xref: /freebsd/sys/kern/vfs_aio.c (revision 91f764172e197c82efa97a66cfbc13d2c744b02b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. John S. Dyson's name may not be used to endorse or promote products
12  *    derived from this software without specific prior written permission.
13  *
14  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
15  * bad that happens because of using this software isn't the responsibility
16  * of the author.  This software is distributed AS-IS.
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/cdefs.h>
24 __FBSDID("$FreeBSD$");
25 
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
29 #include <sys/bio.h>
30 #include <sys/buf.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
39 #include <sys/file.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
45 #include <sys/proc.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sema.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 #include <sys/sx.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
61 #include <sys/conf.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
65 
66 #include <machine/atomic.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 #include <vm/uma.h>
75 #include <sys/aio.h>
76 
77 /*
78  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
79  * overflow. (XXX will be removed soon.)
80  */
81 static u_long jobrefid;
82 
83 /*
84  * Counter for aio_fsync.
85  */
86 static uint64_t jobseqno;
87 
88 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC	32
90 #endif
91 
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC	256
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE
97 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
98 #endif
99 
100 #ifndef MAX_BUF_AIO
101 #define MAX_BUF_AIO		16
102 #endif
103 
104 FEATURE(aio, "Asynchronous I/O");
105 SYSCTL_DECL(_p1003_1b);
106 
107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
109 
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
111     "Async IO management");
112 
113 static int enable_aio_unsafe = 0;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
115     "Permit asynchronous IO on all file types, not just known-safe types");
116 
117 static unsigned int unsafe_warningcnt = 1;
118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
119     &unsafe_warningcnt, 0,
120     "Warnings that will be triggered upon failed IO requests on unsafe files");
121 
122 static int max_aio_procs = MAX_AIO_PROCS;
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
124     "Maximum number of kernel processes to use for handling async IO ");
125 
126 static int num_aio_procs = 0;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
128     "Number of presently active kernel processes for async IO");
129 
130 /*
131  * The code will adjust the actual number of AIO processes towards this
132  * number when it gets a chance.
133  */
134 static int target_aio_procs = TARGET_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
136     0,
137     "Preferred number of ready kernel processes for async IO");
138 
139 static int max_queue_count = MAX_AIO_QUEUE;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
141     "Maximum number of aio requests to queue, globally");
142 
143 static int num_queue_count = 0;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
145     "Number of queued aio requests");
146 
147 static int num_buf_aio = 0;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
149     "Number of aio requests presently handled by the buf subsystem");
150 
151 static int num_unmapped_aio = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
153     0,
154     "Number of aio requests presently handled by unmapped I/O buffers");
155 
156 /* Number of async I/O processes in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
159 
160 static int aiod_lifetime;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
162     "Maximum lifetime for idle aiod");
163 
164 static int max_aio_per_proc = MAX_AIO_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
166     0,
167     "Maximum active aio requests per process");
168 
169 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
171     &max_aio_queue_per_proc, 0,
172     "Maximum queued aio requests per process");
173 
174 static int max_buf_aio = MAX_BUF_AIO;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
176     "Maximum buf aio requests per process");
177 
178 /*
179  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
180  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
181  * vfs.aio.aio_listio_max.
182  */
183 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
184     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
185     0, "Maximum aio requests for a single lio_listio call");
186 
187 #ifdef COMPAT_FREEBSD6
188 typedef struct oaiocb {
189 	int	aio_fildes;		/* File descriptor */
190 	off_t	aio_offset;		/* File offset for I/O */
191 	volatile void *aio_buf;         /* I/O buffer in process space */
192 	size_t	aio_nbytes;		/* Number of bytes for I/O */
193 	struct	osigevent aio_sigevent;	/* Signal to deliver */
194 	int	aio_lio_opcode;		/* LIO opcode */
195 	int	aio_reqprio;		/* Request priority -- ignored */
196 	struct	__aiocb_private	_aiocb_private;
197 } oaiocb_t;
198 #endif
199 
200 /*
201  * Below is a key of locks used to protect each member of struct kaiocb
202  * aioliojob and kaioinfo and any backends.
203  *
204  * * - need not protected
205  * a - locked by kaioinfo lock
206  * b - locked by backend lock, the backend lock can be null in some cases,
207  *     for example, BIO belongs to this type, in this case, proc lock is
208  *     reused.
209  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210  */
211 
212 /*
213  * If the routine that services an AIO request blocks while running in an
214  * AIO kernel process it can starve other I/O requests.  BIO requests
215  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
216  * processes at all.  Socket I/O requests use a separate pool of
217  * kprocs and also force non-blocking I/O.  Other file I/O requests
218  * use the generic fo_read/fo_write operations which can block.  The
219  * fsync and mlock operations can also block while executing.  Ideally
220  * none of these requests would block while executing.
221  *
222  * Note that the service routines cannot toggle O_NONBLOCK in the file
223  * structure directly while handling a request due to races with
224  * userland threads.
225  */
226 
227 /* jobflags */
228 #define	KAIOCB_QUEUEING		0x01
229 #define	KAIOCB_CANCELLED	0x02
230 #define	KAIOCB_CANCELLING	0x04
231 #define	KAIOCB_CHECKSYNC	0x08
232 #define	KAIOCB_CLEARED		0x10
233 #define	KAIOCB_FINISHED		0x20
234 
235 /*
236  * AIO process info
237  */
238 #define AIOP_FREE	0x1			/* proc on free queue */
239 
240 struct aioproc {
241 	int	aioprocflags;			/* (c) AIO proc flags */
242 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
243 	struct	proc *aioproc;			/* (*) the AIO proc */
244 };
245 
246 /*
247  * data-structure for lio signal management
248  */
249 struct aioliojob {
250 	int	lioj_flags;			/* (a) listio flags */
251 	int	lioj_count;			/* (a) count of jobs */
252 	int	lioj_finished_count;		/* (a) count of finished jobs */
253 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
254 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
255 	struct	knlist klist;			/* (a) list of knotes */
256 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
257 };
258 
259 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
260 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
261 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
262 
263 /*
264  * per process aio data structure
265  */
266 struct kaioinfo {
267 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
268 	int	kaio_flags;		/* (a) per process kaio flags */
269 	int	kaio_active_count;	/* (c) number of currently used AIOs */
270 	int	kaio_count;		/* (a) size of AIO queue */
271 	int	kaio_buffer_count;	/* (a) number of bio buffers */
272 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
273 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
274 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
276 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
277 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
278 	struct	task kaio_task;		/* (*) task to kick aio processes */
279 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
280 };
281 
282 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
286 
287 #define KAIO_RUNDOWN	0x1	/* process is being run down */
288 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
289 
290 /*
291  * Operations used to interact with userland aio control blocks.
292  * Different ABIs provide their own operations.
293  */
294 struct aiocb_ops {
295 	int	(*aio_copyin)(struct aiocb *ujob, struct kaiocb *kjob, int ty);
296 	long	(*fetch_status)(struct aiocb *ujob);
297 	long	(*fetch_error)(struct aiocb *ujob);
298 	int	(*store_status)(struct aiocb *ujob, long status);
299 	int	(*store_error)(struct aiocb *ujob, long error);
300 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
302 };
303 
304 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
308 static struct unrhdr *aiod_unr;
309 
310 static void	aio_biocleanup(struct bio *bp);
311 void		aio_init_aioinfo(struct proc *p);
312 static int	aio_onceonly(void);
313 static int	aio_free_entry(struct kaiocb *job);
314 static void	aio_process_rw(struct kaiocb *job);
315 static void	aio_process_sync(struct kaiocb *job);
316 static void	aio_process_mlock(struct kaiocb *job);
317 static void	aio_schedule_fsync(void *context, int pending);
318 static int	aio_newproc(int *);
319 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
320 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
321 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
322 static void	aio_biowakeup(struct bio *bp);
323 static void	aio_proc_rundown(void *arg, struct proc *p);
324 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
325 		    struct image_params *imgp);
326 static int	aio_qbio(struct proc *p, struct kaiocb *job);
327 static void	aio_daemon(void *param);
328 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
329 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
330 static int	aio_kick(struct proc *userp);
331 static void	aio_kick_nowait(struct proc *userp);
332 static void	aio_kick_helper(void *context, int pending);
333 static int	filt_aioattach(struct knote *kn);
334 static void	filt_aiodetach(struct knote *kn);
335 static int	filt_aio(struct knote *kn, long hint);
336 static int	filt_lioattach(struct knote *kn);
337 static void	filt_liodetach(struct knote *kn);
338 static int	filt_lio(struct knote *kn, long hint);
339 
340 /*
341  * Zones for:
342  * 	kaio	Per process async io info
343  *	aiop	async io process data
344  *	aiocb	async io jobs
345  *	aiolio	list io jobs
346  */
347 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
348 
349 /* kqueue filters for aio */
350 static struct filterops aio_filtops = {
351 	.f_isfd = 0,
352 	.f_attach = filt_aioattach,
353 	.f_detach = filt_aiodetach,
354 	.f_event = filt_aio,
355 };
356 static struct filterops lio_filtops = {
357 	.f_isfd = 0,
358 	.f_attach = filt_lioattach,
359 	.f_detach = filt_liodetach,
360 	.f_event = filt_lio
361 };
362 
363 static eventhandler_tag exit_tag, exec_tag;
364 
365 TASKQUEUE_DEFINE_THREAD(aiod_kick);
366 
367 /*
368  * Main operations function for use as a kernel module.
369  */
370 static int
371 aio_modload(struct module *module, int cmd, void *arg)
372 {
373 	int error = 0;
374 
375 	switch (cmd) {
376 	case MOD_LOAD:
377 		aio_onceonly();
378 		break;
379 	case MOD_SHUTDOWN:
380 		break;
381 	default:
382 		error = EOPNOTSUPP;
383 		break;
384 	}
385 	return (error);
386 }
387 
388 static moduledata_t aio_mod = {
389 	"aio",
390 	&aio_modload,
391 	NULL
392 };
393 
394 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
395 MODULE_VERSION(aio, 1);
396 
397 /*
398  * Startup initialization
399  */
400 static int
401 aio_onceonly(void)
402 {
403 
404 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
405 	    EVENTHANDLER_PRI_ANY);
406 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
407 	    NULL, EVENTHANDLER_PRI_ANY);
408 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
409 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
410 	TAILQ_INIT(&aio_freeproc);
411 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
412 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
413 	TAILQ_INIT(&aio_jobs);
414 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
415 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
416 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
417 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
418 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
419 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
420 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
421 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
422 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
423 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
424 	jobrefid = 1;
425 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
426 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
427 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
428 
429 	return (0);
430 }
431 
432 /*
433  * Init the per-process aioinfo structure.  The aioinfo limits are set
434  * per-process for user limit (resource) management.
435  */
436 void
437 aio_init_aioinfo(struct proc *p)
438 {
439 	struct kaioinfo *ki;
440 
441 	ki = uma_zalloc(kaio_zone, M_WAITOK);
442 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
443 	ki->kaio_flags = 0;
444 	ki->kaio_active_count = 0;
445 	ki->kaio_count = 0;
446 	ki->kaio_buffer_count = 0;
447 	TAILQ_INIT(&ki->kaio_all);
448 	TAILQ_INIT(&ki->kaio_done);
449 	TAILQ_INIT(&ki->kaio_jobqueue);
450 	TAILQ_INIT(&ki->kaio_liojoblist);
451 	TAILQ_INIT(&ki->kaio_syncqueue);
452 	TAILQ_INIT(&ki->kaio_syncready);
453 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
454 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
455 	PROC_LOCK(p);
456 	if (p->p_aioinfo == NULL) {
457 		p->p_aioinfo = ki;
458 		PROC_UNLOCK(p);
459 	} else {
460 		PROC_UNLOCK(p);
461 		mtx_destroy(&ki->kaio_mtx);
462 		uma_zfree(kaio_zone, ki);
463 	}
464 
465 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
466 		aio_newproc(NULL);
467 }
468 
469 static int
470 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
471 {
472 	struct thread *td;
473 	int error;
474 
475 	error = sigev_findtd(p, sigev, &td);
476 	if (error)
477 		return (error);
478 	if (!KSI_ONQ(ksi)) {
479 		ksiginfo_set_sigev(ksi, sigev);
480 		ksi->ksi_code = SI_ASYNCIO;
481 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
482 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
483 	}
484 	PROC_UNLOCK(p);
485 	return (error);
486 }
487 
488 /*
489  * Free a job entry.  Wait for completion if it is currently active, but don't
490  * delay forever.  If we delay, we return a flag that says that we have to
491  * restart the queue scan.
