xref: /freebsd/sys/kern/vfs_aio.c (revision f4dc9bf43457515e5c88d1400d4f5ff70a82d9c7)
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 /*
76  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
77  * overflow. (XXX will be removed soon.)
78  */
79 static u_long jobrefid;
80 
81 /*
82  * Counter for aio_fsync.
83  */
84 static uint64_t jobseqno;
85 
86 #ifndef MAX_AIO_PER_PROC
87 #define MAX_AIO_PER_PROC	32
88 #endif
89 
90 #ifndef MAX_AIO_QUEUE_PER_PROC
91 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
92 #endif
93 
94 #ifndef MAX_AIO_QUEUE
95 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
96 #endif
97 
98 #ifndef MAX_BUF_AIO
99 #define MAX_BUF_AIO		16
100 #endif
101 
102 FEATURE(aio, "Asynchronous I/O");
103 
104 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
105 
106 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
107     "Async IO management");
108 
109 static int enable_aio_unsafe = 0;
110 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
111     "Permit asynchronous IO on all file types, not just known-safe types");
112 
113 static int max_aio_procs = MAX_AIO_PROCS;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
115     "Maximum number of kernel processes to use for handling async IO ");
116 
117 static int num_aio_procs = 0;
118 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
119     "Number of presently active kernel processes for async IO");
120 
121 /*
122  * The code will adjust the actual number of AIO processes towards this
123  * number when it gets a chance.
124  */
125 static int target_aio_procs = TARGET_AIO_PROCS;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
127     0,
128     "Preferred number of ready kernel processes for async IO");
129 
130 static int max_queue_count = MAX_AIO_QUEUE;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
132     "Maximum number of aio requests to queue, globally");
133 
134 static int num_queue_count = 0;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
136     "Number of queued aio requests");
137 
138 static int num_buf_aio = 0;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
140     "Number of aio requests presently handled by the buf subsystem");
141 
142 /* Number of async I/O processes in the process of being started */
143 /* XXX This should be local to aio_aqueue() */
144 static int num_aio_resv_start = 0;
145 
146 static int aiod_lifetime;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
148     "Maximum lifetime for idle aiod");
149 
150 static int max_aio_per_proc = MAX_AIO_PER_PROC;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
152     0,
153     "Maximum active aio requests per process (stored in the process)");
154 
155 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
157     &max_aio_queue_per_proc, 0,
158     "Maximum queued aio requests per process (stored in the process)");
159 
160 static int max_buf_aio = MAX_BUF_AIO;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
162     "Maximum buf aio requests per process (stored in the process)");
163 
164 #ifdef COMPAT_FREEBSD6
165 typedef struct oaiocb {
166 	int	aio_fildes;		/* File descriptor */
167 	off_t	aio_offset;		/* File offset for I/O */
168 	volatile void *aio_buf;         /* I/O buffer in process space */
169 	size_t	aio_nbytes;		/* Number of bytes for I/O */
170 	struct	osigevent aio_sigevent;	/* Signal to deliver */
171 	int	aio_lio_opcode;		/* LIO opcode */
172 	int	aio_reqprio;		/* Request priority -- ignored */
173 	struct	__aiocb_private	_aiocb_private;
174 } oaiocb_t;
175 #endif
176 
177 /*
178  * Below is a key of locks used to protect each member of struct kaiocb
179  * aioliojob and kaioinfo and any backends.
180  *
181  * * - need not protected
182  * a - locked by kaioinfo lock
183  * b - locked by backend lock, the backend lock can be null in some cases,
184  *     for example, BIO belongs to this type, in this case, proc lock is
185  *     reused.
186  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
187  */
188 
189 /*
190  * If the routine that services an AIO request blocks while running in an
191  * AIO kernel process it can starve other I/O requests.  BIO requests
192  * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
193  * processes at all.  Socket I/O requests use a separate pool of
194  * kprocs and also force non-blocking I/O.  Other file I/O requests
195  * use the generic fo_read/fo_write operations which can block.  The
196  * fsync and mlock operations can also block while executing.  Ideally
197  * none of these requests would block while executing.
198  *
199  * Note that the service routines cannot toggle O_NONBLOCK in the file
200  * structure directly while handling a request due to races with
201  * userland threads.
202  */
203 
204 /* jobflags */
205 #define	KAIOCB_QUEUEING		0x01
206 #define	KAIOCB_CANCELLED	0x02
207 #define	KAIOCB_CANCELLING	0x04
208 #define	KAIOCB_CHECKSYNC	0x08
209 #define	KAIOCB_CLEARED		0x10
210 #define	KAIOCB_FINISHED		0x20
211 
212 /*
213  * AIO process info
214  */
215 #define AIOP_FREE	0x1			/* proc on free queue */
216 
217 struct aioproc {
218 	int	aioprocflags;			/* (c) AIO proc flags */
219 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
220 	struct	proc *aioproc;			/* (*) the AIO proc */
221 };
222 
223 /*
224  * data-structure for lio signal management
225  */
226 struct aioliojob {
227 	int	lioj_flags;			/* (a) listio flags */
228 	int	lioj_count;			/* (a) listio flags */
229 	int	lioj_finished_count;		/* (a) listio flags */
230 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
231 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
232 	struct	knlist klist;			/* (a) list of knotes */
233 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
234 };
235 
236 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
237 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
238 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
239 
240 /*
241  * per process aio data structure
242  */
243 struct kaioinfo {
244 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
245 	int	kaio_flags;		/* (a) per process kaio flags */
246 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
247 	int	kaio_active_count;	/* (c) number of currently used AIOs */
248 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
249 	int	kaio_count;		/* (a) size of AIO queue */
250 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
251 	int	kaio_buffer_count;	/* (a) number of physio buffers */
252 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
253 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
254 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
255 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
256 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
257 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
258 	struct	task kaio_task;		/* (*) task to kick aio processes */
259 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
260 };
261 
262 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
263 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
264 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
265 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
266 
267 #define KAIO_RUNDOWN	0x1	/* process is being run down */
268 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
269 
270 /*
271  * Operations used to interact with userland aio control blocks.
272  * Different ABIs provide their own operations.
273  */
274 struct aiocb_ops {
275 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
276 	long	(*fetch_status)(struct aiocb *ujob);
277 	long	(*fetch_error)(struct aiocb *ujob);
278 	int	(*store_status)(struct aiocb *ujob, long status);
279 	int	(*store_error)(struct aiocb *ujob, long error);
280 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
281 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
282 };
283 
284 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
285 static struct sema aio_newproc_sem;
286 static struct mtx aio_job_mtx;
287 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
288 static struct unrhdr *aiod_unr;
289 
290 void		aio_init_aioinfo(struct proc *p);
291 static int	aio_onceonly(void);
292 static int	aio_free_entry(struct kaiocb *job);
293 static void	aio_process_rw(struct kaiocb *job);
294 static void	aio_process_sync(struct kaiocb *job);
295 static void	aio_process_mlock(struct kaiocb *job);
296 static void	aio_schedule_fsync(void *context, int pending);
297 static int	aio_newproc(int *);
298 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
299 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
300 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
301 static void	aio_physwakeup(struct bio *bp);
302 static void	aio_proc_rundown(void *arg, struct proc *p);
303 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
304 		    struct image_params *imgp);
305 static int	aio_qphysio(struct proc *p, struct kaiocb *job);
306 static void	aio_daemon(void *param);
307 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
308 static int	aio_kick(struct proc *userp);
309 static void	aio_kick_nowait(struct proc *userp);
310 static void	aio_kick_helper(void *context, int pending);
311 static int	filt_aioattach(struct knote *kn);
312 static void	filt_aiodetach(struct knote *kn);
313 static int	filt_aio(struct knote *kn, long hint);
314 static int	filt_lioattach(struct knote *kn);
315 static void	filt_liodetach(struct knote *kn);
316 static int	filt_lio(struct knote *kn, long hint);
317 
318 /*
319  * Zones for:
320  * 	kaio	Per process async io info
321  *	aiop	async io process data
322  *	aiocb	async io jobs
323  *	aiol	list io job pointer - internal to aio_suspend XXX
324  *	aiolio	list io jobs
325  */
326 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
327 
328 /* kqueue filters for aio */
329 static struct filterops aio_filtops = {
330 	.f_isfd = 0,
331 	.f_attach = filt_aioattach,
332 	.f_detach = filt_aiodetach,
333 	.f_event = filt_aio,
334 };
335 static struct filterops lio_filtops = {
336 	.f_isfd = 0,
337 	.f_attach = filt_lioattach,
338 	.f_detach = filt_liodetach,
339 	.f_event = filt_lio
340 };
341 
342 static eventhandler_tag exit_tag, exec_tag;
343 
344 TASKQUEUE_DEFINE_THREAD(aiod_kick);
345 
346 /*
347  * Main operations function for use as a kernel module.
348  */
349 static int
350 aio_modload(struct module *module, int cmd, void *arg)
351 {
352 	int error = 0;
353 
354 	switch (cmd) {
355 	case MOD_LOAD:
356 		aio_onceonly();
357 		break;
358 	case MOD_SHUTDOWN:
359 		break;
360 	default:
361 		error = EOPNOTSUPP;
362 		break;
363 	}
364 	return (error);
365 }
366 
367 static moduledata_t aio_mod = {
368 	"aio",
369 	&aio_modload,
370 	NULL
371 };
372 
373 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
374 MODULE_VERSION(aio, 1);
375 
376 /*
377  * Startup initialization
378  */
379 static int
380 aio_onceonly(void)
381 {
382 
383 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
384 	    EVENTHANDLER_PRI_ANY);
385 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
386 	    NULL, EVENTHANDLER_PRI_ANY);
387 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
388 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
389 	TAILQ_INIT(&aio_freeproc);
390 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
391 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
392 	TAILQ_INIT(&aio_jobs);
393 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
394 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
395 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
396 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
397 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
398 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
399 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
400 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
401 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
402 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
403 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
404 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
405 	jobrefid = 1;
406 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
407 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
408 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
409 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
410 
411 	return (0);
412 }
413 
414 /*
415  * Init the per-process aioinfo structure.  The aioinfo limits are set
416  * per-process for user limit (resource) management.
