xref: /freebsd/sys/kern/vfs_aio.c (revision d940bfec8c329dd82d8d54efebd81c8aa420503b)
1 /*-
2  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. John S. Dyson's name may not be used to endorse or promote products
10  *    derived from this software without specific prior written permission.
11  *
12  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
13  * bad that happens because of using this software isn't the responsibility
14  * of the author.  This software is distributed AS-IS.
15  */
16 
17 /*
18  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
19  */
20 
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
23 
24 #include "opt_compat.h"
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/capability.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/protosw.h>
49 #include <sys/rwlock.h>
50 #include <sys/sema.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/syscall.h>
54 #include <sys/sysent.h>
55 #include <sys/sysctl.h>
56 #include <sys/sx.h>
57 #include <sys/taskqueue.h>
58 #include <sys/vnode.h>
59 #include <sys/conf.h>
60 #include <sys/event.h>
61 #include <sys/mount.h>
62 
63 #include <machine/atomic.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_extern.h>
67 #include <vm/pmap.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/uma.h>
71 #include <sys/aio.h>
72 
73 #include "opt_vfs_aio.h"
74 
75 /*
76  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
77  * overflow. (XXX will be removed soon.)
78  */
79 static u_long jobrefid;
80 
81 /*
82  * Counter for aio_fsync.
83  */
84 static uint64_t jobseqno;
85 
86 #define JOBST_NULL		0
87 #define JOBST_JOBQSOCK		1
88 #define JOBST_JOBQGLOBAL	2
89 #define JOBST_JOBRUNNING	3
90 #define JOBST_JOBFINISHED	4
91 #define JOBST_JOBQBUF		5
92 #define JOBST_JOBQSYNC		6
93 
94 #ifndef MAX_AIO_PER_PROC
95 #define MAX_AIO_PER_PROC	32
96 #endif
97 
98 #ifndef MAX_AIO_QUEUE_PER_PROC
99 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
100 #endif
101 
102 #ifndef MAX_AIO_PROCS
103 #define MAX_AIO_PROCS		32
104 #endif
105 
106 #ifndef MAX_AIO_QUEUE
107 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
108 #endif
109 
110 #ifndef TARGET_AIO_PROCS
111 #define TARGET_AIO_PROCS	4
112 #endif
113 
114 #ifndef MAX_BUF_AIO
115 #define MAX_BUF_AIO		16
116 #endif
117 
118 #ifndef AIOD_TIMEOUT_DEFAULT
119 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
120 #endif
121 
122 #ifndef AIOD_LIFETIME_DEFAULT
123 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
124 #endif
125 
126 FEATURE(aio, "Asynchronous I/O");
127 
128 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
129 
130 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
131 
132 static int max_aio_procs = MAX_AIO_PROCS;
133 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
134 	CTLFLAG_RW, &max_aio_procs, 0,
135 	"Maximum number of kernel threads to use for handling async IO ");
136 
137 static int num_aio_procs = 0;
138 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
139 	CTLFLAG_RD, &num_aio_procs, 0,
140 	"Number of presently active kernel threads for async IO");
141 
142 /*
143  * The code will adjust the actual number of AIO processes towards this
144  * number when it gets a chance.
145  */
146 static int target_aio_procs = TARGET_AIO_PROCS;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
148 	0, "Preferred number of ready kernel threads for async IO");
149 
150 static int max_queue_count = MAX_AIO_QUEUE;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
152     "Maximum number of aio requests to queue, globally");
153 
154 static int num_queue_count = 0;
155 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
156     "Number of queued aio requests");
157 
158 static int num_buf_aio = 0;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
160     "Number of aio requests presently handled by the buf subsystem");
161 
162 /* Number of async I/O thread in the process of being started */
163 /* XXX This should be local to aio_aqueue() */
164 static int num_aio_resv_start = 0;
165 
166 static int aiod_timeout;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
168     "Timeout value for synchronous aio operations");
169 
170 static int aiod_lifetime;
171 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
172     "Maximum lifetime for idle aiod");
173 
174 static int unloadable = 0;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
176     "Allow unload of aio (not recommended)");
177 
178 
179 static int max_aio_per_proc = MAX_AIO_PER_PROC;
180 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
181     0, "Maximum active aio requests per process (stored in the process)");
182 
183 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
184 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
185     &max_aio_queue_per_proc, 0,
186     "Maximum queued aio requests per process (stored in the process)");
187 
188 static int max_buf_aio = MAX_BUF_AIO;
189 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
190     "Maximum buf aio requests per process (stored in the process)");
191 
192 typedef struct oaiocb {
193 	int	aio_fildes;		/* File descriptor */
194 	off_t	aio_offset;		/* File offset for I/O */
195 	volatile void *aio_buf;         /* I/O buffer in process space */
196 	size_t	aio_nbytes;		/* Number of bytes for I/O */
197 	struct	osigevent aio_sigevent;	/* Signal to deliver */
198 	int	aio_lio_opcode;		/* LIO opcode */
199 	int	aio_reqprio;		/* Request priority -- ignored */
200 	struct	__aiocb_private	_aiocb_private;
201 } oaiocb_t;
202 
203 /*
204  * Below is a key of locks used to protect each member of struct aiocblist
205  * aioliojob and kaioinfo and any backends.
206  *
207  * * - need not protected
208  * a - locked by kaioinfo lock
209  * b - locked by backend lock, the backend lock can be null in some cases,
210  *     for example, BIO belongs to this type, in this case, proc lock is
211  *     reused.
212  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
213  */
214 
215 /*
216  * Current, there is only two backends: BIO and generic file I/O.
217  * socket I/O is served by generic file I/O, this is not a good idea, since
218  * disk file I/O and any other types without O_NONBLOCK flag can block daemon
219  * threads, if there is no thread to serve socket I/O, the socket I/O will be
220  * delayed too long or starved, we should create some threads dedicated to
221  * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
222  * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
223  * structure is not safe because there is race between userland and aio
224  * daemons.
225  */
226 
227 struct aiocblist {
228 	TAILQ_ENTRY(aiocblist) list;	/* (b) internal list of for backend */
229 	TAILQ_ENTRY(aiocblist) plist;	/* (a) list of jobs for each backend */
230 	TAILQ_ENTRY(aiocblist) allist;  /* (a) list of all jobs in proc */
231 	int	jobflags;		/* (a) job flags */
232 	int	jobstate;		/* (b) job state */
233 	int	inputcharge;		/* (*) input blockes */
234 	int	outputcharge;		/* (*) output blockes */
235 	struct	buf *bp;		/* (*) private to BIO backend,
236 				  	 * buffer pointer
237 					 */
238 	struct	proc *userproc;		/* (*) user process */
239 	struct  ucred *cred;		/* (*) active credential when created */
240 	struct	file *fd_file;		/* (*) pointer to file structure */
241 	struct	aioliojob *lio;		/* (*) optional lio job */
242 	struct	aiocb *uuaiocb;		/* (*) pointer in userspace of aiocb */
243 	struct	knlist klist;		/* (a) list of knotes */
244 	struct	aiocb uaiocb;		/* (*) kernel I/O control block */
245 	ksiginfo_t ksi;			/* (a) realtime signal info */
246 	struct	task biotask;		/* (*) private to BIO backend */
247 	uint64_t seqno;			/* (*) job number */
248 	int	pending;		/* (a) number of pending I/O, aio_fsync only */
249 };
250 
251 /* jobflags */
252 #define AIOCBLIST_DONE		0x01
253 #define AIOCBLIST_BUFDONE	0x02
254 #define AIOCBLIST_RUNDOWN	0x04
255 #define AIOCBLIST_CHECKSYNC	0x08
256 
257 /*
258  * AIO process info
259  */
260 #define AIOP_FREE	0x1			/* proc on free queue */
261 
262 struct aiothreadlist {
263 	int aiothreadflags;			/* (c) AIO proc flags */
264 	TAILQ_ENTRY(aiothreadlist) list;	/* (c) list of processes */
265 	struct thread *aiothread;		/* (*) the AIO thread */
266 };
267 
268 /*
269  * data-structure for lio signal management
270  */
271 struct aioliojob {
272 	int	lioj_flags;			/* (a) listio flags */
273 	int	lioj_count;			/* (a) listio flags */
274 	int	lioj_finished_count;		/* (a) listio flags */
275 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
276 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
277 	struct  knlist klist;			/* (a) list of knotes */
278 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
279 };
280 
281 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
282 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
283 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
284 
285 /*
286  * per process aio data structure
287  */
288 struct kaioinfo {
289 	struct mtx	kaio_mtx;	/* the lock to protect this struct */
290 	int	kaio_flags;		/* (a) per process kaio flags */
291 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
292 	int	kaio_active_count;	/* (c) number of currently used AIOs */
293 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
294 	int	kaio_count;		/* (a) size of AIO queue */
295 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
296 	int	kaio_buffer_count;	/* (a) number of physio buffers */
297 	TAILQ_HEAD(,aiocblist) kaio_all;	/* (a) all AIOs in the process */
298 	TAILQ_HEAD(,aiocblist) kaio_done;	/* (a) done queue for process */
299 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
300 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* (a) job queue for process */
301 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* (a) buffer job queue for process */
302 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;  /* (a) queue for aios waiting on sockets,
303 						 *  NOT USED YET.
304 						 */
305 	TAILQ_HEAD(,aiocblist) kaio_syncqueue;	/* (a) queue for aio_fsync */
306 	struct	task	kaio_task;	/* (*) task to kick aio threads */
307 };
308 
309 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
310 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
311 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
312 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
313 
314 #define KAIO_RUNDOWN	0x1	/* process is being run down */
315 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
316 
317 /*
318  * Operations used to interact with userland aio control blocks.
319  * Different ABIs provide their own operations.
320  */
321 struct aiocb_ops {
322 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
323 	long	(*fetch_status)(struct aiocb *ujob);
324 	long	(*fetch_error)(struct aiocb *ujob);
325 	int	(*store_status)(struct aiocb *ujob, long status);
326 	int	(*store_error)(struct aiocb *ujob, long error);
327 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
328 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
329 };
330 
331 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* (c) Idle daemons */
332 static struct sema aio_newproc_sem;
333 static struct mtx aio_job_mtx;
334 static struct mtx aio_sock_mtx;
335 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* (c) Async job list */
336 static struct unrhdr *aiod_unr;
337 
338 void		aio_init_aioinfo(struct proc *p);
339 static int	aio_onceonly(void);
340 static int	aio_free_entry(struct aiocblist *aiocbe);
341 static void	aio_process_rw(struct aiocblist *aiocbe);
342 static void	aio_process_sync(struct aiocblist *aiocbe);
343 static void	aio_process_mlock(struct aiocblist *aiocbe);
344 static int	aio_newproc(int *);
345 int		aio_aqueue(struct thread *td, struct aiocb *job,
346 			struct aioliojob *lio, int type, struct aiocb_ops *ops);
347 static void	aio_physwakeup(struct buf *bp);
348 static void	aio_proc_rundown(void *arg, struct proc *p);
349 static void	aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
350 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
351 static void	biohelper(void *, int);
352 static void	aio_daemon(void *param);
353 static void	aio_swake_cb(struct socket *, struct sockbuf *);
354 static int	aio_unload(void);
355 static void	aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
356 #define DONE_BUF	1
357 #define DONE_QUEUE	2
358 static int	aio_kick(struct proc *userp);
359 static void	aio_kick_nowait(struct proc *userp);
360 static void	aio_kick_helper(void *context, int pending);
361 static int	filt_aioattach(struct knote *kn);
362 static void	filt_aiodetach(struct knote *kn);
363 static int	filt_aio(struct knote *kn, long hint);
364 static int	filt_lioattach(struct knote *kn);
365 static void	filt_liodetach(struct knote *kn);
366 static int	filt_lio(struct knote *kn, long hint);
367 
368 /*
369  * Zones for:
370  * 	kaio	Per process async io info
371  *	aiop	async io thread data
372  *	aiocb	async io jobs
373  *	aiol	list io job pointer - internal to aio_suspend XXX
374  *	aiolio	list io jobs
375  */
376 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
377 
378 /* kqueue filters for aio */
379 static struct filterops aio_filtops = {
380 	.f_isfd = 0,
381 	.f_attach = filt_aioattach,
382 	.f_detach = filt_aiodetach,
383 	.f_event = filt_aio,
384 };
385 static struct filterops lio_filtops = {
386 	.f_isfd = 0,
387 	.f_attach = filt_lioattach,
388 	.f_detach = filt_liodetach,
389 	.f_event = filt_lio
390 };
391 
392 static eventhandler_tag exit_tag, exec_tag;
393 
394 TASKQUEUE_DEFINE_THREAD(aiod_bio);
395 
396 /*
397  * Main operations function for use as a kernel module.
