xref: /freebsd/sys/kern/vfs_aio.c (revision a0ee8cc636cd5c2374ec44ca71226564ea0bca95)
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/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
39 #include <sys/file.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
45 #include <sys/proc.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/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 #include <geom/geom.h>
63 
64 #include <machine/atomic.h>
65 
66 #include <vm/vm.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_extern.h>
69 #include <vm/pmap.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_object.h>
72 #include <vm/uma.h>
73 #include <sys/aio.h>
74 
75 #include "opt_vfs_aio.h"
76 
77 /*
78  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
79  * overflow. (XXX will be removed soon.)
80  */
81 static u_long jobrefid;
82 
83 /*
84  * Counter for aio_fsync.
85  */
86 static uint64_t jobseqno;
87 
88 #define JOBST_NULL		0
89 #define JOBST_JOBQSOCK		1
90 #define JOBST_JOBQGLOBAL	2
91 #define JOBST_JOBRUNNING	3
92 #define JOBST_JOBFINISHED	4
93 #define JOBST_JOBQBUF		5
94 #define JOBST_JOBQSYNC		6
95 
96 #ifndef MAX_AIO_PER_PROC
97 #define MAX_AIO_PER_PROC	32
98 #endif
99 
100 #ifndef MAX_AIO_QUEUE_PER_PROC
101 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
102 #endif
103 
104 #ifndef MAX_AIO_PROCS
105 #define MAX_AIO_PROCS		32
106 #endif
107 
108 #ifndef MAX_AIO_QUEUE
109 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
110 #endif
111 
112 #ifndef TARGET_AIO_PROCS
113 #define TARGET_AIO_PROCS	4
114 #endif
115 
116 #ifndef MAX_BUF_AIO
117 #define MAX_BUF_AIO		16
118 #endif
119 
120 #ifndef AIOD_LIFETIME_DEFAULT
121 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
122 #endif
123 
124 FEATURE(aio, "Asynchronous I/O");
125 
126 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
127 
128 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
129 
130 static int max_aio_procs = MAX_AIO_PROCS;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
132 	CTLFLAG_RW, &max_aio_procs, 0,
133 	"Maximum number of kernel threads to use for handling async IO ");
134 
135 static int num_aio_procs = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
137 	CTLFLAG_RD, &num_aio_procs, 0,
138 	"Number of presently active kernel threads for async IO");
139 
140 /*
141  * The code will adjust the actual number of AIO processes towards this
142  * number when it gets a chance.
143  */
144 static int target_aio_procs = TARGET_AIO_PROCS;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
146 	0, "Preferred number of ready kernel threads for async IO");
147 
148 static int max_queue_count = MAX_AIO_QUEUE;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
150     "Maximum number of aio requests to queue, globally");
151 
152 static int num_queue_count = 0;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
154     "Number of queued aio requests");
155 
156 static int num_buf_aio = 0;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
158     "Number of aio requests presently handled by the buf subsystem");
159 
160 /* Number of async I/O thread in the process of being started */
161 /* XXX This should be local to aio_aqueue() */
162 static int num_aio_resv_start = 0;
163 
164 static int aiod_lifetime;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
166     "Maximum lifetime for idle aiod");
167 
168 static int unloadable = 0;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
170     "Allow unload of aio (not recommended)");
171 
172 
173 static int max_aio_per_proc = MAX_AIO_PER_PROC;
174 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
175     0, "Maximum active aio requests per process (stored in the process)");
176 
177 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
178 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
179     &max_aio_queue_per_proc, 0,
180     "Maximum queued aio requests per process (stored in the process)");
181 
182 static int max_buf_aio = MAX_BUF_AIO;
183 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
184     "Maximum buf aio requests per process (stored in the process)");
185 
186 typedef struct oaiocb {
187 	int	aio_fildes;		/* File descriptor */
188 	off_t	aio_offset;		/* File offset for I/O */
189 	volatile void *aio_buf;         /* I/O buffer in process space */
190 	size_t	aio_nbytes;		/* Number of bytes for I/O */
191 	struct	osigevent aio_sigevent;	/* Signal to deliver */
192 	int	aio_lio_opcode;		/* LIO opcode */
193 	int	aio_reqprio;		/* Request priority -- ignored */
194 	struct	__aiocb_private	_aiocb_private;
195 } oaiocb_t;
196 
197 /*
198  * Below is a key of locks used to protect each member of struct aiocblist
199  * aioliojob and kaioinfo and any backends.
200  *
201  * * - need not protected
202  * a - locked by kaioinfo lock
203  * b - locked by backend lock, the backend lock can be null in some cases,
204  *     for example, BIO belongs to this type, in this case, proc lock is
205  *     reused.
206  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
207  */
208 
209 /*
210  * Current, there is only two backends: BIO and generic file I/O.
211  * socket I/O is served by generic file I/O, this is not a good idea, since
212  * disk file I/O and any other types without O_NONBLOCK flag can block daemon
213  * threads, if there is no thread to serve socket I/O, the socket I/O will be
214  * delayed too long or starved, we should create some threads dedicated to
215  * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
216  * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
217  * structure is not safe because there is race between userland and aio
218  * daemons.
219  */
220 
221 struct aiocblist {
222 	TAILQ_ENTRY(aiocblist) list;	/* (b) internal list of for backend */
223 	TAILQ_ENTRY(aiocblist) plist;	/* (a) list of jobs for each backend */
224 	TAILQ_ENTRY(aiocblist) allist;  /* (a) list of all jobs in proc */
225 	int	jobflags;		/* (a) job flags */
226 	int	jobstate;		/* (b) job state */
227 	int	inputcharge;		/* (*) input blockes */
228 	int	outputcharge;		/* (*) output blockes */
229 	struct	bio *bp;		/* (*) BIO backend BIO pointer */
230 	struct	buf *pbuf;		/* (*) BIO backend buffer pointer */
231 	struct	vm_page *pages[btoc(MAXPHYS)+1]; /* BIO backend pages */
232 	int	npages;			/* BIO backend number of pages */
233 	struct	proc *userproc;		/* (*) user process */
234 	struct  ucred *cred;		/* (*) active credential when created */
235 	struct	file *fd_file;		/* (*) pointer to file structure */
236 	struct	aioliojob *lio;		/* (*) optional lio job */
237 	struct	aiocb *uuaiocb;		/* (*) pointer in userspace of aiocb */
238 	struct	knlist klist;		/* (a) list of knotes */
239 	struct	aiocb uaiocb;		/* (*) kernel I/O control block */
240 	ksiginfo_t ksi;			/* (a) realtime signal info */
241 	uint64_t seqno;			/* (*) job number */
242 	int	pending;		/* (a) number of pending I/O, aio_fsync only */
243 };
244 
245 /* jobflags */
246 #define AIOCBLIST_DONE		0x01
247 #define AIOCBLIST_BUFDONE	0x02
248 #define AIOCBLIST_RUNDOWN	0x04
249 #define AIOCBLIST_CHECKSYNC	0x08
250 
251 /*
252  * AIO process info
253  */
254 #define AIOP_FREE	0x1			/* proc on free queue */
255 
256 struct aiothreadlist {
257 	int aiothreadflags;			/* (c) AIO proc flags */
258 	TAILQ_ENTRY(aiothreadlist) list;	/* (c) list of processes */
259 	struct thread *aiothread;		/* (*) the AIO thread */
260 };
261 
262 /*
263  * data-structure for lio signal management
264  */
265 struct aioliojob {
266 	int	lioj_flags;			/* (a) listio flags */
267 	int	lioj_count;			/* (a) listio flags */
268 	int	lioj_finished_count;		/* (a) listio flags */
269 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
270 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
271 	struct  knlist klist;			/* (a) list of knotes */
272 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
273 };
274 
275 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
276 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
277 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
278 
279 /*
280  * per process aio data structure
281  */
282 struct kaioinfo {
283 	struct mtx	kaio_mtx;	/* the lock to protect this struct */
284 	int	kaio_flags;		/* (a) per process kaio flags */
285 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
286 	int	kaio_active_count;	/* (c) number of currently used AIOs */
287 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
288 	int	kaio_count;		/* (a) size of AIO queue */
289 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
290 	int	kaio_buffer_count;	/* (a) number of physio buffers */
291 	TAILQ_HEAD(,aiocblist) kaio_all;	/* (a) all AIOs in the process */
292 	TAILQ_HEAD(,aiocblist) kaio_done;	/* (a) done queue for process */
293 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
294 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* (a) job queue for process */
295 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* (a) buffer job queue for process */
296 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;  /* (a) queue for aios waiting on sockets,
297 						 *  NOT USED YET.
298 						 */
299 	TAILQ_HEAD(,aiocblist) kaio_syncqueue;	/* (a) queue for aio_fsync */
300 	struct	task	kaio_task;	/* (*) task to kick aio threads */
301 };
302 
303 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
304 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
305 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
306 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
307 
308 #define KAIO_RUNDOWN	0x1	/* process is being run down */
309 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
310 
311 /*
312  * Operations used to interact with userland aio control blocks.
313  * Different ABIs provide their own operations.
314  */
315 struct aiocb_ops {
316 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
317 	long	(*fetch_status)(struct aiocb *ujob);
318 	long	(*fetch_error)(struct aiocb *ujob);
319 	int	(*store_status)(struct aiocb *ujob, long status);
320 	int	(*store_error)(struct aiocb *ujob, long error);
321 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
322 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
323 };
324 
325 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* (c) Idle daemons */
326 static struct sema aio_newproc_sem;
327 static struct mtx aio_job_mtx;
328 static struct mtx aio_sock_mtx;
329 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* (c) Async job list */
330 static struct unrhdr *aiod_unr;
331 
332 void		aio_init_aioinfo(struct proc *p);
333 static int	aio_onceonly(void);
334 static int	aio_free_entry(struct aiocblist *aiocbe);
335 static void	aio_process_rw(struct aiocblist *aiocbe);
336 static void	aio_process_sync(struct aiocblist *aiocbe);
337 static void	aio_process_mlock(struct aiocblist *aiocbe);
338 static int	aio_newproc(int *);
339 int		aio_aqueue(struct thread *td, struct aiocb *job,
340 			struct aioliojob *lio, int type, struct aiocb_ops *ops);
341 static void	aio_physwakeup(struct bio *bp);
342 static void	aio_proc_rundown(void *arg, struct proc *p);
343 static void	aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
344 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
345 static void	aio_daemon(void *param);
346 static void	aio_swake_cb(struct socket *, struct sockbuf *);
347 static int	aio_unload(void);
348 static void	aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
349 #define DONE_BUF	1
350 #define DONE_QUEUE	2
351 static int	aio_kick(struct proc *userp);
352 static void	aio_kick_nowait(struct proc *userp);
353 static void	aio_kick_helper(void *context, int pending);
354 static int	filt_aioattach(struct knote *kn);
355 static void	filt_aiodetach(struct knote *kn);
356 static int	filt_aio(struct knote *kn, long hint);
357 static int	filt_lioattach(struct knote *kn);
358 static void	filt_liodetach(struct knote *kn);
359 static int	filt_lio(struct knote *kn, long hint);
360 
361 /*
362  * Zones for:
363  * 	kaio	Per process async io info
364  *	aiop	async io thread data
365  *	aiocb	async io jobs
366  *	aiol	list io job pointer - internal to aio_suspend XXX
367  *	aiolio	list io jobs
368  */
369 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
370 
371 /* kqueue filters for aio */
372 static struct filterops aio_filtops = {
373 	.f_isfd = 0,
374 	.f_attach = filt_aioattach,
375 	.f_detach = filt_aiodetach,
376 	.f_event = filt_aio,
377 };
378 static struct filterops lio_filtops = {
379 	.f_isfd = 0,
380 	.f_attach = filt_lioattach,
381 	.f_detach = filt_liodetach,
382 	.f_event = filt_lio
383 };
384 
385 static eventhandler_tag exit_tag, exec_tag;
386 
387 TASKQUEUE_DEFINE_THREAD(aiod_kick);
388 
389 /*
390  * Main operations function for use as a kernel module.
