xref: /freebsd/sys/kern/vfs_aio.c (revision 1b6c76a2fe091c74f08427e6c870851025a9cf67)
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/sysctl.h>
42 #include <sys/vnode.h>
43 #include <sys/conf.h>
44 #include <sys/event.h>
45 
46 #include <vm/vm.h>
47 #include <vm/vm_extern.h>
48 #include <vm/pmap.h>
49 #include <vm/vm_map.h>
50 #include <vm/vm_zone.h>
51 #include <sys/aio.h>
52 
53 #include <machine/limits.h>
54 
55 #include "opt_vfs_aio.h"
56 
57 #ifdef VFS_AIO
58 
59 static	long jobrefid;
60 
61 #define JOBST_NULL		0x0
62 #define	JOBST_JOBQPROC		0x1
63 #define JOBST_JOBQGLOBAL	0x2
64 #define JOBST_JOBRUNNING	0x3
65 #define JOBST_JOBFINISHED	0x4
66 #define	JOBST_JOBQBUF		0x5
67 #define	JOBST_JOBBFINISHED	0x6
68 
69 #ifndef MAX_AIO_PER_PROC
70 #define MAX_AIO_PER_PROC	32
71 #endif
72 
73 #ifndef MAX_AIO_QUEUE_PER_PROC
74 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
75 #endif
76 
77 #ifndef MAX_AIO_PROCS
78 #define MAX_AIO_PROCS		32
79 #endif
80 
81 #ifndef MAX_AIO_QUEUE
82 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
83 #endif
84 
85 #ifndef TARGET_AIO_PROCS
86 #define TARGET_AIO_PROCS	4
87 #endif
88 
89 #ifndef MAX_BUF_AIO
90 #define MAX_BUF_AIO		16
91 #endif
92 
93 #ifndef AIOD_TIMEOUT_DEFAULT
94 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
95 #endif
96 
97 #ifndef AIOD_LIFETIME_DEFAULT
98 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
99 #endif
100 
101 static int max_aio_procs = MAX_AIO_PROCS;
102 static int num_aio_procs = 0;
103 static int target_aio_procs = TARGET_AIO_PROCS;
104 static int max_queue_count = MAX_AIO_QUEUE;
105 static int num_queue_count = 0;
106 static int num_buf_aio = 0;
107 static int num_aio_resv_start = 0;
108 static int aiod_timeout;
109 static int aiod_lifetime;
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 /*
151  * AIO process info
152  */
153 #define AIOP_FREE	0x1			/* proc on free queue */
154 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
155 
156 struct aioproclist {
157 	int aioprocflags;			/* AIO proc flags */
158 	TAILQ_ENTRY(aioproclist) list;		/* List of processes */
159 	struct proc *aioproc;			/* The AIO thread */
160 	TAILQ_HEAD (,aiocblist) jobtorun;	/* suggested job to run */
161 };
162 
163 /*
164  * data-structure for lio signal management
165  */
166 struct aio_liojob {
167 	int	lioj_flags;
168 	int	lioj_buffer_count;
169 	int	lioj_buffer_finished_count;
170 	int	lioj_queue_count;
171 	int	lioj_queue_finished_count;
172 	struct	sigevent lioj_signal;	/* signal on all I/O done */
173 	TAILQ_ENTRY	(aio_liojob) lioj_list;
174 	struct	kaioinfo *lioj_ki;
175 };
176 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
177 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
178 
179 /*
180  * per process aio data structure
181  */
182 struct kaioinfo {
183 	int	kaio_flags;		/* per process kaio flags */
184 	int	kaio_maxactive_count;	/* maximum number of AIOs */
185 	int	kaio_active_count;	/* number of currently used AIOs */
186 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
187 	int	kaio_queue_count;	/* size of AIO queue */
188 	int	kaio_ballowed_count;	/* maximum number of buffers */
189 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
190 	int	kaio_buffer_count;	/* number of physio buffers */
191 	int	kaio_buffer_finished_count; /* count of I/O done */
192 	struct 	proc *kaio_p;		/* process that uses this kaio block */
193 	TAILQ_HEAD (,aio_liojob) kaio_liojoblist; /* list of lio jobs */
194 	TAILQ_HEAD (,aiocblist)	kaio_jobqueue;	/* job queue for process */
195 	TAILQ_HEAD (,aiocblist)	kaio_jobdone;	/* done queue for process */
196 	TAILQ_HEAD (,aiocblist)	kaio_bufqueue;	/* buffer job queue for process */
197 	TAILQ_HEAD (,aiocblist)	kaio_bufdone;	/* buffer done queue for process */
198 	TAILQ_HEAD (,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
199 };
200 
201 #define KAIO_RUNDOWN	0x1	/* process is being run down */
202 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
203 
204 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
205 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
206 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
207 
208 static void	aio_init_aioinfo(struct proc *p);
209 static void	aio_onceonly(void *);
210 static int	aio_free_entry(struct aiocblist *aiocbe);
211 static void	aio_process(struct aiocblist *aiocbe);
212 static int	aio_newproc(void);
213 static int	aio_aqueue(struct proc *p, struct aiocb *job, int type);
214 static void	aio_physwakeup(struct buf *bp);
215 static int	aio_fphysio(struct proc *p, struct aiocblist *aiocbe);
216 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
217 static void	aio_daemon(void *uproc);
218 static void	process_signal(void *aioj);
219 
220 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
221 
222 static vm_zone_t kaio_zone = 0, aiop_zone = 0, aiocb_zone = 0, aiol_zone = 0;
223 static vm_zone_t aiolio_zone = 0;
224 
225 /*
226  * Startup initialization
227  */
228 static void
229 aio_onceonly(void *na)
230 {
231 	TAILQ_INIT(&aio_freeproc);
232 	TAILQ_INIT(&aio_activeproc);
233 	TAILQ_INIT(&aio_jobs);
234 	TAILQ_INIT(&aio_bufjobs);
235 	kaio_zone = zinit("AIO", sizeof (struct kaioinfo), 0, 0, 1);
236 	aiop_zone = zinit("AIOP", sizeof (struct aioproclist), 0, 0, 1);
237 	aiocb_zone = zinit("AIOCB", sizeof (struct aiocblist), 0, 0, 1);
238 	aiol_zone = zinit("AIOL", AIO_LISTIO_MAX * sizeof (int), 0, 0, 1);
239 	aiolio_zone = zinit("AIOLIO", AIO_LISTIO_MAX * sizeof (struct
240 	    aio_liojob), 0, 0, 1);
241 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
242 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
243 	jobrefid = 1;
244 }
245 
246 /*
247  * Init the per-process aioinfo structure.  The aioinfo limits are set
248  * per-process for user limit (resource) management.
249  */
250 static void
251 aio_init_aioinfo(struct proc *p)
252 {
253 	struct kaioinfo *ki;
254 	if (p->p_aioinfo == NULL) {
255 		ki = zalloc(kaio_zone);
256 		p->p_aioinfo = ki;
257 		ki->kaio_flags = 0;
258 		ki->kaio_maxactive_count = max_aio_per_proc;
259 		ki->kaio_active_count = 0;
260 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
261 		ki->kaio_queue_count = 0;
262 		ki->kaio_ballowed_count = max_buf_aio;
263 		ki->kaio_buffer_count = 0;
264 		ki->kaio_buffer_finished_count = 0;
265 		ki->kaio_p = p;
266 		TAILQ_INIT(&ki->kaio_jobdone);
267 		TAILQ_INIT(&ki->kaio_jobqueue);
268 		TAILQ_INIT(&ki->kaio_bufdone);
269 		TAILQ_INIT(&ki->kaio_bufqueue);
270 		TAILQ_INIT(&ki->kaio_liojoblist);
271 		TAILQ_INIT(&ki->kaio_sockqueue);
272 	}
273 
274 	while (num_aio_procs < target_aio_procs)
275 		aio_newproc();
276 }
277 
278 /*
279  * Free a job entry.  Wait for completion if it is currently active, but don't
280  * delay forever.  If we delay, we return a flag that says that we have to
281  * restart the queue scan.
