xref: /freebsd/sys/kern/vfs_aio.c (revision cbd30a72ca196976c1c700400ecd424baa1b9c16)
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/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 /* Bigger than AIO_LISTIO_MAX */
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE
97 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
98 #endif
99 
100 #ifndef MAX_BUF_AIO
101 #define MAX_BUF_AIO		16
102 #endif
103 
104 FEATURE(aio, "Asynchronous I/O");
105 
106 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
107 
108 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
109     "Async IO management");
110 
111 static int enable_aio_unsafe = 0;
112 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
113     "Permit asynchronous IO on all file types, not just known-safe types");
114 
115 static unsigned int unsafe_warningcnt = 1;
116 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
117     &unsafe_warningcnt, 0,
118     "Warnings that will be triggered upon failed IO requests on unsafe files");
119 
120 static int max_aio_procs = MAX_AIO_PROCS;
121 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
122     "Maximum number of kernel processes to use for handling async IO ");
123 
124 static int num_aio_procs = 0;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
126     "Number of presently active kernel processes for async IO");
127 
128 /*
129  * The code will adjust the actual number of AIO processes towards this
130  * number when it gets a chance.
131  */
132 static int target_aio_procs = TARGET_AIO_PROCS;
133 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
134     0,
135     "Preferred number of ready kernel processes for async IO");
136 
137 static int max_queue_count = MAX_AIO_QUEUE;
138 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
139     "Maximum number of aio requests to queue, globally");
140 
141 static int num_queue_count = 0;
142 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
143     "Number of queued aio requests");
144 
145 static int num_buf_aio = 0;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
147     "Number of aio requests presently handled by the buf subsystem");
148 
149 /* Number of async I/O processes in the process of being started */
150 /* XXX This should be local to aio_aqueue() */
151 static int num_aio_resv_start = 0;
152 
153 static int aiod_lifetime;
154 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
155     "Maximum lifetime for idle aiod");
156 
157 static int max_aio_per_proc = MAX_AIO_PER_PROC;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
159     0,
160     "Maximum active aio requests per process (stored in the process)");
161 
162 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
163 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
164     &max_aio_queue_per_proc, 0,
165     "Maximum queued aio requests per process (stored in the process)");
166 
167 static int max_buf_aio = MAX_BUF_AIO;
168 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
169     "Maximum buf aio requests per process (stored in the process)");
170 
171 #ifdef COMPAT_FREEBSD6
172 typedef struct oaiocb {
173 	int	aio_fildes;		/* File descriptor */
174 	off_t	aio_offset;		/* File offset for I/O */
175 	volatile void *aio_buf;         /* I/O buffer in process space */
176 	size_t	aio_nbytes;		/* Number of bytes for I/O */
177 	struct	osigevent aio_sigevent;	/* Signal to deliver */
178 	int	aio_lio_opcode;		/* LIO opcode */
179 	int	aio_reqprio;		/* Request priority -- ignored */
180 	struct	__aiocb_private	_aiocb_private;
181 } oaiocb_t;
182 #endif
183 
184 /*
185  * Below is a key of locks used to protect each member of struct kaiocb
186  * aioliojob and kaioinfo and any backends.
187  *
188  * * - need not protected
189  * a - locked by kaioinfo lock
190  * b - locked by backend lock, the backend lock can be null in some cases,
191  *     for example, BIO belongs to this type, in this case, proc lock is
192  *     reused.
193  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
194  */
195 
196 /*
197  * If the routine that services an AIO request blocks while running in an
198  * AIO kernel process it can starve other I/O requests.  BIO requests
199  * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
200  * processes at all.  Socket I/O requests use a separate pool of
201  * kprocs and also force non-blocking I/O.  Other file I/O requests
202  * use the generic fo_read/fo_write operations which can block.  The
203  * fsync and mlock operations can also block while executing.  Ideally
204  * none of these requests would block while executing.
205  *
206  * Note that the service routines cannot toggle O_NONBLOCK in the file
207  * structure directly while handling a request due to races with
208  * userland threads.
