xref: /freebsd/sys/kern/vfs_aio.c (revision aa64588d28258aef88cc33b8043112e8856948d0)
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/eventhandler.h>
32 #include <sys/sysproto.h>
33 #include <sys/filedesc.h>
34 #include <sys/kernel.h>
35 #include <sys/module.h>
36 #include <sys/kthread.h>
37 #include <sys/fcntl.h>
38 #include <sys/file.h>
39 #include <sys/limits.h>
40 #include <sys/lock.h>
41 #include <sys/mutex.h>
42 #include <sys/unistd.h>
43 #include <sys/posix4.h>
44 #include <sys/proc.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/protosw.h>
48 #include <sys/sema.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/syscall.h>
52 #include <sys/sysent.h>
53 #include <sys/sysctl.h>
54 #include <sys/sx.h>
55 #include <sys/taskqueue.h>
56 #include <sys/vnode.h>
57 #include <sys/conf.h>
58 #include <sys/event.h>
59 #include <sys/mount.h>
60 
61 #include <machine/atomic.h>
62 
63 #include <vm/vm.h>
64 #include <vm/vm_extern.h>
65 #include <vm/pmap.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_object.h>
68 #include <vm/uma.h>
69 #include <sys/aio.h>
70 
71 #include "opt_vfs_aio.h"
72 
73 /*
74  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
75  * overflow. (XXX will be removed soon.)
76  */
77 static u_long jobrefid;
78 
79 /*
80  * Counter for aio_fsync.
81  */
82 static uint64_t jobseqno;
83 
84 #define JOBST_NULL		0
85 #define JOBST_JOBQSOCK		1
86 #define JOBST_JOBQGLOBAL	2
87 #define JOBST_JOBRUNNING	3
88 #define JOBST_JOBFINISHED	4
89 #define JOBST_JOBQBUF		5
90 #define JOBST_JOBQSYNC		6
91 
92 #ifndef MAX_AIO_PER_PROC
93 #define MAX_AIO_PER_PROC	32
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE_PER_PROC
97 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
98 #endif
99 
100 #ifndef MAX_AIO_PROCS
101 #define MAX_AIO_PROCS		32
102 #endif
103 
104 #ifndef MAX_AIO_QUEUE
105 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
106 #endif
107 
108 #ifndef TARGET_AIO_PROCS
109 #define TARGET_AIO_PROCS	4
110 #endif
111 
112 #ifndef MAX_BUF_AIO
113 #define MAX_BUF_AIO		16
114 #endif
115 
116 #ifndef AIOD_TIMEOUT_DEFAULT
117 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
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_timeout;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
166     "Timeout value for synchronous aio operations");
167 
168 static int aiod_lifetime;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
170     "Maximum lifetime for idle aiod");
171 
172 static int unloadable = 0;
173 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
174     "Allow unload of aio (not recommended)");
175 
176 
177 static int max_aio_per_proc = MAX_AIO_PER_PROC;
178 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
179     0, "Maximum active aio requests per process (stored in the process)");
180 
181 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
182 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
183     &max_aio_queue_per_proc, 0,
184     "Maximum queued aio requests per process (stored in the process)");
185 
186 static int max_buf_aio = MAX_BUF_AIO;
187 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
188     "Maximum buf aio requests per process (stored in the process)");
189 
190 typedef struct oaiocb {
191 	int	aio_fildes;		/* File descriptor */
192 	off_t	aio_offset;		/* File offset for I/O */
193 	volatile void *aio_buf;         /* I/O buffer in process space */
194 	size_t	aio_nbytes;		/* Number of bytes for I/O */
195 	struct	osigevent aio_sigevent;	/* Signal to deliver */
196 	int	aio_lio_opcode;		/* LIO opcode */
197 	int	aio_reqprio;		/* Request priority -- ignored */
198 	struct	__aiocb_private	_aiocb_private;
199 } oaiocb_t;
200 
201 /*
202  * Below is a key of locks used to protect each member of struct aiocblist
203  * aioliojob and kaioinfo and any backends.
204  *
205  * * - need not protected
206  * a - locked by kaioinfo lock
207  * b - locked by backend lock, the backend lock can be null in some cases,
208  *     for example, BIO belongs to this type, in this case, proc lock is
209  *     reused.
210  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
211  */
212 
213 /*
214  * Current, there is only two backends: BIO and generic file I/O.
215  * socket I/O is served by generic file I/O, this is not a good idea, since
216  * disk file I/O and any other types without O_NONBLOCK flag can block daemon
217  * threads, if there is no thread to serve socket I/O, the socket I/O will be
218  * delayed too long or starved, we should create some threads dedicated to
219  * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
220  * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
221  * structure is not safe because there is race between userland and aio
222  * daemons.
223  */
224 
225 struct aiocblist {
226 	TAILQ_ENTRY(aiocblist) list;	/* (b) internal list of for backend */
227 	TAILQ_ENTRY(aiocblist) plist;	/* (a) list of jobs for each backend */
228 	TAILQ_ENTRY(aiocblist) allist;  /* (a) list of all jobs in proc */
229 	int	jobflags;		/* (a) job flags */
230 	int	jobstate;		/* (b) job state */
231 	int	inputcharge;		/* (*) input blockes */
232 	int	outputcharge;		/* (*) output blockes */
233 	struct	buf *bp;		/* (*) private to BIO backend,
234 				  	 * buffer pointer
235 					 */
236 	struct	proc *userproc;		/* (*) user process */
237 	struct  ucred *cred;		/* (*) active credential when created */
238 	struct	file *fd_file;		/* (*) pointer to file structure */
239 	struct	aioliojob *lio;		/* (*) optional lio job */
240 	struct	aiocb *uuaiocb;		/* (*) pointer in userspace of aiocb */
241 	struct	knlist klist;		/* (a) list of knotes */
242 	struct	aiocb uaiocb;		/* (*) kernel I/O control block */
243 	ksiginfo_t ksi;			/* (a) realtime signal info */
244 	struct	task biotask;		/* (*) private to BIO backend */
245 	uint64_t seqno;			/* (*) job number */
246 	int	pending;		/* (a) number of pending I/O, aio_fsync only */
247 };
248 
249 /* jobflags */
250 #define AIOCBLIST_DONE		0x01
251 #define AIOCBLIST_BUFDONE	0x02
252 #define AIOCBLIST_RUNDOWN	0x04
253 #define AIOCBLIST_CHECKSYNC	0x08
254 
255 /*
256  * AIO process info
257  */
258 #define AIOP_FREE	0x1			/* proc on free queue */
259 
260 struct aiothreadlist {
261 	int aiothreadflags;			/* (c) AIO proc flags */
262 	TAILQ_ENTRY(aiothreadlist) list;	/* (c) list of processes */
263 	struct thread *aiothread;		/* (*) the AIO thread */
264 };
265 
266 /*
267  * data-structure for lio signal management
268  */
269 struct aioliojob {
270 	int	lioj_flags;			/* (a) listio flags */
271 	int	lioj_count;			/* (a) listio flags */
272 	int	lioj_finished_count;		/* (a) listio flags */
273 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
274 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
275 	struct  knlist klist;			/* (a) list of knotes */
276 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
277 };
278 
279 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
280 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
281 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
282 
283 /*
284  * per process aio data structure
285  */
286 struct kaioinfo {
287 	struct mtx	kaio_mtx;	/* the lock to protect this struct */
288 	int	kaio_flags;		/* (a) per process kaio flags */
289 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
290 	int	kaio_active_count;	/* (c) number of currently used AIOs */
291 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
292 	int	kaio_count;		/* (a) size of AIO queue */
293 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
294 	int	kaio_buffer_count;	/* (a) number of physio buffers */
295 	TAILQ_HEAD(,aiocblist) kaio_all;	/* (a) all AIOs in the process */
296 	TAILQ_HEAD(,aiocblist) kaio_done;	/* (a) done queue for process */
297 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
298 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* (a) job queue for process */
299 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* (a) buffer job queue for process */
300 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;  /* (a) queue for aios waiting on sockets,
301 						 *  NOT USED YET.
302 						 */
303 	TAILQ_HEAD(,aiocblist) kaio_syncqueue;	/* (a) queue for aio_fsync */
304 	struct	task	kaio_task;	/* (*) task to kick aio threads */
305 };
306 
307 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
308 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
309 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
310 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
311 
312 #define KAIO_RUNDOWN	0x1	/* process is being run down */
313 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
314 
315 /*
316  * Operations used to interact with userland aio control blocks.
317  * Different ABIs provide their own operations.
318  */
319 struct aiocb_ops {
320 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
321 	long	(*fetch_status)(struct aiocb *ujob);
322 	long	(*fetch_error)(struct aiocb *ujob);
323 	int	(*store_status)(struct aiocb *ujob, long status);
324 	int	(*store_error)(struct aiocb *ujob, long error);
325 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
326 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
327 };
328 
329 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* (c) Idle daemons */
330 static struct sema aio_newproc_sem;
331 static struct mtx aio_job_mtx;
332 static struct mtx aio_sock_mtx;
333 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* (c) Async job list */
334 static struct unrhdr *aiod_unr;
335 
336 void		aio_init_aioinfo(struct proc *p);
337 static int	aio_onceonly(void);
338 static int	aio_free_entry(struct aiocblist *aiocbe);
339 static void	aio_process(struct aiocblist *aiocbe);
340 static int	aio_newproc(int *);
341 int		aio_aqueue(struct thread *td, struct aiocb *job,
342 			struct aioliojob *lio, int type, struct aiocb_ops *ops);
343 static void	aio_physwakeup(struct buf *bp);
344 static void	aio_proc_rundown(void *arg, struct proc *p);
345 static void	aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
346 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
347 static void	biohelper(void *, int);
348 static void	aio_daemon(void *param);
349 static void	aio_swake_cb(struct socket *, struct sockbuf *);
350 static int	aio_unload(void);
351 static void	aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
352 #define DONE_BUF	1
353 #define DONE_QUEUE	2
354 static int	aio_kick(struct proc *userp);
355 static void	aio_kick_nowait(struct proc *userp);
356 static void	aio_kick_helper(void *context, int pending);
357 static int	filt_aioattach(struct knote *kn);
358 static void	filt_aiodetach(struct knote *kn);
359 static int	filt_aio(struct knote *kn, long hint);
360 static int	filt_lioattach(struct knote *kn);
361 static void	filt_liodetach(struct knote *kn);
362 static int	filt_lio(struct knote *kn, long hint);
363 
364 /*
365  * Zones for:
366  * 	kaio	Per process async io info
367  *	aiop	async io thread data
368  *	aiocb	async io jobs
369  *	aiol	list io job pointer - internal to aio_suspend XXX
370  *	aiolio	list io jobs
371  */
372 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
373 
374 /* kqueue filters for aio */
375 static struct filterops aio_filtops = {
376 	.f_isfd = 0,
377 	.f_attach = filt_aioattach,
378 	.f_detach = filt_aiodetach,
379 	.f_event = filt_aio,
380 };
381 static struct filterops lio_filtops = {
382 	.f_isfd = 0,
383 	.f_attach = filt_lioattach,
384 	.f_detach = filt_liodetach,
385 	.f_event = filt_lio
386 };
387 
388 static eventhandler_tag exit_tag, exec_tag;
389 
390 TASKQUEUE_DEFINE_THREAD(aiod_bio);
391 
392 /*
393  * Main operations function for use as a kernel module.