391  */
392 static int
393 aio_modload(struct module *module, int cmd, void *arg)
394 {
395 	int error = 0;
396 
397 	switch (cmd) {
398 	case MOD_LOAD:
399 		aio_onceonly();
400 		break;
401 	case MOD_UNLOAD:
402 		error = aio_unload();
403 		break;
404 	case MOD_SHUTDOWN:
405 		break;
406 	default:
407 		error = EINVAL;
408 		break;
409 	}
410 	return (error);
411 }
412 
413 static moduledata_t aio_mod = {
414 	"aio",
415 	&aio_modload,
416 	NULL
417 };
418 
419 static struct syscall_helper_data aio_syscalls[] = {
420 	SYSCALL_INIT_HELPER(aio_cancel),
421 	SYSCALL_INIT_HELPER(aio_error),
422 	SYSCALL_INIT_HELPER(aio_fsync),
423 	SYSCALL_INIT_HELPER(aio_mlock),
424 	SYSCALL_INIT_HELPER(aio_read),
425 	SYSCALL_INIT_HELPER(aio_return),
426 	SYSCALL_INIT_HELPER(aio_suspend),
427 	SYSCALL_INIT_HELPER(aio_waitcomplete),
428 	SYSCALL_INIT_HELPER(aio_write),
429 	SYSCALL_INIT_HELPER(lio_listio),
430 	SYSCALL_INIT_HELPER(oaio_read),
431 	SYSCALL_INIT_HELPER(oaio_write),
432 	SYSCALL_INIT_HELPER(olio_listio),
433 	SYSCALL_INIT_LAST
434 };
435 
436 #ifdef COMPAT_FREEBSD32
437 #include <sys/mount.h>
438 #include <sys/socket.h>
439 #include <compat/freebsd32/freebsd32.h>
440 #include <compat/freebsd32/freebsd32_proto.h>
441 #include <compat/freebsd32/freebsd32_signal.h>
442 #include <compat/freebsd32/freebsd32_syscall.h>
443 #include <compat/freebsd32/freebsd32_util.h>
444 
445 static struct syscall_helper_data aio32_syscalls[] = {
446 	SYSCALL32_INIT_HELPER(freebsd32_aio_return),
447 	SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
448 	SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
449 	SYSCALL32_INIT_HELPER(freebsd32_aio_error),
450 	SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
451 	SYSCALL32_INIT_HELPER(freebsd32_aio_mlock),
452 	SYSCALL32_INIT_HELPER(freebsd32_aio_read),
453 	SYSCALL32_INIT_HELPER(freebsd32_aio_write),
454 	SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
455 	SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
456 	SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
457 	SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
458 	SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
459 	SYSCALL_INIT_LAST
460 };
461 #endif
462 
463 DECLARE_MODULE(aio, aio_mod,
464 	SI_SUB_VFS, SI_ORDER_ANY);
465 MODULE_VERSION(aio, 1);
466 
467 /*
468  * Startup initialization
469  */
470 static int
471 aio_onceonly(void)
472 {
473 	int error;
474 
475 	/* XXX: should probably just use so->callback */
476 	aio_swake = &aio_swake_cb;
477 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
478 	    EVENTHANDLER_PRI_ANY);
479 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
480 	    EVENTHANDLER_PRI_ANY);
481 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
482 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
483 	TAILQ_INIT(&aio_freeproc);
484 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
485 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
486 	mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
487 	TAILQ_INIT(&aio_jobs);
488 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
489 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
490 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
491 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
492 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
493 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
494 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
495 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
496 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
497 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
498 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
499 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
500 	jobrefid = 1;
501 	async_io_version = _POSIX_VERSION;
502 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
503 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
504 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
505 
506 	error = syscall_helper_register(aio_syscalls, SY_THR_STATIC_KLD);
507 	if (error)
508 		return (error);
509 #ifdef COMPAT_FREEBSD32
510 	error = syscall32_helper_register(aio32_syscalls, SY_THR_STATIC_KLD);
511 	if (error)
512 		return (error);
513 #endif
514 	return (0);
515 }
516 
517 /*
518  * Callback for unload of AIO when used as a module.
519  */
520 static int
521 aio_unload(void)
522 {
523 	int error;
524 
525 	/*
526 	 * XXX: no unloads by default, it's too dangerous.
527 	 * perhaps we could do it if locked out callers and then
528 	 * did an aio_proc_rundown() on each process.
529 	 *
530 	 * jhb: aio_proc_rundown() needs to run on curproc though,
531 	 * so I don't think that would fly.
532 	 */
533 	if (!unloadable)
534 		return (EOPNOTSUPP);
535 
536 #ifdef COMPAT_FREEBSD32
537 	syscall32_helper_unregister(aio32_syscalls);
538 #endif
539 	syscall_helper_unregister(aio_syscalls);
540 
541 	error = kqueue_del_filteropts(EVFILT_AIO);
542 	if (error)
543 		return error;
544 	error = kqueue_del_filteropts(EVFILT_LIO);
545 	if (error)
546 		return error;
547 	async_io_version = 0;
548 	aio_swake = NULL;
549 	taskqueue_free(taskqueue_aiod_kick);
550 	delete_unrhdr(aiod_unr);
551 	uma_zdestroy(kaio_zone);
552 	uma_zdestroy(aiop_zone);
553 	uma_zdestroy(aiocb_zone);
554 	uma_zdestroy(aiol_zone);
555 	uma_zdestroy(aiolio_zone);
556 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
557 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
558 	mtx_destroy(&aio_job_mtx);
559 	mtx_destroy(&aio_sock_mtx);
560 	sema_destroy(&aio_newproc_sem);
561 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
562 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
563 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
564 	return (0);
565 }
566 
567 /*
568  * Init the per-process aioinfo structure.  The aioinfo limits are set
569  * per-process for user limit (resource) management.
570  */
571 void
572 aio_init_aioinfo(struct proc *p)
573 {
574 	struct kaioinfo *ki;
575 
576 	ki = uma_zalloc(kaio_zone, M_WAITOK);
577 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
578 	ki->kaio_flags = 0;
579 	ki->kaio_maxactive_count = max_aio_per_proc;
580 	ki->kaio_active_count = 0;
581 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
582 	ki->kaio_count = 0;
583 	ki->kaio_ballowed_count = max_buf_aio;
584 	ki->kaio_buffer_count = 0;
585 	TAILQ_INIT(&ki->kaio_all);
586 	TAILQ_INIT(&ki->kaio_done);
587 	TAILQ_INIT(&ki->kaio_jobqueue);
588 	TAILQ_INIT(&ki->kaio_bufqueue);
589 	TAILQ_INIT(&ki->kaio_liojoblist);
590 	TAILQ_INIT(&ki->kaio_sockqueue);
591 	TAILQ_INIT(&ki->kaio_syncqueue);
592 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
593 	PROC_LOCK(p);
594 	if (p->p_aioinfo == NULL) {
595 		p->p_aioinfo = ki;
596 		PROC_UNLOCK(p);
597 	} else {
598 		PROC_UNLOCK(p);
599 		mtx_destroy(&ki->kaio_mtx);
600 		uma_zfree(kaio_zone, ki);
601 	}
602 
603 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
604 		aio_newproc(NULL);
605 }
606 
607 static int
608 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
609 {
610 	struct thread *td;
611 	int error;
612 
613 	error = sigev_findtd(p, sigev, &td);
614 	if (error)
615 		return (error);
616 	if (!KSI_ONQ(ksi)) {
617 		ksiginfo_set_sigev(ksi, sigev);
618 		ksi->ksi_code = SI_ASYNCIO;
619 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
620 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
621 	}
622 	PROC_UNLOCK(p);
623 	return (error);
624 }
625 
626 /*
627  * Free a job entry.  Wait for completion if it is currently active, but don't
628  * delay forever.  If we delay, we return a flag that says that we have to
629  * restart the queue scan.
630  */
631 static int
632 aio_free_entry(struct aiocblist *aiocbe)
633 {
634 	struct kaioinfo *ki;
635 	struct aioliojob *lj;
636 	struct proc *p;
637 
638 	p = aiocbe->userproc;
639 	MPASS(curproc == p);
640 	ki = p->p_aioinfo;
641 	MPASS(ki != NULL);
642 
643 	AIO_LOCK_ASSERT(ki, MA_OWNED);
644 	MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
645 
646 	atomic_subtract_int(&num_queue_count, 1);
647 
648 	ki->kaio_count--;
649 	MPASS(ki->kaio_count >= 0);
650 
651 	TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
652 	TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
653 
654 	lj = aiocbe->lio;
655 	if (lj) {
656 		lj->lioj_count--;
657 		lj->lioj_finished_count--;
658 
659 		if (lj->lioj_count == 0) {
660 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
661 			/* lio is going away, we need to destroy any knotes */
662 			knlist_delete(&lj->klist, curthread, 1);
663 			PROC_LOCK(p);
664 			sigqueue_take(&lj->lioj_ksi);
665 			PROC_UNLOCK(p);
666 			uma_zfree(aiolio_zone, lj);
667 		}
668 	}
669 
670 	/* aiocbe is going away, we need to destroy any knotes */
671 	knlist_delete(&aiocbe->klist, curthread, 1);
672 	PROC_LOCK(p);
673 	sigqueue_take(&aiocbe->ksi);
674 	PROC_UNLOCK(p);
675 
676 	MPASS(aiocbe->bp == NULL);
677 	aiocbe->jobstate = JOBST_NULL;
678 	AIO_UNLOCK(ki);
679 
680 	/*
681 	 * The thread argument here is used to find the owning process
682 	 * and is also passed to fo_close() which may pass it to various
683 	 * places such as devsw close() routines.  Because of that, we
684 	 * need a thread pointer from the process owning the job that is
685 	 * persistent and won't disappear out from under us or move to
686 	 * another process.
687 	 *
688 	 * Currently, all the callers of this function call it to remove
689 	 * an aiocblist from the current process' job list either via a
690 	 * syscall or due to the current process calling exit() or
691 	 * execve().  Thus, we know that p == curproc.  We also know that
692 	 * curthread can't exit since we are curthread.
693 	 *
694 	 * Therefore, we use curthread as the thread to pass to
695 	 * knlist_delete().  This does mean that it is possible for the
696 	 * thread pointer at close time to differ from the thread pointer
697 	 * at open time, but this is already true of file descriptors in
698 	 * a multithreaded process.
699 	 */
700 	if (aiocbe->fd_file)
701 		fdrop(aiocbe->fd_file, curthread);
702 	crfree(aiocbe->cred);
703 	uma_zfree(aiocb_zone, aiocbe);
704 	AIO_LOCK(ki);
705 
706 	return (0);
707 }
708 
709 static void
710 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
711 {
712    	aio_proc_rundown(arg, p);
713 }
714 
715 /*
716  * Rundown the jobs for a given process.
717  */
718 static void
719 aio_proc_rundown(void *arg, struct proc *p)
720 {
721 	struct kaioinfo *ki;
722 	struct aioliojob *lj;
723 	struct aiocblist *cbe, *cbn;
724 	struct file *fp;
725 	struct socket *so;
726 	int remove;
727 
728 	KASSERT(curthread->td_proc == p,
729 	    ("%s: called on non-curproc", __func__));
730 	ki = p->p_aioinfo;
731 	if (ki == NULL)
732 		return;
733 
734 	AIO_LOCK(ki);
735 	ki->kaio_flags |= KAIO_RUNDOWN;
736 
737 restart:
738 
739 	/*
740 	 * Try to cancel all pending requests. This code simulates
741 	 * aio_cancel on all pending I/O requests.
