xref: /freebsd/sys/kern/vfs_aio.c (revision 4605a99b51ab72351d7554fbadbb24985f4667b1)
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 aio_fsync.
76  */
77 static uint64_t jobseqno;
78 
79 #ifndef MAX_AIO_PER_PROC
80 #define MAX_AIO_PER_PROC	32
81 #endif
82 
83 #ifndef MAX_AIO_QUEUE_PER_PROC
84 #define MAX_AIO_QUEUE_PER_PROC	256
85 #endif
86 
87 #ifndef MAX_AIO_QUEUE
88 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
89 #endif
90 
91 #ifndef MAX_BUF_AIO
92 #define MAX_BUF_AIO		16
93 #endif
94 
95 FEATURE(aio, "Asynchronous I/O");
96 SYSCTL_DECL(_p1003_1b);
97 
98 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
99 static MALLOC_DEFINE(M_AIO, "aio", "structures for asynchronous I/O");
100 
101 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
102     "Async IO management");
103 
104 static int enable_aio_unsafe = 0;
105 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
106     "Permit asynchronous IO on all file types, not just known-safe types");
107 
108 static unsigned int unsafe_warningcnt = 1;
109 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
110     &unsafe_warningcnt, 0,
111     "Warnings that will be triggered upon failed IO requests on unsafe files");
112 
113 static int max_aio_procs = MAX_AIO_PROCS;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
115     "Maximum number of kernel processes to use for handling async IO ");
116 
117 static int num_aio_procs = 0;
118 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
119     "Number of presently active kernel processes for async IO");
120 
121 /*
122  * The code will adjust the actual number of AIO processes towards this
123  * number when it gets a chance.
124  */
125 static int target_aio_procs = TARGET_AIO_PROCS;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
127     0,
128     "Preferred number of ready kernel processes for async IO");
129 
130 static int max_queue_count = MAX_AIO_QUEUE;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
132     "Maximum number of aio requests to queue, globally");
133 
134 static int num_queue_count = 0;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
136     "Number of queued aio requests");
137 
138 static int num_buf_aio = 0;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
140     "Number of aio requests presently handled by the buf subsystem");
141 
142 static int num_unmapped_aio = 0;
143 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
144     0,
145     "Number of aio requests presently handled by unmapped I/O buffers");
146 
147 /* Number of async I/O processes in the process of being started */
148 /* XXX This should be local to aio_aqueue() */
149 static int num_aio_resv_start = 0;
150 
151 static int aiod_lifetime;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
153     "Maximum lifetime for idle aiod");
154 
155 static int max_aio_per_proc = MAX_AIO_PER_PROC;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
157     0,
158     "Maximum active aio requests per process");
159 
160 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
162     &max_aio_queue_per_proc, 0,
163     "Maximum queued aio requests per process");
164 
165 static int max_buf_aio = MAX_BUF_AIO;
166 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
167     "Maximum buf aio requests per process");
168 
169 /*
170  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
171  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
172  * vfs.aio.aio_listio_max.
173  */
174 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
175     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
176     0, "Maximum aio requests for a single lio_listio call");
177 
178 #ifdef COMPAT_FREEBSD6
179 typedef struct oaiocb {
180 	int	aio_fildes;		/* File descriptor */
181 	off_t	aio_offset;		/* File offset for I/O */
182 	volatile void *aio_buf;         /* I/O buffer in process space */
183 	size_t	aio_nbytes;		/* Number of bytes for I/O */
184 	struct	osigevent aio_sigevent;	/* Signal to deliver */
185 	int	aio_lio_opcode;		/* LIO opcode */
186 	int	aio_reqprio;		/* Request priority -- ignored */
187 	struct	__aiocb_private	_aiocb_private;
188 } oaiocb_t;
189 #endif
190 
191 /*
192  * Below is a key of locks used to protect each member of struct kaiocb
193  * aioliojob and kaioinfo and any backends.
194  *
195  * * - need not protected
196  * a - locked by kaioinfo lock
197  * b - locked by backend lock, the backend lock can be null in some cases,
198  *     for example, BIO belongs to this type, in this case, proc lock is
199  *     reused.
200  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
201  */
202 
203 /*
204  * If the routine that services an AIO request blocks while running in an
205  * AIO kernel process it can starve other I/O requests.  BIO requests
206  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
207  * processes at all.  Socket I/O requests use a separate pool of
208  * kprocs and also force non-blocking I/O.  Other file I/O requests
209  * use the generic fo_read/fo_write operations which can block.  The
210  * fsync and mlock operations can also block while executing.  Ideally
211  * none of these requests would block while executing.
212  *
213  * Note that the service routines cannot toggle O_NONBLOCK in the file
214  * structure directly while handling a request due to races with
215  * userland threads.
216  */
217 
218 /* jobflags */
219 #define	KAIOCB_QUEUEING		0x01
220 #define	KAIOCB_CANCELLED	0x02
221 #define	KAIOCB_CANCELLING	0x04
222 #define	KAIOCB_CHECKSYNC	0x08
223 #define	KAIOCB_CLEARED		0x10
224 #define	KAIOCB_FINISHED		0x20
225 
226 /* ioflags */
227 #define	KAIOCB_IO_FOFFSET	0x01
228 
229 /*
230  * AIO process info
231  */
232 #define AIOP_FREE	0x1			/* proc on free queue */
233 
234 struct aioproc {
235 	int	aioprocflags;			/* (c) AIO proc flags */
236 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
237 	struct	proc *aioproc;			/* (*) the AIO proc */
238 };
239 
240 /*
241  * data-structure for lio signal management
242  */
243 struct aioliojob {
244 	int	lioj_flags;			/* (a) listio flags */
245 	int	lioj_count;			/* (a) count of jobs */
246 	int	lioj_finished_count;		/* (a) count of finished jobs */
247 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
248 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
249 	struct	knlist klist;			/* (a) list of knotes */
250 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
251 };
252 
253 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
254 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
255 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
256 
257 /*
258  * per process aio data structure
259  */
260 struct kaioinfo {
261 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
262 	int	kaio_flags;		/* (a) per process kaio flags */
263 	int	kaio_active_count;	/* (c) number of currently used AIOs */
264 	int	kaio_count;		/* (a) size of AIO queue */
265 	int	kaio_buffer_count;	/* (a) number of bio buffers */
266 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
267 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
268 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
269 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
270 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
271 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
272 	struct	task kaio_task;		/* (*) task to kick aio processes */
273 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
274 };
275 
276 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
277 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
278 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
279 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
280 
281 #define KAIO_RUNDOWN	0x1	/* process is being run down */
282 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
283 
284 /*
285  * Operations used to interact with userland aio control blocks.
286  * Different ABIs provide their own operations.
287  */
288 struct aiocb_ops {
289 	int	(*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty);
290 	long	(*fetch_status)(struct aiocb *ujob);
291 	long	(*fetch_error)(struct aiocb *ujob);
292 	int	(*store_status)(struct aiocb *ujob, long status);
293 	int	(*store_error)(struct aiocb *ujob, long error);
294 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
295 };
296 
297 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
298 static struct sema aio_newproc_sem;
299 static struct mtx aio_job_mtx;
300 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
301 static struct unrhdr *aiod_unr;
302 
303 static void	aio_biocleanup(struct bio *bp);
304 void		aio_init_aioinfo(struct proc *p);
305 static int	aio_onceonly(void);
306 static int	aio_free_entry(struct kaiocb *job);
307 static void	aio_process_rw(struct kaiocb *job);
308 static void	aio_process_sync(struct kaiocb *job);
309 static void	aio_process_mlock(struct kaiocb *job);
310 static void	aio_schedule_fsync(void *context, int pending);
311 static int	aio_newproc(int *);
312 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
313 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
314 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
315 static void	aio_biowakeup(struct bio *bp);
316 static void	aio_proc_rundown(void *arg, struct proc *p);
317 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
318 		    struct image_params *imgp);
319 static int	aio_qbio(struct proc *p, struct kaiocb *job);
320 static void	aio_daemon(void *param);
321 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
322 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
323 static int	aio_kick(struct proc *userp);
324 static void	aio_kick_nowait(struct proc *userp);
325 static void	aio_kick_helper(void *context, int pending);
326 static int	filt_aioattach(struct knote *kn);
327 static void	filt_aiodetach(struct knote *kn);
328 static int	filt_aio(struct knote *kn, long hint);
329 static int	filt_lioattach(struct knote *kn);
330 static void	filt_liodetach(struct knote *kn);
331 static int	filt_lio(struct knote *kn, long hint);
332 
333 /*
334  * Zones for:
335  * 	kaio	Per process async io info
336  *	aiocb	async io jobs
337  *	aiolio	list io jobs
338  */
339 static uma_zone_t kaio_zone, aiocb_zone, aiolio_zone;
340 
341 /* kqueue filters for aio */
342 static struct filterops aio_filtops = {
343 	.f_isfd = 0,
344 	.f_attach = filt_aioattach,
345 	.f_detach = filt_aiodetach,
346 	.f_event = filt_aio,
347 };
348 static struct filterops lio_filtops = {
349 	.f_isfd = 0,
350 	.f_attach = filt_lioattach,
351 	.f_detach = filt_liodetach,
352 	.f_event = filt_lio
353 };
354 
355 static eventhandler_tag exit_tag, exec_tag;
356 
357 TASKQUEUE_DEFINE_THREAD(aiod_kick);
358 
359 /*
360  * Main operations function for use as a kernel module.
361  */
362 static int
363 aio_modload(struct module *module, int cmd, void *arg)
364 {
365 	int error = 0;
366 
367 	switch (cmd) {
368 	case MOD_LOAD:
369 		aio_onceonly();
370 		break;
371 	case MOD_SHUTDOWN:
372 		break;
373 	default:
374 		error = EOPNOTSUPP;
375 		break;
376 	}
377 	return (error);
378 }
379 
380 static moduledata_t aio_mod = {
381 	"aio",
382 	&aio_modload,
383 	NULL
384 };
385 
386 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
387 MODULE_VERSION(aio, 1);
388 
389 /*
390  * Startup initialization
391  */
392 static int
393 aio_onceonly(void)
394 {
395 
396 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
397 	    EVENTHANDLER_PRI_ANY);
398 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
399 	    NULL, EVENTHANDLER_PRI_ANY);
400 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
401 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
402 	TAILQ_INIT(&aio_freeproc);
403 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
404 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
405 	TAILQ_INIT(&aio_jobs);
406 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
407 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
408 	    NULL, NULL, UMA_ALIGN_PTR, 0);
409 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
410 	    NULL, NULL, UMA_ALIGN_PTR, 0);
411 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
412 	    NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
413 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
414 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
415 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
416 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
417 
418 	return (0);
419 }
420 
421 /*
422  * Init the per-process aioinfo structure.  The aioinfo limits are set
423  * per-process for user limit (resource) management.
424  */
425 void
426 aio_init_aioinfo(struct proc *p)
427 {
428 	struct kaioinfo *ki;
429 
430 	ki = uma_zalloc(kaio_zone, M_WAITOK);
431 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
432 	ki->kaio_flags = 0;
433 	ki->kaio_active_count = 0;
434 	ki->kaio_count = 0;
435 	ki->kaio_buffer_count = 0;
436 	TAILQ_INIT(&ki->kaio_all);
437 	TAILQ_INIT(&ki->kaio_done);
438 	TAILQ_INIT(&ki->kaio_jobqueue);
439 	TAILQ_INIT(&ki->kaio_liojoblist);
440 	TAILQ_INIT(&ki->kaio_syncqueue);
441 	TAILQ_INIT(&ki->kaio_syncready);
442 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
443 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
444 	PROC_LOCK(p);
445 	if (p->p_aioinfo == NULL) {
446 		p->p_aioinfo = ki;
447 		PROC_UNLOCK(p);
448 	} else {
449 		PROC_UNLOCK(p);
450 		mtx_destroy(&ki->kaio_mtx);
451 		uma_zfree(kaio_zone, ki);
452 	}
453 
454 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
455 		aio_newproc(NULL);
456 }
457 
458 static int
459 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
460 {
461 	struct thread *td;
462 	int error;
463 
464 	error = sigev_findtd(p, sigev, &td);
465 	if (error)
466 		return (error);
467 	if (!KSI_ONQ(ksi)) {
468 		ksiginfo_set_sigev(ksi, sigev);
469 		ksi->ksi_code = SI_ASYNCIO;
470 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
471 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
472 	}
473 	PROC_UNLOCK(p);
474 	return (error);
475 }
476 
477 /*
478  * Free a job entry.  Wait for completion if it is currently active, but don't
479  * delay forever.  If we delay, we return a flag that says that we have to
480  * restart the queue scan.
