xref: /freebsd/sys/kern/vfs_aio.c (revision a50d73d5782a351ad83e8d1f84d11720a12e70d3)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. John S. Dyson's name may not be used to endorse or promote products
12  *    derived from this software without specific prior written permission.
13  *
14  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
15  * bad that happens because of using this software isn't the responsibility
16  * of the author.  This software is distributed AS-IS.
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/cdefs.h>
24 __FBSDID("$FreeBSD$");
25 
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
29 #include <sys/bio.h>
30 #include <sys/buf.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
39 #include <sys/file.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
45 #include <sys/proc.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sema.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysctl.h>
56 #include <sys/syslog.h>
57 #include <sys/sx.h>
58 #include <sys/taskqueue.h>
59 #include <sys/vnode.h>
60 #include <sys/conf.h>
61 #include <sys/event.h>
62 #include <sys/mount.h>
63 #include <geom/geom.h>
64 
65 #include <machine/atomic.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_extern.h>
70 #include <vm/pmap.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_object.h>
73 #include <vm/uma.h>
74 #include <sys/aio.h>
75 
76 /*
77  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
78  * overflow. (XXX will be removed soon.)
79  */
80 static u_long jobrefid;
81 
82 /*
83  * Counter for aio_fsync.
84  */
85 static uint64_t jobseqno;
86 
87 #ifndef MAX_AIO_PER_PROC
88 #define MAX_AIO_PER_PROC	32
89 #endif
90 
91 #ifndef MAX_AIO_QUEUE_PER_PROC
92 #define MAX_AIO_QUEUE_PER_PROC	256
93 #endif
94 
95 #ifndef MAX_AIO_QUEUE
96 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
97 #endif
98 
99 #ifndef MAX_BUF_AIO
100 #define MAX_BUF_AIO		16
101 #endif
102 
103 FEATURE(aio, "Asynchronous I/O");
104 SYSCTL_DECL(_p1003_1b);
105 
106 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
107 static MALLOC_DEFINE(M_AIO, "aio", "structures for asynchronous I/O");
108 
109 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
110     "Async IO management");
111 
112 static int enable_aio_unsafe = 0;
113 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
114     "Permit asynchronous IO on all file types, not just known-safe types");
115 
116 static unsigned int unsafe_warningcnt = 1;
117 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
118     &unsafe_warningcnt, 0,
119     "Warnings that will be triggered upon failed IO requests on unsafe files");
120 
121 static int max_aio_procs = MAX_AIO_PROCS;
122 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
123     "Maximum number of kernel processes to use for handling async IO ");
124 
125 static int num_aio_procs = 0;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
127     "Number of presently active kernel processes for async IO");
128 
129 /*
130  * The code will adjust the actual number of AIO processes towards this
131  * number when it gets a chance.
132  */
133 static int target_aio_procs = TARGET_AIO_PROCS;
134 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
135     0,
136     "Preferred number of ready kernel processes for async IO");
137 
138 static int max_queue_count = MAX_AIO_QUEUE;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
140     "Maximum number of aio requests to queue, globally");
141 
142 static int num_queue_count = 0;
143 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
144     "Number of queued aio requests");
145 
146 static int num_buf_aio = 0;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
148     "Number of aio requests presently handled by the buf subsystem");
149 
150 static int num_unmapped_aio = 0;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
152     0,
153     "Number of aio requests presently handled by unmapped I/O buffers");
154 
155 /* Number of async I/O processes in the process of being started */
156 /* XXX This should be local to aio_aqueue() */
157 static int num_aio_resv_start = 0;
158 
159 static int aiod_lifetime;
160 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
161     "Maximum lifetime for idle aiod");
162 
163 static int max_aio_per_proc = MAX_AIO_PER_PROC;
164 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
165     0,
166     "Maximum active aio requests per process");
167 
168 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
170     &max_aio_queue_per_proc, 0,
171     "Maximum queued aio requests per process");
172 
173 static int max_buf_aio = MAX_BUF_AIO;
174 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
175     "Maximum buf aio requests per process");
176 
177 /*
178  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
179  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
180  * vfs.aio.aio_listio_max.
181  */
182 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
183     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
184     0, "Maximum aio requests for a single lio_listio call");
185 
186 #ifdef COMPAT_FREEBSD6
187 typedef struct oaiocb {
188 	int	aio_fildes;		/* File descriptor */
189 	off_t	aio_offset;		/* File offset for I/O */
190 	volatile void *aio_buf;         /* I/O buffer in process space */
191 	size_t	aio_nbytes;		/* Number of bytes for I/O */
192 	struct	osigevent aio_sigevent;	/* Signal to deliver */
193 	int	aio_lio_opcode;		/* LIO opcode */
194 	int	aio_reqprio;		/* Request priority -- ignored */
195 	struct	__aiocb_private	_aiocb_private;
196 } oaiocb_t;
197 #endif
198 
199 /*
200  * Below is a key of locks used to protect each member of struct kaiocb
201  * aioliojob and kaioinfo and any backends.
202  *
203  * * - need not protected
204  * a - locked by kaioinfo lock
205  * b - locked by backend lock, the backend lock can be null in some cases,
206  *     for example, BIO belongs to this type, in this case, proc lock is
207  *     reused.
208  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
209  */
210 
211 /*
212  * If the routine that services an AIO request blocks while running in an
213  * AIO kernel process it can starve other I/O requests.  BIO requests
214  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
215  * processes at all.  Socket I/O requests use a separate pool of
216  * kprocs and also force non-blocking I/O.  Other file I/O requests
217  * use the generic fo_read/fo_write operations which can block.  The
218  * fsync and mlock operations can also block while executing.  Ideally
219  * none of these requests would block while executing.
220  *
221  * Note that the service routines cannot toggle O_NONBLOCK in the file
222  * structure directly while handling a request due to races with
223  * userland threads.
224  */
225 
226 /* jobflags */
227 #define	KAIOCB_QUEUEING		0x01
228 #define	KAIOCB_CANCELLED	0x02
229 #define	KAIOCB_CANCELLING	0x04
230 #define	KAIOCB_CHECKSYNC	0x08
231 #define	KAIOCB_CLEARED		0x10
232 #define	KAIOCB_FINISHED		0x20
233 
234 /*
235  * AIO process info
236  */
237 #define AIOP_FREE	0x1			/* proc on free queue */
238 
239 struct aioproc {
240 	int	aioprocflags;			/* (c) AIO proc flags */
241 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
242 	struct	proc *aioproc;			/* (*) the AIO proc */
243 };
244 
245 /*
246  * data-structure for lio signal management
247  */
248 struct aioliojob {
249 	int	lioj_flags;			/* (a) listio flags */
250 	int	lioj_count;			/* (a) count of jobs */
251 	int	lioj_finished_count;		/* (a) count of finished jobs */
252 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
253 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
254 	struct	knlist klist;			/* (a) list of knotes */
255 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
256 };
257 
258 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
259 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
260 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
261 
262 /*
263  * per process aio data structure
264  */
265 struct kaioinfo {
266 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
267 	int	kaio_flags;		/* (a) per process kaio flags */
268 	int	kaio_active_count;	/* (c) number of currently used AIOs */
269 	int	kaio_count;		/* (a) size of AIO queue */
270 	int	kaio_buffer_count;	/* (a) number of bio buffers */
271 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
272 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
273 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
274 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
275 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
276 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
277 	struct	task kaio_task;		/* (*) task to kick aio processes */
278 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
279 };
280 
281 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
282 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
283 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
284 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
285 
286 #define KAIO_RUNDOWN	0x1	/* process is being run down */
287 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
288 
289 /*
290  * Operations used to interact with userland aio control blocks.
291  * Different ABIs provide their own operations.
292  */
293 struct aiocb_ops {
294 	int	(*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty);
295 	long	(*fetch_status)(struct aiocb *ujob);
296 	long	(*fetch_error)(struct aiocb *ujob);
297 	int	(*store_status)(struct aiocb *ujob, long status);
298 	int	(*store_error)(struct aiocb *ujob, long error);
299 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
300 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
301 };
302 
303 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
304 static struct sema aio_newproc_sem;
305 static struct mtx aio_job_mtx;
306 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
307 static struct unrhdr *aiod_unr;
308 
309 static void	aio_biocleanup(struct bio *bp);
310 void		aio_init_aioinfo(struct proc *p);
311 static int	aio_onceonly(void);
312 static int	aio_free_entry(struct kaiocb *job);
313 static void	aio_process_rw(struct kaiocb *job);
314 static void	aio_process_sync(struct kaiocb *job);
315 static void	aio_process_mlock(struct kaiocb *job);
316 static void	aio_schedule_fsync(void *context, int pending);
317 static int	aio_newproc(int *);
318 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
319 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
320 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
321 static void	aio_biowakeup(struct bio *bp);
322 static void	aio_proc_rundown(void *arg, struct proc *p);
323 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
324 		    struct image_params *imgp);
325 static int	aio_qbio(struct proc *p, struct kaiocb *job);
326 static void	aio_daemon(void *param);
327 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
328 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
329 static int	aio_kick(struct proc *userp);
330 static void	aio_kick_nowait(struct proc *userp);
331 static void	aio_kick_helper(void *context, int pending);
332 static int	filt_aioattach(struct knote *kn);
333 static void	filt_aiodetach(struct knote *kn);
334 static int	filt_aio(struct knote *kn, long hint);
335 static int	filt_lioattach(struct knote *kn);
336 static void	filt_liodetach(struct knote *kn);
337 static int	filt_lio(struct knote *kn, long hint);
338 
339 /*
340  * Zones for:
341  * 	kaio	Per process async io info
342  *	aiocb	async io jobs
343  *	aiolio	list io jobs
344  */
345 static uma_zone_t kaio_zone, aiocb_zone, aiolio_zone;
346 
347 /* kqueue filters for aio */
348 static struct filterops aio_filtops = {
349 	.f_isfd = 0,
350 	.f_attach = filt_aioattach,
351 	.f_detach = filt_aiodetach,
352 	.f_event = filt_aio,
353 };
354 static struct filterops lio_filtops = {
355 	.f_isfd = 0,
356 	.f_attach = filt_lioattach,
357 	.f_detach = filt_liodetach,
358 	.f_event = filt_lio
359 };
360 
361 static eventhandler_tag exit_tag, exec_tag;
362 
363 TASKQUEUE_DEFINE_THREAD(aiod_kick);
364 
365 /*
366  * Main operations function for use as a kernel module.
367  */
368 static int
369 aio_modload(struct module *module, int cmd, void *arg)
370 {
371 	int error = 0;
372 
373 	switch (cmd) {
374 	case MOD_LOAD:
375 		aio_onceonly();
376 		break;
377 	case MOD_SHUTDOWN:
378 		break;
379 	default:
380 		error = EOPNOTSUPP;
381 		break;
382 	}
383 	return (error);
384 }
385 
386 static moduledata_t aio_mod = {
387 	"aio",
388 	&aio_modload,
389 	NULL
390 };
391 
392 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
393 MODULE_VERSION(aio, 1);
394 
395 /*
396  * Startup initialization
397  */
398 static int
399 aio_onceonly(void)
400 {
401 
402 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
403 	    EVENTHANDLER_PRI_ANY);
404 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
405 	    NULL, EVENTHANDLER_PRI_ANY);
406 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
407 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
408 	TAILQ_INIT(&aio_freeproc);
409 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
410 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
411 	TAILQ_INIT(&aio_jobs);
412 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
413 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
414 	    NULL, NULL, UMA_ALIGN_PTR, 0);
415 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
416 	    NULL, NULL, UMA_ALIGN_PTR, 0);
417 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
418 	    NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
419 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
420 	jobrefid = 1;
421 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
422 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
423 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
424 
425 	return (0);
426 }
427 
428 /*
429  * Init the per-process aioinfo structure.  The aioinfo limits are set
430  * per-process for user limit (resource) management.
