xref: /freebsd/sys/kern/vfs_aio.c (revision 7d536dc855c85c15bf45f033d108a61b1f3cecc3)
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 	MPASS(job->bp == NULL);
524 	AIO_UNLOCK(ki);
525 
526 	/*
527 	 * The thread argument here is used to find the owning process
528 	 * and is also passed to fo_close() which may pass it to various
529 	 * places such as devsw close() routines.  Because of that, we
530 	 * need a thread pointer from the process owning the job that is
531 	 * persistent and won't disappear out from under us or move to
532 	 * another process.
533 	 *
534 	 * Currently, all the callers of this function call it to remove
535 	 * a kaiocb from the current process' job list either via a
536 	 * syscall or due to the current process calling exit() or
537 	 * execve().  Thus, we know that p == curproc.  We also know that
538 	 * curthread can't exit since we are curthread.
539 	 *
540 	 * Therefore, we use curthread as the thread to pass to
541 	 * knlist_delete().  This does mean that it is possible for the
542 	 * thread pointer at close time to differ from the thread pointer
543 	 * at open time, but this is already true of file descriptors in
544 	 * a multithreaded process.
545 	 */
546 	if (job->fd_file)
547 		fdrop(job->fd_file, curthread);
548 	crfree(job->cred);
549 	uma_zfree(aiocb_zone, job);
550 	AIO_LOCK(ki);
551 
552 	return (0);
553 }
554 
555 static void
556 aio_proc_rundown_exec(void *arg, struct proc *p,
557     struct image_params *imgp __unused)
558 {
559    	aio_proc_rundown(arg, p);
560 }
561 
562 static int
563 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
564 {
565 	aio_cancel_fn_t *func;
566 	int cancelled;
567 
568 	AIO_LOCK_ASSERT(ki, MA_OWNED);
569 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
570 		return (0);
571 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
572 	job->jobflags |= KAIOCB_CANCELLED;
573 
574 	func = job->cancel_fn;
575 
576 	/*
577 	 * If there is no cancel routine, just leave the job marked as
578 	 * cancelled.  The job should be in active use by a caller who
579 	 * should complete it normally or when it fails to install a
580 	 * cancel routine.
581 	 */
582 	if (func == NULL)
583 		return (0);
584 
585 	/*
586 	 * Set the CANCELLING flag so that aio_complete() will defer
587 	 * completions of this job.  This prevents the job from being
588 	 * freed out from under the cancel callback.  After the
589 	 * callback any deferred completion (whether from the callback
590 	 * or any other source) will be completed.
591 	 */
592 	job->jobflags |= KAIOCB_CANCELLING;
593 	AIO_UNLOCK(ki);
594 	func(job);
595 	AIO_LOCK(ki);
596 	job->jobflags &= ~KAIOCB_CANCELLING;
597 	if (job->jobflags & KAIOCB_FINISHED) {
598 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
599 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
600 		aio_bio_done_notify(p, job);
601 	} else {
602 		/*
603 		 * The cancel callback might have scheduled an
604 		 * operation to cancel this request, but it is
605 		 * only counted as cancelled if the request is
606 		 * cancelled when the callback returns.
607 		 */
608 		cancelled = 0;
609 	}
610 	return (cancelled);
611 }
612 
613 /*
614  * Rundown the jobs for a given process.
615  */
616 static void
617 aio_proc_rundown(void *arg, struct proc *p)
618 {
619 	struct kaioinfo *ki;
620 	struct aioliojob *lj;
621 	struct kaiocb *job, *jobn;
622 
623 	KASSERT(curthread->td_proc == p,
624 	    ("%s: called on non-curproc", __func__));
625 	ki = p->p_aioinfo;
626 	if (ki == NULL)
627 		return;
628 
629 	AIO_LOCK(ki);
630 	ki->kaio_flags |= KAIO_RUNDOWN;
631 
632 restart:
633 
634 	/*
635 	 * Try to cancel all pending requests. This code simulates
636 	 * aio_cancel on all pending I/O requests.
637 	 */
638 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
639 		aio_cancel_job(p, ki, job);
640 	}
641 
642 	/* Wait for all running I/O to be finished */
643 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
644 		ki->kaio_flags |= KAIO_WAKEUP;
645 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
646 		goto restart;
647 	}
648 
649 	/* Free all completed I/O requests. */
650 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
651 		aio_free_entry(job);
652 
653 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
654 		if (lj->lioj_count == 0) {
655 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
656 			knlist_delete(&lj->klist, curthread, 1);
657 			PROC_LOCK(p);
658 			sigqueue_take(&lj->lioj_ksi);
659 			PROC_UNLOCK(p);
660 			uma_zfree(aiolio_zone, lj);
661 		} else {
662 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
663 			    lj->lioj_count, lj->lioj_finished_count);
664 		}
665 	}
666 	AIO_UNLOCK(ki);
667 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
668 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
669 	mtx_destroy(&ki->kaio_mtx);
670 	uma_zfree(kaio_zone, ki);
671 	p->p_aioinfo = NULL;
672 }
673 
674 /*
675  * Select a job to run (called by an AIO daemon).
676  */
677 static struct kaiocb *
678 aio_selectjob(struct aioproc *aiop)
679 {
680 	struct kaiocb *job;
681 	struct kaioinfo *ki;
682 	struct proc *userp;
683 
684 	mtx_assert(&aio_job_mtx, MA_OWNED);
685 restart:
686 	TAILQ_FOREACH(job, &aio_jobs, list) {
687 		userp = job->userproc;
688 		ki = userp->p_aioinfo;
689 
690 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
691 			TAILQ_REMOVE(&aio_jobs, job, list);
692 			if (!aio_clear_cancel_function(job))
693 				goto restart;
694 
695 			/* Account for currently active jobs. */
696 			ki->kaio_active_count++;
697 			break;
698 		}
699 	}
700 	return (job);
701 }
702 
703 /*
704  * Move all data to a permanent storage device.  This code
705  * simulates the fsync syscall.
706  */
707 static int
708 aio_fsync_vnode(struct thread *td, struct vnode *vp)
709 {
710 	struct mount *mp;
711 	int error;
712 
713 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
714 		goto drop;
715 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
716 	if (vp->v_object != NULL) {
717 		VM_OBJECT_WLOCK(vp->v_object);
718 		vm_object_page_clean(vp->v_object, 0, 0, 0);
719 		VM_OBJECT_WUNLOCK(vp->v_object);
720 	}
721 	error = VOP_FSYNC(vp, MNT_WAIT, td);
722 
723 	VOP_UNLOCK(vp, 0);
724 	vn_finished_write(mp);
725 drop:
726 	return (error);
727 }
728 
729 /*
730  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
731  * does the I/O request for the non-physio version of the operations.  The
732  * normal vn operations are used, and this code should work in all instances
733  * for every type of file, including pipes, sockets, fifos, and regular files.
734  *
735  * XXX I don't think it works well for socket, pipe, and fifo.
736  */
737 static void
738 aio_process_rw(struct kaiocb *job)
739 {
740 	struct ucred *td_savedcred;
741 	struct thread *td;
742 	struct aiocb *cb;
743 	struct file *fp;
744 	struct uio auio;
745 	struct iovec aiov;
746 	int cnt;
747 	int error;
748 	int oublock_st, oublock_end;
749 	int inblock_st, inblock_end;
750 
751 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
752 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
753 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
754 
755 	aio_switch_vmspace(job);
756 	td = curthread;
757 	td_savedcred = td->td_ucred;
758 	td->td_ucred = job->cred;
759 	cb = &job->uaiocb;
760 	fp = job->fd_file;
761 
762 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
763 	aiov.iov_len = cb->aio_nbytes;
764 
765 	auio.uio_iov = &aiov;
766 	auio.uio_iovcnt = 1;
767 	auio.uio_offset = cb->aio_offset;
768 	auio.uio_resid = cb->aio_nbytes;
769 	cnt = cb->aio_nbytes;
770 	auio.uio_segflg = UIO_USERSPACE;
771 	auio.uio_td = td;
772 
773 	inblock_st = td->td_ru.ru_inblock;
774 	oublock_st = td->td_ru.ru_oublock;
775 	/*
776 	 * aio_aqueue() acquires a reference to the file that is
777 	 * released in aio_free_entry().