394  */
395 static int
396 aio_modload(struct module *module, int cmd, void *arg)
397 {
398 	int error = 0;
399 
400 	switch (cmd) {
401 	case MOD_LOAD:
402 		aio_onceonly();
403 		break;
404 	case MOD_UNLOAD:
405 		error = aio_unload();
406 		break;
407 	case MOD_SHUTDOWN:
408 		break;
409 	default:
410 		error = EINVAL;
411 		break;
412 	}
413 	return (error);
414 }
415 
416 static moduledata_t aio_mod = {
417 	"aio",
418 	&aio_modload,
419 	NULL
420 };
421 
422 static struct syscall_helper_data aio_syscalls[] = {
423 	SYSCALL_INIT_HELPER(aio_cancel),
424 	SYSCALL_INIT_HELPER(aio_error),
425 	SYSCALL_INIT_HELPER(aio_fsync),
426 	SYSCALL_INIT_HELPER(aio_read),
427 	SYSCALL_INIT_HELPER(aio_return),
428 	SYSCALL_INIT_HELPER(aio_suspend),
429 	SYSCALL_INIT_HELPER(aio_waitcomplete),
430 	SYSCALL_INIT_HELPER(aio_write),
431 	SYSCALL_INIT_HELPER(lio_listio),
432 	SYSCALL_INIT_HELPER(oaio_read),
433 	SYSCALL_INIT_HELPER(oaio_write),
434 	SYSCALL_INIT_HELPER(olio_listio),
435 	SYSCALL_INIT_LAST
436 };
437 
438 #ifdef COMPAT_FREEBSD32
439 #include <sys/mount.h>
440 #include <sys/socket.h>
441 #include <compat/freebsd32/freebsd32.h>
442 #include <compat/freebsd32/freebsd32_proto.h>
443 #include <compat/freebsd32/freebsd32_signal.h>
444 #include <compat/freebsd32/freebsd32_syscall.h>
445 #include <compat/freebsd32/freebsd32_util.h>
446 
447 static struct syscall_helper_data aio32_syscalls[] = {
448 	SYSCALL32_INIT_HELPER(freebsd32_aio_return),
449 	SYSCALL32_INIT_HELPER(freebsd32_aio_suspend),
450 	SYSCALL32_INIT_HELPER(freebsd32_aio_cancel),
451 	SYSCALL32_INIT_HELPER(freebsd32_aio_error),
452 	SYSCALL32_INIT_HELPER(freebsd32_aio_fsync),
453 	SYSCALL32_INIT_HELPER(freebsd32_aio_read),
454 	SYSCALL32_INIT_HELPER(freebsd32_aio_write),
455 	SYSCALL32_INIT_HELPER(freebsd32_aio_waitcomplete),
456 	SYSCALL32_INIT_HELPER(freebsd32_lio_listio),
457 	SYSCALL32_INIT_HELPER(freebsd32_oaio_read),
458 	SYSCALL32_INIT_HELPER(freebsd32_oaio_write),
459 	SYSCALL32_INIT_HELPER(freebsd32_olio_listio),
460 	SYSCALL_INIT_LAST
461 };
462 #endif
463 
464 DECLARE_MODULE(aio, aio_mod,
465 	SI_SUB_VFS, SI_ORDER_ANY);
466 MODULE_VERSION(aio, 1);
467 
468 /*
469  * Startup initialization
470  */
471 static int
472 aio_onceonly(void)
473 {
474 	int error;
475 
476 	/* XXX: should probably just use so->callback */
477 	aio_swake = &aio_swake_cb;
478 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
479 	    EVENTHANDLER_PRI_ANY);
480 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
481 	    EVENTHANDLER_PRI_ANY);
482 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
483 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
484 	TAILQ_INIT(&aio_freeproc);
485 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
486 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
487 	mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
488 	TAILQ_INIT(&aio_jobs);
489 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
490 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
491 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
492 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
493 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
494 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
495 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
496 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
497 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
498 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
499 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
500 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
501 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
502 	jobrefid = 1;
503 	async_io_version = _POSIX_VERSION;
504 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
505 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
506 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
507 
508 	error = syscall_helper_register(aio_syscalls);
509 	if (error)
510 		return (error);
511 #ifdef COMPAT_FREEBSD32
512 	error = syscall32_helper_register(aio32_syscalls);
513 	if (error)
514 		return (error);
515 #endif
516 	return (0);
517 }
518 
519 /*
520  * Callback for unload of AIO when used as a module.
521  */
522 static int
523 aio_unload(void)
524 {
525 	int error;
526 
527 	/*
528 	 * XXX: no unloads by default, it's too dangerous.
529 	 * perhaps we could do it if locked out callers and then
530 	 * did an aio_proc_rundown() on each process.
531 	 *
532 	 * jhb: aio_proc_rundown() needs to run on curproc though,
533 	 * so I don't think that would fly.
534 	 */
535 	if (!unloadable)
536 		return (EOPNOTSUPP);
537 
538 #ifdef COMPAT_FREEBSD32
539 	syscall32_helper_unregister(aio32_syscalls);
540 #endif
541 	syscall_helper_unregister(aio_syscalls);
542 
543 	error = kqueue_del_filteropts(EVFILT_AIO);
544 	if (error)
545 		return error;
546 	error = kqueue_del_filteropts(EVFILT_LIO);
547 	if (error)
548 		return error;
549 	async_io_version = 0;
550 	aio_swake = NULL;
551 	taskqueue_free(taskqueue_aiod_bio);
552 	delete_unrhdr(aiod_unr);
553 	uma_zdestroy(kaio_zone);
554 	uma_zdestroy(aiop_zone);
555 	uma_zdestroy(aiocb_zone);
556 	uma_zdestroy(aiol_zone);
557 	uma_zdestroy(aiolio_zone);
558 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
559 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
560 	mtx_destroy(&aio_job_mtx);
561 	mtx_destroy(&aio_sock_mtx);
562 	sema_destroy(&aio_newproc_sem);
563 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
564 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
565 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
566 	return (0);
567 }
568 
569 /*
570  * Init the per-process aioinfo structure.  The aioinfo limits are set
571  * per-process for user limit (resource) management.
572  */
573 void
574 aio_init_aioinfo(struct proc *p)
575 {
576 	struct kaioinfo *ki;
577 
578 	ki = uma_zalloc(kaio_zone, M_WAITOK);
579 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
580 	ki->kaio_flags = 0;
581 	ki->kaio_maxactive_count = max_aio_per_proc;
582 	ki->kaio_active_count = 0;
583 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
584 	ki->kaio_count = 0;
585 	ki->kaio_ballowed_count = max_buf_aio;
586 	ki->kaio_buffer_count = 0;
587 	TAILQ_INIT(&ki->kaio_all);
588 	TAILQ_INIT(&ki->kaio_done);
589 	TAILQ_INIT(&ki->kaio_jobqueue);
590 	TAILQ_INIT(&ki->kaio_bufqueue);
591 	TAILQ_INIT(&ki->kaio_liojoblist);
592 	TAILQ_INIT(&ki->kaio_sockqueue);
593 	TAILQ_INIT(&ki->kaio_syncqueue);
594 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
595 	PROC_LOCK(p);
596 	if (p->p_aioinfo == NULL) {
597 		p->p_aioinfo = ki;
598 		PROC_UNLOCK(p);
599 	} else {
600 		PROC_UNLOCK(p);
601 		mtx_destroy(&ki->kaio_mtx);
602 		uma_zfree(kaio_zone, ki);
603 	}
604 
605 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
606 		aio_newproc(NULL);
607 }
608 
609 static int
610 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
611 {
612 	int ret = 0;
613 
614 	PROC_LOCK(p);
615 	if (!KSI_ONQ(ksi)) {
616 		ksi->ksi_code = SI_ASYNCIO;
617 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
618 		ret = psignal_event(p, sigev, ksi);
619 	}
620 	PROC_UNLOCK(p);
621 	return (ret);
622 }
623 
624 /*
625  * Free a job entry.  Wait for completion if it is currently active, but don't
626  * delay forever.  If we delay, we return a flag that says that we have to
627  * restart the queue scan.
628  */
629 static int
630 aio_free_entry(struct aiocblist *aiocbe)
631 {
632 	struct kaioinfo *ki;
633 	struct aioliojob *lj;
634 	struct proc *p;
635 
636 	p = aiocbe->userproc;
637 	MPASS(curproc == p);
638 	ki = p->p_aioinfo;
639 	MPASS(ki != NULL);
640 
641 	AIO_LOCK_ASSERT(ki, MA_OWNED);
642 	MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
643 
644 	atomic_subtract_int(&num_queue_count, 1);
645 
646 	ki->kaio_count--;
647 	MPASS(ki->kaio_count >= 0);
648 
649 	TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
650 	TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
651 
652 	lj = aiocbe->lio;
653 	if (lj) {
654 		lj->lioj_count--;
655 		lj->lioj_finished_count--;
656 
657 		if (lj->lioj_count == 0) {
658 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
659 			/* lio is going away, we need to destroy any knotes */
660 			knlist_delete(&lj->klist, curthread, 1);
661 			PROC_LOCK(p);
662 			sigqueue_take(&lj->lioj_ksi);
663 			PROC_UNLOCK(p);
664 			uma_zfree(aiolio_zone, lj);
665 		}
666 	}
667 
668 	/* aiocbe is going away, we need to destroy any knotes */
669 	knlist_delete(&aiocbe->klist, curthread, 1);
670 	PROC_LOCK(p);
671 	sigqueue_take(&aiocbe->ksi);
672 	PROC_UNLOCK(p);
673 
674 	MPASS(aiocbe->bp == NULL);
675 	aiocbe->jobstate = JOBST_NULL;
676 	AIO_UNLOCK(ki);
677 
678 	/*
679 	 * The thread argument here is used to find the owning process
680 	 * and is also passed to fo_close() which may pass it to various
681 	 * places such as devsw close() routines.  Because of that, we
682 	 * need a thread pointer from the process owning the job that is
683 	 * persistent and won't disappear out from under us or move to
684 	 * another process.
685 	 *
686 	 * Currently, all the callers of this function call it to remove
687 	 * an aiocblist from the current process' job list either via a
688 	 * syscall or due to the current process calling exit() or
689 	 * execve().  Thus, we know that p == curproc.  We also know that
690 	 * curthread can't exit since we are curthread.
691 	 *
692 	 * Therefore, we use curthread as the thread to pass to
693 	 * knlist_delete().  This does mean that it is possible for the
694 	 * thread pointer at close time to differ from the thread pointer
695 	 * at open time, but this is already true of file descriptors in
696 	 * a multithreaded process.
697 	 */
698 	fdrop(aiocbe->fd_file, curthread);
699 	crfree(aiocbe->cred);
700 	uma_zfree(aiocb_zone, aiocbe);
701 	AIO_LOCK(ki);
702 
703 	return (0);
704 }
705 
706 static void
707 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
708 {
709    	aio_proc_rundown(arg, p);
710 }
711 
712 /*
713  * Rundown the jobs for a given process.
