xref: /freebsd/sys/kern/vfs_aio.c (revision b068bb09a1a82d9fef0e939ad6135443a959e290)
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 /*
233  * AIO process info
234  */
235 #define AIOP_FREE	0x1			/* proc on free queue */
236 
237 struct aioproc {
238 	int	aioprocflags;			/* (c) AIO proc flags */
239 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
240 	struct	proc *aioproc;			/* (*) the AIO proc */
241 };
242 
243 /*
244  * data-structure for lio signal management
245  */
246 struct aioliojob {
247 	int	lioj_flags;			/* (a) listio flags */
248 	int	lioj_count;			/* (a) count of jobs */
249 	int	lioj_finished_count;		/* (a) count of finished jobs */
250 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
251 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
252 	struct	knlist klist;			/* (a) list of knotes */
253 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
254 };
255 
256 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
257 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
258 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
259 
260 /*
261  * per process aio data structure
262  */
263 struct kaioinfo {
264 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
265 	int	kaio_flags;		/* (a) per process kaio flags */
266 	int	kaio_active_count;	/* (c) number of currently used AIOs */
267 	int	kaio_count;		/* (a) size of AIO queue */
268 	int	kaio_buffer_count;	/* (a) number of bio buffers */
269 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
270 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
271 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
272 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
273 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
274 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
275 	struct	task kaio_task;		/* (*) task to kick aio processes */
276 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
277 };
278 
279 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
280 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
281 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
282 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
283 
284 #define KAIO_RUNDOWN	0x1	/* process is being run down */
285 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
286 
287 /*
288  * Operations used to interact with userland aio control blocks.
289  * Different ABIs provide their own operations.
290  */
291 struct aiocb_ops {
292 	int	(*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty);
293 	long	(*fetch_status)(struct aiocb *ujob);
294 	long	(*fetch_error)(struct aiocb *ujob);
295 	int	(*store_status)(struct aiocb *ujob, long status);
296 	int	(*store_error)(struct aiocb *ujob, long error);
297 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
298 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
299 };
300 
301 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
302 static struct sema aio_newproc_sem;
303 static struct mtx aio_job_mtx;
304 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
305 static struct unrhdr *aiod_unr;
306 
307 static void	aio_biocleanup(struct bio *bp);
308 void		aio_init_aioinfo(struct proc *p);
309 static int	aio_onceonly(void);
310 static int	aio_free_entry(struct kaiocb *job);
311 static void	aio_process_rw(struct kaiocb *job);
312 static void	aio_process_sync(struct kaiocb *job);
313 static void	aio_process_mlock(struct kaiocb *job);
314 static void	aio_schedule_fsync(void *context, int pending);
315 static int	aio_newproc(int *);
316 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
317 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
318 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
319 static void	aio_biowakeup(struct bio *bp);
320 static void	aio_proc_rundown(void *arg, struct proc *p);
321 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
322 		    struct image_params *imgp);
323 static int	aio_qbio(struct proc *p, struct kaiocb *job);
324 static void	aio_daemon(void *param);
325 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
326 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
327 static int	aio_kick(struct proc *userp);
328 static void	aio_kick_nowait(struct proc *userp);
329 static void	aio_kick_helper(void *context, int pending);
330 static int	filt_aioattach(struct knote *kn);
331 static void	filt_aiodetach(struct knote *kn);
332 static int	filt_aio(struct knote *kn, long hint);
333 static int	filt_lioattach(struct knote *kn);
334 static void	filt_liodetach(struct knote *kn);
335 static int	filt_lio(struct knote *kn, long hint);
336 
337 /*
338  * Zones for:
339  * 	kaio	Per process async io info
340  *	aiocb	async io jobs
341  *	aiolio	list io jobs
342  */
343 static uma_zone_t kaio_zone, aiocb_zone, aiolio_zone;
344 
345 /* kqueue filters for aio */
346 static struct filterops aio_filtops = {
347 	.f_isfd = 0,
348 	.f_attach = filt_aioattach,
349 	.f_detach = filt_aiodetach,
350 	.f_event = filt_aio,
351 };
352 static struct filterops lio_filtops = {
353 	.f_isfd = 0,
354 	.f_attach = filt_lioattach,
355 	.f_detach = filt_liodetach,
356 	.f_event = filt_lio
357 };
358 
359 static eventhandler_tag exit_tag, exec_tag;
360 
361 TASKQUEUE_DEFINE_THREAD(aiod_kick);
362 
363 /*
364  * Main operations function for use as a kernel module.
365  */
366 static int
367 aio_modload(struct module *module, int cmd, void *arg)
368 {
369 	int error = 0;
370 
371 	switch (cmd) {
372 	case MOD_LOAD:
373 		aio_onceonly();
374 		break;
375 	case MOD_SHUTDOWN:
376 		break;
377 	default:
378 		error = EOPNOTSUPP;
379 		break;
380 	}
381 	return (error);
382 }
383 
384 static moduledata_t aio_mod = {
385 	"aio",
386 	&aio_modload,
387 	NULL
388 };
389 
390 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
391 MODULE_VERSION(aio, 1);
392 
393 /*
394  * Startup initialization
395  */
396 static int
397 aio_onceonly(void)
398 {
399 
400 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
401 	    EVENTHANDLER_PRI_ANY);
402 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
403 	    NULL, EVENTHANDLER_PRI_ANY);
404 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
405 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
406 	TAILQ_INIT(&aio_freeproc);
407 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
408 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
409 	TAILQ_INIT(&aio_jobs);
410 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
411 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
412 	    NULL, NULL, UMA_ALIGN_PTR, 0);
413 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
414 	    NULL, NULL, UMA_ALIGN_PTR, 0);
415 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
416 	    NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
417 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
418 	jobrefid = 1;
419 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
420 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
421 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
422 
423 	return (0);
424 }
425 
426 /*
427  * Init the per-process aioinfo structure.  The aioinfo limits are set
428  * per-process for user limit (resource) management.
429  */
430 void
431 aio_init_aioinfo(struct proc *p)
432 {
433 	struct kaioinfo *ki;
434 
435 	ki = uma_zalloc(kaio_zone, M_WAITOK);
436 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
437 	ki->kaio_flags = 0;
438 	ki->kaio_active_count = 0;
439 	ki->kaio_count = 0;
440 	ki->kaio_buffer_count = 0;
441 	TAILQ_INIT(&ki->kaio_all);
442 	TAILQ_INIT(&ki->kaio_done);
443 	TAILQ_INIT(&ki->kaio_jobqueue);
444 	TAILQ_INIT(&ki->kaio_liojoblist);
445 	TAILQ_INIT(&ki->kaio_syncqueue);
446 	TAILQ_INIT(&ki->kaio_syncready);
447 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
448 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
449 	PROC_LOCK(p);
450 	if (p->p_aioinfo == NULL) {
451 		p->p_aioinfo = ki;
452 		PROC_UNLOCK(p);
453 	} else {
454 		PROC_UNLOCK(p);
455 		mtx_destroy(&ki->kaio_mtx);
456 		uma_zfree(kaio_zone, ki);
457 	}
458 
459 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
460 		aio_newproc(NULL);
461 }
462 
463 static int
464 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
465 {
466 	struct thread *td;
467 	int error;
468 
469 	error = sigev_findtd(p, sigev, &td);
470 	if (error)
471 		return (error);
472 	if (!KSI_ONQ(ksi)) {
473 		ksiginfo_set_sigev(ksi, sigev);
474 		ksi->ksi_code = SI_ASYNCIO;
475 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
476 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
477 	}
478 	PROC_UNLOCK(p);
479 	return (error);
480 }
481 
482 /*
483  * Free a job entry.  Wait for completion if it is currently active, but don't
484  * delay forever.  If we delay, we return a flag that says that we have to
485  * restart the queue scan.
486  */
487 static int
488 aio_free_entry(struct kaiocb *job)
489 {
490 	struct kaioinfo *ki;
491 	struct aioliojob *lj;
492 	struct proc *p;
493 
494 	p = job->userproc;
495 	MPASS(curproc == p);
496 	ki = p->p_aioinfo;
497 	MPASS(ki != NULL);
498 
499 	AIO_LOCK_ASSERT(ki, MA_OWNED);
500 	MPASS(job->jobflags & KAIOCB_FINISHED);
501 
502 	atomic_subtract_int(&num_queue_count, 1);
503 
504 	ki->kaio_count--;
505 	MPASS(ki->kaio_count >= 0);
506 
507 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
508 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
509 
510 	lj = job->lio;
511 	if (lj) {
512 		lj->lioj_count--;
513 		lj->lioj_finished_count--;
514 
515 		if (lj->lioj_count == 0) {
516 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
517 			/* lio is going away, we need to destroy any knotes */
518 			knlist_delete(&lj->klist, curthread, 1);
519 			PROC_LOCK(p);
520 			sigqueue_take(&lj->lioj_ksi);
521 			PROC_UNLOCK(p);
522 			uma_zfree(aiolio_zone, lj);
523 		}
524 	}
525 
526 	/* job is going away, we need to destroy any knotes */
527 	knlist_delete(&job->klist, curthread, 1);
528 	PROC_LOCK(p);
529 	sigqueue_take(&job->ksi);
530 	PROC_UNLOCK(p);
531 
532 	AIO_UNLOCK(ki);
533 
534 	/*
535 	 * The thread argument here is used to find the owning process
536 	 * and is also passed to fo_close() which may pass it to various
537 	 * places such as devsw close() routines.  Because of that, we
538 	 * need a thread pointer from the process owning the job that is
539 	 * persistent and won't disappear out from under us or move to
540 	 * another process.
541 	 *
542 	 * Currently, all the callers of this function call it to remove
543 	 * a kaiocb from the current process' job list either via a
544 	 * syscall or due to the current process calling exit() or
545 	 * execve().  Thus, we know that p == curproc.  We also know that
546 	 * curthread can't exit since we are curthread.
547 	 *
548 	 * Therefore, we use curthread as the thread to pass to
549 	 * knlist_delete().  This does mean that it is possible for the
550 	 * thread pointer at close time to differ from the thread pointer
551 	 * at open time, but this is already true of file descriptors in
552 	 * a multithreaded process.
553 	 */
554 	if (job->fd_file)
555 		fdrop(job->fd_file, curthread);
556 	crfree(job->cred);
557 	if (job->uiop != &job->uio)
558 		free(job->uiop, M_IOV);
559 	uma_zfree(aiocb_zone, job);
560 	AIO_LOCK(ki);
561 
562 	return (0);
563 }
564 
565 static void
566 aio_proc_rundown_exec(void *arg, struct proc *p,
567     struct image_params *imgp __unused)
568 {
569    	aio_proc_rundown(arg, p);
570 }
571 
572 static int
573 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
574 {
575 	aio_cancel_fn_t *func;
576 	int cancelled;
577 
578 	AIO_LOCK_ASSERT(ki, MA_OWNED);
579 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
580 		return (0);
581 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
582 	job->jobflags |= KAIOCB_CANCELLED;
583 
584 	func = job->cancel_fn;
585 
586 	/*
587 	 * If there is no cancel routine, just leave the job marked as
588 	 * cancelled.  The job should be in active use by a caller who
589 	 * should complete it normally or when it fails to install a
590 	 * cancel routine.