778 	 */
779 	if (cb->aio_lio_opcode == LIO_READ) {
780 		auio.uio_rw = UIO_READ;
781 		if (auio.uio_resid == 0)
782 			error = 0;
783 		else
784 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
785 	} else {
786 		if (fp->f_type == DTYPE_VNODE)
787 			bwillwrite();
788 		auio.uio_rw = UIO_WRITE;
789 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
790 	}
791 	inblock_end = td->td_ru.ru_inblock;
792 	oublock_end = td->td_ru.ru_oublock;
793 
794 	job->inputcharge = inblock_end - inblock_st;
795 	job->outputcharge = oublock_end - oublock_st;
796 
797 	if ((error) && (auio.uio_resid != cnt)) {
798 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
799 			error = 0;
800 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
801 			PROC_LOCK(job->userproc);
802 			kern_psignal(job->userproc, SIGPIPE);
803 			PROC_UNLOCK(job->userproc);
804 		}
805 	}
806 
807 	cnt -= auio.uio_resid;
808 	td->td_ucred = td_savedcred;
809 	aio_complete(job, cnt, error);
810 }
811 
812 static void
813 aio_process_sync(struct kaiocb *job)
814 {
815 	struct thread *td = curthread;
816 	struct ucred *td_savedcred = td->td_ucred;
817 	struct file *fp = job->fd_file;
818 	int error = 0;
819 
820 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
821 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
822 
823 	td->td_ucred = job->cred;
824 	if (fp->f_vnode != NULL)
825 		error = aio_fsync_vnode(td, fp->f_vnode);
826 	td->td_ucred = td_savedcred;
827 	aio_complete(job, 0, error);
828 }
829 
830 static void
831 aio_process_mlock(struct kaiocb *job)
832 {
833 	struct aiocb *cb = &job->uaiocb;
834 	int error;
835 
836 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
837 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
838 
839 	aio_switch_vmspace(job);
840 	error = vm_mlock(job->userproc, job->cred,
841 	    __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
842 	aio_complete(job, 0, error);
843 }
844 
845 static void
846 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
847 {
848 	struct aioliojob *lj;
849 	struct kaioinfo *ki;
850 	struct kaiocb *sjob, *sjobn;
851 	int lj_done;
852 	bool schedule_fsync;
853 
854 	ki = userp->p_aioinfo;
855 	AIO_LOCK_ASSERT(ki, MA_OWNED);
856 	lj = job->lio;
857 	lj_done = 0;
858 	if (lj) {
859 		lj->lioj_finished_count++;
860 		if (lj->lioj_count == lj->lioj_finished_count)
861 			lj_done = 1;
862 	}
863 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
864 	MPASS(job->jobflags & KAIOCB_FINISHED);
865 
866 	if (ki->kaio_flags & KAIO_RUNDOWN)
867 		goto notification_done;
868 
869 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
870 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
871 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
872 
873 	KNOTE_LOCKED(&job->klist, 1);
874 
875 	if (lj_done) {
876 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
877 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
878 			KNOTE_LOCKED(&lj->klist, 1);
879 		}
880 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
881 		    == LIOJ_SIGNAL
882 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
883 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
884 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
885 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
886 		}
887 	}
888 
889 notification_done:
890 	if (job->jobflags & KAIOCB_CHECKSYNC) {
891 		schedule_fsync = false;
892 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
893 			if (job->fd_file == sjob->fd_file &&
894 			    job->seqno < sjob->seqno) {
895 				if (--sjob->pending == 0) {
896 					TAILQ_REMOVE(&ki->kaio_syncqueue, sjob,
897 					    list);
898 					if (!aio_clear_cancel_function(sjob))
899 						continue;
900 					TAILQ_INSERT_TAIL(&ki->kaio_syncready,
901 					    sjob, list);
902 					schedule_fsync = true;
903 				}
904 			}
905 		}
906 		if (schedule_fsync)
907 			taskqueue_enqueue(taskqueue_aiod_kick,
908 			    &ki->kaio_sync_task);
909 	}
910 	if (ki->kaio_flags & KAIO_WAKEUP) {
911 		ki->kaio_flags &= ~KAIO_WAKEUP;
912 		wakeup(&userp->p_aioinfo);
913 	}
914 }
915 
916 static void
917 aio_schedule_fsync(void *context, int pending)
918 {
919 	struct kaioinfo *ki;
920 	struct kaiocb *job;
921 
922 	ki = context;
923 	AIO_LOCK(ki);
924 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
925 		job = TAILQ_FIRST(&ki->kaio_syncready);
926 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
927 		AIO_UNLOCK(ki);
928 		aio_schedule(job, aio_process_sync);
929 		AIO_LOCK(ki);
930 	}
931 	AIO_UNLOCK(ki);
932 }
933 
934 bool
935 aio_cancel_cleared(struct kaiocb *job)
936 {
937 	struct kaioinfo *ki;
938 
939 	/*
940 	 * The caller should hold the same queue lock held when
941 	 * aio_clear_cancel_function() was called and set this flag
942 	 * ensuring this check sees an up-to-date value.  However,
943 	 * there is no way to assert that.
944 	 */
945 	ki = job->userproc->p_aioinfo;
946 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
947 }
948 
949 bool
950 aio_clear_cancel_function(struct kaiocb *job)
951 {
952 	struct kaioinfo *ki;
953 
954 	ki = job->userproc->p_aioinfo;
955 	AIO_LOCK(ki);
956 	MPASS(job->cancel_fn != NULL);
957 	if (job->jobflags & KAIOCB_CANCELLING) {
958 		job->jobflags |= KAIOCB_CLEARED;
959 		AIO_UNLOCK(ki);
960 		return (false);
961 	}
962 	job->cancel_fn = NULL;
963 	AIO_UNLOCK(ki);
964 	return (true);
965 }
966 
967 bool
968 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
969 {
970 	struct kaioinfo *ki;
971 
972 	ki = job->userproc->p_aioinfo;
973 	AIO_LOCK(ki);
974 	if (job->jobflags & KAIOCB_CANCELLED) {
975 		AIO_UNLOCK(ki);
976 		return (false);
977 	}
978 	job->cancel_fn = func;
979 	AIO_UNLOCK(ki);
980 	return (true);
981 }
982 
983 void
984 aio_complete(struct kaiocb *job, long status, int error)
985 {
986 	struct kaioinfo *ki;
987 	struct proc *userp;
988 
989 	job->uaiocb._aiocb_private.error = error;
990 	job->uaiocb._aiocb_private.status = status;
991 
992 	userp = job->userproc;
993 	ki = userp->p_aioinfo;
994 
995 	AIO_LOCK(ki);
996 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
997 	    ("duplicate aio_complete"));
998 	job->jobflags |= KAIOCB_FINISHED;
999 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1000 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1001 		aio_bio_done_notify(userp, job);
1002 	}
1003 	AIO_UNLOCK(ki);
1004 }
1005 
1006 void
1007 aio_cancel(struct kaiocb *job)
1008 {
1009 
1010 	aio_complete(job, -1, ECANCELED);
1011 }
1012 
1013 void
1014 aio_switch_vmspace(struct kaiocb *job)
1015 {
1016 
1017 	vmspace_switch_aio(job->userproc->p_vmspace);
1018 }
1019 
1020 /*
1021  * The AIO daemon, most of the actual work is done in aio_process_*,
1022  * but the setup (and address space mgmt) is done in this routine.
1023  */
1024 static void
1025 aio_daemon(void *_id)
1026 {
1027 	struct kaiocb *job;
1028 	struct aioproc *aiop;
1029 	struct kaioinfo *ki;
1030 	struct proc *p;
1031 	struct vmspace *myvm;
1032 	struct thread *td = curthread;
1033 	int id = (intptr_t)_id;
1034 
1035 	/*
1036 	 * Grab an extra reference on the daemon's vmspace so that it
1037 	 * doesn't get freed by jobs that switch to a different
1038 	 * vmspace.
1039 	 */
1040 	p = td->td_proc;
1041 	myvm = vmspace_acquire_ref(p);
1042 
1043 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1044 
1045 	/*
1046 	 * Allocate and ready the aio control info.  There is one aiop structure
1047 	 * per daemon.
1048 	 */
1049 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1050 	aiop->aioproc = p;
1051 	aiop->aioprocflags = 0;
1052 
1053 	/*
1054 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1055 	 * and creating too many daemons.)
1056 	 */
1057 	sema_post(&aio_newproc_sem);
1058 
1059 	mtx_lock(&aio_job_mtx);
1060 	for (;;) {
1061 		/*
1062 		 * Take daemon off of free queue
1063 		 */
1064 		if (aiop->aioprocflags & AIOP_FREE) {
1065 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1066 			aiop->aioprocflags &= ~AIOP_FREE;
1067 		}
1068 
1069 		/*
1070 		 * Check for jobs.
1071 		 */
1072 		while ((job = aio_selectjob(aiop)) != NULL) {
1073 			mtx_unlock(&aio_job_mtx);
1074 
1075 			ki = job->userproc->p_aioinfo;
1076 			job->handle_fn(job);
1077 
1078 			mtx_lock(&aio_job_mtx);
1079 			/* Decrement the active job count. */
1080 			ki->kaio_active_count--;
1081 		}
1082 
1083 		/*
1084 		 * Disconnect from user address space.
1085 		 */
1086 		if (p->p_vmspace != myvm) {
1087 			mtx_unlock(&aio_job_mtx);
1088 			vmspace_switch_aio(myvm);
1089 			mtx_lock(&aio_job_mtx);
1090 			/*
1091 			 * We have to restart to avoid race, we only sleep if
1092 			 * no job can be selected.
1093 			 */
1094 			continue;
1095 		}
1096 
1097 		mtx_assert(&aio_job_mtx, MA_OWNED);
1098 
1099 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1100 		aiop->aioprocflags |= AIOP_FREE;
1101 
1102 		/*
1103 		 * If daemon is inactive for a long time, allow it to exit,
1104 		 * thereby freeing resources.