492  */
493 static int
494 aio_free_entry(struct kaiocb *job)
495 {
496 	struct kaioinfo *ki;
497 	struct aioliojob *lj;
498 	struct proc *p;
499 
500 	p = job->userproc;
501 	MPASS(curproc == p);
502 	ki = p->p_aioinfo;
503 	MPASS(ki != NULL);
504 
505 	AIO_LOCK_ASSERT(ki, MA_OWNED);
506 	MPASS(job->jobflags & KAIOCB_FINISHED);
507 
508 	atomic_subtract_int(&num_queue_count, 1);
509 
510 	ki->kaio_count--;
511 	MPASS(ki->kaio_count >= 0);
512 
513 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
514 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
515 
516 	lj = job->lio;
517 	if (lj) {
518 		lj->lioj_count--;
519 		lj->lioj_finished_count--;
520 
521 		if (lj->lioj_count == 0) {
522 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
523 			/* lio is going away, we need to destroy any knotes */
524 			knlist_delete(&lj->klist, curthread, 1);
525 			PROC_LOCK(p);
526 			sigqueue_take(&lj->lioj_ksi);
527 			PROC_UNLOCK(p);
528 			uma_zfree(aiolio_zone, lj);
529 		}
530 	}
531 
532 	/* job is going away, we need to destroy any knotes */
533 	knlist_delete(&job->klist, curthread, 1);
534 	PROC_LOCK(p);
535 	sigqueue_take(&job->ksi);
536 	PROC_UNLOCK(p);
537 
538 	AIO_UNLOCK(ki);
539 
540 	/*
541 	 * The thread argument here is used to find the owning process
542 	 * and is also passed to fo_close() which may pass it to various
543 	 * places such as devsw close() routines.  Because of that, we
544 	 * need a thread pointer from the process owning the job that is
545 	 * persistent and won't disappear out from under us or move to
546 	 * another process.
547 	 *
548 	 * Currently, all the callers of this function call it to remove
549 	 * a kaiocb from the current process' job list either via a
550 	 * syscall or due to the current process calling exit() or
551 	 * execve().  Thus, we know that p == curproc.  We also know that
552 	 * curthread can't exit since we are curthread.
553 	 *
554 	 * Therefore, we use curthread as the thread to pass to
555 	 * knlist_delete().  This does mean that it is possible for the
556 	 * thread pointer at close time to differ from the thread pointer
557 	 * at open time, but this is already true of file descriptors in
558 	 * a multithreaded process.
559 	 */
560 	if (job->fd_file)
561 		fdrop(job->fd_file, curthread);
562 	crfree(job->cred);
563 	if (job->uiop != &job->uio)
564 		free(job->uiop, M_IOV);
565 	uma_zfree(aiocb_zone, job);
566 	AIO_LOCK(ki);
567 
568 	return (0);
569 }
570 
571 static void
572 aio_proc_rundown_exec(void *arg, struct proc *p,
573     struct image_params *imgp __unused)
574 {
575    	aio_proc_rundown(arg, p);
576 }
577 
578 static int
579 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
580 {
581 	aio_cancel_fn_t *func;
582 	int cancelled;
583 
584 	AIO_LOCK_ASSERT(ki, MA_OWNED);
585 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
586 		return (0);
587 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
588 	job->jobflags |= KAIOCB_CANCELLED;
589 
590 	func = job->cancel_fn;
591 
592 	/*
593 	 * If there is no cancel routine, just leave the job marked as
594 	 * cancelled.  The job should be in active use by a caller who
595 	 * should complete it normally or when it fails to install a
596 	 * cancel routine.
597 	 */
598 	if (func == NULL)
599 		return (0);
600 
601 	/*
602 	 * Set the CANCELLING flag so that aio_complete() will defer
603 	 * completions of this job.  This prevents the job from being
604 	 * freed out from under the cancel callback.  After the
605 	 * callback any deferred completion (whether from the callback
606 	 * or any other source) will be completed.
607 	 */
608 	job->jobflags |= KAIOCB_CANCELLING;
609 	AIO_UNLOCK(ki);
610 	func(job);
611 	AIO_LOCK(ki);
612 	job->jobflags &= ~KAIOCB_CANCELLING;
613 	if (job->jobflags & KAIOCB_FINISHED) {
614 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
615 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
616 		aio_bio_done_notify(p, job);
617 	} else {
618 		/*
619 		 * The cancel callback might have scheduled an
620 		 * operation to cancel this request, but it is
621 		 * only counted as cancelled if the request is
622 		 * cancelled when the callback returns.
623 		 */
624 		cancelled = 0;
625 	}
626 	return (cancelled);
627 }
628 
629 /*
630  * Rundown the jobs for a given process.
631  */
632 static void
633 aio_proc_rundown(void *arg, struct proc *p)
634 {
635 	struct kaioinfo *ki;
636 	struct aioliojob *lj;
637 	struct kaiocb *job, *jobn;
638 
639 	KASSERT(curthread->td_proc == p,
640 	    ("%s: called on non-curproc", __func__));
641 	ki = p->p_aioinfo;
642 	if (ki == NULL)
643 		return;
644 
645 	AIO_LOCK(ki);
646 	ki->kaio_flags |= KAIO_RUNDOWN;
647 
648 restart:
649 
650 	/*
651 	 * Try to cancel all pending requests. This code simulates
652 	 * aio_cancel on all pending I/O requests.
653 	 */
654 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
655 		aio_cancel_job(p, ki, job);
656 	}
657 
658 	/* Wait for all running I/O to be finished */
659 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
660 		ki->kaio_flags |= KAIO_WAKEUP;
661 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
662 		goto restart;
663 	}
664 
665 	/* Free all completed I/O requests. */
666 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
667 		aio_free_entry(job);
668 
669 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
670 		if (lj->lioj_count == 0) {
671 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
672 			knlist_delete(&lj->klist, curthread, 1);
673 			PROC_LOCK(p);
674 			sigqueue_take(&lj->lioj_ksi);
675 			PROC_UNLOCK(p);
676 			uma_zfree(aiolio_zone, lj);
677 		} else {
678 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
679 			    lj->lioj_count, lj->lioj_finished_count);
680 		}
681 	}
682 	AIO_UNLOCK(ki);
683 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
684 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
685 	mtx_destroy(&ki->kaio_mtx);
686 	uma_zfree(kaio_zone, ki);
687 	p->p_aioinfo = NULL;
688 }
689 
690 /*
691  * Select a job to run (called by an AIO daemon).
692  */
693 static struct kaiocb *
694 aio_selectjob(struct aioproc *aiop)
695 {
696 	struct kaiocb *job;
697 	struct kaioinfo *ki;
698 	struct proc *userp;
699 
700 	mtx_assert(&aio_job_mtx, MA_OWNED);
701 restart:
702 	TAILQ_FOREACH(job, &aio_jobs, list) {
703 		userp = job->userproc;
704 		ki = userp->p_aioinfo;
705 
706 		if (ki->kaio_active_count < max_aio_per_proc) {
707 			TAILQ_REMOVE(&aio_jobs, job, list);
708 			if (!aio_clear_cancel_function(job))
709 				goto restart;
710 
711 			/* Account for currently active jobs. */
712 			ki->kaio_active_count++;
713 			break;
714 		}
715 	}
716 	return (job);
717 }
718 
719 /*
720  * Move all data to a permanent storage device.  This code
721  * simulates the fsync and fdatasync syscalls.
722  */
723 static int
724 aio_fsync_vnode(struct thread *td, struct vnode *vp, int op)
725 {
726 	struct mount *mp;
727 	vm_object_t obj;
728 	int error;
729 
730 	for (;;) {
731 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
732 		if (error != 0)
733 			break;
734 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
735 		obj = vp->v_object;
736 		if (obj != NULL) {
737 			VM_OBJECT_WLOCK(obj);
738 			vm_object_page_clean(obj, 0, 0, 0);
739 			VM_OBJECT_WUNLOCK(obj);
740 		}
741 		if (op == LIO_DSYNC)
742 			error = VOP_FDATASYNC(vp, td);
743 		else
744 			error = VOP_FSYNC(vp, MNT_WAIT, td);
745 
746 		VOP_UNLOCK(vp);
747 		vn_finished_write(mp);
748 		if (error != ERELOOKUP)
749 			break;
750 	}
751 	return (error);
752 }
753 
754 /*
755  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
756  * does the I/O request for the non-bio version of the operations.  The normal
757  * vn operations are used, and this code should work in all instances for every
758  * type of file, including pipes, sockets, fifos, and regular files.
759  *
760  * XXX I don't think it works well for socket, pipe, and fifo.
761  */
762 static void
763 aio_process_rw(struct kaiocb *job)
764 {
765 	struct ucred *td_savedcred;
766 	struct thread *td;
767 	struct aiocb *cb;
768 	struct file *fp;
769 	ssize_t cnt;
770 	long msgsnd_st, msgsnd_end;
771 	long msgrcv_st, msgrcv_end;
772 	long oublock_st, oublock_end;
773 	long inblock_st, inblock_end;
774 	int error, opcode;
775 
776 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
777 	    job->uaiocb.aio_lio_opcode == LIO_READV ||
778 	    job->uaiocb.aio_lio_opcode == LIO_WRITE ||
779 	    job->uaiocb.aio_lio_opcode == LIO_WRITEV,
780 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
781 
782 	aio_switch_vmspace(job);
783 	td = curthread;
784 	td_savedcred = td->td_ucred;
785 	td->td_ucred = job->cred;
786 	job->uiop->uio_td = td;
787 	cb = &job->uaiocb;
788 	fp = job->fd_file;
789 
790 	opcode = job->uaiocb.aio_lio_opcode;
791 	cnt = job->uiop->uio_resid;
792 
793 	msgrcv_st = td->td_ru.ru_msgrcv;
794 	msgsnd_st = td->td_ru.ru_msgsnd;
795 	inblock_st = td->td_ru.ru_inblock;
796 	oublock_st = td->td_ru.ru_oublock;
797 
798 	/*
799 	 * aio_aqueue() acquires a reference to the file that is
800 	 * released in aio_free_entry().