742 	 */
743 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
744 		remove = 0;
745 		mtx_lock(&aio_job_mtx);
746 		if (cbe->jobstate == JOBST_JOBQGLOBAL) {
747 			TAILQ_REMOVE(&aio_jobs, cbe, list);
748 			remove = 1;
749 		} else if (cbe->jobstate == JOBST_JOBQSOCK) {
750 			fp = cbe->fd_file;
751 			MPASS(fp->f_type == DTYPE_SOCKET);
752 			so = fp->f_data;
753 			TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
754 			remove = 1;
755 		} else if (cbe->jobstate == JOBST_JOBQSYNC) {
756 			TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
757 			remove = 1;
758 		}
759 		mtx_unlock(&aio_job_mtx);
760 
761 		if (remove) {
762 			cbe->jobstate = JOBST_JOBFINISHED;
763 			cbe->uaiocb._aiocb_private.status = -1;
764 			cbe->uaiocb._aiocb_private.error = ECANCELED;
765 			TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
766 			aio_bio_done_notify(p, cbe, DONE_QUEUE);
767 		}
768 	}
769 
770 	/* Wait for all running I/O to be finished */
771 	if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
772 	    TAILQ_FIRST(&ki->kaio_jobqueue)) {
773 		ki->kaio_flags |= KAIO_WAKEUP;
774 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
775 		goto restart;
776 	}
777 
778 	/* Free all completed I/O requests. */
779 	while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
780 		aio_free_entry(cbe);
781 
782 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
783 		if (lj->lioj_count == 0) {
784 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
785 			knlist_delete(&lj->klist, curthread, 1);
786 			PROC_LOCK(p);
787 			sigqueue_take(&lj->lioj_ksi);
788 			PROC_UNLOCK(p);
789 			uma_zfree(aiolio_zone, lj);
790 		} else {
791 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
792 			    lj->lioj_count, lj->lioj_finished_count);
793 		}
794 	}
795 	AIO_UNLOCK(ki);
796 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
797 	mtx_destroy(&ki->kaio_mtx);
798 	uma_zfree(kaio_zone, ki);
799 	p->p_aioinfo = NULL;
800 }
801 
802 /*
803  * Select a job to run (called by an AIO daemon).
804  */
805 static struct aiocblist *
806 aio_selectjob(struct aiothreadlist *aiop)
807 {
808 	struct aiocblist *aiocbe;
809 	struct kaioinfo *ki;
810 	struct proc *userp;
811 
812 	mtx_assert(&aio_job_mtx, MA_OWNED);
813 	TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
814 		userp = aiocbe->userproc;
815 		ki = userp->p_aioinfo;
816 
817 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
818 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
819 			/* Account for currently active jobs. */
820 			ki->kaio_active_count++;
821 			aiocbe->jobstate = JOBST_JOBRUNNING;
822 			break;
823 		}
824 	}
825 	return (aiocbe);
826 }
827 
828 /*
829  *  Move all data to a permanent storage device, this code
830  *  simulates fsync syscall.
831  */
832 static int
833 aio_fsync_vnode(struct thread *td, struct vnode *vp)
834 {
835 	struct mount *mp;
836 	int error;
837 
838 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
839 		goto drop;
840 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
841 	if (vp->v_object != NULL) {
842 		VM_OBJECT_WLOCK(vp->v_object);
843 		vm_object_page_clean(vp->v_object, 0, 0, 0);
844 		VM_OBJECT_WUNLOCK(vp->v_object);
845 	}
846 	error = VOP_FSYNC(vp, MNT_WAIT, td);
847 
848 	VOP_UNLOCK(vp, 0);
849 	vn_finished_write(mp);
850 drop:
851 	return (error);
852 }
853 
854 /*
855  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
856  * does the I/O request for the non-physio version of the operations.  The
857  * normal vn operations are used, and this code should work in all instances
858  * for every type of file, including pipes, sockets, fifos, and regular files.
859  *
860  * XXX I don't think it works well for socket, pipe, and fifo.
861  */
862 static void
863 aio_process_rw(struct aiocblist *aiocbe)
864 {
865 	struct ucred *td_savedcred;
866 	struct thread *td;
867 	struct aiocb *cb;
868 	struct file *fp;
869 	struct socket *so;
870 	struct uio auio;
871 	struct iovec aiov;
872 	int cnt;
873 	int error;
874 	int oublock_st, oublock_end;
875 	int inblock_st, inblock_end;
876 
877 	KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_READ ||
878 	    aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE,
879 	    ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
880 
881 	td = curthread;
882 	td_savedcred = td->td_ucred;
883 	td->td_ucred = aiocbe->cred;
884 	cb = &aiocbe->uaiocb;
885 	fp = aiocbe->fd_file;
886 
887 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
888 	aiov.iov_len = cb->aio_nbytes;
889 
890 	auio.uio_iov = &aiov;
891 	auio.uio_iovcnt = 1;
892 	auio.uio_offset = cb->aio_offset;
893 	auio.uio_resid = cb->aio_nbytes;
894 	cnt = cb->aio_nbytes;
895 	auio.uio_segflg = UIO_USERSPACE;
896 	auio.uio_td = td;
897 
898 	inblock_st = td->td_ru.ru_inblock;
899 	oublock_st = td->td_ru.ru_oublock;
900 	/*
901 	 * aio_aqueue() acquires a reference to the file that is
902 	 * released in aio_free_entry().
903 	 */
904 	if (cb->aio_lio_opcode == LIO_READ) {
905 		auio.uio_rw = UIO_READ;
906 		if (auio.uio_resid == 0)
907 			error = 0;
908 		else
909 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
910 	} else {
911 		if (fp->f_type == DTYPE_VNODE)
912 			bwillwrite();
913 		auio.uio_rw = UIO_WRITE;
914 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
915 	}
916 	inblock_end = td->td_ru.ru_inblock;
917 	oublock_end = td->td_ru.ru_oublock;
918 
919 	aiocbe->inputcharge = inblock_end - inblock_st;
920 	aiocbe->outputcharge = oublock_end - oublock_st;
921 
922 	if ((error) && (auio.uio_resid != cnt)) {
923 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
924 			error = 0;
925 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
926 			int sigpipe = 1;
927 			if (fp->f_type == DTYPE_SOCKET) {
928 				so = fp->f_data;
929 				if (so->so_options & SO_NOSIGPIPE)
930 					sigpipe = 0;
931 			}
932 			if (sigpipe) {
933 				PROC_LOCK(aiocbe->userproc);
934 				kern_psignal(aiocbe->userproc, SIGPIPE);
935 				PROC_UNLOCK(aiocbe->userproc);
936 			}
937 		}
938 	}
939 
940 	cnt -= auio.uio_resid;
941 	cb->_aiocb_private.error = error;
942 	cb->_aiocb_private.status = cnt;
943 	td->td_ucred = td_savedcred;
944 }
945 
946 static void
947 aio_process_sync(struct aiocblist *aiocbe)
948 {
949 	struct thread *td = curthread;
950 	struct ucred *td_savedcred = td->td_ucred;
951 	struct aiocb *cb = &aiocbe->uaiocb;
952 	struct file *fp = aiocbe->fd_file;
953 	int error = 0;
954 
955 	KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_SYNC,
956 	    ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
957 
958 	td->td_ucred = aiocbe->cred;
959 	if (fp->f_vnode != NULL)
960 		error = aio_fsync_vnode(td, fp->f_vnode);
961 	cb->_aiocb_private.error = error;
962 	cb->_aiocb_private.status = 0;
963 	td->td_ucred = td_savedcred;
964 }
965 
966 static void
967 aio_process_mlock(struct aiocblist *aiocbe)
968 {
969 	struct aiocb *cb = &aiocbe->uaiocb;
970 	int error;
971 
972 	KASSERT(aiocbe->uaiocb.aio_lio_opcode == LIO_MLOCK,
973 	    ("%s: opcode %d", __func__, aiocbe->uaiocb.aio_lio_opcode));
974 
975 	error = vm_mlock(aiocbe->userproc, aiocbe->cred,
976 	    __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
977 	cb->_aiocb_private.error = error;
978 	cb->_aiocb_private.status = 0;
979 }
980 
981 static void
982 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
983 {
984 	struct aioliojob *lj;
985 	struct kaioinfo *ki;
986 	struct aiocblist *scb, *scbn;
987 	int lj_done;
988 
989 	ki = userp->p_aioinfo;
990 	AIO_LOCK_ASSERT(ki, MA_OWNED);
991 	lj = aiocbe->lio;
992 	lj_done = 0;
993 	if (lj) {
994 		lj->lioj_finished_count++;
995 		if (lj->lioj_count == lj->lioj_finished_count)
996 			lj_done = 1;
997 	}
998 	if (type == DONE_QUEUE) {
999 		aiocbe->jobflags |= AIOCBLIST_DONE;
1000 	} else {
1001 		aiocbe->jobflags |= AIOCBLIST_BUFDONE;
1002 	}
1003 	TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
1004 	aiocbe->jobstate = JOBST_JOBFINISHED;
1005 
1006 	if (ki->kaio_flags & KAIO_RUNDOWN)
1007 		goto notification_done;
1008 
1009 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1010 	    aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
1011 		aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
1012 
1013 	KNOTE_LOCKED(&aiocbe->klist, 1);
1014 
1015 	if (lj_done) {
1016 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
1017 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
1018 			KNOTE_LOCKED(&lj->klist, 1);
1019 		}
1020 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
1021 		    == LIOJ_SIGNAL
1022 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
1023 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
1024 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
1025 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1026 		}
1027 	}
1028 
1029 notification_done:
1030 	if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1031 		TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1032 			if (aiocbe->fd_file == scb->fd_file &&
1033 			    aiocbe->seqno < scb->seqno) {
1034 				if (--scb->pending == 0) {
1035 					mtx_lock(&aio_job_mtx);
1036 					scb->jobstate = JOBST_JOBQGLOBAL;
1037 					TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1038 					TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1039 					aio_kick_nowait(userp);
1040 					mtx_unlock(&aio_job_mtx);
1041 				}
1042 			}
1043 		}
1044 	}
1045 	if (ki->kaio_flags & KAIO_WAKEUP) {
1046 		ki->kaio_flags &= ~KAIO_WAKEUP;
1047 		wakeup(&userp->p_aioinfo);
1048 	}
1049 }
1050 
1051 /*
1052  * The AIO daemon, most of the actual work is done in aio_process_*,
1053  * but the setup (and address space mgmt) is done in this routine.
1054  */
1055 static void
1056 aio_daemon(void *_id)
1057 {
1058 	struct aiocblist *aiocbe;
1059 	struct aiothreadlist *aiop;
1060 	struct kaioinfo *ki;
1061 	struct proc *curcp, *mycp, *userp;
1062 	struct vmspace *myvm, *tmpvm;
1063 	struct thread *td = curthread;
1064 	int id = (intptr_t)_id;
1065 
1066 	/*
1067 	 * Local copies of curproc (cp) and vmspace (myvm)
1068 	 */
1069 	mycp = td->td_proc;
1070 	myvm = mycp->p_vmspace;
1071 
1072 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1073 
1074 	/*
1075 	 * Allocate and ready the aio control info.  There is one aiop structure
1076 	 * per daemon.
1077 	 */
1078 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1079 	aiop->aiothread = td;
1080 	aiop->aiothreadflags = 0;
1081 
1082 	/* The daemon resides in its own pgrp. */
1083 	sys_setsid(td, NULL);
1084 
1085 	/*
1086 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1087 	 * and creating too many daemons.)
1088 	 */
1089 	sema_post(&aio_newproc_sem);
1090 
1091 	mtx_lock(&aio_job_mtx);
1092 	for (;;) {
1093 		/*
1094 		 * curcp is the current daemon process context.
1095 		 * userp is the current user process context.
