xref: /dflybsd-src/sys/kern/vfs_aio.c (revision 97fecd825dd1a70c628493b90a9b1b1724f151df)
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_offset = cb->aio_offset;
926 
927 	/* Bring buffer into kernel space. */
928 	if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
929 		error = EFAULT;
930 		goto doerror;
931 	}
932 
933 	crit_enter();
934 
935 	aiocbe->bp = bp;
936 	bp->b_bio1.bio_caller_info2.ptr = aiocbe;
937 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
938 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
939 	aiocbe->jobstate = JOBST_JOBQBUF;
940 	cb->_aiocb_private.status = cb->aio_nbytes;
941 	num_buf_aio++;
942 	bp->b_error = 0;
943 
944 	crit_exit();
945 
946 	/*
947 	 * Perform the transfer.  vn_strategy must be used even though we
948 	 * know we have a device in order to deal with requests which exceed
949 	 * device DMA limitations.
950 	 */
951 	vn_strategy(vp, &bp->b_bio1);
952 
953 	notify = 0;
954 	crit_enter();
955 
956 	/*
957 	 * If we had an error invoking the request, or an error in processing
958 	 * the request before we have returned, we process it as an error in
959 	 * transfer.  Note that such an I/O error is not indicated immediately,
960 	 * but is returned using the aio_error mechanism.  In this case,
961 	 * aio_suspend will return immediately.
962 	 */
963 	if (bp->b_error || (bp->b_flags & B_ERROR)) {
964 		struct aiocb *job = aiocbe->uuaiocb;
965 
966 		aiocbe->uaiocb._aiocb_private.status = 0;
967 		suword(&job->_aiocb_private.status, 0);
968 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
969 		suword(&job->_aiocb_private.error, bp->b_error);
970 
971 		ki->kaio_buffer_finished_count++;
972 
973 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
974 			aiocbe->jobstate = JOBST_JOBBFINISHED;
975 			aiocbe->jobflags |= AIOCBLIST_DONE;
976 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
977 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
978 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
979 			notify = 1;
980 		}
981 	}
982 	crit_exit();
983 	if (notify)
984 		KNOTE(&aiocbe->klist, 0);
985 	return 0;
986 
987 doerror:
988 	ki->kaio_buffer_count--;
989 	if (lj)
990 		lj->lioj_buffer_count--;
991 	aiocbe->bp = NULL;
992 	relpbuf(bp, NULL);
993 	return error;
994 }
995 
996 /*
997  * This waits/tests physio completion.
998  */
999 static int
1000 aio_fphysio(struct aiocblist *iocb)
1001 {
1002 	struct buf *bp;
1003 	int error;
1004 
1005 	bp = iocb->bp;
1006 
1007 	crit_enter();
1008 	while (bp->b_cmd != BUF_CMD_DONE) {
1009 		if (tsleep(bp, 0, "physstr", aiod_timeout)) {
1010 			if (bp->b_cmd != BUF_CMD_DONE) {
1011 				crit_exit();
1012 				return EINPROGRESS;
1013 			} else {
1014 				break;
1015 			}
1016 		}
1017 	}
1018 	crit_exit();
1019 
1020 	/* Release mapping into kernel space. */
1021 	vunmapbuf(bp);
1022 	iocb->bp = 0;
1023 
1024 	error = 0;
1025 
1026 	/* Check for an error. */
1027 	if (bp->b_flags & B_ERROR)
1028 		error = bp->b_error;
1029 
1030 	relpbuf(bp, NULL);
1031 	return (error);
1032 }
1033 #endif /* VFS_AIO */
1034 
1035 /*
1036  * Wake up aio requests that may be serviceable now.
1037  */
1038 void
1039 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1040 {
1041 #ifndef VFS_AIO
1042 	return;
1043 #else
1044 	struct aiocblist *cb,*cbn;
1045 	struct proc *p;
1046 	struct kaioinfo *ki = NULL;
1047 	int opcode, wakecount = 0;
1048 	struct aioproclist *aiop;
1049 
1050 	if (ssb == &so->so_snd) {
1051 		opcode = LIO_WRITE;
1052 		so->so_snd.ssb_flags &= ~SSB_AIO;
1053 	} else {
1054 		opcode = LIO_READ;
1055 		so->so_rcv.ssb_flags &= ~SSB_AIO;
1056 	}
1057 
1058 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1059 		cbn = TAILQ_NEXT(cb, list);
1060 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1061 			p = cb->userproc;
1062 			ki = p->p_aioinfo;
1063 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1064 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1065 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1066 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1067 			wakecount++;
1068 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1069 				panic("invalid queue value");
1070 		}
1071 	}
1072 
1073 	while (wakecount--) {
1074 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1075 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1076 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1077 			aiop->aioprocflags &= ~AIOP_FREE;
1078 			wakeup(aiop->aioproc);
1079 		}
1080 	}
1081 #endif /* VFS_AIO */
1082 }
1083 
1084 #ifdef VFS_AIO
1085 /*
1086  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1087  * technique is done in this code.