591 	 */
592 	if (func == NULL)
593 		return (0);
594 
595 	/*
596 	 * Set the CANCELLING flag so that aio_complete() will defer
597 	 * completions of this job.  This prevents the job from being
598 	 * freed out from under the cancel callback.  After the
599 	 * callback any deferred completion (whether from the callback
600 	 * or any other source) will be completed.
601 	 */
602 	job->jobflags |= KAIOCB_CANCELLING;
603 	AIO_UNLOCK(ki);
604 	func(job);
605 	AIO_LOCK(ki);
606 	job->jobflags &= ~KAIOCB_CANCELLING;
607 	if (job->jobflags & KAIOCB_FINISHED) {
608 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
609 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
610 		aio_bio_done_notify(p, job);
611 	} else {
612 		/*
613 		 * The cancel callback might have scheduled an
614 		 * operation to cancel this request, but it is
615 		 * only counted as cancelled if the request is
616 		 * cancelled when the callback returns.
617 		 */
618 		cancelled = 0;
619 	}
620 	return (cancelled);
621 }
622 
623 /*
624  * Rundown the jobs for a given process.
625  */
626 static void
627 aio_proc_rundown(void *arg, struct proc *p)
628 {
629 	struct kaioinfo *ki;
630 	struct aioliojob *lj;
631 	struct kaiocb *job, *jobn;
632 
633 	KASSERT(curthread->td_proc == p,
634 	    ("%s: called on non-curproc", __func__));
635 	ki = p->p_aioinfo;
636 	if (ki == NULL)
637 		return;
638 
639 	AIO_LOCK(ki);
640 	ki->kaio_flags |= KAIO_RUNDOWN;
641 
642 restart:
643 
644 	/*
645 	 * Try to cancel all pending requests. This code simulates
646 	 * aio_cancel on all pending I/O requests.
647 	 */
648 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
649 		aio_cancel_job(p, ki, job);
650 	}
651 
652 	/* Wait for all running I/O to be finished */
653 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
654 		ki->kaio_flags |= KAIO_WAKEUP;
655 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
656 		goto restart;
657 	}
658 
659 	/* Free all completed I/O requests. */
660 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
661 		aio_free_entry(job);
662 
663 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
664 		if (lj->lioj_count == 0) {
665 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
666 			knlist_delete(&lj->klist, curthread, 1);
667 			PROC_LOCK(p);
668 			sigqueue_take(&lj->lioj_ksi);
669 			PROC_UNLOCK(p);
670 			uma_zfree(aiolio_zone, lj);
671 		} else {
672 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
673 			    lj->lioj_count, lj->lioj_finished_count);
674 		}
675 	}
676 	AIO_UNLOCK(ki);
677 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
678 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
679 	mtx_destroy(&ki->kaio_mtx);
680 	uma_zfree(kaio_zone, ki);
681 	p->p_aioinfo = NULL;
682 }
683 
684 /*
685  * Select a job to run (called by an AIO daemon).
686  */
687 static struct kaiocb *
688 aio_selectjob(struct aioproc *aiop)
689 {
690 	struct kaiocb *job;
691 	struct kaioinfo *ki;
692 	struct proc *userp;
693 
694 	mtx_assert(&aio_job_mtx, MA_OWNED);
695 restart:
696 	TAILQ_FOREACH(job, &aio_jobs, list) {
697 		userp = job->userproc;
698 		ki = userp->p_aioinfo;
699 
700 		if (ki->kaio_active_count < max_aio_per_proc) {
701 			TAILQ_REMOVE(&aio_jobs, job, list);
702 			if (!aio_clear_cancel_function(job))
703 				goto restart;
704 
705 			/* Account for currently active jobs. */
706 			ki->kaio_active_count++;
707 			break;
708 		}
709 	}
710 	return (job);
711 }
712 
713 /*
714  * Move all data to a permanent storage device.  This code
715  * simulates the fsync and fdatasync syscalls.
716  */
717 static int
718 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op)
719 {
720 	struct mount *mp;
721 	int error;
722 
723 	for (;;) {
724 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
725 		if (error != 0)
726 			break;
727 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
728 		vnode_pager_clean_async(vp);
729 		if (op == LIO_DSYNC)
730 			error = VOP_FDATASYNC(vp, td);
731 		else
732 			error = VOP_FSYNC(vp, MNT_WAIT, td);
733 
734 		VOP_UNLOCK(vp);
735 		vn_finished_write(mp);
736 		if (error != ERELOOKUP)
737 			break;
738 	}
739 	return (error);
740 }
741 
742 /*
743  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
744  * does the I/O request for the non-bio version of the operations.  The normal
745  * vn operations are used, and this code should work in all instances for every
746  * type of file, including pipes, sockets, fifos, and regular files.
747  *
748  * XXX I don't think it works well for socket, pipe, and fifo.
749  */
750 static void
751 aio_process_rw(struct kaiocb *job)
752 {
753 	struct ucred *td_savedcred;
754 	struct thread *td;
755 	struct file *fp;
756 	ssize_t cnt;
757 	long msgsnd_st, msgsnd_end;
758 	long msgrcv_st, msgrcv_end;
759 	long oublock_st, oublock_end;
760 	long inblock_st, inblock_end;
761 	int error, opcode;
762 
763 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
764 	    job->uaiocb.aio_lio_opcode == LIO_READV ||
765 	    job->uaiocb.aio_lio_opcode == LIO_WRITE ||
766 	    job->uaiocb.aio_lio_opcode == LIO_WRITEV,
767 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
768 
769 	aio_switch_vmspace(job);
770 	td = curthread;
771 	td_savedcred = td->td_ucred;
772 	td->td_ucred = job->cred;
773 	job->uiop->uio_td = td;
774 	fp = job->fd_file;
775 
776 	opcode = job->uaiocb.aio_lio_opcode;
777 	cnt = job->uiop->uio_resid;
778 
779 	msgrcv_st = td->td_ru.ru_msgrcv;
780 	msgsnd_st = td->td_ru.ru_msgsnd;
781 	inblock_st = td->td_ru.ru_inblock;
782 	oublock_st = td->td_ru.ru_oublock;
783 
784 	/*
785 	 * aio_aqueue() acquires a reference to the file that is
786 	 * released in aio_free_entry().
787 	 */
788 	if (opcode == LIO_READ || opcode == LIO_READV) {
789 		if (job->uiop->uio_resid == 0)
790 			error = 0;
791 		else
792 			error = fo_read(fp, job->uiop, fp->f_cred, FOF_OFFSET,
793 			    td);
794 	} else {
795 		if (fp->f_type == DTYPE_VNODE)
796 			bwillwrite();
797 		error = fo_write(fp, job->uiop, fp->f_cred, FOF_OFFSET, td);
798 	}
799 	msgrcv_end = td->td_ru.ru_msgrcv;
800 	msgsnd_end = td->td_ru.ru_msgsnd;
801 	inblock_end = td->td_ru.ru_inblock;
802 	oublock_end = td->td_ru.ru_oublock;
803 
804 	job->msgrcv = msgrcv_end - msgrcv_st;
805 	job->msgsnd = msgsnd_end - msgsnd_st;
806 	job->inblock = inblock_end - inblock_st;
807 	job->outblock = oublock_end - oublock_st;
808 
809 	if (error != 0 && job->uiop->uio_resid != cnt) {
810 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
811 			error = 0;
812 		if (error == EPIPE && (opcode & LIO_WRITE)) {
813 			PROC_LOCK(job->userproc);
814 			kern_psignal(job->userproc, SIGPIPE);
815 			PROC_UNLOCK(job->userproc);
816 		}
817 	}
818 
819 	cnt -= job->uiop->uio_resid;
820 	td->td_ucred = td_savedcred;
821 	if (error)
822 		aio_complete(job, -1, error);
823 	else
824 		aio_complete(job, cnt, 0);
825 }
826 
827 static void
828 aio_process_sync(struct kaiocb *job)
829 {
830 	struct thread *td = curthread;
831 	struct ucred *td_savedcred = td->td_ucred;
832 	struct file *fp = job->fd_file;
833 	int error = 0;
834 
835 	KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC,
836 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
837 
838 	td->td_ucred = job->cred;
839 	if (fp->f_vnode != NULL) {
840 		error = aio_fsync_vnode(td, fp->f_vnode,
841 		    job->uaiocb.aio_lio_opcode);
842 	}
843 	td->td_ucred = td_savedcred;
844 	if (error)
845 		aio_complete(job, -1, error);
846 	else
847 		aio_complete(job, 0, 0);
848 }
849 
850 static void
851 aio_process_mlock(struct kaiocb *job)
852 {
853 	struct aiocb *cb = &job->uaiocb;
854 	int error;
855 
856 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
857 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
858 
859 	aio_switch_vmspace(job);
860 	error = kern_mlock(job->userproc, job->cred,
861 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
862 	aio_complete(job, error != 0 ? -1 : 0, error);
863 }
864 
865 static void
866 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
867 {
868 	struct aioliojob *lj;
869 	struct kaioinfo *ki;
870 	struct kaiocb *sjob, *sjobn;
871 	int lj_done;
872 	bool schedule_fsync;
873 
874 	ki = userp->p_aioinfo;
875 	AIO_LOCK_ASSERT(ki, MA_OWNED);
876 	lj = job->lio;
877 	lj_done = 0;
878 	if (lj) {
879 		lj->lioj_finished_count++;
880 		if (lj->lioj_count == lj->lioj_finished_count)
881 			lj_done = 1;
882 	}
883 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
884 	MPASS(job->jobflags & KAIOCB_FINISHED);
885 
886 	if (ki->kaio_flags & KAIO_RUNDOWN)
887 		goto notification_done;
888 
889 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
890 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
891 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
892 
893 	KNOTE_LOCKED(&job->klist, 1);
894 
895 	if (lj_done) {
896 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
897 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
898 			KNOTE_LOCKED(&lj->klist, 1);
899 		}
900 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
901 		    == LIOJ_SIGNAL &&
902 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
903 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
904 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
905 			    true);
906 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
907 		}
908 	}
909 
910 notification_done:
911 	if (job->jobflags & KAIOCB_CHECKSYNC) {
912 		schedule_fsync = false;
913 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
914 			if (job->fd_file != sjob->fd_file ||
915 			    job->seqno >= sjob->seqno)
916 				continue;
917 			if (--sjob->pending > 0)
918 				continue;
919 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
920 			if (!aio_clear_cancel_function_locked(sjob))
921 				continue;
922 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
923 			schedule_fsync = true;
924 		}
925 		if (schedule_fsync)
926 			taskqueue_enqueue(taskqueue_aiod_kick,
927 			    &ki->kaio_sync_task);
928 	}
929 	if (ki->kaio_flags & KAIO_WAKEUP) {
930 		ki->kaio_flags &= ~KAIO_WAKEUP;
931 		wakeup(&userp->p_aioinfo);
932 	}
933 }
934 
935 static void
936 aio_schedule_fsync(void *context, int pending)
937 {
938 	struct kaioinfo *ki;
939 	struct kaiocb *job;
940 
941 	ki = context;
942 	AIO_LOCK(ki);
943 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
944 		job = TAILQ_FIRST(&ki->kaio_syncready);
945 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
946 		AIO_UNLOCK(ki);
947 		aio_schedule(job, aio_process_sync);
948 		AIO_LOCK(ki);
949 	}
950 	AIO_UNLOCK(ki);
951 }
952 
953 bool
954 aio_cancel_cleared(struct kaiocb *job)
955 {
956 
957 	/*
958 	 * The caller should hold the same queue lock held when
959 	 * aio_clear_cancel_function() was called and set this flag
960 	 * ensuring this check sees an up-to-date value.  However,
961 	 * there is no way to assert that.