431  */
432 void
433 aio_init_aioinfo(struct proc *p)
434 {
435 	struct kaioinfo *ki;
436 
437 	ki = uma_zalloc(kaio_zone, M_WAITOK);
438 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
439 	ki->kaio_flags = 0;
440 	ki->kaio_active_count = 0;
441 	ki->kaio_count = 0;
442 	ki->kaio_buffer_count = 0;
443 	TAILQ_INIT(&ki->kaio_all);
444 	TAILQ_INIT(&ki->kaio_done);
445 	TAILQ_INIT(&ki->kaio_jobqueue);
446 	TAILQ_INIT(&ki->kaio_liojoblist);
447 	TAILQ_INIT(&ki->kaio_syncqueue);
448 	TAILQ_INIT(&ki->kaio_syncready);
449 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
450 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
451 	PROC_LOCK(p);
452 	if (p->p_aioinfo == NULL) {
453 		p->p_aioinfo = ki;
454 		PROC_UNLOCK(p);
455 	} else {
456 		PROC_UNLOCK(p);
457 		mtx_destroy(&ki->kaio_mtx);
458 		uma_zfree(kaio_zone, ki);
459 	}
460 
461 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
462 		aio_newproc(NULL);
463 }
464 
465 static int
466 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
467 {
468 	struct thread *td;
469 	int error;
470 
471 	error = sigev_findtd(p, sigev, &td);
472 	if (error)
473 		return (error);
474 	if (!KSI_ONQ(ksi)) {
475 		ksiginfo_set_sigev(ksi, sigev);
476 		ksi->ksi_code = SI_ASYNCIO;
477 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
478 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
479 	}
480 	PROC_UNLOCK(p);
481 	return (error);
482 }
483 
484 /*
485  * Free a job entry.  Wait for completion if it is currently active, but don't
486  * delay forever.  If we delay, we return a flag that says that we have to
487  * restart the queue scan.
488  */
489 static int
490 aio_free_entry(struct kaiocb *job)
491 {
492 	struct kaioinfo *ki;
493 	struct aioliojob *lj;
494 	struct proc *p;
495 
496 	p = job->userproc;
497 	MPASS(curproc == p);
498 	ki = p->p_aioinfo;
499 	MPASS(ki != NULL);
500 
501 	AIO_LOCK_ASSERT(ki, MA_OWNED);
502 	MPASS(job->jobflags & KAIOCB_FINISHED);
503 
504 	atomic_subtract_int(&num_queue_count, 1);
505 
506 	ki->kaio_count--;
507 	MPASS(ki->kaio_count >= 0);
508 
509 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
510 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
511 
512 	lj = job->lio;
513 	if (lj) {
514 		lj->lioj_count--;
515 		lj->lioj_finished_count--;
516 
517 		if (lj->lioj_count == 0) {
518 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
519 			/* lio is going away, we need to destroy any knotes */
520 			knlist_delete(&lj->klist, curthread, 1);
521 			PROC_LOCK(p);
522 			sigqueue_take(&lj->lioj_ksi);
523 			PROC_UNLOCK(p);
524 			uma_zfree(aiolio_zone, lj);
525 		}
526 	}
527 
528 	/* job is going away, we need to destroy any knotes */
529 	knlist_delete(&job->klist, curthread, 1);
530 	PROC_LOCK(p);
531 	sigqueue_take(&job->ksi);
532 	PROC_UNLOCK(p);
533 
534 	AIO_UNLOCK(ki);
535 
536 	/*
537 	 * The thread argument here is used to find the owning process
538 	 * and is also passed to fo_close() which may pass it to various
539 	 * places such as devsw close() routines.  Because of that, we
540 	 * need a thread pointer from the process owning the job that is
541 	 * persistent and won't disappear out from under us or move to
542 	 * another process.
543 	 *
544 	 * Currently, all the callers of this function call it to remove
545 	 * a kaiocb from the current process' job list either via a
546 	 * syscall or due to the current process calling exit() or
547 	 * execve().  Thus, we know that p == curproc.  We also know that
548 	 * curthread can't exit since we are curthread.
549 	 *
550 	 * Therefore, we use curthread as the thread to pass to
551 	 * knlist_delete().  This does mean that it is possible for the
552 	 * thread pointer at close time to differ from the thread pointer
553 	 * at open time, but this is already true of file descriptors in
554 	 * a multithreaded process.
555 	 */
556 	if (job->fd_file)
557 		fdrop(job->fd_file, curthread);
558 	crfree(job->cred);
559 	if (job->uiop != &job->uio)
560 		free(job->uiop, M_IOV);
561 	uma_zfree(aiocb_zone, job);
562 	AIO_LOCK(ki);
563 
564 	return (0);
565 }
566 
567 static void
568 aio_proc_rundown_exec(void *arg, struct proc *p,
569     struct image_params *imgp __unused)
570 {
571    	aio_proc_rundown(arg, p);
572 }
573 
574 static int
575 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
576 {
577 	aio_cancel_fn_t *func;
578 	int cancelled;
579 
580 	AIO_LOCK_ASSERT(ki, MA_OWNED);
581 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
582 		return (0);
583 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
584 	job->jobflags |= KAIOCB_CANCELLED;
585 
586 	func = job->cancel_fn;
587 
588 	/*
589 	 * If there is no cancel routine, just leave the job marked as
590 	 * cancelled.  The job should be in active use by a caller who
591 	 * should complete it normally or when it fails to install a
592 	 * cancel routine.
593 	 */
594 	if (func == NULL)
595 		return (0);
596 
597 	/*
598 	 * Set the CANCELLING flag so that aio_complete() will defer
599 	 * completions of this job.  This prevents the job from being
600 	 * freed out from under the cancel callback.  After the
601 	 * callback any deferred completion (whether from the callback
602 	 * or any other source) will be completed.
603 	 */
604 	job->jobflags |= KAIOCB_CANCELLING;
605 	AIO_UNLOCK(ki);
606 	func(job);
607 	AIO_LOCK(ki);
608 	job->jobflags &= ~KAIOCB_CANCELLING;
609 	if (job->jobflags & KAIOCB_FINISHED) {
610 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
611 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
612 		aio_bio_done_notify(p, job);
613 	} else {
614 		/*
615 		 * The cancel callback might have scheduled an
616 		 * operation to cancel this request, but it is
617 		 * only counted as cancelled if the request is
618 		 * cancelled when the callback returns.
619 		 */
620 		cancelled = 0;
621 	}
622 	return (cancelled);
623 }
624 
625 /*
626  * Rundown the jobs for a given process.
627  */
628 static void
629 aio_proc_rundown(void *arg, struct proc *p)
630 {
631 	struct kaioinfo *ki;
632 	struct aioliojob *lj;
633 	struct kaiocb *job, *jobn;
634 
635 	KASSERT(curthread->td_proc == p,
636 	    ("%s: called on non-curproc", __func__));
637 	ki = p->p_aioinfo;
638 	if (ki == NULL)
639 		return;
640 
641 	AIO_LOCK(ki);
642 	ki->kaio_flags |= KAIO_RUNDOWN;
643 
644 restart:
645 
646 	/*
647 	 * Try to cancel all pending requests. This code simulates
648 	 * aio_cancel on all pending I/O requests.
649 	 */
650 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
651 		aio_cancel_job(p, ki, job);
652 	}
653 
654 	/* Wait for all running I/O to be finished */
655 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
656 		ki->kaio_flags |= KAIO_WAKEUP;
657 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
658 		goto restart;
659 	}
660 
661 	/* Free all completed I/O requests. */
662 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
663 		aio_free_entry(job);
664 
665 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
666 		if (lj->lioj_count == 0) {
667 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
668 			knlist_delete(&lj->klist, curthread, 1);
669 			PROC_LOCK(p);
670 			sigqueue_take(&lj->lioj_ksi);
671 			PROC_UNLOCK(p);
672 			uma_zfree(aiolio_zone, lj);
673 		} else {
674 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
675 			    lj->lioj_count, lj->lioj_finished_count);
676 		}
677 	}
678 	AIO_UNLOCK(ki);
679 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
680 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
681 	mtx_destroy(&ki->kaio_mtx);
682 	uma_zfree(kaio_zone, ki);
683 	p->p_aioinfo = NULL;
684 }
685 
686 /*
687  * Select a job to run (called by an AIO daemon).
688  */
689 static struct kaiocb *
690 aio_selectjob(struct aioproc *aiop)
691 {
692 	struct kaiocb *job;
693 	struct kaioinfo *ki;
694 	struct proc *userp;
695 
696 	mtx_assert(&aio_job_mtx, MA_OWNED);
697 restart:
698 	TAILQ_FOREACH(job, &aio_jobs, list) {
699 		userp = job->userproc;
700 		ki = userp->p_aioinfo;
701 
702 		if (ki->kaio_active_count < max_aio_per_proc) {
703 			TAILQ_REMOVE(&aio_jobs, job, list);
704 			if (!aio_clear_cancel_function(job))
705 				goto restart;
706 
707 			/* Account for currently active jobs. */
708 			ki->kaio_active_count++;
709 			break;
710 		}
711 	}
712 	return (job);
713 }
714 
715 /*
716  * Move all data to a permanent storage device.  This code
717  * simulates the fsync and fdatasync syscalls.
718  */
719 static int
720 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op)
721 {
722 	struct mount *mp;
723 	vm_object_t obj;
724 	int error;
725 
726 	for (;;) {
727 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
728 		if (error != 0)
729 			break;
730 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
731 		obj = vp->v_object;
732 		if (obj != NULL) {
733 			VM_OBJECT_WLOCK(obj);
734 			vm_object_page_clean(obj, 0, 0, 0);
735 			VM_OBJECT_WUNLOCK(obj);
736 		}
737 		if (op == LIO_DSYNC)
738 			error = VOP_FDATASYNC(vp, td);
739 		else
740 			error = VOP_FSYNC(vp, MNT_WAIT, td);
741 
742 		VOP_UNLOCK(vp);
743 		vn_finished_write(mp);
744 		if (error != ERELOOKUP)
745 			break;
746 	}
747 	return (error);
748 }
749 
750 /*
751  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
752  * does the I/O request for the non-bio version of the operations.  The normal
753  * vn operations are used, and this code should work in all instances for every
754  * type of file, including pipes, sockets, fifos, and regular files.
755  *
756  * XXX I don't think it works well for socket, pipe, and fifo.
757  */
758 static void
759 aio_process_rw(struct kaiocb *job)
760 {
761 	struct ucred *td_savedcred;
762 	struct thread *td;
763 	struct file *fp;
764 	ssize_t cnt;
765 	long msgsnd_st, msgsnd_end;
766 	long msgrcv_st, msgrcv_end;
767 	long oublock_st, oublock_end;
768 	long inblock_st, inblock_end;
769 	int error, opcode;
770 
771 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
772 	    job->uaiocb.aio_lio_opcode == LIO_READV ||
773 	    job->uaiocb.aio_lio_opcode == LIO_WRITE ||
774 	    job->uaiocb.aio_lio_opcode == LIO_WRITEV,
775 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
776 
777 	aio_switch_vmspace(job);
778 	td = curthread;
779 	td_savedcred = td->td_ucred;
780 	td->td_ucred = job->cred;
781 	job->uiop->uio_td = td;
782 	fp = job->fd_file;
783 
784 	opcode = job->uaiocb.aio_lio_opcode;
785 	cnt = job->uiop->uio_resid;
786 
787 	msgrcv_st = td->td_ru.ru_msgrcv;
788 	msgsnd_st = td->td_ru.ru_msgsnd;
789 	inblock_st = td->td_ru.ru_inblock;
790 	oublock_st = td->td_ru.ru_oublock;
791 
792 	/*
793 	 * aio_aqueue() acquires a reference to the file that is
794 	 * released in aio_free_entry().
