xref: /dflybsd-src/sys/kern/vfs_aio.c (revision 90ea502b8c5d21f908cedff6680ee2bc9e74ce74)
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: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $
17  * $DragonFly: src/sys/kern/vfs_aio.c,v 1.42 2007/07/20 17:21:52 dillon Exp $
18  */
19 
20 /*
21  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
22  */
23 
24 #include <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/buf.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/fcntl.h>
31 #include <sys/file.h>
32 #include <sys/lock.h>
33 #include <sys/unistd.h>
34 #include <sys/proc.h>
35 #include <sys/resourcevar.h>
36 #include <sys/signalvar.h>
37 #include <sys/protosw.h>
38 #include <sys/socketvar.h>
39 #include <sys/sysctl.h>
40 #include <sys/vnode.h>
41 #include <sys/conf.h>
42 #include <sys/event.h>
43 
44 #include <vm/vm.h>
45 #include <vm/vm_extern.h>
46 #include <vm/pmap.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_zone.h>
49 #include <sys/aio.h>
50 #include <sys/file2.h>
51 #include <sys/buf2.h>
52 #include <sys/sysref2.h>
53 #include <sys/thread2.h>
54 
55 #include <machine/limits.h>
56 #include "opt_vfs_aio.h"
57 
58 #ifdef VFS_AIO
59 
60 /*
61  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
62  * overflow.
63  */
64 static	long jobrefid;
65 
66 #define JOBST_NULL		0x0
67 #define JOBST_JOBQGLOBAL	0x2
68 #define JOBST_JOBRUNNING	0x3
69 #define JOBST_JOBFINISHED	0x4
70 #define	JOBST_JOBQBUF		0x5
71 #define	JOBST_JOBBFINISHED	0x6
72 
73 #ifndef MAX_AIO_PER_PROC
74 #define MAX_AIO_PER_PROC	32
75 #endif
76 
77 #ifndef MAX_AIO_QUEUE_PER_PROC
78 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
79 #endif
80 
81 #ifndef MAX_AIO_PROCS
82 #define MAX_AIO_PROCS		32
83 #endif
84 
85 #ifndef MAX_AIO_QUEUE
86 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
87 #endif
88 
89 #ifndef TARGET_AIO_PROCS
90 #define TARGET_AIO_PROCS	4
91 #endif
92 
93 #ifndef MAX_BUF_AIO
94 #define MAX_BUF_AIO		16
95 #endif
96 
97 #ifndef AIOD_TIMEOUT_DEFAULT
98 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
99 #endif
100 
101 #ifndef AIOD_LIFETIME_DEFAULT
102 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
103 #endif
104 
105 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
106 
107 static int max_aio_procs = MAX_AIO_PROCS;
108 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
109 	CTLFLAG_RW, &max_aio_procs, 0,
110 	"Maximum number of kernel threads to use for handling async IO");
111 
112 static int num_aio_procs = 0;
113 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
114 	CTLFLAG_RD, &num_aio_procs, 0,
115 	"Number of presently active kernel threads for async IO");
116 
117 /*
118  * The code will adjust the actual number of AIO processes towards this
119  * number when it gets a chance.
120  */
121 static int target_aio_procs = TARGET_AIO_PROCS;
122 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
123 	0, "Preferred number of ready kernel threads for async IO");
124 
125 static int max_queue_count = MAX_AIO_QUEUE;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
127     "Maximum number of aio requests to queue, globally");
128 
129 static int num_queue_count = 0;
130 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
131     "Number of queued aio requests");
132 
133 static int num_buf_aio = 0;
134 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
135     "Number of aio requests presently handled by the buf subsystem");
136 
137 /* Number of async I/O thread in the process of being started */
138 /* XXX This should be local to _aio_aqueue() */
139 static int num_aio_resv_start = 0;
140 
141 static int aiod_timeout;
142 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
143     "Timeout value for synchronous aio operations");
144 
145 static int aiod_lifetime;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
147     "Maximum lifetime for idle aiod");
148 
149 static int max_aio_per_proc = MAX_AIO_PER_PROC;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
151     0, "Maximum active aio requests per process (stored in the process)");
152 
153 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
154 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
155     &max_aio_queue_per_proc, 0,
156     "Maximum queued aio requests per process (stored in the process)");
157 
158 static int max_buf_aio = MAX_BUF_AIO;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
160     "Maximum buf aio requests per process (stored in the process)");
161 
162 /*
163  * AIO process info
164  */
165 #define AIOP_FREE	0x1			/* proc on free queue */
166 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
167 
168 struct aioproclist {
169 	int aioprocflags;			/* AIO proc flags */
170 	TAILQ_ENTRY(aioproclist) list;		/* List of processes */
171 	struct proc *aioproc;			/* The AIO thread */
172 };
173 
174 /*
175  * data-structure for lio signal management
176  */
177 struct aio_liojob {
178 	int	lioj_flags;
179 	int	lioj_buffer_count;
180 	int	lioj_buffer_finished_count;
181 	int	lioj_queue_count;
182 	int	lioj_queue_finished_count;
183 	struct	sigevent lioj_signal;	/* signal on all I/O done */
184 	TAILQ_ENTRY(aio_liojob) lioj_list;
185 	struct	kaioinfo *lioj_ki;
186 };
187 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
188 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
189 
190 /*
191  * per process aio data structure
192  */
193 struct kaioinfo {
194 	int	kaio_flags;		/* per process kaio flags */
195 	int	kaio_maxactive_count;	/* maximum number of AIOs */
196 	int	kaio_active_count;	/* number of currently used AIOs */
197 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
198 	int	kaio_queue_count;	/* size of AIO queue */
199 	int	kaio_ballowed_count;	/* maximum number of buffers */
200 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
201 	int	kaio_buffer_count;	/* number of physio buffers */
202 	int	kaio_buffer_finished_count; /* count of I/O done */
203 	struct 	proc *kaio_p;		/* process that uses this kaio block */
204 	TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
205 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* job queue for process */
206 	TAILQ_HEAD(,aiocblist) kaio_jobdone;	/* done queue for process */
207 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* buffer job queue for process */
208 	TAILQ_HEAD(,aiocblist) kaio_bufdone;	/* buffer done queue for process */
209 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;	/* queue for aios waiting on sockets */
210 };
211 
212 #define KAIO_RUNDOWN	0x1	/* process is being run down */
213 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
214 
215 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
216 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
217 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
218 static TAILQ_HEAD(,aiocblist) aio_freejobs;		/* Pool of free jobs */
219 
220 static void	aio_init_aioinfo(struct proc *p);
221 static void	aio_onceonly(void *);
222 static int	aio_free_entry(struct aiocblist *aiocbe);
223 static void	aio_process(struct aiocblist *aiocbe);
224 static int	aio_newproc(void);
225 static int	aio_aqueue(struct aiocb *job, int type);
226 static void	aio_physwakeup(struct bio *bio);
227 static int	aio_fphysio(struct aiocblist *aiocbe);
228 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
229 static void	aio_daemon(void *uproc, struct trapframe *frame);
230 static void	process_signal(void *aioj);
231 
232 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
233 
234 /*
235  * Zones for:
236  * 	kaio	Per process async io info
237  *	aiop	async io thread data
238  *	aiocb	async io jobs
239  *	aiol	list io job pointer - internal to aio_suspend XXX
240  *	aiolio	list io jobs
241  */
242 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
243 
244 /*
245  * Startup initialization
246  */
247 static void
248 aio_onceonly(void *na)
249 {
250 	TAILQ_INIT(&aio_freeproc);
251 	TAILQ_INIT(&aio_activeproc);
252 	TAILQ_INIT(&aio_jobs);
253 	TAILQ_INIT(&aio_bufjobs);
254 	TAILQ_INIT(&aio_freejobs);
255 	kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
256 	aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
257 	aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
258 	aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
259 	aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
260 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
261 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
262 	jobrefid = 1;
263 }
264 
265 /*
266  * Init the per-process aioinfo structure.  The aioinfo limits are set
267  * per-process for user limit (resource) management.