1088  */
1089 static int
1090 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1091 {
1092 	struct proc *p = curproc;
1093 	struct filedesc *fdp;
1094 	struct file *fp;
1095 	unsigned int fd;
1096 	struct socket *so;
1097 	int error;
1098 	int opcode, user_opcode;
1099 	struct aiocblist *aiocbe;
1100 	struct aioproclist *aiop;
1101 	struct kaioinfo *ki;
1102 	struct kevent kev;
1103 	struct kqueue *kq;
1104 	struct file *kq_fp;
1105 
1106 	if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1107 		TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1108 	else
1109 		aiocbe = zalloc (aiocb_zone);
1110 
1111 	aiocbe->inputcharge = 0;
1112 	aiocbe->outputcharge = 0;
1113 	callout_init(&aiocbe->timeout);
1114 	SLIST_INIT(&aiocbe->klist);
1115 
1116 	suword(&job->_aiocb_private.status, -1);
1117 	suword(&job->_aiocb_private.error, 0);
1118 	suword(&job->_aiocb_private.kernelinfo, -1);
1119 
1120 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1121 	if (error) {
1122 		suword(&job->_aiocb_private.error, error);
1123 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1124 		return error;
1125 	}
1126 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1127 	    !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1128 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1129 		return EINVAL;
1130 	}
1131 
1132 	/* Save userspace address of the job info. */
1133 	aiocbe->uuaiocb = job;
1134 
1135 	/* Get the opcode. */
1136 	user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1137 	if (type != LIO_NOP)
1138 		aiocbe->uaiocb.aio_lio_opcode = type;
1139 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1140 
1141 	/* Get the fd info for process. */
1142 	fdp = p->p_fd;
1143 
1144 	/*
1145 	 * Range check file descriptor.
1146 	 */
1147 	fd = aiocbe->uaiocb.aio_fildes;
1148 	if (fd >= fdp->fd_nfiles) {
1149 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1150 		if (type == 0)
1151 			suword(&job->_aiocb_private.error, EBADF);
1152 		return EBADF;
1153 	}
1154 
1155 	fp = aiocbe->fd_file = fdp->fd_files[fd].fp;
1156 	if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1157 	    0))) {
1158 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1159 		if (type == 0)
1160 			suword(&job->_aiocb_private.error, EBADF);
1161 		return EBADF;
1162 	}
1163 	fhold(fp);
1164 
1165 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1166 		error = EINVAL;
1167 		goto aqueue_fail;
1168 	}
1169 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1170 	if (error) {
1171 		error = EINVAL;
1172 		goto aqueue_fail;
1173 	}
1174 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1175 	if (jobrefid == LONG_MAX)
1176 		jobrefid = 1;
1177 	else
1178 		jobrefid++;
1179 
1180 	if (opcode == LIO_NOP) {
1181 		fdrop(fp);
1182 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1183 		if (type == 0) {
1184 			suword(&job->_aiocb_private.error, 0);
1185 			suword(&job->_aiocb_private.status, 0);
1186 			suword(&job->_aiocb_private.kernelinfo, 0);
1187 		}
1188 		return 0;
1189 	}
1190 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1191 		if (type == 0)
1192 			suword(&job->_aiocb_private.status, 0);
1193 		error = EINVAL;
1194 		goto aqueue_fail;
1195 	}
1196 
1197 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1198 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1199 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1200 	}
1201 	else {
1202 		/*
1203 		 * This method for requesting kevent-based notification won't
1204 		 * work on the alpha, since we're passing in a pointer
1205 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1206 		 * based method instead.
1207 		 */
1208 		if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1209 		    user_opcode == LIO_WRITE)
1210 			goto no_kqueue;
1211 
1212 		error = copyin((struct kevent *)(uintptr_t)user_opcode,
1213 		    &kev, sizeof(kev));
1214 		if (error)
1215 			goto aqueue_fail;
1216 	}
1217 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1218 	    (kq_fp = fdp->fd_files[kev.ident].fp) == NULL ||
1219 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1220 		error = EBADF;
1221 		goto aqueue_fail;
1222 	}
1223 	kq = (struct kqueue *)kq_fp->f_data;
1224 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1225 	kev.filter = EVFILT_AIO;
1226 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1227 	kev.data = (intptr_t)aiocbe;
1228 	/* XXX lwp kqueue_register takes a thread, but only uses its proc */
1229 	error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread);
1230 aqueue_fail:
1231 	if (error) {
1232 		fdrop(fp);
1233 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1234 		if (type == 0)
1235 			suword(&job->_aiocb_private.error, error);
1236 		goto done;
1237 	}
1238 no_kqueue:
1239 
1240 	suword(&job->_aiocb_private.error, EINPROGRESS);
1241 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1242 	aiocbe->userproc = p;
1243 	aiocbe->jobflags = 0;
1244 	aiocbe->lio = lj;
1245 	ki = p->p_aioinfo;
1246 
1247 	if (fp->f_type == DTYPE_SOCKET) {
1248 		/*
1249 		 * Alternate queueing for socket ops: Reach down into the
1250 		 * descriptor to get the socket data.  Then check to see if the
1251 		 * socket is ready to be read or written (based on the requested
1252 		 * operation).
