xref: /freebsd/sys/kern/vfs_aio.c (revision 783d3ff6d7fae619db8a7990b8a6387de0c677b5)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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/param.h>
24 #include <sys/systm.h>
25 #include <sys/malloc.h>
26 #include <sys/bio.h>
27 #include <sys/buf.h>
28 #include <sys/capsicum.h>
29 #include <sys/eventhandler.h>
30 #include <sys/sysproto.h>
31 #include <sys/filedesc.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/kthread.h>
35 #include <sys/fcntl.h>
36 #include <sys/file.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/unistd.h>
41 #include <sys/posix4.h>
42 #include <sys/proc.h>
43 #include <sys/resourcevar.h>
44 #include <sys/signalvar.h>
45 #include <sys/syscallsubr.h>
46 #include <sys/protosw.h>
47 #include <sys/rwlock.h>
48 #include <sys/sema.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/syscall.h>
52 #include <sys/sysctl.h>
53 #include <sys/syslog.h>
54 #include <sys/sx.h>
55 #include <sys/taskqueue.h>
56 #include <sys/vnode.h>
57 #include <sys/conf.h>
58 #include <sys/event.h>
59 #include <sys/mount.h>
60 #include <geom/geom.h>
61 
62 #include <machine/atomic.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_page.h>
66 #include <vm/vm_extern.h>
67 #include <vm/pmap.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vnode_pager.h>
71 #include <vm/uma.h>
72 #include <sys/aio.h>
73 
74 /*
75  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
76  * overflow. (XXX will be removed soon.)
77  */
78 static u_long jobrefid;
79 
80 /*
81  * Counter for aio_fsync.
82  */
83 static uint64_t jobseqno;
84 
85 #ifndef MAX_AIO_PER_PROC
86 #define MAX_AIO_PER_PROC	32
87 #endif
88 
89 #ifndef MAX_AIO_QUEUE_PER_PROC
90 #define MAX_AIO_QUEUE_PER_PROC	256
91 #endif
92 
93 #ifndef MAX_AIO_QUEUE
94 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
95 #endif
96 
97 #ifndef MAX_BUF_AIO
98 #define MAX_BUF_AIO		16
99 #endif
100 
101 FEATURE(aio, "Asynchronous I/O");
102 SYSCTL_DECL(_p1003_1b);
103 
104 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
105 static MALLOC_DEFINE(M_AIO, "aio", "structures for asynchronous I/O");
106 
107 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
108     "Async IO management");
109 
110 static int enable_aio_unsafe = 0;
111 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
112     "Permit asynchronous IO on all file types, not just known-safe types");
113 
114 static unsigned int unsafe_warningcnt = 1;
115 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
116     &unsafe_warningcnt, 0,
117     "Warnings that will be triggered upon failed IO requests on unsafe files");
118 
119 static int max_aio_procs = MAX_AIO_PROCS;
120 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
121     "Maximum number of kernel processes to use for handling async IO ");
122 
123 static int num_aio_procs = 0;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
125     "Number of presently active kernel processes for async IO");
126 
127 /*
128  * The code will adjust the actual number of AIO processes towards this
129  * number when it gets a chance.
130  */
131 static int target_aio_procs = TARGET_AIO_PROCS;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
133     0,
134     "Preferred number of ready kernel processes for async IO");
135 
136 static int max_queue_count = MAX_AIO_QUEUE;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
138     "Maximum number of aio requests to queue, globally");
139 
140 static int num_queue_count = 0;
141 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
142     "Number of queued aio requests");
143 
144 static int num_buf_aio = 0;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
146     "Number of aio requests presently handled by the buf subsystem");
147 
148 static int num_unmapped_aio = 0;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
150     0,
151     "Number of aio requests presently handled by unmapped I/O buffers");
152 
153 /* Number of async I/O processes in the process of being started */
154 /* XXX This should be local to aio_aqueue() */
155 static int num_aio_resv_start = 0;
156 
157 static int aiod_lifetime;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
159     "Maximum lifetime for idle aiod");
160 
161 static int max_aio_per_proc = MAX_AIO_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
163     0,
164     "Maximum active aio requests per process");
165 
166 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
168     &max_aio_queue_per_proc, 0,
169     "Maximum queued aio requests per process");
170 
171 static int max_buf_aio = MAX_BUF_AIO;
172 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
173     "Maximum buf aio requests per process");
174 
175 /*
176  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
177  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
178  * vfs.aio.aio_listio_max.
179  */
180 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
181     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
182     0, "Maximum aio requests for a single lio_listio call");
183 
184 #ifdef COMPAT_FREEBSD6
185 typedef struct oaiocb {
186 	int	aio_fildes;		/* File descriptor */
187 	off_t	aio_offset;		/* File offset for I/O */
188 	volatile void *aio_buf;         /* I/O buffer in process space */
189 	size_t	aio_nbytes;		/* Number of bytes for I/O */
190 	struct	osigevent aio_sigevent;	/* Signal to deliver */
191 	int	aio_lio_opcode;		/* LIO opcode */
192 	int	aio_reqprio;		/* Request priority -- ignored */
193 	struct	__aiocb_private	_aiocb_private;
194 } oaiocb_t;
195 #endif
196 
197 /*
198  * Below is a key of locks used to protect each member of struct kaiocb
199  * aioliojob and kaioinfo and any backends.
200  *
201  * * - need not protected
202  * a - locked by kaioinfo lock
203  * b - locked by backend lock, the backend lock can be null in some cases,
204  *     for example, BIO belongs to this type, in this case, proc lock is
205  *     reused.
206  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
207  */
208 
209 /*
210  * If the routine that services an AIO request blocks while running in an
211  * AIO kernel process it can starve other I/O requests.  BIO requests
212  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
213  * processes at all.  Socket I/O requests use a separate pool of
214  * kprocs and also force non-blocking I/O.  Other file I/O requests
215  * use the generic fo_read/fo_write operations which can block.  The
216  * fsync and mlock operations can also block while executing.  Ideally
217  * none of these requests would block while executing.
218  *
219  * Note that the service routines cannot toggle O_NONBLOCK in the file
220  * structure directly while handling a request due to races with
221  * userland threads.
222  */
223 
224 /* jobflags */
225 #define	KAIOCB_QUEUEING		0x01
226 #define	KAIOCB_CANCELLED	0x02
227 #define	KAIOCB_CANCELLING	0x04
228 #define	KAIOCB_CHECKSYNC	0x08
229 #define	KAIOCB_CLEARED		0x10
230 #define	KAIOCB_FINISHED		0x20
231 
232 /* ioflags */
233 #define	KAIOCB_IO_FOFFSET	0x01
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  *	aiocb	async io jobs
344  *	aiolio	list io jobs
345  */
346 static uma_zone_t kaio_zone, aiocb_zone, aiolio_zone;
347 
348 /* kqueue filters for aio */
349 static struct filterops aio_filtops = {
350 	.f_isfd = 0,
351 	.f_attach = filt_aioattach,
352 	.f_detach = filt_aiodetach,
353 	.f_event = filt_aio,
354 };
355 static struct filterops lio_filtops = {
356 	.f_isfd = 0,
357 	.f_attach = filt_lioattach,
358 	.f_detach = filt_liodetach,
359 	.f_event = filt_lio
360 };
361 
362 static eventhandler_tag exit_tag, exec_tag;
363 
364 TASKQUEUE_DEFINE_THREAD(aiod_kick);
365 
366 /*
367  * Main operations function for use as a kernel module.
368  */
369 static int
370 aio_modload(struct module *module, int cmd, void *arg)
371 {
372 	int error = 0;
373 
374 	switch (cmd) {
375 	case MOD_LOAD:
376 		aio_onceonly();
377 		break;
378 	case MOD_SHUTDOWN:
379 		break;
380 	default:
381 		error = EOPNOTSUPP;
382 		break;
383 	}
384 	return (error);
385 }
386 
387 static moduledata_t aio_mod = {
388 	"aio",
389 	&aio_modload,
390 	NULL
391 };
392 
393 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
394 MODULE_VERSION(aio, 1);
395 
396 /*
397  * Startup initialization
398  */
399 static int
400 aio_onceonly(void)
401 {
402 
403 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
404 	    EVENTHANDLER_PRI_ANY);
405 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
406 	    NULL, EVENTHANDLER_PRI_ANY);
407 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
408 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
409 	TAILQ_INIT(&aio_freeproc);
410 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
411 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
412 	TAILQ_INIT(&aio_jobs);
413 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
414 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
415 	    NULL, NULL, UMA_ALIGN_PTR, 0);
416 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
417 	    NULL, NULL, UMA_ALIGN_PTR, 0);
418 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
419 	    NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
420 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
421 	jobrefid = 1;
422 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
423 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
424 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
425 
426 	return (0);
427 }
428 
429 /*
430  * Init the per-process aioinfo structure.  The aioinfo limits are set
431  * per-process for user limit (resource) management.
432  */
433 void
434 aio_init_aioinfo(struct proc *p)
435 {
436 	struct kaioinfo *ki;
437 
438 	ki = uma_zalloc(kaio_zone, M_WAITOK);
439 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
440 	ki->kaio_flags = 0;
441 	ki->kaio_active_count = 0;
442 	ki->kaio_count = 0;
443 	ki->kaio_buffer_count = 0;
444 	TAILQ_INIT(&ki->kaio_all);
445 	TAILQ_INIT(&ki->kaio_done);
446 	TAILQ_INIT(&ki->kaio_jobqueue);
447 	TAILQ_INIT(&ki->kaio_liojoblist);
448 	TAILQ_INIT(&ki->kaio_syncqueue);
449 	TAILQ_INIT(&ki->kaio_syncready);
450 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
451 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
452 	PROC_LOCK(p);
453 	if (p->p_aioinfo == NULL) {
454 		p->p_aioinfo = ki;
455 		PROC_UNLOCK(p);
456 	} else {
457 		PROC_UNLOCK(p);
458 		mtx_destroy(&ki->kaio_mtx);
459 		uma_zfree(kaio_zone, ki);
460 	}
461 
462 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
463 		aio_newproc(NULL);
464 }
465 
466 static int
467 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
468 {
469 	struct thread *td;
470 	int error;
471 
472 	error = sigev_findtd(p, sigev, &td);
473 	if (error)
474 		return (error);
475 	if (!KSI_ONQ(ksi)) {
476 		ksiginfo_set_sigev(ksi, sigev);
477 		ksi->ksi_code = SI_ASYNCIO;
478 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
479 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
480 	}
481 	PROC_UNLOCK(p);
482 	return (error);
483 }
484 
485 /*
486  * Free a job entry.  Wait for completion if it is currently active, but don't
487  * delay forever.  If we delay, we return a flag that says that we have to
488  * restart the queue scan.
