xref: /freebsd/sys/kern/vfs_aio.c (revision c17d43407fe04133a94055b0dbc7ea8965654a9f)
1 /*
2  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. John S. Dyson's name may not be used to endorse or promote products
10  *    derived from this software without specific prior written permission.
11  *
12  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
13  * bad that happens because of using this software isn't the responsibility
14  * of the author.  This software is distributed AS-IS.
15  *
16  * $FreeBSD$
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/param.h>
24 #include <sys/systm.h>
25 #include <sys/malloc.h>
26 #include <sys/bio.h>
27 #include <sys/buf.h>
28 #include <sys/sysproto.h>
29 #include <sys/filedesc.h>
30 #include <sys/kernel.h>
31 #include <sys/kthread.h>
32 #include <sys/fcntl.h>
33 #include <sys/file.h>
34 #include <sys/lock.h>
35 #include <sys/mutex.h>
36 #include <sys/unistd.h>
37 #include <sys/proc.h>
38 #include <sys/resourcevar.h>
39 #include <sys/signalvar.h>
40 #include <sys/protosw.h>
41 #include <sys/socketvar.h>
42 #include <sys/syscall.h>
43 #include <sys/sysent.h>
44 #include <sys/sysctl.h>
45 #include <sys/vnode.h>
46 #include <sys/conf.h>
47 #include <sys/event.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_extern.h>
51 #include <vm/pmap.h>
52 #include <vm/vm_map.h>
53 #include <vm/uma.h>
54 #include <sys/aio.h>
55 
56 #include <machine/limits.h>
57 
58 #include "opt_vfs_aio.h"
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 unloadable = 0;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
151     "Allow unload of aio (not recommended)");
152 
153 
154 static int max_aio_per_proc = MAX_AIO_PER_PROC;
155 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
156     0, "Maximum active aio requests per process (stored in the process)");
157 
158 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
160     &max_aio_queue_per_proc, 0,
161     "Maximum queued aio requests per process (stored in the process)");
162 
163 static int max_buf_aio = MAX_BUF_AIO;
164 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
165     "Maximum buf aio requests per process (stored in the process)");
166 
167 struct aiocblist {
168         TAILQ_ENTRY(aiocblist) list;	/* List of jobs */
169         TAILQ_ENTRY(aiocblist) plist;	/* List of jobs for proc */
170         int	jobflags;
171         int	jobstate;
172 	int	inputcharge;
173 	int	outputcharge;
174 	struct	callout_handle timeouthandle;
175         struct	buf *bp;		/* Buffer pointer */
176         struct	proc *userproc;		/* User process */ /* Not td! */
177         struct	file *fd_file;		/* Pointer to file structure */
178 	struct	aiothreadlist *jobaiothread;  /* AIO process descriptor */
179         struct	aio_liojob *lio;	/* Optional lio job */
180         struct	aiocb *uuaiocb;		/* Pointer in userspace of aiocb */
181 	struct	klist klist;		/* list of knotes */
182         struct	aiocb uaiocb;		/* Kernel I/O control block */
183 };
184 
185 /* jobflags */
186 #define AIOCBLIST_RUNDOWN       0x4
187 #define AIOCBLIST_ASYNCFREE     0x8
188 #define AIOCBLIST_DONE          0x10
189 
190 /*
191  * AIO process info
192  */
193 #define AIOP_FREE	0x1			/* proc on free queue */
194 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
195 
196 struct aiothreadlist {
197 	int aiothreadflags;			/* AIO proc flags */
198 	TAILQ_ENTRY(aiothreadlist) list;	/* List of processes */
199 	struct thread *aiothread;		/* The AIO thread */
200 };
201 
202 /*
203  * data-structure for lio signal management
204  */
205 struct aio_liojob {
206 	int	lioj_flags;
207 	int	lioj_buffer_count;
208 	int	lioj_buffer_finished_count;
209 	int	lioj_queue_count;
210 	int	lioj_queue_finished_count;
211 	struct	sigevent lioj_signal;	/* signal on all I/O done */
212 	TAILQ_ENTRY(aio_liojob) lioj_list;
213 	struct	kaioinfo *lioj_ki;
214 };
215 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
216 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
217 
218 /*
219  * per process aio data structure
220  */
221 struct kaioinfo {
222 	int	kaio_flags;		/* per process kaio flags */
223 	int	kaio_maxactive_count;	/* maximum number of AIOs */
224 	int	kaio_active_count;	/* number of currently used AIOs */
225 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
226 	int	kaio_queue_count;	/* size of AIO queue */
227 	int	kaio_ballowed_count;	/* maximum number of buffers */
228 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
229 	int	kaio_buffer_count;	/* number of physio buffers */
230 	int	kaio_buffer_finished_count; /* count of I/O done */
231 	struct 	proc *kaio_p;		/* process that uses this kaio block */
232 	TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
233 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* job queue for process */
234 	TAILQ_HEAD(,aiocblist) kaio_jobdone;	/* done queue for process */
235 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* buffer job queue for process */
236 	TAILQ_HEAD(,aiocblist) kaio_bufdone;	/* buffer done queue for process */
237 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;	/* queue for aios waiting on sockets */
238 };
239 
240 #define KAIO_RUNDOWN	0x1	/* process is being run down */
241 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
242 
243 static TAILQ_HEAD(,aiothreadlist) aio_activeproc;	/* Active daemons */
244 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* Idle daemons */
245 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
246 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
247 
248 static void	aio_init_aioinfo(struct proc *p);
249 static void	aio_onceonly(void);
250 static int	aio_free_entry(struct aiocblist *aiocbe);
251 static void	aio_process(struct aiocblist *aiocbe);
252 static int	aio_newproc(void);
253 static int	aio_aqueue(struct thread *td, struct aiocb *job, int type);
254 static void	aio_physwakeup(struct buf *bp);
255 static void	aio_proc_rundown(struct proc *p);
256 static int	aio_fphysio(struct aiocblist *aiocbe);
257 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
258 static void	aio_daemon(void *uproc);
259 static void	aio_swake_cb(struct socket *, struct sockbuf *);
260 static int	aio_unload(void);
261 static void	process_signal(void *aioj);
262 static int	filt_aioattach(struct knote *kn);
263 static void	filt_aiodetach(struct knote *kn);
264 static int	filt_aio(struct knote *kn, long hint);
265 
266 /*
267  * Zones for:
268  * 	kaio	Per process async io info
269  *	aiop	async io thread data
270  *	aiocb	async io jobs
271  *	aiol	list io job pointer - internal to aio_suspend XXX
272  *	aiolio	list io jobs
273  */
274 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
275 
276 /* kqueue filters for aio */
277 static struct filterops aio_filtops =
278 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
279 
280 /*
281  * Main operations function for use as a kernel module.
282  */
283 static int
284 aio_modload(struct module *module, int cmd, void *arg)
285 {
286 	int error = 0;
287 
288 	switch (cmd) {
289 	case MOD_LOAD:
290 		aio_onceonly();
291 		break;
292 	case MOD_UNLOAD:
293 		error = aio_unload();
294 		break;
295 	case MOD_SHUTDOWN:
296 		break;
297 	default:
298 		error = EINVAL;
299 		break;
300 	}
301 	return (error);
302 }
303 
304 static moduledata_t aio_mod = {
305 	"aio",
306 	&aio_modload,
307 	NULL
308 };
309 
310 SYSCALL_MODULE_HELPER(aio_return);
311 SYSCALL_MODULE_HELPER(aio_suspend);
312 SYSCALL_MODULE_HELPER(aio_cancel);
313 SYSCALL_MODULE_HELPER(aio_error);
314 SYSCALL_MODULE_HELPER(aio_read);
315 SYSCALL_MODULE_HELPER(aio_write);
316 SYSCALL_MODULE_HELPER(aio_waitcomplete);
317 SYSCALL_MODULE_HELPER(lio_listio);
318 
319 DECLARE_MODULE(aio, aio_mod,
320 	SI_SUB_VFS, SI_ORDER_ANY);
321 MODULE_VERSION(aio, 1);
322 
323 /*
324  * Startup initialization
325  */
326 static void
327 aio_onceonly(void)
328 {
329 
330 	/* XXX: should probably just use so->callback */
331 	aio_swake = &aio_swake_cb;
332 	at_exit(aio_proc_rundown);
333 	at_exec(aio_proc_rundown);
334 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
335 	TAILQ_INIT(&aio_freeproc);
336 	TAILQ_INIT(&aio_activeproc);
337 	TAILQ_INIT(&aio_jobs);
338 	TAILQ_INIT(&aio_bufjobs);
339 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
340 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
341 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
342 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
343 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
344 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
345 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
346 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
347 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL,
348 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
349 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
350 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
351 	jobrefid = 1;
352 }
353 
354 /*
355  * Callback for unload of AIO when used as a module.
