xref: /freebsd/sys/kern/vfs_aio.c (revision b9a53e161b94f4eefee30b2c56152a5a8d82cea4)
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/syslog.h>
57 #include <sys/sx.h>
58 #include <sys/taskqueue.h>
59 #include <sys/vnode.h>
60 #include <sys/conf.h>
61 #include <sys/event.h>
62 #include <sys/mount.h>
63 #include <geom/geom.h>
64 
65 #include <machine/atomic.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_extern.h>
70 #include <vm/pmap.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_object.h>
73 #include <vm/uma.h>
74 #include <sys/aio.h>
75 
76 /*
77  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
78  * overflow. (XXX will be removed soon.)
79  */
80 static u_long jobrefid;
81 
82 /*
83  * Counter for aio_fsync.
84  */
85 static uint64_t jobseqno;
86 
87 #ifndef MAX_AIO_PER_PROC
88 #define MAX_AIO_PER_PROC	32
89 #endif
90 
91 #ifndef MAX_AIO_QUEUE_PER_PROC
92 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
93 #endif
94 
95 #ifndef MAX_AIO_QUEUE
96 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
97 #endif
98 
99 #ifndef MAX_BUF_AIO
100 #define MAX_BUF_AIO		16
101 #endif
102 
103 FEATURE(aio, "Asynchronous I/O");
104 
105 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
106 
107 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
108     "Async IO management");
109 
110 static int enable_aio_unsafe = 0;
111 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
112     "Permit asynchronous IO on all file types, not just known-safe types");
113 
114 static unsigned int unsafe_warningcnt = 1;
115 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
116     &unsafe_warningcnt, 0,
117     "Warnings that will be triggered upon failed IO requests on unsafe files");
118 
119 static int max_aio_procs = MAX_AIO_PROCS;
120 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
121     "Maximum number of kernel processes to use for handling async IO ");
122 
123 static int num_aio_procs = 0;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
125     "Number of presently active kernel processes for async IO");
126 
127 /*
128  * The code will adjust the actual number of AIO processes towards this
129  * number when it gets a chance.
130  */
131 static int target_aio_procs = TARGET_AIO_PROCS;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
133     0,
134     "Preferred number of ready kernel processes for async IO");
135 
136 static int max_queue_count = MAX_AIO_QUEUE;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
138     "Maximum number of aio requests to queue, globally");
139 
140 static int num_queue_count = 0;
141 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
142     "Number of queued aio requests");
143 
144 static int num_buf_aio = 0;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
146     "Number of aio requests presently handled by the buf subsystem");
147 
148 /* Number of async I/O processes in the process of being started */
149 /* XXX This should be local to aio_aqueue() */
150 static int num_aio_resv_start = 0;
151 
152 static int aiod_lifetime;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
154     "Maximum lifetime for idle aiod");
155 
156 static int max_aio_per_proc = MAX_AIO_PER_PROC;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
158     0,
159     "Maximum active aio requests per process (stored in the process)");
160 
161 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
163     &max_aio_queue_per_proc, 0,
164     "Maximum queued aio requests per process (stored in the process)");
165 
166 static int max_buf_aio = MAX_BUF_AIO;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
168     "Maximum buf aio requests per process (stored in the process)");
169 
170 #ifdef COMPAT_FREEBSD6
171 typedef struct oaiocb {
172 	int	aio_fildes;		/* File descriptor */
173 	off_t	aio_offset;		/* File offset for I/O */
174 	volatile void *aio_buf;         /* I/O buffer in process space */
175 	size_t	aio_nbytes;		/* Number of bytes for I/O */
176 	struct	osigevent aio_sigevent;	/* Signal to deliver */
177 	int	aio_lio_opcode;		/* LIO opcode */
178 	int	aio_reqprio;		/* Request priority -- ignored */
179 	struct	__aiocb_private	_aiocb_private;
180 } oaiocb_t;
181 #endif
182 
183 /*
184  * Below is a key of locks used to protect each member of struct kaiocb
185  * aioliojob and kaioinfo and any backends.
186  *
187  * * - need not protected
188  * a - locked by kaioinfo lock
189  * b - locked by backend lock, the backend lock can be null in some cases,
190  *     for example, BIO belongs to this type, in this case, proc lock is
191  *     reused.
192  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
193  */
194 
195 /*
196  * If the routine that services an AIO request blocks while running in an
197  * AIO kernel process it can starve other I/O requests.  BIO requests
198  * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
199  * processes at all.  Socket I/O requests use a separate pool of
200  * kprocs and also force non-blocking I/O.  Other file I/O requests
201  * use the generic fo_read/fo_write operations which can block.  The
202  * fsync and mlock operations can also block while executing.  Ideally
203  * none of these requests would block while executing.
204  *
205  * Note that the service routines cannot toggle O_NONBLOCK in the file
206  * structure directly while handling a request due to races with
207  * userland threads.
208  */
209 
210 /* jobflags */
211 #define	KAIOCB_QUEUEING		0x01
212 #define	KAIOCB_CANCELLED	0x02
213 #define	KAIOCB_CANCELLING	0x04
214 #define	KAIOCB_CHECKSYNC	0x08
215 #define	KAIOCB_CLEARED		0x10
216 #define	KAIOCB_FINISHED		0x20
217 
218 /*
219  * AIO process info
220  */
221 #define AIOP_FREE	0x1			/* proc on free queue */
222 
223 struct aioproc {
224 	int	aioprocflags;			/* (c) AIO proc flags */
225 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
226 	struct	proc *aioproc;			/* (*) the AIO proc */
227 };
228 
229 /*
230  * data-structure for lio signal management
231  */
232 struct aioliojob {
233 	int	lioj_flags;			/* (a) listio flags */
234 	int	lioj_count;			/* (a) listio flags */
235 	int	lioj_finished_count;		/* (a) listio flags */
236 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
237 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
238 	struct	knlist klist;			/* (a) list of knotes */
239 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
240 };
241 
242 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
243 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
244 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
245 
246 /*
247  * per process aio data structure
248  */
249 struct kaioinfo {
250 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
251 	int	kaio_flags;		/* (a) per process kaio flags */
252 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
253 	int	kaio_active_count;	/* (c) number of currently used AIOs */
254 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
255 	int	kaio_count;		/* (a) size of AIO queue */
256 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
257 	int	kaio_buffer_count;	/* (a) number of physio buffers */
258 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
259 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
260 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
261 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
262 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
263 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
264 	struct	task kaio_task;		/* (*) task to kick aio processes */
265 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
266 };
267 
268 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
269 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
270 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
271 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
272 
273 #define KAIO_RUNDOWN	0x1	/* process is being run down */
274 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
275 
276 /*
277  * Operations used to interact with userland aio control blocks.
278  * Different ABIs provide their own operations.
279  */
280 struct aiocb_ops {
281 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
282 	long	(*fetch_status)(struct aiocb *ujob);
283 	long	(*fetch_error)(struct aiocb *ujob);
284 	int	(*store_status)(struct aiocb *ujob, long status);
285 	int	(*store_error)(struct aiocb *ujob, long error);
286 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
287 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
288 };
289 
290 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
291 static struct sema aio_newproc_sem;
292 static struct mtx aio_job_mtx;
293 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
294 static struct unrhdr *aiod_unr;
295 
296 void		aio_init_aioinfo(struct proc *p);
297 static int	aio_onceonly(void);
298 static int	aio_free_entry(struct kaiocb *job);
299 static void	aio_process_rw(struct kaiocb *job);
300 static void	aio_process_sync(struct kaiocb *job);
301 static void	aio_process_mlock(struct kaiocb *job);
302 static void	aio_schedule_fsync(void *context, int pending);
303 static int	aio_newproc(int *);
304 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
305 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
306 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
307 static void	aio_physwakeup(struct bio *bp);
308 static void	aio_proc_rundown(void *arg, struct proc *p);
309 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
310 		    struct image_params *imgp);
311 static int	aio_qphysio(struct proc *p, struct kaiocb *job);
312 static void	aio_daemon(void *param);
313 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
314 static int	aio_kick(struct proc *userp);
315 static void	aio_kick_nowait(struct proc *userp);
316 static void	aio_kick_helper(void *context, int pending);
317 static int	filt_aioattach(struct knote *kn);
318 static void	filt_aiodetach(struct knote *kn);
319 static int	filt_aio(struct knote *kn, long hint);
320 static int	filt_lioattach(struct knote *kn);
321 static void	filt_liodetach(struct knote *kn);
322 static int	filt_lio(struct knote *kn, long hint);
323 
324 /*
325  * Zones for:
326  * 	kaio	Per process async io info
327  *	aiop	async io process data
328  *	aiocb	async io jobs
329  *	aiol	list io job pointer - internal to aio_suspend XXX
330  *	aiolio	list io jobs
331  */
332 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
333 
334 /* kqueue filters for aio */
335 static struct filterops aio_filtops = {
336 	.f_isfd = 0,
337 	.f_attach = filt_aioattach,
338 	.f_detach = filt_aiodetach,
339 	.f_event = filt_aio,
340 };
341 static struct filterops lio_filtops = {
342 	.f_isfd = 0,
343 	.f_attach = filt_lioattach,
344 	.f_detach = filt_liodetach,
345 	.f_event = filt_lio
346 };
347 
348 static eventhandler_tag exit_tag, exec_tag;
349 
350 TASKQUEUE_DEFINE_THREAD(aiod_kick);
351 
352 /*
353  * Main operations function for use as a kernel module.
354  */
355 static int
356 aio_modload(struct module *module, int cmd, void *arg)
357 {
358 	int error = 0;
359 
360 	switch (cmd) {
361 	case MOD_LOAD:
362 		aio_onceonly();
363 		break;
364 	case MOD_SHUTDOWN:
365 		break;
366 	default:
367 		error = EOPNOTSUPP;
368 		break;
369 	}
370 	return (error);
371 }
372 
373 static moduledata_t aio_mod = {
374 	"aio",
375 	&aio_modload,
376 	NULL
377 };
378 
379 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
380 MODULE_VERSION(aio, 1);
381 
382 /*
383  * Startup initialization
384  */
385 static int
386 aio_onceonly(void)
387 {
388 
389 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
390 	    EVENTHANDLER_PRI_ANY);
391 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
392 	    NULL, EVENTHANDLER_PRI_ANY);
393 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
394 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
395 	TAILQ_INIT(&aio_freeproc);
396 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
397 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
398 	TAILQ_INIT(&aio_jobs);
399 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
400 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
401 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
402 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
403 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
404 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
405 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
406 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
407 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
408 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
409 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
410 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
411 	jobrefid = 1;
412 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
413 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
414 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
415 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
416 
417 	return (0);
418 }
419 
420 /*
421  * Init the per-process aioinfo structure.  The aioinfo limits are set
422  * per-process for user limit (resource) management.
