xref: /freebsd/sys/kern/vfs_aio.c (revision 2357939bc239bd5334a169b62313806178dd8f30)
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 	mtx_lock(&Giant);
547 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
548 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
549 	    ki->kaio_buffer_finished_count)) {
550 		ki->kaio_flags |= KAIO_RUNDOWN;
551 		if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
552 			break;
553 	}
554 
555 	/*
556 	 * Move any aio ops that are waiting on socket I/O to the normal job
557 	 * queues so they are cleaned up with any others.
558 	 */
559 	s = splnet();
560 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
561 	    aiocbn) {
562 		aiocbn = TAILQ_NEXT(aiocbe, plist);
563 		fp = aiocbe->fd_file;
564 		if (fp != NULL) {
565 			so = fp->f_data;
566 			TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
567 			if (TAILQ_EMPTY(&so->so_aiojobq)) {
568 				so->so_snd.sb_flags &= ~SB_AIO;
569 				so->so_rcv.sb_flags &= ~SB_AIO;
570 			}
571 		}
572 		TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
573 		TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
574 		TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
575 	}
576 	splx(s);
577 
578 restart1:
579 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
580 		aiocbn = TAILQ_NEXT(aiocbe, plist);
581 		if (aio_free_entry(aiocbe))
582 			goto restart1;
583 	}
584 
585 restart2:
586 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
587 	    aiocbn) {
588 		aiocbn = TAILQ_NEXT(aiocbe, plist);
589 		if (aio_free_entry(aiocbe))
590 			goto restart2;
591 	}
592 
593 /*
594  * Note the use of lots of splbio here, trying to avoid splbio for long chains
595  * of I/O.  Probably unnecessary.
596  */
597 restart3:
598 	s = splbio();
599 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
600 		ki->kaio_flags |= KAIO_WAKEUP;
601 		tsleep(p, PRIBIO, "aioprn", 0);
602 		splx(s);
603 		goto restart3;
604 	}
605 	splx(s);
606 
607 restart4:
608 	s = splbio();
609 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
610 		aiocbn = TAILQ_NEXT(aiocbe, plist);
611 		if (aio_free_entry(aiocbe)) {
612 			splx(s);
613 			goto restart4;
614 		}
615 	}
616 	splx(s);
617 
618         /*
619          * If we've slept, jobs might have moved from one queue to another.
620          * Retry rundown if we didn't manage to empty the queues.
621          */
622         if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
623 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
624 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
625 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
626 		goto restart1;
627 
628 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
629 		ljn = TAILQ_NEXT(lj, lioj_list);
630 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
631 		    0)) {
632 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
633 			uma_zfree(aiolio_zone, lj);
634 		} else {
635 #ifdef DIAGNOSTIC
636 			printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
637 			    "QF:%d\n", lj->lioj_buffer_count,
638 			    lj->lioj_buffer_finished_count,
639 			    lj->lioj_queue_count,
640 			    lj->lioj_queue_finished_count);
641 #endif
642 		}
643 	}
644 
645 	uma_zfree(kaio_zone, ki);
646 	p->p_aioinfo = NULL;
647 	mtx_unlock(&Giant);
648 }
649 
650 /*
651  * Select a job to run (called by an AIO daemon).
652  */
653 static struct aiocblist *
654 aio_selectjob(struct aiothreadlist *aiop)
655 {
656 	int s;
657 	struct aiocblist *aiocbe;
658 	struct kaioinfo *ki;
659 	struct proc *userp;
660 
661 	s = splnet();
662 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
663 	    TAILQ_NEXT(aiocbe, list)) {
664 		userp = aiocbe->userproc;
665 		ki = userp->p_aioinfo;
666 
667 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
668 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
669 			splx(s);
670 			return (aiocbe);
671 		}
672 	}
673 	splx(s);
674 
675 	return (NULL);
676 }
677 
678 /*
679  * The AIO processing activity.  This is the code that does the I/O request for
680  * the non-physio version of the operations.  The normal vn operations are used,
681  * and this code should work in all instances for every type of file, including
682  * pipes, sockets, fifos, and regular files.
683  */
684 static void
685 aio_process(struct aiocblist *aiocbe)
686 {
687 	struct ucred *td_savedcred;
688 	struct thread *td;
689 	struct proc *mycp;
690 	struct aiocb *cb;
691 	struct file *fp;
692 	struct uio auio;
693 	struct iovec aiov;
694 	int cnt;
695 	int error;
696 	int oublock_st, oublock_end;
697 	int inblock_st, inblock_end;
698 
699 	td = curthread;
700 	td_savedcred = td->td_ucred;
701 	td->td_ucred = aiocbe->cred;
702 	mycp = td->td_proc;
703 	cb = &aiocbe->uaiocb;
704 	fp = aiocbe->fd_file;
705 
706 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
707 	aiov.iov_len = cb->aio_nbytes;
708 
709 	auio.uio_iov = &aiov;
710 	auio.uio_iovcnt = 1;
711 	auio.uio_offset = cb->aio_offset;
712 	auio.uio_resid = cb->aio_nbytes;
713 	cnt = cb->aio_nbytes;
714 	auio.uio_segflg = UIO_USERSPACE;
715 	auio.uio_td = td;
716 
717 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
718 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
719 	/*
720 	 * _aio_aqueue() acquires a reference to the file that is
721 	 * released in aio_free_entry().
722 	 */
723 	if (cb->aio_lio_opcode == LIO_READ) {
724 		auio.uio_rw = UIO_READ;
725 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
726 	} else {
727 		auio.uio_rw = UIO_WRITE;
728 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
729 	}
730 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
731 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
732 
733 	aiocbe->inputcharge = inblock_end - inblock_st;
734 	aiocbe->outputcharge = oublock_end - oublock_st;
735 
736 	if ((error) && (auio.uio_resid != cnt)) {
737 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
738 			error = 0;
739 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
740 			PROC_LOCK(aiocbe->userproc);
741 			psignal(aiocbe->userproc, SIGPIPE);
742 			PROC_UNLOCK(aiocbe->userproc);
743 		}
744 	}
745 
746 	cnt -= auio.uio_resid;
747 	cb->_aiocb_private.error = error;
748 	cb->_aiocb_private.status = cnt;
749 	td->td_ucred = td_savedcred;
750 }
751 
752 /*
753  * The AIO daemon, most of the actual work is done in aio_process,
754  * but the setup (and address space mgmt) is done in this routine.