801 	 */
802 	if (opcode == LIO_READ || opcode == LIO_READV) {
803 		if (job->uiop->uio_resid == 0)
804 			error = 0;
805 		else
806 			error = fo_read(fp, job->uiop, fp->f_cred, FOF_OFFSET,
807 			    td);
808 	} else {
809 		if (fp->f_type == DTYPE_VNODE)
810 			bwillwrite();
811 		error = fo_write(fp, job->uiop, fp->f_cred, FOF_OFFSET, td);
812 	}
813 	msgrcv_end = td->td_ru.ru_msgrcv;
814 	msgsnd_end = td->td_ru.ru_msgsnd;
815 	inblock_end = td->td_ru.ru_inblock;
816 	oublock_end = td->td_ru.ru_oublock;
817 
818 	job->msgrcv = msgrcv_end - msgrcv_st;
819 	job->msgsnd = msgsnd_end - msgsnd_st;
820 	job->inblock = inblock_end - inblock_st;
821 	job->outblock = oublock_end - oublock_st;
822 
823 	if (error != 0 && job->uiop->uio_resid != cnt) {
824 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
825 			error = 0;
826 		if (error == EPIPE && (opcode & LIO_WRITE)) {
827 			PROC_LOCK(job->userproc);
828 			kern_psignal(job->userproc, SIGPIPE);
829 			PROC_UNLOCK(job->userproc);
830 		}
831 	}
832 
833 	cnt -= job->uiop->uio_resid;
834 	td->td_ucred = td_savedcred;
835 	if (error)
836 		aio_complete(job, -1, error);
837 	else
838 		aio_complete(job, cnt, 0);
839 }
840 
841 static void
842 aio_process_sync(struct kaiocb *job)
843 {
844 	struct thread *td = curthread;
845 	struct ucred *td_savedcred = td->td_ucred;
846 	struct file *fp = job->fd_file;
847 	int error = 0;
848 
849 	KASSERT(job->uaiocb.aio_lio_opcode & LIO_SYNC,
850 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
851 
852 	td->td_ucred = job->cred;
853 	if (fp->f_vnode != NULL) {
854 		error = aio_fsync_vnode(td, fp->f_vnode,
855 		    job->uaiocb.aio_lio_opcode);
856 	}
857 	td->td_ucred = td_savedcred;
858 	if (error)
859 		aio_complete(job, -1, error);
860 	else
861 		aio_complete(job, 0, 0);
862 }
863 
864 static void
865 aio_process_mlock(struct kaiocb *job)
866 {
867 	struct aiocb *cb = &job->uaiocb;
868 	int error;
869 
870 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
871 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
872 
873 	aio_switch_vmspace(job);
874 	error = kern_mlock(job->userproc, job->cred,
875 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
876 	aio_complete(job, error != 0 ? -1 : 0, error);
877 }
878 
879 static void
880 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
881 {
882 	struct aioliojob *lj;
883 	struct kaioinfo *ki;
884 	struct kaiocb *sjob, *sjobn;
885 	int lj_done;
886 	bool schedule_fsync;
887 
888 	ki = userp->p_aioinfo;
889 	AIO_LOCK_ASSERT(ki, MA_OWNED);
890 	lj = job->lio;
891 	lj_done = 0;
892 	if (lj) {
893 		lj->lioj_finished_count++;
894 		if (lj->lioj_count == lj->lioj_finished_count)
895 			lj_done = 1;
896 	}
897 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
898 	MPASS(job->jobflags & KAIOCB_FINISHED);
899 
900 	if (ki->kaio_flags & KAIO_RUNDOWN)
901 		goto notification_done;
902 
903 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
904 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
905 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
906 
907 	KNOTE_LOCKED(&job->klist, 1);
908 
909 	if (lj_done) {
910 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
911 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
912 			KNOTE_LOCKED(&lj->klist, 1);
913 		}
914 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
915 		    == LIOJ_SIGNAL &&
916 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
917 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
918 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
919 			    true);
920 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
921 		}
922 	}
923 
924 notification_done:
925 	if (job->jobflags & KAIOCB_CHECKSYNC) {
926 		schedule_fsync = false;
927 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
928 			if (job->fd_file != sjob->fd_file ||
929 			    job->seqno >= sjob->seqno)
930 				continue;
931 			if (--sjob->pending > 0)
932 				continue;
933 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
934 			if (!aio_clear_cancel_function_locked(sjob))
935 				continue;
936 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
937 			schedule_fsync = true;
938 		}
939 		if (schedule_fsync)
940 			taskqueue_enqueue(taskqueue_aiod_kick,
941 			    &ki->kaio_sync_task);
942 	}
943 	if (ki->kaio_flags & KAIO_WAKEUP) {
944 		ki->kaio_flags &= ~KAIO_WAKEUP;
945 		wakeup(&userp->p_aioinfo);
946 	}
947 }
948 
949 static void
950 aio_schedule_fsync(void *context, int pending)
951 {
952 	struct kaioinfo *ki;
953 	struct kaiocb *job;
954 
955 	ki = context;
956 	AIO_LOCK(ki);
957 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
958 		job = TAILQ_FIRST(&ki->kaio_syncready);
959 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
960 		AIO_UNLOCK(ki);
961 		aio_schedule(job, aio_process_sync);
962 		AIO_LOCK(ki);
963 	}
964 	AIO_UNLOCK(ki);
965 }
966 
967 bool
968 aio_cancel_cleared(struct kaiocb *job)
969 {
970 
971 	/*
972 	 * The caller should hold the same queue lock held when
973 	 * aio_clear_cancel_function() was called and set this flag
974 	 * ensuring this check sees an up-to-date value.  However,
975 	 * there is no way to assert that.
976 	 */
977 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
978 }
979 
980 static bool
981 aio_clear_cancel_function_locked(struct kaiocb *job)
982 {
983 
984 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
985 	MPASS(job->cancel_fn != NULL);
986 	if (job->jobflags & KAIOCB_CANCELLING) {
987 		job->jobflags |= KAIOCB_CLEARED;
988 		return (false);
989 	}
990 	job->cancel_fn = NULL;
991 	return (true);
992 }
993 
994 bool
995 aio_clear_cancel_function(struct kaiocb *job)
996 {
997 	struct kaioinfo *ki;
998 	bool ret;
999 
1000 	ki = job->userproc->p_aioinfo;
1001 	AIO_LOCK(ki);
1002 	ret = aio_clear_cancel_function_locked(job);
1003 	AIO_UNLOCK(ki);
1004 	return (ret);
1005 }
1006 
1007 static bool
1008 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1009 {
1010 
1011 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1012 	if (job->jobflags & KAIOCB_CANCELLED)
1013 		return (false);
1014 	job->cancel_fn = func;
1015 	return (true);
1016 }
1017 
1018 bool
1019 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1020 {
1021 	struct kaioinfo *ki;
1022 	bool ret;
1023 
1024 	ki = job->userproc->p_aioinfo;
1025 	AIO_LOCK(ki);
1026 	ret = aio_set_cancel_function_locked(job, func);
1027 	AIO_UNLOCK(ki);
1028 	return (ret);
1029 }
1030 
1031 void
1032 aio_complete(struct kaiocb *job, long status, int error)
1033 {
1034 	struct kaioinfo *ki;
1035 	struct proc *userp;
1036 
1037 	job->uaiocb._aiocb_private.error = error;
1038 	job->uaiocb._aiocb_private.status = status;
1039 
1040 	userp = job->userproc;
1041 	ki = userp->p_aioinfo;
1042 
1043 	AIO_LOCK(ki);
1044 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1045 	    ("duplicate aio_complete"));
1046 	job->jobflags |= KAIOCB_FINISHED;
1047 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1048 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1049 		aio_bio_done_notify(userp, job);
1050 	}
1051 	AIO_UNLOCK(ki);
1052 }
1053 
1054 void
1055 aio_cancel(struct kaiocb *job)
1056 {
1057 
1058 	aio_complete(job, -1, ECANCELED);
1059 }
1060 
1061 void
1062 aio_switch_vmspace(struct kaiocb *job)
1063 {
1064 
1065 	vmspace_switch_aio(job->userproc->p_vmspace);
1066 }
1067 
1068 /*
1069  * The AIO daemon, most of the actual work is done in aio_process_*,
1070  * but the setup (and address space mgmt) is done in this routine.
1071  */
1072 static void
1073 aio_daemon(void *_id)
1074 {
1075 	struct kaiocb *job;
1076 	struct aioproc *aiop;
1077 	struct kaioinfo *ki;
1078 	struct proc *p;
1079 	struct vmspace *myvm;
1080 	struct thread *td = curthread;
1081 	int id = (intptr_t)_id;
1082 
1083 	/*
1084 	 * Grab an extra reference on the daemon's vmspace so that it
1085 	 * doesn't get freed by jobs that switch to a different
1086 	 * vmspace.
1087 	 */
1088 	p = td->td_proc;
1089 	myvm = vmspace_acquire_ref(p);
1090 
1091 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1092 
1093 	/*
1094 	 * Allocate and ready the aio control info.  There is one aiop structure
1095 	 * per daemon.
1096 	 */
1097 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1098 	aiop->aioproc = p;
1099 	aiop->aioprocflags = 0;
1100 
1101 	/*
1102 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1103 	 * and creating too many daemons.)
1104 	 */
1105 	sema_post(&aio_newproc_sem);
1106 
1107 	mtx_lock(&aio_job_mtx);
1108 	for (;;) {
1109 		/*
1110 		 * Take daemon off of free queue
1111 		 */
1112 		if (aiop->aioprocflags & AIOP_FREE) {
1113 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1114 			aiop->aioprocflags &= ~AIOP_FREE;
1115 		}
1116 
1117 		/*
1118 		 * Check for jobs.
1119 		 */
1120 		while ((job = aio_selectjob(aiop)) != NULL) {
1121 			mtx_unlock(&aio_job_mtx);
1122 
1123 			ki = job->userproc->p_aioinfo;
1124 			job->handle_fn(job);
1125 
1126 			mtx_lock(&aio_job_mtx);
1127 			/* Decrement the active job count. */
1128 			ki->kaio_active_count--;
1129 		}
1130 
1131 		/*
1132 		 * Disconnect from user address space.
1133 		 */
1134 		if (p->p_vmspace != myvm) {
1135 			mtx_unlock(&aio_job_mtx);
1136 			vmspace_switch_aio(myvm);
1137 			mtx_lock(&aio_job_mtx);
1138 			/*
1139 			 * We have to restart to avoid race, we only sleep if
1140 			 * no job can be selected.
1141 			 */
1142 			continue;
1143 		}
1144 
1145 		mtx_assert(&aio_job_mtx, MA_OWNED);
1146 
1147 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1148 		aiop->aioprocflags |= AIOP_FREE;
1149 
1150 		/*
1151 		 * If daemon is inactive for a long time, allow it to exit,
1152 		 * thereby freeing resources.
1153 		 */
1154 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1155 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1156 		    (aiop->aioprocflags & AIOP_FREE) &&
1157 		    num_aio_procs > target_aio_procs)
1158 			break;
1159 	}
1160 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1161 	num_aio_procs--;
1162 	mtx_unlock(&aio_job_mtx);
1163 	uma_zfree(aiop_zone, aiop);
1164 	free_unr(aiod_unr, id);
1165 	vmspace_free(myvm);
1166 
1167 	KASSERT(p->p_vmspace == myvm,
1168 	    ("AIOD: bad vmspace for exiting daemon"));
1169 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1170 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1171 	    refcount_load(&myvm->vm_refcnt)));
1172 	kproc_exit(0);
1173 }
1174 
1175 /*
1176  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1177  * AIO daemon modifies its environment itself.
1178  */
1179 static int
1180 aio_newproc(int *start)
1181 {
1182 	int error;
1183 	struct proc *p;
1184 	int id;
1185 
1186 	id = alloc_unr(aiod_unr);
1187 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1188 		RFNOWAIT, 0, "aiod%d", id);
1189 	if (error == 0) {
1190 		/*
1191 		 * Wait until daemon is started.
1192 		 */
1193 		sema_wait(&aio_newproc_sem);
1194 		mtx_lock(&aio_job_mtx);
1195 		num_aio_procs++;
1196 		if (start != NULL)
1197 			(*start)--;
1198 		mtx_unlock(&aio_job_mtx);
1199 	} else {
1200 		free_unr(aiod_unr, id);
1201 	}
1202 	return (error);
1203 }
1204 
1205 /*
1206  * Try the high-performance, low-overhead bio method for eligible
1207  * VCHR devices.  This method doesn't use an aio helper thread, and
1208  * thus has very low overhead.
1209  *
1210  * Assumes that the caller, aio_aqueue(), has incremented the file
1211  * structure's reference count, preventing its deallocation for the
1212  * duration of this call.