423  */
424 void
425 aio_init_aioinfo(struct proc *p)
426 {
427 	struct kaioinfo *ki;
428 
429 	ki = uma_zalloc(kaio_zone, M_WAITOK);
430 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
431 	ki->kaio_flags = 0;
432 	ki->kaio_maxactive_count = max_aio_per_proc;
433 	ki->kaio_active_count = 0;
434 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
435 	ki->kaio_count = 0;
436 	ki->kaio_ballowed_count = max_buf_aio;
437 	ki->kaio_buffer_count = 0;
438 	TAILQ_INIT(&ki->kaio_all);
439 	TAILQ_INIT(&ki->kaio_done);
440 	TAILQ_INIT(&ki->kaio_jobqueue);
441 	TAILQ_INIT(&ki->kaio_liojoblist);
442 	TAILQ_INIT(&ki->kaio_syncqueue);
443 	TAILQ_INIT(&ki->kaio_syncready);
444 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
445 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
446 	PROC_LOCK(p);
447 	if (p->p_aioinfo == NULL) {
448 		p->p_aioinfo = ki;
449 		PROC_UNLOCK(p);
450 	} else {
451 		PROC_UNLOCK(p);
452 		mtx_destroy(&ki->kaio_mtx);
453 		uma_zfree(kaio_zone, ki);
454 	}
455 
456 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
457 		aio_newproc(NULL);
458 }
459 
460 static int
461 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
462 {
463 	struct thread *td;
464 	int error;
465 
466 	error = sigev_findtd(p, sigev, &td);
467 	if (error)
468 		return (error);
469 	if (!KSI_ONQ(ksi)) {
470 		ksiginfo_set_sigev(ksi, sigev);
471 		ksi->ksi_code = SI_ASYNCIO;
472 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
473 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
474 	}
475 	PROC_UNLOCK(p);
476 	return (error);
477 }
478 
479 /*
480  * Free a job entry.  Wait for completion if it is currently active, but don't
481  * delay forever.  If we delay, we return a flag that says that we have to
482  * restart the queue scan.
483  */
484 static int
485 aio_free_entry(struct kaiocb *job)
486 {
487 	struct kaioinfo *ki;
488 	struct aioliojob *lj;
489 	struct proc *p;
490 
491 	p = job->userproc;
492 	MPASS(curproc == p);
493 	ki = p->p_aioinfo;
494 	MPASS(ki != NULL);
495 
496 	AIO_LOCK_ASSERT(ki, MA_OWNED);
497 	MPASS(job->jobflags & KAIOCB_FINISHED);
498 
499 	atomic_subtract_int(&num_queue_count, 1);
500 
501 	ki->kaio_count--;
502 	MPASS(ki->kaio_count >= 0);
503 
504 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
505 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
506 
507 	lj = job->lio;
508 	if (lj) {
509 		lj->lioj_count--;
510 		lj->lioj_finished_count--;
511 
512 		if (lj->lioj_count == 0) {
513 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
514 			/* lio is going away, we need to destroy any knotes */
515 			knlist_delete(&lj->klist, curthread, 1);
516 			PROC_LOCK(p);
517 			sigqueue_take(&lj->lioj_ksi);
518 			PROC_UNLOCK(p);
519 			uma_zfree(aiolio_zone, lj);
520 		}
521 	}
522 
523 	/* job is going away, we need to destroy any knotes */
524 	knlist_delete(&job->klist, curthread, 1);
525 	PROC_LOCK(p);
526 	sigqueue_take(&job->ksi);
527 	PROC_UNLOCK(p);
528 
529 	AIO_UNLOCK(ki);
530 
531 	/*
532 	 * The thread argument here is used to find the owning process
533 	 * and is also passed to fo_close() which may pass it to various
534 	 * places such as devsw close() routines.  Because of that, we
535 	 * need a thread pointer from the process owning the job that is
536 	 * persistent and won't disappear out from under us or move to
537 	 * another process.
538 	 *
539 	 * Currently, all the callers of this function call it to remove
540 	 * a kaiocb from the current process' job list either via a
541 	 * syscall or due to the current process calling exit() or
542 	 * execve().  Thus, we know that p == curproc.  We also know that
543 	 * curthread can't exit since we are curthread.
544 	 *
545 	 * Therefore, we use curthread as the thread to pass to
546 	 * knlist_delete().  This does mean that it is possible for the
547 	 * thread pointer at close time to differ from the thread pointer
548 	 * at open time, but this is already true of file descriptors in
549 	 * a multithreaded process.
550 	 */
551 	if (job->fd_file)
552 		fdrop(job->fd_file, curthread);
553 	crfree(job->cred);
554 	uma_zfree(aiocb_zone, job);
555 	AIO_LOCK(ki);
556 
557 	return (0);
558 }
559 
560 static void
561 aio_proc_rundown_exec(void *arg, struct proc *p,
562     struct image_params *imgp __unused)
563 {
564    	aio_proc_rundown(arg, p);
565 }
566 
567 static int
568 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
569 {
570 	aio_cancel_fn_t *func;
571 	int cancelled;
572 
573 	AIO_LOCK_ASSERT(ki, MA_OWNED);
574 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
575 		return (0);
576 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
577 	job->jobflags |= KAIOCB_CANCELLED;
578 
579 	func = job->cancel_fn;
580 
581 	/*
582 	 * If there is no cancel routine, just leave the job marked as
583 	 * cancelled.  The job should be in active use by a caller who
584 	 * should complete it normally or when it fails to install a
585 	 * cancel routine.
586 	 */
587 	if (func == NULL)
588 		return (0);
589 
590 	/*
591 	 * Set the CANCELLING flag so that aio_complete() will defer
592 	 * completions of this job.  This prevents the job from being
593 	 * freed out from under the cancel callback.  After the
594 	 * callback any deferred completion (whether from the callback
595 	 * or any other source) will be completed.
596 	 */
597 	job->jobflags |= KAIOCB_CANCELLING;
598 	AIO_UNLOCK(ki);
599 	func(job);
600 	AIO_LOCK(ki);
601 	job->jobflags &= ~KAIOCB_CANCELLING;
602 	if (job->jobflags & KAIOCB_FINISHED) {
603 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
604 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
605 		aio_bio_done_notify(p, job);
606 	} else {
607 		/*
608 		 * The cancel callback might have scheduled an
609 		 * operation to cancel this request, but it is
610 		 * only counted as cancelled if the request is
611 		 * cancelled when the callback returns.
612 		 */
613 		cancelled = 0;
614 	}
615 	return (cancelled);
616 }
617 
618 /*
619  * Rundown the jobs for a given process.
620  */
621 static void
622 aio_proc_rundown(void *arg, struct proc *p)
623 {
624 	struct kaioinfo *ki;
625 	struct aioliojob *lj;
626 	struct kaiocb *job, *jobn;
627 
628 	KASSERT(curthread->td_proc == p,
629 	    ("%s: called on non-curproc", __func__));
630 	ki = p->p_aioinfo;
631 	if (ki == NULL)
632 		return;
633 
634 	AIO_LOCK(ki);
635 	ki->kaio_flags |= KAIO_RUNDOWN;
636 
637 restart:
638 
639 	/*
640 	 * Try to cancel all pending requests. This code simulates
641 	 * aio_cancel on all pending I/O requests.
642 	 */
643 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
644 		aio_cancel_job(p, ki, job);
645 	}
646 
647 	/* Wait for all running I/O to be finished */
648 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
649 		ki->kaio_flags |= KAIO_WAKEUP;
650 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
651 		goto restart;
652 	}
653 
654 	/* Free all completed I/O requests. */
655 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
656 		aio_free_entry(job);
657 
658 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
659 		if (lj->lioj_count == 0) {
660 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
661 			knlist_delete(&lj->klist, curthread, 1);
662 			PROC_LOCK(p);
663 			sigqueue_take(&lj->lioj_ksi);
664 			PROC_UNLOCK(p);
665 			uma_zfree(aiolio_zone, lj);
666 		} else {
667 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
668 			    lj->lioj_count, lj->lioj_finished_count);
669 		}
670 	}
671 	AIO_UNLOCK(ki);
672 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
673 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
674 	mtx_destroy(&ki->kaio_mtx);
675 	uma_zfree(kaio_zone, ki);
676 	p->p_aioinfo = NULL;
677 }
678 
679 /*
680  * Select a job to run (called by an AIO daemon).
681  */
682 static struct kaiocb *
683 aio_selectjob(struct aioproc *aiop)
684 {
685 	struct kaiocb *job;
686 	struct kaioinfo *ki;
687 	struct proc *userp;
688 
689 	mtx_assert(&aio_job_mtx, MA_OWNED);
690 restart:
691 	TAILQ_FOREACH(job, &aio_jobs, list) {
692 		userp = job->userproc;
693 		ki = userp->p_aioinfo;
694 
695 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
696 			TAILQ_REMOVE(&aio_jobs, job, list);
697 			if (!aio_clear_cancel_function(job))
698 				goto restart;
699 
700 			/* Account for currently active jobs. */
701 			ki->kaio_active_count++;
702 			break;
703 		}
704 	}
705 	return (job);
706 }
707 
708 /*
709  * Move all data to a permanent storage device.  This code
710  * simulates the fsync syscall.
711  */
712 static int
713 aio_fsync_vnode(struct thread *td, struct vnode *vp)
714 {
715 	struct mount *mp;
716 	int error;
717 
718 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
719 		goto drop;
720 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
721 	if (vp->v_object != NULL) {
722 		VM_OBJECT_WLOCK(vp->v_object);
723 		vm_object_page_clean(vp->v_object, 0, 0, 0);
724 		VM_OBJECT_WUNLOCK(vp->v_object);
725 	}
726 	error = VOP_FSYNC(vp, MNT_WAIT, td);
727 
728 	VOP_UNLOCK(vp, 0);
729 	vn_finished_write(mp);
730 drop:
731 	return (error);
732 }
733 
734 /*
735  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
736  * does the I/O request for the non-physio version of the operations.  The
737  * normal vn operations are used, and this code should work in all instances
738  * for every type of file, including pipes, sockets, fifos, and regular files.
739  *
740  * XXX I don't think it works well for socket, pipe, and fifo.
741  */
742 static void
743 aio_process_rw(struct kaiocb *job)
744 {
745 	struct ucred *td_savedcred;
746 	struct thread *td;
747 	struct aiocb *cb;
748 	struct file *fp;
749 	struct uio auio;
750 	struct iovec aiov;
751 	ssize_t cnt;
752 	long msgsnd_st, msgsnd_end;
753 	long msgrcv_st, msgrcv_end;
754 	long oublock_st, oublock_end;
755 	long inblock_st, inblock_end;
756 	int error;
757 
758 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
759 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
760 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
761 
762 	aio_switch_vmspace(job);
763 	td = curthread;
764 	td_savedcred = td->td_ucred;
765 	td->td_ucred = job->cred;
766 	cb = &job->uaiocb;
767 	fp = job->fd_file;
768 
769 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
770 	aiov.iov_len = cb->aio_nbytes;
771 
772 	auio.uio_iov = &aiov;
773 	auio.uio_iovcnt = 1;
774 	auio.uio_offset = cb->aio_offset;
775 	auio.uio_resid = cb->aio_nbytes;
776 	cnt = cb->aio_nbytes;
777 	auio.uio_segflg = UIO_USERSPACE;
778 	auio.uio_td = td;
779 
780 	msgrcv_st = td->td_ru.ru_msgrcv;
781 	msgsnd_st = td->td_ru.ru_msgsnd;
782 	inblock_st = td->td_ru.ru_inblock;
783 	oublock_st = td->td_ru.ru_oublock;
784 
785 	/*
786 	 * aio_aqueue() acquires a reference to the file that is
787 	 * released in aio_free_entry().