1105 		 */
1106 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1107 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1108 		    (aiop->aioprocflags & AIOP_FREE) &&
1109 		    num_aio_procs > target_aio_procs)
1110 			break;
1111 	}
1112 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1113 	num_aio_procs--;
1114 	mtx_unlock(&aio_job_mtx);
1115 	uma_zfree(aiop_zone, aiop);
1116 	free_unr(aiod_unr, id);
1117 	vmspace_free(myvm);
1118 
1119 	KASSERT(p->p_vmspace == myvm,
1120 	    ("AIOD: bad vmspace for exiting daemon"));
1121 	KASSERT(myvm->vm_refcnt > 1,
1122 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1123 	kproc_exit(0);
1124 }
1125 
1126 /*
1127  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1128  * AIO daemon modifies its environment itself.
1129  */
1130 static int
1131 aio_newproc(int *start)
1132 {
1133 	int error;
1134 	struct proc *p;
1135 	int id;
1136 
1137 	id = alloc_unr(aiod_unr);
1138 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1139 		RFNOWAIT, 0, "aiod%d", id);
1140 	if (error == 0) {
1141 		/*
1142 		 * Wait until daemon is started.
1143 		 */
1144 		sema_wait(&aio_newproc_sem);
1145 		mtx_lock(&aio_job_mtx);
1146 		num_aio_procs++;
1147 		if (start != NULL)
1148 			(*start)--;
1149 		mtx_unlock(&aio_job_mtx);
1150 	} else {
1151 		free_unr(aiod_unr, id);
1152 	}
1153 	return (error);
1154 }
1155 
1156 /*
1157  * Try the high-performance, low-overhead physio method for eligible
1158  * VCHR devices.  This method doesn't use an aio helper thread, and
1159  * thus has very low overhead.
1160  *
1161  * Assumes that the caller, aio_aqueue(), has incremented the file
1162  * structure's reference count, preventing its deallocation for the
1163  * duration of this call.
1164  */
1165 static int
1166 aio_qphysio(struct proc *p, struct kaiocb *job)
1167 {
1168 	struct aiocb *cb;
1169 	struct file *fp;
1170 	struct bio *bp;
1171 	struct buf *pbuf;
1172 	struct vnode *vp;
1173 	struct cdevsw *csw;
1174 	struct cdev *dev;
1175 	struct kaioinfo *ki;
1176 	int error, ref, unmap, poff;
1177 	vm_prot_t prot;
1178 
1179 	cb = &job->uaiocb;
1180 	fp = job->fd_file;
1181 
1182 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1183 		return (-1);
1184 
1185 	vp = fp->f_vnode;
1186 	if (vp->v_type != VCHR)
1187 		return (-1);
1188 	if (vp->v_bufobj.bo_bsize == 0)
1189 		return (-1);
1190 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1191 		return (-1);
1192 
1193 	ref = 0;
1194 	csw = devvn_refthread(vp, &dev, &ref);
1195 	if (csw == NULL)
1196 		return (ENXIO);
1197 
1198 	if ((csw->d_flags & D_DISK) == 0) {
1199 		error = -1;
1200 		goto unref;
1201 	}
1202 	if (cb->aio_nbytes > dev->si_iosize_max) {
1203 		error = -1;
1204 		goto unref;
1205 	}
1206 
1207 	ki = p->p_aioinfo;
1208 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1209 	unmap = ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed);
1210 	if (unmap) {
1211 		if (cb->aio_nbytes > MAXPHYS) {
1212 			error = -1;
1213 			goto unref;
1214 		}
1215 	} else {
1216 		if (cb->aio_nbytes > MAXPHYS - poff) {
1217 			error = -1;
1218 			goto unref;
1219 		}
1220 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1221 			error = -1;
1222 			goto unref;
1223 		}
1224 	}
1225 	job->bp = bp = g_alloc_bio();
1226 	if (!unmap) {
1227 		job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1228 		BUF_KERNPROC(pbuf);
1229 	}
1230 
1231 	AIO_LOCK(ki);
1232 	if (!unmap)
1233 		ki->kaio_buffer_count++;
1234 	AIO_UNLOCK(ki);
1235 
1236 	bp->bio_length = cb->aio_nbytes;
1237 	bp->bio_bcount = cb->aio_nbytes;
1238 	bp->bio_done = aio_physwakeup;
1239 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1240 	bp->bio_offset = cb->aio_offset;
1241 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1242 	bp->bio_dev = dev;
1243 	bp->bio_caller1 = (void *)job;
1244 
1245 	prot = VM_PROT_READ;
1246 	if (cb->aio_lio_opcode == LIO_READ)
1247 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1248 	if ((job->npages = vm_fault_quick_hold_pages(
1249 	    &curproc->p_vmspace->vm_map,
1250 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1251 	    sizeof(job->pages)/sizeof(job->pages[0]))) < 0) {
1252 		error = EFAULT;
1253 		goto doerror;
1254 	}
1255 	if (!unmap) {
1256 		pmap_qenter((vm_offset_t)pbuf->b_data,
1257 		    job->pages, job->npages);
1258 		bp->bio_data = pbuf->b_data + poff;
1259 	} else {
1260 		bp->bio_ma = job->pages;
1261 		bp->bio_ma_n = job->npages;
1262 		bp->bio_ma_offset = poff;
1263 		bp->bio_data = unmapped_buf;
1264 		bp->bio_flags |= BIO_UNMAPPED;
1265 	}
1266 
1267 	if (!unmap)
1268 		atomic_add_int(&num_buf_aio, 1);
1269 
1270 	/* Perform transfer. */
1271 	csw->d_strategy(bp);
1272 	dev_relthread(dev, ref);
1273 	return (0);
1274 
1275 doerror:
1276 	AIO_LOCK(ki);
1277 	if (!unmap)
1278 		ki->kaio_buffer_count--;
1279 	AIO_UNLOCK(ki);
1280 	if (pbuf) {
1281 		relpbuf(pbuf, NULL);
1282 		job->pbuf = NULL;
1283 	}
1284 	g_destroy_bio(bp);
1285 	job->bp = NULL;
1286 unref:
1287 	dev_relthread(dev, ref);
1288 	return (error);
1289 }
1290 
1291 #ifdef COMPAT_FREEBSD6
1292 static int
1293 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1294 {
1295 
1296 	/*
1297 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1298 	 * supported by AIO with the old sigevent structure.
1299 	 */
1300 	nsig->sigev_notify = osig->sigev_notify;
1301 	switch (nsig->sigev_notify) {
1302 	case SIGEV_NONE:
1303 		break;
1304 	case SIGEV_SIGNAL:
1305 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1306 		break;
1307 	case SIGEV_KEVENT:
1308 		nsig->sigev_notify_kqueue =
1309 		    osig->__sigev_u.__sigev_notify_kqueue;
1310 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1311 		break;
1312 	default:
1313 		return (EINVAL);
1314 	}
1315 	return (0);
1316 }
1317 
1318 static int
1319 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1320 {
1321 	struct oaiocb *ojob;
1322 	int error;
1323 
1324 	bzero(kjob, sizeof(struct aiocb));
1325 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1326 	if (error)
1327 		return (error);
1328 	ojob = (struct oaiocb *)kjob;
1329 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1330 }
1331 #endif
1332 
1333 static int
1334 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1335 {
1336 
1337 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1338 }
1339 
1340 static long
1341 aiocb_fetch_status(struct aiocb *ujob)
1342 {
1343 
1344 	return (fuword(&ujob->_aiocb_private.status));
1345 }
1346 
1347 static long
1348 aiocb_fetch_error(struct aiocb *ujob)
1349 {
1350 
1351 	return (fuword(&ujob->_aiocb_private.error));
1352 }
1353 
1354 static int
1355 aiocb_store_status(struct aiocb *ujob, long status)
1356 {
1357 
1358 	return (suword(&ujob->_aiocb_private.status, status));
1359 }
1360 
1361 static int
1362 aiocb_store_error(struct aiocb *ujob, long error)
1363 {
1364 
1365 	return (suword(&ujob->_aiocb_private.error, error));
1366 }
1367 
1368 static int
1369 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1370 {
1371 
1372 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1373 }
1374 
1375 static int
1376 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1377 {
1378 
1379 	return (suword(ujobp, (long)ujob));
1380 }
1381 
1382 static struct aiocb_ops aiocb_ops = {
1383 	.copyin = aiocb_copyin,
1384 	.fetch_status = aiocb_fetch_status,
1385 	.fetch_error = aiocb_fetch_error,
1386 	.store_status = aiocb_store_status,
1387 	.store_error = aiocb_store_error,
1388 	.store_kernelinfo = aiocb_store_kernelinfo,
1389 	.store_aiocb = aiocb_store_aiocb,
1390 };
1391 
1392 #ifdef COMPAT_FREEBSD6
1393 static struct aiocb_ops aiocb_ops_osigevent = {
1394 	.copyin = aiocb_copyin_old_sigevent,
1395 	.fetch_status = aiocb_fetch_status,
1396 	.fetch_error = aiocb_fetch_error,
1397 	.store_status = aiocb_store_status,
1398 	.store_error = aiocb_store_error,
1399 	.store_kernelinfo = aiocb_store_kernelinfo,
1400 	.store_aiocb = aiocb_store_aiocb,
1401 };
1402 #endif
1403 
1404 /*
1405  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1406  * technique is done in this code.