795 	 */
796 	if (opcode == LIO_READ || opcode == LIO_READV) {
797 		if (job->uiop->uio_resid == 0)
798 			error = 0;
799 		else
800 			error = fo_read(fp, job->uiop, fp->f_cred, FOF_OFFSET,
801 			    td);
802 	} else {
803 		if (fp->f_type == DTYPE_VNODE)
804 			bwillwrite();
805 		error = fo_write(fp, job->uiop, fp->f_cred, FOF_OFFSET, td);
806 	}
807 	msgrcv_end = td->td_ru.ru_msgrcv;
808 	msgsnd_end = td->td_ru.ru_msgsnd;
809 	inblock_end = td->td_ru.ru_inblock;
810 	oublock_end = td->td_ru.ru_oublock;
811 
812 	job->msgrcv = msgrcv_end - msgrcv_st;
813 	job->msgsnd = msgsnd_end - msgsnd_st;
814 	job->inblock = inblock_end - inblock_st;
815 	job->outblock = oublock_end - oublock_st;
816 
817 	if (error != 0 && job->uiop->uio_resid != cnt) {
818 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
819 			error = 0;
820 		if (error == EPIPE && (opcode & LIO_WRITE)) {
821 			PROC_LOCK(job->userproc);
822 			kern_psignal(job->userproc, SIGPIPE);
823 			PROC_UNLOCK(job->userproc);
824 		}
825 	}
826 
827 	cnt -= job->uiop->uio_resid;
828 	td->td_ucred = td_savedcred;
829 	if (error)
830 		aio_complete(job, -1, error);
831 	else
832 		aio_complete(job, cnt, 0);
833 }
834 
835 static void
836 aio_process_sync(struct kaiocb *job)
837 {
838 	struct thread *td = curthread;
839 	struct ucred *td_savedcred = td->td_ucred;
840 	struct file *fp = job->fd_file;
841 	int error = 0;
842 
843 	KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC,
844 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
845 
846 	td->td_ucred = job->cred;
847 	if (fp->f_vnode != NULL) {
848 		error = aio_fsync_vnode(td, fp->f_vnode,
849 		    job->uaiocb.aio_lio_opcode);
850 	}
851 	td->td_ucred = td_savedcred;
852 	if (error)
853 		aio_complete(job, -1, error);
854 	else
855 		aio_complete(job, 0, 0);
856 }
857 
858 static void
859 aio_process_mlock(struct kaiocb *job)
860 {
861 	struct aiocb *cb = &job->uaiocb;
862 	int error;
863 
864 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
865 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
866 
867 	aio_switch_vmspace(job);
868 	error = kern_mlock(job->userproc, job->cred,
869 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
870 	aio_complete(job, error != 0 ? -1 : 0, error);
871 }
872 
873 static void
874 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
875 {
876 	struct aioliojob *lj;
877 	struct kaioinfo *ki;
878 	struct kaiocb *sjob, *sjobn;
879 	int lj_done;
880 	bool schedule_fsync;
881 
882 	ki = userp->p_aioinfo;
883 	AIO_LOCK_ASSERT(ki, MA_OWNED);
884 	lj = job->lio;
885 	lj_done = 0;
886 	if (lj) {
887 		lj->lioj_finished_count++;
888 		if (lj->lioj_count == lj->lioj_finished_count)
889 			lj_done = 1;
890 	}
891 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
892 	MPASS(job->jobflags & KAIOCB_FINISHED);
893 
894 	if (ki->kaio_flags & KAIO_RUNDOWN)
895 		goto notification_done;
896 
897 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
898 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
899 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
900 
901 	KNOTE_LOCKED(&job->klist, 1);
902 
903 	if (lj_done) {
904 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
905 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
906 			KNOTE_LOCKED(&lj->klist, 1);
907 		}
908 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
909 		    == LIOJ_SIGNAL &&
910 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
911 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
912 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
913 			    true);
914 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
915 		}
916 	}
917 
918 notification_done:
919 	if (job->jobflags & KAIOCB_CHECKSYNC) {
920 		schedule_fsync = false;
921 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
922 			if (job->fd_file != sjob->fd_file ||
923 			    job->seqno >= sjob->seqno)
924 				continue;
925 			if (--sjob->pending > 0)
926 				continue;
927 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
928 			if (!aio_clear_cancel_function_locked(sjob))
929 				continue;
930 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
931 			schedule_fsync = true;
932 		}
933 		if (schedule_fsync)
934 			taskqueue_enqueue(taskqueue_aiod_kick,
935 			    &ki->kaio_sync_task);
936 	}
937 	if (ki->kaio_flags & KAIO_WAKEUP) {
938 		ki->kaio_flags &= ~KAIO_WAKEUP;
939 		wakeup(&userp->p_aioinfo);
940 	}
941 }
942 
943 static void
944 aio_schedule_fsync(void *context, int pending)
945 {
946 	struct kaioinfo *ki;
947 	struct kaiocb *job;
948 
949 	ki = context;
950 	AIO_LOCK(ki);
951 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
952 		job = TAILQ_FIRST(&ki->kaio_syncready);
953 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
954 		AIO_UNLOCK(ki);
955 		aio_schedule(job, aio_process_sync);
956 		AIO_LOCK(ki);
957 	}
958 	AIO_UNLOCK(ki);
959 }
960 
961 bool
962 aio_cancel_cleared(struct kaiocb *job)
963 {
964 
965 	/*
966 	 * The caller should hold the same queue lock held when
967 	 * aio_clear_cancel_function() was called and set this flag
968 	 * ensuring this check sees an up-to-date value.  However,
969 	 * there is no way to assert that.
970 	 */
971 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
972 }
973 
974 static bool
975 aio_clear_cancel_function_locked(struct kaiocb *job)
976 {
977 
978 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
979 	MPASS(job->cancel_fn != NULL);
980 	if (job->jobflags & KAIOCB_CANCELLING) {
981 		job->jobflags |= KAIOCB_CLEARED;
982 		return (false);
983 	}
984 	job->cancel_fn = NULL;
985 	return (true);
986 }
987 
988 bool
989 aio_clear_cancel_function(struct kaiocb *job)
990 {
991 	struct kaioinfo *ki;
992 	bool ret;
993 
994 	ki = job->userproc->p_aioinfo;
995 	AIO_LOCK(ki);
996 	ret = aio_clear_cancel_function_locked(job);
997 	AIO_UNLOCK(ki);
998 	return (ret);
999 }
1000 
1001 static bool
1002 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1003 {
1004 
1005 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1006 	if (job->jobflags & KAIOCB_CANCELLED)
1007 		return (false);
1008 	job->cancel_fn = func;
1009 	return (true);
1010 }
1011 
1012 bool
1013 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1014 {
1015 	struct kaioinfo *ki;
1016 	bool ret;
1017 
1018 	ki = job->userproc->p_aioinfo;
1019 	AIO_LOCK(ki);
1020 	ret = aio_set_cancel_function_locked(job, func);
1021 	AIO_UNLOCK(ki);
1022 	return (ret);
1023 }
1024 
1025 void
1026 aio_complete(struct kaiocb *job, long status, int error)
1027 {
1028 	struct kaioinfo *ki;
1029 	struct proc *userp;
1030 
1031 	job->uaiocb._aiocb_private.error = error;
1032 	job->uaiocb._aiocb_private.status = status;
1033 
1034 	userp = job->userproc;
1035 	ki = userp->p_aioinfo;
1036 
1037 	AIO_LOCK(ki);
1038 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1039 	    ("duplicate aio_complete"));
1040 	job->jobflags |= KAIOCB_FINISHED;
1041 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1042 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1043 		aio_bio_done_notify(userp, job);
1044 	}
1045 	AIO_UNLOCK(ki);
1046 }
1047 
1048 void
1049 aio_cancel(struct kaiocb *job)
1050 {
1051 
1052 	aio_complete(job, -1, ECANCELED);
1053 }
1054 
1055 void
1056 aio_switch_vmspace(struct kaiocb *job)
1057 {
1058 
1059 	vmspace_switch_aio(job->userproc->p_vmspace);
1060 }
1061 
1062 /*
1063  * The AIO daemon, most of the actual work is done in aio_process_*,
1064  * but the setup (and address space mgmt) is done in this routine.
1065  */
1066 static void
1067 aio_daemon(void *_id)
1068 {
1069 	struct kaiocb *job;
1070 	struct aioproc *aiop;
1071 	struct kaioinfo *ki;
1072 	struct proc *p;
1073 	struct vmspace *myvm;
1074 	struct thread *td = curthread;
1075 	int id = (intptr_t)_id;
1076 
1077 	/*
1078 	 * Grab an extra reference on the daemon's vmspace so that it
1079 	 * doesn't get freed by jobs that switch to a different
1080 	 * vmspace.
1081 	 */
1082 	p = td->td_proc;
1083 	myvm = vmspace_acquire_ref(p);
1084 
1085 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1086 
1087 	/*
1088 	 * Allocate and ready the aio control info.  There is one aiop structure
1089 	 * per daemon.
1090 	 */
1091 	aiop = malloc(sizeof(*aiop), M_AIO, M_WAITOK);
1092 	aiop->aioproc = p;
1093 	aiop->aioprocflags = 0;
1094 
1095 	/*
1096 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1097 	 * and creating too many daemons.)
1098 	 */
1099 	sema_post(&aio_newproc_sem);
1100 
1101 	mtx_lock(&aio_job_mtx);
1102 	for (;;) {
1103 		/*
1104 		 * Take daemon off of free queue
1105 		 */
1106 		if (aiop->aioprocflags & AIOP_FREE) {
1107 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1108 			aiop->aioprocflags &= ~AIOP_FREE;
1109 		}
1110 
1111 		/*
1112 		 * Check for jobs.
1113 		 */
1114 		while ((job = aio_selectjob(aiop)) != NULL) {
1115 			mtx_unlock(&aio_job_mtx);
1116 
1117 			ki = job->userproc->p_aioinfo;
1118 			job->handle_fn(job);
1119 
1120 			mtx_lock(&aio_job_mtx);
1121 			/* Decrement the active job count. */
1122 			ki->kaio_active_count--;
1123 		}
1124 
1125 		/*
1126 		 * Disconnect from user address space.
1127 		 */
1128 		if (p->p_vmspace != myvm) {
1129 			mtx_unlock(&aio_job_mtx);
1130 			vmspace_switch_aio(myvm);
1131 			mtx_lock(&aio_job_mtx);
1132 			/*
1133 			 * We have to restart to avoid race, we only sleep if
1134 			 * no job can be selected.
1135 			 */
1136 			continue;
1137 		}
1138 
1139 		mtx_assert(&aio_job_mtx, MA_OWNED);
1140 
1141 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1142 		aiop->aioprocflags |= AIOP_FREE;
1143 
1144 		/*
1145 		 * If daemon is inactive for a long time, allow it to exit,
1146 		 * thereby freeing resources.
1147 		 */
1148 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1149 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1150 		    (aiop->aioprocflags & AIOP_FREE) &&
1151 		    num_aio_procs > target_aio_procs)
1152 			break;
1153 	}
1154 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1155 	num_aio_procs--;
1156 	mtx_unlock(&aio_job_mtx);
1157 	free(aiop, M_AIO);
1158 	free_unr(aiod_unr, id);
1159 	vmspace_free(myvm);
1160 
1161 	KASSERT(p->p_vmspace == myvm,
1162 	    ("AIOD: bad vmspace for exiting daemon"));
1163 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1164 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1165 	    refcount_load(&myvm->vm_refcnt)));
1166 	kproc_exit(0);
1167 }
1168 
1169 /*
1170  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1171  * AIO daemon modifies its environment itself.
1172  */
1173 static int
1174 aio_newproc(int *start)
1175 {
1176 	int error;
1177 	struct proc *p;
1178 	int id;
1179 
1180 	id = alloc_unr(aiod_unr);
1181 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1182 		RFNOWAIT, 0, "aiod%d", id);
1183 	if (error == 0) {
1184 		/*
1185 		 * Wait until daemon is started.
1186 		 */
1187 		sema_wait(&aio_newproc_sem);
1188 		mtx_lock(&aio_job_mtx);
1189 		num_aio_procs++;
1190 		if (start != NULL)
1191 			(*start)--;
1192 		mtx_unlock(&aio_job_mtx);
1193 	} else {
1194 		free_unr(aiod_unr, id);
1195 	}
1196 	return (error);
1197 }
1198 
1199 /*
1200  * Try the high-performance, low-overhead bio method for eligible
1201  * VCHR devices.  This method doesn't use an aio helper thread, and
1202  * thus has very low overhead.