714  */
715 static void
716 aio_proc_rundown(void *arg, struct proc *p)
717 {
718 	struct kaioinfo *ki;
719 	struct aioliojob *lj;
720 	struct aiocblist *cbe, *cbn;
721 	struct file *fp;
722 	struct socket *so;
723 	int remove;
724 
725 	KASSERT(curthread->td_proc == p,
726 	    ("%s: called on non-curproc", __func__));
727 	ki = p->p_aioinfo;
728 	if (ki == NULL)
729 		return;
730 
731 	AIO_LOCK(ki);
732 	ki->kaio_flags |= KAIO_RUNDOWN;
733 
734 restart:
735 
736 	/*
737 	 * Try to cancel all pending requests. This code simulates
738 	 * aio_cancel on all pending I/O requests.
739 	 */
740 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
741 		remove = 0;
742 		mtx_lock(&aio_job_mtx);
743 		if (cbe->jobstate == JOBST_JOBQGLOBAL) {
744 			TAILQ_REMOVE(&aio_jobs, cbe, list);
745 			remove = 1;
746 		} else if (cbe->jobstate == JOBST_JOBQSOCK) {
747 			fp = cbe->fd_file;
748 			MPASS(fp->f_type == DTYPE_SOCKET);
749 			so = fp->f_data;
750 			TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
751 			remove = 1;
752 		} else if (cbe->jobstate == JOBST_JOBQSYNC) {
753 			TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
754 			remove = 1;
755 		}
756 		mtx_unlock(&aio_job_mtx);
757 
758 		if (remove) {
759 			cbe->jobstate = JOBST_JOBFINISHED;
760 			cbe->uaiocb._aiocb_private.status = -1;
761 			cbe->uaiocb._aiocb_private.error = ECANCELED;
762 			TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
763 			aio_bio_done_notify(p, cbe, DONE_QUEUE);
764 		}
765 	}
766 
767 	/* Wait for all running I/O to be finished */
768 	if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
769 	    TAILQ_FIRST(&ki->kaio_jobqueue)) {
770 		ki->kaio_flags |= KAIO_WAKEUP;
771 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
772 		goto restart;
773 	}
774 
775 	/* Free all completed I/O requests. */
776 	while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
777 		aio_free_entry(cbe);
778 
779 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
780 		if (lj->lioj_count == 0) {
781 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
782 			knlist_delete(&lj->klist, curthread, 1);
783 			PROC_LOCK(p);
784 			sigqueue_take(&lj->lioj_ksi);
785 			PROC_UNLOCK(p);
786 			uma_zfree(aiolio_zone, lj);
787 		} else {
788 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
789 			    lj->lioj_count, lj->lioj_finished_count);
790 		}
791 	}
792 	AIO_UNLOCK(ki);
793 	taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
794 	mtx_destroy(&ki->kaio_mtx);
795 	uma_zfree(kaio_zone, ki);
796 	p->p_aioinfo = NULL;
797 }
798 
799 /*
800  * Select a job to run (called by an AIO daemon).
801  */
802 static struct aiocblist *
803 aio_selectjob(struct aiothreadlist *aiop)
804 {
805 	struct aiocblist *aiocbe;
806 	struct kaioinfo *ki;
807 	struct proc *userp;
808 
809 	mtx_assert(&aio_job_mtx, MA_OWNED);
810 	TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
811 		userp = aiocbe->userproc;
812 		ki = userp->p_aioinfo;
813 
814 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
815 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
816 			/* Account for currently active jobs. */
817 			ki->kaio_active_count++;
818 			aiocbe->jobstate = JOBST_JOBRUNNING;
819 			break;
820 		}
821 	}
822 	return (aiocbe);
823 }
824 
825 /*
826  *  Move all data to a permanent storage device, this code
827  *  simulates fsync syscall.
828  */
829 static int
830 aio_fsync_vnode(struct thread *td, struct vnode *vp)
831 {
832 	struct mount *mp;
833 	int vfslocked;
834 	int error;
835 
836 	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
837 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
838 		goto drop;
839 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
840 	if (vp->v_object != NULL) {
841 		VM_OBJECT_LOCK(vp->v_object);
842 		vm_object_page_clean(vp->v_object, 0, 0, 0);
843 		VM_OBJECT_UNLOCK(vp->v_object);
844 	}
845 	error = VOP_FSYNC(vp, MNT_WAIT, td);
846 
847 	VOP_UNLOCK(vp, 0);
848 	vn_finished_write(mp);
849 drop:
850 	VFS_UNLOCK_GIANT(vfslocked);
851 	return (error);
852 }
853 
854 /*
855  * The AIO processing activity.  This is the code that does the I/O request for
856  * the non-physio version of the operations.  The normal vn operations are used,
857  * and this code should work in all instances for every type of file, including
858  * 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(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 	td = curthread;
878 	td_savedcred = td->td_ucred;
879 	td->td_ucred = aiocbe->cred;
880 	cb = &aiocbe->uaiocb;
881 	fp = aiocbe->fd_file;
882 
883 	if (cb->aio_lio_opcode == LIO_SYNC) {
884 		error = 0;
885 		cnt = 0;
886 		if (fp->f_vnode != NULL)
887 			error = aio_fsync_vnode(td, fp->f_vnode);
888 		cb->_aiocb_private.error = error;
889 		cb->_aiocb_private.status = 0;
890 		td->td_ucred = td_savedcred;
891 		return;
892 	}
893 
894 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
895 	aiov.iov_len = cb->aio_nbytes;
896 
897 	auio.uio_iov = &aiov;
898 	auio.uio_iovcnt = 1;
899 	auio.uio_offset = cb->aio_offset;
900 	auio.uio_resid = cb->aio_nbytes;
901 	cnt = cb->aio_nbytes;
902 	auio.uio_segflg = UIO_USERSPACE;
903 	auio.uio_td = td;
904 
905 	inblock_st = td->td_ru.ru_inblock;
906 	oublock_st = td->td_ru.ru_oublock;
907 	/*
908 	 * aio_aqueue() acquires a reference to the file that is
909 	 * released in aio_free_entry().
910 	 */
911 	if (cb->aio_lio_opcode == LIO_READ) {
912 		auio.uio_rw = UIO_READ;
913 		if (auio.uio_resid == 0)
914 			error = 0;
915 		else
916 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
917 	} else {
918 		if (fp->f_type == DTYPE_VNODE)
919 			bwillwrite();
920 		auio.uio_rw = UIO_WRITE;
921 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
922 	}
923 	inblock_end = td->td_ru.ru_inblock;
924 	oublock_end = td->td_ru.ru_oublock;
925 
926 	aiocbe->inputcharge = inblock_end - inblock_st;
927 	aiocbe->outputcharge = oublock_end - oublock_st;
928 
929 	if ((error) && (auio.uio_resid != cnt)) {
930 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
931 			error = 0;
932 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
933 			int sigpipe = 1;
934 			if (fp->f_type == DTYPE_SOCKET) {
935 				so = fp->f_data;
936 				if (so->so_options & SO_NOSIGPIPE)
937 					sigpipe = 0;
938 			}
939 			if (sigpipe) {
940 				PROC_LOCK(aiocbe->userproc);
941 				psignal(aiocbe->userproc, SIGPIPE);
942 				PROC_UNLOCK(aiocbe->userproc);
943 			}
944 		}
945 	}
946 
947 	cnt -= auio.uio_resid;
948 	cb->_aiocb_private.error = error;
949 	cb->_aiocb_private.status = cnt;
950 	td->td_ucred = td_savedcred;
951 }
952 
953 static void
954 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
955 {
956 	struct aioliojob *lj;
957 	struct kaioinfo *ki;
958 	struct aiocblist *scb, *scbn;
959 	int lj_done;
960 
961 	ki = userp->p_aioinfo;
962 	AIO_LOCK_ASSERT(ki, MA_OWNED);
963 	lj = aiocbe->lio;
964 	lj_done = 0;
965 	if (lj) {
966 		lj->lioj_finished_count++;
967 		if (lj->lioj_count == lj->lioj_finished_count)
968 			lj_done = 1;
969 	}
970 	if (type == DONE_QUEUE) {
971 		aiocbe->jobflags |= AIOCBLIST_DONE;
972 	} else {
973 		aiocbe->jobflags |= AIOCBLIST_BUFDONE;
974 	}
975 	TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
976 	aiocbe->jobstate = JOBST_JOBFINISHED;
977 
978 	if (ki->kaio_flags & KAIO_RUNDOWN)
979 		goto notification_done;
980 
981 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
982 	    aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
983 		aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
984 
985 	KNOTE_LOCKED(&aiocbe->klist, 1);
986 
987 	if (lj_done) {
988 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
989 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
990 			KNOTE_LOCKED(&lj->klist, 1);
991 		}
992 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
993 		    == LIOJ_SIGNAL
994 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
995 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
996 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
997 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
998 		}
999 	}
1000 
1001 notification_done:
1002 	if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
1003 		TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
1004 			if (aiocbe->fd_file == scb->fd_file &&
1005 			    aiocbe->seqno < scb->seqno) {
1006 				if (--scb->pending == 0) {
1007 					mtx_lock(&aio_job_mtx);
1008 					scb->jobstate = JOBST_JOBQGLOBAL;
1009 					TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
1010 					TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
1011 					aio_kick_nowait(userp);
1012 					mtx_unlock(&aio_job_mtx);
1013 				}
1014 			}
1015 		}
1016 	}
1017 	if (ki->kaio_flags & KAIO_WAKEUP) {
1018 		ki->kaio_flags &= ~KAIO_WAKEUP;
1019 		wakeup(&userp->p_aioinfo);
1020 	}
1021 }
1022 
1023 /*
1024  * The AIO daemon, most of the actual work is done in aio_process,
1025  * but the setup (and address space mgmt) is done in this routine.
1026  */
1027 static void
1028 aio_daemon(void *_id)
1029 {
1030 	struct aiocblist *aiocbe;
1031 	struct aiothreadlist *aiop;
1032 	struct kaioinfo *ki;
1033 	struct proc *curcp, *mycp, *userp;
1034 	struct vmspace *myvm, *tmpvm;
1035 	struct thread *td = curthread;
1036 	int id = (intptr_t)_id;
1037 
1038 	/*
1039 	 * Local copies of curproc (cp) and vmspace (myvm)
1040 	 */
1041 	mycp = td->td_proc;
1042 	myvm = mycp->p_vmspace;
1043 
1044 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
1045 
1046 	/*
1047 	 * Allocate and ready the aio control info.  There is one aiop structure
1048 	 * per daemon.
1049 	 */
1050 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1051 	aiop->aiothread = td;
1052 	aiop->aiothreadflags = 0;
1053 
1054 	/* The daemon resides in its own pgrp. */
1055 	setsid(td, NULL);
1056 
1057 	/*
1058 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1059 	 * and creating too many daemons.)
1060 	 */
1061 	sema_post(&aio_newproc_sem);
1062 
1063 	mtx_lock(&aio_job_mtx);
1064 	for (;;) {
1065 		/*
1066 		 * curcp is the current daemon process context.
1067 		 * userp is the current user process context.
1068 		 */
1069 		curcp = mycp;
1070 
1071 		/*
1072 		 * Take daemon off of free queue
1073 		 */
1074 		if (aiop->aiothreadflags & AIOP_FREE) {
1075 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1076 			aiop->aiothreadflags &= ~AIOP_FREE;
1077 		}
1078 
1079 		/*
1080 		 * Check for jobs.
1081 		 */
1082 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1083 			mtx_unlock(&aio_job_mtx);
1084 			userp = aiocbe->userproc;
1085 
1086 			/*
1087 			 * Connect to process address space for user program.