268  */
269 static void
270 aio_init_aioinfo(struct proc *p)
271 {
272 	struct kaioinfo *ki;
273 	if (p->p_aioinfo == NULL) {
274 		ki = zalloc(kaio_zone);
275 		p->p_aioinfo = ki;
276 		ki->kaio_flags = 0;
277 		ki->kaio_maxactive_count = max_aio_per_proc;
278 		ki->kaio_active_count = 0;
279 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
280 		ki->kaio_queue_count = 0;
281 		ki->kaio_ballowed_count = max_buf_aio;
282 		ki->kaio_buffer_count = 0;
283 		ki->kaio_buffer_finished_count = 0;
284 		ki->kaio_p = p;
285 		TAILQ_INIT(&ki->kaio_jobdone);
286 		TAILQ_INIT(&ki->kaio_jobqueue);
287 		TAILQ_INIT(&ki->kaio_bufdone);
288 		TAILQ_INIT(&ki->kaio_bufqueue);
289 		TAILQ_INIT(&ki->kaio_liojoblist);
290 		TAILQ_INIT(&ki->kaio_sockqueue);
291 	}
292 
293 	while (num_aio_procs < target_aio_procs)
294 		aio_newproc();
295 }
296 
297 /*
298  * Free a job entry.  Wait for completion if it is currently active, but don't
299  * delay forever.  If we delay, we return a flag that says that we have to
300  * restart the queue scan.
301  */
302 static int
303 aio_free_entry(struct aiocblist *aiocbe)
304 {
305 	struct kaioinfo *ki;
306 	struct aio_liojob *lj;
307 	struct proc *p;
308 	int error;
309 
310 	if (aiocbe->jobstate == JOBST_NULL)
311 		panic("aio_free_entry: freeing already free job");
312 
313 	p = aiocbe->userproc;
314 	ki = p->p_aioinfo;
315 	lj = aiocbe->lio;
316 	if (ki == NULL)
317 		panic("aio_free_entry: missing p->p_aioinfo");
318 
319 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
320 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
321 		tsleep(aiocbe, 0, "jobwai", 0);
322 	}
323 	if (aiocbe->bp == NULL) {
324 		if (ki->kaio_queue_count <= 0)
325 			panic("aio_free_entry: process queue size <= 0");
326 		if (num_queue_count <= 0)
327 			panic("aio_free_entry: system wide queue size <= 0");
328 
329 		if (lj) {
330 			lj->lioj_queue_count--;
331 			if (aiocbe->jobflags & AIOCBLIST_DONE)
332 				lj->lioj_queue_finished_count--;
333 		}
334 		ki->kaio_queue_count--;
335 		if (aiocbe->jobflags & AIOCBLIST_DONE)
336 			ki->kaio_queue_finished_count--;
337 		num_queue_count--;
338 	} else {
339 		if (lj) {
340 			lj->lioj_buffer_count--;
341 			if (aiocbe->jobflags & AIOCBLIST_DONE)
342 				lj->lioj_buffer_finished_count--;
343 		}
344 		if (aiocbe->jobflags & AIOCBLIST_DONE)
345 			ki->kaio_buffer_finished_count--;
346 		ki->kaio_buffer_count--;
347 		num_buf_aio--;
348 	}
349 
350 	/* aiocbe is going away, we need to destroy any knotes */
351 	/* XXX lwp knote wants a thread, but only cares about the process */
352 	knote_remove(FIRST_LWP_IN_PROC(p)->lwp_thread, &aiocbe->klist);
353 
354 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
355 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
356 		ki->kaio_flags &= ~KAIO_WAKEUP;
357 		wakeup(p);
358 	}
359 
360 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
361 		if ((error = aio_fphysio(aiocbe)) != 0)
362 			return error;
363 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
364 			panic("aio_free_entry: invalid physio finish-up state");
365 		crit_enter();
366 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
367 		crit_exit();
368 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
369 		crit_enter();
370 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
371 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
372 		crit_exit();
373 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
374 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
375 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
376 		crit_enter();
377 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
378 		crit_exit();
379 		if (aiocbe->bp) {
380 			vunmapbuf(aiocbe->bp);
381 			relpbuf(aiocbe->bp, NULL);
382 			aiocbe->bp = NULL;
383 		}
384 	}
385 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
386 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
387 		zfree(aiolio_zone, lj);
388 	}
389 	aiocbe->jobstate = JOBST_NULL;
390 	callout_stop(&aiocbe->timeout);
391 	fdrop(aiocbe->fd_file);
392 	TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
393 	return 0;
394 }
395 #endif /* VFS_AIO */
396 
397 /*
398  * Rundown the jobs for a given process.
399  */
400 void
401 aio_proc_rundown(struct proc *p)
402 {
403 #ifndef VFS_AIO
404 	return;
405 #else
406 	struct kaioinfo *ki;
407 	struct aio_liojob *lj, *ljn;
408 	struct aiocblist *aiocbe, *aiocbn;
409 	struct file *fp;
410 	struct socket *so;
411 
412 	ki = p->p_aioinfo;
413 	if (ki == NULL)
414 		return;
415 
416 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
417 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
418 	    ki->kaio_buffer_finished_count)) {
419 		ki->kaio_flags |= KAIO_RUNDOWN;
420 		if (tsleep(p, 0, "kaiowt", aiod_timeout))
421 			break;
422 	}
423 
424 	/*
425 	 * Move any aio ops that are waiting on socket I/O to the normal job
426 	 * queues so they are cleaned up with any others.
427 	 */
428 	crit_enter();
429 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
430 	    aiocbn) {
431 		aiocbn = TAILQ_NEXT(aiocbe, plist);
432 		fp = aiocbe->fd_file;
433 		if (fp != NULL) {
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.ssb_flags &= ~SSB_AIO;
438 				so->so_rcv.ssb_flags &= ~SSB_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 	crit_exit();
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 restart3:
463 	crit_enter();
464 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
465 		ki->kaio_flags |= KAIO_WAKEUP;
466 		tsleep(p, 0, "aioprn", 0);
467 		crit_exit();
468 		goto restart3;
469 	}
470 	crit_exit();
471 
472 restart4:
473 	crit_enter();
474 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
475 		aiocbn = TAILQ_NEXT(aiocbe, plist);
476 		if (aio_free_entry(aiocbe)) {
477 			crit_exit();
478 			goto restart4;
479 		}
480 	}
481 	crit_exit();
482 
483         /*
484          * If we've slept, jobs might have moved from one queue to another.
485          * Retry rundown if we didn't manage to empty the queues.
486          */
487         if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
488 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
489 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
490 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
491 		goto restart1;
492 
493 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
494 		ljn = TAILQ_NEXT(lj, lioj_list);
495 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
496 		    0)) {
497 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
498 			zfree(aiolio_zone, lj);
499 		} else {
500 #ifdef DIAGNOSTIC
501 			kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
502 			    "QF:%d\n", lj->lioj_buffer_count,
503 			    lj->lioj_buffer_finished_count,
504 			    lj->lioj_queue_count,
505 			    lj->lioj_queue_finished_count);
506 #endif
507 		}
508 	}
509 
510 	zfree(kaio_zone, ki);
511 	p->p_aioinfo = NULL;
512 #endif /* VFS_AIO */
513 }
514 
515 #ifdef VFS_AIO
516 /*
517  * Select a job to run (called by an AIO daemon).
518  */
519 static struct aiocblist *
520 aio_selectjob(struct aioproclist *aiop)
521 {
522 	struct aiocblist *aiocbe;
523 	struct kaioinfo *ki;
524 	struct proc *userp;
525 
526 	crit_enter();
527 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
528 	    TAILQ_NEXT(aiocbe, list)) {
529 		userp = aiocbe->userproc;
530 		ki = userp->p_aioinfo;
531 
532 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
533 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
534 			crit_exit();
535 			return aiocbe;
536 		}
537 	}
538 	crit_exit();
539 
540 	return NULL;
541 }
542 
543 /*
544  * The AIO processing activity.  This is the code that does the I/O request for
545  * the non-physio version of the operations.  The normal vn operations are used,
546  * and this code should work in all instances for every type of file, including
547  * pipes, sockets, fifos, and regular files.