282  */
283 static int
284 aio_free_entry(struct aiocblist *aiocbe)
285 {
286 	struct kaioinfo *ki;
287 	struct aioproclist *aiop;
288 	struct aio_liojob *lj;
289 	struct proc *p;
290 	int error;
291 	int s;
292 
293 	if (aiocbe->jobstate == JOBST_NULL)
294 		panic("aio_free_entry: freeing already free job");
295 
296 	p = aiocbe->userproc;
297 	ki = p->p_aioinfo;
298 	lj = aiocbe->lio;
299 	if (ki == NULL)
300 		panic("aio_free_entry: missing p->p_aioinfo");
301 
302 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
303 		if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE)
304 			return 0;
305 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
306 		tsleep(aiocbe, PRIBIO, "jobwai", 0);
307 	}
308 	aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
309 
310 	if (aiocbe->bp == NULL) {
311 		if (ki->kaio_queue_count <= 0)
312 			panic("aio_free_entry: process queue size <= 0");
313 		if (num_queue_count <= 0)
314 			panic("aio_free_entry: system wide queue size <= 0");
315 
316 		if (lj) {
317 			lj->lioj_queue_count--;
318 			if (aiocbe->jobflags & AIOCBLIST_DONE)
319 				lj->lioj_queue_finished_count--;
320 		}
321 		ki->kaio_queue_count--;
322 		if (aiocbe->jobflags & AIOCBLIST_DONE)
323 			ki->kaio_queue_finished_count--;
324 		num_queue_count--;
325 	} else {
326 		if (lj) {
327 			lj->lioj_buffer_count--;
328 			if (aiocbe->jobflags & AIOCBLIST_DONE)
329 				lj->lioj_buffer_finished_count--;
330 		}
331 		if (aiocbe->jobflags & AIOCBLIST_DONE)
332 			ki->kaio_buffer_finished_count--;
333 		ki->kaio_buffer_count--;
334 		num_buf_aio--;
335 	}
336 
337 	/* aiocbe is going away, we need to destroy any knotes */
338 	knote_remove(p, &aiocbe->klist);
339 
340 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
341 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
342 		ki->kaio_flags &= ~KAIO_WAKEUP;
343 		wakeup(p);
344 	}
345 
346 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
347 		if ((error = aio_fphysio(p, aiocbe)) != 0)
348 			return error;
349 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
350 			panic("aio_free_entry: invalid physio finish-up state");
351 		s = splbio();
352 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
353 		splx(s);
354 	} else if (aiocbe->jobstate == JOBST_JOBQPROC) {
355 		aiop = aiocbe->jobaioproc;
356 		TAILQ_REMOVE(&aiop->jobtorun, aiocbe, list);
357 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
358 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
359 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
360 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
361 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
362 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
363 		s = splbio();
364 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
365 		splx(s);
366 		if (aiocbe->bp) {
367 			vunmapbuf(aiocbe->bp);
368 			relpbuf(aiocbe->bp, NULL);
369 			aiocbe->bp = NULL;
370 		}
371 	}
372 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
373 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
374 		zfree(aiolio_zone, lj);
375 	}
376 	aiocbe->jobstate = JOBST_NULL;
377 	untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
378 	zfree(aiocb_zone, aiocbe);
379 	return 0;
380 }
381 #endif /* VFS_AIO */
382 
383 /*
384  * Rundown the jobs for a given process.
385  */
386 void
387 aio_proc_rundown(struct proc *p)
388 {
389 #ifndef VFS_AIO
390 	return;
391 #else
392 	int s;
393 	struct kaioinfo *ki;
394 	struct aio_liojob *lj, *ljn;
395 	struct aiocblist *aiocbe, *aiocbn;
396 	struct file *fp;
397 	struct filedesc *fdp;
398 	struct socket *so;
399 
400 	ki = p->p_aioinfo;
401 	if (ki == NULL)
402 		return;
403 
404 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
405 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
406 	    ki->kaio_buffer_finished_count)) {
407 		ki->kaio_flags |= KAIO_RUNDOWN;
408 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
409 			break;
410 	}
411 
412 	/*
413 	 * Move any aio ops that are waiting on socket I/O to the normal job
414 	 * queues so they are cleaned up with any others.
415 	 */
416 	fdp = p->p_fd;
417 
418 	s = splnet();
419 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
420 	    aiocbn) {
421 		aiocbn = TAILQ_NEXT(aiocbe, plist);
422 		fp = fdp->fd_ofiles[aiocbe->uaiocb.aio_fildes];
423 
424 		/*
425 		 * Under some circumstances, the aio_fildes and the file
426 		 * structure don't match.  This would leave aiocbe's in the
427 		 * TAILQ associated with the socket and cause a panic later.
428 		 *
429 		 * Detect and fix.
430 		 */
431 		if ((fp == NULL) || (fp != aiocbe->fd_file))
432 			fp = aiocbe->fd_file;
433 		if (fp) {
434 			so = (struct socket *)fp->f_data;
435 			TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
436 			if (TAILQ_EMPTY(&so->so_aiojobq)) {
437 				so->so_snd.sb_flags &= ~SB_AIO;
438 				so->so_rcv.sb_flags &= ~SB_AIO;
439 			}
440 		}
441 		TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
442 		TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
443 		TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
444 	}
445 	splx(s);
446 
447 restart1:
448 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
449 		aiocbn = TAILQ_NEXT(aiocbe, plist);
450 		if (aio_free_entry(aiocbe))
451 			goto restart1;
452 	}
453 
454 restart2:
455 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
456 	    aiocbn) {
457 		aiocbn = TAILQ_NEXT(aiocbe, plist);
458 		if (aio_free_entry(aiocbe))
459 			goto restart2;
460 	}
461 
462 /*
463  * Note the use of lots of splbio here, trying to avoid splbio for long chains
464  * of I/O.  Probably unnecessary.
465  */
466 restart3:
467 	s = splbio();
468 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
469 		ki->kaio_flags |= KAIO_WAKEUP;
470 		tsleep(p, PRIBIO, "aioprn", 0);
471 		splx(s);
472 		goto restart3;
473 	}
474 	splx(s);
475 
476 restart4:
477 	s = splbio();
478 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
479 		aiocbn = TAILQ_NEXT(aiocbe, plist);
480 		if (aio_free_entry(aiocbe)) {
481 			splx(s);
482 			goto restart4;
483 		}
484 	}
485 	splx(s);
486 
487         /*
488          * If we've slept, jobs might have moved from one queue to another.
489          * Retry rundown if we didn't manage to empty the queues.
490          */
491         if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
492 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
493 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
494 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
495 		goto restart1;
496 
497 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
498 		ljn = TAILQ_NEXT(lj, lioj_list);
499 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
500 		    0)) {
501 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
502 			zfree(aiolio_zone, lj);
503 		} else {
504 #ifdef DIAGNOSTIC
505 			printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
506 			    "QF:%d\n", lj->lioj_buffer_count,
507 			    lj->lioj_buffer_finished_count,
508 			    lj->lioj_queue_count,
509 			    lj->lioj_queue_finished_count);
510 #endif
511 		}
512 	}
513 
514 	zfree(kaio_zone, ki);
515 	p->p_aioinfo = NULL;
516 #endif /* VFS_AIO */
517 }
518 
519 #ifdef VFS_AIO
520 /*
521  * Select a job to run (called by an AIO daemon).
522  */
523 static struct aiocblist *
524 aio_selectjob(struct aioproclist *aiop)
525 {
526 	int s;
527 	struct aiocblist *aiocbe;
528 	struct kaioinfo *ki;
529 	struct proc *userp;
530 
531 	aiocbe = TAILQ_FIRST(&aiop->jobtorun);
532 	if (aiocbe) {
533 		TAILQ_REMOVE(&aiop->jobtorun, aiocbe, list);
534 		return aiocbe;
535 	}
536 
537 	s = splnet();
538 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
539 	    TAILQ_NEXT(aiocbe, list)) {
540 		userp = aiocbe->userproc;
541 		ki = userp->p_aioinfo;
542 
543 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
544 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
545 			splx(s);
546 			return aiocbe;
547 		}
548 	}
549 	splx(s);
550 
551 	return NULL;
552 }
553 
554 /*
555  * The AIO processing activity.  This is the code that does the I/O request for
556  * the non-physio version of the operations.  The normal vn operations are used,
557  * and this code should work in all instances for every type of file, including
558  * pipes, sockets, fifos, and regular files.
559  */
560 static void
561 aio_process(struct aiocblist *aiocbe)
562 {
563 	struct filedesc *fdp;
564 	struct proc *userp, *mycp;
565 	struct aiocb *cb;
566 	struct file *fp;
567 	struct uio auio;
568 	struct iovec aiov;
569 	unsigned int fd;
570 	int cnt;
571 	int error;
572 	off_t offset;
573 	int oublock_st, oublock_end;
574 	int inblock_st, inblock_end;
575 
576 	userp = aiocbe->userproc;
577 	cb = &aiocbe->uaiocb;
578 
579 	mycp = curproc;
580 
581 	fdp = mycp->p_fd;
582 	fd = cb->aio_fildes;
583 	fp = fdp->fd_ofiles[fd];
584 
585 	if ((fp == NULL) || (fp != aiocbe->fd_file)) {
586 		cb->_aiocb_private.error = EBADF;
587 		cb->_aiocb_private.status = -1;
588 		return;
589 	}
590 
591 	aiov.iov_base = (void *)cb->aio_buf;
592 	aiov.iov_len = cb->aio_nbytes;
593 
594 	auio.uio_iov = &aiov;
595 	auio.uio_iovcnt = 1;
596 	auio.uio_offset = offset = cb->aio_offset;
597 	auio.uio_resid = cb->aio_nbytes;
598 	cnt = cb->aio_nbytes;
599 	auio.uio_segflg = UIO_USERSPACE;
600 	auio.uio_procp = mycp;
601 
602 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
603 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
604 	/*
605 	 * Temporarily bump the ref count while reading to avoid the
606 	 * descriptor being ripped out from under us.