489  */
490 static int
491 aio_free_entry(struct kaiocb *job)
492 {
493 	struct kaioinfo *ki;
494 	struct aioliojob *lj;
495 	struct proc *p;
496 
497 	p = job->userproc;
498 	MPASS(curproc == p);
499 	ki = p->p_aioinfo;
500 	MPASS(ki != NULL);
501 
502 	AIO_LOCK_ASSERT(ki, MA_OWNED);
503 	MPASS(job->jobflags & KAIOCB_FINISHED);
504 
505 	atomic_subtract_int(&num_queue_count, 1);
506 
507 	ki->kaio_count--;
508 	MPASS(ki->kaio_count >= 0);
509 
510 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
511 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
512 
513 	lj = job->lio;
514 	if (lj) {
515 		lj->lioj_count--;
516 		lj->lioj_finished_count--;
517 
518 		if (lj->lioj_count == 0) {
519 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
520 			/* lio is going away, we need to destroy any knotes */
521 			knlist_delete(&lj->klist, curthread, 1);
522 			PROC_LOCK(p);
523 			sigqueue_take(&lj->lioj_ksi);
524 			PROC_UNLOCK(p);
525 			uma_zfree(aiolio_zone, lj);
526 		}
527 	}
528 
529 	/* job is going away, we need to destroy any knotes */
530 	knlist_delete(&job->klist, curthread, 1);
531 	PROC_LOCK(p);
532 	sigqueue_take(&job->ksi);
533 	PROC_UNLOCK(p);
534 
535 	AIO_UNLOCK(ki);
536 
537 	/*
538 	 * The thread argument here is used to find the owning process
539 	 * and is also passed to fo_close() which may pass it to various
540 	 * places such as devsw close() routines.  Because of that, we
541 	 * need a thread pointer from the process owning the job that is
542 	 * persistent and won't disappear out from under us or move to
543 	 * another process.
544 	 *
545 	 * Currently, all the callers of this function call it to remove
546 	 * a kaiocb from the current process' job list either via a
547 	 * syscall or due to the current process calling exit() or
548 	 * execve().  Thus, we know that p == curproc.  We also know that
549 	 * curthread can't exit since we are curthread.
550 	 *
551 	 * Therefore, we use curthread as the thread to pass to
552 	 * knlist_delete().  This does mean that it is possible for the
553 	 * thread pointer at close time to differ from the thread pointer
554 	 * at open time, but this is already true of file descriptors in
555 	 * a multithreaded process.
556 	 */
557 	if (job->fd_file)
558 		fdrop(job->fd_file, curthread);
559 	crfree(job->cred);
560 	if (job->uiop != &job->uio)
561 		freeuio(job->uiop);
562 	uma_zfree(aiocb_zone, job);
563 	AIO_LOCK(ki);
564 
565 	return (0);
566 }
567 
568 static void
569 aio_proc_rundown_exec(void *arg, struct proc *p,
570     struct image_params *imgp __unused)
571 {
572    	aio_proc_rundown(arg, p);
573 }
574 
575 static int
576 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
577 {
578 	aio_cancel_fn_t *func;
579 	int cancelled;
580 
581 	AIO_LOCK_ASSERT(ki, MA_OWNED);
582 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
583 		return (0);
584 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
585 	job->jobflags |= KAIOCB_CANCELLED;
586 
587 	func = job->cancel_fn;
588 
589 	/*
590 	 * If there is no cancel routine, just leave the job marked as
591 	 * cancelled.  The job should be in active use by a caller who
592 	 * should complete it normally or when it fails to install a
593 	 * cancel routine.
594 	 */
595 	if (func == NULL)
596 		return (0);
597 
598 	/*
599 	 * Set the CANCELLING flag so that aio_complete() will defer
600 	 * completions of this job.  This prevents the job from being
601 	 * freed out from under the cancel callback.  After the
602 	 * callback any deferred completion (whether from the callback
603 	 * or any other source) will be completed.
604 	 */
605 	job->jobflags |= KAIOCB_CANCELLING;
606 	AIO_UNLOCK(ki);
607 	func(job);
608 	AIO_LOCK(ki);
609 	job->jobflags &= ~KAIOCB_CANCELLING;
610 	if (job->jobflags & KAIOCB_FINISHED) {
611 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
612 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
613 		aio_bio_done_notify(p, job);
614 	} else {
615 		/*
616 		 * The cancel callback might have scheduled an
617 		 * operation to cancel this request, but it is
618 		 * only counted as cancelled if the request is
619 		 * cancelled when the callback returns.
620 		 */
621 		cancelled = 0;
622 	}
623 	return (cancelled);
624 }
625 
626 /*
627  * Rundown the jobs for a given process.
628  */
629 static void
630 aio_proc_rundown(void *arg, struct proc *p)
631 {
632 	struct kaioinfo *ki;
633 	struct aioliojob *lj;
634 	struct kaiocb *job, *jobn;
635 
636 	KASSERT(curthread->td_proc == p,
637 	    ("%s: called on non-curproc", __func__));
638 	ki = p->p_aioinfo;
639 	if (ki == NULL)
640 		return;
641 
642 	AIO_LOCK(ki);
643 	ki->kaio_flags |= KAIO_RUNDOWN;
644 
645 restart:
646 
647 	/*
648 	 * Try to cancel all pending requests. This code simulates
649 	 * aio_cancel on all pending I/O requests.
650 	 */
651 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
652 		aio_cancel_job(p, ki, job);
653 	}
654 
655 	/* Wait for all running I/O to be finished */
656 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
657 		ki->kaio_flags |= KAIO_WAKEUP;
658 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
659 		goto restart;
660 	}
661 
662 	/* Free all completed I/O requests. */
663 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
664 		aio_free_entry(job);
665 
666 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
667 		if (lj->lioj_count == 0) {
668 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
669 			knlist_delete(&lj->klist, curthread, 1);
670 			PROC_LOCK(p);
671 			sigqueue_take(&lj->lioj_ksi);
672 			PROC_UNLOCK(p);
673 			uma_zfree(aiolio_zone, lj);
674 		} else {
675 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
676 			    lj->lioj_count, lj->lioj_finished_count);
677 		}
678 	}
679 	AIO_UNLOCK(ki);
680 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
681 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
682 	mtx_destroy(&ki->kaio_mtx);
683 	uma_zfree(kaio_zone, ki);
684 	p->p_aioinfo = NULL;
685 }
686 
687 /*
688  * Select a job to run (called by an AIO daemon).
689  */
690 static struct kaiocb *
691 aio_selectjob(struct aioproc *aiop)
692 {
693 	struct kaiocb *job;
694 	struct kaioinfo *ki;
695 	struct proc *userp;
696 
697 	mtx_assert(&aio_job_mtx, MA_OWNED);
698 restart:
699 	TAILQ_FOREACH(job, &aio_jobs, list) {
700 		userp = job->userproc;
701 		ki = userp->p_aioinfo;
702 
703 		if (ki->kaio_active_count < max_aio_per_proc) {
704 			TAILQ_REMOVE(&aio_jobs, job, list);
705 			if (!aio_clear_cancel_function(job))
706 				goto restart;
707 
708 			/* Account for currently active jobs. */
709 			ki->kaio_active_count++;
710 			break;
711 		}
712 	}
713 	return (job);
714 }
715 
716 /*
717  * Move all data to a permanent storage device.  This code
718  * simulates the fsync and fdatasync syscalls.
719  */
720 static int
721 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op)
722 {
723 	struct mount *mp;
724 	int error;
725 
726 	for (;;) {
727 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
728 		if (error != 0)
729 			break;
730 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
731 		vnode_pager_clean_async(vp);
732 		if (op == LIO_DSYNC)
733 			error = VOP_FDATASYNC(vp, td);
734 		else
735 			error = VOP_FSYNC(vp, MNT_WAIT, td);
736 
737 		VOP_UNLOCK(vp);
738 		vn_finished_write(mp);
739 		if (error != ERELOOKUP)
740 			break;
741 	}
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 file *fp;
759 	ssize_t cnt;
760 	long msgsnd_st, msgsnd_end;
761 	long msgrcv_st, msgrcv_end;
762 	long oublock_st, oublock_end;
763 	long inblock_st, inblock_end;
764 	int error, opcode;
765 
766 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
767 	    job->uaiocb.aio_lio_opcode == LIO_READV ||
768 	    job->uaiocb.aio_lio_opcode == LIO_WRITE ||
769 	    job->uaiocb.aio_lio_opcode == LIO_WRITEV,
770 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
771 
772 	aio_switch_vmspace(job);
773 	td = curthread;
774 	td_savedcred = td->td_ucred;
775 	td->td_ucred = job->cred;
776 	job->uiop->uio_td = td;
777 	fp = job->fd_file;
778 
779 	opcode = job->uaiocb.aio_lio_opcode;
780 	cnt = job->uiop->uio_resid;
781 
782 	msgrcv_st = td->td_ru.ru_msgrcv;
783 	msgsnd_st = td->td_ru.ru_msgsnd;
784 	inblock_st = td->td_ru.ru_inblock;
785 	oublock_st = td->td_ru.ru_oublock;
786 
787 	/*
788 	 * aio_aqueue() acquires a reference to the file that is
789 	 * released in aio_free_entry().
