xref: /freebsd/sys/kern/vfs_aio.c (revision 7773002178c8dbc52b44e4d705f07706409af8e4)
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/kthread.h>
34 #include <sys/fcntl.h>
35 #include <sys/file.h>
36 #include <sys/limits.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/unistd.h>
40 #include <sys/proc.h>
41 #include <sys/resourcevar.h>
42 #include <sys/signalvar.h>
43 #include <sys/protosw.h>
44 #include <sys/socketvar.h>
45 #include <sys/syscall.h>
46 #include <sys/sysent.h>
47 #include <sys/sysctl.h>
48 #include <sys/sx.h>
49 #include <sys/vnode.h>
50 #include <sys/conf.h>
51 #include <sys/event.h>
52 
53 #include <posix4/posix4.h>
54 #include <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/pmap.h>
57 #include <vm/vm_map.h>
58 #include <vm/uma.h>
59 #include <sys/aio.h>
60 
61 #include "opt_vfs_aio.h"
62 
63 /*
64  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
65  * overflow.
66  */
67 static	long jobrefid;
68 
69 #define JOBST_NULL		0x0
70 #define JOBST_JOBQGLOBAL	0x2
71 #define JOBST_JOBRUNNING	0x3
72 #define JOBST_JOBFINISHED	0x4
73 #define	JOBST_JOBQBUF		0x5
74 #define	JOBST_JOBBFINISHED	0x6
75 
76 #ifndef MAX_AIO_PER_PROC
77 #define MAX_AIO_PER_PROC	32
78 #endif
79 
80 #ifndef MAX_AIO_QUEUE_PER_PROC
81 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
82 #endif
83 
84 #ifndef MAX_AIO_PROCS
85 #define MAX_AIO_PROCS		32
86 #endif
87 
88 #ifndef MAX_AIO_QUEUE
89 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
90 #endif
91 
92 #ifndef TARGET_AIO_PROCS
93 #define TARGET_AIO_PROCS	4
94 #endif
95 
96 #ifndef MAX_BUF_AIO
97 #define MAX_BUF_AIO		16
98 #endif
99 
100 #ifndef AIOD_TIMEOUT_DEFAULT
101 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
102 #endif
103 
104 #ifndef AIOD_LIFETIME_DEFAULT
105 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
106 #endif
107 
108 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
109 
110 static int max_aio_procs = MAX_AIO_PROCS;
111 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
112 	CTLFLAG_RW, &max_aio_procs, 0,
113 	"Maximum number of kernel threads to use for handling async IO ");
114 
115 static int num_aio_procs = 0;
116 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
117 	CTLFLAG_RD, &num_aio_procs, 0,
118 	"Number of presently active kernel threads for async IO");
119 
120 /*
121  * The code will adjust the actual number of AIO processes towards this
122  * number when it gets a chance.
123  */
124 static int target_aio_procs = TARGET_AIO_PROCS;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
126 	0, "Preferred number of ready kernel threads for async IO");
127 
128 static int max_queue_count = MAX_AIO_QUEUE;
129 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
130     "Maximum number of aio requests to queue, globally");
131 
132 static int num_queue_count = 0;
133 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
134     "Number of queued aio requests");
135 
136 static int num_buf_aio = 0;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
138     "Number of aio requests presently handled by the buf subsystem");
139 
140 /* Number of async I/O thread in the process of being started */
141 /* XXX This should be local to _aio_aqueue() */
142 static int num_aio_resv_start = 0;
143 
144 static int aiod_timeout;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
146     "Timeout value for synchronous aio operations");
147 
148 static int aiod_lifetime;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
150     "Maximum lifetime for idle aiod");
151 
152 static int unloadable = 0;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
154     "Allow unload of aio (not recommended)");
155 
156 
157 static int max_aio_per_proc = MAX_AIO_PER_PROC;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
159     0, "Maximum active aio requests per process (stored in the process)");
160 
161 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
163     &max_aio_queue_per_proc, 0,
164     "Maximum queued aio requests per process (stored in the process)");
165 
166 static int max_buf_aio = MAX_BUF_AIO;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
168     "Maximum buf aio requests per process (stored in the process)");
169 
170 struct aiocblist {
171         TAILQ_ENTRY(aiocblist) list;	/* List of jobs */
172         TAILQ_ENTRY(aiocblist) plist;	/* List of jobs for proc */
173         int	jobflags;
174         int	jobstate;
175 	int	inputcharge;
176 	int	outputcharge;
177 	struct	callout_handle timeouthandle;
178         struct	buf *bp;		/* Buffer pointer */
179         struct	proc *userproc;		/* User process */ /* Not td! */
180 	struct  ucred *cred;		/* Active credential when created */
181         struct	file *fd_file;		/* Pointer to file structure */
182         struct	aio_liojob *lio;	/* Optional lio job */
183         struct	aiocb *uuaiocb;		/* Pointer in userspace of aiocb */
184 	struct	klist klist;		/* list of knotes */
185         struct	aiocb uaiocb;		/* Kernel I/O control block */
186 };
187 
188 /* jobflags */
189 #define AIOCBLIST_RUNDOWN       0x4
190 #define AIOCBLIST_DONE          0x10
191 
192 /*
193  * AIO process info
194  */
195 #define AIOP_FREE	0x1			/* proc on free queue */
196 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
197 
198 struct aiothreadlist {
199 	int aiothreadflags;			/* AIO proc flags */
200 	TAILQ_ENTRY(aiothreadlist) list;	/* List of processes */
201 	struct thread *aiothread;		/* The AIO thread */
202 };
203 
204 /*
205  * data-structure for lio signal management
206  */
207 struct aio_liojob {
208 	int	lioj_flags;
209 	int	lioj_buffer_count;
210 	int	lioj_buffer_finished_count;
211 	int	lioj_queue_count;
212 	int	lioj_queue_finished_count;
213 	struct	sigevent lioj_signal;	/* signal on all I/O done */
214 	TAILQ_ENTRY(aio_liojob) lioj_list;
215 	struct	kaioinfo *lioj_ki;
216 };
217 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
218 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
219 
220 /*
221  * per process aio data structure
222  */
223 struct kaioinfo {
224 	int	kaio_flags;		/* per process kaio flags */
225 	int	kaio_maxactive_count;	/* maximum number of AIOs */
226 	int	kaio_active_count;	/* number of currently used AIOs */
227 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
228 	int	kaio_queue_count;	/* size of AIO queue */
229 	int	kaio_ballowed_count;	/* maximum number of buffers */
230 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
231 	int	kaio_buffer_count;	/* number of physio buffers */
232 	int	kaio_buffer_finished_count; /* count of I/O done */
233 	struct 	proc *kaio_p;		/* process that uses this kaio block */
234 	TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
235 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* job queue for process */
236 	TAILQ_HEAD(,aiocblist) kaio_jobdone;	/* done queue for process */
237 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* buffer job queue for process */
238 	TAILQ_HEAD(,aiocblist) kaio_bufdone;	/* buffer done queue for process */
239 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;	/* queue for aios waiting on sockets */
240 };
241 
242 #define KAIO_RUNDOWN	0x1	/* process is being run down */
243 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
244 
245 static TAILQ_HEAD(,aiothreadlist) aio_activeproc;	/* Active daemons */
246 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* Idle daemons */
247 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
248 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
249 
250 static void	aio_init_aioinfo(struct proc *p);
251 static void	aio_onceonly(void);
252 static int	aio_free_entry(struct aiocblist *aiocbe);
253 static void	aio_process(struct aiocblist *aiocbe);
254 static int	aio_newproc(void);
255 static int	aio_aqueue(struct thread *td, struct aiocb *job, int type);
256 static void	aio_physwakeup(struct buf *bp);
257 static void	aio_proc_rundown(void *arg, struct proc *p);
258 static int	aio_fphysio(struct aiocblist *aiocbe);
259 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
260 static void	aio_daemon(void *uproc);
261 static void	aio_swake_cb(struct socket *, struct sockbuf *);
262 static int	aio_unload(void);
263 static void	process_signal(void *aioj);
264 static int	filt_aioattach(struct knote *kn);
265 static void	filt_aiodetach(struct knote *kn);
266 static int	filt_aio(struct knote *kn, long hint);
267 
268 /*
269  * Zones for:
270  * 	kaio	Per process async io info
271  *	aiop	async io thread data
272  *	aiocb	async io jobs
273  *	aiol	list io job pointer - internal to aio_suspend XXX
274  *	aiolio	list io jobs
275  */
276 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
277 
278 /* kqueue filters for aio */
279 static struct filterops aio_filtops =
280 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
281 
282 static eventhandler_tag exit_tag, exec_tag;
283 
284 /*
285  * Main operations function for use as a kernel module.