1407  */
1408 int
1409 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1410 	int type, struct aiocb_ops *ops)
1411 {
1412 	struct proc *p = td->td_proc;
1413 	cap_rights_t rights;
1414 	struct file *fp;
1415 	struct kaiocb *job;
1416 	struct kaioinfo *ki;
1417 	struct kevent kev;
1418 	int opcode;
1419 	int error;
1420 	int fd, kqfd;
1421 	int jid;
1422 	u_short evflags;
1423 
1424 	if (p->p_aioinfo == NULL)
1425 		aio_init_aioinfo(p);
1426 
1427 	ki = p->p_aioinfo;
1428 
1429 	ops->store_status(ujob, -1);
1430 	ops->store_error(ujob, 0);
1431 	ops->store_kernelinfo(ujob, -1);
1432 
1433 	if (num_queue_count >= max_queue_count ||
1434 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1435 		ops->store_error(ujob, EAGAIN);
1436 		return (EAGAIN);
1437 	}
1438 
1439 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1440 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1441 
1442 	error = ops->copyin(ujob, &job->uaiocb);
1443 	if (error) {
1444 		ops->store_error(ujob, error);
1445 		uma_zfree(aiocb_zone, job);
1446 		return (error);
1447 	}
1448 
1449 	/* XXX: aio_nbytes is later casted to signed types. */
1450 	if (job->uaiocb.aio_nbytes > INT_MAX) {
1451 		uma_zfree(aiocb_zone, job);
1452 		return (EINVAL);
1453 	}
1454 
1455 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1456 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1457 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1458 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1459 		ops->store_error(ujob, EINVAL);
1460 		uma_zfree(aiocb_zone, job);
1461 		return (EINVAL);
1462 	}
1463 
1464 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1465 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1466 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1467 		uma_zfree(aiocb_zone, job);
1468 		return (EINVAL);
1469 	}
1470 
1471 	ksiginfo_init(&job->ksi);
1472 
1473 	/* Save userspace address of the job info. */
1474 	job->ujob = ujob;
1475 
1476 	/* Get the opcode. */
1477 	if (type != LIO_NOP)
1478 		job->uaiocb.aio_lio_opcode = type;
1479 	opcode = job->uaiocb.aio_lio_opcode;
1480 
1481 	/*
1482 	 * Validate the opcode and fetch the file object for the specified
1483 	 * file descriptor.
1484 	 *
1485 	 * XXXRW: Moved the opcode validation up here so that we don't
1486 	 * retrieve a file descriptor without knowing what the capabiltity
1487 	 * should be.
1488 	 */
1489 	fd = job->uaiocb.aio_fildes;
1490 	switch (opcode) {
1491 	case LIO_WRITE:
1492 		error = fget_write(td, fd,
1493 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1494 		break;
1495 	case LIO_READ:
1496 		error = fget_read(td, fd,
1497 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1498 		break;
1499 	case LIO_SYNC:
1500 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1501 		break;
1502 	case LIO_MLOCK:
1503 		fp = NULL;
1504 		break;
1505 	case LIO_NOP:
1506 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1507 		break;
1508 	default:
1509 		error = EINVAL;
1510 	}
1511 	if (error) {
1512 		uma_zfree(aiocb_zone, job);
1513 		ops->store_error(ujob, error);
1514 		return (error);
1515 	}
1516 
1517 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1518 		error = EINVAL;
1519 		goto aqueue_fail;
1520 	}
1521 
1522 	if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1523 		error = EINVAL;
1524 		goto aqueue_fail;
1525 	}
1526 
1527 	job->fd_file = fp;
1528 
1529 	mtx_lock(&aio_job_mtx);
1530 	jid = jobrefid++;
1531 	job->seqno = jobseqno++;
1532 	mtx_unlock(&aio_job_mtx);
1533 	error = ops->store_kernelinfo(ujob, jid);
1534 	if (error) {
1535 		error = EINVAL;
1536 		goto aqueue_fail;
1537 	}
1538 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1539 
1540 	if (opcode == LIO_NOP) {
1541 		fdrop(fp, td);
1542 		uma_zfree(aiocb_zone, job);
1543 		return (0);
1544 	}
1545 
1546 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1547 		goto no_kqueue;
1548 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1549 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1550 		error = EINVAL;
1551 		goto aqueue_fail;
1552 	}
1553 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1554 	kev.ident = (uintptr_t)job->ujob;
1555 	kev.filter = EVFILT_AIO;
1556 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1557 	kev.data = (intptr_t)job;
1558 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1559 	error = kqfd_register(kqfd, &kev, td, 1);
1560 	if (error)
1561 		goto aqueue_fail;
1562 
1563 no_kqueue:
1564 
1565 	ops->store_error(ujob, EINPROGRESS);
1566 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1567 	job->userproc = p;
1568 	job->cred = crhold(td->td_ucred);
1569 	job->jobflags = KAIOCB_QUEUEING;
1570 	job->lio = lj;
1571 
1572 	if (opcode == LIO_MLOCK) {
1573 		aio_schedule(job, aio_process_mlock);
1574 		error = 0;
1575 	} else if (fp->f_ops->fo_aio_queue == NULL)
1576 		error = aio_queue_file(fp, job);
1577 	else
1578 		error = fo_aio_queue(fp, job);
1579 	if (error)
1580 		goto aqueue_fail;
1581 
1582 	AIO_LOCK(ki);
1583 	job->jobflags &= ~KAIOCB_QUEUEING;
1584 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1585 	ki->kaio_count++;
1586 	if (lj)
1587 		lj->lioj_count++;
1588 	atomic_add_int(&num_queue_count, 1);
1589 	if (job->jobflags & KAIOCB_FINISHED) {
1590 		/*
1591 		 * The queue callback completed the request synchronously.
1592 		 * The bulk of the completion is deferred in that case
1593 		 * until this point.
1594 		 */
1595 		aio_bio_done_notify(p, job);
1596 	} else
1597 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1598 	AIO_UNLOCK(ki);
1599 	return (0);
1600 
1601 aqueue_fail:
1602 	knlist_delete(&job->klist, curthread, 0);
1603 	if (fp)
1604 		fdrop(fp, td);
1605 	uma_zfree(aiocb_zone, job);
1606 	ops->store_error(ujob, error);
1607 	return (error);
1608 }
1609 
1610 static void
1611 aio_cancel_daemon_job(struct kaiocb *job)
1612 {
1613 
1614 	mtx_lock(&aio_job_mtx);
1615 	if (!aio_cancel_cleared(job))
1616 		TAILQ_REMOVE(&aio_jobs, job, list);
1617 	mtx_unlock(&aio_job_mtx);
1618 	aio_cancel(job);
1619 }
1620 
1621 void
1622 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1623 {
1624 
1625 	mtx_lock(&aio_job_mtx);
1626 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1627 		mtx_unlock(&aio_job_mtx);
1628 		aio_cancel(job);
1629 		return;
1630 	}
1631 	job->handle_fn = func;
1632 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1633 	aio_kick_nowait(job->userproc);
1634 	mtx_unlock(&aio_job_mtx);
1635 }
1636 
1637 static void
1638 aio_cancel_sync(struct kaiocb *job)
1639 {
1640 	struct kaioinfo *ki;
1641 
1642 	ki = job->userproc->p_aioinfo;
1643 	mtx_lock(&aio_job_mtx);
1644 	if (!aio_cancel_cleared(job))
1645 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1646 	mtx_unlock(&aio_job_mtx);
1647 	aio_cancel(job);
1648 }
1649 
1650 int
1651 aio_queue_file(struct file *fp, struct kaiocb *job)
1652 {
1653 	struct aioliojob *lj;
1654 	struct kaioinfo *ki;
1655 	struct kaiocb *job2;
1656 	int error, opcode;
1657 
1658 	lj = job->lio;
1659 	ki = job->userproc->p_aioinfo;
1660 	opcode = job->uaiocb.aio_lio_opcode;
1661 	if (opcode == LIO_SYNC)
1662 		goto queueit;
1663 
1664 	if ((error = aio_qphysio(job->userproc, job)) == 0)
1665 		goto done;
1666 #if 0
1667 	/*
1668 	 * XXX: This means qphysio() failed with EFAULT.  The current
1669 	 * behavior is to retry the operation via fo_read/fo_write.
1670 	 * Wouldn't it be better to just complete the request with an
1671 	 * error here?