481  */
482 static int
483 aio_free_entry(struct kaiocb *job)
484 {
485 	struct kaioinfo *ki;
486 	struct aioliojob *lj;
487 	struct proc *p;
488 
489 	p = job->userproc;
490 	MPASS(curproc == p);
491 	ki = p->p_aioinfo;
492 	MPASS(ki != NULL);
493 
494 	AIO_LOCK_ASSERT(ki, MA_OWNED);
495 	MPASS(job->jobflags & KAIOCB_FINISHED);
496 
497 	atomic_subtract_int(&num_queue_count, 1);
498 
499 	ki->kaio_count--;
500 	MPASS(ki->kaio_count >= 0);
501 
502 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
503 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
504 
505 	lj = job->lio;
506 	if (lj) {
507 		lj->lioj_count--;
508 		lj->lioj_finished_count--;
509 
510 		if (lj->lioj_count == 0) {
511 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
512 			/* lio is going away, we need to destroy any knotes */
513 			knlist_delete(&lj->klist, curthread, 1);
514 			PROC_LOCK(p);
515 			sigqueue_take(&lj->lioj_ksi);
516 			PROC_UNLOCK(p);
517 			uma_zfree(aiolio_zone, lj);
518 		}
519 	}
520 
521 	/* job is going away, we need to destroy any knotes */
522 	knlist_delete(&job->klist, curthread, 1);
523 	PROC_LOCK(p);
524 	sigqueue_take(&job->ksi);
525 	PROC_UNLOCK(p);
526 
527 	AIO_UNLOCK(ki);
528 
529 	/*
530 	 * The thread argument here is used to find the owning process
531 	 * and is also passed to fo_close() which may pass it to various
532 	 * places such as devsw close() routines.  Because of that, we
533 	 * need a thread pointer from the process owning the job that is
534 	 * persistent and won't disappear out from under us or move to
535 	 * another process.
536 	 *
537 	 * Currently, all the callers of this function call it to remove
538 	 * a kaiocb from the current process' job list either via a
539 	 * syscall or due to the current process calling exit() or
540 	 * execve().  Thus, we know that p == curproc.  We also know that
541 	 * curthread can't exit since we are curthread.
542 	 *
543 	 * Therefore, we use curthread as the thread to pass to
544 	 * knlist_delete().  This does mean that it is possible for the
545 	 * thread pointer at close time to differ from the thread pointer
546 	 * at open time, but this is already true of file descriptors in
547 	 * a multithreaded process.
548 	 */
549 	if (job->fd_file)
550 		fdrop(job->fd_file, curthread);
551 	crfree(job->cred);
552 	if (job->uiop != &job->uio)
553 		freeuio(job->uiop);
554 	uma_zfree(aiocb_zone, job);
555 	AIO_LOCK(ki);
556 
557 	return (0);
558 }
559 
560 static void
561 aio_proc_rundown_exec(void *arg, struct proc *p,
562     struct image_params *imgp __unused)
563 {
564    	aio_proc_rundown(arg, p);
565 }
566 
567 static int
568 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
569 {
570 	aio_cancel_fn_t *func;
571 	int cancelled;
572 
573 	AIO_LOCK_ASSERT(ki, MA_OWNED);
574 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
575 		return (0);
576 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
577 	job->jobflags |= KAIOCB_CANCELLED;
578 
579 	func = job->cancel_fn;
580 
581 	/*
582 	 * If there is no cancel routine, just leave the job marked as
583 	 * cancelled.  The job should be in active use by a caller who
584 	 * should complete it normally or when it fails to install a
585 	 * cancel routine.
586 	 */
587 	if (func == NULL)
588 		return (0);
589 
590 	/*
591 	 * Set the CANCELLING flag so that aio_complete() will defer
592 	 * completions of this job.  This prevents the job from being
593 	 * freed out from under the cancel callback.  After the
594 	 * callback any deferred completion (whether from the callback
595 	 * or any other source) will be completed.
596 	 */
597 	job->jobflags |= KAIOCB_CANCELLING;
598 	AIO_UNLOCK(ki);
599 	func(job);
600 	AIO_LOCK(ki);
601 	job->jobflags &= ~KAIOCB_CANCELLING;
602 	if (job->jobflags & KAIOCB_FINISHED) {
603 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
604 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
605 		aio_bio_done_notify(p, job);
606 	} else {
607 		/*
608 		 * The cancel callback might have scheduled an
609 		 * operation to cancel this request, but it is
610 		 * only counted as cancelled if the request is
611 		 * cancelled when the callback returns.
612 		 */
613 		cancelled = 0;
614 	}
615 	return (cancelled);
616 }
617 
618 /*
619  * Rundown the jobs for a given process.
620  */
621 static void
622 aio_proc_rundown(void *arg, struct proc *p)
623 {
624 	struct kaioinfo *ki;
625 	struct aioliojob *lj;
626 	struct kaiocb *job, *jobn;
627 
628 	KASSERT(curthread->td_proc == p,
629 	    ("%s: called on non-curproc", __func__));
630 	ki = p->p_aioinfo;
631 	if (ki == NULL)
632 		return;
633 
634 	AIO_LOCK(ki);
635 	ki->kaio_flags |= KAIO_RUNDOWN;
636 
637 restart:
638 
639 	/*
640 	 * Try to cancel all pending requests. This code simulates
641 	 * aio_cancel on all pending I/O requests.
642 	 */
643 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
644 		aio_cancel_job(p, ki, job);
645 	}
646 
647 	/* Wait for all running I/O to be finished */
648 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
649 		ki->kaio_flags |= KAIO_WAKEUP;
650 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
651 		goto restart;
652 	}
653 
654 	/* Free all completed I/O requests. */
655 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
656 		aio_free_entry(job);
657 
658 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
659 		if (lj->lioj_count == 0) {
660 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
661 			knlist_delete(&lj->klist, curthread, 1);
662 			PROC_LOCK(p);
663 			sigqueue_take(&lj->lioj_ksi);
664 			PROC_UNLOCK(p);
665 			uma_zfree(aiolio_zone, lj);
666 		} else {
667 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
668 			    lj->lioj_count, lj->lioj_finished_count);
669 		}
670 	}
671 	AIO_UNLOCK(ki);
672 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
673 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
674 	mtx_destroy(&ki->kaio_mtx);
675 	uma_zfree(kaio_zone, ki);
676 	p->p_aioinfo = NULL;
677 }
678 
679 /*
680  * Select a job to run (called by an AIO daemon).
681  */
682 static struct kaiocb *
683 aio_selectjob(struct aioproc *aiop)
684 {
685 	struct kaiocb *job;
686 	struct kaioinfo *ki;
687 	struct proc *userp;
688 
689 	mtx_assert(&aio_job_mtx, MA_OWNED);
690 restart:
691 	TAILQ_FOREACH(job, &aio_jobs, list) {
692 		userp = job->userproc;
693 		ki = userp->p_aioinfo;
694 
695 		if (ki->kaio_active_count < max_aio_per_proc) {
696 			TAILQ_REMOVE(&aio_jobs, job, list);
697 			if (!aio_clear_cancel_function(job))
698 				goto restart;
699 
700 			/* Account for currently active jobs. */
701 			ki->kaio_active_count++;
702 			break;
703 		}
704 	}
705 	return (job);
706 }
707 
708 /*
709  * Move all data to a permanent storage device.  This code
710  * simulates the fsync and fdatasync syscalls.
711  */
712 static int
713 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op)
714 {
715 	struct mount *mp;
716 	int error;
717 
718 	for (;;) {
719 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
720 		if (error != 0)
721 			break;
722 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
723 		vnode_pager_clean_async(vp);
724 		if (op == LIO_DSYNC)
725 			error = VOP_FDATASYNC(vp, td);
726 		else
727 			error = VOP_FSYNC(vp, MNT_WAIT, td);
728 
729 		VOP_UNLOCK(vp);
730 		vn_finished_write(mp);
731 		if (error != ERELOOKUP)
732 			break;
733 	}
734 	return (error);
735 }
736 
737 /*
738  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
739  * does the I/O request for the non-bio version of the operations.  The normal
740  * vn operations are used, and this code should work in all instances for every
741  * type of file, including pipes, sockets, fifos, and regular files.
742  *
743  * XXX I don't think it works well for socket, pipe, and fifo.
744  */
745 static void
746 aio_process_rw(struct kaiocb *job)
747 {
748 	struct ucred *td_savedcred;
749 	struct thread *td;
750 	struct file *fp;
751 	ssize_t cnt;
752 	long msgsnd_st, msgsnd_end;
753 	long msgrcv_st, msgrcv_end;
754 	long oublock_st, oublock_end;
755 	long inblock_st, inblock_end;
756 	int error, opcode;
757 
758 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
759 	    job->uaiocb.aio_lio_opcode == LIO_READV ||
760 	    job->uaiocb.aio_lio_opcode == LIO_WRITE ||
761 	    job->uaiocb.aio_lio_opcode == LIO_WRITEV,
762 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
763 
764 	aio_switch_vmspace(job);
765 	td = curthread;
766 	td_savedcred = td->td_ucred;
767 	td->td_ucred = job->cred;
768 	job->uiop->uio_td = td;
769 	fp = job->fd_file;
770 
771 	opcode = job->uaiocb.aio_lio_opcode;
772 	cnt = job->uiop->uio_resid;
773 
774 	msgrcv_st = td->td_ru.ru_msgrcv;
775 	msgsnd_st = td->td_ru.ru_msgsnd;
776 	inblock_st = td->td_ru.ru_inblock;
777 	oublock_st = td->td_ru.ru_oublock;
778 
779 	/*
780 	 * aio_aqueue() acquires a reference to the file that is
781 	 * released in aio_free_entry().