286  */
287 static int
288 aio_modload(struct module *module, int cmd, void *arg)
289 {
290 	int error = 0;
291 
292 	switch (cmd) {
293 	case MOD_LOAD:
294 		aio_onceonly();
295 		break;
296 	case MOD_UNLOAD:
297 		error = aio_unload();
298 		break;
299 	case MOD_SHUTDOWN:
300 		break;
301 	default:
302 		error = EINVAL;
303 		break;
304 	}
305 	return (error);
306 }
307 
308 static moduledata_t aio_mod = {
309 	"aio",
310 	&aio_modload,
311 	NULL
312 };
313 
314 SYSCALL_MODULE_HELPER(aio_return);
315 SYSCALL_MODULE_HELPER(aio_suspend);
316 SYSCALL_MODULE_HELPER(aio_cancel);
317 SYSCALL_MODULE_HELPER(aio_error);
318 SYSCALL_MODULE_HELPER(aio_read);
319 SYSCALL_MODULE_HELPER(aio_write);
320 SYSCALL_MODULE_HELPER(aio_waitcomplete);
321 SYSCALL_MODULE_HELPER(lio_listio);
322 
323 DECLARE_MODULE(aio, aio_mod,
324 	SI_SUB_VFS, SI_ORDER_ANY);
325 MODULE_VERSION(aio, 1);
326 
327 /*
328  * Startup initialization
329  */
330 static void
331 aio_onceonly(void)
332 {
333 
334 	/* XXX: should probably just use so->callback */
335 	aio_swake = &aio_swake_cb;
336 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
337 	    EVENTHANDLER_PRI_ANY);
338 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown, NULL,
339 	    EVENTHANDLER_PRI_ANY);
340 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
341 	TAILQ_INIT(&aio_freeproc);
342 	TAILQ_INIT(&aio_activeproc);
343 	TAILQ_INIT(&aio_jobs);
344 	TAILQ_INIT(&aio_bufjobs);
345 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
346 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
347 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
348 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
349 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
350 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
351 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
352 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
353 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aio_liojob), NULL,
354 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
355 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
356 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
357 	jobrefid = 1;
358 	async_io_version = _POSIX_VERSION;
359 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
360 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
361 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
362 }
363 
364 /*
365  * Callback for unload of AIO when used as a module.
366  */
367 static int
368 aio_unload(void)
369 {
370 
371 	/*
372 	 * XXX: no unloads by default, it's too dangerous.
373 	 * perhaps we could do it if locked out callers and then
374 	 * did an aio_proc_rundown() on each process.
375 	 */
376 	if (!unloadable)
377 		return (EOPNOTSUPP);
378 
379 	async_io_version = 0;
380 	aio_swake = NULL;
381 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
382 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
383 	kqueue_del_filteropts(EVFILT_AIO);
384 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
385 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
386 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
387 	return (0);
388 }
389 
390 /*
391  * Init the per-process aioinfo structure.  The aioinfo limits are set
392  * per-process for user limit (resource) management.
393  */
394 static void
395 aio_init_aioinfo(struct proc *p)
396 {
397 	struct kaioinfo *ki;
398 
399 	if (p->p_aioinfo == NULL) {
400 		ki = uma_zalloc(kaio_zone, M_WAITOK);
401 		p->p_aioinfo = ki;
402 		ki->kaio_flags = 0;
403 		ki->kaio_maxactive_count = max_aio_per_proc;
404 		ki->kaio_active_count = 0;
405 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
406 		ki->kaio_queue_count = 0;
407 		ki->kaio_ballowed_count = max_buf_aio;
408 		ki->kaio_buffer_count = 0;
409 		ki->kaio_buffer_finished_count = 0;
410 		ki->kaio_p = p;
411 		TAILQ_INIT(&ki->kaio_jobdone);
412 		TAILQ_INIT(&ki->kaio_jobqueue);
413 		TAILQ_INIT(&ki->kaio_bufdone);
414 		TAILQ_INIT(&ki->kaio_bufqueue);
415 		TAILQ_INIT(&ki->kaio_liojoblist);
416 		TAILQ_INIT(&ki->kaio_sockqueue);
417 	}
418 
419 	while (num_aio_procs < target_aio_procs)
420 		aio_newproc();
421 }
422 
423 /*
424  * Free a job entry.  Wait for completion if it is currently active, but don't
425  * delay forever.  If we delay, we return a flag that says that we have to
426  * restart the queue scan.
427  */
428 static int
429 aio_free_entry(struct aiocblist *aiocbe)
430 {
431 	struct kaioinfo *ki;
432 	struct aio_liojob *lj;
433 	struct proc *p;
434 	int error;
435 	int s;
436 
437 	if (aiocbe->jobstate == JOBST_NULL)
438 		panic("aio_free_entry: freeing already free job");
439 
440 	p = aiocbe->userproc;
441 	ki = p->p_aioinfo;
442 	lj = aiocbe->lio;
443 	if (ki == NULL)
444 		panic("aio_free_entry: missing p->p_aioinfo");
445 
446 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
447 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
448 		tsleep(aiocbe, PRIBIO, "jobwai", 0);
449 	}
450 	if (aiocbe->bp == NULL) {
451 		if (ki->kaio_queue_count <= 0)
452 			panic("aio_free_entry: process queue size <= 0");
453 		if (num_queue_count <= 0)
454 			panic("aio_free_entry: system wide queue size <= 0");
455 
456 		if (lj) {
457 			lj->lioj_queue_count--;
458 			if (aiocbe->jobflags & AIOCBLIST_DONE)
459 				lj->lioj_queue_finished_count--;
460 		}
461 		ki->kaio_queue_count--;
462 		if (aiocbe->jobflags & AIOCBLIST_DONE)
463 			ki->kaio_queue_finished_count--;
464 		num_queue_count--;
465 	} else {
466 		if (lj) {
467 			lj->lioj_buffer_count--;
468 			if (aiocbe->jobflags & AIOCBLIST_DONE)
469 				lj->lioj_buffer_finished_count--;
470 		}
471 		if (aiocbe->jobflags & AIOCBLIST_DONE)
472 			ki->kaio_buffer_finished_count--;
473 		ki->kaio_buffer_count--;
474 		num_buf_aio--;
475 	}
476 
477 	/* aiocbe is going away, we need to destroy any knotes */
478 	/* XXXKSE Note the thread here is used to eventually find the
479 	 * owning process again, but it is also used to do a fo_close
480 	 * and that requires the thread. (but does it require the
481 	 * OWNING thread? (or maybe the running thread?)
482 	 * There is a semantic problem here...
483 	 */
484 	knote_remove(FIRST_THREAD_IN_PROC(p), &aiocbe->klist); /* XXXKSE */
485 
486 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
487 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
488 		ki->kaio_flags &= ~KAIO_WAKEUP;
489 		wakeup(p);
490 	}
491 
492 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
493 		if ((error = aio_fphysio(aiocbe)) != 0)
494 			return (error);
495 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
496 			panic("aio_free_entry: invalid physio finish-up state");
497 		s = splbio();
498 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
499 		splx(s);
500 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
501 		s = splnet();
502 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
503 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
504 		splx(s);
505 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
506 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
507 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
508 		s = splbio();
509 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
510 		splx(s);
511 		if (aiocbe->bp) {
512 			vunmapbuf(aiocbe->bp);
513 			relpbuf(aiocbe->bp, NULL);
514 			aiocbe->bp = NULL;
515 		}
516 	}
517 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
518 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
519 		uma_zfree(aiolio_zone, lj);
520 	}
521 	aiocbe->jobstate = JOBST_NULL;
522 	untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
523 	fdrop(aiocbe->fd_file, curthread);
524 	crfree(aiocbe->cred);
525 	uma_zfree(aiocb_zone, aiocbe);
526 	return (0);
527 }
528 
529 /*
530  * Rundown the jobs for a given process.