1213  */
1214 static int
1215 aio_qbio(struct proc *p, struct kaiocb *job)
1216 {
1217 	struct aiocb *cb;
1218 	struct file *fp;
1219 	struct buf *pbuf;
1220 	struct vnode *vp;
1221 	struct cdevsw *csw;
1222 	struct cdev *dev;
1223 	struct kaioinfo *ki;
1224 	struct bio **bios = NULL;
1225 	off_t offset;
1226 	int bio_cmd, error, i, iovcnt, opcode, poff, ref;
1227 	vm_prot_t prot;
1228 	bool use_unmapped;
1229 
1230 	cb = &job->uaiocb;
1231 	fp = job->fd_file;
1232 	opcode = cb->aio_lio_opcode;
1233 
1234 	if (!(opcode == LIO_WRITE || opcode == LIO_WRITEV ||
1235 	    opcode == LIO_READ || opcode == LIO_READV))
1236 		return (-1);
1237 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1238 		return (-1);
1239 
1240 	vp = fp->f_vnode;
1241 	if (vp->v_type != VCHR)
1242 		return (-1);
1243 	if (vp->v_bufobj.bo_bsize == 0)
1244 		return (-1);
1245 
1246 	bio_cmd = (opcode & LIO_WRITE) ? BIO_WRITE : BIO_READ;
1247 	iovcnt = job->uiop->uio_iovcnt;
1248 	if (iovcnt > max_buf_aio)
1249 		return (-1);
1250 	for (i = 0; i < iovcnt; i++) {
1251 		if (job->uiop->uio_iov[i].iov_len % vp->v_bufobj.bo_bsize != 0)
1252 			return (-1);
1253 		if (job->uiop->uio_iov[i].iov_len > maxphys) {
1254 			error = -1;
1255 			return (-1);
1256 		}
1257 	}
1258 	offset = cb->aio_offset;
1259 
1260 	ref = 0;
1261 	csw = devvn_refthread(vp, &dev, &ref);
1262 	if (csw == NULL)
1263 		return (ENXIO);
1264 
1265 	if ((csw->d_flags & D_DISK) == 0) {
1266 		error = -1;
1267 		goto unref;
1268 	}
1269 	if (job->uiop->uio_resid > dev->si_iosize_max) {
1270 		error = -1;
1271 		goto unref;
1272 	}
1273 
1274 	ki = p->p_aioinfo;
1275 	job->error = 0;
1276 
1277 	use_unmapped = (dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed;
1278 	if (!use_unmapped) {
1279 		AIO_LOCK(ki);
1280 		if (ki->kaio_buffer_count + iovcnt > max_buf_aio) {
1281 			AIO_UNLOCK(ki);
1282 			error = EAGAIN;
1283 			goto unref;
1284 		}
1285 		ki->kaio_buffer_count += iovcnt;
1286 		AIO_UNLOCK(ki);
1287 	}
1288 
1289 	bios = malloc(sizeof(struct bio *) * iovcnt, M_TEMP, M_WAITOK);
1290 	atomic_store_int(&job->nbio, iovcnt);
1291 	for (i = 0; i < iovcnt; i++) {
1292 		struct vm_page** pages;
1293 		struct bio *bp;
1294 		void *buf;
1295 		size_t nbytes;
1296 		int npages;
1297 
1298 		buf = job->uiop->uio_iov[i].iov_base;
1299 		nbytes = job->uiop->uio_iov[i].iov_len;
1300 
1301 		bios[i] = g_alloc_bio();
1302 		bp = bios[i];
1303 
1304 		poff = (vm_offset_t)buf & PAGE_MASK;
1305 		if (use_unmapped) {
1306 			pbuf = NULL;
1307 			pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1308 			    nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1309 		} else {
1310 			pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1311 			BUF_KERNPROC(pbuf);
1312 			pages = pbuf->b_pages;
1313 		}
1314 
1315 		bp->bio_length = nbytes;
1316 		bp->bio_bcount = nbytes;
1317 		bp->bio_done = aio_biowakeup;
1318 		bp->bio_offset = offset;
1319 		bp->bio_cmd = bio_cmd;
1320 		bp->bio_dev = dev;
1321 		bp->bio_caller1 = job;
1322 		bp->bio_caller2 = pbuf;
1323 
1324 		prot = VM_PROT_READ;
1325 		if (opcode == LIO_READ || opcode == LIO_READV)
1326 			prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1327 		npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1328 		    (vm_offset_t)buf, bp->bio_length, prot, pages,
1329 		    atop(maxphys) + 1);
1330 		if (npages < 0) {
1331 			if (pbuf != NULL)
1332 				uma_zfree(pbuf_zone, pbuf);
1333 			else
1334 				free(pages, M_TEMP);
1335 			error = EFAULT;
1336 			g_destroy_bio(bp);
1337 			i--;
1338 			goto destroy_bios;
1339 		}
1340 		if (pbuf != NULL) {
1341 			pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages);
1342 			bp->bio_data = pbuf->b_data + poff;
1343 			pbuf->b_npages = npages;
1344 			atomic_add_int(&num_buf_aio, 1);
1345 		} else {
1346 			bp->bio_ma = pages;
1347 			bp->bio_ma_n = npages;
1348 			bp->bio_ma_offset = poff;
1349 			bp->bio_data = unmapped_buf;
1350 			bp->bio_flags |= BIO_UNMAPPED;
1351 			atomic_add_int(&num_unmapped_aio, 1);
1352 		}
1353 
1354 		offset += nbytes;
1355 	}
1356 
1357 	/* Perform transfer. */
1358 	for (i = 0; i < iovcnt; i++)
1359 		csw->d_strategy(bios[i]);
1360 	free(bios, M_TEMP);
1361 
1362 	dev_relthread(dev, ref);
1363 	return (0);
1364 
1365 destroy_bios:
1366 	for (; i >= 0; i--)
1367 		aio_biocleanup(bios[i]);
1368 	free(bios, M_TEMP);
1369 unref:
1370 	dev_relthread(dev, ref);
1371 	return (error);
1372 }
1373 
1374 #ifdef COMPAT_FREEBSD6
1375 static int
1376 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1377 {
1378 
1379 	/*
1380 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1381 	 * supported by AIO with the old sigevent structure.
1382 	 */
1383 	nsig->sigev_notify = osig->sigev_notify;
1384 	switch (nsig->sigev_notify) {
1385 	case SIGEV_NONE:
1386 		break;
1387 	case SIGEV_SIGNAL:
1388 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1389 		break;
1390 	case SIGEV_KEVENT:
1391 		nsig->sigev_notify_kqueue =
1392 		    osig->__sigev_u.__sigev_notify_kqueue;
1393 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1394 		break;
1395 	default:
1396 		return (EINVAL);
1397 	}
1398 	return (0);
1399 }
1400 
1401 static int
1402 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
1403     int type __unused)
1404 {
1405 	struct oaiocb *ojob;
1406 	struct aiocb *kcb = &kjob->uaiocb;
1407 	int error;
1408 
1409 	bzero(kcb, sizeof(struct aiocb));
1410 	error = copyin(ujob, kcb, sizeof(struct oaiocb));
1411 	if (error)
1412 		return (error);
1413 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
1414 	ojob = (struct oaiocb *)kcb;
1415 	return (convert_old_sigevent(&ojob->aio_sigevent, &kcb->aio_sigevent));
1416 }
1417 #endif
1418 
1419 static int
1420 aiocb_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
1421 {
1422 	struct aiocb *kcb = &kjob->uaiocb;
1423 	int error;
1424 
1425 	error = copyin(ujob, kcb, sizeof(struct aiocb));
1426 	if (error)
1427 		return (error);
1428 	if (type == LIO_NOP)
1429 		type = kcb->aio_lio_opcode;
1430 	if (type & LIO_VECTORED) {
1431 		/* malloc a uio and copy in the iovec */
1432 		error = copyinuio(__DEVOLATILE(struct iovec*, kcb->aio_iov),
1433 		    kcb->aio_iovcnt, &kjob->uiop);
1434 	}
1435 
1436 	return (error);
1437 }
1438 
1439 static long
1440 aiocb_fetch_status(struct aiocb *ujob)
1441 {
1442 
1443 	return (fuword(&ujob->_aiocb_private.status));
1444 }
1445 
1446 static long
1447 aiocb_fetch_error(struct aiocb *ujob)
1448 {
1449 
1450 	return (fuword(&ujob->_aiocb_private.error));
1451 }
1452 
1453 static int
1454 aiocb_store_status(struct aiocb *ujob, long status)
1455 {
1456 
1457 	return (suword(&ujob->_aiocb_private.status, status));
1458 }
1459 
1460 static int
1461 aiocb_store_error(struct aiocb *ujob, long error)
1462 {
1463 
1464 	return (suword(&ujob->_aiocb_private.error, error));
1465 }
1466 
1467 static int
1468 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1469 {
1470 
1471 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1472 }
1473 
1474 static int
1475 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1476 {
1477 
1478 	return (suword(ujobp, (long)ujob));
1479 }
1480 
1481 static struct aiocb_ops aiocb_ops = {
1482 	.aio_copyin = aiocb_copyin,
1483 	.fetch_status = aiocb_fetch_status,
1484 	.fetch_error = aiocb_fetch_error,
1485 	.store_status = aiocb_store_status,
1486 	.store_error = aiocb_store_error,
1487 	.store_kernelinfo = aiocb_store_kernelinfo,
1488 	.store_aiocb = aiocb_store_aiocb,
1489 };
1490 
1491 #ifdef COMPAT_FREEBSD6
1492 static struct aiocb_ops aiocb_ops_osigevent = {
1493 	.aio_copyin = aiocb_copyin_old_sigevent,
1494 	.fetch_status = aiocb_fetch_status,
1495 	.fetch_error = aiocb_fetch_error,
1496 	.store_status = aiocb_store_status,
1497 	.store_error = aiocb_store_error,
1498 	.store_kernelinfo = aiocb_store_kernelinfo,
1499 	.store_aiocb = aiocb_store_aiocb,
1500 };
1501 #endif
1502 
1503 /*
1504  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1505  * technique is done in this code.
1506  */
1507 int
1508 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1509     int type, struct aiocb_ops *ops)
1510 {
1511 	struct proc *p = td->td_proc;
1512 	struct file *fp = NULL;
1513 	struct kaiocb *job;
1514 	struct kaioinfo *ki;
1515 	struct kevent kev;
1516 	int opcode;
1517 	int error;
1518 	int fd, kqfd;
1519 	int jid;
1520 	u_short evflags;
1521 
1522 	if (p->p_aioinfo == NULL)
1523 		aio_init_aioinfo(p);
1524 
1525 	ki = p->p_aioinfo;
1526 
1527 	ops->store_status(ujob, -1);
1528 	ops->store_error(ujob, 0);
1529 	ops->store_kernelinfo(ujob, -1);
1530 
1531 	if (num_queue_count >= max_queue_count ||
1532 	    ki->kaio_count >= max_aio_queue_per_proc) {
1533 		error = EAGAIN;
1534 		goto err1;
1535 	}
1536 
1537 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1538 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1539 
1540 	error = ops->aio_copyin(ujob, job, type);
1541 	if (error)
1542 		goto err2;
1543 
1544 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1545 		error = EINVAL;
1546 		goto err2;
1547 	}
1548 
1549 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1550 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1551 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1552 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1553 		error = EINVAL;
1554 		goto err2;
1555 	}
1556 
1557 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1558 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1559 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1560 		error = EINVAL;
1561 		goto err2;
1562 	}
1563 
1564 	/* Get the opcode. */
1565 	if (type == LIO_NOP) {
1566 		switch (job->uaiocb.aio_lio_opcode) {
1567 		case LIO_WRITE:
1568 		case LIO_WRITEV:
1569 		case LIO_NOP:
1570 		case LIO_READ:
1571 		case LIO_READV:
1572 			opcode = job->uaiocb.aio_lio_opcode;
1573 			break;
1574 		default:
1575 			error = EINVAL;
1576 			goto err2;
1577 		}
1578 	} else
1579 		opcode = job->uaiocb.aio_lio_opcode = type;
1580 
1581 	ksiginfo_init(&job->ksi);
1582 
1583 	/* Save userspace address of the job info. */
1584 	job->ujob = ujob;
1585 
1586 	/*
1587 	 * Validate the opcode and fetch the file object for the specified
1588 	 * file descriptor.
1589 	 *
1590 	 * XXXRW: Moved the opcode validation up here so that we don't
1591 	 * retrieve a file descriptor without knowing what the capabiltity
1592 	 * should be.
