xref: /freebsd/sys/kern/vfs_aio.c (revision 82397d791966b09d344251bc709cd9db2b3a1902)
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/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 #include <sys/sx.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
61 #include <sys/conf.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
65 
66 #include <machine/atomic.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 #include <vm/uma.h>
75 #include <sys/aio.h>
76 
77 /*
78  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
79  * overflow. (XXX will be removed soon.)
80  */
81 static u_long jobrefid;
82 
83 /*
84  * Counter for aio_fsync.
85  */
86 static uint64_t jobseqno;
87 
88 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC	32
90 #endif
91 
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC	256
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE
97 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
98 #endif
99 
100 #ifndef MAX_BUF_AIO
101 #define MAX_BUF_AIO		16
102 #endif
103 
104 FEATURE(aio, "Asynchronous I/O");
105 SYSCTL_DECL(_p1003_1b);
106 
107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
109 
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
111     "Async IO management");
112 
113 static int enable_aio_unsafe = 0;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
115     "Permit asynchronous IO on all file types, not just known-safe types");
116 
117 static unsigned int unsafe_warningcnt = 1;
118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
119     &unsafe_warningcnt, 0,
120     "Warnings that will be triggered upon failed IO requests on unsafe files");
121 
122 static int max_aio_procs = MAX_AIO_PROCS;
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
124     "Maximum number of kernel processes to use for handling async IO ");
125 
126 static int num_aio_procs = 0;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
128     "Number of presently active kernel processes for async IO");
129 
130 /*
131  * The code will adjust the actual number of AIO processes towards this
132  * number when it gets a chance.
133  */
134 static int target_aio_procs = TARGET_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
136     0,
137     "Preferred number of ready kernel processes for async IO");
138 
139 static int max_queue_count = MAX_AIO_QUEUE;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
141     "Maximum number of aio requests to queue, globally");
142 
143 static int num_queue_count = 0;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
145     "Number of queued aio requests");
146 
147 static int num_buf_aio = 0;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
149     "Number of aio requests presently handled by the buf subsystem");
150 
151 static int num_unmapped_aio = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
153     0,
154     "Number of aio requests presently handled by unmapped I/O buffers");
155 
156 /* Number of async I/O processes in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
159 
160 static int aiod_lifetime;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
162     "Maximum lifetime for idle aiod");
163 
164 static int max_aio_per_proc = MAX_AIO_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
166     0,
167     "Maximum active aio requests per process");
168 
169 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
171     &max_aio_queue_per_proc, 0,
172     "Maximum queued aio requests per process");
173 
174 static int max_buf_aio = MAX_BUF_AIO;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
176     "Maximum buf aio requests per process");
177 
178 /*
179  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
180  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
181  * vfs.aio.aio_listio_max.
182  */
183 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
184     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
185     0, "Maximum aio requests for a single lio_listio call");
186 
187 #ifdef COMPAT_FREEBSD6
188 typedef struct oaiocb {
189 	int	aio_fildes;		/* File descriptor */
190 	off_t	aio_offset;		/* File offset for I/O */
191 	volatile void *aio_buf;         /* I/O buffer in process space */
192 	size_t	aio_nbytes;		/* Number of bytes for I/O */
193 	struct	osigevent aio_sigevent;	/* Signal to deliver */
194 	int	aio_lio_opcode;		/* LIO opcode */
195 	int	aio_reqprio;		/* Request priority -- ignored */
196 	struct	__aiocb_private	_aiocb_private;
197 } oaiocb_t;
198 #endif
199 
200 /*
201  * Below is a key of locks used to protect each member of struct kaiocb
202  * aioliojob and kaioinfo and any backends.
203  *
204  * * - need not protected
205  * a - locked by kaioinfo lock
206  * b - locked by backend lock, the backend lock can be null in some cases,
207  *     for example, BIO belongs to this type, in this case, proc lock is
208  *     reused.
209  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210  */
211 
212 /*
213  * If the routine that services an AIO request blocks while running in an
214  * AIO kernel process it can starve other I/O requests.  BIO requests
215  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
216  * processes at all.  Socket I/O requests use a separate pool of
217  * kprocs and also force non-blocking I/O.  Other file I/O requests
218  * use the generic fo_read/fo_write operations which can block.  The
219  * fsync and mlock operations can also block while executing.  Ideally
220  * none of these requests would block while executing.
221  *
222  * Note that the service routines cannot toggle O_NONBLOCK in the file
223  * structure directly while handling a request due to races with
224  * userland threads.
225  */
226 
227 /* jobflags */
228 #define	KAIOCB_QUEUEING		0x01
229 #define	KAIOCB_CANCELLED	0x02
230 #define	KAIOCB_CANCELLING	0x04
231 #define	KAIOCB_CHECKSYNC	0x08
232 #define	KAIOCB_CLEARED		0x10
233 #define	KAIOCB_FINISHED		0x20
234 
235 /*
236  * AIO process info
237  */
238 #define AIOP_FREE	0x1			/* proc on free queue */
239 
240 struct aioproc {
241 	int	aioprocflags;			/* (c) AIO proc flags */
242 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
243 	struct	proc *aioproc;			/* (*) the AIO proc */
244 };
245 
246 /*
247  * data-structure for lio signal management
248  */
249 struct aioliojob {
250 	int	lioj_flags;			/* (a) listio flags */
251 	int	lioj_count;			/* (a) count of jobs */
252 	int	lioj_finished_count;		/* (a) count of finished jobs */
253 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
254 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
255 	struct	knlist klist;			/* (a) list of knotes */
256 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
257 };
258 
259 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
260 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
261 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
262 
263 /*
264  * per process aio data structure
265  */
266 struct kaioinfo {
267 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
268 	int	kaio_flags;		/* (a) per process kaio flags */
269 	int	kaio_active_count;	/* (c) number of currently used AIOs */
270 	int	kaio_count;		/* (a) size of AIO queue */
271 	int	kaio_buffer_count;	/* (a) number of bio buffers */
272 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
273 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
274 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
276 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
277 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
278 	struct	task kaio_task;		/* (*) task to kick aio processes */
279 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
280 };
281 
282 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
286 
287 #define KAIO_RUNDOWN	0x1	/* process is being run down */
288 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
289 
290 /*
291  * Operations used to interact with userland aio control blocks.
292  * Different ABIs provide their own operations.
293  */
294 struct aiocb_ops {
295 	int	(*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty);
296 	long	(*fetch_status)(struct aiocb *ujob);
297 	long	(*fetch_error)(struct aiocb *ujob);
298 	int	(*store_status)(struct aiocb *ujob, long status);
299 	int	(*store_error)(struct aiocb *ujob, long error);
300 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
302 };
303 
304 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
308 static struct unrhdr *aiod_unr;
309 
310 static void	aio_biocleanup(struct bio *bp);
311 void		aio_init_aioinfo(struct proc *p);
312 static int	aio_onceonly(void);
313 static int	aio_free_entry(struct kaiocb *job);
314 static void	aio_process_rw(struct kaiocb *job);
315 static void	aio_process_sync(struct kaiocb *job);
316 static void	aio_process_mlock(struct kaiocb *job);
317 static void	aio_schedule_fsync(void *context, int pending);
318 static int	aio_newproc(int *);
319 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
320 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
321 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
322 static void	aio_biowakeup(struct bio *bp);
323 static void	aio_proc_rundown(void *arg, struct proc *p);
324 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
325 		    struct image_params *imgp);
326 static int	aio_qbio(struct proc *p, struct kaiocb *job);
327 static void	aio_daemon(void *param);
328 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
329 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
330 static int	aio_kick(struct proc *userp);
331 static void	aio_kick_nowait(struct proc *userp);
332 static void	aio_kick_helper(void *context, int pending);
333 static int	filt_aioattach(struct knote *kn);
334 static void	filt_aiodetach(struct knote *kn);
335 static int	filt_aio(struct knote *kn, long hint);
336 static int	filt_lioattach(struct knote *kn);
337 static void	filt_liodetach(struct knote *kn);
338 static int	filt_lio(struct knote *kn, long hint);
339 
340 /*
341  * Zones for:
342  * 	kaio	Per process async io info
343  *	aiop	async io process data
344  *	aiocb	async io jobs
345  *	aiolio	list io jobs
346  */
347 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
348 
349 /* kqueue filters for aio */
350 static struct filterops aio_filtops = {
351 	.f_isfd = 0,
352 	.f_attach = filt_aioattach,
353 	.f_detach = filt_aiodetach,
354 	.f_event = filt_aio,
355 };
356 static struct filterops lio_filtops = {
357 	.f_isfd = 0,
358 	.f_attach = filt_lioattach,
359 	.f_detach = filt_liodetach,
360 	.f_event = filt_lio
361 };
362 
363 static eventhandler_tag exit_tag, exec_tag;
364 
365 TASKQUEUE_DEFINE_THREAD(aiod_kick);
366 
367 /*
368  * Main operations function for use as a kernel module.
369  */
370 static int
371 aio_modload(struct module *module, int cmd, void *arg)
372 {
373 	int error = 0;
374 
375 	switch (cmd) {
376 	case MOD_LOAD:
377 		aio_onceonly();
378 		break;
379 	case MOD_SHUTDOWN:
380 		break;
381 	default:
382 		error = EOPNOTSUPP;
383 		break;
384 	}
385 	return (error);
386 }
387 
388 static moduledata_t aio_mod = {
389 	"aio",
390 	&aio_modload,
391 	NULL
392 };
393 
394 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
395 MODULE_VERSION(aio, 1);
396 
397 /*
398  * Startup initialization
399  */
400 static int
401 aio_onceonly(void)
402 {
403 
404 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
405 	    EVENTHANDLER_PRI_ANY);
406 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
407 	    NULL, EVENTHANDLER_PRI_ANY);
408 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
409 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
410 	TAILQ_INIT(&aio_freeproc);
411 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
412 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
413 	TAILQ_INIT(&aio_jobs);
414 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
415 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
416 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
417 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
418 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
419 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
420 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
421 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
422 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
423 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
424 	jobrefid = 1;
425 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
426 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
427 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
428 
429 	return (0);
430 }
431 
432 /*
433  * Init the per-process aioinfo structure.  The aioinfo limits are set
434  * per-process for user limit (resource) management.
435  */
436 void
437 aio_init_aioinfo(struct proc *p)
438 {
439 	struct kaioinfo *ki;
440 
441 	ki = uma_zalloc(kaio_zone, M_WAITOK);
442 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
443 	ki->kaio_flags = 0;
444 	ki->kaio_active_count = 0;
445 	ki->kaio_count = 0;
446 	ki->kaio_buffer_count = 0;
447 	TAILQ_INIT(&ki->kaio_all);
448 	TAILQ_INIT(&ki->kaio_done);
449 	TAILQ_INIT(&ki->kaio_jobqueue);
450 	TAILQ_INIT(&ki->kaio_liojoblist);
451 	TAILQ_INIT(&ki->kaio_syncqueue);
452 	TAILQ_INIT(&ki->kaio_syncready);
453 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
454 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
455 	PROC_LOCK(p);
456 	if (p->p_aioinfo == NULL) {
457 		p->p_aioinfo = ki;
458 		PROC_UNLOCK(p);
459 	} else {
460 		PROC_UNLOCK(p);
461 		mtx_destroy(&ki->kaio_mtx);
462 		uma_zfree(kaio_zone, ki);
463 	}
464 
465 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
466 		aio_newproc(NULL);
467 }
468 
469 static int
470 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
471 {
472 	struct thread *td;
473 	int error;
474 
475 	error = sigev_findtd(p, sigev, &td);
476 	if (error)
477 		return (error);
478 	if (!KSI_ONQ(ksi)) {
479 		ksiginfo_set_sigev(ksi, sigev);
480 		ksi->ksi_code = SI_ASYNCIO;
481 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
482 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
483 	}
484 	PROC_UNLOCK(p);
485 	return (error);
486 }
487 
488 /*
489  * Free a job entry.  Wait for completion if it is currently active, but don't
490  * delay forever.  If we delay, we return a flag that says that we have to
491  * restart the queue scan.