531  */
532 static void
533 aio_proc_rundown(void *arg, struct proc *p)
534 {
535 	int s;
536 	struct kaioinfo *ki;
537 	struct aio_liojob *lj, *ljn;
538 	struct aiocblist *aiocbe, *aiocbn;
539 	struct file *fp;
540 	struct socket *so;
541 
542 	ki = p->p_aioinfo;
543 	if (ki == NULL)
544 		return;
545 
546 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
547 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
548 	    ki->kaio_buffer_finished_count)) {
549 		ki->kaio_flags |= KAIO_RUNDOWN;
550 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
551 			break;
552 	}
553 
554 	/*
555 	 * Move any aio ops that are waiting on socket I/O to the normal job
556 	 * queues so they are cleaned up with any others.
557 	 */
558 	s = splnet();
559 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
560 	    aiocbn) {
561 		aiocbn = TAILQ_NEXT(aiocbe, plist);
562 		fp = aiocbe->fd_file;
563 		if (fp != NULL) {
564 			so = 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 ucred *td_savedcred;
686 	struct thread *td;
687 	struct proc *mycp;
688 	struct aiocb *cb;
689 	struct file *fp;
690 	struct uio auio;
691 	struct iovec aiov;
692 	int cnt;
693 	int error;
694 	int oublock_st, oublock_end;
695 	int inblock_st, inblock_end;
696 
697 	td = curthread;
698 	td_savedcred = td->td_ucred;
699 	td->td_ucred = aiocbe->cred;
700 	mycp = td->td_proc;
701 	cb = &aiocbe->uaiocb;
702 	fp = aiocbe->fd_file;
703 
704 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
705 	aiov.iov_len = cb->aio_nbytes;
706 
707 	auio.uio_iov = &aiov;
708 	auio.uio_iovcnt = 1;
709 	auio.uio_offset = cb->aio_offset;
710 	auio.uio_resid = cb->aio_nbytes;
711 	cnt = cb->aio_nbytes;
712 	auio.uio_segflg = UIO_USERSPACE;
713 	auio.uio_td = td;
714 
715 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
716 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
717 	/*
718 	 * _aio_aqueue() acquires a reference to the file that is
719 	 * released in aio_free_entry().
720 	 */
721 	if (cb->aio_lio_opcode == LIO_READ) {
722 		auio.uio_rw = UIO_READ;
723 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
724 	} else {
725 		auio.uio_rw = UIO_WRITE;
726 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
727 	}
728 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
729 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
730 
731 	aiocbe->inputcharge = inblock_end - inblock_st;
732 	aiocbe->outputcharge = oublock_end - oublock_st;
733 
734 	if ((error) && (auio.uio_resid != cnt)) {
735 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
736 			error = 0;
737 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
738 			PROC_LOCK(aiocbe->userproc);
739 			psignal(aiocbe->userproc, SIGPIPE);
740 			PROC_UNLOCK(aiocbe->userproc);
741 		}
742 	}
743 
744 	cnt -= auio.uio_resid;
745 	cb->_aiocb_private.error = error;
746 	cb->_aiocb_private.status = cnt;
747 	td->td_ucred = td_savedcred;
748 }
749 
750 /*
751  * The AIO daemon, most of the actual work is done in aio_process,
752  * but the setup (and address space mgmt) is done in this routine.
753  */
754 static void
755 aio_daemon(void *uproc)
756 {
757 	int s;
758 	struct aio_liojob *lj;
759 	struct aiocb *cb;
760 	struct aiocblist *aiocbe;
761 	struct aiothreadlist *aiop;
762 	struct kaioinfo *ki;
763 	struct proc *curcp, *mycp, *userp;
764 	struct vmspace *myvm, *tmpvm;
765 	struct thread *td = curthread;
766 	struct pgrp *newpgrp;
767 	struct session *newsess;
768 
769 	mtx_lock(&Giant);
770 	/*
771 	 * Local copies of curproc (cp) and vmspace (myvm)
772 	 */
773 	mycp = td->td_proc;
774 	myvm = mycp->p_vmspace;
775 
776 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
777 
778 	/*
779 	 * Allocate and ready the aio control info.  There is one aiop structure
780 	 * per daemon.
781 	 */
782 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
783 	aiop->aiothread = td;
784 	aiop->aiothreadflags |= AIOP_FREE;
785 
786 	s = splnet();
787 
788 	/*
789 	 * Place thread (lightweight process) onto the AIO free thread list.
790 	 */
791 	if (TAILQ_EMPTY(&aio_freeproc))
792 		wakeup(&aio_freeproc);
793 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
794 
795 	splx(s);
796 
797 	/*
798 	 * Get rid of our current filedescriptors.  AIOD's don't need any
799 	 * filedescriptors, except as temporarily inherited from the client.
800 	 */
801 	fdfree(td);
802 
803 	mtx_unlock(&Giant);
804 	/* The daemon resides in its own pgrp. */
805 	MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
806 		M_WAITOK | M_ZERO);
807 	MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
808 		M_WAITOK | M_ZERO);
809 
810 	sx_xlock(&proctree_lock);
811 	enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
812 	sx_xunlock(&proctree_lock);
813 	mtx_lock(&Giant);
814 
815 	/*
816 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
817 	 * and creating too many daemons.)
818 	 */
819 	wakeup(mycp);
820 
821 	for (;;) {
822 		/*
823 		 * curcp is the current daemon process context.
824 		 * userp is the current user process context.
825 		 */
826 		curcp = mycp;
827 
828 		/*
829 		 * Take daemon off of free queue
830 		 */
831 		if (aiop->aiothreadflags & AIOP_FREE) {
832 			s = splnet();
833 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
834 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
835 			aiop->aiothreadflags &= ~AIOP_FREE;
836 			splx(s);
837 		}
838 		aiop->aiothreadflags &= ~AIOP_SCHED;
839 
840 		/*
841 		 * Check for jobs.
842 		 */
843 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
844 			cb = &aiocbe->uaiocb;
845 			userp = aiocbe->userproc;
846 
847 			aiocbe->jobstate = JOBST_JOBRUNNING;
848 
849 			/*
850 			 * Connect to process address space for user program.
851 			 */
852 			if (userp != curcp) {
853 				/*
854 				 * Save the current address space that we are
855 				 * connected to.
856 				 */
857 				tmpvm = mycp->p_vmspace;
858 
859 				/*
860 				 * Point to the new user address space, and
861 				 * refer to it.
862 				 */
863 				mycp->p_vmspace = userp->p_vmspace;
864 				mycp->p_vmspace->vm_refcnt++;
865 
866 				/* Activate the new mapping. */
867 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
868 
869 				/*
870 				 * If the old address space wasn't the daemons
871 				 * own address space, then we need to remove the
872 				 * daemon's reference from the other process
873 				 * that it was acting on behalf of.
874 				 */
875 				if (tmpvm != myvm) {
876 					vmspace_free(tmpvm);
877 				}
878 				curcp = userp;
879 			}
880 
881 			ki = userp->p_aioinfo;
882 			lj = aiocbe->lio;
883 
884 			/* Account for currently active jobs. */
885 			ki->kaio_active_count++;
886 
887 			/* Do the I/O function. */
888 			aio_process(aiocbe);
889 
890 			/* Decrement the active job count. */
891 			ki->kaio_active_count--;
892 
893 			/*
894 			 * Increment the completion count for wakeup/signal
895 			 * comparisons.