1593 	 */
1594 	fd = job->uaiocb.aio_fildes;
1595 	switch (opcode) {
1596 	case LIO_WRITE:
1597 	case LIO_WRITEV:
1598 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1599 		break;
1600 	case LIO_READ:
1601 	case LIO_READV:
1602 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1603 		break;
1604 	case LIO_SYNC:
1605 	case LIO_DSYNC:
1606 		error = fget(td, fd, &cap_fsync_rights, &fp);
1607 		break;
1608 	case LIO_MLOCK:
1609 		break;
1610 	case LIO_NOP:
1611 		error = fget(td, fd, &cap_no_rights, &fp);
1612 		break;
1613 	default:
1614 		error = EINVAL;
1615 	}
1616 	if (error)
1617 		goto err3;
1618 
1619 	if ((opcode & LIO_SYNC) && fp->f_vnode == NULL) {
1620 		error = EINVAL;
1621 		goto err3;
1622 	}
1623 
1624 	if ((opcode == LIO_READ || opcode == LIO_READV ||
1625 	    opcode == LIO_WRITE || opcode == LIO_WRITEV) &&
1626 	    job->uaiocb.aio_offset < 0 &&
1627 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1628 		error = EINVAL;
1629 		goto err3;
1630 	}
1631 
1632 	if (fp != NULL && fp->f_ops == &path_fileops) {
1633 		error = EBADF;
1634 		goto err3;
1635 	}
1636 
1637 	job->fd_file = fp;
1638 
1639 	mtx_lock(&aio_job_mtx);
1640 	jid = jobrefid++;
1641 	job->seqno = jobseqno++;
1642 	mtx_unlock(&aio_job_mtx);
1643 	error = ops->store_kernelinfo(ujob, jid);
1644 	if (error) {
1645 		error = EINVAL;
1646 		goto err3;
1647 	}
1648 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1649 
1650 	if (opcode == LIO_NOP) {
1651 		fdrop(fp, td);
1652 		MPASS(job->uiop == &job->uio || job->uiop == NULL);
1653 		uma_zfree(aiocb_zone, job);
1654 		return (0);
1655 	}
1656 
1657 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1658 		goto no_kqueue;
1659 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1660 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1661 		error = EINVAL;
1662 		goto err3;
1663 	}
1664 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1665 	memset(&kev, 0, sizeof(kev));
1666 	kev.ident = (uintptr_t)job->ujob;
1667 	kev.filter = EVFILT_AIO;
1668 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1669 	kev.data = (intptr_t)job;
1670 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1671 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1672 	if (error)
1673 		goto err3;
1674 
1675 no_kqueue:
1676 
1677 	ops->store_error(ujob, EINPROGRESS);
1678 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1679 	job->userproc = p;
1680 	job->cred = crhold(td->td_ucred);
1681 	job->jobflags = KAIOCB_QUEUEING;
1682 	job->lio = lj;
1683 
1684 	if (opcode & LIO_VECTORED) {
1685 		/* Use the uio copied in by aio_copyin */
1686 		MPASS(job->uiop != &job->uio && job->uiop != NULL);
1687 	} else {
1688 		/* Setup the inline uio */
1689 		job->iov[0].iov_base = (void *)(uintptr_t)job->uaiocb.aio_buf;
1690 		job->iov[0].iov_len = job->uaiocb.aio_nbytes;
1691 		job->uio.uio_iov = job->iov;
1692 		job->uio.uio_iovcnt = 1;
1693 		job->uio.uio_resid = job->uaiocb.aio_nbytes;
1694 		job->uio.uio_segflg = UIO_USERSPACE;
1695 		job->uiop = &job->uio;
1696 	}
1697 	switch (opcode & (LIO_READ | LIO_WRITE)) {
1698 	case LIO_READ:
1699 		job->uiop->uio_rw = UIO_READ;
1700 		break;
1701 	case LIO_WRITE:
1702 		job->uiop->uio_rw = UIO_WRITE;
1703 		break;
1704 	}
1705 	job->uiop->uio_offset = job->uaiocb.aio_offset;
1706 	job->uiop->uio_td = td;
1707 
1708 	if (opcode == LIO_MLOCK) {
1709 		aio_schedule(job, aio_process_mlock);
1710 		error = 0;
1711 	} else if (fp->f_ops->fo_aio_queue == NULL)
1712 		error = aio_queue_file(fp, job);
1713 	else
1714 		error = fo_aio_queue(fp, job);
1715 	if (error)
1716 		goto err3;
1717 
1718 	AIO_LOCK(ki);
1719 	job->jobflags &= ~KAIOCB_QUEUEING;
1720 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1721 	ki->kaio_count++;
1722 	if (lj)
1723 		lj->lioj_count++;
1724 	atomic_add_int(&num_queue_count, 1);
1725 	if (job->jobflags & KAIOCB_FINISHED) {
1726 		/*
1727 		 * The queue callback completed the request synchronously.
1728 		 * The bulk of the completion is deferred in that case
1729 		 * until this point.
1730 		 */
1731 		aio_bio_done_notify(p, job);
1732 	} else
1733 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1734 	AIO_UNLOCK(ki);
1735 	return (0);
1736 
1737 err3:
1738 	if (fp)
1739 		fdrop(fp, td);
1740 	knlist_delete(&job->klist, curthread, 0);
1741 err2:
1742 	if (job->uiop != &job->uio)
1743 		free(job->uiop, M_IOV);
1744 	uma_zfree(aiocb_zone, job);
1745 err1:
1746 	ops->store_error(ujob, error);
1747 	return (error);
1748 }
1749 
1750 static void
1751 aio_cancel_daemon_job(struct kaiocb *job)
1752 {
1753 
1754 	mtx_lock(&aio_job_mtx);
1755 	if (!aio_cancel_cleared(job))
1756 		TAILQ_REMOVE(&aio_jobs, job, list);
1757 	mtx_unlock(&aio_job_mtx);
1758 	aio_cancel(job);
1759 }
1760 
1761 void
1762 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1763 {
1764 
1765 	mtx_lock(&aio_job_mtx);
1766 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1767 		mtx_unlock(&aio_job_mtx);
1768 		aio_cancel(job);
1769 		return;
1770 	}
1771 	job->handle_fn = func;
1772 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1773 	aio_kick_nowait(job->userproc);
1774 	mtx_unlock(&aio_job_mtx);
1775 }
1776 
1777 static void
1778 aio_cancel_sync(struct kaiocb *job)
1779 {
1780 	struct kaioinfo *ki;
1781 
1782 	ki = job->userproc->p_aioinfo;
1783 	AIO_LOCK(ki);
1784 	if (!aio_cancel_cleared(job))
1785 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1786 	AIO_UNLOCK(ki);
1787 	aio_cancel(job);
1788 }
1789 
1790 int
1791 aio_queue_file(struct file *fp, struct kaiocb *job)
1792 {
1793 	struct kaioinfo *ki;
1794 	struct kaiocb *job2;
1795 	struct vnode *vp;
1796 	struct mount *mp;
1797 	int error;
1798 	bool safe;
1799 
1800 	ki = job->userproc->p_aioinfo;
1801 	error = aio_qbio(job->userproc, job);
1802 	if (error >= 0)
1803 		return (error);
1804 	safe = false;
1805 	if (fp->f_type == DTYPE_VNODE) {
1806 		vp = fp->f_vnode;
1807 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1808 			mp = fp->f_vnode->v_mount;
1809 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1810 				safe = true;
1811 		}
1812 	}
1813 	if (!(safe || enable_aio_unsafe)) {
1814 		counted_warning(&unsafe_warningcnt,
1815 		    "is attempting to use unsafe AIO requests");
1816 		return (EOPNOTSUPP);
1817 	}
1818 
1819 	if (job->uaiocb.aio_lio_opcode & (LIO_WRITE | LIO_READ)) {
1820 		aio_schedule(job, aio_process_rw);
1821 		error = 0;
1822 	} else if (job->uaiocb.aio_lio_opcode & LIO_SYNC) {
1823 		AIO_LOCK(ki);
1824 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1825 			if (job2->fd_file == job->fd_file &&
1826 			    ((job2->uaiocb.aio_lio_opcode & LIO_SYNC) == 0) &&
1827 			    job2->seqno < job->seqno) {
1828 				job2->jobflags |= KAIOCB_CHECKSYNC;
1829 				job->pending++;
1830 			}
1831 		}
1832 		if (job->pending != 0) {
1833 			if (!aio_set_cancel_function_locked(job,
1834 				aio_cancel_sync)) {
1835 				AIO_UNLOCK(ki);
1836 				aio_cancel(job);
1837 				return (0);
1838 			}
1839 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1840 			AIO_UNLOCK(ki);
1841 			return (0);
1842 		}
1843 		AIO_UNLOCK(ki);
1844 		aio_schedule(job, aio_process_sync);
1845 		error = 0;
1846 	} else {
1847 		error = EINVAL;
1848 	}
1849 	return (error);
1850 }
1851 
1852 static void
1853 aio_kick_nowait(struct proc *userp)
1854 {
1855 	struct kaioinfo *ki = userp->p_aioinfo;
1856 	struct aioproc *aiop;
1857 
1858 	mtx_assert(&aio_job_mtx, MA_OWNED);
1859 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1860 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1861 		aiop->aioprocflags &= ~AIOP_FREE;
1862 		wakeup(aiop->aioproc);
1863 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1864 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1865 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1866 	}
1867 }
1868 
1869 static int
1870 aio_kick(struct proc *userp)
1871 {
1872 	struct kaioinfo *ki = userp->p_aioinfo;
1873 	struct aioproc *aiop;
1874 	int error, ret = 0;
1875 
1876 	mtx_assert(&aio_job_mtx, MA_OWNED);
1877 retryproc:
1878 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1879 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1880 		aiop->aioprocflags &= ~AIOP_FREE;
1881 		wakeup(aiop->aioproc);
1882 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1883 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1884 		num_aio_resv_start++;
1885 		mtx_unlock(&aio_job_mtx);
1886 		error = aio_newproc(&num_aio_resv_start);
1887 		mtx_lock(&aio_job_mtx);
1888 		if (error) {
1889 			num_aio_resv_start--;
1890 			goto retryproc;
1891 		}
1892 	} else {
1893 		ret = -1;
1894 	}
1895 	return (ret);
1896 }
1897 
1898 static void
1899 aio_kick_helper(void *context, int pending)
1900 {
1901 	struct proc *userp = context;
1902 
1903 	mtx_lock(&aio_job_mtx);
1904 	while (--pending >= 0) {
1905 		if (aio_kick(userp))
1906 			break;
1907 	}
1908 	mtx_unlock(&aio_job_mtx);
1909 }
1910 
1911 /*
1912  * Support the aio_return system call, as a side-effect, kernel resources are
1913  * released.
1914  */
1915 static int
1916 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1917 {
1918 	struct proc *p = td->td_proc;
1919 	struct kaiocb *job;
1920 	struct kaioinfo *ki;
1921 	long status, error;
1922 
1923 	ki = p->p_aioinfo;
1924 	if (ki == NULL)
1925 		return (EINVAL);
1926 	AIO_LOCK(ki);
1927 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1928 		if (job->ujob == ujob)
1929 			break;
1930 	}
1931 	if (job != NULL) {
1932 		MPASS(job->jobflags & KAIOCB_FINISHED);
1933 		status = job->uaiocb._aiocb_private.status;
1934 		error = job->uaiocb._aiocb_private.error;
1935 		td->td_retval[0] = status;
1936 		td->td_ru.ru_oublock += job->outblock;
1937 		td->td_ru.ru_inblock += job->inblock;
1938 		td->td_ru.ru_msgsnd += job->msgsnd;
1939 		td->td_ru.ru_msgrcv += job->msgrcv;
1940 		aio_free_entry(job);
1941 		AIO_UNLOCK(ki);
1942 		ops->store_error(ujob, error);
1943 		ops->store_status(ujob, status);
1944 	} else {
1945 		error = EINVAL;
1946 		AIO_UNLOCK(ki);
1947 	}
1948 	return (error);
1949 }
1950 
1951 int
1952 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1953 {
1954 
1955 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1956 }
1957 
1958 /*
1959  * Allow a process to wakeup when any of the I/O requests are completed.