782 	 */
783 	if (opcode == LIO_READ || opcode == LIO_READV) {
784 		if (job->uiop->uio_resid == 0)
785 			error = 0;
786 		else
787 			error = fo_read(fp, job->uiop, fp->f_cred,
788 			    (job->ioflags & KAIOCB_IO_FOFFSET) != 0 ? 0 :
789 			    FOF_OFFSET, td);
790 	} else {
791 		if (fp->f_type == DTYPE_VNODE)
792 			bwillwrite();
793 		error = fo_write(fp, job->uiop, fp->f_cred, (job->ioflags &
794 		    KAIOCB_IO_FOFFSET) != 0 ? 0 : FOF_OFFSET, td);
795 	}
796 	msgrcv_end = td->td_ru.ru_msgrcv;
797 	msgsnd_end = td->td_ru.ru_msgsnd;
798 	inblock_end = td->td_ru.ru_inblock;
799 	oublock_end = td->td_ru.ru_oublock;
800 
801 	job->msgrcv = msgrcv_end - msgrcv_st;
802 	job->msgsnd = msgsnd_end - msgsnd_st;
803 	job->inblock = inblock_end - inblock_st;
804 	job->outblock = oublock_end - oublock_st;
805 
806 	if (error != 0 && job->uiop->uio_resid != cnt) {
807 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
808 			error = 0;
809 		if (error == EPIPE && (opcode & LIO_WRITE)) {
810 			PROC_LOCK(job->userproc);
811 			kern_psignal(job->userproc, SIGPIPE);
812 			PROC_UNLOCK(job->userproc);
813 		}
814 	}
815 
816 	cnt -= job->uiop->uio_resid;
817 	td->td_ucred = td_savedcred;
818 	if (error)
819 		aio_complete(job, -1, error);
820 	else
821 		aio_complete(job, cnt, 0);
822 }
823 
824 static void
825 aio_process_sync(struct kaiocb *job)
826 {
827 	struct thread *td = curthread;
828 	struct ucred *td_savedcred = td->td_ucred;
829 	struct file *fp = job->fd_file;
830 	int error = 0;
831 
832 	KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC,
833 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
834 
835 	td->td_ucred = job->cred;
836 	if (fp->f_vnode != NULL) {
837 		error = aio_fsync_vnode(td, fp->f_vnode,
838 		    job->uaiocb.aio_lio_opcode);
839 	}
840 	td->td_ucred = td_savedcred;
841 	if (error)
842 		aio_complete(job, -1, error);
843 	else
844 		aio_complete(job, 0, 0);
845 }
846 
847 static void
848 aio_process_mlock(struct kaiocb *job)
849 {
850 	struct aiocb *cb = &job->uaiocb;
851 	int error;
852 
853 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
854 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
855 
856 	aio_switch_vmspace(job);
857 	error = kern_mlock(job->userproc, job->cred,
858 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
859 	aio_complete(job, error != 0 ? -1 : 0, error);
860 }
861 
862 static void
863 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
864 {
865 	struct aioliojob *lj;
866 	struct kaioinfo *ki;
867 	struct kaiocb *sjob, *sjobn;
868 	int lj_done;
869 	bool schedule_fsync;
870 
871 	ki = userp->p_aioinfo;
872 	AIO_LOCK_ASSERT(ki, MA_OWNED);
873 	lj = job->lio;
874 	lj_done = 0;
875 	if (lj) {
876 		lj->lioj_finished_count++;
877 		if (lj->lioj_count == lj->lioj_finished_count)
878 			lj_done = 1;
879 	}
880 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
881 	MPASS(job->jobflags & KAIOCB_FINISHED);
882 
883 	if (ki->kaio_flags & KAIO_RUNDOWN)
884 		goto notification_done;
885 
886 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
887 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
888 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
889 
890 	KNOTE_LOCKED(&job->klist, 1);
891 
892 	if (lj_done) {
893 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
894 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
895 			KNOTE_LOCKED(&lj->klist, 1);
896 		}
897 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
898 		    == LIOJ_SIGNAL &&
899 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
900 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
901 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
902 			    true);
903 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
904 		}
905 	}
906 
907 notification_done:
908 	if (job->jobflags & KAIOCB_CHECKSYNC) {
909 		schedule_fsync = false;
910 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
911 			if (job->fd_file != sjob->fd_file ||
912 			    job->seqno >= sjob->seqno)
913 				continue;
914 			if (--sjob->pending > 0)
915 				continue;
916 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
917 			if (!aio_clear_cancel_function_locked(sjob))
918 				continue;
919 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
920 			schedule_fsync = true;
921 		}
922 		if (schedule_fsync)
923 			taskqueue_enqueue(taskqueue_aiod_kick,
924 			    &ki->kaio_sync_task);
925 	}
926 	if (ki->kaio_flags & KAIO_WAKEUP) {
927 		ki->kaio_flags &= ~KAIO_WAKEUP;
928 		wakeup(&userp->p_aioinfo);
929 	}
930 }
931 
932 static void
933 aio_schedule_fsync(void *context, int pending)
934 {
935 	struct kaioinfo *ki;
936 	struct kaiocb *job;
937 
938 	ki = context;
939 	AIO_LOCK(ki);
940 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
941 		job = TAILQ_FIRST(&ki->kaio_syncready);
942 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
943 		AIO_UNLOCK(ki);
944 		aio_schedule(job, aio_process_sync);
945 		AIO_LOCK(ki);
946 	}
947 	AIO_UNLOCK(ki);
948 }
949 
950 bool
951 aio_cancel_cleared(struct kaiocb *job)
952 {
953 
954 	/*
955 	 * The caller should hold the same queue lock held when
956 	 * aio_clear_cancel_function() was called and set this flag
957 	 * ensuring this check sees an up-to-date value.  However,
958 	 * there is no way to assert that.
959 	 */
960 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
961 }
962 
963 static bool
964 aio_clear_cancel_function_locked(struct kaiocb *job)
965 {
966 
967 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
968 	MPASS(job->cancel_fn != NULL);
969 	if (job->jobflags & KAIOCB_CANCELLING) {
970 		job->jobflags |= KAIOCB_CLEARED;
971 		return (false);
972 	}
973 	job->cancel_fn = NULL;
974 	return (true);
975 }
976 
977 bool
978 aio_clear_cancel_function(struct kaiocb *job)
979 {
980 	struct kaioinfo *ki;
981 	bool ret;
982 
983 	ki = job->userproc->p_aioinfo;
984 	AIO_LOCK(ki);
985 	ret = aio_clear_cancel_function_locked(job);
986 	AIO_UNLOCK(ki);
987 	return (ret);
988 }
989 
990 static bool
991 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
992 {
993 
994 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
995 	if (job->jobflags & KAIOCB_CANCELLED)
996 		return (false);
997 	job->cancel_fn = func;
998 	return (true);
999 }
1000 
1001 bool
1002 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1003 {
1004 	struct kaioinfo *ki;
1005 	bool ret;
1006 
1007 	ki = job->userproc->p_aioinfo;
1008 	AIO_LOCK(ki);
1009 	ret = aio_set_cancel_function_locked(job, func);
1010 	AIO_UNLOCK(ki);
1011 	return (ret);
1012 }
1013 
1014 void
1015 aio_complete(struct kaiocb *job, long status, int error)
1016 {
1017 	struct kaioinfo *ki;
1018 	struct proc *userp;
1019 
1020 	job->uaiocb._aiocb_private.error = error;
1021 	job->uaiocb._aiocb_private.status = status;
1022 
1023 	userp = job->userproc;
1024 	ki = userp->p_aioinfo;
1025 
1026 	AIO_LOCK(ki);
1027 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1028 	    ("duplicate aio_complete"));
1029 	job->jobflags |= KAIOCB_FINISHED;
1030 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1031 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1032 		aio_bio_done_notify(userp, job);
1033 	}
1034 	AIO_UNLOCK(ki);
1035 }
1036 
1037 void
1038 aio_cancel(struct kaiocb *job)
1039 {
1040 
1041 	aio_complete(job, -1, ECANCELED);
1042 }
1043 
1044 void
1045 aio_switch_vmspace(struct kaiocb *job)
1046 {
1047 
1048 	vmspace_switch_aio(job->userproc->p_vmspace);
1049 }
1050 
1051 /*
1052  * The AIO daemon, most of the actual work is done in aio_process_*,
1053  * but the setup (and address space mgmt) is done in this routine.
1054  */
1055 static void
1056 aio_daemon(void *_id)
1057 {
1058 	struct kaiocb *job;
1059 	struct aioproc *aiop;
1060 	struct kaioinfo *ki;
1061 	struct proc *p;
1062 	struct vmspace *myvm;
1063 	struct thread *td = curthread;
1064 	int id = (intptr_t)_id;
1065 
1066 	/*
1067 	 * Grab an extra reference on the daemon's vmspace so that it
1068 	 * doesn't get freed by jobs that switch to a different
1069 	 * vmspace.
1070 	 */
1071 	p = td->td_proc;
1072 	myvm = vmspace_acquire_ref(p);
1073 
1074 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1075 
1076 	/*
1077 	 * Allocate and ready the aio control info.  There is one aiop structure
1078 	 * per daemon.
1079 	 */
1080 	aiop = malloc(sizeof(*aiop), M_AIO, M_WAITOK);
1081 	aiop->aioproc = p;
1082 	aiop->aioprocflags = 0;
1083 
1084 	/*
1085 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1086 	 * and creating too many daemons.)
1087 	 */
1088 	sema_post(&aio_newproc_sem);
1089 
1090 	mtx_lock(&aio_job_mtx);
1091 	for (;;) {
1092 		/*
1093 		 * Take daemon off of free queue
1094 		 */
1095 		if (aiop->aioprocflags & AIOP_FREE) {
1096 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1097 			aiop->aioprocflags &= ~AIOP_FREE;
1098 		}
1099 
1100 		/*
1101 		 * Check for jobs.
1102 		 */
1103 		while ((job = aio_selectjob(aiop)) != NULL) {
1104 			mtx_unlock(&aio_job_mtx);
1105 
1106 			ki = job->userproc->p_aioinfo;
1107 			job->handle_fn(job);
1108 
1109 			mtx_lock(&aio_job_mtx);
1110 			/* Decrement the active job count. */
1111 			ki->kaio_active_count--;
1112 		}
1113 
1114 		/*
1115 		 * Disconnect from user address space.
1116 		 */
1117 		if (p->p_vmspace != myvm) {
1118 			mtx_unlock(&aio_job_mtx);
1119 			vmspace_switch_aio(myvm);
1120 			mtx_lock(&aio_job_mtx);
1121 			/*
1122 			 * We have to restart to avoid race, we only sleep if
1123 			 * no job can be selected.
1124 			 */
1125 			continue;
1126 		}
1127 
1128 		mtx_assert(&aio_job_mtx, MA_OWNED);
1129 
1130 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1131 		aiop->aioprocflags |= AIOP_FREE;
1132 
1133 		/*
1134 		 * If daemon is inactive for a long time, allow it to exit,
1135 		 * thereby freeing resources.
1136 		 */
1137 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1138 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1139 		    (aiop->aioprocflags & AIOP_FREE) &&
1140 		    num_aio_procs > target_aio_procs)
1141 			break;
1142 	}
1143 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1144 	num_aio_procs--;
1145 	mtx_unlock(&aio_job_mtx);
1146 	free(aiop, M_AIO);
1147 	free_unr(aiod_unr, id);
1148 	vmspace_free(myvm);
1149 
1150 	KASSERT(p->p_vmspace == myvm,
1151 	    ("AIOD: bad vmspace for exiting daemon"));
1152 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1153 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1154 	    refcount_load(&myvm->vm_refcnt)));
1155 	kproc_exit(0);
1156 }
1157 
1158 /*
1159  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1160  * AIO daemon modifies its environment itself.
1161  */
1162 static int
1163 aio_newproc(int *start)
1164 {
1165 	int error;
1166 	struct proc *p;
1167 	int id;
1168 
1169 	id = alloc_unr(aiod_unr);
1170 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1171 		RFNOWAIT, 0, "aiod%d", id);
1172 	if (error == 0) {
1173 		/*
1174 		 * Wait until daemon is started.
1175 		 */
1176 		sema_wait(&aio_newproc_sem);
1177 		mtx_lock(&aio_job_mtx);
1178 		num_aio_procs++;
1179 		if (start != NULL)
1180 			(*start)--;
1181 		mtx_unlock(&aio_job_mtx);
1182 	} else {
1183 		free_unr(aiod_unr, id);
1184 	}
1185 	return (error);
1186 }
1187 
1188 /*
1189  * Try the high-performance, low-overhead bio method for eligible
1190  * VCHR devices.  This method doesn't use an aio helper thread, and
1191  * thus has very low overhead.
1192  *
1193  * Assumes that the caller, aio_aqueue(), has incremented the file
1194  * structure's reference count, preventing its deallocation for the
1195  * duration of this call.