1088 			 */
1089 			if (userp != curcp) {
1090 				/*
1091 				 * Save the current address space that we are
1092 				 * connected to.
1093 				 */
1094 				tmpvm = mycp->p_vmspace;
1095 
1096 				/*
1097 				 * Point to the new user address space, and
1098 				 * refer to it.
1099 				 */
1100 				mycp->p_vmspace = userp->p_vmspace;
1101 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1102 
1103 				/* Activate the new mapping. */
1104 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1105 
1106 				/*
1107 				 * If the old address space wasn't the daemons
1108 				 * own address space, then we need to remove the
1109 				 * daemon's reference from the other process
1110 				 * that it was acting on behalf of.
1111 				 */
1112 				if (tmpvm != myvm) {
1113 					vmspace_free(tmpvm);
1114 				}
1115 				curcp = userp;
1116 			}
1117 
1118 			ki = userp->p_aioinfo;
1119 
1120 			/* Do the I/O function. */
1121 			aio_process(aiocbe);
1122 
1123 			mtx_lock(&aio_job_mtx);
1124 			/* Decrement the active job count. */
1125 			ki->kaio_active_count--;
1126 			mtx_unlock(&aio_job_mtx);
1127 
1128 			AIO_LOCK(ki);
1129 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1130 			aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1131 			AIO_UNLOCK(ki);
1132 
1133 			mtx_lock(&aio_job_mtx);
1134 		}
1135 
1136 		/*
1137 		 * Disconnect from user address space.
1138 		 */
1139 		if (curcp != mycp) {
1140 
1141 			mtx_unlock(&aio_job_mtx);
1142 
1143 			/* Get the user address space to disconnect from. */
1144 			tmpvm = mycp->p_vmspace;
1145 
1146 			/* Get original address space for daemon. */
1147 			mycp->p_vmspace = myvm;
1148 
1149 			/* Activate the daemon's address space. */
1150 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1151 #ifdef DIAGNOSTIC
1152 			if (tmpvm == myvm) {
1153 				printf("AIOD: vmspace problem -- %d\n",
1154 				    mycp->p_pid);
1155 			}
1156 #endif
1157 			/* Remove our vmspace reference. */
1158 			vmspace_free(tmpvm);
1159 
1160 			curcp = mycp;
1161 
1162 			mtx_lock(&aio_job_mtx);
1163 			/*
1164 			 * We have to restart to avoid race, we only sleep if
1165 			 * no job can be selected, that should be
1166 			 * curcp == mycp.
1167 			 */
1168 			continue;
1169 		}
1170 
1171 		mtx_assert(&aio_job_mtx, MA_OWNED);
1172 
1173 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1174 		aiop->aiothreadflags |= AIOP_FREE;
1175 
1176 		/*
1177 		 * If daemon is inactive for a long time, allow it to exit,
1178 		 * thereby freeing resources.
1179 		 */
1180 		if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1181 		    aiod_lifetime)) {
1182 			if (TAILQ_EMPTY(&aio_jobs)) {
1183 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1184 				    (num_aio_procs > target_aio_procs)) {
1185 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1186 					num_aio_procs--;
1187 					mtx_unlock(&aio_job_mtx);
1188 					uma_zfree(aiop_zone, aiop);
1189 					free_unr(aiod_unr, id);
1190 #ifdef DIAGNOSTIC
1191 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1192 						printf("AIOD: bad vm refcnt for"
1193 						    " exiting daemon: %d\n",
1194 						    mycp->p_vmspace->vm_refcnt);
1195 					}
1196 #endif
1197 					kproc_exit(0);
1198 				}
1199 			}
1200 		}
1201 	}
1202 	mtx_unlock(&aio_job_mtx);
1203 	panic("shouldn't be here\n");
1204 }
1205 
1206 /*
1207  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1208  * AIO daemon modifies its environment itself.
1209  */
1210 static int
1211 aio_newproc(int *start)
1212 {
1213 	int error;
1214 	struct proc *p;
1215 	int id;
1216 
1217 	id = alloc_unr(aiod_unr);
1218 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1219 		RFNOWAIT, 0, "aiod%d", id);
1220 	if (error == 0) {
1221 		/*
1222 		 * Wait until daemon is started.
1223 		 */
1224 		sema_wait(&aio_newproc_sem);
1225 		mtx_lock(&aio_job_mtx);
1226 		num_aio_procs++;
1227 		if (start != NULL)
1228 			(*start)--;
1229 		mtx_unlock(&aio_job_mtx);
1230 	} else {
1231 		free_unr(aiod_unr, id);
1232 	}
1233 	return (error);
1234 }
1235 
1236 /*
1237  * Try the high-performance, low-overhead physio method for eligible
1238  * VCHR devices.  This method doesn't use an aio helper thread, and
1239  * thus has very low overhead.
1240  *
1241  * Assumes that the caller, aio_aqueue(), has incremented the file
1242  * structure's reference count, preventing its deallocation for the
1243  * duration of this call.
1244  */
1245 static int
1246 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1247 {
1248 	struct aiocb *cb;
1249 	struct file *fp;
1250 	struct buf *bp;
1251 	struct vnode *vp;
1252 	struct kaioinfo *ki;
1253 	struct aioliojob *lj;
1254 	int error;
1255 
1256 	cb = &aiocbe->uaiocb;
1257 	fp = aiocbe->fd_file;
1258 
1259 	if (fp->f_type != DTYPE_VNODE)
1260 		return (-1);
1261 
1262 	vp = fp->f_vnode;
1263 
1264 	/*
1265 	 * If its not a disk, we don't want to return a positive error.
1266 	 * It causes the aio code to not fall through to try the thread
1267 	 * way when you're talking to a regular file.
1268 	 */
1269 	if (!vn_isdisk(vp, &error)) {
1270 		if (error == ENOTBLK)
1271 			return (-1);
1272 		else
1273 			return (error);
1274 	}
1275 
1276 	if (vp->v_bufobj.bo_bsize == 0)
1277 		return (-1);
1278 
1279  	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1280 		return (-1);
1281 
1282 	if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1283 		return (-1);
1284 
1285 	if (cb->aio_nbytes >
1286 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1287 		return (-1);
1288 
1289 	ki = p->p_aioinfo;
1290 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1291 		return (-1);
1292 
1293 	/* Create and build a buffer header for a transfer. */
1294 	bp = (struct buf *)getpbuf(NULL);
1295 	BUF_KERNPROC(bp);
1296 
1297 	AIO_LOCK(ki);
1298 	ki->kaio_count++;
1299 	ki->kaio_buffer_count++;
1300 	lj = aiocbe->lio;
1301 	if (lj)
1302 		lj->lioj_count++;
1303 	AIO_UNLOCK(ki);
1304 
1305 	/*
1306 	 * Get a copy of the kva from the physical buffer.
1307 	 */
1308 	error = 0;
1309 
1310 	bp->b_bcount = cb->aio_nbytes;
1311 	bp->b_bufsize = cb->aio_nbytes;
1312 	bp->b_iodone = aio_physwakeup;
1313 	bp->b_saveaddr = bp->b_data;
1314 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1315 	bp->b_offset = cb->aio_offset;
1316 	bp->b_iooffset = cb->aio_offset;
1317 	bp->b_blkno = btodb(cb->aio_offset);
1318 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1319 
1320 	/*
1321 	 * Bring buffer into kernel space.
1322 	 */
1323 	if (vmapbuf(bp) < 0) {
1324 		error = EFAULT;
1325 		goto doerror;
1326 	}
1327 
1328 	AIO_LOCK(ki);
1329 	aiocbe->bp = bp;
1330 	bp->b_caller1 = (void *)aiocbe;
1331 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1332 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1333 	aiocbe->jobstate = JOBST_JOBQBUF;
1334 	cb->_aiocb_private.status = cb->aio_nbytes;
1335 	AIO_UNLOCK(ki);
1336 
1337 	atomic_add_int(&num_queue_count, 1);
1338 	atomic_add_int(&num_buf_aio, 1);
1339 
1340 	bp->b_error = 0;
1341 
1342 	TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1343 
1344 	/* Perform transfer. */
1345 	dev_strategy(vp->v_rdev, bp);
1346 	return (0);
1347 
1348 doerror:
1349 	AIO_LOCK(ki);
1350 	ki->kaio_count--;
1351 	ki->kaio_buffer_count--;
1352 	if (lj)
1353 		lj->lioj_count--;
1354 	aiocbe->bp = NULL;
1355 	AIO_UNLOCK(ki);
1356 	relpbuf(bp, NULL);
1357 	return (error);
1358 }
1359 
1360 /*
1361  * Wake up aio requests that may be serviceable now.
1362  */
1363 static void
1364 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1365 {
1366 	struct aiocblist *cb, *cbn;
1367 	int opcode;
1368 
1369 	SOCKBUF_LOCK_ASSERT(sb);
1370 	if (sb == &so->so_snd)
1371 		opcode = LIO_WRITE;
1372 	else
1373 		opcode = LIO_READ;
1374 
1375 	sb->sb_flags &= ~SB_AIO;
1376 	mtx_lock(&aio_job_mtx);
1377 	TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1378 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1379 			if (cb->jobstate != JOBST_JOBQSOCK)
1380 				panic("invalid queue value");
1381 			/* XXX
1382 			 * We don't have actual sockets backend yet,
1383 			 * so we simply move the requests to the generic
1384 			 * file I/O backend.
1385 			 */
1386 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1387 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1388 			aio_kick_nowait(cb->userproc);
1389 		}
1390 	}
1391 	mtx_unlock(&aio_job_mtx);
1392 }
1393 
1394 static int
1395 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1396 {
1397 
1398 	/*
1399 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1400 	 * supported by AIO with the old sigevent structure.
1401 	 */
1402 	nsig->sigev_notify = osig->sigev_notify;
1403 	switch (nsig->sigev_notify) {
1404 	case SIGEV_NONE:
1405 		break;
1406 	case SIGEV_SIGNAL:
1407 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1408 		break;
1409 	case SIGEV_KEVENT:
1410 		nsig->sigev_notify_kqueue =
1411 		    osig->__sigev_u.__sigev_notify_kqueue;
1412 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1413 		break;
1414 	default:
1415 		return (EINVAL);
1416 	}
1417 	return (0);
1418 }
1419 
1420 static int
1421 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1422 {
1423 	struct oaiocb *ojob;
1424 	int error;
1425 
1426 	bzero(kjob, sizeof(struct aiocb));
1427 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1428 	if (error)
1429 		return (error);
1430 	ojob = (struct oaiocb *)kjob;
1431 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1432 }
1433 
1434 static int
1435 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1436 {
1437 
1438 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1439 }
1440 
1441 static long
1442 aiocb_fetch_status(struct aiocb *ujob)
1443 {
1444 
1445 	return (fuword(&ujob->_aiocb_private.status));
1446 }
1447 
1448 static long
1449 aiocb_fetch_error(struct aiocb *ujob)
1450 {
1451 
1452 	return (fuword(&ujob->_aiocb_private.error));
1453 }
1454 
1455 static int
1456 aiocb_store_status(struct aiocb *ujob, long status)
1457 {
1458 
1459 	return (suword(&ujob->_aiocb_private.status, status));
1460 }
1461 
1462 static int
1463 aiocb_store_error(struct aiocb *ujob, long error)
1464 {
1465 
1466 	return (suword(&ujob->_aiocb_private.error, error));
1467 }
1468 
1469 static int
1470 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1471 {
1472 
1473 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1474 }
1475 
1476 static int
1477 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1478 {
1479 
1480 	return (suword(ujobp, (long)ujob));
1481 }
1482 
1483 static struct aiocb_ops aiocb_ops = {
1484 	.copyin = aiocb_copyin,
1485 	.fetch_status = aiocb_fetch_status,
1486 	.fetch_error = aiocb_fetch_error,
1487 	.store_status = aiocb_store_status,
1488 	.store_error = aiocb_store_error,
1489 	.store_kernelinfo = aiocb_store_kernelinfo,
1490 	.store_aiocb = aiocb_store_aiocb,
1491 };
1492 
1493 static struct aiocb_ops aiocb_ops_osigevent = {
1494 	.copyin = aiocb_copyin_old_sigevent,
1495 	.fetch_status = aiocb_fetch_status,
1496 	.fetch_error = aiocb_fetch_error,
1497 	.store_status = aiocb_store_status,
1498 	.store_error = aiocb_store_error,
1499 	.store_kernelinfo = aiocb_store_kernelinfo,
1500 	.store_aiocb = aiocb_store_aiocb,
1501 };
1502 
1503 /*
1504  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1505  * technique is done in this code.
