xref: /freebsd/sys/kern/vfs_aio.c (revision 89e3d5671ba13dceca272d5b159c9bd805f3f504)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. John S. Dyson's name may not be used to endorse or promote products
12  *    derived from this software without specific prior written permission.
13  *
14  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
15  * bad that happens because of using this software isn't the responsibility
16  * of the author.  This software is distributed AS-IS.
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/cdefs.h>
24 __FBSDID("$FreeBSD$");
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/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sema.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 #include <sys/sx.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
61 #include <sys/conf.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
65 
66 #include <machine/atomic.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 #include <vm/uma.h>
75 #include <sys/aio.h>
76 
77 /*
78  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
79  * overflow. (XXX will be removed soon.)
80  */
81 static u_long jobrefid;
82 
83 /*
84  * Counter for aio_fsync.
85  */
86 static uint64_t jobseqno;
87 
88 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC	32
90 #endif
91 
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC	256
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE
97 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
98 #endif
99 
100 #ifndef MAX_BUF_AIO
101 #define MAX_BUF_AIO		16
102 #endif
103 
104 FEATURE(aio, "Asynchronous I/O");
105 SYSCTL_DECL(_p1003_1b);
106 
107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
109 
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
111     "Async IO management");
112 
113 static int enable_aio_unsafe = 0;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
115     "Permit asynchronous IO on all file types, not just known-safe types");
116 
117 static unsigned int unsafe_warningcnt = 1;
118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
119     &unsafe_warningcnt, 0,
120     "Warnings that will be triggered upon failed IO requests on unsafe files");
121 
122 static int max_aio_procs = MAX_AIO_PROCS;
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
124     "Maximum number of kernel processes to use for handling async IO ");
125 
126 static int num_aio_procs = 0;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
128     "Number of presently active kernel processes for async IO");
129 
130 /*
131  * The code will adjust the actual number of AIO processes towards this
132  * number when it gets a chance.
133  */
134 static int target_aio_procs = TARGET_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
136     0,
137     "Preferred number of ready kernel processes for async IO");
138 
139 static int max_queue_count = MAX_AIO_QUEUE;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
141     "Maximum number of aio requests to queue, globally");
142 
143 static int num_queue_count = 0;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
145     "Number of queued aio requests");
146 
147 static int num_buf_aio = 0;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
149     "Number of aio requests presently handled by the buf subsystem");
150 
151 static int num_unmapped_aio = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
153     0,
154     "Number of aio requests presently handled by unmapped I/O buffers");
155 
156 /* Number of async I/O processes in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
159 
160 static int aiod_lifetime;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
162     "Maximum lifetime for idle aiod");
163 
164 static int max_aio_per_proc = MAX_AIO_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
166     0,
167     "Maximum active aio requests per process");
168 
169 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
171     &max_aio_queue_per_proc, 0,
172     "Maximum queued aio requests per process");
173 
174 static int max_buf_aio = MAX_BUF_AIO;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
176     "Maximum buf aio requests per process");
177 
178 /*
179  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
180  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
181  * vfs.aio.aio_listio_max.
182  */
183 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
184     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
185     0, "Maximum aio requests for a single lio_listio call");
186 
187 #ifdef COMPAT_FREEBSD6
188 typedef struct oaiocb {
189 	int	aio_fildes;		/* File descriptor */
190 	off_t	aio_offset;		/* File offset for I/O */
191 	volatile void *aio_buf;         /* I/O buffer in process space */
192 	size_t	aio_nbytes;		/* Number of bytes for I/O */
193 	struct	osigevent aio_sigevent;	/* Signal to deliver */
194 	int	aio_lio_opcode;		/* LIO opcode */
195 	int	aio_reqprio;		/* Request priority -- ignored */
196 	struct	__aiocb_private	_aiocb_private;
197 } oaiocb_t;
198 #endif
199 
200 /*
201  * Below is a key of locks used to protect each member of struct kaiocb
202  * aioliojob and kaioinfo and any backends.
203  *
204  * * - need not protected
205  * a - locked by kaioinfo lock
206  * b - locked by backend lock, the backend lock can be null in some cases,
207  *     for example, BIO belongs to this type, in this case, proc lock is
208  *     reused.
209  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210  */
211 
212 /*
213  * If the routine that services an AIO request blocks while running in an
214  * AIO kernel process it can starve other I/O requests.  BIO requests
215  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
216  * processes at all.  Socket I/O requests use a separate pool of
217  * kprocs and also force non-blocking I/O.  Other file I/O requests
218  * use the generic fo_read/fo_write operations which can block.  The
219  * fsync and mlock operations can also block while executing.  Ideally
220  * none of these requests would block while executing.
221  *
222  * Note that the service routines cannot toggle O_NONBLOCK in the file
223  * structure directly while handling a request due to races with
224  * userland threads.
225  */
226 
227 /* jobflags */
228 #define	KAIOCB_QUEUEING		0x01
229 #define	KAIOCB_CANCELLED	0x02
230 #define	KAIOCB_CANCELLING	0x04
231 #define	KAIOCB_CHECKSYNC	0x08
232 #define	KAIOCB_CLEARED		0x10
233 #define	KAIOCB_FINISHED		0x20
234 
235 /*
236  * AIO process info
237  */
238 #define AIOP_FREE	0x1			/* proc on free queue */
239 
240 struct aioproc {
241 	int	aioprocflags;			/* (c) AIO proc flags */
242 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
243 	struct	proc *aioproc;			/* (*) the AIO proc */
244 };
245 
246 /*
247  * data-structure for lio signal management
248  */
249 struct aioliojob {
250 	int	lioj_flags;			/* (a) listio flags */
251 	int	lioj_count;			/* (a) count of jobs */
252 	int	lioj_finished_count;		/* (a) count of finished jobs */
253 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
254 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
255 	struct	knlist klist;			/* (a) list of knotes */
256 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
257 };
258 
259 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
260 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
261 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
262 
263 /*
264  * per process aio data structure
265  */
266 struct kaioinfo {
267 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
268 	int	kaio_flags;		/* (a) per process kaio flags */
269 	int	kaio_active_count;	/* (c) number of currently used AIOs */
270 	int	kaio_count;		/* (a) size of AIO queue */
271 	int	kaio_buffer_count;	/* (a) number of bio buffers */
272 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
273 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
274 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
276 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
277 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
278 	struct	task kaio_task;		/* (*) task to kick aio processes */
279 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
280 };
281 
282 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
286 
287 #define KAIO_RUNDOWN	0x1	/* process is being run down */
288 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
289 
290 /*
291  * Operations used to interact with userland aio control blocks.
292  * Different ABIs provide their own operations.
293  */
294 struct aiocb_ops {
295 	int	(*aio_copyin)(struct aiocb *ujob, struct aiocb *kjob);
296 	long	(*fetch_status)(struct aiocb *ujob);
297 	long	(*fetch_error)(struct aiocb *ujob);
298 	int	(*store_status)(struct aiocb *ujob, long status);
299 	int	(*store_error)(struct aiocb *ujob, long error);
300 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
302 };
303 
304 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
308 static struct unrhdr *aiod_unr;
309 
310 void		aio_init_aioinfo(struct proc *p);
311 static int	aio_onceonly(void);
312 static int	aio_free_entry(struct kaiocb *job);
313 static void	aio_process_rw(struct kaiocb *job);
314 static void	aio_process_sync(struct kaiocb *job);
315 static void	aio_process_mlock(struct kaiocb *job);
316 static void	aio_schedule_fsync(void *context, int pending);
317 static int	aio_newproc(int *);
318 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
319 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
320 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
321 static void	aio_biowakeup(struct bio *bp);
322 static void	aio_proc_rundown(void *arg, struct proc *p);
323 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
324 		    struct image_params *imgp);
325 static int	aio_qbio(struct proc *p, struct kaiocb *job);
326 static void	aio_daemon(void *param);
327 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
328 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
329 static int	aio_kick(struct proc *userp);
330 static void	aio_kick_nowait(struct proc *userp);
331 static void	aio_kick_helper(void *context, int pending);
332 static int	filt_aioattach(struct knote *kn);
333 static void	filt_aiodetach(struct knote *kn);
334 static int	filt_aio(struct knote *kn, long hint);
335 static int	filt_lioattach(struct knote *kn);
336 static void	filt_liodetach(struct knote *kn);
337 static int	filt_lio(struct knote *kn, long hint);
338 
339 /*
340  * Zones for:
341  * 	kaio	Per process async io info
342  *	aiop	async io process data
343  *	aiocb	async io jobs
344  *	aiolio	list io jobs
345  */
346 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
347 
348 /* kqueue filters for aio */
349 static struct filterops aio_filtops = {
350 	.f_isfd = 0,
351 	.f_attach = filt_aioattach,
352 	.f_detach = filt_aiodetach,
353 	.f_event = filt_aio,
354 };
355 static struct filterops lio_filtops = {
356 	.f_isfd = 0,
357 	.f_attach = filt_lioattach,
358 	.f_detach = filt_liodetach,
359 	.f_event = filt_lio
360 };
361 
362 static eventhandler_tag exit_tag, exec_tag;
363 
364 TASKQUEUE_DEFINE_THREAD(aiod_kick);
365 
366 /*
367  * Main operations function for use as a kernel module.
368  */
369 static int
370 aio_modload(struct module *module, int cmd, void *arg)
371 {
372 	int error = 0;
373 
374 	switch (cmd) {
375 	case MOD_LOAD:
376 		aio_onceonly();
377 		break;
378 	case MOD_SHUTDOWN:
379 		break;
380 	default:
381 		error = EOPNOTSUPP;
382 		break;
383 	}
384 	return (error);
385 }
386 
387 static moduledata_t aio_mod = {
388 	"aio",
389 	&aio_modload,
390 	NULL
391 };
392 
393 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
394 MODULE_VERSION(aio, 1);
395 
396 /*
397  * Startup initialization
398  */
399 static int
400 aio_onceonly(void)
401 {
402 
403 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
404 	    EVENTHANDLER_PRI_ANY);
405 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
406 	    NULL, EVENTHANDLER_PRI_ANY);
407 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
408 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
409 	TAILQ_INIT(&aio_freeproc);
410 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
411 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
412 	TAILQ_INIT(&aio_jobs);
413 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
414 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
415 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
416 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
417 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
418 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
419 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
420 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
421 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
422 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
423 	jobrefid = 1;
424 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
425 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
426 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
427 
428 	return (0);
429 }
430 
431 /*
432  * Init the per-process aioinfo structure.  The aioinfo limits are set
433  * per-process for user limit (resource) management.