896 			 */
897 			aiocbe->jobflags |= AIOCBLIST_DONE;
898 			ki->kaio_queue_finished_count++;
899 			if (lj)
900 				lj->lioj_queue_finished_count++;
901 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
902 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
903 				ki->kaio_flags &= ~KAIO_WAKEUP;
904 				wakeup(userp);
905 			}
906 
907 			s = splbio();
908 			if (lj && (lj->lioj_flags &
909 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
910 				if ((lj->lioj_queue_finished_count ==
911 				    lj->lioj_queue_count) &&
912 				    (lj->lioj_buffer_finished_count ==
913 				    lj->lioj_buffer_count)) {
914 					PROC_LOCK(userp);
915 					psignal(userp,
916 					    lj->lioj_signal.sigev_signo);
917 					PROC_UNLOCK(userp);
918 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
919 				}
920 			}
921 			splx(s);
922 
923 			aiocbe->jobstate = JOBST_JOBFINISHED;
924 
925 			s = splnet();
926 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
927 			TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
928 			splx(s);
929 			KNOTE(&aiocbe->klist, 0);
930 
931 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
932 				wakeup(aiocbe);
933 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
934 			}
935 
936 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
937 				PROC_LOCK(userp);
938 				psignal(userp, cb->aio_sigevent.sigev_signo);
939 				PROC_UNLOCK(userp);
940 			}
941 		}
942 
943 		/*
944 		 * Disconnect from user address space.
945 		 */
946 		if (curcp != mycp) {
947 			/* Get the user address space to disconnect from. */
948 			tmpvm = mycp->p_vmspace;
949 
950 			/* Get original address space for daemon. */
951 			mycp->p_vmspace = myvm;
952 
953 			/* Activate the daemon's address space. */
954 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
955 #ifdef DIAGNOSTIC
956 			if (tmpvm == myvm) {
957 				printf("AIOD: vmspace problem -- %d\n",
958 				    mycp->p_pid);
959 			}
960 #endif
961 			/* Remove our vmspace reference. */
962 			vmspace_free(tmpvm);
963 
964 			curcp = mycp;
965 		}
966 
967 		/*
968 		 * If we are the first to be put onto the free queue, wakeup
969 		 * anyone waiting for a daemon.
970 		 */
971 		s = splnet();
972 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
973 		if (TAILQ_EMPTY(&aio_freeproc))
974 			wakeup(&aio_freeproc);
975 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
976 		aiop->aiothreadflags |= AIOP_FREE;
977 		splx(s);
978 
979 		/*
980 		 * If daemon is inactive for a long time, allow it to exit,
981 		 * thereby freeing resources.
982 		 */
983 		if ((aiop->aiothreadflags & AIOP_SCHED) == 0 &&
984 		    tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) {
985 			s = splnet();
986 			if (TAILQ_EMPTY(&aio_jobs)) {
987 				if ((aiop->aiothreadflags & AIOP_FREE) &&
988 				    (num_aio_procs > target_aio_procs)) {
989 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
990 					splx(s);
991 					uma_zfree(aiop_zone, aiop);
992 					num_aio_procs--;
993 #ifdef DIAGNOSTIC
994 					if (mycp->p_vmspace->vm_refcnt <= 1) {
995 						printf("AIOD: bad vm refcnt for"
996 						    " exiting daemon: %d\n",
997 						    mycp->p_vmspace->vm_refcnt);
998 					}
999 #endif
1000 					kthread_exit(0);
1001 				}
1002 			}
1003 			splx(s);
1004 		}
1005 	}
1006 }
1007 
1008 /*
1009  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
1010  * AIO daemon modifies its environment itself.
1011  */
1012 static int
1013 aio_newproc(void)
1014 {
1015 	int error;
1016 	struct proc *p;
1017 
1018 	error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, 0, "aiod%d",
1019 			       num_aio_procs);
1020 	if (error)
1021 		return (error);
1022 
1023 	/*
1024 	 * Wait until daemon is started, but continue on just in case to
1025 	 * handle error conditions.
1026 	 */
1027 	error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1028 
1029 	num_aio_procs++;
1030 
1031 	return (error);
1032 }
1033 
1034 /*
1035  * Try the high-performance, low-overhead physio method for eligible
1036  * VCHR devices.  This method doesn't use an aio helper thread, and
1037  * thus has very low overhead.
1038  *
1039  * Assumes that the caller, _aio_aqueue(), has incremented the file
1040  * structure's reference count, preventing its deallocation for the
1041  * duration of this call.
1042  */
1043 static int
1044 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1045 {
1046 	int error;
1047 	struct aiocb *cb;
1048 	struct file *fp;
1049 	struct buf *bp;
1050 	struct vnode *vp;
1051 	struct kaioinfo *ki;
1052 	struct aio_liojob *lj;
1053 	int s;
1054 	int notify;
1055 
1056 	cb = &aiocbe->uaiocb;
1057 	fp = aiocbe->fd_file;
1058 
1059 	if (fp->f_type != DTYPE_VNODE)
1060 		return (-1);
1061 
1062 	vp = fp->f_vnode;
1063 
1064 	/*
1065 	 * If its not a disk, we don't want to return a positive error.
1066 	 * It causes the aio code to not fall through to try the thread
1067 	 * way when you're talking to a regular file.
1068 	 */
1069 	if (!vn_isdisk(vp, &error)) {
1070 		if (error == ENOTBLK)
1071 			return (-1);
1072 		else
1073 			return (error);
1074 	}
1075 
1076  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
1077 		return (-1);
1078 
1079 	if (cb->aio_nbytes >
1080 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1081 		return (-1);
1082 
1083 	ki = p->p_aioinfo;
1084 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1085 		return (-1);
1086 
1087 	ki->kaio_buffer_count++;
1088 
1089 	lj = aiocbe->lio;
1090 	if (lj)
1091 		lj->lioj_buffer_count++;
1092 
1093 	/* Create and build a buffer header for a transfer. */
1094 	bp = (struct buf *)getpbuf(NULL);
1095 	BUF_KERNPROC(bp);
1096 
1097 	/*
1098 	 * Get a copy of the kva from the physical buffer.
1099 	 */
1100 	bp->b_dev = vp->v_rdev;
1101 	error = 0;
1102 
1103 	bp->b_bcount = cb->aio_nbytes;
1104 	bp->b_bufsize = cb->aio_nbytes;
1105 	bp->b_flags = B_PHYS;
1106 	bp->b_iodone = aio_physwakeup;
1107 	bp->b_saveaddr = bp->b_data;
1108 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1109 	bp->b_offset = cb->aio_offset;
1110 	bp->b_iooffset = cb->aio_offset;
1111 	bp->b_blkno = btodb(cb->aio_offset);
1112 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1113 
1114 	/*
1115 	 * Bring buffer into kernel space.
1116 	 */
1117 	if (vmapbuf(bp) < 0) {
1118 		error = EFAULT;
1119 		goto doerror;
1120 	}
1121 
1122 	s = splbio();
1123 	aiocbe->bp = bp;
1124 	bp->b_caller1 = (void *)aiocbe;
1125 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1126 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1127 	aiocbe->jobstate = JOBST_JOBQBUF;
1128 	cb->_aiocb_private.status = cb->aio_nbytes;
1129 	num_buf_aio++;
1130 	bp->b_error = 0;
1131 
1132 	splx(s);
1133 
1134 	/* Perform transfer. */
1135 	DEV_STRATEGY(bp);
1136 
1137 	notify = 0;
1138 	s = splbio();
1139 
1140 	/*
1141 	 * If we had an error invoking the request, or an error in processing
1142 	 * the request before we have returned, we process it as an error in
1143 	 * transfer.  Note that such an I/O error is not indicated immediately,
1144 	 * but is returned using the aio_error mechanism.  In this case,
1145 	 * aio_suspend will return immediately.
1146 	 */
1147 	if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1148 		struct aiocb *job = aiocbe->uuaiocb;
1149 
1150 		aiocbe->uaiocb._aiocb_private.status = 0;
1151 		suword(&job->_aiocb_private.status, 0);
1152 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1153 		suword(&job->_aiocb_private.error, bp->b_error);
1154 
1155 		ki->kaio_buffer_finished_count++;
1156 
1157 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1158 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1159 			aiocbe->jobflags |= AIOCBLIST_DONE;
1160 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1161 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1162 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1163 			notify = 1;
1164 		}
1165 	}
1166 	splx(s);
1167 	if (notify)
1168 		KNOTE(&aiocbe->klist, 0);
1169 	return (0);
1170 
1171 doerror:
1172 	ki->kaio_buffer_count--;
1173 	if (lj)
1174 		lj->lioj_buffer_count--;
1175 	aiocbe->bp = NULL;
1176 	relpbuf(bp, NULL);
1177 	return (error);
1178 }
1179 
1180 /*
1181  * This waits/tests physio completion.