1506  */
1507 int
1508 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1509 	int type, struct aiocb_ops *ops)
1510 {
1511 	struct proc *p = td->td_proc;
1512 	struct file *fp;
1513 	struct socket *so;
1514 	struct aiocblist *aiocbe, *cb;
1515 	struct kaioinfo *ki;
1516 	struct kevent kev;
1517 	struct sockbuf *sb;
1518 	int opcode;
1519 	int error;
1520 	int fd, kqfd;
1521 	int jid;
1522 
1523 	if (p->p_aioinfo == NULL)
1524 		aio_init_aioinfo(p);
1525 
1526 	ki = p->p_aioinfo;
1527 
1528 	ops->store_status(job, -1);
1529 	ops->store_error(job, 0);
1530 	ops->store_kernelinfo(job, -1);
1531 
1532 	if (num_queue_count >= max_queue_count ||
1533 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1534 		ops->store_error(job, EAGAIN);
1535 		return (EAGAIN);
1536 	}
1537 
1538 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1539 	aiocbe->inputcharge = 0;
1540 	aiocbe->outputcharge = 0;
1541 	knlist_init_mtx(&aiocbe->klist, AIO_MTX(ki));
1542 
1543 	error = ops->copyin(job, &aiocbe->uaiocb);
1544 	if (error) {
1545 		ops->store_error(job, error);
1546 		uma_zfree(aiocb_zone, aiocbe);
1547 		return (error);
1548 	}
1549 
1550 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1551 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1552 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1553 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1554 		ops->store_error(job, EINVAL);
1555 		uma_zfree(aiocb_zone, aiocbe);
1556 		return (EINVAL);
1557 	}
1558 
1559 	if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1560 	     aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1561 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1562 		uma_zfree(aiocb_zone, aiocbe);
1563 		return (EINVAL);
1564 	}
1565 
1566 	ksiginfo_init(&aiocbe->ksi);
1567 
1568 	/* Save userspace address of the job info. */
1569 	aiocbe->uuaiocb = job;
1570 
1571 	/* Get the opcode. */
1572 	if (type != LIO_NOP)
1573 		aiocbe->uaiocb.aio_lio_opcode = type;
1574 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1575 
1576 	/* Fetch the file object for the specified file descriptor. */
1577 	fd = aiocbe->uaiocb.aio_fildes;
1578 	switch (opcode) {
1579 	case LIO_WRITE:
1580 		error = fget_write(td, fd, &fp);
1581 		break;
1582 	case LIO_READ:
1583 		error = fget_read(td, fd, &fp);
1584 		break;
1585 	default:
1586 		error = fget(td, fd, &fp);
1587 	}
1588 	if (error) {
1589 		uma_zfree(aiocb_zone, aiocbe);
1590 		ops->store_error(job, error);
1591 		return (error);
1592 	}
1593 
1594 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1595 		error = EINVAL;
1596 		goto aqueue_fail;
1597 	}
1598 
1599 	if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1600 		error = EINVAL;
1601 		goto aqueue_fail;
1602 	}
1603 
1604 	aiocbe->fd_file = fp;
1605 
1606 	mtx_lock(&aio_job_mtx);
1607 	jid = jobrefid++;
1608 	aiocbe->seqno = jobseqno++;
1609 	mtx_unlock(&aio_job_mtx);
1610 	error = ops->store_kernelinfo(job, jid);
1611 	if (error) {
1612 		error = EINVAL;
1613 		goto aqueue_fail;
1614 	}
1615 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1616 
1617 	if (opcode == LIO_NOP) {
1618 		fdrop(fp, td);
1619 		uma_zfree(aiocb_zone, aiocbe);
1620 		return (0);
1621 	}
1622 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1623 	    (opcode != LIO_SYNC)) {
1624 		error = EINVAL;
1625 		goto aqueue_fail;
1626 	}
1627 
1628 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1629 		goto no_kqueue;
1630 	kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1631 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1632 	kev.filter = EVFILT_AIO;
1633 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1634 	kev.data = (intptr_t)aiocbe;
1635 	kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1636 	error = kqfd_register(kqfd, &kev, td, 1);
1637 aqueue_fail:
1638 	if (error) {
1639 		fdrop(fp, td);
1640 		uma_zfree(aiocb_zone, aiocbe);
1641 		ops->store_error(job, error);
1642 		goto done;
1643 	}
1644 no_kqueue:
1645 
1646 	ops->store_error(job, EINPROGRESS);
1647 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1648 	aiocbe->userproc = p;
1649 	aiocbe->cred = crhold(td->td_ucred);
1650 	aiocbe->jobflags = 0;
1651 	aiocbe->lio = lj;
1652 
1653 	if (opcode == LIO_SYNC)
1654 		goto queueit;
1655 
1656 	if (fp->f_type == DTYPE_SOCKET) {
1657 		/*
1658 		 * Alternate queueing for socket ops: Reach down into the
1659 		 * descriptor to get the socket data.  Then check to see if the
1660 		 * socket is ready to be read or written (based on the requested
1661 		 * operation).
1662 		 *
1663 		 * If it is not ready for io, then queue the aiocbe on the
1664 		 * socket, and set the flags so we get a call when sbnotify()
1665 		 * happens.
1666 		 *
1667 		 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1668 		 * and unlock the snd sockbuf for no reason.
1669 		 */
1670 		so = fp->f_data;
1671 		sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1672 		SOCKBUF_LOCK(sb);
1673 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1674 		    LIO_WRITE) && (!sowriteable(so)))) {
1675 			sb->sb_flags |= SB_AIO;
1676 
1677 			mtx_lock(&aio_job_mtx);
1678 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1679 			mtx_unlock(&aio_job_mtx);
1680 
1681 			AIO_LOCK(ki);
1682 			TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1683 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1684 			aiocbe->jobstate = JOBST_JOBQSOCK;
1685 			ki->kaio_count++;
1686 			if (lj)
1687 				lj->lioj_count++;
1688 			AIO_UNLOCK(ki);
1689 			SOCKBUF_UNLOCK(sb);
1690 			atomic_add_int(&num_queue_count, 1);
1691 			error = 0;
1692 			goto done;
1693 		}
1694 		SOCKBUF_UNLOCK(sb);
1695 	}
1696 
1697 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1698 		goto done;
1699 #if 0
1700 	if (error > 0) {
1701 		aiocbe->uaiocb._aiocb_private.error = error;
1702 		ops->store_error(job, error);
1703 		goto done;
1704 	}
1705 #endif
1706 queueit:
1707 	/* No buffer for daemon I/O. */
1708 	aiocbe->bp = NULL;
1709 	atomic_add_int(&num_queue_count, 1);
1710 
1711 	AIO_LOCK(ki);
1712 	ki->kaio_count++;
1713 	if (lj)
1714 		lj->lioj_count++;
1715 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1716 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1717 	if (opcode == LIO_SYNC) {
1718 		TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1719 			if (cb->fd_file == aiocbe->fd_file &&
1720 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1721 			    cb->seqno < aiocbe->seqno) {
1722 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1723 				aiocbe->pending++;
1724 			}
1725 		}
1726 		TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1727 			if (cb->fd_file == aiocbe->fd_file &&
1728 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1729 			    cb->seqno < aiocbe->seqno) {
1730 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1731 				aiocbe->pending++;
1732 			}
1733 		}
1734 		if (aiocbe->pending != 0) {
1735 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1736 			aiocbe->jobstate = JOBST_JOBQSYNC;
1737 			AIO_UNLOCK(ki);
1738 			goto done;
1739 		}
1740 	}
1741 	mtx_lock(&aio_job_mtx);
1742 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1743 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1744 	aio_kick_nowait(p);
1745 	mtx_unlock(&aio_job_mtx);
1746 	AIO_UNLOCK(ki);
1747 	error = 0;
1748 done:
1749 	return (error);
1750 }
1751 
1752 static void
1753 aio_kick_nowait(struct proc *userp)
1754 {
1755 	struct kaioinfo *ki = userp->p_aioinfo;
1756 	struct aiothreadlist *aiop;
1757 
1758 	mtx_assert(&aio_job_mtx, MA_OWNED);
1759 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1760 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1761 		aiop->aiothreadflags &= ~AIOP_FREE;
1762 		wakeup(aiop->aiothread);
1763 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1764 	    ((ki->kaio_active_count + num_aio_resv_start) <
1765 	    ki->kaio_maxactive_count)) {
1766 		taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1767 	}
1768 }
1769 
1770 static int
1771 aio_kick(struct proc *userp)
1772 {
1773 	struct kaioinfo *ki = userp->p_aioinfo;
1774 	struct aiothreadlist *aiop;
1775 	int error, ret = 0;
1776 
1777 	mtx_assert(&aio_job_mtx, MA_OWNED);
1778 retryproc:
1779 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1780 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1781 		aiop->aiothreadflags &= ~AIOP_FREE;
1782 		wakeup(aiop->aiothread);
1783 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1784 	    ((ki->kaio_active_count + num_aio_resv_start) <
1785 	    ki->kaio_maxactive_count)) {
1786 		num_aio_resv_start++;
1787 		mtx_unlock(&aio_job_mtx);
1788 		error = aio_newproc(&num_aio_resv_start);
1789 		mtx_lock(&aio_job_mtx);
1790 		if (error) {
1791 			num_aio_resv_start--;
1792 			goto retryproc;
1793 		}
1794 	} else {
1795 		ret = -1;
1796 	}
1797 	return (ret);
1798 }
1799 
1800 static void
1801 aio_kick_helper(void *context, int pending)
1802 {
1803 	struct proc *userp = context;
1804 
1805 	mtx_lock(&aio_job_mtx);
1806 	while (--pending >= 0) {
1807 		if (aio_kick(userp))
1808 			break;
1809 	}
1810 	mtx_unlock(&aio_job_mtx);
1811 }
1812 
1813 /*
1814  * Support the aio_return system call, as a side-effect, kernel resources are
1815  * released.
