xref: /freebsd/sys/kern/vfs_aio.c (revision cec50dea12481dc578c0805c887ab2097e1c06c5)
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 
17 /*
18  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
19  */
20 
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
23 
24 #include <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/malloc.h>
27 #include <sys/bio.h>
28 #include <sys/buf.h>
29 #include <sys/eventhandler.h>
30 #include <sys/sysproto.h>
31 #include <sys/filedesc.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/kthread.h>
35 #include <sys/fcntl.h>
36 #include <sys/file.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/unistd.h>
41 #include <sys/proc.h>
42 #include <sys/resourcevar.h>
43 #include <sys/signalvar.h>
44 #include <sys/protosw.h>
45 #include <sys/socketvar.h>
46 #include <sys/syscall.h>
47 #include <sys/sysent.h>
48 #include <sys/sysctl.h>
49 #include <sys/sx.h>
50 #include <sys/vnode.h>
51 #include <sys/conf.h>
52 #include <sys/event.h>
53 
54 #include <posix4/posix4.h>
55 #include <vm/vm.h>
56 #include <vm/vm_extern.h>
57 #include <vm/pmap.h>
58 #include <vm/vm_map.h>
59 #include <vm/uma.h>
60 #include <sys/aio.h>
61 
62 #include "opt_vfs_aio.h"
63 
64 NET_NEEDS_GIANT("aio");
65 
66 /*
67  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
68  * overflow.
69  */
70 static	long jobrefid;
71 
72 #define JOBST_NULL		0x0
73 #define JOBST_JOBQGLOBAL	0x2
74 #define JOBST_JOBRUNNING	0x3
75 #define JOBST_JOBFINISHED	0x4
76 #define	JOBST_JOBQBUF		0x5
77 #define	JOBST_JOBBFINISHED	0x6
78 
79 #ifndef MAX_AIO_PER_PROC
80 #define MAX_AIO_PER_PROC	32
81 #endif
82 
83 #ifndef MAX_AIO_QUEUE_PER_PROC
84 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
85 #endif
86 
87 #ifndef MAX_AIO_PROCS
88 #define MAX_AIO_PROCS		32
89 #endif
90 
91 #ifndef MAX_AIO_QUEUE
92 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
93 #endif
94 
95 #ifndef TARGET_AIO_PROCS
96 #define TARGET_AIO_PROCS	4
97 #endif
98 
99 #ifndef MAX_BUF_AIO
100 #define MAX_BUF_AIO		16
101 #endif
102 
103 #ifndef AIOD_TIMEOUT_DEFAULT
104 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
105 #endif
106 
107 #ifndef AIOD_LIFETIME_DEFAULT
108 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
109 #endif
110 
111 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
112 
113 static int max_aio_procs = MAX_AIO_PROCS;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
115 	CTLFLAG_RW, &max_aio_procs, 0,
116 	"Maximum number of kernel threads to use for handling async IO ");
117 
118 static int num_aio_procs = 0;
119 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
120 	CTLFLAG_RD, &num_aio_procs, 0,
121 	"Number of presently active kernel threads for async IO");
122 
123 /*
124  * The code will adjust the actual number of AIO processes towards this
125  * number when it gets a chance.
126  */
127 static int target_aio_procs = TARGET_AIO_PROCS;
128 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
129 	0, "Preferred number of ready kernel threads for async IO");
130 
131 static int max_queue_count = MAX_AIO_QUEUE;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
133     "Maximum number of aio requests to queue, globally");
134 
135 static int num_queue_count = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
137     "Number of queued aio requests");
138 
139 static int num_buf_aio = 0;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
141     "Number of aio requests presently handled by the buf subsystem");
142 
143 /* Number of async I/O thread in the process of being started */
144 /* XXX This should be local to _aio_aqueue() */
145 static int num_aio_resv_start = 0;
146 
147 static int aiod_timeout;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
149     "Timeout value for synchronous aio operations");
150 
151 static int aiod_lifetime;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
153     "Maximum lifetime for idle aiod");
154 
155 static int unloadable = 0;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
157     "Allow unload of aio (not recommended)");
158 
159 
160 static int max_aio_per_proc = MAX_AIO_PER_PROC;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
162     0, "Maximum active aio requests per process (stored in the process)");
163 
164 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
166     &max_aio_queue_per_proc, 0,
167     "Maximum queued aio requests per process (stored in the process)");
168 
169 static int max_buf_aio = MAX_BUF_AIO;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
171     "Maximum buf aio requests per process (stored in the process)");
172 
173 struct aiocblist {
174 	TAILQ_ENTRY(aiocblist) list;	/* List of jobs */
175 	TAILQ_ENTRY(aiocblist) plist;	/* List of jobs for proc */
176 	int	jobflags;
177 	int	jobstate;
178 	int	inputcharge;
179 	int	outputcharge;
180 	struct	callout_handle timeouthandle;
181 	struct	buf *bp;		/* Buffer pointer */
182 	struct	proc *userproc;		/* User process */ /* Not td! */
183 	struct  ucred *cred;		/* Active credential when created */
184 	struct	file *fd_file;		/* Pointer to file structure */
185 	struct	aio_liojob *lio;	/* Optional lio job */
186 	struct	aiocb *uuaiocb;		/* Pointer in userspace of aiocb */
187 	struct	knlist klist;		/* list of knotes */
188 	struct	aiocb uaiocb;		/* Kernel I/O control block */
189 };
190 
191 /* jobflags */
192 #define AIOCBLIST_RUNDOWN	0x4
193 #define AIOCBLIST_DONE		0x10
194 
195 /*
196  * AIO process info
197  */
198 #define AIOP_FREE	0x1			/* proc on free queue */
199 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
200 
201 struct aiothreadlist {
202 	int aiothreadflags;			/* AIO proc flags */
203 	TAILQ_ENTRY(aiothreadlist) list;	/* List of processes */
204 	struct thread *aiothread;		/* The AIO thread */
205 };
206 
207 /*
208  * data-structure for lio signal management
209  */
210 struct aio_liojob {
211 	int	lioj_flags;
212 	int	lioj_buffer_count;
213 	int	lioj_buffer_finished_count;
214 	int	lioj_queue_count;
215 	int	lioj_queue_finished_count;
216 	struct	sigevent lioj_signal;	/* signal on all I/O done */
217 	TAILQ_ENTRY(aio_liojob) lioj_list;
218 	struct	kaioinfo *lioj_ki;
219 };
220 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
221 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
222 
223 /*
224  * per process aio data structure
225  */
226 struct kaioinfo {
227 	int	kaio_flags;		/* per process kaio flags */
228 	int	kaio_maxactive_count;	/* maximum number of AIOs */
229 	int	kaio_active_count;	/* number of currently used AIOs */
230 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
231 	int	kaio_queue_count;	/* size of AIO queue */
232 	int	kaio_ballowed_count;	/* maximum number of buffers */
233 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
234 	int	kaio_buffer_count;	/* number of physio buffers */
235 	int	kaio_buffer_finished_count; /* count of I/O done */
236 	struct 	proc *kaio_p;		/* process that uses this kaio block */
237 	TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
238 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* job queue for process */
239 	TAILQ_HEAD(,aiocblist) kaio_jobdone;	/* done queue for process */
240 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* buffer job queue for process */
241 	TAILQ_HEAD(,aiocblist) kaio_bufdone;	/* buffer done queue for process */
242 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;	/* queue for aios waiting on sockets */
243 };
244 
245 #define KAIO_RUNDOWN	0x1	/* process is being run down */
246 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
247 
248 static TAILQ_HEAD(,aiothreadlist) aio_activeproc;	/* Active daemons */
249 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* Idle daemons */
250 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
251 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
252 
253 static void	aio_init_aioinfo(struct proc *p);
254 static void	aio_onceonly(void);
255 static int	aio_free_entry(struct aiocblist *aiocbe);
256 static void	aio_process(struct aiocblist *aiocbe);
257 static int	aio_newproc(void);
258 static int	aio_aqueue(struct thread *td, struct aiocb *job, int type);
259 static void	aio_physwakeup(struct buf *bp);
260 static void	aio_proc_rundown(void *arg, struct proc *p);
261 static int	aio_fphysio(struct aiocblist *aiocbe);
262 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
263 static void	aio_daemon(void *uproc);
264 static void	aio_swake_cb(struct socket *, struct sockbuf *);
265 static int	aio_unload(void);
266 static void	process_signal(void *aioj);
267 static int	filt_aioattach(struct knote *kn);
268 static void	filt_aiodetach(struct knote *kn);
269 static int	filt_aio(struct knote *kn, long hint);
270 
271 /*
272  * Zones for:
273  * 	kaio	Per process async io info
274  *	aiop	async io thread data
275  *	aiocb	async io jobs
276  *	aiol	list io job pointer - internal to aio_suspend XXX
277  *	aiolio	list io jobs
278  */
279 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
280 
281 /* kqueue filters for aio */
282 static struct filterops aio_filtops =
283 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
284 
285 static eventhandler_tag exit_tag, exec_tag;
286 
287 /*
288  * Main operations function for use as a kernel module.