788 	 */
789 	if (cb->aio_lio_opcode == LIO_READ) {
790 		auio.uio_rw = UIO_READ;
791 		if (auio.uio_resid == 0)
792 			error = 0;
793 		else
794 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
795 	} else {
796 		if (fp->f_type == DTYPE_VNODE)
797 			bwillwrite();
798 		auio.uio_rw = UIO_WRITE;
799 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
800 	}
801 	msgrcv_end = td->td_ru.ru_msgrcv;
802 	msgsnd_end = td->td_ru.ru_msgsnd;
803 	inblock_end = td->td_ru.ru_inblock;
804 	oublock_end = td->td_ru.ru_oublock;
805 
806 	job->msgrcv = msgrcv_end - msgrcv_st;
807 	job->msgsnd = msgsnd_end - msgsnd_st;
808 	job->inblock = inblock_end - inblock_st;
809 	job->outblock = oublock_end - oublock_st;
810 
811 	if ((error) && (auio.uio_resid != cnt)) {
812 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
813 			error = 0;
814 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
815 			PROC_LOCK(job->userproc);
816 			kern_psignal(job->userproc, SIGPIPE);
817 			PROC_UNLOCK(job->userproc);
818 		}
819 	}
820 
821 	cnt -= auio.uio_resid;
822 	td->td_ucred = td_savedcred;
823 	if (error)
824 		aio_complete(job, -1, error);
825 	else
826 		aio_complete(job, cnt, 0);
827 }
828 
829 static void
830 aio_process_sync(struct kaiocb *job)
831 {
832 	struct thread *td = curthread;
833 	struct ucred *td_savedcred = td->td_ucred;
834 	struct file *fp = job->fd_file;
835 	int error = 0;
836 
837 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
838 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
839 
840 	td->td_ucred = job->cred;
841 	if (fp->f_vnode != NULL)
842 		error = aio_fsync_vnode(td, fp->f_vnode);
843 	td->td_ucred = td_savedcred;
844 	if (error)
845 		aio_complete(job, -1, error);
846 	else
847 		aio_complete(job, 0, 0);
848 }
849 
850 static void
851 aio_process_mlock(struct kaiocb *job)
852 {
853 	struct aiocb *cb = &job->uaiocb;
854 	int error;
855 
856 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
857 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
858 
859 	aio_switch_vmspace(job);
860 	error = vm_mlock(job->userproc, job->cred,
861 	    __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
862 	if (error)
863 		aio_complete(job, -1, error);
864 	else
865 		aio_complete(job, 0, 0);
866 }
867 
868 static void
869 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
870 {
871 	struct aioliojob *lj;
872 	struct kaioinfo *ki;
873 	struct kaiocb *sjob, *sjobn;
874 	int lj_done;
875 	bool schedule_fsync;
876 
877 	ki = userp->p_aioinfo;
878 	AIO_LOCK_ASSERT(ki, MA_OWNED);
879 	lj = job->lio;
880 	lj_done = 0;
881 	if (lj) {
882 		lj->lioj_finished_count++;
883 		if (lj->lioj_count == lj->lioj_finished_count)
884 			lj_done = 1;
885 	}
886 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
887 	MPASS(job->jobflags & KAIOCB_FINISHED);
888 
889 	if (ki->kaio_flags & KAIO_RUNDOWN)
890 		goto notification_done;
891 
892 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
893 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
894 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
895 
896 	KNOTE_LOCKED(&job->klist, 1);
897 
898 	if (lj_done) {
899 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
900 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
901 			KNOTE_LOCKED(&lj->klist, 1);
902 		}
903 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
904 		    == LIOJ_SIGNAL
905 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
906 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
907 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
908 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
909 		}
910 	}
911 
912 notification_done:
913 	if (job->jobflags & KAIOCB_CHECKSYNC) {
914 		schedule_fsync = false;
915 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
916 			if (job->fd_file != sjob->fd_file ||
917 			    job->seqno >= sjob->seqno)
918 				continue;
919 			if (--sjob->pending > 0)
920 				continue;
921 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
922 			if (!aio_clear_cancel_function(sjob))
923 				continue;
924 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
925 			schedule_fsync = true;
926 		}
927 		if (schedule_fsync)
928 			taskqueue_enqueue(taskqueue_aiod_kick,
929 			    &ki->kaio_sync_task);
930 	}
931 	if (ki->kaio_flags & KAIO_WAKEUP) {
932 		ki->kaio_flags &= ~KAIO_WAKEUP;
933 		wakeup(&userp->p_aioinfo);
934 	}
935 }
936 
937 static void
938 aio_schedule_fsync(void *context, int pending)
939 {
940 	struct kaioinfo *ki;
941 	struct kaiocb *job;
942 
943 	ki = context;
944 	AIO_LOCK(ki);
945 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
946 		job = TAILQ_FIRST(&ki->kaio_syncready);
947 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
948 		AIO_UNLOCK(ki);
949 		aio_schedule(job, aio_process_sync);
950 		AIO_LOCK(ki);
951 	}
952 	AIO_UNLOCK(ki);
953 }
954 
955 bool
956 aio_cancel_cleared(struct kaiocb *job)
957 {
958 	struct kaioinfo *ki;
959 
960 	/*
961 	 * The caller should hold the same queue lock held when
962 	 * aio_clear_cancel_function() was called and set this flag
963 	 * ensuring this check sees an up-to-date value.  However,
964 	 * there is no way to assert that.
965 	 */
966 	ki = job->userproc->p_aioinfo;
967 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
968 }
969 
970 bool
971 aio_clear_cancel_function(struct kaiocb *job)
972 {
973 	struct kaioinfo *ki;
974 
975 	ki = job->userproc->p_aioinfo;
976 	AIO_LOCK(ki);
977 	MPASS(job->cancel_fn != NULL);
978 	if (job->jobflags & KAIOCB_CANCELLING) {
979 		job->jobflags |= KAIOCB_CLEARED;
980 		AIO_UNLOCK(ki);
981 		return (false);
982 	}
983 	job->cancel_fn = NULL;
984 	AIO_UNLOCK(ki);
985 	return (true);
986 }
987 
988 bool
989 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
990 {
991 	struct kaioinfo *ki;
992 
993 	ki = job->userproc->p_aioinfo;
994 	AIO_LOCK(ki);
995 	if (job->jobflags & KAIOCB_CANCELLED) {
996 		AIO_UNLOCK(ki);
997 		return (false);
998 	}
999 	job->cancel_fn = func;
1000 	AIO_UNLOCK(ki);
1001 	return (true);
1002 }
1003 
1004 void
1005 aio_complete(struct kaiocb *job, long status, int error)
1006 {
1007 	struct kaioinfo *ki;
1008 	struct proc *userp;
1009 
1010 	job->uaiocb._aiocb_private.error = error;
1011 	job->uaiocb._aiocb_private.status = status;
1012 
1013 	userp = job->userproc;
1014 	ki = userp->p_aioinfo;
1015 
1016 	AIO_LOCK(ki);
1017 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1018 	    ("duplicate aio_complete"));
1019 	job->jobflags |= KAIOCB_FINISHED;
1020 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1021 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1022 		aio_bio_done_notify(userp, job);
1023 	}
1024 	AIO_UNLOCK(ki);
1025 }
1026 
1027 void
1028 aio_cancel(struct kaiocb *job)
1029 {
1030 
1031 	aio_complete(job, -1, ECANCELED);
1032 }
1033 
1034 void
1035 aio_switch_vmspace(struct kaiocb *job)
1036 {
1037 
1038 	vmspace_switch_aio(job->userproc->p_vmspace);
1039 }
1040 
1041 /*
1042  * The AIO daemon, most of the actual work is done in aio_process_*,
1043  * but the setup (and address space mgmt) is done in this routine.
1044  */
1045 static void
1046 aio_daemon(void *_id)
1047 {
1048 	struct kaiocb *job;
1049 	struct aioproc *aiop;
1050 	struct kaioinfo *ki;
1051 	struct proc *p;
1052 	struct vmspace *myvm;
1053 	struct thread *td = curthread;
1054 	int id = (intptr_t)_id;
1055 
1056 	/*
1057 	 * Grab an extra reference on the daemon's vmspace so that it
1058 	 * doesn't get freed by jobs that switch to a different
1059 	 * vmspace.
1060 	 */
1061 	p = td->td_proc;
1062 	myvm = vmspace_acquire_ref(p);
1063 
1064 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1065 
1066 	/*
1067 	 * Allocate and ready the aio control info.  There is one aiop structure
1068 	 * per daemon.
1069 	 */
1070 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1071 	aiop->aioproc = p;
1072 	aiop->aioprocflags = 0;
1073 
1074 	/*
1075 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1076 	 * and creating too many daemons.)
1077 	 */
1078 	sema_post(&aio_newproc_sem);
1079 
1080 	mtx_lock(&aio_job_mtx);
1081 	for (;;) {
1082 		/*
1083 		 * Take daemon off of free queue
1084 		 */
1085 		if (aiop->aioprocflags & AIOP_FREE) {
1086 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1087 			aiop->aioprocflags &= ~AIOP_FREE;
1088 		}
1089 
1090 		/*
1091 		 * Check for jobs.
1092 		 */
1093 		while ((job = aio_selectjob(aiop)) != NULL) {
1094 			mtx_unlock(&aio_job_mtx);
1095 
1096 			ki = job->userproc->p_aioinfo;
1097 			job->handle_fn(job);
1098 
1099 			mtx_lock(&aio_job_mtx);
1100 			/* Decrement the active job count. */
1101 			ki->kaio_active_count--;
1102 		}
1103 
1104 		/*
1105 		 * Disconnect from user address space.
1106 		 */
1107 		if (p->p_vmspace != myvm) {
1108 			mtx_unlock(&aio_job_mtx);
1109 			vmspace_switch_aio(myvm);
1110 			mtx_lock(&aio_job_mtx);
1111 			/*
1112 			 * We have to restart to avoid race, we only sleep if
1113 			 * no job can be selected.
1114 			 */
1115 			continue;
1116 		}
1117 
1118 		mtx_assert(&aio_job_mtx, MA_OWNED);
1119 
1120 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1121 		aiop->aioprocflags |= AIOP_FREE;
1122 
1123 		/*
1124 		 * If daemon is inactive for a long time, allow it to exit,
1125 		 * thereby freeing resources.