1203  *
1204  * Assumes that the caller, aio_aqueue(), has incremented the file
1205  * structure's reference count, preventing its deallocation for the
1206  * duration of this call.
1207  */
1208 static int
1209 aio_qbio(struct proc *p, struct kaiocb *job)
1210 {
1211 	struct aiocb *cb;
1212 	struct file *fp;
1213 	struct buf *pbuf;
1214 	struct vnode *vp;
1215 	struct cdevsw *csw;
1216 	struct cdev *dev;
1217 	struct kaioinfo *ki;
1218 	struct bio **bios = NULL;
1219 	off_t offset;
1220 	int bio_cmd, error, i, iovcnt, opcode, poff, ref;
1221 	vm_prot_t prot;
1222 	bool use_unmapped;
1223 
1224 	cb = &job->uaiocb;
1225 	fp = job->fd_file;
1226 	opcode = cb->aio_lio_opcode;
1227 
1228 	if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV ||
1229 	    opcode == LIO_READ || opcode == LIO_READV))
1230 		return (-1);
1231 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1232 		return (-1);
1233 
1234 	vp = fp->f_vnode;
1235 	if (vp->v_type != VCHR)
1236 		return (-1);
1237 	if (vp->v_bufobj.bo_bsize == 0)
1238 		return (-1);
1239 
1240 	bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ;
1241 	iovcnt = job->uiop->uio_iovcnt;
1242 	if (iovcnt > max_buf_aio)
1243 		return (-1);
1244 	for (i = 0; i < iovcnt; i++) {
1245 		if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0)
1246 			return (-1);
1247 		if (job->uiop->uio_iov[i].iov_len > maxphys) {
1248 			error = -1;
1249 			return (-1);
1250 		}
1251 	}
1252 	offset = cb->aio_offset;
1253 
1254 	ref = 0;
1255 	csw = devvn_refthread(vp, &dev, &ref);
1256 	if (csw == NULL)
1257 		return (ENXIO);
1258 
1259 	if ((csw->d_flags & D_DISK) == 0) {
1260 		error = -1;
1261 		goto unref;
1262 	}
1263 	if (job->uiop->uio_resid > dev->si_iosize_max) {
1264 		error = -1;
1265 		goto unref;
1266 	}
1267 
1268 	ki = p->p_aioinfo;
1269 	job->error = 0;
1270 
1271 	use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed;
1272 	if (!use_unmapped) {
1273 		AIO_LOCK(ki);
1274 		if (ki->kaio_buffer_count + iovcnt > max_buf_aio) {
1275 			AIO_UNLOCK(ki);
1276 			error = EAGAIN;
1277 			goto unref;
1278 		}
1279 		ki->kaio_buffer_count += iovcnt;
1280 		AIO_UNLOCK(ki);
1281 	}
1282 
1283 	bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK);
1284 	refcount_init(&job->nbio, iovcnt);
1285 	for (i = 0; i < iovcnt; i++) {
1286 		struct vm_page** pages;
1287 		struct bio *bp;
1288 		void *buf;
1289 		size_t nbytes;
1290 		int npages;
1291 
1292 		buf = job->uiop->uio_iov[i].iov_base;
1293 		nbytes = job->uiop->uio_iov[i].iov_len;
1294 
1295 		bios[i] = g_alloc_bio();
1296 		bp = bios[i];
1297 
1298 		poff = (vm_offset_t)buf & PAGE_MASK;
1299 		if (use_unmapped) {
1300 			pbuf = NULL;
1301 			pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1302 			    nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1303 		} else {
1304 			pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1305 			BUF_KERNPROC(pbuf);
1306 			pages = pbuf->b_pages;
1307 		}
1308 
1309 		bp->bio_length = nbytes;
1310 		bp->bio_bcount = nbytes;
1311 		bp->bio_done = aio_biowakeup;
1312 		bp->bio_offset = offset;
1313 		bp->bio_cmd = bio_cmd;
1314 		bp->bio_dev = dev;
1315 		bp->bio_caller1 = job;
1316 		bp->bio_caller2 = pbuf;
1317 
1318 		prot = VM_PROT_READ;
1319 		if (opcode == LIO_READ || opcode == LIO_READV)
1320 			prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1321 		npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1322 		    (vm_offset_t)buf, bp->bio_length, prot, pages,
1323 		    atop(maxphys) + 1);
1324 		if (npages < 0) {
1325 			if (pbuf != NULL)
1326 				uma_zfree(pbuf_zone, pbuf);
1327 			else
1328 				free(pages, M_TEMP);
1329 			error = EFAULT;
1330 			g_destroy_bio(bp);
1331 			i--;
1332 			goto destroy_bios;
1333 		}
1334 		if (pbuf != NULL) {
1335 			pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages);
1336 			bp->bio_data = pbuf->b_data + poff;
1337 			pbuf->b_npages = npages;
1338 			atomic_add_int(&num_buf_aio, 1);
1339 		} else {
1340 			bp->bio_ma = pages;
1341 			bp->bio_ma_n = npages;
1342 			bp->bio_ma_offset = poff;
1343 			bp->bio_data = unmapped_buf;
1344 			bp->bio_flags |= BIO_UNMAPPED;
1345 			atomic_add_int(&num_unmapped_aio, 1);
1346 		}
1347 
1348 		offset += nbytes;
1349 	}
1350 
1351 	/* Perform transfer. */
1352 	for (i = 0; i < iovcnt; i++)
1353 		csw->d_strategy(bios[i]);
1354 	free(bios, M_TEMP);
1355 
1356 	dev_relthread(dev, ref);
1357 	return (0);
1358 
1359 destroy_bios:
1360 	for (; i >= 0; i--)
1361 		aio_biocleanup(bios[i]);
1362 	free(bios, M_TEMP);
1363 unref:
1364 	dev_relthread(dev, ref);
1365 	return (error);
1366 }
1367 
1368 #ifdef COMPAT_FREEBSD6
1369 static int
1370 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1371 {
1372 
1373 	/*
1374 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1375 	 * supported by AIO with the old sigevent structure.
1376 	 */
1377 	nsig->sigev_notify = osig->sigev_notify;
1378 	switch (nsig->sigev_notify) {
1379 	case SIGEV_NONE:
1380 		break;
1381 	case SIGEV_SIGNAL:
1382 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1383 		break;
1384 	case SIGEV_KEVENT:
1385 		nsig->sigev_notify_kqueue =
1386 		    osig->__sigev_u.__sigev_notify_kqueue;
1387 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1388 		break;
1389 	default:
1390 		return (EINVAL);
1391 	}
1392 	return (0);
1393 }
1394 
1395 static int
1396 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
1397     int type __unused)
1398 {
1399 	struct oaiocb *ojob;
1400 	struct aiocb *kcb = &kjob->uaiocb;
1401 	int error;
1402 
1403 	bzero(kcb, sizeof(struct aiocb));
1404 	error = copyin(ujob, kcb, sizeof(struct oaiocb));
1405 	if (error)
1406 		return (error);
1407 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
1408 	ojob = (struct oaiocb *)kcb;
1409 	return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent));
1410 }
1411 #endif
1412 
1413 static int
1414 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
1415 {
1416 	struct aiocb *kcb = &kjob->uaiocb;
1417 	int error;
1418 
1419 	error = copyin(ujob, kcb, sizeof(struct aiocb));
1420 	if (error)
1421 		return (error);
1422 	if (type == LIO_NOP)
1423 		type = kcb->aio_lio_opcode;
1424 	if (type & LIO_VECTORED) {
1425 		/* malloc a uio and copy in the iovec */
1426 		error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov),
1427 		    kcb->aio_iovcnt, &kjob->uiop);
1428 	}
1429 
1430 	return (error);
1431 }
1432 
1433 static long
1434 aiocb_fetch_status(struct aiocb *ujob)
1435 {
1436 
1437 	return (fuword(&ujob->_aiocb_private.status));
1438 }
1439 
1440 static long
1441 aiocb_fetch_error(struct aiocb *ujob)
1442 {
1443 
1444 	return (fuword(&ujob->_aiocb_private.error));
1445 }
1446 
1447 static int
1448 aiocb_store_status(struct aiocb *ujob, long status)
1449 {
1450 
1451 	return (suword(&ujob->_aiocb_private.status, status));
1452 }
1453 
1454 static int
1455 aiocb_store_error(struct aiocb *ujob, long error)
1456 {
1457 
1458 	return (suword(&ujob->_aiocb_private.error, error));
1459 }
1460 
1461 static int
1462 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1463 {
1464 
1465 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1466 }
1467 
1468 static int
1469 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1470 {
1471 
1472 	return (suword(ujobp, (long)ujob));
1473 }
1474 
1475 static struct aiocb_ops aiocb_ops = {
1476 	.aio_copyin = aiocb_copyin,
1477 	.fetch_status = aiocb_fetch_status,
1478 	.fetch_error = aiocb_fetch_error,
1479 	.store_status = aiocb_store_status,
1480 	.store_error = aiocb_store_error,
1481 	.store_kernelinfo = aiocb_store_kernelinfo,
1482 	.store_aiocb = aiocb_store_aiocb,
1483 };
1484 
1485 #ifdef COMPAT_FREEBSD6
1486 static struct aiocb_ops aiocb_ops_osigevent = {
1487 	.aio_copyin = aiocb_copyin_old_sigevent,
1488 	.fetch_status = aiocb_fetch_status,
1489 	.fetch_error = aiocb_fetch_error,
1490 	.store_status = aiocb_store_status,
1491 	.store_error = aiocb_store_error,
1492 	.store_kernelinfo = aiocb_store_kernelinfo,
1493 	.store_aiocb = aiocb_store_aiocb,
1494 };
1495 #endif
1496 
1497 /*
1498  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1499  * technique is done in this code.
1500  */
1501 int
1502 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1503     int type, struct aiocb_ops *ops)
1504 {
1505 	struct proc *p = td->td_proc;
1506 	struct file *fp = NULL;
1507 	struct kaiocb *job;
1508 	struct kaioinfo *ki;
1509 	struct kevent kev;
1510 	int opcode;
1511 	int error;
1512 	int fd, kqfd;
1513 	int jid;
1514 	u_short evflags;
1515 
1516 	if (p->p_aioinfo == NULL)
1517 		aio_init_aioinfo(p);
1518 
1519 	ki = p->p_aioinfo;
1520 
1521 	ops->store_status(ujob, -1);
1522 	ops->store_error(ujob, 0);
1523 	ops->store_kernelinfo(ujob, -1);
1524 
1525 	if (num_queue_count >= max_queue_count ||
1526 	    ki->kaio_count >= max_aio_queue_per_proc) {
1527 		error = EAGAIN;
1528 		goto err1;
1529 	}
1530 
1531 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1532 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1533 
1534 	error = ops->aio_copyin(ujob, job, type);
1535 	if (error)
1536 		goto err2;
1537 
1538 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1539 		error = EINVAL;
1540 		goto err2;
1541 	}
1542 
1543 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1544 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1545 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1546 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1547 		error = EINVAL;
1548 		goto err2;
1549 	}
1550 
1551 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1552 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1553 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1554 		error = EINVAL;
1555 		goto err2;
1556 	}
1557 
1558 	/* Get the opcode. */
1559 	if (type == LIO_NOP) {
1560 		switch (job->uaiocb.aio_lio_opcode) {
1561 		case LIO_WRITE:
1562 		case LIO_WRITEV:
1563 		case LIO_NOP:
1564 		case LIO_READ:
1565 		case LIO_READV:
1566 			opcode = job->uaiocb.aio_lio_opcode;
1567 			break;
1568 		default:
1569 			error = EINVAL;
1570 			goto err2;
1571 		}
1572 	} else
1573 		opcode = job->uaiocb.aio_lio_opcode = type;
1574 
1575 	ksiginfo_init(&job->ksi);
1576 
1577 	/* Save userspace address of the job info. */
1578 	job->ujob = ujob;
1579 
1580 	/*
1581 	 * Validate the opcode and fetch the file object for the specified
1582 	 * file descriptor.