398  */
399 static int
400 aio_modload(struct module *module, int cmd, void *arg)
401 {
402 	int error = 0;
403 
404 	switch (cmd) {
405 	case MOD_LOAD:
406 		aio_onceonly();
407 		break;
408 	case MOD_UNLOAD:
409 		error = aio_unload();
410 		break;
411 	case MOD_SHUTDOWN:
412 		break;
413 	default:
414 		error = EINVAL;
415 		break;
416 	}
417 	return (error);
418 }
419 
420 static moduledata_t aio_mod = {
421 	"aio",
422 	&aio_modload,
423 	NULL
424 };
425 
426 static struct syscall_helper_data aio_syscalls[] = {
427 	SYSCALL_INIT_HELPER(aio_cancel),
428 	SYSCALL_INIT_HELPER(aio_error),
429 	SYSCALL_INIT_HELPER(aio_fsync),
430 	SYSCALL_INIT_HELPER(aio_mlock),
431 	SYSCALL_INIT_HELPER(aio_read),
432 	SYSCALL_INIT_HELPER(aio_return),
433 	SYSCALL_INIT_HELPER(aio_suspend),
434 	SYSCALL_INIT_HELPER(aio_waitcomplete),
435 	SYSCALL_INIT_HELPER(aio_write),
436 	SYSCALL_INIT_HELPER(lio_listio),
437 	SYSCALL_INIT_HELPER(oaio_read),
438 	SYSCALL_INIT_HELPER(oaio_write),
439 	SYSCALL_INIT_HELPER(olio_listio),
440 	SYSCALL_INIT_LAST
441 };
442 
443 #ifdef COMPAT_FREEBSD32
444 #include <sys/mount.h>
445 #include <sys/socket.h>
446 #include <compat/freebsd32/freebsd32.h>
447 #include <compat/freebsd32/freebsd32_proto.h>
448 #include <compat/freebsd32/freebsd32_signal.h>
449 #include <compat/freebsd32/freebsd32_syscall.h>
450 #include <compat/freebsd32/freebsd32_util.h>
451 
452 static struct syscall_helper_data aio32_syscalls[] = {
453 	SYSCALL32_INIT_HELPER(freebsd32_aio_return),
454 	SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
455 	SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
456 	SYSCALL32_INIT_HELPER(freebsd32_aio_error),
457 	SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
458 	SYSCALL32_INIT_HELPER(freebsd32_aio_mlock),
459 	SYSCALL32_INIT_HELPER(freebsd32_aio_read),
460 	SYSCALL32_INIT_HELPER(freebsd32_aio_write),
461 	SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
462 	SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
463 	SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
464 	SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
465 	SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
466 	SYSCALL_INIT_LAST
467 };
468 #endif
469 
470 DECLARE_MODULE(aio, aio_mod,
471 	SI_SUB_VFS, SI_ORDER_ANY);
472 MODULE_VERSION(aio, 1);
473 
474 /*
475  * Startup initialization
476  */
477 static int
478 aio_onceonly(void)
479 {
480 	int error;
481 
482 	/* XXX: should probably just use so->callback */
483 	aio_swake = &aio_swake_cb;
484 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
485 	    EVENTHANDLER_PRI_ANY);
486 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
487 	    EVENTHANDLER_PRI_ANY);
488 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
489 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
490 	TAILQ_INIT(&aio_freeproc);
491 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
492 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
493 	mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
494 	TAILQ_INIT(&aio_jobs);
495 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
496 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
497 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
498 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
499 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
500 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
501 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
502 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
503 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
504 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
505 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
506 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
507 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
508 	jobrefid = 1;
509 	async_io_version = _POSIX_VERSION;
510 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
511 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
512 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
513 
514 	error = syscall_helper_register(aio_syscalls);
515 	if (error)
516 		return (error);
517 #ifdef COMPAT_FREEBSD32
518 	error = syscall32_helper_register(aio32_syscalls);
519 	if (error)
520 		return (error);
521 #endif
522 	return (0);
523 }
524 
525 /*
526  * Callback for unload of AIO when used as a module.
527  */
528 static int
529 aio_unload(void)
530 {
531 	int error;
532 
533 	/*
534 	 * XXX: no unloads by default, it's too dangerous.
535 	 * perhaps we could do it if locked out callers and then
536 	 * did an aio_proc_rundown() on each process.
537 	 *
538 	 * jhb: aio_proc_rundown() needs to run on curproc though,
539 	 * so I don't think that would fly.
540 	 */
541 	if (!unloadable)
542 		return (EOPNOTSUPP);
543 
544 #ifdef COMPAT_FREEBSD32
545 	syscall32_helper_unregister(aio32_syscalls);
546 #endif
547 	syscall_helper_unregister(aio_syscalls);
548 
549 	error = kqueue_del_filteropts(EVFILT_AIO);
550 	if (error)
551 		return error;
552 	error = kqueue_del_filteropts(EVFILT_LIO);
553 	if (error)
554 		return error;
555 	async_io_version = 0;
556 	aio_swake = NULL;
557 	taskqueue_free(taskqueue_aiod_bio);
558 	delete_unrhdr(aiod_unr);
559 	uma_zdestroy(kaio_zone);
560 	uma_zdestroy(aiop_zone);
561 	uma_zdestroy(aiocb_zone);
562 	uma_zdestroy(aiol_zone);
563 	uma_zdestroy(aiolio_zone);
564 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
565 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
566 	mtx_destroy(&aio_job_mtx);
567 	mtx_destroy(&aio_sock_mtx);
568 	sema_destroy(&aio_newproc_sem);
569 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
570 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
571 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
572 	return (0);
573 }
574 
575 /*
576  * Init the per-process aioinfo structure.  The aioinfo limits are set
577  * per-process for user limit (resource) management.
578  */
579 void
580 aio_init_aioinfo(struct proc *p)
581 {
582 	struct kaioinfo *ki;
583 
584 	ki = uma_zalloc(kaio_zone, M_WAITOK);
585 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
586 	ki->kaio_flags = 0;
587 	ki->kaio_maxactive_count = max_aio_per_proc;
588 	ki->kaio_active_count = 0;
589 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
590 	ki->kaio_count = 0;
591 	ki->kaio_ballowed_count = max_buf_aio;
592 	ki->kaio_buffer_count = 0;
593 	TAILQ_INIT(&ki->kaio_all);
594 	TAILQ_INIT(&ki->kaio_done);
595 	TAILQ_INIT(&ki->kaio_jobqueue);
596 	TAILQ_INIT(&ki->kaio_bufqueue);
597 	TAILQ_INIT(&ki->kaio_liojoblist);
598 	TAILQ_INIT(&ki->kaio_sockqueue);
599 	TAILQ_INIT(&ki->kaio_syncqueue);
600 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
601 	PROC_LOCK(p);
602 	if (p->p_aioinfo == NULL) {
603 		p->p_aioinfo = ki;
604 		PROC_UNLOCK(p);
605 	} else {
606 		PROC_UNLOCK(p);
607 		mtx_destroy(&ki->kaio_mtx);
608 		uma_zfree(kaio_zone, ki);
609 	}
610 
611 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
612 		aio_newproc(NULL);
613 }
614 
615 static int
616 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
617 {
618 	struct thread *td;
619 	int error;
620 
621 	error = sigev_findtd(p, sigev, &td);
622 	if (error)
623 		return (error);
624 	if (!KSI_ONQ(ksi)) {
625 		ksiginfo_set_sigev(ksi, sigev);
626 		ksi->ksi_code = SI_ASYNCIO;
627 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
628 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
629 	}
630 	PROC_UNLOCK(p);
631 	return (error);
632 }
633 
634 /*
635  * Free a job entry.  Wait for completion if it is currently active, but don't
636  * delay forever.  If we delay, we return a flag that says that we have to
637  * restart the queue scan.
638  */
639 static int
640 aio_free_entry(struct aiocblist *aiocbe)
641 {
642 	struct kaioinfo *ki;
643 	struct aioliojob *lj;
644 	struct proc *p;
645 
646 	p = aiocbe->userproc;
647 	MPASS(curproc == p);
648 	ki = p->p_aioinfo;
649 	MPASS(ki != NULL);
650 
651 	AIO_LOCK_ASSERT(ki, MA_OWNED);
652 	MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
653 
654 	atomic_subtract_int(&num_queue_count, 1);
655 
656 	ki->kaio_count--;
657 	MPASS(ki->kaio_count >= 0);
658 
659 	TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
660 	TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
661 
662 	lj = aiocbe->lio;
663 	if (lj) {
664 		lj->lioj_count--;
665 		lj->lioj_finished_count--;
666 
667 		if (lj->lioj_count == 0) {
668 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
669 			/* lio is going away, we need to destroy any knotes */
670 			knlist_delete(&lj->klist, curthread, 1);
671 			PROC_LOCK(p);
672 			sigqueue_take(&lj->lioj_ksi);
673 			PROC_UNLOCK(p);
674 			uma_zfree(aiolio_zone, lj);
675 		}
676 	}
677 
678 	/* aiocbe is going away, we need to destroy any knotes */
679 	knlist_delete(&aiocbe->klist, curthread, 1);
680 	PROC_LOCK(p);
681 	sigqueue_take(&aiocbe->ksi);
682 	PROC_UNLOCK(p);
683 
684 	MPASS(aiocbe->bp == NULL);
685 	aiocbe->jobstate = JOBST_NULL;
686 	AIO_UNLOCK(ki);
687 
688 	/*
689 	 * The thread argument here is used to find the owning process
690 	 * and is also passed to fo_close() which may pass it to various
691 	 * places such as devsw close() routines.  Because of that, we
692 	 * need a thread pointer from the process owning the job that is
693 	 * persistent and won't disappear out from under us or move to
694 	 * another process.
695 	 *
696 	 * Currently, all the callers of this function call it to remove
697 	 * an aiocblist from the current process' job list either via a
698 	 * syscall or due to the current process calling exit() or
699 	 * execve().  Thus, we know that p == curproc.  We also know that
700 	 * curthread can't exit since we are curthread.
701 	 *
702 	 * Therefore, we use curthread as the thread to pass to
703 	 * knlist_delete().  This does mean that it is possible for the
704 	 * thread pointer at close time to differ from the thread pointer
705 	 * at open time, but this is already true of file descriptors in
706 	 * a multithreaded process.
707 	 */
708 	if (aiocbe->fd_file)
709 		fdrop(aiocbe->fd_file, curthread);
710 	crfree(aiocbe->cred);
711 	uma_zfree(aiocb_zone, aiocbe);
712 	AIO_LOCK(ki);
713 
714 	return (0);
715 }
716 
717 static void
718 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
719 {
720    	aio_proc_rundown(arg, p);
721 }
722 
723 /*
724  * Rundown the jobs for a given process.