607 	 */
608 	fhold(fp);
609 	if (cb->aio_lio_opcode == LIO_READ) {
610 		auio.uio_rw = UIO_READ;
611 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, mycp);
612 	} else {
613 		auio.uio_rw = UIO_WRITE;
614 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, mycp);
615 	}
616 	fdrop(fp, mycp);
617 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
618 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
619 
620 	aiocbe->inputcharge = inblock_end - inblock_st;
621 	aiocbe->outputcharge = oublock_end - oublock_st;
622 
623 	if ((error) && (auio.uio_resid != cnt)) {
624 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
625 			error = 0;
626 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
627 			PROC_LOCK(userp);
628 			psignal(userp, SIGPIPE);
629 			PROC_UNLOCK(userp);
630 		}
631 	}
632 
633 	cnt -= auio.uio_resid;
634 	cb->_aiocb_private.error = error;
635 	cb->_aiocb_private.status = cnt;
636 }
637 
638 /*
639  * The AIO daemon, most of the actual work is done in aio_process,
640  * but the setup (and address space mgmt) is done in this routine.
641  */
642 static void
643 aio_daemon(void *uproc)
644 {
645 	int s;
646 	struct aio_liojob *lj;
647 	struct aiocb *cb;
648 	struct aiocblist *aiocbe;
649 	struct aioproclist *aiop;
650 	struct kaioinfo *ki;
651 	struct proc *curcp, *mycp, *userp;
652 	struct vmspace *myvm, *tmpvm;
653 
654 	mtx_lock(&Giant);
655 	/*
656 	 * Local copies of curproc (cp) and vmspace (myvm)
657 	 */
658 	mycp = curproc;
659 	myvm = mycp->p_vmspace;
660 
661 	if (mycp->p_textvp) {
662 		vrele(mycp->p_textvp);
663 		mycp->p_textvp = NULL;
664 	}
665 
666 	/*
667 	 * Allocate and ready the aio control info.  There is one aiop structure
668 	 * per daemon.
669 	 */
670 	aiop = zalloc(aiop_zone);
671 	aiop->aioproc = mycp;
672 	aiop->aioprocflags |= AIOP_FREE;
673 	TAILQ_INIT(&aiop->jobtorun);
674 
675 	s = splnet();
676 
677 	/*
678 	 * Place thread (lightweight process) onto the AIO free thread list.
679 	 */
680 	if (TAILQ_EMPTY(&aio_freeproc))
681 		wakeup(&aio_freeproc);
682 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
683 
684 	splx(s);
685 
686 	/*
687 	 * Get rid of our current filedescriptors.  AIOD's don't need any
688 	 * filedescriptors, except as temporarily inherited from the client.
689 	 * Credentials are also cloned, and made equivalent to "root".
690 	 */
691 	fdfree(mycp);
692 	mycp->p_fd = NULL;
693 	mycp->p_ucred = crcopy(mycp->p_ucred);
694 	mycp->p_ucred->cr_uid = 0;
695 	uifree(mycp->p_ucred->cr_uidinfo);
696 	mycp->p_ucred->cr_uidinfo = uifind(0);
697 	mycp->p_ucred->cr_ngroups = 1;
698 	mycp->p_ucred->cr_groups[0] = 1;
699 
700 	/* The daemon resides in its own pgrp. */
701 	enterpgrp(mycp, mycp->p_pid, 1);
702 
703 	/* Mark special process type. */
704 	mycp->p_flag |= P_SYSTEM;
705 
706 	/*
707 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
708 	 * creating to many daemons.)
709 	 */
710 	wakeup(mycp);
711 
712 	for (;;) {
713 		/*
714 		 * curcp is the current daemon process context.
715 		 * userp is the current user process context.
716 		 */
717 		curcp = mycp;
718 
719 		/*
720 		 * Take daemon off of free queue
721 		 */
722 		if (aiop->aioprocflags & AIOP_FREE) {
723 			s = splnet();
724 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
725 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
726 			aiop->aioprocflags &= ~AIOP_FREE;
727 			splx(s);
728 		}
729 		aiop->aioprocflags &= ~AIOP_SCHED;
730 
731 		/*
732 		 * Check for jobs.
733 		 */
734 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
735 			cb = &aiocbe->uaiocb;
736 			userp = aiocbe->userproc;
737 
738 			aiocbe->jobstate = JOBST_JOBRUNNING;
739 
740 			/*
741 			 * Connect to process address space for user program.
742 			 */
743 			if (userp != curcp) {
744 				/*
745 				 * Save the current address space that we are
746 				 * connected to.
747 				 */
748 				tmpvm = mycp->p_vmspace;
749 
750 				/*
751 				 * Point to the new user address space, and
752 				 * refer to it.
753 				 */
754 				mycp->p_vmspace = userp->p_vmspace;
755 				mycp->p_vmspace->vm_refcnt++;
756 
757 				/* Activate the new mapping. */
758 				pmap_activate(mycp);
759 
760 				/*
761 				 * If the old address space wasn't the daemons
762 				 * own address space, then we need to remove the
763 				 * daemon's reference from the other process
764 				 * that it was acting on behalf of.
765 				 */
766 				if (tmpvm != myvm) {
767 					vmspace_free(tmpvm);
768 				}
769 
770 				/*
771 				 * Disassociate from previous clients file
772 				 * descriptors, and associate to the new clients
773 				 * descriptors.  Note that the daemon doesn't
774 				 * need to worry about its orginal descriptors,
775 				 * because they were originally freed.
776 				 */
777 				if (mycp->p_fd)
778 					fdfree(mycp);
779 				mycp->p_fd = fdshare(userp);
780 				curcp = userp;
781 			}
782 
783 			ki = userp->p_aioinfo;
784 			lj = aiocbe->lio;
785 
786 			/* Account for currently active jobs. */
787 			ki->kaio_active_count++;
788 
789 			/* Do the I/O function. */
790 			aiocbe->jobaioproc = aiop;
791 			aio_process(aiocbe);
792 
793 			/* Decrement the active job count. */
794 			ki->kaio_active_count--;
795 
796 			/*
797 			 * Increment the completion count for wakeup/signal
798 			 * comparisons.
799 			 */
800 			aiocbe->jobflags |= AIOCBLIST_DONE;
801 			ki->kaio_queue_finished_count++;
802 			if (lj)
803 				lj->lioj_queue_finished_count++;
804 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
805 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
806 				ki->kaio_flags &= ~KAIO_WAKEUP;
807 				wakeup(userp);
808 			}
809 
810 			s = splbio();
811 			if (lj && (lj->lioj_flags &
812 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
813 				if ((lj->lioj_queue_finished_count ==
814 				    lj->lioj_queue_count) &&
815 				    (lj->lioj_buffer_finished_count ==
816 				    lj->lioj_buffer_count)) {
817 					PROC_LOCK(userp);
818 					psignal(userp,
819 					    lj->lioj_signal.sigev_signo);
820 					PROC_UNLOCK(userp);
821 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
822 				}
823 			}
824 			splx(s);
825 
826 			aiocbe->jobstate = JOBST_JOBFINISHED;
827 
828 			/*
829 			 * If the I/O request should be automatically rundown,
830 			 * do the needed cleanup.  Otherwise, place the queue
831 			 * entry for the just finished I/O request into the done
832 			 * queue for the associated client.
833 			 */
834 			s = splnet();
835 			if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE) {
836 				aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
837 				zfree(aiocb_zone, aiocbe);
838 			} else {
839 				TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
840 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe,
841 				    plist);
842 			}
843 			splx(s);
844 			KNOTE(&aiocbe->klist, 0);
845 
846 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
847 				wakeup(aiocbe);
848 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
849 			}
850 
851 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
852 				PROC_LOCK(userp);
853 				psignal(userp, cb->aio_sigevent.sigev_signo);
854 				PROC_UNLOCK(userp);
855 			}
856 		}
857 
858 		/*
859 		 * Disconnect from user address space.
860 		 */
861 		if (curcp != mycp) {
862 			/* Get the user address space to disconnect from. */
863 			tmpvm = mycp->p_vmspace;
864 
865 			/* Get original address space for daemon. */
866 			mycp->p_vmspace = myvm;
867 
868 			/* Activate the daemon's address space. */
869 			pmap_activate(mycp);
870 #ifdef DIAGNOSTIC
871 			if (tmpvm == myvm) {
872 				printf("AIOD: vmspace problem -- %d\n",
873 				    mycp->p_pid);
874 			}
875 #endif
876 			/* Remove our vmspace reference. */
877 			vmspace_free(tmpvm);
878 
879 			/*
880 			 * Disassociate from the user process's file
881 			 * descriptors.