962 	 */
963 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
964 }
965 
966 static bool
967 aio_clear_cancel_function_locked(struct kaiocb *job)
968 {
969 
970 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
971 	MPASS(job->cancel_fn != NULL);
972 	if (job->jobflags & KAIOCB_CANCELLING) {
973 		job->jobflags |= KAIOCB_CLEARED;
974 		return (false);
975 	}
976 	job->cancel_fn = NULL;
977 	return (true);
978 }
979 
980 bool
981 aio_clear_cancel_function(struct kaiocb *job)
982 {
983 	struct kaioinfo *ki;
984 	bool ret;
985 
986 	ki = job->userproc->p_aioinfo;
987 	AIO_LOCK(ki);
988 	ret = aio_clear_cancel_function_locked(job);
989 	AIO_UNLOCK(ki);
990 	return (ret);
991 }
992 
993 static bool
994 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
995 {
996 
997 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
998 	if (job->jobflags & KAIOCB_CANCELLED)
999 		return (false);
1000 	job->cancel_fn = func;
1001 	return (true);
1002 }
1003 
1004 bool
1005 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1006 {
1007 	struct kaioinfo *ki;
1008 	bool ret;
1009 
1010 	ki = job->userproc->p_aioinfo;
1011 	AIO_LOCK(ki);
1012 	ret = aio_set_cancel_function_locked(job, func);
1013 	AIO_UNLOCK(ki);
1014 	return (ret);
1015 }
1016 
1017 void
1018 aio_complete(struct kaiocb *job, long status, int error)
1019 {
1020 	struct kaioinfo *ki;
1021 	struct proc *userp;
1022 
1023 	job->uaiocb._aiocb_private.error = error;
1024 	job->uaiocb._aiocb_private.status = status;
1025 
1026 	userp = job->userproc;
1027 	ki = userp->p_aioinfo;
1028 
1029 	AIO_LOCK(ki);
1030 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1031 	    ("duplicate aio_complete"));
1032 	job->jobflags |= KAIOCB_FINISHED;
1033 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1034 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1035 		aio_bio_done_notify(userp, job);
1036 	}
1037 	AIO_UNLOCK(ki);
1038 }
1039 
1040 void
1041 aio_cancel(struct kaiocb *job)
1042 {
1043 
1044 	aio_complete(job, -1, ECANCELED);
1045 }
1046 
1047 void
1048 aio_switch_vmspace(struct kaiocb *job)
1049 {
1050 
1051 	vmspace_switch_aio(job->userproc->p_vmspace);
1052 }
1053 
1054 /*
1055  * The AIO daemon, most of the actual work is done in aio_process_*,
1056  * but the setup (and address space mgmt) is done in this routine.
1057  */
1058 static void
1059 aio_daemon(void *_id)
1060 {
1061 	struct kaiocb *job;
1062 	struct aioproc *aiop;
1063 	struct kaioinfo *ki;
1064 	struct proc *p;
1065 	struct vmspace *myvm;
1066 	struct thread *td = curthread;
1067 	int id = (intptr_t)_id;
1068 
1069 	/*
1070 	 * Grab an extra reference on the daemon's vmspace so that it
1071 	 * doesn't get freed by jobs that switch to a different
1072 	 * vmspace.
1073 	 */
1074 	p = td->td_proc;
1075 	myvm = vmspace_acquire_ref(p);
1076 
1077 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1078 
1079 	/*
1080 	 * Allocate and ready the aio control info.  There is one aiop structure
1081 	 * per daemon.
1082 	 */
1083 	aiop = malloc(sizeof(*aiop), M_AIO, M_WAITOK);
1084 	aiop->aioproc = p;
1085 	aiop->aioprocflags = 0;
1086 
1087 	/*
1088 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1089 	 * and creating too many daemons.)
1090 	 */
1091 	sema_post(&aio_newproc_sem);
1092 
1093 	mtx_lock(&aio_job_mtx);
1094 	for (;;) {
1095 		/*
1096 		 * Take daemon off of free queue
1097 		 */
1098 		if (aiop->aioprocflags & AIOP_FREE) {
1099 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1100 			aiop->aioprocflags &= ~AIOP_FREE;
1101 		}
1102 
1103 		/*
1104 		 * Check for jobs.
1105 		 */
1106 		while ((job = aio_selectjob(aiop)) != NULL) {
1107 			mtx_unlock(&aio_job_mtx);
1108 
1109 			ki = job->userproc->p_aioinfo;
1110 			job->handle_fn(job);
1111 
1112 			mtx_lock(&aio_job_mtx);
1113 			/* Decrement the active job count. */
1114 			ki->kaio_active_count--;
1115 		}
1116 
1117 		/*
1118 		 * Disconnect from user address space.
1119 		 */
1120 		if (p->p_vmspace != myvm) {
1121 			mtx_unlock(&aio_job_mtx);
1122 			vmspace_switch_aio(myvm);
1123 			mtx_lock(&aio_job_mtx);
1124 			/*
1125 			 * We have to restart to avoid race, we only sleep if
1126 			 * no job can be selected.
1127 			 */
1128 			continue;
1129 		}
1130 
1131 		mtx_assert(&aio_job_mtx, MA_OWNED);
1132 
1133 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1134 		aiop->aioprocflags |= AIOP_FREE;
1135 
1136 		/*
1137 		 * If daemon is inactive for a long time, allow it to exit,
1138 		 * thereby freeing resources.
1139 		 */
1140 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1141 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1142 		    (aiop->aioprocflags & AIOP_FREE) &&
1143 		    num_aio_procs > target_aio_procs)
1144 			break;
1145 	}
1146 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1147 	num_aio_procs--;
1148 	mtx_unlock(&aio_job_mtx);
1149 	free(aiop, M_AIO);
1150 	free_unr(aiod_unr, id);
1151 	vmspace_free(myvm);
1152 
1153 	KASSERT(p->p_vmspace == myvm,
1154 	    ("AIOD: bad vmspace for exiting daemon"));
1155 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1156 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1157 	    refcount_load(&myvm->vm_refcnt)));
1158 	kproc_exit(0);
1159 }
1160 
1161 /*
1162  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1163  * AIO daemon modifies its environment itself.
1164  */
1165 static int
1166 aio_newproc(int *start)
1167 {
1168 	int error;
1169 	struct proc *p;
1170 	int id;
1171 
1172 	id = alloc_unr(aiod_unr);
1173 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1174 		RFNOWAIT, 0, "aiod%d", id);
1175 	if (error == 0) {
1176 		/*
1177 		 * Wait until daemon is started.
1178 		 */
1179 		sema_wait(&aio_newproc_sem);
1180 		mtx_lock(&aio_job_mtx);
1181 		num_aio_procs++;
1182 		if (start != NULL)
1183 			(*start)--;
1184 		mtx_unlock(&aio_job_mtx);
1185 	} else {
1186 		free_unr(aiod_unr, id);
1187 	}
1188 	return (error);
1189 }
1190 
1191 /*
1192  * Try the high-performance, low-overhead bio method for eligible
1193  * VCHR devices.  This method doesn't use an aio helper thread, and
1194  * thus has very low overhead.
1195  *
1196  * Assumes that the caller, aio_aqueue(), has incremented the file
1197  * structure's reference count, preventing its deallocation for the
1198  * duration of this call.
1199  */
1200 static int
1201 aio_qbio(struct proc *p, struct kaiocb *job)
1202 {
1203 	struct aiocb *cb;
1204 	struct file *fp;
1205 	struct buf *pbuf;
1206 	struct vnode *vp;
1207 	struct cdevsw *csw;
1208 	struct cdev *dev;
1209 	struct kaioinfo *ki;
1210 	struct bio **bios = NULL;
1211 	off_t offset;
1212 	int bio_cmd, error, i, iovcnt, opcode, poff, ref;
1213 	vm_prot_t prot;
1214 	bool use_unmapped;
1215 
1216 	cb = &job->uaiocb;
1217 	fp = job->fd_file;
1218 	opcode = cb->aio_lio_opcode;
1219 
1220 	if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV ||
1221 	    opcode == LIO_READ || opcode == LIO_READV))
1222 		return (-1);
1223 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1224 		return (-1);
1225 
1226 	vp = fp->f_vnode;
1227 	if (vp->v_type != VCHR)
1228 		return (-1);
1229 	if (vp->v_bufobj.bo_bsize == 0)
1230 		return (-1);
1231 
1232 	bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ;
1233 	iovcnt = job->uiop->uio_iovcnt;
1234 	if (iovcnt > max_buf_aio)
1235 		return (-1);
1236 	for (i = 0; i < iovcnt; i++) {
1237 		if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0)
1238 			return (-1);
1239 		if (job->uiop->uio_iov[i].iov_len > maxphys) {
1240 			error = -1;
1241 			return (-1);
1242 		}
1243 	}
1244 	offset = cb->aio_offset;
1245 
1246 	ref = 0;
1247 	csw = devvn_refthread(vp, &dev, &ref);
1248 	if (csw == NULL)
1249 		return (ENXIO);
1250 
1251 	if ((csw->d_flags & D_DISK) == 0) {
1252 		error = -1;
1253 		goto unref;
1254 	}
1255 	if (job->uiop->uio_resid > dev->si_iosize_max) {
1256 		error = -1;
1257 		goto unref;
1258 	}
1259 
1260 	ki = p->p_aioinfo;
1261 	job->error = 0;
1262 
1263 	use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed;
1264 	if (!use_unmapped) {
1265 		AIO_LOCK(ki);
1266 		if (ki->kaio_buffer_count + iovcnt > max_buf_aio) {
1267 			AIO_UNLOCK(ki);
1268 			error = EAGAIN;
1269 			goto unref;
1270 		}
1271 		ki->kaio_buffer_count += iovcnt;
1272 		AIO_UNLOCK(ki);
1273 	}
1274 
1275 	bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK);
1276 	refcount_init(&job->nbio, iovcnt);
1277 	for (i = 0; i < iovcnt; i++) {
1278 		struct vm_page** pages;
1279 		struct bio *bp;
1280 		void *buf;
1281 		size_t nbytes;
1282 		int npages;
1283 
1284 		buf = job->uiop->uio_iov[i].iov_base;
1285 		nbytes = job->uiop->uio_iov[i].iov_len;
1286 
1287 		bios[i] = g_alloc_bio();
1288 		bp = bios[i];
1289 
1290 		poff = (vm_offset_t)buf & PAGE_MASK;
1291 		if (use_unmapped) {
1292 			pbuf = NULL;
1293 			pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1294 			    nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1295 		} else {
1296 			pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1297 			BUF_KERNPROC(pbuf);
1298 			pages = pbuf->b_pages;
1299 		}
1300 
1301 		bp->bio_length = nbytes;
1302 		bp->bio_bcount = nbytes;
1303 		bp->bio_done = aio_biowakeup;
1304 		bp->bio_offset = offset;
1305 		bp->bio_cmd = bio_cmd;
1306 		bp->bio_dev = dev;
1307 		bp->bio_caller1 = job;
1308 		bp->bio_caller2 = pbuf;
1309 
1310 		prot = VM_PROT_READ;
1311 		if (opcode == LIO_READ || opcode == LIO_READV)
1312 			prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1313 		npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1314 		    (vm_offset_t)buf, bp->bio_length, prot, pages,
1315 		    atop(maxphys) + 1);
1316 		if (npages < 0) {
1317 			if (pbuf != NULL)
1318 				uma_zfree(pbuf_zone, pbuf);
1319 			else
1320 				free(pages, M_TEMP);
1321 			error = EFAULT;
1322 			g_destroy_bio(bp);
1323 			i--;
1324 			goto destroy_bios;
1325 		}
1326 		if (pbuf != NULL) {
1327 			pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages);
1328 			bp->bio_data = pbuf->b_data + poff;
1329 			pbuf->b_npages = npages;
1330 			atomic_add_int(&num_buf_aio, 1);
1331 		} else {
1332 			bp->bio_ma = pages;
1333 			bp->bio_ma_n = npages;
1334 			bp->bio_ma_offset = poff;
1335 			bp->bio_data = unmapped_buf;
1336 			bp->bio_flags |= BIO_UNMAPPED;
1337 			atomic_add_int(&num_unmapped_aio, 1);
1338 		}
1339 
1340 		offset += nbytes;
1341 	}
1342 
1343 	/* Perform transfer. */
1344 	for (i = 0; i < iovcnt; i++)
1345 		csw->d_strategy(bios[i]);
1346 	free(bios, M_TEMP);
1347 
1348 	dev_relthread(dev, ref);
1349 	return (0);
1350 
1351 destroy_bios:
1352 	for (; i >= 0; i--)
1353 		aio_biocleanup(bios[i]);
1354 	free(bios, M_TEMP);
1355 unref:
1356 	dev_relthread(dev, ref);
1357 	return (error);
1358 }
1359 
1360 #ifdef COMPAT_FREEBSD6
1361 static int
1362 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1363 {
1364 
1365 	/*
1366 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1367 	 * supported by AIO with the old sigevent structure.