755  */
756 static void
757 aio_daemon(void *uproc)
758 {
759 	int s;
760 	struct aio_liojob *lj;
761 	struct aiocb *cb;
762 	struct aiocblist *aiocbe;
763 	struct aiothreadlist *aiop;
764 	struct kaioinfo *ki;
765 	struct proc *curcp, *mycp, *userp;
766 	struct vmspace *myvm, *tmpvm;
767 	struct thread *td = curthread;
768 	struct pgrp *newpgrp;
769 	struct session *newsess;
770 
771 	mtx_lock(&Giant);
772 	/*
773 	 * Local copies of curproc (cp) and vmspace (myvm)
774 	 */
775 	mycp = td->td_proc;
776 	myvm = mycp->p_vmspace;
777 
778 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
779 
780 	/*
781 	 * Allocate and ready the aio control info.  There is one aiop structure
782 	 * per daemon.
783 	 */
784 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
785 	aiop->aiothread = td;
786 	aiop->aiothreadflags |= AIOP_FREE;
787 
788 	s = splnet();
789 
790 	/*
791 	 * Place thread (lightweight process) onto the AIO free thread list.
792 	 */
793 	if (TAILQ_EMPTY(&aio_freeproc))
794 		wakeup(&aio_freeproc);
795 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
796 
797 	splx(s);
798 
799 	/*
800 	 * Get rid of our current filedescriptors.  AIOD's don't need any
801 	 * filedescriptors, except as temporarily inherited from the client.
802 	 */
803 	fdfree(td);
804 
805 	mtx_unlock(&Giant);
806 	/* The daemon resides in its own pgrp. */
807 	MALLOC(newpgrp, struct pgrp *, sizeof(struct pgrp), M_PGRP,
808 		M_WAITOK | M_ZERO);
809 	MALLOC(newsess, struct session *, sizeof(struct session), M_SESSION,
810 		M_WAITOK | M_ZERO);
811 
812 	sx_xlock(&proctree_lock);
813 	enterpgrp(mycp, mycp->p_pid, newpgrp, newsess);
814 	sx_xunlock(&proctree_lock);
815 	mtx_lock(&Giant);
816 
817 	/*
818 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
819 	 * and creating too many daemons.)
820 	 */
821 	wakeup(mycp);
822 
823 	for (;;) {
824 		/*
825 		 * curcp is the current daemon process context.
826 		 * userp is the current user process context.
827 		 */
828 		curcp = mycp;
829 
830 		/*
831 		 * Take daemon off of free queue
832 		 */
833 		if (aiop->aiothreadflags & AIOP_FREE) {
834 			s = splnet();
835 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
836 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
837 			aiop->aiothreadflags &= ~AIOP_FREE;
838 			splx(s);
839 		}
840 		aiop->aiothreadflags &= ~AIOP_SCHED;
841 
842 		/*
843 		 * Check for jobs.
844 		 */
845 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
846 			cb = &aiocbe->uaiocb;
847 			userp = aiocbe->userproc;
848 
849 			aiocbe->jobstate = JOBST_JOBRUNNING;
850 
851 			/*
852 			 * Connect to process address space for user program.
853 			 */
854 			if (userp != curcp) {
855 				/*
856 				 * Save the current address space that we are
857 				 * connected to.
858 				 */
859 				tmpvm = mycp->p_vmspace;
860 
861 				/*
862 				 * Point to the new user address space, and
863 				 * refer to it.
864 				 */
865 				mycp->p_vmspace = userp->p_vmspace;
866 				mycp->p_vmspace->vm_refcnt++;
867 
868 				/* Activate the new mapping. */
869 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
870 
871 				/*
872 				 * If the old address space wasn't the daemons
873 				 * own address space, then we need to remove the
874 				 * daemon's reference from the other process
875 				 * that it was acting on behalf of.
876 				 */
877 				if (tmpvm != myvm) {
878 					vmspace_free(tmpvm);
879 				}
880 				curcp = userp;
881 			}
882 
883 			ki = userp->p_aioinfo;
884 			lj = aiocbe->lio;
885 
886 			/* Account for currently active jobs. */
887 			ki->kaio_active_count++;
888 
889 			/* Do the I/O function. */
890 			aio_process(aiocbe);
891 
892 			/* Decrement the active job count. */
893 			ki->kaio_active_count--;
894 
895 			/*
896 			 * Increment the completion count for wakeup/signal
897 			 * comparisons.
898 			 */
899 			aiocbe->jobflags |= AIOCBLIST_DONE;
900 			ki->kaio_queue_finished_count++;
901 			if (lj)
902 				lj->lioj_queue_finished_count++;
903 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
904 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
905 				ki->kaio_flags &= ~KAIO_WAKEUP;
906 				wakeup(userp);
907 			}
908 
909 			s = splbio();
910 			if (lj && (lj->lioj_flags &
911 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
912 				if ((lj->lioj_queue_finished_count ==
913 				    lj->lioj_queue_count) &&
914 				    (lj->lioj_buffer_finished_count ==
915 				    lj->lioj_buffer_count)) {
916 					PROC_LOCK(userp);
917 					psignal(userp,
918 					    lj->lioj_signal.sigev_signo);
919 					PROC_UNLOCK(userp);
920 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
921 				}
922 			}
923 			splx(s);
924 
925 			aiocbe->jobstate = JOBST_JOBFINISHED;
926 
927 			s = splnet();
928 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
929 			TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
930 			splx(s);
931 			KNOTE(&aiocbe->klist, 0);
932 
933 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
934 				wakeup(aiocbe);
935 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
936 			}
937 
938 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
939 				PROC_LOCK(userp);
940 				psignal(userp, cb->aio_sigevent.sigev_signo);
941 				PROC_UNLOCK(userp);
942 			}
943 		}
944 
945 		/*
946 		 * Disconnect from user address space.
947 		 */
948 		if (curcp != mycp) {
949 			/* Get the user address space to disconnect from. */
950 			tmpvm = mycp->p_vmspace;
951 
952 			/* Get original address space for daemon. */
953 			mycp->p_vmspace = myvm;
954 
955 			/* Activate the daemon's address space. */
956 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
957 #ifdef DIAGNOSTIC
958 			if (tmpvm == myvm) {
959 				printf("AIOD: vmspace problem -- %d\n",
960 				    mycp->p_pid);
961 			}
962 #endif
963 			/* Remove our vmspace reference. */
964 			vmspace_free(tmpvm);
965 
966 			curcp = mycp;
967 		}
968 
969 		/*
970 		 * If we are the first to be put onto the free queue, wakeup
971 		 * anyone waiting for a daemon.
972 		 */
973 		s = splnet();
974 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
975 		if (TAILQ_EMPTY(&aio_freeproc))
976 			wakeup(&aio_freeproc);
977 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
978 		aiop->aiothreadflags |= AIOP_FREE;
979 		splx(s);
980 
981 		/*
982 		 * If daemon is inactive for a long time, allow it to exit,
983 		 * thereby freeing resources.