1196  */
1197 static int
1198 aio_qbio(struct proc *p, struct kaiocb *job)
1199 {
1200 	struct aiocb *cb;
1201 	struct file *fp;
1202 	struct buf *pbuf;
1203 	struct vnode *vp;
1204 	struct cdevsw *csw;
1205 	struct cdev *dev;
1206 	struct kaioinfo *ki;
1207 	struct bio **bios = NULL;
1208 	off_t offset;
1209 	int bio_cmd, error, i, iovcnt, opcode, poff, ref;
1210 	vm_prot_t prot;
1211 	bool use_unmapped;
1212 
1213 	cb = &job->uaiocb;
1214 	fp = job->fd_file;
1215 	opcode = cb->aio_lio_opcode;
1216 
1217 	if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV ||
1218 	    opcode == LIO_READ || opcode == LIO_READV))
1219 		return (-1);
1220 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1221 		return (-1);
1222 
1223 	vp = fp->f_vnode;
1224 	if (vp->v_type != VCHR)
1225 		return (-1);
1226 	if (vp->v_bufobj.bo_bsize == 0)
1227 		return (-1);
1228 
1229 	bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ;
1230 	iovcnt = job->uiop->uio_iovcnt;
1231 	if (iovcnt > max_buf_aio)
1232 		return (-1);
1233 	for (i = 0; i < iovcnt; i++) {
1234 		if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0)
1235 			return (-1);
1236 		if (job->uiop->uio_iov[i].iov_len > maxphys) {
1237 			error = -1;
1238 			return (-1);
1239 		}
1240 	}
1241 	offset = cb->aio_offset;
1242 
1243 	ref = 0;
1244 	csw = devvn_refthread(vp, &dev, &ref);
1245 	if (csw == NULL)
1246 		return (ENXIO);
1247 
1248 	if ((csw->d_flags & D_DISK) == 0) {
1249 		error = -1;
1250 		goto unref;
1251 	}
1252 	if (job->uiop->uio_resid > dev->si_iosize_max) {
1253 		error = -1;
1254 		goto unref;
1255 	}
1256 
1257 	ki = p->p_aioinfo;
1258 	job->error = 0;
1259 
1260 	use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed;
1261 	if (!use_unmapped) {
1262 		AIO_LOCK(ki);
1263 		if (ki->kaio_buffer_count + iovcnt > max_buf_aio) {
1264 			AIO_UNLOCK(ki);
1265 			error = EAGAIN;
1266 			goto unref;
1267 		}
1268 		ki->kaio_buffer_count += iovcnt;
1269 		AIO_UNLOCK(ki);
1270 	}
1271 
1272 	bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK);
1273 	refcount_init(&job->nbio, iovcnt);
1274 	for (i = 0; i < iovcnt; i++) {
1275 		struct vm_page** pages;
1276 		struct bio *bp;
1277 		void *buf;
1278 		size_t nbytes;
1279 		int npages;
1280 
1281 		buf = job->uiop->uio_iov[i].iov_base;
1282 		nbytes = job->uiop->uio_iov[i].iov_len;
1283 
1284 		bios[i] = g_alloc_bio();
1285 		bp = bios[i];
1286 
1287 		poff = (vm_offset_t)buf & PAGE_MASK;
1288 		if (use_unmapped) {
1289 			pbuf = NULL;
1290 			pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1291 			    nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1292 		} else {
1293 			pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1294 			BUF_KERNPROC(pbuf);
1295 			pages = pbuf->b_pages;
1296 		}
1297 
1298 		bp->bio_length = nbytes;
1299 		bp->bio_bcount = nbytes;
1300 		bp->bio_done = aio_biowakeup;
1301 		bp->bio_offset = offset;
1302 		bp->bio_cmd = bio_cmd;
1303 		bp->bio_dev = dev;
1304 		bp->bio_caller1 = job;
1305 		bp->bio_caller2 = pbuf;
1306 
1307 		prot = VM_PROT_READ;
1308 		if (opcode == LIO_READ || opcode == LIO_READV)
1309 			prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1310 		npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1311 		    (vm_offset_t)buf, bp->bio_length, prot, pages,
1312 		    atop(maxphys) + 1);
1313 		if (npages < 0) {
1314 			if (pbuf != NULL)
1315 				uma_zfree(pbuf_zone, pbuf);
1316 			else
1317 				free(pages, M_TEMP);
1318 			error = EFAULT;
1319 			g_destroy_bio(bp);
1320 			i--;
1321 			goto destroy_bios;
1322 		}
1323 		if (pbuf != NULL) {
1324 			pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages);
1325 			bp->bio_data = pbuf->b_data + poff;
1326 			pbuf->b_npages = npages;
1327 			atomic_add_int(&num_buf_aio, 1);
1328 		} else {
1329 			bp->bio_ma = pages;
1330 			bp->bio_ma_n = npages;
1331 			bp->bio_ma_offset = poff;
1332 			bp->bio_data = unmapped_buf;
1333 			bp->bio_flags |= BIO_UNMAPPED;
1334 			atomic_add_int(&num_unmapped_aio, 1);
1335 		}
1336 
1337 		offset += nbytes;
1338 	}
1339 
1340 	/* Perform transfer. */
1341 	for (i = 0; i < iovcnt; i++)
1342 		csw->d_strategy(bios[i]);
1343 	free(bios, M_TEMP);
1344 
1345 	dev_relthread(dev, ref);
1346 	return (0);
1347 
1348 destroy_bios:
1349 	for (; i >= 0; i--)
1350 		aio_biocleanup(bios[i]);
1351 	free(bios, M_TEMP);
1352 unref:
1353 	dev_relthread(dev, ref);
1354 	return (error);
1355 }
1356 
1357 #ifdef COMPAT_FREEBSD6
1358 static int
1359 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1360 {
1361 
1362 	/*
1363 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1364 	 * supported by AIO with the old sigevent structure.
1365 	 */
1366 	nsig->sigev_notify = osig->sigev_notify;
1367 	switch (nsig->sigev_notify) {
1368 	case SIGEV_NONE:
1369 		break;
1370 	case SIGEV_SIGNAL:
1371 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1372 		break;
1373 	case SIGEV_KEVENT:
1374 		nsig->sigev_notify_kqueue =
1375 		    osig->__sigev_u.__sigev_notify_kqueue;
1376 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1377 		break;
1378 	default:
1379 		return (EINVAL);
1380 	}
1381 	return (0);
1382 }
1383 
1384 static int
1385 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
1386     int type __unused)
1387 {
1388 	struct oaiocb *ojob;
1389 	struct aiocb *kcb = &kjob->uaiocb;
1390 	int error;
1391 
1392 	bzero(kcb, sizeof(struct aiocb));
1393 	error = copyin(ujob, kcb, sizeof(struct oaiocb));
1394 	if (error)
1395 		return (error);
1396 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
1397 	ojob = (struct oaiocb *)kcb;
1398 	return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent));
1399 }
1400 #endif
1401 
1402 static int
1403 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
1404 {
1405 	struct aiocb *kcb = &kjob->uaiocb;
1406 	int error;
1407 
1408 	error = copyin(ujob, kcb, sizeof(struct aiocb));
1409 	if (error)
1410 		return (error);
1411 	if (type == LIO_NOP)
1412 		type = kcb->aio_lio_opcode;
1413 	if (type & LIO_VECTORED) {
1414 		/* malloc a uio and copy in the iovec */
1415 		error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov),
1416 		    kcb->aio_iovcnt, &kjob->uiop);
1417 	}
1418 
1419 	return (error);
1420 }
1421 
1422 static long
1423 aiocb_fetch_status(struct aiocb *ujob)
1424 {
1425 
1426 	return (fuword(&ujob->_aiocb_private.status));
1427 }
1428 
1429 static long
1430 aiocb_fetch_error(struct aiocb *ujob)
1431 {
1432 
1433 	return (fuword(&ujob->_aiocb_private.error));
1434 }
1435 
1436 static int
1437 aiocb_store_status(struct aiocb *ujob, long status)
1438 {
1439 
1440 	return (suword(&ujob->_aiocb_private.status, status));
1441 }
1442 
1443 static int
1444 aiocb_store_error(struct aiocb *ujob, long error)
1445 {
1446 
1447 	return (suword(&ujob->_aiocb_private.error, error));
1448 }
1449 
1450 static int
1451 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1452 {
1453 
1454 	return (suword(ujobp, (long)ujob));
1455 }
1456 
1457 static struct aiocb_ops aiocb_ops = {
1458 	.aio_copyin = aiocb_copyin,
1459 	.fetch_status = aiocb_fetch_status,
1460 	.fetch_error = aiocb_fetch_error,
1461 	.store_status = aiocb_store_status,
1462 	.store_error = aiocb_store_error,
1463 	.store_aiocb = aiocb_store_aiocb,
1464 };
1465 
1466 #ifdef COMPAT_FREEBSD6
1467 static struct aiocb_ops aiocb_ops_osigevent = {
1468 	.aio_copyin = aiocb_copyin_old_sigevent,
1469 	.fetch_status = aiocb_fetch_status,
1470 	.fetch_error = aiocb_fetch_error,
1471 	.store_status = aiocb_store_status,
1472 	.store_error = aiocb_store_error,
1473 	.store_aiocb = aiocb_store_aiocb,
1474 };
1475 #endif
1476 
1477 /*
1478  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1479  * technique is done in this code.
1480  */
1481 int
1482 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1483     int type, struct aiocb_ops *ops)
1484 {
1485 	struct proc *p = td->td_proc;
1486 	struct file *fp = NULL;
1487 	struct kaiocb *job;
1488 	struct kaioinfo *ki;
1489 	struct kevent kev;
1490 	int opcode;
1491 	int error;
1492 	int fd, kqfd;
1493 	u_short evflags;
1494 
1495 	if (p->p_aioinfo == NULL)
1496 		aio_init_aioinfo(p);
1497 
1498 	ki = p->p_aioinfo;
1499 
1500 	ops->store_status(ujob, -1);
1501 	ops->store_error(ujob, 0);
1502 
1503 	if (num_queue_count >= max_queue_count ||
1504 	    ki->kaio_count >= max_aio_queue_per_proc) {
1505 		error = EAGAIN;
1506 		goto err1;
1507 	}
1508 
1509 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1510 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1511 
1512 	error = ops->aio_copyin(ujob, job, type);
1513 	if (error)
1514 		goto err2;
1515 
1516 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1517 		error = EINVAL;
1518 		goto err2;
1519 	}
1520 
1521 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1522 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1523 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1524 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1525 		error = EINVAL;
1526 		goto err2;
1527 	}
1528 
1529 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1530 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1531 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1532 		error = EINVAL;
1533 		goto err2;
1534 	}
1535 
1536 	/* Get the opcode. */
1537 	if (type == LIO_NOP) {
1538 		switch (job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET) {
1539 		case LIO_WRITE:
1540 		case LIO_WRITEV:
1541 		case LIO_NOP:
1542 		case LIO_READ:
1543 		case LIO_READV:
1544 			opcode = job->uaiocb.aio_lio_opcode & ~LIO_FOFFSET;
1545 			if ((job->uaiocb.aio_lio_opcode & LIO_FOFFSET) != 0)
1546 				job->ioflags |= KAIOCB_IO_FOFFSET;
1547 			break;
1548 		default:
1549 			error = EINVAL;
1550 			goto err2;
1551 		}
1552 	} else
1553 		opcode = job->uaiocb.aio_lio_opcode = type;
1554 
1555 	ksiginfo_init(&job->ksi);
1556 
1557 	/* Save userspace address of the job info. */
1558 	job->ujob = ujob;
1559 
1560 	/*
1561 	 * Validate the opcode and fetch the file object for the specified
1562 	 * file descriptor.