1182  */
1183 static int
1184 aio_fphysio(struct aiocblist *iocb)
1185 {
1186 	int s;
1187 	struct buf *bp;
1188 	int error;
1189 
1190 	bp = iocb->bp;
1191 
1192 	s = splbio();
1193 	while ((bp->b_flags & B_DONE) == 0) {
1194 		if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1195 			if ((bp->b_flags & B_DONE) == 0) {
1196 				splx(s);
1197 				return (EINPROGRESS);
1198 			} else
1199 				break;
1200 		}
1201 	}
1202 	splx(s);
1203 
1204 	/* Release mapping into kernel space. */
1205 	vunmapbuf(bp);
1206 	iocb->bp = 0;
1207 
1208 	error = 0;
1209 
1210 	/* Check for an error. */
1211 	if (bp->b_ioflags & BIO_ERROR)
1212 		error = bp->b_error;
1213 
1214 	relpbuf(bp, NULL);
1215 	return (error);
1216 }
1217 
1218 /*
1219  * Wake up aio requests that may be serviceable now.
1220  */
1221 static void
1222 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1223 {
1224 	struct aiocblist *cb,*cbn;
1225 	struct proc *p;
1226 	struct kaioinfo *ki = NULL;
1227 	int opcode, wakecount = 0;
1228 	struct aiothreadlist *aiop;
1229 
1230 	if (sb == &so->so_snd) {
1231 		opcode = LIO_WRITE;
1232 		so->so_snd.sb_flags &= ~SB_AIO;
1233 	} else {
1234 		opcode = LIO_READ;
1235 		so->so_rcv.sb_flags &= ~SB_AIO;
1236 	}
1237 
1238 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1239 		cbn = TAILQ_NEXT(cb, list);
1240 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1241 			p = cb->userproc;
1242 			ki = p->p_aioinfo;
1243 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1244 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1245 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1246 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1247 			wakecount++;
1248 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1249 				panic("invalid queue value");
1250 		}
1251 	}
1252 
1253 	while (wakecount--) {
1254 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1255 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1256 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1257 			aiop->aiothreadflags &= ~AIOP_FREE;
1258 			wakeup(aiop->aiothread);
1259 		}
1260 	}
1261 }
1262 
1263 /*
1264  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1265  * technique is done in this code.
1266  */
1267 static int
1268 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1269 {
1270 	struct proc *p = td->td_proc;
1271 	struct filedesc *fdp;
1272 	struct file *fp;
1273 	unsigned int fd;
1274 	struct socket *so;
1275 	int s;
1276 	int error;
1277 	int opcode, user_opcode;
1278 	struct aiocblist *aiocbe;
1279 	struct aiothreadlist *aiop;
1280 	struct kaioinfo *ki;
1281 	struct kevent kev;
1282 	struct kqueue *kq;
1283 	struct file *kq_fp;
1284 
1285 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1286 	aiocbe->inputcharge = 0;
1287 	aiocbe->outputcharge = 0;
1288 	callout_handle_init(&aiocbe->timeouthandle);
1289 	SLIST_INIT(&aiocbe->klist);
1290 
1291 	suword(&job->_aiocb_private.status, -1);
1292 	suword(&job->_aiocb_private.error, 0);
1293 	suword(&job->_aiocb_private.kernelinfo, -1);
1294 
1295 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1296 	if (error) {
1297 		suword(&job->_aiocb_private.error, error);
1298 		uma_zfree(aiocb_zone, aiocbe);
1299 		return (error);
1300 	}
1301 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1302 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1303 		uma_zfree(aiocb_zone, aiocbe);
1304 		return (EINVAL);
1305 	}
1306 
1307 	/* Save userspace address of the job info. */
1308 	aiocbe->uuaiocb = job;
1309 
1310 	/* Get the opcode. */
1311 	user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1312 	if (type != LIO_NOP)
1313 		aiocbe->uaiocb.aio_lio_opcode = type;
1314 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1315 
1316 	/* Get the fd info for process. */
1317 	fdp = p->p_fd;
1318 
1319 	/*
1320 	 * Range check file descriptor.
1321 	 */
1322 	FILEDESC_LOCK(fdp);
1323 	fd = aiocbe->uaiocb.aio_fildes;
1324 	if (fd >= fdp->fd_nfiles) {
1325 		FILEDESC_UNLOCK(fdp);
1326 		uma_zfree(aiocb_zone, aiocbe);
1327 		if (type == 0)
1328 			suword(&job->_aiocb_private.error, EBADF);
1329 		return (EBADF);
1330 	}
1331 
1332 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1333 	if ((fp == NULL) ||
1334 	    ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 0)) ||
1335 	    ((opcode == LIO_READ) && ((fp->f_flag & FREAD) == 0))) {
1336 		FILEDESC_UNLOCK(fdp);
1337 		uma_zfree(aiocb_zone, aiocbe);
1338 		if (type == 0)
1339 			suword(&job->_aiocb_private.error, EBADF);
1340 		return (EBADF);
1341 	}
1342 	fhold(fp);
1343 	FILEDESC_UNLOCK(fdp);
1344 
1345 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1346 		error = EINVAL;
1347 		goto aqueue_fail;
1348 	}
1349 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1350 	if (error) {
1351 		error = EINVAL;
1352 		goto aqueue_fail;
1353 	}
1354 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1355 	if (jobrefid == LONG_MAX)
1356 		jobrefid = 1;
1357 	else
1358 		jobrefid++;
1359 
1360 	if (opcode == LIO_NOP) {
1361 		fdrop(fp, td);
1362 		uma_zfree(aiocb_zone, aiocbe);
1363 		if (type == 0) {
1364 			suword(&job->_aiocb_private.error, 0);
1365 			suword(&job->_aiocb_private.status, 0);
1366 			suword(&job->_aiocb_private.kernelinfo, 0);
1367 		}
1368 		return (0);
1369 	}
1370 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1371 		if (type == 0)
1372 			suword(&job->_aiocb_private.status, 0);
1373 		error = EINVAL;
1374 		goto aqueue_fail;
1375 	}
1376 
1377 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1378 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1379 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1380 	}
1381 	else {
1382 		/*
1383 		 * This method for requesting kevent-based notification won't
1384 		 * work on the alpha, since we're passing in a pointer
1385 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1386 		 * based method instead.
1387 		 */
1388 		if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1389 		    user_opcode == LIO_WRITE)
1390 			goto no_kqueue;
1391 
1392 		error = copyin((struct kevent *)(uintptr_t)user_opcode,
1393 		    &kev, sizeof(kev));
1394 		if (error)
1395 			goto aqueue_fail;
1396 	}
1397 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1398 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1399 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1400 		error = EBADF;
1401 		goto aqueue_fail;
1402 	}
1403 	kq = kq_fp->f_data;
1404 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1405 	kev.filter = EVFILT_AIO;
1406 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1407 	kev.data = (intptr_t)aiocbe;
1408 	error = kqueue_register(kq, &kev, td);
1409 aqueue_fail:
1410 	if (error) {
1411 		fdrop(fp, td);
1412 		uma_zfree(aiocb_zone, aiocbe);
1413 		if (type == 0)
1414 			suword(&job->_aiocb_private.error, error);
1415 		goto done;
1416 	}
1417 no_kqueue:
1418 
1419 	suword(&job->_aiocb_private.error, EINPROGRESS);
1420 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1421 	aiocbe->userproc = p;
1422 	aiocbe->cred = crhold(td->td_ucred);
1423 	aiocbe->jobflags = 0;
1424 	aiocbe->lio = lj;
1425 	ki = p->p_aioinfo;
1426 
1427 	if (fp->f_type == DTYPE_SOCKET) {
1428 		/*
1429 		 * Alternate queueing for socket ops: Reach down into the
1430 		 * descriptor to get the socket data.  Then check to see if the
1431 		 * socket is ready to be read or written (based on the requested
1432 		 * operation).
1433 		 *
1434 		 * If it is not ready for io, then queue the aiocbe on the
1435 		 * socket, and set the flags so we get a call when sbnotify()
1436 		 * happens.