417  */
418 void
419 aio_init_aioinfo(struct proc *p)
420 {
421 	struct kaioinfo *ki;
422 
423 	ki = uma_zalloc(kaio_zone, M_WAITOK);
424 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
425 	ki->kaio_flags = 0;
426 	ki->kaio_maxactive_count = max_aio_per_proc;
427 	ki->kaio_active_count = 0;
428 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
429 	ki->kaio_count = 0;
430 	ki->kaio_ballowed_count = max_buf_aio;
431 	ki->kaio_buffer_count = 0;
432 	TAILQ_INIT(&ki->kaio_all);
433 	TAILQ_INIT(&ki->kaio_done);
434 	TAILQ_INIT(&ki->kaio_jobqueue);
435 	TAILQ_INIT(&ki->kaio_liojoblist);
436 	TAILQ_INIT(&ki->kaio_syncqueue);
437 	TAILQ_INIT(&ki->kaio_syncready);
438 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
439 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
440 	PROC_LOCK(p);
441 	if (p->p_aioinfo == NULL) {
442 		p->p_aioinfo = ki;
443 		PROC_UNLOCK(p);
444 	} else {
445 		PROC_UNLOCK(p);
446 		mtx_destroy(&ki->kaio_mtx);
447 		uma_zfree(kaio_zone, ki);
448 	}
449 
450 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
451 		aio_newproc(NULL);
452 }
453 
454 static int
455 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
456 {
457 	struct thread *td;
458 	int error;
459 
460 	error = sigev_findtd(p, sigev, &td);
461 	if (error)
462 		return (error);
463 	if (!KSI_ONQ(ksi)) {
464 		ksiginfo_set_sigev(ksi, sigev);
465 		ksi->ksi_code = SI_ASYNCIO;
466 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
467 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
468 	}
469 	PROC_UNLOCK(p);
470 	return (error);
471 }
472 
473 /*
474  * Free a job entry.  Wait for completion if it is currently active, but don't
475  * delay forever.  If we delay, we return a flag that says that we have to
476  * restart the queue scan.
477  */
478 static int
479 aio_free_entry(struct kaiocb *job)
480 {
481 	struct kaioinfo *ki;
482 	struct aioliojob *lj;
483 	struct proc *p;
484 
485 	p = job->userproc;
486 	MPASS(curproc == p);
487 	ki = p->p_aioinfo;
488 	MPASS(ki != NULL);
489 
490 	AIO_LOCK_ASSERT(ki, MA_OWNED);
491 	MPASS(job->jobflags & KAIOCB_FINISHED);
492 
493 	atomic_subtract_int(&num_queue_count, 1);
494 
495 	ki->kaio_count--;
496 	MPASS(ki->kaio_count >= 0);
497 
498 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
499 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
500 
501 	lj = job->lio;
502 	if (lj) {
503 		lj->lioj_count--;
504 		lj->lioj_finished_count--;
505 
506 		if (lj->lioj_count == 0) {
507 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
508 			/* lio is going away, we need to destroy any knotes */
509 			knlist_delete(&lj->klist, curthread, 1);
510 			PROC_LOCK(p);
511 			sigqueue_take(&lj->lioj_ksi);
512 			PROC_UNLOCK(p);
513 			uma_zfree(aiolio_zone, lj);
514 		}
515 	}
516 
517 	/* job is going away, we need to destroy any knotes */
518 	knlist_delete(&job->klist, curthread, 1);
519 	PROC_LOCK(p);
520 	sigqueue_take(&job->ksi);
521 	PROC_UNLOCK(p);
522 
523 	AIO_UNLOCK(ki);
524 
525 	/*
526 	 * The thread argument here is used to find the owning process
527 	 * and is also passed to fo_close() which may pass it to various
528 	 * places such as devsw close() routines.  Because of that, we
529 	 * need a thread pointer from the process owning the job that is
530 	 * persistent and won't disappear out from under us or move to
531 	 * another process.
532 	 *
533 	 * Currently, all the callers of this function call it to remove
534 	 * a kaiocb from the current process' job list either via a
535 	 * syscall or due to the current process calling exit() or
536 	 * execve().  Thus, we know that p == curproc.  We also know that
537 	 * curthread can't exit since we are curthread.
538 	 *
539 	 * Therefore, we use curthread as the thread to pass to
540 	 * knlist_delete().  This does mean that it is possible for the
541 	 * thread pointer at close time to differ from the thread pointer
542 	 * at open time, but this is already true of file descriptors in
543 	 * a multithreaded process.
544 	 */
545 	if (job->fd_file)
546 		fdrop(job->fd_file, curthread);
547 	crfree(job->cred);
548 	uma_zfree(aiocb_zone, job);
549 	AIO_LOCK(ki);
550 
551 	return (0);
552 }
553 
554 static void
555 aio_proc_rundown_exec(void *arg, struct proc *p,
556     struct image_params *imgp __unused)
557 {
558    	aio_proc_rundown(arg, p);
559 }
560 
561 static int
562 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
563 {
564 	aio_cancel_fn_t *func;
565 	int cancelled;
566 
567 	AIO_LOCK_ASSERT(ki, MA_OWNED);
568 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
569 		return (0);
570 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
571 	job->jobflags |= KAIOCB_CANCELLED;
572 
573 	func = job->cancel_fn;
574 
575 	/*
576 	 * If there is no cancel routine, just leave the job marked as
577 	 * cancelled.  The job should be in active use by a caller who
578 	 * should complete it normally or when it fails to install a
579 	 * cancel routine.
580 	 */
581 	if (func == NULL)
582 		return (0);
583 
584 	/*
585 	 * Set the CANCELLING flag so that aio_complete() will defer
586 	 * completions of this job.  This prevents the job from being
587 	 * freed out from under the cancel callback.  After the
588 	 * callback any deferred completion (whether from the callback
589 	 * or any other source) will be completed.
590 	 */
591 	job->jobflags |= KAIOCB_CANCELLING;
592 	AIO_UNLOCK(ki);
593 	func(job);
594 	AIO_LOCK(ki);
595 	job->jobflags &= ~KAIOCB_CANCELLING;
596 	if (job->jobflags & KAIOCB_FINISHED) {
597 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
598 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
599 		aio_bio_done_notify(p, job);
600 	} else {
601 		/*
602 		 * The cancel callback might have scheduled an
603 		 * operation to cancel this request, but it is
604 		 * only counted as cancelled if the request is
605 		 * cancelled when the callback returns.
606 		 */
607 		cancelled = 0;
608 	}
609 	return (cancelled);
610 }
611 
612 /*
613  * Rundown the jobs for a given process.
614  */
615 static void
616 aio_proc_rundown(void *arg, struct proc *p)
617 {
618 	struct kaioinfo *ki;
619 	struct aioliojob *lj;
620 	struct kaiocb *job, *jobn;
621 
622 	KASSERT(curthread->td_proc == p,
623 	    ("%s: called on non-curproc", __func__));
624 	ki = p->p_aioinfo;
625 	if (ki == NULL)
626 		return;
627 
628 	AIO_LOCK(ki);
629 	ki->kaio_flags |= KAIO_RUNDOWN;
630 
631 restart:
632 
633 	/*
634 	 * Try to cancel all pending requests. This code simulates
635 	 * aio_cancel on all pending I/O requests.
636 	 */
637 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
638 		aio_cancel_job(p, ki, job);
639 	}
640 
641 	/* Wait for all running I/O to be finished */
642 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
643 		ki->kaio_flags |= KAIO_WAKEUP;
644 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
645 		goto restart;
646 	}
647 
648 	/* Free all completed I/O requests. */
649 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
650 		aio_free_entry(job);
651 
652 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
653 		if (lj->lioj_count == 0) {
654 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
655 			knlist_delete(&lj->klist, curthread, 1);
656 			PROC_LOCK(p);
657 			sigqueue_take(&lj->lioj_ksi);
658 			PROC_UNLOCK(p);
659 			uma_zfree(aiolio_zone, lj);
660 		} else {
661 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
662 			    lj->lioj_count, lj->lioj_finished_count);
663 		}
664 	}
665 	AIO_UNLOCK(ki);
666 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
667 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
668 	mtx_destroy(&ki->kaio_mtx);
669 	uma_zfree(kaio_zone, ki);
670 	p->p_aioinfo = NULL;
671 }
672 
673 /*
674  * Select a job to run (called by an AIO daemon).
675  */
676 static struct kaiocb *
677 aio_selectjob(struct aioproc *aiop)
678 {
679 	struct kaiocb *job;
680 	struct kaioinfo *ki;
681 	struct proc *userp;
682 
683 	mtx_assert(&aio_job_mtx, MA_OWNED);
684 restart:
685 	TAILQ_FOREACH(job, &aio_jobs, list) {
686 		userp = job->userproc;
687 		ki = userp->p_aioinfo;
688 
689 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
690 			TAILQ_REMOVE(&aio_jobs, job, list);
691 			if (!aio_clear_cancel_function(job))
692 				goto restart;
693 
694 			/* Account for currently active jobs. */
695 			ki->kaio_active_count++;
696 			break;
697 		}
698 	}
699 	return (job);
700 }
701 
702 /*
703  * Move all data to a permanent storage device.  This code
704  * simulates the fsync syscall.
705  */
706 static int
707 aio_fsync_vnode(struct thread *td, struct vnode *vp)
708 {
709 	struct mount *mp;
710 	int error;
711 
712 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
713 		goto drop;
714 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
715 	if (vp->v_object != NULL) {
716 		VM_OBJECT_WLOCK(vp->v_object);
717 		vm_object_page_clean(vp->v_object, 0, 0, 0);
718 		VM_OBJECT_WUNLOCK(vp->v_object);
719 	}
720 	error = VOP_FSYNC(vp, MNT_WAIT, td);
721 
722 	VOP_UNLOCK(vp, 0);
723 	vn_finished_write(mp);
724 drop:
725 	return (error);
726 }
727 
728 /*
729  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
730  * does the I/O request for the non-physio version of the operations.  The
731  * normal vn operations are used, and this code should work in all instances
732  * for every type of file, including pipes, sockets, fifos, and regular files.
733  *
734  * XXX I don't think it works well for socket, pipe, and fifo.
735  */
736 static void
737 aio_process_rw(struct kaiocb *job)
738 {
739 	struct ucred *td_savedcred;
740 	struct thread *td;
741 	struct aiocb *cb;
742 	struct file *fp;
743 	struct uio auio;
744 	struct iovec aiov;
745 	ssize_t cnt;
746 	long msgsnd_st, msgsnd_end;
747 	long msgrcv_st, msgrcv_end;
748 	long oublock_st, oublock_end;
749 	long inblock_st, inblock_end;
750 	int error;
751 
752 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
753 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
754 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
755 
756 	aio_switch_vmspace(job);
757 	td = curthread;
758 	td_savedcred = td->td_ucred;
759 	td->td_ucred = job->cred;
760 	cb = &job->uaiocb;
761 	fp = job->fd_file;
762 
763 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
764 	aiov.iov_len = cb->aio_nbytes;
765 
766 	auio.uio_iov = &aiov;
767 	auio.uio_iovcnt = 1;
768 	auio.uio_offset = cb->aio_offset;
769 	auio.uio_resid = cb->aio_nbytes;
770 	cnt = cb->aio_nbytes;
771 	auio.uio_segflg = UIO_USERSPACE;
772 	auio.uio_td = td;
773 
774 	msgrcv_st = td->td_ru.ru_msgrcv;
775 	msgsnd_st = td->td_ru.ru_msgsnd;
776 	inblock_st = td->td_ru.ru_inblock;
777 	oublock_st = td->td_ru.ru_oublock;
778 
779 	/*
780 	 * aio_aqueue() acquires a reference to the file that is
781 	 * released in aio_free_entry().
