemon off of free queue
712 		 */
713 		if (aiop->aioprocflags & AIOP_FREE) {
714 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
715 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
716 			aiop->aioprocflags &= ~AIOP_FREE;
717 		}
718 		aiop->aioprocflags &= ~AIOP_SCHED;
719 
720 		/*
721 		 * Check for jobs
722 		 */
723 		while ( aiocbe = aio_selectjob(aiop)) {
724 			struct proc *userp;
725 			struct aiocb *cb;
726 			struct kaioinfo *ki;
727 			struct aio_liojob *lj;
728 
729 			cb = &aiocbe->uaiocb;
730 			userp = aiocbe->userproc;
731 
732 			aiocbe->jobstate = JOBST_JOBRUNNING;
733 
734 			/*
735 			 * Connect to process address space for user program
736 			 */
737 			if (userp != curcp) {
738 				struct vmspace *tmpvm;
739 				/*
740 				 * Save the current address space that we are connected to.
741 				 */
742 				tmpvm = mycp->p_vmspace;
743 				/*
744 				 * Point to the new user address space, and refer to it.
745 				 */
746 				mycp->p_vmspace = userp->p_vmspace;
747 				mycp->p_vmspace->vm_refcnt++;
748 				/*
749 				 * Activate the new mapping.
750 				 */
751 				pmap_activate(mycp);
752 				/*
753 				 * If the old address space wasn't the daemons own address
754 				 * space, then we need to remove the daemon's reference from
755 				 * the other process that it was acting on behalf of.
756 				 */
757 				if (tmpvm != myvm) {
758 					vmspace_free(tmpvm);
759 				}
760 				/*
761 				 * Disassociate from previous clients file descriptors, and
762 				 * associate to the new clients descriptors.  Note that
763 				 * the daemon doesn't need to worry about its orginal
764 				 * descriptors, because they were originally freed.
765 				 */
766 				if (mycp->p_fd)
767 					fdfree(mycp);
768 				mycp->p_fd = fdshare(userp);
769 				curcp = userp;
770 			}
771 
772 			ki = userp->p_aioinfo;
773 			lj = aiocbe->lio;
774 
775 			/*
776 			 * Account for currently active jobs
777 			 */
778 			ki->kaio_active_count++;
779 
780 			/*
781 			 * Do the I/O function
782 			 */
783 			aiocbe->jobaioproc = aiop;
784 			aio_process(aiocbe);
785 
786 			/*
787 			 * decrement the active job count
788 			 */
789 			ki->kaio_active_count--;
790 
791 			/*
792 			 * increment the completion count for wakeup/signal comparisons
793 			 */
794 			aiocbe->jobflags |= AIOCBLIST_DONE;
795 			ki->kaio_queue_finished_count++;
796 			if (lj) {
797 				lj->lioj_queue_finished_count++;
798 			}
799 			if ((ki->kaio_flags & KAIO_WAKEUP) ||
800 				(ki->kaio_flags & KAIO_RUNDOWN) &&
801 				(ki->kaio_active_count == 0)) {
802 				ki->kaio_flags &= ~KAIO_WAKEUP;
803 				wakeup(userp);
804 			}
805 
806 			s = splbio();
807 			if (lj && (lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
808 				LIOJ_SIGNAL) {
809 				if ((lj->lioj_queue_finished_count == lj->lioj_queue_count) &&
810 					(lj->lioj_buffer_finished_count == lj->lioj_buffer_count)) {
811 						psignal(userp, lj->lioj_signal.sigev_signo);
812 						lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
813 				}
814 			}
815 			splx(s);
816 
817 			aiocbe->jobstate = JOBST_JOBFINISHED;
818 
819 			/*
820 			 * If the I/O request should be automatically rundown, do the
821 			 * needed cleanup.  Otherwise, place the queue entry for
822 			 * the just finished I/O request into the done queue for the
823 			 * associated client.
824 			 */
825 			if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE) {
826 				aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
827 				TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
828 			} else {
829 				TAILQ_REMOVE(&ki->kaio_jobqueue,
830 					aiocbe, plist);
831 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone,
832 					aiocbe, plist);
833 			}
834 
835 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
836 				wakeup(aiocbe);
837 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
838 			}
839 
840 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
841 				psignal(userp, cb->aio_sigevent.sigev_signo);
842 			}
843 		}
844 
845 		/*
846 		 * Disconnect from user address space
847 		 */
848 		if (curcp != mycp) {
849 			struct vmspace *tmpvm;
850 			/*
851 			 * Get the user address space to disconnect from.
852 			 */
853 			tmpvm = mycp->p_vmspace;
854 			/*
855 			 * Get original address space for daemon.
856 			 */
857 			mycp->p_vmspace = myvm;
858 			/*
859 			 * Activate the daemon's address space.
860 			 */
861 			pmap_activate(mycp);
862 #if defined(DIAGNOSTIC)
863 			if (tmpvm == myvm)
864 				printf("AIOD: vmspace problem -- %d\n", mycp->p_pid);
865 #endif
866 			/*
867 			 * remove our vmspace reference.
868 			 */
869 			vmspace_free(tmpvm);
870 			/*
871 			 * disassociate from the user process's file descriptors.
872 			 */
873 			if (mycp->p_fd)
874 				fdfree(mycp);
875 			mycp->p_fd = NULL;
876 			curcp = mycp;
877 		}
878 
879 		/*
880 		 * If we are the first to be put onto the free queue, wakeup
881 		 * anyone waiting for a daemon.