790 	 */
791 	if (opcode == LIO_READ || opcode == LIO_READV) {
792 		if (job->uiop->uio_resid == 0)
793 			error = 0;
794 		else
795 			error = fo_read(fp, job->uiop, fp->f_cred,
796 			    (job->ioflags & KAIOCB_IO_FOFFSET) != 0 ? 0 :
797 			    FOF_OFFSET, td);
798 	} else {
799 		if (fp->f_type == DTYPE_VNODE)
800 			bwillwrite();
801 		error = fo_write(fp, job->uiop, fp->f_cred, (job->ioflags &
802 		    KAIOCB_IO_FOFFSET) != 0 ? 0 : 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 && (opcode & LIO_WRITE)) {
818 			PROC_LOCK(job->userproc);
819 			kern_psignal(job->userproc, SIGPIPE);
820 			PROC_UNLOCK(job->userproc);
821 		}
822 	}
823 
824 	cnt -= job->uiop->uio_resid;
825 	td->td_ucred = td_savedcred;
826 	if (error)
827 		aio_complete(job, -1, error);
828 	else
829 		aio_complete(job, cnt, 0);
830 }
831 
832 static void
833 aio_process_sync(struct kaiocb *job)
834 {
835 	struct thread *td = curthread;
836 	struct ucred *td_savedcred = td->td_ucred;
837 	struct file *fp = job->fd_file;
838 	int error = 0;
839 
840 	KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC,
841 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
842 
843 	td->td_ucred = job->cred;
844 	if (fp->f_vnode != NULL) {
845 		error = aio_fsync_vnode(td, fp->f_vnode,
846 		    job->uaiocb.aio_lio_opcode);
847 	}
848 	td->td_ucred = td_savedcred;
849 	if (error)
850 		aio_complete(job, -1, error);
851 	else
852 		aio_complete(job, 0, 0);
853 }
854 
855 static void
856 aio_process_mlock(struct kaiocb *job)
857 {
858 	struct aiocb *cb = &job->uaiocb;
859 	int error;
860 
861 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
862 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
863 
864 	aio_switch_vmspace(job);
865 	error = kern_mlock(job->userproc, job->cred,
866 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
867 	aio_complete(job, error != 0 ? -1 : 0, error);
868 }
869 
870 static void
871 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
872 {
873 	struct aioliojob *lj;
874 	struct kaioinfo *ki;
875 	struct kaiocb *sjob, *sjobn;
876 	int lj_done;
877 	bool schedule_fsync;
878 
879 	ki = userp->p_aioinfo;
880 	AIO_LOCK_ASSERT(ki, MA_OWNED);
881 	lj = job->lio;
882 	lj_done = 0;
883 	if (lj) {
884 		lj->lioj_finished_count++;
885 		if (lj->lioj_count == lj->lioj_finished_count)
886 			lj_done = 1;
887 	}
888 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
889 	MPASS(job->jobflags & KAIOCB_FINISHED);
890 
891 	if (ki->kaio_flags & KAIO_RUNDOWN)
892 		goto notification_done;
893 
894 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
895 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
896 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
897 
898 	KNOTE_LOCKED(&job->klist, 1);
899 
900 	if (lj_done) {
901 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
902 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
903 			KNOTE_LOCKED(&lj->klist, 1);
904 		}
905 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
906 		    == LIOJ_SIGNAL &&
907 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
908 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
909 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
910 			    true);
911 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
912 		}
913 	}
914 
915 notification_done:
916 	if (job->jobflags & KAIOCB_CHECKSYNC) {
917 		schedule_fsync = false;
918 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
919 			if (job->fd_file != sjob->fd_file ||
920 			    job->seqno >= sjob->seqno)
921 				continue;
922 			if (--sjob->pending > 0)
923 				continue;
924 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
925 			if (!aio_clear_cancel_function_locked(sjob))
926 				continue;
927 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
928 			schedule_fsync = true;
929 		}
930 		if (schedule_fsync)
931 			taskqueue_enqueue(taskqueue_aiod_kick,
932 			    &ki->kaio_sync_task);
933 	}
934 	if (ki->kaio_flags & KAIO_WAKEUP) {
935 		ki->kaio_flags &= ~KAIO_WAKEUP;
936 		wakeup(&userp->p_aioinfo);
937 	}
938 }
939 
940 static void
941 aio_schedule_fsync(void *context, int pending)
942 {
943 	struct kaioinfo *ki;
944 	struct kaiocb *job;
945 
946 	ki = context;
947 	AIO_LOCK(ki);
948 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
949 		job = TAILQ_FIRST(&ki->kaio_syncready);
950 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
951 		AIO_UNLOCK(ki);
952 		aio_schedule(job, aio_process_sync);
953 		AIO_LOCK(ki);
954 	}
955 	AIO_UNLOCK(ki);
956 }
957 
958 bool
959 aio_cancel_cleared(struct kaiocb *job)
960 {
961 
962 	/*
963 	 * The caller should hold the same queue lock held when
964 	 * aio_clear_cancel_function() was called and set this flag
965 	 * ensuring this check sees an up-to-date value.  However,
966 	 * there is no way to assert that.
967 	 */
968 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
969 }
970 
971 static bool
972 aio_clear_cancel_function_locked(struct kaiocb *job)
973 {
974 
975 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
976 	MPASS(job->cancel_fn != NULL);
977 	if (job->jobflags & KAIOCB_CANCELLING) {
978 		job->jobflags |= KAIOCB_CLEARED;
979 		return (false);
980 	}
981 	job->cancel_fn = NULL;
982 	return (true);
983 }
984 
985 bool
986 aio_clear_cancel_function(struct kaiocb *job)
987 {
988 	struct kaioinfo *ki;
989 	bool ret;
990 
991 	ki = job->userproc->p_aioinfo;
992 	AIO_LOCK(ki);
993 	ret = aio_clear_cancel_function_locked(job);
994 	AIO_UNLOCK(ki);
995 	return (ret);
996 }
997 
998 static bool
999 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1000 {
1001 
1002 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1003 	if (job->jobflags & KAIOCB_CANCELLED)
1004 		return (false);
1005 	job->cancel_fn = func;
1006 	return (true);
1007 }
1008 
1009 bool
1010 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1011 {
1012 	struct kaioinfo *ki;
1013 	bool ret;
1014 
1015 	ki = job->userproc->p_aioinfo;
1016 	AIO_LOCK(ki);
1017 	ret = aio_set_cancel_function_locked(job, func);
1018 	AIO_UNLOCK(ki);
1019 	return (ret);
1020 }
1021 
1022 void
1023 aio_complete(struct kaiocb *job, long status, int error)
1024 {
1025 	struct kaioinfo *ki;
1026 	struct proc *userp;
1027 
1028 	job->uaiocb._aiocb_private.error = error;
1029 	job->uaiocb._aiocb_private.status = status;
1030 
1031 	userp = job->userproc;
1032 	ki = userp->p_aioinfo;
1033 
1034 	AIO_LOCK(ki);
1035 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1036 	    ("duplicate aio_complete"));
1037 	job->jobflags |= KAIOCB_FINISHED;
1038 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1039 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1040 		aio_bio_done_notify(userp, job);
1041 	}
1042 	AIO_UNLOCK(ki);
1043 }
1044 
1045 void
1046 aio_cancel(struct kaiocb *job)
1047 {
1048 
1049 	aio_complete(job, -1, ECANCELED);
1050 }
1051 
1052 void
1053 aio_switch_vmspace(struct kaiocb *job)
1054 {
1055 
1056 	vmspace_switch_aio(job->userproc->p_vmspace);
1057 }
1058 
1059 /*
1060  * The AIO daemon, most of the actual work is done in aio_process_*,
1061  * but the setup (and address space mgmt) is done in this routine.
1062  */
1063 static void
1064 aio_daemon(void *_id)
1065 {
1066 	struct kaiocb *job;
1067 	struct aioproc *aiop;
1068 	struct kaioinfo *ki;
1069 	struct proc *p;
1070 	struct vmspace *myvm;
1071 	struct thread *td = curthread;
1072 	int id = (intptr_t)_id;
1073 
1074 	/*
1075 	 * Grab an extra reference on the daemon's vmspace so that it
1076 	 * doesn't get freed by jobs that switch to a different
1077 	 * vmspace.
1078 	 */
1079 	p = td->td_proc;
1080 	myvm = vmspace_acquire_ref(p);
1081 
1082 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1083 
1084 	/*
1085 	 * Allocate and ready the aio control info.  There is one aiop structure
1086 	 * per daemon.
1087 	 */
1088 	aiop = malloc(sizeof(*aiop), M_AIO, M_WAITOK);
1089 	aiop->aioproc = p;
1090 	aiop->aioprocflags = 0;
1091 
1092 	/*
1093 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1094 	 * and creating too many daemons.)
1095 	 */
1096 	sema_post(&aio_newproc_sem);
1097 
1098 	mtx_lock(&aio_job_mtx);
1099 	for (;;) {
1100 		/*
1101 		 * Take daemon off of free queue
1102 		 */
1103 		if (aiop->aioprocflags & AIOP_FREE) {
1104 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1105 			aiop->aioprocflags &= ~AIOP_FREE;
1106 		}
1107 
1108 		/*
1109 		 * Check for jobs.
1110 		 */
1111 		while ((job = aio_selectjob(aiop)) != NULL) {
1112 			mtx_unlock(&aio_job_mtx);
1113 
1114 			ki = job->userproc->p_aioinfo;
1115 			job->handle_fn(job);
1116 
1117 			mtx_lock(&aio_job_mtx);
1118 			/* Decrement the active job count. */
1119 			ki->kaio_active_count--;
1120 		}
1121 
1122 		/*
1123 		 * Disconnect from user address space.
1124 		 */
1125 		if (p->p_vmspace != myvm) {
1126 			mtx_unlock(&aio_job_mtx);
1127 			vmspace_switch_aio(myvm);
1128 			mtx_lock(&aio_job_mtx);
1129 			/*
1130 			 * We have to restart to avoid race, we only sleep if
1131 			 * no job can be selected.
1132 			 */
1133 			continue;
1134 		}
1135 
1136 		mtx_assert(&aio_job_mtx, MA_OWNED);
1137 
1138 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1139 		aiop->aioprocflags |= AIOP_FREE;
1140 
1141 		/*
1142 		 * If daemon is inactive for a long time, allow it to exit,
1143 		 * thereby freeing resources.
1144 		 */
1145 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1146 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1147 		    (aiop->aioprocflags & AIOP_FREE) &&
1148 		    num_aio_procs > target_aio_procs)
1149 			break;
1150 	}
1151 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1152 	num_aio_procs--;
1153 	mtx_unlock(&aio_job_mtx);
1154 	free(aiop, M_AIO);
1155 	free_unr(aiod_unr, id);
1156 	vmspace_free(myvm);
1157 
1158 	KASSERT(p->p_vmspace == myvm,
1159 	    ("AIOD: bad vmspace for exiting daemon"));
1160 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1161 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1162 	    refcount_load(&myvm->vm_refcnt)));
1163 	kproc_exit(0);
1164 }
1165 
1166 /*
1167  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1168  * AIO daemon modifies its environment itself.
1169  */
1170 static int
1171 aio_newproc(int *start)
1172 {
1173 	int error;
1174 	struct proc *p;
1175 	int id;
1176 
1177 	id = alloc_unr(aiod_unr);
1178 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1179 		RFNOWAIT, 0, "aiod%d", id);
1180 	if (error == 0) {
1181 		/*
1182 		 * Wait until daemon is started.
1183 		 */
1184 		sema_wait(&aio_newproc_sem);
1185 		mtx_lock(&aio_job_mtx);
1186 		num_aio_procs++;
1187 		if (start != NULL)
1188 			(*start)--;
1189 		mtx_unlock(&aio_job_mtx);
1190 	} else {
1191 		free_unr(aiod_unr, id);
1192 	}
1193 	return (error);
1194 }
1195 
1196 /*
1197  * Try the high-performance, low-overhead bio method for eligible
1198  * VCHR devices.  This method doesn't use an aio helper thread, and
1199  * thus has very low overhead.
1200  *
1201  * Assumes that the caller, aio_aqueue(), has incremented the file
1202  * structure's reference count, preventing its deallocation for the
1203  * duration of this call.