1960  */
1961 static int
1962 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1963     struct timespec *ts)
1964 {
1965 	struct proc *p = td->td_proc;
1966 	struct timeval atv;
1967 	struct kaioinfo *ki;
1968 	struct kaiocb *firstjob, *job;
1969 	int error, i, timo;
1970 
1971 	timo = 0;
1972 	if (ts) {
1973 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1974 			return (EINVAL);
1975 
1976 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1977 		if (itimerfix(&atv))
1978 			return (EINVAL);
1979 		timo = tvtohz(&atv);
1980 	}
1981 
1982 	ki = p->p_aioinfo;
1983 	if (ki == NULL)
1984 		return (EAGAIN);
1985 
1986 	if (njoblist == 0)
1987 		return (0);
1988 
1989 	AIO_LOCK(ki);
1990 	for (;;) {
1991 		firstjob = NULL;
1992 		error = 0;
1993 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1994 			for (i = 0; i < njoblist; i++) {
1995 				if (job->ujob == ujoblist[i]) {
1996 					if (firstjob == NULL)
1997 						firstjob = job;
1998 					if (job->jobflags & KAIOCB_FINISHED)
1999 						goto RETURN;
2000 				}
2001 			}
2002 		}
2003 		/* All tasks were finished. */
2004 		if (firstjob == NULL)
2005 			break;
2006 
2007 		ki->kaio_flags |= KAIO_WAKEUP;
2008 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2009 		    "aiospn", timo);
2010 		if (error == ERESTART)
2011 			error = EINTR;
2012 		if (error)
2013 			break;
2014 	}
2015 RETURN:
2016 	AIO_UNLOCK(ki);
2017 	return (error);
2018 }
2019 
2020 int
2021 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2022 {
2023 	struct timespec ts, *tsp;
2024 	struct aiocb **ujoblist;
2025 	int error;
2026 
2027 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2028 		return (EINVAL);
2029 
2030 	if (uap->timeout) {
2031 		/* Get timespec struct. */
2032 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2033 			return (error);
2034 		tsp = &ts;
2035 	} else
2036 		tsp = NULL;
2037 
2038 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2039 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2040 	if (error == 0)
2041 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2042 	free(ujoblist, M_AIOS);
2043 	return (error);
2044 }
2045 
2046 /*
2047  * aio_cancel cancels any non-bio aio operations not currently in progress.
2048  */
2049 int
2050 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2051 {
2052 	struct proc *p = td->td_proc;
2053 	struct kaioinfo *ki;
2054 	struct kaiocb *job, *jobn;
2055 	struct file *fp;
2056 	int error;
2057 	int cancelled = 0;
2058 	int notcancelled = 0;
2059 	struct vnode *vp;
2060 
2061 	/* Lookup file object. */
2062 	error = fget(td, uap->fd, &cap_no_rights, &fp);
2063 	if (error)
2064 		return (error);
2065 
2066 	ki = p->p_aioinfo;
2067 	if (ki == NULL)
2068 		goto done;
2069 
2070 	if (fp->f_type == DTYPE_VNODE) {
2071 		vp = fp->f_vnode;
2072 		if (vn_isdisk(vp)) {
2073 			fdrop(fp, td);
2074 			td->td_retval[0] = AIO_NOTCANCELED;
2075 			return (0);
2076 		}
2077 	}
2078 
2079 	AIO_LOCK(ki);
2080 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2081 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2082 		    ((uap->aiocbp == NULL) ||
2083 		     (uap->aiocbp == job->ujob))) {
2084 			if (aio_cancel_job(p, ki, job)) {
2085 				cancelled++;
2086 			} else {
2087 				notcancelled++;
2088 			}
2089 			if (uap->aiocbp != NULL)
2090 				break;
2091 		}
2092 	}
2093 	AIO_UNLOCK(ki);
2094 
2095 done:
2096 	fdrop(fp, td);
2097 
2098 	if (uap->aiocbp != NULL) {
2099 		if (cancelled) {
2100 			td->td_retval[0] = AIO_CANCELED;
2101 			return (0);
2102 		}
2103 	}
2104 
2105 	if (notcancelled) {
2106 		td->td_retval[0] = AIO_NOTCANCELED;
2107 		return (0);
2108 	}
2109 
2110 	if (cancelled) {
2111 		td->td_retval[0] = AIO_CANCELED;
2112 		return (0);
2113 	}
2114 
2115 	td->td_retval[0] = AIO_ALLDONE;
2116 
2117 	return (0);
2118 }
2119 
2120 /*
2121  * aio_error is implemented in the kernel level for compatibility purposes
2122  * only.  For a user mode async implementation, it would be best to do it in
2123  * a userland subroutine.
2124  */
2125 static int
2126 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2127 {
2128 	struct proc *p = td->td_proc;
2129 	struct kaiocb *job;
2130 	struct kaioinfo *ki;
2131 	int status;
2132 
2133 	ki = p->p_aioinfo;
2134 	if (ki == NULL) {
2135 		td->td_retval[0] = EINVAL;
2136 		return (0);
2137 	}
2138 
2139 	AIO_LOCK(ki);
2140 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2141 		if (job->ujob == ujob) {
2142 			if (job->jobflags & KAIOCB_FINISHED)
2143 				td->td_retval[0] =
2144 					job->uaiocb._aiocb_private.error;
2145 			else
2146 				td->td_retval[0] = EINPROGRESS;
2147 			AIO_UNLOCK(ki);
2148 			return (0);
2149 		}
2150 	}
2151 	AIO_UNLOCK(ki);
2152 
2153 	/*
2154 	 * Hack for failure of aio_aqueue.
2155 	 */
2156 	status = ops->fetch_status(ujob);
2157 	if (status == -1) {
2158 		td->td_retval[0] = ops->fetch_error(ujob);
2159 		return (0);
2160 	}
2161 
2162 	td->td_retval[0] = EINVAL;
2163 	return (0);
2164 }
2165 
2166 int
2167 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2168 {
2169 
2170 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2171 }
2172 
2173 /* syscall - asynchronous read from a file (REALTIME) */
2174 #ifdef COMPAT_FREEBSD6
2175 int
2176 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2177 {
2178 
2179 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2180 	    &aiocb_ops_osigevent));
2181 }
2182 #endif
2183 
2184 int
2185 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2186 {
2187 
2188 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2189 }
2190 
2191 int
2192 sys_aio_readv(struct thread *td, struct aio_readv_args *uap)
2193 {
2194 
2195 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READV, &aiocb_ops));
2196 }
2197 
2198 /* syscall - asynchronous write to a file (REALTIME) */
2199 #ifdef COMPAT_FREEBSD6
2200 int
2201 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2202 {
2203 
2204 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2205 	    &aiocb_ops_osigevent));
2206 }
2207 #endif
2208 
2209 int
2210 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2211 {
2212 
2213 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2214 }
2215 
2216 int
2217 sys_aio_writev(struct thread *td, struct aio_writev_args *uap)
2218 {
2219 
2220 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITEV, &aiocb_ops));
2221 }
2222 
2223 int
2224 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2225 {
2226 
2227 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2228 }
2229 
2230 static int
2231 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2232     struct aiocb **acb_list, int nent, struct sigevent *sig,
2233     struct aiocb_ops *ops)
2234 {
2235 	struct proc *p = td->td_proc;
2236 	struct aiocb *job;
2237 	struct kaioinfo *ki;
2238 	struct aioliojob *lj;
2239 	struct kevent kev;
2240 	int error;
2241 	int nagain, nerror;
2242 	int i;
2243 
2244 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2245 		return (EINVAL);
2246 
2247 	if (nent < 0 || nent > max_aio_queue_per_proc)
2248 		return (EINVAL);
2249 
2250 	if (p->p_aioinfo == NULL)
2251 		aio_init_aioinfo(p);
2252 
2253 	ki = p->p_aioinfo;
2254 
2255 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2256 	lj->lioj_flags = 0;
2257 	lj->lioj_count = 0;
2258 	lj->lioj_finished_count = 0;
2259 	lj->lioj_signal.sigev_notify = SIGEV_NONE;
2260 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2261 	ksiginfo_init(&lj->lioj_ksi);
2262 
2263 	/*
2264 	 * Setup signal.
2265 	 */
2266 	if (sig && (mode == LIO_NOWAIT)) {
2267 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2268 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2269 			/* Assume only new style KEVENT */
2270 			memset(&kev, 0, sizeof(kev));
2271 			kev.filter = EVFILT_LIO;
2272 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2273 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2274 			kev.data = (intptr_t)lj;
2275 			/* pass user defined sigval data */
2276 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2277 			error = kqfd_register(
2278 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2279 			    M_WAITOK);
2280 			if (error) {
2281 				uma_zfree(aiolio_zone, lj);
2282 				return (error);
2283 			}
2284 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2285 			;
2286 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2287 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2288 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2289 					uma_zfree(aiolio_zone, lj);
2290 					return EINVAL;
2291 				}
2292 				lj->lioj_flags |= LIOJ_SIGNAL;
2293 		} else {
2294 			uma_zfree(aiolio_zone, lj);
2295 			return EINVAL;
2296 		}
2297 	}
2298 
2299 	AIO_LOCK(ki);
2300 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2301 	/*
2302 	 * Add extra aiocb count to avoid the lio to be freed
2303 	 * by other threads doing aio_waitcomplete or aio_return,
2304 	 * and prevent event from being sent until we have queued
2305 	 * all tasks.
2306 	 */
2307 	lj->lioj_count = 1;
2308 	AIO_UNLOCK(ki);
2309 
2310 	/*
2311 	 * Get pointers to the list of I/O requests.