882 			 */
883 			if (mycp->p_fd)
884 				fdfree(mycp);
885 			mycp->p_fd = NULL;
886 			curcp = mycp;
887 		}
888 
889 		/*
890 		 * If we are the first to be put onto the free queue, wakeup
891 		 * anyone waiting for a daemon.
892 		 */
893 		s = splnet();
894 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
895 		if (TAILQ_EMPTY(&aio_freeproc))
896 			wakeup(&aio_freeproc);
897 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
898 		aiop->aioprocflags |= AIOP_FREE;
899 		splx(s);
900 
901 		/*
902 		 * If daemon is inactive for a long time, allow it to exit,
903 		 * thereby freeing resources.
904 		 */
905 		if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
906 		    PRIBIO, "aiordy", aiod_lifetime)) {
907 			s = splnet();
908 			if ((TAILQ_FIRST(&aio_jobs) == NULL) &&
909 			    (TAILQ_FIRST(&aiop->jobtorun) == NULL)) {
910 				if ((aiop->aioprocflags & AIOP_FREE) &&
911 				    (num_aio_procs > target_aio_procs)) {
912 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
913 					splx(s);
914 					zfree(aiop_zone, aiop);
915 					num_aio_procs--;
916 #ifdef DIAGNOSTIC
917 					if (mycp->p_vmspace->vm_refcnt <= 1) {
918 						printf("AIOD: bad vm refcnt for"
919 						    " exiting daemon: %d\n",
920 						    mycp->p_vmspace->vm_refcnt);
921 					}
922 #endif
923 					kthread_exit(0);
924 				}
925 			}
926 			splx(s);
927 		}
928 	}
929 }
930 
931 /*
932  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
933  * AIO daemon modifies its environment itself.
934  */
935 static int
936 aio_newproc()
937 {
938 	int error;
939 	struct proc *p;
940 
941 	error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, "aiod%d",
942 			       num_aio_procs);
943 	if (error)
944 		return error;
945 
946 	/*
947 	 * Wait until daemon is started, but continue on just in case to
948 	 * handle error conditions.
949 	 */
950 	error = tsleep(p, PZERO, "aiosta", aiod_timeout);
951 
952 	num_aio_procs++;
953 
954 	return error;
955 }
956 
957 /*
958  * Try the high-performance, low-overhead physio method for eligible
959  * VCHR devices.  This method doesn't use an aio helper thread, and
960  * thus has very low overhead.
961  *
962  * Assumes that the caller, _aio_aqueue(), has incremented the file
963  * structure's reference count, preventing its deallocation for the
964  * duration of this call.
965  */
966 static int
967 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
968 {
969 	int error;
970 	struct aiocb *cb;
971 	struct file *fp;
972 	struct buf *bp;
973 	struct vnode *vp;
974 	struct kaioinfo *ki;
975 	struct filedesc *fdp;
976 	struct aio_liojob *lj;
977 	int fd;
978 	int s;
979 	int notify;
980 
981 	cb = &aiocbe->uaiocb;
982 	fdp = p->p_fd;
983 	fd = cb->aio_fildes;
984 	fp = fdp->fd_ofiles[fd];
985 
986 	if (fp->f_type != DTYPE_VNODE)
987 		return (-1);
988 
989 	vp = (struct vnode *)fp->f_data;
990 
991 	/*
992 	 * If its not a disk, we don't want to return a positive error.
993 	 * It causes the aio code to not fall through to try the thread
994 	 * way when you're talking to a regular file.
995 	 */
996 	if (!vn_isdisk(vp, &error)) {
997 		if (error == ENOTBLK)
998 			return (-1);
999 		else
1000 			return (error);
1001 	}
1002 
1003  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
1004 		return (-1);
1005 
1006 	if (cb->aio_nbytes >
1007 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1008 		return (-1);
1009 
1010 	ki = p->p_aioinfo;
1011 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1012 		return (-1);
1013 
1014 	ki->kaio_buffer_count++;
1015 
1016 	lj = aiocbe->lio;
1017 	if (lj)
1018 		lj->lioj_buffer_count++;
1019 
1020 	/* Create and build a buffer header for a transfer. */
1021 	bp = (struct buf *)getpbuf(NULL);
1022 	BUF_KERNPROC(bp);
1023 
1024 	/*
1025 	 * Get a copy of the kva from the physical buffer.
1026 	 */
1027 	bp->b_caller1 = p;
1028 	bp->b_dev = vp->v_rdev;
1029 	error = bp->b_error = 0;
1030 
1031 	bp->b_bcount = cb->aio_nbytes;
1032 	bp->b_bufsize = cb->aio_nbytes;
1033 	bp->b_flags = B_PHYS;
1034 	bp->b_iodone = aio_physwakeup;
1035 	bp->b_saveaddr = bp->b_data;
1036 	bp->b_data = (void *)cb->aio_buf;
1037 	bp->b_blkno = btodb(cb->aio_offset);
1038 
1039 	if (cb->aio_lio_opcode == LIO_WRITE) {
1040 		bp->b_iocmd = BIO_WRITE;
1041 		if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_READ)) {
1042 			error = EFAULT;
1043 			goto doerror;
1044 		}
1045 	} else {
1046 		bp->b_iocmd = BIO_READ;
1047 		if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_WRITE)) {
1048 			error = EFAULT;
1049 			goto doerror;
1050 		}
1051 	}
1052 
1053 	/* Bring buffer into kernel space. */
1054 	vmapbuf(bp);
1055 
1056 	s = splbio();
1057 	aiocbe->bp = bp;
1058 	bp->b_spc = (void *)aiocbe;
1059 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1060 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1061 	aiocbe->jobstate = JOBST_JOBQBUF;
1062 	cb->_aiocb_private.status = cb->aio_nbytes;
1063 	num_buf_aio++;
1064 	bp->b_error = 0;
1065 
1066 	splx(s);
1067 
1068 	/* Perform transfer. */
1069 	DEV_STRATEGY(bp, 0);
1070 
1071 	notify = 0;
1072 	s = splbio();
1073 
1074 	/*
1075 	 * If we had an error invoking the request, or an error in processing
1076 	 * the request before we have returned, we process it as an error in
1077 	 * transfer.  Note that such an I/O error is not indicated immediately,
1078 	 * but is returned using the aio_error mechanism.  In this case,
1079 	 * aio_suspend will return immediately.
1080 	 */
1081 	if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1082 		struct aiocb *job = aiocbe->uuaiocb;
1083 
1084 		aiocbe->uaiocb._aiocb_private.status = 0;
1085 		suword(&job->_aiocb_private.status, 0);
1086 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1087 		suword(&job->_aiocb_private.error, bp->b_error);
1088 
1089 		ki->kaio_buffer_finished_count++;
1090 
1091 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1092 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1093 			aiocbe->jobflags |= AIOCBLIST_DONE;
1094 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1095 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1096 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1097 			notify = 1;
1098 		}
1099 	}
1100 	splx(s);
1101 	if (notify)
1102 		KNOTE(&aiocbe->klist, 0);
1103 	return 0;
1104 
1105 doerror:
1106 	ki->kaio_buffer_count--;
1107 	if (lj)
1108 		lj->lioj_buffer_count--;
1109 	aiocbe->bp = NULL;
1110 	relpbuf(bp, NULL);
1111 	return error;
1112 }
1113 
1114 /*
1115  * This waits/tests physio completion.
1116  */
1117 static int
1118 aio_fphysio(struct proc *p, struct aiocblist *iocb)
1119 {
1120 	int s;
1121 	struct buf *bp;
1122 	int error;
1123 
1124 	bp = iocb->bp;
1125 
1126 	s = splbio();
1127 	while ((bp->b_flags & B_DONE) == 0) {
1128 		if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1129 			if ((bp->b_flags & B_DONE) == 0) {
1130 				splx(s);
1131 				return EINPROGRESS;
1132 			} else
1133 				break;
1134 		}
1135 	}
1136 	splx(s);
1137 
1138 	/* Release mapping into kernel space. */
1139 	vunmapbuf(bp);
1140 	iocb->bp = 0;
1141 
1142 	error = 0;
1143 
1144 	/* Check for an error. */
1145 	if (bp->b_ioflags & BIO_ERROR)
1146 		error = bp->b_error;
1147 
1148 	relpbuf(bp, NULL);
1149 	return (error);
1150 }
1151 #endif /* VFS_AIO */
1152 
1153 /*
1154  * Wake up aio requests that may be serviceable now.