725  */
726 static void
727 aio_proc_rundown(void *arg, struct proc *p)
728 {
729 	struct kaioinfo *ki;
730 	struct aioliojob *lj;
731 	struct aiocblist *cbe, *cbn;
732 	struct file *fp;
733 	struct socket *so;
734 	int remove;
735 
736 	KASSERT(curthread->td_proc == p,
737 	    ("%s: called on non-curproc", __func__));
738 	ki = p->p_aioinfo;
739 	if (ki == NULL)
740 		return;
741 
742 	AIO_LOCK(ki);
743 	ki->kaio_flags |= KAIO_RUNDOWN;
744 
745 restart:
746 
747 	/*
748 	 * Try to cancel all pending requests. This code simulates
749 	 * aio_cancel on all pending I/O requests.
750 	 */
751 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
752 		remove = 0;
753 		mtx_lock(&aio_job_mtx);
754 		if (cbe->jobstate == JOBST_JOBQGLOBAL) {
755 			TAILQ_REMOVE(&aio_jobs, cbe, list);
756 			remove = 1;
757 		} else if (cbe->jobstate == JOBST_JOBQSOCK) {
758 			fp = cbe->fd_file;
759 			MPASS(fp->f_type == DTYPE_SOCKET);
760 			so = fp->f_data;
761 			TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
762 			remove = 1;
763 		} else if (cbe->jobstate == JOBST_JOBQSYNC) {
764 			TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
765 			remove = 1;
766 		}
767 		mtx_unlock(&aio_job_mtx);
768 
769 		if (remove) {
770 			cbe->jobstate = JOBST_JOBFINISHED;
771 			cbe->uaiocb._aiocb_private.status = -1;
772 			cbe->uaiocb._aiocb_private.error = ECANCELED;
773 			TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
774 			aio_bio_done_notify(p, cbe, DONE_QUEUE);
775 		}
776 	}
777 
778 	/* Wait for all running I/O to be finished */
779 	if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
780 	    TAILQ_FIRST(&ki->kaio_jobqueue)) {
781 		ki->kaio_flags |= KAIO_WAKEUP;
782 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
783 		goto restart;
784 	}
785 
786 	/* Free all completed I/O requests. */
787 	while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
788 		aio_free_entry(cbe);
789 
790 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
791 		if (lj->lioj_count == 0) {
792 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
793 			knlist_delete(&lj->klist, curthread, 1);
794 			PROC_LOCK(p);
795 			sigqueue_take(&lj->lioj_ksi);
796 			PROC_UNLOCK(p);
797 			uma_zfree(aiolio_zone, lj);
798 		} else {
799 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
800 			    lj->lioj_count, lj->lioj_finished_count);
801 		}
802 	}
803 	AIO_UNLOCK(ki);
804 	taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
805 	mtx_destroy(&ki->kaio_mtx);
806 	uma_zfree(kaio_zone, ki);
807 	p->p_aioinfo = NULL;
808 }
809 
810 /*
811  * Select a job to run (called by an AIO daemon).
812  */
813 static struct aiocblist *
814 aio_selectjob(struct aiothreadlist *aiop)
815 {
816 	struct aiocblist *aiocbe;
817 	struct kaioinfo *ki;
818 	struct proc *userp;
819 
820 	mtx_assert(&aio_job_mtx, MA_OWNED);
821 	TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
822 		userp = aiocbe->userproc;
823 		ki = userp->p_aioinfo;
824 
825 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
826 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
827 			/* Account for currently active jobs. */
828 			ki->kaio_active_count++;
829 			aiocbe->jobstate = JOBST_JOBRUNNING;
830 			break;
831 		}
832 	}
833 	return (aiocbe);
834 }
835 
836 /*
837  *  Move all data to a permanent storage device, this code
838  *  simulates fsync syscall.
839  */
840 static int
841 aio_fsync_vnode(struct thread *td, struct vnode *vp)
842 {
843 	struct mount *mp;
844 	int error;
845 
846 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
847 		goto drop;
848 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
849 	if (vp->v_object != NULL) {
850 		VM_OBJECT_WLOCK(vp->v_object);
851 		vm_object_page_clean(vp->v_object, 0, 0, 0);
852 		VM_OBJECT_WUNLOCK(vp->v_object);
853 	}
854 	error = VOP_FSYNC(vp, MNT_WAIT, td);
855 
856 	VOP_UNLOCK(vp, 0);
857 	vn_finished_write(mp);
858 drop:
859 	return (error);
860 }
861 
862 /*
863  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
864  * does the I/O request for the non-physio version of the operations.  The
865  * normal vn operations are used, and this code should work in all instances
866  * for every type of file, including pipes, sockets, fifos, and regular files.
867  *
868  * XXX I don't think it works well for socket, pipe, and fifo.
869  */
870 static void
871 aio_process_rw(struct aiocblist *aiocbe)
872 {
873 	struct ucred *td_savedcred;
874 	struct thread *td;
875 	struct aiocb *cb;
876 	struct file *fp;
877 	struct socket *so;
878 	struct uio auio;
879 	struct iovec aiov;
880 	int cnt;
881 	int error;
882 	int oublock_st, oublock_end;
883 	int inblock_st, inblock_end;
884 
885 	KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_READ ||
886 	    aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE,
887 	    ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
888 
889 	td = curthread;
890 	td_savedcred = td->td_ucred;
891 	td->td_ucred = aiocbe->cred;
892 	cb = &aiocbe->uaiocb;
893 	fp = aiocbe->fd_file;
894 
895 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
896 	aiov.iov_len = cb->aio_nbytes;
897 
898 	auio.uio_iov = &aiov;
899 	auio.uio_iovcnt = 1;
900 	auio.uio_offset = cb->aio_offset;
901 	auio.uio_resid = cb->aio_nbytes;
902 	cnt = cb->aio_nbytes;
903 	auio.uio_segflg = UIO_USERSPACE;
904 	auio.uio_td = td;
905 
906 	inblock_st = td->td_ru.ru_inblock;
907 	oublock_st = td->td_ru.ru_oublock;
908 	/*
909 	 * aio_aqueue() acquires a reference to the file that is
910 	 * released in aio_free_entry().
911 	 */
912 	if (cb->aio_lio_opcode == LIO_READ) {
913 		auio.uio_rw = UIO_READ;
914 		if (auio.uio_resid == 0)
915 			error = 0;
916 		else
917 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
918 	} else {
919 		if (fp->f_type == DTYPE_VNODE)
920 			bwillwrite();
921 		auio.uio_rw = UIO_WRITE;
922 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
923 	}
924 	inblock_end = td->td_ru.ru_inblock;
925 	oublock_end = td->td_ru.ru_oublock;
926 
927 	aiocbe->inputcharge = inblock_end - inblock_st;
928 	aiocbe->outputcharge = oublock_end - oublock_st;
929 
930 	if ((error) && (auio.uio_resid != cnt)) {
931 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
932 			error = 0;
933 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
934 			int sigpipe = 1;
935 			if (fp->f_type == DTYPE_SOCKET) {
936 				so = fp->f_data;
937 				if (so->so_options & SO_NOSIGPIPE)
938 					sigpipe = 0;
939 			}
940 			if (sigpipe) {
941 				PROC_LOCK(aiocbe->userproc);
942 				kern_psignal(aiocbe->userproc, SIGPIPE);
943 				PROC_UNLOCK(aiocbe->userproc);
944 			}
945 		}
946 	}
947 
948 	cnt -= auio.uio_resid;
949 	cb->_aiocb_private.error = error;
950 	cb->_aiocb_private.status = cnt;
951 	td->td_ucred = td_savedcred;
952 }
953 
954 static void
955 aio_process_sync(struct aiocblist *aiocbe)
956 {
957 	struct thread *td = curthread;
958 	struct ucred *td_savedcred = td->td_ucred;
959 	struct aiocb *cb = &aiocbe->uaiocb;
960 	struct file *fp = aiocbe->fd_file;
961 	int error = 0;
962 
963 	KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_SYNC,
964 	    ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
965 
966 	td->td_ucred = aiocbe->cred;
967 	if (fp->f_vnode != NULL)
968 		error = aio_fsync_vnode(td, fp->f_vnode);
969 	cb->_aiocb_private.error = error;
970 	cb->_aiocb_private.status = 0;
971 	td->td_ucred = td_savedcred;
972 }
973 
974 static void
975 aio_process_mlock(struct aiocblist *aiocbe)
976 {
977 	struct aiocb *cb = &aiocbe->uaiocb;
978 	int error;
979 
980 	KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_MLOCK,
981 	    ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
982 
983 	error = vm_mlock(aiocbe->userproc, aiocbe->cred,
984 	    __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
985 	cb->_aiocb_private.error = error;
986 	cb->_aiocb_private.status = 0;
987 }
988 
989 static void
990 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
991 {
992 	struct aioliojob *lj;
993 	struct kaioinfo *ki;
994 	struct aiocblist *scb, *scbn;
995 	int lj_done;
996 
997 	ki = userp->p_aioinfo;
998 	AIO_LOCK_ASSERT(ki, MA_OWNED);
999 	lj = aiocbe->lio;
1000 	lj_done = 0;
1001 	if (lj) {
1002 		lj->lioj_finished_count++;
1003 		if (lj->lioj_count == lj->lioj_finished_count)
1004 			lj_done = 1;
1005 	}
1006 	if (type == DONE_QUEUE) {
1007 		aiocbe->jobflags |= AIOCBLIST_DONE;
1008 	} else {
1009 		aiocbe->jobflags |= AIOCBLIST_BUFDONE;
1010 	}
1011 	TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
1012 	aiocbe->jobstate = JOBST_JOBFINISHED;
1013 
1014 	if (ki->kaio_flags & KAIO_RUNDOWN)
1015 		goto notification_done;
1016 
1017 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1018 	    aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
1019 		aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
1020 
1021 	KNOTE_LOCKED(&aiocbe->klist, 1);
1022 
1023 	if (lj_done) {
1024 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
1025 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
1026 			KNOTE_LOCKED(&lj->klist, 1);
1027 		}
1028 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
1029 		    == LIOJ_SIGNAL
1030 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
1031 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
1032 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
1033 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1034 		}
1035 	}
1036 
1037 notification_done:
1038 	if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1039 		TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1040 			if (aiocbe->fd_file == scb->fd_file &&
1041 			    aiocbe->seqno < scb->seqno) {
1042 				if (--scb->pending == 0) {
1043 					mtx_lock(&aio_job_mtx);
1044 					scb->jobstate = JOBST_JOBQGLOBAL;
1045 					TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1046 					TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1047 					aio_kick_nowait(userp);
1048 					mtx_unlock(&aio_job_mtx);
1049 				}
1050 			}
1051 		}
1052 	}
1053 	if (ki->kaio_flags & KAIO_WAKEUP) {
1054 		ki->kaio_flags &= ~KAIO_WAKEUP;
1055 		wakeup(&userp->p_aioinfo);
1056 	}
1057 }
1058 
1059 /*
1060  * The AIO daemon, most of the actual work is done in aio_process_*,
1061  * but the setup (and address space mgmt) is done in this routine.
1062  */
1063 static void
1064 aio_daemon(void *_id)
1065 {
1066 	struct aiocblist *aiocbe;
1067 	struct aiothreadlist *aiop;
1068 	struct kaioinfo *ki;
1069 	struct proc *curcp, *mycp, *userp;
1070 	struct vmspace *myvm, *tmpvm;
1071 	struct thread *td = curthread;
1072 	int id = (intptr_t)_id;
1073 
1074 	/*
1075 	 * Local copies of curproc (cp) and vmspace (myvm)
1076 	 */
1077 	mycp = td->td_proc;
1078 	myvm = mycp->p_vmspace;
1079 
1080 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1081 
1082 	/*
1083 	 * Allocate and ready the aio control info.  There is one aiop structure
1084 	 * per daemon.