1583 	 *
1584 	 * XXXRW: Moved the opcode validation up here so that we don't
1585 	 * retrieve a file descriptor without knowing what the capabiltity
1586 	 * should be.
1587 	 */
1588 	fd = job->uaiocb.aio_fildes;
1589 	switch (opcode) {
1590 	case LIO_WRITE:
1591 	case LIO_WRITEV:
1592 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1593 		break;
1594 	case LIO_READ:
1595 	case LIO_READV:
1596 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1597 		break;
1598 	case LIO_SYNC:
1599 	case LIO_DSYNC:
1600 		error = fget(td, fd, &cap_fsync_rights, &fp);
1601 		break;
1602 	case LIO_MLOCK:
1603 		break;
1604 	case LIO_NOP:
1605 		error = fget(td, fd, &cap_no_rights, &fp);
1606 		break;
1607 	default:
1608 		error = EINVAL;
1609 	}
1610 	if (error)
1611 		goto err3;
1612 
1613 	if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) {
1614 		error = EINVAL;
1615 		goto err3;
1616 	}
1617 
1618 	if ((opcode == LIO_READ || opcode == LIO_READV ||
1619 	    opcode == LIO_WRITE || opcode == LIO_WRITEV) &&
1620 	    job->uaiocb.aio_offset < 0 &&
1621 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1622 		error = EINVAL;
1623 		goto err3;
1624 	}
1625 
1626 	if (fp != NULL && fp->f_ops == &path_fileops) {
1627 		error = EBADF;
1628 		goto err3;
1629 	}
1630 
1631 	job->fd_file = fp;
1632 
1633 	mtx_lock(&aio_job_mtx);
1634 	jid = jobrefid++;
1635 	job->seqno = jobseqno++;
1636 	mtx_unlock(&aio_job_mtx);
1637 	error = ops->store_kernelinfo(ujob, jid);
1638 	if (error) {
1639 		error = EINVAL;
1640 		goto err3;
1641 	}
1642 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1643 
1644 	if (opcode == LIO_NOP) {
1645 		fdrop(fp, td);
1646 		MPASS(job->uiop == &job->uio || job->uiop == NULL);
1647 		uma_zfree(aiocb_zone, job);
1648 		return (0);
1649 	}
1650 
1651 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1652 		goto no_kqueue;
1653 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1654 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1655 		error = EINVAL;
1656 		goto err3;
1657 	}
1658 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1659 	memset(&kev, 0, sizeof(kev));
1660 	kev.ident = (uintptr_t)job->ujob;
1661 	kev.filter = EVFILT_AIO;
1662 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1663 	kev.data = (intptr_t)job;
1664 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1665 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1666 	if (error)
1667 		goto err3;
1668 
1669 no_kqueue:
1670 
1671 	ops->store_error(ujob, EINPROGRESS);
1672 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1673 	job->userproc = p;
1674 	job->cred = crhold(td->td_ucred);
1675 	job->jobflags = KAIOCB_QUEUEING;
1676 	job->lio = lj;
1677 
1678 	if (opcode & LIO_VECTORED) {
1679 		/* Use the uio copied in by aio_copyin */
1680 		MPASS(job->uiop != &job->uio && job->uiop != NULL);
1681 	} else {
1682 		/* Setup the inline uio */
1683 		job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf;
1684 		job->iov[0].iov_len = job->uaiocb.aio_nbytes;
1685 		job->uio.uio_iov = job->iov;
1686 		job->uio.uio_iovcnt = 1;
1687 		job->uio.uio_resid = job->uaiocb.aio_nbytes;
1688 		job->uio.uio_segflg = UIO_USERSPACE;
1689 		job->uiop = &job->uio;
1690 	}
1691 	switch (opcode & (LIO_READ | LIO_WRITE)) {
1692 	case LIO_READ:
1693 		job->uiop->uio_rw = UIO_READ;
1694 		break;
1695 	case LIO_WRITE:
1696 		job->uiop->uio_rw = UIO_WRITE;
1697 		break;
1698 	}
1699 	job->uiop->uio_offset = job->uaiocb.aio_offset;
1700 	job->uiop->uio_td = td;
1701 
1702 	if (opcode == LIO_MLOCK) {
1703 		aio_schedule(job, aio_process_mlock);
1704 		error = 0;
1705 	} else if (fp->f_ops->fo_aio_queue == NULL)
1706 		error = aio_queue_file(fp, job);
1707 	else
1708 		error = fo_aio_queue(fp, job);
1709 	if (error)
1710 		goto err4;
1711 
1712 	AIO_LOCK(ki);
1713 	job->jobflags &= ~KAIOCB_QUEUEING;
1714 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1715 	ki->kaio_count++;
1716 	if (lj)
1717 		lj->lioj_count++;
1718 	atomic_add_int(&num_queue_count, 1);
1719 	if (job->jobflags & KAIOCB_FINISHED) {
1720 		/*
1721 		 * The queue callback completed the request synchronously.
1722 		 * The bulk of the completion is deferred in that case
1723 		 * until this point.
1724 		 */
1725 		aio_bio_done_notify(p, job);
1726 	} else
1727 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1728 	AIO_UNLOCK(ki);
1729 	return (0);
1730 
1731 err4:
1732 	crfree(job->cred);
1733 err3:
1734 	if (fp)
1735 		fdrop(fp, td);
1736 	knlist_delete(&job->klist, curthread, 0);
1737 err2:
1738 	if (job->uiop != &job->uio)
1739 		free(job->uiop, M_IOV);
1740 	uma_zfree(aiocb_zone, job);
1741 err1:
1742 	ops->store_error(ujob, error);
1743 	return (error);
1744 }
1745 
1746 static void
1747 aio_cancel_daemon_job(struct kaiocb *job)
1748 {
1749 
1750 	mtx_lock(&aio_job_mtx);
1751 	if (!aio_cancel_cleared(job))
1752 		TAILQ_REMOVE(&aio_jobs, job, list);
1753 	mtx_unlock(&aio_job_mtx);
1754 	aio_cancel(job);
1755 }
1756 
1757 void
1758 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1759 {
1760 
1761 	mtx_lock(&aio_job_mtx);
1762 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1763 		mtx_unlock(&aio_job_mtx);
1764 		aio_cancel(job);
1765 		return;
1766 	}
1767 	job->handle_fn = func;
1768 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1769 	aio_kick_nowait(job->userproc);
1770 	mtx_unlock(&aio_job_mtx);
1771 }
1772 
1773 static void
1774 aio_cancel_sync(struct kaiocb *job)
1775 {
1776 	struct kaioinfo *ki;
1777 
1778 	ki = job->userproc->p_aioinfo;
1779 	AIO_LOCK(ki);
1780 	if (!aio_cancel_cleared(job))
1781 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1782 	AIO_UNLOCK(ki);
1783 	aio_cancel(job);
1784 }
1785 
1786 int
1787 aio_queue_file(struct file *fp, struct kaiocb *job)
1788 {
1789 	struct kaioinfo *ki;
1790 	struct kaiocb *job2;
1791 	struct vnode *vp;
1792 	struct mount *mp;
1793 	int error;
1794 	bool safe;
1795 
1796 	ki = job->userproc->p_aioinfo;
1797 	error = aio_qbio(job->userproc, job);
1798 	if (error >= 0)
1799 		return (error);
1800 	safe = false;
1801 	if (fp->f_type == DTYPE_VNODE) {
1802 		vp = fp->f_vnode;
1803 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1804 			mp = fp->f_vnode->v_mount;
1805 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1806 				safe = true;
1807 		}
1808 	}
1809 	if (!(safe || enable_aio_unsafe)) {
1810 		counted_warning(&unsafe_warningcnt,
1811 		    "is attempting to use unsafe AIO requests");
1812 		return (EOPNOTSUPP);
1813 	}
1814 
1815 	if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) {
1816 		aio_schedule(job, aio_process_rw);
1817 		error = 0;
1818 	} else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) {
1819 		AIO_LOCK(ki);
1820 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1821 			if (job2->fd_file == job->fd_file &&
1822 			    ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) &&
1823 			    job2->seqno < job->seqno) {
1824 				job2->jobflags |= KAIOCB_CHECKSYNC;
1825 				job->pending++;
1826 			}
1827 		}
1828 		if (job->pending != 0) {
1829 			if (!aio_set_cancel_function_locked(job,
1830 				aio_cancel_sync)) {
1831 				AIO_UNLOCK(ki);
1832 				aio_cancel(job);
1833 				return (0);
1834 			}
1835 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1836 			AIO_UNLOCK(ki);
1837 			return (0);
1838 		}
1839 		AIO_UNLOCK(ki);
1840 		aio_schedule(job, aio_process_sync);
1841 		error = 0;
1842 	} else {
1843 		error = EINVAL;
1844 	}
1845 	return (error);
1846 }
1847 
1848 static void
1849 aio_kick_nowait(struct proc *userp)
1850 {
1851 	struct kaioinfo *ki = userp->p_aioinfo;
1852 	struct aioproc *aiop;
1853 
1854 	mtx_assert(&aio_job_mtx, MA_OWNED);
1855 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1856 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1857 		aiop->aioprocflags &= ~AIOP_FREE;
1858 		wakeup(aiop->aioproc);
1859 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1860 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1861 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1862 	}
1863 }
1864 
1865 static int
1866 aio_kick(struct proc *userp)
1867 {
1868 	struct kaioinfo *ki = userp->p_aioinfo;
1869 	struct aioproc *aiop;
1870 	int error, ret = 0;
1871 
1872 	mtx_assert(&aio_job_mtx, MA_OWNED);
1873 retryproc:
1874 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1875 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1876 		aiop->aioprocflags &= ~AIOP_FREE;
1877 		wakeup(aiop->aioproc);
1878 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1879 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1880 		num_aio_resv_start++;
1881 		mtx_unlock(&aio_job_mtx);
1882 		error = aio_newproc(&num_aio_resv_start);
1883 		mtx_lock(&aio_job_mtx);
1884 		if (error) {
1885 			num_aio_resv_start--;
1886 			goto retryproc;
1887 		}
1888 	} else {
1889 		ret = -1;
1890 	}
1891 	return (ret);
1892 }
1893 
1894 static void
1895 aio_kick_helper(void *context, int pending)
1896 {
1897 	struct proc *userp = context;
1898 
1899 	mtx_lock(&aio_job_mtx);
1900 	while (--pending >= 0) {
1901 		if (aio_kick(userp))
1902 			break;
1903 	}
1904 	mtx_unlock(&aio_job_mtx);
1905 }
1906 
1907 /*
1908  * Support the aio_return system call, as a side-effect, kernel resources are
1909  * released.
1910  */
1911 static int
1912 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1913 {
1914 	struct proc *p = td->td_proc;
1915 	struct kaiocb *job;
1916 	struct kaioinfo *ki;
1917 	long status, error;
1918 
1919 	ki = p->p_aioinfo;
1920 	if (ki == NULL)
1921 		return (EINVAL);
1922 	AIO_LOCK(ki);
1923 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1924 		if (job->ujob == ujob)
1925 			break;
1926 	}
1927 	if (job != NULL) {
1928 		MPASS(job->jobflags & KAIOCB_FINISHED);
1929 		status = job->uaiocb._aiocb_private.status;
1930 		error = job->uaiocb._aiocb_private.error;
1931 		td->td_retval[0] = status;
1932 		td->td_ru.ru_oublock += job->outblock;
1933 		td->td_ru.ru_inblock += job->inblock;
1934 		td->td_ru.ru_msgsnd += job->msgsnd;
1935 		td->td_ru.ru_msgrcv += job->msgrcv;
1936 		aio_free_entry(job);
1937 		AIO_UNLOCK(ki);
1938 		ops->store_error(ujob, error);
1939 		ops->store_status(ujob, status);
1940 	} else {
1941 		error = EINVAL;
1942 		AIO_UNLOCK(ki);
1943 	}
1944 	return (error);
1945 }
1946 
1947 int
1948 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1949 {
1950 
1951 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1952 }
1953 
1954 /*
1955  * Allow a process to wakeup when any of the I/O requests are completed.