1368 	 */
1369 	nsig->sigev_notify = osig->sigev_notify;
1370 	switch (nsig->sigev_notify) {
1371 	case SIGEV_NONE:
1372 		break;
1373 	case SIGEV_SIGNAL:
1374 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1375 		break;
1376 	case SIGEV_KEVENT:
1377 		nsig->sigev_notify_kqueue =
1378 		    osig->__sigev_u.__sigev_notify_kqueue;
1379 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1380 		break;
1381 	default:
1382 		return (EINVAL);
1383 	}
1384 	return (0);
1385 }
1386 
1387 static int
1388 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
1389     int type __unused)
1390 {
1391 	struct oaiocb *ojob;
1392 	struct aiocb *kcb = &kjob->uaiocb;
1393 	int error;
1394 
1395 	bzero(kcb, sizeof(struct aiocb));
1396 	error = copyin(ujob, kcb, sizeof(struct oaiocb));
1397 	if (error)
1398 		return (error);
1399 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
1400 	ojob = (struct oaiocb *)kcb;
1401 	return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent));
1402 }
1403 #endif
1404 
1405 static int
1406 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
1407 {
1408 	struct aiocb *kcb = &kjob->uaiocb;
1409 	int error;
1410 
1411 	error = copyin(ujob, kcb, sizeof(struct aiocb));
1412 	if (error)
1413 		return (error);
1414 	if (type == LIO_NOP)
1415 		type = kcb->aio_lio_opcode;
1416 	if (type & LIO_VECTORED) {
1417 		/* malloc a uio and copy in the iovec */
1418 		error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov),
1419 		    kcb->aio_iovcnt, &kjob->uiop);
1420 	}
1421 
1422 	return (error);
1423 }
1424 
1425 static long
1426 aiocb_fetch_status(struct aiocb *ujob)
1427 {
1428 
1429 	return (fuword(&ujob->_aiocb_private.status));
1430 }
1431 
1432 static long
1433 aiocb_fetch_error(struct aiocb *ujob)
1434 {
1435 
1436 	return (fuword(&ujob->_aiocb_private.error));
1437 }
1438 
1439 static int
1440 aiocb_store_status(struct aiocb *ujob, long status)
1441 {
1442 
1443 	return (suword(&ujob->_aiocb_private.status, status));
1444 }
1445 
1446 static int
1447 aiocb_store_error(struct aiocb *ujob, long error)
1448 {
1449 
1450 	return (suword(&ujob->_aiocb_private.error, error));
1451 }
1452 
1453 static int
1454 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1455 {
1456 
1457 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1458 }
1459 
1460 static int
1461 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1462 {
1463 
1464 	return (suword(ujobp, (long)ujob));
1465 }
1466 
1467 static struct aiocb_ops aiocb_ops = {
1468 	.aio_copyin = aiocb_copyin,
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_kernelinfo = aiocb_store_kernelinfo,
1474 	.store_aiocb = aiocb_store_aiocb,
1475 };
1476 
1477 #ifdef COMPAT_FREEBSD6
1478 static struct aiocb_ops aiocb_ops_osigevent = {
1479 	.aio_copyin = aiocb_copyin_old_sigevent,
1480 	.fetch_status = aiocb_fetch_status,
1481 	.fetch_error = aiocb_fetch_error,
1482 	.store_status = aiocb_store_status,
1483 	.store_error = aiocb_store_error,
1484 	.store_kernelinfo = aiocb_store_kernelinfo,
1485 	.store_aiocb = aiocb_store_aiocb,
1486 };
1487 #endif
1488 
1489 /*
1490  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1491  * technique is done in this code.
1492  */
1493 int
1494 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1495     int type, struct aiocb_ops *ops)
1496 {
1497 	struct proc *p = td->td_proc;
1498 	struct file *fp = NULL;
1499 	struct kaiocb *job;
1500 	struct kaioinfo *ki;
1501 	struct kevent kev;
1502 	int opcode;
1503 	int error;
1504 	int fd, kqfd;
1505 	int jid;
1506 	u_short evflags;
1507 
1508 	if (p->p_aioinfo == NULL)
1509 		aio_init_aioinfo(p);
1510 
1511 	ki = p->p_aioinfo;
1512 
1513 	ops->store_status(ujob, -1);
1514 	ops->store_error(ujob, 0);
1515 	ops->store_kernelinfo(ujob, -1);
1516 
1517 	if (num_queue_count >= max_queue_count ||
1518 	    ki->kaio_count >= max_aio_queue_per_proc) {
1519 		error = EAGAIN;
1520 		goto err1;
1521 	}
1522 
1523 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1524 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1525 
1526 	error = ops->aio_copyin(ujob, job, type);
1527 	if (error)
1528 		goto err2;
1529 
1530 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1531 		error = EINVAL;
1532 		goto err2;
1533 	}
1534 
1535 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1536 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1537 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1538 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1539 		error = EINVAL;
1540 		goto err2;
1541 	}
1542 
1543 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1544 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1545 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1546 		error = EINVAL;
1547 		goto err2;
1548 	}
1549 
1550 	/* Get the opcode. */
1551 	if (type == LIO_NOP) {
1552 		switch (job->uaiocb.aio_lio_opcode) {
1553 		case LIO_WRITE:
1554 		case LIO_WRITEV:
1555 		case LIO_NOP:
1556 		case LIO_READ:
1557 		case LIO_READV:
1558 			opcode = job->uaiocb.aio_lio_opcode;
1559 			break;
1560 		default:
1561 			error = EINVAL;
1562 			goto err2;
1563 		}
1564 	} else
1565 		opcode = job->uaiocb.aio_lio_opcode = type;
1566 
1567 	ksiginfo_init(&job->ksi);
1568 
1569 	/* Save userspace address of the job info. */
1570 	job->ujob = ujob;
1571 
1572 	/*
1573 	 * Validate the opcode and fetch the file object for the specified
1574 	 * file descriptor.
1575 	 *
1576 	 * XXXRW: Moved the opcode validation up here so that we don't
1577 	 * retrieve a file descriptor without knowing what the capabiltity
1578 	 * should be.
1579 	 */
1580 	fd = job->uaiocb.aio_fildes;
1581 	switch (opcode) {
1582 	case LIO_WRITE:
1583 	case LIO_WRITEV:
1584 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1585 		break;
1586 	case LIO_READ:
1587 	case LIO_READV:
1588 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1589 		break;
1590 	case LIO_SYNC:
1591 	case LIO_DSYNC:
1592 		error = fget(td, fd, &cap_fsync_rights, &fp);
1593 		break;
1594 	case LIO_MLOCK:
1595 		break;
1596 	case LIO_NOP:
1597 		error = fget(td, fd, &cap_no_rights, &fp);
1598 		break;
1599 	default:
1600 		error = EINVAL;
1601 	}
1602 	if (error)
1603 		goto err3;
1604 
1605 	if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) {
1606 		error = EINVAL;
1607 		goto err3;
1608 	}
1609 
1610 	if ((opcode == LIO_READ || opcode == LIO_READV ||
1611 	    opcode == LIO_WRITE || opcode == LIO_WRITEV) &&
1612 	    job->uaiocb.aio_offset < 0 &&
1613 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1614 		error = EINVAL;
1615 		goto err3;
1616 	}
1617 
1618 	if (fp != NULL && fp->f_ops == &path_fileops) {
1619 		error = EBADF;
1620 		goto err3;
1621 	}
1622 
1623 	job->fd_file = fp;
1624 
1625 	mtx_lock(&aio_job_mtx);
1626 	jid = jobrefid++;
1627 	job->seqno = jobseqno++;
1628 	mtx_unlock(&aio_job_mtx);
1629 	error = ops->store_kernelinfo(ujob, jid);
1630 	if (error) {
1631 		error = EINVAL;
1632 		goto err3;
1633 	}
1634 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1635 
1636 	if (opcode == LIO_NOP) {
1637 		fdrop(fp, td);
1638 		MPASS(job->uiop == &job->uio || job->uiop == NULL);
1639 		uma_zfree(aiocb_zone, job);
1640 		return (0);
1641 	}
1642 
1643 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1644 		goto no_kqueue;
1645 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1646 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1647 		error = EINVAL;
1648 		goto err3;
1649 	}
1650 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1651 	memset(&kev, 0, sizeof(kev));
1652 	kev.ident = (uintptr_t)job->ujob;
1653 	kev.filter = EVFILT_AIO;
1654 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1655 	kev.data = (intptr_t)job;
1656 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1657 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1658 	if (error)
1659 		goto err3;
1660 
1661 no_kqueue:
1662 
1663 	ops->store_error(ujob, EINPROGRESS);
1664 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1665 	job->userproc = p;
1666 	job->cred = crhold(td->td_ucred);
1667 	job->jobflags = KAIOCB_QUEUEING;
1668 	job->lio = lj;
1669 
1670 	if (opcode & LIO_VECTORED) {
1671 		/* Use the uio copied in by aio_copyin */
1672 		MPASS(job->uiop != &job->uio && job->uiop != NULL);
1673 	} else {
1674 		/* Setup the inline uio */
1675 		job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf;
1676 		job->iov[0].iov_len = job->uaiocb.aio_nbytes;
1677 		job->uio.uio_iov = job->iov;
1678 		job->uio.uio_iovcnt = 1;
1679 		job->uio.uio_resid = job->uaiocb.aio_nbytes;
1680 		job->uio.uio_segflg = UIO_USERSPACE;
1681 		job->uiop = &job->uio;
1682 	}
1683 	switch (opcode & (LIO_READ | LIO_WRITE)) {
1684 	case LIO_READ:
1685 		job->uiop->uio_rw = UIO_READ;
1686 		break;
1687 	case LIO_WRITE:
1688 		job->uiop->uio_rw = UIO_WRITE;
1689 		break;
1690 	}
1691 	job->uiop->uio_offset = job->uaiocb.aio_offset;
1692 	job->uiop->uio_td = td;
1693 
1694 	if (opcode == LIO_MLOCK) {
1695 		aio_schedule(job, aio_process_mlock);
1696 		error = 0;
1697 	} else if (fp->f_ops->fo_aio_queue == NULL)
1698 		error = aio_queue_file(fp, job);
1699 	else
1700 		error = fo_aio_queue(fp, job);
1701 	if (error)
1702 		goto err4;
1703 
1704 	AIO_LOCK(ki);
1705 	job->jobflags &= ~KAIOCB_QUEUEING;
1706 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1707 	ki->kaio_count++;
1708 	if (lj)
1709 		lj->lioj_count++;
1710 	atomic_add_int(&num_queue_count, 1);
1711 	if (job->jobflags & KAIOCB_FINISHED) {
1712 		/*
1713 		 * The queue callback completed the request synchronously.