1096 		 */
1097 		curcp = mycp;
1098 
1099 		/*
1100 		 * Take daemon off of free queue
1101 		 */
1102 		if (aiop->aiothreadflags & AIOP_FREE) {
1103 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1104 			aiop->aiothreadflags &= ~AIOP_FREE;
1105 		}
1106 
1107 		/*
1108 		 * Check for jobs.
1109 		 */
1110 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1111 			mtx_unlock(&aio_job_mtx);
1112 			userp = aiocbe->userproc;
1113 
1114 			/*
1115 			 * Connect to process address space for user program.
1116 			 */
1117 			if (userp != curcp) {
1118 				/*
1119 				 * Save the current address space that we are
1120 				 * connected to.
1121 				 */
1122 				tmpvm = mycp->p_vmspace;
1123 
1124 				/*
1125 				 * Point to the new user address space, and
1126 				 * refer to it.
1127 				 */
1128 				mycp->p_vmspace = userp->p_vmspace;
1129 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1130 
1131 				/* Activate the new mapping. */
1132 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1133 
1134 				/*
1135 				 * If the old address space wasn't the daemons
1136 				 * own address space, then we need to remove the
1137 				 * daemon's reference from the other process
1138 				 * that it was acting on behalf of.
1139 				 */
1140 				if (tmpvm != myvm) {
1141 					vmspace_free(tmpvm);
1142 				}
1143 				curcp = userp;
1144 			}
1145 
1146 			ki = userp->p_aioinfo;
1147 
1148 			/* Do the I/O function. */
1149 			switch(aiocbe->uaiocb.aio_lio_opcode) {
1150 			case LIO_READ:
1151 			case LIO_WRITE:
1152 				aio_process_rw(aiocbe);
1153 				break;
1154 			case LIO_SYNC:
1155 				aio_process_sync(aiocbe);
1156 				break;
1157 			case LIO_MLOCK:
1158 				aio_process_mlock(aiocbe);
1159 				break;
1160 			}
1161 
1162 			mtx_lock(&aio_job_mtx);
1163 			/* Decrement the active job count. */
1164 			ki->kaio_active_count--;
1165 			mtx_unlock(&aio_job_mtx);
1166 
1167 			AIO_LOCK(ki);
1168 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1169 			aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1170 			AIO_UNLOCK(ki);
1171 
1172 			mtx_lock(&aio_job_mtx);
1173 		}
1174 
1175 		/*
1176 		 * Disconnect from user address space.
1177 		 */
1178 		if (curcp != mycp) {
1179 
1180 			mtx_unlock(&aio_job_mtx);
1181 
1182 			/* Get the user address space to disconnect from. */
1183 			tmpvm = mycp->p_vmspace;
1184 
1185 			/* Get original address space for daemon. */
1186 			mycp->p_vmspace = myvm;
1187 
1188 			/* Activate the daemon's address space. */
1189 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1190 #ifdef DIAGNOSTIC
1191 			if (tmpvm == myvm) {
1192 				printf("AIOD: vmspace problem -- %d\n",
1193 				    mycp->p_pid);
1194 			}
1195 #endif
1196 			/* Remove our vmspace reference. */
1197 			vmspace_free(tmpvm);
1198 
1199 			curcp = mycp;
1200 
1201 			mtx_lock(&aio_job_mtx);
1202 			/*
1203 			 * We have to restart to avoid race, we only sleep if
1204 			 * no job can be selected, that should be
1205 			 * curcp == mycp.
1206 			 */
1207 			continue;
1208 		}
1209 
1210 		mtx_assert(&aio_job_mtx, MA_OWNED);
1211 
1212 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1213 		aiop->aiothreadflags |= AIOP_FREE;
1214 
1215 		/*
1216 		 * If daemon is inactive for a long time, allow it to exit,
1217 		 * thereby freeing resources.
1218 		 */
1219 		if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1220 		    aiod_lifetime)) {
1221 			if (TAILQ_EMPTY(&aio_jobs)) {
1222 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1223 				    (num_aio_procs > target_aio_procs)) {
1224 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1225 					num_aio_procs--;
1226 					mtx_unlock(&aio_job_mtx);
1227 					uma_zfree(aiop_zone, aiop);
1228 					free_unr(aiod_unr, id);
1229 #ifdef DIAGNOSTIC
1230 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1231 						printf("AIOD: bad vm refcnt for"
1232 						    " exiting daemon: %d\n",
1233 						    mycp->p_vmspace->vm_refcnt);
1234 					}
1235 #endif
1236 					kproc_exit(0);
1237 				}
1238 			}
1239 		}
1240 	}
1241 	mtx_unlock(&aio_job_mtx);
1242 	panic("shouldn't be here\n");
1243 }
1244 
1245 /*
1246  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1247  * AIO daemon modifies its environment itself.
1248  */
1249 static int
1250 aio_newproc(int *start)
1251 {
1252 	int error;
1253 	struct proc *p;
1254 	int id;
1255 
1256 	id = alloc_unr(aiod_unr);
1257 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1258 		RFNOWAIT, 0, "aiod%d", id);
1259 	if (error == 0) {
1260 		/*
1261 		 * Wait until daemon is started.
1262 		 */
1263 		sema_wait(&aio_newproc_sem);
1264 		mtx_lock(&aio_job_mtx);
1265 		num_aio_procs++;
1266 		if (start != NULL)
1267 			(*start)--;
1268 		mtx_unlock(&aio_job_mtx);
1269 	} else {
1270 		free_unr(aiod_unr, id);
1271 	}
1272 	return (error);
1273 }
1274 
1275 /*
1276  * Try the high-performance, low-overhead physio method for eligible
1277  * VCHR devices.  This method doesn't use an aio helper thread, and
1278  * thus has very low overhead.
1279  *
1280  * Assumes that the caller, aio_aqueue(), has incremented the file
1281  * structure's reference count, preventing its deallocation for the
1282  * duration of this call.
1283  */
1284 static int
1285 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1286 {
1287 	struct aiocb *cb;
1288 	struct file *fp;
1289 	struct bio *bp;
1290 	struct buf *pbuf;
1291 	struct vnode *vp;
1292 	struct cdevsw *csw;
1293 	struct cdev *dev;
1294 	struct kaioinfo *ki;
1295 	struct aioliojob *lj;
1296 	int error, ref, unmap, poff;
1297 	vm_prot_t prot;
1298 
1299 	cb = &aiocbe->uaiocb;
1300 	fp = aiocbe->fd_file;
1301 
1302 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1303 		return (-1);
1304 
1305 	vp = fp->f_vnode;
1306 	if (vp->v_type != VCHR)
1307 		return (-1);
1308 	if (vp->v_bufobj.bo_bsize == 0)
1309 		return (-1);
1310 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1311 		return (-1);
1312 
1313 	ref = 0;
1314 	csw = devvn_refthread(vp, &dev, &ref);
1315 	if (csw == NULL)
1316 		return (ENXIO);
1317 
1318 	if ((csw->d_flags & D_DISK) == 0) {
1319 		error = -1;
1320 		goto unref;
1321 	}
1322 	if (cb->aio_nbytes > dev->si_iosize_max) {
1323 		error = -1;
1324 		goto unref;
1325 	}
1326 
1327 	ki = p->p_aioinfo;
1328 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1329 	unmap = ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed);
1330 	if (unmap) {
1331 		if (cb->aio_nbytes > MAXPHYS) {
1332 			error = -1;
1333 			goto unref;
1334 		}
1335 	} else {
1336 		if (cb->aio_nbytes > MAXPHYS - poff) {
1337 			error = -1;
1338 			goto unref;
1339 		}
1340 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1341 			error = -1;
1342 			goto unref;
1343 		}
1344 	}
1345 	aiocbe->bp = bp = g_alloc_bio();
1346 	if (!unmap) {
1347 		aiocbe->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1348 		BUF_KERNPROC(pbuf);
1349 	}
1350 
1351 	AIO_LOCK(ki);
1352 	ki->kaio_count++;
1353 	if (!unmap)
1354 		ki->kaio_buffer_count++;
1355 	lj = aiocbe->lio;
1356 	if (lj)
1357 		lj->lioj_count++;
1358 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1359 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1360 	aiocbe->jobstate = JOBST_JOBQBUF;
1361 	cb->_aiocb_private.status = cb->aio_nbytes;
1362 	AIO_UNLOCK(ki);
1363 
1364 	bp->bio_length = cb->aio_nbytes;
1365 	bp->bio_bcount = cb->aio_nbytes;
1366 	bp->bio_done = aio_physwakeup;
1367 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1368 	bp->bio_offset = cb->aio_offset;
1369 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1370 	bp->bio_dev = dev;
1371 	bp->bio_caller1 = (void *)aiocbe;
1372 
1373 	prot = VM_PROT_READ;
1374 	if (cb->aio_lio_opcode == LIO_READ)
1375 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1376 	if ((aiocbe->npages = vm_fault_quick_hold_pages(
1377 	    &curproc->p_vmspace->vm_map,
1378 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, aiocbe->pages,
1379 	    sizeof(aiocbe->pages)/sizeof(aiocbe->pages[0]))) < 0) {
1380 		error = EFAULT;
1381 		goto doerror;
1382 	}
1383 	if (!unmap) {
1384 		pmap_qenter((vm_offset_t)pbuf->b_data,
1385 		    aiocbe->pages, aiocbe->npages);
1386 		bp->bio_data = pbuf->b_data + poff;
1387 	} else {
1388 		bp->bio_ma = aiocbe->pages;
1389 		bp->bio_ma_n = aiocbe->npages;
1390 		bp->bio_ma_offset = poff;
1391 		bp->bio_data = unmapped_buf;
1392 		bp->bio_flags |= BIO_UNMAPPED;
1393 	}
1394 
1395 	atomic_add_int(&num_queue_count, 1);
1396 	if (!unmap)
1397 		atomic_add_int(&num_buf_aio, 1);
1398 
1399 	/* Perform transfer. */
1400 	csw->d_strategy(bp);
1401 	dev_relthread(dev, ref);
1402 	return (0);
1403 
1404 doerror:
1405 	AIO_LOCK(ki);
1406 	aiocbe->jobstate = JOBST_NULL;
1407 	TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1408 	TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
1409 	ki->kaio_count--;
1410 	if (!unmap)
1411 		ki->kaio_buffer_count--;
1412 	if (lj)
1413 		lj->lioj_count--;
1414 	AIO_UNLOCK(ki);
1415 	if (pbuf) {
1416 		relpbuf(pbuf, NULL);
1417 		aiocbe->pbuf = NULL;
1418 	}
1419 	g_destroy_bio(bp);
1420 	aiocbe->bp = NULL;
1421 unref:
1422 	dev_relthread(dev, ref);
1423 	return (error);
1424 }
1425 
1426 /*
1427  * Wake up aio requests that may be serviceable now.
1428  */
1429 static void
1430 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1431 {
1432 	struct aiocblist *cb, *cbn;
1433 	int opcode;
1434 
1435 	SOCKBUF_LOCK_ASSERT(sb);
1436 	if (sb == &so->so_snd)
1437 		opcode = LIO_WRITE;
1438 	else
1439 		opcode = LIO_READ;
1440 
1441 	sb->sb_flags &= ~SB_AIO;
1442 	mtx_lock(&aio_job_mtx);
1443 	TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1444 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1445 			if (cb->jobstate != JOBST_JOBQSOCK)
1446 				panic("invalid queue value");
1447 			/* XXX
1448 			 * We don't have actual sockets backend yet,
1449 			 * so we simply move the requests to the generic
1450 			 * file I/O backend.
1451 			 */
1452 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1453 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1454 			aio_kick_nowait(cb->userproc);
1455 		}
1456 	}
1457 	mtx_unlock(&aio_job_mtx);
1458 }
1459 
1460 static int
1461 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1462 {
1463 
1464 	/*
1465 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1466 	 * supported by AIO with the old sigevent structure.