1126 		 */
1127 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1128 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1129 		    (aiop->aioprocflags & AIOP_FREE) &&
1130 		    num_aio_procs > target_aio_procs)
1131 			break;
1132 	}
1133 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1134 	num_aio_procs--;
1135 	mtx_unlock(&aio_job_mtx);
1136 	uma_zfree(aiop_zone, aiop);
1137 	free_unr(aiod_unr, id);
1138 	vmspace_free(myvm);
1139 
1140 	KASSERT(p->p_vmspace == myvm,
1141 	    ("AIOD: bad vmspace for exiting daemon"));
1142 	KASSERT(myvm->vm_refcnt > 1,
1143 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1144 	kproc_exit(0);
1145 }
1146 
1147 /*
1148  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1149  * AIO daemon modifies its environment itself.
1150  */
1151 static int
1152 aio_newproc(int *start)
1153 {
1154 	int error;
1155 	struct proc *p;
1156 	int id;
1157 
1158 	id = alloc_unr(aiod_unr);
1159 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1160 		RFNOWAIT, 0, "aiod%d", id);
1161 	if (error == 0) {
1162 		/*
1163 		 * Wait until daemon is started.
1164 		 */
1165 		sema_wait(&aio_newproc_sem);
1166 		mtx_lock(&aio_job_mtx);
1167 		num_aio_procs++;
1168 		if (start != NULL)
1169 			(*start)--;
1170 		mtx_unlock(&aio_job_mtx);
1171 	} else {
1172 		free_unr(aiod_unr, id);
1173 	}
1174 	return (error);
1175 }
1176 
1177 /*
1178  * Try the high-performance, low-overhead physio method for eligible
1179  * VCHR devices.  This method doesn't use an aio helper thread, and
1180  * thus has very low overhead.
1181  *
1182  * Assumes that the caller, aio_aqueue(), has incremented the file
1183  * structure's reference count, preventing its deallocation for the
1184  * duration of this call.
1185  */
1186 static int
1187 aio_qphysio(struct proc *p, struct kaiocb *job)
1188 {
1189 	struct aiocb *cb;
1190 	struct file *fp;
1191 	struct bio *bp;
1192 	struct buf *pbuf;
1193 	struct vnode *vp;
1194 	struct cdevsw *csw;
1195 	struct cdev *dev;
1196 	struct kaioinfo *ki;
1197 	int error, ref, poff;
1198 	vm_prot_t prot;
1199 
1200 	cb = &job->uaiocb;
1201 	fp = job->fd_file;
1202 
1203 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1204 		return (-1);
1205 
1206 	vp = fp->f_vnode;
1207 	if (vp->v_type != VCHR)
1208 		return (-1);
1209 	if (vp->v_bufobj.bo_bsize == 0)
1210 		return (-1);
1211 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1212 		return (-1);
1213 
1214 	ref = 0;
1215 	csw = devvn_refthread(vp, &dev, &ref);
1216 	if (csw == NULL)
1217 		return (ENXIO);
1218 
1219 	if ((csw->d_flags & D_DISK) == 0) {
1220 		error = -1;
1221 		goto unref;
1222 	}
1223 	if (cb->aio_nbytes > dev->si_iosize_max) {
1224 		error = -1;
1225 		goto unref;
1226 	}
1227 
1228 	ki = p->p_aioinfo;
1229 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1230 	if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1231 		if (cb->aio_nbytes > MAXPHYS) {
1232 			error = -1;
1233 			goto unref;
1234 		}
1235 
1236 		pbuf = NULL;
1237 	} else {
1238 		if (cb->aio_nbytes > MAXPHYS - poff) {
1239 			error = -1;
1240 			goto unref;
1241 		}
1242 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1243 			error = -1;
1244 			goto unref;
1245 		}
1246 
1247 		job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1248 		BUF_KERNPROC(pbuf);
1249 		AIO_LOCK(ki);
1250 		ki->kaio_buffer_count++;
1251 		AIO_UNLOCK(ki);
1252 	}
1253 	job->bp = bp = g_alloc_bio();
1254 
1255 	bp->bio_length = cb->aio_nbytes;
1256 	bp->bio_bcount = cb->aio_nbytes;
1257 	bp->bio_done = aio_physwakeup;
1258 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1259 	bp->bio_offset = cb->aio_offset;
1260 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1261 	bp->bio_dev = dev;
1262 	bp->bio_caller1 = (void *)job;
1263 
1264 	prot = VM_PROT_READ;
1265 	if (cb->aio_lio_opcode == LIO_READ)
1266 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1267 	job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1268 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1269 	    nitems(job->pages));
1270 	if (job->npages < 0) {
1271 		error = EFAULT;
1272 		goto doerror;
1273 	}
1274 	if (pbuf != NULL) {
1275 		pmap_qenter((vm_offset_t)pbuf->b_data,
1276 		    job->pages, job->npages);
1277 		bp->bio_data = pbuf->b_data + poff;
1278 		atomic_add_int(&num_buf_aio, 1);
1279 	} else {
1280 		bp->bio_ma = job->pages;
1281 		bp->bio_ma_n = job->npages;
1282 		bp->bio_ma_offset = poff;
1283 		bp->bio_data = unmapped_buf;
1284 		bp->bio_flags |= BIO_UNMAPPED;
1285 	}
1286 
1287 	/* Perform transfer. */
1288 	csw->d_strategy(bp);
1289 	dev_relthread(dev, ref);
1290 	return (0);
1291 
1292 doerror:
1293 	if (pbuf != NULL) {
1294 		AIO_LOCK(ki);
1295 		ki->kaio_buffer_count--;
1296 		AIO_UNLOCK(ki);
1297 		relpbuf(pbuf, NULL);
1298 		job->pbuf = NULL;
1299 	}
1300 	g_destroy_bio(bp);
1301 	job->bp = NULL;
1302 unref:
1303 	dev_relthread(dev, ref);
1304 	return (error);
1305 }
1306 
1307 #ifdef COMPAT_FREEBSD6
1308 static int
1309 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1310 {
1311 
1312 	/*
1313 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1314 	 * supported by AIO with the old sigevent structure.
1315 	 */
1316 	nsig->sigev_notify = osig->sigev_notify;
1317 	switch (nsig->sigev_notify) {
1318 	case SIGEV_NONE:
1319 		break;
1320 	case SIGEV_SIGNAL:
1321 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1322 		break;
1323 	case SIGEV_KEVENT:
1324 		nsig->sigev_notify_kqueue =
1325 		    osig->__sigev_u.__sigev_notify_kqueue;
1326 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1327 		break;
1328 	default:
1329 		return (EINVAL);
1330 	}
1331 	return (0);
1332 }
1333 
1334 static int
1335 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1336 {
1337 	struct oaiocb *ojob;
1338 	int error;
1339 
1340 	bzero(kjob, sizeof(struct aiocb));
1341 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1342 	if (error)
1343 		return (error);
1344 	ojob = (struct oaiocb *)kjob;
1345 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1346 }
1347 #endif
1348 
1349 static int
1350 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1351 {
1352 
1353 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1354 }
1355 
1356 static long
1357 aiocb_fetch_status(struct aiocb *ujob)
1358 {
1359 
1360 	return (fuword(&ujob->_aiocb_private.status));
1361 }
1362 
1363 static long
1364 aiocb_fetch_error(struct aiocb *ujob)
1365 {
1366 
1367 	return (fuword(&ujob->_aiocb_private.error));
1368 }
1369 
1370 static int
1371 aiocb_store_status(struct aiocb *ujob, long status)
1372 {
1373 
1374 	return (suword(&ujob->_aiocb_private.status, status));
1375 }
1376 
1377 static int
1378 aiocb_store_error(struct aiocb *ujob, long error)
1379 {
1380 
1381 	return (suword(&ujob->_aiocb_private.error, error));
1382 }
1383 
1384 static int
1385 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1386 {
1387 
1388 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1389 }
1390 
1391 static int
1392 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1393 {
1394 
1395 	return (suword(ujobp, (long)ujob));
1396 }
1397 
1398 static struct aiocb_ops aiocb_ops = {
1399 	.copyin = aiocb_copyin,
1400 	.fetch_status = aiocb_fetch_status,
1401 	.fetch_error = aiocb_fetch_error,
1402 	.store_status = aiocb_store_status,
1403 	.store_error = aiocb_store_error,
1404 	.store_kernelinfo = aiocb_store_kernelinfo,
1405 	.store_aiocb = aiocb_store_aiocb,
1406 };
1407 
1408 #ifdef COMPAT_FREEBSD6
1409 static struct aiocb_ops aiocb_ops_osigevent = {
1410 	.copyin = aiocb_copyin_old_sigevent,
1411 	.fetch_status = aiocb_fetch_status,
1412 	.fetch_error = aiocb_fetch_error,
1413 	.store_status = aiocb_store_status,
1414 	.store_error = aiocb_store_error,
1415 	.store_kernelinfo = aiocb_store_kernelinfo,
1416 	.store_aiocb = aiocb_store_aiocb,
1417 };
1418 #endif
1419 
1420 /*
1421  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1422  * technique is done in this code.
1423  */
1424 int
1425 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1426 	int type, struct aiocb_ops *ops)
1427 {
1428 	struct proc *p = td->td_proc;
1429 	cap_rights_t rights;
1430 	struct file *fp;
1431 	struct kaiocb *job;
1432 	struct kaioinfo *ki;
1433 	struct kevent kev;
1434 	int opcode;
1435 	int error;
1436 	int fd, kqfd;
1437 	int jid;
1438 	u_short evflags;
1439 
1440 	if (p->p_aioinfo == NULL)
1441 		aio_init_aioinfo(p);
1442 
1443 	ki = p->p_aioinfo;
1444 
1445 	ops->store_status(ujob, -1);
1446 	ops->store_error(ujob, 0);
1447 	ops->store_kernelinfo(ujob, -1);
1448 
1449 	if (num_queue_count >= max_queue_count ||
1450 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1451 		ops->store_error(ujob, EAGAIN);
1452 		return (EAGAIN);
1453 	}
1454 
1455 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1456 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1457 
1458 	error = ops->copyin(ujob, &job->uaiocb);
1459 	if (error) {
1460 		ops->store_error(ujob, error);
1461 		uma_zfree(aiocb_zone, job);
1462 		return (error);
1463 	}
1464 
1465 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1466 		uma_zfree(aiocb_zone, job);
1467 		return (EINVAL);
1468 	}
1469 
1470 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1471 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1472 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1473 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1474 		ops->store_error(ujob, EINVAL);
1475 		uma_zfree(aiocb_zone, job);
1476 		return (EINVAL);
1477 	}
1478 
1479 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1480 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1481 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1482 		uma_zfree(aiocb_zone, job);
1483 		return (EINVAL);
1484 	}
1485 
1486 	ksiginfo_init(&job->ksi);
1487 
1488 	/* Save userspace address of the job info. */
1489 	job->ujob = ujob;
1490 
1491 	/* Get the opcode. */
1492 	if (type != LIO_NOP)
1493 		job->uaiocb.aio_lio_opcode = type;
1494 	opcode = job->uaiocb.aio_lio_opcode;
1495 
1496 	/*
1497 	 * Validate the opcode and fetch the file object for the specified
1498 	 * file descriptor.