548  */
549 static void
550 aio_process(struct aiocblist *aiocbe)
551 {
552 	struct thread *mytd;
553 	struct aiocb *cb;
554 	struct file *fp;
555 	struct uio auio;
556 	struct iovec aiov;
557 	int cnt;
558 	int error;
559 	int oublock_st, oublock_end;
560 	int inblock_st, inblock_end;
561 
562 	mytd = curthread;
563 	cb = &aiocbe->uaiocb;
564 	fp = aiocbe->fd_file;
565 
566 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
567 	aiov.iov_len = cb->aio_nbytes;
568 
569 	auio.uio_iov = &aiov;
570 	auio.uio_iovcnt = 1;
571 	auio.uio_offset = cb->aio_offset;
572 	auio.uio_resid = cb->aio_nbytes;
573 	cnt = cb->aio_nbytes;
574 	auio.uio_segflg = UIO_USERSPACE;
575 	auio.uio_td = mytd;
576 
577 	inblock_st = mytd->td_lwp->lwp_ru.ru_inblock;
578 	oublock_st = mytd->td_lwp->lwp_ru.ru_oublock;
579 	/*
580 	 * _aio_aqueue() acquires a reference to the file that is
581 	 * released in aio_free_entry().
582 	 */
583 	if (cb->aio_lio_opcode == LIO_READ) {
584 		auio.uio_rw = UIO_READ;
585 		error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET);
586 	} else {
587 		auio.uio_rw = UIO_WRITE;
588 		error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET);
589 	}
590 	inblock_end = mytd->td_lwp->lwp_ru.ru_inblock;
591 	oublock_end = mytd->td_lwp->lwp_ru.ru_oublock;
592 
593 	aiocbe->inputcharge = inblock_end - inblock_st;
594 	aiocbe->outputcharge = oublock_end - oublock_st;
595 
596 	if ((error) && (auio.uio_resid != cnt)) {
597 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
598 			error = 0;
599 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
600 			ksignal(aiocbe->userproc, SIGPIPE);
601 	}
602 
603 	cnt -= auio.uio_resid;
604 	cb->_aiocb_private.error = error;
605 	cb->_aiocb_private.status = cnt;
606 }
607 
608 /*
609  * The AIO daemon, most of the actual work is done in aio_process,
610  * but the setup (and address space mgmt) is done in this routine.
611  *
612  * The MP lock is held on entry.
613  */
614 static void
615 aio_daemon(void *uproc, struct trapframe *frame)
616 {
617 	struct aio_liojob *lj;
618 	struct aiocb *cb;
619 	struct aiocblist *aiocbe;
620 	struct aioproclist *aiop;
621 	struct kaioinfo *ki;
622 	struct proc *mycp, *userp;
623 	struct vmspace *curvm;
624 	struct lwp *mylwp;
625 	struct ucred *cr;
626 
627 	mylwp = curthread->td_lwp;
628 	mycp = mylwp->lwp_proc;
629 
630 	if (mycp->p_textvp) {
631 		vrele(mycp->p_textvp);
632 		mycp->p_textvp = NULL;
633 	}
634 
635 	/*
636 	 * Allocate and ready the aio control info.  There is one aiop structure
637 	 * per daemon.
638 	 */
639 	aiop = zalloc(aiop_zone);
640 	aiop->aioproc = mycp;
641 	aiop->aioprocflags |= AIOP_FREE;
642 
643 	crit_enter();
644 
645 	/*
646 	 * Place thread (lightweight process) onto the AIO free thread list.
647 	 */
648 	if (TAILQ_EMPTY(&aio_freeproc))
649 		wakeup(&aio_freeproc);
650 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
651 
652 	crit_exit();
653 
654 	/* Make up a name for the daemon. */
655 	strcpy(mycp->p_comm, "aiod");
656 
657 	/*
658 	 * Get rid of our current filedescriptors.  AIOD's don't need any
659 	 * filedescriptors, except as temporarily inherited from the client.
660 	 * Credentials are also cloned, and made equivalent to "root".
661 	 */
662 	fdfree(mycp, NULL);
663 	cr = cratom(&mycp->p_ucred);
664 	cr->cr_uid = 0;
665 	uireplace(&cr->cr_uidinfo, uifind(0));
666 	cr->cr_ngroups = 1;
667 	cr->cr_groups[0] = 1;
668 
669 	/* The daemon resides in its own pgrp. */
670 	enterpgrp(mycp, mycp->p_pid, 1);
671 
672 	/* Mark special process type. */
673 	mycp->p_flag |= P_SYSTEM | P_KTHREADP;
674 
675 	/*
676 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
677 	 * and creating too many daemons.)
678 	 */
679 	wakeup(mycp);
680 	curvm = NULL;
681 
682 	for (;;) {
683 		/*
684 		 * Take daemon off of free queue
685 		 */
686 		if (aiop->aioprocflags & AIOP_FREE) {
687 			crit_enter();
688 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
689 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
690 			aiop->aioprocflags &= ~AIOP_FREE;
691 			crit_exit();
692 		}
693 		aiop->aioprocflags &= ~AIOP_SCHED;
694 
695 		/*
696 		 * Check for jobs.
697 		 */
698 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
699 			cb = &aiocbe->uaiocb;
700 			userp = aiocbe->userproc;
701 
702 			aiocbe->jobstate = JOBST_JOBRUNNING;
703 
704 			/*
705 			 * Connect to process address space for user program.
706 			 */
707 			if (curvm != userp->p_vmspace) {
708 				pmap_setlwpvm(mylwp, userp->p_vmspace);
709 				if (curvm)
710 					sysref_put(&curvm->vm_sysref);
711 				curvm = userp->p_vmspace;
712 				sysref_get(&curvm->vm_sysref);
713 			}
714 
715 			ki = userp->p_aioinfo;
716 			lj = aiocbe->lio;
717 
718 			/* Account for currently active jobs. */
719 			ki->kaio_active_count++;
720 
721 			/* Do the I/O function. */
722 			aio_process(aiocbe);
723 
724 			/* Decrement the active job count. */
725 			ki->kaio_active_count--;
726 
727 			/*
728 			 * Increment the completion count for wakeup/signal
729 			 * comparisons.
730 			 */
731 			aiocbe->jobflags |= AIOCBLIST_DONE;
732 			ki->kaio_queue_finished_count++;
733 			if (lj)
734 				lj->lioj_queue_finished_count++;
735 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
736 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
737 				ki->kaio_flags &= ~KAIO_WAKEUP;
738 				wakeup(userp);
739 			}
740 
741 			crit_enter();
742 			if (lj && (lj->lioj_flags &
743 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
744 				if ((lj->lioj_queue_finished_count ==
745 				    lj->lioj_queue_count) &&
746 				    (lj->lioj_buffer_finished_count ==
747 				    lj->lioj_buffer_count)) {
748 						ksignal(userp,
749 						    lj->lioj_signal.sigev_signo);
750 						lj->lioj_flags |=
751 						    LIOJ_SIGNAL_POSTED;
752 				}
753 			}
754 			crit_exit();
755 
756 			aiocbe->jobstate = JOBST_JOBFINISHED;
757 
758 			crit_enter();
759 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
760 			TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
761 			crit_exit();
762 			KNOTE(&aiocbe->klist, 0);
763 
764 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
765 				wakeup(aiocbe);
766 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
767 			}
768 
769 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
770 				ksignal(userp, cb->aio_sigevent.sigev_signo);
771 			}
772 		}
773 
774 		/*
775 		 * Disconnect from user address space.
776 		 */
777 		if (curvm) {
778 			/* swap our original address space back in */
779 			pmap_setlwpvm(mylwp, mycp->p_vmspace);
780 			sysref_put(&curvm->vm_sysref);
781 			curvm = NULL;
782 		}
783 
784 		/*
785 		 * If we are the first to be put onto the free queue, wakeup
786 		 * anyone waiting for a daemon.
787 		 */
788 		crit_enter();
789 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
790 		if (TAILQ_EMPTY(&aio_freeproc))
791 			wakeup(&aio_freeproc);
792 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
793 		aiop->aioprocflags |= AIOP_FREE;
794 		crit_exit();
795 
796 		/*
797 		 * If daemon is inactive for a long time, allow it to exit,
798 		 * thereby freeing resources.
799 		 */
800 		if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
801 		    0, "aiordy", aiod_lifetime)) {
802 			crit_enter();
803 			if (TAILQ_EMPTY(&aio_jobs)) {
804 				if ((aiop->aioprocflags & AIOP_FREE) &&
805 				    (num_aio_procs > target_aio_procs)) {
806 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
807 					crit_exit();
808 					zfree(aiop_zone, aiop);
809 					num_aio_procs--;
810 #ifdef DIAGNOSTIC
811 					if (mycp->p_vmspace->vm_sysref.refcnt <= 1) {
812 						kprintf("AIOD: bad vm refcnt for"
813 						    " exiting daemon: %d\n",
814 						    mycp->p_vmspace->vm_sysref.refcnt);
815 					}
816 #endif
817 					exit1(0);
818 				}
819 			}
820 			crit_exit();
821 		}
822 	}
823 }
824 
825 /*
826  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
827  * AIO daemon modifies its environment itself.