356  */
357 static int
358 aio_unload(void)
359 {
360 
361 	/*
362 	 * XXX: no unloads by default, it's too dangerous.
363 	 * perhaps we could do it if locked out callers and then
364 	 * did an aio_proc_rundown() on each process.
365 	 */
366 	if (!unloadable)
367 		return (EOPNOTSUPP);
368 
369 	aio_swake = NULL;
370 	rm_at_exit(aio_proc_rundown);
371 	rm_at_exec(aio_proc_rundown);
372 	kqueue_del_filteropts(EVFILT_AIO);
373 	return (0);
374 }
375 
376 /*
377  * Init the per-process aioinfo structure.  The aioinfo limits are set
378  * per-process for user limit (resource) management.
379  */
380 static void
381 aio_init_aioinfo(struct proc *p)
382 {
383 	struct kaioinfo *ki;
384 	if (p->p_aioinfo == NULL) {
385 		ki = uma_zalloc(kaio_zone, M_WAITOK);
386 		p->p_aioinfo = ki;
387 		ki->kaio_flags = 0;
388 		ki->kaio_maxactive_count = max_aio_per_proc;
389 		ki->kaio_active_count = 0;
390 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
391 		ki->kaio_queue_count = 0;
392 		ki->kaio_ballowed_count = max_buf_aio;
393 		ki->kaio_buffer_count = 0;
394 		ki->kaio_buffer_finished_count = 0;
395 		ki->kaio_p = p;
396 		TAILQ_INIT(&ki->kaio_jobdone);
397 		TAILQ_INIT(&ki->kaio_jobqueue);
398 		TAILQ_INIT(&ki->kaio_bufdone);
399 		TAILQ_INIT(&ki->kaio_bufqueue);
400 		TAILQ_INIT(&ki->kaio_liojoblist);
401 		TAILQ_INIT(&ki->kaio_sockqueue);
402 	}
403 
404 	while (num_aio_procs < target_aio_procs)
405 		aio_newproc();
406 }
407 
408 /*
409  * Free a job entry.  Wait for completion if it is currently active, but don't
410  * delay forever.  If we delay, we return a flag that says that we have to
411  * restart the queue scan.
412  */
413 static int
414 aio_free_entry(struct aiocblist *aiocbe)
415 {
416 	struct kaioinfo *ki;
417 	struct aio_liojob *lj;
418 	struct proc *p;
419 	int error;
420 	int s;
421 
422 	if (aiocbe->jobstate == JOBST_NULL)
423 		panic("aio_free_entry: freeing already free job");
424 
425 	p = aiocbe->userproc;
426 	ki = p->p_aioinfo;
427 	lj = aiocbe->lio;
428 	if (ki == NULL)
429 		panic("aio_free_entry: missing p->p_aioinfo");
430 
431 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
432 		if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE)
433 			return 0;
434 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
435 		tsleep(aiocbe, PRIBIO, "jobwai", 0);
436 	}
437 	aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
438 
439 	if (aiocbe->bp == NULL) {
440 		if (ki->kaio_queue_count <= 0)
441 			panic("aio_free_entry: process queue size <= 0");
442 		if (num_queue_count <= 0)
443 			panic("aio_free_entry: system wide queue size <= 0");
444 
445 		if (lj) {
446 			lj->lioj_queue_count--;
447 			if (aiocbe->jobflags & AIOCBLIST_DONE)
448 				lj->lioj_queue_finished_count--;
449 		}
450 		ki->kaio_queue_count--;
451 		if (aiocbe->jobflags & AIOCBLIST_DONE)
452 			ki->kaio_queue_finished_count--;
453 		num_queue_count--;
454 	} else {
455 		if (lj) {
456 			lj->lioj_buffer_count--;
457 			if (aiocbe->jobflags & AIOCBLIST_DONE)
458 				lj->lioj_buffer_finished_count--;
459 		}
460 		if (aiocbe->jobflags & AIOCBLIST_DONE)
461 			ki->kaio_buffer_finished_count--;
462 		ki->kaio_buffer_count--;
463 		num_buf_aio--;
464 	}
465 
466 	/* aiocbe is going away, we need to destroy any knotes */
467 	/* XXXKSE Note the thread here is used to eventually find the
468 	 * owning process again, but it is also used to do a fo_close
469 	 * and that requires the thread. (but does it require the
470 	 * OWNING thread? (or maybe the running thread?)
471 	 * There is a semantic problem here...
472 	 */
473 	knote_remove(FIRST_THREAD_IN_PROC(p), &aiocbe->klist); /* XXXKSE */
474 
475 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
476 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
477 		ki->kaio_flags &= ~KAIO_WAKEUP;
478 		wakeup(p);
479 	}
480 
481 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
482 		if ((error = aio_fphysio(aiocbe)) != 0)
483 			return error;
484 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
485 			panic("aio_free_entry: invalid physio finish-up state");
486 		s = splbio();
487 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
488 		splx(s);
489 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
490 		s = splnet();
491 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
492 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
493 		splx(s);
494 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
495 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
496 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
497 		s = splbio();
498 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
499 		splx(s);
500 		if (aiocbe->bp) {
501 			vunmapbuf(aiocbe->bp);
502 			relpbuf(aiocbe->bp, NULL);
503 			aiocbe->bp = NULL;
504 		}
505 	}
506 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
507 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
508 		uma_zfree(aiolio_zone, lj);
509 	}
510 	aiocbe->jobstate = JOBST_NULL;
511 	untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
512 	uma_zfree(aiocb_zone, aiocbe);
513 	return 0;
514 }
515 
516 /*
517  * Rundown the jobs for a given process.
518  */
519 static void
520 aio_proc_rundown(struct proc *p)
521 {
522 	int s;
523 	struct kaioinfo *ki;
524 	struct aio_liojob *lj, *ljn;
525 	struct aiocblist *aiocbe, *aiocbn;
526 	struct file *fp;
527 	struct filedesc *fdp;
528 	struct socket *so;
529 
530 	ki = p->p_aioinfo;
531 	if (ki == NULL)
532 		return;
533 
534 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
535 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
536 	    ki->kaio_buffer_finished_count)) {
537 		ki->kaio_flags |= KAIO_RUNDOWN;
538 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
539 			break;
540 	}
541 
542 	/*
543 	 * Move any aio ops that are waiting on socket I/O to the normal job
544 	 * queues so they are cleaned up with any others.
545 	 */
546 	fdp = p->p_fd;
547 
548 	s = splnet();
549 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
550 	    aiocbn) {
551 		aiocbn = TAILQ_NEXT(aiocbe, plist);
552 		fp = fdp->fd_ofiles[aiocbe->uaiocb.aio_fildes];
553 
554 		/*
555 		 * Under some circumstances, the aio_fildes and the file
556 		 * structure don't match.  This would leave aiocbe's in the
557 		 * TAILQ associated with the socket and cause a panic later.
558 		 *
559 		 * Detect and fix.
560 		 */
561 		if ((fp == NULL) || (fp != aiocbe->fd_file))
562 			fp = aiocbe->fd_file;
563 		if (fp) {
564 			so = (struct socket *)fp->f_data;
565 			TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
566 			if (TAILQ_EMPTY(&so->so_aiojobq)) {
567 				so->so_snd.sb_flags &= ~SB_AIO;
568 				so->so_rcv.sb_flags &= ~SB_AIO;
569 			}
570 		}
571 		TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
572 		TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
573 		TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
574 	}
575 	splx(s);
576 
577 restart1:
578 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
579 		aiocbn = TAILQ_NEXT(aiocbe, plist);
580 		if (aio_free_entry(aiocbe))
581 			goto restart1;
582 	}
583 
584 restart2:
585 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
586 	    aiocbn) {
587 		aiocbn = TAILQ_NEXT(aiocbe, plist);
588 		if (aio_free_entry(aiocbe))
589 			goto restart2;
590 	}
591 
592 /*
593  * Note the use of lots of splbio here, trying to avoid splbio for long chains
594  * of I/O.  Probably unnecessary.
595  */
596 restart3:
597 	s = splbio();
598 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
599 		ki->kaio_flags |= KAIO_WAKEUP;
600 		tsleep(p, PRIBIO, "aioprn", 0);
601 		splx(s);
602 		goto restart3;
603 	}
604 	splx(s);
605 
606 restart4:
607 	s = splbio();
608 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
609 		aiocbn = TAILQ_NEXT(aiocbe, plist);
610 		if (aio_free_entry(aiocbe)) {
611 			splx(s);
612 			goto restart4;
613 		}
614 	}
615 	splx(s);
616 
617         /*
618          * If we've slept, jobs might have moved from one queue to another.