984 		 */
985 		if ((aiop->aiothreadflags & AIOP_SCHED) == 0 &&
986 		    tsleep(aiop->aiothread, PRIBIO, "aiordy", aiod_lifetime)) {
987 			s = splnet();
988 			if (TAILQ_EMPTY(&aio_jobs)) {
989 				if ((aiop->aiothreadflags & AIOP_FREE) &&
990 				    (num_aio_procs > target_aio_procs)) {
991 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
992 					splx(s);
993 					uma_zfree(aiop_zone, aiop);
994 					num_aio_procs--;
995 #ifdef DIAGNOSTIC
996 					if (mycp->p_vmspace->vm_refcnt <= 1) {
997 						printf("AIOD: bad vm refcnt for"
998 						    " exiting daemon: %d\n",
999 						    mycp->p_vmspace->vm_refcnt);
1000 					}
1001 #endif
1002 					kthread_exit(0);
1003 				}
1004 			}
1005 			splx(s);
1006 		}
1007 	}
1008 }
1009 
1010 /*
1011  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
1012  * AIO daemon modifies its environment itself.
1013  */
1014 static int
1015 aio_newproc(void)
1016 {
1017 	int error;
1018 	struct proc *p;
1019 
1020 	error = kthread_create(aio_daemon, curproc, &p, RFNOWAIT, 0, "aiod%d",
1021 			       num_aio_procs);
1022 	if (error)
1023 		return (error);
1024 
1025 	/*
1026 	 * Wait until daemon is started, but continue on just in case to
1027 	 * handle error conditions.
1028 	 */
1029 	error = tsleep(p, PZERO, "aiosta", aiod_timeout);
1030 
1031 	num_aio_procs++;
1032 
1033 	return (error);
1034 }
1035 
1036 /*
1037  * Try the high-performance, low-overhead physio method for eligible
1038  * VCHR devices.  This method doesn't use an aio helper thread, and
1039  * thus has very low overhead.
1040  *
1041  * Assumes that the caller, _aio_aqueue(), has incremented the file
1042  * structure's reference count, preventing its deallocation for the
1043  * duration of this call.
1044  */
1045 static int
1046 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1047 {
1048 	int error;
1049 	struct aiocb *cb;
1050 	struct file *fp;
1051 	struct buf *bp;
1052 	struct vnode *vp;
1053 	struct kaioinfo *ki;
1054 	struct aio_liojob *lj;
1055 	int s;
1056 	int notify;
1057 
1058 	cb = &aiocbe->uaiocb;
1059 	fp = aiocbe->fd_file;
1060 
1061 	if (fp->f_type != DTYPE_VNODE)
1062 		return (-1);
1063 
1064 	vp = fp->f_vnode;
1065 
1066 	/*
1067 	 * If its not a disk, we don't want to return a positive error.
1068 	 * It causes the aio code to not fall through to try the thread
1069 	 * way when you're talking to a regular file.
1070 	 */
1071 	if (!vn_isdisk(vp, &error)) {
1072 		if (error == ENOTBLK)
1073 			return (-1);
1074 		else
1075 			return (error);
1076 	}
1077 
1078  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
1079 		return (-1);
1080 
1081 	if (cb->aio_nbytes >
1082 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1083 		return (-1);
1084 
1085 	ki = p->p_aioinfo;
1086 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1087 		return (-1);
1088 
1089 	ki->kaio_buffer_count++;
1090 
1091 	lj = aiocbe->lio;
1092 	if (lj)
1093 		lj->lioj_buffer_count++;
1094 
1095 	/* Create and build a buffer header for a transfer. */
1096 	bp = (struct buf *)getpbuf(NULL);
1097 	BUF_KERNPROC(bp);
1098 
1099 	/*
1100 	 * Get a copy of the kva from the physical buffer.
1101 	 */
1102 	bp->b_dev = vp->v_rdev;
1103 	error = 0;
1104 
1105 	bp->b_bcount = cb->aio_nbytes;
1106 	bp->b_bufsize = cb->aio_nbytes;
1107 	bp->b_iodone = aio_physwakeup;
1108 	bp->b_saveaddr = bp->b_data;
1109 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1110 	bp->b_offset = cb->aio_offset;
1111 	bp->b_iooffset = cb->aio_offset;
1112 	bp->b_blkno = btodb(cb->aio_offset);
1113 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1114 
1115 	/*
1116 	 * Bring buffer into kernel space.
1117 	 */
1118 	if (vmapbuf(bp) < 0) {
1119 		error = EFAULT;
1120 		goto doerror;
1121 	}
1122 
1123 	s = splbio();
1124 	aiocbe->bp = bp;
1125 	bp->b_caller1 = (void *)aiocbe;
1126 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
1127 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1128 	aiocbe->jobstate = JOBST_JOBQBUF;
1129 	cb->_aiocb_private.status = cb->aio_nbytes;
1130 	num_buf_aio++;
1131 	bp->b_error = 0;
1132 
1133 	splx(s);
1134 
1135 	/* Perform transfer. */
1136 	DEV_STRATEGY(bp);
1137 
1138 	notify = 0;
1139 	s = splbio();
1140 
1141 	/*
1142 	 * If we had an error invoking the request, or an error in processing
1143 	 * the request before we have returned, we process it as an error in
1144 	 * transfer.  Note that such an I/O error is not indicated immediately,
1145 	 * but is returned using the aio_error mechanism.  In this case,
1146 	 * aio_suspend will return immediately.
1147 	 */
1148 	if (bp->b_error || (bp->b_ioflags & BIO_ERROR)) {
1149 		struct aiocb *job = aiocbe->uuaiocb;
1150 
1151 		aiocbe->uaiocb._aiocb_private.status = 0;
1152 		suword(&job->_aiocb_private.status, 0);
1153 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1154 		suword(&job->_aiocb_private.error, bp->b_error);
1155 
1156 		ki->kaio_buffer_finished_count++;
1157 
1158 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1159 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1160 			aiocbe->jobflags |= AIOCBLIST_DONE;
1161 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1162 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1163 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1164 			notify = 1;
1165 		}
1166 	}
1167 	splx(s);
1168 	if (notify)
1169 		KNOTE(&aiocbe->klist, 0);
1170 	return (0);
1171 
1172 doerror:
1173 	ki->kaio_buffer_count--;
1174 	if (lj)
1175 		lj->lioj_buffer_count--;
1176 	aiocbe->bp = NULL;
1177 	relpbuf(bp, NULL);
1178 	return (error);
1179 }
1180 
1181 /*
1182  * This waits/tests physio completion.