1467 	 */
1468 	nsig->sigev_notify = osig->sigev_notify;
1469 	switch (nsig->sigev_notify) {
1470 	case SIGEV_NONE:
1471 		break;
1472 	case SIGEV_SIGNAL:
1473 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1474 		break;
1475 	case SIGEV_KEVENT:
1476 		nsig->sigev_notify_kqueue =
1477 		    osig->__sigev_u.__sigev_notify_kqueue;
1478 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1479 		break;
1480 	default:
1481 		return (EINVAL);
1482 	}
1483 	return (0);
1484 }
1485 
1486 static int
1487 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1488 {
1489 	struct oaiocb *ojob;
1490 	int error;
1491 
1492 	bzero(kjob, sizeof(struct aiocb));
1493 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1494 	if (error)
1495 		return (error);
1496 	ojob = (struct oaiocb *)kjob;
1497 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1498 }
1499 
1500 static int
1501 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1502 {
1503 
1504 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1505 }
1506 
1507 static long
1508 aiocb_fetch_status(struct aiocb *ujob)
1509 {
1510 
1511 	return (fuword(&ujob->_aiocb_private.status));
1512 }
1513 
1514 static long
1515 aiocb_fetch_error(struct aiocb *ujob)
1516 {
1517 
1518 	return (fuword(&ujob->_aiocb_private.error));
1519 }
1520 
1521 static int
1522 aiocb_store_status(struct aiocb *ujob, long status)
1523 {
1524 
1525 	return (suword(&ujob->_aiocb_private.status, status));
1526 }
1527 
1528 static int
1529 aiocb_store_error(struct aiocb *ujob, long error)
1530 {
1531 
1532 	return (suword(&ujob->_aiocb_private.error, error));
1533 }
1534 
1535 static int
1536 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1537 {
1538 
1539 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1540 }
1541 
1542 static int
1543 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1544 {
1545 
1546 	return (suword(ujobp, (long)ujob));
1547 }
1548 
1549 static struct aiocb_ops aiocb_ops = {
1550 	.copyin = aiocb_copyin,
1551 	.fetch_status = aiocb_fetch_status,
1552 	.fetch_error = aiocb_fetch_error,
1553 	.store_status = aiocb_store_status,
1554 	.store_error = aiocb_store_error,
1555 	.store_kernelinfo = aiocb_store_kernelinfo,
1556 	.store_aiocb = aiocb_store_aiocb,
1557 };
1558 
1559 static struct aiocb_ops aiocb_ops_osigevent = {
1560 	.copyin = aiocb_copyin_old_sigevent,
1561 	.fetch_status = aiocb_fetch_status,
1562 	.fetch_error = aiocb_fetch_error,
1563 	.store_status = aiocb_store_status,
1564 	.store_error = aiocb_store_error,
1565 	.store_kernelinfo = aiocb_store_kernelinfo,
1566 	.store_aiocb = aiocb_store_aiocb,
1567 };
1568 
1569 /*
1570  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1571  * technique is done in this code.
1572  */
1573 int
1574 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1575 	int type, struct aiocb_ops *ops)
1576 {
1577 	struct proc *p = td->td_proc;
1578 	cap_rights_t rights;
1579 	struct file *fp;
1580 	struct socket *so;
1581 	struct aiocblist *aiocbe, *cb;
1582 	struct kaioinfo *ki;
1583 	struct kevent kev;
1584 	struct sockbuf *sb;
1585 	int opcode;
1586 	int error;
1587 	int fd, kqfd;
1588 	int jid;
1589 	u_short evflags;
1590 
1591 	if (p->p_aioinfo == NULL)
1592 		aio_init_aioinfo(p);
1593 
1594 	ki = p->p_aioinfo;
1595 
1596 	ops->store_status(job, -1);
1597 	ops->store_error(job, 0);
1598 	ops->store_kernelinfo(job, -1);
1599 
1600 	if (num_queue_count >= max_queue_count ||
1601 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1602 		ops->store_error(job, EAGAIN);
1603 		return (EAGAIN);
1604 	}
1605 
1606 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1607 	knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1608 
1609 	error = ops->copyin(job, &aiocbe->uaiocb);
1610 	if (error) {
1611 		ops->store_error(job, error);
1612 		uma_zfree(aiocb_zone, aiocbe);
1613 		return (error);
1614 	}
1615 
1616 	/* XXX: aio_nbytes is later casted to signed types. */
1617 	if (aiocbe->uaiocb.aio_nbytes > INT_MAX) {
1618 		uma_zfree(aiocb_zone, aiocbe);
1619 		return (EINVAL);
1620 	}
1621 
1622 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1623 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1624 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1625 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1626 		ops->store_error(job, EINVAL);
1627 		uma_zfree(aiocb_zone, aiocbe);
1628 		return (EINVAL);
1629 	}
1630 
1631 	if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1632 	     aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1633 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1634 		uma_zfree(aiocb_zone, aiocbe);
1635 		return (EINVAL);
1636 	}
1637 
1638 	ksiginfo_init(&aiocbe->ksi);
1639 
1640 	/* Save userspace address of the job info. */
1641 	aiocbe->uuaiocb = job;
1642 
1643 	/* Get the opcode. */
1644 	if (type != LIO_NOP)
1645 		aiocbe->uaiocb.aio_lio_opcode = type;
1646 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1647 
1648 	/*
1649 	 * Validate the opcode and fetch the file object for the specified
1650 	 * file descriptor.
1651 	 *
1652 	 * XXXRW: Moved the opcode validation up here so that we don't
1653 	 * retrieve a file descriptor without knowing what the capabiltity
1654 	 * should be.
1655 	 */
1656 	fd = aiocbe->uaiocb.aio_fildes;
1657 	switch (opcode) {
1658 	case LIO_WRITE:
1659 		error = fget_write(td, fd,
1660 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1661 		break;
1662 	case LIO_READ:
1663 		error = fget_read(td, fd,
1664 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1665 		break;
1666 	case LIO_SYNC:
1667 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1668 		break;
1669 	case LIO_MLOCK:
1670 		fp = NULL;
1671 		break;
1672 	case LIO_NOP:
1673 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1674 		break;
1675 	default:
1676 		error = EINVAL;
1677 	}
1678 	if (error) {
1679 		uma_zfree(aiocb_zone, aiocbe);
1680 		ops->store_error(job, error);
1681 		return (error);
1682 	}
1683 
1684 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1685 		error = EINVAL;
1686 		goto aqueue_fail;
1687 	}
1688 
1689 	if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1690 		error = EINVAL;
1691 		goto aqueue_fail;
1692 	}
1693 
1694 	aiocbe->fd_file = fp;
1695 
1696 	mtx_lock(&aio_job_mtx);
1697 	jid = jobrefid++;
1698 	aiocbe->seqno = jobseqno++;
1699 	mtx_unlock(&aio_job_mtx);
1700 	error = ops->store_kernelinfo(job, jid);
1701 	if (error) {
1702 		error = EINVAL;
1703 		goto aqueue_fail;
1704 	}
1705 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1706 
1707 	if (opcode == LIO_NOP) {
1708 		fdrop(fp, td);
1709 		uma_zfree(aiocb_zone, aiocbe);
1710 		return (0);
1711 	}
1712 
1713 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1714 		goto no_kqueue;
1715 	evflags = aiocbe->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1716 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1717 		error = EINVAL;
1718 		goto aqueue_fail;
1719 	}
1720 	kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1721 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1722 	kev.filter = EVFILT_AIO;
1723 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1724 	kev.data = (intptr_t)aiocbe;
1725 	kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1726 	error = kqfd_register(kqfd, &kev, td, 1);
1727 aqueue_fail:
1728 	if (error) {
1729 		if (fp)
1730 			fdrop(fp, td);
1731 		uma_zfree(aiocb_zone, aiocbe);
1732 		ops->store_error(job, error);
1733 		goto done;
1734 	}
1735 no_kqueue:
1736 
1737 	ops->store_error(job, EINPROGRESS);
1738 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1739 	aiocbe->userproc = p;
1740 	aiocbe->cred = crhold(td->td_ucred);
1741 	aiocbe->jobflags = 0;
1742 	aiocbe->lio = lj;
1743 
1744 	if (opcode == LIO_SYNC)
1745 		goto queueit;
1746 
1747 	if (fp && fp->f_type == DTYPE_SOCKET) {
1748 		/*
1749 		 * Alternate queueing for socket ops: Reach down into the
1750 		 * descriptor to get the socket data.  Then check to see if the
1751 		 * socket is ready to be read or written (based on the requested
1752 		 * operation).
1753 		 *
1754 		 * If it is not ready for io, then queue the aiocbe on the
1755 		 * socket, and set the flags so we get a call when sbnotify()
1756 		 * happens.
1757 		 *
1758 		 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1759 		 * and unlock the snd sockbuf for no reason.
1760 		 */
1761 		so = fp->f_data;
1762 		sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1763 		SOCKBUF_LOCK(sb);
1764 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1765 		    LIO_WRITE) && (!sowriteable(so)))) {
1766 			sb->sb_flags |= SB_AIO;
1767 
1768 			mtx_lock(&aio_job_mtx);
1769 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1770 			mtx_unlock(&aio_job_mtx);
1771 
1772 			AIO_LOCK(ki);
1773 			TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1774 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1775 			aiocbe->jobstate = JOBST_JOBQSOCK;
1776 			ki->kaio_count++;
1777 			if (lj)
1778 				lj->lioj_count++;
1779 			AIO_UNLOCK(ki);
1780 			SOCKBUF_UNLOCK(sb);
1781 			atomic_add_int(&num_queue_count, 1);
1782 			error = 0;
1783 			goto done;
1784 		}
1785 		SOCKBUF_UNLOCK(sb);
1786 	}
1787 
1788 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1789 		goto done;
1790 #if 0
1791 	if (error > 0) {
1792 		aiocbe->uaiocb._aiocb_private.error = error;
1793 		ops->store_error(job, error);
1794 		goto done;
1795 	}
1796 #endif
1797 queueit:
1798 	atomic_add_int(&num_queue_count, 1);
1799 
1800 	AIO_LOCK(ki);
1801 	ki->kaio_count++;
1802 	if (lj)
1803 		lj->lioj_count++;
1804 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1805 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1806 	if (opcode == LIO_SYNC) {
1807 		TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1808 			if (cb->fd_file == aiocbe->fd_file &&
1809 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1810 			    cb->seqno < aiocbe->seqno) {
1811 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1812 				aiocbe->pending++;
1813 			}
1814 		}
1815 		TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1816 			if (cb->fd_file == aiocbe->fd_file &&
1817 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1818 			    cb->seqno < aiocbe->seqno) {
1819 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1820 				aiocbe->pending++;
1821 			}
1822 		}
1823 		if (aiocbe->pending != 0) {
1824 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1825 			aiocbe->jobstate = JOBST_JOBQSYNC;
1826 			AIO_UNLOCK(ki);
1827 			goto done;
1828 		}
1829 	}
1830 	mtx_lock(&aio_job_mtx);
1831 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1832 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1833 	aio_kick_nowait(p);
1834 	mtx_unlock(&aio_job_mtx);
1835 	AIO_UNLOCK(ki);
1836 	error = 0;
1837 done:
1838 	return (error);
1839 }
1840 
1841 static void
1842 aio_kick_nowait(struct proc *userp)
1843 {
1844 	struct kaioinfo *ki = userp->p_aioinfo;
1845 	struct aiothreadlist *aiop;
1846 
1847 	mtx_assert(&aio_job_mtx, MA_OWNED);
1848 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1849 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1850 		aiop->aiothreadflags &= ~AIOP_FREE;
1851 		wakeup(aiop->aiothread);
1852 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1853 	    ((ki->kaio_active_count + num_aio_resv_start) <
1854 	    ki->kaio_maxactive_count)) {
1855 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1856 	}
1857 }
1858 
1859 static int
1860 aio_kick(struct proc *userp)
1861 {
1862 	struct kaioinfo *ki = userp->p_aioinfo;
1863 	struct aiothreadlist *aiop;
1864 	int error, ret = 0;
1865 
1866 	mtx_assert(&aio_job_mtx, MA_OWNED);
1867 retryproc:
1868 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1869 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1870 		aiop->aiothreadflags &= ~AIOP_FREE;
1871 		wakeup(aiop->aiothread);
1872 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1873 	    ((ki->kaio_active_count + num_aio_resv_start) <
1874 	    ki->kaio_maxactive_count)) {
1875 		num_aio_resv_start++;
1876 		mtx_unlock(&aio_job_mtx);
1877 		error = aio_newproc(&num_aio_resv_start);
1878 		mtx_lock(&aio_job_mtx);
1879 		if (error) {
1880 			num_aio_resv_start--;
1881 			goto retryproc;
1882 		}
1883 	} else {
1884 		ret = -1;
1885 	}
1886 	return (ret);
1887 }
1888 
1889 static void
1890 aio_kick_helper(void *context, int pending)
1891 {
1892 	struct proc *userp = context;
1893 
1894 	mtx_lock(&aio_job_mtx);
1895 	while (--pending >= 0) {
1896 		if (aio_kick(userp))
1897 			break;
1898 	}
1899 	mtx_unlock(&aio_job_mtx);
1900 }
1901 
1902 /*
1903  * Support the aio_return system call, as a side-effect, kernel resources are
1904  * released.