1204  */
1205 static int
1206 aio_qbio(struct proc *p, struct kaiocb *job)
1207 {
1208 	struct aiocb *cb;
1209 	struct file *fp;
1210 	struct buf *pbuf;
1211 	struct vnode *vp;
1212 	struct cdevsw *csw;
1213 	struct cdev *dev;
1214 	struct kaioinfo *ki;
1215 	struct bio **bios = NULL;
1216 	off_t offset;
1217 	int bio_cmd, error, i, iovcnt, opcode, poff, ref;
1218 	vm_prot_t prot;
1219 	bool use_unmapped;
1220 
1221 	cb = &job->uaiocb;
1222 	fp = job->fd_file;
1223 	opcode = cb->aio_lio_opcode;
1224 
1225 	if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV ||
1226 	    opcode == LIO_READ || opcode == LIO_READV))
1227 		return (-1);
1228 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1229 		return (-1);
1230 
1231 	vp = fp->f_vnode;
1232 	if (vp->v_type != VCHR)
1233 		return (-1);
1234 	if (vp->v_bufobj.bo_bsize == 0)
1235 		return (-1);
1236 
1237 	bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : 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 	refcount_init(&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_NOP)
1420 		type = kcb->aio_lio_opcode;
1421 	if (type & LIO_VECTORED) {
1422 		/* malloc a uio and copy in the iovec */
1423 		error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov),
1424 		    kcb->aio_iovcnt, &kjob->uiop);
1425 	}
1426 
1427 	return (error);
1428 }
1429 
1430 static long
1431 aiocb_fetch_status(struct aiocb *ujob)
1432 {
1433 
1434 	return (fuword(&ujob->_aiocb_private.status));
1435 }
1436 
1437 static long
1438 aiocb_fetch_error(struct aiocb *ujob)
1439 {
1440 
1441 	return (fuword(&ujob->_aiocb_private.error));
1442 }
1443 
1444 static int
1445 aiocb_store_status(struct aiocb *ujob, long status)
1446 {
1447 
1448 	return (suword(&ujob->_aiocb_private.status, status));
1449 }
1450 
1451 static int
1452 aiocb_store_error(struct aiocb *ujob, long error)
1453 {
1454 
1455 	return (suword(&ujob->_aiocb_private.error, error));
1456 }
1457 
1458 static int
1459 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1460 {
1461 
1462 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1463 }
1464 
1465 static int
1466 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1467 {
1468 
1469 	return (suword(ujobp, (long)ujob));
1470 }
1471 
1472 static struct aiocb_ops aiocb_ops = {
1473 	.aio_copyin = aiocb_copyin,
1474 	.fetch_status = aiocb_fetch_status,
1475 	.fetch_error = aiocb_fetch_error,
1476 	.store_status = aiocb_store_status,
1477 	.store_error = aiocb_store_error,
1478 	.store_kernelinfo = aiocb_store_kernelinfo,
1479 	.store_aiocb = aiocb_store_aiocb,
1480 };
1481 
1482 #ifdef COMPAT_FREEBSD6
1483 static struct aiocb_ops aiocb_ops_osigevent = {
1484 	.aio_copyin = aiocb_copyin_old_sigevent,
1485 	.fetch_status = aiocb_fetch_status,
1486 	.fetch_error = aiocb_fetch_error,
1487 	.store_status = aiocb_store_status,
1488 	.store_error = aiocb_store_error,
1489 	.store_kernelinfo = aiocb_store_kernelinfo,
1490 	.store_aiocb = aiocb_store_aiocb,
1491 };
1492 #endif
1493 
1494 /*
1495  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1496  * technique is done in this code.
1497  */
1498 int
1499 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1500     int type, struct aiocb_ops *ops)
1501 {
1502 	struct proc *p = td->td_proc;
1503 	struct file *fp = NULL;
1504 	struct kaiocb *job;
1505 	struct kaioinfo *ki;
1506 	struct kevent kev;
1507 	int opcode;
1508 	int error;
1509 	int fd, kqfd;
1510 	int jid;
1511 	u_short evflags;
1512 
1513 	if (p->p_aioinfo == NULL)
1514 		aio_init_aioinfo(p);
1515 
1516 	ki = p->p_aioinfo;
1517 
1518 	ops->store_status(ujob, -1);
1519 	ops->store_error(ujob, 0);
1520 	ops->store_kernelinfo(ujob, -1);
1521 
1522 	if (num_queue_count >= max_queue_count ||
1523 	    ki->kaio_count >= max_aio_queue_per_proc) {
1524 		error = EAGAIN;
1525 		goto err1;
1526 	}
1527 
1528 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1529 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1530 
1531 	error = ops->aio_copyin(ujob, job, type);
1532 	if (error)
1533 		goto err2;
1534 
1535 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1536 		error = EINVAL;
1537 		goto err2;
1538 	}
1539 
1540 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1541 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1542 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1543 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1544 		error = EINVAL;
1545 		goto err2;
1546 	}
1547 
1548 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1549 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1550 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1551 		error = EINVAL;
1552 		goto err2;
1553 	}
1554 
1555 	/* Get the opcode. */
1556 	if (type == LIO_NOP) {
1557 		switch (job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET) {
1558 		case LIO_WRITE:
1559 		case LIO_WRITEV:
1560 		case LIO_NOP:
1561 		case LIO_READ:
1562 		case LIO_READV:
1563 			opcode = job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET;
1564 			if ((job->uaiocb.aio_lio_opcode & LIO_FOFFSET) != 0)
1565 				job->ioflags |= KAIOCB_IO_FOFFSET;
1566 			break;
1567 		default:
1568 			error = EINVAL;
1569 			goto err2;
1570 		}
1571 	} else
1572 		opcode = job->uaiocb.aio_lio_opcode = type;
1573 
1574 	ksiginfo_init(&job->ksi);
1575 
1576 	/* Save userspace address of the job info. */
1577 	job->ujob = ujob;
1578 
1579 	/*
1580 	 * Validate the opcode and fetch the file object for the specified
1581 	 * file descriptor.
1582 	 *
1583 	 * XXXRW: Moved the opcode validation up here so that we don't
1584 	 * retrieve a file descriptor without knowing what the capabiltity
1585 	 * should be.
1586 	 */
1587 	fd = job->uaiocb.aio_fildes;
1588 	switch (opcode) {
1589 	case LIO_WRITE:
1590 	case LIO_WRITEV:
1591 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1592 		break;
1593 	case LIO_READ:
1594 	case LIO_READV:
1595 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1596 		break;
1597 	case LIO_SYNC:
1598 	case LIO_DSYNC:
1599 		error = fget(td, fd, &cap_fsync_rights, &fp);
1600 		break;
1601 	case LIO_MLOCK:
1602 		break;
1603 	case LIO_NOP:
1604 		error = fget(td, fd, &cap_no_rights, &fp);
1605 		break;
1606 	default:
1607 		error = EINVAL;
1608 	}
1609 	if (error)
1610 		goto err3;
1611 
1612 	if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) {
1613 		error = EINVAL;
1614 		goto err3;
1615 	}
1616 
1617 	if ((opcode == LIO_READ || opcode == LIO_READV ||
1618 	    opcode == LIO_WRITE || opcode == LIO_WRITEV) &&
1619 	    job->uaiocb.aio_offset < 0 &&
1620 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1621 		error = EINVAL;
1622 		goto err3;
1623 	}
1624 
1625 	if (fp != NULL && fp->f_ops == &path_fileops) {
1626 		error = EBADF;
1627 		goto err3;
1628 	}
1629 
1630 	job->fd_file = fp;
1631 
1632 	mtx_lock(&aio_job_mtx);
1633 	jid = jobrefid++;
1634 	job->seqno = jobseqno++;
1635 	mtx_unlock(&aio_job_mtx);
1636 	error = ops->store_kernelinfo(ujob, jid);
1637 	if (error) {
1638 		error = EINVAL;
1639 		goto err3;
1640 	}
1641 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1642 
1643 	if (opcode == LIO_NOP) {
1644 		fdrop(fp, td);
1645 		MPASS(job->uiop == &job->uio || job->uiop == NULL);
1646 		uma_zfree(aiocb_zone, job);
1647 		return (0);
1648 	}
1649 
1650 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1651 		goto no_kqueue;
1652 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1653 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1654 		error = EINVAL;
1655 		goto err3;
1656 	}
1657 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1658 	memset(&kev, 0, sizeof(kev));
1659 	kev.ident = (uintptr_t)job->ujob;
1660 	kev.filter = EVFILT_AIO;
1661 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1662 	kev.data = (intptr_t)job;
1663 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1664 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1665 	if (error)
1666 		goto err3;
1667 
1668 no_kqueue:
1669 
1670 	ops->store_error(ujob, EINPROGRESS);
1671 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1672 	job->userproc = p;
1673 	job->cred = crhold(td->td_ucred);
1674 	job->jobflags = KAIOCB_QUEUEING;
1675 	job->lio = lj;
1676 
1677 	if (opcode & LIO_VECTORED) {
1678 		/* Use the uio copied in by aio_copyin */
1679 		MPASS(job->uiop != &job->uio && job->uiop != NULL);
1680 	} else {
1681 		/* Setup the inline uio */
1682 		job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf;
1683 		job->iov[0].iov_len = job->uaiocb.aio_nbytes;
1684 		job->uio.uio_iov = job->iov;
1685 		job->uio.uio_iovcnt = 1;
1686 		job->uio.uio_resid = job->uaiocb.aio_nbytes;
1687 		job->uio.uio_segflg = UIO_USERSPACE;
1688 		job->uiop = &job->uio;
1689 	}
1690 	switch (opcode & (LIO_READ | LIO_WRITE)) {
1691 	case LIO_READ:
1692 		job->uiop->uio_rw = UIO_READ;
1693 		break;
1694 	case LIO_WRITE:
1695 		job->uiop->uio_rw = UIO_WRITE;
1696 		break;
1697 	}
1698 	job->uiop->uio_offset = job->uaiocb.aio_offset;
1699 	job->uiop->uio_td = td;
1700 
1701 	if (opcode == LIO_MLOCK) {
1702 		aio_schedule(job, aio_process_mlock);
1703 		error = 0;
1704 	} else if (fp->f_ops->fo_aio_queue == NULL)
1705 		error = aio_queue_file(fp, job);
1706 	else
1707 		error = fo_aio_queue(fp, job);
1708 	if (error)
1709 		goto err4;
1710 
1711 	AIO_LOCK(ki);
1712 	job->jobflags &= ~KAIOCB_QUEUEING;
1713 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1714 	ki->kaio_count++;
1715 	if (lj)
1716 		lj->lioj_count++;
1717 	atomic_add_int(&num_queue_count, 1);
1718 	if (job->jobflags & KAIOCB_FINISHED) {
1719 		/*
1720 		 * The queue callback completed the request synchronously.