882 		 */
883 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
884 		if (TAILQ_EMPTY(&aio_freeproc))
885 			wakeup(&aio_freeproc);
886 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
887 		aiop->aioprocflags |= AIOP_FREE;
888 
889 		/*
890 		 * If daemon is inactive for a long time, allow it to exit, thereby
891 		 * freeing resources.
892 		 */
893 		if (((aiop->aioprocflags & AIOP_SCHED) == 0) &&
894 			tsleep(mycp, PRIBIO, "aiordy", aiod_lifetime)) {
895 			if ((TAILQ_FIRST(&aio_jobs) == NULL) &&
896 				(TAILQ_FIRST(&aiop->jobtorun) == NULL)) {
897 				if ((aiop->aioprocflags & AIOP_FREE) &&
898 					(num_aio_procs > target_aio_procs)) {
899 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
900 					zfree(aiop_zone, aiop);
901 					num_aio_procs--;
902 #if defined(DIAGNOSTIC)
903 					if (mycp->p_vmspace->vm_refcnt <= 1)
904 						printf("AIOD: bad vm refcnt for exiting daemon: %d\n",
905 							mycp->p_vmspace->vm_refcnt);
906 #endif
907 					exit1(mycp, 0);
908 				}
909 			}
910 		}
911 	}
912 }
913 
914 /*
915  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.
916  * The AIO daemon modifies its environment itself.
917  */
918 static int
919 aio_newproc()
920 {
921 	int error;
922 	struct rfork_args rfa;
923 	struct proc *p, *np;
924 
925 	rfa.flags = RFPROC | RFCFDG;
926 
927 	p = curproc;
928 	if (error = rfork(p, &rfa))
929 		return error;
930 
931 	np = pfind(p->p_retval[0]);
932 	cpu_set_fork_handler(np, aio_daemon, p);
933 
934 	/*
935 	 * Wait until daemon is started, but continue on just in case (to
936 	 * handle error conditions.
937 	 */
938 	error = tsleep(np, PZERO, "aiosta", aiod_timeout);
939 	num_aio_procs++;
940 
941 	return error;
942 
943 }
944 
945 /*
946  * Try the high-performance physio method for eligible VCHR devices.  This
947  * routine doesn't require the use of any additional threads, and have
948  * overhead.
949  */
950 int
951 aio_qphysio(p, aiocbe)
952 	struct proc *p;
953 	struct aiocblist *aiocbe;
954 {
955 	int error;
956 	struct aiocb *cb;
957 	struct file *fp;
958 	struct buf *bp;
959 	int bflags;
960 	struct vnode *vp;
961 	struct kaioinfo *ki;
962 	struct filedesc *fdp;
963 	struct aio_liojob *lj;
964 	int fd;
965 	int majordev;
966 	int s;
967 	int cnt;
968 	dev_t dev;
969 	int rw;
970 	d_strategy_t *fstrategy;
971 	struct cdevsw *cdev;
972 	struct cdevsw *bdev;
973 
974 	cb = &aiocbe->uaiocb;
975 	fdp = p->p_fd;
976 	fd = cb->aio_fildes;
977 	fp = fdp->fd_ofiles[fd];
978 
979 	if (fp->f_type != DTYPE_VNODE) {
980 		return -1;
981 	}
982 
983 	vp = (struct vnode *)fp->f_data;
984 	if (vp->v_type != VCHR || ((cb->aio_nbytes & (DEV_BSIZE - 1)) != 0)) {
985 		return -1;
986 	}
987 
988 	if ((cb->aio_nbytes > MAXPHYS) && (num_buf_aio >= max_buf_aio)) {
989 		return -1;
990 	}
991 
992 	if ((vp->v_specinfo == NULL) || (vp->v_flag & VISTTY)) {
993 		return -1;
994 	}
995 
996 	majordev = major(vp->v_rdev);
997 	if (majordev == NODEV) {
998 		return -1;
999 	}
1000 
1001 	cdev = cdevsw[major(vp->v_rdev)];
1002 	if (cdev == NULL) {
1003 		return -1;
1004 	}
1005 
1006 	if (cdev->d_bmaj == -1) {
1007 		return -1;
1008 	}
1009 	bdev = cdev;
1010 
1011 	ki = p->p_aioinfo;
1012 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1013 		return -1;
1014 	}
1015 
1016 	cnt = cb->aio_nbytes;
1017 	if (cnt > MAXPHYS) {
1018 		return -1;
1019 	}
1020 
1021 	dev = makedev(bdev->d_bmaj, minor(vp->v_rdev));
1022 
1023 	/*
1024 	 * Physical I/O is charged directly to the process, so we don't have
1025 	 * to fake it.