1155  */
1156 void
1157 aio_swake(struct socket *so, struct sockbuf *sb)
1158 {
1159 #ifndef VFS_AIO
1160 	return;
1161 #else
1162 	struct aiocblist *cb,*cbn;
1163 	struct proc *p;
1164 	struct kaioinfo *ki = NULL;
1165 	int opcode, wakecount = 0;
1166 	struct aioproclist *aiop;
1167 
1168 	if (sb == &so->so_snd) {
1169 		opcode = LIO_WRITE;
1170 		so->so_snd.sb_flags &= ~SB_AIO;
1171 	} else {
1172 		opcode = LIO_READ;
1173 		so->so_rcv.sb_flags &= ~SB_AIO;
1174 	}
1175 
1176 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1177 		cbn = TAILQ_NEXT(cb, list);
1178 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1179 			p = cb->userproc;
1180 			ki = p->p_aioinfo;
1181 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1182 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1183 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1184 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1185 			wakecount++;
1186 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1187 				panic("invalid queue value");
1188 		}
1189 	}
1190 
1191 	while (wakecount--) {
1192 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1193 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1194 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1195 			aiop->aioprocflags &= ~AIOP_FREE;
1196 			wakeup(aiop->aioproc);
1197 		}
1198 	}
1199 #endif /* VFS_AIO */
1200 }
1201 
1202 #ifdef VFS_AIO
1203 /*
1204  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1205  * technique is done in this code.
1206  */
1207 static int
1208 _aio_aqueue(struct proc *p, struct aiocb *job, struct aio_liojob *lj, int type)
1209 {
1210 	struct filedesc *fdp;
1211 	struct file *fp;
1212 	unsigned int fd;
1213 	struct socket *so;
1214 	int s;
1215 	int error;
1216 	int opcode;
1217 	struct aiocblist *aiocbe;
1218 	struct aioproclist *aiop;
1219 	struct kaioinfo *ki;
1220 	struct kevent kev;
1221 	struct kqueue *kq;
1222 	struct file *kq_fp;
1223 
1224 	aiocbe = zalloc(aiocb_zone);
1225 	aiocbe->inputcharge = 0;
1226 	aiocbe->outputcharge = 0;
1227 	callout_handle_init(&aiocbe->timeouthandle);
1228 	SLIST_INIT(&aiocbe->klist);
1229 
1230 	suword(&job->_aiocb_private.status, -1);
1231 	suword(&job->_aiocb_private.error, 0);
1232 	suword(&job->_aiocb_private.kernelinfo, -1);
1233 
1234 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1235 	if (error) {
1236 		suword(&job->_aiocb_private.error, error);
1237 		zfree(aiocb_zone, aiocbe);
1238 		return error;
1239 	}
1240 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1241 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1242 		zfree(aiocb_zone, aiocbe);
1243 		return EINVAL;
1244 	}
1245 
1246 	/* Save userspace address of the job info. */
1247 	aiocbe->uuaiocb = job;
1248 
1249 	/* Get the opcode. */
1250 	if (type != LIO_NOP)
1251 		aiocbe->uaiocb.aio_lio_opcode = type;
1252 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1253 
1254 	/* Get the fd info for process. */
1255 	fdp = p->p_fd;
1256 
1257 	/*
1258 	 * Range check file descriptor.
1259 	 */
1260 	fd = aiocbe->uaiocb.aio_fildes;
1261 	if (fd >= fdp->fd_nfiles) {
1262 		zfree(aiocb_zone, aiocbe);
1263 		if (type == 0)
1264 			suword(&job->_aiocb_private.error, EBADF);
1265 		return EBADF;
1266 	}
1267 
1268 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1269 	if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1270 	    0))) {
1271 		zfree(aiocb_zone, aiocbe);
1272 		if (type == 0)
1273 			suword(&job->_aiocb_private.error, EBADF);
1274 		return EBADF;
1275 	}
1276 
1277 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1278 		zfree(aiocb_zone, aiocbe);
1279 		if (type == 0)
1280 			suword(&job->_aiocb_private.error, EINVAL);
1281 		return EINVAL;
1282 	}
1283 
1284 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1285 	if (error) {
1286 		zfree(aiocb_zone, aiocbe);
1287 		if (type == 0)
1288 			suword(&job->_aiocb_private.error, EINVAL);
1289 		return error;
1290 	}
1291 
1292 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1293 	if (jobrefid == LONG_MAX)
1294 		jobrefid = 1;
1295 	else
1296 		jobrefid++;
1297 
1298 	if (opcode == LIO_NOP) {
1299 		zfree(aiocb_zone, aiocbe);
1300 		if (type == 0) {
1301 			suword(&job->_aiocb_private.error, 0);
1302 			suword(&job->_aiocb_private.status, 0);
1303 			suword(&job->_aiocb_private.kernelinfo, 0);
1304 		}
1305 		return 0;
1306 	}
1307 
1308 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1309 		zfree(aiocb_zone, aiocbe);
1310 		if (type == 0) {
1311 			suword(&job->_aiocb_private.status, 0);
1312 			suword(&job->_aiocb_private.error, EINVAL);
1313 		}
1314 		return EINVAL;
1315 	}
1316 
1317 	fhold(fp);
1318 
1319 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1320 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1321 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1322 	}
1323 	else {
1324 		/*
1325 		 * This method for requesting kevent-based notification won't
1326 		 * work on the alpha, since we're passing in a pointer
1327 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1328 		 * based method instead.
1329 		 */
1330 		struct kevent *kevp;
1331 
1332 		kevp = (struct kevent *)job->aio_lio_opcode;
1333 		if (kevp == NULL)
1334 			goto no_kqueue;
1335 
1336 		error = copyin(kevp, &kev, sizeof(kev));
1337 		if (error)
1338 			goto aqueue_fail;
1339 	}
1340 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1341 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1342 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1343 		error = EBADF;
1344 		goto aqueue_fail;
1345 	}
1346 	kq = (struct kqueue *)kq_fp->f_data;
1347 	kev.ident = (uintptr_t)aiocbe;
1348 	kev.filter = EVFILT_AIO;
1349 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1350 	error = kqueue_register(kq, &kev, p);
1351 aqueue_fail:
1352 	if (error) {
1353 		zfree(aiocb_zone, aiocbe);
1354 		if (type == 0)
1355 			suword(&job->_aiocb_private.error, error);
1356 		goto done;
1357 	}
1358 no_kqueue:
1359 
1360 	suword(&job->_aiocb_private.error, EINPROGRESS);
1361 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1362 	aiocbe->userproc = p;
1363 	aiocbe->jobflags = 0;
1364 	aiocbe->lio = lj;
1365 	ki = p->p_aioinfo;
1366 
1367 	if (fp->f_type == DTYPE_SOCKET) {
1368 		/*
1369 		 * Alternate queueing for socket ops: Reach down into the
1370 		 * descriptor to get the socket data.  Then check to see if the
1371 		 * socket is ready to be read or written (based on the requested
1372 		 * operation).
1373 		 *
1374 		 * If it is not ready for io, then queue the aiocbe on the
1375 		 * socket, and set the flags so we get a call when sbnotify()
1376 		 * happens.
1377 		 */
1378 		so = (struct socket *)fp->f_data;
1379 		s = splnet();
1380 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1381 		    LIO_WRITE) && (!sowriteable(so)))) {
1382 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1383 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1384 			if (opcode == LIO_READ)
1385 				so->so_rcv.sb_flags |= SB_AIO;
1386 			else
1387 				so->so_snd.sb_flags |= SB_AIO;
1388 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1389 			ki->kaio_queue_count++;
1390 			num_queue_count++;
1391 			splx(s);
1392 			error = 0;
1393 			goto done;
1394 		}
1395 		splx(s);
1396 	}
1397 
1398 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1399 		goto done;
1400 	if (error > 0) {
1401 		suword(&job->_aiocb_private.status, 0);
1402 		aiocbe->uaiocb._aiocb_private.error = error;
1403 		suword(&job->_aiocb_private.error, error);
1404 		goto done;
1405 	}
1406 
1407 	/* No buffer for daemon I/O. */
1408 	aiocbe->bp = NULL;
1409 
1410 	ki->kaio_queue_count++;
1411 	if (lj)
1412 		lj->lioj_queue_count++;
1413 	s = splnet();
1414 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1415 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1416 	splx(s);
1417 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1418 
1419 	num_queue_count++;
1420 	error = 0;
1421 
1422 	/*
1423 	 * If we don't have a free AIO process, and we are below our quota, then
1424 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1425 	 * pick-up this job.  If we don't sucessfully create the new process
1426 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1427 	 * which is likely not the correct thing to do.