1672 	 */
1673 	if (error > 0)
1674 		goto done;
1675 #endif
1676 queueit:
1677 	if (!enable_aio_unsafe)
1678 		return (EOPNOTSUPP);
1679 
1680 	if (opcode == LIO_SYNC) {
1681 		AIO_LOCK(ki);
1682 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1683 			if (job2->fd_file == job->fd_file &&
1684 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1685 			    job2->seqno < job->seqno) {
1686 				job2->jobflags |= KAIOCB_CHECKSYNC;
1687 				job->pending++;
1688 			}
1689 		}
1690 		if (job->pending != 0) {
1691 			if (!aio_set_cancel_function(job, aio_cancel_sync)) {
1692 				AIO_UNLOCK(ki);
1693 				aio_cancel(job);
1694 				return (0);
1695 			}
1696 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1697 			AIO_UNLOCK(ki);
1698 			return (0);
1699 		}
1700 		AIO_UNLOCK(ki);
1701 	}
1702 
1703 	switch (opcode) {
1704 	case LIO_READ:
1705 	case LIO_WRITE:
1706 		aio_schedule(job, aio_process_rw);
1707 		error = 0;
1708 		break;
1709 	case LIO_SYNC:
1710 		aio_schedule(job, aio_process_sync);
1711 		error = 0;
1712 		break;
1713 	default:
1714 		error = EINVAL;
1715 	}
1716 done:
1717 	return (error);
1718 }
1719 
1720 static void
1721 aio_kick_nowait(struct proc *userp)
1722 {
1723 	struct kaioinfo *ki = userp->p_aioinfo;
1724 	struct aioproc *aiop;
1725 
1726 	mtx_assert(&aio_job_mtx, MA_OWNED);
1727 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1728 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1729 		aiop->aioprocflags &= ~AIOP_FREE;
1730 		wakeup(aiop->aioproc);
1731 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1732 	    ki->kaio_active_count + num_aio_resv_start <
1733 	    ki->kaio_maxactive_count) {
1734 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1735 	}
1736 }
1737 
1738 static int
1739 aio_kick(struct proc *userp)
1740 {
1741 	struct kaioinfo *ki = userp->p_aioinfo;
1742 	struct aioproc *aiop;
1743 	int error, ret = 0;
1744 
1745 	mtx_assert(&aio_job_mtx, MA_OWNED);
1746 retryproc:
1747 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1748 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1749 		aiop->aioprocflags &= ~AIOP_FREE;
1750 		wakeup(aiop->aioproc);
1751 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1752 	    ki->kaio_active_count + num_aio_resv_start <
1753 	    ki->kaio_maxactive_count) {
1754 		num_aio_resv_start++;
1755 		mtx_unlock(&aio_job_mtx);
1756 		error = aio_newproc(&num_aio_resv_start);
1757 		mtx_lock(&aio_job_mtx);
1758 		if (error) {
1759 			num_aio_resv_start--;
1760 			goto retryproc;
1761 		}
1762 	} else {
1763 		ret = -1;
1764 	}
1765 	return (ret);
1766 }
1767 
1768 static void
1769 aio_kick_helper(void *context, int pending)
1770 {
1771 	struct proc *userp = context;
1772 
1773 	mtx_lock(&aio_job_mtx);
1774 	while (--pending >= 0) {
1775 		if (aio_kick(userp))
1776 			break;
1777 	}
1778 	mtx_unlock(&aio_job_mtx);
1779 }
1780 
1781 /*
1782  * Support the aio_return system call, as a side-effect, kernel resources are
1783  * released.
1784  */
1785 static int
1786 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1787 {
1788 	struct proc *p = td->td_proc;
1789 	struct kaiocb *job;
1790 	struct kaioinfo *ki;
1791 	int status, error;
1792 
1793 	ki = p->p_aioinfo;
1794 	if (ki == NULL)
1795 		return (EINVAL);
1796 	AIO_LOCK(ki);
1797 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1798 		if (job->ujob == ujob)
1799 			break;
1800 	}
1801 	if (job != NULL) {
1802 		MPASS(job->jobflags & KAIOCB_FINISHED);
1803 		status = job->uaiocb._aiocb_private.status;
1804 		error = job->uaiocb._aiocb_private.error;
1805 		td->td_retval[0] = status;
1806 		if (job->uaiocb.aio_lio_opcode == LIO_WRITE) {
1807 			td->td_ru.ru_oublock += job->outputcharge;
1808 			job->outputcharge = 0;
1809 		} else if (job->uaiocb.aio_lio_opcode == LIO_READ) {
1810 			td->td_ru.ru_inblock += job->inputcharge;
1811 			job->inputcharge = 0;
1812 		}
1813 		aio_free_entry(job);
1814 		AIO_UNLOCK(ki);
1815 		ops->store_error(ujob, error);
1816 		ops->store_status(ujob, status);
1817 	} else {
1818 		error = EINVAL;
1819 		AIO_UNLOCK(ki);
1820 	}
1821 	return (error);
1822 }
1823 
1824 int
1825 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1826 {
1827 
1828 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1829 }
1830 
1831 /*
1832  * Allow a process to wakeup when any of the I/O requests are completed.
1833  */
1834 static int
1835 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1836     struct timespec *ts)
1837 {
1838 	struct proc *p = td->td_proc;
1839 	struct timeval atv;
1840 	struct kaioinfo *ki;
1841 	struct kaiocb *firstjob, *job;
1842 	int error, i, timo;
1843 
1844 	timo = 0;
1845 	if (ts) {
1846 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1847 			return (EINVAL);
1848 
1849 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1850 		if (itimerfix(&atv))
1851 			return (EINVAL);
1852 		timo = tvtohz(&atv);
1853 	}
1854 
1855 	ki = p->p_aioinfo;
1856 	if (ki == NULL)
1857 		return (EAGAIN);
1858 
1859 	if (njoblist == 0)
1860 		return (0);
1861 
1862 	AIO_LOCK(ki);
1863 	for (;;) {
1864 		firstjob = NULL;
1865 		error = 0;
1866 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1867 			for (i = 0; i < njoblist; i++) {
1868 				if (job->ujob == ujoblist[i]) {
1869 					if (firstjob == NULL)
1870 						firstjob = job;
1871 					if (job->jobflags & KAIOCB_FINISHED)
1872 						goto RETURN;
1873 				}
1874 			}
1875 		}
1876 		/* All tasks were finished. */
1877 		if (firstjob == NULL)
1878 			break;
1879 
1880 		ki->kaio_flags |= KAIO_WAKEUP;
1881 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1882 		    "aiospn", timo);
1883 		if (error == ERESTART)
1884 			error = EINTR;
1885 		if (error)
1886 			break;
1887 	}
1888 RETURN:
1889 	AIO_UNLOCK(ki);
1890 	return (error);
1891 }
1892 
1893 int
1894 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1895 {
1896 	struct timespec ts, *tsp;
1897 	struct aiocb **ujoblist;
1898 	int error;
1899 
1900 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1901 		return (EINVAL);
1902 
1903 	if (uap->timeout) {
1904 		/* Get timespec struct. */
1905 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1906 			return (error);
1907 		tsp = &ts;
1908 	} else
1909 		tsp = NULL;
1910 
1911 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1912 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1913 	if (error == 0)
1914 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1915 	uma_zfree(aiol_zone, ujoblist);
1916 	return (error);
1917 }
1918 
1919 /*
1920  * aio_cancel cancels any non-physio aio operations not currently in
1921  * progress.
1922  */
1923 int
1924 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1925 {
1926 	struct proc *p = td->td_proc;
1927 	struct kaioinfo *ki;
1928 	struct kaiocb *job, *jobn;
1929 	struct file *fp;
1930 	cap_rights_t rights;
1931 	int error;
1932 	int cancelled = 0;
1933 	int notcancelled = 0;
1934 	struct vnode *vp;
1935 
1936 	/* Lookup file object. */
1937 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1938 	if (error)
1939 		return (error);
1940 
1941 	ki = p->p_aioinfo;
1942 	if (ki == NULL)
1943 		goto done;
1944 
1945 	if (fp->f_type == DTYPE_VNODE) {
1946 		vp = fp->f_vnode;
1947 		if (vn_isdisk(vp, &error)) {
1948 			fdrop(fp, td);
1949 			td->td_retval[0] = AIO_NOTCANCELED;
1950 			return (0);
1951 		}
1952 	}
1953 
1954 	AIO_LOCK(ki);
1955 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1956 		if ((uap->fd == job->uaiocb.aio_fildes) &&
1957 		    ((uap->aiocbp == NULL) ||
1958 		     (uap->aiocbp == job->ujob))) {
1959 			if (aio_cancel_job(p, ki, job)) {
1960 				cancelled++;
1961 			} else {
1962 				notcancelled++;
1963 			}
1964 			if (uap->aiocbp != NULL)
1965 				break;
1966 		}
1967 	}
1968 	AIO_UNLOCK(ki);
1969 
1970 done:
1971 	fdrop(fp, td);
1972 
1973 	if (uap->aiocbp != NULL) {
1974 		if (cancelled) {
1975 			td->td_retval[0] = AIO_CANCELED;
1976 			return (0);
1977 		}
1978 	}
1979 
1980 	if (notcancelled) {
1981 		td->td_retval[0] = AIO_NOTCANCELED;
1982 		return (0);
1983 	}
1984 
1985 	if (cancelled) {
1986 		td->td_retval[0] = AIO_CANCELED;
1987 		return (0);
1988 	}
1989 
1990 	td->td_retval[0] = AIO_ALLDONE;
1991 
1992 	return (0);
1993 }
1994 
1995 /*
1996  * aio_error is implemented in the kernel level for compatibility purposes
1997  * only.  For a user mode async implementation, it would be best to do it in
1998  * a userland subroutine.