782 	 */
783 	if (cb->aio_lio_opcode == LIO_READ) {
784 		auio.uio_rw = UIO_READ;
785 		if (auio.uio_resid == 0)
786 			error = 0;
787 		else
788 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
789 	} else {
790 		if (fp->f_type == DTYPE_VNODE)
791 			bwillwrite();
792 		auio.uio_rw = UIO_WRITE;
793 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
794 	}
795 	msgrcv_end = td->td_ru.ru_msgrcv;
796 	msgsnd_end = td->td_ru.ru_msgsnd;
797 	inblock_end = td->td_ru.ru_inblock;
798 	oublock_end = td->td_ru.ru_oublock;
799 
800 	job->msgrcv = msgrcv_end - msgrcv_st;
801 	job->msgsnd = msgsnd_end - msgsnd_st;
802 	job->inblock = inblock_end - inblock_st;
803 	job->outblock = oublock_end - oublock_st;
804 
805 	if ((error) && (auio.uio_resid != cnt)) {
806 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
807 			error = 0;
808 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
809 			PROC_LOCK(job->userproc);
810 			kern_psignal(job->userproc, SIGPIPE);
811 			PROC_UNLOCK(job->userproc);
812 		}
813 	}
814 
815 	cnt -= auio.uio_resid;
816 	td->td_ucred = td_savedcred;
817 	if (error)
818 		aio_complete(job, -1, error);
819 	else
820 		aio_complete(job, cnt, 0);
821 }
822 
823 static void
824 aio_process_sync(struct kaiocb *job)
825 {
826 	struct thread *td = curthread;
827 	struct ucred *td_savedcred = td->td_ucred;
828 	struct file *fp = job->fd_file;
829 	int error = 0;
830 
831 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
832 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
833 
834 	td->td_ucred = job->cred;
835 	if (fp->f_vnode != NULL)
836 		error = aio_fsync_vnode(td, fp->f_vnode);
837 	td->td_ucred = td_savedcred;
838 	if (error)
839 		aio_complete(job, -1, error);
840 	else
841 		aio_complete(job, 0, 0);
842 }
843 
844 static void
845 aio_process_mlock(struct kaiocb *job)
846 {
847 	struct aiocb *cb = &job->uaiocb;
848 	int error;
849 
850 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
851 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
852 
853 	aio_switch_vmspace(job);
854 	error = vm_mlock(job->userproc, job->cred,
855 	    __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
856 	if (error)
857 		aio_complete(job, -1, error);
858 	else
859 		aio_complete(job, 0, 0);
860 }
861 
862 static void
863 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
864 {
865 	struct aioliojob *lj;
866 	struct kaioinfo *ki;
867 	struct kaiocb *sjob, *sjobn;
868 	int lj_done;
869 	bool schedule_fsync;
870 
871 	ki = userp->p_aioinfo;
872 	AIO_LOCK_ASSERT(ki, MA_OWNED);
873 	lj = job->lio;
874 	lj_done = 0;
875 	if (lj) {
876 		lj->lioj_finished_count++;
877 		if (lj->lioj_count == lj->lioj_finished_count)
878 			lj_done = 1;
879 	}
880 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
881 	MPASS(job->jobflags & KAIOCB_FINISHED);
882 
883 	if (ki->kaio_flags & KAIO_RUNDOWN)
884 		goto notification_done;
885 
886 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
887 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
888 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
889 
890 	KNOTE_LOCKED(&job->klist, 1);
891 
892 	if (lj_done) {
893 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
894 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
895 			KNOTE_LOCKED(&lj->klist, 1);
896 		}
897 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
898 		    == LIOJ_SIGNAL
899 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
900 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
901 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
902 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
903 		}
904 	}
905 
906 notification_done:
907 	if (job->jobflags & KAIOCB_CHECKSYNC) {
908 		schedule_fsync = false;
909 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
910 			if (job->fd_file == sjob->fd_file &&
911 			    job->seqno < sjob->seqno) {
912 				if (--sjob->pending == 0) {
913 					TAILQ_REMOVE(&ki->kaio_syncqueue, sjob,
914 					    list);
915 					if (!aio_clear_cancel_function(sjob))
916 						continue;
917 					TAILQ_INSERT_TAIL(&ki->kaio_syncready,
918 					    sjob, list);
919 					schedule_fsync = true;
920 				}
921 			}
922 		}
923 		if (schedule_fsync)
924 			taskqueue_enqueue(taskqueue_aiod_kick,
925 			    &ki->kaio_sync_task);
926 	}
927 	if (ki->kaio_flags & KAIO_WAKEUP) {
928 		ki->kaio_flags &= ~KAIO_WAKEUP;
929 		wakeup(&userp->p_aioinfo);
930 	}
931 }
932 
933 static void
934 aio_schedule_fsync(void *context, int pending)
935 {
936 	struct kaioinfo *ki;
937 	struct kaiocb *job;
938 
939 	ki = context;
940 	AIO_LOCK(ki);
941 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
942 		job = TAILQ_FIRST(&ki->kaio_syncready);
943 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
944 		AIO_UNLOCK(ki);
945 		aio_schedule(job, aio_process_sync);
946 		AIO_LOCK(ki);
947 	}
948 	AIO_UNLOCK(ki);
949 }
950 
951 bool
952 aio_cancel_cleared(struct kaiocb *job)
953 {
954 	struct kaioinfo *ki;
955 
956 	/*
957 	 * The caller should hold the same queue lock held when
958 	 * aio_clear_cancel_function() was called and set this flag
959 	 * ensuring this check sees an up-to-date value.  However,
960 	 * there is no way to assert that.
961 	 */
962 	ki = job->userproc->p_aioinfo;
963 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
964 }
965 
966 bool
967 aio_clear_cancel_function(struct kaiocb *job)
968 {
969 	struct kaioinfo *ki;
970 
971 	ki = job->userproc->p_aioinfo;
972 	AIO_LOCK(ki);
973 	MPASS(job->cancel_fn != NULL);
974 	if (job->jobflags & KAIOCB_CANCELLING) {
975 		job->jobflags |= KAIOCB_CLEARED;
976 		AIO_UNLOCK(ki);
977 		return (false);
978 	}
979 	job->cancel_fn = NULL;
980 	AIO_UNLOCK(ki);
981 	return (true);
982 }
983 
984 bool
985 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
986 {
987 	struct kaioinfo *ki;
988 
989 	ki = job->userproc->p_aioinfo;
990 	AIO_LOCK(ki);
991 	if (job->jobflags & KAIOCB_CANCELLED) {
992 		AIO_UNLOCK(ki);
993 		return (false);
994 	}
995 	job->cancel_fn = func;
996 	AIO_UNLOCK(ki);
997 	return (true);
998 }
999 
1000 void
1001 aio_complete(struct kaiocb *job, long status, int error)
1002 {
1003 	struct kaioinfo *ki;
1004 	struct proc *userp;
1005 
1006 	job->uaiocb._aiocb_private.error = error;
1007 	job->uaiocb._aiocb_private.status = status;
1008 
1009 	userp = job->userproc;
1010 	ki = userp->p_aioinfo;
1011 
1012 	AIO_LOCK(ki);
1013 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1014 	    ("duplicate aio_complete"));
1015 	job->jobflags |= KAIOCB_FINISHED;
1016 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1017 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1018 		aio_bio_done_notify(userp, job);
1019 	}
1020 	AIO_UNLOCK(ki);
1021 }
1022 
1023 void
1024 aio_cancel(struct kaiocb *job)
1025 {
1026 
1027 	aio_complete(job, -1, ECANCELED);
1028 }
1029 
1030 void
1031 aio_switch_vmspace(struct kaiocb *job)
1032 {
1033 
1034 	vmspace_switch_aio(job->userproc->p_vmspace);
1035 }
1036 
1037 /*
1038  * The AIO daemon, most of the actual work is done in aio_process_*,
1039  * but the setup (and address space mgmt) is done in this routine.
1040  */
1041 static void
1042 aio_daemon(void *_id)
1043 {
1044 	struct kaiocb *job;
1045 	struct aioproc *aiop;
1046 	struct kaioinfo *ki;
1047 	struct proc *p;
1048 	struct vmspace *myvm;
1049 	struct thread *td = curthread;
1050 	int id = (intptr_t)_id;
1051 
1052 	/*
1053 	 * Grab an extra reference on the daemon's vmspace so that it
1054 	 * doesn't get freed by jobs that switch to a different
1055 	 * vmspace.
1056 	 */
1057 	p = td->td_proc;
1058 	myvm = vmspace_acquire_ref(p);
1059 
1060 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1061 
1062 	/*
1063 	 * Allocate and ready the aio control info.  There is one aiop structure
1064 	 * per daemon.
1065 	 */
1066 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1067 	aiop->aioproc = p;
1068 	aiop->aioprocflags = 0;
1069 
1070 	/*
1071 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1072 	 * and creating too many daemons.)
1073 	 */
1074 	sema_post(&aio_newproc_sem);
1075 
1076 	mtx_lock(&aio_job_mtx);
1077 	for (;;) {
1078 		/*
1079 		 * Take daemon off of free queue
1080 		 */
1081 		if (aiop->aioprocflags & AIOP_FREE) {
1082 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1083 			aiop->aioprocflags &= ~AIOP_FREE;
1084 		}
1085 
1086 		/*
1087 		 * Check for jobs.
1088 		 */
1089 		while ((job = aio_selectjob(aiop)) != NULL) {
1090 			mtx_unlock(&aio_job_mtx);
1091 
1092 			ki = job->userproc->p_aioinfo;
1093 			job->handle_fn(job);
1094 
1095 			mtx_lock(&aio_job_mtx);
1096 			/* Decrement the active job count. */
1097 			ki->kaio_active_count--;
1098 		}
1099 
1100 		/*
1101 		 * Disconnect from user address space.
1102 		 */
1103 		if (p->p_vmspace != myvm) {
1104 			mtx_unlock(&aio_job_mtx);
1105 			vmspace_switch_aio(myvm);
1106 			mtx_lock(&aio_job_mtx);
1107 			/*
1108 			 * We have to restart to avoid race, we only sleep if
1109 			 * no job can be selected.
1110 			 */
1111 			continue;
1112 		}
1113 
1114 		mtx_assert(&aio_job_mtx, MA_OWNED);
1115 
1116 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1117 		aiop->aioprocflags |= AIOP_FREE;
1118 
1119 		/*
1120 		 * If daemon is inactive for a long time, allow it to exit,
1121 		 * thereby freeing resources.