1905  */
1906 static int
1907 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1908 {
1909 	struct proc *p = td->td_proc;
1910 	struct aiocblist *cb;
1911 	struct kaioinfo *ki;
1912 	int status, error;
1913 
1914 	ki = p->p_aioinfo;
1915 	if (ki == NULL)
1916 		return (EINVAL);
1917 	AIO_LOCK(ki);
1918 	TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1919 		if (cb->uuaiocb == uaiocb)
1920 			break;
1921 	}
1922 	if (cb != NULL) {
1923 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
1924 		status = cb->uaiocb._aiocb_private.status;
1925 		error = cb->uaiocb._aiocb_private.error;
1926 		td->td_retval[0] = status;
1927 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1928 			td->td_ru.ru_oublock += cb->outputcharge;
1929 			cb->outputcharge = 0;
1930 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1931 			td->td_ru.ru_inblock += cb->inputcharge;
1932 			cb->inputcharge = 0;
1933 		}
1934 		aio_free_entry(cb);
1935 		AIO_UNLOCK(ki);
1936 		ops->store_error(uaiocb, error);
1937 		ops->store_status(uaiocb, status);
1938 	} else {
1939 		error = EINVAL;
1940 		AIO_UNLOCK(ki);
1941 	}
1942 	return (error);
1943 }
1944 
1945 int
1946 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1947 {
1948 
1949 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1950 }
1951 
1952 /*
1953  * Allow a process to wakeup when any of the I/O requests are completed.
1954  */
1955 static int
1956 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1957     struct timespec *ts)
1958 {
1959 	struct proc *p = td->td_proc;
1960 	struct timeval atv;
1961 	struct kaioinfo *ki;
1962 	struct aiocblist *cb, *cbfirst;
1963 	int error, i, timo;
1964 
1965 	timo = 0;
1966 	if (ts) {
1967 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1968 			return (EINVAL);
1969 
1970 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1971 		if (itimerfix(&atv))
1972 			return (EINVAL);
1973 		timo = tvtohz(&atv);
1974 	}
1975 
1976 	ki = p->p_aioinfo;
1977 	if (ki == NULL)
1978 		return (EAGAIN);
1979 
1980 	if (njoblist == 0)
1981 		return (0);
1982 
1983 	AIO_LOCK(ki);
1984 	for (;;) {
1985 		cbfirst = NULL;
1986 		error = 0;
1987 		TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1988 			for (i = 0; i < njoblist; i++) {
1989 				if (cb->uuaiocb == ujoblist[i]) {
1990 					if (cbfirst == NULL)
1991 						cbfirst = cb;
1992 					if (cb->jobstate == JOBST_JOBFINISHED)
1993 						goto RETURN;
1994 				}
1995 			}
1996 		}
1997 		/* All tasks were finished. */
1998 		if (cbfirst == NULL)
1999 			break;
2000 
2001 		ki->kaio_flags |= KAIO_WAKEUP;
2002 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2003 		    "aiospn", timo);
2004 		if (error == ERESTART)
2005 			error = EINTR;
2006 		if (error)
2007 			break;
2008 	}
2009 RETURN:
2010 	AIO_UNLOCK(ki);
2011 	return (error);
2012 }
2013 
2014 int
2015 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
2016 {
2017 	struct timespec ts, *tsp;
2018 	struct aiocb **ujoblist;
2019 	int error;
2020 
2021 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2022 		return (EINVAL);
2023 
2024 	if (uap->timeout) {
2025 		/* Get timespec struct. */
2026 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
2027 			return (error);
2028 		tsp = &ts;
2029 	} else
2030 		tsp = NULL;
2031 
2032 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2033 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
2034 	if (error == 0)
2035 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2036 	uma_zfree(aiol_zone, ujoblist);
2037 	return (error);
2038 }
2039 
2040 /*
2041  * aio_cancel cancels any non-physio aio operations not currently in
2042  * progress.
2043  */
2044 int
2045 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
2046 {
2047 	struct proc *p = td->td_proc;
2048 	struct kaioinfo *ki;
2049 	struct aiocblist *cbe, *cbn;
2050 	struct file *fp;
2051 	struct socket *so;
2052 	cap_rights_t rights;
2053 	int error;
2054 	int remove;
2055 	int cancelled = 0;
2056 	int notcancelled = 0;
2057 	struct vnode *vp;
2058 
2059 	/* Lookup file object. */
2060 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
2061 	if (error)
2062 		return (error);
2063 
2064 	ki = p->p_aioinfo;
2065 	if (ki == NULL)
2066 		goto done;
2067 
2068 	if (fp->f_type == DTYPE_VNODE) {
2069 		vp = fp->f_vnode;
2070 		if (vn_isdisk(vp, &error)) {
2071 			fdrop(fp, td);
2072 			td->td_retval[0] = AIO_NOTCANCELED;
2073 			return (0);
2074 		}
2075 	}
2076 
2077 	AIO_LOCK(ki);
2078 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
2079 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
2080 		    ((uap->aiocbp == NULL) ||
2081 		     (uap->aiocbp == cbe->uuaiocb))) {
2082 			remove = 0;
2083 
2084 			mtx_lock(&aio_job_mtx);
2085 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
2086 				TAILQ_REMOVE(&aio_jobs, cbe, list);
2087 				remove = 1;
2088 			} else if (cbe->jobstate == JOBST_JOBQSOCK) {
2089 				MPASS(fp->f_type == DTYPE_SOCKET);
2090 				so = fp->f_data;
2091 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2092 				remove = 1;
2093 			} else if (cbe->jobstate == JOBST_JOBQSYNC) {
2094 				TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2095 				remove = 1;
2096 			}
2097 			mtx_unlock(&aio_job_mtx);
2098 
2099 			if (remove) {
2100 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2101 				cbe->uaiocb._aiocb_private.status = -1;
2102 				cbe->uaiocb._aiocb_private.error = ECANCELED;
2103 				aio_bio_done_notify(p, cbe, DONE_QUEUE);
2104 				cancelled++;
2105 			} else {
2106 				notcancelled++;
2107 			}
2108 			if (uap->aiocbp != NULL)
2109 				break;
2110 		}
2111 	}
2112 	AIO_UNLOCK(ki);
2113 
2114 done:
2115 	fdrop(fp, td);
2116 
2117 	if (uap->aiocbp != NULL) {
2118 		if (cancelled) {
2119 			td->td_retval[0] = AIO_CANCELED;
2120 			return (0);
2121 		}
2122 	}
2123 
2124 	if (notcancelled) {
2125 		td->td_retval[0] = AIO_NOTCANCELED;
2126 		return (0);
2127 	}
2128 
2129 	if (cancelled) {
2130 		td->td_retval[0] = AIO_CANCELED;
2131 		return (0);
2132 	}
2133 
2134 	td->td_retval[0] = AIO_ALLDONE;
2135 
2136 	return (0);
2137 }
2138 
2139 /*
2140  * aio_error is implemented in the kernel level for compatibility purposes
2141  * only.  For a user mode async implementation, it would be best to do it in
2142  * a userland subroutine.
2143  */
2144 static int
2145 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2146 {
2147 	struct proc *p = td->td_proc;
2148 	struct aiocblist *cb;
2149 	struct kaioinfo *ki;
2150 	int status;
2151 
2152 	ki = p->p_aioinfo;
2153 	if (ki == NULL) {
2154 		td->td_retval[0] = EINVAL;
2155 		return (0);
2156 	}
2157 
2158 	AIO_LOCK(ki);
2159 	TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2160 		if (cb->uuaiocb == aiocbp) {
2161 			if (cb->jobstate == JOBST_JOBFINISHED)
2162 				td->td_retval[0] =
2163 					cb->uaiocb._aiocb_private.error;
2164 			else
2165 				td->td_retval[0] = EINPROGRESS;
2166 			AIO_UNLOCK(ki);
2167 			return (0);
2168 		}
2169 	}
2170 	AIO_UNLOCK(ki);
2171 
2172 	/*
2173 	 * Hack for failure of aio_aqueue.
2174 	 */
2175 	status = ops->fetch_status(aiocbp);
2176 	if (status == -1) {
2177 		td->td_retval[0] = ops->fetch_error(aiocbp);
2178 		return (0);
2179 	}
2180 
2181 	td->td_retval[0] = EINVAL;
2182 	return (0);
2183 }
2184 
2185 int
2186 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2187 {
2188 
2189 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2190 }
2191 
2192 /* syscall - asynchronous read from a file (REALTIME) */
2193 int
2194 sys_oaio_read(struct thread *td, struct oaio_read_args *uap)
2195 {
2196 
2197 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2198 	    &aiocb_ops_osigevent));
2199 }
2200 
2201 int
2202 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2203 {
2204 
2205 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2206 }
2207 
2208 /* syscall - asynchronous write to a file (REALTIME) */
2209 int
2210 sys_oaio_write(struct thread *td, struct oaio_write_args *uap)
2211 {
2212 
2213 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2214 	    &aiocb_ops_osigevent));
2215 }
2216 
2217 int
2218 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2219 {
2220 
2221 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2222 }
2223 
2224 int
2225 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2226 {
2227 
2228 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2229 }
2230 
2231 static int
2232 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2233     struct aiocb **acb_list, int nent, struct sigevent *sig,
2234     struct aiocb_ops *ops)
2235 {
2236 	struct proc *p = td->td_proc;
2237 	struct aiocb *iocb;
2238 	struct kaioinfo *ki;
2239 	struct aioliojob *lj;
2240 	struct kevent kev;
2241 	int error;
2242 	int nerror;
2243 	int i;
2244 
2245 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2246 		return (EINVAL);
2247 
2248 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2249 		return (EINVAL);
2250 
2251 	if (p->p_aioinfo == NULL)
2252 		aio_init_aioinfo(p);
2253 
2254 	ki = p->p_aioinfo;
2255 
2256 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2257 	lj->lioj_flags = 0;
2258 	lj->lioj_count = 0;
2259 	lj->lioj_finished_count = 0;
2260 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2261 	ksiginfo_init(&lj->lioj_ksi);
2262 
2263 	/*
2264 	 * Setup signal.