619          * Retry rundown if we didn't manage to empty the queues.
620          */
621         if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
622 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
623 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
624 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
625 		goto restart1;
626 
627 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
628 		ljn = TAILQ_NEXT(lj, lioj_list);
629 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
630 		    0)) {
631 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
632 			uma_zfree(aiolio_zone, lj);
633 		} else {
634 #ifdef DIAGNOSTIC
635 			printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
636 			    "QF:%d\n", lj->lioj_buffer_count,
637 			    lj->lioj_buffer_finished_count,
638 			    lj->lioj_queue_count,
639 			    lj->lioj_queue_finished_count);
640 #endif
641 		}
642 	}
643 
644 	uma_zfree(kaio_zone, ki);
645 	p->p_aioinfo = NULL;
646 }
647 
648 /*
649  * Select a job to run (called by an AIO daemon).
650  */
651 static struct aiocblist *
652 aio_selectjob(struct aiothreadlist *aiop)
653 {
654 	int s;
655 	struct aiocblist *aiocbe;
656 	struct kaioinfo *ki;
657 	struct proc *userp;
658 
659 	s = splnet();
660 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
661 	    TAILQ_NEXT(aiocbe, list)) {
662 		userp = aiocbe->userproc;
663 		ki = userp->p_aioinfo;
664 
665 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
666 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
667 			splx(s);
668 			return aiocbe;
669 		}
670 	}
671 	splx(s);
672 
673 	return NULL;
674 }
675 
676 /*
677  * The AIO processing activity.  This is the code that does the I/O request for
678  * the non-physio version of the operations.  The normal vn operations are used,
679  * and this code should work in all instances for every type of file, including
680  * pipes, sockets, fifos, and regular files.
681  */
682 static void
683 aio_process(struct aiocblist *aiocbe)
684 {
685 	struct filedesc *fdp;
686 	struct thread *td;
687 	struct proc *userp;
688 	struct proc *mycp;
689 	struct aiocb *cb;
690 	struct file *fp;
691 	struct uio auio;
692 	struct iovec aiov;
693 	unsigned int fd;
694 	int cnt;
695 	int error;
696 	off_t offset;
697 	int oublock_st, oublock_end;
698 	int inblock_st, inblock_end;
699 
700 	userp = aiocbe->userproc;
701 	td = curthread;
702 	mycp = td->td_proc;
703 	cb = &aiocbe->uaiocb;
704 
705 	fdp = mycp->p_fd;
706 	fd = cb->aio_fildes;
707 	fp = fdp->fd_ofiles[fd];
708 
709 	if ((fp == NULL) || (fp != aiocbe->fd_file)) {
710 		cb->_aiocb_private.error = EBADF;
711 		cb->_aiocb_private.status = -1;
712 		return;
713 	}
714 
715 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
716 	aiov.iov_len = cb->aio_nbytes;
717 
718 	auio.uio_iov = &aiov;
719 	auio.uio_iovcnt = 1;
720 	auio.uio_offset = offset = cb->aio_offset;
721 	auio.uio_resid = cb->aio_nbytes;
722 	cnt = cb->aio_nbytes;
723 	auio.uio_segflg = UIO_USERSPACE;
724 	auio.uio_td = td;
725 
726 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
727 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
728 	/*
729 	 * Temporarily bump the ref count while reading to avoid the
730 	 * descriptor being ripped out from under us.
731 	 */
732 	fhold(fp);
733 	if (cb->aio_lio_opcode == LIO_READ) {
734 		auio.uio_rw = UIO_READ;
735 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
736 	} else {
737 		auio.uio_rw = UIO_WRITE;
738 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
739 	}
740 	fdrop(fp, td);
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(userp);
752 			psignal(userp, SIGPIPE);
753 			PROC_UNLOCK(userp);
754 		}
755 	}
756 
757 	cnt -= auio.uio_resid;
758 	cb->_aiocb_private.error = error;
759 	cb->_aiocb_private.status = cnt;
760 }
761 
762 /*
763  * The AIO daemon, most of the actual work is done in aio_process,
764  * but the setup (and address space mgmt) is done in this routine.
765  */
766 static void
767 aio_daemon(void *uproc)
768 {
769 	int s;
770 	struct aio_liojob *lj;
771 	struct aiocb *cb;
772 	struct aiocblist *aiocbe;
773 	struct aiothreadlist *aiop;
774 	struct kaioinfo *ki;
775 	struct proc *curcp, *mycp, *userp;
776 	struct vmspace *myvm, *tmpvm;
777 	struct thread *td = curthread;
778 	struct pgrp *newpgrp;
779 	struct session *newsess;
780 
781 	mtx_lock(&Giant);
782 	/*
783 	 * Local copies of curproc (cp) and vmspace (myvm)
784 	 */
785 	mycp = td->td_proc;
786 	myvm = mycp->p_vmspace;
787 
788 	if (mycp->p_textvp) {
789 		vrele(mycp->p_textvp);
790 		mycp->p_textvp = NULL;
791 	}
792 
793 	/*
794 	 * Allocate and ready the aio control info.  There is one aiop structure
795 	 * per daemon.
796 	 */
797 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
798 	aiop->aiothread = td;
799 	aiop->aiothreadflags |= AIOP_FREE;
800 
801 	s = splnet();
802 
803 	/*
804 	 * Place thread (lightweight process) onto the AIO free thread list.
805 	 */
806 	if (TAILQ_EMPTY(&aio_freeproc))
807 		wakeup(&aio_freeproc);
808 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
809 
810 	splx(s);
811 
812 	/*
813 	 * Get rid of our current filedescriptors.  AIOD's don't need any
814 	 * filedescriptors, except as temporarily inherited from the client.
815 	 */
816 	fdfree(td);
817 	mycp->p_fd = NULL;
818 
819 	/* The daemon resides in its own pgrp. */
820 	MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
821 		M_WAITOK | M_ZERO);
822 	MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
823 		M_WAITOK | M_ZERO);
824 
825 	PGRPSESS_XLOCK();
826 	enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
827 	PGRPSESS_XUNLOCK();
828 
829 	/* Mark special process type. */
830 	mycp->p_flag |= P_SYSTEM;
831 
832 	/*
833 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
834 	 * and creating too many daemons.)
835 	 */
836 	wakeup(mycp);
837 
838 	for (;;) {
839 		/*
840 		 * curcp is the current daemon process context.
841 		 * userp is the current user process context.
842 		 */
843 		curcp = mycp;
844 
845 		/*
846 		 * Take daemon off of free queue
847 		 */
848 		if (aiop->aiothreadflags & AIOP_FREE) {
849 			s = splnet();
850 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
851 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
852 			aiop->aiothreadflags &= ~AIOP_FREE;
853 			splx(s);
854 		}
855 		aiop->aiothreadflags &= ~AIOP_SCHED;
856 
857 		/*
858 		 * Check for jobs.
859 		 */
860 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
861 			cb = &aiocbe->uaiocb;
862 			userp = aiocbe->userproc;
863 
864 			aiocbe->jobstate = JOBST_JOBRUNNING;
865 
866 			/*
867 			 * Connect to process address space for user program.
868 			 */
869 			if (userp != curcp) {
870 				/*
871 				 * Save the current address space that we are
872 				 * connected to.
873 				 */
874 				tmpvm = mycp->p_vmspace;
875 
876 				/*
877 				 * Point to the new user address space, and
878 				 * refer to it.
879 				 */
880 				mycp->p_vmspace = userp->p_vmspace;
881 				mycp->p_vmspace->vm_refcnt++;
882 
883 				/* Activate the new mapping. */
884 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
885 
886 				/*
887 				 * If the old address space wasn't the daemons
888 				 * own address space, then we need to remove the
889 				 * daemon's reference from the other process
890 				 * that it was acting on behalf of.
891 				 */
892 				if (tmpvm != myvm) {
893 					vmspace_free(tmpvm);
894 				}
895 
896 				/*
897 				 * Disassociate from previous clients file
898 				 * descriptors, and associate to the new clients
899 				 * descriptors.  Note that the daemon doesn't
900 				 * need to worry about its orginal descriptors,
901 				 * because they were originally freed.
902 				 */
903 				if (mycp->p_fd)
904 					fdfree(td);
905 				mycp->p_fd = fdshare(userp);
906 				curcp = userp;
907 			}
908 
909 			ki = userp->p_aioinfo;
910 			lj = aiocbe->lio;
911 
912 			/* Account for currently active jobs. */
913 			ki->kaio_active_count++;
914 
915 			/* Do the I/O function. */
916 			aiocbe->jobaiothread = aiop;
917 			aio_process(aiocbe);
918 
919 			/* Decrement the active job count. */
920 			ki->kaio_active_count--;
921 
922 			/*
923 			 * Increment the completion count for wakeup/signal
924 			 * comparisons.