1499 	 *
1500 	 * XXXRW: Moved the opcode validation up here so that we don't
1501 	 * retrieve a file descriptor without knowing what the capabiltity
1502 	 * should be.
1503 	 */
1504 	fd = job->uaiocb.aio_fildes;
1505 	switch (opcode) {
1506 	case LIO_WRITE:
1507 		error = fget_write(td, fd,
1508 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1509 		break;
1510 	case LIO_READ:
1511 		error = fget_read(td, fd,
1512 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1513 		break;
1514 	case LIO_SYNC:
1515 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1516 		break;
1517 	case LIO_MLOCK:
1518 		fp = NULL;
1519 		break;
1520 	case LIO_NOP:
1521 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1522 		break;
1523 	default:
1524 		error = EINVAL;
1525 	}
1526 	if (error) {
1527 		uma_zfree(aiocb_zone, job);
1528 		ops->store_error(ujob, error);
1529 		return (error);
1530 	}
1531 
1532 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1533 		error = EINVAL;
1534 		goto aqueue_fail;
1535 	}
1536 
1537 	if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1538 		error = EINVAL;
1539 		goto aqueue_fail;
1540 	}
1541 
1542 	job->fd_file = fp;
1543 
1544 	mtx_lock(&aio_job_mtx);
1545 	jid = jobrefid++;
1546 	job->seqno = jobseqno++;
1547 	mtx_unlock(&aio_job_mtx);
1548 	error = ops->store_kernelinfo(ujob, jid);
1549 	if (error) {
1550 		error = EINVAL;
1551 		goto aqueue_fail;
1552 	}
1553 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1554 
1555 	if (opcode == LIO_NOP) {
1556 		fdrop(fp, td);
1557 		uma_zfree(aiocb_zone, job);
1558 		return (0);
1559 	}
1560 
1561 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1562 		goto no_kqueue;
1563 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1564 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1565 		error = EINVAL;
1566 		goto aqueue_fail;
1567 	}
1568 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1569 	kev.ident = (uintptr_t)job->ujob;
1570 	kev.filter = EVFILT_AIO;
1571 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1572 	kev.data = (intptr_t)job;
1573 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1574 	error = kqfd_register(kqfd, &kev, td, 1);
1575 	if (error)
1576 		goto aqueue_fail;
1577 
1578 no_kqueue:
1579 
1580 	ops->store_error(ujob, EINPROGRESS);
1581 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1582 	job->userproc = p;
1583 	job->cred = crhold(td->td_ucred);
1584 	job->jobflags = KAIOCB_QUEUEING;
1585 	job->lio = lj;
1586 
1587 	if (opcode == LIO_MLOCK) {
1588 		aio_schedule(job, aio_process_mlock);
1589 		error = 0;
1590 	} else if (fp->f_ops->fo_aio_queue == NULL)
1591 		error = aio_queue_file(fp, job);
1592 	else
1593 		error = fo_aio_queue(fp, job);
1594 	if (error)
1595 		goto aqueue_fail;
1596 
1597 	AIO_LOCK(ki);
1598 	job->jobflags &= ~KAIOCB_QUEUEING;
1599 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1600 	ki->kaio_count++;
1601 	if (lj)
1602 		lj->lioj_count++;
1603 	atomic_add_int(&num_queue_count, 1);
1604 	if (job->jobflags & KAIOCB_FINISHED) {
1605 		/*
1606 		 * The queue callback completed the request synchronously.
1607 		 * The bulk of the completion is deferred in that case
1608 		 * until this point.
1609 		 */
1610 		aio_bio_done_notify(p, job);
1611 	} else
1612 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1613 	AIO_UNLOCK(ki);
1614 	return (0);
1615 
1616 aqueue_fail:
1617 	knlist_delete(&job->klist, curthread, 0);
1618 	if (fp)
1619 		fdrop(fp, td);
1620 	uma_zfree(aiocb_zone, job);
1621 	ops->store_error(ujob, error);
1622 	return (error);
1623 }
1624 
1625 static void
1626 aio_cancel_daemon_job(struct kaiocb *job)
1627 {
1628 
1629 	mtx_lock(&aio_job_mtx);
1630 	if (!aio_cancel_cleared(job))
1631 		TAILQ_REMOVE(&aio_jobs, job, list);
1632 	mtx_unlock(&aio_job_mtx);
1633 	aio_cancel(job);
1634 }
1635 
1636 void
1637 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1638 {
1639 
1640 	mtx_lock(&aio_job_mtx);
1641 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1642 		mtx_unlock(&aio_job_mtx);
1643 		aio_cancel(job);
1644 		return;
1645 	}
1646 	job->handle_fn = func;
1647 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1648 	aio_kick_nowait(job->userproc);
1649 	mtx_unlock(&aio_job_mtx);
1650 }
1651 
1652 static void
1653 aio_cancel_sync(struct kaiocb *job)
1654 {
1655 	struct kaioinfo *ki;
1656 
1657 	ki = job->userproc->p_aioinfo;
1658 	mtx_lock(&aio_job_mtx);
1659 	if (!aio_cancel_cleared(job))
1660 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1661 	mtx_unlock(&aio_job_mtx);
1662 	aio_cancel(job);
1663 }
1664 
1665 int
1666 aio_queue_file(struct file *fp, struct kaiocb *job)
1667 {
1668 	struct aioliojob *lj;
1669 	struct kaioinfo *ki;
1670 	struct kaiocb *job2;
1671 	struct vnode *vp;
1672 	struct mount *mp;
1673 	int error, opcode;
1674 	bool safe;
1675 
1676 	lj = job->lio;
1677 	ki = job->userproc->p_aioinfo;
1678 	opcode = job->uaiocb.aio_lio_opcode;
1679 	if (opcode == LIO_SYNC)
1680 		goto queueit;
1681 
1682 	if ((error = aio_qphysio(job->userproc, job)) == 0)
1683 		goto done;
1684 #if 0
1685 	/*
1686 	 * XXX: This means qphysio() failed with EFAULT.  The current
1687 	 * behavior is to retry the operation via fo_read/fo_write.
1688 	 * Wouldn't it be better to just complete the request with an
1689 	 * error here?
1690 	 */
1691 	if (error > 0)
1692 		goto done;
1693 #endif
1694 queueit:
1695 	safe = false;
1696 	if (fp->f_type == DTYPE_VNODE) {
1697 		vp = fp->f_vnode;
1698 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1699 			mp = fp->f_vnode->v_mount;
1700 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1701 				safe = true;
1702 		}
1703 	}
1704 	if (!(safe || enable_aio_unsafe)) {
1705 		counted_warning(&unsafe_warningcnt,
1706 		    "is attempting to use unsafe AIO requests");
1707 		return (EOPNOTSUPP);
1708 	}
1709 
1710 	if (opcode == LIO_SYNC) {
1711 		AIO_LOCK(ki);
1712 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1713 			if (job2->fd_file == job->fd_file &&
1714 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1715 			    job2->seqno < job->seqno) {
1716 				job2->jobflags |= KAIOCB_CHECKSYNC;
1717 				job->pending++;
1718 			}
1719 		}
1720 		if (job->pending != 0) {
1721 			if (!aio_set_cancel_function(job, aio_cancel_sync)) {
1722 				AIO_UNLOCK(ki);
1723 				aio_cancel(job);
1724 				return (0);
1725 			}
1726 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1727 			AIO_UNLOCK(ki);
1728 			return (0);
1729 		}
1730 		AIO_UNLOCK(ki);
1731 	}
1732 
1733 	switch (opcode) {
1734 	case LIO_READ:
1735 	case LIO_WRITE:
1736 		aio_schedule(job, aio_process_rw);
1737 		error = 0;
1738 		break;
1739 	case LIO_SYNC:
1740 		aio_schedule(job, aio_process_sync);
1741 		error = 0;
1742 		break;
1743 	default:
1744 		error = EINVAL;
1745 	}
1746 done:
1747 	return (error);
1748 }
1749 
1750 static void
1751 aio_kick_nowait(struct proc *userp)
1752 {
1753 	struct kaioinfo *ki = userp->p_aioinfo;
1754 	struct aioproc *aiop;
1755 
1756 	mtx_assert(&aio_job_mtx, MA_OWNED);
1757 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1758 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1759 		aiop->aioprocflags &= ~AIOP_FREE;
1760 		wakeup(aiop->aioproc);
1761 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1762 	    ki->kaio_active_count + num_aio_resv_start <
1763 	    ki->kaio_maxactive_count) {
1764 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1765 	}
1766 }
1767 
1768 static int
1769 aio_kick(struct proc *userp)
1770 {
1771 	struct kaioinfo *ki = userp->p_aioinfo;
1772 	struct aioproc *aiop;
1773 	int error, ret = 0;
1774 
1775 	mtx_assert(&aio_job_mtx, MA_OWNED);
1776 retryproc:
1777 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1778 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1779 		aiop->aioprocflags &= ~AIOP_FREE;
1780 		wakeup(aiop->aioproc);
1781 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1782 	    ki->kaio_active_count + num_aio_resv_start <
1783 	    ki->kaio_maxactive_count) {
1784 		num_aio_resv_start++;
1785 		mtx_unlock(&aio_job_mtx);
1786 		error = aio_newproc(&num_aio_resv_start);
1787 		mtx_lock(&aio_job_mtx);
1788 		if (error) {
1789 			num_aio_resv_start--;
1790 			goto retryproc;
1791 		}
1792 	} else {
1793 		ret = -1;
1794 	}
1795 	return (ret);
1796 }
1797 
1798 static void
1799 aio_kick_helper(void *context, int pending)
1800 {
1801 	struct proc *userp = context;
1802 
1803 	mtx_lock(&aio_job_mtx);
1804 	while (--pending >= 0) {
1805 		if (aio_kick(userp))
1806 			break;
1807 	}
1808 	mtx_unlock(&aio_job_mtx);
1809 }
1810 
1811 /*
1812  * Support the aio_return system call, as a side-effect, kernel resources are
1813  * released.
1814  */
1815 static int
1816 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1817 {
1818 	struct proc *p = td->td_proc;
1819 	struct kaiocb *job;
1820 	struct kaioinfo *ki;
1821 	long status, error;
1822 
1823 	ki = p->p_aioinfo;
1824 	if (ki == NULL)
1825 		return (EINVAL);
1826 	AIO_LOCK(ki);
1827 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1828 		if (job->ujob == ujob)
1829 			break;
1830 	}
1831 	if (job != NULL) {
1832 		MPASS(job->jobflags & KAIOCB_FINISHED);
1833 		status = job->uaiocb._aiocb_private.status;
1834 		error = job->uaiocb._aiocb_private.error;
1835 		td->td_retval[0] = status;
1836 		td->td_ru.ru_oublock += job->outblock;
1837 		td->td_ru.ru_inblock += job->inblock;
1838 		td->td_ru.ru_msgsnd += job->msgsnd;
1839 		td->td_ru.ru_msgrcv += job->msgrcv;
1840 		aio_free_entry(job);
1841 		AIO_UNLOCK(ki);
1842 		ops->store_error(ujob, error);
1843 		ops->store_status(ujob, status);
1844 	} else {
1845 		error = EINVAL;
1846 		AIO_UNLOCK(ki);
1847 	}
1848 	return (error);
1849 }
1850 
1851 int
1852 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1853 {
1854 
1855 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1856 }
1857 
1858 /*
1859  * Allow a process to wakeup when any of the I/O requests are completed.