289  */
290 static int
291 aio_modload(struct module *module, int cmd, void *arg)
292 {
293 	int error = 0;
294 
295 	switch (cmd) {
296 	case MOD_LOAD:
297 		aio_onceonly();
298 		break;
299 	case MOD_UNLOAD:
300 		error = aio_unload();
301 		break;
302 	case MOD_SHUTDOWN:
303 		break;
304 	default:
305 		error = EINVAL;
306 		break;
307 	}
308 	return (error);
309 }
310 
311 static moduledata_t aio_mod = {
312 	"aio",
313 	&aio_modload,
314 	NULL
315 };
316 
317 SYSCALL_MODULE_HELPER(aio_return);
318 SYSCALL_MODULE_HELPER(aio_suspend);
319 SYSCALL_MODULE_HELPER(aio_cancel);
320 SYSCALL_MODULE_HELPER(aio_error);
321 SYSCALL_MODULE_HELPER(aio_read);
322 SYSCALL_MODULE_HELPER(aio_write);
323 SYSCALL_MODULE_HELPER(aio_waitcomplete);
324 SYSCALL_MODULE_HELPER(lio_listio);
325 
326 DECLARE_MODULE(aio, aio_mod,
327 	SI_SUB_VFS, SI_ORDER_ANY);
328 MODULE_VERSION(aio, 1);
329 
330 /*
331  * Startup initialization
332  */
333 static void
334 aio_onceonly(void)
335 {
336 
337 	/* XXX: should probably just use so->callback */
338 	aio_swake = &aio_swake_cb;
339 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
340 	    EVENTHANDLER_PRI_ANY);
341 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown, NULL,
342 	    EVENTHANDLER_PRI_ANY);
343 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
344 	TAILQ_INIT(&aio_freeproc);
345 	TAILQ_INIT(&aio_activeproc);
346 	TAILQ_INIT(&aio_jobs);
347 	TAILQ_INIT(&aio_bufjobs);
348 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
349 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
350 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
351 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
352 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
353 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
354 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
355 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
356 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL,
357 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
358 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
359 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
360 	jobrefid = 1;
361 	async_io_version = _POSIX_VERSION;
362 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
363 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
364 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
365 }
366 
367 /*
368  * Callback for unload of AIO when used as a module.
369  */
370 static int
371 aio_unload(void)
372 {
373 	int error;
374 
375 	/*
376 	 * XXX: no unloads by default, it's too dangerous.
377 	 * perhaps we could do it if locked out callers and then
378 	 * did an aio_proc_rundown() on each process.
379 	 */
380 	if (!unloadable)
381 		return (EOPNOTSUPP);
382 
383 	error = kqueue_del_filteropts(EVFILT_AIO);
384 	if (error)
385 		return error;
386 
387 	async_io_version = 0;
388 	aio_swake = NULL;
389 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
390 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
391 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
392 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
393 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
394 	return (0);
395 }
396 
397 /*
398  * Init the per-process aioinfo structure.  The aioinfo limits are set
399  * per-process for user limit (resource) management.
400  */
401 static void
402 aio_init_aioinfo(struct proc *p)
403 {
404 	struct kaioinfo *ki;
405 
406 	if (p->p_aioinfo == NULL) {
407 		ki = uma_zalloc(kaio_zone, M_WAITOK);
408 		p->p_aioinfo = ki;
409 		ki->kaio_flags = 0;
410 		ki->kaio_maxactive_count = max_aio_per_proc;
411 		ki->kaio_active_count = 0;
412 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
413 		ki->kaio_queue_count = 0;
414 		ki->kaio_ballowed_count = max_buf_aio;
415 		ki->kaio_buffer_count = 0;
416 		ki->kaio_buffer_finished_count = 0;
417 		ki->kaio_p = p;
418 		TAILQ_INIT(&ki->kaio_jobdone);
419 		TAILQ_INIT(&ki->kaio_jobqueue);
420 		TAILQ_INIT(&ki->kaio_bufdone);
421 		TAILQ_INIT(&ki->kaio_bufqueue);
422 		TAILQ_INIT(&ki->kaio_liojoblist);
423 		TAILQ_INIT(&ki->kaio_sockqueue);
424 	}
425 
426 	while (num_aio_procs < target_aio_procs)
427 		aio_newproc();
428 }
429 
430 /*
431  * Free a job entry.  Wait for completion if it is currently active, but don't
432  * delay forever.  If we delay, we return a flag that says that we have to
433  * restart the queue scan.
434  */
435 static int
436 aio_free_entry(struct aiocblist *aiocbe)
437 {
438 	struct kaioinfo *ki;
439 	struct aio_liojob *lj;
440 	struct proc *p;
441 	int error;
442 	int s;
443 
444 	if (aiocbe->jobstate == JOBST_NULL)
445 		panic("aio_free_entry: freeing already free job");
446 
447 	p = aiocbe->userproc;
448 	ki = p->p_aioinfo;
449 	lj = aiocbe->lio;
450 	if (ki == NULL)
451 		panic("aio_free_entry: missing p->p_aioinfo");
452 
453 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
454 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
455 		tsleep(aiocbe, PRIBIO, "jobwai", 0);
456 	}
457 	if (aiocbe->bp == NULL) {
458 		if (ki->kaio_queue_count <= 0)
459 			panic("aio_free_entry: process queue size <= 0");
460 		if (num_queue_count <= 0)
461 			panic("aio_free_entry: system wide queue size <= 0");
462 
463 		if (lj) {
464 			lj->lioj_queue_count--;
465 			if (aiocbe->jobflags & AIOCBLIST_DONE)
466 				lj->lioj_queue_finished_count--;
467 		}
468 		ki->kaio_queue_count--;
469 		if (aiocbe->jobflags & AIOCBLIST_DONE)
470 			ki->kaio_queue_finished_count--;
471 		num_queue_count--;
472 	} else {
473 		if (lj) {
474 			lj->lioj_buffer_count--;
475 			if (aiocbe->jobflags & AIOCBLIST_DONE)
476 				lj->lioj_buffer_finished_count--;
477 		}
478 		if (aiocbe->jobflags & AIOCBLIST_DONE)
479 			ki->kaio_buffer_finished_count--;
480 		ki->kaio_buffer_count--;
481 		num_buf_aio--;
482 	}
483 
484 	/* aiocbe is going away, we need to destroy any knotes */
485 	/* XXXKSE Note the thread here is used to eventually find the
486 	 * owning process again, but it is also used to do a fo_close
487 	 * and that requires the thread. (but does it require the
488 	 * OWNING thread? (or maybe the running thread?)
489 	 * There is a semantic problem here...
490 	 */
491 	knlist_clear(&aiocbe->klist, 0); /* XXXKSE */
492 
493 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
494 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
495 		ki->kaio_flags &= ~KAIO_WAKEUP;
496 		wakeup(p);
497 	}
498 
499 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
500 		if ((error = aio_fphysio(aiocbe)) != 0)
501 			return (error);
502 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
503 			panic("aio_free_entry: invalid physio finish-up state");
504 		s = splbio();
505 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
506 		splx(s);
507 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
508 		s = splnet();
509 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
510 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
511 		splx(s);
512 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
513 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
514 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
515 		s = splbio();
516 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
517 		splx(s);
518 		if (aiocbe->bp) {
519 			vunmapbuf(aiocbe->bp);
520 			relpbuf(aiocbe->bp, NULL);
521 			aiocbe->bp = NULL;
522 		}
523 	}
524 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
525 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
526 		uma_zfree(aiolio_zone, lj);
527 	}
528 	aiocbe->jobstate = JOBST_NULL;
529 	untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
530 	fdrop(aiocbe->fd_file, curthread);
531 	crfree(aiocbe->cred);
532 	uma_zfree(aiocb_zone, aiocbe);
533 	return (0);
534 }
535 
536 /*
537  * Rundown the jobs for a given process.
538  */
539 static void
540 aio_proc_rundown(void *arg, struct proc *p)
541 {
542 	int s;
543 	struct kaioinfo *ki;
544 	struct aio_liojob *lj, *ljn;
545 	struct aiocblist *aiocbe, *aiocbn;
546 	struct file *fp;
547 	struct socket *so;
548 
549 	ki = p->p_aioinfo;
550 	if (ki == NULL)
551 		return;
552 
553 	mtx_lock(&Giant);
554 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
555 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
556 	    ki->kaio_buffer_finished_count)) {
557 		ki->kaio_flags |= KAIO_RUNDOWN;
558 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
559 			break;
560 	}
561 
562 	/*
563 	 * Move any aio ops that are waiting on socket I/O to the normal job
564 	 * queues so they are cleaned up with any others.