2265 	 */
2266 	if (sig && (mode == LIO_NOWAIT)) {
2267 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2268 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2269 			/* Assume only new style KEVENT */
2270 			kev.filter = EVFILT_LIO;
2271 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2272 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2273 			kev.data = (intptr_t)lj;
2274 			/* pass user defined sigval data */
2275 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2276 			error = kqfd_register(
2277 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2278 			if (error) {
2279 				uma_zfree(aiolio_zone, lj);
2280 				return (error);
2281 			}
2282 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2283 			;
2284 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2285 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2286 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2287 					uma_zfree(aiolio_zone, lj);
2288 					return EINVAL;
2289 				}
2290 				lj->lioj_flags |= LIOJ_SIGNAL;
2291 		} else {
2292 			uma_zfree(aiolio_zone, lj);
2293 			return EINVAL;
2294 		}
2295 	}
2296 
2297 	AIO_LOCK(ki);
2298 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2299 	/*
2300 	 * Add extra aiocb count to avoid the lio to be freed
2301 	 * by other threads doing aio_waitcomplete or aio_return,
2302 	 * and prevent event from being sent until we have queued
2303 	 * all tasks.
2304 	 */
2305 	lj->lioj_count = 1;
2306 	AIO_UNLOCK(ki);
2307 
2308 	/*
2309 	 * Get pointers to the list of I/O requests.
2310 	 */
2311 	nerror = 0;
2312 	for (i = 0; i < nent; i++) {
2313 		iocb = acb_list[i];
2314 		if (iocb != NULL) {
2315 			error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2316 			if (error != 0)
2317 				nerror++;
2318 		}
2319 	}
2320 
2321 	error = 0;
2322 	AIO_LOCK(ki);
2323 	if (mode == LIO_WAIT) {
2324 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2325 			ki->kaio_flags |= KAIO_WAKEUP;
2326 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2327 			    PRIBIO | PCATCH, "aiospn", 0);
2328 			if (error == ERESTART)
2329 				error = EINTR;
2330 			if (error)
2331 				break;
2332 		}
2333 	} else {
2334 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2335 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2336 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2337 				KNOTE_LOCKED(&lj->klist, 1);
2338 			}
2339 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2340 			    == LIOJ_SIGNAL
2341 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2342 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2343 				aio_sendsig(p, &lj->lioj_signal,
2344 					    &lj->lioj_ksi);
2345 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2346 			}
2347 		}
2348 	}
2349 	lj->lioj_count--;
2350 	if (lj->lioj_count == 0) {
2351 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2352 		knlist_delete(&lj->klist, curthread, 1);
2353 		PROC_LOCK(p);
2354 		sigqueue_take(&lj->lioj_ksi);
2355 		PROC_UNLOCK(p);
2356 		AIO_UNLOCK(ki);
2357 		uma_zfree(aiolio_zone, lj);
2358 	} else
2359 		AIO_UNLOCK(ki);
2360 
2361 	if (nerror)
2362 		return (EIO);
2363 	return (error);
2364 }
2365 
2366 /* syscall - list directed I/O (REALTIME) */
2367 int
2368 sys_olio_listio(struct thread *td, struct olio_listio_args *uap)
2369 {
2370 	struct aiocb **acb_list;
2371 	struct sigevent *sigp, sig;
2372 	struct osigevent osig;
2373 	int error, nent;
2374 
2375 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2376 		return (EINVAL);
2377 
2378 	nent = uap->nent;
2379 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2380 		return (EINVAL);
2381 
2382 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2383 		error = copyin(uap->sig, &osig, sizeof(osig));
2384 		if (error)
2385 			return (error);
2386 		error = convert_old_sigevent(&osig, &sig);
2387 		if (error)
2388 			return (error);
2389 		sigp = &sig;
2390 	} else
2391 		sigp = NULL;
2392 
2393 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2394 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2395 	if (error == 0)
2396 		error = kern_lio_listio(td, uap->mode,
2397 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2398 		    &aiocb_ops_osigevent);
2399 	free(acb_list, M_LIO);
2400 	return (error);
2401 }
2402 
2403 /* syscall - list directed I/O (REALTIME) */
2404 int
2405 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2406 {
2407 	struct aiocb **acb_list;
2408 	struct sigevent *sigp, sig;
2409 	int error, nent;
2410 
2411 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2412 		return (EINVAL);
2413 
2414 	nent = uap->nent;
2415 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2416 		return (EINVAL);
2417 
2418 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2419 		error = copyin(uap->sig, &sig, sizeof(sig));
2420 		if (error)
2421 			return (error);
2422 		sigp = &sig;
2423 	} else
2424 		sigp = NULL;
2425 
2426 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2427 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2428 	if (error == 0)
2429 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2430 		    nent, sigp, &aiocb_ops);
2431 	free(acb_list, M_LIO);
2432 	return (error);
2433 }
2434 
2435 static void
2436 aio_physwakeup(struct bio *bp)
2437 {
2438 	struct aiocblist *aiocbe = (struct aiocblist *)bp->bio_caller1;
2439 	struct proc *userp;
2440 	struct kaioinfo *ki;
2441 	int nblks;
2442 
2443 	/* Release mapping into kernel space. */
2444 	if (aiocbe->pbuf) {
2445 		pmap_qremove((vm_offset_t)aiocbe->pbuf->b_data, aiocbe->npages);
2446 		relpbuf(aiocbe->pbuf, NULL);
2447 		aiocbe->pbuf = NULL;
2448 		atomic_subtract_int(&num_buf_aio, 1);
2449 	}
2450 	vm_page_unhold_pages(aiocbe->pages, aiocbe->npages);
2451 
2452 	bp = aiocbe->bp;
2453 	aiocbe->bp = NULL;
2454 	userp = aiocbe->userproc;
2455 	ki = userp->p_aioinfo;
2456 	AIO_LOCK(ki);
2457 	aiocbe->uaiocb._aiocb_private.status -= bp->bio_resid;
2458 	aiocbe->uaiocb._aiocb_private.error = 0;
2459 	if (bp->bio_flags & BIO_ERROR)
2460 		aiocbe->uaiocb._aiocb_private.error = bp->bio_error;
2461 	nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2462 	if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2463 		aiocbe->outputcharge += nblks;
2464 	else
2465 		aiocbe->inputcharge += nblks;
2466 	TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2467 	ki->kaio_buffer_count--;
2468 	aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2469 	AIO_UNLOCK(ki);
2470 
2471 	g_destroy_bio(bp);
2472 }
2473 
2474 /* syscall - wait for the next completion of an aio request */
2475 static int
2476 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2477     struct timespec *ts, struct aiocb_ops *ops)
2478 {
2479 	struct proc *p = td->td_proc;
2480 	struct timeval atv;
2481 	struct kaioinfo *ki;
2482 	struct aiocblist *cb;
2483 	struct aiocb *uuaiocb;
2484 	int error, status, timo;
2485 
2486 	ops->store_aiocb(aiocbp, NULL);
2487 
2488 	if (ts == NULL) {
2489 		timo = 0;
2490 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2491 		timo = -1;
2492 	} else {
2493 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2494 			return (EINVAL);
2495 
2496 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2497 		if (itimerfix(&atv))
2498 			return (EINVAL);
2499 		timo = tvtohz(&atv);
2500 	}
2501 
2502 	if (p->p_aioinfo == NULL)
2503 		aio_init_aioinfo(p);
2504 	ki = p->p_aioinfo;
2505 
2506 	error = 0;
2507 	cb = NULL;
2508 	AIO_LOCK(ki);
2509 	while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2510 		if (timo == -1) {
2511 			error = EWOULDBLOCK;
2512 			break;
2513 		}
2514 		ki->kaio_flags |= KAIO_WAKEUP;
2515 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2516 		    "aiowc", timo);
2517 		if (timo && error == ERESTART)
2518 			error = EINTR;
2519 		if (error)
2520 			break;
2521 	}
2522 
2523 	if (cb != NULL) {
2524 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
2525 		uuaiocb = cb->uuaiocb;
2526 		status = cb->uaiocb._aiocb_private.status;
2527 		error = cb->uaiocb._aiocb_private.error;
2528 		td->td_retval[0] = status;
2529 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2530 			td->td_ru.ru_oublock += cb->outputcharge;
2531 			cb->outputcharge = 0;
2532 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2533 			td->td_ru.ru_inblock += cb->inputcharge;
2534 			cb->inputcharge = 0;
2535 		}
2536 		aio_free_entry(cb);
2537 		AIO_UNLOCK(ki);
2538 		ops->store_aiocb(aiocbp, uuaiocb);
2539 		ops->store_error(uuaiocb, error);
2540 		ops->store_status(uuaiocb, status);
2541 	} else
2542 		AIO_UNLOCK(ki);
2543 
2544 	return (error);
2545 }
2546 
2547 int
2548 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2549 {
2550 	struct timespec ts, *tsp;
2551 	int error;
2552 
2553 	if (uap->timeout) {
2554 		/* Get timespec struct. */
2555 		error = copyin(uap->timeout, &ts, sizeof(ts));
2556 		if (error)
2557 			return (error);
2558 		tsp = &ts;
2559 	} else
2560 		tsp = NULL;
2561 
2562 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2563 }
2564 
2565 static int
2566 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2567     struct aiocb_ops *ops)
2568 {
2569 	struct proc *p = td->td_proc;
2570 	struct kaioinfo *ki;
2571 
2572 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2573 		return (EINVAL);
2574 	ki = p->p_aioinfo;
2575 	if (ki == NULL)
2576 		aio_init_aioinfo(p);
2577 	return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2578 }
2579 
2580 int
2581 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2582 {
2583 
2584 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2585 }
2586 
2587 /* kqueue attach function */
2588 static int
2589 filt_aioattach(struct knote *kn)
2590 {
2591 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2592 
2593 	/*
2594 	 * The aiocbe pointer must be validated before using it, so
2595 	 * registration is restricted to the kernel; the user cannot
2596 	 * set EV_FLAG1.
2597 	 */
2598 	if ((kn->kn_flags & EV_FLAG1) == 0)
2599 		return (EPERM);
2600 	kn->kn_ptr.p_aio = aiocbe;
2601 	kn->kn_flags &= ~EV_FLAG1;
2602 
2603 	knlist_add(&aiocbe->klist, kn, 0);
2604 
2605 	return (0);
2606 }
2607 
2608 /* kqueue detach function */
2609 static void
2610 filt_aiodetach(struct knote *kn)
2611 {
2612 	struct knlist *knl;
2613 
2614 	knl = &kn->kn_ptr.p_aio->klist;
2615 	knl->kl_lock(knl->kl_lockarg);
2616 	if (!knlist_empty(knl))
2617 		knlist_remove(knl, kn, 1);
2618 	knl->kl_unlock(knl->kl_lockarg);
2619 }
2620 
2621 /* kqueue filter function */
2622 /*ARGSUSED*/
2623 static int
2624 filt_aio(struct knote *kn, long hint)
2625 {
2626 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2627 
2628 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2629 	if (aiocbe->jobstate != JOBST_JOBFINISHED)
2630 		return (0);
2631 	kn->kn_flags |= EV_EOF;
2632 	return (1);
2633 }
2634 
2635 /* kqueue attach function */
2636 static int
2637 filt_lioattach(struct knote *kn)
2638 {
2639 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2640 
2641 	/*
2642 	 * The aioliojob pointer must be validated before using it, so
2643 	 * registration is restricted to the kernel; the user cannot
2644 	 * set EV_FLAG1.