1253 		 *
1254 		 * If it is not ready for io, then queue the aiocbe on the
1255 		 * socket, and set the flags so we get a call when ssb_notify()
1256 		 * happens.
1257 		 */
1258 		so = (struct socket *)fp->f_data;
1259 		crit_enter();
1260 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1261 		    LIO_WRITE) && (!sowriteable(so)))) {
1262 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1263 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1264 			if (opcode == LIO_READ)
1265 				so->so_rcv.ssb_flags |= SSB_AIO;
1266 			else
1267 				so->so_snd.ssb_flags |= SSB_AIO;
1268 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1269 			ki->kaio_queue_count++;
1270 			num_queue_count++;
1271 			crit_exit();
1272 			error = 0;
1273 			goto done;
1274 		}
1275 		crit_exit();
1276 	}
1277 
1278 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1279 		goto done;
1280 	if (error > 0) {
1281 		suword(&job->_aiocb_private.status, 0);
1282 		aiocbe->uaiocb._aiocb_private.error = error;
1283 		suword(&job->_aiocb_private.error, error);
1284 		goto done;
1285 	}
1286 
1287 	/* No buffer for daemon I/O. */
1288 	aiocbe->bp = NULL;
1289 
1290 	ki->kaio_queue_count++;
1291 	if (lj)
1292 		lj->lioj_queue_count++;
1293 	crit_enter();
1294 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1295 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1296 	crit_exit();
1297 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1298 
1299 	num_queue_count++;
1300 	error = 0;
1301 
1302 	/*
1303 	 * If we don't have a free AIO process, and we are below our quota, then
1304 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1305 	 * pick-up this job.  If we don't successfully create the new process
1306 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1307 	 * which is likely not the correct thing to do.
1308 	 */
1309 	crit_enter();
1310 retryproc:
1311 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1312 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1313 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1314 		aiop->aioprocflags &= ~AIOP_FREE;
1315 		wakeup(aiop->aioproc);
1316 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1317 	    ((ki->kaio_active_count + num_aio_resv_start) <
1318 	    ki->kaio_maxactive_count)) {
1319 		num_aio_resv_start++;
1320 		if ((error = aio_newproc()) == 0) {
1321 			num_aio_resv_start--;
1322 			goto retryproc;
1323 		}
1324 		num_aio_resv_start--;
1325 	}
1326 	crit_exit();
1327 done:
1328 	return error;
1329 }
1330 
1331 /*
1332  * This routine queues an AIO request, checking for quotas.
1333  */
1334 static int
1335 aio_aqueue(struct aiocb *job, int type)
1336 {
1337 	struct proc *p = curproc;
1338 	struct kaioinfo *ki;
1339 
1340 	if (p->p_aioinfo == NULL)
1341 		aio_init_aioinfo(p);
1342 
1343 	if (num_queue_count >= max_queue_count)
1344 		return EAGAIN;
1345 
1346 	ki = p->p_aioinfo;
1347 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1348 		return EAGAIN;
1349 
1350 	return _aio_aqueue(job, NULL, type);
1351 }
1352 #endif /* VFS_AIO */
1353 
1354 /*
1355  * Support the aio_return system call, as a side-effect, kernel resources are
1356  * released.