434  */
435 void
436 aio_init_aioinfo(struct proc *p)
437 {
438 	struct kaioinfo *ki;
439 
440 	ki = uma_zalloc(kaio_zone, M_WAITOK);
441 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
442 	ki->kaio_flags = 0;
443 	ki->kaio_active_count = 0;
444 	ki->kaio_count = 0;
445 	ki->kaio_buffer_count = 0;
446 	TAILQ_INIT(&ki->kaio_all);
447 	TAILQ_INIT(&ki->kaio_done);
448 	TAILQ_INIT(&ki->kaio_jobqueue);
449 	TAILQ_INIT(&ki->kaio_liojoblist);
450 	TAILQ_INIT(&ki->kaio_syncqueue);
451 	TAILQ_INIT(&ki->kaio_syncready);
452 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
453 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
454 	PROC_LOCK(p);
455 	if (p->p_aioinfo == NULL) {
456 		p->p_aioinfo = ki;
457 		PROC_UNLOCK(p);
458 	} else {
459 		PROC_UNLOCK(p);
460 		mtx_destroy(&ki->kaio_mtx);
461 		uma_zfree(kaio_zone, ki);
462 	}
463 
464 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
465 		aio_newproc(NULL);
466 }
467 
468 static int
469 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi, bool ext)
470 {
471 	struct thread *td;
472 	int error;
473 
474 	error = sigev_findtd(p, sigev, &td);
475 	if (error)
476 		return (error);
477 	if (!KSI_ONQ(ksi)) {
478 		ksiginfo_set_sigev(ksi, sigev);
479 		ksi->ksi_code = SI_ASYNCIO;
480 		ksi->ksi_flags |= ext ? (KSI_EXT | KSI_INS) : 0;
481 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
482 	}
483 	PROC_UNLOCK(p);
484 	return (error);
485 }
486 
487 /*
488  * Free a job entry.  Wait for completion if it is currently active, but don't
489  * delay forever.  If we delay, we return a flag that says that we have to
490  * restart the queue scan.
491  */
492 static int
493 aio_free_entry(struct kaiocb *job)
494 {
495 	struct kaioinfo *ki;
496 	struct aioliojob *lj;
497 	struct proc *p;
498 
499 	p = job->userproc;
500 	MPASS(curproc == p);
501 	ki = p->p_aioinfo;
502 	MPASS(ki != NULL);
503 
504 	AIO_LOCK_ASSERT(ki, MA_OWNED);
505 	MPASS(job->jobflags & KAIOCB_FINISHED);
506 
507 	atomic_subtract_int(&num_queue_count, 1);
508 
509 	ki->kaio_count--;
510 	MPASS(ki->kaio_count >= 0);
511 
512 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
513 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
514 
515 	lj = job->lio;
516 	if (lj) {
517 		lj->lioj_count--;
518 		lj->lioj_finished_count--;
519 
520 		if (lj->lioj_count == 0) {
521 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
522 			/* lio is going away, we need to destroy any knotes */
523 			knlist_delete(&lj->klist, curthread, 1);
524 			PROC_LOCK(p);
525 			sigqueue_take(&lj->lioj_ksi);
526 			PROC_UNLOCK(p);
527 			uma_zfree(aiolio_zone, lj);
528 		}
529 	}
530 
531 	/* job is going away, we need to destroy any knotes */
532 	knlist_delete(&job->klist, curthread, 1);
533 	PROC_LOCK(p);
534 	sigqueue_take(&job->ksi);
535 	PROC_UNLOCK(p);
536 
537 	AIO_UNLOCK(ki);
538 
539 	/*
540 	 * The thread argument here is used to find the owning process
541 	 * and is also passed to fo_close() which may pass it to various
542 	 * places such as devsw close() routines.  Because of that, we
543 	 * need a thread pointer from the process owning the job that is
544 	 * persistent and won't disappear out from under us or move to
545 	 * another process.
546 	 *
547 	 * Currently, all the callers of this function call it to remove
548 	 * a kaiocb from the current process' job list either via a
549 	 * syscall or due to the current process calling exit() or
550 	 * execve().  Thus, we know that p == curproc.  We also know that
551 	 * curthread can't exit since we are curthread.
552 	 *
553 	 * Therefore, we use curthread as the thread to pass to
554 	 * knlist_delete().  This does mean that it is possible for the
555 	 * thread pointer at close time to differ from the thread pointer
556 	 * at open time, but this is already true of file descriptors in
557 	 * a multithreaded process.
558 	 */
559 	if (job->fd_file)
560 		fdrop(job->fd_file, curthread);
561 	crfree(job->cred);
562 	uma_zfree(aiocb_zone, job);
563 	AIO_LOCK(ki);
564 
565 	return (0);
566 }
567 
568 static void
569 aio_proc_rundown_exec(void *arg, struct proc *p,
570     struct image_params *imgp __unused)
571 {
572    	aio_proc_rundown(arg, p);
573 }
574 
575 static int
576 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
577 {
578 	aio_cancel_fn_t *func;
579 	int cancelled;
580 
581 	AIO_LOCK_ASSERT(ki, MA_OWNED);
582 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
583 		return (0);
584 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
585 	job->jobflags |= KAIOCB_CANCELLED;
586 
587 	func = job->cancel_fn;
588 
589 	/*
590 	 * If there is no cancel routine, just leave the job marked as
591 	 * cancelled.  The job should be in active use by a caller who
592 	 * should complete it normally or when it fails to install a
593 	 * cancel routine.
594 	 */
595 	if (func == NULL)
596 		return (0);
597 
598 	/*
599 	 * Set the CANCELLING flag so that aio_complete() will defer
600 	 * completions of this job.  This prevents the job from being
601 	 * freed out from under the cancel callback.  After the
602 	 * callback any deferred completion (whether from the callback
603 	 * or any other source) will be completed.
604 	 */
605 	job->jobflags |= KAIOCB_CANCELLING;
606 	AIO_UNLOCK(ki);
607 	func(job);
608 	AIO_LOCK(ki);
609 	job->jobflags &= ~KAIOCB_CANCELLING;
610 	if (job->jobflags & KAIOCB_FINISHED) {
611 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
612 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
613 		aio_bio_done_notify(p, job);
614 	} else {
615 		/*
616 		 * The cancel callback might have scheduled an
617 		 * operation to cancel this request, but it is
618 		 * only counted as cancelled if the request is
619 		 * cancelled when the callback returns.
620 		 */
621 		cancelled = 0;
622 	}
623 	return (cancelled);
624 }
625 
626 /*
627  * Rundown the jobs for a given process.
628  */
629 static void
630 aio_proc_rundown(void *arg, struct proc *p)
631 {
632 	struct kaioinfo *ki;
633 	struct aioliojob *lj;
634 	struct kaiocb *job, *jobn;
635 
636 	KASSERT(curthread->td_proc == p,
637 	    ("%s: called on non-curproc", __func__));
638 	ki = p->p_aioinfo;
639 	if (ki == NULL)
640 		return;
641 
642 	AIO_LOCK(ki);
643 	ki->kaio_flags |= KAIO_RUNDOWN;
644 
645 restart:
646 
647 	/*
648 	 * Try to cancel all pending requests. This code simulates
649 	 * aio_cancel on all pending I/O requests.
650 	 */
651 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
652 		aio_cancel_job(p, ki, job);
653 	}
654 
655 	/* Wait for all running I/O to be finished */
656 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
657 		ki->kaio_flags |= KAIO_WAKEUP;
658 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
659 		goto restart;
660 	}
661 
662 	/* Free all completed I/O requests. */
663 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
664 		aio_free_entry(job);
665 
666 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
667 		if (lj->lioj_count == 0) {
668 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
669 			knlist_delete(&lj->klist, curthread, 1);
670 			PROC_LOCK(p);
671 			sigqueue_take(&lj->lioj_ksi);
672 			PROC_UNLOCK(p);
673 			uma_zfree(aiolio_zone, lj);
674 		} else {
675 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
676 			    lj->lioj_count, lj->lioj_finished_count);
677 		}
678 	}
679 	AIO_UNLOCK(ki);
680 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
681 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
682 	mtx_destroy(&ki->kaio_mtx);
683 	uma_zfree(kaio_zone, ki);
684 	p->p_aioinfo = NULL;
685 }
686 
687 /*
688  * Select a job to run (called by an AIO daemon).
689  */
690 static struct kaiocb *
691 aio_selectjob(struct aioproc *aiop)
692 {
693 	struct kaiocb *job;
694 	struct kaioinfo *ki;
695 	struct proc *userp;
696 
697 	mtx_assert(&aio_job_mtx, MA_OWNED);
698 restart:
699 	TAILQ_FOREACH(job, &aio_jobs, list) {
700 		userp = job->userproc;
701 		ki = userp->p_aioinfo;
702 
703 		if (ki->kaio_active_count < max_aio_per_proc) {
704 			TAILQ_REMOVE(&aio_jobs, job, list);
705 			if (!aio_clear_cancel_function(job))
706 				goto restart;
707 
708 			/* Account for currently active jobs. */
709 			ki->kaio_active_count++;
710 			break;
711 		}
712 	}
713 	return (job);
714 }
715 
716 /*
717  * Move all data to a permanent storage device.  This code
718  * simulates the fsync syscall.
719  */
720 static int
721 aio_fsync_vnode(struct thread *td, struct vnode *vp)
722 {
723 	struct mount *mp;
724 	int error;
725 
726 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
727 		goto drop;
728 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
729 	if (vp->v_object != NULL) {
730 		VM_OBJECT_WLOCK(vp->v_object);
731 		vm_object_page_clean(vp->v_object, 0, 0, 0);
732 		VM_OBJECT_WUNLOCK(vp->v_object);
733 	}
734 	error = VOP_FSYNC(vp, MNT_WAIT, td);
735 
736 	VOP_UNLOCK(vp);
737 	vn_finished_write(mp);
738 drop:
739 	return (error);
740 }
741 
742 /*
743  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
744  * does the I/O request for the non-bio version of the operations.  The normal
745  * vn operations are used, and this code should work in all instances for every
746  * type of file, including pipes, sockets, fifos, and regular files.
747  *
748  * XXX I don't think it works well for socket, pipe, and fifo.
749  */
750 static void
751 aio_process_rw(struct kaiocb *job)
752 {
753 	struct ucred *td_savedcred;
754 	struct thread *td;
755 	struct aiocb *cb;
756 	struct file *fp;
757 	struct uio auio;
758 	struct iovec aiov;
759 	ssize_t cnt;
760 	long msgsnd_st, msgsnd_end;
761 	long msgrcv_st, msgrcv_end;
762 	long oublock_st, oublock_end;
763 	long inblock_st, inblock_end;
764 	int error;
765 
766 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
767 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
768 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
769 
770 	aio_switch_vmspace(job);
771 	td = curthread;
772 	td_savedcred = td->td_ucred;
773 	td->td_ucred = job->cred;
774 	cb = &job->uaiocb;
775 	fp = job->fd_file;
776 
777 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
778 	aiov.iov_len = cb->aio_nbytes;
779 
780 	auio.uio_iov = &aiov;
781 	auio.uio_iovcnt = 1;
782 	auio.uio_offset = cb->aio_offset;
783 	auio.uio_resid = cb->aio_nbytes;
784 	cnt = cb->aio_nbytes;
785 	auio.uio_segflg = UIO_USERSPACE;
786 	auio.uio_td = td;
787 
788 	msgrcv_st = td->td_ru.ru_msgrcv;
789 	msgsnd_st = td->td_ru.ru_msgsnd;
790 	inblock_st = td->td_ru.ru_inblock;
791 	oublock_st = td->td_ru.ru_oublock;
792 
793 	/*
794 	 * aio_aqueue() acquires a reference to the file that is
795 	 * released in aio_free_entry().