1026 	 */
1027 	aiocbe->inputcharge = 0;
1028 	aiocbe->outputcharge = 0;
1029 
1030 	ki->kaio_buffer_count++;
1031 
1032 	lj = aiocbe->lio;
1033 	if (lj) {
1034 		lj->lioj_buffer_count++;
1035 	}
1036 
1037 	/* create and build a buffer header for a transfer */
1038 	bp = (struct buf *)getpbuf();
1039 
1040 	/*
1041 	 * get a copy of the kva from the physical buffer
1042 	 */
1043 	bp->b_proc = p;
1044 	bp->b_dev = dev;
1045 	error = bp->b_error = 0;
1046 
1047 	if (cb->aio_lio_opcode == LIO_WRITE) {
1048 		rw = 0;
1049 		bflags = B_WRITE;
1050 	} else {
1051 		rw = 1;
1052 		bflags = B_READ;
1053 	}
1054 
1055 	bp->b_bcount = cb->aio_nbytes;
1056 	bp->b_bufsize = cb->aio_nbytes;
1057 	bp->b_flags = B_BUSY | B_PHYS | B_CALL | bflags;
1058 	bp->b_iodone = aio_physwakeup;
1059 	bp->b_saveaddr = bp->b_data;
1060 	bp->b_data = (void *) cb->aio_buf;
1061 	bp->b_blkno = btodb(cb->aio_offset);
1062 
1063 	if (rw && !useracc(bp->b_data, bp->b_bufsize, B_WRITE)) {
1064 		error = EFAULT;
1065 		goto doerror;
1066 	}
1067 	if (!rw && !useracc(bp->b_data, bp->b_bufsize, B_READ)) {
1068 		error = EFAULT;
1069 		goto doerror;
1070 	}
1071 
1072 	/* bring buffer into kernel space */
1073 	vmapbuf(bp);
1074 
1075 	s = splbio();
1076 	aiocbe->bp = bp;
1077 	bp->b_spc = (void *)aiocbe;
1078 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1079 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1080 	aiocbe->jobstate = JOBST_JOBQBUF;
1081 	cb->_aiocb_private.status = cb->aio_nbytes;
1082 	num_buf_aio++;
1083 	fstrategy = bdev->d_strategy;
1084 	bp->b_error = 0;
1085 
1086 	splx(s);
1087 	/* perform transfer */
1088 	(*fstrategy)(bp);
1089 
1090 	s = splbio();
1091 	/*
1092 	 * If we had an error invoking the request, or an error in processing
1093 	 * the request before we have returned, we process it as an error
1094 	 * in transfer.  Note that such an I/O error is not indicated immediately,
1095 	 * but is returned using the aio_error mechanism.  In this case, aio_suspend
1096 	 * will return immediately.
1097 	 */
1098 	if (bp->b_error || (bp->b_flags & B_ERROR)) {
1099 		struct aiocb *job = aiocbe->uuaiocb;
1100 
1101 		aiocbe->uaiocb._aiocb_private.status = 0;
1102 		suword(&job->_aiocb_private.status, 0);
1103 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1104 		suword(&job->_aiocb_private.error, bp->b_error);
1105 
1106 		ki->kaio_buffer_finished_count++;
1107 
1108 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1109 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1110 			aiocbe->jobflags |= AIOCBLIST_DONE;
1111 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1112 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1113 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1114 		}
1115 	}
1116 	splx(s);
1117 	return 0;
1118 
1119 doerror:
1120 	ki->kaio_buffer_count--;
1121 	if (lj) {
1122 		lj->lioj_buffer_count--;
1123 	}
1124 	aiocbe->bp = NULL;
1125 	relpbuf(bp);
1126 	return error;
1127 }
1128 
1129 /*
1130  * This waits/tests physio completion.
1131  */
1132 int
1133 aio_fphysio(p, iocb, flgwait)
1134 	struct proc *p;
1135 	struct aiocblist *iocb;
1136 	int flgwait;
1137 {
1138 	int s;
1139 	struct buf *bp;
1140 	int error;
1141 
1142 	bp = iocb->bp;
1143 
1144 	s = splbio();
1145 	if (flgwait == 0) {
1146 		if ((bp->b_flags & B_DONE) == 0) {
1147 			splx(s);
1148 			return EINPROGRESS;
1149 		}
1150 	}
1151 
1152 	while ((bp->b_flags & B_DONE) == 0) {
1153 		if (tsleep((caddr_t)bp, PRIBIO, "physstr", aiod_timeout)) {
1154 			if ((bp->b_flags & B_DONE) == 0) {
1155 				splx(s);
1156 				return EINPROGRESS;
1157 			} else {
1158 				break;
1159 			}
1160 		}
1161 	}
1162 
1163 	/* release mapping into kernel space */
1164 	vunmapbuf(bp);
1165 	iocb->bp = 0;
1166 
1167 	error = 0;
1168 	/*
1169 	 * check for an error
1170 	 */
1171 	if (bp->b_flags & B_ERROR) {
1172 		error = bp->b_error;
1173 	}
1174 
1175 	relpbuf(bp);
1176 	return (error);
1177 }
1178 
1179 /*
1180  * Queue a new AIO request.  Choosing either the threaded or direct physio
1181  * VCHR technique is done in this code.