1860  */
1861 static int
1862 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1863     struct timespec *ts)
1864 {
1865 	struct proc *p = td->td_proc;
1866 	struct timeval atv;
1867 	struct kaioinfo *ki;
1868 	struct kaiocb *firstjob, *job;
1869 	int error, i, timo;
1870 
1871 	timo = 0;
1872 	if (ts) {
1873 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1874 			return (EINVAL);
1875 
1876 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1877 		if (itimerfix(&atv))
1878 			return (EINVAL);
1879 		timo = tvtohz(&atv);
1880 	}
1881 
1882 	ki = p->p_aioinfo;
1883 	if (ki == NULL)
1884 		return (EAGAIN);
1885 
1886 	if (njoblist == 0)
1887 		return (0);
1888 
1889 	AIO_LOCK(ki);
1890 	for (;;) {
1891 		firstjob = NULL;
1892 		error = 0;
1893 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1894 			for (i = 0; i < njoblist; i++) {
1895 				if (job->ujob == ujoblist[i]) {
1896 					if (firstjob == NULL)
1897 						firstjob = job;
1898 					if (job->jobflags & KAIOCB_FINISHED)
1899 						goto RETURN;
1900 				}
1901 			}
1902 		}
1903 		/* All tasks were finished. */
1904 		if (firstjob == NULL)
1905 			break;
1906 
1907 		ki->kaio_flags |= KAIO_WAKEUP;
1908 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1909 		    "aiospn", timo);
1910 		if (error == ERESTART)
1911 			error = EINTR;
1912 		if (error)
1913 			break;
1914 	}
1915 RETURN:
1916 	AIO_UNLOCK(ki);
1917 	return (error);
1918 }
1919 
1920 int
1921 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1922 {
1923 	struct timespec ts, *tsp;
1924 	struct aiocb **ujoblist;
1925 	int error;
1926 
1927 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1928 		return (EINVAL);
1929 
1930 	if (uap->timeout) {
1931 		/* Get timespec struct. */
1932 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1933 			return (error);
1934 		tsp = &ts;
1935 	} else
1936 		tsp = NULL;
1937 
1938 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1939 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1940 	if (error == 0)
1941 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1942 	uma_zfree(aiol_zone, ujoblist);
1943 	return (error);
1944 }
1945 
1946 /*
1947  * aio_cancel cancels any non-physio aio operations not currently in
1948  * progress.
1949  */
1950 int
1951 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1952 {
1953 	struct proc *p = td->td_proc;
1954 	struct kaioinfo *ki;
1955 	struct kaiocb *job, *jobn;
1956 	struct file *fp;
1957 	cap_rights_t rights;
1958 	int error;
1959 	int cancelled = 0;
1960 	int notcancelled = 0;
1961 	struct vnode *vp;
1962 
1963 	/* Lookup file object. */
1964 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1965 	if (error)
1966 		return (error);
1967 
1968 	ki = p->p_aioinfo;
1969 	if (ki == NULL)
1970 		goto done;
1971 
1972 	if (fp->f_type == DTYPE_VNODE) {
1973 		vp = fp->f_vnode;
1974 		if (vn_isdisk(vp, &error)) {
1975 			fdrop(fp, td);
1976 			td->td_retval[0] = AIO_NOTCANCELED;
1977 			return (0);
1978 		}
1979 	}
1980 
1981 	AIO_LOCK(ki);
1982 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1983 		if ((uap->fd == job->uaiocb.aio_fildes) &&
1984 		    ((uap->aiocbp == NULL) ||
1985 		     (uap->aiocbp == job->ujob))) {
1986 			if (aio_cancel_job(p, ki, job)) {
1987 				cancelled++;
1988 			} else {
1989 				notcancelled++;
1990 			}
1991 			if (uap->aiocbp != NULL)
1992 				break;
1993 		}
1994 	}
1995 	AIO_UNLOCK(ki);
1996 
1997 done:
1998 	fdrop(fp, td);
1999 
2000 	if (uap->aiocbp != NULL) {
2001 		if (cancelled) {
2002 			td->td_retval[0] = AIO_CANCELED;
2003 			return (0);
2004 		}
2005 	}
2006 
2007 	if (notcancelled) {
2008 		td->td_retval[0] = AIO_NOTCANCELED;
2009 		return (0);
2010 	}
2011 
2012 	if (cancelled) {
2013 		td->td_retval[0] = AIO_CANCELED;
2014 		return (0);
2015 	}
2016 
2017 	td->td_retval[0] = AIO_ALLDONE;
2018 
2019 	return (0);
2020 }
2021 
2022 /*
2023  * aio_error is implemented in the kernel level for compatibility purposes
2024  * only.  For a user mode async implementation, it would be best to do it in
2025  * a userland subroutine.
2026  */
2027 static int
2028 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2029 {
2030 	struct proc *p = td->td_proc;
2031 	struct kaiocb *job;
2032 	struct kaioinfo *ki;
2033 	int status;
2034 
2035 	ki = p->p_aioinfo;
2036 	if (ki == NULL) {
2037 		td->td_retval[0] = EINVAL;
2038 		return (0);
2039 	}
2040 
2041 	AIO_LOCK(ki);
2042 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2043 		if (job->ujob == ujob) {
2044 			if (job->jobflags & KAIOCB_FINISHED)
2045 				td->td_retval[0] =
2046 					job->uaiocb._aiocb_private.error;
2047 			else
2048 				td->td_retval[0] = EINPROGRESS;
2049 			AIO_UNLOCK(ki);
2050 			return (0);
2051 		}
2052 	}
2053 	AIO_UNLOCK(ki);
2054 
2055 	/*
2056 	 * Hack for failure of aio_aqueue.
2057 	 */
2058 	status = ops->fetch_status(ujob);
2059 	if (status == -1) {
2060 		td->td_retval[0] = ops->fetch_error(ujob);
2061 		return (0);
2062 	}
2063 
2064 	td->td_retval[0] = EINVAL;
2065 	return (0);
2066 }
2067 
2068 int
2069 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2070 {
2071 
2072 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2073 }
2074 
2075 /* syscall - asynchronous read from a file (REALTIME) */
2076 #ifdef COMPAT_FREEBSD6
2077 int
2078 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2079 {
2080 
2081 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2082 	    &aiocb_ops_osigevent));
2083 }
2084 #endif
2085 
2086 int
2087 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2088 {
2089 
2090 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2091 }
2092 
2093 /* syscall - asynchronous write to a file (REALTIME) */
2094 #ifdef COMPAT_FREEBSD6
2095 int
2096 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2097 {
2098 
2099 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2100 	    &aiocb_ops_osigevent));
2101 }
2102 #endif
2103 
2104 int
2105 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2106 {
2107 
2108 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2109 }
2110 
2111 int
2112 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2113 {
2114 
2115 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2116 }
2117 
2118 static int
2119 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2120     struct aiocb **acb_list, int nent, struct sigevent *sig,
2121     struct aiocb_ops *ops)
2122 {
2123 	struct proc *p = td->td_proc;
2124 	struct aiocb *job;
2125 	struct kaioinfo *ki;
2126 	struct aioliojob *lj;
2127 	struct kevent kev;
2128 	int error;
2129 	int nerror;
2130 	int i;
2131 
2132 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2133 		return (EINVAL);
2134 
2135 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2136 		return (EINVAL);
2137 
2138 	if (p->p_aioinfo == NULL)
2139 		aio_init_aioinfo(p);
2140 
2141 	ki = p->p_aioinfo;
2142 
2143 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2144 	lj->lioj_flags = 0;
2145 	lj->lioj_count = 0;
2146 	lj->lioj_finished_count = 0;
2147 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2148 	ksiginfo_init(&lj->lioj_ksi);
2149 
2150 	/*
2151 	 * Setup signal.
2152 	 */
2153 	if (sig && (mode == LIO_NOWAIT)) {
2154 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2155 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2156 			/* Assume only new style KEVENT */
2157 			kev.filter = EVFILT_LIO;
2158 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2159 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2160 			kev.data = (intptr_t)lj;
2161 			/* pass user defined sigval data */
2162 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2163 			error = kqfd_register(
2164 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2165 			if (error) {
2166 				uma_zfree(aiolio_zone, lj);
2167 				return (error);
2168 			}
2169 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2170 			;
2171 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2172 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2173 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2174 					uma_zfree(aiolio_zone, lj);
2175 					return EINVAL;
2176 				}
2177 				lj->lioj_flags |= LIOJ_SIGNAL;
2178 		} else {
2179 			uma_zfree(aiolio_zone, lj);
2180 			return EINVAL;
2181 		}
2182 	}
2183 
2184 	AIO_LOCK(ki);
2185 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2186 	/*
2187 	 * Add extra aiocb count to avoid the lio to be freed
2188 	 * by other threads doing aio_waitcomplete or aio_return,
2189 	 * and prevent event from being sent until we have queued
2190 	 * all tasks.
2191 	 */
2192 	lj->lioj_count = 1;
2193 	AIO_UNLOCK(ki);
2194 
2195 	/*
2196 	 * Get pointers to the list of I/O requests.