925 			 */
926 			aiocbe->jobflags |= AIOCBLIST_DONE;
927 			ki->kaio_queue_finished_count++;
928 			if (lj)
929 				lj->lioj_queue_finished_count++;
930 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
931 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
932 				ki->kaio_flags &= ~KAIO_WAKEUP;
933 				wakeup(userp);
934 			}
935 
936 			s = splbio();
937 			if (lj && (lj->lioj_flags &
938 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
939 				if ((lj->lioj_queue_finished_count ==
940 				    lj->lioj_queue_count) &&
941 				    (lj->lioj_buffer_finished_count ==
942 				    lj->lioj_buffer_count)) {
943 					PROC_LOCK(userp);
944 					psignal(userp,
945 					    lj->lioj_signal.sigev_signo);
946 					PROC_UNLOCK(userp);
947 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
948 				}
949 			}
950 			splx(s);
951 
952 			aiocbe->jobstate = JOBST_JOBFINISHED;
953 
954 			/*
955 			 * If the I/O request should be automatically rundown,
956 			 * do the needed cleanup.  Otherwise, place the queue
957 			 * entry for the just finished I/O request into the done
958 			 * queue for the associated client.
959 			 */
960 			s = splnet();
961 			if (aiocbe->jobflags & AIOCBLIST_ASYNCFREE) {
962 				aiocbe->jobflags &= ~AIOCBLIST_ASYNCFREE;
963 				uma_zfree(aiocb_zone, aiocbe);
964 			} else {
965 				TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
966 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe,
967 				    plist);
968 			}
969 			splx(s);
970 			KNOTE(&aiocbe->klist, 0);
971 
972 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
973 				wakeup(aiocbe);
974 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
975 			}
976 
977 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
978 				PROC_LOCK(userp);
979 				psignal(userp, cb->aio_sigevent.sigev_signo);
980 				PROC_UNLOCK(userp);
981 			}
982 		}
983 
984 		/*
985 		 * Disconnect from user address space.
986 		 */
987 		if (curcp != mycp) {
988 			/* Get the user address space to disconnect from. */
989 			tmpvm = mycp->p_vmspace;
990 
991 			/* Get original address space for daemon. */
992 			mycp->p_vmspace = myvm;
993 
994 			/* Activate the daemon's address space. */
995 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
996 #ifdef DIAGNOSTIC
997 			if (tmpvm == myvm) {
998 				printf("AIOD: vmspace problem -- %d\n",
999 				    mycp->p_pid);
1000 			}
1001 #endif
1002 			/* Remove our vmspace reference. */
1003 			vmspace_free(tmpvm);
1004 
1005 			/*
1006 			 * Disassociate from the user process's file
1007 			 * descriptors.
1008 			 */
1009 			if (mycp->p_fd)
1010 				fdfree(td);
1011 			mycp->p_fd = NULL;
1012 			curcp = mycp;
1013 		}
1014 
1015 		/*
1016 		 * If we are the first to be put onto the free queue, wakeup
1017 		 * anyone waiting for a daemon.
1018 		 */
1019 		s = splnet();
1020 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
1021 		if (TAILQ_EMPTY(&aio_freeproc))
1022 			wakeup(&aio_freeproc);
1023 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1024 		aiop->aiothreadflags |= AIOP_FREE;
1025 		splx(s);
1026 
1027 		/*
1028 		 * If daemon is inactive for a long time, allow it to exit,
1029 		 * thereby freeing resources.
1030 		 */
1031 		if ((aiop->aiothreadflags & AIOP_SCHED) == 0 &&
1032 		    tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) {
1033 			s = splnet();
1034 			if (TAILQ_EMPTY(&aio_jobs)) {
1035 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1036 				    (num_aio_procs > target_aio_procs)) {
1037 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1038 					splx(s);
1039 					uma_zfree(aiop_zone, aiop);
1040 					num_aio_procs--;
1041 #ifdef DIAGNOSTIC
1042 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1043 						printf("AIOD: bad vm refcnt for"
1044 						    " exiting daemon: %d\n",
1045 						    mycp->p_vmspace->vm_refcnt);
1046 					}
1047 #endif
1048 					kthread_exit(0);
1049 				}
1050 			}
1051 			splx(s);
1052 		}
1053 	}
1054 }
1055 
1056 /*
1057  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
1058  * AIO daemon modifies its environment itself.
1059  */
1060 static int
1061 aio_newproc()
1062 {
1063 	int error;
1064 	struct proc *p;
1065 
1066 	error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, "aiod%d",
1067 			       num_aio_procs);
1068 	if (error)
1069 		return error;
1070 
1071 	/*
1072 	 * Wait until daemon is started, but continue on just in case to
1073 	 * handle error conditions.
1074 	 */
1075 	error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1076 
1077 	num_aio_procs++;
1078 
1079 	return error;
1080 }
1081 
1082 /*
1083  * Try the high-performance, low-overhead physio method for eligible
1084  * VCHR devices.  This method doesn't use an aio helper thread, and
1085  * thus has very low overhead.
1086  *
1087  * Assumes that the caller, _aio_aqueue(), has incremented the file
1088  * structure's reference count, preventing its deallocation for the
1089  * duration of this call.
1090  */
1091 static int
1092 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1093 {
1094 	int error;
1095 	struct aiocb *cb;
1096 	struct file *fp;
1097 	struct buf *bp;
1098 	struct vnode *vp;
1099 	struct kaioinfo *ki;
1100 	struct aio_liojob *lj;
1101 	int s;
1102 	int notify;
1103 
1104 	cb = &aiocbe->uaiocb;
1105 	fp = aiocbe->fd_file;
1106 
1107 	if (fp->f_type != DTYPE_VNODE)
1108 		return (-1);
1109 
1110 	vp = (struct vnode *)fp->f_data;
1111 
1112 	/*
1113 	 * If its not a disk, we don't want to return a positive error.
1114 	 * It causes the aio code to not fall through to try the thread
1115 	 * way when you're talking to a regular file.
1116 	 */
1117 	if (!vn_isdisk(vp, &error)) {
1118 		if (error == ENOTBLK)
1119 			return (-1);
1120 		else
1121 			return (error);
1122 	}
1123 
1124  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
1125 		return (-1);
1126 
1127 	if (cb->aio_nbytes >
1128 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1129 		return (-1);
1130 
1131 	ki = p->p_aioinfo;
1132 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1133 		return (-1);
1134 
1135 	ki->kaio_buffer_count++;
1136 
1137 	lj = aiocbe->lio;
1138 	if (lj)
1139 		lj->lioj_buffer_count++;
1140 
1141 	/* Create and build a buffer header for a transfer. */
1142 	bp = (struct buf *)getpbuf(NULL);
1143 	BUF_KERNPROC(bp);
1144 
1145 	/*
1146 	 * Get a copy of the kva from the physical buffer.
1147 	 */
1148 	bp->b_caller1 = p;
1149 	bp->b_dev = vp->v_rdev;
1150 	error = bp->b_error = 0;
1151 
1152 	bp->b_bcount = cb->aio_nbytes;
1153 	bp->b_bufsize = cb->aio_nbytes;
1154 	bp->b_flags = B_PHYS;
1155 	bp->b_iodone = aio_physwakeup;
1156 	bp->b_saveaddr = bp->b_data;
1157 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1158 	bp->b_blkno = btodb(cb->aio_offset);
1159 
1160 	if (cb->aio_lio_opcode == LIO_WRITE) {
1161 		bp->b_iocmd = BIO_WRITE;
1162 		if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_READ)) {
1163 			error = EFAULT;
1164 			goto doerror;
1165 		}
1166 	} else {
1167 		bp->b_iocmd = BIO_READ;
1168 		if (!useracc(bp->b_data, bp->b_bufsize, VM_PROT_WRITE)) {
1169 			error = EFAULT;
1170 			goto doerror;
1171 		}
1172 	}
1173 
1174 	/* Bring buffer into kernel space. */
1175 	vmapbuf(bp);
1176 
1177 	s = splbio();
1178 	aiocbe->bp = bp;
1179 	bp->b_spc = (void *)aiocbe;
1180 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1181 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1182 	aiocbe->jobstate = JOBST_JOBQBUF;
1183 	cb->_aiocb_private.status = cb->aio_nbytes;
1184 	num_buf_aio++;
1185 	bp->b_error = 0;
1186 
1187 	splx(s);
1188 
1189 	/* Perform transfer. */
1190 	DEV_STRATEGY(bp, 0);
1191 
1192 	notify = 0;
1193 	s = splbio();
1194 
1195 	/*
1196 	 * If we had an error invoking the request, or an error in processing
1197 	 * the request before we have returned, we process it as an error in
1198 	 * transfer.  Note that such an I/O error is not indicated immediately,
1199 	 * but is returned using the aio_error mechanism.  In this case,
1200 	 * aio_suspend will return immediately.