492  */
493 static int
494 aio_free_entry(struct kaiocb *job)
495 {
496 	struct kaioinfo *ki;
497 	struct aioliojob *lj;
498 	struct proc *p;
499 
500 	p = job->userproc;
501 	MPASS(curproc == p);
502 	ki = p->p_aioinfo;
503 	MPASS(ki != NULL);
504 
505 	AIO_LOCK_ASSERT(ki, MA_OWNED);
506 	MPASS(job->jobflags & KAIOCB_FINISHED);
507 
508 	atomic_subtract_int(&num_queue_count, 1);
509 
510 	ki->kaio_count--;
511 	MPASS(ki->kaio_count >= 0);
512 
513 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
514 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
515 
516 	lj = job->lio;
517 	if (lj) {
518 		lj->lioj_count--;
519 		lj->lioj_finished_count--;
520 
521 		if (lj->lioj_count == 0) {
522 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
523 			/* lio is going away, we need to destroy any knotes */
524 			knlist_delete(&lj->klist, curthread, 1);
525 			PROC_LOCK(p);
526 			sigqueue_take(&lj->lioj_ksi);
527 			PROC_UNLOCK(p);
528 			uma_zfree(aiolio_zone, lj);
529 		}
530 	}
531 
532 	/* job is going away, we need to destroy any knotes */
533 	knlist_delete(&job->klist, curthread, 1);
534 	PROC_LOCK(p);
535 	sigqueue_take(&job->ksi);
536 	PROC_UNLOCK(p);
537 
538 	AIO_UNLOCK(ki);
539 
540 	/*
541 	 * The thread argument here is used to find the owning process
542 	 * and is also passed to fo_close() which may pass it to various
543 	 * places such as devsw close() routines.  Because of that, we
544 	 * need a thread pointer from the process owning the job that is
545 	 * persistent and won't disappear out from under us or move to
546 	 * another process.
547 	 *
548 	 * Currently, all the callers of this function call it to remove
549 	 * a kaiocb from the current process' job list either via a
550 	 * syscall or due to the current process calling exit() or
551 	 * execve().  Thus, we know that p == curproc.  We also know that
552 	 * curthread can't exit since we are curthread.
553 	 *
554 	 * Therefore, we use curthread as the thread to pass to
555 	 * knlist_delete().  This does mean that it is possible for the
556 	 * thread pointer at close time to differ from the thread pointer
557 	 * at open time, but this is already true of file descriptors in
558 	 * a multithreaded process.
559 	 */
560 	if (job->fd_file)
561 		fdrop(job->fd_file, curthread);
562 	crfree(job->cred);
563 	if (job->uiop != &job->uio)
564 		free(job->uiop, M_IOV);
565 	uma_zfree(aiocb_zone, job);
566 	AIO_LOCK(ki);
567 
568 	return (0);
569 }
570 
571 static void
572 aio_proc_rundown_exec(void *arg, struct proc *p,
573     struct image_params *imgp __unused)
574 {
575    	aio_proc_rundown(arg, p);
576 }
577 
578 static int
579 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
580 {
581 	aio_cancel_fn_t *func;
582 	int cancelled;
583 
584 	AIO_LOCK_ASSERT(ki, MA_OWNED);
585 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
586 		return (0);
587 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
588 	job->jobflags |= KAIOCB_CANCELLED;
589 
590 	func = job->cancel_fn;
591 
592 	/*
593 	 * If there is no cancel routine, just leave the job marked as
594 	 * cancelled.  The job should be in active use by a caller who
595 	 * should complete it normally or when it fails to install a
596 	 * cancel routine.
597 	 */
598 	if (func == NULL)
599 		return (0);
600 
601 	/*
602 	 * Set the CANCELLING flag so that aio_complete() will defer
603 	 * completions of this job.  This prevents the job from being
604 	 * freed out from under the cancel callback.  After the
605 	 * callback any deferred completion (whether from the callback
606 	 * or any other source) will be completed.
607 	 */
608 	job->jobflags |= KAIOCB_CANCELLING;
609 	AIO_UNLOCK(ki);
610 	func(job);
611 	AIO_LOCK(ki);
612 	job->jobflags &= ~KAIOCB_CANCELLING;
613 	if (job->jobflags & KAIOCB_FINISHED) {
614 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
615 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
616 		aio_bio_done_notify(p, job);
617 	} else {
618 		/*
619 		 * The cancel callback might have scheduled an
620 		 * operation to cancel this request, but it is
621 		 * only counted as cancelled if the request is
622 		 * cancelled when the callback returns.
623 		 */
624 		cancelled = 0;
625 	}
626 	return (cancelled);
627 }
628 
629 /*
630  * Rundown the jobs for a given process.
631  */
632 static void
633 aio_proc_rundown(void *arg, struct proc *p)
634 {
635 	struct kaioinfo *ki;
636 	struct aioliojob *lj;
637 	struct kaiocb *job, *jobn;
638 
639 	KASSERT(curthread->td_proc == p,
640 	    ("%s: called on non-curproc", __func__));
641 	ki = p->p_aioinfo;
642 	if (ki == NULL)
643 		return;
644 
645 	AIO_LOCK(ki);
646 	ki->kaio_flags |= KAIO_RUNDOWN;
647 
648 restart:
649 
650 	/*
651 	 * Try to cancel all pending requests. This code simulates
652 	 * aio_cancel on all pending I/O requests.
653 	 */
654 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
655 		aio_cancel_job(p, ki, job);
656 	}
657 
658 	/* Wait for all running I/O to be finished */
659 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
660 		ki->kaio_flags |= KAIO_WAKEUP;
661 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
662 		goto restart;
663 	}
664 
665 	/* Free all completed I/O requests. */
666 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
667 		aio_free_entry(job);
668 
669 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
670 		if (lj->lioj_count == 0) {
671 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
672 			knlist_delete(&lj->klist, curthread, 1);
673 			PROC_LOCK(p);
674 			sigqueue_take(&lj->lioj_ksi);
675 			PROC_UNLOCK(p);
676 			uma_zfree(aiolio_zone, lj);
677 		} else {
678 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
679 			    lj->lioj_count, lj->lioj_finished_count);
680 		}
681 	}
682 	AIO_UNLOCK(ki);
683 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
684 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
685 	mtx_destroy(&ki->kaio_mtx);
686 	uma_zfree(kaio_zone, ki);
687 	p->p_aioinfo = NULL;
688 }
689 
690 /*
691  * Select a job to run (called by an AIO daemon).
692  */
693 static struct kaiocb *
694 aio_selectjob(struct aioproc *aiop)
695 {
696 	struct kaiocb *job;
697 	struct kaioinfo *ki;
698 	struct proc *userp;
699 
700 	mtx_assert(&aio_job_mtx, MA_OWNED);
701 restart:
702 	TAILQ_FOREACH(job, &aio_jobs, list) {
703 		userp = job->userproc;
704 		ki = userp->p_aioinfo;
705 
706 		if (ki->kaio_active_count < max_aio_per_proc) {
707 			TAILQ_REMOVE(&aio_jobs, job, list);
708 			if (!aio_clear_cancel_function(job))
709 				goto restart;
710 
711 			/* Account for currently active jobs. */
712 			ki->kaio_active_count++;
713 			break;
714 		}
715 	}
716 	return (job);
717 }
718 
719 /*
720  * Move all data to a permanent storage device.  This code
721  * simulates the fsync syscall.
722  */
723 static int
724 aio_fsync_vnode(struct thread *td, struct vnode *vp)
725 {
726 	struct mount *mp;
727 	int error;
728 
729 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
730 		goto drop;
731 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
732 	if (vp->v_object != NULL) {
733 		VM_OBJECT_WLOCK(vp->v_object);
734 		vm_object_page_clean(vp->v_object, 0, 0, 0);
735 		VM_OBJECT_WUNLOCK(vp->v_object);
736 	}
737 	error = VOP_FSYNC(vp, MNT_WAIT, td);
738 
739 	VOP_UNLOCK(vp);
740 	vn_finished_write(mp);
741 drop:
742 	return (error);
743 }
744 
745 /*
746  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
747  * does the I/O request for the non-bio version of the operations.  The normal
748  * vn operations are used, and this code should work in all instances for every
749  * type of file, including pipes, sockets, fifos, and regular files.
750  *
751  * XXX I don't think it works well for socket, pipe, and fifo.
752  */
753 static void
754 aio_process_rw(struct kaiocb *job)
755 {
756 	struct ucred *td_savedcred;
757 	struct thread *td;
758 	struct aiocb *cb;
759 	struct file *fp;
760 	ssize_t cnt;
761 	long msgsnd_st, msgsnd_end;
762 	long msgrcv_st, msgrcv_end;
763 	long oublock_st, oublock_end;
764 	long inblock_st, inblock_end;
765 	int error, opcode;
766 
767 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
768 	    job->uaiocb.aio_lio_opcode == LIO_READV ||
769 	    job->uaiocb.aio_lio_opcode == LIO_WRITE ||
770 	    job->uaiocb.aio_lio_opcode == LIO_WRITEV,
771 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
772 
773 	aio_switch_vmspace(job);
774 	td = curthread;
775 	td_savedcred = td->td_ucred;
776 	td->td_ucred = job->cred;
777 	job->uiop->uio_td = td;
778 	cb = &job->uaiocb;
779 	fp = job->fd_file;
780 
781 	opcode = job->uaiocb.aio_lio_opcode;
782 	cnt = job->uiop->uio_resid;
783 
784 	msgrcv_st = td->td_ru.ru_msgrcv;
785 	msgsnd_st = td->td_ru.ru_msgsnd;
786 	inblock_st = td->td_ru.ru_inblock;
787 	oublock_st = td->td_ru.ru_oublock;
788 
789 	/*
790 	 * aio_aqueue() acquires a reference to the file that is
791 	 * released in aio_free_entry().