1714 		 * The bulk of the completion is deferred in that case
1715 		 * until this point.
1716 		 */
1717 		aio_bio_done_notify(p, job);
1718 	} else
1719 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1720 	AIO_UNLOCK(ki);
1721 	return (0);
1722 
1723 err4:
1724 	crfree(job->cred);
1725 err3:
1726 	if (fp)
1727 		fdrop(fp, td);
1728 	knlist_delete(&job->klist, curthread, 0);
1729 err2:
1730 	if (job->uiop != &job->uio)
1731 		free(job->uiop, M_IOV);
1732 	uma_zfree(aiocb_zone, job);
1733 err1:
1734 	ops->store_error(ujob, error);
1735 	return (error);
1736 }
1737 
1738 static void
1739 aio_cancel_daemon_job(struct kaiocb *job)
1740 {
1741 
1742 	mtx_lock(&aio_job_mtx);
1743 	if (!aio_cancel_cleared(job))
1744 		TAILQ_REMOVE(&aio_jobs, job, list);
1745 	mtx_unlock(&aio_job_mtx);
1746 	aio_cancel(job);
1747 }
1748 
1749 void
1750 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1751 {
1752 
1753 	mtx_lock(&aio_job_mtx);
1754 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1755 		mtx_unlock(&aio_job_mtx);
1756 		aio_cancel(job);
1757 		return;
1758 	}
1759 	job->handle_fn = func;
1760 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1761 	aio_kick_nowait(job->userproc);
1762 	mtx_unlock(&aio_job_mtx);
1763 }
1764 
1765 static void
1766 aio_cancel_sync(struct kaiocb *job)
1767 {
1768 	struct kaioinfo *ki;
1769 
1770 	ki = job->userproc->p_aioinfo;
1771 	AIO_LOCK(ki);
1772 	if (!aio_cancel_cleared(job))
1773 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1774 	AIO_UNLOCK(ki);
1775 	aio_cancel(job);
1776 }
1777 
1778 int
1779 aio_queue_file(struct file *fp, struct kaiocb *job)
1780 {
1781 	struct kaioinfo *ki;
1782 	struct kaiocb *job2;
1783 	struct vnode *vp;
1784 	struct mount *mp;
1785 	int error;
1786 	bool safe;
1787 
1788 	ki = job->userproc->p_aioinfo;
1789 	error = aio_qbio(job->userproc, job);
1790 	if (error >= 0)
1791 		return (error);
1792 	safe = false;
1793 	if (fp->f_type == DTYPE_VNODE) {
1794 		vp = fp->f_vnode;
1795 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1796 			mp = fp->f_vnode->v_mount;
1797 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1798 				safe = true;
1799 		}
1800 	}
1801 	if (!(safe || enable_aio_unsafe)) {
1802 		counted_warning(&unsafe_warningcnt,
1803 		    "is attempting to use unsafe AIO requests");
1804 		return (EOPNOTSUPP);
1805 	}
1806 
1807 	if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) {
1808 		aio_schedule(job, aio_process_rw);
1809 		error = 0;
1810 	} else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) {
1811 		AIO_LOCK(ki);
1812 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1813 			if (job2->fd_file == job->fd_file &&
1814 			    ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) &&
1815 			    job2->seqno < job->seqno) {
1816 				job2->jobflags |= KAIOCB_CHECKSYNC;
1817 				job->pending++;
1818 			}
1819 		}
1820 		if (job->pending != 0) {
1821 			if (!aio_set_cancel_function_locked(job,
1822 				aio_cancel_sync)) {
1823 				AIO_UNLOCK(ki);
1824 				aio_cancel(job);
1825 				return (0);
1826 			}
1827 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1828 			AIO_UNLOCK(ki);
1829 			return (0);
1830 		}
1831 		AIO_UNLOCK(ki);
1832 		aio_schedule(job, aio_process_sync);
1833 		error = 0;
1834 	} else {
1835 		error = EINVAL;
1836 	}
1837 	return (error);
1838 }
1839 
1840 static void
1841 aio_kick_nowait(struct proc *userp)
1842 {
1843 	struct kaioinfo *ki = userp->p_aioinfo;
1844 	struct aioproc *aiop;
1845 
1846 	mtx_assert(&aio_job_mtx, MA_OWNED);
1847 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1848 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1849 		aiop->aioprocflags &= ~AIOP_FREE;
1850 		wakeup(aiop->aioproc);
1851 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1852 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1853 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1854 	}
1855 }
1856 
1857 static int
1858 aio_kick(struct proc *userp)
1859 {
1860 	struct kaioinfo *ki = userp->p_aioinfo;
1861 	struct aioproc *aiop;
1862 	int error, ret = 0;
1863 
1864 	mtx_assert(&aio_job_mtx, MA_OWNED);
1865 retryproc:
1866 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1867 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1868 		aiop->aioprocflags &= ~AIOP_FREE;
1869 		wakeup(aiop->aioproc);
1870 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1871 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1872 		num_aio_resv_start++;
1873 		mtx_unlock(&aio_job_mtx);
1874 		error = aio_newproc(&num_aio_resv_start);
1875 		mtx_lock(&aio_job_mtx);
1876 		if (error) {
1877 			num_aio_resv_start--;
1878 			goto retryproc;
1879 		}
1880 	} else {
1881 		ret = -1;
1882 	}
1883 	return (ret);
1884 }
1885 
1886 static void
1887 aio_kick_helper(void *context, int pending)
1888 {
1889 	struct proc *userp = context;
1890 
1891 	mtx_lock(&aio_job_mtx);
1892 	while (--pending >= 0) {
1893 		if (aio_kick(userp))
1894 			break;
1895 	}
1896 	mtx_unlock(&aio_job_mtx);
1897 }
1898 
1899 /*
1900  * Support the aio_return system call, as a side-effect, kernel resources are
1901  * released.
1902  */
1903 static int
1904 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1905 {
1906 	struct proc *p = td->td_proc;
1907 	struct kaiocb *job;
1908 	struct kaioinfo *ki;
1909 	long status, error;
1910 
1911 	ki = p->p_aioinfo;
1912 	if (ki == NULL)
1913 		return (EINVAL);
1914 	AIO_LOCK(ki);
1915 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1916 		if (job->ujob == ujob)
1917 			break;
1918 	}
1919 	if (job != NULL) {
1920 		MPASS(job->jobflags & KAIOCB_FINISHED);
1921 		status = job->uaiocb._aiocb_private.status;
1922 		error = job->uaiocb._aiocb_private.error;
1923 		td->td_retval[0] = status;
1924 		td->td_ru.ru_oublock += job->outblock;
1925 		td->td_ru.ru_inblock += job->inblock;
1926 		td->td_ru.ru_msgsnd += job->msgsnd;
1927 		td->td_ru.ru_msgrcv += job->msgrcv;
1928 		aio_free_entry(job);
1929 		AIO_UNLOCK(ki);
1930 		ops->store_error(ujob, error);
1931 		ops->store_status(ujob, status);
1932 	} else {
1933 		error = EINVAL;
1934 		AIO_UNLOCK(ki);
1935 	}
1936 	return (error);
1937 }
1938 
1939 int
1940 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1941 {
1942 
1943 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1944 }
1945 
1946 /*
1947  * Allow a process to wakeup when any of the I/O requests are completed.
1948  */
1949 static int
1950 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1951     struct timespec *ts)
1952 {
1953 	struct proc *p = td->td_proc;
1954 	struct timeval atv;
1955 	struct kaioinfo *ki;
1956 	struct kaiocb *firstjob, *job;
1957 	int error, i, timo;
1958 
1959 	timo = 0;
1960 	if (ts) {
1961 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1962 			return (EINVAL);
1963 
1964 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1965 		if (itimerfix(&atv))
1966 			return (EINVAL);
1967 		timo = tvtohz(&atv);
1968 	}
1969 
1970 	ki = p->p_aioinfo;
1971 	if (ki == NULL)
1972 		return (EAGAIN);
1973 
1974 	if (njoblist == 0)
1975 		return (0);
1976 
1977 	AIO_LOCK(ki);
1978 	for (;;) {
1979 		firstjob = NULL;
1980 		error = 0;
1981 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1982 			for (i = 0; i < njoblist; i++) {
1983 				if (job->ujob == ujoblist[i]) {
1984 					if (firstjob == NULL)
1985 						firstjob = job;
1986 					if (job->jobflags & KAIOCB_FINISHED)
1987 						goto RETURN;
1988 				}
1989 			}
1990 		}
1991 		/* All tasks were finished. */
1992 		if (firstjob == NULL)
1993 			break;
1994 
1995 		ki->kaio_flags |= KAIO_WAKEUP;
1996 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1997 		    "aiospn", timo);
1998 		if (error == ERESTART)
1999 			error = EINTR;
2000 		if (error)
2001 			break;
2002 	}
2003 RETURN:
2004 	AIO_UNLOCK(ki);
2005 	return (error);
2006 }
2007 
2008 int
2009 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2010 {
2011 	struct timespec ts, *tsp;
2012 	struct aiocb **ujoblist;
2013 	int error;
2014 
2015 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2016 		return (EINVAL);
2017 
2018 	if (uap->timeout) {
2019 		/* Get timespec struct. */
2020 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2021 			return (error);
2022 		tsp = &ts;
2023 	} else
2024 		tsp = NULL;
2025 
2026 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2027 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2028 	if (error == 0)
2029 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2030 	free(ujoblist, M_AIO);
2031 	return (error);
2032 }
2033 
2034 /*
2035  * aio_cancel cancels any non-bio aio operations not currently in progress.