1122 		 */
1123 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1124 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1125 		    (aiop->aioprocflags & AIOP_FREE) &&
1126 		    num_aio_procs > target_aio_procs)
1127 			break;
1128 	}
1129 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1130 	num_aio_procs--;
1131 	mtx_unlock(&aio_job_mtx);
1132 	uma_zfree(aiop_zone, aiop);
1133 	free_unr(aiod_unr, id);
1134 	vmspace_free(myvm);
1135 
1136 	KASSERT(p->p_vmspace == myvm,
1137 	    ("AIOD: bad vmspace for exiting daemon"));
1138 	KASSERT(myvm->vm_refcnt > 1,
1139 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1140 	kproc_exit(0);
1141 }
1142 
1143 /*
1144  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1145  * AIO daemon modifies its environment itself.
1146  */
1147 static int
1148 aio_newproc(int *start)
1149 {
1150 	int error;
1151 	struct proc *p;
1152 	int id;
1153 
1154 	id = alloc_unr(aiod_unr);
1155 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1156 		RFNOWAIT, 0, "aiod%d", id);
1157 	if (error == 0) {
1158 		/*
1159 		 * Wait until daemon is started.
1160 		 */
1161 		sema_wait(&aio_newproc_sem);
1162 		mtx_lock(&aio_job_mtx);
1163 		num_aio_procs++;
1164 		if (start != NULL)
1165 			(*start)--;
1166 		mtx_unlock(&aio_job_mtx);
1167 	} else {
1168 		free_unr(aiod_unr, id);
1169 	}
1170 	return (error);
1171 }
1172 
1173 /*
1174  * Try the high-performance, low-overhead physio method for eligible
1175  * VCHR devices.  This method doesn't use an aio helper thread, and
1176  * thus has very low overhead.
1177  *
1178  * Assumes that the caller, aio_aqueue(), has incremented the file
1179  * structure's reference count, preventing its deallocation for the
1180  * duration of this call.
1181  */
1182 static int
1183 aio_qphysio(struct proc *p, struct kaiocb *job)
1184 {
1185 	struct aiocb *cb;
1186 	struct file *fp;
1187 	struct bio *bp;
1188 	struct buf *pbuf;
1189 	struct vnode *vp;
1190 	struct cdevsw *csw;
1191 	struct cdev *dev;
1192 	struct kaioinfo *ki;
1193 	int error, ref, poff;
1194 	vm_prot_t prot;
1195 
1196 	cb = &job->uaiocb;
1197 	fp = job->fd_file;
1198 
1199 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1200 		return (-1);
1201 
1202 	vp = fp->f_vnode;
1203 	if (vp->v_type != VCHR)
1204 		return (-1);
1205 	if (vp->v_bufobj.bo_bsize == 0)
1206 		return (-1);
1207 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1208 		return (-1);
1209 
1210 	ref = 0;
1211 	csw = devvn_refthread(vp, &dev, &ref);
1212 	if (csw == NULL)
1213 		return (ENXIO);
1214 
1215 	if ((csw->d_flags & D_DISK) == 0) {
1216 		error = -1;
1217 		goto unref;
1218 	}
1219 	if (cb->aio_nbytes > dev->si_iosize_max) {
1220 		error = -1;
1221 		goto unref;
1222 	}
1223 
1224 	ki = p->p_aioinfo;
1225 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1226 	if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1227 		if (cb->aio_nbytes > MAXPHYS) {
1228 			error = -1;
1229 			goto unref;
1230 		}
1231 
1232 		pbuf = NULL;
1233 	} else {
1234 		if (cb->aio_nbytes > MAXPHYS - poff) {
1235 			error = -1;
1236 			goto unref;
1237 		}
1238 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1239 			error = -1;
1240 			goto unref;
1241 		}
1242 
1243 		job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1244 		BUF_KERNPROC(pbuf);
1245 		AIO_LOCK(ki);
1246 		ki->kaio_buffer_count++;
1247 		AIO_UNLOCK(ki);
1248 	}
1249 	job->bp = bp = g_alloc_bio();
1250 
1251 	bp->bio_length = cb->aio_nbytes;
1252 	bp->bio_bcount = cb->aio_nbytes;
1253 	bp->bio_done = aio_physwakeup;
1254 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1255 	bp->bio_offset = cb->aio_offset;
1256 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1257 	bp->bio_dev = dev;
1258 	bp->bio_caller1 = (void *)job;
1259 
1260 	prot = VM_PROT_READ;
1261 	if (cb->aio_lio_opcode == LIO_READ)
1262 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1263 	job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1264 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1265 	    nitems(job->pages));
1266 	if (job->npages < 0) {
1267 		error = EFAULT;
1268 		goto doerror;
1269 	}
1270 	if (pbuf != NULL) {
1271 		pmap_qenter((vm_offset_t)pbuf->b_data,
1272 		    job->pages, job->npages);
1273 		bp->bio_data = pbuf->b_data + poff;
1274 		atomic_add_int(&num_buf_aio, 1);
1275 	} else {
1276 		bp->bio_ma = job->pages;
1277 		bp->bio_ma_n = job->npages;
1278 		bp->bio_ma_offset = poff;
1279 		bp->bio_data = unmapped_buf;
1280 		bp->bio_flags |= BIO_UNMAPPED;
1281 	}
1282 
1283 	/* Perform transfer. */
1284 	csw->d_strategy(bp);
1285 	dev_relthread(dev, ref);
1286 	return (0);
1287 
1288 doerror:
1289 	if (pbuf != NULL) {
1290 		AIO_LOCK(ki);
1291 		ki->kaio_buffer_count--;
1292 		AIO_UNLOCK(ki);
1293 		relpbuf(pbuf, NULL);
1294 		job->pbuf = NULL;
1295 	}
1296 	g_destroy_bio(bp);
1297 	job->bp = NULL;
1298 unref:
1299 	dev_relthread(dev, ref);
1300 	return (error);
1301 }
1302 
1303 #ifdef COMPAT_FREEBSD6
1304 static int
1305 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1306 {
1307 
1308 	/*
1309 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1310 	 * supported by AIO with the old sigevent structure.
1311 	 */
1312 	nsig->sigev_notify = osig->sigev_notify;
1313 	switch (nsig->sigev_notify) {
1314 	case SIGEV_NONE:
1315 		break;
1316 	case SIGEV_SIGNAL:
1317 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1318 		break;
1319 	case SIGEV_KEVENT:
1320 		nsig->sigev_notify_kqueue =
1321 		    osig->__sigev_u.__sigev_notify_kqueue;
1322 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1323 		break;
1324 	default:
1325 		return (EINVAL);
1326 	}
1327 	return (0);
1328 }
1329 
1330 static int
1331 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1332 {
1333 	struct oaiocb *ojob;
1334 	int error;
1335 
1336 	bzero(kjob, sizeof(struct aiocb));
1337 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1338 	if (error)
1339 		return (error);
1340 	ojob = (struct oaiocb *)kjob;
1341 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1342 }
1343 #endif
1344 
1345 static int
1346 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1347 {
1348 
1349 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1350 }
1351 
1352 static long
1353 aiocb_fetch_status(struct aiocb *ujob)
1354 {
1355 
1356 	return (fuword(&ujob->_aiocb_private.status));
1357 }
1358 
1359 static long
1360 aiocb_fetch_error(struct aiocb *ujob)
1361 {
1362 
1363 	return (fuword(&ujob->_aiocb_private.error));
1364 }
1365 
1366 static int
1367 aiocb_store_status(struct aiocb *ujob, long status)
1368 {
1369 
1370 	return (suword(&ujob->_aiocb_private.status, status));
1371 }
1372 
1373 static int
1374 aiocb_store_error(struct aiocb *ujob, long error)
1375 {
1376 
1377 	return (suword(&ujob->_aiocb_private.error, error));
1378 }
1379 
1380 static int
1381 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1382 {
1383 
1384 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1385 }
1386 
1387 static int
1388 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1389 {
1390 
1391 	return (suword(ujobp, (long)ujob));
1392 }
1393 
1394 static struct aiocb_ops aiocb_ops = {
1395 	.copyin = aiocb_copyin,
1396 	.fetch_status = aiocb_fetch_status,
1397 	.fetch_error = aiocb_fetch_error,
1398 	.store_status = aiocb_store_status,
1399 	.store_error = aiocb_store_error,
1400 	.store_kernelinfo = aiocb_store_kernelinfo,
1401 	.store_aiocb = aiocb_store_aiocb,
1402 };
1403 
1404 #ifdef COMPAT_FREEBSD6
1405 static struct aiocb_ops aiocb_ops_osigevent = {
1406 	.copyin = aiocb_copyin_old_sigevent,
1407 	.fetch_status = aiocb_fetch_status,
1408 	.fetch_error = aiocb_fetch_error,
1409 	.store_status = aiocb_store_status,
1410 	.store_error = aiocb_store_error,
1411 	.store_kernelinfo = aiocb_store_kernelinfo,
1412 	.store_aiocb = aiocb_store_aiocb,
1413 };
1414 #endif
1415 
1416 /*
1417  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1418  * technique is done in this code.
1419  */
1420 int
1421 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1422 	int type, struct aiocb_ops *ops)
1423 {
1424 	struct proc *p = td->td_proc;
1425 	cap_rights_t rights;
1426 	struct file *fp;
1427 	struct kaiocb *job;
1428 	struct kaioinfo *ki;
1429 	struct kevent kev;
1430 	int opcode;
1431 	int error;
1432 	int fd, kqfd;
1433 	int jid;
1434 	u_short evflags;
1435 
1436 	if (p->p_aioinfo == NULL)
1437 		aio_init_aioinfo(p);
1438 
1439 	ki = p->p_aioinfo;
1440 
1441 	ops->store_status(ujob, -1);
1442 	ops->store_error(ujob, 0);
1443 	ops->store_kernelinfo(ujob, -1);
1444 
1445 	if (num_queue_count >= max_queue_count ||
1446 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1447 		ops->store_error(ujob, EAGAIN);
1448 		return (EAGAIN);
1449 	}
1450 
1451 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1452 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1453 
1454 	error = ops->copyin(ujob, &job->uaiocb);
1455 	if (error) {
1456 		ops->store_error(ujob, error);
1457 		uma_zfree(aiocb_zone, job);
1458 		return (error);
1459 	}
1460 
1461 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1462 		uma_zfree(aiocb_zone, job);
1463 		return (EINVAL);
1464 	}
1465 
1466 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1467 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1468 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1469 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1470 		ops->store_error(ujob, EINVAL);
1471 		uma_zfree(aiocb_zone, job);
1472 		return (EINVAL);
1473 	}
1474 
1475 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1476 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1477 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1478 		uma_zfree(aiocb_zone, job);
1479 		return (EINVAL);
1480 	}
1481 
1482 	ksiginfo_init(&job->ksi);
1483 
1484 	/* Save userspace address of the job info. */
1485 	job->ujob = ujob;
1486 
1487 	/* Get the opcode. */
1488 	if (type != LIO_NOP)
1489 		job->uaiocb.aio_lio_opcode = type;
1490 	opcode = job->uaiocb.aio_lio_opcode;
1491 
1492 	/*
1493 	 * Validate the opcode and fetch the file object for the specified
1494 	 * file descriptor.