1721 		 * The bulk of the completion is deferred in that case
1722 		 * until this point.
1723 		 */
1724 		aio_bio_done_notify(p, job);
1725 	} else
1726 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1727 	AIO_UNLOCK(ki);
1728 	return (0);
1729 
1730 err4:
1731 	crfree(job->cred);
1732 err3:
1733 	if (fp)
1734 		fdrop(fp, td);
1735 	knlist_delete(&job->klist, curthread, 0);
1736 err2:
1737 	if (job->uiop != &job->uio)
1738 		freeuio(job->uiop);
1739 	uma_zfree(aiocb_zone, job);
1740 err1:
1741 	ops->store_error(ujob, error);
1742 	return (error);
1743 }
1744 
1745 static void
1746 aio_cancel_daemon_job(struct kaiocb *job)
1747 {
1748 
1749 	mtx_lock(&aio_job_mtx);
1750 	if (!aio_cancel_cleared(job))
1751 		TAILQ_REMOVE(&aio_jobs, job, list);
1752 	mtx_unlock(&aio_job_mtx);
1753 	aio_cancel(job);
1754 }
1755 
1756 void
1757 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1758 {
1759 
1760 	mtx_lock(&aio_job_mtx);
1761 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1762 		mtx_unlock(&aio_job_mtx);
1763 		aio_cancel(job);
1764 		return;
1765 	}
1766 	job->handle_fn = func;
1767 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1768 	aio_kick_nowait(job->userproc);
1769 	mtx_unlock(&aio_job_mtx);
1770 }
1771 
1772 static void
1773 aio_cancel_sync(struct kaiocb *job)
1774 {
1775 	struct kaioinfo *ki;
1776 
1777 	ki = job->userproc->p_aioinfo;
1778 	AIO_LOCK(ki);
1779 	if (!aio_cancel_cleared(job))
1780 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1781 	AIO_UNLOCK(ki);
1782 	aio_cancel(job);
1783 }
1784 
1785 int
1786 aio_queue_file(struct file *fp, struct kaiocb *job)
1787 {
1788 	struct kaioinfo *ki;
1789 	struct kaiocb *job2;
1790 	struct vnode *vp;
1791 	struct mount *mp;
1792 	int error;
1793 	bool safe;
1794 
1795 	ki = job->userproc->p_aioinfo;
1796 	error = aio_qbio(job->userproc, job);
1797 	if (error >= 0)
1798 		return (error);
1799 	safe = false;
1800 	if (fp->f_type == DTYPE_VNODE) {
1801 		vp = fp->f_vnode;
1802 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1803 			mp = fp->f_vnode->v_mount;
1804 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1805 				safe = true;
1806 		}
1807 	}
1808 	if (!(safe || enable_aio_unsafe)) {
1809 		counted_warning(&unsafe_warningcnt,
1810 		    "is attempting to use unsafe AIO requests");
1811 		return (EOPNOTSUPP);
1812 	}
1813 
1814 	if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) {
1815 		aio_schedule(job, aio_process_rw);
1816 		error = 0;
1817 	} else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) {
1818 		AIO_LOCK(ki);
1819 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1820 			if (job2->fd_file == job->fd_file &&
1821 			    ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) &&
1822 			    job2->seqno < job->seqno) {
1823 				job2->jobflags |= KAIOCB_CHECKSYNC;
1824 				job->pending++;
1825 			}
1826 		}
1827 		if (job->pending != 0) {
1828 			if (!aio_set_cancel_function_locked(job,
1829 				aio_cancel_sync)) {
1830 				AIO_UNLOCK(ki);
1831 				aio_cancel(job);
1832 				return (0);
1833 			}
1834 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1835 			AIO_UNLOCK(ki);
1836 			return (0);
1837 		}
1838 		AIO_UNLOCK(ki);
1839 		aio_schedule(job, aio_process_sync);
1840 		error = 0;
1841 	} else {
1842 		error = EINVAL;
1843 	}
1844 	return (error);
1845 }
1846 
1847 static void
1848 aio_kick_nowait(struct proc *userp)
1849 {
1850 	struct kaioinfo *ki = userp->p_aioinfo;
1851 	struct aioproc *aiop;
1852 
1853 	mtx_assert(&aio_job_mtx, MA_OWNED);
1854 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1855 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1856 		aiop->aioprocflags &= ~AIOP_FREE;
1857 		wakeup(aiop->aioproc);
1858 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1859 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1860 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1861 	}
1862 }
1863 
1864 static int
1865 aio_kick(struct proc *userp)
1866 {
1867 	struct kaioinfo *ki = userp->p_aioinfo;
1868 	struct aioproc *aiop;
1869 	int error, ret = 0;
1870 
1871 	mtx_assert(&aio_job_mtx, MA_OWNED);
1872 retryproc:
1873 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1874 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1875 		aiop->aioprocflags &= ~AIOP_FREE;
1876 		wakeup(aiop->aioproc);
1877 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1878 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1879 		num_aio_resv_start++;
1880 		mtx_unlock(&aio_job_mtx);
1881 		error = aio_newproc(&num_aio_resv_start);
1882 		mtx_lock(&aio_job_mtx);
1883 		if (error) {
1884 			num_aio_resv_start--;
1885 			goto retryproc;
1886 		}
1887 	} else {
1888 		ret = -1;
1889 	}
1890 	return (ret);
1891 }
1892 
1893 static void
1894 aio_kick_helper(void *context, int pending)
1895 {
1896 	struct proc *userp = context;
1897 
1898 	mtx_lock(&aio_job_mtx);
1899 	while (--pending >= 0) {
1900 		if (aio_kick(userp))
1901 			break;
1902 	}
1903 	mtx_unlock(&aio_job_mtx);
1904 }
1905 
1906 /*
1907  * Support the aio_return system call, as a side-effect, kernel resources are
1908  * released.
1909  */
1910 static int
1911 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1912 {
1913 	struct proc *p = td->td_proc;
1914 	struct kaiocb *job;
1915 	struct kaioinfo *ki;
1916 	long status, error;
1917 
1918 	ki = p->p_aioinfo;
1919 	if (ki == NULL)
1920 		return (EINVAL);
1921 	AIO_LOCK(ki);
1922 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1923 		if (job->ujob == ujob)
1924 			break;
1925 	}
1926 	if (job != NULL) {
1927 		MPASS(job->jobflags & KAIOCB_FINISHED);
1928 		status = job->uaiocb._aiocb_private.status;
1929 		error = job->uaiocb._aiocb_private.error;
1930 		td->td_retval[0] = status;
1931 		td->td_ru.ru_oublock += job->outblock;
1932 		td->td_ru.ru_inblock += job->inblock;
1933 		td->td_ru.ru_msgsnd += job->msgsnd;
1934 		td->td_ru.ru_msgrcv += job->msgrcv;
1935 		aio_free_entry(job);
1936 		AIO_UNLOCK(ki);
1937 		ops->store_error(ujob, error);
1938 		ops->store_status(ujob, status);
1939 	} else {
1940 		error = EINVAL;
1941 		AIO_UNLOCK(ki);
1942 	}
1943 	return (error);
1944 }
1945 
1946 int
1947 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1948 {
1949 
1950 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1951 }
1952 
1953 /*
1954  * Allow a process to wakeup when any of the I/O requests are completed.
1955  */
1956 static int
1957 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1958     struct timespec *ts)
1959 {
1960 	struct proc *p = td->td_proc;
1961 	struct timeval atv;
1962 	struct kaioinfo *ki;
1963 	struct kaiocb *firstjob, *job;
1964 	int error, i, timo;
1965 
1966 	timo = 0;
1967 	if (ts) {
1968 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1969 			return (EINVAL);
1970 
1971 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1972 		if (itimerfix(&atv))
1973 			return (EINVAL);
1974 		timo = tvtohz(&atv);
1975 	}
1976 
1977 	ki = p->p_aioinfo;
1978 	if (ki == NULL)
1979 		return (EAGAIN);
1980 
1981 	if (njoblist == 0)
1982 		return (0);
1983 
1984 	AIO_LOCK(ki);
1985 	for (;;) {
1986 		firstjob = NULL;
1987 		error = 0;
1988 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1989 			for (i = 0; i < njoblist; i++) {
1990 				if (job->ujob == ujoblist[i]) {
1991 					if (firstjob == NULL)
1992 						firstjob = job;
1993 					if (job->jobflags & KAIOCB_FINISHED)
1994 						goto RETURN;
1995 				}
1996 			}
1997 		}
1998 		/* All tasks were finished. */
1999 		if (firstjob == NULL)
2000 			break;
2001 
2002 		ki->kaio_flags |= KAIO_WAKEUP;
2003 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2004 		    "aiospn", timo);
2005 		if (error == ERESTART)
2006 			error = EINTR;
2007 		if (error)
2008 			break;
2009 	}
2010 RETURN:
2011 	AIO_UNLOCK(ki);
2012 	return (error);
2013 }
2014 
2015 int
2016 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2017 {
2018 	struct timespec ts, *tsp;
2019 	struct aiocb **ujoblist;
2020 	int error;
2021 
2022 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2023 		return (EINVAL);
2024 
2025 	if (uap->timeout) {
2026 		/* Get timespec struct. */
2027 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2028 			return (error);
2029 		tsp = &ts;
2030 	} else
2031 		tsp = NULL;
2032 
2033 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2034 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2035 	if (error == 0)
2036 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2037 	free(ujoblist, M_AIO);
2038 	return (error);
2039 }
2040 
2041 /*
2042  * aio_cancel cancels any non-bio aio operations not currently in progress.
2043  */
2044 int
2045 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2046 {
2047 	struct proc *p = td->td_proc;
2048 	struct kaioinfo *ki;
2049 	struct kaiocb *job, *jobn;
2050 	struct file *fp;
2051 	int error;
2052 	int cancelled = 0;
2053 	int notcancelled = 0;
2054 	struct vnode *vp;
2055 
2056 	/* Lookup file object. */
2057 	error = fget(td, uap->fd, &cap_no_rights, &fp);
2058 	if (error)
2059 		return (error);
2060 
2061 	ki = p->p_aioinfo;
2062 	if (ki == NULL)
2063 		goto done;
2064 
2065 	if (fp->f_type == DTYPE_VNODE) {
2066 		vp = fp->f_vnode;
2067 		if (vn_isdisk(vp)) {
2068 			fdrop(fp, td);
2069 			td->td_retval[0] = AIO_NOTCANCELED;
2070 			return (0);
2071 		}
2072 	}
2073 
2074 	AIO_LOCK(ki);
2075 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2076 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2077 		    ((uap->aiocbp == NULL) ||
2078 		     (uap->aiocbp == job->ujob))) {
2079 			if (aio_cancel_job(p, ki, job)) {
2080 				cancelled++;
2081 			} else {
2082 				notcancelled++;
2083 			}
2084 			if (uap->aiocbp != NULL)
2085 				break;
2086 		}
2087 	}
2088 	AIO_UNLOCK(ki);
2089 
2090 done:
2091 	fdrop(fp, td);
2092 
2093 	if (uap->aiocbp != NULL) {
2094 		if (cancelled) {
2095 			td->td_retval[0] = AIO_CANCELED;
2096 			return (0);
2097 		}
2098 	}
2099 
2100 	if (notcancelled) {
2101 		td->td_retval[0] = AIO_NOTCANCELED;
2102 		return (0);
2103 	}
2104 
2105 	if (cancelled) {
2106 		td->td_retval[0] = AIO_CANCELED;
2107 		return (0);
2108 	}
2109 
2110 	td->td_retval[0] = AIO_ALLDONE;
2111 
2112 	return (0);
2113 }
2114 
2115 /*
2116  * aio_error is implemented in the kernel level for compatibility purposes
2117  * only.  For a user mode async implementation, it would be best to do it in
2118  * a userland subroutine.