2036  */
2037 int
2038 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2039 {
2040 	struct proc *p = td->td_proc;
2041 	struct kaioinfo *ki;
2042 	struct kaiocb *job, *jobn;
2043 	struct file *fp;
2044 	int error;
2045 	int cancelled = 0;
2046 	int notcancelled = 0;
2047 	struct vnode *vp;
2048 
2049 	/* Lookup file object. */
2050 	error = fget(td, uap->fd, &cap_no_rights, &fp);
2051 	if (error)
2052 		return (error);
2053 
2054 	ki = p->p_aioinfo;
2055 	if (ki == NULL)
2056 		goto done;
2057 
2058 	if (fp->f_type == DTYPE_VNODE) {
2059 		vp = fp->f_vnode;
2060 		if (vn_isdisk(vp)) {
2061 			fdrop(fp, td);
2062 			td->td_retval[0] = AIO_NOTCANCELED;
2063 			return (0);
2064 		}
2065 	}
2066 
2067 	AIO_LOCK(ki);
2068 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2069 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2070 		    ((uap->aiocbp == NULL) ||
2071 		     (uap->aiocbp == job->ujob))) {
2072 			if (aio_cancel_job(p, ki, job)) {
2073 				cancelled++;
2074 			} else {
2075 				notcancelled++;
2076 			}
2077 			if (uap->aiocbp != NULL)
2078 				break;
2079 		}
2080 	}
2081 	AIO_UNLOCK(ki);
2082 
2083 done:
2084 	fdrop(fp, td);
2085 
2086 	if (uap->aiocbp != NULL) {
2087 		if (cancelled) {
2088 			td->td_retval[0] = AIO_CANCELED;
2089 			return (0);
2090 		}
2091 	}
2092 
2093 	if (notcancelled) {
2094 		td->td_retval[0] = AIO_NOTCANCELED;
2095 		return (0);
2096 	}
2097 
2098 	if (cancelled) {
2099 		td->td_retval[0] = AIO_CANCELED;
2100 		return (0);
2101 	}
2102 
2103 	td->td_retval[0] = AIO_ALLDONE;
2104 
2105 	return (0);
2106 }
2107 
2108 /*
2109  * aio_error is implemented in the kernel level for compatibility purposes
2110  * only.  For a user mode async implementation, it would be best to do it in
2111  * a userland subroutine.
2112  */
2113 static int
2114 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2115 {
2116 	struct proc *p = td->td_proc;
2117 	struct kaiocb *job;
2118 	struct kaioinfo *ki;
2119 	int status;
2120 
2121 	ki = p->p_aioinfo;
2122 	if (ki == NULL) {
2123 		td->td_retval[0] = EINVAL;
2124 		return (0);
2125 	}
2126 
2127 	AIO_LOCK(ki);
2128 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2129 		if (job->ujob == ujob) {
2130 			if (job->jobflags & KAIOCB_FINISHED)
2131 				td->td_retval[0] =
2132 					job->uaiocb._aiocb_private.error;
2133 			else
2134 				td->td_retval[0] = EINPROGRESS;
2135 			AIO_UNLOCK(ki);
2136 			return (0);
2137 		}
2138 	}
2139 	AIO_UNLOCK(ki);
2140 
2141 	/*
2142 	 * Hack for failure of aio_aqueue.
2143 	 */
2144 	status = ops->fetch_status(ujob);
2145 	if (status == -1) {
2146 		td->td_retval[0] = ops->fetch_error(ujob);
2147 		return (0);
2148 	}
2149 
2150 	td->td_retval[0] = EINVAL;
2151 	return (0);
2152 }
2153 
2154 int
2155 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2156 {
2157 
2158 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2159 }
2160 
2161 /* syscall - asynchronous read from a file (REALTIME) */
2162 #ifdef COMPAT_FREEBSD6
2163 int
2164 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2165 {
2166 
2167 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2168 	    &aiocb_ops_osigevent));
2169 }
2170 #endif
2171 
2172 int
2173 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2174 {
2175 
2176 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2177 }
2178 
2179 int
2180 sys_aio_readv(struct thread *td, struct aio_readv_args *uap)
2181 {
2182 
2183 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops));
2184 }
2185 
2186 /* syscall - asynchronous write to a file (REALTIME) */
2187 #ifdef COMPAT_FREEBSD6
2188 int
2189 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2190 {
2191 
2192 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2193 	    &aiocb_ops_osigevent));
2194 }
2195 #endif
2196 
2197 int
2198 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2199 {
2200 
2201 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2202 }
2203 
2204 int
2205 sys_aio_writev(struct thread *td, struct aio_writev_args *uap)
2206 {
2207 
2208 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops));
2209 }
2210 
2211 int
2212 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2213 {
2214 
2215 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2216 }
2217 
2218 static int
2219 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2220     struct aiocb **acb_list, int nent, struct sigevent *sig,
2221     struct aiocb_ops *ops)
2222 {
2223 	struct proc *p = td->td_proc;
2224 	struct aiocb *job;
2225 	struct kaioinfo *ki;
2226 	struct aioliojob *lj;
2227 	struct kevent kev;
2228 	int error;
2229 	int nagain, nerror;
2230 	int i;
2231 
2232 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2233 		return (EINVAL);
2234 
2235 	if (nent < 0 || nent > max_aio_queue_per_proc)
2236 		return (EINVAL);
2237 
2238 	if (p->p_aioinfo == NULL)
2239 		aio_init_aioinfo(p);
2240 
2241 	ki = p->p_aioinfo;
2242 
2243 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2244 	lj->lioj_flags = 0;
2245 	lj->lioj_count = 0;
2246 	lj->lioj_finished_count = 0;
2247 	lj->lioj_signal.sigev_notify = SIGEV_NONE;
2248 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2249 	ksiginfo_init(&lj->lioj_ksi);
2250 
2251 	/*
2252 	 * Setup signal.
2253 	 */
2254 	if (sig && (mode == LIO_NOWAIT)) {
2255 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2256 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2257 			/* Assume only new style KEVENT */
2258 			memset(&kev, 0, sizeof(kev));
2259 			kev.filter = EVFILT_LIO;
2260 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2261 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2262 			kev.data = (intptr_t)lj;
2263 			/* pass user defined sigval data */
2264 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2265 			error = kqfd_register(
2266 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2267 			    M_WAITOK);
2268 			if (error) {
2269 				uma_zfree(aiolio_zone, lj);
2270 				return (error);
2271 			}
2272 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2273 			;
2274 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2275 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2276 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2277 					uma_zfree(aiolio_zone, lj);
2278 					return EINVAL;
2279 				}
2280 				lj->lioj_flags |= LIOJ_SIGNAL;
2281 		} else {
2282 			uma_zfree(aiolio_zone, lj);
2283 			return EINVAL;
2284 		}
2285 	}
2286 
2287 	AIO_LOCK(ki);
2288 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2289 	/*
2290 	 * Add extra aiocb count to avoid the lio to be freed
2291 	 * by other threads doing aio_waitcomplete or aio_return,
2292 	 * and prevent event from being sent until we have queued
2293 	 * all tasks.
2294 	 */
2295 	lj->lioj_count = 1;
2296 	AIO_UNLOCK(ki);
2297 
2298 	/*
2299 	 * Get pointers to the list of I/O requests.
2300 	 */
2301 	nagain = 0;
2302 	nerror = 0;
2303 	for (i = 0; i < nent; i++) {
2304 		job = acb_list[i];
2305 		if (job != NULL) {
2306 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2307 			if (error == EAGAIN)
2308 				nagain++;
2309 			else if (error != 0)
2310 				nerror++;
2311 		}
2312 	}
2313 
2314 	error = 0;
2315 	AIO_LOCK(ki);
2316 	if (mode == LIO_WAIT) {
2317 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2318 			ki->kaio_flags |= KAIO_WAKEUP;
2319 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2320 			    PRIBIO | PCATCH, "aiospn", 0);
2321 			if (error == ERESTART)
2322 				error = EINTR;
2323 			if (error)
2324 				break;
2325 		}
2326 	} else {
2327 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2328 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2329 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2330 				KNOTE_LOCKED(&lj->klist, 1);
2331 			}
2332 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2333 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2334 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2335 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2336 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2337 				    lj->lioj_count != 1);
2338 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2339 			}
2340 		}
2341 	}
2342 	lj->lioj_count--;
2343 	if (lj->lioj_count == 0) {
2344 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2345 		knlist_delete(&lj->klist, curthread, 1);
2346 		PROC_LOCK(p);
2347 		sigqueue_take(&lj->lioj_ksi);
2348 		PROC_UNLOCK(p);
2349 		AIO_UNLOCK(ki);
2350 		uma_zfree(aiolio_zone, lj);
2351 	} else
2352 		AIO_UNLOCK(ki);
2353 
2354 	if (nerror)
2355 		return (EIO);
2356 	else if (nagain)
2357 		return (EAGAIN);
2358 	else
2359 		return (error);
2360 }
2361 
2362 /* syscall - list directed I/O (REALTIME) */
2363 #ifdef COMPAT_FREEBSD6
2364 int
2365 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2366 {
2367 	struct aiocb **acb_list;
2368 	struct sigevent *sigp, sig;
2369 	struct osigevent osig;
2370 	int error, nent;
2371 
2372 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2373 		return (EINVAL);
2374 
2375 	nent = uap->nent;
2376 	if (nent < 0 || nent > max_aio_queue_per_proc)
2377 		return (EINVAL);
2378 
2379 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2380 		error = copyin(uap->sig, &osig, sizeof(osig));
2381 		if (error)
2382 			return (error);
2383 		error = convert_old_sigevent(&osig, &sig);
2384 		if (error)
2385 			return (error);
2386 		sigp = &sig;
2387 	} else
2388 		sigp = NULL;
2389 
2390 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2391 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2392 	if (error == 0)
2393 		error = kern_lio_listio(td, uap->mode,
2394 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2395 		    &aiocb_ops_osigevent);
2396 	free(acb_list, M_LIO);
2397 	return (error);
2398 }
2399 #endif
2400 
2401 /* syscall - list directed I/O (REALTIME) */
2402 int
2403 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2404 {
2405 	struct aiocb **acb_list;
2406 	struct sigevent *sigp, sig;
2407 	int error, nent;
2408 
2409 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2410 		return (EINVAL);
2411 
2412 	nent = uap->nent;
2413 	if (nent < 0 || nent > max_aio_queue_per_proc)
2414 		return (EINVAL);
2415 
2416 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2417 		error = copyin(uap->sig, &sig, sizeof(sig));
2418 		if (error)
2419 			return (error);
2420 		sigp = &sig;
2421 	} else
2422 		sigp = NULL;
2423 
2424 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2425 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2426 	if (error == 0)
2427 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2428 		    nent, sigp, &aiocb_ops);
2429 	free(acb_list, M_LIO);
2430 	return (error);
2431 }
2432 
2433 static void
2434 aio_biocleanup(struct bio *bp)
2435 {
2436 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2437 	struct kaioinfo *ki;
2438 	struct buf *pbuf = (struct buf *)bp->bio_caller2;
2439 
2440 	/* Release mapping into kernel space. */
2441 	if (pbuf != NULL) {
2442 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2443 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2444 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2445 		uma_zfree(pbuf_zone, pbuf);
2446 		atomic_subtract_int(&num_buf_aio, 1);
2447 		ki = job->userproc->p_aioinfo;
2448 		AIO_LOCK(ki);
2449 		ki->kaio_buffer_count--;
2450 		AIO_UNLOCK(ki);
2451 	} else {
2452 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2453 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2454 		free(bp->bio_ma, M_TEMP);
2455 		atomic_subtract_int(&num_unmapped_aio, 1);
2456 	}
2457 	g_destroy_bio(bp);
2458 }
2459 
2460 static void
2461 aio_biowakeup(struct bio *bp)
2462 {
2463 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2464 	size_t nbytes;
2465 	long bcount = bp->bio_bcount;
2466 	long resid = bp->bio_resid;
2467 	int opcode, nblks;
2468 	int bio_error = bp->bio_error;
2469 	uint16_t flags = bp->bio_flags;
2470 
2471 	opcode = job->uaiocb.aio_lio_opcode;
2472 
2473 	aio_biocleanup(bp);
2474 
2475 	nbytes = bcount - resid;
2476 	atomic_add_acq_long(&job->nbytes, nbytes);
2477 	nblks = btodb(nbytes);
2478 
2479 	/*
2480 	 * If multiple bios experienced an error, the job will reflect the
2481 	 * error of whichever failed bio completed last.