828  */
829 static int
830 aio_newproc(void)
831 {
832 	int error;
833 	struct lwp *lp, *nlp;
834 	struct proc *np;
835 
836 	lp = &lwp0;
837 	error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
838 	if (error)
839 		return error;
840 	nlp = ONLY_LWP_IN_PROC(np);
841 	cpu_set_fork_handler(nlp, aio_daemon, curproc);
842 	start_forked_proc(lp, np);
843 
844 	/*
845 	 * Wait until daemon is started, but continue on just in case to
846 	 * handle error conditions.
847 	 */
848 	error = tsleep(np, 0, "aiosta", aiod_timeout);
849 	num_aio_procs++;
850 
851 	return error;
852 }
853 
854 /*
855  * Try the high-performance, low-overhead physio method for eligible
856  * VCHR devices.  This method doesn't use an aio helper thread, and
857  * thus has very low overhead.
858  *
859  * Assumes that the caller, _aio_aqueue(), has incremented the file
860  * structure's reference count, preventing its deallocation for the
861  * duration of this call.
862  */
863 static int
864 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
865 {
866 	int error;
867 	struct aiocb *cb;
868 	struct file *fp;
869 	struct buf *bp;
870 	struct vnode *vp;
871 	struct kaioinfo *ki;
872 	struct aio_liojob *lj;
873 	int notify;
874 
875 	cb = &aiocbe->uaiocb;
876 	fp = aiocbe->fd_file;
877 
878 	if (fp->f_type != DTYPE_VNODE)
879 		return (-1);
880 
881 	vp = (struct vnode *)fp->f_data;
882 
883 	/*
884 	 * If its not a disk, we don't want to return a positive error.
885 	 * It causes the aio code to not fall through to try the thread
886 	 * way when you're talking to a regular file.
887 	 */
888 	if (!vn_isdisk(vp, &error)) {
889 		if (error == ENOTBLK)
890 			return (-1);
891 		else
892 			return (error);
893 	}
894 
895  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
896 		return (-1);
897 
898 	if (cb->aio_nbytes >
899 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
900 		return (-1);
901 
902 	ki = p->p_aioinfo;
903 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
904 		return (-1);
905 
906 	ki->kaio_buffer_count++;
907 
908 	lj = aiocbe->lio;
909 	if (lj)
910 		lj->lioj_buffer_count++;
911 
912 	/* Create and build a buffer header for a transfer. */
913 	bp = getpbuf(NULL);
914 	BUF_KERNPROC(bp);
915 
916 	/*
917 	 * Get a copy of the kva from the physical buffer.
918 	 */
919 	bp->b_bio1.bio_caller_info1.ptr = p;
920 	error = 0;
921 
922 	bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
923 		    BUF_CMD_WRITE : BUF_CMD_READ;
924 	bp->b_bio1.bio_done = aio_physwakeup;
925 	bp->b_bio1.bio_flags |= BIO_SYNC;
926 	bp->b_bio1.bio_offset = cb->aio_offset;
927 
928 	/* Bring buffer into kernel space. */
929 	if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
930 		error = EFAULT;
931 		goto doerror;
932 	}
933 
934 	crit_enter();
935 
936 	aiocbe->bp = bp;
937 	bp->b_bio1.bio_caller_info2.ptr = aiocbe;
938 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
939 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
940 	aiocbe->jobstate = JOBST_JOBQBUF;
941 	cb->_aiocb_private.status = cb->aio_nbytes;
942 	num_buf_aio++;
943 	bp->b_error = 0;
944 
945 	crit_exit();
946 
947 	/*
948 	 * Perform the transfer.  vn_strategy must be used even though we
949 	 * know we have a device in order to deal with requests which exceed
950 	 * device DMA limitations.
951 	 */
952 	vn_strategy(vp, &bp->b_bio1);
953 
954 	notify = 0;
955 	crit_enter();
956 
957 #if 0
958 	/*
959 	 * If we had an error invoking the request, or an error in processing
960 	 * the request before we have returned, we process it as an error in
961 	 * transfer.  Note that such an I/O error is not indicated immediately,
962 	 * but is returned using the aio_error mechanism.  In this case,
963 	 * aio_suspend will return immediately.
964 	 */
965 	if (bp->b_error || (bp->b_flags & B_ERROR)) {
966 		struct aiocb *job = aiocbe->uuaiocb;
967 
968 		aiocbe->uaiocb._aiocb_private.status = 0;
969 		suword(&job->_aiocb_private.status, 0);
970 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
971 		suword(&job->_aiocb_private.error, bp->b_error);
972 
973 		ki->kaio_buffer_finished_count++;
974 
975 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
976 			aiocbe->jobstate = JOBST_JOBBFINISHED;
977 			aiocbe->jobflags |= AIOCBLIST_DONE;
978 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
979 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
980 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
981 			notify = 1;
982 		}
983 	}
984 #endif
985 	crit_exit();
986 	if (notify)
987 		KNOTE(&aiocbe->klist, 0);
988 	return 0;
989 
990 doerror:
991 	ki->kaio_buffer_count--;
992 	if (lj)
993 		lj->lioj_buffer_count--;
994 	aiocbe->bp = NULL;
995 	relpbuf(bp, NULL);
996 	return error;
997 }
998 
999 /*
1000  * This waits/tests physio completion.
1001  */
1002 static int
1003 aio_fphysio(struct aiocblist *iocb)
1004 {
1005 	struct buf *bp;
1006 	int error;
1007 
1008 	bp = iocb->bp;
1009 
1010 	error = biowait_timeout(&bp->b_bio1, "physstr", aiod_timeout);
1011 	if (error == EWOULDBLOCK)
1012 		return EINPROGRESS;
1013 
1014 	/* Release mapping into kernel space. */
1015 	vunmapbuf(bp);
1016 	iocb->bp = 0;
1017 
1018 	error = 0;
1019 
1020 	/* Check for an error. */
1021 	if (bp->b_flags & B_ERROR)
1022 		error = bp->b_error;
1023 
1024 	relpbuf(bp, NULL);
1025 	return (error);
1026 }
1027 #endif /* VFS_AIO */
1028 
1029 /*
1030  * Wake up aio requests that may be serviceable now.
1031  */
1032 void
1033 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1034 {
1035 #ifndef VFS_AIO
1036 	return;
1037 #else
1038 	struct aiocblist *cb,*cbn;
1039 	struct proc *p;
1040 	struct kaioinfo *ki = NULL;
1041 	int opcode, wakecount = 0;
1042 	struct aioproclist *aiop;
1043 
1044 	if (ssb == &so->so_snd) {
1045 		opcode = LIO_WRITE;
1046 		so->so_snd.ssb_flags &= ~SSB_AIO;
1047 	} else {
1048 		opcode = LIO_READ;
1049 		so->so_rcv.ssb_flags &= ~SSB_AIO;
1050 	}
1051 
1052 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1053 		cbn = TAILQ_NEXT(cb, list);
1054 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1055 			p = cb->userproc;
1056 			ki = p->p_aioinfo;
1057 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1058 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1059 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1060 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1061 			wakecount++;
1062 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1063 				panic("invalid queue value");
1064 		}
1065 	}
1066 
1067 	while (wakecount--) {
1068 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1069 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1070 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1071 			aiop->aioprocflags &= ~AIOP_FREE;
1072 			wakeup(aiop->aioproc);
1073 		}
1074 	}
1075 #endif /* VFS_AIO */
1076 }
1077 
1078 #ifdef VFS_AIO
1079 /*
1080  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1081  * technique is done in this code.