792 	 */
793 	if (opcode == LIO_READ || opcode == LIO_READV) {
794 		if (job->uiop->uio_resid == 0)
795 			error = 0;
796 		else
797 			error = fo_read(fp, job->uiop, fp->f_cred, FOF_OFFSET,
798 			    td);
799 	} else {
800 		if (fp->f_type == DTYPE_VNODE)
801 			bwillwrite();
802 		error = fo_write(fp, job->uiop, fp->f_cred, FOF_OFFSET, td);
803 	}
804 	msgrcv_end = td->td_ru.ru_msgrcv;
805 	msgsnd_end = td->td_ru.ru_msgsnd;
806 	inblock_end = td->td_ru.ru_inblock;
807 	oublock_end = td->td_ru.ru_oublock;
808 
809 	job->msgrcv = msgrcv_end - msgrcv_st;
810 	job->msgsnd = msgsnd_end - msgsnd_st;
811 	job->inblock = inblock_end - inblock_st;
812 	job->outblock = oublock_end - oublock_st;
813 
814 	if (error != 0 && job->uiop->uio_resid != cnt) {
815 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
816 			error = 0;
817 		if (error == EPIPE &&
818 		    (opcode == LIO_WRITE || opcode == LIO_WRITEV)) {
819 			PROC_LOCK(job->userproc);
820 			kern_psignal(job->userproc, SIGPIPE);
821 			PROC_UNLOCK(job->userproc);
822 		}
823 	}
824 
825 	cnt -= job->uiop->uio_resid;
826 	td->td_ucred = td_savedcred;
827 	if (error)
828 		aio_complete(job, -1, error);
829 	else
830 		aio_complete(job, cnt, 0);
831 }
832 
833 static void
834 aio_process_sync(struct kaiocb *job)
835 {
836 	struct thread *td = curthread;
837 	struct ucred *td_savedcred = td->td_ucred;
838 	struct file *fp = job->fd_file;
839 	int error = 0;
840 
841 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
842 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
843 
844 	td->td_ucred = job->cred;
845 	if (fp->f_vnode != NULL)
846 		error = aio_fsync_vnode(td, fp->f_vnode);
847 	td->td_ucred = td_savedcred;
848 	if (error)
849 		aio_complete(job, -1, error);
850 	else
851 		aio_complete(job, 0, 0);
852 }
853 
854 static void
855 aio_process_mlock(struct kaiocb *job)
856 {
857 	struct aiocb *cb = &job->uaiocb;
858 	int error;
859 
860 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
861 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
862 
863 	aio_switch_vmspace(job);
864 	error = kern_mlock(job->userproc, job->cred,
865 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
866 	aio_complete(job, error != 0 ? -1 : 0, error);
867 }
868 
869 static void
870 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
871 {
872 	struct aioliojob *lj;
873 	struct kaioinfo *ki;
874 	struct kaiocb *sjob, *sjobn;
875 	int lj_done;
876 	bool schedule_fsync;
877 
878 	ki = userp->p_aioinfo;
879 	AIO_LOCK_ASSERT(ki, MA_OWNED);
880 	lj = job->lio;
881 	lj_done = 0;
882 	if (lj) {
883 		lj->lioj_finished_count++;
884 		if (lj->lioj_count == lj->lioj_finished_count)
885 			lj_done = 1;
886 	}
887 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
888 	MPASS(job->jobflags & KAIOCB_FINISHED);
889 
890 	if (ki->kaio_flags & KAIO_RUNDOWN)
891 		goto notification_done;
892 
893 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
894 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
895 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
896 
897 	KNOTE_LOCKED(&job->klist, 1);
898 
899 	if (lj_done) {
900 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
901 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
902 			KNOTE_LOCKED(&lj->klist, 1);
903 		}
904 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
905 		    == LIOJ_SIGNAL &&
906 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
907 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
908 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
909 			    true);
910 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
911 		}
912 	}
913 
914 notification_done:
915 	if (job->jobflags & KAIOCB_CHECKSYNC) {
916 		schedule_fsync = false;
917 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
918 			if (job->fd_file != sjob->fd_file ||
919 			    job->seqno >= sjob->seqno)
920 				continue;
921 			if (--sjob->pending > 0)
922 				continue;
923 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
924 			if (!aio_clear_cancel_function_locked(sjob))
925 				continue;
926 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
927 			schedule_fsync = true;
928 		}
929 		if (schedule_fsync)
930 			taskqueue_enqueue(taskqueue_aiod_kick,
931 			    &ki->kaio_sync_task);
932 	}
933 	if (ki->kaio_flags & KAIO_WAKEUP) {
934 		ki->kaio_flags &= ~KAIO_WAKEUP;
935 		wakeup(&userp->p_aioinfo);
936 	}
937 }
938 
939 static void
940 aio_schedule_fsync(void *context, int pending)
941 {
942 	struct kaioinfo *ki;
943 	struct kaiocb *job;
944 
945 	ki = context;
946 	AIO_LOCK(ki);
947 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
948 		job = TAILQ_FIRST(&ki->kaio_syncready);
949 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
950 		AIO_UNLOCK(ki);
951 		aio_schedule(job, aio_process_sync);
952 		AIO_LOCK(ki);
953 	}
954 	AIO_UNLOCK(ki);
955 }
956 
957 bool
958 aio_cancel_cleared(struct kaiocb *job)
959 {
960 
961 	/*
962 	 * The caller should hold the same queue lock held when
963 	 * aio_clear_cancel_function() was called and set this flag
964 	 * ensuring this check sees an up-to-date value.  However,
965 	 * there is no way to assert that.
966 	 */
967 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
968 }
969 
970 static bool
971 aio_clear_cancel_function_locked(struct kaiocb *job)
972 {
973 
974 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
975 	MPASS(job->cancel_fn != NULL);
976 	if (job->jobflags & KAIOCB_CANCELLING) {
977 		job->jobflags |= KAIOCB_CLEARED;
978 		return (false);
979 	}
980 	job->cancel_fn = NULL;
981 	return (true);
982 }
983 
984 bool
985 aio_clear_cancel_function(struct kaiocb *job)
986 {
987 	struct kaioinfo *ki;
988 	bool ret;
989 
990 	ki = job->userproc->p_aioinfo;
991 	AIO_LOCK(ki);
992 	ret = aio_clear_cancel_function_locked(job);
993 	AIO_UNLOCK(ki);
994 	return (ret);
995 }
996 
997 static bool
998 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
999 {
1000 
1001 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1002 	if (job->jobflags & KAIOCB_CANCELLED)
1003 		return (false);
1004 	job->cancel_fn = func;
1005 	return (true);
1006 }
1007 
1008 bool
1009 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1010 {
1011 	struct kaioinfo *ki;
1012 	bool ret;
1013 
1014 	ki = job->userproc->p_aioinfo;
1015 	AIO_LOCK(ki);
1016 	ret = aio_set_cancel_function_locked(job, func);
1017 	AIO_UNLOCK(ki);
1018 	return (ret);
1019 }
1020 
1021 void
1022 aio_complete(struct kaiocb *job, long status, int error)
1023 {
1024 	struct kaioinfo *ki;
1025 	struct proc *userp;
1026 
1027 	job->uaiocb._aiocb_private.error = error;
1028 	job->uaiocb._aiocb_private.status = status;
1029 
1030 	userp = job->userproc;
1031 	ki = userp->p_aioinfo;
1032 
1033 	AIO_LOCK(ki);
1034 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1035 	    ("duplicate aio_complete"));
1036 	job->jobflags |= KAIOCB_FINISHED;
1037 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1038 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1039 		aio_bio_done_notify(userp, job);
1040 	}
1041 	AIO_UNLOCK(ki);
1042 }
1043 
1044 void
1045 aio_cancel(struct kaiocb *job)
1046 {
1047 
1048 	aio_complete(job, -1, ECANCELED);
1049 }
1050 
1051 void
1052 aio_switch_vmspace(struct kaiocb *job)
1053 {
1054 
1055 	vmspace_switch_aio(job->userproc->p_vmspace);
1056 }
1057 
1058 /*
1059  * The AIO daemon, most of the actual work is done in aio_process_*,
1060  * but the setup (and address space mgmt) is done in this routine.
1061  */
1062 static void
1063 aio_daemon(void *_id)
1064 {
1065 	struct kaiocb *job;
1066 	struct aioproc *aiop;
1067 	struct kaioinfo *ki;
1068 	struct proc *p;
1069 	struct vmspace *myvm;
1070 	struct thread *td = curthread;
1071 	int id = (intptr_t)_id;
1072 
1073 	/*
1074 	 * Grab an extra reference on the daemon's vmspace so that it
1075 	 * doesn't get freed by jobs that switch to a different
1076 	 * vmspace.
1077 	 */
1078 	p = td->td_proc;
1079 	myvm = vmspace_acquire_ref(p);
1080 
1081 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1082 
1083 	/*
1084 	 * Allocate and ready the aio control info.  There is one aiop structure
1085 	 * per daemon.
1086 	 */
1087 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1088 	aiop->aioproc = p;
1089 	aiop->aioprocflags = 0;
1090 
1091 	/*
1092 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1093 	 * and creating too many daemons.)
1094 	 */
1095 	sema_post(&aio_newproc_sem);
1096 
1097 	mtx_lock(&aio_job_mtx);
1098 	for (;;) {
1099 		/*
1100 		 * Take daemon off of free queue
1101 		 */
1102 		if (aiop->aioprocflags & AIOP_FREE) {
1103 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1104 			aiop->aioprocflags &= ~AIOP_FREE;
1105 		}
1106 
1107 		/*
1108 		 * Check for jobs.
1109 		 */
1110 		while ((job = aio_selectjob(aiop)) != NULL) {
1111 			mtx_unlock(&aio_job_mtx);
1112 
1113 			ki = job->userproc->p_aioinfo;
1114 			job->handle_fn(job);
1115 
1116 			mtx_lock(&aio_job_mtx);
1117 			/* Decrement the active job count. */
1118 			ki->kaio_active_count--;
1119 		}
1120 
1121 		/*
1122 		 * Disconnect from user address space.
1123 		 */
1124 		if (p->p_vmspace != myvm) {
1125 			mtx_unlock(&aio_job_mtx);
1126 			vmspace_switch_aio(myvm);
1127 			mtx_lock(&aio_job_mtx);
1128 			/*
1129 			 * We have to restart to avoid race, we only sleep if
1130 			 * no job can be selected.
1131 			 */
1132 			continue;
1133 		}
1134 
1135 		mtx_assert(&aio_job_mtx, MA_OWNED);
1136 
1137 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1138 		aiop->aioprocflags |= AIOP_FREE;
1139 
1140 		/*
1141 		 * If daemon is inactive for a long time, allow it to exit,
1142 		 * thereby freeing resources.
1143 		 */
1144 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1145 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1146 		    (aiop->aioprocflags & AIOP_FREE) &&
1147 		    num_aio_procs > target_aio_procs)
1148 			break;
1149 	}
1150 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1151 	num_aio_procs--;
1152 	mtx_unlock(&aio_job_mtx);
1153 	uma_zfree(aiop_zone, aiop);
1154 	free_unr(aiod_unr, id);
1155 	vmspace_free(myvm);
1156 
1157 	KASSERT(p->p_vmspace == myvm,
1158 	    ("AIOD: bad vmspace for exiting daemon"));
1159 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1160 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1161 	    refcount_load(&myvm->vm_refcnt)));
1162 	kproc_exit(0);
1163 }
1164 
1165 /*
1166  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1167  * AIO daemon modifies its environment itself.
1168  */
1169 static int
1170 aio_newproc(int *start)
1171 {
1172 	int error;
1173 	struct proc *p;
1174 	int id;
1175 
1176 	id = alloc_unr(aiod_unr);
1177 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1178 		RFNOWAIT, 0, "aiod%d", id);
1179 	if (error == 0) {
1180 		/*
1181 		 * Wait until daemon is started.
1182 		 */
1183 		sema_wait(&aio_newproc_sem);
1184 		mtx_lock(&aio_job_mtx);
1185 		num_aio_procs++;
1186 		if (start != NULL)
1187 			(*start)--;
1188 		mtx_unlock(&aio_job_mtx);
1189 	} else {
1190 		free_unr(aiod_unr, id);
1191 	}
1192 	return (error);
1193 }
1194 
1195 /*
1196  * Try the high-performance, low-overhead bio method for eligible
1197  * VCHR devices.  This method doesn't use an aio helper thread, and
1198  * thus has very low overhead.
1199  *
1200  * Assumes that the caller, aio_aqueue(), has incremented the file
1201  * structure's reference count, preventing its deallocation for the
1202  * duration of this call.