2312 	 */
2313 	nagain = 0;
2314 	nerror = 0;
2315 	for (i = 0; i < nent; i++) {
2316 		job = acb_list[i];
2317 		if (job != NULL) {
2318 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2319 			if (error == EAGAIN)
2320 				nagain++;
2321 			else if (error != 0)
2322 				nerror++;
2323 		}
2324 	}
2325 
2326 	error = 0;
2327 	AIO_LOCK(ki);
2328 	if (mode == LIO_WAIT) {
2329 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2330 			ki->kaio_flags |= KAIO_WAKEUP;
2331 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2332 			    PRIBIO | PCATCH, "aiospn", 0);
2333 			if (error == ERESTART)
2334 				error = EINTR;
2335 			if (error)
2336 				break;
2337 		}
2338 	} else {
2339 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2340 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2341 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2342 				KNOTE_LOCKED(&lj->klist, 1);
2343 			}
2344 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2345 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2346 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2347 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2348 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2349 				    lj->lioj_count != 1);
2350 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2351 			}
2352 		}
2353 	}
2354 	lj->lioj_count--;
2355 	if (lj->lioj_count == 0) {
2356 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2357 		knlist_delete(&lj->klist, curthread, 1);
2358 		PROC_LOCK(p);
2359 		sigqueue_take(&lj->lioj_ksi);
2360 		PROC_UNLOCK(p);
2361 		AIO_UNLOCK(ki);
2362 		uma_zfree(aiolio_zone, lj);
2363 	} else
2364 		AIO_UNLOCK(ki);
2365 
2366 	if (nerror)
2367 		return (EIO);
2368 	else if (nagain)
2369 		return (EAGAIN);
2370 	else
2371 		return (error);
2372 }
2373 
2374 /* syscall - list directed I/O (REALTIME) */
2375 #ifdef COMPAT_FREEBSD6
2376 int
2377 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2378 {
2379 	struct aiocb **acb_list;
2380 	struct sigevent *sigp, sig;
2381 	struct osigevent osig;
2382 	int error, nent;
2383 
2384 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2385 		return (EINVAL);
2386 
2387 	nent = uap->nent;
2388 	if (nent < 0 || nent > max_aio_queue_per_proc)
2389 		return (EINVAL);
2390 
2391 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2392 		error = copyin(uap->sig, &osig, sizeof(osig));
2393 		if (error)
2394 			return (error);
2395 		error = convert_old_sigevent(&osig, &sig);
2396 		if (error)
2397 			return (error);
2398 		sigp = &sig;
2399 	} else
2400 		sigp = NULL;
2401 
2402 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2403 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2404 	if (error == 0)
2405 		error = kern_lio_listio(td, uap->mode,
2406 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2407 		    &aiocb_ops_osigevent);
2408 	free(acb_list, M_LIO);
2409 	return (error);
2410 }
2411 #endif
2412 
2413 /* syscall - list directed I/O (REALTIME) */
2414 int
2415 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2416 {
2417 	struct aiocb **acb_list;
2418 	struct sigevent *sigp, sig;
2419 	int error, nent;
2420 
2421 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2422 		return (EINVAL);
2423 
2424 	nent = uap->nent;
2425 	if (nent < 0 || nent > max_aio_queue_per_proc)
2426 		return (EINVAL);
2427 
2428 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2429 		error = copyin(uap->sig, &sig, sizeof(sig));
2430 		if (error)
2431 			return (error);
2432 		sigp = &sig;
2433 	} else
2434 		sigp = NULL;
2435 
2436 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2437 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2438 	if (error == 0)
2439 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2440 		    nent, sigp, &aiocb_ops);
2441 	free(acb_list, M_LIO);
2442 	return (error);
2443 }
2444 
2445 static void
2446 aio_biocleanup(struct bio *bp)
2447 {
2448 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2449 	struct kaioinfo *ki;
2450 	struct buf *pbuf = (struct buf *)bp->bio_caller2;
2451 
2452 	/* Release mapping into kernel space. */
2453 	if (pbuf != NULL) {
2454 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2455 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2456 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2457 		uma_zfree(pbuf_zone, pbuf);
2458 		atomic_subtract_int(&num_buf_aio, 1);
2459 		ki = job->userproc->p_aioinfo;
2460 		AIO_LOCK(ki);
2461 		ki->kaio_buffer_count--;
2462 		AIO_UNLOCK(ki);
2463 	} else {
2464 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2465 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2466 		free(bp->bio_ma, M_TEMP);
2467 		atomic_subtract_int(&num_unmapped_aio, 1);
2468 	}
2469 	g_destroy_bio(bp);
2470 }
2471 
2472 static void
2473 aio_biowakeup(struct bio *bp)
2474 {
2475 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2476 	size_t nbytes;
2477 	long bcount = bp->bio_bcount;
2478 	long resid = bp->bio_resid;
2479 	int error, opcode, nblks;
2480 	int bio_error = bp->bio_error;
2481 	uint16_t flags = bp->bio_flags;
2482 
2483 	opcode = job->uaiocb.aio_lio_opcode;
2484 
2485 	aio_biocleanup(bp);
2486 
2487 	nbytes =bcount - resid;
2488 	atomic_add_acq_long(&job->nbytes, nbytes);
2489 	nblks = btodb(nbytes);
2490 	error = 0;
2491 	/*
2492 	 * If multiple bios experienced an error, the job will reflect the
2493 	 * error of whichever failed bio completed last.
2494 	 */
2495 	if (flags & BIO_ERROR)
2496 		atomic_set_int(&job->error, bio_error);
2497 	if (opcode & LIO_WRITE)
2498 		atomic_add_int(&job->outblock, nblks);
2499 	else
2500 		atomic_add_int(&job->inblock, nblks);
2501 	atomic_subtract_int(&job->nbio, 1);
2502 
2503 
2504 	if (atomic_load_int(&job->nbio) == 0) {
2505 		if (atomic_load_int(&job->error))
2506 			aio_complete(job, -1, job->error);
2507 		else
2508 			aio_complete(job, atomic_load_long(&job->nbytes), 0);
2509 	}
2510 }
2511 
2512 /* syscall - wait for the next completion of an aio request */
2513 static int
2514 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2515     struct timespec *ts, struct aiocb_ops *ops)
2516 {
2517 	struct proc *p = td->td_proc;
2518 	struct timeval atv;
2519 	struct kaioinfo *ki;
2520 	struct kaiocb *job;
2521 	struct aiocb *ujob;
2522 	long error, status;
2523 	int timo;
2524 
2525 	ops->store_aiocb(ujobp, NULL);
2526 
2527 	if (ts == NULL) {
2528 		timo = 0;
2529 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2530 		timo = -1;
2531 	} else {
2532 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2533 			return (EINVAL);
2534 
2535 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2536 		if (itimerfix(&atv))
2537 			return (EINVAL);
2538 		timo = tvtohz(&atv);
2539 	}
2540 
2541 	if (p->p_aioinfo == NULL)
2542 		aio_init_aioinfo(p);
2543 	ki = p->p_aioinfo;
2544 
2545 	error = 0;
2546 	job = NULL;
2547 	AIO_LOCK(ki);
2548 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2549 		if (timo == -1) {
2550 			error = EWOULDBLOCK;
2551 			break;
2552 		}
2553 		ki->kaio_flags |= KAIO_WAKEUP;
2554 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2555 		    "aiowc", timo);
2556 		if (timo && error == ERESTART)
2557 			error = EINTR;
2558 		if (error)
2559 			break;
2560 	}
2561 
2562 	if (job != NULL) {
2563 		MPASS(job->jobflags & KAIOCB_FINISHED);
2564 		ujob = job->ujob;
2565 		status = job->uaiocb._aiocb_private.status;
2566 		error = job->uaiocb._aiocb_private.error;
2567 		td->td_retval[0] = status;
2568 		td->td_ru.ru_oublock += job->outblock;
2569 		td->td_ru.ru_inblock += job->inblock;
2570 		td->td_ru.ru_msgsnd += job->msgsnd;
2571 		td->td_ru.ru_msgrcv += job->msgrcv;
2572 		aio_free_entry(job);
2573 		AIO_UNLOCK(ki);
2574 		ops->store_aiocb(ujobp, ujob);
2575 		ops->store_error(ujob, error);
2576 		ops->store_status(ujob, status);
2577 	} else
2578 		AIO_UNLOCK(ki);
2579 
2580 	return (error);
2581 }
2582 
2583 int
2584 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2585 {
2586 	struct timespec ts, *tsp;
2587 	int error;
2588 
2589 	if (uap->timeout) {
2590 		/* Get timespec struct. */
2591 		error = copyin(uap->timeout, &ts, sizeof(ts));
2592 		if (error)
2593 			return (error);
2594 		tsp = &ts;
2595 	} else
2596 		tsp = NULL;
2597 
2598 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2599 }
2600 
2601 static int
2602 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2603     struct aiocb_ops *ops)
2604 {
2605 	int listop;
2606 
2607 	switch (op) {
2608 	case O_SYNC:
2609 		listop = LIO_SYNC;
2610 		break;
2611 	case O_DSYNC:
2612 		listop = LIO_DSYNC;
2613 		break;
2614 	default:
2615 		return (EINVAL);
2616 	}
2617 
2618 	return (aio_aqueue(td, ujob, NULL, listop, ops));
2619 }
2620 
2621 int
2622 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2623 {
2624 
2625 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2626 }
2627 
2628 /* kqueue attach function */
2629 static int
2630 filt_aioattach(struct knote *kn)
2631 {
2632 	struct kaiocb *job;
2633 
2634 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2635 
2636 	/*
2637 	 * The job pointer must be validated before using it, so
2638 	 * registration is restricted to the kernel; the user cannot
2639 	 * set EV_FLAG1.
2640 	 */
2641 	if ((kn->kn_flags & EV_FLAG1) == 0)
2642 		return (EPERM);
2643 	kn->kn_ptr.p_aio = job;
2644 	kn->kn_flags &= ~EV_FLAG1;
2645 
2646 	knlist_add(&job->klist, kn, 0);
2647 
2648 	return (0);
2649 }
2650 
2651 /* kqueue detach function */
2652 static void
2653 filt_aiodetach(struct knote *kn)
2654 {
2655 	struct knlist *knl;
2656 
2657 	knl = &kn->kn_ptr.p_aio->klist;
2658 	knl->kl_lock(knl->kl_lockarg);
2659 	if (!knlist_empty(knl))
2660 		knlist_remove(knl, kn, 1);
2661 	knl->kl_unlock(knl->kl_lockarg);
2662 }
2663 
2664 /* kqueue filter function */
2665 /*ARGSUSED*/
2666 static int
2667 filt_aio(struct knote *kn, long hint)
2668 {
2669 	struct kaiocb *job = kn->kn_ptr.p_aio;
2670 
2671 	kn->kn_data = job->uaiocb._aiocb_private.error;
2672 	if (!(job->jobflags & KAIOCB_FINISHED))
2673 		return (0);
2674 	kn->kn_flags |= EV_EOF;
2675 	return (1);
2676 }
2677 
2678 /* kqueue attach function */
2679 static int
2680 filt_lioattach(struct knote *kn)
2681 {
2682 	struct aioliojob *lj;
2683 
2684 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2685 
2686 	/*
2687 	 * The aioliojob pointer must be validated before using it, so
2688 	 * registration is restricted to the kernel; the user cannot
2689 	 * set EV_FLAG1.
2690 	 */
2691 	if ((kn->kn_flags & EV_FLAG1) == 0)
2692 		return (EPERM);
2693 	kn->kn_ptr.p_lio = lj;
2694 	kn->kn_flags &= ~EV_FLAG1;
2695 
2696 	knlist_add(&lj->klist, kn, 0);
2697 
2698 	return (0);
2699 }
2700 
2701 /* kqueue detach function */
2702 static void
2703 filt_liodetach(struct knote *kn)
2704 {
2705 	struct knlist *knl;
2706 
2707 	knl = &kn->kn_ptr.p_lio->klist;
2708 	knl->kl_lock(knl->kl_lockarg);
2709 	if (!knlist_empty(knl))
2710 		knlist_remove(knl, kn, 1);
2711 	knl->kl_unlock(knl->kl_lockarg);
2712 }
2713 
2714 /* kqueue filter function */
2715 /*ARGSUSED*/
2716 static int
2717 filt_lio(struct knote *kn, long hint)
2718 {
2719 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2720 
2721 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2722 }
2723 
2724 #ifdef COMPAT_FREEBSD32
2725 #include <sys/mount.h>
2726 #include <sys/socket.h>
2727 #include <compat/freebsd32/freebsd32.h>
2728 #include <compat/freebsd32/freebsd32_proto.h>
2729 #include <compat/freebsd32/freebsd32_signal.h>
2730 #include <compat/freebsd32/freebsd32_syscall.h>
2731 #include <compat/freebsd32/freebsd32_util.h>
2732 
2733 struct __aiocb_private32 {
2734 	int32_t	status;
2735 	int32_t	error;
2736 	uint32_t kernelinfo;
2737 };
2738 
2739 #ifdef COMPAT_FREEBSD6
2740 typedef struct oaiocb32 {
2741 	int	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 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2746 	int	aio_lio_opcode;		/* LIO opcode */
2747 	int	aio_reqprio;		/* Request priority -- ignored */
2748 	struct	__aiocb_private32 _aiocb_private;
2749 } oaiocb32_t;
2750 #endif
2751 
2752 typedef struct aiocb32 {
2753 	int32_t	aio_fildes;		/* File descriptor */
2754 	uint64_t aio_offset __packed;	/* File offset for I/O */
2755 	uint32_t aio_buf;	/* I/O buffer in process space */
2756 	uint32_t aio_nbytes;	/* Number of bytes for I/O */
2757 	int	__spare__[2];
2758 	uint32_t __spare2__;
2759 	int	aio_lio_opcode;		/* LIO opcode */
2760 	int	aio_reqprio;		/* Request priority -- ignored */
2761 	struct	__aiocb_private32 _aiocb_private;
2762 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2763 } aiocb32_t;
2764 
2765 #ifdef COMPAT_FREEBSD6
2766 static int
2767 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2768 {
2769 
2770 	/*
2771 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2772 	 * supported by AIO with the old sigevent structure.