1999  */
2000 static int
2001 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2002 {
2003 	struct proc *p = td->td_proc;
2004 	struct kaiocb *job;
2005 	struct kaioinfo *ki;
2006 	int status;
2007 
2008 	ki = p->p_aioinfo;
2009 	if (ki == NULL) {
2010 		td->td_retval[0] = EINVAL;
2011 		return (0);
2012 	}
2013 
2014 	AIO_LOCK(ki);
2015 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2016 		if (job->ujob == ujob) {
2017 			if (job->jobflags & KAIOCB_FINISHED)
2018 				td->td_retval[0] =
2019 					job->uaiocb._aiocb_private.error;
2020 			else
2021 				td->td_retval[0] = EINPROGRESS;
2022 			AIO_UNLOCK(ki);
2023 			return (0);
2024 		}
2025 	}
2026 	AIO_UNLOCK(ki);
2027 
2028 	/*
2029 	 * Hack for failure of aio_aqueue.
2030 	 */
2031 	status = ops->fetch_status(ujob);
2032 	if (status == -1) {
2033 		td->td_retval[0] = ops->fetch_error(ujob);
2034 		return (0);
2035 	}
2036 
2037 	td->td_retval[0] = EINVAL;
2038 	return (0);
2039 }
2040 
2041 int
2042 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2043 {
2044 
2045 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2046 }
2047 
2048 /* syscall - asynchronous read from a file (REALTIME) */
2049 #ifdef COMPAT_FREEBSD6
2050 int
2051 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2052 {
2053 
2054 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2055 	    &aiocb_ops_osigevent));
2056 }
2057 #endif
2058 
2059 int
2060 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2061 {
2062 
2063 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2064 }
2065 
2066 /* syscall - asynchronous write to a file (REALTIME) */
2067 #ifdef COMPAT_FREEBSD6
2068 int
2069 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2070 {
2071 
2072 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2073 	    &aiocb_ops_osigevent));
2074 }
2075 #endif
2076 
2077 int
2078 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2079 {
2080 
2081 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2082 }
2083 
2084 int
2085 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2086 {
2087 
2088 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2089 }
2090 
2091 static int
2092 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2093     struct aiocb **acb_list, int nent, struct sigevent *sig,
2094     struct aiocb_ops *ops)
2095 {
2096 	struct proc *p = td->td_proc;
2097 	struct aiocb *job;
2098 	struct kaioinfo *ki;
2099 	struct aioliojob *lj;
2100 	struct kevent kev;
2101 	int error;
2102 	int nerror;
2103 	int i;
2104 
2105 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2106 		return (EINVAL);
2107 
2108 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2109 		return (EINVAL);
2110 
2111 	if (p->p_aioinfo == NULL)
2112 		aio_init_aioinfo(p);
2113 
2114 	ki = p->p_aioinfo;
2115 
2116 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2117 	lj->lioj_flags = 0;
2118 	lj->lioj_count = 0;
2119 	lj->lioj_finished_count = 0;
2120 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2121 	ksiginfo_init(&lj->lioj_ksi);
2122 
2123 	/*
2124 	 * Setup signal.
2125 	 */
2126 	if (sig && (mode == LIO_NOWAIT)) {
2127 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2128 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2129 			/* Assume only new style KEVENT */
2130 			kev.filter = EVFILT_LIO;
2131 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2132 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2133 			kev.data = (intptr_t)lj;
2134 			/* pass user defined sigval data */
2135 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2136 			error = kqfd_register(
2137 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2138 			if (error) {
2139 				uma_zfree(aiolio_zone, lj);
2140 				return (error);
2141 			}
2142 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2143 			;
2144 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2145 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2146 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2147 					uma_zfree(aiolio_zone, lj);
2148 					return EINVAL;
2149 				}
2150 				lj->lioj_flags |= LIOJ_SIGNAL;
2151 		} else {
2152 			uma_zfree(aiolio_zone, lj);
2153 			return EINVAL;
2154 		}
2155 	}
2156 
2157 	AIO_LOCK(ki);
2158 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2159 	/*
2160 	 * Add extra aiocb count to avoid the lio to be freed
2161 	 * by other threads doing aio_waitcomplete or aio_return,
2162 	 * and prevent event from being sent until we have queued
2163 	 * all tasks.
2164 	 */
2165 	lj->lioj_count = 1;
2166 	AIO_UNLOCK(ki);
2167 
2168 	/*
2169 	 * Get pointers to the list of I/O requests.
2170 	 */
2171 	nerror = 0;
2172 	for (i = 0; i < nent; i++) {
2173 		job = acb_list[i];
2174 		if (job != NULL) {
2175 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2176 			if (error != 0)
2177 				nerror++;
2178 		}
2179 	}
2180 
2181 	error = 0;
2182 	AIO_LOCK(ki);
2183 	if (mode == LIO_WAIT) {
2184 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2185 			ki->kaio_flags |= KAIO_WAKEUP;
2186 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2187 			    PRIBIO | PCATCH, "aiospn", 0);
2188 			if (error == ERESTART)
2189 				error = EINTR;
2190 			if (error)
2191 				break;
2192 		}
2193 	} else {
2194 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2195 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2196 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2197 				KNOTE_LOCKED(&lj->klist, 1);
2198 			}
2199 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2200 			    == LIOJ_SIGNAL
2201 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2202 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2203 				aio_sendsig(p, &lj->lioj_signal,
2204 					    &lj->lioj_ksi);
2205 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2206 			}
2207 		}
2208 	}
2209 	lj->lioj_count--;
2210 	if (lj->lioj_count == 0) {
2211 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2212 		knlist_delete(&lj->klist, curthread, 1);
2213 		PROC_LOCK(p);
2214 		sigqueue_take(&lj->lioj_ksi);
2215 		PROC_UNLOCK(p);
2216 		AIO_UNLOCK(ki);
2217 		uma_zfree(aiolio_zone, lj);
2218 	} else
2219 		AIO_UNLOCK(ki);
2220 
2221 	if (nerror)
2222 		return (EIO);
2223 	return (error);
2224 }
2225 
2226 /* syscall - list directed I/O (REALTIME) */
2227 #ifdef COMPAT_FREEBSD6
2228 int
2229 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2230 {
2231 	struct aiocb **acb_list;
2232 	struct sigevent *sigp, sig;
2233 	struct osigevent osig;
2234 	int error, nent;
2235 
2236 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2237 		return (EINVAL);
2238 
2239 	nent = uap->nent;
2240 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2241 		return (EINVAL);
2242 
2243 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2244 		error = copyin(uap->sig, &osig, sizeof(osig));
2245 		if (error)
2246 			return (error);
2247 		error = convert_old_sigevent(&osig, &sig);
2248 		if (error)
2249 			return (error);
2250 		sigp = &sig;
2251 	} else
2252 		sigp = NULL;
2253 
2254 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2255 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2256 	if (error == 0)
2257 		error = kern_lio_listio(td, uap->mode,
2258 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2259 		    &aiocb_ops_osigevent);
2260 	free(acb_list, M_LIO);
2261 	return (error);
2262 }
2263 #endif
2264 
2265 /* syscall - list directed I/O (REALTIME) */
2266 int
2267 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2268 {
2269 	struct aiocb **acb_list;
2270 	struct sigevent *sigp, sig;
2271 	int error, nent;
2272 
2273 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2274 		return (EINVAL);
2275 
2276 	nent = uap->nent;
2277 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2278 		return (EINVAL);
2279 
2280 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2281 		error = copyin(uap->sig, &sig, sizeof(sig));
2282 		if (error)
2283 			return (error);
2284 		sigp = &sig;
2285 	} else
2286 		sigp = NULL;
2287 
2288 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2289 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2290 	if (error == 0)
2291 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2292 		    nent, sigp, &aiocb_ops);
2293 	free(acb_list, M_LIO);