2119  */
2120 static int
2121 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2122 {
2123 	struct proc *p = td->td_proc;
2124 	struct kaiocb *job;
2125 	struct kaioinfo *ki;
2126 	int status;
2127 
2128 	ki = p->p_aioinfo;
2129 	if (ki == NULL) {
2130 		td->td_retval[0] = EINVAL;
2131 		return (0);
2132 	}
2133 
2134 	AIO_LOCK(ki);
2135 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2136 		if (job->ujob == ujob) {
2137 			if (job->jobflags & KAIOCB_FINISHED)
2138 				td->td_retval[0] =
2139 					job->uaiocb._aiocb_private.error;
2140 			else
2141 				td->td_retval[0] = EINPROGRESS;
2142 			AIO_UNLOCK(ki);
2143 			return (0);
2144 		}
2145 	}
2146 	AIO_UNLOCK(ki);
2147 
2148 	/*
2149 	 * Hack for failure of aio_aqueue.
2150 	 */
2151 	status = ops->fetch_status(ujob);
2152 	if (status == -1) {
2153 		td->td_retval[0] = ops->fetch_error(ujob);
2154 		return (0);
2155 	}
2156 
2157 	td->td_retval[0] = EINVAL;
2158 	return (0);
2159 }
2160 
2161 int
2162 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2163 {
2164 
2165 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2166 }
2167 
2168 /* syscall - asynchronous read from a file (REALTIME) */
2169 #ifdef COMPAT_FREEBSD6
2170 int
2171 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2172 {
2173 
2174 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2175 	    &aiocb_ops_osigevent));
2176 }
2177 #endif
2178 
2179 int
2180 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2181 {
2182 
2183 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2184 }
2185 
2186 int
2187 sys_aio_readv(struct thread *td, struct aio_readv_args *uap)
2188 {
2189 
2190 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops));
2191 }
2192 
2193 /* syscall - asynchronous write to a file (REALTIME) */
2194 #ifdef COMPAT_FREEBSD6
2195 int
2196 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2197 {
2198 
2199 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2200 	    &aiocb_ops_osigevent));
2201 }
2202 #endif
2203 
2204 int
2205 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2206 {
2207 
2208 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2209 }
2210 
2211 int
2212 sys_aio_writev(struct thread *td, struct aio_writev_args *uap)
2213 {
2214 
2215 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops));
2216 }
2217 
2218 int
2219 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2220 {
2221 
2222 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2223 }
2224 
2225 static int
2226 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2227     struct aiocb **acb_list, int nent, struct sigevent *sig,
2228     struct aiocb_ops *ops)
2229 {
2230 	struct proc *p = td->td_proc;
2231 	struct aiocb *job;
2232 	struct kaioinfo *ki;
2233 	struct aioliojob *lj;
2234 	struct kevent kev;
2235 	int error;
2236 	int nagain, nerror;
2237 	int i;
2238 
2239 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2240 		return (EINVAL);
2241 
2242 	if (nent < 0 || nent > max_aio_queue_per_proc)
2243 		return (EINVAL);
2244 
2245 	if (p->p_aioinfo == NULL)
2246 		aio_init_aioinfo(p);
2247 
2248 	ki = p->p_aioinfo;
2249 
2250 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2251 	lj->lioj_flags = 0;
2252 	lj->lioj_count = 0;
2253 	lj->lioj_finished_count = 0;
2254 	lj->lioj_signal.sigev_notify = SIGEV_NONE;
2255 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2256 	ksiginfo_init(&lj->lioj_ksi);
2257 
2258 	/*
2259 	 * Setup signal.
2260 	 */
2261 	if (sig && (mode == LIO_NOWAIT)) {
2262 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2263 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2264 			/* Assume only new style KEVENT */
2265 			memset(&kev, 0, sizeof(kev));
2266 			kev.filter = EVFILT_LIO;
2267 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2268 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2269 			kev.data = (intptr_t)lj;
2270 			/* pass user defined sigval data */
2271 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2272 			error = kqfd_register(
2273 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2274 			    M_WAITOK);
2275 			if (error) {
2276 				uma_zfree(aiolio_zone, lj);
2277 				return (error);
2278 			}
2279 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2280 			;
2281 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2282 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2283 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2284 					uma_zfree(aiolio_zone, lj);
2285 					return EINVAL;
2286 				}
2287 				lj->lioj_flags |= LIOJ_SIGNAL;
2288 		} else {
2289 			uma_zfree(aiolio_zone, lj);
2290 			return EINVAL;
2291 		}
2292 	}
2293 
2294 	AIO_LOCK(ki);
2295 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2296 	/*
2297 	 * Add extra aiocb count to avoid the lio to be freed
2298 	 * by other threads doing aio_waitcomplete or aio_return,
2299 	 * and prevent event from being sent until we have queued
2300 	 * all tasks.
2301 	 */
2302 	lj->lioj_count = 1;
2303 	AIO_UNLOCK(ki);
2304 
2305 	/*
2306 	 * Get pointers to the list of I/O requests.
2307 	 */
2308 	nagain = 0;
2309 	nerror = 0;
2310 	for (i = 0; i < nent; i++) {
2311 		job = acb_list[i];
2312 		if (job != NULL) {
2313 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2314 			if (error == EAGAIN)
2315 				nagain++;
2316 			else if (error != 0)
2317 				nerror++;
2318 		}
2319 	}
2320 
2321 	error = 0;
2322 	AIO_LOCK(ki);
2323 	if (mode == LIO_WAIT) {
2324 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2325 			ki->kaio_flags |= KAIO_WAKEUP;
2326 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2327 			    PRIBIO | PCATCH, "aiospn", 0);
2328 			if (error == ERESTART)
2329 				error = EINTR;
2330 			if (error)
2331 				break;
2332 		}
2333 	} else {
2334 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2335 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2336 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2337 				KNOTE_LOCKED(&lj->klist, 1);
2338 			}
2339 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2340 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2341 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2342 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2343 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2344 				    lj->lioj_count != 1);
2345 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2346 			}
2347 		}
2348 	}
2349 	lj->lioj_count--;
2350 	if (lj->lioj_count == 0) {
2351 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2352 		knlist_delete(&lj->klist, curthread, 1);
2353 		PROC_LOCK(p);
2354 		sigqueue_take(&lj->lioj_ksi);
2355 		PROC_UNLOCK(p);
2356 		AIO_UNLOCK(ki);
2357 		uma_zfree(aiolio_zone, lj);
2358 	} else
2359 		AIO_UNLOCK(ki);
2360 
2361 	if (nerror)
2362 		return (EIO);
2363 	else if (nagain)
2364 		return (EAGAIN);
2365 	else
2366 		return (error);
2367 }
2368 
2369 /* syscall - list directed I/O (REALTIME) */
2370 #ifdef COMPAT_FREEBSD6
2371 int
2372 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2373 {
2374 	struct aiocb **acb_list;
2375 	struct sigevent *sigp, sig;
2376 	struct osigevent osig;
2377 	int error, nent;
2378 
2379 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2380 		return (EINVAL);
2381 
2382 	nent = uap->nent;
2383 	if (nent < 0 || nent > max_aio_queue_per_proc)
2384 		return (EINVAL);
2385 
2386 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2387 		error = copyin(uap->sig, &osig, sizeof(osig));
2388 		if (error)
2389 			return (error);
2390 		error = convert_old_sigevent(&osig, &sig);
2391 		if (error)
2392 			return (error);
2393 		sigp = &sig;
2394 	} else
2395 		sigp = NULL;
2396 
2397 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2398 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2399 	if (error == 0)
2400 		error = kern_lio_listio(td, uap->mode,
2401 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2402 		    &aiocb_ops_osigevent);
2403 	free(acb_list, M_LIO);
2404 	return (error);
2405 }
2406 #endif
2407 
2408 /* syscall - list directed I/O (REALTIME) */
2409 int
2410 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2411 {
2412 	struct aiocb **acb_list;
2413 	struct sigevent *sigp, sig;
2414 	int error, nent;
2415 
2416 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2417 		return (EINVAL);
2418 
2419 	nent = uap->nent;
2420 	if (nent < 0 || nent > max_aio_queue_per_proc)
2421 		return (EINVAL);
2422 
2423 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2424 		error = copyin(uap->sig, &sig, sizeof(sig));
2425 		if (error)
2426 			return (error);
2427 		sigp = &sig;
2428 	} else
2429 		sigp = NULL;
2430 
2431 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2432 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2433 	if (error == 0)
2434 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2435 		    nent, sigp, &aiocb_ops);
2436 	free(acb_list, M_LIO);
2437 	return (error);
2438 }
2439 
2440 static void
2441 aio_biocleanup(struct bio *bp)
2442 {
2443 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2444 	struct kaioinfo *ki;
2445 	struct buf *pbuf = (struct buf *)bp->bio_caller2;
2446 
2447 	/* Release mapping into kernel space. */
2448 	if (pbuf != NULL) {
2449 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2450 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2451 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2452 		uma_zfree(pbuf_zone, pbuf);
2453 		atomic_subtract_int(&num_buf_aio, 1);
2454 		ki = job->userproc->p_aioinfo;
2455 		AIO_LOCK(ki);
2456 		ki->kaio_buffer_count--;
2457 		AIO_UNLOCK(ki);
2458 	} else {
2459 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2460 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2461 		free(bp->bio_ma, M_TEMP);
2462 		atomic_subtract_int(&num_unmapped_aio, 1);
2463 	}
2464 	g_destroy_bio(bp);
2465 }
2466 
2467 static void
2468 aio_biowakeup(struct bio *bp)
2469 {
2470 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2471 	size_t nbytes;
2472 	long bcount = bp->bio_bcount;
2473 	long resid = bp->bio_resid;
2474 	int opcode, nblks;
2475 	int bio_error = bp->bio_error;
2476 	uint16_t flags = bp->bio_flags;
2477 
2478 	opcode = job->uaiocb.aio_lio_opcode;
2479 
2480 	aio_biocleanup(bp);
2481 
2482 	nbytes = bcount - resid;
2483 	atomic_add_acq_long(&job->nbytes, nbytes);
2484 	nblks = btodb(nbytes);
2485 
2486 	/*
2487 	 * If multiple bios experienced an error, the job will reflect the
2488 	 * error of whichever failed bio completed last.