1495 	 *
1496 	 * XXXRW: Moved the opcode validation up here so that we don't
1497 	 * retrieve a file descriptor without knowing what the capabiltity
1498 	 * should be.
1499 	 */
1500 	fd = job->uaiocb.aio_fildes;
1501 	switch (opcode) {
1502 	case LIO_WRITE:
1503 		error = fget_write(td, fd,
1504 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1505 		break;
1506 	case LIO_READ:
1507 		error = fget_read(td, fd,
1508 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1509 		break;
1510 	case LIO_SYNC:
1511 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1512 		break;
1513 	case LIO_MLOCK:
1514 		fp = NULL;
1515 		break;
1516 	case LIO_NOP:
1517 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1518 		break;
1519 	default:
1520 		error = EINVAL;
1521 	}
1522 	if (error) {
1523 		uma_zfree(aiocb_zone, job);
1524 		ops->store_error(ujob, error);
1525 		return (error);
1526 	}
1527 
1528 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1529 		error = EINVAL;
1530 		goto aqueue_fail;
1531 	}
1532 
1533 	if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1534 		error = EINVAL;
1535 		goto aqueue_fail;
1536 	}
1537 
1538 	job->fd_file = fp;
1539 
1540 	mtx_lock(&aio_job_mtx);
1541 	jid = jobrefid++;
1542 	job->seqno = jobseqno++;
1543 	mtx_unlock(&aio_job_mtx);
1544 	error = ops->store_kernelinfo(ujob, jid);
1545 	if (error) {
1546 		error = EINVAL;
1547 		goto aqueue_fail;
1548 	}
1549 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1550 
1551 	if (opcode == LIO_NOP) {
1552 		fdrop(fp, td);
1553 		uma_zfree(aiocb_zone, job);
1554 		return (0);
1555 	}
1556 
1557 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1558 		goto no_kqueue;
1559 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1560 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1561 		error = EINVAL;
1562 		goto aqueue_fail;
1563 	}
1564 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1565 	kev.ident = (uintptr_t)job->ujob;
1566 	kev.filter = EVFILT_AIO;
1567 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1568 	kev.data = (intptr_t)job;
1569 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1570 	error = kqfd_register(kqfd, &kev, td, 1);
1571 	if (error)
1572 		goto aqueue_fail;
1573 
1574 no_kqueue:
1575 
1576 	ops->store_error(ujob, EINPROGRESS);
1577 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1578 	job->userproc = p;
1579 	job->cred = crhold(td->td_ucred);
1580 	job->jobflags = KAIOCB_QUEUEING;
1581 	job->lio = lj;
1582 
1583 	if (opcode == LIO_MLOCK) {
1584 		aio_schedule(job, aio_process_mlock);
1585 		error = 0;
1586 	} else if (fp->f_ops->fo_aio_queue == NULL)
1587 		error = aio_queue_file(fp, job);
1588 	else
1589 		error = fo_aio_queue(fp, job);
1590 	if (error)
1591 		goto aqueue_fail;
1592 
1593 	AIO_LOCK(ki);
1594 	job->jobflags &= ~KAIOCB_QUEUEING;
1595 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1596 	ki->kaio_count++;
1597 	if (lj)
1598 		lj->lioj_count++;
1599 	atomic_add_int(&num_queue_count, 1);
1600 	if (job->jobflags & KAIOCB_FINISHED) {
1601 		/*
1602 		 * The queue callback completed the request synchronously.
1603 		 * The bulk of the completion is deferred in that case
1604 		 * until this point.
1605 		 */
1606 		aio_bio_done_notify(p, job);
1607 	} else
1608 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1609 	AIO_UNLOCK(ki);
1610 	return (0);
1611 
1612 aqueue_fail:
1613 	knlist_delete(&job->klist, curthread, 0);
1614 	if (fp)
1615 		fdrop(fp, td);
1616 	uma_zfree(aiocb_zone, job);
1617 	ops->store_error(ujob, error);
1618 	return (error);
1619 }
1620 
1621 static void
1622 aio_cancel_daemon_job(struct kaiocb *job)
1623 {
1624 
1625 	mtx_lock(&aio_job_mtx);
1626 	if (!aio_cancel_cleared(job))
1627 		TAILQ_REMOVE(&aio_jobs, job, list);
1628 	mtx_unlock(&aio_job_mtx);
1629 	aio_cancel(job);
1630 }
1631 
1632 void
1633 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1634 {
1635 
1636 	mtx_lock(&aio_job_mtx);
1637 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1638 		mtx_unlock(&aio_job_mtx);
1639 		aio_cancel(job);
1640 		return;
1641 	}
1642 	job->handle_fn = func;
1643 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1644 	aio_kick_nowait(job->userproc);
1645 	mtx_unlock(&aio_job_mtx);
1646 }
1647 
1648 static void
1649 aio_cancel_sync(struct kaiocb *job)
1650 {
1651 	struct kaioinfo *ki;
1652 
1653 	ki = job->userproc->p_aioinfo;
1654 	mtx_lock(&aio_job_mtx);
1655 	if (!aio_cancel_cleared(job))
1656 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1657 	mtx_unlock(&aio_job_mtx);
1658 	aio_cancel(job);
1659 }
1660 
1661 int
1662 aio_queue_file(struct file *fp, struct kaiocb *job)
1663 {
1664 	struct aioliojob *lj;
1665 	struct kaioinfo *ki;
1666 	struct kaiocb *job2;
1667 	int error, opcode;
1668 
1669 	lj = job->lio;
1670 	ki = job->userproc->p_aioinfo;
1671 	opcode = job->uaiocb.aio_lio_opcode;
1672 	if (opcode == LIO_SYNC)
1673 		goto queueit;
1674 
1675 	if ((error = aio_qphysio(job->userproc, job)) == 0)
1676 		goto done;
1677 #if 0
1678 	/*
1679 	 * XXX: This means qphysio() failed with EFAULT.  The current
1680 	 * behavior is to retry the operation via fo_read/fo_write.
1681 	 * Wouldn't it be better to just complete the request with an
1682 	 * error here?
1683 	 */
1684 	if (error > 0)
1685 		goto done;
1686 #endif
1687 queueit:
1688 	if (!enable_aio_unsafe)
1689 		return (EOPNOTSUPP);
1690 
1691 	if (opcode == LIO_SYNC) {
1692 		AIO_LOCK(ki);
1693 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1694 			if (job2->fd_file == job->fd_file &&
1695 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1696 			    job2->seqno < job->seqno) {
1697 				job2->jobflags |= KAIOCB_CHECKSYNC;
1698 				job->pending++;
1699 			}
1700 		}
1701 		if (job->pending != 0) {
1702 			if (!aio_set_cancel_function(job, aio_cancel_sync)) {
1703 				AIO_UNLOCK(ki);
1704 				aio_cancel(job);
1705 				return (0);
1706 			}
1707 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1708 			AIO_UNLOCK(ki);
1709 			return (0);
1710 		}
1711 		AIO_UNLOCK(ki);
1712 	}
1713 
1714 	switch (opcode) {
1715 	case LIO_READ:
1716 	case LIO_WRITE:
1717 		aio_schedule(job, aio_process_rw);
1718 		error = 0;
1719 		break;
1720 	case LIO_SYNC:
1721 		aio_schedule(job, aio_process_sync);
1722 		error = 0;
1723 		break;
1724 	default:
1725 		error = EINVAL;
1726 	}
1727 done:
1728 	return (error);
1729 }
1730 
1731 static void
1732 aio_kick_nowait(struct proc *userp)
1733 {
1734 	struct kaioinfo *ki = userp->p_aioinfo;
1735 	struct aioproc *aiop;
1736 
1737 	mtx_assert(&aio_job_mtx, MA_OWNED);
1738 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1739 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1740 		aiop->aioprocflags &= ~AIOP_FREE;
1741 		wakeup(aiop->aioproc);
1742 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1743 	    ki->kaio_active_count + num_aio_resv_start <
1744 	    ki->kaio_maxactive_count) {
1745 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1746 	}
1747 }
1748 
1749 static int
1750 aio_kick(struct proc *userp)
1751 {
1752 	struct kaioinfo *ki = userp->p_aioinfo;
1753 	struct aioproc *aiop;
1754 	int error, ret = 0;
1755 
1756 	mtx_assert(&aio_job_mtx, MA_OWNED);
1757 retryproc:
1758 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1759 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1760 		aiop->aioprocflags &= ~AIOP_FREE;
1761 		wakeup(aiop->aioproc);
1762 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1763 	    ki->kaio_active_count + num_aio_resv_start <
1764 	    ki->kaio_maxactive_count) {
1765 		num_aio_resv_start++;
1766 		mtx_unlock(&aio_job_mtx);
1767 		error = aio_newproc(&num_aio_resv_start);
1768 		mtx_lock(&aio_job_mtx);
1769 		if (error) {
1770 			num_aio_resv_start--;
1771 			goto retryproc;
1772 		}
1773 	} else {
1774 		ret = -1;
1775 	}
1776 	return (ret);
1777 }
1778 
1779 static void
1780 aio_kick_helper(void *context, int pending)
1781 {
1782 	struct proc *userp = context;
1783 
1784 	mtx_lock(&aio_job_mtx);
1785 	while (--pending >= 0) {
1786 		if (aio_kick(userp))
1787 			break;
1788 	}
1789 	mtx_unlock(&aio_job_mtx);
1790 }
1791 
1792 /*
1793  * Support the aio_return system call, as a side-effect, kernel resources are
1794  * released.
1795  */
1796 static int
1797 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1798 {
1799 	struct proc *p = td->td_proc;
1800 	struct kaiocb *job;
1801 	struct kaioinfo *ki;
1802 	long status, error;
1803 
1804 	ki = p->p_aioinfo;
1805 	if (ki == NULL)
1806 		return (EINVAL);
1807 	AIO_LOCK(ki);
1808 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1809 		if (job->ujob == ujob)
1810 			break;
1811 	}
1812 	if (job != NULL) {
1813 		MPASS(job->jobflags & KAIOCB_FINISHED);
1814 		status = job->uaiocb._aiocb_private.status;
1815 		error = job->uaiocb._aiocb_private.error;
1816 		td->td_retval[0] = status;
1817 		td->td_ru.ru_oublock += job->outblock;
1818 		td->td_ru.ru_inblock += job->inblock;
1819 		td->td_ru.ru_msgsnd += job->msgsnd;
1820 		td->td_ru.ru_msgrcv += job->msgrcv;
1821 		aio_free_entry(job);
1822 		AIO_UNLOCK(ki);
1823 		ops->store_error(ujob, error);
1824 		ops->store_status(ujob, status);
1825 	} else {
1826 		error = EINVAL;
1827 		AIO_UNLOCK(ki);
1828 	}
1829 	return (error);
1830 }
1831 
1832 int
1833 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1834 {
1835 
1836 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1837 }
1838 
1839 /*
1840  * Allow a process to wakeup when any of the I/O requests are completed.