2482 	 */
2483 	if (flags & BIO_ERROR)
2484 		atomic_store_int(&job->error, bio_error);
2485 	if (opcode & LIO_WRITE)
2486 		atomic_add_int(&job->outblock, nblks);
2487 	else
2488 		atomic_add_int(&job->inblock, nblks);
2489 
2490 	if (refcount_release(&job->nbio)) {
2491 		bio_error = atomic_load_int(&job->error);
2492 		if (bio_error != 0)
2493 			aio_complete(job, -1, bio_error);
2494 		else
2495 			aio_complete(job, atomic_load_long(&job->nbytes), 0);
2496 	}
2497 }
2498 
2499 /* syscall - wait for the next completion of an aio request */
2500 static int
2501 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2502     struct timespec *ts, struct aiocb_ops *ops)
2503 {
2504 	struct proc *p = td->td_proc;
2505 	struct timeval atv;
2506 	struct kaioinfo *ki;
2507 	struct kaiocb *job;
2508 	struct aiocb *ujob;
2509 	long error, status;
2510 	int timo;
2511 
2512 	ops->store_aiocb(ujobp, NULL);
2513 
2514 	if (ts == NULL) {
2515 		timo = 0;
2516 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2517 		timo = -1;
2518 	} else {
2519 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2520 			return (EINVAL);
2521 
2522 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2523 		if (itimerfix(&atv))
2524 			return (EINVAL);
2525 		timo = tvtohz(&atv);
2526 	}
2527 
2528 	if (p->p_aioinfo == NULL)
2529 		aio_init_aioinfo(p);
2530 	ki = p->p_aioinfo;
2531 
2532 	error = 0;
2533 	job = NULL;
2534 	AIO_LOCK(ki);
2535 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2536 		if (timo == -1) {
2537 			error = EWOULDBLOCK;
2538 			break;
2539 		}
2540 		ki->kaio_flags |= KAIO_WAKEUP;
2541 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2542 		    "aiowc", timo);
2543 		if (timo && error == ERESTART)
2544 			error = EINTR;
2545 		if (error)
2546 			break;
2547 	}
2548 
2549 	if (job != NULL) {
2550 		MPASS(job->jobflags & KAIOCB_FINISHED);
2551 		ujob = job->ujob;
2552 		status = job->uaiocb._aiocb_private.status;
2553 		error = job->uaiocb._aiocb_private.error;
2554 		td->td_retval[0] = status;
2555 		td->td_ru.ru_oublock += job->outblock;
2556 		td->td_ru.ru_inblock += job->inblock;
2557 		td->td_ru.ru_msgsnd += job->msgsnd;
2558 		td->td_ru.ru_msgrcv += job->msgrcv;
2559 		aio_free_entry(job);
2560 		AIO_UNLOCK(ki);
2561 		ops->store_aiocb(ujobp, ujob);
2562 		ops->store_error(ujob, error);
2563 		ops->store_status(ujob, status);
2564 	} else
2565 		AIO_UNLOCK(ki);
2566 
2567 	return (error);
2568 }
2569 
2570 int
2571 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2572 {
2573 	struct timespec ts, *tsp;
2574 	int error;
2575 
2576 	if (uap->timeout) {
2577 		/* Get timespec struct. */
2578 		error = copyin(uap->timeout, &ts, sizeof(ts));
2579 		if (error)
2580 			return (error);
2581 		tsp = &ts;
2582 	} else
2583 		tsp = NULL;
2584 
2585 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2586 }
2587 
2588 static int
2589 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2590     struct aiocb_ops *ops)
2591 {
2592 	int listop;
2593 
2594 	switch (op) {
2595 	case O_SYNC:
2596 		listop = LIO_SYNC;
2597 		break;
2598 	case O_DSYNC:
2599 		listop = LIO_DSYNC;
2600 		break;
2601 	default:
2602 		return (EINVAL);
2603 	}
2604 
2605 	return (aio_aqueue(td, ujob, NULL, listop, ops));
2606 }
2607 
2608 int
2609 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2610 {
2611 
2612 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2613 }
2614 
2615 /* kqueue attach function */
2616 static int
2617 filt_aioattach(struct knote *kn)
2618 {
2619 	struct kaiocb *job;
2620 
2621 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2622 
2623 	/*
2624 	 * The job pointer must be validated before using it, so
2625 	 * registration is restricted to the kernel; the user cannot
2626 	 * set EV_FLAG1.
2627 	 */
2628 	if ((kn->kn_flags & EV_FLAG1) == 0)
2629 		return (EPERM);
2630 	kn->kn_ptr.p_aio = job;
2631 	kn->kn_flags &= ~EV_FLAG1;
2632 
2633 	knlist_add(&job->klist, kn, 0);
2634 
2635 	return (0);
2636 }
2637 
2638 /* kqueue detach function */
2639 static void
2640 filt_aiodetach(struct knote *kn)
2641 {
2642 	struct knlist *knl;
2643 
2644 	knl = &kn->kn_ptr.p_aio->klist;
2645 	knl->kl_lock(knl->kl_lockarg);
2646 	if (!knlist_empty(knl))
2647 		knlist_remove(knl, kn, 1);
2648 	knl->kl_unlock(knl->kl_lockarg);
2649 }
2650 
2651 /* kqueue filter function */
2652 /*ARGSUSED*/
2653 static int
2654 filt_aio(struct knote *kn, long hint)
2655 {
2656 	struct kaiocb *job = kn->kn_ptr.p_aio;
2657 
2658 	kn->kn_data = job->uaiocb._aiocb_private.error;
2659 	if (!(job->jobflags & KAIOCB_FINISHED))
2660 		return (0);
2661 	kn->kn_flags |= EV_EOF;
2662 	return (1);
2663 }
2664 
2665 /* kqueue attach function */
2666 static int
2667 filt_lioattach(struct knote *kn)
2668 {
2669 	struct aioliojob *lj;
2670 
2671 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2672 
2673 	/*
2674 	 * The aioliojob pointer must be validated before using it, so
2675 	 * registration is restricted to the kernel; the user cannot
2676 	 * set EV_FLAG1.
2677 	 */
2678 	if ((kn->kn_flags & EV_FLAG1) == 0)
2679 		return (EPERM);
2680 	kn->kn_ptr.p_lio = lj;
2681 	kn->kn_flags &= ~EV_FLAG1;
2682 
2683 	knlist_add(&lj->klist, kn, 0);
2684 
2685 	return (0);
2686 }
2687 
2688 /* kqueue detach function */
2689 static void
2690 filt_liodetach(struct knote *kn)
2691 {
2692 	struct knlist *knl;
2693 
2694 	knl = &kn->kn_ptr.p_lio->klist;
2695 	knl->kl_lock(knl->kl_lockarg);
2696 	if (!knlist_empty(knl))
2697 		knlist_remove(knl, kn, 1);
2698 	knl->kl_unlock(knl->kl_lockarg);
2699 }
2700 
2701 /* kqueue filter function */
2702 /*ARGSUSED*/
2703 static int
2704 filt_lio(struct knote *kn, long hint)
2705 {
2706 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2707 
2708 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2709 }
2710 
2711 #ifdef COMPAT_FREEBSD32
2712 #include <sys/mount.h>
2713 #include <sys/socket.h>
2714 #include <sys/sysent.h>
2715 #include <compat/freebsd32/freebsd32.h>
2716 #include <compat/freebsd32/freebsd32_proto.h>
2717 #include <compat/freebsd32/freebsd32_signal.h>
2718 #include <compat/freebsd32/freebsd32_syscall.h>
2719 #include <compat/freebsd32/freebsd32_util.h>
2720 
2721 struct __aiocb_private32 {
2722 	int32_t	status;
2723 	int32_t	error;
2724 	uint32_t kernelinfo;
2725 };
2726 
2727 #ifdef COMPAT_FREEBSD6
2728 typedef struct oaiocb32 {
2729 	int	aio_fildes;		/* File descriptor */
2730 	uint64_t aio_offset __packed;	/* File offset for I/O */
2731 	uint32_t aio_buf;		/* I/O buffer in process space */
2732 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2733 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2734 	int	aio_lio_opcode;		/* LIO opcode */
2735 	int	aio_reqprio;		/* Request priority -- ignored */
2736 	struct	__aiocb_private32 _aiocb_private;
2737 } oaiocb32_t;
2738 #endif
2739 
2740 typedef struct aiocb32 {
2741 	int32_t	aio_fildes;		/* File descriptor */
2742 	uint64_t aio_offset __packed;	/* File offset for I/O */
2743 	uint32_t aio_buf;	/* I/O buffer in process space */
2744 	uint32_t aio_nbytes;	/* Number of bytes for I/O */
2745 	int	__spare__[2];
2746 	uint32_t __spare2__;
2747 	int	aio_lio_opcode;		/* LIO opcode */
2748 	int	aio_reqprio;		/* Request priority -- ignored */
2749 	struct	__aiocb_private32 _aiocb_private;
2750 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2751 } aiocb32_t;
2752 
2753 #ifdef COMPAT_FREEBSD6
2754 static int
2755 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2756 {
2757 
2758 	/*
2759 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2760 	 * supported by AIO with the old sigevent structure.