1956  */
1957 static int
1958 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1959     struct timespec *ts)
1960 {
1961 	struct proc *p = td->td_proc;
1962 	struct timeval atv;
1963 	struct kaioinfo *ki;
1964 	struct kaiocb *firstjob, *job;
1965 	int error, i, timo;
1966 
1967 	timo = 0;
1968 	if (ts) {
1969 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1970 			return (EINVAL);
1971 
1972 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1973 		if (itimerfix(&atv))
1974 			return (EINVAL);
1975 		timo = tvtohz(&atv);
1976 	}
1977 
1978 	ki = p->p_aioinfo;
1979 	if (ki == NULL)
1980 		return (EAGAIN);
1981 
1982 	if (njoblist == 0)
1983 		return (0);
1984 
1985 	AIO_LOCK(ki);
1986 	for (;;) {
1987 		firstjob = NULL;
1988 		error = 0;
1989 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1990 			for (i = 0; i < njoblist; i++) {
1991 				if (job->ujob == ujoblist[i]) {
1992 					if (firstjob == NULL)
1993 						firstjob = job;
1994 					if (job->jobflags & KAIOCB_FINISHED)
1995 						goto RETURN;
1996 				}
1997 			}
1998 		}
1999 		/* All tasks were finished. */
2000 		if (firstjob == NULL)
2001 			break;
2002 
2003 		ki->kaio_flags |= KAIO_WAKEUP;
2004 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2005 		    "aiospn", timo);
2006 		if (error == ERESTART)
2007 			error = EINTR;
2008 		if (error)
2009 			break;
2010 	}
2011 RETURN:
2012 	AIO_UNLOCK(ki);
2013 	return (error);
2014 }
2015 
2016 int
2017 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2018 {
2019 	struct timespec ts, *tsp;
2020 	struct aiocb **ujoblist;
2021 	int error;
2022 
2023 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2024 		return (EINVAL);
2025 
2026 	if (uap->timeout) {
2027 		/* Get timespec struct. */
2028 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2029 			return (error);
2030 		tsp = &ts;
2031 	} else
2032 		tsp = NULL;
2033 
2034 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2035 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2036 	if (error == 0)
2037 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2038 	free(ujoblist, M_AIO);
2039 	return (error);
2040 }
2041 
2042 /*
2043  * aio_cancel cancels any non-bio aio operations not currently in progress.
2044  */
2045 int
2046 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2047 {
2048 	struct proc *p = td->td_proc;
2049 	struct kaioinfo *ki;
2050 	struct kaiocb *job, *jobn;
2051 	struct file *fp;
2052 	int error;
2053 	int cancelled = 0;
2054 	int notcancelled = 0;
2055 	struct vnode *vp;
2056 
2057 	/* Lookup file object. */
2058 	error = fget(td, uap->fd, &cap_no_rights, &fp);
2059 	if (error)
2060 		return (error);
2061 
2062 	ki = p->p_aioinfo;
2063 	if (ki == NULL)
2064 		goto done;
2065 
2066 	if (fp->f_type == DTYPE_VNODE) {
2067 		vp = fp->f_vnode;
2068 		if (vn_isdisk(vp)) {
2069 			fdrop(fp, td);
2070 			td->td_retval[0] = AIO_NOTCANCELED;
2071 			return (0);
2072 		}
2073 	}
2074 
2075 	AIO_LOCK(ki);
2076 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2077 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2078 		    ((uap->aiocbp == NULL) ||
2079 		     (uap->aiocbp == job->ujob))) {
2080 			if (aio_cancel_job(p, ki, job)) {
2081 				cancelled++;
2082 			} else {
2083 				notcancelled++;
2084 			}
2085 			if (uap->aiocbp != NULL)
2086 				break;
2087 		}
2088 	}
2089 	AIO_UNLOCK(ki);
2090 
2091 done:
2092 	fdrop(fp, td);
2093 
2094 	if (uap->aiocbp != NULL) {
2095 		if (cancelled) {
2096 			td->td_retval[0] = AIO_CANCELED;
2097 			return (0);
2098 		}
2099 	}
2100 
2101 	if (notcancelled) {
2102 		td->td_retval[0] = AIO_NOTCANCELED;
2103 		return (0);
2104 	}
2105 
2106 	if (cancelled) {
2107 		td->td_retval[0] = AIO_CANCELED;
2108 		return (0);
2109 	}
2110 
2111 	td->td_retval[0] = AIO_ALLDONE;
2112 
2113 	return (0);
2114 }
2115 
2116 /*
2117  * aio_error is implemented in the kernel level for compatibility purposes
2118  * only.  For a user mode async implementation, it would be best to do it in
2119  * a userland subroutine.
2120  */
2121 static int
2122 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2123 {
2124 	struct proc *p = td->td_proc;
2125 	struct kaiocb *job;
2126 	struct kaioinfo *ki;
2127 	int status;
2128 
2129 	ki = p->p_aioinfo;
2130 	if (ki == NULL) {
2131 		td->td_retval[0] = EINVAL;
2132 		return (0);
2133 	}
2134 
2135 	AIO_LOCK(ki);
2136 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2137 		if (job->ujob == ujob) {
2138 			if (job->jobflags & KAIOCB_FINISHED)
2139 				td->td_retval[0] =
2140 					job->uaiocb._aiocb_private.error;
2141 			else
2142 				td->td_retval[0] = EINPROGRESS;
2143 			AIO_UNLOCK(ki);
2144 			return (0);
2145 		}
2146 	}
2147 	AIO_UNLOCK(ki);
2148 
2149 	/*
2150 	 * Hack for failure of aio_aqueue.
2151 	 */
2152 	status = ops->fetch_status(ujob);
2153 	if (status == -1) {
2154 		td->td_retval[0] = ops->fetch_error(ujob);
2155 		return (0);
2156 	}
2157 
2158 	td->td_retval[0] = EINVAL;
2159 	return (0);
2160 }
2161 
2162 int
2163 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2164 {
2165 
2166 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2167 }
2168 
2169 /* syscall - asynchronous read from a file (REALTIME) */
2170 #ifdef COMPAT_FREEBSD6
2171 int
2172 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2173 {
2174 
2175 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2176 	    &aiocb_ops_osigevent));
2177 }
2178 #endif
2179 
2180 int
2181 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2182 {
2183 
2184 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2185 }
2186 
2187 int
2188 sys_aio_readv(struct thread *td, struct aio_readv_args *uap)
2189 {
2190 
2191 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops));
2192 }
2193 
2194 /* syscall - asynchronous write to a file (REALTIME) */
2195 #ifdef COMPAT_FREEBSD6
2196 int
2197 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2198 {
2199 
2200 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2201 	    &aiocb_ops_osigevent));
2202 }
2203 #endif
2204 
2205 int
2206 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2207 {
2208 
2209 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2210 }
2211 
2212 int
2213 sys_aio_writev(struct thread *td, struct aio_writev_args *uap)
2214 {
2215 
2216 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops));
2217 }
2218 
2219 int
2220 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2221 {
2222 
2223 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2224 }
2225 
2226 static int
2227 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2228     struct aiocb **acb_list, int nent, struct sigevent *sig,
2229     struct aiocb_ops *ops)
2230 {
2231 	struct proc *p = td->td_proc;
2232 	struct aiocb *job;
2233 	struct kaioinfo *ki;
2234 	struct aioliojob *lj;
2235 	struct kevent kev;
2236 	int error;
2237 	int nagain, nerror;
2238 	int i;
2239 
2240 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2241 		return (EINVAL);
2242 
2243 	if (nent < 0 || nent > max_aio_queue_per_proc)
2244 		return (EINVAL);
2245 
2246 	if (p->p_aioinfo == NULL)
2247 		aio_init_aioinfo(p);
2248 
2249 	ki = p->p_aioinfo;
2250 
2251 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2252 	lj->lioj_flags = 0;
2253 	lj->lioj_count = 0;
2254 	lj->lioj_finished_count = 0;
2255 	lj->lioj_signal.sigev_notify = SIGEV_NONE;
2256 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2257 	ksiginfo_init(&lj->lioj_ksi);
2258 
2259 	/*
2260 	 * Setup signal.
2261 	 */
2262 	if (sig && (mode == LIO_NOWAIT)) {
2263 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2264 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2265 			/* Assume only new style KEVENT */
2266 			memset(&kev, 0, sizeof(kev));
2267 			kev.filter = EVFILT_LIO;
2268 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2269 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2270 			kev.data = (intptr_t)lj;
2271 			/* pass user defined sigval data */
2272 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2273 			error = kqfd_register(
2274 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2275 			    M_WAITOK);
2276 			if (error) {
2277 				uma_zfree(aiolio_zone, lj);
2278 				return (error);
2279 			}
2280 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2281 			;
2282 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2283 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2284 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2285 					uma_zfree(aiolio_zone, lj);
2286 					return EINVAL;
2287 				}
2288 				lj->lioj_flags |= LIOJ_SIGNAL;
2289 		} else {
2290 			uma_zfree(aiolio_zone, lj);
2291 			return EINVAL;
2292 		}
2293 	}
2294 
2295 	AIO_LOCK(ki);
2296 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2297 	/*
2298 	 * Add extra aiocb count to avoid the lio to be freed
2299 	 * by other threads doing aio_waitcomplete or aio_return,
2300 	 * and prevent event from being sent until we have queued
2301 	 * all tasks.
2302 	 */
2303 	lj->lioj_count = 1;
2304 	AIO_UNLOCK(ki);
2305 
2306 	/*
2307 	 * Get pointers to the list of I/O requests.