1357  */
1358 int
1359 sys_aio_return(struct aio_return_args *uap)
1360 {
1361 #ifndef VFS_AIO
1362 	return ENOSYS;
1363 #else
1364 	struct proc *p = curproc;
1365 	struct lwp *lp = curthread->td_lwp;
1366 	long jobref;
1367 	struct aiocblist *cb, *ncb;
1368 	struct aiocb *ujob;
1369 	struct kaioinfo *ki;
1370 
1371 	ki = p->p_aioinfo;
1372 	if (ki == NULL)
1373 		return EINVAL;
1374 
1375 	ujob = uap->aiocbp;
1376 
1377 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1378 	if (jobref == -1 || jobref == 0)
1379 		return EINVAL;
1380 
1381 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1382 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1383 		    jobref) {
1384 			if (ujob == cb->uuaiocb) {
1385 				uap->sysmsg_result =
1386 				    cb->uaiocb._aiocb_private.status;
1387 			} else
1388 				uap->sysmsg_result = EFAULT;
1389 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1390 				lp->lwp_ru.ru_oublock += cb->outputcharge;
1391 				cb->outputcharge = 0;
1392 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1393 				lp->lwp_ru.ru_inblock += cb->inputcharge;
1394 				cb->inputcharge = 0;
1395 			}
1396 			aio_free_entry(cb);
1397 			return 0;
1398 		}
1399 	}
1400 	crit_enter();
1401 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1402 		ncb = TAILQ_NEXT(cb, plist);
1403 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1404 		    == jobref) {
1405 			crit_exit();
1406 			if (ujob == cb->uuaiocb) {
1407 				uap->sysmsg_result =
1408 				    cb->uaiocb._aiocb_private.status;
1409 			} else
1410 				uap->sysmsg_result = EFAULT;
1411 			aio_free_entry(cb);
1412 			return 0;
1413 		}
1414 	}
1415 	crit_exit();
1416 
1417 	return (EINVAL);
1418 #endif /* VFS_AIO */
1419 }
1420 
1421 /*
1422  * Allow a process to wakeup when any of the I/O requests are completed.
1423  */
1424 int
1425 sys_aio_suspend(struct aio_suspend_args *uap)
1426 {
1427 #ifndef VFS_AIO
1428 	return ENOSYS;
1429 #else
1430 	struct proc *p = curproc;
1431 	struct timeval atv;
1432 	struct timespec ts;
1433 	struct aiocb *const *cbptr, *cbp;
1434 	struct kaioinfo *ki;
1435 	struct aiocblist *cb;
1436 	int i;
1437 	int njoblist;
1438 	int error, timo;
1439 	long *ijoblist;
1440 	struct aiocb **ujoblist;
1441 
1442 	if (uap->nent > AIO_LISTIO_MAX)
1443 		return EINVAL;
1444 
1445 	timo = 0;
1446 	if (uap->timeout) {
1447 		/* Get timespec struct. */
1448 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1449 			return error;
1450 
1451 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1452 			return (EINVAL);
1453 
1454 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1455 		if (itimerfix(&atv))
1456 			return (EINVAL);
1457 		timo = tvtohz_high(&atv);
1458 	}
1459 
1460 	ki = p->p_aioinfo;
1461 	if (ki == NULL)
1462 		return EAGAIN;
1463 
1464 	njoblist = 0;
1465 	ijoblist = zalloc(aiol_zone);
1466 	ujoblist = zalloc(aiol_zone);
1467 	cbptr = uap->aiocbp;
1468 
1469 	for (i = 0; i < uap->nent; i++) {
1470 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1471 		if (cbp == 0)
1472 			continue;
1473 		ujoblist[njoblist] = cbp;
1474 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1475 		njoblist++;
1476 	}
1477 
1478 	if (njoblist == 0) {
1479 		zfree(aiol_zone, ijoblist);
1480 		zfree(aiol_zone, ujoblist);
1481 		return 0;
1482 	}
1483 
1484 	error = 0;
1485 	for (;;) {
1486 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1487 			for (i = 0; i < njoblist; i++) {
1488 				if (((intptr_t)
1489 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1490 				    ijoblist[i]) {
1491 					if (ujoblist[i] != cb->uuaiocb)
1492 						error = EINVAL;
1493 					zfree(aiol_zone, ijoblist);
1494 					zfree(aiol_zone, ujoblist);
1495 					return error;
1496 				}
1497 			}
1498 		}
1499 
1500 		crit_enter();
1501 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1502 		    TAILQ_NEXT(cb, plist)) {
1503 			for (i = 0; i < njoblist; i++) {
1504 				if (((intptr_t)
1505 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1506 				    ijoblist[i]) {
1507 					crit_exit();
1508 					if (ujoblist[i] != cb->uuaiocb)
1509 						error = EINVAL;
1510 					zfree(aiol_zone, ijoblist);
1511 					zfree(aiol_zone, ujoblist);
1512 					return error;
1513 				}
1514 			}
1515 		}
1516 
1517 		ki->kaio_flags |= KAIO_WAKEUP;
1518 		error = tsleep(p, PCATCH, "aiospn", timo);
1519 		crit_exit();
1520 
1521 		if (error == ERESTART || error == EINTR) {
1522 			zfree(aiol_zone, ijoblist);
1523 			zfree(aiol_zone, ujoblist);
1524 			return EINTR;
1525 		} else if (error == EWOULDBLOCK) {
1526 			zfree(aiol_zone, ijoblist);
1527 			zfree(aiol_zone, ujoblist);
1528 			return EAGAIN;
1529 		}
1530 	}
1531 
1532 /* NOTREACHED */
1533 	return EINVAL;
1534 #endif /* VFS_AIO */
1535 }
1536 
1537 /*
1538  * aio_cancel cancels any non-physio aio operations not currently in
1539  * progress.