1201 	 */
1202 	if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1203 		struct aiocb *job = aiocbe->uuaiocb;
1204 
1205 		aiocbe->uaiocb._aiocb_private.status = 0;
1206 		suword(&job->_aiocb_private.status, 0);
1207 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1208 		suword(&job->_aiocb_private.error, bp->b_error);
1209 
1210 		ki->kaio_buffer_finished_count++;
1211 
1212 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1213 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1214 			aiocbe->jobflags |= AIOCBLIST_DONE;
1215 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1216 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1217 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1218 			notify = 1;
1219 		}
1220 	}
1221 	splx(s);
1222 	if (notify)
1223 		KNOTE(&aiocbe->klist, 0);
1224 	return 0;
1225 
1226 doerror:
1227 	ki->kaio_buffer_count--;
1228 	if (lj)
1229 		lj->lioj_buffer_count--;
1230 	aiocbe->bp = NULL;
1231 	relpbuf(bp, NULL);
1232 	return error;
1233 }
1234 
1235 /*
1236  * This waits/tests physio completion.
1237  */
1238 static int
1239 aio_fphysio(struct aiocblist *iocb)
1240 {
1241 	int s;
1242 	struct buf *bp;
1243 	int error;
1244 
1245 	bp = iocb->bp;
1246 
1247 	s = splbio();
1248 	while ((bp->b_flags & B_DONE) == 0) {
1249 		if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1250 			if ((bp->b_flags & B_DONE) == 0) {
1251 				splx(s);
1252 				return EINPROGRESS;
1253 			} else
1254 				break;
1255 		}
1256 	}
1257 	splx(s);
1258 
1259 	/* Release mapping into kernel space. */
1260 	vunmapbuf(bp);
1261 	iocb->bp = 0;
1262 
1263 	error = 0;
1264 
1265 	/* Check for an error. */
1266 	if (bp->b_ioflags & BIO_ERROR)
1267 		error = bp->b_error;
1268 
1269 	relpbuf(bp, NULL);
1270 	return (error);
1271 }
1272 
1273 /*
1274  * Wake up aio requests that may be serviceable now.
1275  */
1276 static void
1277 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1278 {
1279 	struct aiocblist *cb,*cbn;
1280 	struct proc *p;
1281 	struct kaioinfo *ki = NULL;
1282 	int opcode, wakecount = 0;
1283 	struct aiothreadlist *aiop;
1284 
1285 	if (sb == &so->so_snd) {
1286 		opcode = LIO_WRITE;
1287 		so->so_snd.sb_flags &= ~SB_AIO;
1288 	} else {
1289 		opcode = LIO_READ;
1290 		so->so_rcv.sb_flags &= ~SB_AIO;
1291 	}
1292 
1293 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1294 		cbn = TAILQ_NEXT(cb, list);
1295 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1296 			p = cb->userproc;
1297 			ki = p->p_aioinfo;
1298 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1299 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1300 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1301 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1302 			wakecount++;
1303 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1304 				panic("invalid queue value");
1305 		}
1306 	}
1307 
1308 	while (wakecount--) {
1309 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1310 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1311 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1312 			aiop->aiothreadflags &= ~AIOP_FREE;
1313 			wakeup(aiop->aiothread);
1314 		}
1315 	}
1316 }
1317 
1318 /*
1319  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1320  * technique is done in this code.
1321  */
1322 static int
1323 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1324 {
1325 	struct proc *p = td->td_proc;
1326 	struct filedesc *fdp;
1327 	struct file *fp;
1328 	unsigned int fd;
1329 	struct socket *so;
1330 	int s;
1331 	int error;
1332 	int opcode;
1333 	struct aiocblist *aiocbe;
1334 	struct aiothreadlist *aiop;
1335 	struct kaioinfo *ki;
1336 	struct kevent kev;
1337 	struct kqueue *kq;
1338 	struct file *kq_fp;
1339 
1340 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1341 	aiocbe->inputcharge = 0;
1342 	aiocbe->outputcharge = 0;
1343 	callout_handle_init(&aiocbe->timeouthandle);
1344 	SLIST_INIT(&aiocbe->klist);
1345 
1346 	suword(&job->_aiocb_private.status, -1);
1347 	suword(&job->_aiocb_private.error, 0);
1348 	suword(&job->_aiocb_private.kernelinfo, -1);
1349 
1350 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1351 	if (error) {
1352 		suword(&job->_aiocb_private.error, error);
1353 		uma_zfree(aiocb_zone, aiocbe);
1354 		return error;
1355 	}
1356 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1357 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1358 		uma_zfree(aiocb_zone, aiocbe);
1359 		return EINVAL;
1360 	}
1361 
1362 	/* Save userspace address of the job info. */
1363 	aiocbe->uuaiocb = job;
1364 
1365 	/* Get the opcode. */
1366 	if (type != LIO_NOP)
1367 		aiocbe->uaiocb.aio_lio_opcode = type;
1368 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1369 
1370 	/* Get the fd info for process. */
1371 	fdp = p->p_fd;
1372 
1373 	/*
1374 	 * Range check file descriptor.
1375 	 */
1376 	fd = aiocbe->uaiocb.aio_fildes;
1377 	if (fd >= fdp->fd_nfiles) {
1378 		uma_zfree(aiocb_zone, aiocbe);
1379 		if (type == 0)
1380 			suword(&job->_aiocb_private.error, EBADF);
1381 		return EBADF;
1382 	}
1383 
1384 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1385 	if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1386 	    0))) {
1387 		uma_zfree(aiocb_zone, aiocbe);
1388 		if (type == 0)
1389 			suword(&job->_aiocb_private.error, EBADF);
1390 		return EBADF;
1391 	}
1392 
1393 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1394 		uma_zfree(aiocb_zone, aiocbe);
1395 		if (type == 0)
1396 			suword(&job->_aiocb_private.error, EINVAL);
1397 		return EINVAL;
1398 	}
1399 
1400 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1401 	if (error) {
1402 		uma_zfree(aiocb_zone, aiocbe);
1403 		if (type == 0)
1404 			suword(&job->_aiocb_private.error, EINVAL);
1405 		return error;
1406 	}
1407 
1408 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1409 	if (jobrefid == LONG_MAX)
1410 		jobrefid = 1;
1411 	else
1412 		jobrefid++;
1413 
1414 	if (opcode == LIO_NOP) {
1415 		uma_zfree(aiocb_zone, aiocbe);
1416 		if (type == 0) {
1417 			suword(&job->_aiocb_private.error, 0);
1418 			suword(&job->_aiocb_private.status, 0);
1419 			suword(&job->_aiocb_private.kernelinfo, 0);
1420 		}
1421 		return 0;
1422 	}
1423 
1424 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1425 		uma_zfree(aiocb_zone, aiocbe);
1426 		if (type == 0) {
1427 			suword(&job->_aiocb_private.status, 0);
1428 			suword(&job->_aiocb_private.error, EINVAL);
1429 		}
1430 		return EINVAL;
1431 	}
1432 
1433 	fhold(fp);
1434 
1435 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1436 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1437 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1438 	}
1439 	else {
1440 		/*
1441 		 * This method for requesting kevent-based notification won't
1442 		 * work on the alpha, since we're passing in a pointer
1443 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1444 		 * based method instead.
1445 		 */
1446 		struct kevent *kevp;
1447 
1448 		kevp = (struct kevent *)(uintptr_t)job->aio_lio_opcode;
1449 		if (kevp == NULL)
1450 			goto no_kqueue;
1451 
1452 		error = copyin(kevp, &kev, sizeof(kev));
1453 		if (error)
1454 			goto aqueue_fail;
1455 	}
1456 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1457 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1458 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1459 		error = EBADF;
1460 		goto aqueue_fail;
1461 	}
1462 	kq = (struct kqueue *)kq_fp->f_data;
1463 	kev.ident = (uintptr_t)aiocbe;
1464 	kev.filter = EVFILT_AIO;
1465 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1466 	error = kqueue_register(kq, &kev, td);
1467 aqueue_fail:
1468 	if (error) {
1469 		uma_zfree(aiocb_zone, aiocbe);
1470 		if (type == 0)
1471 			suword(&job->_aiocb_private.error, error);
1472 		goto done;
1473 	}
1474 no_kqueue:
1475 
1476 	suword(&job->_aiocb_private.error, EINPROGRESS);
1477 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1478 	aiocbe->userproc = p;
1479 	aiocbe->jobflags = 0;
1480 	aiocbe->lio = lj;
1481 	ki = p->p_aioinfo;
1482 
1483 	if (fp->f_type == DTYPE_SOCKET) {
1484 		/*
1485 		 * Alternate queueing for socket ops: Reach down into the
1486 		 * descriptor to get the socket data.  Then check to see if the
1487 		 * socket is ready to be read or written (based on the requested
1488 		 * operation).