1203  */
1204 static int
1205 aio_qbio(struct proc *p, struct kaiocb *job)
1206 {
1207 	struct aiocb *cb;
1208 	struct file *fp;
1209 	struct buf *pbuf;
1210 	struct vnode *vp;
1211 	struct cdevsw *csw;
1212 	struct cdev *dev;
1213 	struct kaioinfo *ki;
1214 	struct bio **bios = NULL;
1215 	off_t offset;
1216 	int bio_cmd, error, i, iovcnt, opcode, poff, ref;
1217 	vm_prot_t prot;
1218 	bool use_unmapped;
1219 
1220 	cb = &job->uaiocb;
1221 	fp = job->fd_file;
1222 	opcode = cb->aio_lio_opcode;
1223 
1224 	if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV ||
1225 	    opcode == LIO_READ || opcode == LIO_READV))
1226 		return (-1);
1227 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1228 		return (-1);
1229 
1230 	vp = fp->f_vnode;
1231 	if (vp->v_type != VCHR)
1232 		return (-1);
1233 	if (vp->v_bufobj.bo_bsize == 0)
1234 		return (-1);
1235 
1236 	bio_cmd = opcode == LIO_WRITE || opcode == LIO_WRITEV ? BIO_WRITE :
1237 	    BIO_READ;
1238 	iovcnt = job->uiop->uio_iovcnt;
1239 	if (iovcnt > max_buf_aio)
1240 		return (-1);
1241 	for (i = 0; i < iovcnt; i++) {
1242 		if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0)
1243 			return (-1);
1244 		if (job->uiop->uio_iov[i].iov_len > maxphys) {
1245 			error = -1;
1246 			return (-1);
1247 		}
1248 	}
1249 	offset = cb->aio_offset;
1250 
1251 	ref = 0;
1252 	csw = devvn_refthread(vp, &dev, &ref);
1253 	if (csw == NULL)
1254 		return (ENXIO);
1255 
1256 	if ((csw->d_flags & D_DISK) == 0) {
1257 		error = -1;
1258 		goto unref;
1259 	}
1260 	if (job->uiop->uio_resid > dev->si_iosize_max) {
1261 		error = -1;
1262 		goto unref;
1263 	}
1264 
1265 	ki = p->p_aioinfo;
1266 	job->error = 0;
1267 
1268 	use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed;
1269 	if (!use_unmapped) {
1270 		AIO_LOCK(ki);
1271 		if (ki->kaio_buffer_count + iovcnt > max_buf_aio) {
1272 			AIO_UNLOCK(ki);
1273 			error = EAGAIN;
1274 			goto unref;
1275 		}
1276 		ki->kaio_buffer_count += iovcnt;
1277 		AIO_UNLOCK(ki);
1278 	}
1279 
1280 	bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK);
1281 	atomic_store_int(&job->nbio, iovcnt);
1282 	for (i = 0; i < iovcnt; i++) {
1283 		struct vm_page** pages;
1284 		struct bio *bp;
1285 		void *buf;
1286 		size_t nbytes;
1287 		int npages;
1288 
1289 		buf = job->uiop->uio_iov[i].iov_base;
1290 		nbytes = job->uiop->uio_iov[i].iov_len;
1291 
1292 		bios[i] = g_alloc_bio();
1293 		bp = bios[i];
1294 
1295 		poff = (vm_offset_t)buf & PAGE_MASK;
1296 		if (use_unmapped) {
1297 			pbuf = NULL;
1298 			pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1299 			    nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1300 		} else {
1301 			pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1302 			BUF_KERNPROC(pbuf);
1303 			pages = pbuf->b_pages;
1304 		}
1305 
1306 		bp->bio_length = nbytes;
1307 		bp->bio_bcount = nbytes;
1308 		bp->bio_done = aio_biowakeup;
1309 		bp->bio_offset = offset;
1310 		bp->bio_cmd = bio_cmd;
1311 		bp->bio_dev = dev;
1312 		bp->bio_caller1 = job;
1313 		bp->bio_caller2 = pbuf;
1314 
1315 		prot = VM_PROT_READ;
1316 		if (opcode == LIO_READ || opcode == LIO_READV)
1317 			prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1318 		npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1319 		    (vm_offset_t)buf, bp->bio_length, prot, pages,
1320 		    atop(maxphys) + 1);
1321 		if (npages < 0) {
1322 			if (pbuf != NULL)
1323 				uma_zfree(pbuf_zone, pbuf);
1324 			else
1325 				free(pages, M_TEMP);
1326 			error = EFAULT;
1327 			g_destroy_bio(bp);
1328 			i--;
1329 			goto destroy_bios;
1330 		}
1331 		if (pbuf != NULL) {
1332 			pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages);
1333 			bp->bio_data = pbuf->b_data + poff;
1334 			pbuf->b_npages = npages;
1335 			atomic_add_int(&num_buf_aio, 1);
1336 		} else {
1337 			bp->bio_ma = pages;
1338 			bp->bio_ma_n = npages;
1339 			bp->bio_ma_offset = poff;
1340 			bp->bio_data = unmapped_buf;
1341 			bp->bio_flags |= BIO_UNMAPPED;
1342 			atomic_add_int(&num_unmapped_aio, 1);
1343 		}
1344 
1345 		offset += nbytes;
1346 	}
1347 
1348 	/* Perform transfer. */
1349 	for (i = 0; i < iovcnt; i++)
1350 		csw->d_strategy(bios[i]);
1351 	free(bios, M_TEMP);
1352 
1353 	dev_relthread(dev, ref);
1354 	return (0);
1355 
1356 destroy_bios:
1357 	for (; i >= 0; i--)
1358 		aio_biocleanup(bios[i]);
1359 	free(bios, M_TEMP);
1360 unref:
1361 	dev_relthread(dev, ref);
1362 	return (error);
1363 }
1364 
1365 #ifdef COMPAT_FREEBSD6
1366 static int
1367 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1368 {
1369 
1370 	/*
1371 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1372 	 * supported by AIO with the old sigevent structure.
1373 	 */
1374 	nsig->sigev_notify = osig->sigev_notify;
1375 	switch (nsig->sigev_notify) {
1376 	case SIGEV_NONE:
1377 		break;
1378 	case SIGEV_SIGNAL:
1379 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1380 		break;
1381 	case SIGEV_KEVENT:
1382 		nsig->sigev_notify_kqueue =
1383 		    osig->__sigev_u.__sigev_notify_kqueue;
1384 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1385 		break;
1386 	default:
1387 		return (EINVAL);
1388 	}
1389 	return (0);
1390 }
1391 
1392 static int
1393 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
1394     int type __unused)
1395 {
1396 	struct oaiocb *ojob;
1397 	struct aiocb *kcb = &kjob->uaiocb;
1398 	int error;
1399 
1400 	bzero(kcb, sizeof(struct aiocb));
1401 	error = copyin(ujob, kcb, sizeof(struct oaiocb));
1402 	if (error)
1403 		return (error);
1404 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
1405 	ojob = (struct oaiocb *)kcb;
1406 	return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent));
1407 }
1408 #endif
1409 
1410 static int
1411 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
1412 {
1413 	struct aiocb *kcb = &kjob->uaiocb;
1414 	int error;
1415 
1416 	error = copyin(ujob, kcb, sizeof(struct aiocb));
1417 	if (error)
1418 		return (error);
1419 	if (type == LIO_READV || type == LIO_WRITEV) {
1420 		/* malloc a uio and copy in the iovec */
1421 		error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov),
1422 		    kcb->aio_iovcnt, &kjob->uiop);
1423 	}
1424 
1425 	return (error);
1426 }
1427 
1428 static long
1429 aiocb_fetch_status(struct aiocb *ujob)
1430 {
1431 
1432 	return (fuword(&ujob->_aiocb_private.status));
1433 }
1434 
1435 static long
1436 aiocb_fetch_error(struct aiocb *ujob)
1437 {
1438 
1439 	return (fuword(&ujob->_aiocb_private.error));
1440 }
1441 
1442 static int
1443 aiocb_store_status(struct aiocb *ujob, long status)
1444 {
1445 
1446 	return (suword(&ujob->_aiocb_private.status, status));
1447 }
1448 
1449 static int
1450 aiocb_store_error(struct aiocb *ujob, long error)
1451 {
1452 
1453 	return (suword(&ujob->_aiocb_private.error, error));
1454 }
1455 
1456 static int
1457 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1458 {
1459 
1460 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1461 }
1462 
1463 static int
1464 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1465 {
1466 
1467 	return (suword(ujobp, (long)ujob));
1468 }
1469 
1470 static struct aiocb_ops aiocb_ops = {
1471 	.aio_copyin = aiocb_copyin,
1472 	.fetch_status = aiocb_fetch_status,
1473 	.fetch_error = aiocb_fetch_error,
1474 	.store_status = aiocb_store_status,
1475 	.store_error = aiocb_store_error,
1476 	.store_kernelinfo = aiocb_store_kernelinfo,
1477 	.store_aiocb = aiocb_store_aiocb,
1478 };
1479 
1480 #ifdef COMPAT_FREEBSD6
1481 static struct aiocb_ops aiocb_ops_osigevent = {
1482 	.aio_copyin = aiocb_copyin_old_sigevent,
1483 	.fetch_status = aiocb_fetch_status,
1484 	.fetch_error = aiocb_fetch_error,
1485 	.store_status = aiocb_store_status,
1486 	.store_error = aiocb_store_error,
1487 	.store_kernelinfo = aiocb_store_kernelinfo,
1488 	.store_aiocb = aiocb_store_aiocb,
1489 };
1490 #endif
1491 
1492 /*
1493  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1494  * technique is done in this code.
1495  */
1496 int
1497 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1498     int type, struct aiocb_ops *ops)
1499 {
1500 	struct proc *p = td->td_proc;
1501 	struct file *fp = NULL;
1502 	struct kaiocb *job;
1503 	struct kaioinfo *ki;
1504 	struct kevent kev;
1505 	int opcode;
1506 	int error;
1507 	int fd, kqfd;
1508 	int jid;
1509 	u_short evflags;
1510 
1511 	if (p->p_aioinfo == NULL)
1512 		aio_init_aioinfo(p);
1513 
1514 	ki = p->p_aioinfo;
1515 
1516 	ops->store_status(ujob, -1);
1517 	ops->store_error(ujob, 0);
1518 	ops->store_kernelinfo(ujob, -1);
1519 
1520 	if (num_queue_count >= max_queue_count ||
1521 	    ki->kaio_count >= max_aio_queue_per_proc) {
1522 		error = EAGAIN;
1523 		goto err1;
1524 	}
1525 
1526 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1527 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1528 
1529 	error = ops->aio_copyin(ujob, job, type);
1530 	if (error)
1531 		goto err2;
1532 
1533 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1534 		error = EINVAL;
1535 		goto err2;
1536 	}
1537 
1538 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1539 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1540 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1541 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1542 		error = EINVAL;
1543 		goto err2;
1544 	}
1545 
1546 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1547 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1548 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1549 		error = EINVAL;
1550 		goto err2;
1551 	}
1552 
1553 	ksiginfo_init(&job->ksi);
1554 
1555 	/* Save userspace address of the job info. */
1556 	job->ujob = ujob;
1557 
1558 	/* Get the opcode. */
1559 	if (type != LIO_NOP)
1560 		job->uaiocb.aio_lio_opcode = type;
1561 	opcode = job->uaiocb.aio_lio_opcode;
1562 
1563 	/*
1564 	 * Validate the opcode and fetch the file object for the specified
1565 	 * file descriptor.
1566 	 *
1567 	 * XXXRW: Moved the opcode validation up here so that we don't
1568 	 * retrieve a file descriptor without knowing what the capabiltity
1569 	 * should be.