1540  */
1541 int
1542 sys_aio_cancel(struct aio_cancel_args *uap)
1543 {
1544 #ifndef VFS_AIO
1545 	return ENOSYS;
1546 #else
1547 	struct proc *p = curproc;
1548 	struct kaioinfo *ki;
1549 	struct aiocblist *cbe, *cbn;
1550 	struct file *fp;
1551 	struct filedesc *fdp;
1552 	struct socket *so;
1553 	struct proc *po;
1554 	int error;
1555 	int cancelled=0;
1556 	int notcancelled=0;
1557 	struct vnode *vp;
1558 
1559 	fdp = p->p_fd;
1560 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1561 	    (fp = fdp->fd_files[uap->fd].fp) == NULL)
1562 		return (EBADF);
1563 
1564         if (fp->f_type == DTYPE_VNODE) {
1565 		vp = (struct vnode *)fp->f_data;
1566 
1567 		if (vn_isdisk(vp,&error)) {
1568 			uap->sysmsg_result = AIO_NOTCANCELED;
1569         	        return 0;
1570 		}
1571 	} else if (fp->f_type == DTYPE_SOCKET) {
1572 		so = (struct socket *)fp->f_data;
1573 
1574 		crit_enter();
1575 
1576 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1577 			cbn = TAILQ_NEXT(cbe, list);
1578 			if ((uap->aiocbp == NULL) ||
1579 				(uap->aiocbp == cbe->uuaiocb) ) {
1580 				po = cbe->userproc;
1581 				ki = po->p_aioinfo;
1582 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1583 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1584 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1585 				if (ki->kaio_flags & KAIO_WAKEUP) {
1586 					wakeup(po);
1587 				}
1588 				cbe->jobstate = JOBST_JOBFINISHED;
1589 				cbe->uaiocb._aiocb_private.status=-1;
1590 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1591 				cancelled++;
1592 /* XXX cancelled, knote? */
1593 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1594 				    SIGEV_SIGNAL)
1595 					ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1596 				if (uap->aiocbp)
1597 					break;
1598 			}
1599 		}
1600 		crit_exit();
1601 
1602 		if ((cancelled) && (uap->aiocbp)) {
1603 			uap->sysmsg_result = AIO_CANCELED;
1604 			return 0;
1605 		}
1606 	}
1607 	ki=p->p_aioinfo;
1608 	if (ki == NULL)
1609 		goto done;
1610 	crit_enter();
1611 
1612 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1613 		cbn = TAILQ_NEXT(cbe, plist);
1614 
1615 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1616 		    ((uap->aiocbp == NULL ) ||
1617 		     (uap->aiocbp == cbe->uuaiocb))) {
1618 
1619 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1620 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1621                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1622                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1623                                     plist);
1624 				cancelled++;
1625 				ki->kaio_queue_finished_count++;
1626 				cbe->jobstate = JOBST_JOBFINISHED;
1627 				cbe->uaiocb._aiocb_private.status = -1;
1628 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1629 /* XXX cancelled, knote? */
1630 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1631 				    SIGEV_SIGNAL)
1632 					ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1633 			} else {
1634 				notcancelled++;
1635 			}
1636 		}
1637 	}
1638 	crit_exit();
1639 done:
1640 	if (notcancelled) {
1641 		uap->sysmsg_result = AIO_NOTCANCELED;
1642 		return 0;
1643 	}
1644 	if (cancelled) {
1645 		uap->sysmsg_result = AIO_CANCELED;
1646 		return 0;
1647 	}
1648 	uap->sysmsg_result = AIO_ALLDONE;
1649 
1650 	return 0;
1651 #endif /* VFS_AIO */
1652 }
1653 
1654 /*
1655  * aio_error is implemented in the kernel level for compatibility purposes only.
1656  * For a user mode async implementation, it would be best to do it in a userland
1657  * subroutine.