565 	 */
566 	s = splnet();
567 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
568 	    aiocbn) {
569 		aiocbn = TAILQ_NEXT(aiocbe, plist);
570 		fp = aiocbe->fd_file;
571 		if (fp != NULL) {
572 			so = fp->f_data;
573 			TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
574 			if (TAILQ_EMPTY(&so->so_aiojobq)) {
575 				SOCKBUF_LOCK(&so->so_snd);
576 				so->so_snd.sb_flags &= ~SB_AIO;
577 				SOCKBUF_UNLOCK(&so->so_snd);
578 				SOCKBUF_LOCK(&so->so_rcv);
579 				so->so_rcv.sb_flags &= ~SB_AIO;
580 				SOCKBUF_UNLOCK(&so->so_rcv);
581 			}
582 		}
583 		TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
584 		TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
585 		TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
586 	}
587 	splx(s);
588 
589 restart1:
590 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
591 		aiocbn = TAILQ_NEXT(aiocbe, plist);
592 		if (aio_free_entry(aiocbe))
593 			goto restart1;
594 	}
595 
596 restart2:
597 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
598 	    aiocbn) {
599 		aiocbn = TAILQ_NEXT(aiocbe, plist);
600 		if (aio_free_entry(aiocbe))
601 			goto restart2;
602 	}
603 
604 /*
605  * Note the use of lots of splbio here, trying to avoid splbio for long chains
606  * of I/O.  Probably unnecessary.
607  */
608 restart3:
609 	s = splbio();
610 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
611 		ki->kaio_flags |= KAIO_WAKEUP;
612 		tsleep(p, PRIBIO, "aioprn", 0);
613 		splx(s);
614 		goto restart3;
615 	}
616 	splx(s);
617 
618 restart4:
619 	s = splbio();
620 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
621 		aiocbn = TAILQ_NEXT(aiocbe, plist);
622 		if (aio_free_entry(aiocbe)) {
623 			splx(s);
624 			goto restart4;
625 		}
626 	}
627 	splx(s);
628 
629 	/*
630 	 * If we've slept, jobs might have moved from one queue to another.
631 	 * Retry rundown if we didn't manage to empty the queues.
632 	 */
633 	if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
634 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
635 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
636 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
637 		goto restart1;
638 
639 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
640 		ljn = TAILQ_NEXT(lj, lioj_list);
641 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
642 		    0)) {
643 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
644 			uma_zfree(aiolio_zone, lj);
645 		} else {
646 #ifdef DIAGNOSTIC
647 			printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
648 			    "QF:%d\n", lj->lioj_buffer_count,
649 			    lj->lioj_buffer_finished_count,
650 			    lj->lioj_queue_count,
651 			    lj->lioj_queue_finished_count);
652 #endif
653 		}
654 	}
655 
656 	uma_zfree(kaio_zone, ki);
657 	p->p_aioinfo = NULL;
658 	mtx_unlock(&Giant);
659 }
660 
661 /*
662  * Select a job to run (called by an AIO daemon).
663  */
664 static struct aiocblist *
665 aio_selectjob(struct aiothreadlist *aiop)
666 {
667 	int s;
668 	struct aiocblist *aiocbe;
669 	struct kaioinfo *ki;
670 	struct proc *userp;
671 
672 	s = splnet();
673 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
674 	    TAILQ_NEXT(aiocbe, list)) {
675 		userp = aiocbe->userproc;
676 		ki = userp->p_aioinfo;
677 
678 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
679 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
680 			splx(s);
681 			return (aiocbe);
682 		}
683 	}
684 	splx(s);
685 
686 	return (NULL);
687 }
688 
689 /*
690  * The AIO processing activity.  This is the code that does the I/O request for
691  * the non-physio version of the operations.  The normal vn operations are used,
692  * and this code should work in all instances for every type of file, including
693  * pipes, sockets, fifos, and regular files.
694  */
695 static void
696 aio_process(struct aiocblist *aiocbe)
697 {
698 	struct ucred *td_savedcred;
699 	struct thread *td;
700 	struct proc *mycp;
701 	struct aiocb *cb;
702 	struct file *fp;
703 	struct uio auio;
704 	struct iovec aiov;
705 	int cnt;
706 	int error;
707 	int oublock_st, oublock_end;
708 	int inblock_st, inblock_end;
709 
710 	td = curthread;
711 	td_savedcred = td->td_ucred;
712 	td->td_ucred = aiocbe->cred;
713 	mycp = td->td_proc;
714 	cb = &aiocbe->uaiocb;
715 	fp = aiocbe->fd_file;
716 
717 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
718 	aiov.iov_len = cb->aio_nbytes;
719 
720 	auio.uio_iov = &aiov;
721 	auio.uio_iovcnt = 1;
722 	auio.uio_offset = cb->aio_offset;
723 	auio.uio_resid = cb->aio_nbytes;
724 	cnt = cb->aio_nbytes;
725 	auio.uio_segflg = UIO_USERSPACE;
726 	auio.uio_td = td;
727 
728 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
729 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
730 	/*
731 	 * _aio_aqueue() acquires a reference to the file that is
732 	 * released in aio_free_entry().
733 	 */
734 	if (cb->aio_lio_opcode == LIO_READ) {
735 		auio.uio_rw = UIO_READ;
736 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
737 	} else {
738 		auio.uio_rw = UIO_WRITE;
739 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
740 	}
741 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
742 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
743 
744 	aiocbe->inputcharge = inblock_end - inblock_st;
745 	aiocbe->outputcharge = oublock_end - oublock_st;
746 
747 	if ((error) && (auio.uio_resid != cnt)) {
748 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
749 			error = 0;
750 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
751 			PROC_LOCK(aiocbe->userproc);
752 			psignal(aiocbe->userproc, SIGPIPE);
753 			PROC_UNLOCK(aiocbe->userproc);
754 		}
755 	}
756 
757 	cnt -= auio.uio_resid;
758 	cb->_aiocb_private.error = error;
759 	cb->_aiocb_private.status = cnt;
760 	td->td_ucred = td_savedcred;
761 }
762 
763 /*
764  * The AIO daemon, most of the actual work is done in aio_process,
765  * but the setup (and address space mgmt) is done in this routine.
766  */
767 static void
768 aio_daemon(void *uproc)
769 {
770 	int s;
771 	struct aio_liojob *lj;
772 	struct aiocb *cb;
773 	struct aiocblist *aiocbe;
774 	struct aiothreadlist *aiop;
775 	struct kaioinfo *ki;
776 	struct proc *curcp, *mycp, *userp;
777 	struct vmspace *myvm, *tmpvm;
778 	struct thread *td = curthread;
779 	struct pgrp *newpgrp;
780 	struct session *newsess;
781 
782 	mtx_lock(&Giant);
783 	/*
784 	 * Local copies of curproc (cp) and vmspace (myvm)
785 	 */
786 	mycp = td->td_proc;
787 	myvm = mycp->p_vmspace;
788 
789 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
790 
791 	/*
792 	 * Allocate and ready the aio control info.  There is one aiop structure
793 	 * per daemon.
794 	 */
795 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
796 	aiop->aiothread = td;
797 	aiop->aiothreadflags |= AIOP_FREE;
798 
799 	s = splnet();
800 
801 	/*
802 	 * Place thread (lightweight process) onto the AIO free thread list.
803 	 */
804 	if (TAILQ_EMPTY(&aio_freeproc))
805 		wakeup(&aio_freeproc);
806 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
807 
808 	splx(s);
809 
810 	/*
811 	 * Get rid of our current filedescriptors.  AIOD's don't need any
812 	 * filedescriptors, except as temporarily inherited from the client.
813 	 */
814 	fdfree(td);
815 
816 	mtx_unlock(&Giant);
817 	/* The daemon resides in its own pgrp. */
818 	MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
819 		M_WAITOK | M_ZERO);
820 	MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
821 		M_WAITOK | M_ZERO);
822 
823 	sx_xlock(&proctree_lock);
824 	enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
825 	sx_xunlock(&proctree_lock);
826 	mtx_lock(&Giant);
827 
828 	/*
829 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
830 	 * and creating too many daemons.)
831 	 */
832 	wakeup(mycp);
833 
834 	for (;;) {
835 		/*
836 		 * curcp is the current daemon process context.
837 		 * userp is the current user process context.
838 		 */
839 		curcp = mycp;
840 
841 		/*
842 		 * Take daemon off of free queue
843 		 */
844 		if (aiop->aiothreadflags & AIOP_FREE) {
845 			s = splnet();
846 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
847 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
848 			aiop->aiothreadflags &= ~AIOP_FREE;
849 			splx(s);
850 		}
851 		aiop->aiothreadflags &= ~AIOP_SCHED;
852 
853 		/*
854 		 * Check for jobs.