1183  */
1184 static int
1185 aio_fphysio(struct aiocblist *iocb)
1186 {
1187 	int s;
1188 	struct buf *bp;
1189 	int error;
1190 
1191 	bp = iocb->bp;
1192 
1193 	s = splbio();
1194 	while ((bp->b_flags & B_DONE) == 0) {
1195 		if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1196 			if ((bp->b_flags & B_DONE) == 0) {
1197 				splx(s);
1198 				return (EINPROGRESS);
1199 			} else
1200 				break;
1201 		}
1202 	}
1203 	splx(s);
1204 
1205 	/* Release mapping into kernel space. */
1206 	vunmapbuf(bp);
1207 	iocb->bp = 0;
1208 
1209 	error = 0;
1210 
1211 	/* Check for an error. */
1212 	if (bp->b_ioflags & BIO_ERROR)
1213 		error = bp->b_error;
1214 
1215 	relpbuf(bp, NULL);
1216 	return (error);
1217 }
1218 
1219 /*
1220  * Wake up aio requests that may be serviceable now.
1221  */
1222 static void
1223 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1224 {
1225 	struct aiocblist *cb,*cbn;
1226 	struct proc *p;
1227 	struct kaioinfo *ki = NULL;
1228 	int opcode, wakecount = 0;
1229 	struct aiothreadlist *aiop;
1230 
1231 	if (sb == &so->so_snd) {
1232 		opcode = LIO_WRITE;
1233 		so->so_snd.sb_flags &= ~SB_AIO;
1234 	} else {
1235 		opcode = LIO_READ;
1236 		so->so_rcv.sb_flags &= ~SB_AIO;
1237 	}
1238 
1239 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1240 		cbn = TAILQ_NEXT(cb, list);
1241 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1242 			p = cb->userproc;
1243 			ki = p->p_aioinfo;
1244 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1245 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1246 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1247 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1248 			wakecount++;
1249 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1250 				panic("invalid queue value");
1251 		}
1252 	}
1253 
1254 	while (wakecount--) {
1255 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1256 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1257 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1258 			aiop->aiothreadflags &= ~AIOP_FREE;
1259 			wakeup(aiop->aiothread);
1260 		}
1261 	}
1262 }
1263 
1264 /*
1265  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1266  * technique is done in this code.
1267  */
1268 static int
1269 _aio_aqueue(struct thread *td, struct aiocb *job, struct aio_liojob *lj, int type)
1270 {
1271 	struct proc *p = td->td_proc;
1272 	struct filedesc *fdp;
1273 	struct file *fp;
1274 	unsigned int fd;
1275 	struct socket *so;
1276 	int s;
1277 	int error;
1278 	int opcode, user_opcode;
1279 	struct aiocblist *aiocbe;
1280 	struct aiothreadlist *aiop;
1281 	struct kaioinfo *ki;
1282 	struct kevent kev;
1283 	struct kqueue *kq;
1284 	struct file *kq_fp;
1285 
1286 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK);
1287 	aiocbe->inputcharge = 0;
1288 	aiocbe->outputcharge = 0;
1289 	callout_handle_init(&aiocbe->timeouthandle);
1290 	SLIST_INIT(&aiocbe->klist);
1291 
1292 	suword(&job->_aiocb_private.status, -1);
1293 	suword(&job->_aiocb_private.error, 0);
1294 	suword(&job->_aiocb_private.kernelinfo, -1);
1295 
1296 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1297 	if (error) {
1298 		suword(&job->_aiocb_private.error, error);
1299 		uma_zfree(aiocb_zone, aiocbe);
1300 		return (error);
1301 	}
1302 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1303 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1304 		uma_zfree(aiocb_zone, aiocbe);
1305 		return (EINVAL);
1306 	}
1307 
1308 	/* Save userspace address of the job info. */
1309 	aiocbe->uuaiocb = job;
1310 
1311 	/* Get the opcode. */
1312 	user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1313 	if (type != LIO_NOP)
1314 		aiocbe->uaiocb.aio_lio_opcode = type;
1315 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1316 
1317 	/* Get the fd info for process. */
1318 	fdp = p->p_fd;
1319 
1320 	/*
1321 	 * Range check file descriptor.
1322 	 */
1323 	FILEDESC_LOCK(fdp);
1324 	fd = aiocbe->uaiocb.aio_fildes;
1325 	if (fd >= fdp->fd_nfiles) {
1326 		FILEDESC_UNLOCK(fdp);
1327 		uma_zfree(aiocb_zone, aiocbe);
1328 		if (type == 0)
1329 			suword(&job->_aiocb_private.error, EBADF);
1330 		return (EBADF);
1331 	}
1332 
1333 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1334 	if ((fp == NULL) ||
1335 	    ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) == 0)) ||
1336 	    ((opcode == LIO_READ) && ((fp->f_flag & FREAD) == 0))) {
1337 		FILEDESC_UNLOCK(fdp);
1338 		uma_zfree(aiocb_zone, aiocbe);
1339 		if (type == 0)
1340 			suword(&job->_aiocb_private.error, EBADF);
1341 		return (EBADF);
1342 	}
1343 	fhold(fp);
1344 	FILEDESC_UNLOCK(fdp);
1345 
1346 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1347 		error = EINVAL;
1348 		goto aqueue_fail;
1349 	}
1350 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1351 	if (error) {
1352 		error = EINVAL;
1353 		goto aqueue_fail;
1354 	}
1355 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1356 	if (jobrefid == LONG_MAX)
1357 		jobrefid = 1;
1358 	else
1359 		jobrefid++;
1360 
1361 	if (opcode == LIO_NOP) {
1362 		fdrop(fp, td);
1363 		uma_zfree(aiocb_zone, aiocbe);
1364 		if (type == 0) {
1365 			suword(&job->_aiocb_private.error, 0);
1366 			suword(&job->_aiocb_private.status, 0);
1367 			suword(&job->_aiocb_private.kernelinfo, 0);
1368 		}
1369 		return (0);
1370 	}
1371 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1372 		if (type == 0)
1373 			suword(&job->_aiocb_private.status, 0);
1374 		error = EINVAL;
1375 		goto aqueue_fail;
1376 	}
1377 
1378 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1379 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1380 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1381 	}
1382 	else {
1383 		/*
1384 		 * This method for requesting kevent-based notification won't
1385 		 * work on the alpha, since we're passing in a pointer
1386 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1387 		 * based method instead.