1658  */
1659 int
1660 sys_aio_error(struct aio_error_args *uap)
1661 {
1662 #ifndef VFS_AIO
1663 	return ENOSYS;
1664 #else
1665 	struct proc *p = curproc;
1666 	struct aiocblist *cb;
1667 	struct kaioinfo *ki;
1668 	long jobref;
1669 
1670 	ki = p->p_aioinfo;
1671 	if (ki == NULL)
1672 		return EINVAL;
1673 
1674 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1675 	if ((jobref == -1) || (jobref == 0))
1676 		return EINVAL;
1677 
1678 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1679 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1680 		    jobref) {
1681 			uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1682 			return 0;
1683 		}
1684 	}
1685 
1686 	crit_enter();
1687 
1688 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1689 	    plist)) {
1690 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1691 		    jobref) {
1692 			uap->sysmsg_result = EINPROGRESS;
1693 			crit_exit();
1694 			return 0;
1695 		}
1696 	}
1697 
1698 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1699 	    plist)) {
1700 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1701 		    jobref) {
1702 			uap->sysmsg_result = EINPROGRESS;
1703 			crit_exit();
1704 			return 0;
1705 		}
1706 	}
1707 	crit_exit();
1708 
1709 	crit_enter();
1710 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1711 	    plist)) {
1712 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1713 		    jobref) {
1714 			uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1715 			crit_exit();
1716 			return 0;
1717 		}
1718 	}
1719 
1720 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1721 	    plist)) {
1722 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1723 		    jobref) {
1724 			uap->sysmsg_result = EINPROGRESS;
1725 			crit_exit();
1726 			return 0;
1727 		}
1728 	}
1729 	crit_exit();
1730 
1731 #if (0)
1732 	/*
1733 	 * Hack for lio.
1734 	 */
1735 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1736 	if (status == -1)
1737 		return fuword(&uap->aiocbp->_aiocb_private.error);
1738 #endif
1739 	return EINVAL;
1740 #endif /* VFS_AIO */
1741 }
1742 
1743 /* syscall - asynchronous read from a file (REALTIME) */
1744 int
1745 sys_aio_read(struct aio_read_args *uap)
1746 {
1747 #ifndef VFS_AIO
1748 	return ENOSYS;
1749 #else
1750 	return aio_aqueue(uap->aiocbp, LIO_READ);
1751 #endif /* VFS_AIO */
1752 }
1753 
1754 /* syscall - asynchronous write to a file (REALTIME) */
1755 int
1756 sys_aio_write(struct aio_write_args *uap)
1757 {
1758 #ifndef VFS_AIO
1759 	return ENOSYS;
1760 #else
1761 	return aio_aqueue(uap->aiocbp, LIO_WRITE);
1762 #endif /* VFS_AIO */
1763 }
1764 
1765 /* syscall - XXX undocumented */
1766 int
1767 sys_lio_listio(struct lio_listio_args *uap)
1768 {
1769 #ifndef VFS_AIO
1770 	return ENOSYS;
1771 #else
1772 	struct proc *p = curproc;
1773 	struct lwp *lp = curthread->td_lwp;
1774 	int nent, nentqueued;
1775 	struct aiocb *iocb, * const *cbptr;
1776 	struct aiocblist *cb;
1777 	struct kaioinfo *ki;
1778 	struct aio_liojob *lj;
1779 	int error, runningcode;
1780 	int nerror;
1781 	int i;
1782 
1783 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1784 		return EINVAL;
1785 
1786 	nent = uap->nent;
1787 	if (nent > AIO_LISTIO_MAX)
1788 		return EINVAL;
1789 
1790 	if (p->p_aioinfo == NULL)
1791 		aio_init_aioinfo(p);
1792 
1793 	if ((nent + num_queue_count) > max_queue_count)
1794 		return EAGAIN;
1795 
1796 	ki = p->p_aioinfo;
1797 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1798 		return EAGAIN;
1799 
1800 	lj = zalloc(aiolio_zone);
1801 	if (!lj)
1802 		return EAGAIN;
1803 
1804 	lj->lioj_flags = 0;
1805 	lj->lioj_buffer_count = 0;
1806 	lj->lioj_buffer_finished_count = 0;
1807 	lj->lioj_queue_count = 0;
1808 	lj->lioj_queue_finished_count = 0;
1809 	lj->lioj_ki = ki;
1810 
1811 	/*
1812 	 * Setup signal.
1813 	 */
1814 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1815 		error = copyin(uap->sig, &lj->lioj_signal,
1816 		    sizeof(lj->lioj_signal));
1817 		if (error) {
1818 			zfree(aiolio_zone, lj);
1819 			return error;
1820 		}
1821 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1822 			zfree(aiolio_zone, lj);
1823 			return EINVAL;
1824 		}
1825 		lj->lioj_flags |= LIOJ_SIGNAL;
1826 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1827 	} else
1828 		lj->lioj_flags &= ~LIOJ_SIGNAL;
1829 
1830 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1831 	/*
1832 	 * Get pointers to the list of I/O requests.
1833 	 */
1834 	nerror = 0;
1835 	nentqueued = 0;
1836 	cbptr = uap->acb_list;
1837 	for (i = 0; i < uap->nent; i++) {
1838 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1839 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1840 			error = _aio_aqueue(iocb, lj, 0);
1841 			if (error == 0)
1842 				nentqueued++;
1843 			else
1844 				nerror++;
1845 		}
1846 	}
1847 
1848 	/*
1849 	 * If we haven't queued any, then just return error.