1489 		 *
1490 		 * If it is not ready for io, then queue the aiocbe on the
1491 		 * socket, and set the flags so we get a call when sbnotify()
1492 		 * happens.
1493 		 */
1494 		so = (struct socket *)fp->f_data;
1495 		s = splnet();
1496 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1497 		    LIO_WRITE) && (!sowriteable(so)))) {
1498 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1499 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1500 			if (opcode == LIO_READ)
1501 				so->so_rcv.sb_flags |= SB_AIO;
1502 			else
1503 				so->so_snd.sb_flags |= SB_AIO;
1504 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1505 			ki->kaio_queue_count++;
1506 			num_queue_count++;
1507 			splx(s);
1508 			error = 0;
1509 			goto done;
1510 		}
1511 		splx(s);
1512 	}
1513 
1514 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1515 		goto done;
1516 	if (error > 0) {
1517 		suword(&job->_aiocb_private.status, 0);
1518 		aiocbe->uaiocb._aiocb_private.error = error;
1519 		suword(&job->_aiocb_private.error, error);
1520 		goto done;
1521 	}
1522 
1523 	/* No buffer for daemon I/O. */
1524 	aiocbe->bp = NULL;
1525 
1526 	ki->kaio_queue_count++;
1527 	if (lj)
1528 		lj->lioj_queue_count++;
1529 	s = splnet();
1530 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1531 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1532 	splx(s);
1533 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1534 
1535 	num_queue_count++;
1536 	error = 0;
1537 
1538 	/*
1539 	 * If we don't have a free AIO process, and we are below our quota, then
1540 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1541 	 * pick-up this job.  If we don't sucessfully create the new process
1542 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1543 	 * which is likely not the correct thing to do.
1544 	 */
1545 	s = splnet();
1546 retryproc:
1547 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1548 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1549 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1550 		aiop->aiothreadflags &= ~AIOP_FREE;
1551 		wakeup(aiop->aiothread);
1552 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1553 	    ((ki->kaio_active_count + num_aio_resv_start) <
1554 	    ki->kaio_maxactive_count)) {
1555 		num_aio_resv_start++;
1556 		if ((error = aio_newproc()) == 0) {
1557 			num_aio_resv_start--;
1558 			goto retryproc;
1559 		}
1560 		num_aio_resv_start--;
1561 	}
1562 	splx(s);
1563 done:
1564 	fdrop(fp, td);
1565 	return error;
1566 }
1567 
1568 /*
1569  * This routine queues an AIO request, checking for quotas.
1570  */
1571 static int
1572 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1573 {
1574 	struct proc *p = td->td_proc;
1575 	struct kaioinfo *ki;
1576 
1577 	if (p->p_aioinfo == NULL)
1578 		aio_init_aioinfo(p);
1579 
1580 	if (num_queue_count >= max_queue_count)
1581 		return EAGAIN;
1582 
1583 	ki = p->p_aioinfo;
1584 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1585 		return EAGAIN;
1586 
1587 	return _aio_aqueue(td, job, NULL, type);
1588 }
1589 
1590 /*
1591  * Support the aio_return system call, as a side-effect, kernel resources are
1592  * released.
1593  */
1594 int
1595 aio_return(struct thread *td, struct aio_return_args *uap)
1596 {
1597 	struct proc *p = td->td_proc;
1598 	int s;
1599 	int jobref;
1600 	struct aiocblist *cb, *ncb;
1601 	struct aiocb *ujob;
1602 	struct kaioinfo *ki;
1603 
1604 	ki = p->p_aioinfo;
1605 	if (ki == NULL)
1606 		return EINVAL;
1607 
1608 	ujob = uap->aiocbp;
1609 
1610 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1611 	if (jobref == -1 || jobref == 0)
1612 		return EINVAL;
1613 
1614 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1615 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1616 		    jobref) {
1617 			if (ujob == cb->uuaiocb) {
1618 				td->td_retval[0] =
1619 				    cb->uaiocb._aiocb_private.status;
1620 			} else
1621 				td->td_retval[0] = EFAULT;
1622 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1623 				p->p_stats->p_ru.ru_oublock +=
1624 				    cb->outputcharge;
1625 				cb->outputcharge = 0;
1626 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1627 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1628 				cb->inputcharge = 0;
1629 			}
1630 			aio_free_entry(cb);
1631 			return 0;
1632 		}
1633 	}
1634 	s = splbio();
1635 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1636 		ncb = TAILQ_NEXT(cb, plist);
1637 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1638 		    == jobref) {
1639 			splx(s);
1640 			if (ujob == cb->uuaiocb) {
1641 				td->td_retval[0] =
1642 				    cb->uaiocb._aiocb_private.status;
1643 			} else
1644 				td->td_retval[0] = EFAULT;
1645 			aio_free_entry(cb);
1646 			return 0;
1647 		}
1648 	}
1649 	splx(s);
1650 
1651 	return (EINVAL);
1652 }
1653 
1654 /*
1655  * Allow a process to wakeup when any of the I/O requests are completed.
1656  */
1657 int
1658 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1659 {
1660 	struct proc *p = td->td_proc;
1661 	struct timeval atv;
1662 	struct timespec ts;
1663 	struct aiocb *const *cbptr, *cbp;
1664 	struct kaioinfo *ki;
1665 	struct aiocblist *cb;
1666 	int i;
1667 	int njoblist;
1668 	int error, s, timo;
1669 	int *ijoblist;
1670 	struct aiocb **ujoblist;
1671 
1672 	if (uap->nent > AIO_LISTIO_MAX)
1673 		return EINVAL;
1674 
1675 	timo = 0;
1676 	if (uap->timeout) {
1677 		/* Get timespec struct. */
1678 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1679 			return error;
1680 
1681 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1682 			return (EINVAL);
1683 
1684 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1685 		if (itimerfix(&atv))
1686 			return (EINVAL);
1687 		timo = tvtohz(&atv);
1688 	}
1689 
1690 	ki = p->p_aioinfo;
1691 	if (ki == NULL)
1692 		return EAGAIN;
1693 
1694 	njoblist = 0;
1695 	ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1696 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1697 	cbptr = uap->aiocbp;
1698 
1699 	for (i = 0; i < uap->nent; i++) {
1700 		cbp = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
1701 		if (cbp == 0)
1702 			continue;
1703 		ujoblist[njoblist] = cbp;
1704 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1705 		njoblist++;
1706 	}
1707 
1708 	if (njoblist == 0) {
1709 		uma_zfree(aiol_zone, ijoblist);
1710 		uma_zfree(aiol_zone, ujoblist);
1711 		return 0;
1712 	}
1713 
1714 	error = 0;
1715 	for (;;) {
1716 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1717 			for (i = 0; i < njoblist; i++) {
1718 				if (((intptr_t)
1719 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1720 				    ijoblist[i]) {
1721 					if (ujoblist[i] != cb->uuaiocb)
1722 						error = EINVAL;
1723 					uma_zfree(aiol_zone, ijoblist);
1724 					uma_zfree(aiol_zone, ujoblist);
1725 					return error;
1726 				}
1727 			}
1728 		}
1729 
1730 		s = splbio();
1731 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1732 		    TAILQ_NEXT(cb, plist)) {
1733 			for (i = 0; i < njoblist; i++) {
1734 				if (((intptr_t)
1735 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1736 				    ijoblist[i]) {
1737 					splx(s);
1738 					if (ujoblist[i] != cb->uuaiocb)
1739 						error = EINVAL;
1740 					uma_zfree(aiol_zone, ijoblist);
1741 					uma_zfree(aiol_zone, ujoblist);
1742 					return error;
1743 				}
1744 			}
1745 		}
1746 
1747 		ki->kaio_flags |= KAIO_WAKEUP;
1748 		error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1749 		splx(s);
1750 
1751 		if (error == ERESTART || error == EINTR) {
1752 			uma_zfree(aiol_zone, ijoblist);
1753 			uma_zfree(aiol_zone, ujoblist);
1754 			return EINTR;
1755 		} else if (error == EWOULDBLOCK) {
1756 			uma_zfree(aiol_zone, ijoblist);
1757 			uma_zfree(aiol_zone, ujoblist);
1758 			return EAGAIN;
1759 		}
1760 	}
1761 
1762 /* NOTREACHED */
1763 	return EINVAL;
1764 }
1765 
1766 /*
1767  * aio_cancel cancels any non-physio aio operations not currently in
1768  * progress.