1428 	 */
1429 retryproc:
1430 	s = splnet();
1431 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1432 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1433 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1434 		aiop->aioprocflags &= ~AIOP_FREE;
1435 		wakeup(aiop->aioproc);
1436 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1437 	    ((ki->kaio_active_count + num_aio_resv_start) <
1438 	    ki->kaio_maxactive_count)) {
1439 		num_aio_resv_start++;
1440 		if ((error = aio_newproc()) == 0) {
1441 			num_aio_resv_start--;
1442 			p->p_retval[0] = 0;
1443 			goto retryproc;
1444 		}
1445 		num_aio_resv_start--;
1446 	}
1447 	splx(s);
1448 done:
1449 	fdrop(fp, p);
1450 	return error;
1451 }
1452 
1453 /*
1454  * This routine queues an AIO request, checking for quotas.
1455  */
1456 static int
1457 aio_aqueue(struct proc *p, struct aiocb *job, int type)
1458 {
1459 	struct kaioinfo *ki;
1460 
1461 	if (p->p_aioinfo == NULL)
1462 		aio_init_aioinfo(p);
1463 
1464 	if (num_queue_count >= max_queue_count)
1465 		return EAGAIN;
1466 
1467 	ki = p->p_aioinfo;
1468 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1469 		return EAGAIN;
1470 
1471 	return _aio_aqueue(p, job, NULL, type);
1472 }
1473 #endif /* VFS_AIO */
1474 
1475 /*
1476  * Support the aio_return system call, as a side-effect, kernel resources are
1477  * released.
1478  */
1479 int
1480 aio_return(struct proc *p, struct aio_return_args *uap)
1481 {
1482 #ifndef VFS_AIO
1483 	return ENOSYS;
1484 #else
1485 	int s;
1486 	int jobref;
1487 	struct aiocblist *cb, *ncb;
1488 	struct aiocb *ujob;
1489 	struct kaioinfo *ki;
1490 
1491 	ki = p->p_aioinfo;
1492 	if (ki == NULL)
1493 		return EINVAL;
1494 
1495 	ujob = uap->aiocbp;
1496 
1497 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1498 	if (jobref == -1 || jobref == 0)
1499 		return EINVAL;
1500 
1501 	s = splnet();
1502 	for (cb = TAILQ_FIRST(&ki->kaio_jobdone); cb; cb = TAILQ_NEXT(cb,
1503 	    plist)) {
1504 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1505 		    jobref) {
1506 			splx(s);
1507 			if (ujob == cb->uuaiocb) {
1508 				p->p_retval[0] =
1509 				    cb->uaiocb._aiocb_private.status;
1510 			} else
1511 				p->p_retval[0] = EFAULT;
1512 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1513 				curproc->p_stats->p_ru.ru_oublock +=
1514 				    cb->outputcharge;
1515 				cb->outputcharge = 0;
1516 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1517 				curproc->p_stats->p_ru.ru_inblock +=
1518 				    cb->inputcharge;
1519 				cb->inputcharge = 0;
1520 			}
1521 			aio_free_entry(cb);
1522 			return 0;
1523 		}
1524 	}
1525 	splx(s);
1526 
1527 	s = splbio();
1528 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1529 		ncb = TAILQ_NEXT(cb, plist);
1530 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1531 		    == jobref) {
1532 			splx(s);
1533 			if (ujob == cb->uuaiocb) {
1534 				p->p_retval[0] =
1535 				    cb->uaiocb._aiocb_private.status;
1536 			} else
1537 				p->p_retval[0] = EFAULT;
1538 			aio_free_entry(cb);
1539 			return 0;
1540 		}
1541 	}
1542 	splx(s);
1543 
1544 	return (EINVAL);
1545 #endif /* VFS_AIO */
1546 }
1547 
1548 /*
1549  * Allow a process to wakeup when any of the I/O requests are completed.
1550  */
1551 int
1552 aio_suspend(struct proc *p, struct aio_suspend_args *uap)
1553 {
1554 #ifndef VFS_AIO
1555 	return ENOSYS;
1556 #else
1557 	struct timeval atv;
1558 	struct timespec ts;
1559 	struct aiocb *const *cbptr, *cbp;
1560 	struct kaioinfo *ki;
1561 	struct aiocblist *cb;
1562 	int i;
1563 	int njoblist;
1564 	int error, s, timo;
1565 	int *ijoblist;
1566 	struct aiocb **ujoblist;
1567 
1568 	if (uap->nent >= AIO_LISTIO_MAX)
1569 		return EINVAL;
1570 
1571 	timo = 0;
1572 	if (uap->timeout) {
1573 		/* Get timespec struct. */
1574 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1575 			return error;
1576 
1577 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1578 			return (EINVAL);
1579 
1580 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1581 		if (itimerfix(&atv))
1582 			return (EINVAL);
1583 		timo = tvtohz(&atv);
1584 	}
1585 
1586 	ki = p->p_aioinfo;
1587 	if (ki == NULL)
1588 		return EAGAIN;
1589 
1590 	njoblist = 0;
1591 	ijoblist = zalloc(aiol_zone);
1592 	ujoblist = zalloc(aiol_zone);
1593 	cbptr = uap->aiocbp;
1594 
1595 	for (i = 0; i < uap->nent; i++) {
1596 		cbp = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
1597 		if (cbp == 0)
1598 			continue;
1599 		ujoblist[njoblist] = cbp;
1600 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1601 		njoblist++;
1602 	}
1603 
1604 	if (njoblist == 0) {
1605 		zfree(aiol_zone, ijoblist);
1606 		zfree(aiol_zone, ujoblist);
1607 		return 0;
1608 	}
1609 
1610 	error = 0;
1611 	for (;;) {
1612 		for (cb = TAILQ_FIRST(&ki->kaio_jobdone); cb; cb =
1613 		    TAILQ_NEXT(cb, plist)) {
1614 			for (i = 0; i < njoblist; i++) {
1615 				if (((intptr_t)
1616 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1617 				    ijoblist[i]) {
1618 					if (ujoblist[i] != cb->uuaiocb)
1619 						error = EINVAL;
1620 					zfree(aiol_zone, ijoblist);
1621 					zfree(aiol_zone, ujoblist);
1622 					return error;
1623 				}
1624 			}
1625 		}
1626 
1627 		s = splbio();
1628 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1629 		    TAILQ_NEXT(cb, plist)) {
1630 			for (i = 0; i < njoblist; i++) {
1631 				if (((intptr_t)
1632 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1633 				    ijoblist[i]) {
1634 					splx(s);
1635 					if (ujoblist[i] != cb->uuaiocb)
1636 						error = EINVAL;
1637 					zfree(aiol_zone, ijoblist);
1638 					zfree(aiol_zone, ujoblist);
1639 					return error;
1640 				}
1641 			}
1642 		}
1643 
1644 		ki->kaio_flags |= KAIO_WAKEUP;
1645 		error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1646 		splx(s);
1647 
1648 		if (error == ERESTART || error == EINTR) {
1649 			zfree(aiol_zone, ijoblist);
1650 			zfree(aiol_zone, ujoblist);
1651 			return EINTR;
1652 		} else if (error == EWOULDBLOCK) {
1653 			zfree(aiol_zone, ijoblist);
1654 			zfree(aiol_zone, ujoblist);
1655 			return EAGAIN;
1656 		}
1657 	}
1658 
1659 /* NOTREACHED */
1660 	return EINVAL;
1661 #endif /* VFS_AIO */
1662 }
1663 
1664 /*
1665  * aio_cancel cancels any non-physio aio operations not currently in
1666  * progress.