1182  */
1183 static int
1184 _aio_aqueue(struct proc *p, struct aiocb *job, struct aio_liojob *lj, int type)
1185 {
1186 	struct filedesc *fdp;
1187 	struct file *fp;
1188 	unsigned int fd;
1189 
1190 	int error;
1191 	int opcode;
1192 	struct aiocblist *aiocbe;
1193 	struct aioproclist *aiop;
1194 	struct kaioinfo *ki;
1195 
1196 	if (aiocbe = TAILQ_FIRST(&aio_freejobs)) {
1197 		TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1198 	} else {
1199 		aiocbe = zalloc (aiocb_zone);
1200 	}
1201 
1202 	aiocbe->inputcharge = 0;
1203 	aiocbe->outputcharge = 0;
1204 
1205 	suword(&job->_aiocb_private.status, -1);
1206 	suword(&job->_aiocb_private.error, 0);
1207 	suword(&job->_aiocb_private.kernelinfo, -1);
1208 
1209 	error = copyin((caddr_t)job,
1210 		(caddr_t) &aiocbe->uaiocb, sizeof aiocbe->uaiocb);
1211 	if (error) {
1212 		suword(&job->_aiocb_private.error, error);
1213 
1214 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1215 		return error;
1216 	}
1217 
1218 	/*
1219 	 * Save userspace address of the job info
1220 	 */
1221 	aiocbe->uuaiocb = job;
1222 
1223 	/*
1224 	 * Get the opcode
1225 	 */
1226 	if (type != LIO_NOP) {
1227 		aiocbe->uaiocb.aio_lio_opcode = type;
1228 	}
1229 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1230 
1231 	/*
1232 	 * Get the fd info for process
1233 	 */
1234 	fdp = p->p_fd;
1235 
1236 	/*
1237 	 * Range check file descriptor
1238 	 */
1239 	fd = aiocbe->uaiocb.aio_fildes;
1240 	if (fd >= fdp->fd_nfiles) {
1241 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1242 		if (type == 0) {
1243 			suword(&job->_aiocb_private.error, EBADF);
1244 		}
1245 		return EBADF;
1246 	}
1247 
1248 	fp = fdp->fd_ofiles[fd];
1249 	if ((fp == NULL) ||
1250 		((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 0))) {
1251 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1252 		if (type == 0) {
1253 			suword(&job->_aiocb_private.error, EBADF);
1254 		}
1255 		return EBADF;
1256 	}
1257 
1258 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1259 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1260 		if (type == 0) {
1261 			suword(&job->_aiocb_private.error, EINVAL);
1262 		}
1263 		return EINVAL;
1264 	}
1265 
1266 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1267 	if (error) {
1268 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1269 		if (type == 0) {
1270 			suword(&job->_aiocb_private.error, EINVAL);
1271 		}
1272 		return error;
1273 	}
1274 
1275 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1276 	if (jobrefid == LONG_MAX)
1277 		jobrefid = 1;
1278 	else
1279 		jobrefid++;
1280 
1281 	if (opcode == LIO_NOP) {
1282 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1283 		if (type == 0) {
1284 			suword(&job->_aiocb_private.error, 0);
1285 			suword(&job->_aiocb_private.status, 0);
1286 			suword(&job->_aiocb_private.kernelinfo, 0);
1287 		}
1288 		return 0;
1289 	}
1290 
1291 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1292 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1293 		if (type == 0) {
1294 			suword(&job->_aiocb_private.status, 0);
1295 			suword(&job->_aiocb_private.error, EINVAL);
1296 		}
1297 		return EINVAL;
1298 	}
1299 
1300 	suword(&job->_aiocb_private.error, EINPROGRESS);
1301 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1302 	aiocbe->userproc = p;
1303 	aiocbe->jobflags = 0;
1304 	aiocbe->lio = lj;
1305 	ki = p->p_aioinfo;
1306 
1307 	if ((error = aio_qphysio(p, aiocbe)) == 0) {
1308 		return 0;
1309 	} else if (error > 0) {
1310 		suword(&job->_aiocb_private.status, 0);
1311 		aiocbe->uaiocb._aiocb_private.error = error;
1312 		suword(&job->_aiocb_private.error, error);
1313 		return error;
1314 	}
1315 
1316 	/*
1317 	 * No buffer for daemon I/O
1318 	 */
1319 	aiocbe->bp = NULL;
1320 
1321 	ki->kaio_queue_count++;
1322 	if (lj) {
1323 		lj->lioj_queue_count++;
1324 	}
1325 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1326 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1327 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1328 
1329 	num_queue_count++;
1330 	error = 0;
1331 
1332 	/*
1333 	 * If we don't have a free AIO process, and we are below our
1334 	 * quota, then start one.  Otherwise, depend on the subsequent
1335 	 * I/O completions to pick-up this job.  If we don't sucessfully
1336 	 * create the new process (thread) due to resource issues, we
1337 	 * return an error for now (EAGAIN), which is likely not the
1338 	 * correct thing to do.
1339 	 */
1340 retryproc:
1341 	if (aiop = TAILQ_FIRST(&aio_freeproc)) {
1342 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1343 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1344 		aiop->aioprocflags &= ~AIOP_FREE;
1345 		wakeup(aiop->aioproc);
1346 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1347 			((ki->kaio_active_count + num_aio_resv_start) <
1348 				ki->kaio_maxactive_count)) {
1349 		num_aio_resv_start++;
1350 		if ((error = aio_newproc()) == 0) {
1351 			num_aio_resv_start--;
1352 			p->p_retval[0] = 0;
1353 			goto retryproc;
1354 		}
1355 		num_aio_resv_start--;
1356 	}
1357 	return error;
1358 }
1359 
1360 /*
1361  * This routine queues an AIO request, checking for quotas.
1362  */
1363 static int
1364 aio_aqueue(struct proc *p, struct aiocb *job, int type)
1365 {
1366 	struct kaioinfo *ki;
1367 
1368 	if (p->p_aioinfo == NULL) {
1369 		aio_init_aioinfo(p);
1370 	}
1371 
1372 	if (num_queue_count >= max_queue_count)
1373 		return EAGAIN;
1374 
1375 	ki = p->p_aioinfo;
1376 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1377 		return EAGAIN;
1378 
1379 	return _aio_aqueue(p, job, NULL, type);
1380 }
1381 
1382 /*
1383  * Support the aio_return system call, as a side-effect, kernel
1384  * resources are released.