1570 	 */
1571 	fd = job->uaiocb.aio_fildes;
1572 	switch (opcode) {
1573 	case LIO_WRITE:
1574 	case LIO_WRITEV:
1575 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1576 		break;
1577 	case LIO_READ:
1578 	case LIO_READV:
1579 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1580 		break;
1581 	case LIO_SYNC:
1582 		error = fget(td, fd, &cap_fsync_rights, &fp);
1583 		break;
1584 	case LIO_MLOCK:
1585 		break;
1586 	case LIO_NOP:
1587 		error = fget(td, fd, &cap_no_rights, &fp);
1588 		break;
1589 	default:
1590 		error = EINVAL;
1591 	}
1592 	if (error)
1593 		goto err3;
1594 
1595 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1596 		error = EINVAL;
1597 		goto err3;
1598 	}
1599 
1600 	if ((opcode == LIO_READ || opcode == LIO_READV ||
1601 	    opcode == LIO_WRITE || opcode == LIO_WRITEV) &&
1602 	    job->uaiocb.aio_offset < 0 &&
1603 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1604 		error = EINVAL;
1605 		goto err3;
1606 	}
1607 
1608 	job->fd_file = fp;
1609 
1610 	mtx_lock(&aio_job_mtx);
1611 	jid = jobrefid++;
1612 	job->seqno = jobseqno++;
1613 	mtx_unlock(&aio_job_mtx);
1614 	error = ops->store_kernelinfo(ujob, jid);
1615 	if (error) {
1616 		error = EINVAL;
1617 		goto err3;
1618 	}
1619 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1620 
1621 	if (opcode == LIO_NOP) {
1622 		fdrop(fp, td);
1623 		MPASS(job->uiop == &job->uio || job->uiop == NULL);
1624 		uma_zfree(aiocb_zone, job);
1625 		return (0);
1626 	}
1627 
1628 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1629 		goto no_kqueue;
1630 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1631 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1632 		error = EINVAL;
1633 		goto err3;
1634 	}
1635 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1636 	memset(&kev, 0, sizeof(kev));
1637 	kev.ident = (uintptr_t)job->ujob;
1638 	kev.filter = EVFILT_AIO;
1639 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1640 	kev.data = (intptr_t)job;
1641 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1642 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1643 	if (error)
1644 		goto err3;
1645 
1646 no_kqueue:
1647 
1648 	ops->store_error(ujob, EINPROGRESS);
1649 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1650 	job->userproc = p;
1651 	job->cred = crhold(td->td_ucred);
1652 	job->jobflags = KAIOCB_QUEUEING;
1653 	job->lio = lj;
1654 
1655 	switch (opcode) {
1656 	case LIO_READV:
1657 	case LIO_WRITEV:
1658 		/* Use the uio copied in by aio_copyin */
1659 		MPASS(job->uiop != &job->uio && job->uiop != NULL);
1660 		break;
1661 	case LIO_READ:
1662 	case LIO_WRITE:
1663 		/* Setup the inline uio */
1664 		job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf;
1665 		job->iov[0].iov_len = job->uaiocb.aio_nbytes;
1666 		job->uio.uio_iov = job->iov;
1667 		job->uio.uio_iovcnt = 1;
1668 		job->uio.uio_resid = job->uaiocb.aio_nbytes;
1669 		job->uio.uio_segflg = UIO_USERSPACE;
1670 		/* FALLTHROUGH */
1671 	default:
1672 		job->uiop = &job->uio;
1673 		break;
1674 	}
1675 	switch (opcode) {
1676 	case LIO_READ:
1677 	case LIO_READV:
1678 		job->uiop->uio_rw = UIO_READ;
1679 		break;
1680 	case LIO_WRITE:
1681 	case LIO_WRITEV:
1682 		job->uiop->uio_rw = UIO_WRITE;
1683 		break;
1684 	}
1685 	job->uiop->uio_offset = job->uaiocb.aio_offset;
1686 	job->uiop->uio_td = td;
1687 
1688 	if (opcode == LIO_MLOCK) {
1689 		aio_schedule(job, aio_process_mlock);
1690 		error = 0;
1691 	} else if (fp->f_ops->fo_aio_queue == NULL)
1692 		error = aio_queue_file(fp, job);
1693 	else
1694 		error = fo_aio_queue(fp, job);
1695 	if (error)
1696 		goto err3;
1697 
1698 	AIO_LOCK(ki);
1699 	job->jobflags &= ~KAIOCB_QUEUEING;
1700 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1701 	ki->kaio_count++;
1702 	if (lj)
1703 		lj->lioj_count++;
1704 	atomic_add_int(&num_queue_count, 1);
1705 	if (job->jobflags & KAIOCB_FINISHED) {
1706 		/*
1707 		 * The queue callback completed the request synchronously.
1708 		 * The bulk of the completion is deferred in that case
1709 		 * until this point.
1710 		 */
1711 		aio_bio_done_notify(p, job);
1712 	} else
1713 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1714 	AIO_UNLOCK(ki);
1715 	return (0);
1716 
1717 err3:
1718 	if (fp)
1719 		fdrop(fp, td);
1720 	knlist_delete(&job->klist, curthread, 0);
1721 err2:
1722 	if (job->uiop != &job->uio)
1723 		free(job->uiop, M_IOV);
1724 	uma_zfree(aiocb_zone, job);
1725 err1:
1726 	ops->store_error(ujob, error);
1727 	return (error);
1728 }
1729 
1730 static void
1731 aio_cancel_daemon_job(struct kaiocb *job)
1732 {
1733 
1734 	mtx_lock(&aio_job_mtx);
1735 	if (!aio_cancel_cleared(job))
1736 		TAILQ_REMOVE(&aio_jobs, job, list);
1737 	mtx_unlock(&aio_job_mtx);
1738 	aio_cancel(job);
1739 }
1740 
1741 void
1742 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1743 {
1744 
1745 	mtx_lock(&aio_job_mtx);
1746 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1747 		mtx_unlock(&aio_job_mtx);
1748 		aio_cancel(job);
1749 		return;
1750 	}
1751 	job->handle_fn = func;
1752 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1753 	aio_kick_nowait(job->userproc);
1754 	mtx_unlock(&aio_job_mtx);
1755 }
1756 
1757 static void
1758 aio_cancel_sync(struct kaiocb *job)
1759 {
1760 	struct kaioinfo *ki;
1761 
1762 	ki = job->userproc->p_aioinfo;
1763 	AIO_LOCK(ki);
1764 	if (!aio_cancel_cleared(job))
1765 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1766 	AIO_UNLOCK(ki);
1767 	aio_cancel(job);
1768 }
1769 
1770 int
1771 aio_queue_file(struct file *fp, struct kaiocb *job)
1772 {
1773 	struct kaioinfo *ki;
1774 	struct kaiocb *job2;
1775 	struct vnode *vp;
1776 	struct mount *mp;
1777 	int error;
1778 	bool safe;
1779 
1780 	ki = job->userproc->p_aioinfo;
1781 	error = aio_qbio(job->userproc, job);
1782 	if (error >= 0)
1783 		return (error);
1784 	safe = false;
1785 	if (fp->f_type == DTYPE_VNODE) {
1786 		vp = fp->f_vnode;
1787 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1788 			mp = fp->f_vnode->v_mount;
1789 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1790 				safe = true;
1791 		}
1792 	}
1793 	if (!(safe || enable_aio_unsafe)) {
1794 		counted_warning(&unsafe_warningcnt,
1795 		    "is attempting to use unsafe AIO requests");
1796 		return (EOPNOTSUPP);
1797 	}
1798 
1799 	switch (job->uaiocb.aio_lio_opcode) {
1800 	case LIO_READ:
1801 	case LIO_READV:
1802 	case LIO_WRITE:
1803 	case LIO_WRITEV:
1804 		aio_schedule(job, aio_process_rw);
1805 		error = 0;
1806 		break;
1807 	case LIO_SYNC:
1808 		AIO_LOCK(ki);
1809 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1810 			if (job2->fd_file == job->fd_file &&
1811 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1812 			    job2->seqno < job->seqno) {
1813 				job2->jobflags |= KAIOCB_CHECKSYNC;
1814 				job->pending++;
1815 			}
1816 		}
1817 		if (job->pending != 0) {
1818 			if (!aio_set_cancel_function_locked(job,
1819 				aio_cancel_sync)) {
1820 				AIO_UNLOCK(ki);
1821 				aio_cancel(job);
1822 				return (0);
1823 			}
1824 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1825 			AIO_UNLOCK(ki);
1826 			return (0);
1827 		}
1828 		AIO_UNLOCK(ki);
1829 		aio_schedule(job, aio_process_sync);
1830 		error = 0;
1831 		break;
1832 	default:
1833 		error = EINVAL;
1834 	}
1835 	return (error);
1836 }
1837 
1838 static void
1839 aio_kick_nowait(struct proc *userp)
1840 {
1841 	struct kaioinfo *ki = userp->p_aioinfo;
1842 	struct aioproc *aiop;
1843 
1844 	mtx_assert(&aio_job_mtx, MA_OWNED);
1845 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1846 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1847 		aiop->aioprocflags &= ~AIOP_FREE;
1848 		wakeup(aiop->aioproc);
1849 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1850 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1851 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1852 	}
1853 }
1854 
1855 static int
1856 aio_kick(struct proc *userp)
1857 {
1858 	struct kaioinfo *ki = userp->p_aioinfo;
1859 	struct aioproc *aiop;
1860 	int error, ret = 0;
1861 
1862 	mtx_assert(&aio_job_mtx, MA_OWNED);
1863 retryproc:
1864 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1865 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1866 		aiop->aioprocflags &= ~AIOP_FREE;
1867 		wakeup(aiop->aioproc);
1868 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1869 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1870 		num_aio_resv_start++;
1871 		mtx_unlock(&aio_job_mtx);
1872 		error = aio_newproc(&num_aio_resv_start);
1873 		mtx_lock(&aio_job_mtx);
1874 		if (error) {
1875 			num_aio_resv_start--;
1876 			goto retryproc;
1877 		}
1878 	} else {
1879 		ret = -1;
1880 	}
1881 	return (ret);
1882 }
1883 
1884 static void
1885 aio_kick_helper(void *context, int pending)
1886 {
1887 	struct proc *userp = context;
1888 
1889 	mtx_lock(&aio_job_mtx);
1890 	while (--pending >= 0) {
1891 		if (aio_kick(userp))
1892 			break;
1893 	}
1894 	mtx_unlock(&aio_job_mtx);
1895 }
1896 
1897 /*
1898  * Support the aio_return system call, as a side-effect, kernel resources are
1899  * released.
1900  */
1901 static int
1902 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1903 {
1904 	struct proc *p = td->td_proc;
1905 	struct kaiocb *job;
1906 	struct kaioinfo *ki;
1907 	long status, error;
1908 
1909 	ki = p->p_aioinfo;
1910 	if (ki == NULL)
1911 		return (EINVAL);
1912 	AIO_LOCK(ki);
1913 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1914 		if (job->ujob == ujob)
1915 			break;
1916 	}
1917 	if (job != NULL) {
1918 		MPASS(job->jobflags & KAIOCB_FINISHED);
1919 		status = job->uaiocb._aiocb_private.status;
1920 		error = job->uaiocb._aiocb_private.error;
1921 		td->td_retval[0] = status;
1922 		td->td_ru.ru_oublock += job->outblock;
1923 		td->td_ru.ru_inblock += job->inblock;
1924 		td->td_ru.ru_msgsnd += job->msgsnd;
1925 		td->td_ru.ru_msgrcv += job->msgrcv;
1926 		aio_free_entry(job);
1927 		AIO_UNLOCK(ki);
1928 		ops->store_error(ujob, error);
1929 		ops->store_status(ujob, status);
1930 	} else {
1931 		error = EINVAL;
1932 		AIO_UNLOCK(ki);
1933 	}
1934 	return (error);
1935 }
1936 
1937 int
1938 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1939 {
1940 
1941 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1942 }
1943 
1944 /*
1945  * Allow a process to wakeup when any of the I/O requests are completed.