1816  */
1817 static int
1818 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1819 {
1820 	struct proc *p = td->td_proc;
1821 	struct aiocblist *cb;
1822 	struct kaioinfo *ki;
1823 	int status, error;
1824 
1825 	ki = p->p_aioinfo;
1826 	if (ki == NULL)
1827 		return (EINVAL);
1828 	AIO_LOCK(ki);
1829 	TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1830 		if (cb->uuaiocb == uaiocb)
1831 			break;
1832 	}
1833 	if (cb != NULL) {
1834 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
1835 		status = cb->uaiocb._aiocb_private.status;
1836 		error = cb->uaiocb._aiocb_private.error;
1837 		td->td_retval[0] = status;
1838 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1839 			td->td_ru.ru_oublock += cb->outputcharge;
1840 			cb->outputcharge = 0;
1841 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1842 			td->td_ru.ru_inblock += cb->inputcharge;
1843 			cb->inputcharge = 0;
1844 		}
1845 		aio_free_entry(cb);
1846 		AIO_UNLOCK(ki);
1847 		ops->store_error(uaiocb, error);
1848 		ops->store_status(uaiocb, status);
1849 	} else {
1850 		error = EINVAL;
1851 		AIO_UNLOCK(ki);
1852 	}
1853 	return (error);
1854 }
1855 
1856 int
1857 aio_return(struct thread *td, struct aio_return_args *uap)
1858 {
1859 
1860 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1861 }
1862 
1863 /*
1864  * Allow a process to wakeup when any of the I/O requests are completed.
1865  */
1866 static int
1867 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1868     struct timespec *ts)
1869 {
1870 	struct proc *p = td->td_proc;
1871 	struct timeval atv;
1872 	struct kaioinfo *ki;
1873 	struct aiocblist *cb, *cbfirst;
1874 	int error, i, timo;
1875 
1876 	timo = 0;
1877 	if (ts) {
1878 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1879 			return (EINVAL);
1880 
1881 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1882 		if (itimerfix(&atv))
1883 			return (EINVAL);
1884 		timo = tvtohz(&atv);
1885 	}
1886 
1887 	ki = p->p_aioinfo;
1888 	if (ki == NULL)
1889 		return (EAGAIN);
1890 
1891 	if (njoblist == 0)
1892 		return (0);
1893 
1894 	AIO_LOCK(ki);
1895 	for (;;) {
1896 		cbfirst = NULL;
1897 		error = 0;
1898 		TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1899 			for (i = 0; i < njoblist; i++) {
1900 				if (cb->uuaiocb == ujoblist[i]) {
1901 					if (cbfirst == NULL)
1902 						cbfirst = cb;
1903 					if (cb->jobstate == JOBST_JOBFINISHED)
1904 						goto RETURN;
1905 				}
1906 			}
1907 		}
1908 		/* All tasks were finished. */
1909 		if (cbfirst == NULL)
1910 			break;
1911 
1912 		ki->kaio_flags |= KAIO_WAKEUP;
1913 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1914 		    "aiospn", timo);
1915 		if (error == ERESTART)
1916 			error = EINTR;
1917 		if (error)
1918 			break;
1919 	}
1920 RETURN:
1921 	AIO_UNLOCK(ki);
1922 	return (error);
1923 }
1924 
1925 int
1926 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1927 {
1928 	struct timespec ts, *tsp;
1929 	struct aiocb **ujoblist;
1930 	int error;
1931 
1932 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1933 		return (EINVAL);
1934 
1935 	if (uap->timeout) {
1936 		/* Get timespec struct. */
1937 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1938 			return (error);
1939 		tsp = &ts;
1940 	} else
1941 		tsp = NULL;
1942 
1943 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1944 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1945 	if (error == 0)
1946 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1947 	uma_zfree(aiol_zone, ujoblist);
1948 	return (error);
1949 }
1950 
1951 /*
1952  * aio_cancel cancels any non-physio aio operations not currently in
1953  * progress.
1954  */
1955 int
1956 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1957 {
1958 	struct proc *p = td->td_proc;
1959 	struct kaioinfo *ki;
1960 	struct aiocblist *cbe, *cbn;
1961 	struct file *fp;
1962 	struct socket *so;
1963 	int error;
1964 	int remove;
1965 	int cancelled = 0;
1966 	int notcancelled = 0;
1967 	struct vnode *vp;
1968 
1969 	/* Lookup file object. */
1970 	error = fget(td, uap->fd, &fp);
1971 	if (error)
1972 		return (error);
1973 
1974 	ki = p->p_aioinfo;
1975 	if (ki == NULL)
1976 		goto done;
1977 
1978 	if (fp->f_type == DTYPE_VNODE) {
1979 		vp = fp->f_vnode;
1980 		if (vn_isdisk(vp, &error)) {
1981 			fdrop(fp, td);
1982 			td->td_retval[0] = AIO_NOTCANCELED;
1983 			return (0);
1984 		}
1985 	}
1986 
1987 	AIO_LOCK(ki);
1988 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1989 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1990 		    ((uap->aiocbp == NULL) ||
1991 		     (uap->aiocbp == cbe->uuaiocb))) {
1992 			remove = 0;
1993 
1994 			mtx_lock(&aio_job_mtx);
1995 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1996 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1997 				remove = 1;
1998 			} else if (cbe->jobstate == JOBST_JOBQSOCK) {
1999 				MPASS(fp->f_type == DTYPE_SOCKET);
2000 				so = fp->f_data;
2001 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
2002 				remove = 1;
2003 			} else if (cbe->jobstate == JOBST_JOBQSYNC) {
2004 				TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
2005 				remove = 1;
2006 			}
2007 			mtx_unlock(&aio_job_mtx);
2008 
2009 			if (remove) {
2010 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
2011 				cbe->uaiocb._aiocb_private.status = -1;
2012 				cbe->uaiocb._aiocb_private.error = ECANCELED;
2013 				aio_bio_done_notify(p, cbe, DONE_QUEUE);
2014 				cancelled++;
2015 			} else {
2016 				notcancelled++;
2017 			}
2018 			if (uap->aiocbp != NULL)
2019 				break;
2020 		}
2021 	}
2022 	AIO_UNLOCK(ki);
2023 
2024 done:
2025 	fdrop(fp, td);
2026 
2027 	if (uap->aiocbp != NULL) {
2028 		if (cancelled) {
2029 			td->td_retval[0] = AIO_CANCELED;
2030 			return (0);
2031 		}
2032 	}
2033 
2034 	if (notcancelled) {
2035 		td->td_retval[0] = AIO_NOTCANCELED;
2036 		return (0);
2037 	}
2038 
2039 	if (cancelled) {
2040 		td->td_retval[0] = AIO_CANCELED;
2041 		return (0);
2042 	}
2043 
2044 	td->td_retval[0] = AIO_ALLDONE;
2045 
2046 	return (0);
2047 }
2048 
2049 /*
2050  * aio_error is implemented in the kernel level for compatibility purposes
2051  * only.  For a user mode async implementation, it would be best to do it in
2052  * a userland subroutine.
2053  */
2054 static int
2055 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2056 {
2057 	struct proc *p = td->td_proc;
2058 	struct aiocblist *cb;
2059 	struct kaioinfo *ki;
2060 	int status;
2061 
2062 	ki = p->p_aioinfo;
2063 	if (ki == NULL) {
2064 		td->td_retval[0] = EINVAL;
2065 		return (0);
2066 	}
2067 
2068 	AIO_LOCK(ki);
2069 	TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2070 		if (cb->uuaiocb == aiocbp) {
2071 			if (cb->jobstate == JOBST_JOBFINISHED)
2072 				td->td_retval[0] =
2073 					cb->uaiocb._aiocb_private.error;
2074 			else
2075 				td->td_retval[0] = EINPROGRESS;
2076 			AIO_UNLOCK(ki);
2077 			return (0);
2078 		}
2079 	}
2080 	AIO_UNLOCK(ki);
2081 
2082 	/*
2083 	 * Hack for failure of aio_aqueue.
2084 	 */
2085 	status = ops->fetch_status(aiocbp);
2086 	if (status == -1) {
2087 		td->td_retval[0] = ops->fetch_error(aiocbp);
2088 		return (0);
2089 	}
2090 
2091 	td->td_retval[0] = EINVAL;
2092 	return (0);
2093 }
2094 
2095 int
2096 aio_error(struct thread *td, struct aio_error_args *uap)
2097 {
2098 
2099 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2100 }
2101 
2102 /* syscall - asynchronous read from a file (REALTIME) */
2103 int
2104 oaio_read(struct thread *td, struct oaio_read_args *uap)
2105 {
2106 
2107 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2108 	    &aiocb_ops_osigevent));
2109 }
2110 
2111 int
2112 aio_read(struct thread *td, struct aio_read_args *uap)
2113 {
2114 
2115 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2116 }
2117 
2118 /* syscall - asynchronous write to a file (REALTIME) */
2119 int
2120 oaio_write(struct thread *td, struct oaio_write_args *uap)
2121 {
2122 
2123 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2124 	    &aiocb_ops_osigevent));
2125 }
2126 
2127 int
2128 aio_write(struct thread *td, struct aio_write_args *uap)
2129 {
2130 
2131 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2132 }
2133 
2134 static int
2135 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2136     struct aiocb **acb_list, int nent, struct sigevent *sig,
2137     struct aiocb_ops *ops)
2138 {
2139 	struct proc *p = td->td_proc;
2140 	struct aiocb *iocb;
2141 	struct kaioinfo *ki;
2142 	struct aioliojob *lj;
2143 	struct kevent kev;
2144 	int error;
2145 	int nerror;
2146 	int i;
2147 
2148 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2149 		return (EINVAL);
2150 
2151 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2152 		return (EINVAL);
2153 
2154 	if (p->p_aioinfo == NULL)
2155 		aio_init_aioinfo(p);
2156 
2157 	ki = p->p_aioinfo;
2158 
2159 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2160 	lj->lioj_flags = 0;
2161 	lj->lioj_count = 0;
2162 	lj->lioj_finished_count = 0;
2163 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2164 	ksiginfo_init(&lj->lioj_ksi);
2165 
2166 	/*
2167 	 * Setup signal.
2168 	 */
2169 	if (sig && (mode == LIO_NOWAIT)) {
2170 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2171 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2172 			/* Assume only new style KEVENT */
2173 			kev.filter = EVFILT_LIO;
2174 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2175 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2176 			kev.data = (intptr_t)lj;
2177 			/* pass user defined sigval data */
2178 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2179 			error = kqfd_register(
2180 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2181 			if (error) {
2182 				uma_zfree(aiolio_zone, lj);
2183 				return (error);
2184 			}
2185 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2186 			;
2187 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2188 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2189 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2190 					uma_zfree(aiolio_zone, lj);
2191 					return EINVAL;
2192 				}
2193 				lj->lioj_flags |= LIOJ_SIGNAL;
2194 		} else {
2195 			uma_zfree(aiolio_zone, lj);
2196 			return EINVAL;
2197 		}
2198 	}
2199 
2200 	AIO_LOCK(ki);
2201 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2202 	/*
2203 	 * Add extra aiocb count to avoid the lio to be freed
2204 	 * by other threads doing aio_waitcomplete or aio_return,
2205 	 * and prevent event from being sent until we have queued
2206 	 * all tasks.
2207 	 */
2208 	lj->lioj_count = 1;
2209 	AIO_UNLOCK(ki);
2210 
2211 	/*
2212 	 * Get pointers to the list of I/O requests.