1082  */
1083 static int
1084 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1085 {
1086 	struct proc *p = curproc;
1087 	struct file *fp;
1088 	unsigned int fd;
1089 	struct socket *so;
1090 	int error;
1091 	int opcode, user_opcode;
1092 	struct aiocblist *aiocbe;
1093 	struct aioproclist *aiop;
1094 	struct kaioinfo *ki;
1095 	struct kevent kev;
1096 	struct kqueue *kq;
1097 	struct file *kq_fp;
1098 	int fflags;
1099 
1100 	if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1101 		TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1102 	else
1103 		aiocbe = zalloc (aiocb_zone);
1104 
1105 	aiocbe->inputcharge = 0;
1106 	aiocbe->outputcharge = 0;
1107 	callout_init(&aiocbe->timeout);
1108 	SLIST_INIT(&aiocbe->klist);
1109 
1110 	suword(&job->_aiocb_private.status, -1);
1111 	suword(&job->_aiocb_private.error, 0);
1112 	suword(&job->_aiocb_private.kernelinfo, -1);
1113 
1114 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1115 	if (error) {
1116 		suword(&job->_aiocb_private.error, error);
1117 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1118 		return error;
1119 	}
1120 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1121 	    !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1122 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1123 		return EINVAL;
1124 	}
1125 
1126 	/* Save userspace address of the job info. */
1127 	aiocbe->uuaiocb = job;
1128 
1129 	/* Get the opcode. */
1130 	user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1131 	if (type != LIO_NOP)
1132 		aiocbe->uaiocb.aio_lio_opcode = type;
1133 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1134 
1135 	/*
1136 	 * Range check file descriptor.
1137 	 */
1138 	fflags = (opcode == LIO_WRITE) ? FWRITE : FREAD;
1139 	fd = aiocbe->uaiocb.aio_fildes;
1140 	fp = holdfp(p->p_fd, fd, fflags);
1141 	if (fp == NULL) {
1142 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1143 		if (type == 0)
1144 			suword(&job->_aiocb_private.error, EBADF);
1145 		return EBADF;
1146 	}
1147 
1148 	aiocbe->fd_file = fp;
1149 
1150 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1151 		error = EINVAL;
1152 		goto aqueue_fail;
1153 	}
1154 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1155 	if (error) {
1156 		error = EINVAL;
1157 		goto aqueue_fail;
1158 	}
1159 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1160 	if (jobrefid == LONG_MAX)
1161 		jobrefid = 1;
1162 	else
1163 		jobrefid++;
1164 
1165 	if (opcode == LIO_NOP) {
1166 		fdrop(fp);
1167 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1168 		if (type == 0) {
1169 			suword(&job->_aiocb_private.error, 0);
1170 			suword(&job->_aiocb_private.status, 0);
1171 			suword(&job->_aiocb_private.kernelinfo, 0);
1172 		}
1173 		return 0;
1174 	}
1175 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1176 		if (type == 0)
1177 			suword(&job->_aiocb_private.status, 0);
1178 		error = EINVAL;
1179 		goto aqueue_fail;
1180 	}
1181 
1182 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1183 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1184 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1185 	}
1186 	else {
1187 		/*
1188 		 * This method for requesting kevent-based notification won't
1189 		 * work on the alpha, since we're passing in a pointer
1190 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1191 		 * based method instead.
1192 		 */
1193 		if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1194 		    user_opcode == LIO_WRITE)
1195 			goto no_kqueue;
1196 
1197 		error = copyin((struct kevent *)(uintptr_t)user_opcode,
1198 		    &kev, sizeof(kev));
1199 		if (error)
1200 			goto aqueue_fail;
1201 	}
1202 	kq_fp = holdfp(p->p_fd, (int)kev.ident, -1);
1203 	if (kq_fp == NULL || kq_fp->f_type != DTYPE_KQUEUE) {
1204 		if (kq_fp) {
1205 			fdrop(kq_fp);
1206 			kq_fp = NULL;
1207 		}
1208 		error = EBADF;
1209 		goto aqueue_fail;
1210 	}
1211 	kq = (struct kqueue *)kq_fp->f_data;
1212 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1213 	kev.filter = EVFILT_AIO;
1214 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1215 	kev.data = (intptr_t)aiocbe;
1216 	/* XXX lwp kqueue_register takes a thread, but only uses its proc */
1217 	error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread);
1218 	fdrop(kq_fp);
1219 aqueue_fail:
1220 	if (error) {
1221 		fdrop(fp);
1222 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1223 		if (type == 0)
1224 			suword(&job->_aiocb_private.error, error);
1225 		goto done;
1226 	}
1227 no_kqueue:
1228 
1229 	suword(&job->_aiocb_private.error, EINPROGRESS);
1230 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1231 	aiocbe->userproc = p;
1232 	aiocbe->jobflags = 0;
1233 	aiocbe->lio = lj;
1234 	ki = p->p_aioinfo;
1235 
1236 	if (fp->f_type == DTYPE_SOCKET) {
1237 		/*
1238 		 * Alternate queueing for socket ops: Reach down into the
1239 		 * descriptor to get the socket data.  Then check to see if the
1240 		 * socket is ready to be read or written (based on the requested
1241 		 * operation).
1242 		 *
1243 		 * If it is not ready for io, then queue the aiocbe on the
1244 		 * socket, and set the flags so we get a call when ssb_notify()
1245 		 * happens.
1246 		 */
1247 		so = (struct socket *)fp->f_data;
1248 		crit_enter();
1249 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1250 		    LIO_WRITE) && (!sowriteable(so)))) {
1251 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1252 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1253 			if (opcode == LIO_READ)
1254 				so->so_rcv.ssb_flags |= SSB_AIO;
1255 			else
1256 				so->so_snd.ssb_flags |= SSB_AIO;
1257 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1258 			ki->kaio_queue_count++;
1259 			num_queue_count++;
1260 			crit_exit();
1261 			error = 0;
1262 			goto done;
1263 		}
1264 		crit_exit();
1265 	}
1266 
1267 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1268 		goto done;
1269 	if (error > 0) {
1270 		suword(&job->_aiocb_private.status, 0);
1271 		aiocbe->uaiocb._aiocb_private.error = error;
1272 		suword(&job->_aiocb_private.error, error);
1273 		goto done;
1274 	}
1275 
1276 	/* No buffer for daemon I/O. */
1277 	aiocbe->bp = NULL;
1278 
1279 	ki->kaio_queue_count++;
1280 	if (lj)
1281 		lj->lioj_queue_count++;
1282 	crit_enter();
1283 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1284 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1285 	crit_exit();
1286 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1287 
1288 	num_queue_count++;
1289 	error = 0;
1290 
1291 	/*
1292 	 * If we don't have a free AIO process, and we are below our quota, then
1293 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1294 	 * pick-up this job.  If we don't successfully create the new process
1295 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1296 	 * which is likely not the correct thing to do.
1297 	 */
1298 	crit_enter();
1299 retryproc:
1300 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1301 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1302 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1303 		aiop->aioprocflags &= ~AIOP_FREE;
1304 		wakeup(aiop->aioproc);
1305 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1306 	    ((ki->kaio_active_count + num_aio_resv_start) <
1307 	    ki->kaio_maxactive_count)) {
1308 		num_aio_resv_start++;
1309 		if ((error = aio_newproc()) == 0) {
1310 			num_aio_resv_start--;
1311 			goto retryproc;
1312 		}
1313 		num_aio_resv_start--;
1314 	}
1315 	crit_exit();
1316 done:
1317 	return error;
1318 }
1319 
1320 /*
1321  * This routine queues an AIO request, checking for quotas.
1322  */
1323 static int
1324 aio_aqueue(struct aiocb *job, int type)
1325 {
1326 	struct proc *p = curproc;
1327 	struct kaioinfo *ki;
1328 
1329 	if (p->p_aioinfo == NULL)
1330 		aio_init_aioinfo(p);
1331 
1332 	if (num_queue_count >= max_queue_count)
1333 		return EAGAIN;
1334 
1335 	ki = p->p_aioinfo;
1336 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1337 		return EAGAIN;
1338 
1339 	return _aio_aqueue(job, NULL, type);
1340 }
1341 #endif /* VFS_AIO */
1342 
1343 /*
1344  * Support the aio_return system call, as a side-effect, kernel resources are
1345  * released.