1437 		 */
1438 		so = fp->f_data;
1439 		s = splnet();
1440 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1441 		    LIO_WRITE) && (!sowriteable(so)))) {
1442 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1443 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1444 			if (opcode == LIO_READ)
1445 				so->so_rcv.sb_flags |= SB_AIO;
1446 			else
1447 				so->so_snd.sb_flags |= SB_AIO;
1448 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1449 			ki->kaio_queue_count++;
1450 			num_queue_count++;
1451 			splx(s);
1452 			error = 0;
1453 			goto done;
1454 		}
1455 		splx(s);
1456 	}
1457 
1458 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1459 		goto done;
1460 	if (error > 0) {
1461 		suword(&job->_aiocb_private.status, 0);
1462 		aiocbe->uaiocb._aiocb_private.error = error;
1463 		suword(&job->_aiocb_private.error, error);
1464 		goto done;
1465 	}
1466 
1467 	/* No buffer for daemon I/O. */
1468 	aiocbe->bp = NULL;
1469 
1470 	ki->kaio_queue_count++;
1471 	if (lj)
1472 		lj->lioj_queue_count++;
1473 	s = splnet();
1474 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1475 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1476 	splx(s);
1477 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1478 
1479 	num_queue_count++;
1480 	error = 0;
1481 
1482 	/*
1483 	 * If we don't have a free AIO process, and we are below our quota, then
1484 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1485 	 * pick-up this job.  If we don't sucessfully create the new process
1486 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1487 	 * which is likely not the correct thing to do.
1488 	 */
1489 	s = splnet();
1490 retryproc:
1491 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1492 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1493 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1494 		aiop->aiothreadflags &= ~AIOP_FREE;
1495 		wakeup(aiop->aiothread);
1496 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1497 	    ((ki->kaio_active_count + num_aio_resv_start) <
1498 	    ki->kaio_maxactive_count)) {
1499 		num_aio_resv_start++;
1500 		if ((error = aio_newproc()) == 0) {
1501 			num_aio_resv_start--;
1502 			goto retryproc;
1503 		}
1504 		num_aio_resv_start--;
1505 	}
1506 	splx(s);
1507 done:
1508 	return (error);
1509 }
1510 
1511 /*
1512  * This routine queues an AIO request, checking for quotas.
1513  */
1514 static int
1515 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1516 {
1517 	struct proc *p = td->td_proc;
1518 	struct kaioinfo *ki;
1519 
1520 	if (p->p_aioinfo == NULL)
1521 		aio_init_aioinfo(p);
1522 
1523 	if (num_queue_count >= max_queue_count)
1524 		return (EAGAIN);
1525 
1526 	ki = p->p_aioinfo;
1527 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1528 		return (EAGAIN);
1529 
1530 	return _aio_aqueue(td, job, NULL, type);
1531 }
1532 
1533 /*
1534  * Support the aio_return system call, as a side-effect, kernel resources are
1535  * released.
1536  */
1537 int
1538 aio_return(struct thread *td, struct aio_return_args *uap)
1539 {
1540 	struct proc *p = td->td_proc;
1541 	int s;
1542 	long jobref;
1543 	struct aiocblist *cb, *ncb;
1544 	struct aiocb *ujob;
1545 	struct kaioinfo *ki;
1546 
1547 	ujob = uap->aiocbp;
1548 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1549 	if (jobref == -1 || jobref == 0)
1550 		return (EINVAL);
1551 
1552 	ki = p->p_aioinfo;
1553 	if (ki == NULL)
1554 		return (EINVAL);
1555 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1556 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1557 		    jobref) {
1558 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1559 				p->p_stats->p_ru.ru_oublock +=
1560 				    cb->outputcharge;
1561 				cb->outputcharge = 0;
1562 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1563 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1564 				cb->inputcharge = 0;
1565 			}
1566 			goto done;
1567 		}
1568 	}
1569 	s = splbio();
1570 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1571 		ncb = TAILQ_NEXT(cb, plist);
1572 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1573 		    == jobref) {
1574 			break;
1575 		}
1576 	}
1577 	splx(s);
1578  done:
1579 	if (cb != NULL) {
1580 		if (ujob == cb->uuaiocb) {
1581 			td->td_retval[0] =
1582 			    cb->uaiocb._aiocb_private.status;
1583 		} else
1584 			td->td_retval[0] = EFAULT;
1585 		aio_free_entry(cb);
1586 		return (0);
1587 	}
1588 	return (EINVAL);
1589 }
1590 
1591 /*
1592  * Allow a process to wakeup when any of the I/O requests are completed.
1593  */
1594 int
1595 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1596 {
1597 	struct proc *p = td->td_proc;
1598 	struct timeval atv;
1599 	struct timespec ts;
1600 	struct aiocb *const *cbptr, *cbp;
1601 	struct kaioinfo *ki;
1602 	struct aiocblist *cb;
1603 	int i;
1604 	int njoblist;
1605 	int error, s, timo;
1606 	long *ijoblist;
1607 	struct aiocb **ujoblist;
1608 
1609 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1610 		return (EINVAL);
1611 
1612 	timo = 0;
1613 	if (uap->timeout) {
1614 		/* Get timespec struct. */
1615 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1616 			return (error);
1617 
1618 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1619 			return (EINVAL);
1620 
1621 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1622 		if (itimerfix(&atv))
1623 			return (EINVAL);
1624 		timo = tvtohz(&atv);
1625 	}
1626 
1627 	ki = p->p_aioinfo;
1628 	if (ki == NULL)
1629 		return (EAGAIN);
1630 
1631 	njoblist = 0;
1632 	ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1633 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1634 	cbptr = uap->aiocbp;
1635 
1636 	for (i = 0; i < uap->nent; i++) {
1637 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1638 		if (cbp == 0)
1639 			continue;
1640 		ujoblist[njoblist] = cbp;
1641 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1642 		njoblist++;
1643 	}
1644 
1645 	if (njoblist == 0) {
1646 		uma_zfree(aiol_zone, ijoblist);
1647 		uma_zfree(aiol_zone, ujoblist);
1648 		return (0);
1649 	}
1650 
1651 	error = 0;
1652 	for (;;) {
1653 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1654 			for (i = 0; i < njoblist; i++) {
1655 				if (((intptr_t)
1656 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1657 				    ijoblist[i]) {
1658 					if (ujoblist[i] != cb->uuaiocb)
1659 						error = EINVAL;
1660 					uma_zfree(aiol_zone, ijoblist);
1661 					uma_zfree(aiol_zone, ujoblist);
1662 					return (error);
1663 				}
1664 			}
1665 		}
1666 
1667 		s = splbio();
1668 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1669 		    TAILQ_NEXT(cb, plist)) {
1670 			for (i = 0; i < njoblist; i++) {
1671 				if (((intptr_t)
1672 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1673 				    ijoblist[i]) {
1674 					splx(s);
1675 					if (ujoblist[i] != cb->uuaiocb)
1676 						error = EINVAL;
1677 					uma_zfree(aiol_zone, ijoblist);
1678 					uma_zfree(aiol_zone, ujoblist);
1679 					return (error);
1680 				}
1681 			}
1682 		}
1683 
1684 		ki->kaio_flags |= KAIO_WAKEUP;
1685 		error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1686 		splx(s);
1687 
1688 		if (error == ERESTART || error == EINTR) {
1689 			uma_zfree(aiol_zone, ijoblist);
1690 			uma_zfree(aiol_zone, ujoblist);
1691 			return (EINTR);
1692 		} else if (error == EWOULDBLOCK) {
1693 			uma_zfree(aiol_zone, ijoblist);
1694 			uma_zfree(aiol_zone, ujoblist);
1695 			return (EAGAIN);
1696 		}
1697 	}
1698 
1699 /* NOTREACHED */
1700 	return (EINVAL);
1701 }
1702 
1703 /*
1704  * aio_cancel cancels any non-physio aio operations not currently in
1705  * progress.