1667  */
1668 int
1669 aio_cancel(struct proc *p, struct aio_cancel_args *uap)
1670 {
1671 #ifndef VFS_AIO
1672 	return ENOSYS;
1673 #else
1674 	struct kaioinfo *ki;
1675 	struct aiocblist *cbe, *cbn;
1676 	struct file *fp;
1677 	struct filedesc *fdp;
1678 	struct socket *so;
1679 	struct proc *po;
1680 	int s,error;
1681 	int cancelled=0;
1682 	int notcancelled=0;
1683 	struct vnode *vp;
1684 
1685 	fdp = p->p_fd;
1686 
1687 	fp = fdp->fd_ofiles[uap->fd];
1688 
1689 	if (fp == NULL) {
1690 		return EBADF;
1691 	}
1692 
1693         if (fp->f_type == DTYPE_VNODE) {
1694 		vp = (struct vnode *)fp->f_data;
1695 
1696 		if (vn_isdisk(vp,&error)) {
1697 			p->p_retval[0] = AIO_NOTCANCELED;
1698         	        return 0;
1699 		}
1700 	} else if (fp->f_type == DTYPE_SOCKET) {
1701 		so = (struct socket *)fp->f_data;
1702 
1703 		s = splnet();
1704 
1705 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1706 			cbn = TAILQ_NEXT(cbe, list);
1707 			if ((uap->aiocbp == NULL) ||
1708 				(uap->aiocbp == cbe->uuaiocb) ) {
1709 				po = cbe->userproc;
1710 				ki = po->p_aioinfo;
1711 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1712 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1713 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1714 				if (ki->kaio_flags & KAIO_WAKEUP) {
1715 					wakeup(po);
1716 				}
1717 				cbe->jobstate = JOBST_JOBFINISHED;
1718 				cbe->uaiocb._aiocb_private.status=-1;
1719 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1720 				cancelled++;
1721 /* XXX cancelled, knote? */
1722 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1723 				    SIGEV_SIGNAL) {
1724 					PROC_LOCK(cbe->userproc);
1725 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1726 					PROC_UNLOCK(cbe->userproc);
1727 				}
1728 				if (uap->aiocbp)
1729 					break;
1730 			}
1731 		}
1732 
1733 		splx(s);
1734 
1735 		if ((cancelled) && (uap->aiocbp)) {
1736 			p->p_retval[0] = AIO_CANCELED;
1737 			return 0;
1738 		}
1739 
1740 	}
1741 
1742 	ki=p->p_aioinfo;
1743 
1744 	s = splnet();
1745 
1746 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1747 		cbn = TAILQ_NEXT(cbe, plist);
1748 
1749 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1750 		    ((uap->aiocbp == NULL ) ||
1751 		     (uap->aiocbp == cbe->uuaiocb))) {
1752 
1753 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1754 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1755                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1756                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1757                                     plist);
1758 				cancelled++;
1759 				ki->kaio_queue_finished_count++;
1760 				cbe->jobstate = JOBST_JOBFINISHED;
1761 				cbe->uaiocb._aiocb_private.status = -1;
1762 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1763 /* XXX cancelled, knote? */
1764 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1765 				    SIGEV_SIGNAL) {
1766 					PROC_LOCK(cbe->userproc);
1767 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1768 					PROC_UNLOCK(cbe->userproc);
1769 				}
1770 			} else {
1771 				notcancelled++;
1772 			}
1773 		}
1774 	}
1775 
1776 	splx(s);
1777 
1778 
1779 	if (notcancelled) {
1780 		p->p_retval[0] = AIO_NOTCANCELED;
1781 		return 0;
1782 	}
1783 
1784 	if (cancelled) {
1785 		p->p_retval[0] = AIO_CANCELED;
1786 		return 0;
1787 	}
1788 
1789 	p->p_retval[0] = AIO_ALLDONE;
1790 
1791 	return 0;
1792 #endif /* VFS_AIO */
1793 }
1794 
1795 /*
1796  * aio_error is implemented in the kernel level for compatibility purposes only.
1797  * For a user mode async implementation, it would be best to do it in a userland
1798  * subroutine.
1799  */
1800 int
1801 aio_error(struct proc *p, struct aio_error_args *uap)
1802 {
1803 #ifndef VFS_AIO
1804 	return ENOSYS;
1805 #else
1806 	int s;
1807 	struct aiocblist *cb;
1808 	struct kaioinfo *ki;
1809 	int jobref;
1810 
1811 	ki = p->p_aioinfo;
1812 	if (ki == NULL)
1813 		return EINVAL;
1814 
1815 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1816 	if ((jobref == -1) || (jobref == 0))
1817 		return EINVAL;
1818 
1819 	for (cb = TAILQ_FIRST(&ki->kaio_jobdone); cb; cb = TAILQ_NEXT(cb,
1820 	    plist)) {
1821 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1822 		    jobref) {
1823 			p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1824 			return 0;
1825 		}
1826 	}
1827 
1828 	s = splnet();
1829 
1830 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1831 	    plist)) {
1832 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1833 		    jobref) {
1834 			p->p_retval[0] = EINPROGRESS;
1835 			splx(s);
1836 			return 0;
1837 		}
1838 	}
1839 
1840 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1841 	    plist)) {
1842 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1843 		    jobref) {
1844 			p->p_retval[0] = EINPROGRESS;
1845 			splx(s);
1846 			return 0;
1847 		}
1848 	}
1849 	splx(s);
1850 
1851 	s = splbio();
1852 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1853 	    plist)) {
1854 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1855 		    jobref) {
1856 			p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1857 			splx(s);
1858 			return 0;
1859 		}
1860 	}
1861 
1862 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1863 	    plist)) {
1864 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1865 		    jobref) {
1866 			p->p_retval[0] = EINPROGRESS;
1867 			splx(s);
1868 			return 0;
1869 		}
1870 	}
1871 	splx(s);
1872 
1873 #if (0)
1874 	/*
1875 	 * Hack for lio.
1876 	 */
1877 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1878 	if (status == -1)
1879 		return fuword(&uap->aiocbp->_aiocb_private.error);
1880 #endif
1881 	return EINVAL;
1882 #endif /* VFS_AIO */
1883 }
1884 
1885 int
1886 aio_read(struct proc *p, struct aio_read_args *uap)
1887 {
1888 #ifndef VFS_AIO
1889 	return ENOSYS;
1890 #else
1891 	return aio_aqueue(p, uap->aiocbp, LIO_READ);
1892 #endif /* VFS_AIO */
1893 }
1894 
1895 int
1896 aio_write(struct proc *p, struct aio_write_args *uap)
1897 {
1898 #ifndef VFS_AIO
1899 	return ENOSYS;
1900 #else
1901 	return aio_aqueue(p, uap->aiocbp, LIO_WRITE);
1902 #endif /* VFS_AIO */
1903 }
1904 
1905 int
1906 lio_listio(struct proc *p, struct lio_listio_args *uap)
1907 {
1908 #ifndef VFS_AIO
1909 	return ENOSYS;
1910 #else
1911 	int nent, nentqueued;
1912 	struct aiocb *iocb, * const *cbptr;
1913 	struct aiocblist *cb;
1914 	struct kaioinfo *ki;
1915 	struct aio_liojob *lj;
1916 	int error, runningcode;
1917 	int nerror;
1918 	int i;
1919 	int s;
1920 
1921 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1922 		return EINVAL;
1923 
1924 	nent = uap->nent;
1925 	if (nent > AIO_LISTIO_MAX)
1926 		return EINVAL;
1927 
1928 	if (p->p_aioinfo == NULL)
1929 		aio_init_aioinfo(p);
1930 
1931 	if ((nent + num_queue_count) > max_queue_count)
1932 		return EAGAIN;
1933 
1934 	ki = p->p_aioinfo;
1935 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1936 		return EAGAIN;
1937 
1938 	lj = zalloc(aiolio_zone);
1939 	if (!lj)
1940 		return EAGAIN;
1941 
1942 	lj->lioj_flags = 0;
1943 	lj->lioj_buffer_count = 0;
1944 	lj->lioj_buffer_finished_count = 0;
1945 	lj->lioj_queue_count = 0;
1946 	lj->lioj_queue_finished_count = 0;
1947 	lj->lioj_ki = ki;
1948 
1949 	/*
1950 	 * Setup signal.
1951 	 */
1952 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1953 		error = copyin(uap->sig, &lj->lioj_signal,
1954 			       sizeof(lj->lioj_signal));
1955 		if (error) {
1956 			zfree(aiolio_zone, lj);
1957 			return error;
1958 		}
1959 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1960 			zfree(aiolio_zone, lj);
1961 			return EINVAL;
1962 		}
1963 		lj->lioj_flags |= LIOJ_SIGNAL;
1964 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1965 	} else
1966 		lj->lioj_flags &= ~LIOJ_SIGNAL;
1967 
1968 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1969 	/*
1970 	 * Get pointers to the list of I/O requests.
1971 	 */
1972 	nerror = 0;
1973 	nentqueued = 0;
1974 	cbptr = uap->acb_list;
1975 	for (i = 0; i < uap->nent; i++) {
1976 		iocb = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
1977 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != NULL)) {
1978 			error = _aio_aqueue(p, iocb, lj, 0);
1979 			if (error == 0)
1980 				nentqueued++;
1981 			else
1982 				nerror++;
1983 		}
1984 	}
1985 
1986 	/*
1987 	 * If we haven't queued any, then just return error.
1988 	 */
1989 	if (nentqueued == 0)
1990 		return 0;
1991 
1992 	/*
1993 	 * Calculate the appropriate error return.
1994 	 */
1995 	runningcode = 0;
1996 	if (nerror)
1997 		runningcode = EIO;
1998 
1999 	if (uap->mode == LIO_WAIT) {
2000 		int command, found, jobref;
2001 
2002 		for (;;) {
2003 			found = 0;
2004 			for (i = 0; i < uap->nent; i++) {
2005 				/*
2006 				 * Fetch address of the control buf pointer in
2007 				 * user space.
2008 				 */
2009 				iocb = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
2010 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2011 				    == 0))
2012 					continue;
2013 
2014 				/*
2015 				 * Fetch the associated command from user space.