1346  *
1347  * MPALMOSTSAFE
1348  */
1349 int
1350 sys_aio_return(struct aio_return_args *uap)
1351 {
1352 #ifndef VFS_AIO
1353 	return (ENOSYS);
1354 #else
1355 	struct proc *p = curproc;
1356 	struct lwp *lp = curthread->td_lwp;
1357 	long jobref;
1358 	struct aiocblist *cb, *ncb;
1359 	struct aiocb *ujob;
1360 	struct kaioinfo *ki;
1361 	int error;
1362 
1363 	ki = p->p_aioinfo;
1364 	if (ki == NULL)
1365 		return EINVAL;
1366 
1367 	ujob = uap->aiocbp;
1368 
1369 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1370 	if (jobref == -1 || jobref == 0)
1371 		return EINVAL;
1372 
1373 	get_mplock();
1374 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1375 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1376 		    jobref) {
1377 			if (ujob == cb->uuaiocb) {
1378 				uap->sysmsg_result =
1379 				    cb->uaiocb._aiocb_private.status;
1380 			} else {
1381 				uap->sysmsg_result = EFAULT;
1382 			}
1383 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1384 				lp->lwp_ru.ru_oublock += cb->outputcharge;
1385 				cb->outputcharge = 0;
1386 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1387 				lp->lwp_ru.ru_inblock += cb->inputcharge;
1388 				cb->inputcharge = 0;
1389 			}
1390 			aio_free_entry(cb);
1391 			error = 0;
1392 			goto done;
1393 		}
1394 	}
1395 	crit_enter();
1396 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1397 		ncb = TAILQ_NEXT(cb, plist);
1398 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1399 		    == jobref) {
1400 			crit_exit();
1401 			if (ujob == cb->uuaiocb) {
1402 				uap->sysmsg_result =
1403 				    cb->uaiocb._aiocb_private.status;
1404 			} else {
1405 				uap->sysmsg_result = EFAULT;
1406 			}
1407 			aio_free_entry(cb);
1408 			error = 0;
1409 			goto done;
1410 		}
1411 	}
1412 	crit_exit();
1413 	error = EINVAL;
1414 done:
1415 	rel_mplock();
1416 	return (error);
1417 #endif /* VFS_AIO */
1418 }
1419 
1420 /*
1421  * Allow a process to wakeup when any of the I/O requests are completed.
1422  *
1423  * MPALMOSTSAFE
1424  */
1425 int
1426 sys_aio_suspend(struct aio_suspend_args *uap)
1427 {
1428 #ifndef VFS_AIO
1429 	return ENOSYS;
1430 #else
1431 	struct proc *p = curproc;
1432 	struct timeval atv;
1433 	struct timespec ts;
1434 	struct aiocb *const *cbptr, *cbp;
1435 	struct kaioinfo *ki;
1436 	struct aiocblist *cb;
1437 	int i;
1438 	int njoblist;
1439 	int error, timo;
1440 	long *ijoblist;
1441 	struct aiocb **ujoblist;
1442 
1443 	if ((u_int)uap->nent > AIO_LISTIO_MAX)
1444 		return EINVAL;
1445 
1446 	timo = 0;
1447 	if (uap->timeout) {
1448 		/* Get timespec struct. */
1449 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1450 			return error;
1451 
1452 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1453 			return (EINVAL);
1454 
1455 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1456 		if (itimerfix(&atv))
1457 			return (EINVAL);
1458 		timo = tvtohz_high(&atv);
1459 	}
1460 
1461 	ki = p->p_aioinfo;
1462 	if (ki == NULL)
1463 		return EAGAIN;
1464 
1465 	get_mplock();
1466 
1467 	njoblist = 0;
1468 	ijoblist = zalloc(aiol_zone);
1469 	ujoblist = zalloc(aiol_zone);
1470 	cbptr = uap->aiocbp;
1471 
1472 	for (i = 0; i < uap->nent; i++) {
1473 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1474 		if (cbp == 0)
1475 			continue;
1476 		ujoblist[njoblist] = cbp;
1477 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1478 		njoblist++;
1479 	}
1480 
1481 	if (njoblist == 0) {
1482 		zfree(aiol_zone, ijoblist);
1483 		zfree(aiol_zone, ujoblist);
1484 		error = 0;
1485 		goto done;
1486 	}
1487 
1488 	error = 0;
1489 	for (;;) {
1490 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1491 			for (i = 0; i < njoblist; i++) {
1492 				if (((intptr_t)
1493 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1494 				    ijoblist[i]) {
1495 					if (ujoblist[i] != cb->uuaiocb)
1496 						error = EINVAL;
1497 					zfree(aiol_zone, ijoblist);
1498 					zfree(aiol_zone, ujoblist);
1499 					goto done;
1500 				}
1501 			}
1502 		}
1503 
1504 		crit_enter();
1505 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1506 		    TAILQ_NEXT(cb, plist)) {
1507 			for (i = 0; i < njoblist; i++) {
1508 				if (((intptr_t)
1509 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1510 				    ijoblist[i]) {
1511 					crit_exit();
1512 					if (ujoblist[i] != cb->uuaiocb)
1513 						error = EINVAL;
1514 					zfree(aiol_zone, ijoblist);
1515 					zfree(aiol_zone, ujoblist);
1516 					goto done;
1517 				}
1518 			}
1519 		}
1520 
1521 		ki->kaio_flags |= KAIO_WAKEUP;
1522 		error = tsleep(p, PCATCH, "aiospn", timo);
1523 		crit_exit();
1524 
1525 		if (error == ERESTART || error == EINTR) {
1526 			zfree(aiol_zone, ijoblist);
1527 			zfree(aiol_zone, ujoblist);
1528 			error = EINTR;
1529 			goto done;
1530 		} else if (error == EWOULDBLOCK) {
1531 			zfree(aiol_zone, ijoblist);
1532 			zfree(aiol_zone, ujoblist);
1533 			error = EAGAIN;
1534 			goto done;
1535 		}
1536 	}
1537 
1538 /* NOTREACHED */
1539 	error = EINVAL;
1540 done:
1541 	rel_mplock();
1542 	return (error);
1543 #endif /* VFS_AIO */
1544 }
1545 
1546 /*
1547  * aio_cancel cancels any non-physio aio operations not currently in
1548  * progress.
1549  *
1550  * MPALMOSTSAFE
1551  */
1552 int
1553 sys_aio_cancel(struct aio_cancel_args *uap)
1554 {
1555 #ifndef VFS_AIO
1556 	return ENOSYS;
1557 #else
1558 	struct proc *p = curproc;
1559 	struct kaioinfo *ki;
1560 	struct aiocblist *cbe, *cbn;
1561 	struct file *fp;
1562 	struct socket *so;
1563 	struct proc *po;
1564 	int error;
1565 	int cancelled=0;
1566 	int notcancelled=0;
1567 	struct vnode *vp;
1568 
1569 	fp = holdfp(p->p_fd, uap->fd, -1);
1570 	if (fp == NULL)
1571 		return (EBADF);
1572 
1573 	get_mplock();
1574 
1575         if (fp->f_type == DTYPE_VNODE) {
1576 		vp = (struct vnode *)fp->f_data;
1577 
1578 		if (vn_isdisk(vp,&error)) {
1579 			uap->sysmsg_result = AIO_NOTCANCELED;
1580 			error = 0;
1581 			goto done2;
1582 		}
1583 	} else if (fp->f_type == DTYPE_SOCKET) {
1584 		so = (struct socket *)fp->f_data;
1585 
1586 		crit_enter();
1587 
1588 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1589 			cbn = TAILQ_NEXT(cbe, list);
1590 			if ((uap->aiocbp == NULL) ||
1591 				(uap->aiocbp == cbe->uuaiocb) ) {
1592 				po = cbe->userproc;
1593 				ki = po->p_aioinfo;
1594 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1595 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1596 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1597 				if (ki->kaio_flags & KAIO_WAKEUP) {
1598 					wakeup(po);
1599 				}
1600 				cbe->jobstate = JOBST_JOBFINISHED;
1601 				cbe->uaiocb._aiocb_private.status=-1;
1602 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1603 				cancelled++;
1604 /* XXX cancelled, knote? */
1605 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1606 				    SIGEV_SIGNAL)
1607 					ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1608 				if (uap->aiocbp)
1609 					break;
1610 			}
1611 		}
1612 		crit_exit();
1613 
1614 		if ((cancelled) && (uap->aiocbp)) {
1615 			uap->sysmsg_result = AIO_CANCELED;
1616 			error = 0;
1617 			goto done2;
1618 		}
1619 	}
1620 	ki=p->p_aioinfo;
1621 	if (ki == NULL)
1622 		goto done;
1623 	crit_enter();
1624 
1625 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1626 		cbn = TAILQ_NEXT(cbe, plist);
1627 
1628 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1629 		    ((uap->aiocbp == NULL ) ||
1630 		     (uap->aiocbp == cbe->uuaiocb))) {
1631 
1632 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1633 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1634                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1635                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1636                                     plist);
1637 				cancelled++;
1638 				ki->kaio_queue_finished_count++;
1639 				cbe->jobstate = JOBST_JOBFINISHED;
1640 				cbe->uaiocb._aiocb_private.status = -1;
1641 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1642 /* XXX cancelled, knote? */
1643 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1644 				    SIGEV_SIGNAL)
1645 					ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1646 			} else {
1647 				notcancelled++;
1648 			}
1649 		}
1650 	}
1651 	crit_exit();
1652 done:
1653 	if (notcancelled)
1654 		uap->sysmsg_result = AIO_NOTCANCELED;
1655 	else if (cancelled)
1656 		uap->sysmsg_result = AIO_CANCELED;
1657 	else
1658 		uap->sysmsg_result = AIO_ALLDONE;
1659 	error = 0;
1660 done2:
1661 	rel_mplock();
1662 	fdrop(fp);
1663 	return error;
1664 #endif /* VFS_AIO */
1665 }
1666 
1667 /*
1668  * aio_error is implemented in the kernel level for compatibility purposes only.