2761 	 */
2762 	CP(*osig, *nsig, sigev_notify);
2763 	switch (nsig->sigev_notify) {
2764 	case SIGEV_NONE:
2765 		break;
2766 	case SIGEV_SIGNAL:
2767 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2768 		break;
2769 	case SIGEV_KEVENT:
2770 		nsig->sigev_notify_kqueue =
2771 		    osig->__sigev_u.__sigev_notify_kqueue;
2772 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2773 		break;
2774 	default:
2775 		return (EINVAL);
2776 	}
2777 	return (0);
2778 }
2779 
2780 static int
2781 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
2782     int type __unused)
2783 {
2784 	struct oaiocb32 job32;
2785 	struct aiocb *kcb = &kjob->uaiocb;
2786 	int error;
2787 
2788 	bzero(kcb, sizeof(struct aiocb));
2789 	error = copyin(ujob, &job32, sizeof(job32));
2790 	if (error)
2791 		return (error);
2792 
2793 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
2794 
2795 	CP(job32, *kcb, aio_fildes);
2796 	CP(job32, *kcb, aio_offset);
2797 	PTRIN_CP(job32, *kcb, aio_buf);
2798 	CP(job32, *kcb, aio_nbytes);
2799 	CP(job32, *kcb, aio_lio_opcode);
2800 	CP(job32, *kcb, aio_reqprio);
2801 	CP(job32, *kcb, _aiocb_private.status);
2802 	CP(job32, *kcb, _aiocb_private.error);
2803 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2804 	return (convert_old_sigevent32(&job32.aio_sigevent,
2805 	    &kcb->aio_sigevent));
2806 }
2807 #endif
2808 
2809 static int
2810 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
2811 {
2812 	struct aiocb32 job32;
2813 	struct aiocb *kcb = &kjob->uaiocb;
2814 	struct iovec32 *iov32;
2815 	int error;
2816 
2817 	error = copyin(ujob, &job32, sizeof(job32));
2818 	if (error)
2819 		return (error);
2820 	CP(job32, *kcb, aio_fildes);
2821 	CP(job32, *kcb, aio_offset);
2822 	CP(job32, *kcb, aio_lio_opcode);
2823 	if (type == LIO_NOP)
2824 		type = kcb->aio_lio_opcode;
2825 	if (type & LIO_VECTORED) {
2826 		iov32 = PTRIN(job32.aio_iov);
2827 		CP(job32, *kcb, aio_iovcnt);
2828 		/* malloc a uio and copy in the iovec */
2829 		error = freebsd32_copyinuio(iov32,
2830 		    kcb->aio_iovcnt, &kjob->uiop);
2831 		if (error)
2832 			return (error);
2833 	} else {
2834 		PTRIN_CP(job32, *kcb, aio_buf);
2835 		CP(job32, *kcb, aio_nbytes);
2836 	}
2837 	CP(job32, *kcb, aio_reqprio);
2838 	CP(job32, *kcb, _aiocb_private.status);
2839 	CP(job32, *kcb, _aiocb_private.error);
2840 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2841 	error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent);
2842 
2843 	return (error);
2844 }
2845 
2846 static long
2847 aiocb32_fetch_status(struct aiocb *ujob)
2848 {
2849 	struct aiocb32 *ujob32;
2850 
2851 	ujob32 = (struct aiocb32 *)ujob;
2852 	return (fuword32(&ujob32->_aiocb_private.status));
2853 }
2854 
2855 static long
2856 aiocb32_fetch_error(struct aiocb *ujob)
2857 {
2858 	struct aiocb32 *ujob32;
2859 
2860 	ujob32 = (struct aiocb32 *)ujob;
2861 	return (fuword32(&ujob32->_aiocb_private.error));
2862 }
2863 
2864 static int
2865 aiocb32_store_status(struct aiocb *ujob, long status)
2866 {
2867 	struct aiocb32 *ujob32;
2868 
2869 	ujob32 = (struct aiocb32 *)ujob;
2870 	return (suword32(&ujob32->_aiocb_private.status, status));
2871 }
2872 
2873 static int
2874 aiocb32_store_error(struct aiocb *ujob, long error)
2875 {
2876 	struct aiocb32 *ujob32;
2877 
2878 	ujob32 = (struct aiocb32 *)ujob;
2879 	return (suword32(&ujob32->_aiocb_private.error, error));
2880 }
2881 
2882 static int
2883 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2884 {
2885 	struct aiocb32 *ujob32;
2886 
2887 	ujob32 = (struct aiocb32 *)ujob;
2888 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2889 }
2890 
2891 static int
2892 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2893 {
2894 
2895 	return (suword32(ujobp, (long)ujob));
2896 }
2897 
2898 static struct aiocb_ops aiocb32_ops = {
2899 	.aio_copyin = aiocb32_copyin,
2900 	.fetch_status = aiocb32_fetch_status,
2901 	.fetch_error = aiocb32_fetch_error,
2902 	.store_status = aiocb32_store_status,
2903 	.store_error = aiocb32_store_error,
2904 	.store_kernelinfo = aiocb32_store_kernelinfo,
2905 	.store_aiocb = aiocb32_store_aiocb,
2906 };
2907 
2908 #ifdef COMPAT_FREEBSD6
2909 static struct aiocb_ops aiocb32_ops_osigevent = {
2910 	.aio_copyin = aiocb32_copyin_old_sigevent,
2911 	.fetch_status = aiocb32_fetch_status,
2912 	.fetch_error = aiocb32_fetch_error,
2913 	.store_status = aiocb32_store_status,
2914 	.store_error = aiocb32_store_error,
2915 	.store_kernelinfo = aiocb32_store_kernelinfo,
2916 	.store_aiocb = aiocb32_store_aiocb,
2917 };
2918 #endif
2919 
2920 int
2921 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2922 {
2923 
2924 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2925 }
2926 
2927 int
2928 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2929 {
2930 	struct timespec32 ts32;
2931 	struct timespec ts, *tsp;
2932 	struct aiocb **ujoblist;
2933 	uint32_t *ujoblist32;
2934 	int error, i;
2935 
2936 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2937 		return (EINVAL);
2938 
2939 	if (uap->timeout) {
2940 		/* Get timespec struct. */
2941 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2942 			return (error);
2943 		CP(ts32, ts, tv_sec);
2944 		CP(ts32, ts, tv_nsec);
2945 		tsp = &ts;
2946 	} else
2947 		tsp = NULL;
2948 
2949 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIO, M_WAITOK);
2950 	ujoblist32 = (uint32_t *)ujoblist;
2951 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2952 	    sizeof(ujoblist32[0]));
2953 	if (error == 0) {
2954 		for (i = uap->nent - 1; i >= 0; i--)
2955 			ujoblist[i] = PTRIN(ujoblist32[i]);
2956 
2957 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2958 	}
2959 	free(ujoblist, M_AIO);
2960 	return (error);
2961 }
2962 
2963 int
2964 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2965 {
2966 
2967 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2968 }
2969 
2970 #ifdef COMPAT_FREEBSD6
2971 int
2972 freebsd6_freebsd32_aio_read(struct thread *td,
2973     struct freebsd6_freebsd32_aio_read_args *uap)
2974 {
2975 
2976 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2977 	    &aiocb32_ops_osigevent));
2978 }
2979 #endif
2980 
2981 int
2982 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2983 {
2984 
2985 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2986 	    &aiocb32_ops));
2987 }
2988 
2989 int
2990 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap)
2991 {
2992 
2993 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV,
2994 	    &aiocb32_ops));
2995 }
2996 
2997 #ifdef COMPAT_FREEBSD6
2998 int
2999 freebsd6_freebsd32_aio_write(struct thread *td,
3000     struct freebsd6_freebsd32_aio_write_args *uap)
3001 {
3002 
3003 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3004 	    &aiocb32_ops_osigevent));
3005 }
3006 #endif
3007 
3008 int
3009 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
3010 {
3011 
3012 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3013 	    &aiocb32_ops));
3014 }
3015 
3016 int
3017 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap)
3018 {
3019 
3020 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV,
3021 	    &aiocb32_ops));
3022 }
3023 
3024 int
3025 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
3026 {
3027 
3028 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
3029 	    &aiocb32_ops));
3030 }
3031 
3032 int
3033 freebsd32_aio_waitcomplete(struct thread *td,
3034     struct freebsd32_aio_waitcomplete_args *uap)
3035 {
3036 	struct timespec32 ts32;
3037 	struct timespec ts, *tsp;
3038 	int error;
3039 
3040 	if (uap->timeout) {
3041 		/* Get timespec struct. */
3042 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
3043 		if (error)
3044 			return (error);
3045 		CP(ts32, ts, tv_sec);
3046 		CP(ts32, ts, tv_nsec);
3047 		tsp = &ts;
3048 	} else
3049 		tsp = NULL;
3050 
3051 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
3052 	    &aiocb32_ops));
3053 }
3054 
3055 int
3056 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
3057 {
3058 
3059 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
3060 	    &aiocb32_ops));
3061 }
3062 
3063 #ifdef COMPAT_FREEBSD6
3064 int
3065 freebsd6_freebsd32_lio_listio(struct thread *td,
3066     struct freebsd6_freebsd32_lio_listio_args *uap)
3067 {
3068 	struct aiocb **acb_list;
3069 	struct sigevent *sigp, sig;
3070 	struct osigevent32 osig;
3071 	uint32_t *acb_list32;
3072 	int error, i, nent;
3073 
3074 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3075 		return (EINVAL);
3076 
3077 	nent = uap->nent;
3078 	if (nent < 0 || nent > max_aio_queue_per_proc)
3079 		return (EINVAL);
3080 
3081 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3082 		error = copyin(uap->sig, &osig, sizeof(osig));
3083 		if (error)
3084 			return (error);
3085 		error = convert_old_sigevent32(&osig, &sig);
3086 		if (error)
3087 			return (error);
3088 		sigp = &sig;
3089 	} else
3090 		sigp = NULL;
3091 
3092 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3093 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3094 	if (error) {
3095 		free(acb_list32, M_LIO);
3096 		return (error);
3097 	}
3098 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3099 	for (i = 0; i < nent; i++)
3100 		acb_list[i] = PTRIN(acb_list32[i]);
3101 	free(acb_list32, M_LIO);
3102 
3103 	error = kern_lio_listio(td, uap->mode,
3104 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3105 	    &aiocb32_ops_osigevent);
3106 	free(acb_list, M_LIO);
3107 	return (error);
3108 }
3109 #endif
3110 
3111 int
3112 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3113 {
3114 	struct aiocb **acb_list;
3115 	struct sigevent *sigp, sig;
3116 	struct sigevent32 sig32;
3117 	uint32_t *acb_list32;
3118 	int error, i, nent;
3119 
3120 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3121 		return (EINVAL);
3122 
3123 	nent = uap->nent;
3124 	if (nent < 0 || nent > max_aio_queue_per_proc)
3125 		return (EINVAL);
3126 
3127 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3128 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3129 		if (error)
3130 			return (error);
3131 		error = convert_sigevent32(&sig32, &sig);
3132 		if (error)
3133 			return (error);
3134 		sigp = &sig;
3135 	} else
3136 		sigp = NULL;
3137 
3138 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3139 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3140 	if (error) {
3141 		free(acb_list32, M_LIO);
3142 		return (error);
3143 	}
3144 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3145 	for (i = 0; i < nent; i++)
3146 		acb_list[i] = PTRIN(acb_list32[i]);
3147 	free(acb_list32, M_LIO);
3148 
3149 	error = kern_lio_listio(td, uap->mode,
3150 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3151 	    &aiocb32_ops);
3152 	free(acb_list, M_LIO);
3153 	return (error);
3154 }
3155 
3156 #endif
3157