1085 	 */
1086 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1087 	aiop->aiothread = td;
1088 	aiop->aiothreadflags = 0;
1089 
1090 	/* The daemon resides in its own pgrp. */
1091 	sys_setsid(td, NULL);
1092 
1093 	/*
1094 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1095 	 * and creating too many daemons.)
1096 	 */
1097 	sema_post(&aio_newproc_sem);
1098 
1099 	mtx_lock(&aio_job_mtx);
1100 	for (;;) {
1101 		/*
1102 		 * curcp is the current daemon process context.
1103 		 * userp is the current user process context.
1104 		 */
1105 		curcp = mycp;
1106 
1107 		/*
1108 		 * Take daemon off of free queue
1109 		 */
1110 		if (aiop->aiothreadflags & AIOP_FREE) {
1111 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1112 			aiop->aiothreadflags &= ~AIOP_FREE;
1113 		}
1114 
1115 		/*
1116 		 * Check for jobs.
1117 		 */
1118 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1119 			mtx_unlock(&aio_job_mtx);
1120 			userp = aiocbe->userproc;
1121 
1122 			/*
1123 			 * Connect to process address space for user program.
1124 			 */
1125 			if (userp != curcp) {
1126 				/*
1127 				 * Save the current address space that we are
1128 				 * connected to.
1129 				 */
1130 				tmpvm = mycp->p_vmspace;
1131 
1132 				/*
1133 				 * Point to the new user address space, and
1134 				 * refer to it.
1135 				 */
1136 				mycp->p_vmspace = userp->p_vmspace;
1137 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1138 
1139 				/* Activate the new mapping. */
1140 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1141 
1142 				/*
1143 				 * If the old address space wasn't the daemons
1144 				 * own address space, then we need to remove the
1145 				 * daemon's reference from the other process
1146 				 * that it was acting on behalf of.
1147 				 */
1148 				if (tmpvm != myvm) {
1149 					vmspace_free(tmpvm);
1150 				}
1151 				curcp = userp;
1152 			}
1153 
1154 			ki = userp->p_aioinfo;
1155 
1156 			/* Do the I/O function. */
1157 			switch(aiocbe->uaiocb.aio_lio_opcode) {
1158 			case LIO_READ:
1159 			case LIO_WRITE:
1160 				aio_process_rw(aiocbe);
1161 				break;
1162 			case LIO_SYNC:
1163 				aio_process_sync(aiocbe);
1164 				break;
1165 			case LIO_MLOCK:
1166 				aio_process_mlock(aiocbe);
1167 				break;
1168 			}
1169 
1170 			mtx_lock(&aio_job_mtx);
1171 			/* Decrement the active job count. */
1172 			ki->kaio_active_count--;
1173 			mtx_unlock(&aio_job_mtx);
1174 
1175 			AIO_LOCK(ki);
1176 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1177 			aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1178 			AIO_UNLOCK(ki);
1179 
1180 			mtx_lock(&aio_job_mtx);
1181 		}
1182 
1183 		/*
1184 		 * Disconnect from user address space.
1185 		 */
1186 		if (curcp != mycp) {
1187 
1188 			mtx_unlock(&aio_job_mtx);
1189 
1190 			/* Get the user address space to disconnect from. */
1191 			tmpvm = mycp->p_vmspace;
1192 
1193 			/* Get original address space for daemon. */
1194 			mycp->p_vmspace = myvm;
1195 
1196 			/* Activate the daemon's address space. */
1197 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1198 #ifdef DIAGNOSTIC
1199 			if (tmpvm == myvm) {
1200 				printf("AIOD: vmspace problem -- %d\n",
1201 				    mycp->p_pid);
1202 			}
1203 #endif
1204 			/* Remove our vmspace reference. */
1205 			vmspace_free(tmpvm);
1206 
1207 			curcp = mycp;
1208 
1209 			mtx_lock(&aio_job_mtx);
1210 			/*
1211 			 * We have to restart to avoid race, we only sleep if
1212 			 * no job can be selected, that should be
1213 			 * curcp == mycp.
1214 			 */
1215 			continue;
1216 		}
1217 
1218 		mtx_assert(&aio_job_mtx, MA_OWNED);
1219 
1220 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1221 		aiop->aiothreadflags |= AIOP_FREE;
1222 
1223 		/*
1224 		 * If daemon is inactive for a long time, allow it to exit,
1225 		 * thereby freeing resources.
1226 		 */
1227 		if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1228 		    aiod_lifetime)) {
1229 			if (TAILQ_EMPTY(&aio_jobs)) {
1230 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1231 				    (num_aio_procs > target_aio_procs)) {
1232 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1233 					num_aio_procs--;
1234 					mtx_unlock(&aio_job_mtx);
1235 					uma_zfree(aiop_zone, aiop);
1236 					free_unr(aiod_unr, id);
1237 #ifdef DIAGNOSTIC
1238 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1239 						printf("AIOD: bad vm refcnt for"
1240 						    " exiting daemon: %d\n",
1241 						    mycp->p_vmspace->vm_refcnt);
1242 					}
1243 #endif
1244 					kproc_exit(0);
1245 				}
1246 			}
1247 		}
1248 	}
1249 	mtx_unlock(&aio_job_mtx);
1250 	panic("shouldn't be here\n");
1251 }
1252 
1253 /*
1254  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1255  * AIO daemon modifies its environment itself.
1256  */
1257 static int
1258 aio_newproc(int *start)
1259 {
1260 	int error;
1261 	struct proc *p;
1262 	int id;
1263 
1264 	id = alloc_unr(aiod_unr);
1265 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1266 		RFNOWAIT, 0, "aiod%d", id);
1267 	if (error == 0) {
1268 		/*
1269 		 * Wait until daemon is started.
1270 		 */
1271 		sema_wait(&aio_newproc_sem);
1272 		mtx_lock(&aio_job_mtx);
1273 		num_aio_procs++;
1274 		if (start != NULL)
1275 			(*start)--;
1276 		mtx_unlock(&aio_job_mtx);
1277 	} else {
1278 		free_unr(aiod_unr, id);
1279 	}
1280 	return (error);
1281 }
1282 
1283 /*
1284  * Try the high-performance, low-overhead physio method for eligible
1285  * VCHR devices.  This method doesn't use an aio helper thread, and
1286  * thus has very low overhead.
1287  *
1288  * Assumes that the caller, aio_aqueue(), has incremented the file
1289  * structure's reference count, preventing its deallocation for the
1290  * duration of this call.
1291  */
1292 static int
1293 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1294 {
1295 	struct aiocb *cb;
1296 	struct file *fp;
1297 	struct buf *bp;
1298 	struct vnode *vp;
1299 	struct cdevsw *csw;
1300 	struct cdev *dev;
1301 	struct kaioinfo *ki;
1302 	struct aioliojob *lj;
1303 	int error, ref;
1304 
1305 	cb = &aiocbe->uaiocb;
1306 	fp = aiocbe->fd_file;
1307 
1308 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1309 		return (-1);
1310 
1311 	vp = fp->f_vnode;
1312 
1313 	/*
1314 	 * If its not a disk, we don't want to return a positive error.
1315 	 * It causes the aio code to not fall through to try the thread
1316 	 * way when you're talking to a regular file.
1317 	 */
1318 	if (!vn_isdisk(vp, &error)) {
1319 		if (error == ENOTBLK)
1320 			return (-1);
1321 		else
1322 			return (error);
1323 	}
1324 
1325 	if (vp->v_bufobj.bo_bsize == 0)
1326 		return (-1);
1327 
1328  	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1329 		return (-1);
1330 
1331 	if (cb->aio_nbytes >
1332 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1333 		return (-1);
1334 
1335 	ki = p->p_aioinfo;
1336 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1337 		return (-1);
1338 
1339 	ref = 0;
1340 	csw = devvn_refthread(vp, &dev, &ref);
1341 	if (csw == NULL)
1342 		return (ENXIO);
1343 	if (cb->aio_nbytes > dev->si_iosize_max) {
1344 		error = -1;
1345 		goto unref;
1346 	}
1347 
1348 	/* Create and build a buffer header for a transfer. */
1349 	bp = (struct buf *)getpbuf(NULL);
1350 	BUF_KERNPROC(bp);
1351 
1352 	AIO_LOCK(ki);
1353 	ki->kaio_count++;
1354 	ki->kaio_buffer_count++;
1355 	lj = aiocbe->lio;
1356 	if (lj)
1357 		lj->lioj_count++;
1358 	AIO_UNLOCK(ki);
1359 
1360 	/*
1361 	 * Get a copy of the kva from the physical buffer.
1362 	 */
1363 	error = 0;
1364 
1365 	bp->b_bcount = cb->aio_nbytes;
1366 	bp->b_bufsize = cb->aio_nbytes;
1367 	bp->b_iodone = aio_physwakeup;
1368 	bp->b_saveaddr = bp->b_data;
1369 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1370 	bp->b_offset = cb->aio_offset;
1371 	bp->b_iooffset = cb->aio_offset;
1372 	bp->b_blkno = btodb(cb->aio_offset);
1373 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1374 
1375 	/*
1376 	 * Bring buffer into kernel space.
1377 	 */
1378 	if (vmapbuf(bp, (dev->si_flags & SI_UNMAPPED) == 0) < 0) {
1379 		error = EFAULT;
1380 		goto doerror;
1381 	}
1382 
1383 	AIO_LOCK(ki);
1384 	aiocbe->bp = bp;
1385 	bp->b_caller1 = (void *)aiocbe;
1386 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1387 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1388 	aiocbe->jobstate = JOBST_JOBQBUF;
1389 	cb->_aiocb_private.status = cb->aio_nbytes;
1390 	AIO_UNLOCK(ki);
1391 
1392 	atomic_add_int(&num_queue_count, 1);
1393 	atomic_add_int(&num_buf_aio, 1);
1394 
1395 	bp->b_error = 0;
1396 
1397 	TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1398 
1399 	/* Perform transfer. */
1400 	dev_strategy_csw(dev, csw, bp);
1401 	dev_relthread(dev, ref);
1402 	return (0);
1403 
1404 doerror:
1405 	AIO_LOCK(ki);
1406 	ki->kaio_count--;
1407 	ki->kaio_buffer_count--;
1408 	if (lj)
1409 		lj->lioj_count--;
1410 	aiocbe->bp = NULL;
1411 	AIO_UNLOCK(ki);
1412 	relpbuf(bp, NULL);
1413 unref:
1414 	dev_relthread(dev, ref);
1415 	return (error);
1416 }
1417 
1418 /*
1419  * Wake up aio requests that may be serviceable now.
1420  */
1421 static void
1422 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1423 {
1424 	struct aiocblist *cb, *cbn;
1425 	int opcode;
1426 
1427 	SOCKBUF_LOCK_ASSERT(sb);
1428 	if (sb == &so->so_snd)
1429 		opcode = LIO_WRITE;
1430 	else
1431 		opcode = LIO_READ;
1432 
1433 	sb->sb_flags &= ~SB_AIO;
1434 	mtx_lock(&aio_job_mtx);
1435 	TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1436 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1437 			if (cb->jobstate != JOBST_JOBQSOCK)
1438 				panic("invalid queue value");
1439 			/* XXX
1440 			 * We don't have actual sockets backend yet,
1441 			 * so we simply move the requests to the generic
1442 			 * file I/O backend.
1443 			 */
1444 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1445 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1446 			aio_kick_nowait(cb->userproc);
1447 		}
1448 	}
1449 	mtx_unlock(&aio_job_mtx);
1450 }
1451 
1452 static int
1453 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1454 {
1455 
1456 	/*
1457 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1458 	 * supported by AIO with the old sigevent structure.