2489 	 */
2490 	if (flags & BIO_ERROR)
2491 		atomic_store_int(&job->error, bio_error);
2492 	if (opcode & LIO_WRITE)
2493 		atomic_add_int(&job->outblock, nblks);
2494 	else
2495 		atomic_add_int(&job->inblock, nblks);
2496 
2497 	if (refcount_release(&job->nbio)) {
2498 		bio_error = atomic_load_int(&job->error);
2499 		if (bio_error != 0)
2500 			aio_complete(job, -1, bio_error);
2501 		else
2502 			aio_complete(job, atomic_load_long(&job->nbytes), 0);
2503 	}
2504 }
2505 
2506 /* syscall - wait for the next completion of an aio request */
2507 static int
2508 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2509     struct timespec *ts, struct aiocb_ops *ops)
2510 {
2511 	struct proc *p = td->td_proc;
2512 	struct timeval atv;
2513 	struct kaioinfo *ki;
2514 	struct kaiocb *job;
2515 	struct aiocb *ujob;
2516 	long error, status;
2517 	int timo;
2518 
2519 	ops->store_aiocb(ujobp, NULL);
2520 
2521 	if (ts == NULL) {
2522 		timo = 0;
2523 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2524 		timo = -1;
2525 	} else {
2526 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2527 			return (EINVAL);
2528 
2529 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2530 		if (itimerfix(&atv))
2531 			return (EINVAL);
2532 		timo = tvtohz(&atv);
2533 	}
2534 
2535 	if (p->p_aioinfo == NULL)
2536 		aio_init_aioinfo(p);
2537 	ki = p->p_aioinfo;
2538 
2539 	error = 0;
2540 	job = NULL;
2541 	AIO_LOCK(ki);
2542 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2543 		if (timo == -1) {
2544 			error = EWOULDBLOCK;
2545 			break;
2546 		}
2547 		ki->kaio_flags |= KAIO_WAKEUP;
2548 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2549 		    "aiowc", timo);
2550 		if (timo && error == ERESTART)
2551 			error = EINTR;
2552 		if (error)
2553 			break;
2554 	}
2555 
2556 	if (job != NULL) {
2557 		MPASS(job->jobflags & KAIOCB_FINISHED);
2558 		ujob = job->ujob;
2559 		status = job->uaiocb._aiocb_private.status;
2560 		error = job->uaiocb._aiocb_private.error;
2561 		td->td_retval[0] = status;
2562 		td->td_ru.ru_oublock += job->outblock;
2563 		td->td_ru.ru_inblock += job->inblock;
2564 		td->td_ru.ru_msgsnd += job->msgsnd;
2565 		td->td_ru.ru_msgrcv += job->msgrcv;
2566 		aio_free_entry(job);
2567 		AIO_UNLOCK(ki);
2568 		ops->store_aiocb(ujobp, ujob);
2569 		ops->store_error(ujob, error);
2570 		ops->store_status(ujob, status);
2571 	} else
2572 		AIO_UNLOCK(ki);
2573 
2574 	return (error);
2575 }
2576 
2577 int
2578 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2579 {
2580 	struct timespec ts, *tsp;
2581 	int error;
2582 
2583 	if (uap->timeout) {
2584 		/* Get timespec struct. */
2585 		error = copyin(uap->timeout, &ts, sizeof(ts));
2586 		if (error)
2587 			return (error);
2588 		tsp = &ts;
2589 	} else
2590 		tsp = NULL;
2591 
2592 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2593 }
2594 
2595 static int
2596 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2597     struct aiocb_ops *ops)
2598 {
2599 	int listop;
2600 
2601 	switch (op) {
2602 	case O_SYNC:
2603 		listop = LIO_SYNC;
2604 		break;
2605 	case O_DSYNC:
2606 		listop = LIO_DSYNC;
2607 		break;
2608 	default:
2609 		return (EINVAL);
2610 	}
2611 
2612 	return (aio_aqueue(td, ujob, NULL, listop, ops));
2613 }
2614 
2615 int
2616 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2617 {
2618 
2619 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2620 }
2621 
2622 /* kqueue attach function */
2623 static int
2624 filt_aioattach(struct knote *kn)
2625 {
2626 	struct kaiocb *job;
2627 
2628 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2629 
2630 	/*
2631 	 * The job pointer must be validated before using it, so
2632 	 * registration is restricted to the kernel; the user cannot
2633 	 * set EV_FLAG1.
2634 	 */
2635 	if ((kn->kn_flags & EV_FLAG1) == 0)
2636 		return (EPERM);
2637 	kn->kn_ptr.p_aio = job;
2638 	kn->kn_flags &= ~EV_FLAG1;
2639 
2640 	knlist_add(&job->klist, kn, 0);
2641 
2642 	return (0);
2643 }
2644 
2645 /* kqueue detach function */
2646 static void
2647 filt_aiodetach(struct knote *kn)
2648 {
2649 	struct knlist *knl;
2650 
2651 	knl = &kn->kn_ptr.p_aio->klist;
2652 	knl->kl_lock(knl->kl_lockarg);
2653 	if (!knlist_empty(knl))
2654 		knlist_remove(knl, kn, 1);
2655 	knl->kl_unlock(knl->kl_lockarg);
2656 }
2657 
2658 /* kqueue filter function */
2659 /*ARGSUSED*/
2660 static int
2661 filt_aio(struct knote *kn, long hint)
2662 {
2663 	struct kaiocb *job = kn->kn_ptr.p_aio;
2664 
2665 	kn->kn_data = job->uaiocb._aiocb_private.error;
2666 	if (!(job->jobflags & KAIOCB_FINISHED))
2667 		return (0);
2668 	kn->kn_flags |= EV_EOF;
2669 	return (1);
2670 }
2671 
2672 /* kqueue attach function */
2673 static int
2674 filt_lioattach(struct knote *kn)
2675 {
2676 	struct aioliojob *lj;
2677 
2678 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2679 
2680 	/*
2681 	 * The aioliojob pointer must be validated before using it, so
2682 	 * registration is restricted to the kernel; the user cannot
2683 	 * set EV_FLAG1.
2684 	 */
2685 	if ((kn->kn_flags & EV_FLAG1) == 0)
2686 		return (EPERM);
2687 	kn->kn_ptr.p_lio = lj;
2688 	kn->kn_flags &= ~EV_FLAG1;
2689 
2690 	knlist_add(&lj->klist, kn, 0);
2691 
2692 	return (0);
2693 }
2694 
2695 /* kqueue detach function */
2696 static void
2697 filt_liodetach(struct knote *kn)
2698 {
2699 	struct knlist *knl;
2700 
2701 	knl = &kn->kn_ptr.p_lio->klist;
2702 	knl->kl_lock(knl->kl_lockarg);
2703 	if (!knlist_empty(knl))
2704 		knlist_remove(knl, kn, 1);
2705 	knl->kl_unlock(knl->kl_lockarg);
2706 }
2707 
2708 /* kqueue filter function */
2709 /*ARGSUSED*/
2710 static int
2711 filt_lio(struct knote *kn, long hint)
2712 {
2713 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2714 
2715 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2716 }
2717 
2718 #ifdef COMPAT_FREEBSD32
2719 #include <sys/mount.h>
2720 #include <sys/socket.h>
2721 #include <sys/sysent.h>
2722 #include <compat/freebsd32/freebsd32.h>
2723 #include <compat/freebsd32/freebsd32_proto.h>
2724 #include <compat/freebsd32/freebsd32_signal.h>
2725 #include <compat/freebsd32/freebsd32_syscall.h>
2726 #include <compat/freebsd32/freebsd32_util.h>
2727 
2728 struct __aiocb_private32 {
2729 	int32_t	status;
2730 	int32_t	error;
2731 	uint32_t kernelinfo;
2732 };
2733 
2734 #ifdef COMPAT_FREEBSD6
2735 typedef struct oaiocb32 {
2736 	int	aio_fildes;		/* File descriptor */
2737 	uint64_t aio_offset __packed;	/* File offset for I/O */
2738 	uint32_t aio_buf;		/* I/O buffer in process space */
2739 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2740 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2741 	int	aio_lio_opcode;		/* LIO opcode */
2742 	int	aio_reqprio;		/* Request priority -- ignored */
2743 	struct	__aiocb_private32 _aiocb_private;
2744 } oaiocb32_t;
2745 #endif
2746 
2747 typedef struct aiocb32 {
2748 	int32_t	aio_fildes;		/* File descriptor */
2749 	uint64_t aio_offset __packed;	/* File offset for I/O */
2750 	uint32_t aio_buf;	/* I/O buffer in process space */
2751 	uint32_t aio_nbytes;	/* Number of bytes for I/O */
2752 	int	__spare__[2];
2753 	uint32_t __spare2__;
2754 	int	aio_lio_opcode;		/* LIO opcode */
2755 	int	aio_reqprio;		/* Request priority -- ignored */
2756 	struct	__aiocb_private32 _aiocb_private;
2757 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2758 } aiocb32_t;
2759 
2760 #ifdef COMPAT_FREEBSD6
2761 static int
2762 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2763 {
2764 
2765 	/*
2766 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2767 	 * supported by AIO with the old sigevent structure.