209  */
210 
211 /* jobflags */
212 #define	KAIOCB_QUEUEING		0x01
213 #define	KAIOCB_CANCELLED	0x02
214 #define	KAIOCB_CANCELLING	0x04
215 #define	KAIOCB_CHECKSYNC	0x08
216 #define	KAIOCB_CLEARED		0x10
217 #define	KAIOCB_FINISHED		0x20
218 
219 /*
220  * AIO process info
221  */
222 #define AIOP_FREE	0x1			/* proc on free queue */
223 
224 struct aioproc {
225 	int	aioprocflags;			/* (c) AIO proc flags */
226 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
227 	struct	proc *aioproc;			/* (*) the AIO proc */
228 };
229 
230 /*
231  * data-structure for lio signal management
232  */
233 struct aioliojob {
234 	int	lioj_flags;			/* (a) listio flags */
235 	int	lioj_count;			/* (a) listio flags */
236 	int	lioj_finished_count;		/* (a) listio flags */
237 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
238 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
239 	struct	knlist klist;			/* (a) list of knotes */
240 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
241 };
242 
243 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
244 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
245 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
246 
247 /*
248  * per process aio data structure
249  */
250 struct kaioinfo {
251 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
252 	int	kaio_flags;		/* (a) per process kaio flags */
253 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
254 	int	kaio_active_count;	/* (c) number of currently used AIOs */
255 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
256 	int	kaio_count;		/* (a) size of AIO queue */
257 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
258 	int	kaio_buffer_count;	/* (a) number of physio buffers */
259 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
260 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
261 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
262 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
263 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
264 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
265 	struct	task kaio_task;		/* (*) task to kick aio processes */
266 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
267 };
268 
269 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
270 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
271 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
272 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
273 
274 #define KAIO_RUNDOWN	0x1	/* process is being run down */
275 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
276 
277 /*
278  * Operations used to interact with userland aio control blocks.
279  * Different ABIs provide their own operations.
280  */
281 struct aiocb_ops {
282 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
283 	long	(*fetch_status)(struct aiocb *ujob);
284 	long	(*fetch_error)(struct aiocb *ujob);
285 	int	(*store_status)(struct aiocb *ujob, long status);
286 	int	(*store_error)(struct aiocb *ujob, long error);
287 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
288 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
289 };
290 
291 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
292 static struct sema aio_newproc_sem;
293 static struct mtx aio_job_mtx;
294 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
295 static struct unrhdr *aiod_unr;
296 
297 void		aio_init_aioinfo(struct proc *p);
298 static int	aio_onceonly(void);
299 static int	aio_free_entry(struct kaiocb *job);
300 static void	aio_process_rw(struct kaiocb *job);
301 static void	aio_process_sync(struct kaiocb *job);
302 static void	aio_process_mlock(struct kaiocb *job);
303 static void	aio_schedule_fsync(void *context, int pending);
304 static int	aio_newproc(int *);
305 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
306 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
307 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
308 static void	aio_physwakeup(struct bio *bp);
309 static void	aio_proc_rundown(void *arg, struct proc *p);
310 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
311 		    struct image_params *imgp);
312 static int	aio_qphysio(struct proc *p, struct kaiocb *job);
313 static void	aio_daemon(void *param);
314 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
315 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
316 static int	aio_kick(struct proc *userp);
317 static void	aio_kick_nowait(struct proc *userp);
318 static void	aio_kick_helper(void *context, int pending);
319 static int	filt_aioattach(struct knote *kn);
320 static void	filt_aiodetach(struct knote *kn);
321 static int	filt_aio(struct knote *kn, long hint);
322 static int	filt_lioattach(struct knote *kn);
323 static void	filt_liodetach(struct knote *kn);
324 static int	filt_lio(struct knote *kn, long hint);
325 
326 /*
327  * Zones for:
328  * 	kaio	Per process async io info
329  *	aiop	async io process data
330  *	aiocb	async io jobs
331  *	aiol	list io job pointer - internal to aio_suspend XXX
332  *	aiolio	list io jobs
333  */
334 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
335 
336 /* kqueue filters for aio */
337 static struct filterops aio_filtops = {
338 	.f_isfd = 0,
339 	.f_attach = filt_aioattach,
340 	.f_detach = filt_aiodetach,
341 	.f_event = filt_aio,
342 };
343 static struct filterops lio_filtops = {
344 	.f_isfd = 0,
345 	.f_attach = filt_lioattach,
346 	.f_detach = filt_liodetach,
347 	.f_event = filt_lio
348 };
349 
350 static eventhandler_tag exit_tag, exec_tag;
351 
352 TASKQUEUE_DEFINE_THREAD(aiod_kick);
353 
354 /*
355  * Main operations function for use as a kernel module.