1669  * For a user mode async implementation, it would be best to do it in a userland
1670  * subroutine.
1671  *
1672  * MPALMOSTSAFE
1673  */
1674 int
1675 sys_aio_error(struct aio_error_args *uap)
1676 {
1677 #ifndef VFS_AIO
1678 	return ENOSYS;
1679 #else
1680 	struct proc *p = curproc;
1681 	struct aiocblist *cb;
1682 	struct kaioinfo *ki;
1683 	long jobref;
1684 	int error;
1685 
1686 	ki = p->p_aioinfo;
1687 	if (ki == NULL)
1688 		return EINVAL;
1689 
1690 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1691 	if ((jobref == -1) || (jobref == 0))
1692 		return EINVAL;
1693 
1694 	get_mplock();
1695 	error = 0;
1696 
1697 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1698 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1699 		    jobref) {
1700 			uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1701 			goto done;
1702 		}
1703 	}
1704 
1705 	crit_enter();
1706 
1707 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1708 	    plist)) {
1709 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1710 		    jobref) {
1711 			uap->sysmsg_result = EINPROGRESS;
1712 			crit_exit();
1713 			goto done;
1714 		}
1715 	}
1716 
1717 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1718 	    plist)) {
1719 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1720 		    jobref) {
1721 			uap->sysmsg_result = EINPROGRESS;
1722 			crit_exit();
1723 			goto done;
1724 		}
1725 	}
1726 	crit_exit();
1727 
1728 	crit_enter();
1729 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1730 	    plist)) {
1731 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1732 		    jobref) {
1733 			uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1734 			crit_exit();
1735 			goto done;
1736 		}
1737 	}
1738 
1739 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1740 	    plist)) {
1741 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1742 		    jobref) {
1743 			uap->sysmsg_result = EINPROGRESS;
1744 			crit_exit();
1745 			goto done;
1746 		}
1747 	}
1748 	crit_exit();
1749 	error = EINVAL;
1750 done:
1751 	rel_mplock();
1752 	return (error);
1753 #endif /* VFS_AIO */
1754 }
1755 
1756 /*
1757  * syscall - asynchronous read from a file (REALTIME)
1758  *
1759  * MPALMOSTSAFE
1760  */
1761 int
1762 sys_aio_read(struct aio_read_args *uap)
1763 {
1764 #ifndef VFS_AIO
1765 	return ENOSYS;
1766 #else
1767 	int error;
1768 
1769 	get_mplock();
1770 	error =  aio_aqueue(uap->aiocbp, LIO_READ);
1771 	rel_mplock();
1772 	return (error);
1773 #endif /* VFS_AIO */
1774 }
1775 
1776 /*
1777  * syscall - asynchronous write to a file (REALTIME)
1778  *
1779  * MPALMOSTSAFE
1780  */
1781 int
1782 sys_aio_write(struct aio_write_args *uap)
1783 {
1784 #ifndef VFS_AIO
1785 	return ENOSYS;
1786 #else
1787 	int error;
1788 
1789 	get_mplock();
1790 	error = aio_aqueue(uap->aiocbp, LIO_WRITE);
1791 	rel_mplock();
1792 	return (error);
1793 #endif /* VFS_AIO */
1794 }
1795 
1796 /*
1797  * syscall - XXX undocumented
1798  *
1799  * MPALMOSTSAFE
1800  */
1801 int
1802 sys_lio_listio(struct lio_listio_args *uap)
1803 {
1804 #ifndef VFS_AIO
1805 	return ENOSYS;
1806 #else
1807 	struct proc *p = curproc;
1808 	struct lwp *lp = curthread->td_lwp;
1809 	int nent, nentqueued;
1810 	struct aiocb *iocb, * const *cbptr;
1811 	struct aiocblist *cb;
1812 	struct kaioinfo *ki;
1813 	struct aio_liojob *lj;
1814 	int error, runningcode;
1815 	int nerror;
1816 	int i;
1817 
1818 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1819 		return EINVAL;
1820 
1821 	nent = uap->nent;
1822 	if (nent > AIO_LISTIO_MAX)
1823 		return EINVAL;
1824 
1825 	get_mplock();
1826 
1827 	if (p->p_aioinfo == NULL)
1828 		aio_init_aioinfo(p);
1829 
1830 	if ((nent + num_queue_count) > max_queue_count) {
1831 		error = EAGAIN;
1832 		goto done;
1833 	}
1834 
1835 	ki = p->p_aioinfo;
1836 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) {
1837 		error = EAGAIN;
1838 		goto done;
1839 	}
1840 
1841 	lj = zalloc(aiolio_zone);
1842 	if (lj == NULL) {
1843 		error = EAGAIN;
1844 		goto done;
1845 	}
1846 
1847 	lj->lioj_flags = 0;
1848 	lj->lioj_buffer_count = 0;
1849 	lj->lioj_buffer_finished_count = 0;
1850 	lj->lioj_queue_count = 0;
1851 	lj->lioj_queue_finished_count = 0;
1852 	lj->lioj_ki = ki;
1853 
1854 	/*
1855 	 * Setup signal.
1856 	 */
1857 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1858 		error = copyin(uap->sig, &lj->lioj_signal,
1859 		    sizeof(lj->lioj_signal));
1860 		if (error) {
1861 			zfree(aiolio_zone, lj);
1862 			goto done;
1863 		}
1864 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1865 			zfree(aiolio_zone, lj);
1866 			error = EINVAL;
1867 			goto done;
1868 		}
1869 		lj->lioj_flags |= LIOJ_SIGNAL;
1870 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1871 	} else
1872 		lj->lioj_flags &= ~LIOJ_SIGNAL;
1873 
1874 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1875 	/*
1876 	 * Get pointers to the list of I/O requests.
1877 	 */
1878 	nerror = 0;
1879 	nentqueued = 0;
1880 	cbptr = uap->acb_list;
1881 	for (i = 0; i < uap->nent; i++) {
1882 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1883 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1884 			error = _aio_aqueue(iocb, lj, 0);
1885 			if (error == 0)
1886 				nentqueued++;
1887 			else
1888 				nerror++;
1889 		}
1890 	}
1891 
1892 	/*
1893 	 * If we haven't queued any, then just return error.
1894 	 */
1895 	if (nentqueued == 0) {
1896 		error = 0;
1897 		goto done;
1898 	}
1899 
1900 	/*
1901 	 * Calculate the appropriate error return.
1902 	 */
1903 	runningcode = 0;
1904 	if (nerror)
1905 		runningcode = EIO;
1906 
1907 	if (uap->mode == LIO_WAIT) {
1908 		int command, found, jobref;
1909 
1910 		for (;;) {
1911 			found = 0;
1912 			for (i = 0; i < uap->nent; i++) {
1913 				/*
1914 				 * Fetch address of the control buf pointer in
1915 				 * user space.
1916 				 */
1917 				iocb = (struct aiocb *)
1918 				    (intptr_t)fuword(&cbptr[i]);
1919 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1920 				    == 0))
1921 					continue;
1922 
1923 				/*
1924 				 * Fetch the associated command from user space.