2213 	 */
2214 	nerror = 0;
2215 	for (i = 0; i < nent; i++) {
2216 		iocb = acb_list[i];
2217 		if (iocb != NULL) {
2218 			error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2219 			if (error != 0)
2220 				nerror++;
2221 		}
2222 	}
2223 
2224 	error = 0;
2225 	AIO_LOCK(ki);
2226 	if (mode == LIO_WAIT) {
2227 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2228 			ki->kaio_flags |= KAIO_WAKEUP;
2229 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2230 			    PRIBIO | PCATCH, "aiospn", 0);
2231 			if (error == ERESTART)
2232 				error = EINTR;
2233 			if (error)
2234 				break;
2235 		}
2236 	} else {
2237 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2238 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2239 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2240 				KNOTE_LOCKED(&lj->klist, 1);
2241 			}
2242 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2243 			    == LIOJ_SIGNAL
2244 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2245 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2246 				aio_sendsig(p, &lj->lioj_signal,
2247 					    &lj->lioj_ksi);
2248 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2249 			}
2250 		}
2251 	}
2252 	lj->lioj_count--;
2253 	if (lj->lioj_count == 0) {
2254 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2255 		knlist_delete(&lj->klist, curthread, 1);
2256 		PROC_LOCK(p);
2257 		sigqueue_take(&lj->lioj_ksi);
2258 		PROC_UNLOCK(p);
2259 		AIO_UNLOCK(ki);
2260 		uma_zfree(aiolio_zone, lj);
2261 	} else
2262 		AIO_UNLOCK(ki);
2263 
2264 	if (nerror)
2265 		return (EIO);
2266 	return (error);
2267 }
2268 
2269 /* syscall - list directed I/O (REALTIME) */
2270 int
2271 olio_listio(struct thread *td, struct olio_listio_args *uap)
2272 {
2273 	struct aiocb **acb_list;
2274 	struct sigevent *sigp, sig;
2275 	struct osigevent osig;
2276 	int error, nent;
2277 
2278 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2279 		return (EINVAL);
2280 
2281 	nent = uap->nent;
2282 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2283 		return (EINVAL);
2284 
2285 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2286 		error = copyin(uap->sig, &osig, sizeof(osig));
2287 		if (error)
2288 			return (error);
2289 		error = convert_old_sigevent(&osig, &sig);
2290 		if (error)
2291 			return (error);
2292 		sigp = &sig;
2293 	} else
2294 		sigp = NULL;
2295 
2296 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2297 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2298 	if (error == 0)
2299 		error = kern_lio_listio(td, uap->mode,
2300 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2301 		    &aiocb_ops_osigevent);
2302 	free(acb_list, M_LIO);
2303 	return (error);
2304 }
2305 
2306 /* syscall - list directed I/O (REALTIME) */
2307 int
2308 lio_listio(struct thread *td, struct lio_listio_args *uap)
2309 {
2310 	struct aiocb **acb_list;
2311 	struct sigevent *sigp, sig;
2312 	int error, nent;
2313 
2314 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2315 		return (EINVAL);
2316 
2317 	nent = uap->nent;
2318 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2319 		return (EINVAL);
2320 
2321 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2322 		error = copyin(uap->sig, &sig, sizeof(sig));
2323 		if (error)
2324 			return (error);
2325 		sigp = &sig;
2326 	} else
2327 		sigp = NULL;
2328 
2329 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2330 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2331 	if (error == 0)
2332 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2333 		    nent, sigp, &aiocb_ops);
2334 	free(acb_list, M_LIO);
2335 	return (error);
2336 }
2337 
2338 /*
2339  * Called from interrupt thread for physio, we should return as fast
2340  * as possible, so we schedule a biohelper task.
2341  */
2342 static void
2343 aio_physwakeup(struct buf *bp)
2344 {
2345 	struct aiocblist *aiocbe;
2346 
2347 	aiocbe = (struct aiocblist *)bp->b_caller1;
2348 	taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2349 }
2350 
2351 /*
2352  * Task routine to perform heavy tasks, process wakeup, and signals.
2353  */
2354 static void
2355 biohelper(void *context, int pending)
2356 {
2357 	struct aiocblist *aiocbe = context;
2358 	struct buf *bp;
2359 	struct proc *userp;
2360 	struct kaioinfo *ki;
2361 	int nblks;
2362 
2363 	bp = aiocbe->bp;
2364 	userp = aiocbe->userproc;
2365 	ki = userp->p_aioinfo;
2366 	AIO_LOCK(ki);
2367 	aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2368 	aiocbe->uaiocb._aiocb_private.error = 0;
2369 	if (bp->b_ioflags & BIO_ERROR)
2370 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2371 	nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2372 	if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2373 		aiocbe->outputcharge += nblks;
2374 	else
2375 		aiocbe->inputcharge += nblks;
2376 	aiocbe->bp = NULL;
2377 	TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2378 	ki->kaio_buffer_count--;
2379 	aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2380 	AIO_UNLOCK(ki);
2381 
2382 	/* Release mapping into kernel space. */
2383 	vunmapbuf(bp);
2384 	relpbuf(bp, NULL);
2385 	atomic_subtract_int(&num_buf_aio, 1);
2386 }
2387 
2388 /* syscall - wait for the next completion of an aio request */
2389 static int
2390 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2391     struct timespec *ts, struct aiocb_ops *ops)
2392 {
2393 	struct proc *p = td->td_proc;
2394 	struct timeval atv;
2395 	struct kaioinfo *ki;
2396 	struct aiocblist *cb;
2397 	struct aiocb *uuaiocb;
2398 	int error, status, timo;
2399 
2400 	ops->store_aiocb(aiocbp, NULL);
2401 
2402 	timo = 0;
2403 	if (ts) {
2404 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2405 			return (EINVAL);
2406 
2407 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2408 		if (itimerfix(&atv))
2409 			return (EINVAL);
2410 		timo = tvtohz(&atv);
2411 	}
2412 
2413 	if (p->p_aioinfo == NULL)
2414 		aio_init_aioinfo(p);
2415 	ki = p->p_aioinfo;
2416 
2417 	error = 0;
2418 	cb = NULL;
2419 	AIO_LOCK(ki);
2420 	while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2421 		ki->kaio_flags |= KAIO_WAKEUP;
2422 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2423 		    "aiowc", timo);
2424 		if (timo && error == ERESTART)
2425 			error = EINTR;
2426 		if (error)
2427 			break;
2428 	}
2429 
2430 	if (cb != NULL) {
2431 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
2432 		uuaiocb = cb->uuaiocb;
2433 		status = cb->uaiocb._aiocb_private.status;
2434 		error = cb->uaiocb._aiocb_private.error;
2435 		td->td_retval[0] = status;
2436 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2437 			td->td_ru.ru_oublock += cb->outputcharge;
2438 			cb->outputcharge = 0;
2439 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2440 			td->td_ru.ru_inblock += cb->inputcharge;
2441 			cb->inputcharge = 0;
2442 		}
2443 		aio_free_entry(cb);
2444 		AIO_UNLOCK(ki);
2445 		ops->store_aiocb(aiocbp, uuaiocb);
2446 		ops->store_error(uuaiocb, error);
2447 		ops->store_status(uuaiocb, status);
2448 	} else
2449 		AIO_UNLOCK(ki);
2450 
2451 	return (error);
2452 }
2453 
2454 int
2455 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2456 {
2457 	struct timespec ts, *tsp;
2458 	int error;
2459 
2460 	if (uap->timeout) {
2461 		/* Get timespec struct. */
2462 		error = copyin(uap->timeout, &ts, sizeof(ts));
2463 		if (error)
2464 			return (error);
2465 		tsp = &ts;
2466 	} else
2467 		tsp = NULL;
2468 
2469 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2470 }
2471 
2472 static int
2473 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2474     struct aiocb_ops *ops)
2475 {
2476 	struct proc *p = td->td_proc;
2477 	struct kaioinfo *ki;
2478 
2479 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2480 		return (EINVAL);
2481 	ki = p->p_aioinfo;
2482 	if (ki == NULL)
2483 		aio_init_aioinfo(p);
2484 	return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2485 }
2486 
2487 int
2488 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2489 {
2490 
2491 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2492 }
2493 
2494 /* kqueue attach function */
2495 static int
2496 filt_aioattach(struct knote *kn)
2497 {
2498 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2499 
2500 	/*
2501 	 * The aiocbe pointer must be validated before using it, so
2502 	 * registration is restricted to the kernel; the user cannot
2503 	 * set EV_FLAG1.
2504 	 */
2505 	if ((kn->kn_flags & EV_FLAG1) == 0)
2506 		return (EPERM);
2507 	kn->kn_ptr.p_aio = aiocbe;
2508 	kn->kn_flags &= ~EV_FLAG1;
2509 
2510 	knlist_add(&aiocbe->klist, kn, 0);
2511 
2512 	return (0);
2513 }
2514 
2515 /* kqueue detach function */
2516 static void
2517 filt_aiodetach(struct knote *kn)
2518 {
2519 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2520 
2521 	if (!knlist_empty(&aiocbe->klist))
2522 		knlist_remove(&aiocbe->klist, kn, 0);
2523 }
2524 
2525 /* kqueue filter function */
2526 /*ARGSUSED*/
2527 static int
2528 filt_aio(struct knote *kn, long hint)
2529 {
2530 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2531 
2532 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2533 	if (aiocbe->jobstate != JOBST_JOBFINISHED)
2534 		return (0);
2535 	kn->kn_flags |= EV_EOF;
2536 	return (1);
2537 }
2538 
2539 /* kqueue attach function */
2540 static int
2541 filt_lioattach(struct knote *kn)
2542 {
2543 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2544 
2545 	/*
2546 	 * The aioliojob pointer must be validated before using it, so
2547 	 * registration is restricted to the kernel; the user cannot
2548 	 * set EV_FLAG1.
2549 	 */
2550 	if ((kn->kn_flags & EV_FLAG1) == 0)
2551 		return (EPERM);
2552 	kn->kn_ptr.p_lio = lj;
2553 	kn->kn_flags &= ~EV_FLAG1;
2554 
2555 	knlist_add(&lj->klist, kn, 0);
2556 
2557 	return (0);
2558 }
2559 
2560 /* kqueue detach function */
2561 static void
2562 filt_liodetach(struct knote *kn)
2563 {
2564 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2565 
2566 	if (!knlist_empty(&lj->klist))
2567 		knlist_remove(&lj->klist, kn, 0);
2568 }
2569 
2570 /* kqueue filter function */
2571 /*ARGSUSED*/
2572 static int
2573 filt_lio(struct knote *kn, long hint)
2574 {
2575 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2576 
2577 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2578 }
2579 
2580 #ifdef COMPAT_FREEBSD32
2581 
2582 struct __aiocb_private32 {
2583 	int32_t	status;
2584 	int32_t	error;
2585 	uint32_t kernelinfo;
2586 };
2587 
2588 typedef struct oaiocb32 {
2589 	int	aio_fildes;		/* File descriptor */
2590 	uint64_t aio_offset __packed;	/* File offset for I/O */
2591 	uint32_t aio_buf;		/* I/O buffer in process space */
2592 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2593 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2594 	int	aio_lio_opcode;		/* LIO opcode */
2595 	int	aio_reqprio;		/* Request priority -- ignored */
2596 	struct	__aiocb_private32 _aiocb_private;
2597 } oaiocb32_t;
2598 
2599 typedef struct aiocb32 {
2600 	int32_t	aio_fildes;		/* File descriptor */
2601 	uint64_t aio_offset __packed;	/* File offset for I/O */
2602 	uint32_t aio_buf;		/* I/O buffer in process space */
2603 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2604 	int	__spare__[2];
2605 	uint32_t __spare2__;
2606 	int	aio_lio_opcode;		/* LIO opcode */
2607 	int	aio_reqprio;		/* Request priority -- ignored */
2608 	struct __aiocb_private32 _aiocb_private;
2609 	struct sigevent32 aio_sigevent;	/* Signal to deliver */
2610 } aiocb32_t;
2611 
2612 static int
2613 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2614 {
2615 
2616 	/*
2617 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2618 	 * supported by AIO with the old sigevent structure.