1706  */
1707 int
1708 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1709 {
1710 	struct proc *p = td->td_proc;
1711 	struct kaioinfo *ki;
1712 	struct aiocblist *cbe, *cbn;
1713 	struct file *fp;
1714 	struct filedesc *fdp;
1715 	struct socket *so;
1716 	struct proc *po;
1717 	int s,error;
1718 	int cancelled=0;
1719 	int notcancelled=0;
1720 	struct vnode *vp;
1721 
1722 	fdp = p->p_fd;
1723 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1724 	    (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1725 		return (EBADF);
1726 
1727         if (fp->f_type == DTYPE_VNODE) {
1728 		vp = fp->f_vnode;
1729 
1730 		if (vn_isdisk(vp,&error)) {
1731 			td->td_retval[0] = AIO_NOTCANCELED;
1732         	        return (0);
1733 		}
1734 	} else if (fp->f_type == DTYPE_SOCKET) {
1735 		so = fp->f_data;
1736 
1737 		s = splnet();
1738 
1739 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1740 			cbn = TAILQ_NEXT(cbe, list);
1741 			if ((uap->aiocbp == NULL) ||
1742 				(uap->aiocbp == cbe->uuaiocb) ) {
1743 				po = cbe->userproc;
1744 				ki = po->p_aioinfo;
1745 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1746 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1747 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1748 				if (ki->kaio_flags & KAIO_WAKEUP) {
1749 					wakeup(po);
1750 				}
1751 				cbe->jobstate = JOBST_JOBFINISHED;
1752 				cbe->uaiocb._aiocb_private.status=-1;
1753 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1754 				cancelled++;
1755 /* XXX cancelled, knote? */
1756 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1757 				    SIGEV_SIGNAL) {
1758 					PROC_LOCK(cbe->userproc);
1759 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1760 					PROC_UNLOCK(cbe->userproc);
1761 				}
1762 				if (uap->aiocbp)
1763 					break;
1764 			}
1765 		}
1766 		splx(s);
1767 
1768 		if ((cancelled) && (uap->aiocbp)) {
1769 			td->td_retval[0] = AIO_CANCELED;
1770 			return (0);
1771 		}
1772 	}
1773 	ki=p->p_aioinfo;
1774 	if (ki == NULL)
1775 		goto done;
1776 	s = splnet();
1777 
1778 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1779 		cbn = TAILQ_NEXT(cbe, plist);
1780 
1781 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1782 		    ((uap->aiocbp == NULL ) ||
1783 		     (uap->aiocbp == cbe->uuaiocb))) {
1784 
1785 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1786 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1787                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1788                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1789                                     plist);
1790 				cancelled++;
1791 				ki->kaio_queue_finished_count++;
1792 				cbe->jobstate = JOBST_JOBFINISHED;
1793 				cbe->uaiocb._aiocb_private.status = -1;
1794 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1795 /* XXX cancelled, knote? */
1796 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1797 				    SIGEV_SIGNAL) {
1798 					PROC_LOCK(cbe->userproc);
1799 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1800 					PROC_UNLOCK(cbe->userproc);
1801 				}
1802 			} else {
1803 				notcancelled++;
1804 			}
1805 		}
1806 	}
1807 	splx(s);
1808 done:
1809 	if (notcancelled) {
1810 		td->td_retval[0] = AIO_NOTCANCELED;
1811 		return (0);
1812 	}
1813 	if (cancelled) {
1814 		td->td_retval[0] = AIO_CANCELED;
1815 		return (0);
1816 	}
1817 	td->td_retval[0] = AIO_ALLDONE;
1818 
1819 	return (0);
1820 }
1821 
1822 /*
1823  * aio_error is implemented in the kernel level for compatibility purposes only.
1824  * For a user mode async implementation, it would be best to do it in a userland
1825  * subroutine.
1826  */
1827 int
1828 aio_error(struct thread *td, struct aio_error_args *uap)
1829 {
1830 	struct proc *p = td->td_proc;
1831 	int s;
1832 	struct aiocblist *cb;
1833 	struct kaioinfo *ki;
1834 	long jobref;
1835 
1836 	ki = p->p_aioinfo;
1837 	if (ki == NULL)
1838 		return (EINVAL);
1839 
1840 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1841 	if ((jobref == -1) || (jobref == 0))
1842 		return (EINVAL);
1843 
1844 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1845 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1846 		    jobref) {
1847 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1848 			return (0);
1849 		}
1850 	}
1851 
1852 	s = splnet();
1853 
1854 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1855 	    plist)) {
1856 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1857 		    jobref) {
1858 			td->td_retval[0] = EINPROGRESS;
1859 			splx(s);
1860 			return (0);
1861 		}
1862 	}
1863 
1864 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1865 	    plist)) {
1866 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1867 		    jobref) {
1868 			td->td_retval[0] = EINPROGRESS;
1869 			splx(s);
1870 			return (0);
1871 		}
1872 	}
1873 	splx(s);
1874 
1875 	s = splbio();
1876 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1877 	    plist)) {
1878 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1879 		    jobref) {
1880 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1881 			splx(s);
1882 			return (0);
1883 		}
1884 	}
1885 
1886 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); 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 #if (0)
1898 	/*
1899 	 * Hack for lio.
1900 	 */
1901 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1902 	if (status == -1)
1903 		return fuword(&uap->aiocbp->_aiocb_private.error);
1904 #endif
1905 	return (EINVAL);
1906 }
1907 
1908 /* syscall - asynchronous read from a file (REALTIME) */
1909 int
1910 aio_read(struct thread *td, struct aio_read_args *uap)
1911 {
1912 
1913 	return aio_aqueue(td, uap->aiocbp, LIO_READ);
1914 }
1915 
1916 /* syscall - asynchronous write to a file (REALTIME) */
1917 int
1918 aio_write(struct thread *td, struct aio_write_args *uap)
1919 {
1920 
1921 	return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1922 }
1923 
1924 /* syscall - list directed I/O (REALTIME) */
1925 int
1926 lio_listio(struct thread *td, struct lio_listio_args *uap)
1927 {
1928 	struct proc *p = td->td_proc;
1929 	int nent, nentqueued;
1930 	struct aiocb *iocb, * const *cbptr;
1931 	struct aiocblist *cb;
1932 	struct kaioinfo *ki;
1933 	struct aio_liojob *lj;
1934 	int error, runningcode;
1935 	int nerror;
1936 	int i;
1937 	int s;
1938 
1939 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1940 		return (EINVAL);
1941 
1942 	nent = uap->nent;
1943 	if (nent < 0 || nent > AIO_LISTIO_MAX)
1944 		return (EINVAL);
1945 
1946 	if (p->p_aioinfo == NULL)
1947 		aio_init_aioinfo(p);
1948 
1949 	if ((nent + num_queue_count) > max_queue_count)
1950 		return (EAGAIN);
1951 
1952 	ki = p->p_aioinfo;
1953 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1954 		return (EAGAIN);
1955 
1956 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
1957 	if (!lj)
1958 		return (EAGAIN);
1959 
1960 	lj->lioj_flags = 0;
1961 	lj->lioj_buffer_count = 0;
1962 	lj->lioj_buffer_finished_count = 0;
1963 	lj->lioj_queue_count = 0;
1964 	lj->lioj_queue_finished_count = 0;
1965 	lj->lioj_ki = ki;
1966 
1967 	/*
1968 	 * Setup signal.
1969 	 */
1970 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1971 		error = copyin(uap->sig, &lj->lioj_signal,
1972 			       sizeof(lj->lioj_signal));
1973 		if (error) {
1974 			uma_zfree(aiolio_zone, lj);
1975 			return (error);
1976 		}
1977 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1978 			uma_zfree(aiolio_zone, lj);
1979 			return (EINVAL);
1980 		}
1981 		lj->lioj_flags |= LIOJ_SIGNAL;
1982 	}
1983 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1984 	/*
1985 	 * Get pointers to the list of I/O requests.
1986 	 */
1987 	nerror = 0;
1988 	nentqueued = 0;
1989 	cbptr = uap->acb_list;
1990 	for (i = 0; i < uap->nent; i++) {
1991 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1992 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1993 			error = _aio_aqueue(td, iocb, lj, 0);
1994 			if (error == 0)
1995 				nentqueued++;
1996 			else
1997 				nerror++;
1998 		}
1999 	}
2000 
2001 	/*
2002 	 * If we haven't queued any, then just return error.
2003 	 */
2004 	if (nentqueued == 0)
2005 		return (0);
2006 
2007 	/*
2008 	 * Calculate the appropriate error return.