1946  */
1947 static int
1948 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1949     struct timespec *ts)
1950 {
1951 	struct proc *p = td->td_proc;
1952 	struct timeval atv;
1953 	struct kaioinfo *ki;
1954 	struct kaiocb *firstjob, *job;
1955 	int error, i, timo;
1956 
1957 	timo = 0;
1958 	if (ts) {
1959 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1960 			return (EINVAL);
1961 
1962 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1963 		if (itimerfix(&atv))
1964 			return (EINVAL);
1965 		timo = tvtohz(&atv);
1966 	}
1967 
1968 	ki = p->p_aioinfo;
1969 	if (ki == NULL)
1970 		return (EAGAIN);
1971 
1972 	if (njoblist == 0)
1973 		return (0);
1974 
1975 	AIO_LOCK(ki);
1976 	for (;;) {
1977 		firstjob = NULL;
1978 		error = 0;
1979 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1980 			for (i = 0; i < njoblist; i++) {
1981 				if (job->ujob == ujoblist[i]) {
1982 					if (firstjob == NULL)
1983 						firstjob = job;
1984 					if (job->jobflags & KAIOCB_FINISHED)
1985 						goto RETURN;
1986 				}
1987 			}
1988 		}
1989 		/* All tasks were finished. */
1990 		if (firstjob == NULL)
1991 			break;
1992 
1993 		ki->kaio_flags |= KAIO_WAKEUP;
1994 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1995 		    "aiospn", timo);
1996 		if (error == ERESTART)
1997 			error = EINTR;
1998 		if (error)
1999 			break;
2000 	}
2001 RETURN:
2002 	AIO_UNLOCK(ki);
2003 	return (error);
2004 }
2005 
2006 int
2007 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2008 {
2009 	struct timespec ts, *tsp;
2010 	struct aiocb **ujoblist;
2011 	int error;
2012 
2013 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2014 		return (EINVAL);
2015 
2016 	if (uap->timeout) {
2017 		/* Get timespec struct. */
2018 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2019 			return (error);
2020 		tsp = &ts;
2021 	} else
2022 		tsp = NULL;
2023 
2024 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2025 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2026 	if (error == 0)
2027 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2028 	free(ujoblist, M_AIOS);
2029 	return (error);
2030 }
2031 
2032 /*
2033  * aio_cancel cancels any non-bio aio operations not currently in progress.
2034  */
2035 int
2036 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2037 {
2038 	struct proc *p = td->td_proc;
2039 	struct kaioinfo *ki;
2040 	struct kaiocb *job, *jobn;
2041 	struct file *fp;
2042 	int error;
2043 	int cancelled = 0;
2044 	int notcancelled = 0;
2045 	struct vnode *vp;
2046 
2047 	/* Lookup file object. */
2048 	error = fget(td, uap->fd, &cap_no_rights, &fp);
2049 	if (error)
2050 		return (error);
2051 
2052 	ki = p->p_aioinfo;
2053 	if (ki == NULL)
2054 		goto done;
2055 
2056 	if (fp->f_type == DTYPE_VNODE) {
2057 		vp = fp->f_vnode;
2058 		if (vn_isdisk(vp)) {
2059 			fdrop(fp, td);
2060 			td->td_retval[0] = AIO_NOTCANCELED;
2061 			return (0);
2062 		}
2063 	}
2064 
2065 	AIO_LOCK(ki);
2066 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2067 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2068 		    ((uap->aiocbp == NULL) ||
2069 		     (uap->aiocbp == job->ujob))) {
2070 			if (aio_cancel_job(p, ki, job)) {
2071 				cancelled++;
2072 			} else {
2073 				notcancelled++;
2074 			}
2075 			if (uap->aiocbp != NULL)
2076 				break;
2077 		}
2078 	}
2079 	AIO_UNLOCK(ki);
2080 
2081 done:
2082 	fdrop(fp, td);
2083 
2084 	if (uap->aiocbp != NULL) {
2085 		if (cancelled) {
2086 			td->td_retval[0] = AIO_CANCELED;
2087 			return (0);
2088 		}
2089 	}
2090 
2091 	if (notcancelled) {
2092 		td->td_retval[0] = AIO_NOTCANCELED;
2093 		return (0);
2094 	}
2095 
2096 	if (cancelled) {
2097 		td->td_retval[0] = AIO_CANCELED;
2098 		return (0);
2099 	}
2100 
2101 	td->td_retval[0] = AIO_ALLDONE;
2102 
2103 	return (0);
2104 }
2105 
2106 /*
2107  * aio_error is implemented in the kernel level for compatibility purposes
2108  * only.  For a user mode async implementation, it would be best to do it in
2109  * a userland subroutine.
2110  */
2111 static int
2112 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2113 {
2114 	struct proc *p = td->td_proc;
2115 	struct kaiocb *job;
2116 	struct kaioinfo *ki;
2117 	int status;
2118 
2119 	ki = p->p_aioinfo;
2120 	if (ki == NULL) {
2121 		td->td_retval[0] = EINVAL;
2122 		return (0);
2123 	}
2124 
2125 	AIO_LOCK(ki);
2126 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2127 		if (job->ujob == ujob) {
2128 			if (job->jobflags & KAIOCB_FINISHED)
2129 				td->td_retval[0] =
2130 					job->uaiocb._aiocb_private.error;
2131 			else
2132 				td->td_retval[0] = EINPROGRESS;
2133 			AIO_UNLOCK(ki);
2134 			return (0);
2135 		}
2136 	}
2137 	AIO_UNLOCK(ki);
2138 
2139 	/*
2140 	 * Hack for failure of aio_aqueue.
2141 	 */
2142 	status = ops->fetch_status(ujob);
2143 	if (status == -1) {
2144 		td->td_retval[0] = ops->fetch_error(ujob);
2145 		return (0);
2146 	}
2147 
2148 	td->td_retval[0] = EINVAL;
2149 	return (0);
2150 }
2151 
2152 int
2153 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2154 {
2155 
2156 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2157 }
2158 
2159 /* syscall - asynchronous read from a file (REALTIME) */
2160 #ifdef COMPAT_FREEBSD6
2161 int
2162 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2163 {
2164 
2165 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2166 	    &aiocb_ops_osigevent));
2167 }
2168 #endif
2169 
2170 int
2171 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2172 {
2173 
2174 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2175 }
2176 
2177 int
2178 sys_aio_readv(struct thread *td, struct aio_readv_args *uap)
2179 {
2180 
2181 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops));
2182 }
2183 
2184 /* syscall - asynchronous write to a file (REALTIME) */
2185 #ifdef COMPAT_FREEBSD6
2186 int
2187 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2188 {
2189 
2190 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2191 	    &aiocb_ops_osigevent));
2192 }
2193 #endif
2194 
2195 int
2196 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2197 {
2198 
2199 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2200 }
2201 
2202 int
2203 sys_aio_writev(struct thread *td, struct aio_writev_args *uap)
2204 {
2205 
2206 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops));
2207 }
2208 
2209 int
2210 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2211 {
2212 
2213 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2214 }
2215 
2216 static int
2217 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2218     struct aiocb **acb_list, int nent, struct sigevent *sig,
2219     struct aiocb_ops *ops)
2220 {
2221 	struct proc *p = td->td_proc;
2222 	struct aiocb *job;
2223 	struct kaioinfo *ki;
2224 	struct aioliojob *lj;
2225 	struct kevent kev;
2226 	int error;
2227 	int nagain, nerror;
2228 	int i;
2229 
2230 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2231 		return (EINVAL);
2232 
2233 	if (nent < 0 || nent > max_aio_queue_per_proc)
2234 		return (EINVAL);
2235 
2236 	if (p->p_aioinfo == NULL)
2237 		aio_init_aioinfo(p);
2238 
2239 	ki = p->p_aioinfo;
2240 
2241 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2242 	lj->lioj_flags = 0;
2243 	lj->lioj_count = 0;
2244 	lj->lioj_finished_count = 0;
2245 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2246 	ksiginfo_init(&lj->lioj_ksi);
2247 
2248 	/*
2249 	 * Setup signal.
2250 	 */
2251 	if (sig && (mode == LIO_NOWAIT)) {
2252 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2253 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2254 			/* Assume only new style KEVENT */
2255 			memset(&kev, 0, sizeof(kev));
2256 			kev.filter = EVFILT_LIO;
2257 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2258 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2259 			kev.data = (intptr_t)lj;
2260 			/* pass user defined sigval data */
2261 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2262 			error = kqfd_register(
2263 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2264 			    M_WAITOK);
2265 			if (error) {
2266 				uma_zfree(aiolio_zone, lj);
2267 				return (error);
2268 			}
2269 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2270 			;
2271 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2272 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2273 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2274 					uma_zfree(aiolio_zone, lj);
2275 					return EINVAL;
2276 				}
2277 				lj->lioj_flags |= LIOJ_SIGNAL;
2278 		} else {
2279 			uma_zfree(aiolio_zone, lj);
2280 			return EINVAL;
2281 		}
2282 	}
2283 
2284 	AIO_LOCK(ki);
2285 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2286 	/*
2287 	 * Add extra aiocb count to avoid the lio to be freed
2288 	 * by other threads doing aio_waitcomplete or aio_return,
2289 	 * and prevent event from being sent until we have queued
2290 	 * all tasks.
2291 	 */
2292 	lj->lioj_count = 1;
2293 	AIO_UNLOCK(ki);
2294 
2295 	/*
2296 	 * Get pointers to the list of I/O requests.