1925 				 */
1926 				command = fuword(&iocb->aio_lio_opcode);
1927 				if (command == LIO_NOP) {
1928 					found++;
1929 					continue;
1930 				}
1931 
1932 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1933 
1934 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1935 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1936 					    == jobref) {
1937 						if (cb->uaiocb.aio_lio_opcode
1938 						    == LIO_WRITE) {
1939 							lp->lwp_ru.ru_oublock +=
1940 							    cb->outputcharge;
1941 							cb->outputcharge = 0;
1942 						} else if (cb->uaiocb.aio_lio_opcode
1943 						    == LIO_READ) {
1944 							lp->lwp_ru.ru_inblock +=
1945 							    cb->inputcharge;
1946 							cb->inputcharge = 0;
1947 						}
1948 						found++;
1949 						break;
1950 					}
1951 				}
1952 
1953 				crit_enter();
1954 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1955 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1956 					    == jobref) {
1957 						found++;
1958 						break;
1959 					}
1960 				}
1961 				crit_exit();
1962 			}
1963 
1964 			/*
1965 			 * If all I/Os have been disposed of, then we can
1966 			 * return.
1967 			 */
1968 			if (found == nentqueued) {
1969 				error = runningcode;
1970 				goto done;
1971 			}
1972 
1973 			ki->kaio_flags |= KAIO_WAKEUP;
1974 			error = tsleep(p, PCATCH, "aiospn", 0);
1975 
1976 			if (error == EINTR) {
1977 				goto done;
1978 			} else if (error == EWOULDBLOCK) {
1979 				error = EAGAIN;
1980 				goto done;
1981 			}
1982 		}
1983 	}
1984 
1985 	error = runningcode;
1986 done:
1987 	rel_mplock();
1988 	return (error);
1989 #endif /* VFS_AIO */
1990 }
1991 
1992 #ifdef VFS_AIO
1993 /*
1994  * This is a weird hack so that we can post a signal.  It is safe to do so from
1995  * a timeout routine, but *not* from an interrupt routine.
1996  */
1997 static void
1998 process_signal(void *aioj)
1999 {
2000 	struct aiocblist *aiocbe = aioj;
2001 	struct aio_liojob *lj = aiocbe->lio;
2002 	struct aiocb *cb = &aiocbe->uaiocb;
2003 
2004 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2005 	    (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2006 		ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2007 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2008 	}
2009 
2010 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2011 		ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2012 }
2013 
2014 /*
2015  * Interrupt handler for physio, performs the necessary process wakeups, and
2016  * signals.
2017  */
2018 static void
2019 aio_physwakeup(struct bio *bio)
2020 {
2021 	struct buf *bp = bio->bio_buf;
2022 	struct aiocblist *aiocbe;
2023 	struct proc *p;
2024 	struct kaioinfo *ki;
2025 	struct aio_liojob *lj;
2026 
2027 	aiocbe = bio->bio_caller_info2.ptr;
2028 
2029 	if (aiocbe) {
2030 		p = bio->bio_caller_info1.ptr;
2031 
2032 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2033 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2034 		aiocbe->uaiocb._aiocb_private.error = 0;
2035 		aiocbe->jobflags |= AIOCBLIST_DONE;
2036 
2037 		if (bp->b_flags & B_ERROR)
2038 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2039 
2040 		lj = aiocbe->lio;
2041 		if (lj) {
2042 			lj->lioj_buffer_finished_count++;
2043 
2044 			/*
2045 			 * wakeup/signal if all of the interrupt jobs are done.
2046 			 */
2047 			if (lj->lioj_buffer_finished_count ==
2048 			    lj->lioj_buffer_count) {
2049 				/*
2050 				 * Post a signal if it is called for.
2051 				 */
2052 				if ((lj->lioj_flags &
2053 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2054 				    LIOJ_SIGNAL) {
2055 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2056 					callout_reset(&aiocbe->timeout, 0,
2057 							process_signal, aiocbe);
2058 				}
2059 			}
2060 		}
2061 
2062 		ki = p->p_aioinfo;
2063 		if (ki) {
2064 			ki->kaio_buffer_finished_count++;
2065 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2066 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2067 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2068 
2069 			KNOTE(&aiocbe->klist, 0);
2070 			/* Do the wakeup. */
2071 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2072 				ki->kaio_flags &= ~KAIO_WAKEUP;
2073 				wakeup(p);
2074 			}
2075 		}
2076 
2077 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2078 			callout_reset(&aiocbe->timeout, 0,
2079 					process_signal, aiocbe);
2080 		}
2081 	}
2082 	biodone_sync(bio);
2083 }
2084 #endif /* VFS_AIO */
2085 
2086 /*
2087  * syscall - wait for the next completion of an aio request
2088  *
2089  * MPALMOSTSAFE
2090  */
2091 int
2092 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2093 {
2094 #ifndef VFS_AIO
2095 	return ENOSYS;
2096 #else
2097 	struct proc *p = curproc;
2098 	struct lwp *lp = curthread->td_lwp;
2099 	struct timeval atv;
2100 	struct timespec ts;
2101 	struct kaioinfo *ki;
2102 	struct aiocblist *cb = NULL;
2103 	int error, timo;
2104 
2105 	suword(uap->aiocbp, (int)NULL);
2106 
2107 	timo = 0;
2108 	if (uap->timeout) {
2109 		/* Get timespec struct. */
2110 		error = copyin(uap->timeout, &ts, sizeof(ts));
2111 		if (error)
2112 			return error;
2113 
2114 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2115 			return (EINVAL);
2116 
2117 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2118 		if (itimerfix(&atv))
2119 			return (EINVAL);
2120 		timo = tvtohz_high(&atv);
2121 	}
2122 
2123 	ki = p->p_aioinfo;
2124 	if (ki == NULL)
2125 		return EAGAIN;
2126 
2127 	get_mplock();
2128 
2129 	for (;;) {
2130 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2131 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2132 			uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2133 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2134 				lp->lwp_ru.ru_oublock +=
2135 				    cb->outputcharge;
2136 				cb->outputcharge = 0;
2137 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2138 				lp->lwp_ru.ru_inblock += cb->inputcharge;
2139 				cb->inputcharge = 0;
2140 			}
2141 			aio_free_entry(cb);
2142 			error = cb->uaiocb._aiocb_private.error;
2143 			break;
2144 		}
2145 
2146 		crit_enter();
2147  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2148 			crit_exit();
2149 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2150 			uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2151 			aio_free_entry(cb);
2152 			error = cb->uaiocb._aiocb_private.error;
2153 			break;
2154 		}
2155 
2156 		ki->kaio_flags |= KAIO_WAKEUP;
2157 		error = tsleep(p, PCATCH, "aiowc", timo);
2158 		crit_exit();
2159 
2160 		if (error == ERESTART) {
2161 			error = EINTR;
2162 			break;
2163 		}
2164 		if (error < 0)
2165 			break;
2166 		if (error == EINTR)
2167 			break;
2168 		if (error == EWOULDBLOCK) {
2169 			error = EAGAIN;
2170 			break;
2171 		}
2172 	}
2173 	rel_mplock();
2174 	return (error);
2175 #endif /* VFS_AIO */
2176 }
2177 
2178 #ifndef VFS_AIO
2179 static int
2180 filt_aioattach(struct knote *kn)
2181 {
2182 
2183 	return (ENXIO);
2184 }
2185 
2186 struct filterops aio_filtops =
2187 	{ 0, filt_aioattach, NULL, NULL };
2188 
2189 #else
2190 /* kqueue attach function */
2191 static int
2192 filt_aioattach(struct knote *kn)
2193 {
2194 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2195 
2196 	/*
2197 	 * The aiocbe pointer must be validated before using it, so
2198 	 * registration is restricted to the kernel; the user cannot
2199 	 * set EV_FLAG1.
2200 	 */
2201 	if ((kn->kn_flags & EV_FLAG1) == 0)
2202 		return (EPERM);
2203 	kn->kn_flags &= ~EV_FLAG1;
2204 
2205 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2206 
2207 	return (0);
2208 }
2209 
2210 /* kqueue detach function */
2211 static void
2212 filt_aiodetach(struct knote *kn)
2213 {
2214 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2215 
2216 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2217 }
2218 
2219 /* kqueue filter function */
2220 /*ARGSUSED*/
2221 static int
2222 filt_aio(struct knote *kn, long hint)
2223 {
2224 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2225 
2226 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2227 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2228 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2229 		return (0);
2230 	kn->kn_flags |= EV_EOF;
2231 	return (1);
2232 }
2233 
2234 struct filterops aio_filtops =
2235 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
2236 #endif /* VFS_AIO */
2237