1388 		 */
1389 		if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1390 		    user_opcode == LIO_WRITE)
1391 			goto no_kqueue;
1392 
1393 		error = copyin((struct kevent *)(uintptr_t)user_opcode,
1394 		    &kev, sizeof(kev));
1395 		if (error)
1396 			goto aqueue_fail;
1397 	}
1398 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1399 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1400 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1401 		error = EBADF;
1402 		goto aqueue_fail;
1403 	}
1404 	kq = kq_fp->f_data;
1405 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1406 	kev.filter = EVFILT_AIO;
1407 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1408 	kev.data = (intptr_t)aiocbe;
1409 	error = kqueue_register(kq, &kev, td);
1410 aqueue_fail:
1411 	if (error) {
1412 		fdrop(fp, td);
1413 		uma_zfree(aiocb_zone, aiocbe);
1414 		if (type == 0)
1415 			suword(&job->_aiocb_private.error, error);
1416 		goto done;
1417 	}
1418 no_kqueue:
1419 
1420 	suword(&job->_aiocb_private.error, EINPROGRESS);
1421 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1422 	aiocbe->userproc = p;
1423 	aiocbe->cred = crhold(td->td_ucred);
1424 	aiocbe->jobflags = 0;
1425 	aiocbe->lio = lj;
1426 	ki = p->p_aioinfo;
1427 
1428 	if (fp->f_type == DTYPE_SOCKET) {
1429 		/*
1430 		 * Alternate queueing for socket ops: Reach down into the
1431 		 * descriptor to get the socket data.  Then check to see if the
1432 		 * socket is ready to be read or written (based on the requested
1433 		 * operation).
1434 		 *
1435 		 * If it is not ready for io, then queue the aiocbe on the
1436 		 * socket, and set the flags so we get a call when sbnotify()
1437 		 * happens.
1438 		 */
1439 		so = fp->f_data;
1440 		s = splnet();
1441 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1442 		    LIO_WRITE) && (!sowriteable(so)))) {
1443 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1444 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1445 			if (opcode == LIO_READ)
1446 				so->so_rcv.sb_flags |= SB_AIO;
1447 			else
1448 				so->so_snd.sb_flags |= SB_AIO;
1449 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1450 			ki->kaio_queue_count++;
1451 			num_queue_count++;
1452 			splx(s);
1453 			error = 0;
1454 			goto done;
1455 		}
1456 		splx(s);
1457 	}
1458 
1459 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1460 		goto done;
1461 	if (error > 0) {
1462 		suword(&job->_aiocb_private.status, 0);
1463 		aiocbe->uaiocb._aiocb_private.error = error;
1464 		suword(&job->_aiocb_private.error, error);
1465 		goto done;
1466 	}
1467 
1468 	/* No buffer for daemon I/O. */
1469 	aiocbe->bp = NULL;
1470 
1471 	ki->kaio_queue_count++;
1472 	if (lj)
1473 		lj->lioj_queue_count++;
1474 	s = splnet();
1475 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1476 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1477 	splx(s);
1478 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1479 
1480 	num_queue_count++;
1481 	error = 0;
1482 
1483 	/*
1484 	 * If we don't have a free AIO process, and we are below our quota, then
1485 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1486 	 * pick-up this job.  If we don't sucessfully create the new process
1487 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1488 	 * which is likely not the correct thing to do.
1489 	 */
1490 	s = splnet();
1491 retryproc:
1492 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1493 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1494 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1495 		aiop->aiothreadflags &= ~AIOP_FREE;
1496 		wakeup(aiop->aiothread);
1497 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1498 	    ((ki->kaio_active_count + num_aio_resv_start) <
1499 	    ki->kaio_maxactive_count)) {
1500 		num_aio_resv_start++;
1501 		if ((error = aio_newproc()) == 0) {
1502 			num_aio_resv_start--;
1503 			goto retryproc;
1504 		}
1505 		num_aio_resv_start--;
1506 	}
1507 	splx(s);
1508 done:
1509 	return (error);
1510 }
1511 
1512 /*
1513  * This routine queues an AIO request, checking for quotas.
1514  */
1515 static int
1516 aio_aqueue(struct thread *td, struct aiocb *job, int type)
1517 {
1518 	struct proc *p = td->td_proc;
1519 	struct kaioinfo *ki;
1520 
1521 	if (p->p_aioinfo == NULL)
1522 		aio_init_aioinfo(p);
1523 
1524 	if (num_queue_count >= max_queue_count)
1525 		return (EAGAIN);
1526 
1527 	ki = p->p_aioinfo;
1528 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1529 		return (EAGAIN);
1530 
1531 	return _aio_aqueue(td, job, NULL, type);
1532 }
1533 
1534 /*
1535  * Support the aio_return system call, as a side-effect, kernel resources are
1536  * released.
1537  */
1538 int
1539 aio_return(struct thread *td, struct aio_return_args *uap)
1540 {
1541 	struct proc *p = td->td_proc;
1542 	int s;
1543 	long jobref;
1544 	struct aiocblist *cb, *ncb;
1545 	struct aiocb *ujob;
1546 	struct kaioinfo *ki;
1547 
1548 	ujob = uap->aiocbp;
1549 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1550 	if (jobref == -1 || jobref == 0)
1551 		return (EINVAL);
1552 
1553 	ki = p->p_aioinfo;
1554 	if (ki == NULL)
1555 		return (EINVAL);
1556 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1557 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1558 		    jobref) {
1559 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1560 				p->p_stats->p_ru.ru_oublock +=
1561 				    cb->outputcharge;
1562 				cb->outputcharge = 0;
1563 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1564 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1565 				cb->inputcharge = 0;
1566 			}
1567 			goto done;
1568 		}
1569 	}
1570 	s = splbio();
1571 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1572 		ncb = TAILQ_NEXT(cb, plist);
1573 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1574 		    == jobref) {
1575 			break;
1576 		}
1577 	}
1578 	splx(s);
1579  done:
1580 	if (cb != NULL) {
1581 		if (ujob == cb->uuaiocb) {
1582 			td->td_retval[0] =
1583 			    cb->uaiocb._aiocb_private.status;
1584 		} else
1585 			td->td_retval[0] = EFAULT;
1586 		aio_free_entry(cb);
1587 		return (0);
1588 	}
1589 	return (EINVAL);
1590 }
1591 
1592 /*
1593  * Allow a process to wakeup when any of the I/O requests are completed.