1841  */
1842 static int
1843 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1844     struct timespec *ts)
1845 {
1846 	struct proc *p = td->td_proc;
1847 	struct timeval atv;
1848 	struct kaioinfo *ki;
1849 	struct kaiocb *firstjob, *job;
1850 	int error, i, timo;
1851 
1852 	timo = 0;
1853 	if (ts) {
1854 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1855 			return (EINVAL);
1856 
1857 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1858 		if (itimerfix(&atv))
1859 			return (EINVAL);
1860 		timo = tvtohz(&atv);
1861 	}
1862 
1863 	ki = p->p_aioinfo;
1864 	if (ki == NULL)
1865 		return (EAGAIN);
1866 
1867 	if (njoblist == 0)
1868 		return (0);
1869 
1870 	AIO_LOCK(ki);
1871 	for (;;) {
1872 		firstjob = NULL;
1873 		error = 0;
1874 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1875 			for (i = 0; i < njoblist; i++) {
1876 				if (job->ujob == ujoblist[i]) {
1877 					if (firstjob == NULL)
1878 						firstjob = job;
1879 					if (job->jobflags & KAIOCB_FINISHED)
1880 						goto RETURN;
1881 				}
1882 			}
1883 		}
1884 		/* All tasks were finished. */
1885 		if (firstjob == NULL)
1886 			break;
1887 
1888 		ki->kaio_flags |= KAIO_WAKEUP;
1889 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1890 		    "aiospn", timo);
1891 		if (error == ERESTART)
1892 			error = EINTR;
1893 		if (error)
1894 			break;
1895 	}
1896 RETURN:
1897 	AIO_UNLOCK(ki);
1898 	return (error);
1899 }
1900 
1901 int
1902 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1903 {
1904 	struct timespec ts, *tsp;
1905 	struct aiocb **ujoblist;
1906 	int error;
1907 
1908 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1909 		return (EINVAL);
1910 
1911 	if (uap->timeout) {
1912 		/* Get timespec struct. */
1913 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1914 			return (error);
1915 		tsp = &ts;
1916 	} else
1917 		tsp = NULL;
1918 
1919 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1920 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1921 	if (error == 0)
1922 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1923 	uma_zfree(aiol_zone, ujoblist);
1924 	return (error);
1925 }
1926 
1927 /*
1928  * aio_cancel cancels any non-physio aio operations not currently in
1929  * progress.
1930  */
1931 int
1932 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1933 {
1934 	struct proc *p = td->td_proc;
1935 	struct kaioinfo *ki;
1936 	struct kaiocb *job, *jobn;
1937 	struct file *fp;
1938 	cap_rights_t rights;
1939 	int error;
1940 	int cancelled = 0;
1941 	int notcancelled = 0;
1942 	struct vnode *vp;
1943 
1944 	/* Lookup file object. */
1945 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1946 	if (error)
1947 		return (error);
1948 
1949 	ki = p->p_aioinfo;
1950 	if (ki == NULL)
1951 		goto done;
1952 
1953 	if (fp->f_type == DTYPE_VNODE) {
1954 		vp = fp->f_vnode;
1955 		if (vn_isdisk(vp, &error)) {
1956 			fdrop(fp, td);
1957 			td->td_retval[0] = AIO_NOTCANCELED;
1958 			return (0);
1959 		}
1960 	}
1961 
1962 	AIO_LOCK(ki);
1963 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1964 		if ((uap->fd == job->uaiocb.aio_fildes) &&
1965 		    ((uap->aiocbp == NULL) ||
1966 		     (uap->aiocbp == job->ujob))) {
1967 			if (aio_cancel_job(p, ki, job)) {
1968 				cancelled++;
1969 			} else {
1970 				notcancelled++;
1971 			}
1972 			if (uap->aiocbp != NULL)
1973 				break;
1974 		}
1975 	}
1976 	AIO_UNLOCK(ki);
1977 
1978 done:
1979 	fdrop(fp, td);
1980 
1981 	if (uap->aiocbp != NULL) {
1982 		if (cancelled) {
1983 			td->td_retval[0] = AIO_CANCELED;
1984 			return (0);
1985 		}
1986 	}
1987 
1988 	if (notcancelled) {
1989 		td->td_retval[0] = AIO_NOTCANCELED;
1990 		return (0);
1991 	}
1992 
1993 	if (cancelled) {
1994 		td->td_retval[0] = AIO_CANCELED;
1995 		return (0);
1996 	}
1997 
1998 	td->td_retval[0] = AIO_ALLDONE;
1999 
2000 	return (0);
2001 }
2002 
2003 /*
2004  * aio_error is implemented in the kernel level for compatibility purposes
2005  * only.  For a user mode async implementation, it would be best to do it in
2006  * a userland subroutine.
2007  */
2008 static int
2009 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2010 {
2011 	struct proc *p = td->td_proc;
2012 	struct kaiocb *job;
2013 	struct kaioinfo *ki;
2014 	int status;
2015 
2016 	ki = p->p_aioinfo;
2017 	if (ki == NULL) {
2018 		td->td_retval[0] = EINVAL;
2019 		return (0);
2020 	}
2021 
2022 	AIO_LOCK(ki);
2023 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2024 		if (job->ujob == ujob) {
2025 			if (job->jobflags & KAIOCB_FINISHED)
2026 				td->td_retval[0] =
2027 					job->uaiocb._aiocb_private.error;
2028 			else
2029 				td->td_retval[0] = EINPROGRESS;
2030 			AIO_UNLOCK(ki);
2031 			return (0);
2032 		}
2033 	}
2034 	AIO_UNLOCK(ki);
2035 
2036 	/*
2037 	 * Hack for failure of aio_aqueue.
2038 	 */
2039 	status = ops->fetch_status(ujob);
2040 	if (status == -1) {
2041 		td->td_retval[0] = ops->fetch_error(ujob);
2042 		return (0);
2043 	}
2044 
2045 	td->td_retval[0] = EINVAL;
2046 	return (0);
2047 }
2048 
2049 int
2050 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2051 {
2052 
2053 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2054 }
2055 
2056 /* syscall - asynchronous read from a file (REALTIME) */
2057 #ifdef COMPAT_FREEBSD6
2058 int
2059 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2060 {
2061 
2062 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2063 	    &aiocb_ops_osigevent));
2064 }
2065 #endif
2066 
2067 int
2068 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2069 {
2070 
2071 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2072 }
2073 
2074 /* syscall - asynchronous write to a file (REALTIME) */
2075 #ifdef COMPAT_FREEBSD6
2076 int
2077 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2078 {
2079 
2080 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2081 	    &aiocb_ops_osigevent));
2082 }
2083 #endif
2084 
2085 int
2086 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2087 {
2088 
2089 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2090 }
2091 
2092 int
2093 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2094 {
2095 
2096 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2097 }
2098 
2099 static int
2100 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2101     struct aiocb **acb_list, int nent, struct sigevent *sig,
2102     struct aiocb_ops *ops)
2103 {
2104 	struct proc *p = td->td_proc;
2105 	struct aiocb *job;
2106 	struct kaioinfo *ki;
2107 	struct aioliojob *lj;
2108 	struct kevent kev;
2109 	int error;
2110 	int nerror;
2111 	int i;
2112 
2113 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2114 		return (EINVAL);
2115 
2116 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2117 		return (EINVAL);
2118 
2119 	if (p->p_aioinfo == NULL)
2120 		aio_init_aioinfo(p);
2121 
2122 	ki = p->p_aioinfo;
2123 
2124 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2125 	lj->lioj_flags = 0;
2126 	lj->lioj_count = 0;
2127 	lj->lioj_finished_count = 0;
2128 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2129 	ksiginfo_init(&lj->lioj_ksi);
2130 
2131 	/*
2132 	 * Setup signal.
2133 	 */
2134 	if (sig && (mode == LIO_NOWAIT)) {
2135 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2136 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2137 			/* Assume only new style KEVENT */
2138 			kev.filter = EVFILT_LIO;
2139 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2140 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2141 			kev.data = (intptr_t)lj;
2142 			/* pass user defined sigval data */
2143 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2144 			error = kqfd_register(
2145 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2146 			if (error) {
2147 				uma_zfree(aiolio_zone, lj);
2148 				return (error);
2149 			}
2150 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2151 			;
2152 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2153 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2154 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2155 					uma_zfree(aiolio_zone, lj);
2156 					return EINVAL;
2157 				}
2158 				lj->lioj_flags |= LIOJ_SIGNAL;
2159 		} else {
2160 			uma_zfree(aiolio_zone, lj);
2161 			return EINVAL;
2162 		}
2163 	}
2164 
2165 	AIO_LOCK(ki);
2166 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2167 	/*
2168 	 * Add extra aiocb count to avoid the lio to be freed
2169 	 * by other threads doing aio_waitcomplete or aio_return,
2170 	 * and prevent event from being sent until we have queued
2171 	 * all tasks.
2172 	 */
2173 	lj->lioj_count = 1;
2174 	AIO_UNLOCK(ki);
2175 
2176 	/*
2177 	 * Get pointers to the list of I/O requests.