356  */
357 static int
358 aio_modload(struct module *module, int cmd, void *arg)
359 {
360 	int error = 0;
361 
362 	switch (cmd) {
363 	case MOD_LOAD:
364 		aio_onceonly();
365 		break;
366 	case MOD_SHUTDOWN:
367 		break;
368 	default:
369 		error = EOPNOTSUPP;
370 		break;
371 	}
372 	return (error);
373 }
374 
375 static moduledata_t aio_mod = {
376 	"aio",
377 	&aio_modload,
378 	NULL
379 };
380 
381 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
382 MODULE_VERSION(aio, 1);
383 
384 /*
385  * Startup initialization
386  */
387 static int
388 aio_onceonly(void)
389 {
390 
391 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
392 	    EVENTHANDLER_PRI_ANY);
393 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
394 	    NULL, EVENTHANDLER_PRI_ANY);
395 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
396 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
397 	TAILQ_INIT(&aio_freeproc);
398 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
399 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
400 	TAILQ_INIT(&aio_jobs);
401 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
402 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
403 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
404 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
405 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
406 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
407 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
408 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
409 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
410 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
411 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
412 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
413 	jobrefid = 1;
414 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
415 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
416 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
417 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
418 
419 	return (0);
420 }
421 
422 /*
423  * Init the per-process aioinfo structure.  The aioinfo limits are set
424  * per-process for user limit (resource) management.
425  */
426 void
427 aio_init_aioinfo(struct proc *p)
428 {
429 	struct kaioinfo *ki;
430 
431 	ki = uma_zalloc(kaio_zone, M_WAITOK);
432 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
433 	ki->kaio_flags = 0;
434 	ki->kaio_maxactive_count = max_aio_per_proc;
435 	ki->kaio_active_count = 0;
436 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
437 	ki->kaio_count = 0;
438 	ki->kaio_ballowed_count = max_buf_aio;
439 	ki->kaio_buffer_count = 0;
440 	TAILQ_INIT(&ki->kaio_all);
441 	TAILQ_INIT(&ki->kaio_done);
442 	TAILQ_INIT(&ki->kaio_jobqueue);
443 	TAILQ_INIT(&ki->kaio_liojoblist);
444 	TAILQ_INIT(&ki->kaio_syncqueue);
445 	TAILQ_INIT(&ki->kaio_syncready);
446 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
447 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
448 	PROC_LOCK(p);
449 	if (p->p_aioinfo == NULL) {
450 		p->p_aioinfo = ki;
451 		PROC_UNLOCK(p);
452 	} else {
453 		PROC_UNLOCK(p);
454 		mtx_destroy(&ki->kaio_mtx);
455 		uma_zfree(kaio_zone, ki);
456 	}
457 
458 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
459 		aio_newproc(NULL);
460 }
461 
462 static int
463 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
464 {
465 	struct thread *td;
466 	int error;
467 
468 	error = sigev_findtd(p, sigev, &td);
469 	if (error)
470 		return (error);
471 	if (!KSI_ONQ(ksi)) {
472 		ksiginfo_set_sigev(ksi, sigev);
473 		ksi->ksi_code = SI_ASYNCIO;
474 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
475 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
476 	}
477 	PROC_UNLOCK(p);
478 	return (error);
479 }
480 
481 /*
482  * Free a job entry.  Wait for completion if it is currently active, but don't
483  * delay forever.  If we delay, we return a flag that says that we have to
484  * restart the queue scan.