2294 	return (error);
2295 }
2296 
2297 static void
2298 aio_physwakeup(struct bio *bp)
2299 {
2300 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2301 	struct proc *userp;
2302 	struct kaioinfo *ki;
2303 	size_t nbytes;
2304 	int error, nblks;
2305 
2306 	/* Release mapping into kernel space. */
2307 	userp = job->userproc;
2308 	ki = userp->p_aioinfo;
2309 	if (job->pbuf) {
2310 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2311 		relpbuf(job->pbuf, NULL);
2312 		job->pbuf = NULL;
2313 		atomic_subtract_int(&num_buf_aio, 1);
2314 		AIO_LOCK(ki);
2315 		ki->kaio_buffer_count--;
2316 		AIO_UNLOCK(ki);
2317 	}
2318 	vm_page_unhold_pages(job->pages, job->npages);
2319 
2320 	bp = job->bp;
2321 	job->bp = NULL;
2322 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2323 	error = 0;
2324 	if (bp->bio_flags & BIO_ERROR)
2325 		error = bp->bio_error;
2326 	nblks = btodb(nbytes);
2327 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2328 		job->outputcharge += nblks;
2329 	else
2330 		job->inputcharge += nblks;
2331 
2332 	aio_complete(job, nbytes, error);
2333 
2334 	g_destroy_bio(bp);
2335 }
2336 
2337 /* syscall - wait for the next completion of an aio request */
2338 static int
2339 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2340     struct timespec *ts, struct aiocb_ops *ops)
2341 {
2342 	struct proc *p = td->td_proc;
2343 	struct timeval atv;
2344 	struct kaioinfo *ki;
2345 	struct kaiocb *job;
2346 	struct aiocb *ujob;
2347 	int error, status, timo;
2348 
2349 	ops->store_aiocb(ujobp, NULL);
2350 
2351 	if (ts == NULL) {
2352 		timo = 0;
2353 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2354 		timo = -1;
2355 	} else {
2356 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2357 			return (EINVAL);
2358 
2359 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2360 		if (itimerfix(&atv))
2361 			return (EINVAL);
2362 		timo = tvtohz(&atv);
2363 	}
2364 
2365 	if (p->p_aioinfo == NULL)
2366 		aio_init_aioinfo(p);
2367 	ki = p->p_aioinfo;
2368 
2369 	error = 0;
2370 	job = NULL;
2371 	AIO_LOCK(ki);
2372 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2373 		if (timo == -1) {
2374 			error = EWOULDBLOCK;
2375 			break;
2376 		}
2377 		ki->kaio_flags |= KAIO_WAKEUP;
2378 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2379 		    "aiowc", timo);
2380 		if (timo && error == ERESTART)
2381 			error = EINTR;
2382 		if (error)
2383 			break;
2384 	}
2385 
2386 	if (job != NULL) {
2387 		MPASS(job->jobflags & KAIOCB_FINISHED);
2388 		ujob = job->ujob;
2389 		status = job->uaiocb._aiocb_private.status;
2390 		error = job->uaiocb._aiocb_private.error;
2391 		td->td_retval[0] = status;
2392 		if (job->uaiocb.aio_lio_opcode == LIO_WRITE) {
2393 			td->td_ru.ru_oublock += job->outputcharge;
2394 			job->outputcharge = 0;
2395 		} else if (job->uaiocb.aio_lio_opcode == LIO_READ) {
2396 			td->td_ru.ru_inblock += job->inputcharge;
2397 			job->inputcharge = 0;
2398 		}
2399 		aio_free_entry(job);
2400 		AIO_UNLOCK(ki);
2401 		ops->store_aiocb(ujobp, ujob);
2402 		ops->store_error(ujob, error);
2403 		ops->store_status(ujob, status);
2404 	} else
2405 		AIO_UNLOCK(ki);
2406 
2407 	return (error);
2408 }
2409 
2410 int
2411 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2412 {
2413 	struct timespec ts, *tsp;
2414 	int error;
2415 
2416 	if (uap->timeout) {
2417 		/* Get timespec struct. */
2418 		error = copyin(uap->timeout, &ts, sizeof(ts));
2419 		if (error)
2420 			return (error);
2421 		tsp = &ts;
2422 	} else
2423 		tsp = NULL;
2424 
2425 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2426 }
2427 
2428 static int
2429 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2430     struct aiocb_ops *ops)
2431 {
2432 	struct proc *p = td->td_proc;
2433 	struct kaioinfo *ki;
2434 
2435 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2436 		return (EINVAL);
2437 	ki = p->p_aioinfo;
2438 	if (ki == NULL)
2439 		aio_init_aioinfo(p);
2440 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2441 }
2442 
2443 int
2444 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2445 {
2446 
2447 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2448 }
2449 
2450 /* kqueue attach function */
2451 static int
2452 filt_aioattach(struct knote *kn)
2453 {
2454 	struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2455 
2456 	/*
2457 	 * The job pointer must be validated before using it, so
2458 	 * registration is restricted to the kernel; the user cannot
2459 	 * set EV_FLAG1.
2460 	 */
2461 	if ((kn->kn_flags & EV_FLAG1) == 0)
2462 		return (EPERM);
2463 	kn->kn_ptr.p_aio = job;
2464 	kn->kn_flags &= ~EV_FLAG1;
2465 
2466 	knlist_add(&job->klist, kn, 0);
2467 
2468 	return (0);
2469 }
2470 
2471 /* kqueue detach function */
2472 static void
2473 filt_aiodetach(struct knote *kn)
2474 {
2475 	struct knlist *knl;
2476 
2477 	knl = &kn->kn_ptr.p_aio->klist;
2478 	knl->kl_lock(knl->kl_lockarg);
2479 	if (!knlist_empty(knl))
2480 		knlist_remove(knl, kn, 1);
2481 	knl->kl_unlock(knl->kl_lockarg);
2482 }
2483 
2484 /* kqueue filter function */
2485 /*ARGSUSED*/
2486 static int
2487 filt_aio(struct knote *kn, long hint)
2488 {
2489 	struct kaiocb *job = kn->kn_ptr.p_aio;
2490 
2491 	kn->kn_data = job->uaiocb._aiocb_private.error;
2492 	if (!(job->jobflags & KAIOCB_FINISHED))
2493 		return (0);
2494 	kn->kn_flags |= EV_EOF;
2495 	return (1);
2496 }
2497 
2498 /* kqueue attach function */
2499 static int
2500 filt_lioattach(struct knote *kn)
2501 {
2502 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2503 
2504 	/*
2505 	 * The aioliojob pointer must be validated before using it, so
2506 	 * registration is restricted to the kernel; the user cannot
2507 	 * set EV_FLAG1.
2508 	 */
2509 	if ((kn->kn_flags & EV_FLAG1) == 0)
2510 		return (EPERM);
2511 	kn->kn_ptr.p_lio = lj;
2512 	kn->kn_flags &= ~EV_FLAG1;
2513 
2514 	knlist_add(&lj->klist, kn, 0);
2515 
2516 	return (0);
2517 }
2518 
2519 /* kqueue detach function */
2520 static void
2521 filt_liodetach(struct knote *kn)
2522 {
2523 	struct knlist *knl;
2524 
2525 	knl = &kn->kn_ptr.p_lio->klist;
2526 	knl->kl_lock(knl->kl_lockarg);
2527 	if (!knlist_empty(knl))
2528 		knlist_remove(knl, kn, 1);
2529 	knl->kl_unlock(knl->kl_lockarg);
2530 }
2531 
2532 /* kqueue filter function */
2533 /*ARGSUSED*/
2534 static int
2535 filt_lio(struct knote *kn, long hint)
2536 {
2537 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2538 
2539 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2540 }
2541 
2542 #ifdef COMPAT_FREEBSD32
2543 #include <sys/mount.h>
2544 #include <sys/socket.h>
2545 #include <compat/freebsd32/freebsd32.h>
2546 #include <compat/freebsd32/freebsd32_proto.h>
2547 #include <compat/freebsd32/freebsd32_signal.h>
2548 #include <compat/freebsd32/freebsd32_syscall.h>
2549 #include <compat/freebsd32/freebsd32_util.h>
2550 
2551 struct __aiocb_private32 {
2552 	int32_t	status;
2553 	int32_t	error;
2554 	uint32_t kernelinfo;
2555 };
2556 
2557 #ifdef COMPAT_FREEBSD6
2558 typedef struct oaiocb32 {
2559 	int	aio_fildes;		/* File descriptor */
2560 	uint64_t aio_offset __packed;	/* File offset for I/O */
2561 	uint32_t aio_buf;		/* I/O buffer in process space */
2562 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2563 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2564 	int	aio_lio_opcode;		/* LIO opcode */
2565 	int	aio_reqprio;		/* Request priority -- ignored */
2566 	struct	__aiocb_private32 _aiocb_private;
2567 } oaiocb32_t;
2568 #endif
2569 
2570 typedef struct aiocb32 {
2571 	int32_t	aio_fildes;		/* File descriptor */
2572 	uint64_t aio_offset __packed;	/* File offset for I/O */
2573 	uint32_t aio_buf;		/* I/O buffer in process space */
2574 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2575 	int	__spare__[2];
2576 	uint32_t __spare2__;
2577 	int	aio_lio_opcode;		/* LIO opcode */
2578 	int	aio_reqprio;		/* Request priority -- ignored */
2579 	struct	__aiocb_private32 _aiocb_private;
2580 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2581 } aiocb32_t;
2582 
2583 #ifdef COMPAT_FREEBSD6
2584 static int
2585 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2586 {
2587 
2588 	/*
2589 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2590 	 * supported by AIO with the old sigevent structure.