1594  */
1595 int
1596 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1597 {
1598 	struct proc *p = td->td_proc;
1599 	struct timeval atv;
1600 	struct timespec ts;
1601 	struct aiocb *const *cbptr, *cbp;
1602 	struct kaioinfo *ki;
1603 	struct aiocblist *cb;
1604 	int i;
1605 	int njoblist;
1606 	int error, s, timo;
1607 	long *ijoblist;
1608 	struct aiocb **ujoblist;
1609 
1610 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1611 		return (EINVAL);
1612 
1613 	timo = 0;
1614 	if (uap->timeout) {
1615 		/* Get timespec struct. */
1616 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1617 			return (error);
1618 
1619 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1620 			return (EINVAL);
1621 
1622 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1623 		if (itimerfix(&atv))
1624 			return (EINVAL);
1625 		timo = tvtohz(&atv);
1626 	}
1627 
1628 	ki = p->p_aioinfo;
1629 	if (ki == NULL)
1630 		return (EAGAIN);
1631 
1632 	njoblist = 0;
1633 	ijoblist = uma_zalloc(aiol_zone, M_WAITOK);
1634 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1635 	cbptr = uap->aiocbp;
1636 
1637 	for (i = 0; i < uap->nent; i++) {
1638 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1639 		if (cbp == 0)
1640 			continue;
1641 		ujoblist[njoblist] = cbp;
1642 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1643 		njoblist++;
1644 	}
1645 
1646 	if (njoblist == 0) {
1647 		uma_zfree(aiol_zone, ijoblist);
1648 		uma_zfree(aiol_zone, ujoblist);
1649 		return (0);
1650 	}
1651 
1652 	error = 0;
1653 	for (;;) {
1654 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1655 			for (i = 0; i < njoblist; i++) {
1656 				if (((intptr_t)
1657 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1658 				    ijoblist[i]) {
1659 					if (ujoblist[i] != cb->uuaiocb)
1660 						error = EINVAL;
1661 					uma_zfree(aiol_zone, ijoblist);
1662 					uma_zfree(aiol_zone, ujoblist);
1663 					return (error);
1664 				}
1665 			}
1666 		}
1667 
1668 		s = splbio();
1669 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1670 		    TAILQ_NEXT(cb, plist)) {
1671 			for (i = 0; i < njoblist; i++) {
1672 				if (((intptr_t)
1673 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1674 				    ijoblist[i]) {
1675 					splx(s);
1676 					if (ujoblist[i] != cb->uuaiocb)
1677 						error = EINVAL;
1678 					uma_zfree(aiol_zone, ijoblist);
1679 					uma_zfree(aiol_zone, ujoblist);
1680 					return (error);
1681 				}
1682 			}
1683 		}
1684 
1685 		ki->kaio_flags |= KAIO_WAKEUP;
1686 		error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1687 		splx(s);
1688 
1689 		if (error == ERESTART || error == EINTR) {
1690 			uma_zfree(aiol_zone, ijoblist);
1691 			uma_zfree(aiol_zone, ujoblist);
1692 			return (EINTR);
1693 		} else if (error == EWOULDBLOCK) {
1694 			uma_zfree(aiol_zone, ijoblist);
1695 			uma_zfree(aiol_zone, ujoblist);
1696 			return (EAGAIN);
1697 		}
1698 	}
1699 
1700 /* NOTREACHED */
1701 	return (EINVAL);
1702 }
1703 
1704 /*
1705  * aio_cancel cancels any non-physio aio operations not currently in
1706  * progress.
1707  */
1708 int
1709 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1710 {
1711 	struct proc *p = td->td_proc;
1712 	struct kaioinfo *ki;
1713 	struct aiocblist *cbe, *cbn;
1714 	struct file *fp;
1715 	struct filedesc *fdp;
1716 	struct socket *so;
1717 	struct proc *po;
1718 	int s,error;
1719 	int cancelled=0;
1720 	int notcancelled=0;
1721 	struct vnode *vp;
1722 
1723 	fdp = p->p_fd;
1724 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1725 	    (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1726 		return (EBADF);
1727 
1728         if (fp->f_type == DTYPE_VNODE) {
1729 		vp = fp->f_vnode;
1730 
1731 		if (vn_isdisk(vp,&error)) {
1732 			td->td_retval[0] = AIO_NOTCANCELED;
1733         	        return (0);
1734 		}
1735 	} else if (fp->f_type == DTYPE_SOCKET) {
1736 		so = fp->f_data;
1737 
1738 		s = splnet();
1739 
1740 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1741 			cbn = TAILQ_NEXT(cbe, list);
1742 			if ((uap->aiocbp == NULL) ||
1743 				(uap->aiocbp == cbe->uuaiocb) ) {
1744 				po = cbe->userproc;
1745 				ki = po->p_aioinfo;
1746 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1747 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1748 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1749 				if (ki->kaio_flags & KAIO_WAKEUP) {
1750 					wakeup(po);
1751 				}
1752 				cbe->jobstate = JOBST_JOBFINISHED;
1753 				cbe->uaiocb._aiocb_private.status=-1;
1754 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1755 				cancelled++;
1756 /* XXX cancelled, knote? */
1757 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1758 				    SIGEV_SIGNAL) {
1759 					PROC_LOCK(cbe->userproc);
1760 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1761 					PROC_UNLOCK(cbe->userproc);
1762 				}
1763 				if (uap->aiocbp)
1764 					break;
1765 			}
1766 		}
1767 		splx(s);
1768 
1769 		if ((cancelled) && (uap->aiocbp)) {
1770 			td->td_retval[0] = AIO_CANCELED;
1771 			return (0);
1772 		}
1773 	}
1774 	ki=p->p_aioinfo;
1775 	if (ki == NULL)
1776 		goto done;
1777 	s = splnet();
1778 
1779 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1780 		cbn = TAILQ_NEXT(cbe, plist);
1781 
1782 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1783 		    ((uap->aiocbp == NULL ) ||
1784 		     (uap->aiocbp == cbe->uuaiocb))) {
1785 
1786 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1787 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1788                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1789                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1790                                     plist);
1791 				cancelled++;
1792 				ki->kaio_queue_finished_count++;
1793 				cbe->jobstate = JOBST_JOBFINISHED;
1794 				cbe->uaiocb._aiocb_private.status = -1;
1795 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1796 /* XXX cancelled, knote? */
1797 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1798 				    SIGEV_SIGNAL) {
1799 					PROC_LOCK(cbe->userproc);
1800 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1801 					PROC_UNLOCK(cbe->userproc);
1802 				}
1803 			} else {
1804 				notcancelled++;
1805 			}
1806 		}
1807 	}
1808 	splx(s);
1809 done:
1810 	if (notcancelled) {
1811 		td->td_retval[0] = AIO_NOTCANCELED;
1812 		return (0);
1813 	}
1814 	if (cancelled) {
1815 		td->td_retval[0] = AIO_CANCELED;
1816 		return (0);
1817 	}
1818 	td->td_retval[0] = AIO_ALLDONE;
1819 
1820 	return (0);
1821 }
1822 
1823 /*
1824  * aio_error is implemented in the kernel level for compatibility purposes only.
1825  * For a user mode async implementation, it would be best to do it in a userland
1826  * subroutine.