1459 	 */
1460 	nsig->sigev_notify = osig->sigev_notify;
1461 	switch (nsig->sigev_notify) {
1462 	case SIGEV_NONE:
1463 		break;
1464 	case SIGEV_SIGNAL:
1465 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1466 		break;
1467 	case SIGEV_KEVENT:
1468 		nsig->sigev_notify_kqueue =
1469 		    osig->__sigev_u.__sigev_notify_kqueue;
1470 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1471 		break;
1472 	default:
1473 		return (EINVAL);
1474 	}
1475 	return (0);
1476 }
1477 
1478 static int
1479 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1480 {
1481 	struct oaiocb *ojob;
1482 	int error;
1483 
1484 	bzero(kjob, sizeof(struct aiocb));
1485 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1486 	if (error)
1487 		return (error);
1488 	ojob = (struct oaiocb *)kjob;
1489 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1490 }
1491 
1492 static int
1493 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1494 {
1495 
1496 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1497 }
1498 
1499 static long
1500 aiocb_fetch_status(struct aiocb *ujob)
1501 {
1502 
1503 	return (fuword(&ujob->_aiocb_private.status));
1504 }
1505 
1506 static long
1507 aiocb_fetch_error(struct aiocb *ujob)
1508 {
1509 
1510 	return (fuword(&ujob->_aiocb_private.error));
1511 }
1512 
1513 static int
1514 aiocb_store_status(struct aiocb *ujob, long status)
1515 {
1516 
1517 	return (suword(&ujob->_aiocb_private.status, status));
1518 }
1519 
1520 static int
1521 aiocb_store_error(struct aiocb *ujob, long error)
1522 {
1523 
1524 	return (suword(&ujob->_aiocb_private.error, error));
1525 }
1526 
1527 static int
1528 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1529 {
1530 
1531 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1532 }
1533 
1534 static int
1535 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1536 {
1537 
1538 	return (suword(ujobp, (long)ujob));
1539 }
1540 
1541 static struct aiocb_ops aiocb_ops = {
1542 	.copyin = aiocb_copyin,
1543 	.fetch_status = aiocb_fetch_status,
1544 	.fetch_error = aiocb_fetch_error,
1545 	.store_status = aiocb_store_status,
1546 	.store_error = aiocb_store_error,
1547 	.store_kernelinfo = aiocb_store_kernelinfo,
1548 	.store_aiocb = aiocb_store_aiocb,
1549 };
1550 
1551 static struct aiocb_ops aiocb_ops_osigevent = {
1552 	.copyin = aiocb_copyin_old_sigevent,
1553 	.fetch_status = aiocb_fetch_status,
1554 	.fetch_error = aiocb_fetch_error,
1555 	.store_status = aiocb_store_status,
1556 	.store_error = aiocb_store_error,
1557 	.store_kernelinfo = aiocb_store_kernelinfo,
1558 	.store_aiocb = aiocb_store_aiocb,
1559 };
1560 
1561 /*
1562  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1563  * technique is done in this code.
1564  */
1565 int
1566 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1567 	int type, struct aiocb_ops *ops)
1568 {
1569 	struct proc *p = td->td_proc;
1570 	struct file *fp;
1571 	struct socket *so;
1572 	struct aiocblist *aiocbe, *cb;
1573 	struct kaioinfo *ki;
1574 	struct kevent kev;
1575 	struct sockbuf *sb;
1576 	int opcode;
1577 	int error;
1578 	int fd, kqfd;
1579 	int jid;
1580 	u_short evflags;
1581 
1582 	if (p->p_aioinfo == NULL)
1583 		aio_init_aioinfo(p);
1584 
1585 	ki = p->p_aioinfo;
1586 
1587 	ops->store_status(job, -1);
1588 	ops->store_error(job, 0);
1589 	ops->store_kernelinfo(job, -1);
1590 
1591 	if (num_queue_count >= max_queue_count ||
1592 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1593 		ops->store_error(job, EAGAIN);
1594 		return (EAGAIN);
1595 	}
1596 
1597 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1598 	knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1599 
1600 	error = ops->copyin(job, &aiocbe->uaiocb);
1601 	if (error) {
1602 		ops->store_error(job, error);
1603 		uma_zfree(aiocb_zone, aiocbe);
1604 		return (error);
1605 	}
1606 
1607 	/* XXX: aio_nbytes is later casted to signed types. */
1608 	if (aiocbe->uaiocb.aio_nbytes > INT_MAX) {
1609 		uma_zfree(aiocb_zone, aiocbe);
1610 		return (EINVAL);
1611 	}
1612 
1613 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1614 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1615 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1616 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1617 		ops->store_error(job, EINVAL);
1618 		uma_zfree(aiocb_zone, aiocbe);
1619 		return (EINVAL);
1620 	}
1621 
1622 	if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1623 	     aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1624 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1625 		uma_zfree(aiocb_zone, aiocbe);
1626 		return (EINVAL);
1627 	}
1628 
1629 	ksiginfo_init(&aiocbe->ksi);
1630 
1631 	/* Save userspace address of the job info. */
1632 	aiocbe->uuaiocb = job;
1633 
1634 	/* Get the opcode. */
1635 	if (type != LIO_NOP)
1636 		aiocbe->uaiocb.aio_lio_opcode = type;
1637 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1638 
1639 	/*
1640 	 * Validate the opcode and fetch the file object for the specified
1641 	 * file descriptor.
1642 	 *
1643 	 * XXXRW: Moved the opcode validation up here so that we don't
1644 	 * retrieve a file descriptor without knowing what the capabiltity
1645 	 * should be.
1646 	 */
1647 	fd = aiocbe->uaiocb.aio_fildes;
1648 	switch (opcode) {
1649 	case LIO_WRITE:
1650 		error = fget_write(td, fd, CAP_PWRITE, &fp);
1651 		break;
1652 	case LIO_READ:
1653 		error = fget_read(td, fd, CAP_PREAD, &fp);
1654 		break;
1655 	case LIO_SYNC:
1656 		error = fget(td, fd, CAP_FSYNC, &fp);
1657 		break;
1658 	case LIO_MLOCK:
1659 		fp = NULL;
1660 		break;
1661 	case LIO_NOP:
1662 		error = fget(td, fd, CAP_NONE, &fp);
1663 		break;
1664 	default:
1665 		error = EINVAL;
1666 	}
1667 	if (error) {
1668 		uma_zfree(aiocb_zone, aiocbe);
1669 		ops->store_error(job, error);
1670 		return (error);
1671 	}
1672 
1673 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1674 		error = EINVAL;
1675 		goto aqueue_fail;
1676 	}
1677 
1678 	if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1679 		error = EINVAL;
1680 		goto aqueue_fail;
1681 	}
1682 
1683 	aiocbe->fd_file = fp;
1684 
1685 	mtx_lock(&aio_job_mtx);
1686 	jid = jobrefid++;
1687 	aiocbe->seqno = jobseqno++;
1688 	mtx_unlock(&aio_job_mtx);
1689 	error = ops->store_kernelinfo(job, jid);
1690 	if (error) {
1691 		error = EINVAL;
1692 		goto aqueue_fail;
1693 	}
1694 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1695 
1696 	if (opcode == LIO_NOP) {
1697 		fdrop(fp, td);
1698 		uma_zfree(aiocb_zone, aiocbe);
1699 		return (0);
1700 	}
1701 
1702 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1703 		goto no_kqueue;
1704 	evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1705 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1706 		error = EINVAL;
1707 		goto aqueue_fail;
1708 	}
1709 	kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1710 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1711 	kev.filter = EVFILT_AIO;
1712 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1713 	kev.data = (intptr_t)aiocbe;
1714 	kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1715 	error = kqfd_register(kqfd, &kev, td, 1);
1716 aqueue_fail:
1717 	if (error) {
1718 		if (fp)
1719 			fdrop(fp, td);
1720 		uma_zfree(aiocb_zone, aiocbe);
1721 		ops->store_error(job, error);
1722 		goto done;
1723 	}
1724 no_kqueue:
1725 
1726 	ops->store_error(job, EINPROGRESS);
1727 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1728 	aiocbe->userproc = p;
1729 	aiocbe->cred = crhold(td->td_ucred);
1730 	aiocbe->jobflags = 0;
1731 	aiocbe->lio = lj;
1732 
1733 	if (opcode == LIO_SYNC)
1734 		goto queueit;
1735 
1736 	if (fp && fp->f_type == DTYPE_SOCKET) {
1737 		/*
1738 		 * Alternate queueing for socket ops: Reach down into the
1739 		 * descriptor to get the socket data.  Then check to see if the
1740 		 * socket is ready to be read or written (based on the requested
1741 		 * operation).
1742 		 *
1743 		 * If it is not ready for io, then queue the aiocbe on the
1744 		 * socket, and set the flags so we get a call when sbnotify()
1745 		 * happens.
1746 		 *
1747 		 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1748 		 * and unlock the snd sockbuf for no reason.
1749 		 */
1750 		so = fp->f_data;
1751 		sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1752 		SOCKBUF_LOCK(sb);
1753 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1754 		    LIO_WRITE) && (!sowriteable(so)))) {
1755 			sb->sb_flags |= SB_AIO;
1756 
1757 			mtx_lock(&aio_job_mtx);
1758 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1759 			mtx_unlock(&aio_job_mtx);
1760 
1761 			AIO_LOCK(ki);
1762 			TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1763 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1764 			aiocbe->jobstate = JOBST_JOBQSOCK;
1765 			ki->kaio_count++;
1766 			if (lj)
1767 				lj->lioj_count++;
1768 			AIO_UNLOCK(ki);
1769 			SOCKBUF_UNLOCK(sb);
1770 			atomic_add_int(&num_queue_count, 1);
1771 			error = 0;
1772 			goto done;
1773 		}
1774 		SOCKBUF_UNLOCK(sb);
1775 	}
1776 
1777 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1778 		goto done;
1779 #if 0
1780 	if (error > 0) {
1781 		aiocbe->uaiocb._aiocb_private.error = error;
1782 		ops->store_error(job, error);
1783 		goto done;
1784 	}
1785 #endif
1786 queueit:
1787 	/* No buffer for daemon I/O. */
1788 	aiocbe->bp = NULL;
1789 	atomic_add_int(&num_queue_count, 1);
1790 
1791 	AIO_LOCK(ki);
1792 	ki->kaio_count++;
1793 	if (lj)
1794 		lj->lioj_count++;
1795 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1796 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1797 	if (opcode == LIO_SYNC) {
1798 		TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1799 			if (cb->fd_file == aiocbe->fd_file &&
1800 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1801 			    cb->seqno < aiocbe->seqno) {
1802 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1803 				aiocbe->pending++;
1804 			}
1805 		}
1806 		TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1807 			if (cb->fd_file == aiocbe->fd_file &&
1808 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1809 			    cb->seqno < aiocbe->seqno) {
1810 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1811 				aiocbe->pending++;
1812 			}
1813 		}
1814 		if (aiocbe->pending != 0) {
1815 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1816 			aiocbe->jobstate = JOBST_JOBQSYNC;
1817 			AIO_UNLOCK(ki);
1818 			goto done;
1819 		}
1820 	}
1821 	mtx_lock(&aio_job_mtx);
1822 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1823 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1824 	aio_kick_nowait(p);
1825 	mtx_unlock(&aio_job_mtx);
1826 	AIO_UNLOCK(ki);
1827 	error = 0;
1828 done:
1829 	return (error);
1830 }
1831 
1832 static void
1833 aio_kick_nowait(struct proc *userp)
1834 {
1835 	struct kaioinfo *ki = userp->p_aioinfo;
1836 	struct aiothreadlist *aiop;
1837 
1838 	mtx_assert(&aio_job_mtx, MA_OWNED);
1839 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1840 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1841 		aiop->aiothreadflags &= ~AIOP_FREE;
1842 		wakeup(aiop->aiothread);
1843 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1844 	    ((ki->kaio_active_count + num_aio_resv_start) <
1845 	    ki->kaio_maxactive_count)) {
1846 		taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1847 	}
1848 }
1849 
1850 static int
1851 aio_kick(struct proc *userp)
1852 {
1853 	struct kaioinfo *ki = userp->p_aioinfo;
1854 	struct aiothreadlist *aiop;
1855 	int error, ret = 0;
1856 
1857 	mtx_assert(&aio_job_mtx, MA_OWNED);
1858 retryproc:
1859 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1860 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1861 		aiop->aiothreadflags &= ~AIOP_FREE;
1862 		wakeup(aiop->aiothread);
1863 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1864 	    ((ki->kaio_active_count + num_aio_resv_start) <
1865 	    ki->kaio_maxactive_count)) {
1866 		num_aio_resv_start++;
1867 		mtx_unlock(&aio_job_mtx);
1868 		error = aio_newproc(&num_aio_resv_start);
1869 		mtx_lock(&aio_job_mtx);
1870 		if (error) {
1871 			num_aio_resv_start--;
1872 			goto retryproc;
1873 		}
1874 	} else {
1875 		ret = -1;
1876 	}
1877 	return (ret);
1878 }
1879 
1880 static void
1881 aio_kick_helper(void *context, int pending)
1882 {
1883 	struct proc *userp = context;
1884 
1885 	mtx_lock(&aio_job_mtx);
1886 	while (--pending >= 0) {
1887 		if (aio_kick(userp))
1888 			break;
1889 	}
1890 	mtx_unlock(&aio_job_mtx);
1891 }
1892 
1893 /*
1894  * Support the aio_return system call, as a side-effect, kernel resources are
1895  * released.