2178 	 */
2179 	nerror = 0;
2180 	for (i = 0; i < nent; i++) {
2181 		job = acb_list[i];
2182 		if (job != NULL) {
2183 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2184 			if (error != 0)
2185 				nerror++;
2186 		}
2187 	}
2188 
2189 	error = 0;
2190 	AIO_LOCK(ki);
2191 	if (mode == LIO_WAIT) {
2192 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2193 			ki->kaio_flags |= KAIO_WAKEUP;
2194 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2195 			    PRIBIO | PCATCH, "aiospn", 0);
2196 			if (error == ERESTART)
2197 				error = EINTR;
2198 			if (error)
2199 				break;
2200 		}
2201 	} else {
2202 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2203 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2204 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2205 				KNOTE_LOCKED(&lj->klist, 1);
2206 			}
2207 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2208 			    == LIOJ_SIGNAL
2209 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2210 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2211 				aio_sendsig(p, &lj->lioj_signal,
2212 					    &lj->lioj_ksi);
2213 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2214 			}
2215 		}
2216 	}
2217 	lj->lioj_count--;
2218 	if (lj->lioj_count == 0) {
2219 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2220 		knlist_delete(&lj->klist, curthread, 1);
2221 		PROC_LOCK(p);
2222 		sigqueue_take(&lj->lioj_ksi);
2223 		PROC_UNLOCK(p);
2224 		AIO_UNLOCK(ki);
2225 		uma_zfree(aiolio_zone, lj);
2226 	} else
2227 		AIO_UNLOCK(ki);
2228 
2229 	if (nerror)
2230 		return (EIO);
2231 	return (error);
2232 }
2233 
2234 /* syscall - list directed I/O (REALTIME) */
2235 #ifdef COMPAT_FREEBSD6
2236 int
2237 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2238 {
2239 	struct aiocb **acb_list;
2240 	struct sigevent *sigp, sig;
2241 	struct osigevent osig;
2242 	int error, nent;
2243 
2244 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2245 		return (EINVAL);
2246 
2247 	nent = uap->nent;
2248 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2249 		return (EINVAL);
2250 
2251 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2252 		error = copyin(uap->sig, &osig, sizeof(osig));
2253 		if (error)
2254 			return (error);
2255 		error = convert_old_sigevent(&osig, &sig);
2256 		if (error)
2257 			return (error);
2258 		sigp = &sig;
2259 	} else
2260 		sigp = NULL;
2261 
2262 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2263 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2264 	if (error == 0)
2265 		error = kern_lio_listio(td, uap->mode,
2266 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2267 		    &aiocb_ops_osigevent);
2268 	free(acb_list, M_LIO);
2269 	return (error);
2270 }
2271 #endif
2272 
2273 /* syscall - list directed I/O (REALTIME) */
2274 int
2275 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2276 {
2277 	struct aiocb **acb_list;
2278 	struct sigevent *sigp, sig;
2279 	int error, nent;
2280 
2281 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2282 		return (EINVAL);
2283 
2284 	nent = uap->nent;
2285 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2286 		return (EINVAL);
2287 
2288 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2289 		error = copyin(uap->sig, &sig, sizeof(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, uap->acb_list, acb_list,
2300 		    nent, sigp, &aiocb_ops);
2301 	free(acb_list, M_LIO);
2302 	return (error);
2303 }
2304 
2305 static void
2306 aio_physwakeup(struct bio *bp)
2307 {
2308 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2309 	struct proc *userp;
2310 	struct kaioinfo *ki;
2311 	size_t nbytes;
2312 	int error, nblks;
2313 
2314 	/* Release mapping into kernel space. */
2315 	userp = job->userproc;
2316 	ki = userp->p_aioinfo;
2317 	if (job->pbuf) {
2318 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2319 		relpbuf(job->pbuf, NULL);
2320 		job->pbuf = NULL;
2321 		atomic_subtract_int(&num_buf_aio, 1);
2322 		AIO_LOCK(ki);
2323 		ki->kaio_buffer_count--;
2324 		AIO_UNLOCK(ki);
2325 	}
2326 	vm_page_unhold_pages(job->pages, job->npages);
2327 
2328 	bp = job->bp;
2329 	job->bp = NULL;
2330 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2331 	error = 0;
2332 	if (bp->bio_flags & BIO_ERROR)
2333 		error = bp->bio_error;
2334 	nblks = btodb(nbytes);
2335 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2336 		job->outblock += nblks;
2337 	else
2338 		job->inblock += nblks;
2339 
2340 	if (error)
2341 		aio_complete(job, -1, error);
2342 	else
2343 		aio_complete(job, nbytes, 0);
2344 
2345 	g_destroy_bio(bp);
2346 }
2347 
2348 /* syscall - wait for the next completion of an aio request */
2349 static int
2350 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2351     struct timespec *ts, struct aiocb_ops *ops)
2352 {
2353 	struct proc *p = td->td_proc;
2354 	struct timeval atv;
2355 	struct kaioinfo *ki;
2356 	struct kaiocb *job;
2357 	struct aiocb *ujob;
2358 	long error, status;
2359 	int timo;
2360 
2361 	ops->store_aiocb(ujobp, NULL);
2362 
2363 	if (ts == NULL) {
2364 		timo = 0;
2365 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2366 		timo = -1;
2367 	} else {
2368 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2369 			return (EINVAL);
2370 
2371 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2372 		if (itimerfix(&atv))
2373 			return (EINVAL);
2374 		timo = tvtohz(&atv);
2375 	}
2376 
2377 	if (p->p_aioinfo == NULL)
2378 		aio_init_aioinfo(p);
2379 	ki = p->p_aioinfo;
2380 
2381 	error = 0;
2382 	job = NULL;
2383 	AIO_LOCK(ki);
2384 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2385 		if (timo == -1) {
2386 			error = EWOULDBLOCK;
2387 			break;
2388 		}
2389 		ki->kaio_flags |= KAIO_WAKEUP;
2390 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2391 		    "aiowc", timo);
2392 		if (timo && error == ERESTART)
2393 			error = EINTR;
2394 		if (error)
2395 			break;
2396 	}
2397 
2398 	if (job != NULL) {
2399 		MPASS(job->jobflags & KAIOCB_FINISHED);
2400 		ujob = job->ujob;
2401 		status = job->uaiocb._aiocb_private.status;
2402 		error = job->uaiocb._aiocb_private.error;
2403 		td->td_retval[0] = status;
2404 		td->td_ru.ru_oublock += job->outblock;
2405 		td->td_ru.ru_inblock += job->inblock;
2406 		td->td_ru.ru_msgsnd += job->msgsnd;
2407 		td->td_ru.ru_msgrcv += job->msgrcv;
2408 		aio_free_entry(job);
2409 		AIO_UNLOCK(ki);
2410 		ops->store_aiocb(ujobp, ujob);
2411 		ops->store_error(ujob, error);
2412 		ops->store_status(ujob, status);
2413 	} else
2414 		AIO_UNLOCK(ki);
2415 
2416 	return (error);
2417 }
2418 
2419 int
2420 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2421 {
2422 	struct timespec ts, *tsp;
2423 	int error;
2424 
2425 	if (uap->timeout) {
2426 		/* Get timespec struct. */
2427 		error = copyin(uap->timeout, &ts, sizeof(ts));
2428 		if (error)
2429 			return (error);
2430 		tsp = &ts;
2431 	} else
2432 		tsp = NULL;
2433 
2434 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2435 }
2436 
2437 static int
2438 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2439     struct aiocb_ops *ops)
2440 {
2441 	struct proc *p = td->td_proc;
2442 	struct kaioinfo *ki;
2443 
2444 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2445 		return (EINVAL);
2446 	ki = p->p_aioinfo;
2447 	if (ki == NULL)
2448 		aio_init_aioinfo(p);
2449 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2450 }
2451 
2452 int
2453 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2454 {
2455 
2456 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2457 }
2458 
2459 /* kqueue attach function */
2460 static int
2461 filt_aioattach(struct knote *kn)
2462 {
2463 	struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2464 
2465 	/*
2466 	 * The job pointer must be validated before using it, so
2467 	 * registration is restricted to the kernel; the user cannot
2468 	 * set EV_FLAG1.
2469 	 */
2470 	if ((kn->kn_flags & EV_FLAG1) == 0)
2471 		return (EPERM);
2472 	kn->kn_ptr.p_aio = job;
2473 	kn->kn_flags &= ~EV_FLAG1;
2474 
2475 	knlist_add(&job->klist, kn, 0);
2476 
2477 	return (0);
2478 }
2479 
2480 /* kqueue detach function */
2481 static void
2482 filt_aiodetach(struct knote *kn)
2483 {
2484 	struct knlist *knl;
2485 
2486 	knl = &kn->kn_ptr.p_aio->klist;
2487 	knl->kl_lock(knl->kl_lockarg);
2488 	if (!knlist_empty(knl))
2489 		knlist_remove(knl, kn, 1);
2490 	knl->kl_unlock(knl->kl_lockarg);
2491 }
2492 
2493 /* kqueue filter function */
2494 /*ARGSUSED*/
2495 static int
2496 filt_aio(struct knote *kn, long hint)
2497 {
2498 	struct kaiocb *job = kn->kn_ptr.p_aio;
2499 
2500 	kn->kn_data = job->uaiocb._aiocb_private.error;
2501 	if (!(job->jobflags & KAIOCB_FINISHED))
2502 		return (0);
2503 	kn->kn_flags |= EV_EOF;
2504 	return (1);
2505 }
2506 
2507 /* kqueue attach function */
2508 static int
2509 filt_lioattach(struct knote *kn)
2510 {
2511 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2512 
2513 	/*
2514 	 * The aioliojob pointer must be validated before using it, so
2515 	 * registration is restricted to the kernel; the user cannot
2516 	 * set EV_FLAG1.
2517 	 */
2518 	if ((kn->kn_flags & EV_FLAG1) == 0)
2519 		return (EPERM);
2520 	kn->kn_ptr.p_lio = lj;
2521 	kn->kn_flags &= ~EV_FLAG1;
2522 
2523 	knlist_add(&lj->klist, kn, 0);
2524 
2525 	return (0);
2526 }
2527 
2528 /* kqueue detach function */
2529 static void
2530 filt_liodetach(struct knote *kn)
2531 {
2532 	struct knlist *knl;
2533 
2534 	knl = &kn->kn_ptr.p_lio->klist;
2535 	knl->kl_lock(knl->kl_lockarg);
2536 	if (!knlist_empty(knl))
2537 		knlist_remove(knl, kn, 1);
2538 	knl->kl_unlock(knl->kl_lockarg);
2539 }
2540 
2541 /* kqueue filter function */
2542 /*ARGSUSED*/
2543 static int
2544 filt_lio(struct knote *kn, long hint)
2545 {
2546 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2547 
2548 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2549 }
2550 
2551 #ifdef COMPAT_FREEBSD32
2552 #include <sys/mount.h>
2553 #include <sys/socket.h>
2554 #include <compat/freebsd32/freebsd32.h>
2555 #include <compat/freebsd32/freebsd32_proto.h>
2556 #include <compat/freebsd32/freebsd32_signal.h>
2557 #include <compat/freebsd32/freebsd32_syscall.h>
2558 #include <compat/freebsd32/freebsd32_util.h>
2559 
2560 struct __aiocb_private32 {
2561 	int32_t	status;
2562 	int32_t	error;
2563 	uint32_t kernelinfo;
2564 };
2565 
2566 #ifdef COMPAT_FREEBSD6
2567 typedef struct oaiocb32 {
2568 	int	aio_fildes;		/* File descriptor */
2569 	uint64_t aio_offset __packed;	/* File offset for I/O */
2570 	uint32_t aio_buf;		/* I/O buffer in process space */
2571 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2572 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2573 	int	aio_lio_opcode;		/* LIO opcode */
2574 	int	aio_reqprio;		/* Request priority -- ignored */
2575 	struct	__aiocb_private32 _aiocb_private;
2576 } oaiocb32_t;
2577 #endif
2578 
2579 typedef struct aiocb32 {
2580 	int32_t	aio_fildes;		/* File descriptor */
2581 	uint64_t aio_offset __packed;	/* File offset for I/O */
2582 	uint32_t aio_buf;		/* I/O buffer in process space */
2583 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2584 	int	__spare__[2];
2585 	uint32_t __spare2__;
2586 	int	aio_lio_opcode;		/* LIO opcode */
2587 	int	aio_reqprio;		/* Request priority -- ignored */
2588 	struct	__aiocb_private32 _aiocb_private;
2589 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2590 } aiocb32_t;
2591 
2592 #ifdef COMPAT_FREEBSD6
2593 static int
2594 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2595 {
2596 
2597 	/*
2598 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2599 	 * supported by AIO with the old sigevent structure.