1827  */
1828 int
1829 aio_error(struct thread *td, struct aio_error_args *uap)
1830 {
1831 	struct proc *p = td->td_proc;
1832 	int s;
1833 	struct aiocblist *cb;
1834 	struct kaioinfo *ki;
1835 	long jobref;
1836 
1837 	ki = p->p_aioinfo;
1838 	if (ki == NULL)
1839 		return (EINVAL);
1840 
1841 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1842 	if ((jobref == -1) || (jobref == 0))
1843 		return (EINVAL);
1844 
1845 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1846 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1847 		    jobref) {
1848 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1849 			return (0);
1850 		}
1851 	}
1852 
1853 	s = splnet();
1854 
1855 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1856 	    plist)) {
1857 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1858 		    jobref) {
1859 			td->td_retval[0] = EINPROGRESS;
1860 			splx(s);
1861 			return (0);
1862 		}
1863 	}
1864 
1865 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1866 	    plist)) {
1867 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1868 		    jobref) {
1869 			td->td_retval[0] = EINPROGRESS;
1870 			splx(s);
1871 			return (0);
1872 		}
1873 	}
1874 	splx(s);
1875 
1876 	s = splbio();
1877 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1878 	    plist)) {
1879 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1880 		    jobref) {
1881 			td->td_retval[0] = cb->uaiocb._aiocb_private.error;
1882 			splx(s);
1883 			return (0);
1884 		}
1885 	}
1886 
1887 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1888 	    plist)) {
1889 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1890 		    jobref) {
1891 			td->td_retval[0] = EINPROGRESS;
1892 			splx(s);
1893 			return (0);
1894 		}
1895 	}
1896 	splx(s);
1897 
1898 #if (0)
1899 	/*
1900 	 * Hack for lio.
1901 	 */
1902 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1903 	if (status == -1)
1904 		return fuword(&uap->aiocbp->_aiocb_private.error);
1905 #endif
1906 	return (EINVAL);
1907 }
1908 
1909 /* syscall - asynchronous read from a file (REALTIME) */
1910 int
1911 aio_read(struct thread *td, struct aio_read_args *uap)
1912 {
1913 
1914 	return aio_aqueue(td, uap->aiocbp, LIO_READ);
1915 }
1916 
1917 /* syscall - asynchronous write to a file (REALTIME) */
1918 int
1919 aio_write(struct thread *td, struct aio_write_args *uap)
1920 {
1921 
1922 	return aio_aqueue(td, uap->aiocbp, LIO_WRITE);
1923 }
1924 
1925 /* syscall - list directed I/O (REALTIME) */
1926 int
1927 lio_listio(struct thread *td, struct lio_listio_args *uap)
1928 {
1929 	struct proc *p = td->td_proc;
1930 	int nent, nentqueued;
1931 	struct aiocb *iocb, * const *cbptr;
1932 	struct aiocblist *cb;
1933 	struct kaioinfo *ki;
1934 	struct aio_liojob *lj;
1935 	int error, runningcode;
1936 	int nerror;
1937 	int i;
1938 	int s;
1939 
1940 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1941 		return (EINVAL);
1942 
1943 	nent = uap->nent;
1944 	if (nent < 0 || nent > AIO_LISTIO_MAX)
1945 		return (EINVAL);
1946 
1947 	if (p->p_aioinfo == NULL)
1948 		aio_init_aioinfo(p);
1949 
1950 	if ((nent + num_queue_count) > max_queue_count)
1951 		return (EAGAIN);
1952 
1953 	ki = p->p_aioinfo;
1954 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1955 		return (EAGAIN);
1956 
1957 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
1958 	if (!lj)
1959 		return (EAGAIN);
1960 
1961 	lj->lioj_flags = 0;
1962 	lj->lioj_buffer_count = 0;
1963 	lj->lioj_buffer_finished_count = 0;
1964 	lj->lioj_queue_count = 0;
1965 	lj->lioj_queue_finished_count = 0;
1966 	lj->lioj_ki = ki;
1967 
1968 	/*
1969 	 * Setup signal.
1970 	 */
1971 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1972 		error = copyin(uap->sig, &lj->lioj_signal,
1973 			       sizeof(lj->lioj_signal));
1974 		if (error) {
1975 			uma_zfree(aiolio_zone, lj);
1976 			return (error);
1977 		}
1978 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1979 			uma_zfree(aiolio_zone, lj);
1980 			return (EINVAL);
1981 		}
1982 		lj->lioj_flags |= LIOJ_SIGNAL;
1983 	}
1984 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1985 	/*
1986 	 * Get pointers to the list of I/O requests.
1987 	 */
1988 	nerror = 0;
1989 	nentqueued = 0;
1990 	cbptr = uap->acb_list;
1991 	for (i = 0; i < uap->nent; i++) {
1992 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1993 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1994 			error = _aio_aqueue(td, iocb, lj, 0);
1995 			if (error == 0)
1996 				nentqueued++;
1997 			else
1998 				nerror++;
1999 		}
2000 	}
2001 
2002 	/*
2003 	 * If we haven't queued any, then just return error.
2004 	 */
2005 	if (nentqueued == 0)
2006 		return (0);
2007 
2008 	/*
2009 	 * Calculate the appropriate error return.
2010 	 */
2011 	runningcode = 0;
2012 	if (nerror)
2013 		runningcode = EIO;
2014 
2015 	if (uap->mode == LIO_WAIT) {
2016 		int command, found, jobref;
2017 
2018 		for (;;) {
2019 			found = 0;
2020 			for (i = 0; i < uap->nent; i++) {
2021 				/*
2022 				 * Fetch address of the control buf pointer in
2023 				 * user space.
2024 				 */
2025 				iocb = (struct aiocb *)
2026 				    (intptr_t)fuword(&cbptr[i]);
2027 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
2028 				    == 0))
2029 					continue;
2030 
2031 				/*
2032 				 * Fetch the associated command from user space.
2033 				 */
2034 				command = fuword(&iocb->aio_lio_opcode);
2035 				if (command == LIO_NOP) {
2036 					found++;
2037 					continue;
2038 				}
2039 
2040 				jobref =
2041 				    fuword(&iocb->_aiocb_private.kernelinfo);
2042 
2043 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
2044 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2045 					    == jobref) {
2046 						if (cb->uaiocb.aio_lio_opcode
2047 						    == LIO_WRITE) {
2048 							p->p_stats->p_ru.ru_oublock
2049 							    +=
2050 							    cb->outputcharge;
2051 							cb->outputcharge = 0;
2052 						} else if (cb->uaiocb.aio_lio_opcode
2053 						    == LIO_READ) {
2054 							p->p_stats->p_ru.ru_inblock
2055 							    += cb->inputcharge;
2056 							cb->inputcharge = 0;
2057 						}
2058 						found++;
2059 						break;
2060 					}
2061 				}
2062 
2063 				s = splbio();
2064 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
2065 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
2066 					    == jobref) {
2067 						found++;
2068 						break;
2069 					}
2070 				}
2071 				splx(s);
2072 			}
2073 
2074 			/*
2075 			 * If all I/Os have been disposed of, then we can
2076 			 * return.