2619 	 */
2620 	CP(*osig, *nsig, sigev_notify);
2621 	switch (nsig->sigev_notify) {
2622 	case SIGEV_NONE:
2623 		break;
2624 	case SIGEV_SIGNAL:
2625 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2626 		break;
2627 	case SIGEV_KEVENT:
2628 		nsig->sigev_notify_kqueue =
2629 		    osig->__sigev_u.__sigev_notify_kqueue;
2630 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2631 		break;
2632 	default:
2633 		return (EINVAL);
2634 	}
2635 	return (0);
2636 }
2637 
2638 static int
2639 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2640 {
2641 	struct oaiocb32 job32;
2642 	int error;
2643 
2644 	bzero(kjob, sizeof(struct aiocb));
2645 	error = copyin(ujob, &job32, sizeof(job32));
2646 	if (error)
2647 		return (error);
2648 
2649 	CP(job32, *kjob, aio_fildes);
2650 	CP(job32, *kjob, aio_offset);
2651 	PTRIN_CP(job32, *kjob, aio_buf);
2652 	CP(job32, *kjob, aio_nbytes);
2653 	CP(job32, *kjob, aio_lio_opcode);
2654 	CP(job32, *kjob, aio_reqprio);
2655 	CP(job32, *kjob, _aiocb_private.status);
2656 	CP(job32, *kjob, _aiocb_private.error);
2657 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2658 	return (convert_old_sigevent32(&job32.aio_sigevent,
2659 	    &kjob->aio_sigevent));
2660 }
2661 
2662 static int
2663 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig)
2664 {
2665 
2666 	CP(*sig32, *sig, sigev_notify);
2667 	switch (sig->sigev_notify) {
2668 	case SIGEV_NONE:
2669 		break;
2670 	case SIGEV_THREAD_ID:
2671 		CP(*sig32, *sig, sigev_notify_thread_id);
2672 		/* FALLTHROUGH */
2673 	case SIGEV_SIGNAL:
2674 		CP(*sig32, *sig, sigev_signo);
2675 		break;
2676 	case SIGEV_KEVENT:
2677 		CP(*sig32, *sig, sigev_notify_kqueue);
2678 		PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr);
2679 		break;
2680 	default:
2681 		return (EINVAL);
2682 	}
2683 	return (0);
2684 }
2685 
2686 static int
2687 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2688 {
2689 	struct aiocb32 job32;
2690 	int error;
2691 
2692 	error = copyin(ujob, &job32, sizeof(job32));
2693 	if (error)
2694 		return (error);
2695 	CP(job32, *kjob, aio_fildes);
2696 	CP(job32, *kjob, aio_offset);
2697 	PTRIN_CP(job32, *kjob, aio_buf);
2698 	CP(job32, *kjob, aio_nbytes);
2699 	CP(job32, *kjob, aio_lio_opcode);
2700 	CP(job32, *kjob, aio_reqprio);
2701 	CP(job32, *kjob, _aiocb_private.status);
2702 	CP(job32, *kjob, _aiocb_private.error);
2703 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2704 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2705 }
2706 
2707 static long
2708 aiocb32_fetch_status(struct aiocb *ujob)
2709 {
2710 	struct aiocb32 *ujob32;
2711 
2712 	ujob32 = (struct aiocb32 *)ujob;
2713 	return (fuword32(&ujob32->_aiocb_private.status));
2714 }
2715 
2716 static long
2717 aiocb32_fetch_error(struct aiocb *ujob)
2718 {
2719 	struct aiocb32 *ujob32;
2720 
2721 	ujob32 = (struct aiocb32 *)ujob;
2722 	return (fuword32(&ujob32->_aiocb_private.error));
2723 }
2724 
2725 static int
2726 aiocb32_store_status(struct aiocb *ujob, long status)
2727 {
2728 	struct aiocb32 *ujob32;
2729 
2730 	ujob32 = (struct aiocb32 *)ujob;
2731 	return (suword32(&ujob32->_aiocb_private.status, status));
2732 }
2733 
2734 static int
2735 aiocb32_store_error(struct aiocb *ujob, long error)
2736 {
2737 	struct aiocb32 *ujob32;
2738 
2739 	ujob32 = (struct aiocb32 *)ujob;
2740 	return (suword32(&ujob32->_aiocb_private.error, error));
2741 }
2742 
2743 static int
2744 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2745 {
2746 	struct aiocb32 *ujob32;
2747 
2748 	ujob32 = (struct aiocb32 *)ujob;
2749 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2750 }
2751 
2752 static int
2753 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2754 {
2755 
2756 	return (suword32(ujobp, (long)ujob));
2757 }
2758 
2759 static struct aiocb_ops aiocb32_ops = {
2760 	.copyin = aiocb32_copyin,
2761 	.fetch_status = aiocb32_fetch_status,
2762 	.fetch_error = aiocb32_fetch_error,
2763 	.store_status = aiocb32_store_status,
2764 	.store_error = aiocb32_store_error,
2765 	.store_kernelinfo = aiocb32_store_kernelinfo,
2766 	.store_aiocb = aiocb32_store_aiocb,
2767 };
2768 
2769 static struct aiocb_ops aiocb32_ops_osigevent = {
2770 	.copyin = aiocb32_copyin_old_sigevent,
2771 	.fetch_status = aiocb32_fetch_status,
2772 	.fetch_error = aiocb32_fetch_error,
2773 	.store_status = aiocb32_store_status,
2774 	.store_error = aiocb32_store_error,
2775 	.store_kernelinfo = aiocb32_store_kernelinfo,
2776 	.store_aiocb = aiocb32_store_aiocb,
2777 };
2778 
2779 int
2780 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2781 {
2782 
2783 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2784 }
2785 
2786 int
2787 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2788 {
2789 	struct timespec32 ts32;
2790 	struct timespec ts, *tsp;
2791 	struct aiocb **ujoblist;
2792 	uint32_t *ujoblist32;
2793 	int error, i;
2794 
2795 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2796 		return (EINVAL);
2797 
2798 	if (uap->timeout) {
2799 		/* Get timespec struct. */
2800 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2801 			return (error);
2802 		CP(ts32, ts, tv_sec);
2803 		CP(ts32, ts, tv_nsec);
2804 		tsp = &ts;
2805 	} else
2806 		tsp = NULL;
2807 
2808 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2809 	ujoblist32 = (uint32_t *)ujoblist;
2810 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2811 	    sizeof(ujoblist32[0]));
2812 	if (error == 0) {
2813 		for (i = uap->nent; i > 0; i--)
2814 			ujoblist[i] = PTRIN(ujoblist32[i]);
2815 
2816 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2817 	}
2818 	uma_zfree(aiol_zone, ujoblist);
2819 	return (error);
2820 }
2821 
2822 int
2823 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2824 {
2825 
2826 	return (aio_cancel(td, (struct aio_cancel_args *)uap));
2827 }
2828 
2829 int
2830 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2831 {
2832 
2833 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2834 }
2835 
2836 int
2837 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2838 {
2839 
2840 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2841 	    &aiocb32_ops_osigevent));
2842 }
2843 
2844 int
2845 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2846 {
2847 
2848 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2849 	    &aiocb32_ops));
2850 }
2851 
2852 int
2853 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2854 {
2855 
2856 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2857 	    &aiocb32_ops_osigevent));
2858 }
2859 
2860 int
2861 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2862 {
2863 
2864 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2865 	    &aiocb32_ops));
2866 }
2867 
2868 int
2869 freebsd32_aio_waitcomplete(struct thread *td,
2870     struct freebsd32_aio_waitcomplete_args *uap)
2871 {
2872 	struct timespec32 ts32;
2873 	struct timespec ts, *tsp;
2874 	int error;
2875 
2876 	if (uap->timeout) {
2877 		/* Get timespec struct. */
2878 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2879 		if (error)
2880 			return (error);
2881 		CP(ts32, ts, tv_sec);
2882 		CP(ts32, ts, tv_nsec);
2883 		tsp = &ts;
2884 	} else
2885 		tsp = NULL;
2886 
2887 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2888 	    &aiocb32_ops));
2889 }
2890 
2891 int
2892 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2893 {
2894 
2895 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2896 	    &aiocb32_ops));
2897 }
2898 
2899 int
2900 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2901 {
2902 	struct aiocb **acb_list;
2903 	struct sigevent *sigp, sig;
2904 	struct osigevent32 osig;
2905 	uint32_t *acb_list32;
2906 	int error, i, nent;
2907 
2908 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2909 		return (EINVAL);
2910 
2911 	nent = uap->nent;
2912 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2913 		return (EINVAL);
2914 
2915 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2916 		error = copyin(uap->sig, &osig, sizeof(osig));
2917 		if (error)
2918 			return (error);
2919 		error = convert_old_sigevent32(&osig, &sig);
2920 		if (error)
2921 			return (error);
2922 		sigp = &sig;
2923 	} else
2924 		sigp = NULL;
2925 
2926 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2927 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2928 	if (error) {
2929 		free(acb_list32, M_LIO);
2930 		return (error);
2931 	}
2932 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2933 	for (i = 0; i < nent; i++)
2934 		acb_list[i] = PTRIN(acb_list32[i]);
2935 	free(acb_list32, M_LIO);
2936 
2937 	error = kern_lio_listio(td, uap->mode,
2938 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2939 	    &aiocb32_ops_osigevent);
2940 	free(acb_list, M_LIO);
2941 	return (error);
2942 }
2943 
2944 int
2945 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2946 {
2947 	struct aiocb **acb_list;
2948 	struct sigevent *sigp, sig;
2949 	struct sigevent32 sig32;
2950 	uint32_t *acb_list32;
2951 	int error, i, nent;
2952 
2953 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2954 		return (EINVAL);
2955 
2956 	nent = uap->nent;
2957 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2958 		return (EINVAL);
2959 
2960 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2961 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2962 		if (error)
2963 			return (error);
2964 		error = convert_sigevent32(&sig32, &sig);
2965 		if (error)
2966 			return (error);
2967 		sigp = &sig;
2968 	} else
2969 		sigp = NULL;
2970 
2971 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2972 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2973 	if (error) {
2974 		free(acb_list32, M_LIO);
2975 		return (error);
2976 	}
2977 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2978 	for (i = 0; i < nent; i++)
2979 		acb_list[i] = PTRIN(acb_list32[i]);
2980 	free(acb_list32, M_LIO);
2981 
2982 	error = kern_lio_listio(td, uap->mode,
2983 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2984 	    &aiocb32_ops);
2985 	free(acb_list, M_LIO);
2986 	return (error);
2987 }
2988 
2989 #endif
2990