2308 	 */
2309 	nagain = 0;
2310 	nerror = 0;
2311 	for (i = 0; i < nent; i++) {
2312 		job = acb_list[i];
2313 		if (job != NULL) {
2314 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2315 			if (error == EAGAIN)
2316 				nagain++;
2317 			else if (error != 0)
2318 				nerror++;
2319 		}
2320 	}
2321 
2322 	error = 0;
2323 	AIO_LOCK(ki);
2324 	if (mode == LIO_WAIT) {
2325 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2326 			ki->kaio_flags |= KAIO_WAKEUP;
2327 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2328 			    PRIBIO | PCATCH, "aiospn", 0);
2329 			if (error == ERESTART)
2330 				error = EINTR;
2331 			if (error)
2332 				break;
2333 		}
2334 	} else {
2335 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2336 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2337 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2338 				KNOTE_LOCKED(&lj->klist, 1);
2339 			}
2340 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2341 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2342 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2343 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2344 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2345 				    lj->lioj_count != 1);
2346 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2347 			}
2348 		}
2349 	}
2350 	lj->lioj_count--;
2351 	if (lj->lioj_count == 0) {
2352 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2353 		knlist_delete(&lj->klist, curthread, 1);
2354 		PROC_LOCK(p);
2355 		sigqueue_take(&lj->lioj_ksi);
2356 		PROC_UNLOCK(p);
2357 		AIO_UNLOCK(ki);
2358 		uma_zfree(aiolio_zone, lj);
2359 	} else
2360 		AIO_UNLOCK(ki);
2361 
2362 	if (nerror)
2363 		return (EIO);
2364 	else if (nagain)
2365 		return (EAGAIN);
2366 	else
2367 		return (error);
2368 }
2369 
2370 /* syscall - list directed I/O (REALTIME) */
2371 #ifdef COMPAT_FREEBSD6
2372 int
2373 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2374 {
2375 	struct aiocb **acb_list;
2376 	struct sigevent *sigp, sig;
2377 	struct osigevent osig;
2378 	int error, nent;
2379 
2380 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2381 		return (EINVAL);
2382 
2383 	nent = uap->nent;
2384 	if (nent < 0 || nent > max_aio_queue_per_proc)
2385 		return (EINVAL);
2386 
2387 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2388 		error = copyin(uap->sig, &osig, sizeof(osig));
2389 		if (error)
2390 			return (error);
2391 		error = convert_old_sigevent(&osig, &sig);
2392 		if (error)
2393 			return (error);
2394 		sigp = &sig;
2395 	} else
2396 		sigp = NULL;
2397 
2398 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2399 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2400 	if (error == 0)
2401 		error = kern_lio_listio(td, uap->mode,
2402 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2403 		    &aiocb_ops_osigevent);
2404 	free(acb_list, M_LIO);
2405 	return (error);
2406 }
2407 #endif
2408 
2409 /* syscall - list directed I/O (REALTIME) */
2410 int
2411 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2412 {
2413 	struct aiocb **acb_list;
2414 	struct sigevent *sigp, sig;
2415 	int error, nent;
2416 
2417 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2418 		return (EINVAL);
2419 
2420 	nent = uap->nent;
2421 	if (nent < 0 || nent > max_aio_queue_per_proc)
2422 		return (EINVAL);
2423 
2424 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2425 		error = copyin(uap->sig, &sig, sizeof(sig));
2426 		if (error)
2427 			return (error);
2428 		sigp = &sig;
2429 	} else
2430 		sigp = NULL;
2431 
2432 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2433 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2434 	if (error == 0)
2435 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2436 		    nent, sigp, &aiocb_ops);
2437 	free(acb_list, M_LIO);
2438 	return (error);
2439 }
2440 
2441 static void
2442 aio_biocleanup(struct bio *bp)
2443 {
2444 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2445 	struct kaioinfo *ki;
2446 	struct buf *pbuf = (struct buf *)bp->bio_caller2;
2447 
2448 	/* Release mapping into kernel space. */
2449 	if (pbuf != NULL) {
2450 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2451 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2452 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2453 		uma_zfree(pbuf_zone, pbuf);
2454 		atomic_subtract_int(&num_buf_aio, 1);
2455 		ki = job->userproc->p_aioinfo;
2456 		AIO_LOCK(ki);
2457 		ki->kaio_buffer_count--;
2458 		AIO_UNLOCK(ki);
2459 	} else {
2460 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2461 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2462 		free(bp->bio_ma, M_TEMP);
2463 		atomic_subtract_int(&num_unmapped_aio, 1);
2464 	}
2465 	g_destroy_bio(bp);
2466 }
2467 
2468 static void
2469 aio_biowakeup(struct bio *bp)
2470 {
2471 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2472 	size_t nbytes;
2473 	long bcount = bp->bio_bcount;
2474 	long resid = bp->bio_resid;
2475 	int opcode, nblks;
2476 	int bio_error = bp->bio_error;
2477 	uint16_t flags = bp->bio_flags;
2478 
2479 	opcode = job->uaiocb.aio_lio_opcode;
2480 
2481 	aio_biocleanup(bp);
2482 
2483 	nbytes = bcount - resid;
2484 	atomic_add_acq_long(&job->nbytes, nbytes);
2485 	nblks = btodb(nbytes);
2486 
2487 	/*
2488 	 * If multiple bios experienced an error, the job will reflect the
2489 	 * error of whichever failed bio completed last.
2490 	 */
2491 	if (flags & BIO_ERROR)
2492 		atomic_store_int(&job->error, bio_error);
2493 	if (opcode & LIO_WRITE)
2494 		atomic_add_int(&job->outblock, nblks);
2495 	else
2496 		atomic_add_int(&job->inblock, nblks);
2497 
2498 	if (refcount_release(&job->nbio)) {
2499 		bio_error = atomic_load_int(&job->error);
2500 		if (bio_error != 0)
2501 			aio_complete(job, -1, bio_error);
2502 		else
2503 			aio_complete(job, atomic_load_long(&job->nbytes), 0);
2504 	}
2505 }
2506 
2507 /* syscall - wait for the next completion of an aio request */
2508 static int
2509 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2510     struct timespec *ts, struct aiocb_ops *ops)
2511 {
2512 	struct proc *p = td->td_proc;
2513 	struct timeval atv;
2514 	struct kaioinfo *ki;
2515 	struct kaiocb *job;
2516 	struct aiocb *ujob;
2517 	long error, status;
2518 	int timo;
2519 
2520 	ops->store_aiocb(ujobp, NULL);
2521 
2522 	if (ts == NULL) {
2523 		timo = 0;
2524 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2525 		timo = -1;
2526 	} else {
2527 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2528 			return (EINVAL);
2529 
2530 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2531 		if (itimerfix(&atv))
2532 			return (EINVAL);
2533 		timo = tvtohz(&atv);
2534 	}
2535 
2536 	if (p->p_aioinfo == NULL)
2537 		aio_init_aioinfo(p);
2538 	ki = p->p_aioinfo;
2539 
2540 	error = 0;
2541 	job = NULL;
2542 	AIO_LOCK(ki);
2543 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2544 		if (timo == -1) {
2545 			error = EWOULDBLOCK;
2546 			break;
2547 		}
2548 		ki->kaio_flags |= KAIO_WAKEUP;
2549 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2550 		    "aiowc", timo);
2551 		if (timo && error == ERESTART)
2552 			error = EINTR;
2553 		if (error)
2554 			break;
2555 	}
2556 
2557 	if (job != NULL) {
2558 		MPASS(job->jobflags & KAIOCB_FINISHED);
2559 		ujob = job->ujob;
2560 		status = job->uaiocb._aiocb_private.status;
2561 		error = job->uaiocb._aiocb_private.error;
2562 		td->td_retval[0] = status;
2563 		td->td_ru.ru_oublock += job->outblock;
2564 		td->td_ru.ru_inblock += job->inblock;
2565 		td->td_ru.ru_msgsnd += job->msgsnd;
2566 		td->td_ru.ru_msgrcv += job->msgrcv;
2567 		aio_free_entry(job);
2568 		AIO_UNLOCK(ki);
2569 		ops->store_aiocb(ujobp, ujob);
2570 		ops->store_error(ujob, error);
2571 		ops->store_status(ujob, status);
2572 	} else
2573 		AIO_UNLOCK(ki);
2574 
2575 	return (error);
2576 }
2577 
2578 int
2579 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2580 {
2581 	struct timespec ts, *tsp;
2582 	int error;
2583 
2584 	if (uap->timeout) {
2585 		/* Get timespec struct. */
2586 		error = copyin(uap->timeout, &ts, sizeof(ts));
2587 		if (error)
2588 			return (error);
2589 		tsp = &ts;
2590 	} else
2591 		tsp = NULL;
2592 
2593 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2594 }
2595 
2596 static int
2597 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2598     struct aiocb_ops *ops)
2599 {
2600 	int listop;
2601 
2602 	switch (op) {
2603 	case O_SYNC:
2604 		listop = LIO_SYNC;
2605 		break;
2606 	case O_DSYNC:
2607 		listop = LIO_DSYNC;
2608 		break;
2609 	default:
2610 		return (EINVAL);
2611 	}
2612 
2613 	return (aio_aqueue(td, ujob, NULL, listop, ops));
2614 }
2615 
2616 int
2617 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2618 {
2619 
2620 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2621 }
2622 
2623 /* kqueue attach function */
2624 static int
2625 filt_aioattach(struct knote *kn)
2626 {
2627 	struct kaiocb *job;
2628 
2629 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2630 
2631 	/*
2632 	 * The job pointer must be validated before using it, so
2633 	 * registration is restricted to the kernel; the user cannot
2634 	 * set EV_FLAG1.
2635 	 */
2636 	if ((kn->kn_flags & EV_FLAG1) == 0)
2637 		return (EPERM);
2638 	kn->kn_ptr.p_aio = job;
2639 	kn->kn_flags &= ~EV_FLAG1;
2640 
2641 	knlist_add(&job->klist, kn, 0);
2642 
2643 	return (0);
2644 }
2645 
2646 /* kqueue detach function */
2647 static void
2648 filt_aiodetach(struct knote *kn)
2649 {
2650 	struct knlist *knl;
2651 
2652 	knl = &kn->kn_ptr.p_aio->klist;
2653 	knl->kl_lock(knl->kl_lockarg);
2654 	if (!knlist_empty(knl))
2655 		knlist_remove(knl, kn, 1);
2656 	knl->kl_unlock(knl->kl_lockarg);
2657 }
2658 
2659 /* kqueue filter function */
2660 /*ARGSUSED*/
2661 static int
2662 filt_aio(struct knote *kn, long hint)
2663 {
2664 	struct kaiocb *job = kn->kn_ptr.p_aio;
2665 
2666 	kn->kn_data = job->uaiocb._aiocb_private.error;
2667 	if (!(job->jobflags & KAIOCB_FINISHED))
2668 		return (0);
2669 	kn->kn_flags |= EV_EOF;
2670 	return (1);
2671 }
2672 
2673 /* kqueue attach function */
2674 static int
2675 filt_lioattach(struct knote *kn)
2676 {
2677 	struct aioliojob *lj;
2678 
2679 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2680 
2681 	/*
2682 	 * The aioliojob pointer must be validated before using it, so
2683 	 * registration is restricted to the kernel; the user cannot
2684 	 * set EV_FLAG1.
2685 	 */
2686 	if ((kn->kn_flags & EV_FLAG1) == 0)
2687 		return (EPERM);
2688 	kn->kn_ptr.p_lio = lj;
2689 	kn->kn_flags &= ~EV_FLAG1;
2690 
2691 	knlist_add(&lj->klist, kn, 0);
2692 
2693 	return (0);
2694 }
2695 
2696 /* kqueue detach function */
2697 static void
2698 filt_liodetach(struct knote *kn)
2699 {
2700 	struct knlist *knl;
2701 
2702 	knl = &kn->kn_ptr.p_lio->klist;
2703 	knl->kl_lock(knl->kl_lockarg);
2704 	if (!knlist_empty(knl))
2705 		knlist_remove(knl, kn, 1);
2706 	knl->kl_unlock(knl->kl_lockarg);
2707 }
2708 
2709 /* kqueue filter function */
2710 /*ARGSUSED*/
2711 static int
2712 filt_lio(struct knote *kn, long hint)
2713 {
2714 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2715 
2716 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2717 }
2718 
2719 #ifdef COMPAT_FREEBSD32
2720 #include <sys/mount.h>
2721 #include <sys/socket.h>
2722 #include <sys/sysent.h>
2723 #include <compat/freebsd32/freebsd32.h>
2724 #include <compat/freebsd32/freebsd32_proto.h>
2725 #include <compat/freebsd32/freebsd32_signal.h>
2726 #include <compat/freebsd32/freebsd32_syscall.h>
2727 #include <compat/freebsd32/freebsd32_util.h>
2728 
2729 struct __aiocb_private32 {
2730 	int32_t	status;
2731 	int32_t	error;
2732 	uint32_t kernelinfo;
2733 };
2734 
2735 #ifdef COMPAT_FREEBSD6
2736 typedef struct oaiocb32 {
2737 	int	aio_fildes;		/* File descriptor */
2738 	uint64_t aio_offset __packed;	/* File offset for I/O */
2739 	uint32_t aio_buf;		/* I/O buffer in process space */
2740 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2741 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2742 	int	aio_lio_opcode;		/* LIO opcode */
2743 	int	aio_reqprio;		/* Request priority -- ignored */
2744 	struct	__aiocb_private32 _aiocb_private;
2745 } oaiocb32_t;
2746 #endif
2747 
2748 typedef struct aiocb32 {
2749 	int32_t	aio_fildes;		/* File descriptor */
2750 	uint64_t aio_offset __packed;	/* File offset for I/O */
2751 	uint32_t aio_buf;	/* I/O buffer in process space */
2752 	uint32_t aio_nbytes;	/* Number of bytes for I/O */
2753 	int	__spare__[2];
2754 	uint32_t __spare2__;
2755 	int	aio_lio_opcode;		/* LIO opcode */
2756 	int	aio_reqprio;		/* Request priority -- ignored */
2757 	struct	__aiocb_private32 _aiocb_private;
2758 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2759 } aiocb32_t;
2760 
2761 #ifdef COMPAT_FREEBSD6
2762 static int
2763 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2764 {
2765 
2766 	/*
2767 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2768 	 * supported by AIO with the old sigevent structure.