2297 	 */
2298 	nagain = 0;
2299 	nerror = 0;
2300 	for (i = 0; i < nent; i++) {
2301 		job = acb_list[i];
2302 		if (job != NULL) {
2303 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2304 			if (error == EAGAIN)
2305 				nagain++;
2306 			else if (error != 0)
2307 				nerror++;
2308 		}
2309 	}
2310 
2311 	error = 0;
2312 	AIO_LOCK(ki);
2313 	if (mode == LIO_WAIT) {
2314 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2315 			ki->kaio_flags |= KAIO_WAKEUP;
2316 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2317 			    PRIBIO | PCATCH, "aiospn", 0);
2318 			if (error == ERESTART)
2319 				error = EINTR;
2320 			if (error)
2321 				break;
2322 		}
2323 	} else {
2324 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2325 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2326 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2327 				KNOTE_LOCKED(&lj->klist, 1);
2328 			}
2329 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2330 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2331 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2332 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2333 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2334 				    lj->lioj_count != 1);
2335 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2336 			}
2337 		}
2338 	}
2339 	lj->lioj_count--;
2340 	if (lj->lioj_count == 0) {
2341 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2342 		knlist_delete(&lj->klist, curthread, 1);
2343 		PROC_LOCK(p);
2344 		sigqueue_take(&lj->lioj_ksi);
2345 		PROC_UNLOCK(p);
2346 		AIO_UNLOCK(ki);
2347 		uma_zfree(aiolio_zone, lj);
2348 	} else
2349 		AIO_UNLOCK(ki);
2350 
2351 	if (nerror)
2352 		return (EIO);
2353 	else if (nagain)
2354 		return (EAGAIN);
2355 	else
2356 		return (error);
2357 }
2358 
2359 /* syscall - list directed I/O (REALTIME) */
2360 #ifdef COMPAT_FREEBSD6
2361 int
2362 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2363 {
2364 	struct aiocb **acb_list;
2365 	struct sigevent *sigp, sig;
2366 	struct osigevent osig;
2367 	int error, nent;
2368 
2369 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2370 		return (EINVAL);
2371 
2372 	nent = uap->nent;
2373 	if (nent < 0 || nent > max_aio_queue_per_proc)
2374 		return (EINVAL);
2375 
2376 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2377 		error = copyin(uap->sig, &osig, sizeof(osig));
2378 		if (error)
2379 			return (error);
2380 		error = convert_old_sigevent(&osig, &sig);
2381 		if (error)
2382 			return (error);
2383 		sigp = &sig;
2384 	} else
2385 		sigp = NULL;
2386 
2387 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2388 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2389 	if (error == 0)
2390 		error = kern_lio_listio(td, uap->mode,
2391 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2392 		    &aiocb_ops_osigevent);
2393 	free(acb_list, M_LIO);
2394 	return (error);
2395 }
2396 #endif
2397 
2398 /* syscall - list directed I/O (REALTIME) */
2399 int
2400 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2401 {
2402 	struct aiocb **acb_list;
2403 	struct sigevent *sigp, sig;
2404 	int error, nent;
2405 
2406 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2407 		return (EINVAL);
2408 
2409 	nent = uap->nent;
2410 	if (nent < 0 || nent > max_aio_queue_per_proc)
2411 		return (EINVAL);
2412 
2413 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2414 		error = copyin(uap->sig, &sig, sizeof(sig));
2415 		if (error)
2416 			return (error);
2417 		sigp = &sig;
2418 	} else
2419 		sigp = NULL;
2420 
2421 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2422 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2423 	if (error == 0)
2424 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2425 		    nent, sigp, &aiocb_ops);
2426 	free(acb_list, M_LIO);
2427 	return (error);
2428 }
2429 
2430 static void
2431 aio_biocleanup(struct bio *bp)
2432 {
2433 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2434 	struct kaioinfo *ki;
2435 	struct buf *pbuf = (struct buf *)bp->bio_caller2;
2436 
2437 	/* Release mapping into kernel space. */
2438 	if (pbuf != NULL) {
2439 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2440 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2441 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2442 		uma_zfree(pbuf_zone, pbuf);
2443 		atomic_subtract_int(&num_buf_aio, 1);
2444 		ki = job->userproc->p_aioinfo;
2445 		AIO_LOCK(ki);
2446 		ki->kaio_buffer_count--;
2447 		AIO_UNLOCK(ki);
2448 	} else {
2449 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2450 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2451 		free(bp->bio_ma, M_TEMP);
2452 		atomic_subtract_int(&num_unmapped_aio, 1);
2453 	}
2454 	g_destroy_bio(bp);
2455 }
2456 
2457 static void
2458 aio_biowakeup(struct bio *bp)
2459 {
2460 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2461 	size_t nbytes;
2462 	long bcount = bp->bio_bcount;
2463 	long resid = bp->bio_resid;
2464 	int error, opcode, nblks;
2465 	int bio_error = bp->bio_error;
2466 	uint16_t flags = bp->bio_flags;
2467 
2468 	opcode = job->uaiocb.aio_lio_opcode;
2469 
2470 	aio_biocleanup(bp);
2471 
2472 	nbytes =bcount - resid;
2473 	atomic_add_acq_long(&job->nbytes, nbytes);
2474 	nblks = btodb(nbytes);
2475 	error = 0;
2476 	/*
2477 	 * If multiple bios experienced an error, the job will reflect the
2478 	 * error of whichever failed bio completed last.
2479 	 */
2480 	if (flags & BIO_ERROR)
2481 		atomic_set_int(&job->error, bio_error);
2482 	if (opcode == LIO_WRITE || opcode == LIO_WRITEV)
2483 		atomic_add_int(&job->outblock, nblks);
2484 	else
2485 		atomic_add_int(&job->inblock, nblks);
2486 	atomic_subtract_int(&job->nbio, 1);
2487 
2488 
2489 	if (atomic_load_int(&job->nbio) == 0) {
2490 		if (atomic_load_int(&job->error))
2491 			aio_complete(job, -1, job->error);
2492 		else
2493 			aio_complete(job, atomic_load_long(&job->nbytes), 0);
2494 	}
2495 }
2496 
2497 /* syscall - wait for the next completion of an aio request */
2498 static int
2499 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2500     struct timespec *ts, struct aiocb_ops *ops)
2501 {
2502 	struct proc *p = td->td_proc;
2503 	struct timeval atv;
2504 	struct kaioinfo *ki;
2505 	struct kaiocb *job;
2506 	struct aiocb *ujob;
2507 	long error, status;
2508 	int timo;
2509 
2510 	ops->store_aiocb(ujobp, NULL);
2511 
2512 	if (ts == NULL) {
2513 		timo = 0;
2514 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2515 		timo = -1;
2516 	} else {
2517 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2518 			return (EINVAL);
2519 
2520 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2521 		if (itimerfix(&atv))
2522 			return (EINVAL);
2523 		timo = tvtohz(&atv);
2524 	}
2525 
2526 	if (p->p_aioinfo == NULL)
2527 		aio_init_aioinfo(p);
2528 	ki = p->p_aioinfo;
2529 
2530 	error = 0;
2531 	job = NULL;
2532 	AIO_LOCK(ki);
2533 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2534 		if (timo == -1) {
2535 			error = EWOULDBLOCK;
2536 			break;
2537 		}
2538 		ki->kaio_flags |= KAIO_WAKEUP;
2539 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2540 		    "aiowc", timo);
2541 		if (timo && error == ERESTART)
2542 			error = EINTR;
2543 		if (error)
2544 			break;
2545 	}
2546 
2547 	if (job != NULL) {
2548 		MPASS(job->jobflags & KAIOCB_FINISHED);
2549 		ujob = job->ujob;
2550 		status = job->uaiocb._aiocb_private.status;
2551 		error = job->uaiocb._aiocb_private.error;
2552 		td->td_retval[0] = status;
2553 		td->td_ru.ru_oublock += job->outblock;
2554 		td->td_ru.ru_inblock += job->inblock;
2555 		td->td_ru.ru_msgsnd += job->msgsnd;
2556 		td->td_ru.ru_msgrcv += job->msgrcv;
2557 		aio_free_entry(job);
2558 		AIO_UNLOCK(ki);
2559 		ops->store_aiocb(ujobp, ujob);
2560 		ops->store_error(ujob, error);
2561 		ops->store_status(ujob, status);
2562 	} else
2563 		AIO_UNLOCK(ki);
2564 
2565 	return (error);
2566 }
2567 
2568 int
2569 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2570 {
2571 	struct timespec ts, *tsp;
2572 	int error;
2573 
2574 	if (uap->timeout) {
2575 		/* Get timespec struct. */
2576 		error = copyin(uap->timeout, &ts, sizeof(ts));
2577 		if (error)
2578 			return (error);
2579 		tsp = &ts;
2580 	} else
2581 		tsp = NULL;
2582 
2583 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2584 }
2585 
2586 static int
2587 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2588     struct aiocb_ops *ops)
2589 {
2590 
2591 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2592 		return (EINVAL);
2593 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2594 }
2595 
2596 int
2597 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2598 {
2599 
2600 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2601 }
2602 
2603 /* kqueue attach function */
2604 static int
2605 filt_aioattach(struct knote *kn)
2606 {
2607 	struct kaiocb *job;
2608 
2609 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2610 
2611 	/*
2612 	 * The job pointer must be validated before using it, so
2613 	 * registration is restricted to the kernel; the user cannot
2614 	 * set EV_FLAG1.
2615 	 */
2616 	if ((kn->kn_flags & EV_FLAG1) == 0)
2617 		return (EPERM);
2618 	kn->kn_ptr.p_aio = job;
2619 	kn->kn_flags &= ~EV_FLAG1;
2620 
2621 	knlist_add(&job->klist, kn, 0);
2622 
2623 	return (0);
2624 }
2625 
2626 /* kqueue detach function */
2627 static void
2628 filt_aiodetach(struct knote *kn)
2629 {
2630 	struct knlist *knl;
2631 
2632 	knl = &kn->kn_ptr.p_aio->klist;
2633 	knl->kl_lock(knl->kl_lockarg);
2634 	if (!knlist_empty(knl))
2635 		knlist_remove(knl, kn, 1);
2636 	knl->kl_unlock(knl->kl_lockarg);
2637 }
2638 
2639 /* kqueue filter function */
2640 /*ARGSUSED*/
2641 static int
2642 filt_aio(struct knote *kn, long hint)
2643 {
2644 	struct kaiocb *job = kn->kn_ptr.p_aio;
2645 
2646 	kn->kn_data = job->uaiocb._aiocb_private.error;
2647 	if (!(job->jobflags & KAIOCB_FINISHED))
2648 		return (0);
2649 	kn->kn_flags |= EV_EOF;
2650 	return (1);
2651 }
2652 
2653 /* kqueue attach function */
2654 static int
2655 filt_lioattach(struct knote *kn)
2656 {
2657 	struct aioliojob *lj;
2658 
2659 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2660 
2661 	/*
2662 	 * The aioliojob pointer must be validated before using it, so
2663 	 * registration is restricted to the kernel; the user cannot
2664 	 * set EV_FLAG1.
2665 	 */
2666 	if ((kn->kn_flags & EV_FLAG1) == 0)
2667 		return (EPERM);
2668 	kn->kn_ptr.p_lio = lj;
2669 	kn->kn_flags &= ~EV_FLAG1;
2670 
2671 	knlist_add(&lj->klist, kn, 0);
2672 
2673 	return (0);
2674 }
2675 
2676 /* kqueue detach function */
2677 static void
2678 filt_liodetach(struct knote *kn)
2679 {
2680 	struct knlist *knl;
2681 
2682 	knl = &kn->kn_ptr.p_lio->klist;
2683 	knl->kl_lock(knl->kl_lockarg);
2684 	if (!knlist_empty(knl))
2685 		knlist_remove(knl, kn, 1);
2686 	knl->kl_unlock(knl->kl_lockarg);
2687 }
2688 
2689 /* kqueue filter function */
2690 /*ARGSUSED*/
2691 static int
2692 filt_lio(struct knote *kn, long hint)
2693 {
2694 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2695 
2696 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2697 }
2698 
2699 #ifdef COMPAT_FREEBSD32
2700 #include <sys/mount.h>
2701 #include <sys/socket.h>
2702 #include <compat/freebsd32/freebsd32.h>
2703 #include <compat/freebsd32/freebsd32_proto.h>
2704 #include <compat/freebsd32/freebsd32_signal.h>
2705 #include <compat/freebsd32/freebsd32_syscall.h>
2706 #include <compat/freebsd32/freebsd32_util.h>
2707 
2708 struct __aiocb_private32 {
2709 	int32_t	status;
2710 	int32_t	error;
2711 	uint32_t kernelinfo;
2712 };
2713 
2714 #ifdef COMPAT_FREEBSD6
2715 typedef struct oaiocb32 {
2716 	int	aio_fildes;		/* File descriptor */
2717 	uint64_t aio_offset __packed;	/* File offset for I/O */
2718 	uint32_t aio_buf;		/* I/O buffer in process space */
2719 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2720 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2721 	int	aio_lio_opcode;		/* LIO opcode */
2722 	int	aio_reqprio;		/* Request priority -- ignored */
2723 	struct	__aiocb_private32 _aiocb_private;
2724 } oaiocb32_t;
2725 #endif
2726 
2727 typedef struct aiocb32 {
2728 	int32_t	aio_fildes;		/* File descriptor */
2729 	uint64_t aio_offset __packed;	/* File offset for I/O */
2730 	uint32_t aio_buf;	/* I/O buffer in process space */
2731 	uint32_t aio_nbytes;	/* Number of bytes for I/O */
2732 	int	__spare__[2];
2733 	uint32_t __spare2__;
2734 	int	aio_lio_opcode;		/* LIO opcode */
2735 	int	aio_reqprio;		/* Request priority -- ignored */
2736 	struct	__aiocb_private32 _aiocb_private;
2737 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2738 } aiocb32_t;
2739 
2740 #ifdef COMPAT_FREEBSD6
2741 static int
2742 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2743 {
2744 
2745 	/*
2746 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2747 	 * supported by AIO with the old sigevent structure.