796 	 */
797 	if (cb->aio_lio_opcode == LIO_READ) {
798 		auio.uio_rw = UIO_READ;
799 		if (auio.uio_resid == 0)
800 			error = 0;
801 		else
802 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
803 	} else {
804 		if (fp->f_type == DTYPE_VNODE)
805 			bwillwrite();
806 		auio.uio_rw = UIO_WRITE;
807 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
808 	}
809 	msgrcv_end = td->td_ru.ru_msgrcv;
810 	msgsnd_end = td->td_ru.ru_msgsnd;
811 	inblock_end = td->td_ru.ru_inblock;
812 	oublock_end = td->td_ru.ru_oublock;
813 
814 	job->msgrcv = msgrcv_end - msgrcv_st;
815 	job->msgsnd = msgsnd_end - msgsnd_st;
816 	job->inblock = inblock_end - inblock_st;
817 	job->outblock = oublock_end - oublock_st;
818 
819 	if ((error) && (auio.uio_resid != cnt)) {
820 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
821 			error = 0;
822 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
823 			PROC_LOCK(job->userproc);
824 			kern_psignal(job->userproc, SIGPIPE);
825 			PROC_UNLOCK(job->userproc);
826 		}
827 	}
828 
829 	cnt -= auio.uio_resid;
830 	td->td_ucred = td_savedcred;
831 	if (error)
832 		aio_complete(job, -1, error);
833 	else
834 		aio_complete(job, cnt, 0);
835 }
836 
837 static void
838 aio_process_sync(struct kaiocb *job)
839 {
840 	struct thread *td = curthread;
841 	struct ucred *td_savedcred = td->td_ucred;
842 	struct file *fp = job->fd_file;
843 	int error = 0;
844 
845 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
846 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
847 
848 	td->td_ucred = job->cred;
849 	if (fp->f_vnode != NULL)
850 		error = aio_fsync_vnode(td, fp->f_vnode);
851 	td->td_ucred = td_savedcred;
852 	if (error)
853 		aio_complete(job, -1, error);
854 	else
855 		aio_complete(job, 0, 0);
856 }
857 
858 static void
859 aio_process_mlock(struct kaiocb *job)
860 {
861 	struct aiocb *cb = &job->uaiocb;
862 	int error;
863 
864 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
865 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
866 
867 	aio_switch_vmspace(job);
868 	error = kern_mlock(job->userproc, job->cred,
869 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
870 	aio_complete(job, error != 0 ? -1 : 0, error);
871 }
872 
873 static void
874 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
875 {
876 	struct aioliojob *lj;
877 	struct kaioinfo *ki;
878 	struct kaiocb *sjob, *sjobn;
879 	int lj_done;
880 	bool schedule_fsync;
881 
882 	ki = userp->p_aioinfo;
883 	AIO_LOCK_ASSERT(ki, MA_OWNED);
884 	lj = job->lio;
885 	lj_done = 0;
886 	if (lj) {
887 		lj->lioj_finished_count++;
888 		if (lj->lioj_count == lj->lioj_finished_count)
889 			lj_done = 1;
890 	}
891 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
892 	MPASS(job->jobflags & KAIOCB_FINISHED);
893 
894 	if (ki->kaio_flags & KAIO_RUNDOWN)
895 		goto notification_done;
896 
897 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
898 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
899 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi, true);
900 
901 	KNOTE_LOCKED(&job->klist, 1);
902 
903 	if (lj_done) {
904 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
905 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
906 			KNOTE_LOCKED(&lj->klist, 1);
907 		}
908 		if ((lj->lioj_flags & (LIOJ_SIGNAL | LIOJ_SIGNAL_POSTED))
909 		    == LIOJ_SIGNAL &&
910 		    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
911 		    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
912 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi,
913 			    true);
914 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
915 		}
916 	}
917 
918 notification_done:
919 	if (job->jobflags & KAIOCB_CHECKSYNC) {
920 		schedule_fsync = false;
921 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
922 			if (job->fd_file != sjob->fd_file ||
923 			    job->seqno >= sjob->seqno)
924 				continue;
925 			if (--sjob->pending > 0)
926 				continue;
927 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
928 			if (!aio_clear_cancel_function_locked(sjob))
929 				continue;
930 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
931 			schedule_fsync = true;
932 		}
933 		if (schedule_fsync)
934 			taskqueue_enqueue(taskqueue_aiod_kick,
935 			    &ki->kaio_sync_task);
936 	}
937 	if (ki->kaio_flags & KAIO_WAKEUP) {
938 		ki->kaio_flags &= ~KAIO_WAKEUP;
939 		wakeup(&userp->p_aioinfo);
940 	}
941 }
942 
943 static void
944 aio_schedule_fsync(void *context, int pending)
945 {
946 	struct kaioinfo *ki;
947 	struct kaiocb *job;
948 
949 	ki = context;
950 	AIO_LOCK(ki);
951 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
952 		job = TAILQ_FIRST(&ki->kaio_syncready);
953 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
954 		AIO_UNLOCK(ki);
955 		aio_schedule(job, aio_process_sync);
956 		AIO_LOCK(ki);
957 	}
958 	AIO_UNLOCK(ki);
959 }
960 
961 bool
962 aio_cancel_cleared(struct kaiocb *job)
963 {
964 
965 	/*
966 	 * The caller should hold the same queue lock held when
967 	 * aio_clear_cancel_function() was called and set this flag
968 	 * ensuring this check sees an up-to-date value.  However,
969 	 * there is no way to assert that.
970 	 */
971 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
972 }
973 
974 static bool
975 aio_clear_cancel_function_locked(struct kaiocb *job)
976 {
977 
978 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
979 	MPASS(job->cancel_fn != NULL);
980 	if (job->jobflags & KAIOCB_CANCELLING) {
981 		job->jobflags |= KAIOCB_CLEARED;
982 		return (false);
983 	}
984 	job->cancel_fn = NULL;
985 	return (true);
986 }
987 
988 bool
989 aio_clear_cancel_function(struct kaiocb *job)
990 {
991 	struct kaioinfo *ki;
992 	bool ret;
993 
994 	ki = job->userproc->p_aioinfo;
995 	AIO_LOCK(ki);
996 	ret = aio_clear_cancel_function_locked(job);
997 	AIO_UNLOCK(ki);
998 	return (ret);
999 }
1000 
1001 static bool
1002 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1003 {
1004 
1005 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1006 	if (job->jobflags & KAIOCB_CANCELLED)
1007 		return (false);
1008 	job->cancel_fn = func;
1009 	return (true);
1010 }
1011 
1012 bool
1013 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1014 {
1015 	struct kaioinfo *ki;
1016 	bool ret;
1017 
1018 	ki = job->userproc->p_aioinfo;
1019 	AIO_LOCK(ki);
1020 	ret = aio_set_cancel_function_locked(job, func);
1021 	AIO_UNLOCK(ki);
1022 	return (ret);
1023 }
1024 
1025 void
1026 aio_complete(struct kaiocb *job, long status, int error)
1027 {
1028 	struct kaioinfo *ki;
1029 	struct proc *userp;
1030 
1031 	job->uaiocb._aiocb_private.error = error;
1032 	job->uaiocb._aiocb_private.status = status;
1033 
1034 	userp = job->userproc;
1035 	ki = userp->p_aioinfo;
1036 
1037 	AIO_LOCK(ki);
1038 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1039 	    ("duplicate aio_complete"));
1040 	job->jobflags |= KAIOCB_FINISHED;
1041 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1042 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1043 		aio_bio_done_notify(userp, job);
1044 	}
1045 	AIO_UNLOCK(ki);
1046 }
1047 
1048 void
1049 aio_cancel(struct kaiocb *job)
1050 {
1051 
1052 	aio_complete(job, -1, ECANCELED);
1053 }
1054 
1055 void
1056 aio_switch_vmspace(struct kaiocb *job)
1057 {
1058 
1059 	vmspace_switch_aio(job->userproc->p_vmspace);
1060 }
1061 
1062 /*
1063  * The AIO daemon, most of the actual work is done in aio_process_*,
1064  * but the setup (and address space mgmt) is done in this routine.
1065  */
1066 static void
1067 aio_daemon(void *_id)
1068 {
1069 	struct kaiocb *job;
1070 	struct aioproc *aiop;
1071 	struct kaioinfo *ki;
1072 	struct proc *p;
1073 	struct vmspace *myvm;
1074 	struct thread *td = curthread;
1075 	int id = (intptr_t)_id;
1076 
1077 	/*
1078 	 * Grab an extra reference on the daemon's vmspace so that it
1079 	 * doesn't get freed by jobs that switch to a different
1080 	 * vmspace.
1081 	 */
1082 	p = td->td_proc;
1083 	myvm = vmspace_acquire_ref(p);
1084 
1085 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1086 
1087 	/*
1088 	 * Allocate and ready the aio control info.  There is one aiop structure
1089 	 * per daemon.
1090 	 */
1091 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1092 	aiop->aioproc = p;
1093 	aiop->aioprocflags = 0;
1094 
1095 	/*
1096 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1097 	 * and creating too many daemons.)
1098 	 */
1099 	sema_post(&aio_newproc_sem);
1100 
1101 	mtx_lock(&aio_job_mtx);
1102 	for (;;) {
1103 		/*
1104 		 * Take daemon off of free queue
1105 		 */
1106 		if (aiop->aioprocflags & AIOP_FREE) {
1107 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1108 			aiop->aioprocflags &= ~AIOP_FREE;
1109 		}
1110 
1111 		/*
1112 		 * Check for jobs.
1113 		 */
1114 		while ((job = aio_selectjob(aiop)) != NULL) {
1115 			mtx_unlock(&aio_job_mtx);
1116 
1117 			ki = job->userproc->p_aioinfo;
1118 			job->handle_fn(job);
1119 
1120 			mtx_lock(&aio_job_mtx);
1121 			/* Decrement the active job count. */
1122 			ki->kaio_active_count--;
1123 		}
1124 
1125 		/*
1126 		 * Disconnect from user address space.
1127 		 */
1128 		if (p->p_vmspace != myvm) {
1129 			mtx_unlock(&aio_job_mtx);
1130 			vmspace_switch_aio(myvm);
1131 			mtx_lock(&aio_job_mtx);
1132 			/*
1133 			 * We have to restart to avoid race, we only sleep if
1134 			 * no job can be selected.
1135 			 */
1136 			continue;
1137 		}
1138 
1139 		mtx_assert(&aio_job_mtx, MA_OWNED);
1140 
1141 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1142 		aiop->aioprocflags |= AIOP_FREE;
1143 
1144 		/*
1145 		 * If daemon is inactive for a long time, allow it to exit,
1146 		 * thereby freeing resources.
1147 		 */
1148 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1149 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1150 		    (aiop->aioprocflags & AIOP_FREE) &&
1151 		    num_aio_procs > target_aio_procs)
1152 			break;
1153 	}
1154 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1155 	num_aio_procs--;
1156 	mtx_unlock(&aio_job_mtx);
1157 	uma_zfree(aiop_zone, aiop);
1158 	free_unr(aiod_unr, id);
1159 	vmspace_free(myvm);
1160 
1161 	KASSERT(p->p_vmspace == myvm,
1162 	    ("AIOD: bad vmspace for exiting daemon"));
1163 	KASSERT(refcount_load(&myvm->vm_refcnt) > 1,
1164 	    ("AIOD: bad vm refcnt for exiting daemon: %d",
1165 	    refcount_load(&myvm->vm_refcnt)));
1166 	kproc_exit(0);
1167 }
1168 
1169 /*
1170  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1171  * AIO daemon modifies its environment itself.