1896  */
1897 static int
1898 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1899 {
1900 	struct proc *p = td->td_proc;
1901 	struct aiocblist *cb;
1902 	struct kaioinfo *ki;
1903 	int status, error;
1904 
1905 	ki = p->p_aioinfo;
1906 	if (ki == NULL)
1907 		return (EINVAL);
1908 	AIO_LOCK(ki);
1909 	TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1910 		if (cb->uuaiocb == uaiocb)
1911 			break;
1912 	}
1913 	if (cb != NULL) {
1914 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
1915 		status = cb->uaiocb._aiocb_private.status;
1916 		error = cb->uaiocb._aiocb_private.error;
1917 		td->td_retval[0] = status;
1918 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1919 			td->td_ru.ru_oublock += cb->outputcharge;
1920 			cb->outputcharge = 0;
1921 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1922 			td->td_ru.ru_inblock += cb->inputcharge;
1923 			cb->inputcharge = 0;
1924 		}
1925 		aio_free_entry(cb);
1926 		AIO_UNLOCK(ki);
1927 		ops->store_error(uaiocb, error);
1928 		ops->store_status(uaiocb, status);
1929 	} else {
1930 		error = EINVAL;
1931 		AIO_UNLOCK(ki);
1932 	}
1933 	return (error);
1934 }
1935 
1936 int
1937 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1938 {
1939 
1940 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1941 }
1942 
1943 /*
1944  * Allow a process to wakeup when any of the I/O requests are completed.
1945  */
1946 static int
1947 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1948     struct timespec *ts)
1949 {
1950 	struct proc *p = td->td_proc;
1951 	struct timeval atv;
1952 	struct kaioinfo *ki;
1953 	struct aiocblist *cb, *cbfirst;
1954 	int error, i, timo;
1955 
1956 	timo = 0;
1957 	if (ts) {
1958 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1959 			return (EINVAL);
1960 
1961 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1962 		if (itimerfix(&atv))
1963 			return (EINVAL);
1964 		timo = tvtohz(&atv);
1965 	}
1966 
1967 	ki = p->p_aioinfo;
1968 	if (ki == NULL)
1969 		return (EAGAIN);
1970 
1971 	if (njoblist == 0)
1972 		return (0);
1973 
1974 	AIO_LOCK(ki);
1975 	for (;;) {
1976 		cbfirst = NULL;
1977 		error = 0;
1978 		TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1979 			for (i = 0; i < njoblist; i++) {
1980 				if (cb->uuaiocb == ujoblist[i]) {
1981 					if (cbfirst == NULL)
1982 						cbfirst = cb;
1983 					if (cb->jobstate == JOBST_JOBFINISHED)
1984 						goto RETURN;
1985 				}
1986 			}
1987 		}
1988 		/* All tasks were finished. */
1989 		if (cbfirst == NULL)
1990 			break;
1991 
1992 		ki->kaio_flags |= KAIO_WAKEUP;
1993 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1994 		    "aiospn", timo);
1995 		if (error == ERESTART)
1996 			error = EINTR;
1997 		if (error)
1998 			break;
1999 	}
2000 RETURN:
2001 	AIO_UNLOCK(ki);
2002 	return (error);
2003 }
2004 
2005 int
2006 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2007 {
2008 	struct timespec ts, *tsp;
2009 	struct aiocb **ujoblist;
2010 	int error;
2011 
2012 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2013 		return (EINVAL);
2014 
2015 	if (uap->timeout) {
2016 		/* Get timespec struct. */
2017 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2018 			return (error);
2019 		tsp = &ts;
2020 	} else
2021 		tsp = NULL;
2022 
2023 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2024 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2025 	if (error == 0)
2026 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2027 	uma_zfree(aiol_zone, ujoblist);
2028 	return (error);
2029 }
2030 
2031 /*
2032  * aio_cancel cancels any non-physio aio operations not currently in
2033  * progress.
2034  */
2035 int
2036 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2037 {
2038 	struct proc *p = td->td_proc;
2039 	struct kaioinfo *ki;
2040 	struct aiocblist *cbe, *cbn;
2041 	struct file *fp;
2042 	struct socket *so;
2043 	int error;
2044 	int remove;
2045 	int cancelled = 0;
2046 	int notcancelled = 0;
2047 	struct vnode *vp;
2048 
2049 	/* Lookup file object. */
2050 	error = fget(td, uap->fd, 0, &fp);
2051 	if (error)
2052 		return (error);
2053 
2054 	ki = p->p_aioinfo;
2055 	if (ki == NULL)
2056 		goto done;
2057 
2058 	if (fp->f_type == DTYPE_VNODE) {
2059 		vp = fp->f_vnode;
2060 		if (vn_isdisk(vp, &error)) {
2061 			fdrop(fp, td);
2062 			td->td_retval[0] = AIO_NOTCANCELED;
2063 			return (0);
2064 		}
2065 	}
2066 
2067 	AIO_LOCK(ki);
2068 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
2069 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
2070 		    ((uap->aiocbp == NULL) ||
2071 		     (uap->aiocbp == cbe->uuaiocb))) {
2072 			remove = 0;
2073 
2074 			mtx_lock(&aio_job_mtx);
2075 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
2076 				TAILQ_REMOVE(&aio_jobs, cbe, list);
2077 				remove = 1;
2078 			} else if (cbe->jobstate == JOBST_JOBQSOCK) {
2079 				MPASS(fp->f_type == DTYPE_SOCKET);
2080 				so = fp->f_data;
2081 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2082 				remove = 1;
2083 			} else if (cbe->jobstate == JOBST_JOBQSYNC) {
2084 				TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2085 				remove = 1;
2086 			}
2087 			mtx_unlock(&aio_job_mtx);
2088 
2089 			if (remove) {
2090 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2091 				cbe->uaiocb._aiocb_private.status = -1;
2092 				cbe->uaiocb._aiocb_private.error = ECANCELED;
2093 				aio_bio_done_notify(p, cbe, DONE_QUEUE);
2094 				cancelled++;
2095 			} else {
2096 				notcancelled++;
2097 			}
2098 			if (uap->aiocbp != NULL)
2099 				break;
2100 		}
2101 	}
2102 	AIO_UNLOCK(ki);
2103 
2104 done:
2105 	fdrop(fp, td);
2106 
2107 	if (uap->aiocbp != NULL) {
2108 		if (cancelled) {
2109 			td->td_retval[0] = AIO_CANCELED;
2110 			return (0);
2111 		}
2112 	}
2113 
2114 	if (notcancelled) {
2115 		td->td_retval[0] = AIO_NOTCANCELED;
2116 		return (0);
2117 	}
2118 
2119 	if (cancelled) {
2120 		td->td_retval[0] = AIO_CANCELED;
2121 		return (0);
2122 	}
2123 
2124 	td->td_retval[0] = AIO_ALLDONE;
2125 
2126 	return (0);
2127 }
2128 
2129 /*
2130  * aio_error is implemented in the kernel level for compatibility purposes
2131  * only.  For a user mode async implementation, it would be best to do it in
2132  * a userland subroutine.
2133  */
2134 static int
2135 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2136 {
2137 	struct proc *p = td->td_proc;
2138 	struct aiocblist *cb;
2139 	struct kaioinfo *ki;
2140 	int status;
2141 
2142 	ki = p->p_aioinfo;
2143 	if (ki == NULL) {
2144 		td->td_retval[0] = EINVAL;
2145 		return (0);
2146 	}
2147 
2148 	AIO_LOCK(ki);
2149 	TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2150 		if (cb->uuaiocb == aiocbp) {
2151 			if (cb->jobstate == JOBST_JOBFINISHED)
2152 				td->td_retval[0] =
2153 					cb->uaiocb._aiocb_private.error;
2154 			else
2155 				td->td_retval[0] = EINPROGRESS;
2156 			AIO_UNLOCK(ki);
2157 			return (0);
2158 		}
2159 	}
2160 	AIO_UNLOCK(ki);
2161 
2162 	/*
2163 	 * Hack for failure of aio_aqueue.
2164 	 */
2165 	status = ops->fetch_status(aiocbp);
2166 	if (status == -1) {
2167 		td->td_retval[0] = ops->fetch_error(aiocbp);
2168 		return (0);
2169 	}
2170 
2171 	td->td_retval[0] = EINVAL;
2172 	return (0);
2173 }
2174 
2175 int
2176 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2177 {
2178 
2179 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2180 }
2181 
2182 /* syscall - asynchronous read from a file (REALTIME) */
2183 int
2184 sys_oaio_read(struct thread *td, struct oaio_read_args *uap)
2185 {
2186 
2187 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2188 	    &aiocb_ops_osigevent));
2189 }
2190 
2191 int
2192 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2193 {
2194 
2195 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2196 }
2197 
2198 /* syscall - asynchronous write to a file (REALTIME) */
2199 int
2200 sys_oaio_write(struct thread *td, struct oaio_write_args *uap)
2201 {
2202 
2203 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2204 	    &aiocb_ops_osigevent));
2205 }
2206 
2207 int
2208 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2209 {
2210 
2211 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2212 }
2213 
2214 int
2215 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2216 {
2217 
2218 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2219 }
2220 
2221 static int
2222 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2223     struct aiocb **acb_list, int nent, struct sigevent *sig,
2224     struct aiocb_ops *ops)
2225 {
2226 	struct proc *p = td->td_proc;
2227 	struct aiocb *iocb;
2228 	struct kaioinfo *ki;
2229 	struct aioliojob *lj;
2230 	struct kevent kev;
2231 	int error;
2232 	int nerror;
2233 	int i;
2234 
2235 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2236 		return (EINVAL);
2237 
2238 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2239 		return (EINVAL);
2240 
2241 	if (p->p_aioinfo == NULL)
2242 		aio_init_aioinfo(p);
2243 
2244 	ki = p->p_aioinfo;
2245 
2246 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2247 	lj->lioj_flags = 0;
2248 	lj->lioj_count = 0;
2249 	lj->lioj_finished_count = 0;
2250 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2251 	ksiginfo_init(&lj->lioj_ksi);
2252 
2253 	/*
2254 	 * Setup signal.