2769 	 */
2770 	CP(*osig, *nsig, sigev_notify);
2771 	switch (nsig->sigev_notify) {
2772 	case SIGEV_NONE:
2773 		break;
2774 	case SIGEV_SIGNAL:
2775 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2776 		break;
2777 	case SIGEV_KEVENT:
2778 		nsig->sigev_notify_kqueue =
2779 		    osig->__sigev_u.__sigev_notify_kqueue;
2780 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2781 		break;
2782 	default:
2783 		return (EINVAL);
2784 	}
2785 	return (0);
2786 }
2787 
2788 static int
2789 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
2790     int type __unused)
2791 {
2792 	struct oaiocb32 job32;
2793 	struct aiocb *kcb = &kjob->uaiocb;
2794 	int error;
2795 
2796 	bzero(kcb, sizeof(struct aiocb));
2797 	error = copyin(ujob, &job32, sizeof(job32));
2798 	if (error)
2799 		return (error);
2800 
2801 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
2802 
2803 	CP(job32, *kcb, aio_fildes);
2804 	CP(job32, *kcb, aio_offset);
2805 	PTRIN_CP(job32, *kcb, aio_buf);
2806 	CP(job32, *kcb, aio_nbytes);
2807 	CP(job32, *kcb, aio_lio_opcode);
2808 	CP(job32, *kcb, aio_reqprio);
2809 	CP(job32, *kcb, _aiocb_private.status);
2810 	CP(job32, *kcb, _aiocb_private.error);
2811 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2812 	return (convert_old_sigevent32(&job32.aio_sigevent,
2813 	    &kcb->aio_sigevent));
2814 }
2815 #endif
2816 
2817 static int
2818 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
2819 {
2820 	struct aiocb32 job32;
2821 	struct aiocb *kcb = &kjob->uaiocb;
2822 	struct iovec32 *iov32;
2823 	int error;
2824 
2825 	error = copyin(ujob, &job32, sizeof(job32));
2826 	if (error)
2827 		return (error);
2828 	CP(job32, *kcb, aio_fildes);
2829 	CP(job32, *kcb, aio_offset);
2830 	CP(job32, *kcb, aio_lio_opcode);
2831 	if (type == LIO_NOP)
2832 		type = kcb->aio_lio_opcode;
2833 	if (type & LIO_VECTORED) {
2834 		iov32 = PTRIN(job32.aio_iov);
2835 		CP(job32, *kcb, aio_iovcnt);
2836 		/* malloc a uio and copy in the iovec */
2837 		error = freebsd32_copyinuio(iov32,
2838 		    kcb->aio_iovcnt, &kjob->uiop);
2839 		if (error)
2840 			return (error);
2841 	} else {
2842 		PTRIN_CP(job32, *kcb, aio_buf);
2843 		CP(job32, *kcb, aio_nbytes);
2844 	}
2845 	CP(job32, *kcb, aio_reqprio);
2846 	CP(job32, *kcb, _aiocb_private.status);
2847 	CP(job32, *kcb, _aiocb_private.error);
2848 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2849 	error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent);
2850 
2851 	return (error);
2852 }
2853 
2854 static long
2855 aiocb32_fetch_status(struct aiocb *ujob)
2856 {
2857 	struct aiocb32 *ujob32;
2858 
2859 	ujob32 = (struct aiocb32 *)ujob;
2860 	return (fuword32(&ujob32->_aiocb_private.status));
2861 }
2862 
2863 static long
2864 aiocb32_fetch_error(struct aiocb *ujob)
2865 {
2866 	struct aiocb32 *ujob32;
2867 
2868 	ujob32 = (struct aiocb32 *)ujob;
2869 	return (fuword32(&ujob32->_aiocb_private.error));
2870 }
2871 
2872 static int
2873 aiocb32_store_status(struct aiocb *ujob, long status)
2874 {
2875 	struct aiocb32 *ujob32;
2876 
2877 	ujob32 = (struct aiocb32 *)ujob;
2878 	return (suword32(&ujob32->_aiocb_private.status, status));
2879 }
2880 
2881 static int
2882 aiocb32_store_error(struct aiocb *ujob, long error)
2883 {
2884 	struct aiocb32 *ujob32;
2885 
2886 	ujob32 = (struct aiocb32 *)ujob;
2887 	return (suword32(&ujob32->_aiocb_private.error, error));
2888 }
2889 
2890 static int
2891 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2892 {
2893 	struct aiocb32 *ujob32;
2894 
2895 	ujob32 = (struct aiocb32 *)ujob;
2896 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2897 }
2898 
2899 static int
2900 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2901 {
2902 
2903 	return (suword32(ujobp, (long)ujob));
2904 }
2905 
2906 static struct aiocb_ops aiocb32_ops = {
2907 	.aio_copyin = aiocb32_copyin,
2908 	.fetch_status = aiocb32_fetch_status,
2909 	.fetch_error = aiocb32_fetch_error,
2910 	.store_status = aiocb32_store_status,
2911 	.store_error = aiocb32_store_error,
2912 	.store_kernelinfo = aiocb32_store_kernelinfo,
2913 	.store_aiocb = aiocb32_store_aiocb,
2914 };
2915 
2916 #ifdef COMPAT_FREEBSD6
2917 static struct aiocb_ops aiocb32_ops_osigevent = {
2918 	.aio_copyin = aiocb32_copyin_old_sigevent,
2919 	.fetch_status = aiocb32_fetch_status,
2920 	.fetch_error = aiocb32_fetch_error,
2921 	.store_status = aiocb32_store_status,
2922 	.store_error = aiocb32_store_error,
2923 	.store_kernelinfo = aiocb32_store_kernelinfo,
2924 	.store_aiocb = aiocb32_store_aiocb,
2925 };
2926 #endif
2927 
2928 int
2929 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2930 {
2931 
2932 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2933 }
2934 
2935 int
2936 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2937 {
2938 	struct timespec32 ts32;
2939 	struct timespec ts, *tsp;
2940 	struct aiocb **ujoblist;
2941 	uint32_t *ujoblist32;
2942 	int error, i;
2943 
2944 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2945 		return (EINVAL);
2946 
2947 	if (uap->timeout) {
2948 		/* Get timespec struct. */
2949 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2950 			return (error);
2951 		CP(ts32, ts, tv_sec);
2952 		CP(ts32, ts, tv_nsec);
2953 		tsp = &ts;
2954 	} else
2955 		tsp = NULL;
2956 
2957 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2958 	ujoblist32 = (uint32_t *)ujoblist;
2959 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2960 	    sizeof(ujoblist32[0]));
2961 	if (error == 0) {
2962 		for (i = uap->nent - 1; i >= 0; i--)
2963 			ujoblist[i] = PTRIN(ujoblist32[i]);
2964 
2965 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2966 	}
2967 	free(ujoblist, M_AIO);
2968 	return (error);
2969 }
2970 
2971 int
2972 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2973 {
2974 
2975 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2976 }
2977 
2978 #ifdef COMPAT_FREEBSD6
2979 int
2980 freebsd6_freebsd32_aio_read(struct thread *td,
2981     struct freebsd6_freebsd32_aio_read_args *uap)
2982 {
2983 
2984 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2985 	    &aiocb32_ops_osigevent));
2986 }
2987 #endif
2988 
2989 int
2990 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2991 {
2992 
2993 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2994 	    &aiocb32_ops));
2995 }
2996 
2997 int
2998 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap)
2999 {
3000 
3001 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV,
3002 	    &aiocb32_ops));
3003 }
3004 
3005 #ifdef COMPAT_FREEBSD6
3006 int
3007 freebsd6_freebsd32_aio_write(struct thread *td,
3008     struct freebsd6_freebsd32_aio_write_args *uap)
3009 {
3010 
3011 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3012 	    &aiocb32_ops_osigevent));
3013 }
3014 #endif
3015 
3016 int
3017 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
3018 {
3019 
3020 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3021 	    &aiocb32_ops));
3022 }
3023 
3024 int
3025 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap)
3026 {
3027 
3028 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV,
3029 	    &aiocb32_ops));
3030 }
3031 
3032 int
3033 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
3034 {
3035 
3036 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
3037 	    &aiocb32_ops));
3038 }
3039 
3040 int
3041 freebsd32_aio_waitcomplete(struct thread *td,
3042     struct freebsd32_aio_waitcomplete_args *uap)
3043 {
3044 	struct timespec32 ts32;
3045 	struct timespec ts, *tsp;
3046 	int error;
3047 
3048 	if (uap->timeout) {
3049 		/* Get timespec struct. */
3050 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
3051 		if (error)
3052 			return (error);
3053 		CP(ts32, ts, tv_sec);
3054 		CP(ts32, ts, tv_nsec);
3055 		tsp = &ts;
3056 	} else
3057 		tsp = NULL;
3058 
3059 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
3060 	    &aiocb32_ops));
3061 }
3062 
3063 int
3064 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
3065 {
3066 
3067 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
3068 	    &aiocb32_ops));
3069 }
3070 
3071 #ifdef COMPAT_FREEBSD6
3072 int
3073 freebsd6_freebsd32_lio_listio(struct thread *td,
3074     struct freebsd6_freebsd32_lio_listio_args *uap)
3075 {
3076 	struct aiocb **acb_list;
3077 	struct sigevent *sigp, sig;
3078 	struct osigevent32 osig;
3079 	uint32_t *acb_list32;
3080 	int error, i, nent;
3081 
3082 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3083 		return (EINVAL);
3084 
3085 	nent = uap->nent;
3086 	if (nent < 0 || nent > max_aio_queue_per_proc)
3087 		return (EINVAL);
3088 
3089 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3090 		error = copyin(uap->sig, &osig, sizeof(osig));
3091 		if (error)
3092 			return (error);
3093 		error = convert_old_sigevent32(&osig, &sig);
3094 		if (error)
3095 			return (error);
3096 		sigp = &sig;
3097 	} else
3098 		sigp = NULL;
3099 
3100 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3101 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3102 	if (error) {
3103 		free(acb_list32, M_LIO);
3104 		return (error);
3105 	}
3106 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3107 	for (i = 0; i < nent; i++)
3108 		acb_list[i] = PTRIN(acb_list32[i]);
3109 	free(acb_list32, M_LIO);
3110 
3111 	error = kern_lio_listio(td, uap->mode,
3112 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3113 	    &aiocb32_ops_osigevent);
3114 	free(acb_list, M_LIO);
3115 	return (error);
3116 }
3117 #endif
3118 
3119 int
3120 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3121 {
3122 	struct aiocb **acb_list;
3123 	struct sigevent *sigp, sig;
3124 	struct sigevent32 sig32;
3125 	uint32_t *acb_list32;
3126 	int error, i, nent;
3127 
3128 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3129 		return (EINVAL);
3130 
3131 	nent = uap->nent;
3132 	if (nent < 0 || nent > max_aio_queue_per_proc)
3133 		return (EINVAL);
3134 
3135 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3136 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3137 		if (error)
3138 			return (error);
3139 		error = convert_sigevent32(&sig32, &sig);
3140 		if (error)
3141 			return (error);
3142 		sigp = &sig;
3143 	} else
3144 		sigp = NULL;
3145 
3146 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3147 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3148 	if (error) {
3149 		free(acb_list32, M_LIO);
3150 		return (error);
3151 	}
3152 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3153 	for (i = 0; i < nent; i++)
3154 		acb_list[i] = PTRIN(acb_list32[i]);
3155 	free(acb_list32, M_LIO);
3156 
3157 	error = kern_lio_listio(td, uap->mode,
3158 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3159 	    &aiocb32_ops);
3160 	free(acb_list, M_LIO);
3161 	return (error);
3162 }
3163 
3164 #endif
3165