1172  */
1173 static int
1174 aio_newproc(int *start)
1175 {
1176 	int error;
1177 	struct proc *p;
1178 	int id;
1179 
1180 	id = alloc_unr(aiod_unr);
1181 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1182 		RFNOWAIT, 0, "aiod%d", id);
1183 	if (error == 0) {
1184 		/*
1185 		 * Wait until daemon is started.
1186 		 */
1187 		sema_wait(&aio_newproc_sem);
1188 		mtx_lock(&aio_job_mtx);
1189 		num_aio_procs++;
1190 		if (start != NULL)
1191 			(*start)--;
1192 		mtx_unlock(&aio_job_mtx);
1193 	} else {
1194 		free_unr(aiod_unr, id);
1195 	}
1196 	return (error);
1197 }
1198 
1199 /*
1200  * Try the high-performance, low-overhead bio method for eligible
1201  * VCHR devices.  This method doesn't use an aio helper thread, and
1202  * thus has very low overhead.
1203  *
1204  * Assumes that the caller, aio_aqueue(), has incremented the file
1205  * structure's reference count, preventing its deallocation for the
1206  * duration of this call.
1207  */
1208 static int
1209 aio_qbio(struct proc *p, struct kaiocb *job)
1210 {
1211 	struct aiocb *cb;
1212 	struct file *fp;
1213 	struct bio *bp;
1214 	struct buf *pbuf;
1215 	struct vnode *vp;
1216 	struct cdevsw *csw;
1217 	struct cdev *dev;
1218 	struct kaioinfo *ki;
1219 	struct vm_page **pages;
1220 	int error, npages, poff, ref;
1221 	vm_prot_t prot;
1222 
1223 	cb = &job->uaiocb;
1224 	fp = job->fd_file;
1225 
1226 	if (!(cb->aio_lio_opcode == LIO_WRITE ||
1227 	    cb->aio_lio_opcode == LIO_READ))
1228 		return (-1);
1229 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1230 		return (-1);
1231 
1232 	vp = fp->f_vnode;
1233 	if (vp->v_type != VCHR)
1234 		return (-1);
1235 	if (vp->v_bufobj.bo_bsize == 0)
1236 		return (-1);
1237 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1238 		return (-1);
1239 
1240 	ref = 0;
1241 	csw = devvn_refthread(vp, &dev, &ref);
1242 	if (csw == NULL)
1243 		return (ENXIO);
1244 
1245 	if ((csw->d_flags & D_DISK) == 0) {
1246 		error = -1;
1247 		goto unref;
1248 	}
1249 	if (cb->aio_nbytes > dev->si_iosize_max) {
1250 		error = -1;
1251 		goto unref;
1252 	}
1253 
1254 	ki = p->p_aioinfo;
1255 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1256 	if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1257 		if (cb->aio_nbytes > maxphys) {
1258 			error = -1;
1259 			goto unref;
1260 		}
1261 
1262 		pbuf = NULL;
1263 		pages = malloc(sizeof(vm_page_t) * (atop(round_page(
1264 		    cb->aio_nbytes)) + 1), M_TEMP, M_WAITOK | M_ZERO);
1265 	} else {
1266 		if (cb->aio_nbytes > maxphys) {
1267 			error = -1;
1268 			goto unref;
1269 		}
1270 		if (ki->kaio_buffer_count >= max_buf_aio) {
1271 			error = EAGAIN;
1272 			goto unref;
1273 		}
1274 
1275 		pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
1276 		BUF_KERNPROC(pbuf);
1277 		AIO_LOCK(ki);
1278 		ki->kaio_buffer_count++;
1279 		AIO_UNLOCK(ki);
1280 		pages = pbuf->b_pages;
1281 	}
1282 	bp = g_alloc_bio();
1283 
1284 	bp->bio_length = cb->aio_nbytes;
1285 	bp->bio_bcount = cb->aio_nbytes;
1286 	bp->bio_done = aio_biowakeup;
1287 	bp->bio_offset = cb->aio_offset;
1288 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1289 	bp->bio_dev = dev;
1290 	bp->bio_caller1 = job;
1291 	bp->bio_caller2 = pbuf;
1292 
1293 	prot = VM_PROT_READ;
1294 	if (cb->aio_lio_opcode == LIO_READ)
1295 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1296 	npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1297 	    (vm_offset_t)cb->aio_buf, bp->bio_length, prot, pages,
1298 	    atop(maxphys) + 1);
1299 	if (npages < 0) {
1300 		error = EFAULT;
1301 		goto doerror;
1302 	}
1303 	if (pbuf != NULL) {
1304 		pmap_qenter((vm_offset_t)pbuf->b_data, pages, npages);
1305 		bp->bio_data = pbuf->b_data + poff;
1306 		atomic_add_int(&num_buf_aio, 1);
1307 		pbuf->b_npages = npages;
1308 	} else {
1309 		bp->bio_ma = pages;
1310 		bp->bio_ma_n = npages;
1311 		bp->bio_ma_offset = poff;
1312 		bp->bio_data = unmapped_buf;
1313 		bp->bio_flags |= BIO_UNMAPPED;
1314 		atomic_add_int(&num_unmapped_aio, 1);
1315 	}
1316 
1317 	/* Perform transfer. */
1318 	csw->d_strategy(bp);
1319 	dev_relthread(dev, ref);
1320 	return (0);
1321 
1322 doerror:
1323 	if (pbuf != NULL) {
1324 		AIO_LOCK(ki);
1325 		ki->kaio_buffer_count--;
1326 		AIO_UNLOCK(ki);
1327 		uma_zfree(pbuf_zone, pbuf);
1328 	} else {
1329 		free(pages, M_TEMP);
1330 	}
1331 	g_destroy_bio(bp);
1332 unref:
1333 	dev_relthread(dev, ref);
1334 	return (error);
1335 }
1336 
1337 #ifdef COMPAT_FREEBSD6
1338 static int
1339 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1340 {
1341 
1342 	/*
1343 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1344 	 * supported by AIO with the old sigevent structure.
1345 	 */
1346 	nsig->sigev_notify = osig->sigev_notify;
1347 	switch (nsig->sigev_notify) {
1348 	case SIGEV_NONE:
1349 		break;
1350 	case SIGEV_SIGNAL:
1351 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1352 		break;
1353 	case SIGEV_KEVENT:
1354 		nsig->sigev_notify_kqueue =
1355 		    osig->__sigev_u.__sigev_notify_kqueue;
1356 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1357 		break;
1358 	default:
1359 		return (EINVAL);
1360 	}
1361 	return (0);
1362 }
1363 
1364 static int
1365 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1366 {
1367 	struct oaiocb *ojob;
1368 	int error;
1369 
1370 	bzero(kjob, sizeof(struct aiocb));
1371 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1372 	if (error)
1373 		return (error);
1374 	ojob = (struct oaiocb *)kjob;
1375 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1376 }
1377 #endif
1378 
1379 static int
1380 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1381 {
1382 
1383 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1384 }
1385 
1386 static long
1387 aiocb_fetch_status(struct aiocb *ujob)
1388 {
1389 
1390 	return (fuword(&ujob->_aiocb_private.status));
1391 }
1392 
1393 static long
1394 aiocb_fetch_error(struct aiocb *ujob)
1395 {
1396 
1397 	return (fuword(&ujob->_aiocb_private.error));
1398 }
1399 
1400 static int
1401 aiocb_store_status(struct aiocb *ujob, long status)
1402 {
1403 
1404 	return (suword(&ujob->_aiocb_private.status, status));
1405 }
1406 
1407 static int
1408 aiocb_store_error(struct aiocb *ujob, long error)
1409 {
1410 
1411 	return (suword(&ujob->_aiocb_private.error, error));
1412 }
1413 
1414 static int
1415 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1416 {
1417 
1418 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1419 }
1420 
1421 static int
1422 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1423 {
1424 
1425 	return (suword(ujobp, (long)ujob));
1426 }
1427 
1428 static struct aiocb_ops aiocb_ops = {
1429 	.aio_copyin = aiocb_copyin,
1430 	.fetch_status = aiocb_fetch_status,
1431 	.fetch_error = aiocb_fetch_error,
1432 	.store_status = aiocb_store_status,
1433 	.store_error = aiocb_store_error,
1434 	.store_kernelinfo = aiocb_store_kernelinfo,
1435 	.store_aiocb = aiocb_store_aiocb,
1436 };
1437 
1438 #ifdef COMPAT_FREEBSD6
1439 static struct aiocb_ops aiocb_ops_osigevent = {
1440 	.aio_copyin = aiocb_copyin_old_sigevent,
1441 	.fetch_status = aiocb_fetch_status,
1442 	.fetch_error = aiocb_fetch_error,
1443 	.store_status = aiocb_store_status,
1444 	.store_error = aiocb_store_error,
1445 	.store_kernelinfo = aiocb_store_kernelinfo,
1446 	.store_aiocb = aiocb_store_aiocb,
1447 };
1448 #endif
1449 
1450 /*
1451  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1452  * technique is done in this code.
1453  */
1454 int
1455 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1456     int type, struct aiocb_ops *ops)
1457 {
1458 	struct proc *p = td->td_proc;
1459 	struct file *fp;
1460 	struct kaiocb *job;
1461 	struct kaioinfo *ki;
1462 	struct kevent kev;
1463 	int opcode;
1464 	int error;
1465 	int fd, kqfd;
1466 	int jid;
1467 	u_short evflags;
1468 
1469 	if (p->p_aioinfo == NULL)
1470 		aio_init_aioinfo(p);
1471 
1472 	ki = p->p_aioinfo;
1473 
1474 	ops->store_status(ujob, -1);
1475 	ops->store_error(ujob, 0);
1476 	ops->store_kernelinfo(ujob, -1);
1477 
1478 	if (num_queue_count >= max_queue_count ||
1479 	    ki->kaio_count >= max_aio_queue_per_proc) {
1480 		ops->store_error(ujob, EAGAIN);
1481 		return (EAGAIN);
1482 	}
1483 
1484 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1485 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1486 
1487 	error = ops->aio_copyin(ujob, &job->uaiocb);
1488 	if (error) {
1489 		ops->store_error(ujob, error);
1490 		uma_zfree(aiocb_zone, job);
1491 		return (error);
1492 	}
1493 
1494 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1495 		uma_zfree(aiocb_zone, job);
1496 		return (EINVAL);
1497 	}
1498 
1499 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1500 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1501 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1502 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1503 		ops->store_error(ujob, EINVAL);
1504 		uma_zfree(aiocb_zone, job);
1505 		return (EINVAL);
1506 	}
1507 
1508 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1509 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1510 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1511 		uma_zfree(aiocb_zone, job);
1512 		return (EINVAL);
1513 	}
1514 
1515 	ksiginfo_init(&job->ksi);
1516 
1517 	/* Save userspace address of the job info. */
1518 	job->ujob = ujob;
1519 
1520 	/* Get the opcode. */
1521 	if (type != LIO_NOP)
1522 		job->uaiocb.aio_lio_opcode = type;
1523 	opcode = job->uaiocb.aio_lio_opcode;
1524 
1525 	/*
1526 	 * Validate the opcode and fetch the file object for the specified
1527 	 * file descriptor.