2255 	 */
2256 	if (sig && (mode == LIO_NOWAIT)) {
2257 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2258 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2259 			/* Assume only new style KEVENT */
2260 			kev.filter = EVFILT_LIO;
2261 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2262 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2263 			kev.data = (intptr_t)lj;
2264 			/* pass user defined sigval data */
2265 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2266 			error = kqfd_register(
2267 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2268 			if (error) {
2269 				uma_zfree(aiolio_zone, lj);
2270 				return (error);
2271 			}
2272 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2273 			;
2274 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2275 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2276 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2277 					uma_zfree(aiolio_zone, lj);
2278 					return EINVAL;
2279 				}
2280 				lj->lioj_flags |= LIOJ_SIGNAL;
2281 		} else {
2282 			uma_zfree(aiolio_zone, lj);
2283 			return EINVAL;
2284 		}
2285 	}
2286 
2287 	AIO_LOCK(ki);
2288 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2289 	/*
2290 	 * Add extra aiocb count to avoid the lio to be freed
2291 	 * by other threads doing aio_waitcomplete or aio_return,
2292 	 * and prevent event from being sent until we have queued
2293 	 * all tasks.
2294 	 */
2295 	lj->lioj_count = 1;
2296 	AIO_UNLOCK(ki);
2297 
2298 	/*
2299 	 * Get pointers to the list of I/O requests.
2300 	 */
2301 	nerror = 0;
2302 	for (i = 0; i < nent; i++) {
2303 		iocb = acb_list[i];
2304 		if (iocb != NULL) {
2305 			error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2306 			if (error != 0)
2307 				nerror++;
2308 		}
2309 	}
2310 
2311 	error = 0;
2312 	AIO_LOCK(ki);
2313 	if (mode == LIO_WAIT) {
2314 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2315 			ki->kaio_flags |= KAIO_WAKEUP;
2316 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2317 			    PRIBIO | PCATCH, "aiospn", 0);
2318 			if (error == ERESTART)
2319 				error = EINTR;
2320 			if (error)
2321 				break;
2322 		}
2323 	} else {
2324 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2325 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2326 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2327 				KNOTE_LOCKED(&lj->klist, 1);
2328 			}
2329 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2330 			    == LIOJ_SIGNAL
2331 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2332 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2333 				aio_sendsig(p, &lj->lioj_signal,
2334 					    &lj->lioj_ksi);
2335 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2336 			}
2337 		}
2338 	}
2339 	lj->lioj_count--;
2340 	if (lj->lioj_count == 0) {
2341 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2342 		knlist_delete(&lj->klist, curthread, 1);
2343 		PROC_LOCK(p);
2344 		sigqueue_take(&lj->lioj_ksi);
2345 		PROC_UNLOCK(p);
2346 		AIO_UNLOCK(ki);
2347 		uma_zfree(aiolio_zone, lj);
2348 	} else
2349 		AIO_UNLOCK(ki);
2350 
2351 	if (nerror)
2352 		return (EIO);
2353 	return (error);
2354 }
2355 
2356 /* syscall - list directed I/O (REALTIME) */
2357 int
2358 sys_olio_listio(struct thread *td, struct olio_listio_args *uap)
2359 {
2360 	struct aiocb **acb_list;
2361 	struct sigevent *sigp, sig;
2362 	struct osigevent osig;
2363 	int error, nent;
2364 
2365 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2366 		return (EINVAL);
2367 
2368 	nent = uap->nent;
2369 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2370 		return (EINVAL);
2371 
2372 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2373 		error = copyin(uap->sig, &osig, sizeof(osig));
2374 		if (error)
2375 			return (error);
2376 		error = convert_old_sigevent(&osig, &sig);
2377 		if (error)
2378 			return (error);
2379 		sigp = &sig;
2380 	} else
2381 		sigp = NULL;
2382 
2383 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2384 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2385 	if (error == 0)
2386 		error = kern_lio_listio(td, uap->mode,
2387 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2388 		    &aiocb_ops_osigevent);
2389 	free(acb_list, M_LIO);
2390 	return (error);
2391 }
2392 
2393 /* syscall - list directed I/O (REALTIME) */
2394 int
2395 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2396 {
2397 	struct aiocb **acb_list;
2398 	struct sigevent *sigp, sig;
2399 	int error, nent;
2400 
2401 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2402 		return (EINVAL);
2403 
2404 	nent = uap->nent;
2405 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2406 		return (EINVAL);
2407 
2408 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2409 		error = copyin(uap->sig, &sig, sizeof(sig));
2410 		if (error)
2411 			return (error);
2412 		sigp = &sig;
2413 	} else
2414 		sigp = NULL;
2415 
2416 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2417 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2418 	if (error == 0)
2419 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2420 		    nent, sigp, &aiocb_ops);
2421 	free(acb_list, M_LIO);
2422 	return (error);
2423 }
2424 
2425 /*
2426  * Called from interrupt thread for physio, we should return as fast
2427  * as possible, so we schedule a biohelper task.
2428  */
2429 static void
2430 aio_physwakeup(struct buf *bp)
2431 {
2432 	struct aiocblist *aiocbe;
2433 
2434 	aiocbe = (struct aiocblist *)bp->b_caller1;
2435 	taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2436 }
2437 
2438 /*
2439  * Task routine to perform heavy tasks, process wakeup, and signals.
2440  */
2441 static void
2442 biohelper(void *context, int pending)
2443 {
2444 	struct aiocblist *aiocbe = context;
2445 	struct buf *bp;
2446 	struct proc *userp;
2447 	struct kaioinfo *ki;
2448 	int nblks;
2449 
2450 	bp = aiocbe->bp;
2451 	userp = aiocbe->userproc;
2452 	ki = userp->p_aioinfo;
2453 	AIO_LOCK(ki);
2454 	aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2455 	aiocbe->uaiocb._aiocb_private.error = 0;
2456 	if (bp->b_ioflags & BIO_ERROR)
2457 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2458 	nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2459 	if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2460 		aiocbe->outputcharge += nblks;
2461 	else
2462 		aiocbe->inputcharge += nblks;
2463 	aiocbe->bp = NULL;
2464 	TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2465 	ki->kaio_buffer_count--;
2466 	aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2467 	AIO_UNLOCK(ki);
2468 
2469 	/* Release mapping into kernel space. */
2470 	vunmapbuf(bp);
2471 	relpbuf(bp, NULL);
2472 	atomic_subtract_int(&num_buf_aio, 1);
2473 }
2474 
2475 /* syscall - wait for the next completion of an aio request */
2476 static int
2477 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2478     struct timespec *ts, struct aiocb_ops *ops)
2479 {
2480 	struct proc *p = td->td_proc;
2481 	struct timeval atv;
2482 	struct kaioinfo *ki;
2483 	struct aiocblist *cb;
2484 	struct aiocb *uuaiocb;
2485 	int error, status, timo;
2486 
2487 	ops->store_aiocb(aiocbp, NULL);
2488 
2489 	timo = 0;
2490 	if (ts) {
2491 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2492 			return (EINVAL);
2493 
2494 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2495 		if (itimerfix(&atv))
2496 			return (EINVAL);
2497 		timo = tvtohz(&atv);
2498 	}
2499 
2500 	if (p->p_aioinfo == NULL)
2501 		aio_init_aioinfo(p);
2502 	ki = p->p_aioinfo;
2503 
2504 	error = 0;
2505 	cb = NULL;
2506 	AIO_LOCK(ki);
2507 	while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2508 		ki->kaio_flags |= KAIO_WAKEUP;
2509 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2510 		    "aiowc", timo);
2511 		if (timo && error == ERESTART)
2512 			error = EINTR;
2513 		if (error)
2514 			break;
2515 	}
2516 
2517 	if (cb != NULL) {
2518 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
2519 		uuaiocb = cb->uuaiocb;
2520 		status = cb->uaiocb._aiocb_private.status;
2521 		error = cb->uaiocb._aiocb_private.error;
2522 		td->td_retval[0] = status;
2523 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2524 			td->td_ru.ru_oublock += cb->outputcharge;
2525 			cb->outputcharge = 0;
2526 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2527 			td->td_ru.ru_inblock += cb->inputcharge;
2528 			cb->inputcharge = 0;
2529 		}
2530 		aio_free_entry(cb);
2531 		AIO_UNLOCK(ki);
2532 		ops->store_aiocb(aiocbp, uuaiocb);
2533 		ops->store_error(uuaiocb, error);
2534 		ops->store_status(uuaiocb, status);
2535 	} else
2536 		AIO_UNLOCK(ki);
2537 
2538 	return (error);
2539 }
2540 
2541 int
2542 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2543 {
2544 	struct timespec ts, *tsp;
2545 	int error;
2546 
2547 	if (uap->timeout) {
2548 		/* Get timespec struct. */
2549 		error = copyin(uap->timeout, &ts, sizeof(ts));
2550 		if (error)
2551 			return (error);
2552 		tsp = &ts;
2553 	} else
2554 		tsp = NULL;
2555 
2556 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2557 }
2558 
2559 static int
2560 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2561     struct aiocb_ops *ops)
2562 {
2563 	struct proc *p = td->td_proc;
2564 	struct kaioinfo *ki;
2565 
2566 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2567 		return (EINVAL);
2568 	ki = p->p_aioinfo;
2569 	if (ki == NULL)
2570 		aio_init_aioinfo(p);
2571 	return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2572 }
2573 
2574 int
2575 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2576 {
2577 
2578 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2579 }
2580 
2581 /* kqueue attach function */
2582 static int
2583 filt_aioattach(struct knote *kn)
2584 {
2585 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2586 
2587 	/*
2588 	 * The aiocbe pointer must be validated before using it, so
2589 	 * registration is restricted to the kernel; the user cannot
2590 	 * set EV_FLAG1.
2591 	 */
2592 	if ((kn->kn_flags & EV_FLAG1) == 0)
2593 		return (EPERM);
2594 	kn->kn_ptr.p_aio = aiocbe;
2595 	kn->kn_flags &= ~EV_FLAG1;
2596 
2597 	knlist_add(&aiocbe->klist, kn, 0);
2598 
2599 	return (0);
2600 }
2601 
2602 /* kqueue detach function */
2603 static void
2604 filt_aiodetach(struct knote *kn)
2605 {
2606 	struct knlist *knl;
2607 
2608 	knl = &kn->kn_ptr.p_aio->klist;
2609 	knl->kl_lock(knl->kl_lockarg);
2610 	if (!knlist_empty(knl))
2611 		knlist_remove(knl, kn, 1);
2612 	knl->kl_unlock(knl->kl_lockarg);
2613 }
2614 
2615 /* kqueue filter function */
2616 /*ARGSUSED*/
2617 static int
2618 filt_aio(struct knote *kn, long hint)
2619 {
2620 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2621 
2622 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2623 	if (aiocbe->jobstate != JOBST_JOBFINISHED)
2624 		return (0);
2625 	kn->kn_flags |= EV_EOF;
2626 	return (1);
2627 }
2628 
2629 /* kqueue attach function */
2630 static int
2631 filt_lioattach(struct knote *kn)
2632 {
2633 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2634 
2635 	/*
2636 	 * The aioliojob pointer must be validated before using it, so
2637 	 * registration is restricted to the kernel; the user cannot
2638 	 * set EV_FLAG1.