2009 	 */
2010 	runningcode = 0;
2011 	if (nerror)
2012 		runningcode = EIO;
2013 
2014 	if (uap->mode == LIO_WAIT) {
2015 		int command, found, jobref;
2016 
2017 		for (;;) {
2018 			found = 0;
2019 			for (i = 0; i < uap->nent; i++) {
2020 				/*
2021 				 * Fetch address of the control buf pointer in
2022 				 * user space.
2023 				 */
2024 				iocb = (struct aiocb *)
2025 				    (intptr_t)fuword(&cbptr[i]);
2026 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2027 				    == 0))
2028 					continue;
2029 
2030 				/*
2031 				 * Fetch the associated command from user space.
2032 				 */
2033 				command = fuword(&iocb->aio_lio_opcode);
2034 				if (command == LIO_NOP) {
2035 					found++;
2036 					continue;
2037 				}
2038 
2039 				jobref =
2040 				    fuword(&iocb->_aiocb_private.kernelinfo);
2041 
2042 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2043 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2044 					    == jobref) {
2045 						if (cb->uaiocb.aio_lio_opcode
2046 						    == LIO_WRITE) {
2047 							p->p_stats->p_ru.ru_oublock
2048 							    +=
2049 							    cb->outputcharge;
2050 							cb->outputcharge = 0;
2051 						} else if (cb->uaiocb.aio_lio_opcode
2052 						    == LIO_READ) {
2053 							p->p_stats->p_ru.ru_inblock
2054 							    += cb->inputcharge;
2055 							cb->inputcharge = 0;
2056 						}
2057 						found++;
2058 						break;
2059 					}
2060 				}
2061 
2062 				s = splbio();
2063 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2064 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2065 					    == jobref) {
2066 						found++;
2067 						break;
2068 					}
2069 				}
2070 				splx(s);
2071 			}
2072 
2073 			/*
2074 			 * If all I/Os have been disposed of, then we can
2075 			 * return.
2076 			 */
2077 			if (found == nentqueued)
2078 				return (runningcode);
2079 
2080 			ki->kaio_flags |= KAIO_WAKEUP;
2081 			error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2082 
2083 			if (error == EINTR)
2084 				return (EINTR);
2085 			else if (error == EWOULDBLOCK)
2086 				return (EAGAIN);
2087 		}
2088 	}
2089 
2090 	return (runningcode);
2091 }
2092 
2093 /*
2094  * This is a weird hack so that we can post a signal.  It is safe to do so from
2095  * a timeout routine, but *not* from an interrupt routine.
2096  */
2097 static void
2098 process_signal(void *aioj)
2099 {
2100 	struct aiocblist *aiocbe = aioj;
2101 	struct aio_liojob *lj = aiocbe->lio;
2102 	struct aiocb *cb = &aiocbe->uaiocb;
2103 
2104 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2105 		(lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2106 		PROC_LOCK(lj->lioj_ki->kaio_p);
2107 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2108 		PROC_UNLOCK(lj->lioj_ki->kaio_p);
2109 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2110 	}
2111 
2112 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2113 		PROC_LOCK(aiocbe->userproc);
2114 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2115 		PROC_UNLOCK(aiocbe->userproc);
2116 	}
2117 }
2118 
2119 /*
2120  * Interrupt handler for physio, performs the necessary process wakeups, and
2121  * signals.
2122  */
2123 static void
2124 aio_physwakeup(struct buf *bp)
2125 {
2126 	struct aiocblist *aiocbe;
2127 	struct proc *p;
2128 	struct kaioinfo *ki;
2129 	struct aio_liojob *lj;
2130 
2131 	wakeup(bp);
2132 
2133 	aiocbe = (struct aiocblist *)bp->b_caller1;
2134 	if (aiocbe) {
2135 		p = aiocbe->userproc;
2136 
2137 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2138 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2139 		aiocbe->uaiocb._aiocb_private.error = 0;
2140 		aiocbe->jobflags |= AIOCBLIST_DONE;
2141 
2142 		if (bp->b_ioflags & BIO_ERROR)
2143 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2144 
2145 		lj = aiocbe->lio;
2146 		if (lj) {
2147 			lj->lioj_buffer_finished_count++;
2148 
2149 			/*
2150 			 * wakeup/signal if all of the interrupt jobs are done.
2151 			 */
2152 			if (lj->lioj_buffer_finished_count ==
2153 			    lj->lioj_buffer_count) {
2154 				/*
2155 				 * Post a signal if it is called for.
2156 				 */
2157 				if ((lj->lioj_flags &
2158 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2159 				    LIOJ_SIGNAL) {
2160 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2161 					aiocbe->timeouthandle =
2162 						timeout(process_signal,
2163 							aiocbe, 0);
2164 				}
2165 			}
2166 		}
2167 
2168 		ki = p->p_aioinfo;
2169 		if (ki) {
2170 			ki->kaio_buffer_finished_count++;
2171 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2172 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2173 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2174 
2175 			KNOTE(&aiocbe->klist, 0);
2176 			/* Do the wakeup. */
2177 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2178 				ki->kaio_flags &= ~KAIO_WAKEUP;
2179 				wakeup(p);
2180 			}
2181 		}
2182 
2183 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2184 			aiocbe->timeouthandle =
2185 				timeout(process_signal, aiocbe, 0);
2186 	}
2187 }
2188 
2189 /* syscall - wait for the next completion of an aio request */
2190 int
2191 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2192 {
2193 	struct proc *p = td->td_proc;
2194 	struct timeval atv;
2195 	struct timespec ts;
2196 	struct kaioinfo *ki;
2197 	struct aiocblist *cb = NULL;
2198 	int error, s, timo;
2199 
2200 	suword(uap->aiocbp, (int)NULL);
2201 
2202 	timo = 0;
2203 	if (uap->timeout) {
2204 		/* Get timespec struct. */
2205 		error = copyin(uap->timeout, &ts, sizeof(ts));
2206 		if (error)
2207 			return (error);
2208 
2209 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2210 			return (EINVAL);
2211 
2212 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2213 		if (itimerfix(&atv))
2214 			return (EINVAL);
2215 		timo = tvtohz(&atv);
2216 	}
2217 
2218 	ki = p->p_aioinfo;
2219 	if (ki == NULL)
2220 		return (EAGAIN);
2221 
2222 	for (;;) {
2223 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2224 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2225 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2226 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2227 				p->p_stats->p_ru.ru_oublock +=
2228 				    cb->outputcharge;
2229 				cb->outputcharge = 0;
2230 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2231 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2232 				cb->inputcharge = 0;
2233 			}
2234 			aio_free_entry(cb);
2235 			return (cb->uaiocb._aiocb_private.error);
2236 		}
2237 
2238 		s = splbio();
2239  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2240 			splx(s);
2241 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2242 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2243 			aio_free_entry(cb);
2244 			return (cb->uaiocb._aiocb_private.error);
2245 		}
2246 
2247 		ki->kaio_flags |= KAIO_WAKEUP;
2248 		error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2249 		splx(s);
2250 
2251 		if (error == ERESTART)
2252 			return (EINTR);
2253 		else if (error < 0)
2254 			return (error);
2255 		else if (error == EINTR)
2256 			return (EINTR);
2257 		else if (error == EWOULDBLOCK)
2258 			return (EAGAIN);
2259 	}
2260 }
2261 
2262 /* kqueue attach function */
2263 static int
2264 filt_aioattach(struct knote *kn)
2265 {
2266 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2267 
2268 	/*
2269 	 * The aiocbe pointer must be validated before using it, so
2270 	 * registration is restricted to the kernel; the user cannot
2271 	 * set EV_FLAG1.
2272 	 */
2273 	if ((kn->kn_flags & EV_FLAG1) == 0)
2274 		return (EPERM);
2275 	kn->kn_flags &= ~EV_FLAG1;
2276 
2277 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2278 
2279 	return (0);
2280 }
2281 
2282 /* kqueue detach function */
2283 static void
2284 filt_aiodetach(struct knote *kn)
2285 {
2286 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2287 
2288 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2289 }
2290 
2291 /* kqueue filter function */
2292 /*ARGSUSED*/
2293 static int
2294 filt_aio(struct knote *kn, long hint)
2295 {
2296 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2297 
2298 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2299 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2300 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2301 		return (0);
2302 	kn->kn_flags |= EV_EOF;
2303 	return (1);
2304 }
2305