2773 	 */
2774 	CP(*osig, *nsig, sigev_notify);
2775 	switch (nsig->sigev_notify) {
2776 	case SIGEV_NONE:
2777 		break;
2778 	case SIGEV_SIGNAL:
2779 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2780 		break;
2781 	case SIGEV_KEVENT:
2782 		nsig->sigev_notify_kqueue =
2783 		    osig->__sigev_u.__sigev_notify_kqueue;
2784 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2785 		break;
2786 	default:
2787 		return (EINVAL);
2788 	}
2789 	return (0);
2790 }
2791 
2792 static int
2793 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct kaiocb *kjob,
2794     int type __unused)
2795 {
2796 	struct oaiocb32 job32;
2797 	struct aiocb *kcb = &kjob->uaiocb;
2798 	int error;
2799 
2800 	bzero(kcb, sizeof(struct aiocb));
2801 	error = copyin(ujob, &job32, sizeof(job32));
2802 	if (error)
2803 		return (error);
2804 
2805 	/* No need to copyin aio_iov, because it did not exist in FreeBSD 6 */
2806 
2807 	CP(job32, *kcb, aio_fildes);
2808 	CP(job32, *kcb, aio_offset);
2809 	PTRIN_CP(job32, *kcb, aio_buf);
2810 	CP(job32, *kcb, aio_nbytes);
2811 	CP(job32, *kcb, aio_lio_opcode);
2812 	CP(job32, *kcb, aio_reqprio);
2813 	CP(job32, *kcb, _aiocb_private.status);
2814 	CP(job32, *kcb, _aiocb_private.error);
2815 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2816 	return (convert_old_sigevent32(&job32.aio_sigevent,
2817 	    &kcb->aio_sigevent));
2818 }
2819 #endif
2820 
2821 static int
2822 aiocb32_copyin(struct aiocb *ujob, struct kaiocb *kjob, int type)
2823 {
2824 	struct aiocb32 job32;
2825 	struct aiocb *kcb = &kjob->uaiocb;
2826 	struct iovec32 *iov32;
2827 	int error;
2828 
2829 	error = copyin(ujob, &job32, sizeof(job32));
2830 	if (error)
2831 		return (error);
2832 	CP(job32, *kcb, aio_fildes);
2833 	CP(job32, *kcb, aio_offset);
2834 	CP(job32, *kcb, aio_lio_opcode);
2835 	if (type == LIO_NOP)
2836 		type = kcb->aio_lio_opcode;
2837 	if (type & LIO_VECTORED) {
2838 		iov32 = PTRIN(job32.aio_iov);
2839 		CP(job32, *kcb, aio_iovcnt);
2840 		/* malloc a uio and copy in the iovec */
2841 		error = freebsd32_copyinuio(iov32,
2842 		    kcb->aio_iovcnt, &kjob->uiop);
2843 		if (error)
2844 			return (error);
2845 	} else {
2846 		PTRIN_CP(job32, *kcb, aio_buf);
2847 		CP(job32, *kcb, aio_nbytes);
2848 	}
2849 	CP(job32, *kcb, aio_reqprio);
2850 	CP(job32, *kcb, _aiocb_private.status);
2851 	CP(job32, *kcb, _aiocb_private.error);
2852 	PTRIN_CP(job32, *kcb, _aiocb_private.kernelinfo);
2853 	error = convert_sigevent32(&job32.aio_sigevent, &kcb->aio_sigevent);
2854 
2855 	return (error);
2856 }
2857 
2858 static long
2859 aiocb32_fetch_status(struct aiocb *ujob)
2860 {
2861 	struct aiocb32 *ujob32;
2862 
2863 	ujob32 = (struct aiocb32 *)ujob;
2864 	return (fuword32(&ujob32->_aiocb_private.status));
2865 }
2866 
2867 static long
2868 aiocb32_fetch_error(struct aiocb *ujob)
2869 {
2870 	struct aiocb32 *ujob32;
2871 
2872 	ujob32 = (struct aiocb32 *)ujob;
2873 	return (fuword32(&ujob32->_aiocb_private.error));
2874 }
2875 
2876 static int
2877 aiocb32_store_status(struct aiocb *ujob, long status)
2878 {
2879 	struct aiocb32 *ujob32;
2880 
2881 	ujob32 = (struct aiocb32 *)ujob;
2882 	return (suword32(&ujob32->_aiocb_private.status, status));
2883 }
2884 
2885 static int
2886 aiocb32_store_error(struct aiocb *ujob, long error)
2887 {
2888 	struct aiocb32 *ujob32;
2889 
2890 	ujob32 = (struct aiocb32 *)ujob;
2891 	return (suword32(&ujob32->_aiocb_private.error, error));
2892 }
2893 
2894 static int
2895 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2896 {
2897 	struct aiocb32 *ujob32;
2898 
2899 	ujob32 = (struct aiocb32 *)ujob;
2900 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2901 }
2902 
2903 static int
2904 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2905 {
2906 
2907 	return (suword32(ujobp, (long)ujob));
2908 }
2909 
2910 static struct aiocb_ops aiocb32_ops = {
2911 	.aio_copyin = aiocb32_copyin,
2912 	.fetch_status = aiocb32_fetch_status,
2913 	.fetch_error = aiocb32_fetch_error,
2914 	.store_status = aiocb32_store_status,
2915 	.store_error = aiocb32_store_error,
2916 	.store_kernelinfo = aiocb32_store_kernelinfo,
2917 	.store_aiocb = aiocb32_store_aiocb,
2918 };
2919 
2920 #ifdef COMPAT_FREEBSD6
2921 static struct aiocb_ops aiocb32_ops_osigevent = {
2922 	.aio_copyin = aiocb32_copyin_old_sigevent,
2923 	.fetch_status = aiocb32_fetch_status,
2924 	.fetch_error = aiocb32_fetch_error,
2925 	.store_status = aiocb32_store_status,
2926 	.store_error = aiocb32_store_error,
2927 	.store_kernelinfo = aiocb32_store_kernelinfo,
2928 	.store_aiocb = aiocb32_store_aiocb,
2929 };
2930 #endif
2931 
2932 int
2933 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2934 {
2935 
2936 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2937 }
2938 
2939 int
2940 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2941 {
2942 	struct timespec32 ts32;
2943 	struct timespec ts, *tsp;
2944 	struct aiocb **ujoblist;
2945 	uint32_t *ujoblist32;
2946 	int error, i;
2947 
2948 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2949 		return (EINVAL);
2950 
2951 	if (uap->timeout) {
2952 		/* Get timespec struct. */
2953 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2954 			return (error);
2955 		CP(ts32, ts, tv_sec);
2956 		CP(ts32, ts, tv_nsec);
2957 		tsp = &ts;
2958 	} else
2959 		tsp = NULL;
2960 
2961 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2962 	ujoblist32 = (uint32_t *)ujoblist;
2963 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2964 	    sizeof(ujoblist32[0]));
2965 	if (error == 0) {
2966 		for (i = uap->nent - 1; i >= 0; i--)
2967 			ujoblist[i] = PTRIN(ujoblist32[i]);
2968 
2969 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2970 	}
2971 	free(ujoblist, M_AIOS);
2972 	return (error);
2973 }
2974 
2975 int
2976 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2977 {
2978 
2979 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2980 }
2981 
2982 #ifdef COMPAT_FREEBSD6
2983 int
2984 freebsd6_freebsd32_aio_read(struct thread *td,
2985     struct freebsd6_freebsd32_aio_read_args *uap)
2986 {
2987 
2988 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2989 	    &aiocb32_ops_osigevent));
2990 }
2991 #endif
2992 
2993 int
2994 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2995 {
2996 
2997 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2998 	    &aiocb32_ops));
2999 }
3000 
3001 int
3002 freebsd32_aio_readv(struct thread *td, struct freebsd32_aio_readv_args *uap)
3003 {
3004 
3005 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READV,
3006 	    &aiocb32_ops));
3007 }
3008 
3009 #ifdef COMPAT_FREEBSD6
3010 int
3011 freebsd6_freebsd32_aio_write(struct thread *td,
3012     struct freebsd6_freebsd32_aio_write_args *uap)
3013 {
3014 
3015 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3016 	    &aiocb32_ops_osigevent));
3017 }
3018 #endif
3019 
3020 int
3021 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
3022 {
3023 
3024 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
3025 	    &aiocb32_ops));
3026 }
3027 
3028 int
3029 freebsd32_aio_writev(struct thread *td, struct freebsd32_aio_writev_args *uap)
3030 {
3031 
3032 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITEV,
3033 	    &aiocb32_ops));
3034 }
3035 
3036 int
3037 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
3038 {
3039 
3040 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
3041 	    &aiocb32_ops));
3042 }
3043 
3044 int
3045 freebsd32_aio_waitcomplete(struct thread *td,
3046     struct freebsd32_aio_waitcomplete_args *uap)
3047 {
3048 	struct timespec32 ts32;
3049 	struct timespec ts, *tsp;
3050 	int error;
3051 
3052 	if (uap->timeout) {
3053 		/* Get timespec struct. */
3054 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
3055 		if (error)
3056 			return (error);
3057 		CP(ts32, ts, tv_sec);
3058 		CP(ts32, ts, tv_nsec);
3059 		tsp = &ts;
3060 	} else
3061 		tsp = NULL;
3062 
3063 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
3064 	    &aiocb32_ops));
3065 }
3066 
3067 int
3068 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
3069 {
3070 
3071 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
3072 	    &aiocb32_ops));
3073 }
3074 
3075 #ifdef COMPAT_FREEBSD6
3076 int
3077 freebsd6_freebsd32_lio_listio(struct thread *td,
3078     struct freebsd6_freebsd32_lio_listio_args *uap)
3079 {
3080 	struct aiocb **acb_list;
3081 	struct sigevent *sigp, sig;
3082 	struct osigevent32 osig;
3083 	uint32_t *acb_list32;
3084 	int error, i, nent;
3085 
3086 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3087 		return (EINVAL);
3088 
3089 	nent = uap->nent;
3090 	if (nent < 0 || nent > max_aio_queue_per_proc)
3091 		return (EINVAL);
3092 
3093 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3094 		error = copyin(uap->sig, &osig, sizeof(osig));
3095 		if (error)
3096 			return (error);
3097 		error = convert_old_sigevent32(&osig, &sig);
3098 		if (error)
3099 			return (error);
3100 		sigp = &sig;
3101 	} else
3102 		sigp = NULL;
3103 
3104 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3105 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3106 	if (error) {
3107 		free(acb_list32, M_LIO);
3108 		return (error);
3109 	}
3110 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3111 	for (i = 0; i < nent; i++)
3112 		acb_list[i] = PTRIN(acb_list32[i]);
3113 	free(acb_list32, M_LIO);
3114 
3115 	error = kern_lio_listio(td, uap->mode,
3116 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3117 	    &aiocb32_ops_osigevent);
3118 	free(acb_list, M_LIO);
3119 	return (error);
3120 }
3121 #endif
3122 
3123 int
3124 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3125 {
3126 	struct aiocb **acb_list;
3127 	struct sigevent *sigp, sig;
3128 	struct sigevent32 sig32;
3129 	uint32_t *acb_list32;
3130 	int error, i, nent;
3131 
3132 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3133 		return (EINVAL);
3134 
3135 	nent = uap->nent;
3136 	if (nent < 0 || nent > max_aio_queue_per_proc)
3137 		return (EINVAL);
3138 
3139 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3140 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3141 		if (error)
3142 			return (error);
3143 		error = convert_sigevent32(&sig32, &sig);
3144 		if (error)
3145 			return (error);
3146 		sigp = &sig;
3147 	} else
3148 		sigp = NULL;
3149 
3150 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3151 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3152 	if (error) {
3153 		free(acb_list32, M_LIO);
3154 		return (error);
3155 	}
3156 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3157 	for (i = 0; i < nent; i++)
3158 		acb_list[i] = PTRIN(acb_list32[i]);
3159 	free(acb_list32, M_LIO);
3160 
3161 	error = kern_lio_listio(td, uap->mode,
3162 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3163 	    &aiocb32_ops);
3164 	free(acb_list, M_LIO);
3165 	return (error);
3166 }
3167 
3168 #endif
3169