1563 	 *
1564 	 * XXXRW: Moved the opcode validation up here so that we don't
1565 	 * retrieve a file descriptor without knowing what the capabiltity
1566 	 * should be.
1567 	 */
1568 	fd = job->uaiocb.aio_fildes;
1569 	switch (opcode) {
1570 	case LIO_WRITE:
1571 	case LIO_WRITEV:
1572 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1573 		break;
1574 	case LIO_READ:
1575 	case LIO_READV:
1576 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1577 		break;
1578 	case LIO_SYNC:
1579 	case LIO_DSYNC:
1580 		error = fget(td, fd, &cap_fsync_rights, &fp);
1581 		break;
1582 	case LIO_MLOCK:
1583 		break;
1584 	case LIO_NOP:
1585 		error = fget(td, fd, &cap_no_rights, &fp);
1586 		break;
1587 	default:
1588 		error = EINVAL;
1589 	}
1590 	if (error)
1591 		goto err3;
1592 
1593 	if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) {
1594 		error = EINVAL;
1595 		goto err3;
1596 	}
1597 
1598 	if ((opcode == LIO_READ || opcode == LIO_READV ||
1599 	    opcode == LIO_WRITE || opcode == LIO_WRITEV) &&
1600 	    job->uaiocb.aio_offset < 0 &&
1601 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1602 		error = EINVAL;
1603 		goto err3;
1604 	}
1605 
1606 	if (fp != NULL && fp->f_ops == &path_fileops) {
1607 		error = EBADF;
1608 		goto err3;
1609 	}
1610 
1611 	job->fd_file = fp;
1612 
1613 	mtx_lock(&aio_job_mtx);
1614 	job->seqno = jobseqno++;
1615 	mtx_unlock(&aio_job_mtx);
1616 	if (opcode == LIO_NOP) {
1617 		fdrop(fp, td);
1618 		MPASS(job->uiop == &job->uio || job->uiop == NULL);
1619 		uma_zfree(aiocb_zone, job);
1620 		return (0);
1621 	}
1622 
1623 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1624 		goto no_kqueue;
1625 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1626 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1627 		error = EINVAL;
1628 		goto err3;
1629 	}
1630 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1631 	memset(&kev, 0, sizeof(kev));
1632 	kev.ident = (uintptr_t)job->ujob;
1633 	kev.filter = EVFILT_AIO;
1634 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1635 	kev.data = (intptr_t)job;
1636 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1637 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1638 	if (error)
1639 		goto err3;
1640 
1641 no_kqueue:
1642 
1643 	ops->store_error(ujob, EINPROGRESS);
1644 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1645 	job->userproc = p;
1646 	job->cred = crhold(td->td_ucred);
1647 	job->jobflags = KAIOCB_QUEUEING;
1648 	job->lio = lj;
1649 
1650 	if (opcode & LIO_VECTORED) {
1651 		/* Use the uio copied in by aio_copyin */
1652 		MPASS(job->uiop != &job->uio && job->uiop != NULL);
1653 	} else {
1654 		/* Setup the inline uio */
1655 		job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf;
1656 		job->iov[0].iov_len = job->uaiocb.aio_nbytes;
1657 		job->uio.uio_iov = job->iov;
1658 		job->uio.uio_iovcnt = 1;
1659 		job->uio.uio_resid = job->uaiocb.aio_nbytes;
1660 		job->uio.uio_segflg = UIO_USERSPACE;
1661 		job->uiop = &job->uio;
1662 	}
1663 	switch (opcode & (LIO_READ | LIO_WRITE)) {
1664 	case LIO_READ:
1665 		job->uiop->uio_rw = UIO_READ;
1666 		break;
1667 	case LIO_WRITE:
1668 		job->uiop->uio_rw = UIO_WRITE;
1669 		break;
1670 	}
1671 	job->uiop->uio_offset = job->uaiocb.aio_offset;
1672 	job->uiop->uio_td = td;
1673 
1674 	if (opcode == LIO_MLOCK) {
1675 		aio_schedule(job, aio_process_mlock);
1676 		error = 0;
1677 	} else if (fp->f_ops->fo_aio_queue == NULL)
1678 		error = aio_queue_file(fp, job);
1679 	else
1680 		error = fo_aio_queue(fp, job);
1681 	if (error)
1682 		goto err4;
1683 
1684 	AIO_LOCK(ki);
1685 	job->jobflags &= ~KAIOCB_QUEUEING;
1686 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1687 	ki->kaio_count++;
1688 	if (lj)
1689 		lj->lioj_count++;
1690 	atomic_add_int(&num_queue_count, 1);
1691 	if (job->jobflags & KAIOCB_FINISHED) {
1692 		/*
1693 		 * The queue callback completed the request synchronously.
1694 		 * The bulk of the completion is deferred in that case
1695 		 * until this point.
1696 		 */
1697 		aio_bio_done_notify(p, job);
1698 	} else
1699 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1700 	AIO_UNLOCK(ki);
1701 	return (0);
1702 
1703 err4:
1704 	crfree(job->cred);
1705 err3:
1706 	if (fp)
1707 		fdrop(fp, td);
1708 	knlist_delete(&job->klist, curthread, 0);
1709 err2:
1710 	if (job->uiop != &job->uio)
1711 		freeuio(job->uiop);
1712 	uma_zfree(aiocb_zone, job);
1713 err1:
1714 	ops->store_error(ujob, error);
1715 	return (error);
1716 }
1717 
1718 static void
1719 aio_cancel_daemon_job(struct kaiocb *job)
1720 {
1721 
1722 	mtx_lock(&aio_job_mtx);
1723 	if (!aio_cancel_cleared(job))
1724 		TAILQ_REMOVE(&aio_jobs, job, list);
1725 	mtx_unlock(&aio_job_mtx);
1726 	aio_cancel(job);
1727 }
1728 
1729 void
1730 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1731 {
1732 
1733 	mtx_lock(&aio_job_mtx);
1734 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1735 		mtx_unlock(&aio_job_mtx);
1736 		aio_cancel(job);
1737 		return;
1738 	}
1739 	job->handle_fn = func;
1740 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1741 	aio_kick_nowait(job->userproc);
1742 	mtx_unlock(&aio_job_mtx);
1743 }
1744 
1745 static void
1746 aio_cancel_sync(struct kaiocb *job)
1747 {
1748 	struct kaioinfo *ki;
1749 
1750 	ki = job->userproc->p_aioinfo;
1751 	AIO_LOCK(ki);
1752 	if (!aio_cancel_cleared(job))
1753 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1754 	AIO_UNLOCK(ki);
1755 	aio_cancel(job);
1756 }
1757 
1758 int
1759 aio_queue_file(struct file *fp, struct kaiocb *job)
1760 {
1761 	struct kaioinfo *ki;
1762 	struct kaiocb *job2;
1763 	struct vnode *vp;
1764 	struct mount *mp;
1765 	int error;
1766 	bool safe;
1767 
1768 	ki = job->userproc->p_aioinfo;
1769 	error = aio_qbio(job->userproc, job);
1770 	if (error >= 0)
1771 		return (error);
1772 	safe = false;
1773 	if (fp->f_type == DTYPE_VNODE) {
1774 		vp = fp->f_vnode;
1775 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1776 			mp = fp->f_vnode->v_mount;
1777 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1778 				safe = true;
1779 		}
1780 	}
1781 	if (!(safe || enable_aio_unsafe)) {
1782 		counted_warning(&unsafe_warningcnt,
1783 		    "is attempting to use unsafe AIO requests");
1784 		return (EOPNOTSUPP);
1785 	}
1786 
1787 	if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) {
1788 		aio_schedule(job, aio_process_rw);
1789 		error = 0;
1790 	} else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) {
1791 		AIO_LOCK(ki);
1792 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1793 			if (job2->fd_file == job->fd_file &&
1794 			    ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) &&
1795 			    job2->seqno < job->seqno) {
1796 				job2->jobflags |= KAIOCB_CHECKSYNC;
1797 				job->pending++;
1798 			}
1799 		}
1800 		if (job->pending != 0) {
1801 			if (!aio_set_cancel_function_locked(job,
1802 				aio_cancel_sync)) {
1803 				AIO_UNLOCK(ki);
1804 				aio_cancel(job);
1805 				return (0);
1806 			}
1807 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1808 			AIO_UNLOCK(ki);
1809 			return (0);
1810 		}
1811 		AIO_UNLOCK(ki);
1812 		aio_schedule(job, aio_process_sync);
1813 		error = 0;
1814 	} else {
1815 		error = EINVAL;
1816 	}
1817 	return (error);
1818 }
1819 
1820 static void
1821 aio_kick_nowait(struct proc *userp)
1822 {
1823 	struct kaioinfo *ki = userp->p_aioinfo;
1824 	struct aioproc *aiop;
1825 
1826 	mtx_assert(&aio_job_mtx, MA_OWNED);
1827 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1828 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1829 		aiop->aioprocflags &= ~AIOP_FREE;
1830 		wakeup(aiop->aioproc);
1831 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1832 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1833 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1834 	}
1835 }
1836 
1837 static int
1838 aio_kick(struct proc *userp)
1839 {
1840 	struct kaioinfo *ki = userp->p_aioinfo;
1841 	struct aioproc *aiop;
1842 	int error, ret = 0;
1843 
1844 	mtx_assert(&aio_job_mtx, MA_OWNED);
1845 retryproc:
1846 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1847 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1848 		aiop->aioprocflags &= ~AIOP_FREE;
1849 		wakeup(aiop->aioproc);
1850 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1851 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1852 		num_aio_resv_start++;
1853 		mtx_unlock(&aio_job_mtx);
1854 		error = aio_newproc(&num_aio_resv_start);
1855 		mtx_lock(&aio_job_mtx);
1856 		if (error) {
1857 			num_aio_resv_start--;
1858 			goto retryproc;
1859 		}
1860 	} else {
1861 		ret = -1;
1862 	}
1863 	return (ret);
1864 }
1865 
1866 static void
1867 aio_kick_helper(void *context, int pending)
1868 {
1869 	struct proc *userp = context;
1870 
1871 	mtx_lock(&aio_job_mtx);
1872 	while (--pending >= 0) {
1873 		if (aio_kick(userp))
1874 			break;
1875 	}
1876 	mtx_unlock(&aio_job_mtx);
1877 }
1878 
1879 /*
1880  * Support the aio_return system call, as a side-effect, kernel resources are
1881  * released.
1882  */
1883 static int
1884 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1885 {
1886 	struct proc *p = td->td_proc;
1887 	struct kaiocb *job;
1888 	struct kaioinfo *ki;
1889 	long status, error;
1890 
1891 	ki = p->p_aioinfo;
1892 	if (ki == NULL)
1893 		return (EINVAL);
1894 	AIO_LOCK(ki);
1895 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1896 		if (job->ujob == ujob)
1897 			break;
1898 	}
1899 	if (job != NULL) {
1900 		MPASS(job->jobflags & KAIOCB_FINISHED);
1901 		status = job->uaiocb._aiocb_private.status;
1902 		error = job->uaiocb._aiocb_private.error;
1903 		td->td_retval[0] = status;
1904 		td->td_ru.ru_oublock += job->outblock;
1905 		td->td_ru.ru_inblock += job->inblock;
1906 		td->td_ru.ru_msgsnd += job->msgsnd;
1907 		td->td_ru.ru_msgrcv += job->msgrcv;
1908 		aio_free_entry(job);
1909 		AIO_UNLOCK(ki);
1910 		ops->store_error(ujob, error);
1911 		ops->store_status(ujob, status);
1912 	} else {
1913 		error = EINVAL;
1914 		AIO_UNLOCK(ki);
1915 	}
1916 	return (error);
1917 }
1918 
1919 int
1920 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1921 {
1922 
1923 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1924 }
1925 
1926 /*
1927  * Allow a process to wakeup when any of the I/O requests are completed.