2645 	 */
2646 	if ((kn->kn_flags & EV_FLAG1) == 0)
2647 		return (EPERM);
2648 	kn->kn_ptr.p_lio = lj;
2649 	kn->kn_flags &= ~EV_FLAG1;
2650 
2651 	knlist_add(&lj->klist, kn, 0);
2652 
2653 	return (0);
2654 }
2655 
2656 /* kqueue detach function */
2657 static void
2658 filt_liodetach(struct knote *kn)
2659 {
2660 	struct knlist *knl;
2661 
2662 	knl = &kn->kn_ptr.p_lio->klist;
2663 	knl->kl_lock(knl->kl_lockarg);
2664 	if (!knlist_empty(knl))
2665 		knlist_remove(knl, kn, 1);
2666 	knl->kl_unlock(knl->kl_lockarg);
2667 }
2668 
2669 /* kqueue filter function */
2670 /*ARGSUSED*/
2671 static int
2672 filt_lio(struct knote *kn, long hint)
2673 {
2674 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2675 
2676 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2677 }
2678 
2679 #ifdef COMPAT_FREEBSD32
2680 
2681 struct __aiocb_private32 {
2682 	int32_t	status;
2683 	int32_t	error;
2684 	uint32_t kernelinfo;
2685 };
2686 
2687 typedef struct oaiocb32 {
2688 	int	aio_fildes;		/* File descriptor */
2689 	uint64_t aio_offset __packed;	/* File offset for I/O */
2690 	uint32_t aio_buf;		/* I/O buffer in process space */
2691 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2692 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2693 	int	aio_lio_opcode;		/* LIO opcode */
2694 	int	aio_reqprio;		/* Request priority -- ignored */
2695 	struct	__aiocb_private32 _aiocb_private;
2696 } oaiocb32_t;
2697 
2698 typedef struct aiocb32 {
2699 	int32_t	aio_fildes;		/* File descriptor */
2700 	uint64_t aio_offset __packed;	/* File offset for I/O */
2701 	uint32_t aio_buf;		/* I/O buffer in process space */
2702 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2703 	int	__spare__[2];
2704 	uint32_t __spare2__;
2705 	int	aio_lio_opcode;		/* LIO opcode */
2706 	int	aio_reqprio;		/* Request priority -- ignored */
2707 	struct __aiocb_private32 _aiocb_private;
2708 	struct sigevent32 aio_sigevent;	/* Signal to deliver */
2709 } aiocb32_t;
2710 
2711 static int
2712 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2713 {
2714 
2715 	/*
2716 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2717 	 * supported by AIO with the old sigevent structure.
2718 	 */
2719 	CP(*osig, *nsig, sigev_notify);
2720 	switch (nsig->sigev_notify) {
2721 	case SIGEV_NONE:
2722 		break;
2723 	case SIGEV_SIGNAL:
2724 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2725 		break;
2726 	case SIGEV_KEVENT:
2727 		nsig->sigev_notify_kqueue =
2728 		    osig->__sigev_u.__sigev_notify_kqueue;
2729 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2730 		break;
2731 	default:
2732 		return (EINVAL);
2733 	}
2734 	return (0);
2735 }
2736 
2737 static int
2738 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2739 {
2740 	struct oaiocb32 job32;
2741 	int error;
2742 
2743 	bzero(kjob, sizeof(struct aiocb));
2744 	error = copyin(ujob, &job32, sizeof(job32));
2745 	if (error)
2746 		return (error);
2747 
2748 	CP(job32, *kjob, aio_fildes);
2749 	CP(job32, *kjob, aio_offset);
2750 	PTRIN_CP(job32, *kjob, aio_buf);
2751 	CP(job32, *kjob, aio_nbytes);
2752 	CP(job32, *kjob, aio_lio_opcode);
2753 	CP(job32, *kjob, aio_reqprio);
2754 	CP(job32, *kjob, _aiocb_private.status);
2755 	CP(job32, *kjob, _aiocb_private.error);
2756 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2757 	return (convert_old_sigevent32(&job32.aio_sigevent,
2758 	    &kjob->aio_sigevent));
2759 }
2760 
2761 static int
2762 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2763 {
2764 	struct aiocb32 job32;
2765 	int error;
2766 
2767 	error = copyin(ujob, &job32, sizeof(job32));
2768 	if (error)
2769 		return (error);
2770 	CP(job32, *kjob, aio_fildes);
2771 	CP(job32, *kjob, aio_offset);
2772 	PTRIN_CP(job32, *kjob, aio_buf);
2773 	CP(job32, *kjob, aio_nbytes);
2774 	CP(job32, *kjob, aio_lio_opcode);
2775 	CP(job32, *kjob, aio_reqprio);
2776 	CP(job32, *kjob, _aiocb_private.status);
2777 	CP(job32, *kjob, _aiocb_private.error);
2778 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2779 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2780 }
2781 
2782 static long
2783 aiocb32_fetch_status(struct aiocb *ujob)
2784 {
2785 	struct aiocb32 *ujob32;
2786 
2787 	ujob32 = (struct aiocb32 *)ujob;
2788 	return (fuword32(&ujob32->_aiocb_private.status));
2789 }
2790 
2791 static long
2792 aiocb32_fetch_error(struct aiocb *ujob)
2793 {
2794 	struct aiocb32 *ujob32;
2795 
2796 	ujob32 = (struct aiocb32 *)ujob;
2797 	return (fuword32(&ujob32->_aiocb_private.error));
2798 }
2799 
2800 static int
2801 aiocb32_store_status(struct aiocb *ujob, long status)
2802 {
2803 	struct aiocb32 *ujob32;
2804 
2805 	ujob32 = (struct aiocb32 *)ujob;
2806 	return (suword32(&ujob32->_aiocb_private.status, status));
2807 }
2808 
2809 static int
2810 aiocb32_store_error(struct aiocb *ujob, long error)
2811 {
2812 	struct aiocb32 *ujob32;
2813 
2814 	ujob32 = (struct aiocb32 *)ujob;
2815 	return (suword32(&ujob32->_aiocb_private.error, error));
2816 }
2817 
2818 static int
2819 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2820 {
2821 	struct aiocb32 *ujob32;
2822 
2823 	ujob32 = (struct aiocb32 *)ujob;
2824 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2825 }
2826 
2827 static int
2828 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2829 {
2830 
2831 	return (suword32(ujobp, (long)ujob));
2832 }
2833 
2834 static struct aiocb_ops aiocb32_ops = {
2835 	.copyin = aiocb32_copyin,
2836 	.fetch_status = aiocb32_fetch_status,
2837 	.fetch_error = aiocb32_fetch_error,
2838 	.store_status = aiocb32_store_status,
2839 	.store_error = aiocb32_store_error,
2840 	.store_kernelinfo = aiocb32_store_kernelinfo,
2841 	.store_aiocb = aiocb32_store_aiocb,
2842 };
2843 
2844 static struct aiocb_ops aiocb32_ops_osigevent = {
2845 	.copyin = aiocb32_copyin_old_sigevent,
2846 	.fetch_status = aiocb32_fetch_status,
2847 	.fetch_error = aiocb32_fetch_error,
2848 	.store_status = aiocb32_store_status,
2849 	.store_error = aiocb32_store_error,
2850 	.store_kernelinfo = aiocb32_store_kernelinfo,
2851 	.store_aiocb = aiocb32_store_aiocb,
2852 };
2853 
2854 int
2855 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2856 {
2857 
2858 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2859 }
2860 
2861 int
2862 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2863 {
2864 	struct timespec32 ts32;
2865 	struct timespec ts, *tsp;
2866 	struct aiocb **ujoblist;
2867 	uint32_t *ujoblist32;
2868 	int error, i;
2869 
2870 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2871 		return (EINVAL);
2872 
2873 	if (uap->timeout) {
2874 		/* Get timespec struct. */
2875 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2876 			return (error);
2877 		CP(ts32, ts, tv_sec);
2878 		CP(ts32, ts, tv_nsec);
2879 		tsp = &ts;
2880 	} else
2881 		tsp = NULL;
2882 
2883 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2884 	ujoblist32 = (uint32_t *)ujoblist;
2885 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2886 	    sizeof(ujoblist32[0]));
2887 	if (error == 0) {
2888 		for (i = uap->nent; i > 0; i--)
2889 			ujoblist[i] = PTRIN(ujoblist32[i]);
2890 
2891 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2892 	}
2893 	uma_zfree(aiol_zone, ujoblist);
2894 	return (error);
2895 }
2896 
2897 int
2898 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2899 {
2900 
2901 	return (sys_aio_cancel(td, (struct aio_cancel_args *)uap));
2902 }
2903 
2904 int
2905 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2906 {
2907 
2908 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2909 }
2910 
2911 int
2912 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2913 {
2914 
2915 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2916 	    &aiocb32_ops_osigevent));
2917 }
2918 
2919 int
2920 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2921 {
2922 
2923 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2924 	    &aiocb32_ops));
2925 }
2926 
2927 int
2928 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2929 {
2930 
2931 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2932 	    &aiocb32_ops_osigevent));
2933 }
2934 
2935 int
2936 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2937 {
2938 
2939 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2940 	    &aiocb32_ops));
2941 }
2942 
2943 int
2944 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2945 {
2946 
2947 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2948 	    &aiocb32_ops));
2949 }
2950 
2951 int
2952 freebsd32_aio_waitcomplete(struct thread *td,
2953     struct freebsd32_aio_waitcomplete_args *uap)
2954 {
2955 	struct timespec32 ts32;
2956 	struct timespec ts, *tsp;
2957 	int error;
2958 
2959 	if (uap->timeout) {
2960 		/* Get timespec struct. */
2961 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2962 		if (error)
2963 			return (error);
2964 		CP(ts32, ts, tv_sec);
2965 		CP(ts32, ts, tv_nsec);
2966 		tsp = &ts;
2967 	} else
2968 		tsp = NULL;
2969 
2970 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2971 	    &aiocb32_ops));
2972 }
2973 
2974 int
2975 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2976 {
2977 
2978 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2979 	    &aiocb32_ops));
2980 }
2981 
2982 int
2983 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2984 {
2985 	struct aiocb **acb_list;
2986 	struct sigevent *sigp, sig;
2987 	struct osigevent32 osig;
2988 	uint32_t *acb_list32;
2989 	int error, i, nent;
2990 
2991 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2992 		return (EINVAL);
2993 
2994 	nent = uap->nent;
2995 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2996 		return (EINVAL);
2997 
2998 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2999 		error = copyin(uap->sig, &osig, sizeof(osig));
3000 		if (error)
3001 			return (error);
3002 		error = convert_old_sigevent32(&osig, &sig);
3003 		if (error)
3004 			return (error);
3005 		sigp = &sig;
3006 	} else
3007 		sigp = NULL;
3008 
3009 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3010 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3011 	if (error) {
3012 		free(acb_list32, M_LIO);
3013 		return (error);
3014 	}
3015 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3016 	for (i = 0; i < nent; i++)
3017 		acb_list[i] = PTRIN(acb_list32[i]);
3018 	free(acb_list32, M_LIO);
3019 
3020 	error = kern_lio_listio(td, uap->mode,
3021 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3022 	    &aiocb32_ops_osigevent);
3023 	free(acb_list, M_LIO);
3024 	return (error);
3025 }
3026 
3027 int
3028 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
3029 {
3030 	struct aiocb **acb_list;
3031 	struct sigevent *sigp, sig;
3032 	struct sigevent32 sig32;
3033 	uint32_t *acb_list32;
3034 	int error, i, nent;
3035 
3036 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
3037 		return (EINVAL);
3038 
3039 	nent = uap->nent;
3040 	if (nent < 0 || nent > AIO_LISTIO_MAX)
3041 		return (EINVAL);
3042 
3043 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
3044 		error = copyin(uap->sig, &sig32, sizeof(sig32));
3045 		if (error)
3046 			return (error);
3047 		error = convert_sigevent32(&sig32, &sig);
3048 		if (error)
3049 			return (error);
3050 		sigp = &sig;
3051 	} else
3052 		sigp = NULL;
3053 
3054 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
3055 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
3056 	if (error) {
3057 		free(acb_list32, M_LIO);
3058 		return (error);
3059 	}
3060 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
3061 	for (i = 0; i < nent; i++)
3062 		acb_list[i] = PTRIN(acb_list32[i]);
3063 	free(acb_list32, M_LIO);
3064 
3065 	error = kern_lio_listio(td, uap->mode,
3066 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
3067 	    &aiocb32_ops);
3068 	free(acb_list, M_LIO);
3069 	return (error);
3070 }
3071 
3072 #endif
3073