2591 	 */
2592 	CP(*osig, *nsig, sigev_notify);
2593 	switch (nsig->sigev_notify) {
2594 	case SIGEV_NONE:
2595 		break;
2596 	case SIGEV_SIGNAL:
2597 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2598 		break;
2599 	case SIGEV_KEVENT:
2600 		nsig->sigev_notify_kqueue =
2601 		    osig->__sigev_u.__sigev_notify_kqueue;
2602 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2603 		break;
2604 	default:
2605 		return (EINVAL);
2606 	}
2607 	return (0);
2608 }
2609 
2610 static int
2611 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2612 {
2613 	struct oaiocb32 job32;
2614 	int error;
2615 
2616 	bzero(kjob, sizeof(struct aiocb));
2617 	error = copyin(ujob, &job32, sizeof(job32));
2618 	if (error)
2619 		return (error);
2620 
2621 	CP(job32, *kjob, aio_fildes);
2622 	CP(job32, *kjob, aio_offset);
2623 	PTRIN_CP(job32, *kjob, aio_buf);
2624 	CP(job32, *kjob, aio_nbytes);
2625 	CP(job32, *kjob, aio_lio_opcode);
2626 	CP(job32, *kjob, aio_reqprio);
2627 	CP(job32, *kjob, _aiocb_private.status);
2628 	CP(job32, *kjob, _aiocb_private.error);
2629 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2630 	return (convert_old_sigevent32(&job32.aio_sigevent,
2631 	    &kjob->aio_sigevent));
2632 }
2633 #endif
2634 
2635 static int
2636 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2637 {
2638 	struct aiocb32 job32;
2639 	int error;
2640 
2641 	error = copyin(ujob, &job32, sizeof(job32));
2642 	if (error)
2643 		return (error);
2644 	CP(job32, *kjob, aio_fildes);
2645 	CP(job32, *kjob, aio_offset);
2646 	PTRIN_CP(job32, *kjob, aio_buf);
2647 	CP(job32, *kjob, aio_nbytes);
2648 	CP(job32, *kjob, aio_lio_opcode);
2649 	CP(job32, *kjob, aio_reqprio);
2650 	CP(job32, *kjob, _aiocb_private.status);
2651 	CP(job32, *kjob, _aiocb_private.error);
2652 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2653 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2654 }
2655 
2656 static long
2657 aiocb32_fetch_status(struct aiocb *ujob)
2658 {
2659 	struct aiocb32 *ujob32;
2660 
2661 	ujob32 = (struct aiocb32 *)ujob;
2662 	return (fuword32(&ujob32->_aiocb_private.status));
2663 }
2664 
2665 static long
2666 aiocb32_fetch_error(struct aiocb *ujob)
2667 {
2668 	struct aiocb32 *ujob32;
2669 
2670 	ujob32 = (struct aiocb32 *)ujob;
2671 	return (fuword32(&ujob32->_aiocb_private.error));
2672 }
2673 
2674 static int
2675 aiocb32_store_status(struct aiocb *ujob, long status)
2676 {
2677 	struct aiocb32 *ujob32;
2678 
2679 	ujob32 = (struct aiocb32 *)ujob;
2680 	return (suword32(&ujob32->_aiocb_private.status, status));
2681 }
2682 
2683 static int
2684 aiocb32_store_error(struct aiocb *ujob, long error)
2685 {
2686 	struct aiocb32 *ujob32;
2687 
2688 	ujob32 = (struct aiocb32 *)ujob;
2689 	return (suword32(&ujob32->_aiocb_private.error, error));
2690 }
2691 
2692 static int
2693 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2694 {
2695 	struct aiocb32 *ujob32;
2696 
2697 	ujob32 = (struct aiocb32 *)ujob;
2698 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2699 }
2700 
2701 static int
2702 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2703 {
2704 
2705 	return (suword32(ujobp, (long)ujob));
2706 }
2707 
2708 static struct aiocb_ops aiocb32_ops = {
2709 	.copyin = aiocb32_copyin,
2710 	.fetch_status = aiocb32_fetch_status,
2711 	.fetch_error = aiocb32_fetch_error,
2712 	.store_status = aiocb32_store_status,
2713 	.store_error = aiocb32_store_error,
2714 	.store_kernelinfo = aiocb32_store_kernelinfo,
2715 	.store_aiocb = aiocb32_store_aiocb,
2716 };
2717 
2718 #ifdef COMPAT_FREEBSD6
2719 static struct aiocb_ops aiocb32_ops_osigevent = {
2720 	.copyin = aiocb32_copyin_old_sigevent,
2721 	.fetch_status = aiocb32_fetch_status,
2722 	.fetch_error = aiocb32_fetch_error,
2723 	.store_status = aiocb32_store_status,
2724 	.store_error = aiocb32_store_error,
2725 	.store_kernelinfo = aiocb32_store_kernelinfo,
2726 	.store_aiocb = aiocb32_store_aiocb,
2727 };
2728 #endif
2729 
2730 int
2731 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2732 {
2733 
2734 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2735 }
2736 
2737 int
2738 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2739 {
2740 	struct timespec32 ts32;
2741 	struct timespec ts, *tsp;
2742 	struct aiocb **ujoblist;
2743 	uint32_t *ujoblist32;
2744 	int error, i;
2745 
2746 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2747 		return (EINVAL);
2748 
2749 	if (uap->timeout) {
2750 		/* Get timespec struct. */
2751 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2752 			return (error);
2753 		CP(ts32, ts, tv_sec);
2754 		CP(ts32, ts, tv_nsec);
2755 		tsp = &ts;
2756 	} else
2757 		tsp = NULL;
2758 
2759 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2760 	ujoblist32 = (uint32_t *)ujoblist;
2761 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2762 	    sizeof(ujoblist32[0]));
2763 	if (error == 0) {
2764 		for (i = uap->nent; i > 0; i--)
2765 			ujoblist[i] = PTRIN(ujoblist32[i]);
2766 
2767 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2768 	}
2769 	uma_zfree(aiol_zone, ujoblist);
2770 	return (error);
2771 }
2772 
2773 int
2774 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2775 {
2776 
2777 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2778 }
2779 
2780 #ifdef COMPAT_FREEBSD6
2781 int
2782 freebsd6_freebsd32_aio_read(struct thread *td,
2783     struct freebsd6_freebsd32_aio_read_args *uap)
2784 {
2785 
2786 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2787 	    &aiocb32_ops_osigevent));
2788 }
2789 #endif
2790 
2791 int
2792 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2793 {
2794 
2795 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2796 	    &aiocb32_ops));
2797 }
2798 
2799 #ifdef COMPAT_FREEBSD6
2800 int
2801 freebsd6_freebsd32_aio_write(struct thread *td,
2802     struct freebsd6_freebsd32_aio_write_args *uap)
2803 {
2804 
2805 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2806 	    &aiocb32_ops_osigevent));
2807 }
2808 #endif
2809 
2810 int
2811 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2812 {
2813 
2814 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2815 	    &aiocb32_ops));
2816 }
2817 
2818 int
2819 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2820 {
2821 
2822 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2823 	    &aiocb32_ops));
2824 }
2825 
2826 int
2827 freebsd32_aio_waitcomplete(struct thread *td,
2828     struct freebsd32_aio_waitcomplete_args *uap)
2829 {
2830 	struct timespec32 ts32;
2831 	struct timespec ts, *tsp;
2832 	int error;
2833 
2834 	if (uap->timeout) {
2835 		/* Get timespec struct. */
2836 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2837 		if (error)
2838 			return (error);
2839 		CP(ts32, ts, tv_sec);
2840 		CP(ts32, ts, tv_nsec);
2841 		tsp = &ts;
2842 	} else
2843 		tsp = NULL;
2844 
2845 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2846 	    &aiocb32_ops));
2847 }
2848 
2849 int
2850 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2851 {
2852 
2853 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2854 	    &aiocb32_ops));
2855 }
2856 
2857 #ifdef COMPAT_FREEBSD6
2858 int
2859 freebsd6_freebsd32_lio_listio(struct thread *td,
2860     struct freebsd6_freebsd32_lio_listio_args *uap)
2861 {
2862 	struct aiocb **acb_list;
2863 	struct sigevent *sigp, sig;
2864 	struct osigevent32 osig;
2865 	uint32_t *acb_list32;
2866 	int error, i, nent;
2867 
2868 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2869 		return (EINVAL);
2870 
2871 	nent = uap->nent;
2872 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2873 		return (EINVAL);
2874 
2875 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2876 		error = copyin(uap->sig, &osig, sizeof(osig));
2877 		if (error)
2878 			return (error);
2879 		error = convert_old_sigevent32(&osig, &sig);
2880 		if (error)
2881 			return (error);
2882 		sigp = &sig;
2883 	} else
2884 		sigp = NULL;
2885 
2886 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2887 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2888 	if (error) {
2889 		free(acb_list32, M_LIO);
2890 		return (error);
2891 	}
2892 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2893 	for (i = 0; i < nent; i++)
2894 		acb_list[i] = PTRIN(acb_list32[i]);
2895 	free(acb_list32, M_LIO);
2896 
2897 	error = kern_lio_listio(td, uap->mode,
2898 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2899 	    &aiocb32_ops_osigevent);
2900 	free(acb_list, M_LIO);
2901 	return (error);
2902 }
2903 #endif
2904 
2905 int
2906 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2907 {
2908 	struct aiocb **acb_list;
2909 	struct sigevent *sigp, sig;
2910 	struct sigevent32 sig32;
2911 	uint32_t *acb_list32;
2912 	int error, i, nent;
2913 
2914 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2915 		return (EINVAL);
2916 
2917 	nent = uap->nent;
2918 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2919 		return (EINVAL);
2920 
2921 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2922 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2923 		if (error)
2924 			return (error);
2925 		error = convert_sigevent32(&sig32, &sig);
2926 		if (error)
2927 			return (error);
2928 		sigp = &sig;
2929 	} else
2930 		sigp = NULL;
2931 
2932 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2933 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2934 	if (error) {
2935 		free(acb_list32, M_LIO);
2936 		return (error);
2937 	}
2938 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2939 	for (i = 0; i < nent; i++)
2940 		acb_list[i] = PTRIN(acb_list32[i]);
2941 	free(acb_list32, M_LIO);
2942 
2943 	error = kern_lio_listio(td, uap->mode,
2944 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2945 	    &aiocb32_ops);
2946 	free(acb_list, M_LIO);
2947 	return (error);
2948 }
2949 
2950 #endif
2951