1928  */
1929 static int
1930 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1931     struct timespec *ts)
1932 {
1933 	struct proc *p = td->td_proc;
1934 	struct timeval atv;
1935 	struct kaioinfo *ki;
1936 	struct kaiocb *firstjob, *job;
1937 	int error, i, timo;
1938 
1939 	timo = 0;
1940 	if (ts) {
1941 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1942 			return (EINVAL);
1943 
1944 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1945 		if (itimerfix(&atv))
1946 			return (EINVAL);
1947 		timo = tvtohz(&atv);
1948 	}
1949 
1950 	ki = p->p_aioinfo;
1951 	if (ki == NULL)
1952 		return (EAGAIN);
1953 
1954 	if (njoblist == 0)
1955 		return (0);
1956 
1957 	AIO_LOCK(ki);
1958 	for (;;) {
1959 		firstjob = NULL;
1960 		error = 0;
1961 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1962 			for (i = 0; i < njoblist; i++) {
1963 				if (job->ujob == ujoblist[i]) {
1964 					if (firstjob == NULL)
1965 						firstjob = job;
1966 					if (job->jobflags & KAIOCB_FINISHED)
1967 						goto RETURN;
1968 				}
1969 			}
1970 		}
1971 		/* All tasks were finished. */
1972 		if (firstjob == NULL)
1973 			break;
1974 
1975 		ki->kaio_flags |= KAIO_WAKEUP;
1976 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1977 		    "aiospn", timo);
1978 		if (error == ERESTART)
1979 			error = EINTR;
1980 		if (error)
1981 			break;
1982 	}
1983 RETURN:
1984 	AIO_UNLOCK(ki);
1985 	return (error);
1986 }
1987 
1988 int
1989 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1990 {
1991 	struct timespec ts, *tsp;
1992 	struct aiocb **ujoblist;
1993 	int error;
1994 
1995 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1996 		return (EINVAL);
1997 
1998 	if (uap->timeout) {
1999 		/* Get timespec struct. */
2000 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2001 			return (error);
2002 		tsp = &ts;
2003 	} else
2004 		tsp = NULL;
2005 
2006 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2007 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2008 	if (error == 0)
2009 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2010 	free(ujoblist, M_AIO);
2011 	return (error);
2012 }
2013 
2014 /*
2015  * aio_cancel cancels any non-bio aio operations not currently in progress.
2016  */
2017 int
2018 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2019 {
2020 	struct proc *p = td->td_proc;
2021 	struct kaioinfo *ki;
2022 	struct kaiocb *job, *jobn;
2023 	struct file *fp;
2024 	int error;
2025 	int cancelled = 0;
2026 	int notcancelled = 0;
2027 	struct vnode *vp;
2028 
2029 	/* Lookup file object. */
2030 	error = fget(td, uap->fd, &cap_no_rights, &fp);
2031 	if (error)
2032 		return (error);
2033 
2034 	ki = p->p_aioinfo;
2035 	if (ki == NULL)
2036 		goto done;
2037 
2038 	if (fp->f_type == DTYPE_VNODE) {
2039 		vp = fp->f_vnode;
2040 		if (vn_isdisk(vp)) {
2041 			fdrop(fp, td);
2042 			td->td_retval[0] = AIO_NOTCANCELED;
2043 			return (0);
2044 		}
2045 	}
2046 
2047 	AIO_LOCK(ki);
2048 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2049 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2050 		    ((uap->aiocbp == NULL) ||
2051 		     (uap->aiocbp == job->ujob))) {
2052 			if (aio_cancel_job(p, ki, job)) {
2053 				cancelled++;
2054 			} else {
2055 				notcancelled++;
2056 			}
2057 			if (uap->aiocbp != NULL)
2058 				break;
2059 		}
2060 	}
2061 	AIO_UNLOCK(ki);
2062 
2063 done:
2064 	fdrop(fp, td);
2065 
2066 	if (uap->aiocbp != NULL) {
2067 		if (cancelled) {
2068 			td->td_retval[0] = AIO_CANCELED;
2069 			return (0);
2070 		}
2071 	}
2072 
2073 	if (notcancelled) {
2074 		td->td_retval[0] = AIO_NOTCANCELED;
2075 		return (0);
2076 	}
2077 
2078 	if (cancelled) {
2079 		td->td_retval[0] = AIO_CANCELED;
2080 		return (0);
2081 	}
2082 
2083 	td->td_retval[0] = AIO_ALLDONE;
2084 
2085 	return (0);
2086 }
2087 
2088 /*
2089  * aio_error is implemented in the kernel level for compatibility purposes
2090  * only.  For a user mode async implementation, it would be best to do it in
2091  * a userland subroutine.
2092  */
2093 static int
2094 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2095 {
2096 	struct proc *p = td->td_proc;
2097 	struct kaiocb *job;
2098 	struct kaioinfo *ki;
2099 	int status;
2100 
2101 	ki = p->p_aioinfo;
2102 	if (ki == NULL) {
2103 		td->td_retval[0] = EINVAL;
2104 		return (0);
2105 	}
2106 
2107 	AIO_LOCK(ki);
2108 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2109 		if (job->ujob == ujob) {
2110 			if (job->jobflags & KAIOCB_FINISHED)
2111 				td->td_retval[0] =
2112 					job->uaiocb._aiocb_private.error;
2113 			else
2114 				td->td_retval[0] = EINPROGRESS;
2115 			AIO_UNLOCK(ki);
2116 			return (0);
2117 		}
2118 	}
2119 	AIO_UNLOCK(ki);
2120 
2121 	/*
2122 	 * Hack for failure of aio_aqueue.
2123 	 */
2124 	status = ops->fetch_status(ujob);
2125 	if (status == -1) {
2126 		td->td_retval[0] = ops->fetch_error(ujob);
2127 		return (0);
2128 	}
2129 
2130 	td->td_retval[0] = EINVAL;
2131 	return (0);
2132 }
2133 
2134 int
2135 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2136 {
2137 
2138 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2139 }
2140 
2141 /* syscall - asynchronous read from a file (REALTIME) */
2142 #ifdef COMPAT_FREEBSD6
2143 int
2144 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2145 {
2146 
2147 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2148 	    &aiocb_ops_osigevent));
2149 }
2150 #endif
2151 
2152 int
2153 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2154 {
2155 
2156 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2157 }
2158 
2159 int
2160 sys_aio_readv(struct thread *td, struct aio_readv_args *uap)
2161 {
2162 
2163 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops));
2164 }
2165 
2166 /* syscall - asynchronous write to a file (REALTIME) */
2167 #ifdef COMPAT_FREEBSD6
2168 int
2169 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2170 {
2171 
2172 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2173 	    &aiocb_ops_osigevent));
2174 }
2175 #endif
2176 
2177 int
2178 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2179 {
2180 
2181 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2182 }
2183 
2184 int
2185 sys_aio_writev(struct thread *td, struct aio_writev_args *uap)
2186 {
2187 
2188 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops));
2189 }
2190 
2191 int
2192 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2193 {
2194 
2195 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2196 }
2197 
2198 static int
2199 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2200     struct aiocb **acb_list, int nent, struct sigevent *sig,
2201     struct aiocb_ops *ops)
2202 {
2203 	struct proc *p = td->td_proc;
2204 	struct aiocb *job;
2205 	struct kaioinfo *ki;
2206 	struct aioliojob *lj;
2207 	struct kevent kev;
2208 	int error;
2209 	int nagain, nerror;
2210 	int i;
2211 
2212 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2213 		return (EINVAL);
2214 
2215 	if (nent < 0 || nent > max_aio_queue_per_proc)
2216 		return (EINVAL);
2217 
2218 	if (p->p_aioinfo == NULL)
2219 		aio_init_aioinfo(p);
2220 
2221 	ki = p->p_aioinfo;
2222 
2223 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2224 	lj->lioj_flags = 0;
2225 	lj->lioj_count = 0;
2226 	lj->lioj_finished_count = 0;
2227 	lj->lioj_signal.sigev_notify = SIGEV_NONE;
2228 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2229 	ksiginfo_init(&lj->lioj_ksi);
2230 
2231 	/*
2232 	 * Setup signal.
2233 	 */
2234 	if (sig && (mode == LIO_NOWAIT)) {
2235 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2236 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2237 			/* Assume only new style KEVENT */
2238 			memset(&kev, 0, sizeof(kev));
2239 			kev.filter = EVFILT_LIO;
2240 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2241 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2242 			kev.data = (intptr_t)lj;
2243 			/* pass user defined sigval data */
2244 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2245 			error = kqfd_register(
2246 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2247 			    M_WAITOK);
2248 			if (error) {
2249 				uma_zfree(aiolio_zone, lj);
2250 				return (error);
2251 			}
2252 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2253 			;
2254 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2255 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2256 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2257 					uma_zfree(aiolio_zone, lj);
2258 					return EINVAL;
2259 				}
2260 				lj->lioj_flags |= LIOJ_SIGNAL;
2261 		} else {
2262 			uma_zfree(aiolio_zone, lj);
2263 			return EINVAL;
2264 		}
2265 	}
2266 
2267 	AIO_LOCK(ki);
2268 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2269 	/*
2270 	 * Add extra aiocb count to avoid the lio to be freed
2271 	 * by other threads doing aio_waitcomplete or aio_return,
2272 	 * and prevent event from being sent until we have queued
2273 	 * all tasks.
2274 	 */
2275 	lj->lioj_count = 1;
2276 	AIO_UNLOCK(ki);
2277 
2278 	/*
2279 	 * Get pointers to the list of I/O requests.