855 		 */
856 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
857 			cb = &aiocbe->uaiocb;
858 			userp = aiocbe->userproc;
859 
860 			aiocbe->jobstate = JOBST_JOBRUNNING;
861 
862 			/*
863 			 * Connect to process address space for user program.
864 			 */
865 			if (userp != curcp) {
866 				/*
867 				 * Save the current address space that we are
868 				 * connected to.
869 				 */
870 				tmpvm = mycp->p_vmspace;
871 
872 				/*
873 				 * Point to the new user address space, and
874 				 * refer to it.
875 				 */
876 				mycp->p_vmspace = userp->p_vmspace;
877 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
878 
879 				/* Activate the new mapping. */
880 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
881 
882 				/*
883 				 * If the old address space wasn't the daemons
884 				 * own address space, then we need to remove the
885 				 * daemon's reference from the other process
886 				 * that it was acting on behalf of.
887 				 */
888 				if (tmpvm != myvm) {
889 					vmspace_free(tmpvm);
890 				}
891 				curcp = userp;
892 			}
893 
894 			ki = userp->p_aioinfo;
895 			lj = aiocbe->lio;
896 
897 			/* Account for currently active jobs. */
898 			ki->kaio_active_count++;
899 
900 			/* Do the I/O function. */
901 			aio_process(aiocbe);
902 
903 			/* Decrement the active job count. */
904 			ki->kaio_active_count--;
905 
906 			/*
907 			 * Increment the completion count for wakeup/signal
908 			 * comparisons.
909 			 */
910 			aiocbe->jobflags |= AIOCBLIST_DONE;
911 			ki->kaio_queue_finished_count++;
912 			if (lj)
913 				lj->lioj_queue_finished_count++;
914 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
915 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
916 				ki->kaio_flags &= ~KAIO_WAKEUP;
917 				wakeup(userp);
918 			}
919 
920 			s = splbio();
921 			if (lj && (lj->lioj_flags &
922 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
923 				if ((lj->lioj_queue_finished_count ==
924 				    lj->lioj_queue_count) &&
925 				    (lj->lioj_buffer_finished_count ==
926 				    lj->lioj_buffer_count)) {
927 					PROC_LOCK(userp);
928 					psignal(userp,
929 					    lj->lioj_signal.sigev_signo);
930 					PROC_UNLOCK(userp);
931 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
932 				}
933 			}
934 			splx(s);
935 
936 			aiocbe->jobstate = JOBST_JOBFINISHED;
937 
938 			s = splnet();
939 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
940 			TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
941 			splx(s);
942 			KNOTE_UNLOCKED(&aiocbe->klist, 0);
943 
944 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
945 				wakeup(aiocbe);
946 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
947 			}
948 
949 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
950 				PROC_LOCK(userp);
951 				psignal(userp, cb->aio_sigevent.sigev_signo);
952 				PROC_UNLOCK(userp);
953 			}
954 		}
955 
956 		/*
957 		 * Disconnect from user address space.
958 		 */
959 		if (curcp != mycp) {
960 			/* Get the user address space to disconnect from. */
961 			tmpvm = mycp->p_vmspace;
962 
963 			/* Get original address space for daemon. */
964 			mycp->p_vmspace = myvm;
965 
966 			/* Activate the daemon's address space. */
967 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
968 #ifdef DIAGNOSTIC
969 			if (tmpvm == myvm) {
970 				printf("AIOD: vmspace problem -- %d\n",
971 				    mycp->p_pid);
972 			}
973 #endif
974 			/* Remove our vmspace reference. */
975 			vmspace_free(tmpvm);
976 
977 			curcp = mycp;
978 		}
979 
980 		/*
981 		 * If we are the first to be put onto the free queue, wakeup
982 		 * anyone waiting for a daemon.
983 		 */
984 		s = splnet();
985 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
986 		if (TAILQ_EMPTY(&aio_freeproc))
987 			wakeup(&aio_freeproc);
988 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
989 		aiop->aiothreadflags |= AIOP_FREE;
990 		splx(s);
991 
992 		/*
993 		 * If daemon is inactive for a long time, allow it to exit,
994 		 * thereby freeing resources.
995 		 */
996 		if ((aiop->aiothreadflags & AIOP_SCHED) == 0 &&
997 		    tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) {
998 			s = splnet();
999 			if (TAILQ_EMPTY(&aio_jobs)) {
1000 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1001 				    (num_aio_procs > target_aio_procs)) {
1002 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1003 					splx(s);
1004 					uma_zfree(aiop_zone, aiop);
1005 					num_aio_procs--;
1006 #ifdef DIAGNOSTIC
1007 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1008 						printf("AIOD: bad vm refcnt for"
1009 						    " exiting daemon: %d\n",
1010 						    mycp->p_vmspace->vm_refcnt);
1011 					}
1012 #endif
1013 					kthread_exit(0);
1014 				}
1015 			}
1016 			splx(s);
1017 		}
1018 	}
1019 }
1020 
1021 /*
1022  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
1023  * AIO daemon modifies its environment itself.
1024  */
1025 static int
1026 aio_newproc(void)
1027 {
1028 	int error;
1029 	struct proc *p;
1030 
1031 	error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, 0, "aiod%d",
1032 	    num_aio_procs);
1033 	if (error)
1034 		return (error);
1035 
1036 	/*
1037 	 * Wait until daemon is started, but continue on just in case to
1038 	 * handle error conditions.
1039 	 */
1040 	error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1041 
1042 	num_aio_procs++;
1043 
1044 	return (error);
1045 }
1046 
1047 /*
1048  * Try the high-performance, low-overhead physio method for eligible
1049  * VCHR devices.  This method doesn't use an aio helper thread, and
1050  * thus has very low overhead.
1051  *
1052  * Assumes that the caller, _aio_aqueue(), has incremented the file
1053  * structure's reference count, preventing its deallocation for the
1054  * duration of this call.
1055  */
1056 static int
1057 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1058 {
1059 	int error;
1060 	struct aiocb *cb;
1061 	struct file *fp;
1062 	struct buf *bp;
1063 	struct vnode *vp;
1064 	struct kaioinfo *ki;
1065 	struct aio_liojob *lj;
1066 	int s;
1067 	int notify;
1068 
1069 	cb = &aiocbe->uaiocb;
1070 	fp = aiocbe->fd_file;
1071 
1072 	if (fp->f_type != DTYPE_VNODE)
1073 		return (-1);
1074 
1075 	vp = fp->f_vnode;
1076 
1077 	/*
1078 	 * If its not a disk, we don't want to return a positive error.
1079 	 * It causes the aio code to not fall through to try the thread
1080 	 * way when you're talking to a regular file.
1081 	 */
1082 	if (!vn_isdisk(vp, &error)) {
1083 		if (error == ENOTBLK)
1084 			return (-1);
1085 		else
1086 			return (error);
1087 	}
1088 
1089  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
1090 		return (-1);
1091 
1092 	if (cb->aio_nbytes >
1093 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1094 		return (-1);
1095 
1096 	ki = p->p_aioinfo;
1097 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1098 		return (-1);
1099 
1100 	ki->kaio_buffer_count++;
1101 
1102 	lj = aiocbe->lio;
1103 	if (lj)
1104 		lj->lioj_buffer_count++;
1105 
1106 	/* Create and build a buffer header for a transfer. */
1107 	bp = (struct buf *)getpbuf(NULL);
1108 	BUF_KERNPROC(bp);
1109 
1110 	/*
1111 	 * Get a copy of the kva from the physical buffer.
1112 	 */
1113 	bp->b_dev = vp->v_rdev;
1114 	error = 0;
1115 
1116 	bp->b_bcount = cb->aio_nbytes;
1117 	bp->b_bufsize = cb->aio_nbytes;
1118 	bp->b_iodone = aio_physwakeup;
1119 	bp->b_saveaddr = bp->b_data;
1120 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1121 	bp->b_offset = cb->aio_offset;
1122 	bp->b_iooffset = cb->aio_offset;
1123 	bp->b_blkno = btodb(cb->aio_offset);
1124 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1125 
1126 	/*
1127 	 * Bring buffer into kernel space.
1128 	 */
1129 	if (vmapbuf(bp) < 0) {
1130 		error = EFAULT;
1131 		goto doerror;
1132 	}
1133 
1134 	s = splbio();
1135 	aiocbe->bp = bp;
1136 	bp->b_caller1 = (void *)aiocbe;
1137 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1138 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1139 	aiocbe->jobstate = JOBST_JOBQBUF;
1140 	cb->_aiocb_private.status = cb->aio_nbytes;
1141 	num_buf_aio++;
1142 	bp->b_error = 0;
1143 
1144 	splx(s);
1145 
1146 	/* Perform transfer. */
1147 	DEV_STRATEGY(bp);
1148 
1149 	notify = 0;
1150 	s = splbio();
1151 
1152 	/*
1153 	 * If we had an error invoking the request, or an error in processing
1154 	 * the request before we have returned, we process it as an error in
1155 	 * transfer.  Note that such an I/O error is not indicated immediately,
1156 	 * but is returned using the aio_error mechanism.  In this case,
1157 	 * aio_suspend will return immediately.