1385  */
1386 int
1387 aio_return(struct proc *p, struct aio_return_args *uap)
1388 {
1389 	int s;
1390 	int jobref;
1391 	struct aiocblist *cb, *ncb;
1392 	struct aiocb *ujob;
1393 	struct kaioinfo *ki;
1394 
1395 	ki = p->p_aioinfo;
1396 	if (ki == NULL) {
1397 		return EINVAL;
1398 	}
1399 
1400 	ujob = uap->aiocbp;
1401 
1402 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1403 	if (jobref == -1 || jobref == 0)
1404 		return EINVAL;
1405 
1406 	for (cb = TAILQ_FIRST(&ki->kaio_jobdone);
1407 		cb;
1408 		cb = TAILQ_NEXT(cb, plist)) {
1409 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == jobref) {
1410 			if (ujob == cb->uuaiocb) {
1411 				p->p_retval[0] = cb->uaiocb._aiocb_private.status;
1412 			} else {
1413 				p->p_retval[0] = EFAULT;
1414 			}
1415 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1416 				curproc->p_stats->p_ru.ru_oublock += cb->outputcharge;
1417 				cb->outputcharge = 0;
1418 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1419 				curproc->p_stats->p_ru.ru_inblock += cb->inputcharge;
1420 				cb->inputcharge = 0;
1421 			}
1422 			aio_free_entry(cb);
1423 			return 0;
1424 		}
1425 	}
1426 
1427 	s = splbio();
1428 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone);
1429 		cb;
1430 		cb = ncb) {
1431 		ncb = TAILQ_NEXT(cb, plist);
1432 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == jobref) {
1433 			splx(s);
1434 			if (ujob == cb->uuaiocb) {
1435 				p->p_retval[0] = cb->uaiocb._aiocb_private.status;
1436 			} else {
1437 				p->p_retval[0] = EFAULT;
1438 			}
1439 			aio_free_entry(cb);
1440 			return 0;
1441 		}
1442 	}
1443 	splx(s);
1444 
1445 	return (EINVAL);
1446 }
1447 
1448 /*
1449  * Allow a process to wakeup when any of the I/O requests are
1450  * completed.
1451  */
1452 int
1453 aio_suspend(struct proc *p, struct aio_suspend_args *uap)
1454 {
1455 	struct timeval atv;
1456 	struct timespec ts;
1457 	struct aiocb *const *cbptr, *cbp;
1458 	struct kaioinfo *ki;
1459 	struct aiocblist *cb;
1460 	int i;
1461 	int njoblist;
1462 	int error, s, timo;
1463 	int *ijoblist;
1464 	struct aiocb **ujoblist;
1465 
1466 	if (uap->nent >= AIO_LISTIO_MAX)
1467 		return EINVAL;
1468 
1469 	timo = 0;
1470 	if (uap->timeout) {
1471 		/*
1472 		 * Get timespec struct
1473 		 */
1474 		if (error = copyin((caddr_t) uap->timeout, (caddr_t) &ts, sizeof ts)) {
1475 			return error;
1476 		}
1477 
1478 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1479 			return (EINVAL);
1480 
1481 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1482 		if (itimerfix(&atv))
1483 			return (EINVAL);
1484 		timo = tvtohz(&atv);
1485 	}
1486 
1487 	ki = p->p_aioinfo;
1488 	if (ki == NULL)
1489 		return EAGAIN;
1490 
1491 	njoblist = 0;
1492 	ijoblist = zalloc(aiol_zone);
1493 	ujoblist = zalloc(aiol_zone);
1494 	cbptr = uap->aiocbp;
1495 
1496 	for(i = 0; i < uap->nent; i++) {
1497 		cbp = (struct aiocb *) (intptr_t) fuword((caddr_t) &cbptr[i]);
1498 		if (cbp == 0)
1499 			continue;
1500 		ujoblist[njoblist] = cbp;
1501 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1502 		njoblist++;
1503 	}
1504 	if (njoblist == 0) {
1505 		zfree(aiol_zone, ijoblist);
1506 		zfree(aiol_zone, ujoblist);
1507 		return 0;
1508 	}
1509 
1510 	error = 0;
1511 	while (1) {
1512 		for (cb = TAILQ_FIRST(&ki->kaio_jobdone);
1513 			cb; cb = TAILQ_NEXT(cb, plist)) {
1514 			for(i = 0; i < njoblist; i++) {
1515 				if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1516 					ijoblist[i]) {
1517 					if (ujoblist[i] != cb->uuaiocb)
1518 						error = EINVAL;
1519 					zfree(aiol_zone, ijoblist);
1520 					zfree(aiol_zone, ujoblist);
1521 					return error;
1522 				}
1523 			}
1524 		}
1525 
1526 		s = splbio();
1527 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone);
1528 			cb; cb = TAILQ_NEXT(cb, plist)) {
1529 			for(i = 0; i < njoblist; i++) {
1530 				if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1531 					ijoblist[i]) {
1532 					splx(s);
1533 					if (ujoblist[i] != cb->uuaiocb)
1534 						error = EINVAL;
1535 					zfree(aiol_zone, ijoblist);
1536 					zfree(aiol_zone, ujoblist);
1537 					return error;
1538 				}
1539 			}
1540 		}
1541 
1542 		ki->kaio_flags |= KAIO_WAKEUP;
1543 		error = tsleep(p, PRIBIO|PCATCH, "aiospn", timo);
1544 		splx(s);
1545 
1546 		if (error == EINTR) {
1547 			zfree(aiol_zone, ijoblist);
1548 			zfree(aiol_zone, ujoblist);
1549 			return EINTR;
1550 		} else if (error == EWOULDBLOCK) {
1551 			zfree(aiol_zone, ijoblist);
1552 			zfree(aiol_zone, ujoblist);
1553 			return EAGAIN;
1554 		}
1555 	}
1556 
1557 /* NOTREACHED */
1558 	return EINVAL;
1559 }
1560 
1561 /*
1562  * aio_cancel at the kernel level is a NOOP right now.  It
1563  * might be possible to support it partially in user mode, or
1564  * in kernel mode later on.