2280 	 */
2281 	nagain = 0;
2282 	nerror = 0;
2283 	for (i = 0; i < nent; i++) {
2284 		job = acb_list[i];
2285 		if (job != NULL) {
2286 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2287 			if (error == EAGAIN)
2288 				nagain++;
2289 			else if (error != 0)
2290 				nerror++;
2291 		}
2292 	}
2293 
2294 	error = 0;
2295 	AIO_LOCK(ki);
2296 	if (mode == LIO_WAIT) {
2297 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2298 			ki->kaio_flags |= KAIO_WAKEUP;
2299 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2300 			    PRIBIO | PCATCH, "aiospn", 0);
2301 			if (error == ERESTART)
2302 				error = EINTR;
2303 			if (error)
2304 				break;
2305 		}
2306 	} else {
2307 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2308 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2309 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2310 				KNOTE_LOCKED(&lj->klist, 1);
2311 			}
2312 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2313 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2314 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2315 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2316 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2317 				    lj->lioj_count != 1);
2318 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2319 			}
2320 		}
2321 	}
2322 	lj->lioj_count--;
2323 	if (lj->lioj_count == 0) {
2324 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2325 		knlist_delete(&lj->klist, curthread, 1);
2326 		PROC_LOCK(p);
2327 		sigqueue_take(&lj->lioj_ksi);
2328 		PROC_UNLOCK(p);
2329 		AIO_UNLOCK(ki);
2330 		uma_zfree(aiolio_zone, lj);
2331 	} else
2332 		AIO_UNLOCK(ki);
2333 
2334 	if (nerror)
2335 		return (EIO);
2336 	else if (nagain)
2337 		return (EAGAIN);
2338 	else
2339 		return (error);
2340 }
2341 
2342 /* syscall - list directed I/O (REALTIME) */
2343 #ifdef COMPAT_FREEBSD6
2344 int
2345 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2346 {
2347 	struct aiocb **acb_list;
2348 	struct sigevent *sigp, sig;
2349 	struct osigevent osig;
2350 	int error, nent;
2351 
2352 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2353 		return (EINVAL);
2354 
2355 	nent = uap->nent;
2356 	if (nent < 0 || nent > max_aio_queue_per_proc)
2357 		return (EINVAL);
2358 
2359 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2360 		error = copyin(uap->sig, &osig, sizeof(osig));
2361 		if (error)
2362 			return (error);
2363 		error = convert_old_sigevent(&osig, &sig);
2364 		if (error)
2365 			return (error);
2366 		sigp = &sig;
2367 	} else
2368 		sigp = NULL;
2369 
2370 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2371 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2372 	if (error == 0)
2373 		error = kern_lio_listio(td, uap->mode,
2374 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2375 		    &aiocb_ops_osigevent);
2376 	free(acb_list, M_LIO);
2377 	return (error);
2378 }
2379 #endif
2380 
2381 /* syscall - list directed I/O (REALTIME) */
2382 int
2383 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2384 {
2385 	struct aiocb **acb_list;
2386 	struct sigevent *sigp, sig;
2387 	int error, nent;
2388 
2389 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2390 		return (EINVAL);
2391 
2392 	nent = uap->nent;
2393 	if (nent < 0 || nent > max_aio_queue_per_proc)
2394 		return (EINVAL);
2395 
2396 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2397 		error = copyin(uap->sig, &sig, sizeof(sig));
2398 		if (error)
2399 			return (error);
2400 		sigp = &sig;
2401 	} else
2402 		sigp = NULL;
2403 
2404 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2405 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2406 	if (error == 0)
2407 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2408 		    nent, sigp, &aiocb_ops);
2409 	free(acb_list, M_LIO);
2410 	return (error);
2411 }
2412 
2413 static void
2414 aio_biocleanup(struct bio *bp)
2415 {
2416 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2417 	struct kaioinfo *ki;
2418 	struct buf *pbuf = (struct buf *)bp->bio_caller2;
2419 
2420 	/* Release mapping into kernel space. */
2421 	if (pbuf != NULL) {
2422 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2423 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2424 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2425 		uma_zfree(pbuf_zone, pbuf);
2426 		atomic_subtract_int(&num_buf_aio, 1);
2427 		ki = job->userproc->p_aioinfo;
2428 		AIO_LOCK(ki);
2429 		ki->kaio_buffer_count--;
2430 		AIO_UNLOCK(ki);
2431 	} else {
2432 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2433 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2434 		free(bp->bio_ma, M_TEMP);
2435 		atomic_subtract_int(&num_unmapped_aio, 1);
2436 	}
2437 	g_destroy_bio(bp);
2438 }
2439 
2440 static void
2441 aio_biowakeup(struct bio *bp)
2442 {
2443 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2444 	size_t nbytes;
2445 	long bcount = bp->bio_bcount;
2446 	long resid = bp->bio_resid;
2447 	int opcode, nblks;
2448 	int bio_error = bp->bio_error;
2449 	uint16_t flags = bp->bio_flags;
2450 
2451 	opcode = job->uaiocb.aio_lio_opcode;
2452 
2453 	aio_biocleanup(bp);
2454 
2455 	nbytes = bcount - resid;
2456 	atomic_add_acq_long(&job->nbytes, nbytes);
2457 	nblks = btodb(nbytes);
2458 
2459 	/*
2460 	 * If multiple bios experienced an error, the job will reflect the
2461 	 * error of whichever failed bio completed last.
2462 	 */
2463 	if (flags & BIO_ERROR)
2464 		atomic_store_int(&job->error, bio_error);
2465 	if (opcode & LIO_WRITE)
2466 		atomic_add_int(&job->outblock, nblks);
2467 	else
2468 		atomic_add_int(&job->inblock, nblks);
2469 
2470 	if (refcount_release(&job->nbio)) {
2471 		bio_error = atomic_load_int(&job->error);
2472 		if (bio_error != 0)
2473 			aio_complete(job, -1, bio_error);
2474 		else
2475 			aio_complete(job, atomic_load_long(&job->nbytes), 0);
2476 	}
2477 }
2478 
2479 /* syscall - wait for the next completion of an aio request */
2480 static int
2481 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2482     struct timespec *ts, struct aiocb_ops *ops)
2483 {
2484 	struct proc *p = td->td_proc;
2485 	struct timeval atv;
2486 	struct kaioinfo *ki;
2487 	struct kaiocb *job;
2488 	struct aiocb *ujob;
2489 	long error, status;
2490 	int timo;
2491 
2492 	ops->store_aiocb(ujobp, NULL);
2493 
2494 	if (ts == NULL) {
2495 		timo = 0;
2496 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2497 		timo = -1;
2498 	} else {
2499 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2500 			return (EINVAL);
2501 
2502 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2503 		if (itimerfix(&atv))
2504 			return (EINVAL);
2505 		timo = tvtohz(&atv);
2506 	}
2507 
2508 	if (p->p_aioinfo == NULL)
2509 		aio_init_aioinfo(p);
2510 	ki = p->p_aioinfo;
2511 
2512 	error = 0;
2513 	job = NULL;
2514 	AIO_LOCK(ki);
2515 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2516 		if (timo == -1) {
2517 			error = EWOULDBLOCK;
2518 			break;
2519 		}
2520 		ki->kaio_flags |= KAIO_WAKEUP;
2521 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2522 		    "aiowc", timo);
2523 		if (timo && error == ERESTART)
2524 			error = EINTR;
2525 		if (error)
2526 			break;
2527 	}
2528 
2529 	if (job != NULL) {
2530 		MPASS(job->jobflags & KAIOCB_FINISHED);
2531 		ujob = job->ujob;
2532 		status = job->uaiocb._aiocb_private.status;
2533 		error = job->uaiocb._aiocb_private.error;
2534 		td->td_retval[0] = status;
2535 		td->td_ru.ru_oublock += job->outblock;
2536 		td->td_ru.ru_inblock += job->inblock;
2537 		td->td_ru.ru_msgsnd += job->msgsnd;
2538 		td->td_ru.ru_msgrcv += job->msgrcv;
2539 		aio_free_entry(job);
2540 		AIO_UNLOCK(ki);
2541 		ops->store_aiocb(ujobp, ujob);
2542 		ops->store_error(ujob, error);
2543 		ops->store_status(ujob, status);
2544 	} else
2545 		AIO_UNLOCK(ki);
2546 
2547 	return (error);
2548 }
2549 
2550 int
2551 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2552 {
2553 	struct timespec ts, *tsp;
2554 	int error;
2555 
2556 	if (uap->timeout) {
2557 		/* Get timespec struct. */
2558 		error = copyin(uap->timeout, &ts, sizeof(ts));
2559 		if (error)
2560 			return (error);
2561 		tsp = &ts;
2562 	} else
2563 		tsp = NULL;
2564 
2565 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2566 }
2567 
2568 static int
2569 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2570     struct aiocb_ops *ops)
2571 {
2572 	int listop;
2573 
2574 	switch (op) {
2575 	case O_SYNC:
2576 		listop = LIO_SYNC;
2577 		break;
2578 	case O_DSYNC:
2579 		listop = LIO_DSYNC;
2580 		break;
2581 	default:
2582 		return (EINVAL);
2583 	}
2584 
2585 	return (aio_aqueue(td, ujob, NULL, listop, ops));
2586 }
2587 
2588 int
2589 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2590 {
2591 
2592 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2593 }
2594 
2595 /* kqueue attach function */
2596 static int
2597 filt_aioattach(struct knote *kn)
2598 {
2599 	struct kaiocb *job;
2600 
2601 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2602 
2603 	/*
2604 	 * The job pointer must be validated before using it, so
2605 	 * registration is restricted to the kernel; the user cannot
2606 	 * set EV_FLAG1.
2607 	 */
2608 	if ((kn->kn_flags & EV_FLAG1) == 0)
2609 		return (EPERM);
2610 	kn->kn_ptr.p_aio = job;
2611 	kn->kn_flags &= ~EV_FLAG1;
2612 
2613 	knlist_add(&job->klist, kn, 0);
2614 
2615 	return (0);
2616 }
2617 
2618 /* kqueue detach function */
2619 static void
2620 filt_aiodetach(struct knote *kn)
2621 {
2622 	struct knlist *knl;
2623 
2624 	knl = &kn->kn_ptr.p_aio->klist;
2625 	knl->kl_lock(knl->kl_lockarg);
2626 	if (!knlist_empty(knl))
2627 		knlist_remove(knl, kn, 1);
2628 	knl->kl_unlock(knl->kl_lockarg);
2629 }
2630 
2631 /* kqueue filter function */
2632 /*ARGSUSED*/
2633 static int
2634 filt_aio(struct knote *kn, long hint)
2635 {
2636 	struct kaiocb *job = kn->kn_ptr.p_aio;
2637 
2638 	kn->kn_data = job->uaiocb._aiocb_private.error;
2639 	if (!(job->jobflags & KAIOCB_FINISHED))
2640 		return (0);
2641 	kn->kn_flags |= EV_EOF;
2642 	return (1);
2643 }
2644 
2645 /* kqueue attach function */
2646 static int
2647 filt_lioattach(struct knote *kn)
2648 {
2649 	struct aioliojob *lj;
2650 
2651 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2652 
2653 	/*
2654 	 * The aioliojob pointer must be validated before using it, so
2655 	 * registration is restricted to the kernel; the user cannot
2656 	 * set EV_FLAG1.
2657 	 */
2658 	if ((kn->kn_flags & EV_FLAG1) == 0)
2659 		return (EPERM);
2660 	kn->kn_ptr.p_lio = lj;
2661 	kn->kn_flags &= ~EV_FLAG1;
2662 
2663 	knlist_add(&lj->klist, kn, 0);
2664 
2665 	return (0);
2666 }
2667 
2668 /* kqueue detach function */
2669 static void
2670 filt_liodetach(struct knote *kn)
2671 {
2672 	struct knlist *knl;
2673 
2674 	knl = &kn->kn_ptr.p_lio->klist;
2675 	knl->kl_lock(knl->kl_lockarg);
2676 	if (!knlist_empty(knl))
2677 		knlist_remove(knl, kn, 1);
2678 	knl->kl_unlock(knl->kl_lockarg);
2679 }
2680 
2681 /* kqueue filter function */
2682 /*ARGSUSED*/
2683 static int
2684 filt_lio(struct knote *kn, long hint)
2685 {
2686 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2687 
2688 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2689 }
2690 
2691 #ifdef COMPAT_FREEBSD32
2692 #include <sys/mount.h>
2693 #include <sys/socket.h>
2694 #include <sys/sysent.h>
2695 #include <compat/freebsd32/freebsd32.h>
2696 #include <compat/freebsd32/freebsd32_proto.h>
2697 #include <compat/freebsd32/freebsd32_signal.h>
2698 #include <compat/freebsd32/freebsd32_syscall.h>
2699 #include <compat/freebsd32/freebsd32_util.h>
2700 
2701 struct __aiocb_private32 {
2702 	int32_t	status;
2703 	int32_t	error;
2704 	uint32_t spare;
2705 };
2706 
2707 #ifdef COMPAT_FREEBSD6
2708 typedef struct oaiocb32 {
2709 	int	aio_fildes;		/* File descriptor */
2710 	uint64_t aio_offset __packed;	/* File offset for I/O */
2711 	uint32_t aio_buf;		/* I/O buffer in process space */
2712 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2713 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2714 	int	aio_lio_opcode;		/* LIO opcode */
2715 	int	aio_reqprio;		/* Request priority -- ignored */
2716 	struct	__aiocb_private32 _aiocb_private;
2717 } oaiocb32_t;
2718 #endif
2719 
2720 typedef struct aiocb32 {
2721 	int32_t	aio_fildes;		/* File descriptor */
2722 	uint64_t aio_offset __packed;	/* File offset for I/O */
2723 	uint32_t aio_buf;	/* I/O buffer in process space */
2724 	uint32_t aio_nbytes;	/* Number of bytes for I/O */
2725 	int	__spare__[2];
2726 	uint32_t __spare2__;
2727 	int	aio_lio_opcode;		/* LIO opcode */
2728 	int	aio_reqprio;		/* Request priority -- ignored */
2729 	struct	__aiocb_private32 _aiocb_private;
2730 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2731 } aiocb32_t;
2732 
2733 #ifdef COMPAT_FREEBSD6
2734 static int
2735 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2736 {
2737 
2738 	/*
2739 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2740 	 * supported by AIO with the old sigevent structure.