2077 			 */
2078 			if (found == nentqueued)
2079 				return (runningcode);
2080 
2081 			ki->kaio_flags |= KAIO_WAKEUP;
2082 			error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
2083 
2084 			if (error == EINTR)
2085 				return (EINTR);
2086 			else if (error == EWOULDBLOCK)
2087 				return (EAGAIN);
2088 		}
2089 	}
2090 
2091 	return (runningcode);
2092 }
2093 
2094 /*
2095  * This is a weird hack so that we can post a signal.  It is safe to do so from
2096  * a timeout routine, but *not* from an interrupt routine.
2097  */
2098 static void
2099 process_signal(void *aioj)
2100 {
2101 	struct aiocblist *aiocbe = aioj;
2102 	struct aio_liojob *lj = aiocbe->lio;
2103 	struct aiocb *cb = &aiocbe->uaiocb;
2104 
2105 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2106 		(lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2107 		PROC_LOCK(lj->lioj_ki->kaio_p);
2108 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2109 		PROC_UNLOCK(lj->lioj_ki->kaio_p);
2110 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2111 	}
2112 
2113 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2114 		PROC_LOCK(aiocbe->userproc);
2115 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2116 		PROC_UNLOCK(aiocbe->userproc);
2117 	}
2118 }
2119 
2120 /*
2121  * Interrupt handler for physio, performs the necessary process wakeups, and
2122  * signals.
2123  */
2124 static void
2125 aio_physwakeup(struct buf *bp)
2126 {
2127 	struct aiocblist *aiocbe;
2128 	struct proc *p;
2129 	struct kaioinfo *ki;
2130 	struct aio_liojob *lj;
2131 
2132 	wakeup(bp);
2133 
2134 	aiocbe = (struct aiocblist *)bp->b_caller1;
2135 	if (aiocbe) {
2136 		p = aiocbe->userproc;
2137 
2138 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2139 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2140 		aiocbe->uaiocb._aiocb_private.error = 0;
2141 		aiocbe->jobflags |= AIOCBLIST_DONE;
2142 
2143 		if (bp->b_ioflags & BIO_ERROR)
2144 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2145 
2146 		lj = aiocbe->lio;
2147 		if (lj) {
2148 			lj->lioj_buffer_finished_count++;
2149 
2150 			/*
2151 			 * wakeup/signal if all of the interrupt jobs are done.
2152 			 */
2153 			if (lj->lioj_buffer_finished_count ==
2154 			    lj->lioj_buffer_count) {
2155 				/*
2156 				 * Post a signal if it is called for.
2157 				 */
2158 				if ((lj->lioj_flags &
2159 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2160 				    LIOJ_SIGNAL) {
2161 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2162 					aiocbe->timeouthandle =
2163 						timeout(process_signal,
2164 							aiocbe, 0);
2165 				}
2166 			}
2167 		}
2168 
2169 		ki = p->p_aioinfo;
2170 		if (ki) {
2171 			ki->kaio_buffer_finished_count++;
2172 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2173 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2174 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2175 
2176 			KNOTE(&aiocbe->klist, 0);
2177 			/* Do the wakeup. */
2178 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2179 				ki->kaio_flags &= ~KAIO_WAKEUP;
2180 				wakeup(p);
2181 			}
2182 		}
2183 
2184 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2185 			aiocbe->timeouthandle =
2186 				timeout(process_signal, aiocbe, 0);
2187 	}
2188 }
2189 
2190 /* syscall - wait for the next completion of an aio request */
2191 int
2192 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2193 {
2194 	struct proc *p = td->td_proc;
2195 	struct timeval atv;
2196 	struct timespec ts;
2197 	struct kaioinfo *ki;
2198 	struct aiocblist *cb = NULL;
2199 	int error, s, timo;
2200 
2201 	suword(uap->aiocbp, (int)NULL);
2202 
2203 	timo = 0;
2204 	if (uap->timeout) {
2205 		/* Get timespec struct. */
2206 		error = copyin(uap->timeout, &ts, sizeof(ts));
2207 		if (error)
2208 			return (error);
2209 
2210 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2211 			return (EINVAL);
2212 
2213 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2214 		if (itimerfix(&atv))
2215 			return (EINVAL);
2216 		timo = tvtohz(&atv);
2217 	}
2218 
2219 	ki = p->p_aioinfo;
2220 	if (ki == NULL)
2221 		return (EAGAIN);
2222 
2223 	for (;;) {
2224 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2225 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2226 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2227 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2228 				p->p_stats->p_ru.ru_oublock +=
2229 				    cb->outputcharge;
2230 				cb->outputcharge = 0;
2231 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2232 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2233 				cb->inputcharge = 0;
2234 			}
2235 			aio_free_entry(cb);
2236 			return (cb->uaiocb._aiocb_private.error);
2237 		}
2238 
2239 		s = splbio();
2240  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2241 			splx(s);
2242 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2243 			td->td_retval[0] = cb->uaiocb._aiocb_private.status;
2244 			aio_free_entry(cb);
2245 			return (cb->uaiocb._aiocb_private.error);
2246 		}
2247 
2248 		ki->kaio_flags |= KAIO_WAKEUP;
2249 		error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2250 		splx(s);
2251 
2252 		if (error == ERESTART)
2253 			return (EINTR);
2254 		else if (error < 0)
2255 			return (error);
2256 		else if (error == EINTR)
2257 			return (EINTR);
2258 		else if (error == EWOULDBLOCK)
2259 			return (EAGAIN);
2260 	}
2261 }
2262 
2263 /* kqueue attach function */
2264 static int
2265 filt_aioattach(struct knote *kn)
2266 {
2267 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2268 
2269 	/*
2270 	 * The aiocbe pointer must be validated before using it, so
2271 	 * registration is restricted to the kernel; the user cannot
2272 	 * set EV_FLAG1.
2273 	 */
2274 	if ((kn->kn_flags & EV_FLAG1) == 0)
2275 		return (EPERM);
2276 	kn->kn_flags &= ~EV_FLAG1;
2277 
2278 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2279 
2280 	return (0);
2281 }
2282 
2283 /* kqueue detach function */
2284 static void
2285 filt_aiodetach(struct knote *kn)
2286 {
2287 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2288 
2289 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2290 }
2291 
2292 /* kqueue filter function */
2293 /*ARGSUSED*/
2294 static int
2295 filt_aio(struct knote *kn, long hint)
2296 {
2297 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2298 
2299 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2300 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2301 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2302 		return (0);
2303 	kn->kn_flags |= EV_EOF;
2304 	return (1);
2305 }
2306