1769  */
1770 int
1771 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1772 {
1773 	struct proc *p = td->td_proc;
1774 	struct kaioinfo *ki;
1775 	struct aiocblist *cbe, *cbn;
1776 	struct file *fp;
1777 	struct filedesc *fdp;
1778 	struct socket *so;
1779 	struct proc *po;
1780 	int s,error;
1781 	int cancelled=0;
1782 	int notcancelled=0;
1783 	struct vnode *vp;
1784 
1785 	fdp = p->p_fd;
1786 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1787 	    (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1788 		return (EBADF);
1789 
1790         if (fp->f_type == DTYPE_VNODE) {
1791 		vp = (struct vnode *)fp->f_data;
1792 
1793 		if (vn_isdisk(vp,&error)) {
1794 			td->td_retval[0] = AIO_NOTCANCELED;
1795         	        return 0;
1796 		}
1797 	} else if (fp->f_type == DTYPE_SOCKET) {
1798 		so = (struct socket *)fp->f_data;
1799 
1800 		s = splnet();
1801 
1802 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1803 			cbn = TAILQ_NEXT(cbe, list);
1804 			if ((uap->aiocbp == NULL) ||
1805 				(uap->aiocbp == cbe->uuaiocb) ) {
1806 				po = cbe->userproc;
1807 				ki = po->p_aioinfo;
1808 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1809 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1810 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1811 				if (ki->kaio_flags & KAIO_WAKEUP) {
1812 					wakeup(po);
1813 				}
1814 				cbe->jobstate = JOBST_JOBFINISHED;
1815 				cbe->uaiocb._aiocb_private.status=-1;
1816 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1817 				cancelled++;
1818 /* XXX cancelled, knote? */
1819 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1820 				    SIGEV_SIGNAL) {
1821 					PROC_LOCK(cbe->userproc);
1822 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1823 					PROC_UNLOCK(cbe->userproc);
1824 				}
1825 				if (uap->aiocbp)
1826 					break;
1827 			}
1828 		}
1829 		splx(s);
1830 
1831 		if ((cancelled) && (uap->aiocbp)) {
1832 			td->td_retval[0] = AIO_CANCELED;
1833 			return 0;
1834 		}
1835 	}
1836 	ki=p->p_aioinfo;
1837 	s = splnet();
1838 
1839 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1840 		cbn = TAILQ_NEXT(cbe, plist);
1841 
1842 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1843 		    ((uap->aiocbp == NULL ) ||
1844 		     (uap->aiocbp == cbe->uuaiocb))) {
1845 
1846 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1847 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1848                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1849                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1850                                     plist);
1851 				cancelled++;
1852 				ki->kaio_queue_finished_count++;
1853 				cbe->jobstate = JOBST_JOBFINISHED;
1854 				cbe->uaiocb._aiocb_private.status = -1;
1855 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1856 /* XXX cancelled, knote? */
1857 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1858 				    SIGEV_SIGNAL) {
1859 					PROC_LOCK(cbe->userproc);
1860 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1861 					PROC_UNLOCK(cbe->userproc);
1862 				}
1863 			} else {
1864 				notcancelled++;
1865 			}
1866 		}
1867 	}
1868 	splx(s);
1869 
1870 	if (notcancelled) {
1871 		td->td_retval[0] = AIO_NOTCANCELED;
1872 		return 0;
1873 	}
1874 	if (cancelled) {
1875 		td->td_retval[0] = AIO_CANCELED;
1876 		return 0;
1877 	}
1878 	td->td_retval[0] = AIO_ALLDONE;
1879 
1880 	return 0;
1881 }
1882 
1883 /*
1884  * aio_error is implemented in the kernel level for compatibility purposes only.
1885  * For a user mode async implementation, it would be best to do it in a userland
1886  * subroutine.
1887  */
1888 int
1889 aio_error(struct thread *td, struct aio_error_args *uap)
1890 {
1891 	struct proc *p = td->td_proc;
1892 	int s;
1893 	struct aiocblist *cb;
1894 	struct kaioinfo *ki;
1895 	int jobref;
1896 
1897 	ki = p->p_aioinfo;
1898 	if (ki == NULL)
1899 		return EINVAL;
1900 
1901 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1902 	if ((jobref == -1) || (jobref == 0))
1903 		return EINVAL;
1904 
1905 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1906 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1907 		    jobref) {
1908 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1909 			return 0;
1910 		}
1911 	}
1912 
1913 	s = splnet();
1914 
1915 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1916 	    plist)) {
1917 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1918 		    jobref) {
1919 			td->td_retval[0] = EINPROGRESS;
1920 			splx(s);
1921 			return 0;
1922 		}
1923 	}
1924 
1925 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1926 	    plist)) {
1927 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1928 		    jobref) {
1929 			td->td_retval[0] = EINPROGRESS;
1930 			splx(s);
1931 			return 0;
1932 		}
1933 	}
1934 	splx(s);
1935 
1936 	s = splbio();
1937 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1938 	    plist)) {
1939 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1940 		    jobref) {
1941 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1942 			splx(s);
1943 			return 0;
1944 		}
1945 	}
1946 
1947 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1948 	    plist)) {
1949 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1950 		    jobref) {
1951 			td->td_retval[0] = EINPROGRESS;
1952 			splx(s);
1953 			return 0;
1954 		}
1955 	}
1956 	splx(s);
1957 
1958 #if (0)
1959 	/*
1960 	 * Hack for lio.
1961 	 */
1962 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1963 	if (status == -1)
1964 		return fuword(&uap->aiocbp->_aiocb_private.error);
1965 #endif
1966 	return EINVAL;
1967 }
1968 
1969 /* syscall - asynchronous read from a file (REALTIME) */
1970 int
1971 aio_read(struct thread *td, struct aio_read_args *uap)
1972 {
1973 
1974 	return aio_aqueue(td, uap->aiocbp, LIO_READ);
1975 }
1976 
1977 /* syscall - asynchronous write to a file (REALTIME) */
1978 int
1979 aio_write(struct thread *td, struct aio_write_args *uap)
1980 {
1981 
1982 	return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1983 }
1984 
1985 /* syscall - XXX undocumented */
1986 int
1987 lio_listio(struct thread *td, struct lio_listio_args *uap)
1988 {
1989 	struct proc *p = td->td_proc;
1990 	int nent, nentqueued;
1991 	struct aiocb *iocb, * const *cbptr;
1992 	struct aiocblist *cb;
1993 	struct kaioinfo *ki;
1994 	struct aio_liojob *lj;
1995 	int error, runningcode;
1996 	int nerror;
1997 	int i;
1998 	int s;
1999 
2000 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2001 		return EINVAL;
2002 
2003 	nent = uap->nent;
2004 	if (nent > AIO_LISTIO_MAX)
2005 		return EINVAL;
2006 
2007 	if (p->p_aioinfo == NULL)
2008 		aio_init_aioinfo(p);
2009 
2010 	if ((nent + num_queue_count) > max_queue_count)
2011 		return EAGAIN;
2012 
2013 	ki = p->p_aioinfo;
2014 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
2015 		return EAGAIN;
2016 
2017 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2018 	if (!lj)
2019 		return EAGAIN;
2020 
2021 	lj->lioj_flags = 0;
2022 	lj->lioj_buffer_count = 0;
2023 	lj->lioj_buffer_finished_count = 0;
2024 	lj->lioj_queue_count = 0;
2025 	lj->lioj_queue_finished_count = 0;
2026 	lj->lioj_ki = ki;
2027 
2028 	/*
2029 	 * Setup signal.
2030 	 */
2031 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2032 		error = copyin(uap->sig, &lj->lioj_signal,
2033 			       sizeof(lj->lioj_signal));
2034 		if (error) {
2035 			uma_zfree(aiolio_zone, lj);
2036 			return error;
2037 		}
2038 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2039 			uma_zfree(aiolio_zone, lj);
2040 			return EINVAL;
2041 		}
2042 		lj->lioj_flags |= LIOJ_SIGNAL;
2043 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
2044 	} else
2045 		lj->lioj_flags &= ~LIOJ_SIGNAL;
2046 
2047 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2048 	/*
2049 	 * Get pointers to the list of I/O requests.
2050 	 */
2051 	nerror = 0;
2052 	nentqueued = 0;
2053 	cbptr = uap->acb_list;
2054 	for (i = 0; i < uap->nent; i++) {
2055 		iocb = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
2056 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != NULL)) {
2057 			error = _aio_aqueue(td, iocb, lj, 0);
2058 			if (error == 0)
2059 				nentqueued++;
2060 			else
2061 				nerror++;
2062 		}
2063 	}
2064 
2065 	/*
2066 	 * If we haven't queued any, then just return error.