1528 	 *
1529 	 * XXXRW: Moved the opcode validation up here so that we don't
1530 	 * retrieve a file descriptor without knowing what the capabiltity
1531 	 * should be.
1532 	 */
1533 	fd = job->uaiocb.aio_fildes;
1534 	switch (opcode) {
1535 	case LIO_WRITE:
1536 		error = fget_write(td, fd, &cap_pwrite_rights, &fp);
1537 		break;
1538 	case LIO_READ:
1539 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1540 		break;
1541 	case LIO_SYNC:
1542 		error = fget(td, fd, &cap_fsync_rights, &fp);
1543 		break;
1544 	case LIO_MLOCK:
1545 		fp = NULL;
1546 		break;
1547 	case LIO_NOP:
1548 		error = fget(td, fd, &cap_no_rights, &fp);
1549 		break;
1550 	default:
1551 		error = EINVAL;
1552 	}
1553 	if (error) {
1554 		uma_zfree(aiocb_zone, job);
1555 		ops->store_error(ujob, error);
1556 		return (error);
1557 	}
1558 
1559 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1560 		error = EINVAL;
1561 		goto aqueue_fail;
1562 	}
1563 
1564 	if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1565 	    job->uaiocb.aio_offset < 0 &&
1566 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1567 		error = EINVAL;
1568 		goto aqueue_fail;
1569 	}
1570 
1571 	job->fd_file = fp;
1572 
1573 	mtx_lock(&aio_job_mtx);
1574 	jid = jobrefid++;
1575 	job->seqno = jobseqno++;
1576 	mtx_unlock(&aio_job_mtx);
1577 	error = ops->store_kernelinfo(ujob, jid);
1578 	if (error) {
1579 		error = EINVAL;
1580 		goto aqueue_fail;
1581 	}
1582 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1583 
1584 	if (opcode == LIO_NOP) {
1585 		fdrop(fp, td);
1586 		uma_zfree(aiocb_zone, job);
1587 		return (0);
1588 	}
1589 
1590 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1591 		goto no_kqueue;
1592 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1593 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1594 		error = EINVAL;
1595 		goto aqueue_fail;
1596 	}
1597 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1598 	memset(&kev, 0, sizeof(kev));
1599 	kev.ident = (uintptr_t)job->ujob;
1600 	kev.filter = EVFILT_AIO;
1601 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1602 	kev.data = (intptr_t)job;
1603 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1604 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1605 	if (error)
1606 		goto aqueue_fail;
1607 
1608 no_kqueue:
1609 
1610 	ops->store_error(ujob, EINPROGRESS);
1611 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1612 	job->userproc = p;
1613 	job->cred = crhold(td->td_ucred);
1614 	job->jobflags = KAIOCB_QUEUEING;
1615 	job->lio = lj;
1616 
1617 	if (opcode == LIO_MLOCK) {
1618 		aio_schedule(job, aio_process_mlock);
1619 		error = 0;
1620 	} else if (fp->f_ops->fo_aio_queue == NULL)
1621 		error = aio_queue_file(fp, job);
1622 	else
1623 		error = fo_aio_queue(fp, job);
1624 	if (error)
1625 		goto aqueue_fail;
1626 
1627 	AIO_LOCK(ki);
1628 	job->jobflags &= ~KAIOCB_QUEUEING;
1629 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1630 	ki->kaio_count++;
1631 	if (lj)
1632 		lj->lioj_count++;
1633 	atomic_add_int(&num_queue_count, 1);
1634 	if (job->jobflags & KAIOCB_FINISHED) {
1635 		/*
1636 		 * The queue callback completed the request synchronously.
1637 		 * The bulk of the completion is deferred in that case
1638 		 * until this point.
1639 		 */
1640 		aio_bio_done_notify(p, job);
1641 	} else
1642 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1643 	AIO_UNLOCK(ki);
1644 	return (0);
1645 
1646 aqueue_fail:
1647 	knlist_delete(&job->klist, curthread, 0);
1648 	if (fp)
1649 		fdrop(fp, td);
1650 	uma_zfree(aiocb_zone, job);
1651 	ops->store_error(ujob, error);
1652 	return (error);
1653 }
1654 
1655 static void
1656 aio_cancel_daemon_job(struct kaiocb *job)
1657 {
1658 
1659 	mtx_lock(&aio_job_mtx);
1660 	if (!aio_cancel_cleared(job))
1661 		TAILQ_REMOVE(&aio_jobs, job, list);
1662 	mtx_unlock(&aio_job_mtx);
1663 	aio_cancel(job);
1664 }
1665 
1666 void
1667 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1668 {
1669 
1670 	mtx_lock(&aio_job_mtx);
1671 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1672 		mtx_unlock(&aio_job_mtx);
1673 		aio_cancel(job);
1674 		return;
1675 	}
1676 	job->handle_fn = func;
1677 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1678 	aio_kick_nowait(job->userproc);
1679 	mtx_unlock(&aio_job_mtx);
1680 }
1681 
1682 static void
1683 aio_cancel_sync(struct kaiocb *job)
1684 {
1685 	struct kaioinfo *ki;
1686 
1687 	ki = job->userproc->p_aioinfo;
1688 	AIO_LOCK(ki);
1689 	if (!aio_cancel_cleared(job))
1690 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1691 	AIO_UNLOCK(ki);
1692 	aio_cancel(job);
1693 }
1694 
1695 int
1696 aio_queue_file(struct file *fp, struct kaiocb *job)
1697 {
1698 	struct kaioinfo *ki;
1699 	struct kaiocb *job2;
1700 	struct vnode *vp;
1701 	struct mount *mp;
1702 	int error;
1703 	bool safe;
1704 
1705 	ki = job->userproc->p_aioinfo;
1706 	error = aio_qbio(job->userproc, job);
1707 	if (error >= 0)
1708 		return (error);
1709 	safe = false;
1710 	if (fp->f_type == DTYPE_VNODE) {
1711 		vp = fp->f_vnode;
1712 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1713 			mp = fp->f_vnode->v_mount;
1714 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1715 				safe = true;
1716 		}
1717 	}
1718 	if (!(safe || enable_aio_unsafe)) {
1719 		counted_warning(&unsafe_warningcnt,
1720 		    "is attempting to use unsafe AIO requests");
1721 		return (EOPNOTSUPP);
1722 	}
1723 
1724 	switch (job->uaiocb.aio_lio_opcode) {
1725 	case LIO_READ:
1726 	case LIO_WRITE:
1727 		aio_schedule(job, aio_process_rw);
1728 		error = 0;
1729 		break;
1730 	case LIO_SYNC:
1731 		AIO_LOCK(ki);
1732 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1733 			if (job2->fd_file == job->fd_file &&
1734 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1735 			    job2->seqno < job->seqno) {
1736 				job2->jobflags |= KAIOCB_CHECKSYNC;
1737 				job->pending++;
1738 			}
1739 		}
1740 		if (job->pending != 0) {
1741 			if (!aio_set_cancel_function_locked(job,
1742 				aio_cancel_sync)) {
1743 				AIO_UNLOCK(ki);
1744 				aio_cancel(job);
1745 				return (0);
1746 			}
1747 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1748 			AIO_UNLOCK(ki);
1749 			return (0);
1750 		}
1751 		AIO_UNLOCK(ki);
1752 		aio_schedule(job, aio_process_sync);
1753 		error = 0;
1754 		break;
1755 	default:
1756 		error = EINVAL;
1757 	}
1758 	return (error);
1759 }
1760 
1761 static void
1762 aio_kick_nowait(struct proc *userp)
1763 {
1764 	struct kaioinfo *ki = userp->p_aioinfo;
1765 	struct aioproc *aiop;
1766 
1767 	mtx_assert(&aio_job_mtx, MA_OWNED);
1768 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1769 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1770 		aiop->aioprocflags &= ~AIOP_FREE;
1771 		wakeup(aiop->aioproc);
1772 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1773 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1774 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1775 	}
1776 }
1777 
1778 static int
1779 aio_kick(struct proc *userp)
1780 {
1781 	struct kaioinfo *ki = userp->p_aioinfo;
1782 	struct aioproc *aiop;
1783 	int error, ret = 0;
1784 
1785 	mtx_assert(&aio_job_mtx, MA_OWNED);
1786 retryproc:
1787 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1788 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1789 		aiop->aioprocflags &= ~AIOP_FREE;
1790 		wakeup(aiop->aioproc);
1791 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1792 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1793 		num_aio_resv_start++;
1794 		mtx_unlock(&aio_job_mtx);
1795 		error = aio_newproc(&num_aio_resv_start);
1796 		mtx_lock(&aio_job_mtx);
1797 		if (error) {
1798 			num_aio_resv_start--;
1799 			goto retryproc;
1800 		}
1801 	} else {
1802 		ret = -1;
1803 	}
1804 	return (ret);
1805 }
1806 
1807 static void
1808 aio_kick_helper(void *context, int pending)
1809 {
1810 	struct proc *userp = context;
1811 
1812 	mtx_lock(&aio_job_mtx);
1813 	while (--pending >= 0) {
1814 		if (aio_kick(userp))
1815 			break;
1816 	}
1817 	mtx_unlock(&aio_job_mtx);
1818 }
1819 
1820 /*
1821  * Support the aio_return system call, as a side-effect, kernel resources are
1822  * released.
1823  */
1824 static int
1825 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1826 {
1827 	struct proc *p = td->td_proc;
1828 	struct kaiocb *job;
1829 	struct kaioinfo *ki;
1830 	long status, error;
1831 
1832 	ki = p->p_aioinfo;
1833 	if (ki == NULL)
1834 		return (EINVAL);
1835 	AIO_LOCK(ki);
1836 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1837 		if (job->ujob == ujob)
1838 			break;
1839 	}
1840 	if (job != NULL) {
1841 		MPASS(job->jobflags & KAIOCB_FINISHED);
1842 		status = job->uaiocb._aiocb_private.status;
1843 		error = job->uaiocb._aiocb_private.error;
1844 		td->td_retval[0] = status;
1845 		td->td_ru.ru_oublock += job->outblock;
1846 		td->td_ru.ru_inblock += job->inblock;
1847 		td->td_ru.ru_msgsnd += job->msgsnd;
1848 		td->td_ru.ru_msgrcv += job->msgrcv;
1849 		aio_free_entry(job);
1850 		AIO_UNLOCK(ki);
1851 		ops->store_error(ujob, error);
1852 		ops->store_status(ujob, status);
1853 	} else {
1854 		error = EINVAL;
1855 		AIO_UNLOCK(ki);
1856 	}
1857 	return (error);
1858 }
1859 
1860 int
1861 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1862 {
1863 
1864 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1865 }
1866 
1867 /*
1868  * Allow a process to wakeup when any of the I/O requests are completed.