1158 	 */
1159 	if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1160 		struct aiocb *job = aiocbe->uuaiocb;
1161 
1162 		aiocbe->uaiocb._aiocb_private.status = 0;
1163 		suword(&job->_aiocb_private.status, 0);
1164 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1165 		suword(&job->_aiocb_private.error, bp->b_error);
1166 
1167 		ki->kaio_buffer_finished_count++;
1168 
1169 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1170 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1171 			aiocbe->jobflags |= AIOCBLIST_DONE;
1172 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1173 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1174 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1175 			notify = 1;
1176 		}
1177 	}
1178 	splx(s);
1179 	if (notify)
1180 		KNOTE_UNLOCKED(&aiocbe->klist, 0);
1181 	return (0);
1182 
1183 doerror:
1184 	ki->kaio_buffer_count--;
1185 	if (lj)
1186 		lj->lioj_buffer_count--;
1187 	aiocbe->bp = NULL;
1188 	relpbuf(bp, NULL);
1189 	return (error);
1190 }
1191 
1192 /*
1193  * This waits/tests physio completion.
1194  */
1195 static int
1196 aio_fphysio(struct aiocblist *iocb)
1197 {
1198 	int s;
1199 	struct buf *bp;
1200 	int error;
1201 
1202 	bp = iocb->bp;
1203 
1204 	s = splbio();
1205 	while ((bp->b_flags & B_DONE) == 0) {
1206 		if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1207 			if ((bp->b_flags & B_DONE) == 0) {
1208 				splx(s);
1209 				return (EINPROGRESS);
1210 			} else
1211 				break;
1212 		}
1213 	}
1214 	splx(s);
1215 
1216 	/* Release mapping into kernel space. */
1217 	vunmapbuf(bp);
1218 	iocb->bp = 0;
1219 
1220 	error = 0;
1221 
1222 	/* Check for an error. */
1223 	if (bp->b_ioflags & BIO_ERROR)
1224 		error = bp->b_error;
1225 
1226 	relpbuf(bp, NULL);
1227 	return (error);
1228 }
1229 
1230 /*
1231  * Wake up aio requests that may be serviceable now.
1232  */
1233 static void
1234 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1235 {
1236 	struct aiocblist *cb,*cbn;
1237 	struct proc *p;
1238 	struct kaioinfo *ki = NULL;
1239 	int opcode, wakecount = 0;
1240 	struct aiothreadlist *aiop;
1241 
1242 	if (sb == &so->so_snd) {
1243 		opcode = LIO_WRITE;
1244 		SOCKBUF_LOCK(&so->so_snd);
1245 		so->so_snd.sb_flags &= ~SB_AIO;
1246 		SOCKBUF_UNLOCK(&so->so_snd);
1247 	} else {
1248 		opcode = LIO_READ;
1249 		SOCKBUF_LOCK(&so->so_rcv);
1250 		so->so_rcv.sb_flags &= ~SB_AIO;
1251 		SOCKBUF_UNLOCK(&so->so_rcv);
1252 	}
1253 
1254 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1255 		cbn = TAILQ_NEXT(cb, list);
1256 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1257 			p = cb->userproc;
1258 			ki = p->p_aioinfo;
1259 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1260 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1261 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1262 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1263 			wakecount++;
1264 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1265 				panic("invalid queue value");
1266 		}
1267 	}
1268 
1269 	while (wakecount--) {
1270 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1271 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1272 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1273 			aiop->aiothreadflags &= ~AIOP_FREE;
1274 			wakeup(aiop->aiothread);
1275 		}
1276 	}
1277 }
1278 
1279 /*
1280  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1281  * technique is done in this code.
1282  */
1283 static int
1284 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1285 {
1286 	struct proc *p = td->td_proc;
1287 	struct filedesc *fdp;
1288 	struct file *fp;
1289 	unsigned int fd;
1290 	struct socket *so;
1291 	int s;
1292 	int error;
1293 	int opcode, user_opcode;
1294 	struct aiocblist *aiocbe;
1295 	struct aiothreadlist *aiop;
1296 	struct kaioinfo *ki;
1297 	struct kevent kev;
1298 	struct kqueue *kq;
1299 	struct file *kq_fp;
1300 
1301 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1302 	aiocbe->inputcharge = 0;
1303 	aiocbe->outputcharge = 0;
1304 	callout_handle_init(&aiocbe->timeouthandle);
1305 	/* XXX - need a lock */
1306 	knlist_init(&aiocbe->klist, NULL);
1307 
1308 	suword(&job->_aiocb_private.status, -1);
1309 	suword(&job->_aiocb_private.error, 0);
1310 	suword(&job->_aiocb_private.kernelinfo, -1);
1311 
1312 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1313 	if (error) {
1314 		suword(&job->_aiocb_private.error, error);
1315 		uma_zfree(aiocb_zone, aiocbe);
1316 		return (error);
1317 	}
1318 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1319 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1320 		uma_zfree(aiocb_zone, aiocbe);
1321 		return (EINVAL);
1322 	}
1323 
1324 	/* Save userspace address of the job info. */
1325 	aiocbe->uuaiocb = job;
1326 
1327 	/* Get the opcode. */
1328 	user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1329 	if (type != LIO_NOP)
1330 		aiocbe->uaiocb.aio_lio_opcode = type;
1331 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1332 
1333 	/* Get the fd info for process. */
1334 	fdp = p->p_fd;
1335 
1336 	/*
1337 	 * Range check file descriptor.
1338 	 */
1339 	FILEDESC_LOCK(fdp);
1340 	fd = aiocbe->uaiocb.aio_fildes;
1341 	if (fd >= fdp->fd_nfiles) {
1342 		FILEDESC_UNLOCK(fdp);
1343 		uma_zfree(aiocb_zone, aiocbe);
1344 		if (type == 0)
1345 			suword(&job->_aiocb_private.error, EBADF);
1346 		return (EBADF);
1347 	}
1348 
1349 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1350 	if ((fp == NULL) ||
1351 	    ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 0)) ||
1352 	    ((opcode == LIO_READ) && ((fp->f_flag & FREAD) == 0))) {
1353 		FILEDESC_UNLOCK(fdp);
1354 		uma_zfree(aiocb_zone, aiocbe);
1355 		if (type == 0)
1356 			suword(&job->_aiocb_private.error, EBADF);
1357 		return (EBADF);
1358 	}
1359 	fhold(fp);
1360 	FILEDESC_UNLOCK(fdp);
1361 
1362 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1363 		error = EINVAL;
1364 		goto aqueue_fail;
1365 	}
1366 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1367 	if (error) {
1368 		error = EINVAL;
1369 		goto aqueue_fail;
1370 	}
1371 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1372 	if (jobrefid == LONG_MAX)
1373 		jobrefid = 1;
1374 	else
1375 		jobrefid++;
1376 
1377 	if (opcode == LIO_NOP) {
1378 		fdrop(fp, td);
1379 		uma_zfree(aiocb_zone, aiocbe);
1380 		if (type == 0) {
1381 			suword(&job->_aiocb_private.error, 0);
1382 			suword(&job->_aiocb_private.status, 0);
1383 			suword(&job->_aiocb_private.kernelinfo, 0);
1384 		}
1385 		return (0);
1386 	}
1387 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1388 		if (type == 0)
1389 			suword(&job->_aiocb_private.status, 0);
1390 		error = EINVAL;
1391 		goto aqueue_fail;
1392 	}
1393 
1394 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1395 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1396 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1397 	}
1398 	else {
1399 		/*
1400 		 * This method for requesting kevent-based notification won't
1401 		 * work on the alpha, since we're passing in a pointer
1402 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1403 		 * based method instead.
1404 		 */
1405 		if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1406 		    user_opcode == LIO_WRITE)
1407 			goto no_kqueue;
1408 
1409 		error = copyin((struct kevent *)(uintptr_t)user_opcode,
1410 		    &kev, sizeof(kev));
1411 		if (error)
1412 			goto aqueue_fail;
1413 	}
1414 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1415 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1416 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1417 		error = EBADF;
1418 		goto aqueue_fail;
1419 	}
1420 	kq = kq_fp->f_data;
1421 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1422 	kev.filter = EVFILT_AIO;
1423 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1424 	kev.data = (intptr_t)aiocbe;
1425 	error = kqueue_register(kq, &kev, td, 1);
1426 aqueue_fail:
1427 	if (error) {
1428 		fdrop(fp, td);
1429 		uma_zfree(aiocb_zone, aiocbe);
1430 		if (type == 0)
1431 			suword(&job->_aiocb_private.error, error);
1432 		goto done;
1433 	}
1434 no_kqueue:
1435 
1436 	suword(&job->_aiocb_private.error, EINPROGRESS);
1437 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1438 	aiocbe->userproc = p;
1439 	aiocbe->cred = crhold(td->td_ucred);
1440 	aiocbe->jobflags = 0;
1441 	aiocbe->lio = lj;
1442 	ki = p->p_aioinfo;
1443 
1444 	if (fp->f_type == DTYPE_SOCKET) {
1445 		/*
1446 		 * Alternate queueing for socket ops: Reach down into the
1447 		 * descriptor to get the socket data.  Then check to see if the
1448 		 * socket is ready to be read or written (based on the requested
1449 		 * operation).