2067 	 */
2068 	if (nentqueued == 0)
2069 		return 0;
2070 
2071 	/*
2072 	 * Calculate the appropriate error return.
2073 	 */
2074 	runningcode = 0;
2075 	if (nerror)
2076 		runningcode = EIO;
2077 
2078 	if (uap->mode == LIO_WAIT) {
2079 		int command, found, jobref;
2080 
2081 		for (;;) {
2082 			found = 0;
2083 			for (i = 0; i < uap->nent; i++) {
2084 				/*
2085 				 * Fetch address of the control buf pointer in
2086 				 * user space.
2087 				 */
2088 				iocb = (struct aiocb *)(intptr_t)fuword((caddr_t)&cbptr[i]);
2089 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2090 				    == 0))
2091 					continue;
2092 
2093 				/*
2094 				 * Fetch the associated command from user space.
2095 				 */
2096 				command = fuword(&iocb->aio_lio_opcode);
2097 				if (command == LIO_NOP) {
2098 					found++;
2099 					continue;
2100 				}
2101 
2102 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
2103 
2104 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2105 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2106 					    == jobref) {
2107 						if (cb->uaiocb.aio_lio_opcode
2108 						    == LIO_WRITE) {
2109 							p->p_stats->p_ru.ru_oublock
2110 							    +=
2111 							    cb->outputcharge;
2112 							cb->outputcharge = 0;
2113 						} else if (cb->uaiocb.aio_lio_opcode
2114 						    == LIO_READ) {
2115 							p->p_stats->p_ru.ru_inblock
2116 							    += cb->inputcharge;
2117 							cb->inputcharge = 0;
2118 						}
2119 						found++;
2120 						break;
2121 					}
2122 				}
2123 
2124 				s = splbio();
2125 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2126 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2127 					    == jobref) {
2128 						found++;
2129 						break;
2130 					}
2131 				}
2132 				splx(s);
2133 			}
2134 
2135 			/*
2136 			 * If all I/Os have been disposed of, then we can
2137 			 * return.
2138 			 */
2139 			if (found == nentqueued)
2140 				return runningcode;
2141 
2142 			ki->kaio_flags |= KAIO_WAKEUP;
2143 			error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2144 
2145 			if (error == EINTR)
2146 				return EINTR;
2147 			else if (error == EWOULDBLOCK)
2148 				return EAGAIN;
2149 		}
2150 	}
2151 
2152 	return runningcode;
2153 }
2154 
2155 /*
2156  * This is a weird hack so that we can post a signal.  It is safe to do so from
2157  * a timeout routine, but *not* from an interrupt routine.
2158  */
2159 static void
2160 process_signal(void *aioj)
2161 {
2162 	struct aiocblist *aiocbe = aioj;
2163 	struct aio_liojob *lj = aiocbe->lio;
2164 	struct aiocb *cb = &aiocbe->uaiocb;
2165 
2166 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2167 		(lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2168 		PROC_LOCK(lj->lioj_ki->kaio_p);
2169 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2170 		PROC_UNLOCK(lj->lioj_ki->kaio_p);
2171 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2172 	}
2173 
2174 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2175 		PROC_LOCK(aiocbe->userproc);
2176 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2177 		PROC_UNLOCK(aiocbe->userproc);
2178 	}
2179 }
2180 
2181 /*
2182  * Interrupt handler for physio, performs the necessary process wakeups, and
2183  * signals.
2184  */
2185 static void
2186 aio_physwakeup(struct buf *bp)
2187 {
2188 	struct aiocblist *aiocbe;
2189 	struct proc *p;
2190 	struct kaioinfo *ki;
2191 	struct aio_liojob *lj;
2192 
2193 	wakeup(bp);
2194 
2195 	aiocbe = (struct aiocblist *)bp->b_spc;
2196 	if (aiocbe) {
2197 		p = bp->b_caller1;
2198 
2199 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2200 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2201 		aiocbe->uaiocb._aiocb_private.error = 0;
2202 		aiocbe->jobflags |= AIOCBLIST_DONE;
2203 
2204 		if (bp->b_ioflags & BIO_ERROR)
2205 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2206 
2207 		lj = aiocbe->lio;
2208 		if (lj) {
2209 			lj->lioj_buffer_finished_count++;
2210 
2211 			/*
2212 			 * wakeup/signal if all of the interrupt jobs are done.
2213 			 */
2214 			if (lj->lioj_buffer_finished_count ==
2215 			    lj->lioj_buffer_count) {
2216 				/*
2217 				 * Post a signal if it is called for.
2218 				 */
2219 				if ((lj->lioj_flags &
2220 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2221 				    LIOJ_SIGNAL) {
2222 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2223 					aiocbe->timeouthandle =
2224 						timeout(process_signal,
2225 							aiocbe, 0);
2226 				}
2227 			}
2228 		}
2229 
2230 		ki = p->p_aioinfo;
2231 		if (ki) {
2232 			ki->kaio_buffer_finished_count++;
2233 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2234 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2235 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2236 
2237 			KNOTE(&aiocbe->klist, 0);
2238 			/* Do the wakeup. */
2239 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2240 				ki->kaio_flags &= ~KAIO_WAKEUP;
2241 				wakeup(p);
2242 			}
2243 		}
2244 
2245 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2246 			aiocbe->timeouthandle =
2247 				timeout(process_signal, aiocbe, 0);
2248 	}
2249 }
2250 
2251 /* syscall - wait for the next completion of an aio request */
2252 int
2253 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2254 {
2255 	struct proc *p = td->td_proc;
2256 	struct timeval atv;
2257 	struct timespec ts;
2258 	struct aiocb **cbptr;
2259 	struct kaioinfo *ki;
2260 	struct aiocblist *cb = NULL;
2261 	int error, s, timo;
2262 
2263 	suword(uap->aiocbp, (int)NULL);
2264 
2265 	timo = 0;
2266 	if (uap->timeout) {
2267 		/* Get timespec struct. */
2268 		error = copyin(uap->timeout, &ts, sizeof(ts));
2269 		if (error)
2270 			return error;
2271 
2272 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2273 			return (EINVAL);
2274 
2275 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2276 		if (itimerfix(&atv))
2277 			return (EINVAL);
2278 		timo = tvtohz(&atv);
2279 	}
2280 
2281 	ki = p->p_aioinfo;
2282 	if (ki == NULL)
2283 		return EAGAIN;
2284 
2285 	cbptr = uap->aiocbp;
2286 
2287 	for (;;) {
2288 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2289 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2290 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2291 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2292 				p->p_stats->p_ru.ru_oublock +=
2293 				    cb->outputcharge;
2294 				cb->outputcharge = 0;
2295 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2296 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2297 				cb->inputcharge = 0;
2298 			}
2299 			aio_free_entry(cb);
2300 			return cb->uaiocb._aiocb_private.error;
2301 		}
2302 
2303 		s = splbio();
2304  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2305 			splx(s);
2306 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2307 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2308 			aio_free_entry(cb);
2309 			return cb->uaiocb._aiocb_private.error;
2310 		}
2311 
2312 		ki->kaio_flags |= KAIO_WAKEUP;
2313 		error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2314 		splx(s);
2315 
2316 		if (error == ERESTART)
2317 			return EINTR;
2318 		else if (error < 0)
2319 			return error;
2320 		else if (error == EINTR)
2321 			return EINTR;
2322 		else if (error == EWOULDBLOCK)
2323 			return EAGAIN;
2324 	}
2325 }
2326 
2327 /* kqueue attach function */
2328 static int
2329 filt_aioattach(struct knote *kn)
2330 {
2331 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2332 
2333 	/*
2334 	 * The aiocbe pointer must be validated before using it, so
2335 	 * registration is restricted to the kernel; the user cannot
2336 	 * set EV_FLAG1.
2337 	 */
2338 	if ((kn->kn_flags & EV_FLAG1) == 0)
2339 		return (EPERM);
2340 	kn->kn_flags &= ~EV_FLAG1;
2341 
2342 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2343 
2344 	return (0);
2345 }
2346 
2347 /* kqueue detach function */
2348 static void
2349 filt_aiodetach(struct knote *kn)
2350 {
2351 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2352 
2353 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2354 }
2355 
2356 /* kqueue filter function */
2357 /*ARGSUSED*/
2358 static int
2359 filt_aio(struct knote *kn, long hint)
2360 {
2361 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_id;
2362 
2363 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2364 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2365 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2366 		return (0);
2367 	kn->kn_flags |= EV_EOF;
2368 	return (1);
2369 }
2370