1869  */
1870 static int
1871 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1872     struct timespec *ts)
1873 {
1874 	struct proc *p = td->td_proc;
1875 	struct timeval atv;
1876 	struct kaioinfo *ki;
1877 	struct kaiocb *firstjob, *job;
1878 	int error, i, timo;
1879 
1880 	timo = 0;
1881 	if (ts) {
1882 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1883 			return (EINVAL);
1884 
1885 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1886 		if (itimerfix(&atv))
1887 			return (EINVAL);
1888 		timo = tvtohz(&atv);
1889 	}
1890 
1891 	ki = p->p_aioinfo;
1892 	if (ki == NULL)
1893 		return (EAGAIN);
1894 
1895 	if (njoblist == 0)
1896 		return (0);
1897 
1898 	AIO_LOCK(ki);
1899 	for (;;) {
1900 		firstjob = NULL;
1901 		error = 0;
1902 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1903 			for (i = 0; i < njoblist; i++) {
1904 				if (job->ujob == ujoblist[i]) {
1905 					if (firstjob == NULL)
1906 						firstjob = job;
1907 					if (job->jobflags & KAIOCB_FINISHED)
1908 						goto RETURN;
1909 				}
1910 			}
1911 		}
1912 		/* All tasks were finished. */
1913 		if (firstjob == NULL)
1914 			break;
1915 
1916 		ki->kaio_flags |= KAIO_WAKEUP;
1917 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1918 		    "aiospn", timo);
1919 		if (error == ERESTART)
1920 			error = EINTR;
1921 		if (error)
1922 			break;
1923 	}
1924 RETURN:
1925 	AIO_UNLOCK(ki);
1926 	return (error);
1927 }
1928 
1929 int
1930 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1931 {
1932 	struct timespec ts, *tsp;
1933 	struct aiocb **ujoblist;
1934 	int error;
1935 
1936 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1937 		return (EINVAL);
1938 
1939 	if (uap->timeout) {
1940 		/* Get timespec struct. */
1941 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1942 			return (error);
1943 		tsp = &ts;
1944 	} else
1945 		tsp = NULL;
1946 
1947 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
1948 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1949 	if (error == 0)
1950 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1951 	free(ujoblist, M_AIOS);
1952 	return (error);
1953 }
1954 
1955 /*
1956  * aio_cancel cancels any non-bio aio operations not currently in progress.
1957  */
1958 int
1959 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1960 {
1961 	struct proc *p = td->td_proc;
1962 	struct kaioinfo *ki;
1963 	struct kaiocb *job, *jobn;
1964 	struct file *fp;
1965 	int error;
1966 	int cancelled = 0;
1967 	int notcancelled = 0;
1968 	struct vnode *vp;
1969 
1970 	/* Lookup file object. */
1971 	error = fget(td, uap->fd, &cap_no_rights, &fp);
1972 	if (error)
1973 		return (error);
1974 
1975 	ki = p->p_aioinfo;
1976 	if (ki == NULL)
1977 		goto done;
1978 
1979 	if (fp->f_type == DTYPE_VNODE) {
1980 		vp = fp->f_vnode;
1981 		if (vn_isdisk(vp)) {
1982 			fdrop(fp, td);
1983 			td->td_retval[0] = AIO_NOTCANCELED;
1984 			return (0);
1985 		}
1986 	}
1987 
1988 	AIO_LOCK(ki);
1989 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1990 		if ((uap->fd == job->uaiocb.aio_fildes) &&
1991 		    ((uap->aiocbp == NULL) ||
1992 		     (uap->aiocbp == job->ujob))) {
1993 			if (aio_cancel_job(p, ki, job)) {
1994 				cancelled++;
1995 			} else {
1996 				notcancelled++;
1997 			}
1998 			if (uap->aiocbp != NULL)
1999 				break;
2000 		}
2001 	}
2002 	AIO_UNLOCK(ki);
2003 
2004 done:
2005 	fdrop(fp, td);
2006 
2007 	if (uap->aiocbp != NULL) {
2008 		if (cancelled) {
2009 			td->td_retval[0] = AIO_CANCELED;
2010 			return (0);
2011 		}
2012 	}
2013 
2014 	if (notcancelled) {
2015 		td->td_retval[0] = AIO_NOTCANCELED;
2016 		return (0);
2017 	}
2018 
2019 	if (cancelled) {
2020 		td->td_retval[0] = AIO_CANCELED;
2021 		return (0);
2022 	}
2023 
2024 	td->td_retval[0] = AIO_ALLDONE;
2025 
2026 	return (0);
2027 }
2028 
2029 /*
2030  * aio_error is implemented in the kernel level for compatibility purposes
2031  * only.  For a user mode async implementation, it would be best to do it in
2032  * a userland subroutine.
2033  */
2034 static int
2035 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2036 {
2037 	struct proc *p = td->td_proc;
2038 	struct kaiocb *job;
2039 	struct kaioinfo *ki;
2040 	int status;
2041 
2042 	ki = p->p_aioinfo;
2043 	if (ki == NULL) {
2044 		td->td_retval[0] = EINVAL;
2045 		return (0);
2046 	}
2047 
2048 	AIO_LOCK(ki);
2049 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2050 		if (job->ujob == ujob) {
2051 			if (job->jobflags & KAIOCB_FINISHED)
2052 				td->td_retval[0] =
2053 					job->uaiocb._aiocb_private.error;
2054 			else
2055 				td->td_retval[0] = EINPROGRESS;
2056 			AIO_UNLOCK(ki);
2057 			return (0);
2058 		}
2059 	}
2060 	AIO_UNLOCK(ki);
2061 
2062 	/*
2063 	 * Hack for failure of aio_aqueue.
2064 	 */
2065 	status = ops->fetch_status(ujob);
2066 	if (status == -1) {
2067 		td->td_retval[0] = ops->fetch_error(ujob);
2068 		return (0);
2069 	}
2070 
2071 	td->td_retval[0] = EINVAL;
2072 	return (0);
2073 }
2074 
2075 int
2076 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2077 {
2078 
2079 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2080 }
2081 
2082 /* syscall - asynchronous read from a file (REALTIME) */
2083 #ifdef COMPAT_FREEBSD6
2084 int
2085 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2086 {
2087 
2088 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2089 	    &aiocb_ops_osigevent));
2090 }
2091 #endif
2092 
2093 int
2094 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2095 {
2096 
2097 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2098 }
2099 
2100 /* syscall - asynchronous write to a file (REALTIME) */
2101 #ifdef COMPAT_FREEBSD6
2102 int
2103 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2104 {
2105 
2106 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2107 	    &aiocb_ops_osigevent));
2108 }
2109 #endif
2110 
2111 int
2112 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2113 {
2114 
2115 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2116 }
2117 
2118 int
2119 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2120 {
2121 
2122 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2123 }
2124 
2125 static int
2126 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2127     struct aiocb **acb_list, int nent, struct sigevent *sig,
2128     struct aiocb_ops *ops)
2129 {
2130 	struct proc *p = td->td_proc;
2131 	struct aiocb *job;
2132 	struct kaioinfo *ki;
2133 	struct aioliojob *lj;
2134 	struct kevent kev;
2135 	int error;
2136 	int nagain, nerror;
2137 	int i;
2138 
2139 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2140 		return (EINVAL);
2141 
2142 	if (nent < 0 || nent > max_aio_queue_per_proc)
2143 		return (EINVAL);
2144 
2145 	if (p->p_aioinfo == NULL)
2146 		aio_init_aioinfo(p);
2147 
2148 	ki = p->p_aioinfo;
2149 
2150 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2151 	lj->lioj_flags = 0;
2152 	lj->lioj_count = 0;
2153 	lj->lioj_finished_count = 0;
2154 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2155 	ksiginfo_init(&lj->lioj_ksi);
2156 
2157 	/*
2158 	 * Setup signal.
2159 	 */
2160 	if (sig && (mode == LIO_NOWAIT)) {
2161 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2162 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2163 			/* Assume only new style KEVENT */
2164 			memset(&kev, 0, sizeof(kev));
2165 			kev.filter = EVFILT_LIO;
2166 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2167 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2168 			kev.data = (intptr_t)lj;
2169 			/* pass user defined sigval data */
2170 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2171 			error = kqfd_register(
2172 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2173 			    M_WAITOK);
2174 			if (error) {
2175 				uma_zfree(aiolio_zone, lj);
2176 				return (error);
2177 			}
2178 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2179 			;
2180 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2181 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2182 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2183 					uma_zfree(aiolio_zone, lj);
2184 					return EINVAL;
2185 				}
2186 				lj->lioj_flags |= LIOJ_SIGNAL;
2187 		} else {
2188 			uma_zfree(aiolio_zone, lj);
2189 			return EINVAL;
2190 		}
2191 	}
2192 
2193 	AIO_LOCK(ki);
2194 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2195 	/*
2196 	 * Add extra aiocb count to avoid the lio to be freed
2197 	 * by other threads doing aio_waitcomplete or aio_return,
2198 	 * and prevent event from being sent until we have queued
2199 	 * all tasks.
2200 	 */
2201 	lj->lioj_count = 1;
2202 	AIO_UNLOCK(ki);
2203 
2204 	/*
2205 	 * Get pointers to the list of I/O requests.