1565  */
1566 int
1567 aio_cancel(struct proc *p, struct aio_cancel_args *uap)
1568 {
1569       return ENOSYS;
1570 }
1571 
1572 /*
1573  * aio_error is implemented in the kernel level for compatibility
1574  * purposes only.  For a user mode async implementation, it would be
1575  * best to do it in a userland subroutine.
1576  */
1577 int
1578 aio_error(struct proc *p, struct aio_error_args *uap)
1579 {
1580 	int s;
1581 	struct aiocblist *cb;
1582 	struct kaioinfo *ki;
1583 	int jobref;
1584 
1585 	ki = p->p_aioinfo;
1586 	if (ki == NULL)
1587 		return EINVAL;
1588 
1589 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1590 	if ((jobref == -1) || (jobref == 0))
1591 		return EINVAL;
1592 
1593 	for (cb = TAILQ_FIRST(&ki->kaio_jobdone);
1594 		cb;
1595 		cb = TAILQ_NEXT(cb, plist)) {
1596 
1597 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == jobref) {
1598 			p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1599 			return 0;
1600 		}
1601 	}
1602 
1603 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue);
1604 		cb;
1605 		cb = TAILQ_NEXT(cb, plist)) {
1606 
1607 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == jobref) {
1608 			p->p_retval[0] = EINPROGRESS;
1609 			return 0;
1610 		}
1611 	}
1612 
1613 	s = splbio();
1614 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone);
1615 		cb;
1616 		cb = TAILQ_NEXT(cb, plist)) {
1617 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == jobref) {
1618 			p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1619 			splx(s);
1620 			return 0;
1621 		}
1622 	}
1623 
1624 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue);
1625 		cb;
1626 		cb = TAILQ_NEXT(cb, plist)) {
1627 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) == jobref) {
1628 			p->p_retval[0] = EINPROGRESS;
1629 			splx(s);
1630 			return 0;
1631 		}
1632 	}
1633 	splx(s);
1634 
1635 
1636 	/*
1637 	 * Hack for lio
1638 	 */
1639 /*
1640 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1641 	if (status == -1) {
1642 		return fuword(&uap->aiocbp->_aiocb_private.error);
1643 	}
1644 */
1645 	return EINVAL;
1646 }
1647 
1648 int
1649 aio_read(struct proc *p, struct aio_read_args *uap)
1650 {
1651 	struct filedesc *fdp;
1652 	struct file *fp;
1653 	struct uio auio;
1654 	struct iovec aiov;
1655 	unsigned int fd;
1656 	int cnt;
1657 	struct aiocb iocb;
1658 	int error, pmodes;
1659 
1660 	pmodes = fuword(&uap->aiocbp->_aiocb_private.privatemodes);
1661 	if ((pmodes & AIO_PMODE_SYNC) == 0) {
1662 		return aio_aqueue(p, (struct aiocb *) uap->aiocbp, LIO_READ);
1663 	}
1664 
1665 	/*
1666 	 * Get control block
1667 	 */
1668 	if (error = copyin((caddr_t) uap->aiocbp, (caddr_t) &iocb, sizeof iocb))
1669 		return error;
1670 
1671 	/*
1672 	 * Get the fd info for process
1673 	 */
1674 	fdp = p->p_fd;
1675 
1676 	/*
1677 	 * Range check file descriptor
1678 	 */
1679 	fd = iocb.aio_fildes;
1680 	if (fd >= fdp->fd_nfiles)
1681 		return EBADF;
1682 	fp = fdp->fd_ofiles[fd];
1683 	if ((fp == NULL) || ((fp->f_flag & FREAD) == 0))
1684 		return EBADF;
1685 	if (iocb.aio_offset == -1LL)
1686 		return EINVAL;
1687 
1688 	auio.uio_resid = iocb.aio_nbytes;
1689 	if (auio.uio_resid < 0)
1690 		return (EINVAL);
1691 
1692 	/*
1693 	 * Process sync simply -- queue async request.
1694 	 */
1695 	if ((iocb._aiocb_private.privatemodes & AIO_PMODE_SYNC) == 0) {
1696 		return aio_aqueue(p, (struct aiocb *) uap->aiocbp, LIO_READ);
1697 	}
1698 
1699 	aiov.iov_base = (void *) iocb.aio_buf;
1700 	aiov.iov_len = iocb.aio_nbytes;
1701 
1702 	auio.uio_iov = &aiov;
1703 	auio.uio_iovcnt = 1;
1704 	auio.uio_offset = iocb.aio_offset;
1705 	auio.uio_rw = UIO_READ;
1706 	auio.uio_segflg = UIO_USERSPACE;
1707 	auio.uio_procp = p;
1708 
1709 	cnt = iocb.aio_nbytes;
1710 	error = (*fp->f_ops->fo_read)(fp, &auio, fp->f_cred);
1711 	if (error &&
1712 		(auio.uio_resid != cnt) &&
1713 		(error == ERESTART || error == EINTR || error == EWOULDBLOCK))
1714 			error = 0;
1715 	cnt -= auio.uio_resid;
1716 	p->p_retval[0] = cnt;
1717 	return error;
1718 }
1719 
1720 int
1721 aio_write(struct proc *p, struct aio_write_args *uap)
1722 {
1723 	struct filedesc *fdp;
1724 	struct file *fp;
1725 	struct uio auio;
1726 	struct iovec aiov;
1727 	unsigned int fd;
1728 	int cnt;
1729 	struct aiocb iocb;
1730 	int error;
1731 	int pmodes;
1732 
1733 	/*
1734 	 * Process sync simply -- queue async request.