2741 	 */
2742 	CP(*osig, *nsig, sigev_notify);
2743 	switch (nsig->sigev_notify) {
2744 	case SIGEV_NONE:
2745 		break;
2746 	case SIGEV_SIGNAL:
2747 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2748 		break;
2749 	case SIGEV_KEVENT:
2750 		nsig->sigev_notify_kqueue =
2751 		    osig->__sigev_u.__sigev_notify_kqueue;
2752 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2753 		break;
2754 	default:
2755 		return (EINVAL);
2756 	}
2757 	return (0);
2758 }
2759 
2760 static int
2761 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
2762     int type __unused)
2763 {
2764 	struct oaiocb32 job32;
2765 	struct aiocb *kcb = &kjob->uaiocb;
2766 	int error;
2767 
2768 	bzero(kcb, sizeof(struct aiocb));
2769 	error = copyin(ujob, &job32, sizeof(job32));
2770 	if (error)
2771 		return (error);
2772 
2773 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
2774 
2775 	CP(job32, *kcb, aio_fildes);
2776 	CP(job32, *kcb, aio_offset);
2777 	PTRIN_CP(job32, *kcb, aio_buf);
2778 	CP(job32, *kcb, aio_nbytes);
2779 	CP(job32, *kcb, aio_lio_opcode);
2780 	CP(job32, *kcb, aio_reqprio);
2781 	CP(job32, *kcb, _aiocb_private.status);
2782 	CP(job32, *kcb, _aiocb_private.error);
2783 	return (convert_old_sigevent32(&job32.aio_sigevent,
2784 	    &kcb->aio_sigevent));
2785 }
2786 #endif
2787 
2788 static int
2789 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
2790 {
2791 	struct aiocb32 job32;
2792 	struct aiocb *kcb = &kjob->uaiocb;
2793 	struct iovec32 *iov32;
2794 	int error;
2795 
2796 	error = copyin(ujob, &job32, sizeof(job32));
2797 	if (error)
2798 		return (error);
2799 	CP(job32, *kcb, aio_fildes);
2800 	CP(job32, *kcb, aio_offset);
2801 	CP(job32, *kcb, aio_lio_opcode);
2802 	if (type == LIO_NOP)
2803 		type = kcb->aio_lio_opcode;
2804 	if (type & LIO_VECTORED) {
2805 		iov32 = PTRIN(job32.aio_iov);
2806 		CP(job32, *kcb, aio_iovcnt);
2807 		/* malloc a uio and copy in the iovec */
2808 		error = freebsd32_copyinuio(iov32,
2809 		    kcb->aio_iovcnt, &kjob->uiop);
2810 		if (error)
2811 			return (error);
2812 	} else {
2813 		PTRIN_CP(job32, *kcb, aio_buf);
2814 		CP(job32, *kcb, aio_nbytes);
2815 	}
2816 	CP(job32, *kcb, aio_reqprio);
2817 	CP(job32, *kcb, _aiocb_private.status);
2818 	CP(job32, *kcb, _aiocb_private.error);
2819 	error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent);
2820 
2821 	return (error);
2822 }
2823 
2824 static long
2825 aiocb32_fetch_status(struct aiocb *ujob)
2826 {
2827 	struct aiocb32 *ujob32;
2828 
2829 	ujob32 = (struct aiocb32 *)ujob;
2830 	return (fuword32(&ujob32->_aiocb_private.status));
2831 }
2832 
2833 static long
2834 aiocb32_fetch_error(struct aiocb *ujob)
2835 {
2836 	struct aiocb32 *ujob32;
2837 
2838 	ujob32 = (struct aiocb32 *)ujob;
2839 	return (fuword32(&ujob32->_aiocb_private.error));
2840 }
2841 
2842 static int
2843 aiocb32_store_status(struct aiocb *ujob, long status)
2844 {
2845 	struct aiocb32 *ujob32;
2846 
2847 	ujob32 = (struct aiocb32 *)ujob;
2848 	return (suword32(&ujob32->_aiocb_private.status, status));
2849 }
2850 
2851 static int
2852 aiocb32_store_error(struct aiocb *ujob, long error)
2853 {
2854 	struct aiocb32 *ujob32;
2855 
2856 	ujob32 = (struct aiocb32 *)ujob;
2857 	return (suword32(&ujob32->_aiocb_private.error, error));
2858 }
2859 
2860 static int
2861 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2862 {
2863 
2864 	return (suword32(ujobp, (long)ujob));
2865 }
2866 
2867 static struct aiocb_ops aiocb32_ops = {
2868 	.aio_copyin = aiocb32_copyin,
2869 	.fetch_status = aiocb32_fetch_status,
2870 	.fetch_error = aiocb32_fetch_error,
2871 	.store_status = aiocb32_store_status,
2872 	.store_error = aiocb32_store_error,
2873 	.store_aiocb = aiocb32_store_aiocb,
2874 };
2875 
2876 #ifdef COMPAT_FREEBSD6
2877 static struct aiocb_ops aiocb32_ops_osigevent = {
2878 	.aio_copyin = aiocb32_copyin_old_sigevent,
2879 	.fetch_status = aiocb32_fetch_status,
2880 	.fetch_error = aiocb32_fetch_error,
2881 	.store_status = aiocb32_store_status,
2882 	.store_error = aiocb32_store_error,
2883 	.store_aiocb = aiocb32_store_aiocb,
2884 };
2885 #endif
2886 
2887 int
2888 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2889 {
2890 
2891 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2892 }
2893 
2894 int
2895 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2896 {
2897 	struct timespec32 ts32;
2898 	struct timespec ts, *tsp;
2899 	struct aiocb **ujoblist;
2900 	uint32_t *ujoblist32;
2901 	int error, i;
2902 
2903 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2904 		return (EINVAL);
2905 
2906 	if (uap->timeout) {
2907 		/* Get timespec struct. */
2908 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2909 			return (error);
2910 		CP(ts32, ts, tv_sec);
2911 		CP(ts32, ts, tv_nsec);
2912 		tsp = &ts;
2913 	} else
2914 		tsp = NULL;
2915 
2916 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2917 	ujoblist32 = (uint32_t *)ujoblist;
2918 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2919 	    sizeof(ujoblist32[0]));
2920 	if (error == 0) {
2921 		for (i = uap->nent - 1; i >= 0; i--)
2922 			ujoblist[i] = PTRIN(ujoblist32[i]);
2923 
2924 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2925 	}
2926 	free(ujoblist, M_AIO);
2927 	return (error);
2928 }
2929 
2930 int
2931 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2932 {
2933 
2934 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2935 }
2936 
2937 #ifdef COMPAT_FREEBSD6
2938 int
2939 freebsd6_freebsd32_aio_read(struct thread *td,
2940     struct freebsd6_freebsd32_aio_read_args *uap)
2941 {
2942 
2943 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2944 	    &aiocb32_ops_osigevent));
2945 }
2946 #endif
2947 
2948 int
2949 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2950 {
2951 
2952 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2953 	    &aiocb32_ops));
2954 }
2955 
2956 int
2957 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap)
2958 {
2959 
2960 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV,
2961 	    &aiocb32_ops));
2962 }
2963 
2964 #ifdef COMPAT_FREEBSD6
2965 int
2966 freebsd6_freebsd32_aio_write(struct thread *td,
2967     struct freebsd6_freebsd32_aio_write_args *uap)
2968 {
2969 
2970 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2971 	    &aiocb32_ops_osigevent));
2972 }
2973 #endif
2974 
2975 int
2976 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2977 {
2978 
2979 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2980 	    &aiocb32_ops));
2981 }
2982 
2983 int
2984 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap)
2985 {
2986 
2987 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV,
2988 	    &aiocb32_ops));
2989 }
2990 
2991 int
2992 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2993 {
2994 
2995 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2996 	    &aiocb32_ops));
2997 }
2998 
2999 int
3000 freebsd32_aio_waitcomplete(struct thread *td,
3001     struct freebsd32_aio_waitcomplete_args *uap)
3002 {
3003 	struct timespec32 ts32;
3004 	struct timespec ts, *tsp;
3005 	int error;
3006 
3007 	if (uap->timeout) {
3008 		/* Get timespec struct. */
3009 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
3010 		if (error)
3011 			return (error);
3012 		CP(ts32, ts, tv_sec);
3013 		CP(ts32, ts, tv_nsec);
3014 		tsp = &ts;
3015 	} else
3016 		tsp = NULL;
3017 
3018 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
3019 	    &aiocb32_ops));
3020 }
3021 
3022 int
3023 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
3024 {
3025 
3026 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
3027 	    &aiocb32_ops));
3028 }
3029 
3030 #ifdef COMPAT_FREEBSD6
3031 int
3032 freebsd6_freebsd32_lio_listio(struct thread *td,
3033     struct freebsd6_freebsd32_lio_listio_args *uap)
3034 {
3035 	struct aiocb **acb_list;
3036 	struct sigevent *sigp, sig;
3037 	struct osigevent32 osig;
3038 	uint32_t *acb_list32;
3039 	int error, i, nent;
3040 
3041 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3042 		return (EINVAL);
3043 
3044 	nent = uap->nent;
3045 	if (nent < 0 || nent > max_aio_queue_per_proc)
3046 		return (EINVAL);
3047 
3048 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3049 		error = copyin(uap->sig, &osig, sizeof(osig));
3050 		if (error)
3051 			return (error);
3052 		error = convert_old_sigevent32(&osig, &sig);
3053 		if (error)
3054 			return (error);
3055 		sigp = &sig;
3056 	} else
3057 		sigp = NULL;
3058 
3059 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3060 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3061 	if (error) {
3062 		free(acb_list32, M_LIO);
3063 		return (error);
3064 	}
3065 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3066 	for (i = 0; i < nent; i++)
3067 		acb_list[i] = PTRIN(acb_list32[i]);
3068 	free(acb_list32, M_LIO);
3069 
3070 	error = kern_lio_listio(td, uap->mode,
3071 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3072 	    &aiocb32_ops_osigevent);
3073 	free(acb_list, M_LIO);
3074 	return (error);
3075 }
3076 #endif
3077 
3078 int
3079 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3080 {
3081 	struct aiocb **acb_list;
3082 	struct sigevent *sigp, sig;
3083 	struct sigevent32 sig32;
3084 	uint32_t *acb_list32;
3085 	int error, i, nent;
3086 
3087 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3088 		return (EINVAL);
3089 
3090 	nent = uap->nent;
3091 	if (nent < 0 || nent > max_aio_queue_per_proc)
3092 		return (EINVAL);
3093 
3094 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3095 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3096 		if (error)
3097 			return (error);
3098 		error = convert_sigevent32(&sig32, &sig);
3099 		if (error)
3100 			return (error);
3101 		sigp = &sig;
3102 	} else
3103 		sigp = NULL;
3104 
3105 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3106 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3107 	if (error) {
3108 		free(acb_list32, M_LIO);
3109 		return (error);
3110 	}
3111 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3112 	for (i = 0; i < nent; i++)
3113 		acb_list[i] = PTRIN(acb_list32[i]);
3114 	free(acb_list32, M_LIO);
3115 
3116 	error = kern_lio_listio(td, uap->mode,
3117 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3118 	    &aiocb32_ops);
3119 	free(acb_list, M_LIO);
3120 	return (error);
3121 }
3122 
3123 #endif
3124