1850 	 */
1851 	if (nentqueued == 0)
1852 		return 0;
1853 
1854 	/*
1855 	 * Calculate the appropriate error return.
1856 	 */
1857 	runningcode = 0;
1858 	if (nerror)
1859 		runningcode = EIO;
1860 
1861 	if (uap->mode == LIO_WAIT) {
1862 		int command, found, jobref;
1863 
1864 		for (;;) {
1865 			found = 0;
1866 			for (i = 0; i < uap->nent; i++) {
1867 				/*
1868 				 * Fetch address of the control buf pointer in
1869 				 * user space.
1870 				 */
1871 				iocb = (struct aiocb *)
1872 				    (intptr_t)fuword(&cbptr[i]);
1873 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1874 				    == 0))
1875 					continue;
1876 
1877 				/*
1878 				 * Fetch the associated command from user space.
1879 				 */
1880 				command = fuword(&iocb->aio_lio_opcode);
1881 				if (command == LIO_NOP) {
1882 					found++;
1883 					continue;
1884 				}
1885 
1886 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1887 
1888 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1889 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1890 					    == jobref) {
1891 						if (cb->uaiocb.aio_lio_opcode
1892 						    == LIO_WRITE) {
1893 							lp->lwp_ru.ru_oublock +=
1894 							    cb->outputcharge;
1895 							cb->outputcharge = 0;
1896 						} else if (cb->uaiocb.aio_lio_opcode
1897 						    == LIO_READ) {
1898 							lp->lwp_ru.ru_inblock +=
1899 							    cb->inputcharge;
1900 							cb->inputcharge = 0;
1901 						}
1902 						found++;
1903 						break;
1904 					}
1905 				}
1906 
1907 				crit_enter();
1908 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1909 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1910 					    == jobref) {
1911 						found++;
1912 						break;
1913 					}
1914 				}
1915 				crit_exit();
1916 			}
1917 
1918 			/*
1919 			 * If all I/Os have been disposed of, then we can
1920 			 * return.
1921 			 */
1922 			if (found == nentqueued)
1923 				return runningcode;
1924 
1925 			ki->kaio_flags |= KAIO_WAKEUP;
1926 			error = tsleep(p, PCATCH, "aiospn", 0);
1927 
1928 			if (error == EINTR)
1929 				return EINTR;
1930 			else if (error == EWOULDBLOCK)
1931 				return EAGAIN;
1932 		}
1933 	}
1934 
1935 	return runningcode;
1936 #endif /* VFS_AIO */
1937 }
1938 
1939 #ifdef VFS_AIO
1940 /*
1941  * This is a weird hack so that we can post a signal.  It is safe to do so from
1942  * a timeout routine, but *not* from an interrupt routine.
1943  */
1944 static void
1945 process_signal(void *aioj)
1946 {
1947 	struct aiocblist *aiocbe = aioj;
1948 	struct aio_liojob *lj = aiocbe->lio;
1949 	struct aiocb *cb = &aiocbe->uaiocb;
1950 
1951 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
1952 	    (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
1953 		ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1954 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1955 	}
1956 
1957 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
1958 		ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
1959 }
1960 
1961 /*
1962  * Interrupt handler for physio, performs the necessary process wakeups, and
1963  * signals.
1964  */
1965 static void
1966 aio_physwakeup(struct bio *bio)
1967 {
1968 	struct buf *bp = bio->bio_buf;
1969 	struct aiocblist *aiocbe;
1970 	struct proc *p;
1971 	struct kaioinfo *ki;
1972 	struct aio_liojob *lj;
1973 
1974 	aiocbe = bio->bio_caller_info2.ptr;
1975 
1976 	if (aiocbe) {
1977 		p = bio->bio_caller_info1.ptr;
1978 
1979 		aiocbe->jobstate = JOBST_JOBBFINISHED;
1980 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
1981 		aiocbe->uaiocb._aiocb_private.error = 0;
1982 		aiocbe->jobflags |= AIOCBLIST_DONE;
1983 
1984 		if (bp->b_flags & B_ERROR)
1985 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1986 
1987 		lj = aiocbe->lio;
1988 		if (lj) {
1989 			lj->lioj_buffer_finished_count++;
1990 
1991 			/*
1992 			 * wakeup/signal if all of the interrupt jobs are done.
1993 			 */
1994 			if (lj->lioj_buffer_finished_count ==
1995 			    lj->lioj_buffer_count) {
1996 				/*
1997 				 * Post a signal if it is called for.