1450 		 *
1451 		 * If it is not ready for io, then queue the aiocbe on the
1452 		 * socket, and set the flags so we get a call when sbnotify()
1453 		 * happens.
1454 		 */
1455 		so = fp->f_data;
1456 		s = splnet();
1457 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1458 		    LIO_WRITE) && (!sowriteable(so)))) {
1459 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1460 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1461 			if (opcode == LIO_READ) {
1462 				SOCKBUF_LOCK(&so->so_rcv);
1463 				so->so_rcv.sb_flags |= SB_AIO;
1464 				SOCKBUF_UNLOCK(&so->so_rcv);
1465 			} else {
1466 				SOCKBUF_LOCK(&so->so_snd);
1467 				so->so_snd.sb_flags |= SB_AIO;
1468 				SOCKBUF_UNLOCK(&so->so_snd);
1469 			}
1470 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1471 			ki->kaio_queue_count++;
1472 			num_queue_count++;
1473 			splx(s);
1474 			error = 0;
1475 			goto done;
1476 		}
1477 		splx(s);
1478 	}
1479 
1480 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1481 		goto done;
1482 	if (error > 0) {
1483 		suword(&job->_aiocb_private.status, 0);
1484 		aiocbe->uaiocb._aiocb_private.error = error;
1485 		suword(&job->_aiocb_private.error, error);
1486 		goto done;
1487 	}
1488 
1489 	/* No buffer for daemon I/O. */
1490 	aiocbe->bp = NULL;
1491 
1492 	ki->kaio_queue_count++;
1493 	if (lj)
1494 		lj->lioj_queue_count++;
1495 	s = splnet();
1496 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1497 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1498 	splx(s);
1499 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1500 
1501 	num_queue_count++;
1502 	error = 0;
1503 
1504 	/*
1505 	 * If we don't have a free AIO process, and we are below our quota, then
1506 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1507 	 * pick-up this job.  If we don't sucessfully create the new process
1508 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1509 	 * which is likely not the correct thing to do.
1510 	 */
1511 	s = splnet();
1512 retryproc:
1513 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1514 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1515 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1516 		aiop->aiothreadflags &= ~AIOP_FREE;
1517 		wakeup(aiop->aiothread);
1518 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1519 	    ((ki->kaio_active_count + num_aio_resv_start) <
1520 	    ki->kaio_maxactive_count)) {
1521 		num_aio_resv_start++;
1522 		if ((error = aio_newproc()) == 0) {
1523 			num_aio_resv_start--;
1524 			goto retryproc;
1525 		}
1526 		num_aio_resv_start--;
1527 	}
1528 	splx(s);
1529 done:
1530 	return (error);
1531 }
1532 
1533 /*
1534  * This routine queues an AIO request, checking for quotas.
1535  */
1536 static int
1537 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1538 {
1539 	struct proc *p = td->td_proc;
1540 	struct kaioinfo *ki;
1541 
1542 	if (p->p_aioinfo == NULL)
1543 		aio_init_aioinfo(p);
1544 
1545 	if (num_queue_count >= max_queue_count)
1546 		return (EAGAIN);
1547 
1548 	ki = p->p_aioinfo;
1549 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1550 		return (EAGAIN);
1551 
1552 	return _aio_aqueue(td, job, NULL, type);
1553 }
1554 
1555 /*
1556  * Support the aio_return system call, as a side-effect, kernel resources are
1557  * released.
1558  */
1559 int
1560 aio_return(struct thread *td, struct aio_return_args *uap)
1561 {
1562 	struct proc *p = td->td_proc;
1563 	int s;
1564 	long jobref;
1565 	struct aiocblist *cb, *ncb;
1566 	struct aiocb *ujob;
1567 	struct kaioinfo *ki;
1568 
1569 	ujob = uap->aiocbp;
1570 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1571 	if (jobref == -1 || jobref == 0)
1572 		return (EINVAL);
1573 
1574 	ki = p->p_aioinfo;
1575 	if (ki == NULL)
1576 		return (EINVAL);
1577 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1578 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1579 		    jobref) {
1580 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1581 				p->p_stats->p_ru.ru_oublock +=
1582 				    cb->outputcharge;
1583 				cb->outputcharge = 0;
1584 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1585 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1586 				cb->inputcharge = 0;
1587 			}
1588 			goto done;
1589 		}
1590 	}
1591 	s = splbio();
1592 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1593 		ncb = TAILQ_NEXT(cb, plist);
1594 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1595 		    == jobref) {
1596 			break;
1597 		}
1598 	}
1599 	splx(s);
1600  done:
1601 	if (cb != NULL) {
1602 		if (ujob == cb->uuaiocb) {
1603 			td->td_retval[0] =
1604 			    cb->uaiocb._aiocb_private.status;
1605 		} else
1606 			td->td_retval[0] = EFAULT;
1607 		aio_free_entry(cb);
1608 		return (0);
1609 	}
1610 	return (EINVAL);
1611 }
1612 
1613 /*
1614  * Allow a process to wakeup when any of the I/O requests are completed.
1615  */
1616 int
1617 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1618 {
1619 	struct proc *p = td->td_proc;
1620 	struct timeval atv;
1621 	struct timespec ts;
1622 	struct aiocb *const *cbptr, *cbp;
1623 	struct kaioinfo *ki;
1624 	struct aiocblist *cb;
1625 	int i;
1626 	int njoblist;
1627 	int error, s, timo;
1628 	long *ijoblist;
1629 	struct aiocb **ujoblist;
1630 
1631 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1632 		return (EINVAL);
1633 
1634 	timo = 0;
1635 	if (uap->timeout) {
1636 		/* Get timespec struct. */
1637 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1638 			return (error);
1639 
1640 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1641 			return (EINVAL);
1642 
1643 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1644 		if (itimerfix(&atv))
1645 			return (EINVAL);
1646 		timo = tvtohz(&atv);
1647 	}
1648 
1649 	ki = p->p_aioinfo;
1650 	if (ki == NULL)
1651 		return (EAGAIN);
1652 
1653 	njoblist = 0;
1654 	ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1655 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1656 	cbptr = uap->aiocbp;
1657 
1658 	for (i = 0; i < uap->nent; i++) {
1659 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1660 		if (cbp == 0)
1661 			continue;
1662 		ujoblist[njoblist] = cbp;
1663 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1664 		njoblist++;
1665 	}
1666 
1667 	if (njoblist == 0) {
1668 		uma_zfree(aiol_zone, ijoblist);
1669 		uma_zfree(aiol_zone, ujoblist);
1670 		return (0);
1671 	}
1672 
1673 	error = 0;
1674 	for (;;) {
1675 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1676 			for (i = 0; i < njoblist; i++) {
1677 				if (((intptr_t)
1678 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1679 				    ijoblist[i]) {
1680 					if (ujoblist[i] != cb->uuaiocb)
1681 						error = EINVAL;
1682 					uma_zfree(aiol_zone, ijoblist);
1683 					uma_zfree(aiol_zone, ujoblist);
1684 					return (error);
1685 				}
1686 			}
1687 		}
1688 
1689 		s = splbio();
1690 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1691 		    TAILQ_NEXT(cb, plist)) {
1692 			for (i = 0; i < njoblist; i++) {
1693 				if (((intptr_t)
1694 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1695 				    ijoblist[i]) {
1696 					splx(s);
1697 					if (ujoblist[i] != cb->uuaiocb)
1698 						error = EINVAL;
1699 					uma_zfree(aiol_zone, ijoblist);
1700 					uma_zfree(aiol_zone, ujoblist);
1701 					return (error);
1702 				}
1703 			}
1704 		}
1705 
1706 		ki->kaio_flags |= KAIO_WAKEUP;
1707 		error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1708 		splx(s);
1709 
1710 		if (error == ERESTART || error == EINTR) {
1711 			uma_zfree(aiol_zone, ijoblist);
1712 			uma_zfree(aiol_zone, ujoblist);
1713 			return (EINTR);
1714 		} else if (error == EWOULDBLOCK) {
1715 			uma_zfree(aiol_zone, ijoblist);
1716 			uma_zfree(aiol_zone, ujoblist);
1717 			return (EAGAIN);
1718 		}
1719 	}
1720 
1721 /* NOTREACHED */
1722 	return (EINVAL);
1723 }
1724 
1725 /*
1726  * aio_cancel cancels any non-physio aio operations not currently in
1727  * progress.