2016 				 */
2017 				command = fuword(&iocb->aio_lio_opcode);
2018 				if (command == LIO_NOP) {
2019 					found++;
2020 					continue;
2021 				}
2022 
2023 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
2024 
2025 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2026 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2027 					    == jobref) {
2028 						if (cb->uaiocb.aio_lio_opcode
2029 						    == LIO_WRITE) {
2030 							curproc->p_stats->p_ru.ru_oublock
2031 							    +=
2032 							    cb->outputcharge;
2033 							cb->outputcharge = 0;
2034 						} else if (cb->uaiocb.aio_lio_opcode
2035 						    == LIO_READ) {
2036 							curproc->p_stats->p_ru.ru_inblock
2037 							    += cb->inputcharge;
2038 							cb->inputcharge = 0;
2039 						}
2040 						found++;
2041 						break;
2042 					}
2043 				}
2044 
2045 				s = splbio();
2046 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2047 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2048 					    == jobref) {
2049 						found++;
2050 						break;
2051 					}
2052 				}
2053 				splx(s);
2054 			}
2055 
2056 			/*
2057 			 * If all I/Os have been disposed of, then we can
2058 			 * return.
2059 			 */
2060 			if (found == nentqueued)
2061 				return runningcode;
2062 
2063 			ki->kaio_flags |= KAIO_WAKEUP;
2064 			error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2065 
2066 			if (error == EINTR)
2067 				return EINTR;
2068 			else if (error == EWOULDBLOCK)
2069 				return EAGAIN;
2070 		}
2071 	}
2072 
2073 	return runningcode;
2074 #endif /* VFS_AIO */
2075 }
2076 
2077 #ifdef VFS_AIO
2078 /*
2079  * This is a weird hack so that we can post a signal.  It is safe to do so from
2080  * a timeout routine, but *not* from an interrupt routine.
2081  */
2082 static void
2083 process_signal(void *aioj)
2084 {
2085 	struct aiocblist *aiocbe = aioj;
2086 	struct aio_liojob *lj = aiocbe->lio;
2087 	struct aiocb *cb = &aiocbe->uaiocb;
2088 
2089 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2090 		(lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2091 		PROC_LOCK(lj->lioj_ki->kaio_p);
2092 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2093 		PROC_UNLOCK(lj->lioj_ki->kaio_p);
2094 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2095 	}
2096 
2097 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2098 		PROC_LOCK(aiocbe->userproc);
2099 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2100 		PROC_UNLOCK(aiocbe->userproc);
2101 	}
2102 }
2103 
2104 /*
2105  * Interrupt handler for physio, performs the necessary process wakeups, and
2106  * signals.
2107  */
2108 static void
2109 aio_physwakeup(struct buf *bp)
2110 {
2111 	struct aiocblist *aiocbe;
2112 	struct proc *p;
2113 	struct kaioinfo *ki;
2114 	struct aio_liojob *lj;
2115 
2116 	wakeup(bp);
2117 
2118 	aiocbe = (struct aiocblist *)bp->b_spc;
2119 	if (aiocbe) {
2120 		p = bp->b_caller1;
2121 
2122 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2123 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2124 		aiocbe->uaiocb._aiocb_private.error = 0;
2125 		aiocbe->jobflags |= AIOCBLIST_DONE;
2126 
2127 		if (bp->b_ioflags & BIO_ERROR)
2128 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2129 
2130 		lj = aiocbe->lio;
2131 		if (lj) {
2132 			lj->lioj_buffer_finished_count++;
2133 
2134 			/*
2135 			 * wakeup/signal if all of the interrupt jobs are done.
2136 			 */
2137 			if (lj->lioj_buffer_finished_count ==
2138 			    lj->lioj_buffer_count) {
2139 				/*
2140 				 * Post a signal if it is called for.
2141 				 */
2142 				if ((lj->lioj_flags &
2143 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2144 				    LIOJ_SIGNAL) {
2145 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2146 					aiocbe->timeouthandle =
2147 						timeout(process_signal,
2148 							aiocbe, 0);
2149 				}
2150 			}
2151 		}
2152 
2153 		ki = p->p_aioinfo;
2154 		if (ki) {
2155 			ki->kaio_buffer_finished_count++;
2156 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2157 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2158 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2159 
2160 			KNOTE(&aiocbe->klist, 0);
2161 			/* Do the wakeup. */
2162 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2163 				ki->kaio_flags &= ~KAIO_WAKEUP;
2164 				wakeup(p);
2165 			}
2166 		}
2167 
2168 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2169 			aiocbe->timeouthandle =
2170 				timeout(process_signal, aiocbe, 0);
2171 	}
2172 }
2173 #endif /* VFS_AIO */
2174 
2175 int
2176 aio_waitcomplete(struct proc *p, struct aio_waitcomplete_args *uap)
2177 {
2178 #ifndef VFS_AIO
2179 	return ENOSYS;
2180 #else
2181 	struct timeval atv;
2182 	struct timespec ts;
2183 	struct aiocb **cbptr;
2184 	struct kaioinfo *ki;
2185 	struct aiocblist *cb = NULL;
2186 	int error, s, timo;
2187 
2188 	suword(uap->aiocbp, (int)NULL);
2189 
2190 	timo = 0;
2191 	if (uap->timeout) {
2192 		/* Get timespec struct. */
2193 		error = copyin(uap->timeout, &ts, sizeof(ts));
2194 		if (error)
2195 			return error;
2196 
2197 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2198 			return (EINVAL);
2199 
2200 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2201 		if (itimerfix(&atv))
2202 			return (EINVAL);
2203 		timo = tvtohz(&atv);
2204 	}
2205 
2206 	ki = p->p_aioinfo;
2207 	if (ki == NULL)
2208 		return EAGAIN;
2209 
2210 	cbptr = uap->aiocbp;
2211 
2212 	for (;;) {
2213 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2214 			suword(uap->aiocbp, (int)cb->uuaiocb);
2215 			p->p_retval[0] = cb->uaiocb._aiocb_private.status;
2216 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2217 				curproc->p_stats->p_ru.ru_oublock +=
2218 				    cb->outputcharge;
2219 				cb->outputcharge = 0;
2220 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2221 				curproc->p_stats->p_ru.ru_inblock +=
2222 				    cb->inputcharge;
2223 				cb->inputcharge = 0;
2224 			}
2225 			aio_free_entry(cb);
2226 			return cb->uaiocb._aiocb_private.error;
2227 		}
2228 
2229 		s = splbio();
2230  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2231 			splx(s);
2232 			suword(uap->aiocbp, (int)cb->uuaiocb);
2233 			p->p_retval[0] = cb->uaiocb._aiocb_private.status;
2234 			aio_free_entry(cb);
2235 			return cb->uaiocb._aiocb_private.error;
2236 		}
2237 
2238 		ki->kaio_flags |= KAIO_WAKEUP;
2239 		error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2240 		splx(s);
2241 
2242 		if (error == ERESTART)
2243 			return EINTR;
2244 		else if (error < 0)
2245 			return error;
2246 		else if (error == EINTR)
2247 			return EINTR;
2248 		else if (error == EWOULDBLOCK)
2249 			return EAGAIN;
2250 	}
2251 #endif /* VFS_AIO */
2252 }
2253 
2254 
2255 #ifndef VFS_AIO
2256 static int
2257 filt_aioattach(struct knote *kn)
2258 {
2259 
2260 	return (ENXIO);
2261 }
2262 
2263 struct filterops aio_filtops =
2264 	{ 0, filt_aioattach, NULL, NULL };
2265 
2266 #else
2267 static int
2268 filt_aioattach(struct knote *kn)
2269 {
2270 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2271 
2272 	/*
2273 	 * The aiocbe pointer must be validated before using it, so
2274 	 * registration is restricted to the kernel; the user cannot
2275 	 * set EV_FLAG1.
2276 	 */
2277 	if ((kn->kn_flags & EV_FLAG1) == 0)
2278 		return (EPERM);
2279 	kn->kn_flags &= ~EV_FLAG1;
2280 
2281 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2282 
2283 	return (0);
2284 }
2285 
2286 static void
2287 filt_aiodetach(struct knote *kn)
2288 {
2289 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2290 	int s = splhigh();	 /* XXX no clue, so overkill */
2291 
2292 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2293 	splx(s);
2294 }
2295 
2296 /*ARGSUSED*/
2297 static int
2298 filt_aio(struct knote *kn, long hint)
2299 {
2300 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2301 
2302 	kn->kn_data = 0;		/* XXX data returned? */
2303 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2304 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2305 		return (0);
2306 	kn->kn_flags |= EV_EOF;
2307 	return (1);
2308 }
2309 
2310 struct filterops aio_filtops =
2311 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
2312 #endif /* VFS_AIO */
2313