2748 	 */
2749 	CP(*osig, *nsig, sigev_notify);
2750 	switch (nsig->sigev_notify) {
2751 	case SIGEV_NONE:
2752 		break;
2753 	case SIGEV_SIGNAL:
2754 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2755 		break;
2756 	case SIGEV_KEVENT:
2757 		nsig->sigev_notify_kqueue =
2758 		    osig->__sigev_u.__sigev_notify_kqueue;
2759 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2760 		break;
2761 	default:
2762 		return (EINVAL);
2763 	}
2764 	return (0);
2765 }
2766 
2767 static int
2768 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
2769     int type __unused)
2770 {
2771 	struct oaiocb32 job32;
2772 	struct aiocb *kcb = &kjob->uaiocb;
2773 	int error;
2774 
2775 	bzero(kcb, sizeof(struct aiocb));
2776 	error = copyin(ujob, &job32, sizeof(job32));
2777 	if (error)
2778 		return (error);
2779 
2780 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
2781 
2782 	CP(job32, *kcb, aio_fildes);
2783 	CP(job32, *kcb, aio_offset);
2784 	PTRIN_CP(job32, *kcb, aio_buf);
2785 	CP(job32, *kcb, aio_nbytes);
2786 	CP(job32, *kcb, aio_lio_opcode);
2787 	CP(job32, *kcb, aio_reqprio);
2788 	CP(job32, *kcb, _aiocb_private.status);
2789 	CP(job32, *kcb, _aiocb_private.error);
2790 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2791 	return (convert_old_sigevent32(&job32.aio_sigevent,
2792 	    &kcb->aio_sigevent));
2793 }
2794 #endif
2795 
2796 static int
2797 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
2798 {
2799 	struct aiocb32 job32;
2800 	struct aiocb *kcb = &kjob->uaiocb;
2801 	struct iovec32 *iov32;
2802 	int error;
2803 
2804 	error = copyin(ujob, &job32, sizeof(job32));
2805 	if (error)
2806 		return (error);
2807 	CP(job32, *kcb, aio_fildes);
2808 	CP(job32, *kcb, aio_offset);
2809 	CP(job32, *kcb, aio_lio_opcode);
2810 	if (type == LIO_READV || type == LIO_WRITEV) {
2811 		iov32 = PTRIN(job32.aio_iov);
2812 		CP(job32, *kcb, aio_iovcnt);
2813 		/* malloc a uio and copy in the iovec */
2814 		error = freebsd32_copyinuio(iov32,
2815 		    kcb->aio_iovcnt, &kjob->uiop);
2816 		if (error)
2817 			return (error);
2818 	} else {
2819 		PTRIN_CP(job32, *kcb, aio_buf);
2820 		CP(job32, *kcb, aio_nbytes);
2821 	}
2822 	CP(job32, *kcb, aio_reqprio);
2823 	CP(job32, *kcb, _aiocb_private.status);
2824 	CP(job32, *kcb, _aiocb_private.error);
2825 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2826 	error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent);
2827 
2828 	return (error);
2829 }
2830 
2831 static long
2832 aiocb32_fetch_status(struct aiocb *ujob)
2833 {
2834 	struct aiocb32 *ujob32;
2835 
2836 	ujob32 = (struct aiocb32 *)ujob;
2837 	return (fuword32(&ujob32->_aiocb_private.status));
2838 }
2839 
2840 static long
2841 aiocb32_fetch_error(struct aiocb *ujob)
2842 {
2843 	struct aiocb32 *ujob32;
2844 
2845 	ujob32 = (struct aiocb32 *)ujob;
2846 	return (fuword32(&ujob32->_aiocb_private.error));
2847 }
2848 
2849 static int
2850 aiocb32_store_status(struct aiocb *ujob, long status)
2851 {
2852 	struct aiocb32 *ujob32;
2853 
2854 	ujob32 = (struct aiocb32 *)ujob;
2855 	return (suword32(&ujob32->_aiocb_private.status, status));
2856 }
2857 
2858 static int
2859 aiocb32_store_error(struct aiocb *ujob, long error)
2860 {
2861 	struct aiocb32 *ujob32;
2862 
2863 	ujob32 = (struct aiocb32 *)ujob;
2864 	return (suword32(&ujob32->_aiocb_private.error, error));
2865 }
2866 
2867 static int
2868 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2869 {
2870 	struct aiocb32 *ujob32;
2871 
2872 	ujob32 = (struct aiocb32 *)ujob;
2873 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2874 }
2875 
2876 static int
2877 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2878 {
2879 
2880 	return (suword32(ujobp, (long)ujob));
2881 }
2882 
2883 static struct aiocb_ops aiocb32_ops = {
2884 	.aio_copyin = aiocb32_copyin,
2885 	.fetch_status = aiocb32_fetch_status,
2886 	.fetch_error = aiocb32_fetch_error,
2887 	.store_status = aiocb32_store_status,
2888 	.store_error = aiocb32_store_error,
2889 	.store_kernelinfo = aiocb32_store_kernelinfo,
2890 	.store_aiocb = aiocb32_store_aiocb,
2891 };
2892 
2893 #ifdef COMPAT_FREEBSD6
2894 static struct aiocb_ops aiocb32_ops_osigevent = {
2895 	.aio_copyin = aiocb32_copyin_old_sigevent,
2896 	.fetch_status = aiocb32_fetch_status,
2897 	.fetch_error = aiocb32_fetch_error,
2898 	.store_status = aiocb32_store_status,
2899 	.store_error = aiocb32_store_error,
2900 	.store_kernelinfo = aiocb32_store_kernelinfo,
2901 	.store_aiocb = aiocb32_store_aiocb,
2902 };
2903 #endif
2904 
2905 int
2906 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2907 {
2908 
2909 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2910 }
2911 
2912 int
2913 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2914 {
2915 	struct timespec32 ts32;
2916 	struct timespec ts, *tsp;
2917 	struct aiocb **ujoblist;
2918 	uint32_t *ujoblist32;
2919 	int error, i;
2920 
2921 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2922 		return (EINVAL);
2923 
2924 	if (uap->timeout) {
2925 		/* Get timespec struct. */
2926 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2927 			return (error);
2928 		CP(ts32, ts, tv_sec);
2929 		CP(ts32, ts, tv_nsec);
2930 		tsp = &ts;
2931 	} else
2932 		tsp = NULL;
2933 
2934 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2935 	ujoblist32 = (uint32_t *)ujoblist;
2936 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2937 	    sizeof(ujoblist32[0]));
2938 	if (error == 0) {
2939 		for (i = uap->nent - 1; i >= 0; i--)
2940 			ujoblist[i] = PTRIN(ujoblist32[i]);
2941 
2942 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2943 	}
2944 	free(ujoblist, M_AIOS);
2945 	return (error);
2946 }
2947 
2948 int
2949 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2950 {
2951 
2952 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2953 }
2954 
2955 #ifdef COMPAT_FREEBSD6
2956 int
2957 freebsd6_freebsd32_aio_read(struct thread *td,
2958     struct freebsd6_freebsd32_aio_read_args *uap)
2959 {
2960 
2961 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2962 	    &aiocb32_ops_osigevent));
2963 }
2964 #endif
2965 
2966 int
2967 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2968 {
2969 
2970 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2971 	    &aiocb32_ops));
2972 }
2973 
2974 int
2975 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap)
2976 {
2977 
2978 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV,
2979 	    &aiocb32_ops));
2980 }
2981 
2982 #ifdef COMPAT_FREEBSD6
2983 int
2984 freebsd6_freebsd32_aio_write(struct thread *td,
2985     struct freebsd6_freebsd32_aio_write_args *uap)
2986 {
2987 
2988 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2989 	    &aiocb32_ops_osigevent));
2990 }
2991 #endif
2992 
2993 int
2994 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2995 {
2996 
2997 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2998 	    &aiocb32_ops));
2999 }
3000 
3001 int
3002 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap)
3003 {
3004 
3005 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV,
3006 	    &aiocb32_ops));
3007 }
3008 
3009 int
3010 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
3011 {
3012 
3013 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
3014 	    &aiocb32_ops));
3015 }
3016 
3017 int
3018 freebsd32_aio_waitcomplete(struct thread *td,
3019     struct freebsd32_aio_waitcomplete_args *uap)
3020 {
3021 	struct timespec32 ts32;
3022 	struct timespec ts, *tsp;
3023 	int error;
3024 
3025 	if (uap->timeout) {
3026 		/* Get timespec struct. */
3027 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
3028 		if (error)
3029 			return (error);
3030 		CP(ts32, ts, tv_sec);
3031 		CP(ts32, ts, tv_nsec);
3032 		tsp = &ts;
3033 	} else
3034 		tsp = NULL;
3035 
3036 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
3037 	    &aiocb32_ops));
3038 }
3039 
3040 int
3041 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
3042 {
3043 
3044 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
3045 	    &aiocb32_ops));
3046 }
3047 
3048 #ifdef COMPAT_FREEBSD6
3049 int
3050 freebsd6_freebsd32_lio_listio(struct thread *td,
3051     struct freebsd6_freebsd32_lio_listio_args *uap)
3052 {
3053 	struct aiocb **acb_list;
3054 	struct sigevent *sigp, sig;
3055 	struct osigevent32 osig;
3056 	uint32_t *acb_list32;
3057 	int error, i, nent;
3058 
3059 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3060 		return (EINVAL);
3061 
3062 	nent = uap->nent;
3063 	if (nent < 0 || nent > max_aio_queue_per_proc)
3064 		return (EINVAL);
3065 
3066 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3067 		error = copyin(uap->sig, &osig, sizeof(osig));
3068 		if (error)
3069 			return (error);
3070 		error = convert_old_sigevent32(&osig, &sig);
3071 		if (error)
3072 			return (error);
3073 		sigp = &sig;
3074 	} else
3075 		sigp = NULL;
3076 
3077 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3078 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3079 	if (error) {
3080 		free(acb_list32, M_LIO);
3081 		return (error);
3082 	}
3083 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3084 	for (i = 0; i < nent; i++)
3085 		acb_list[i] = PTRIN(acb_list32[i]);
3086 	free(acb_list32, M_LIO);
3087 
3088 	error = kern_lio_listio(td, uap->mode,
3089 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3090 	    &aiocb32_ops_osigevent);
3091 	free(acb_list, M_LIO);
3092 	return (error);
3093 }
3094 #endif
3095 
3096 int
3097 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3098 {
3099 	struct aiocb **acb_list;
3100 	struct sigevent *sigp, sig;
3101 	struct sigevent32 sig32;
3102 	uint32_t *acb_list32;
3103 	int error, i, nent;
3104 
3105 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3106 		return (EINVAL);
3107 
3108 	nent = uap->nent;
3109 	if (nent < 0 || nent > max_aio_queue_per_proc)
3110 		return (EINVAL);
3111 
3112 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3113 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3114 		if (error)
3115 			return (error);
3116 		error = convert_sigevent32(&sig32, &sig);
3117 		if (error)
3118 			return (error);
3119 		sigp = &sig;
3120 	} else
3121 		sigp = NULL;
3122 
3123 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3124 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3125 	if (error) {
3126 		free(acb_list32, M_LIO);
3127 		return (error);
3128 	}
3129 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3130 	for (i = 0; i < nent; i++)
3131 		acb_list[i] = PTRIN(acb_list32[i]);
3132 	free(acb_list32, M_LIO);
3133 
3134 	error = kern_lio_listio(td, uap->mode,
3135 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3136 	    &aiocb32_ops);
3137 	free(acb_list, M_LIO);
3138 	return (error);
3139 }
3140 
3141 #endif
3142