1728  */
1729 int
1730 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1731 {
1732 	struct proc *p = td->td_proc;
1733 	struct kaioinfo *ki;
1734 	struct aiocblist *cbe, *cbn;
1735 	struct file *fp;
1736 	struct filedesc *fdp;
1737 	struct socket *so;
1738 	struct proc *po;
1739 	int s,error;
1740 	int cancelled=0;
1741 	int notcancelled=0;
1742 	struct vnode *vp;
1743 
1744 	fdp = p->p_fd;
1745 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1746 	    (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1747 		return (EBADF);
1748 
1749 	if (fp->f_type == DTYPE_VNODE) {
1750 		vp = fp->f_vnode;
1751 
1752 		if (vn_isdisk(vp,&error)) {
1753 			td->td_retval[0] = AIO_NOTCANCELED;
1754 			return (0);
1755 		}
1756 	} else if (fp->f_type == DTYPE_SOCKET) {
1757 		so = fp->f_data;
1758 
1759 		s = splnet();
1760 
1761 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1762 			cbn = TAILQ_NEXT(cbe, list);
1763 			if ((uap->aiocbp == NULL) ||
1764 				(uap->aiocbp == cbe->uuaiocb) ) {
1765 				po = cbe->userproc;
1766 				ki = po->p_aioinfo;
1767 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1768 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1769 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1770 				if (ki->kaio_flags & KAIO_WAKEUP) {
1771 					wakeup(po);
1772 				}
1773 				cbe->jobstate = JOBST_JOBFINISHED;
1774 				cbe->uaiocb._aiocb_private.status=-1;
1775 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1776 				cancelled++;
1777 /* XXX cancelled, knote? */
1778 				if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1779 				    SIGEV_SIGNAL) {
1780 					PROC_LOCK(cbe->userproc);
1781 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1782 					PROC_UNLOCK(cbe->userproc);
1783 				}
1784 				if (uap->aiocbp)
1785 					break;
1786 			}
1787 		}
1788 		splx(s);
1789 
1790 		if ((cancelled) && (uap->aiocbp)) {
1791 			td->td_retval[0] = AIO_CANCELED;
1792 			return (0);
1793 		}
1794 	}
1795 	ki=p->p_aioinfo;
1796 	if (ki == NULL)
1797 		goto done;
1798 	s = splnet();
1799 
1800 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1801 		cbn = TAILQ_NEXT(cbe, plist);
1802 
1803 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1804 		    ((uap->aiocbp == NULL ) ||
1805 		     (uap->aiocbp == cbe->uuaiocb))) {
1806 
1807 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1808 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1809 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1810 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1811 				    plist);
1812 				cancelled++;
1813 				ki->kaio_queue_finished_count++;
1814 				cbe->jobstate = JOBST_JOBFINISHED;
1815 				cbe->uaiocb._aiocb_private.status = -1;
1816 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1817 /* XXX cancelled, knote? */
1818 				if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1819 				    SIGEV_SIGNAL) {
1820 					PROC_LOCK(cbe->userproc);
1821 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1822 					PROC_UNLOCK(cbe->userproc);
1823 				}
1824 			} else {
1825 				notcancelled++;
1826 			}
1827 		}
1828 	}
1829 	splx(s);
1830 done:
1831 	if (notcancelled) {
1832 		td->td_retval[0] = AIO_NOTCANCELED;
1833 		return (0);
1834 	}
1835 	if (cancelled) {
1836 		td->td_retval[0] = AIO_CANCELED;
1837 		return (0);
1838 	}
1839 	td->td_retval[0] = AIO_ALLDONE;
1840 
1841 	return (0);
1842 }
1843 
1844 /*
1845  * aio_error is implemented in the kernel level for compatibility purposes only.
1846  * For a user mode async implementation, it would be best to do it in a userland
1847  * subroutine.
1848  */
1849 int
1850 aio_error(struct thread *td, struct aio_error_args *uap)
1851 {
1852 	struct proc *p = td->td_proc;
1853 	int s;
1854 	struct aiocblist *cb;
1855 	struct kaioinfo *ki;
1856 	long jobref;
1857 
1858 	ki = p->p_aioinfo;
1859 	if (ki == NULL)
1860 		return (EINVAL);
1861 
1862 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1863 	if ((jobref == -1) || (jobref == 0))
1864 		return (EINVAL);
1865 
1866 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1867 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1868 		    jobref) {
1869 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1870 			return (0);
1871 		}
1872 	}
1873 
1874 	s = splnet();
1875 
1876 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1877 	    plist)) {
1878 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1879 		    jobref) {
1880 			td->td_retval[0] = EINPROGRESS;
1881 			splx(s);
1882 			return (0);
1883 		}
1884 	}
1885 
1886 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1887 	    plist)) {
1888 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1889 		    jobref) {
1890 			td->td_retval[0] = EINPROGRESS;
1891 			splx(s);
1892 			return (0);
1893 		}
1894 	}
1895 	splx(s);
1896 
1897 	s = splbio();
1898 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1899 	    plist)) {
1900 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1901 		    jobref) {
1902 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1903 			splx(s);
1904 			return (0);
1905 		}
1906 	}
1907 
1908 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1909 	    plist)) {
1910 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1911 		    jobref) {
1912 			td->td_retval[0] = EINPROGRESS;
1913 			splx(s);
1914 			return (0);
1915 		}
1916 	}
1917 	splx(s);
1918 
1919 #if (0)
1920 	/*
1921 	 * Hack for lio.
1922 	 */
1923 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1924 	if (status == -1)
1925 		return fuword(&uap->aiocbp->_aiocb_private.error);
1926 #endif
1927 	return (EINVAL);
1928 }
1929 
1930 /* syscall - asynchronous read from a file (REALTIME) */
1931 int
1932 aio_read(struct thread *td, struct aio_read_args *uap)
1933 {
1934 
1935 	return aio_aqueue(td, uap->aiocbp, LIO_READ);
1936 }
1937 
1938 /* syscall - asynchronous write to a file (REALTIME) */
1939 int
1940 aio_write(struct thread *td, struct aio_write_args *uap)
1941 {
1942 
1943 	return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1944 }
1945 
1946 /* syscall - list directed I/O (REALTIME) */
1947 int
1948 lio_listio(struct thread *td, struct lio_listio_args *uap)
1949 {
1950 	struct proc *p = td->td_proc;
1951 	int nent, nentqueued;
1952 	struct aiocb *iocb, * const *cbptr;
1953 	struct aiocblist *cb;
1954 	struct kaioinfo *ki;
1955 	struct aio_liojob *lj;
1956 	int error, runningcode;
1957 	int nerror;
1958 	int i;
1959 	int s;
1960 
1961 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1962 		return (EINVAL);
1963 
1964 	nent = uap->nent;
1965 	if (nent < 0 || nent > AIO_LISTIO_MAX)
1966 		return (EINVAL);
1967 
1968 	if (p->p_aioinfo == NULL)
1969 		aio_init_aioinfo(p);
1970 
1971 	if ((nent + num_queue_count) > max_queue_count)
1972 		return (EAGAIN);
1973 
1974 	ki = p->p_aioinfo;
1975 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1976 		return (EAGAIN);
1977 
1978 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
1979 	if (!lj)
1980 		return (EAGAIN);
1981 
1982 	lj->lioj_flags = 0;
1983 	lj->lioj_buffer_count = 0;
1984 	lj->lioj_buffer_finished_count = 0;
1985 	lj->lioj_queue_count = 0;
1986 	lj->lioj_queue_finished_count = 0;
1987 	lj->lioj_ki = ki;
1988 
1989 	/*
1990 	 * Setup signal.
1991 	 */
1992 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1993 		error = copyin(uap->sig, &lj->lioj_signal,
1994 		    sizeof(lj->lioj_signal));
1995 		if (error) {
1996 			uma_zfree(aiolio_zone, lj);
1997 			return (error);
1998 		}
1999 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2000 			uma_zfree(aiolio_zone, lj);
2001 			return (EINVAL);
2002 		}
2003 		lj->lioj_flags |= LIOJ_SIGNAL;
2004 	}
2005 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2006 	/*
2007 	 * Get pointers to the list of I/O requests.
2008 	 */
2009 	nerror = 0;
2010 	nentqueued = 0;
2011 	cbptr = uap->acb_list;
2012 	for (i = 0; i < uap->nent; i++) {
2013 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
2014 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
2015 			error = _aio_aqueue(td, iocb, lj, 0);
2016 			if (error == 0)
2017 				nentqueued++;
2018 			else
2019 				nerror++;
2020 		}
2021 	}
2022 
2023 	/*
2024 	 * If we haven't queued any, then just return error.