1735 	 */
1736 	pmodes = fuword(&uap->aiocbp->_aiocb_private.privatemodes);
1737 	if ((pmodes & AIO_PMODE_SYNC) == 0) {
1738 		return aio_aqueue(p, (struct aiocb *) uap->aiocbp, LIO_WRITE);
1739 	}
1740 
1741 	if (error = copyin((caddr_t) uap->aiocbp, (caddr_t) &iocb, sizeof iocb))
1742 		return error;
1743 
1744 	/*
1745 	 * Get the fd info for process
1746 	 */
1747 	fdp = p->p_fd;
1748 
1749 	/*
1750 	 * Range check file descriptor
1751 	 */
1752 	fd = iocb.aio_fildes;
1753 	if (fd >= fdp->fd_nfiles)
1754 		return EBADF;
1755 	fp = fdp->fd_ofiles[fd];
1756 	if ((fp == NULL) || ((fp->f_flag & FWRITE) == 0))
1757 		return EBADF;
1758 	if (iocb.aio_offset == -1LL)
1759 		return EINVAL;
1760 
1761 	aiov.iov_base = (void *) iocb.aio_buf;
1762 	aiov.iov_len = iocb.aio_nbytes;
1763 	auio.uio_iov = &aiov;
1764 	auio.uio_iovcnt = 1;
1765 	auio.uio_offset = iocb.aio_offset;
1766 
1767 	auio.uio_resid = iocb.aio_nbytes;
1768 	if (auio.uio_resid < 0)
1769 		return (EINVAL);
1770 
1771 	auio.uio_rw = UIO_WRITE;
1772 	auio.uio_segflg = UIO_USERSPACE;
1773 	auio.uio_procp = p;
1774 
1775 	cnt = iocb.aio_nbytes;
1776 	error = (*fp->f_ops->fo_write)(fp, &auio, fp->f_cred);
1777 	if (error) {
1778 		if (auio.uio_resid != cnt) {
1779 			if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
1780 				error = 0;
1781 			if (error == EPIPE)
1782 				psignal(p, SIGPIPE);
1783 		}
1784 	}
1785 	cnt -= auio.uio_resid;
1786 	p->p_retval[0] = cnt;
1787 	return error;
1788 }
1789 
1790 int
1791 lio_listio(struct proc *p, struct lio_listio_args *uap)
1792 {
1793 	int nent, nentqueued;
1794 	struct aiocb *iocb, * const *cbptr;
1795 	struct aiocblist *cb;
1796 	struct kaioinfo *ki;
1797 	struct aio_liojob *lj;
1798 	int error, runningcode;
1799 	int nerror;
1800 	int i;
1801 	int s;
1802 
1803 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT)) {
1804 		return EINVAL;
1805 	}
1806 
1807 	nent = uap->nent;
1808 	if (nent > AIO_LISTIO_MAX) {
1809 		return EINVAL;
1810 	}
1811 
1812 	if (p->p_aioinfo == NULL) {
1813 		aio_init_aioinfo(p);
1814 	}
1815 
1816 	if ((nent + num_queue_count) > max_queue_count) {
1817 		return EAGAIN;
1818 	}
1819 
1820 	ki = p->p_aioinfo;
1821 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) {
1822 		return EAGAIN;
1823 	}
1824 
1825 	lj = zalloc(aiolio_zone);
1826 	if (!lj) {
1827 		return EAGAIN;
1828 	}
1829 
1830 	lj->lioj_flags = 0;
1831 	lj->lioj_buffer_count = 0;
1832 	lj->lioj_buffer_finished_count = 0;
1833 	lj->lioj_queue_count = 0;
1834 	lj->lioj_queue_finished_count = 0;
1835 	lj->lioj_ki = ki;
1836 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1837 
1838 	/*
1839 	 * Setup signal
1840 	 */
1841 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1842 		error = copyin(uap->sig, &lj->lioj_signal, sizeof lj->lioj_signal);
1843 		if (error)
1844 			return error;
1845 		lj->lioj_flags |= LIOJ_SIGNAL;
1846 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1847 	} else {
1848 		lj->lioj_flags &= ~LIOJ_SIGNAL;
1849 	}
1850 
1851 /*
1852  * get pointers to the list of I/O requests
1853  */
1854 
1855 	nerror = 0;
1856 	nentqueued = 0;
1857 	cbptr = uap->acb_list;
1858 	for(i = 0; i < uap->nent; i++) {
1859 		iocb = (struct aiocb *) (intptr_t) fuword((caddr_t) &cbptr[i]);
1860 		if (((intptr_t) iocb != -1) && ((intptr_t) iocb != NULL)) {
1861 			error = _aio_aqueue(p, iocb, lj, 0);
1862 			if (error == 0) {
1863 				nentqueued++;
1864 			} else {
1865 				nerror++;
1866 			}
1867 		}
1868 	}
1869 
1870 	/*
1871 	 * If we haven't queued any, then just return error
1872 	 */
1873 	if (nentqueued == 0) {
1874 		return 0;
1875 	}
1876 
1877 	/*
1878 	 * Calculate the appropriate error return
1879 	 */
1880 	runningcode = 0;
1881 	if (nerror)
1882 		runningcode = EIO;
1883 
1884 	if (uap->mode == LIO_WAIT) {
1885 		while (1) {
1886 			int found;
1887 			found = 0;
1888 			for(i = 0; i < uap->nent; i++) {
1889 				int jobref, command;
1890 
1891 				/*
1892 				 * Fetch address of the control buf pointer in user space
1893 				 */
1894 				iocb = (struct aiocb *) (intptr_t) fuword((caddr_t) &cbptr[i]);
1895 				if (((intptr_t) iocb == -1) || ((intptr_t) iocb == 0))