2206 	 */
2207 	nagain = 0;
2208 	nerror = 0;
2209 	for (i = 0; i < nent; i++) {
2210 		job = acb_list[i];
2211 		if (job != NULL) {
2212 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2213 			if (error == EAGAIN)
2214 				nagain++;
2215 			else if (error != 0)
2216 				nerror++;
2217 		}
2218 	}
2219 
2220 	error = 0;
2221 	AIO_LOCK(ki);
2222 	if (mode == LIO_WAIT) {
2223 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2224 			ki->kaio_flags |= KAIO_WAKEUP;
2225 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2226 			    PRIBIO | PCATCH, "aiospn", 0);
2227 			if (error == ERESTART)
2228 				error = EINTR;
2229 			if (error)
2230 				break;
2231 		}
2232 	} else {
2233 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2234 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2235 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2236 				KNOTE_LOCKED(&lj->klist, 1);
2237 			}
2238 			if ((lj->lioj_flags & (LIOJ_SIGNAL |
2239 			    LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL &&
2240 			    (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2241 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2242 				aio_sendsig(p, &lj->lioj_signal, &lj->lioj_ksi,
2243 				    lj->lioj_count != 1);
2244 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2245 			}
2246 		}
2247 	}
2248 	lj->lioj_count--;
2249 	if (lj->lioj_count == 0) {
2250 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2251 		knlist_delete(&lj->klist, curthread, 1);
2252 		PROC_LOCK(p);
2253 		sigqueue_take(&lj->lioj_ksi);
2254 		PROC_UNLOCK(p);
2255 		AIO_UNLOCK(ki);
2256 		uma_zfree(aiolio_zone, lj);
2257 	} else
2258 		AIO_UNLOCK(ki);
2259 
2260 	if (nerror)
2261 		return (EIO);
2262 	else if (nagain)
2263 		return (EAGAIN);
2264 	else
2265 		return (error);
2266 }
2267 
2268 /* syscall - list directed I/O (REALTIME) */
2269 #ifdef COMPAT_FREEBSD6
2270 int
2271 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2272 {
2273 	struct aiocb **acb_list;
2274 	struct sigevent *sigp, sig;
2275 	struct osigevent osig;
2276 	int error, nent;
2277 
2278 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2279 		return (EINVAL);
2280 
2281 	nent = uap->nent;
2282 	if (nent < 0 || nent > max_aio_queue_per_proc)
2283 		return (EINVAL);
2284 
2285 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2286 		error = copyin(uap->sig, &osig, sizeof(osig));
2287 		if (error)
2288 			return (error);
2289 		error = convert_old_sigevent(&osig, &sig);
2290 		if (error)
2291 			return (error);
2292 		sigp = &sig;
2293 	} else
2294 		sigp = NULL;
2295 
2296 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2297 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2298 	if (error == 0)
2299 		error = kern_lio_listio(td, uap->mode,
2300 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2301 		    &aiocb_ops_osigevent);
2302 	free(acb_list, M_LIO);
2303 	return (error);
2304 }
2305 #endif
2306 
2307 /* syscall - list directed I/O (REALTIME) */
2308 int
2309 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2310 {
2311 	struct aiocb **acb_list;
2312 	struct sigevent *sigp, sig;
2313 	int error, nent;
2314 
2315 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2316 		return (EINVAL);
2317 
2318 	nent = uap->nent;
2319 	if (nent < 0 || nent > max_aio_queue_per_proc)
2320 		return (EINVAL);
2321 
2322 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2323 		error = copyin(uap->sig, &sig, sizeof(sig));
2324 		if (error)
2325 			return (error);
2326 		sigp = &sig;
2327 	} else
2328 		sigp = NULL;
2329 
2330 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2331 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2332 	if (error == 0)
2333 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2334 		    nent, sigp, &aiocb_ops);
2335 	free(acb_list, M_LIO);
2336 	return (error);
2337 }
2338 
2339 static void
2340 aio_biowakeup(struct bio *bp)
2341 {
2342 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2343 	struct kaioinfo *ki;
2344 	struct buf *pbuf = (struct buf*)bp->bio_caller2;
2345 	size_t nbytes;
2346 	int error, nblks;
2347 
2348 	/* Release mapping into kernel space. */
2349 	if (pbuf != NULL) {
2350 		MPASS(pbuf->b_npages <= atop(maxphys) + 1);
2351 		pmap_qremove((vm_offset_t)pbuf->b_data, pbuf->b_npages);
2352 		vm_page_unhold_pages(pbuf->b_pages, pbuf->b_npages);
2353 		uma_zfree(pbuf_zone, pbuf);
2354 		atomic_subtract_int(&num_buf_aio, 1);
2355 		ki = job->userproc->p_aioinfo;
2356 		AIO_LOCK(ki);
2357 		ki->kaio_buffer_count--;
2358 		AIO_UNLOCK(ki);
2359 	} else {
2360 		MPASS(bp->bio_ma_n <= atop(maxphys) + 1);
2361 		vm_page_unhold_pages(bp->bio_ma, bp->bio_ma_n);
2362 		free(bp->bio_ma, M_TEMP);
2363 		atomic_subtract_int(&num_unmapped_aio, 1);
2364 	}
2365 
2366 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2367 	error = 0;
2368 	if (bp->bio_flags & BIO_ERROR)
2369 		error = bp->bio_error;
2370 	nblks = btodb(nbytes);
2371 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2372 		job->outblock += nblks;
2373 	else
2374 		job->inblock += nblks;
2375 
2376 	if (error)
2377 		aio_complete(job, -1, error);
2378 	else
2379 		aio_complete(job, nbytes, 0);
2380 
2381 	g_destroy_bio(bp);
2382 }
2383 
2384 /* syscall - wait for the next completion of an aio request */
2385 static int
2386 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2387     struct timespec *ts, struct aiocb_ops *ops)
2388 {
2389 	struct proc *p = td->td_proc;
2390 	struct timeval atv;
2391 	struct kaioinfo *ki;
2392 	struct kaiocb *job;
2393 	struct aiocb *ujob;
2394 	long error, status;
2395 	int timo;
2396 
2397 	ops->store_aiocb(ujobp, NULL);
2398 
2399 	if (ts == NULL) {
2400 		timo = 0;
2401 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2402 		timo = -1;
2403 	} else {
2404 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2405 			return (EINVAL);
2406 
2407 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2408 		if (itimerfix(&atv))
2409 			return (EINVAL);
2410 		timo = tvtohz(&atv);
2411 	}
2412 
2413 	if (p->p_aioinfo == NULL)
2414 		aio_init_aioinfo(p);
2415 	ki = p->p_aioinfo;
2416 
2417 	error = 0;
2418 	job = NULL;
2419 	AIO_LOCK(ki);
2420 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2421 		if (timo == -1) {
2422 			error = EWOULDBLOCK;
2423 			break;
2424 		}
2425 		ki->kaio_flags |= KAIO_WAKEUP;
2426 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2427 		    "aiowc", timo);
2428 		if (timo && error == ERESTART)
2429 			error = EINTR;
2430 		if (error)
2431 			break;
2432 	}
2433 
2434 	if (job != NULL) {
2435 		MPASS(job->jobflags & KAIOCB_FINISHED);
2436 		ujob = job->ujob;
2437 		status = job->uaiocb._aiocb_private.status;
2438 		error = job->uaiocb._aiocb_private.error;
2439 		td->td_retval[0] = status;
2440 		td->td_ru.ru_oublock += job->outblock;
2441 		td->td_ru.ru_inblock += job->inblock;
2442 		td->td_ru.ru_msgsnd += job->msgsnd;
2443 		td->td_ru.ru_msgrcv += job->msgrcv;
2444 		aio_free_entry(job);
2445 		AIO_UNLOCK(ki);
2446 		ops->store_aiocb(ujobp, ujob);
2447 		ops->store_error(ujob, error);
2448 		ops->store_status(ujob, status);
2449 	} else
2450 		AIO_UNLOCK(ki);
2451 
2452 	return (error);
2453 }
2454 
2455 int
2456 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2457 {
2458 	struct timespec ts, *tsp;
2459 	int error;
2460 
2461 	if (uap->timeout) {
2462 		/* Get timespec struct. */
2463 		error = copyin(uap->timeout, &ts, sizeof(ts));
2464 		if (error)
2465 			return (error);
2466 		tsp = &ts;
2467 	} else
2468 		tsp = NULL;
2469 
2470 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2471 }
2472 
2473 static int
2474 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2475     struct aiocb_ops *ops)
2476 {
2477 
2478 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2479 		return (EINVAL);
2480 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2481 }
2482 
2483 int
2484 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2485 {
2486 
2487 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2488 }
2489 
2490 /* kqueue attach function */
2491 static int
2492 filt_aioattach(struct knote *kn)
2493 {
2494 	struct kaiocb *job;
2495 
2496 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2497 
2498 	/*
2499 	 * The job pointer must be validated before using it, so
2500 	 * registration is restricted to the kernel; the user cannot
2501 	 * set EV_FLAG1.
2502 	 */
2503 	if ((kn->kn_flags & EV_FLAG1) == 0)
2504 		return (EPERM);
2505 	kn->kn_ptr.p_aio = job;
2506 	kn->kn_flags &= ~EV_FLAG1;
2507 
2508 	knlist_add(&job->klist, kn, 0);
2509 
2510 	return (0);
2511 }
2512 
2513 /* kqueue detach function */
2514 static void
2515 filt_aiodetach(struct knote *kn)
2516 {
2517 	struct knlist *knl;
2518 
2519 	knl = &kn->kn_ptr.p_aio->klist;
2520 	knl->kl_lock(knl->kl_lockarg);
2521 	if (!knlist_empty(knl))
2522 		knlist_remove(knl, kn, 1);
2523 	knl->kl_unlock(knl->kl_lockarg);
2524 }
2525 
2526 /* kqueue filter function */
2527 /*ARGSUSED*/
2528 static int
2529 filt_aio(struct knote *kn, long hint)
2530 {
2531 	struct kaiocb *job = kn->kn_ptr.p_aio;
2532 
2533 	kn->kn_data = job->uaiocb._aiocb_private.error;
2534 	if (!(job->jobflags & KAIOCB_FINISHED))
2535 		return (0);
2536 	kn->kn_flags |= EV_EOF;
2537 	return (1);
2538 }
2539 
2540 /* kqueue attach function */
2541 static int
2542 filt_lioattach(struct knote *kn)
2543 {
2544 	struct aioliojob *lj;
2545 
2546 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2547 
2548 	/*
2549 	 * The aioliojob pointer must be validated before using it, so
2550 	 * registration is restricted to the kernel; the user cannot
2551 	 * set EV_FLAG1.
2552 	 */
2553 	if ((kn->kn_flags & EV_FLAG1) == 0)
2554 		return (EPERM);
2555 	kn->kn_ptr.p_lio = lj;
2556 	kn->kn_flags &= ~EV_FLAG1;
2557 
2558 	knlist_add(&lj->klist, kn, 0);
2559 
2560 	return (0);
2561 }
2562 
2563 /* kqueue detach function */
2564 static void
2565 filt_liodetach(struct knote *kn)
2566 {
2567 	struct knlist *knl;
2568 
2569 	knl = &kn->kn_ptr.p_lio->klist;
2570 	knl->kl_lock(knl->kl_lockarg);
2571 	if (!knlist_empty(knl))
2572 		knlist_remove(knl, kn, 1);
2573 	knl->kl_unlock(knl->kl_lockarg);
2574 }
2575 
2576 /* kqueue filter function */
2577 /*ARGSUSED*/
2578 static int
2579 filt_lio(struct knote *kn, long hint)
2580 {
2581 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2582 
2583 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2584 }
2585 
2586 #ifdef COMPAT_FREEBSD32
2587 #include <sys/mount.h>
2588 #include <sys/socket.h>
2589 #include <compat/freebsd32/freebsd32.h>
2590 #include <compat/freebsd32/freebsd32_proto.h>
2591 #include <compat/freebsd32/freebsd32_signal.h>
2592 #include <compat/freebsd32/freebsd32_syscall.h>
2593 #include <compat/freebsd32/freebsd32_util.h>
2594 
2595 struct __aiocb_private32 {
2596 	int32_t	status;
2597 	int32_t	error;
2598 	uint32_t kernelinfo;
2599 };
2600 
2601 #ifdef COMPAT_FREEBSD6
2602 typedef struct oaiocb32 {
2603 	int	aio_fildes;		/* File descriptor */
2604 	uint64_t aio_offset __packed;	/* File offset for I/O */
2605 	uint32_t aio_buf;		/* I/O buffer in process space */
2606 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2607 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2608 	int	aio_lio_opcode;		/* LIO opcode */
2609 	int	aio_reqprio;		/* Request priority -- ignored */
2610 	struct	__aiocb_private32 _aiocb_private;
2611 } oaiocb32_t;
2612 #endif
2613 
2614 typedef struct aiocb32 {
2615 	int32_t	aio_fildes;		/* File descriptor */
2616 	uint64_t aio_offset __packed;	/* File offset for I/O */
2617 	uint32_t aio_buf;		/* I/O buffer in process space */
2618 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2619 	int	__spare__[2];
2620 	uint32_t __spare2__;
2621 	int	aio_lio_opcode;		/* LIO opcode */
2622 	int	aio_reqprio;		/* Request priority -- ignored */
2623 	struct	__aiocb_private32 _aiocb_private;
2624 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2625 } aiocb32_t;
2626 
2627 #ifdef COMPAT_FREEBSD6
2628 static int
2629 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2630 {
2631 
2632 	/*
2633 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2634 	 * supported by AIO with the old sigevent structure.