2639 	 */
2640 	if ((kn->kn_flags & EV_FLAG1) == 0)
2641 		return (EPERM);
2642 	kn->kn_ptr.p_lio = lj;
2643 	kn->kn_flags &= ~EV_FLAG1;
2644 
2645 	knlist_add(&lj->klist, kn, 0);
2646 
2647 	return (0);
2648 }
2649 
2650 /* kqueue detach function */
2651 static void
2652 filt_liodetach(struct knote *kn)
2653 {
2654 	struct knlist *knl;
2655 
2656 	knl = &kn->kn_ptr.p_lio->klist;
2657 	knl->kl_lock(knl->kl_lockarg);
2658 	if (!knlist_empty(knl))
2659 		knlist_remove(knl, kn, 1);
2660 	knl->kl_unlock(knl->kl_lockarg);
2661 }
2662 
2663 /* kqueue filter function */
2664 /*ARGSUSED*/
2665 static int
2666 filt_lio(struct knote *kn, long hint)
2667 {
2668 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2669 
2670 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2671 }
2672 
2673 #ifdef COMPAT_FREEBSD32
2674 
2675 struct __aiocb_private32 {
2676 	int32_t	status;
2677 	int32_t	error;
2678 	uint32_t kernelinfo;
2679 };
2680 
2681 typedef struct oaiocb32 {
2682 	int	aio_fildes;		/* File descriptor */
2683 	uint64_t aio_offset __packed;	/* File offset for I/O */
2684 	uint32_t aio_buf;		/* I/O buffer in process space */
2685 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2686 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2687 	int	aio_lio_opcode;		/* LIO opcode */
2688 	int	aio_reqprio;		/* Request priority -- ignored */
2689 	struct	__aiocb_private32 _aiocb_private;
2690 } oaiocb32_t;
2691 
2692 typedef struct aiocb32 {
2693 	int32_t	aio_fildes;		/* File descriptor */
2694 	uint64_t aio_offset __packed;	/* File offset for I/O */
2695 	uint32_t aio_buf;		/* I/O buffer in process space */
2696 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2697 	int	__spare__[2];
2698 	uint32_t __spare2__;
2699 	int	aio_lio_opcode;		/* LIO opcode */
2700 	int	aio_reqprio;		/* Request priority -- ignored */
2701 	struct __aiocb_private32 _aiocb_private;
2702 	struct sigevent32 aio_sigevent;	/* Signal to deliver */
2703 } aiocb32_t;
2704 
2705 static int
2706 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2707 {
2708 
2709 	/*
2710 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2711 	 * supported by AIO with the old sigevent structure.
2712 	 */
2713 	CP(*osig, *nsig, sigev_notify);
2714 	switch (nsig->sigev_notify) {
2715 	case SIGEV_NONE:
2716 		break;
2717 	case SIGEV_SIGNAL:
2718 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2719 		break;
2720 	case SIGEV_KEVENT:
2721 		nsig->sigev_notify_kqueue =
2722 		    osig->__sigev_u.__sigev_notify_kqueue;
2723 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2724 		break;
2725 	default:
2726 		return (EINVAL);
2727 	}
2728 	return (0);
2729 }
2730 
2731 static int
2732 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2733 {
2734 	struct oaiocb32 job32;
2735 	int error;
2736 
2737 	bzero(kjob, sizeof(struct aiocb));
2738 	error = copyin(ujob, &job32, sizeof(job32));
2739 	if (error)
2740 		return (error);
2741 
2742 	CP(job32, *kjob, aio_fildes);
2743 	CP(job32, *kjob, aio_offset);
2744 	PTRIN_CP(job32, *kjob, aio_buf);
2745 	CP(job32, *kjob, aio_nbytes);
2746 	CP(job32, *kjob, aio_lio_opcode);
2747 	CP(job32, *kjob, aio_reqprio);
2748 	CP(job32, *kjob, _aiocb_private.status);
2749 	CP(job32, *kjob, _aiocb_private.error);
2750 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2751 	return (convert_old_sigevent32(&job32.aio_sigevent,
2752 	    &kjob->aio_sigevent));
2753 }
2754 
2755 static int
2756 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2757 {
2758 	struct aiocb32 job32;
2759 	int error;
2760 
2761 	error = copyin(ujob, &job32, sizeof(job32));
2762 	if (error)
2763 		return (error);
2764 	CP(job32, *kjob, aio_fildes);
2765 	CP(job32, *kjob, aio_offset);
2766 	PTRIN_CP(job32, *kjob, aio_buf);
2767 	CP(job32, *kjob, aio_nbytes);
2768 	CP(job32, *kjob, aio_lio_opcode);
2769 	CP(job32, *kjob, aio_reqprio);
2770 	CP(job32, *kjob, _aiocb_private.status);
2771 	CP(job32, *kjob, _aiocb_private.error);
2772 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2773 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2774 }
2775 
2776 static long
2777 aiocb32_fetch_status(struct aiocb *ujob)
2778 {
2779 	struct aiocb32 *ujob32;
2780 
2781 	ujob32 = (struct aiocb32 *)ujob;
2782 	return (fuword32(&ujob32->_aiocb_private.status));
2783 }
2784 
2785 static long
2786 aiocb32_fetch_error(struct aiocb *ujob)
2787 {
2788 	struct aiocb32 *ujob32;
2789 
2790 	ujob32 = (struct aiocb32 *)ujob;
2791 	return (fuword32(&ujob32->_aiocb_private.error));
2792 }
2793 
2794 static int
2795 aiocb32_store_status(struct aiocb *ujob, long status)
2796 {
2797 	struct aiocb32 *ujob32;
2798 
2799 	ujob32 = (struct aiocb32 *)ujob;
2800 	return (suword32(&ujob32->_aiocb_private.status, status));
2801 }
2802 
2803 static int
2804 aiocb32_store_error(struct aiocb *ujob, long error)
2805 {
2806 	struct aiocb32 *ujob32;
2807 
2808 	ujob32 = (struct aiocb32 *)ujob;
2809 	return (suword32(&ujob32->_aiocb_private.error, error));
2810 }
2811 
2812 static int
2813 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2814 {
2815 	struct aiocb32 *ujob32;
2816 
2817 	ujob32 = (struct aiocb32 *)ujob;
2818 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2819 }
2820 
2821 static int
2822 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2823 {
2824 
2825 	return (suword32(ujobp, (long)ujob));
2826 }
2827 
2828 static struct aiocb_ops aiocb32_ops = {
2829 	.copyin = aiocb32_copyin,
2830 	.fetch_status = aiocb32_fetch_status,
2831 	.fetch_error = aiocb32_fetch_error,
2832 	.store_status = aiocb32_store_status,
2833 	.store_error = aiocb32_store_error,
2834 	.store_kernelinfo = aiocb32_store_kernelinfo,
2835 	.store_aiocb = aiocb32_store_aiocb,
2836 };
2837 
2838 static struct aiocb_ops aiocb32_ops_osigevent = {
2839 	.copyin = aiocb32_copyin_old_sigevent,
2840 	.fetch_status = aiocb32_fetch_status,
2841 	.fetch_error = aiocb32_fetch_error,
2842 	.store_status = aiocb32_store_status,
2843 	.store_error = aiocb32_store_error,
2844 	.store_kernelinfo = aiocb32_store_kernelinfo,
2845 	.store_aiocb = aiocb32_store_aiocb,
2846 };
2847 
2848 int
2849 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2850 {
2851 
2852 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2853 }
2854 
2855 int
2856 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2857 {
2858 	struct timespec32 ts32;
2859 	struct timespec ts, *tsp;
2860 	struct aiocb **ujoblist;
2861 	uint32_t *ujoblist32;
2862 	int error, i;
2863 
2864 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2865 		return (EINVAL);
2866 
2867 	if (uap->timeout) {
2868 		/* Get timespec struct. */
2869 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2870 			return (error);
2871 		CP(ts32, ts, tv_sec);
2872 		CP(ts32, ts, tv_nsec);
2873 		tsp = &ts;
2874 	} else
2875 		tsp = NULL;
2876 
2877 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2878 	ujoblist32 = (uint32_t *)ujoblist;
2879 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2880 	    sizeof(ujoblist32[0]));
2881 	if (error == 0) {
2882 		for (i = uap->nent; i > 0; i--)
2883 			ujoblist[i] = PTRIN(ujoblist32[i]);
2884 
2885 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2886 	}
2887 	uma_zfree(aiol_zone, ujoblist);
2888 	return (error);
2889 }
2890 
2891 int
2892 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2893 {
2894 
2895 	return (sys_aio_cancel(td, (struct aio_cancel_args *)uap));
2896 }
2897 
2898 int
2899 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2900 {
2901 
2902 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2903 }
2904 
2905 int
2906 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2907 {
2908 
2909 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2910 	    &aiocb32_ops_osigevent));
2911 }
2912 
2913 int
2914 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2915 {
2916 
2917 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2918 	    &aiocb32_ops));
2919 }
2920 
2921 int
2922 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2923 {
2924 
2925 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2926 	    &aiocb32_ops_osigevent));
2927 }
2928 
2929 int
2930 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2931 {
2932 
2933 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2934 	    &aiocb32_ops));
2935 }
2936 
2937 int
2938 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2939 {
2940 
2941 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2942 	    &aiocb32_ops));
2943 }
2944 
2945 int
2946 freebsd32_aio_waitcomplete(struct thread *td,
2947     struct freebsd32_aio_waitcomplete_args *uap)
2948 {
2949 	struct timespec32 ts32;
2950 	struct timespec ts, *tsp;
2951 	int error;
2952 
2953 	if (uap->timeout) {
2954 		/* Get timespec struct. */
2955 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2956 		if (error)
2957 			return (error);
2958 		CP(ts32, ts, tv_sec);
2959 		CP(ts32, ts, tv_nsec);
2960 		tsp = &ts;
2961 	} else
2962 		tsp = NULL;
2963 
2964 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2965 	    &aiocb32_ops));
2966 }
2967 
2968 int
2969 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2970 {
2971 
2972 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2973 	    &aiocb32_ops));
2974 }
2975 
2976 int
2977 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2978 {
2979 	struct aiocb **acb_list;
2980 	struct sigevent *sigp, sig;
2981 	struct osigevent32 osig;
2982 	uint32_t *acb_list32;
2983 	int error, i, nent;
2984 
2985 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2986 		return (EINVAL);
2987 
2988 	nent = uap->nent;
2989 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2990 		return (EINVAL);
2991 
2992 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2993 		error = copyin(uap->sig, &osig, sizeof(osig));
2994 		if (error)
2995 			return (error);
2996 		error = convert_old_sigevent32(&osig, &sig);
2997 		if (error)
2998 			return (error);
2999 		sigp = &sig;
3000 	} else
3001 		sigp = NULL;
3002 
3003 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3004 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3005 	if (error) {
3006 		free(acb_list32, M_LIO);
3007 		return (error);
3008 	}
3009 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3010 	for (i = 0; i < nent; i++)
3011 		acb_list[i] = PTRIN(acb_list32[i]);
3012 	free(acb_list32, M_LIO);
3013 
3014 	error = kern_lio_listio(td, uap->mode,
3015 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3016 	    &aiocb32_ops_osigevent);
3017 	free(acb_list, M_LIO);
3018 	return (error);
3019 }
3020 
3021 int
3022 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3023 {
3024 	struct aiocb **acb_list;
3025 	struct sigevent *sigp, sig;
3026 	struct sigevent32 sig32;
3027 	uint32_t *acb_list32;
3028 	int error, i, nent;
3029 
3030 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3031 		return (EINVAL);
3032 
3033 	nent = uap->nent;
3034 	if (nent < 0 || nent > AIO_LISTIO_MAX)
3035 		return (EINVAL);
3036 
3037 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3038 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3039 		if (error)
3040 			return (error);
3041 		error = convert_sigevent32(&sig32, &sig);
3042 		if (error)
3043 			return (error);
3044 		sigp = &sig;
3045 	} else
3046 		sigp = NULL;
3047 
3048 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3049 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3050 	if (error) {
3051 		free(acb_list32, M_LIO);
3052 		return (error);
3053 	}
3054 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3055 	for (i = 0; i < nent; i++)
3056 		acb_list[i] = PTRIN(acb_list32[i]);
3057 	free(acb_list32, M_LIO);
3058 
3059 	error = kern_lio_listio(td, uap->mode,
3060 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3061 	    &aiocb32_ops);
3062 	free(acb_list, M_LIO);
3063 	return (error);
3064 }
3065 
3066 #endif
3067