1896 					continue;
1897 
1898 				/*
1899 				 * Fetch the associated command from user space
1900 				 */
1901 				command = fuword(&iocb->aio_lio_opcode);
1902 				if (command == LIO_NOP) {
1903 					found++;
1904 					continue;
1905 				}
1906 
1907 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1908 
1909 				for (cb = TAILQ_FIRST(&ki->kaio_jobdone);
1910 					cb;
1911 					cb = TAILQ_NEXT(cb, plist)) {
1912 					if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1913 						jobref) {
1914 						if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1915 							curproc->p_stats->p_ru.ru_oublock +=
1916 								cb->outputcharge;
1917 							cb->outputcharge = 0;
1918 						} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1919 							curproc->p_stats->p_ru.ru_inblock +=
1920 								cb->inputcharge;
1921 							cb->inputcharge = 0;
1922 						}
1923 						found++;
1924 						break;
1925 					}
1926 				}
1927 
1928 				s = splbio();
1929 				for (cb = TAILQ_FIRST(&ki->kaio_bufdone);
1930 					cb;
1931 					cb = TAILQ_NEXT(cb, plist)) {
1932 					if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1933 						jobref) {
1934 						found++;
1935 						break;
1936 					}
1937 				}
1938 				splx(s);
1939 
1940 			}
1941 
1942 			/*
1943 			 * If all I/Os have been disposed of, then we can return
1944 			 */
1945 			if (found == nentqueued) {
1946 				return runningcode;
1947 			}
1948 
1949 			ki->kaio_flags |= KAIO_WAKEUP;
1950 			error = tsleep(p, PRIBIO|PCATCH, "aiospn", 0);
1951 
1952 			if (error == EINTR) {
1953 				return EINTR;
1954 			} else if (error == EWOULDBLOCK) {
1955 				return EAGAIN;
1956 			}
1957 
1958 		}
1959 	}
1960 
1961 	return runningcode;
1962 }
1963 
1964 /*
1965  * This is a wierd hack so that we can post a signal.  It is safe
1966  * to do so from a timeout routine, but *not* from an interrupt routine.
1967  */
1968 static void
1969 process_signal(void *ljarg)
1970 {
1971 	struct aio_liojob *lj = ljarg;
1972 	if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) {
1973 		if (lj->lioj_queue_count == lj->lioj_queue_finished_count) {
1974 			psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1975 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1976 		}
1977 	}
1978 }
1979 
1980 /*
1981  * Interrupt handler for physio, performs the necessary process wakeups,
1982  * and signals.
1983  */
1984 static void
1985 aio_physwakeup(bp)
1986 	struct buf *bp;
1987 {
1988 	struct aiocblist *aiocbe;
1989 	struct proc *p;
1990 	struct kaioinfo *ki;
1991 	struct aio_liojob *lj;
1992 	int s;
1993 	s = splbio();
1994 
1995 	wakeup((caddr_t) bp);
1996 	bp->b_flags &= ~B_CALL;
1997 	bp->b_flags |= B_DONE;
1998 
1999 	aiocbe = (struct aiocblist *)bp->b_spc;
2000 	if (aiocbe) {
2001 		p = bp->b_proc;
2002 
2003 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2004 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2005 		aiocbe->uaiocb._aiocb_private.error = 0;
2006 		aiocbe->jobflags |= AIOCBLIST_DONE;
2007 
2008 		if (bp->b_flags & B_ERROR) {
2009 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2010 		}
2011 
2012 		lj = aiocbe->lio;
2013 		if (lj) {
2014 			lj->lioj_buffer_finished_count++;
2015 			/*
2016 			 * wakeup/signal if all of the interrupt jobs are done
2017 			 */
2018 			if (lj->lioj_buffer_finished_count == lj->lioj_buffer_count) {
2019 				/*
2020 				 * post a signal if it is called for
2021 				 */
2022 				if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2023 					LIOJ_SIGNAL) {
2024 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2025 					timeout(process_signal, lj, 0);
2026 				}
2027 			}
2028 		}
2029 
2030 		ki = p->p_aioinfo;
2031 		if (ki) {
2032 			ki->kaio_buffer_finished_count++;
2033 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2034 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2035 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2036 			/*
2037 			 * and do the wakeup
2038 			 */
2039 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2040 				ki->kaio_flags &= ~KAIO_WAKEUP;
2041 				wakeup(p);
2042 			}
2043 		}
2044 	}
2045 	splx(s);
2046 }
2047