2600 	 */
2601 	CP(*osig, *nsig, sigev_notify);
2602 	switch (nsig->sigev_notify) {
2603 	case SIGEV_NONE:
2604 		break;
2605 	case SIGEV_SIGNAL:
2606 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2607 		break;
2608 	case SIGEV_KEVENT:
2609 		nsig->sigev_notify_kqueue =
2610 		    osig->__sigev_u.__sigev_notify_kqueue;
2611 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2612 		break;
2613 	default:
2614 		return (EINVAL);
2615 	}
2616 	return (0);
2617 }
2618 
2619 static int
2620 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2621 {
2622 	struct oaiocb32 job32;
2623 	int error;
2624 
2625 	bzero(kjob, sizeof(struct aiocb));
2626 	error = copyin(ujob, &job32, sizeof(job32));
2627 	if (error)
2628 		return (error);
2629 
2630 	CP(job32, *kjob, aio_fildes);
2631 	CP(job32, *kjob, aio_offset);
2632 	PTRIN_CP(job32, *kjob, aio_buf);
2633 	CP(job32, *kjob, aio_nbytes);
2634 	CP(job32, *kjob, aio_lio_opcode);
2635 	CP(job32, *kjob, aio_reqprio);
2636 	CP(job32, *kjob, _aiocb_private.status);
2637 	CP(job32, *kjob, _aiocb_private.error);
2638 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2639 	return (convert_old_sigevent32(&job32.aio_sigevent,
2640 	    &kjob->aio_sigevent));
2641 }
2642 #endif
2643 
2644 static int
2645 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2646 {
2647 	struct aiocb32 job32;
2648 	int error;
2649 
2650 	error = copyin(ujob, &job32, sizeof(job32));
2651 	if (error)
2652 		return (error);
2653 	CP(job32, *kjob, aio_fildes);
2654 	CP(job32, *kjob, aio_offset);
2655 	PTRIN_CP(job32, *kjob, aio_buf);
2656 	CP(job32, *kjob, aio_nbytes);
2657 	CP(job32, *kjob, aio_lio_opcode);
2658 	CP(job32, *kjob, aio_reqprio);
2659 	CP(job32, *kjob, _aiocb_private.status);
2660 	CP(job32, *kjob, _aiocb_private.error);
2661 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2662 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2663 }
2664 
2665 static long
2666 aiocb32_fetch_status(struct aiocb *ujob)
2667 {
2668 	struct aiocb32 *ujob32;
2669 
2670 	ujob32 = (struct aiocb32 *)ujob;
2671 	return (fuword32(&ujob32->_aiocb_private.status));
2672 }
2673 
2674 static long
2675 aiocb32_fetch_error(struct aiocb *ujob)
2676 {
2677 	struct aiocb32 *ujob32;
2678 
2679 	ujob32 = (struct aiocb32 *)ujob;
2680 	return (fuword32(&ujob32->_aiocb_private.error));
2681 }
2682 
2683 static int
2684 aiocb32_store_status(struct aiocb *ujob, long status)
2685 {
2686 	struct aiocb32 *ujob32;
2687 
2688 	ujob32 = (struct aiocb32 *)ujob;
2689 	return (suword32(&ujob32->_aiocb_private.status, status));
2690 }
2691 
2692 static int
2693 aiocb32_store_error(struct aiocb *ujob, long error)
2694 {
2695 	struct aiocb32 *ujob32;
2696 
2697 	ujob32 = (struct aiocb32 *)ujob;
2698 	return (suword32(&ujob32->_aiocb_private.error, error));
2699 }
2700 
2701 static int
2702 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2703 {
2704 	struct aiocb32 *ujob32;
2705 
2706 	ujob32 = (struct aiocb32 *)ujob;
2707 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2708 }
2709 
2710 static int
2711 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2712 {
2713 
2714 	return (suword32(ujobp, (long)ujob));
2715 }
2716 
2717 static struct aiocb_ops aiocb32_ops = {
2718 	.copyin = aiocb32_copyin,
2719 	.fetch_status = aiocb32_fetch_status,
2720 	.fetch_error = aiocb32_fetch_error,
2721 	.store_status = aiocb32_store_status,
2722 	.store_error = aiocb32_store_error,
2723 	.store_kernelinfo = aiocb32_store_kernelinfo,
2724 	.store_aiocb = aiocb32_store_aiocb,
2725 };
2726 
2727 #ifdef COMPAT_FREEBSD6
2728 static struct aiocb_ops aiocb32_ops_osigevent = {
2729 	.copyin = aiocb32_copyin_old_sigevent,
2730 	.fetch_status = aiocb32_fetch_status,
2731 	.fetch_error = aiocb32_fetch_error,
2732 	.store_status = aiocb32_store_status,
2733 	.store_error = aiocb32_store_error,
2734 	.store_kernelinfo = aiocb32_store_kernelinfo,
2735 	.store_aiocb = aiocb32_store_aiocb,
2736 };
2737 #endif
2738 
2739 int
2740 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2741 {
2742 
2743 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2744 }
2745 
2746 int
2747 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2748 {
2749 	struct timespec32 ts32;
2750 	struct timespec ts, *tsp;
2751 	struct aiocb **ujoblist;
2752 	uint32_t *ujoblist32;
2753 	int error, i;
2754 
2755 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2756 		return (EINVAL);
2757 
2758 	if (uap->timeout) {
2759 		/* Get timespec struct. */
2760 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2761 			return (error);
2762 		CP(ts32, ts, tv_sec);
2763 		CP(ts32, ts, tv_nsec);
2764 		tsp = &ts;
2765 	} else
2766 		tsp = NULL;
2767 
2768 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2769 	ujoblist32 = (uint32_t *)ujoblist;
2770 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2771 	    sizeof(ujoblist32[0]));
2772 	if (error == 0) {
2773 		for (i = uap->nent; i > 0; i--)
2774 			ujoblist[i] = PTRIN(ujoblist32[i]);
2775 
2776 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2777 	}
2778 	uma_zfree(aiol_zone, ujoblist);
2779 	return (error);
2780 }
2781 
2782 int
2783 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2784 {
2785 
2786 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2787 }
2788 
2789 #ifdef COMPAT_FREEBSD6
2790 int
2791 freebsd6_freebsd32_aio_read(struct thread *td,
2792     struct freebsd6_freebsd32_aio_read_args *uap)
2793 {
2794 
2795 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2796 	    &aiocb32_ops_osigevent));
2797 }
2798 #endif
2799 
2800 int
2801 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2802 {
2803 
2804 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2805 	    &aiocb32_ops));
2806 }
2807 
2808 #ifdef COMPAT_FREEBSD6
2809 int
2810 freebsd6_freebsd32_aio_write(struct thread *td,
2811     struct freebsd6_freebsd32_aio_write_args *uap)
2812 {
2813 
2814 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2815 	    &aiocb32_ops_osigevent));
2816 }
2817 #endif
2818 
2819 int
2820 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2821 {
2822 
2823 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2824 	    &aiocb32_ops));
2825 }
2826 
2827 int
2828 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2829 {
2830 
2831 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2832 	    &aiocb32_ops));
2833 }
2834 
2835 int
2836 freebsd32_aio_waitcomplete(struct thread *td,
2837     struct freebsd32_aio_waitcomplete_args *uap)
2838 {
2839 	struct timespec32 ts32;
2840 	struct timespec ts, *tsp;
2841 	int error;
2842 
2843 	if (uap->timeout) {
2844 		/* Get timespec struct. */
2845 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2846 		if (error)
2847 			return (error);
2848 		CP(ts32, ts, tv_sec);
2849 		CP(ts32, ts, tv_nsec);
2850 		tsp = &ts;
2851 	} else
2852 		tsp = NULL;
2853 
2854 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2855 	    &aiocb32_ops));
2856 }
2857 
2858 int
2859 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2860 {
2861 
2862 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2863 	    &aiocb32_ops));
2864 }
2865 
2866 #ifdef COMPAT_FREEBSD6
2867 int
2868 freebsd6_freebsd32_lio_listio(struct thread *td,
2869     struct freebsd6_freebsd32_lio_listio_args *uap)
2870 {
2871 	struct aiocb **acb_list;
2872 	struct sigevent *sigp, sig;
2873 	struct osigevent32 osig;
2874 	uint32_t *acb_list32;
2875 	int error, i, nent;
2876 
2877 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2878 		return (EINVAL);
2879 
2880 	nent = uap->nent;
2881 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2882 		return (EINVAL);
2883 
2884 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2885 		error = copyin(uap->sig, &osig, sizeof(osig));
2886 		if (error)
2887 			return (error);
2888 		error = convert_old_sigevent32(&osig, &sig);
2889 		if (error)
2890 			return (error);
2891 		sigp = &sig;
2892 	} else
2893 		sigp = NULL;
2894 
2895 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2896 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2897 	if (error) {
2898 		free(acb_list32, M_LIO);
2899 		return (error);
2900 	}
2901 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2902 	for (i = 0; i < nent; i++)
2903 		acb_list[i] = PTRIN(acb_list32[i]);
2904 	free(acb_list32, M_LIO);
2905 
2906 	error = kern_lio_listio(td, uap->mode,
2907 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2908 	    &aiocb32_ops_osigevent);
2909 	free(acb_list, M_LIO);
2910 	return (error);
2911 }
2912 #endif
2913 
2914 int
2915 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2916 {
2917 	struct aiocb **acb_list;
2918 	struct sigevent *sigp, sig;
2919 	struct sigevent32 sig32;
2920 	uint32_t *acb_list32;
2921 	int error, i, nent;
2922 
2923 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2924 		return (EINVAL);
2925 
2926 	nent = uap->nent;
2927 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2928 		return (EINVAL);
2929 
2930 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2931 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2932 		if (error)
2933 			return (error);
2934 		error = convert_sigevent32(&sig32, &sig);
2935 		if (error)
2936 			return (error);
2937 		sigp = &sig;
2938 	} else
2939 		sigp = NULL;
2940 
2941 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2942 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2943 	if (error) {
2944 		free(acb_list32, M_LIO);
2945 		return (error);
2946 	}
2947 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2948 	for (i = 0; i < nent; i++)
2949 		acb_list[i] = PTRIN(acb_list32[i]);
2950 	free(acb_list32, M_LIO);
2951 
2952 	error = kern_lio_listio(td, uap->mode,
2953 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2954 	    &aiocb32_ops);
2955 	free(acb_list, M_LIO);
2956 	return (error);
2957 }
2958 
2959 #endif
2960