2197 	 */
2198 	nerror = 0;
2199 	for (i = 0; i < nent; i++) {
2200 		job = acb_list[i];
2201 		if (job != NULL) {
2202 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2203 			if (error != 0)
2204 				nerror++;
2205 		}
2206 	}
2207 
2208 	error = 0;
2209 	AIO_LOCK(ki);
2210 	if (mode == LIO_WAIT) {
2211 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2212 			ki->kaio_flags |= KAIO_WAKEUP;
2213 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2214 			    PRIBIO | PCATCH, "aiospn", 0);
2215 			if (error == ERESTART)
2216 				error = EINTR;
2217 			if (error)
2218 				break;
2219 		}
2220 	} else {
2221 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2222 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2223 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2224 				KNOTE_LOCKED(&lj->klist, 1);
2225 			}
2226 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2227 			    == LIOJ_SIGNAL
2228 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2229 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2230 				aio_sendsig(p, &lj->lioj_signal,
2231 					    &lj->lioj_ksi);
2232 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2233 			}
2234 		}
2235 	}
2236 	lj->lioj_count--;
2237 	if (lj->lioj_count == 0) {
2238 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2239 		knlist_delete(&lj->klist, curthread, 1);
2240 		PROC_LOCK(p);
2241 		sigqueue_take(&lj->lioj_ksi);
2242 		PROC_UNLOCK(p);
2243 		AIO_UNLOCK(ki);
2244 		uma_zfree(aiolio_zone, lj);
2245 	} else
2246 		AIO_UNLOCK(ki);
2247 
2248 	if (nerror)
2249 		return (EIO);
2250 	return (error);
2251 }
2252 
2253 /* syscall - list directed I/O (REALTIME) */
2254 #ifdef COMPAT_FREEBSD6
2255 int
2256 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2257 {
2258 	struct aiocb **acb_list;
2259 	struct sigevent *sigp, sig;
2260 	struct osigevent osig;
2261 	int error, nent;
2262 
2263 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2264 		return (EINVAL);
2265 
2266 	nent = uap->nent;
2267 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2268 		return (EINVAL);
2269 
2270 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2271 		error = copyin(uap->sig, &osig, sizeof(osig));
2272 		if (error)
2273 			return (error);
2274 		error = convert_old_sigevent(&osig, &sig);
2275 		if (error)
2276 			return (error);
2277 		sigp = &sig;
2278 	} else
2279 		sigp = NULL;
2280 
2281 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2282 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2283 	if (error == 0)
2284 		error = kern_lio_listio(td, uap->mode,
2285 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2286 		    &aiocb_ops_osigevent);
2287 	free(acb_list, M_LIO);
2288 	return (error);
2289 }
2290 #endif
2291 
2292 /* syscall - list directed I/O (REALTIME) */
2293 int
2294 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2295 {
2296 	struct aiocb **acb_list;
2297 	struct sigevent *sigp, sig;
2298 	int error, nent;
2299 
2300 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2301 		return (EINVAL);
2302 
2303 	nent = uap->nent;
2304 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2305 		return (EINVAL);
2306 
2307 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2308 		error = copyin(uap->sig, &sig, sizeof(sig));
2309 		if (error)
2310 			return (error);
2311 		sigp = &sig;
2312 	} else
2313 		sigp = NULL;
2314 
2315 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2316 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2317 	if (error == 0)
2318 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2319 		    nent, sigp, &aiocb_ops);
2320 	free(acb_list, M_LIO);
2321 	return (error);
2322 }
2323 
2324 static void
2325 aio_physwakeup(struct bio *bp)
2326 {
2327 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2328 	struct proc *userp;
2329 	struct kaioinfo *ki;
2330 	size_t nbytes;
2331 	int error, nblks;
2332 
2333 	/* Release mapping into kernel space. */
2334 	userp = job->userproc;
2335 	ki = userp->p_aioinfo;
2336 	if (job->pbuf) {
2337 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2338 		relpbuf(job->pbuf, NULL);
2339 		job->pbuf = NULL;
2340 		atomic_subtract_int(&num_buf_aio, 1);
2341 		AIO_LOCK(ki);
2342 		ki->kaio_buffer_count--;
2343 		AIO_UNLOCK(ki);
2344 	}
2345 	vm_page_unhold_pages(job->pages, job->npages);
2346 
2347 	bp = job->bp;
2348 	job->bp = NULL;
2349 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2350 	error = 0;
2351 	if (bp->bio_flags & BIO_ERROR)
2352 		error = bp->bio_error;
2353 	nblks = btodb(nbytes);
2354 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2355 		job->outblock += nblks;
2356 	else
2357 		job->inblock += nblks;
2358 
2359 	if (error)
2360 		aio_complete(job, -1, error);
2361 	else
2362 		aio_complete(job, nbytes, 0);
2363 
2364 	g_destroy_bio(bp);
2365 }
2366 
2367 /* syscall - wait for the next completion of an aio request */
2368 static int
2369 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2370     struct timespec *ts, struct aiocb_ops *ops)
2371 {
2372 	struct proc *p = td->td_proc;
2373 	struct timeval atv;
2374 	struct kaioinfo *ki;
2375 	struct kaiocb *job;
2376 	struct aiocb *ujob;
2377 	long error, status;
2378 	int timo;
2379 
2380 	ops->store_aiocb(ujobp, NULL);
2381 
2382 	if (ts == NULL) {
2383 		timo = 0;
2384 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2385 		timo = -1;
2386 	} else {
2387 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2388 			return (EINVAL);
2389 
2390 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2391 		if (itimerfix(&atv))
2392 			return (EINVAL);
2393 		timo = tvtohz(&atv);
2394 	}
2395 
2396 	if (p->p_aioinfo == NULL)
2397 		aio_init_aioinfo(p);
2398 	ki = p->p_aioinfo;
2399 
2400 	error = 0;
2401 	job = NULL;
2402 	AIO_LOCK(ki);
2403 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2404 		if (timo == -1) {
2405 			error = EWOULDBLOCK;
2406 			break;
2407 		}
2408 		ki->kaio_flags |= KAIO_WAKEUP;
2409 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2410 		    "aiowc", timo);
2411 		if (timo && error == ERESTART)
2412 			error = EINTR;
2413 		if (error)
2414 			break;
2415 	}
2416 
2417 	if (job != NULL) {
2418 		MPASS(job->jobflags & KAIOCB_FINISHED);
2419 		ujob = job->ujob;
2420 		status = job->uaiocb._aiocb_private.status;
2421 		error = job->uaiocb._aiocb_private.error;
2422 		td->td_retval[0] = status;
2423 		td->td_ru.ru_oublock += job->outblock;
2424 		td->td_ru.ru_inblock += job->inblock;
2425 		td->td_ru.ru_msgsnd += job->msgsnd;
2426 		td->td_ru.ru_msgrcv += job->msgrcv;
2427 		aio_free_entry(job);
2428 		AIO_UNLOCK(ki);
2429 		ops->store_aiocb(ujobp, ujob);
2430 		ops->store_error(ujob, error);
2431 		ops->store_status(ujob, status);
2432 	} else
2433 		AIO_UNLOCK(ki);
2434 
2435 	return (error);
2436 }
2437 
2438 int
2439 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2440 {
2441 	struct timespec ts, *tsp;
2442 	int error;
2443 
2444 	if (uap->timeout) {
2445 		/* Get timespec struct. */
2446 		error = copyin(uap->timeout, &ts, sizeof(ts));
2447 		if (error)
2448 			return (error);
2449 		tsp = &ts;
2450 	} else
2451 		tsp = NULL;
2452 
2453 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2454 }
2455 
2456 static int
2457 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2458     struct aiocb_ops *ops)
2459 {
2460 	struct proc *p = td->td_proc;
2461 	struct kaioinfo *ki;
2462 
2463 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2464 		return (EINVAL);
2465 	ki = p->p_aioinfo;
2466 	if (ki == NULL)
2467 		aio_init_aioinfo(p);
2468 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2469 }
2470 
2471 int
2472 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2473 {
2474 
2475 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2476 }
2477 
2478 /* kqueue attach function */
2479 static int
2480 filt_aioattach(struct knote *kn)
2481 {
2482 	struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2483 
2484 	/*
2485 	 * The job pointer must be validated before using it, so
2486 	 * registration is restricted to the kernel; the user cannot
2487 	 * set EV_FLAG1.
2488 	 */
2489 	if ((kn->kn_flags & EV_FLAG1) == 0)
2490 		return (EPERM);
2491 	kn->kn_ptr.p_aio = job;
2492 	kn->kn_flags &= ~EV_FLAG1;
2493 
2494 	knlist_add(&job->klist, kn, 0);
2495 
2496 	return (0);
2497 }
2498 
2499 /* kqueue detach function */
2500 static void
2501 filt_aiodetach(struct knote *kn)
2502 {
2503 	struct knlist *knl;
2504 
2505 	knl = &kn->kn_ptr.p_aio->klist;
2506 	knl->kl_lock(knl->kl_lockarg);
2507 	if (!knlist_empty(knl))
2508 		knlist_remove(knl, kn, 1);
2509 	knl->kl_unlock(knl->kl_lockarg);
2510 }
2511 
2512 /* kqueue filter function */
2513 /*ARGSUSED*/
2514 static int
2515 filt_aio(struct knote *kn, long hint)
2516 {
2517 	struct kaiocb *job = kn->kn_ptr.p_aio;
2518 
2519 	kn->kn_data = job->uaiocb._aiocb_private.error;
2520 	if (!(job->jobflags & KAIOCB_FINISHED))
2521 		return (0);
2522 	kn->kn_flags |= EV_EOF;
2523 	return (1);
2524 }
2525 
2526 /* kqueue attach function */
2527 static int
2528 filt_lioattach(struct knote *kn)
2529 {
2530 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2531 
2532 	/*
2533 	 * The aioliojob pointer must be validated before using it, so
2534 	 * registration is restricted to the kernel; the user cannot
2535 	 * set EV_FLAG1.
2536 	 */
2537 	if ((kn->kn_flags & EV_FLAG1) == 0)
2538 		return (EPERM);
2539 	kn->kn_ptr.p_lio = lj;
2540 	kn->kn_flags &= ~EV_FLAG1;
2541 
2542 	knlist_add(&lj->klist, kn, 0);
2543 
2544 	return (0);
2545 }
2546 
2547 /* kqueue detach function */
2548 static void
2549 filt_liodetach(struct knote *kn)
2550 {
2551 	struct knlist *knl;
2552 
2553 	knl = &kn->kn_ptr.p_lio->klist;
2554 	knl->kl_lock(knl->kl_lockarg);
2555 	if (!knlist_empty(knl))
2556 		knlist_remove(knl, kn, 1);
2557 	knl->kl_unlock(knl->kl_lockarg);
2558 }
2559 
2560 /* kqueue filter function */
2561 /*ARGSUSED*/
2562 static int
2563 filt_lio(struct knote *kn, long hint)
2564 {
2565 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2566 
2567 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2568 }
2569 
2570 #ifdef COMPAT_FREEBSD32
2571 #include <sys/mount.h>
2572 #include <sys/socket.h>
2573 #include <compat/freebsd32/freebsd32.h>
2574 #include <compat/freebsd32/freebsd32_proto.h>
2575 #include <compat/freebsd32/freebsd32_signal.h>
2576 #include <compat/freebsd32/freebsd32_syscall.h>
2577 #include <compat/freebsd32/freebsd32_util.h>
2578 
2579 struct __aiocb_private32 {
2580 	int32_t	status;
2581 	int32_t	error;
2582 	uint32_t kernelinfo;
2583 };
2584 
2585 #ifdef COMPAT_FREEBSD6
2586 typedef struct oaiocb32 {
2587 	int	aio_fildes;		/* File descriptor */
2588 	uint64_t aio_offset __packed;	/* File offset for I/O */
2589 	uint32_t aio_buf;		/* I/O buffer in process space */
2590 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2591 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2592 	int	aio_lio_opcode;		/* LIO opcode */
2593 	int	aio_reqprio;		/* Request priority -- ignored */
2594 	struct	__aiocb_private32 _aiocb_private;
2595 } oaiocb32_t;
2596 #endif
2597 
2598 typedef struct aiocb32 {
2599 	int32_t	aio_fildes;		/* File descriptor */
2600 	uint64_t aio_offset __packed;	/* File offset for I/O */
2601 	uint32_t aio_buf;		/* I/O buffer in process space */
2602 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2603 	int	__spare__[2];
2604 	uint32_t __spare2__;
2605 	int	aio_lio_opcode;		/* LIO opcode */
2606 	int	aio_reqprio;		/* Request priority -- ignored */
2607 	struct	__aiocb_private32 _aiocb_private;
2608 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2609 } aiocb32_t;
2610 
2611 #ifdef COMPAT_FREEBSD6
2612 static int
2613 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2614 {
2615 
2616 	/*
2617 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2618 	 * supported by AIO with the old sigevent structure.