485  */
486 static int
487 aio_free_entry(struct kaiocb *job)
488 {
489 	struct kaioinfo *ki;
490 	struct aioliojob *lj;
491 	struct proc *p;
492 
493 	p = job->userproc;
494 	MPASS(curproc == p);
495 	ki = p->p_aioinfo;
496 	MPASS(ki != NULL);
497 
498 	AIO_LOCK_ASSERT(ki, MA_OWNED);
499 	MPASS(job->jobflags & KAIOCB_FINISHED);
500 
501 	atomic_subtract_int(&num_queue_count, 1);
502 
503 	ki->kaio_count--;
504 	MPASS(ki->kaio_count >= 0);
505 
506 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
507 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
508 
509 	lj = job->lio;
510 	if (lj) {
511 		lj->lioj_count--;
512 		lj->lioj_finished_count--;
513 
514 		if (lj->lioj_count == 0) {
515 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
516 			/* lio is going away, we need to destroy any knotes */
517 			knlist_delete(&lj->klist, curthread, 1);
518 			PROC_LOCK(p);
519 			sigqueue_take(&lj->lioj_ksi);
520 			PROC_UNLOCK(p);
521 			uma_zfree(aiolio_zone, lj);
522 		}
523 	}
524 
525 	/* job is going away, we need to destroy any knotes */
526 	knlist_delete(&job->klist, curthread, 1);
527 	PROC_LOCK(p);
528 	sigqueue_take(&job->ksi);
529 	PROC_UNLOCK(p);
530 
531 	AIO_UNLOCK(ki);
532 
533 	/*
534 	 * The thread argument here is used to find the owning process
535 	 * and is also passed to fo_close() which may pass it to various
536 	 * places such as devsw close() routines.  Because of that, we
537 	 * need a thread pointer from the process owning the job that is
538 	 * persistent and won't disappear out from under us or move to
539 	 * another process.
540 	 *
541 	 * Currently, all the callers of this function call it to remove
542 	 * a kaiocb from the current process' job list either via a
543 	 * syscall or due to the current process calling exit() or
544 	 * execve().  Thus, we know that p == curproc.  We also know that
545 	 * curthread can't exit since we are curthread.
546 	 *
547 	 * Therefore, we use curthread as the thread to pass to
548 	 * knlist_delete().  This does mean that it is possible for the
549 	 * thread pointer at close time to differ from the thread pointer
550 	 * at open time, but this is already true of file descriptors in
551 	 * a multithreaded process.
552 	 */
553 	if (job->fd_file)
554 		fdrop(job->fd_file, curthread);
555 	crfree(job->cred);
556 	uma_zfree(aiocb_zone, job);
557 	AIO_LOCK(ki);
558 
559 	return (0);
560 }
561 
562 static void
563 aio_proc_rundown_exec(void *arg, struct proc *p,
564     struct image_params *imgp __unused)
565 {
566    	aio_proc_rundown(arg, p);
567 }
568 
569 static int
570 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
571 {
572 	aio_cancel_fn_t *func;
573 	int cancelled;
574 
575 	AIO_LOCK_ASSERT(ki, MA_OWNED);
576 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
577 		return (0);
578 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
579 	job->jobflags |= KAIOCB_CANCELLED;
580 
581 	func = job->cancel_fn;
582 
583 	/*
584 	 * If there is no cancel routine, just leave the job marked as
585 	 * cancelled.  The job should be in active use by a caller who
586 	 * should complete it normally or when it fails to install a
587 	 * cancel routine.
588 	 */
589 	if (func == NULL)
590 		return (0);
591 
592 	/*
593 	 * Set the CANCELLING flag so that aio_complete() will defer
594 	 * completions of this job.  This prevents the job from being
595 	 * freed out from under the cancel callback.  After the
596 	 * callback any deferred completion (whether from the callback
597 	 * or any other source) will be completed.
598 	 */
599 	job->jobflags |= KAIOCB_CANCELLING;
600 	AIO_UNLOCK(ki);
601 	func(job);
602 	AIO_LOCK(ki);
603 	job->jobflags &= ~KAIOCB_CANCELLING;
604 	if (job->jobflags & KAIOCB_FINISHED) {
605 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
606 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
607 		aio_bio_done_notify(p, job);
608 	} else {
609 		/*
610 		 * The cancel callback might have scheduled an
611 		 * operation to cancel this request, but it is
612 		 * only counted as cancelled if the request is
613 		 * cancelled when the callback returns.
614 		 */
615 		cancelled = 0;
616 	}
617 	return (cancelled);
618 }
619 
620 /*
621  * Rundown the jobs for a given process.