2635 	 */
2636 	CP(*osig, *nsig, sigev_notify);
2637 	switch (nsig->sigev_notify) {
2638 	case SIGEV_NONE:
2639 		break;
2640 	case SIGEV_SIGNAL:
2641 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2642 		break;
2643 	case SIGEV_KEVENT:
2644 		nsig->sigev_notify_kqueue =
2645 		    osig->__sigev_u.__sigev_notify_kqueue;
2646 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2647 		break;
2648 	default:
2649 		return (EINVAL);
2650 	}
2651 	return (0);
2652 }
2653 
2654 static int
2655 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2656 {
2657 	struct oaiocb32 job32;
2658 	int error;
2659 
2660 	bzero(kjob, sizeof(struct aiocb));
2661 	error = copyin(ujob, &job32, sizeof(job32));
2662 	if (error)
2663 		return (error);
2664 
2665 	CP(job32, *kjob, aio_fildes);
2666 	CP(job32, *kjob, aio_offset);
2667 	PTRIN_CP(job32, *kjob, aio_buf);
2668 	CP(job32, *kjob, aio_nbytes);
2669 	CP(job32, *kjob, aio_lio_opcode);
2670 	CP(job32, *kjob, aio_reqprio);
2671 	CP(job32, *kjob, _aiocb_private.status);
2672 	CP(job32, *kjob, _aiocb_private.error);
2673 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2674 	return (convert_old_sigevent32(&job32.aio_sigevent,
2675 	    &kjob->aio_sigevent));
2676 }
2677 #endif
2678 
2679 static int
2680 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2681 {
2682 	struct aiocb32 job32;
2683 	int error;
2684 
2685 	error = copyin(ujob, &job32, sizeof(job32));
2686 	if (error)
2687 		return (error);
2688 	CP(job32, *kjob, aio_fildes);
2689 	CP(job32, *kjob, aio_offset);
2690 	PTRIN_CP(job32, *kjob, aio_buf);
2691 	CP(job32, *kjob, aio_nbytes);
2692 	CP(job32, *kjob, aio_lio_opcode);
2693 	CP(job32, *kjob, aio_reqprio);
2694 	CP(job32, *kjob, _aiocb_private.status);
2695 	CP(job32, *kjob, _aiocb_private.error);
2696 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2697 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2698 }
2699 
2700 static long
2701 aiocb32_fetch_status(struct aiocb *ujob)
2702 {
2703 	struct aiocb32 *ujob32;
2704 
2705 	ujob32 = (struct aiocb32 *)ujob;
2706 	return (fuword32(&ujob32->_aiocb_private.status));
2707 }
2708 
2709 static long
2710 aiocb32_fetch_error(struct aiocb *ujob)
2711 {
2712 	struct aiocb32 *ujob32;
2713 
2714 	ujob32 = (struct aiocb32 *)ujob;
2715 	return (fuword32(&ujob32->_aiocb_private.error));
2716 }
2717 
2718 static int
2719 aiocb32_store_status(struct aiocb *ujob, long status)
2720 {
2721 	struct aiocb32 *ujob32;
2722 
2723 	ujob32 = (struct aiocb32 *)ujob;
2724 	return (suword32(&ujob32->_aiocb_private.status, status));
2725 }
2726 
2727 static int
2728 aiocb32_store_error(struct aiocb *ujob, long error)
2729 {
2730 	struct aiocb32 *ujob32;
2731 
2732 	ujob32 = (struct aiocb32 *)ujob;
2733 	return (suword32(&ujob32->_aiocb_private.error, error));
2734 }
2735 
2736 static int
2737 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2738 {
2739 	struct aiocb32 *ujob32;
2740 
2741 	ujob32 = (struct aiocb32 *)ujob;
2742 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2743 }
2744 
2745 static int
2746 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2747 {
2748 
2749 	return (suword32(ujobp, (long)ujob));
2750 }
2751 
2752 static struct aiocb_ops aiocb32_ops = {
2753 	.aio_copyin = aiocb32_copyin,
2754 	.fetch_status = aiocb32_fetch_status,
2755 	.fetch_error = aiocb32_fetch_error,
2756 	.store_status = aiocb32_store_status,
2757 	.store_error = aiocb32_store_error,
2758 	.store_kernelinfo = aiocb32_store_kernelinfo,
2759 	.store_aiocb = aiocb32_store_aiocb,
2760 };
2761 
2762 #ifdef COMPAT_FREEBSD6
2763 static struct aiocb_ops aiocb32_ops_osigevent = {
2764 	.aio_copyin = aiocb32_copyin_old_sigevent,
2765 	.fetch_status = aiocb32_fetch_status,
2766 	.fetch_error = aiocb32_fetch_error,
2767 	.store_status = aiocb32_store_status,
2768 	.store_error = aiocb32_store_error,
2769 	.store_kernelinfo = aiocb32_store_kernelinfo,
2770 	.store_aiocb = aiocb32_store_aiocb,
2771 };
2772 #endif
2773 
2774 int
2775 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2776 {
2777 
2778 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2779 }
2780 
2781 int
2782 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2783 {
2784 	struct timespec32 ts32;
2785 	struct timespec ts, *tsp;
2786 	struct aiocb **ujoblist;
2787 	uint32_t *ujoblist32;
2788 	int error, i;
2789 
2790 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2791 		return (EINVAL);
2792 
2793 	if (uap->timeout) {
2794 		/* Get timespec struct. */
2795 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2796 			return (error);
2797 		CP(ts32, ts, tv_sec);
2798 		CP(ts32, ts, tv_nsec);
2799 		tsp = &ts;
2800 	} else
2801 		tsp = NULL;
2802 
2803 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2804 	ujoblist32 = (uint32_t *)ujoblist;
2805 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2806 	    sizeof(ujoblist32[0]));
2807 	if (error == 0) {
2808 		for (i = uap->nent - 1; i >= 0; i--)
2809 			ujoblist[i] = PTRIN(ujoblist32[i]);
2810 
2811 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2812 	}
2813 	free(ujoblist, M_AIOS);
2814 	return (error);
2815 }
2816 
2817 int
2818 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2819 {
2820 
2821 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2822 }
2823 
2824 #ifdef COMPAT_FREEBSD6
2825 int
2826 freebsd6_freebsd32_aio_read(struct thread *td,
2827     struct freebsd6_freebsd32_aio_read_args *uap)
2828 {
2829 
2830 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2831 	    &aiocb32_ops_osigevent));
2832 }
2833 #endif
2834 
2835 int
2836 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2837 {
2838 
2839 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2840 	    &aiocb32_ops));
2841 }
2842 
2843 #ifdef COMPAT_FREEBSD6
2844 int
2845 freebsd6_freebsd32_aio_write(struct thread *td,
2846     struct freebsd6_freebsd32_aio_write_args *uap)
2847 {
2848 
2849 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2850 	    &aiocb32_ops_osigevent));
2851 }
2852 #endif
2853 
2854 int
2855 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2856 {
2857 
2858 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2859 	    &aiocb32_ops));
2860 }
2861 
2862 int
2863 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2864 {
2865 
2866 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2867 	    &aiocb32_ops));
2868 }
2869 
2870 int
2871 freebsd32_aio_waitcomplete(struct thread *td,
2872     struct freebsd32_aio_waitcomplete_args *uap)
2873 {
2874 	struct timespec32 ts32;
2875 	struct timespec ts, *tsp;
2876 	int error;
2877 
2878 	if (uap->timeout) {
2879 		/* Get timespec struct. */
2880 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2881 		if (error)
2882 			return (error);
2883 		CP(ts32, ts, tv_sec);
2884 		CP(ts32, ts, tv_nsec);
2885 		tsp = &ts;
2886 	} else
2887 		tsp = NULL;
2888 
2889 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2890 	    &aiocb32_ops));
2891 }
2892 
2893 int
2894 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2895 {
2896 
2897 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2898 	    &aiocb32_ops));
2899 }
2900 
2901 #ifdef COMPAT_FREEBSD6
2902 int
2903 freebsd6_freebsd32_lio_listio(struct thread *td,
2904     struct freebsd6_freebsd32_lio_listio_args *uap)
2905 {
2906 	struct aiocb **acb_list;
2907 	struct sigevent *sigp, sig;
2908 	struct osigevent32 osig;
2909 	uint32_t *acb_list32;
2910 	int error, i, nent;
2911 
2912 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2913 		return (EINVAL);
2914 
2915 	nent = uap->nent;
2916 	if (nent < 0 || nent > max_aio_queue_per_proc)
2917 		return (EINVAL);
2918 
2919 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2920 		error = copyin(uap->sig, &osig, sizeof(osig));
2921 		if (error)
2922 			return (error);
2923 		error = convert_old_sigevent32(&osig, &sig);
2924 		if (error)
2925 			return (error);
2926 		sigp = &sig;
2927 	} else
2928 		sigp = NULL;
2929 
2930 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2931 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2932 	if (error) {
2933 		free(acb_list32, M_LIO);
2934 		return (error);
2935 	}
2936 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2937 	for (i = 0; i < nent; i++)
2938 		acb_list[i] = PTRIN(acb_list32[i]);
2939 	free(acb_list32, M_LIO);
2940 
2941 	error = kern_lio_listio(td, uap->mode,
2942 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2943 	    &aiocb32_ops_osigevent);
2944 	free(acb_list, M_LIO);
2945 	return (error);
2946 }
2947 #endif
2948 
2949 int
2950 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2951 {
2952 	struct aiocb **acb_list;
2953 	struct sigevent *sigp, sig;
2954 	struct sigevent32 sig32;
2955 	uint32_t *acb_list32;
2956 	int error, i, nent;
2957 
2958 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2959 		return (EINVAL);
2960 
2961 	nent = uap->nent;
2962 	if (nent < 0 || nent > max_aio_queue_per_proc)
2963 		return (EINVAL);
2964 
2965 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2966 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2967 		if (error)
2968 			return (error);
2969 		error = convert_sigevent32(&sig32, &sig);
2970 		if (error)
2971 			return (error);
2972 		sigp = &sig;
2973 	} else
2974 		sigp = NULL;
2975 
2976 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2977 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2978 	if (error) {
2979 		free(acb_list32, M_LIO);
2980 		return (error);
2981 	}
2982 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2983 	for (i = 0; i < nent; i++)
2984 		acb_list[i] = PTRIN(acb_list32[i]);
2985 	free(acb_list32, M_LIO);
2986 
2987 	error = kern_lio_listio(td, uap->mode,
2988 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2989 	    &aiocb32_ops);
2990 	free(acb_list, M_LIO);
2991 	return (error);
2992 }
2993 
2994 #endif
2995