2768 	 */
2769 	CP(*osig, *nsig, sigev_notify);
2770 	switch (nsig->sigev_notify) {
2771 	case SIGEV_NONE:
2772 		break;
2773 	case SIGEV_SIGNAL:
2774 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2775 		break;
2776 	case SIGEV_KEVENT:
2777 		nsig->sigev_notify_kqueue =
2778 		    osig->__sigev_u.__sigev_notify_kqueue;
2779 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2780 		break;
2781 	default:
2782 		return (EINVAL);
2783 	}
2784 	return (0);
2785 }
2786 
2787 static int
2788 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
2789     int type __unused)
2790 {
2791 	struct oaiocb32 job32;
2792 	struct aiocb *kcb = &kjob->uaiocb;
2793 	int error;
2794 
2795 	bzero(kcb, sizeof(struct aiocb));
2796 	error = copyin(ujob, &job32, sizeof(job32));
2797 	if (error)
2798 		return (error);
2799 
2800 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
2801 
2802 	CP(job32, *kcb, aio_fildes);
2803 	CP(job32, *kcb, aio_offset);
2804 	PTRIN_CP(job32, *kcb, aio_buf);
2805 	CP(job32, *kcb, aio_nbytes);
2806 	CP(job32, *kcb, aio_lio_opcode);
2807 	CP(job32, *kcb, aio_reqprio);
2808 	CP(job32, *kcb, _aiocb_private.status);
2809 	CP(job32, *kcb, _aiocb_private.error);
2810 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2811 	return (convert_old_sigevent32(&job32.aio_sigevent,
2812 	    &kcb->aio_sigevent));
2813 }
2814 #endif
2815 
2816 static int
2817 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
2818 {
2819 	struct aiocb32 job32;
2820 	struct aiocb *kcb = &kjob->uaiocb;
2821 	struct iovec32 *iov32;
2822 	int error;
2823 
2824 	error = copyin(ujob, &job32, sizeof(job32));
2825 	if (error)
2826 		return (error);
2827 	CP(job32, *kcb, aio_fildes);
2828 	CP(job32, *kcb, aio_offset);
2829 	CP(job32, *kcb, aio_lio_opcode);
2830 	if (type == LIO_NOP)
2831 		type = kcb->aio_lio_opcode;
2832 	if (type & LIO_VECTORED) {
2833 		iov32 = PTRIN(job32.aio_iov);
2834 		CP(job32, *kcb, aio_iovcnt);
2835 		/* malloc a uio and copy in the iovec */
2836 		error = freebsd32_copyinuio(iov32,
2837 		    kcb->aio_iovcnt, &kjob->uiop);
2838 		if (error)
2839 			return (error);
2840 	} else {
2841 		PTRIN_CP(job32, *kcb, aio_buf);
2842 		CP(job32, *kcb, aio_nbytes);
2843 	}
2844 	CP(job32, *kcb, aio_reqprio);
2845 	CP(job32, *kcb, _aiocb_private.status);
2846 	CP(job32, *kcb, _aiocb_private.error);
2847 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2848 	error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent);
2849 
2850 	return (error);
2851 }
2852 
2853 static long
2854 aiocb32_fetch_status(struct aiocb *ujob)
2855 {
2856 	struct aiocb32 *ujob32;
2857 
2858 	ujob32 = (struct aiocb32 *)ujob;
2859 	return (fuword32(&ujob32->_aiocb_private.status));
2860 }
2861 
2862 static long
2863 aiocb32_fetch_error(struct aiocb *ujob)
2864 {
2865 	struct aiocb32 *ujob32;
2866 
2867 	ujob32 = (struct aiocb32 *)ujob;
2868 	return (fuword32(&ujob32->_aiocb_private.error));
2869 }
2870 
2871 static int
2872 aiocb32_store_status(struct aiocb *ujob, long status)
2873 {
2874 	struct aiocb32 *ujob32;
2875 
2876 	ujob32 = (struct aiocb32 *)ujob;
2877 	return (suword32(&ujob32->_aiocb_private.status, status));
2878 }
2879 
2880 static int
2881 aiocb32_store_error(struct aiocb *ujob, long error)
2882 {
2883 	struct aiocb32 *ujob32;
2884 
2885 	ujob32 = (struct aiocb32 *)ujob;
2886 	return (suword32(&ujob32->_aiocb_private.error, error));
2887 }
2888 
2889 static int
2890 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2891 {
2892 	struct aiocb32 *ujob32;
2893 
2894 	ujob32 = (struct aiocb32 *)ujob;
2895 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2896 }
2897 
2898 static int
2899 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2900 {
2901 
2902 	return (suword32(ujobp, (long)ujob));
2903 }
2904 
2905 static struct aiocb_ops aiocb32_ops = {
2906 	.aio_copyin = aiocb32_copyin,
2907 	.fetch_status = aiocb32_fetch_status,
2908 	.fetch_error = aiocb32_fetch_error,
2909 	.store_status = aiocb32_store_status,
2910 	.store_error = aiocb32_store_error,
2911 	.store_kernelinfo = aiocb32_store_kernelinfo,
2912 	.store_aiocb = aiocb32_store_aiocb,
2913 };
2914 
2915 #ifdef COMPAT_FREEBSD6
2916 static struct aiocb_ops aiocb32_ops_osigevent = {
2917 	.aio_copyin = aiocb32_copyin_old_sigevent,
2918 	.fetch_status = aiocb32_fetch_status,
2919 	.fetch_error = aiocb32_fetch_error,
2920 	.store_status = aiocb32_store_status,
2921 	.store_error = aiocb32_store_error,
2922 	.store_kernelinfo = aiocb32_store_kernelinfo,
2923 	.store_aiocb = aiocb32_store_aiocb,
2924 };
2925 #endif
2926 
2927 int
2928 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2929 {
2930 
2931 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2932 }
2933 
2934 int
2935 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2936 {
2937 	struct timespec32 ts32;
2938 	struct timespec ts, *tsp;
2939 	struct aiocb **ujoblist;
2940 	uint32_t *ujoblist32;
2941 	int error, i;
2942 
2943 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2944 		return (EINVAL);
2945 
2946 	if (uap->timeout) {
2947 		/* Get timespec struct. */
2948 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2949 			return (error);
2950 		CP(ts32, ts, tv_sec);
2951 		CP(ts32, ts, tv_nsec);
2952 		tsp = &ts;
2953 	} else
2954 		tsp = NULL;
2955 
2956 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2957 	ujoblist32 = (uint32_t *)ujoblist;
2958 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2959 	    sizeof(ujoblist32[0]));
2960 	if (error == 0) {
2961 		for (i = uap->nent - 1; i >= 0; i--)
2962 			ujoblist[i] = PTRIN(ujoblist32[i]);
2963 
2964 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2965 	}
2966 	free(ujoblist, M_AIO);
2967 	return (error);
2968 }
2969 
2970 int
2971 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2972 {
2973 
2974 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2975 }
2976 
2977 #ifdef COMPAT_FREEBSD6
2978 int
2979 freebsd6_freebsd32_aio_read(struct thread *td,
2980     struct freebsd6_freebsd32_aio_read_args *uap)
2981 {
2982 
2983 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2984 	    &aiocb32_ops_osigevent));
2985 }
2986 #endif
2987 
2988 int
2989 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2990 {
2991 
2992 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2993 	    &aiocb32_ops));
2994 }
2995 
2996 int
2997 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap)
2998 {
2999 
3000 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV,
3001 	    &aiocb32_ops));
3002 }
3003 
3004 #ifdef COMPAT_FREEBSD6
3005 int
3006 freebsd6_freebsd32_aio_write(struct thread *td,
3007     struct freebsd6_freebsd32_aio_write_args *uap)
3008 {
3009 
3010 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3011 	    &aiocb32_ops_osigevent));
3012 }
3013 #endif
3014 
3015 int
3016 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
3017 {
3018 
3019 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3020 	    &aiocb32_ops));
3021 }
3022 
3023 int
3024 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap)
3025 {
3026 
3027 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV,
3028 	    &aiocb32_ops));
3029 }
3030 
3031 int
3032 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
3033 {
3034 
3035 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
3036 	    &aiocb32_ops));
3037 }
3038 
3039 int
3040 freebsd32_aio_waitcomplete(struct thread *td,
3041     struct freebsd32_aio_waitcomplete_args *uap)
3042 {
3043 	struct timespec32 ts32;
3044 	struct timespec ts, *tsp;
3045 	int error;
3046 
3047 	if (uap->timeout) {
3048 		/* Get timespec struct. */
3049 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
3050 		if (error)
3051 			return (error);
3052 		CP(ts32, ts, tv_sec);
3053 		CP(ts32, ts, tv_nsec);
3054 		tsp = &ts;
3055 	} else
3056 		tsp = NULL;
3057 
3058 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
3059 	    &aiocb32_ops));
3060 }
3061 
3062 int
3063 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
3064 {
3065 
3066 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
3067 	    &aiocb32_ops));
3068 }
3069 
3070 #ifdef COMPAT_FREEBSD6
3071 int
3072 freebsd6_freebsd32_lio_listio(struct thread *td,
3073     struct freebsd6_freebsd32_lio_listio_args *uap)
3074 {
3075 	struct aiocb **acb_list;
3076 	struct sigevent *sigp, sig;
3077 	struct osigevent32 osig;
3078 	uint32_t *acb_list32;
3079 	int error, i, nent;
3080 
3081 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3082 		return (EINVAL);
3083 
3084 	nent = uap->nent;
3085 	if (nent < 0 || nent > max_aio_queue_per_proc)
3086 		return (EINVAL);
3087 
3088 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3089 		error = copyin(uap->sig, &osig, sizeof(osig));
3090 		if (error)
3091 			return (error);
3092 		error = convert_old_sigevent32(&osig, &sig);
3093 		if (error)
3094 			return (error);
3095 		sigp = &sig;
3096 	} else
3097 		sigp = NULL;
3098 
3099 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3100 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3101 	if (error) {
3102 		free(acb_list32, M_LIO);
3103 		return (error);
3104 	}
3105 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3106 	for (i = 0; i < nent; i++)
3107 		acb_list[i] = PTRIN(acb_list32[i]);
3108 	free(acb_list32, M_LIO);
3109 
3110 	error = kern_lio_listio(td, uap->mode,
3111 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3112 	    &aiocb32_ops_osigevent);
3113 	free(acb_list, M_LIO);
3114 	return (error);
3115 }
3116 #endif
3117 
3118 int
3119 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3120 {
3121 	struct aiocb **acb_list;
3122 	struct sigevent *sigp, sig;
3123 	struct sigevent32 sig32;
3124 	uint32_t *acb_list32;
3125 	int error, i, nent;
3126 
3127 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3128 		return (EINVAL);
3129 
3130 	nent = uap->nent;
3131 	if (nent < 0 || nent > max_aio_queue_per_proc)
3132 		return (EINVAL);
3133 
3134 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3135 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3136 		if (error)
3137 			return (error);
3138 		error = convert_sigevent32(&sig32, &sig);
3139 		if (error)
3140 			return (error);
3141 		sigp = &sig;
3142 	} else
3143 		sigp = NULL;
3144 
3145 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3146 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3147 	if (error) {
3148 		free(acb_list32, M_LIO);
3149 		return (error);
3150 	}
3151 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3152 	for (i = 0; i < nent; i++)
3153 		acb_list[i] = PTRIN(acb_list32[i]);
3154 	free(acb_list32, M_LIO);
3155 
3156 	error = kern_lio_listio(td, uap->mode,
3157 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3158 	    &aiocb32_ops);
3159 	free(acb_list, M_LIO);
3160 	return (error);
3161 }
3162 
3163 #endif
3164