2025 	 */
2026 	if (nentqueued == 0)
2027 		return (0);
2028 
2029 	/*
2030 	 * Calculate the appropriate error return.
2031 	 */
2032 	runningcode = 0;
2033 	if (nerror)
2034 		runningcode = EIO;
2035 
2036 	if (uap->mode == LIO_WAIT) {
2037 		int command, found, jobref;
2038 
2039 		for (;;) {
2040 			found = 0;
2041 			for (i = 0; i < uap->nent; i++) {
2042 				/*
2043 				 * Fetch address of the control buf pointer in
2044 				 * user space.
2045 				 */
2046 				iocb = (struct aiocb *)
2047 				    (intptr_t)fuword(&cbptr[i]);
2048 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2049 				    == 0))
2050 					continue;
2051 
2052 				/*
2053 				 * Fetch the associated command from user space.
2054 				 */
2055 				command = fuword(&iocb->aio_lio_opcode);
2056 				if (command == LIO_NOP) {
2057 					found++;
2058 					continue;
2059 				}
2060 
2061 				jobref =
2062 				    fuword(&iocb->_aiocb_private.kernelinfo);
2063 
2064 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2065 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2066 					    == jobref) {
2067 						if (cb->uaiocb.aio_lio_opcode
2068 						    == LIO_WRITE) {
2069 							p->p_stats->p_ru.ru_oublock
2070 							    +=
2071 							    cb->outputcharge;
2072 							cb->outputcharge = 0;
2073 						} else if (cb->uaiocb.aio_lio_opcode
2074 						    == LIO_READ) {
2075 							p->p_stats->p_ru.ru_inblock
2076 							    += cb->inputcharge;
2077 							cb->inputcharge = 0;
2078 						}
2079 						found++;
2080 						break;
2081 					}
2082 				}
2083 
2084 				s = splbio();
2085 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2086 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2087 					    == jobref) {
2088 						found++;
2089 						break;
2090 					}
2091 				}
2092 				splx(s);
2093 			}
2094 
2095 			/*
2096 			 * If all I/Os have been disposed of, then we can
2097 			 * return.
2098 			 */
2099 			if (found == nentqueued)
2100 				return (runningcode);
2101 
2102 			ki->kaio_flags |= KAIO_WAKEUP;
2103 			error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2104 
2105 			if (error == EINTR)
2106 				return (EINTR);
2107 			else if (error == EWOULDBLOCK)
2108 				return (EAGAIN);
2109 		}
2110 	}
2111 
2112 	return (runningcode);
2113 }
2114 
2115 /*
2116  * This is a weird hack so that we can post a signal.  It is safe to do so from
2117  * a timeout routine, but *not* from an interrupt routine.
2118  */
2119 static void
2120 process_signal(void *aioj)
2121 {
2122 	struct aiocblist *aiocbe = aioj;
2123 	struct aio_liojob *lj = aiocbe->lio;
2124 	struct aiocb *cb = &aiocbe->uaiocb;
2125 
2126 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2127 		(lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2128 		PROC_LOCK(lj->lioj_ki->kaio_p);
2129 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2130 		PROC_UNLOCK(lj->lioj_ki->kaio_p);
2131 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2132 	}
2133 
2134 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2135 		PROC_LOCK(aiocbe->userproc);
2136 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2137 		PROC_UNLOCK(aiocbe->userproc);
2138 	}
2139 }
2140 
2141 /*
2142  * Interrupt handler for physio, performs the necessary process wakeups, and
2143  * signals.
2144  */
2145 static void
2146 aio_physwakeup(struct buf *bp)
2147 {
2148 	struct aiocblist *aiocbe;
2149 	struct proc *p;
2150 	struct kaioinfo *ki;
2151 	struct aio_liojob *lj;
2152 
2153 	wakeup(bp);
2154 
2155 	aiocbe = (struct aiocblist *)bp->b_caller1;
2156 	if (aiocbe) {
2157 		p = aiocbe->userproc;
2158 
2159 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2160 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2161 		aiocbe->uaiocb._aiocb_private.error = 0;
2162 		aiocbe->jobflags |= AIOCBLIST_DONE;
2163 
2164 		if (bp->b_ioflags & BIO_ERROR)
2165 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2166 
2167 		lj = aiocbe->lio;
2168 		if (lj) {
2169 			lj->lioj_buffer_finished_count++;
2170 
2171 			/*
2172 			 * wakeup/signal if all of the interrupt jobs are done.
2173 			 */
2174 			if (lj->lioj_buffer_finished_count ==
2175 			    lj->lioj_buffer_count) {
2176 				/*
2177 				 * Post a signal if it is called for.
2178 				 */
2179 				if ((lj->lioj_flags &
2180 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2181 				    LIOJ_SIGNAL) {
2182 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2183 					aiocbe->timeouthandle =
2184 						timeout(process_signal,
2185 							aiocbe, 0);
2186 				}
2187 			}
2188 		}
2189 
2190 		ki = p->p_aioinfo;
2191 		if (ki) {
2192 			ki->kaio_buffer_finished_count++;
2193 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2194 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2195 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2196 
2197 			KNOTE_UNLOCKED(&aiocbe->klist, 0);
2198 			/* Do the wakeup. */
2199 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2200 				ki->kaio_flags &= ~KAIO_WAKEUP;
2201 				wakeup(p);
2202 			}
2203 		}
2204 
2205 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2206 			aiocbe->timeouthandle =
2207 				timeout(process_signal, aiocbe, 0);
2208 	}
2209 }
2210 
2211 /* syscall - wait for the next completion of an aio request */
2212 int
2213 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2214 {
2215 	struct proc *p = td->td_proc;
2216 	struct timeval atv;
2217 	struct timespec ts;
2218 	struct kaioinfo *ki;
2219 	struct aiocblist *cb = NULL;
2220 	int error, s, timo;
2221 
2222 	suword(uap->aiocbp, (int)NULL);
2223 
2224 	timo = 0;
2225 	if (uap->timeout) {
2226 		/* Get timespec struct. */
2227 		error = copyin(uap->timeout, &ts, sizeof(ts));
2228 		if (error)
2229 			return (error);
2230 
2231 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2232 			return (EINVAL);
2233 
2234 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2235 		if (itimerfix(&atv))
2236 			return (EINVAL);
2237 		timo = tvtohz(&atv);
2238 	}
2239 
2240 	ki = p->p_aioinfo;
2241 	if (ki == NULL)
2242 		return (EAGAIN);
2243 
2244 	for (;;) {
2245 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2246 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2247 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2248 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2249 				p->p_stats->p_ru.ru_oublock +=
2250 				    cb->outputcharge;
2251 				cb->outputcharge = 0;
2252 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2253 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2254 				cb->inputcharge = 0;
2255 			}
2256 			aio_free_entry(cb);
2257 			return (cb->uaiocb._aiocb_private.error);
2258 		}
2259 
2260 		s = splbio();
2261  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2262 			splx(s);
2263 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2264 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2265 			aio_free_entry(cb);
2266 			return (cb->uaiocb._aiocb_private.error);
2267 		}
2268 
2269 		ki->kaio_flags |= KAIO_WAKEUP;
2270 		error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2271 		splx(s);
2272 
2273 		if (error == ERESTART)
2274 			return (EINTR);
2275 		else if (error < 0)
2276 			return (error);
2277 		else if (error == EINTR)
2278 			return (EINTR);
2279 		else if (error == EWOULDBLOCK)
2280 			return (EAGAIN);
2281 	}
2282 }
2283 
2284 /* kqueue attach function */
2285 static int
2286 filt_aioattach(struct knote *kn)
2287 {
2288 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2289 
2290 	/*
2291 	 * The aiocbe pointer must be validated before using it, so
2292 	 * registration is restricted to the kernel; the user cannot
2293 	 * set EV_FLAG1.
2294 	 */
2295 	if ((kn->kn_flags & EV_FLAG1) == 0)
2296 		return (EPERM);
2297 	kn->kn_flags &= ~EV_FLAG1;
2298 
2299 	knlist_add(&aiocbe->klist, kn, 0);
2300 
2301 	return (0);
2302 }
2303 
2304 /* kqueue detach function */
2305 static void
2306 filt_aiodetach(struct knote *kn)
2307 {
2308 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2309 
2310 	knlist_remove(&aiocbe->klist, kn, 0);
2311 }
2312 
2313 /* kqueue filter function */
2314 /*ARGSUSED*/
2315 static int
2316 filt_aio(struct knote *kn, long hint)
2317 {
2318 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2319 
2320 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2321 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2322 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2323 		return (0);
2324 	kn->kn_flags |= EV_EOF;
2325 	return (1);
2326 }
2327