2619 	 */
2620 	CP(*osig, *nsig, sigev_notify);
2621 	switch (nsig->sigev_notify) {
2622 	case SIGEV_NONE:
2623 		break;
2624 	case SIGEV_SIGNAL:
2625 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2626 		break;
2627 	case SIGEV_KEVENT:
2628 		nsig->sigev_notify_kqueue =
2629 		    osig->__sigev_u.__sigev_notify_kqueue;
2630 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2631 		break;
2632 	default:
2633 		return (EINVAL);
2634 	}
2635 	return (0);
2636 }
2637 
2638 static int
2639 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2640 {
2641 	struct oaiocb32 job32;
2642 	int error;
2643 
2644 	bzero(kjob, sizeof(struct aiocb));
2645 	error = copyin(ujob, &job32, sizeof(job32));
2646 	if (error)
2647 		return (error);
2648 
2649 	CP(job32, *kjob, aio_fildes);
2650 	CP(job32, *kjob, aio_offset);
2651 	PTRIN_CP(job32, *kjob, aio_buf);
2652 	CP(job32, *kjob, aio_nbytes);
2653 	CP(job32, *kjob, aio_lio_opcode);
2654 	CP(job32, *kjob, aio_reqprio);
2655 	CP(job32, *kjob, _aiocb_private.status);
2656 	CP(job32, *kjob, _aiocb_private.error);
2657 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2658 	return (convert_old_sigevent32(&job32.aio_sigevent,
2659 	    &kjob->aio_sigevent));
2660 }
2661 #endif
2662 
2663 static int
2664 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2665 {
2666 	struct aiocb32 job32;
2667 	int error;
2668 
2669 	error = copyin(ujob, &job32, sizeof(job32));
2670 	if (error)
2671 		return (error);
2672 	CP(job32, *kjob, aio_fildes);
2673 	CP(job32, *kjob, aio_offset);
2674 	PTRIN_CP(job32, *kjob, aio_buf);
2675 	CP(job32, *kjob, aio_nbytes);
2676 	CP(job32, *kjob, aio_lio_opcode);
2677 	CP(job32, *kjob, aio_reqprio);
2678 	CP(job32, *kjob, _aiocb_private.status);
2679 	CP(job32, *kjob, _aiocb_private.error);
2680 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2681 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2682 }
2683 
2684 static long
2685 aiocb32_fetch_status(struct aiocb *ujob)
2686 {
2687 	struct aiocb32 *ujob32;
2688 
2689 	ujob32 = (struct aiocb32 *)ujob;
2690 	return (fuword32(&ujob32->_aiocb_private.status));
2691 }
2692 
2693 static long
2694 aiocb32_fetch_error(struct aiocb *ujob)
2695 {
2696 	struct aiocb32 *ujob32;
2697 
2698 	ujob32 = (struct aiocb32 *)ujob;
2699 	return (fuword32(&ujob32->_aiocb_private.error));
2700 }
2701 
2702 static int
2703 aiocb32_store_status(struct aiocb *ujob, long status)
2704 {
2705 	struct aiocb32 *ujob32;
2706 
2707 	ujob32 = (struct aiocb32 *)ujob;
2708 	return (suword32(&ujob32->_aiocb_private.status, status));
2709 }
2710 
2711 static int
2712 aiocb32_store_error(struct aiocb *ujob, long error)
2713 {
2714 	struct aiocb32 *ujob32;
2715 
2716 	ujob32 = (struct aiocb32 *)ujob;
2717 	return (suword32(&ujob32->_aiocb_private.error, error));
2718 }
2719 
2720 static int
2721 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2722 {
2723 	struct aiocb32 *ujob32;
2724 
2725 	ujob32 = (struct aiocb32 *)ujob;
2726 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2727 }
2728 
2729 static int
2730 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2731 {
2732 
2733 	return (suword32(ujobp, (long)ujob));
2734 }
2735 
2736 static struct aiocb_ops aiocb32_ops = {
2737 	.copyin = aiocb32_copyin,
2738 	.fetch_status = aiocb32_fetch_status,
2739 	.fetch_error = aiocb32_fetch_error,
2740 	.store_status = aiocb32_store_status,
2741 	.store_error = aiocb32_store_error,
2742 	.store_kernelinfo = aiocb32_store_kernelinfo,
2743 	.store_aiocb = aiocb32_store_aiocb,
2744 };
2745 
2746 #ifdef COMPAT_FREEBSD6
2747 static struct aiocb_ops aiocb32_ops_osigevent = {
2748 	.copyin = aiocb32_copyin_old_sigevent,
2749 	.fetch_status = aiocb32_fetch_status,
2750 	.fetch_error = aiocb32_fetch_error,
2751 	.store_status = aiocb32_store_status,
2752 	.store_error = aiocb32_store_error,
2753 	.store_kernelinfo = aiocb32_store_kernelinfo,
2754 	.store_aiocb = aiocb32_store_aiocb,
2755 };
2756 #endif
2757 
2758 int
2759 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2760 {
2761 
2762 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2763 }
2764 
2765 int
2766 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2767 {
2768 	struct timespec32 ts32;
2769 	struct timespec ts, *tsp;
2770 	struct aiocb **ujoblist;
2771 	uint32_t *ujoblist32;
2772 	int error, i;
2773 
2774 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2775 		return (EINVAL);
2776 
2777 	if (uap->timeout) {
2778 		/* Get timespec struct. */
2779 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2780 			return (error);
2781 		CP(ts32, ts, tv_sec);
2782 		CP(ts32, ts, tv_nsec);
2783 		tsp = &ts;
2784 	} else
2785 		tsp = NULL;
2786 
2787 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2788 	ujoblist32 = (uint32_t *)ujoblist;
2789 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2790 	    sizeof(ujoblist32[0]));
2791 	if (error == 0) {
2792 		for (i = uap->nent; i > 0; i--)
2793 			ujoblist[i] = PTRIN(ujoblist32[i]);
2794 
2795 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2796 	}
2797 	uma_zfree(aiol_zone, ujoblist);
2798 	return (error);
2799 }
2800 
2801 int
2802 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2803 {
2804 
2805 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2806 }
2807 
2808 #ifdef COMPAT_FREEBSD6
2809 int
2810 freebsd6_freebsd32_aio_read(struct thread *td,
2811     struct freebsd6_freebsd32_aio_read_args *uap)
2812 {
2813 
2814 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2815 	    &aiocb32_ops_osigevent));
2816 }
2817 #endif
2818 
2819 int
2820 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2821 {
2822 
2823 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2824 	    &aiocb32_ops));
2825 }
2826 
2827 #ifdef COMPAT_FREEBSD6
2828 int
2829 freebsd6_freebsd32_aio_write(struct thread *td,
2830     struct freebsd6_freebsd32_aio_write_args *uap)
2831 {
2832 
2833 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2834 	    &aiocb32_ops_osigevent));
2835 }
2836 #endif
2837 
2838 int
2839 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2840 {
2841 
2842 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2843 	    &aiocb32_ops));
2844 }
2845 
2846 int
2847 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2848 {
2849 
2850 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2851 	    &aiocb32_ops));
2852 }
2853 
2854 int
2855 freebsd32_aio_waitcomplete(struct thread *td,
2856     struct freebsd32_aio_waitcomplete_args *uap)
2857 {
2858 	struct timespec32 ts32;
2859 	struct timespec ts, *tsp;
2860 	int error;
2861 
2862 	if (uap->timeout) {
2863 		/* Get timespec struct. */
2864 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2865 		if (error)
2866 			return (error);
2867 		CP(ts32, ts, tv_sec);
2868 		CP(ts32, ts, tv_nsec);
2869 		tsp = &ts;
2870 	} else
2871 		tsp = NULL;
2872 
2873 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2874 	    &aiocb32_ops));
2875 }
2876 
2877 int
2878 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2879 {
2880 
2881 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2882 	    &aiocb32_ops));
2883 }
2884 
2885 #ifdef COMPAT_FREEBSD6
2886 int
2887 freebsd6_freebsd32_lio_listio(struct thread *td,
2888     struct freebsd6_freebsd32_lio_listio_args *uap)
2889 {
2890 	struct aiocb **acb_list;
2891 	struct sigevent *sigp, sig;
2892 	struct osigevent32 osig;
2893 	uint32_t *acb_list32;
2894 	int error, i, nent;
2895 
2896 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2897 		return (EINVAL);
2898 
2899 	nent = uap->nent;
2900 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2901 		return (EINVAL);
2902 
2903 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2904 		error = copyin(uap->sig, &osig, sizeof(osig));
2905 		if (error)
2906 			return (error);
2907 		error = convert_old_sigevent32(&osig, &sig);
2908 		if (error)
2909 			return (error);
2910 		sigp = &sig;
2911 	} else
2912 		sigp = NULL;
2913 
2914 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2915 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2916 	if (error) {
2917 		free(acb_list32, M_LIO);
2918 		return (error);
2919 	}
2920 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2921 	for (i = 0; i < nent; i++)
2922 		acb_list[i] = PTRIN(acb_list32[i]);
2923 	free(acb_list32, M_LIO);
2924 
2925 	error = kern_lio_listio(td, uap->mode,
2926 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2927 	    &aiocb32_ops_osigevent);
2928 	free(acb_list, M_LIO);
2929 	return (error);
2930 }
2931 #endif
2932 
2933 int
2934 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2935 {
2936 	struct aiocb **acb_list;
2937 	struct sigevent *sigp, sig;
2938 	struct sigevent32 sig32;
2939 	uint32_t *acb_list32;
2940 	int error, i, nent;
2941 
2942 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2943 		return (EINVAL);
2944 
2945 	nent = uap->nent;
2946 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2947 		return (EINVAL);
2948 
2949 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2950 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2951 		if (error)
2952 			return (error);
2953 		error = convert_sigevent32(&sig32, &sig);
2954 		if (error)
2955 			return (error);
2956 		sigp = &sig;
2957 	} else
2958 		sigp = NULL;
2959 
2960 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2961 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2962 	if (error) {
2963 		free(acb_list32, M_LIO);
2964 		return (error);
2965 	}
2966 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2967 	for (i = 0; i < nent; i++)
2968 		acb_list[i] = PTRIN(acb_list32[i]);
2969 	free(acb_list32, M_LIO);
2970 
2971 	error = kern_lio_listio(td, uap->mode,
2972 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2973 	    &aiocb32_ops);
2974 	free(acb_list, M_LIO);
2975 	return (error);
2976 }
2977 
2978 #endif
2979