1998 				 */
1999 				if ((lj->lioj_flags &
2000 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2001 				    LIOJ_SIGNAL) {
2002 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2003 					callout_reset(&aiocbe->timeout, 0,
2004 							process_signal, aiocbe);
2005 				}
2006 			}
2007 		}
2008 
2009 		ki = p->p_aioinfo;
2010 		if (ki) {
2011 			ki->kaio_buffer_finished_count++;
2012 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2013 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2014 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2015 
2016 			KNOTE(&aiocbe->klist, 0);
2017 			/* Do the wakeup. */
2018 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2019 				ki->kaio_flags &= ~KAIO_WAKEUP;
2020 				wakeup(p);
2021 			}
2022 		}
2023 
2024 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2025 			callout_reset(&aiocbe->timeout, 0,
2026 					process_signal, aiocbe);
2027 		}
2028 	}
2029 	bp->b_cmd = BUF_CMD_DONE;
2030 	wakeup(bp);
2031 }
2032 #endif /* VFS_AIO */
2033 
2034 /* syscall - wait for the next completion of an aio request */
2035 int
2036 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2037 {
2038 #ifndef VFS_AIO
2039 	return ENOSYS;
2040 #else
2041 	struct proc *p = curproc;
2042 	struct lwp *lp = curthread->td_lwp;
2043 	struct timeval atv;
2044 	struct timespec ts;
2045 	struct kaioinfo *ki;
2046 	struct aiocblist *cb = NULL;
2047 	int error, timo;
2048 
2049 	suword(uap->aiocbp, (int)NULL);
2050 
2051 	timo = 0;
2052 	if (uap->timeout) {
2053 		/* Get timespec struct. */
2054 		error = copyin(uap->timeout, &ts, sizeof(ts));
2055 		if (error)
2056 			return error;
2057 
2058 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2059 			return (EINVAL);
2060 
2061 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2062 		if (itimerfix(&atv))
2063 			return (EINVAL);
2064 		timo = tvtohz_high(&atv);
2065 	}
2066 
2067 	ki = p->p_aioinfo;
2068 	if (ki == NULL)
2069 		return EAGAIN;
2070 
2071 	for (;;) {
2072 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2073 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2074 			uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2075 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2076 				lp->lwp_ru.ru_oublock +=
2077 				    cb->outputcharge;
2078 				cb->outputcharge = 0;
2079 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2080 				lp->lwp_ru.ru_inblock += cb->inputcharge;
2081 				cb->inputcharge = 0;
2082 			}
2083 			aio_free_entry(cb);
2084 			return cb->uaiocb._aiocb_private.error;
2085 		}
2086 
2087 		crit_enter();
2088  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2089 			crit_exit();
2090 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2091 			uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2092 			aio_free_entry(cb);
2093 			return cb->uaiocb._aiocb_private.error;
2094 		}
2095 
2096 		ki->kaio_flags |= KAIO_WAKEUP;
2097 		error = tsleep(p, PCATCH, "aiowc", timo);
2098 		crit_exit();
2099 
2100 		if (error == ERESTART)
2101 			return EINTR;
2102 		else if (error < 0)
2103 			return error;
2104 		else if (error == EINTR)
2105 			return EINTR;
2106 		else if (error == EWOULDBLOCK)
2107 			return EAGAIN;
2108 	}
2109 #endif /* VFS_AIO */
2110 }
2111 
2112 #ifndef VFS_AIO
2113 static int
2114 filt_aioattach(struct knote *kn)
2115 {
2116 
2117 	return (ENXIO);
2118 }
2119 
2120 struct filterops aio_filtops =
2121 	{ 0, filt_aioattach, NULL, NULL };
2122 
2123 #else
2124 /* kqueue attach function */
2125 static int
2126 filt_aioattach(struct knote *kn)
2127 {
2128 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2129 
2130 	/*
2131 	 * The aiocbe pointer must be validated before using it, so
2132 	 * registration is restricted to the kernel; the user cannot
2133 	 * set EV_FLAG1.
2134 	 */
2135 	if ((kn->kn_flags & EV_FLAG1) == 0)
2136 		return (EPERM);
2137 	kn->kn_flags &= ~EV_FLAG1;
2138 
2139 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2140 
2141 	return (0);
2142 }
2143 
2144 /* kqueue detach function */
2145 static void
2146 filt_aiodetach(struct knote *kn)
2147 {
2148 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2149 
2150 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2151 }
2152 
2153 /* kqueue filter function */
2154 /*ARGSUSED*/
2155 static int
2156 filt_aio(struct knote *kn, long hint)
2157 {
2158 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2159 
2160 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2161 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2162 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2163 		return (0);
2164 	kn->kn_flags |= EV_EOF;
2165 	return (1);
2166 }
2167 
2168 struct filterops aio_filtops =
2169 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
2170 #endif /* VFS_AIO */
2171