622  */
623 static void
624 aio_proc_rundown(void *arg, struct proc *p)
625 {
626 	struct kaioinfo *ki;
627 	struct aioliojob *lj;
628 	struct kaiocb *job, *jobn;
629 
630 	KASSERT(curthread->td_proc == p,
631 	    ("%s: called on non-curproc", __func__));
632 	ki = p->p_aioinfo;
633 	if (ki == NULL)
634 		return;
635 
636 	AIO_LOCK(ki);
637 	ki->kaio_flags |= KAIO_RUNDOWN;
638 
639 restart:
640 
641 	/*
642 	 * Try to cancel all pending requests. This code simulates
643 	 * aio_cancel on all pending I/O requests.
644 	 */
645 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
646 		aio_cancel_job(p, ki, job);
647 	}
648 
649 	/* Wait for all running I/O to be finished */
650 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
651 		ki->kaio_flags |= KAIO_WAKEUP;
652 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
653 		goto restart;
654 	}
655 
656 	/* Free all completed I/O requests. */
657 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
658 		aio_free_entry(job);
659 
660 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
661 		if (lj->lioj_count == 0) {
662 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
663 			knlist_delete(&lj->klist, curthread, 1);
664 			PROC_LOCK(p);
665 			sigqueue_take(&lj->lioj_ksi);
666 			PROC_UNLOCK(p);
667 			uma_zfree(aiolio_zone, lj);
668 		} else {
669 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
670 			    lj->lioj_count, lj->lioj_finished_count);
671 		}
672 	}
673 	AIO_UNLOCK(ki);
674 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
675 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
676 	mtx_destroy(&ki->kaio_mtx);
677 	uma_zfree(kaio_zone, ki);
678 	p->p_aioinfo = NULL;
679 }
680 
681 /*
682  * Select a job to run (called by an AIO daemon).
683  */
684 static struct kaiocb *
685 aio_selectjob(struct aioproc *aiop)
686 {
687 	struct kaiocb *job;
688 	struct kaioinfo *ki;
689 	struct proc *userp;
690 
691 	mtx_assert(&aio_job_mtx, MA_OWNED);
692 restart:
693 	TAILQ_FOREACH(job, &aio_jobs, list) {
694 		userp = job->userproc;
695 		ki = userp->p_aioinfo;
696 
697 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
698 			TAILQ_REMOVE(&aio_jobs, job, list);
699 			if (!aio_clear_cancel_function(job))
700 				goto restart;
701 
702 			/* Account for currently active jobs. */
703 			ki->kaio_active_count++;
704 			break;
705 		}
706 	}
707 	return (job);
708 }
709 
710 /*
711  * Move all data to a permanent storage device.  This code
712  * simulates the fsync syscall.
713  */
714 static int
715 aio_fsync_vnode(struct thread *td, struct vnode *vp)
716 {
717 	struct mount *mp;
718 	int error;
719 
720 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
721 		goto drop;
722 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
723 	if (vp->v_object != NULL) {
724 		VM_OBJECT_WLOCK(vp->v_object);
725 		vm_object_page_clean(vp->v_object, 0, 0, 0);
726 		VM_OBJECT_WUNLOCK(vp->v_object);
727 	}
728 	error = VOP_FSYNC(vp, MNT_WAIT, td);
729 
730 	VOP_UNLOCK(vp, 0);
731 	vn_finished_write(mp);
732 drop:
733 	return (error);
734 }
735 
736 /*
737  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
738  * does the I/O request for the non-physio version of the operations.  The
739  * normal vn operations are used, and this code should work in all instances
740  * for every type of file, including pipes, sockets, fifos, and regular files.
741  *
742  * XXX I don't think it works well for socket, pipe, and fifo.
743  */
744 static void
745 aio_process_rw(struct kaiocb *job)
746 {
747 	struct ucred *td_savedcred;
748 	struct thread *td;
749 	struct aiocb *cb;
750 	struct file *fp;
751 	struct uio auio;
752 	struct iovec aiov;
753 	ssize_t cnt;
754 	long msgsnd_st, msgsnd_end;
755 	long msgrcv_st, msgrcv_end;
756 	long oublock_st, oublock_end;
757 	long inblock_st, inblock_end;
758 	int error;
759 
760 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
761 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
762 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
763 
764 	aio_switch_vmspace(job);
765 	td = curthread;
766 	td_savedcred = td->td_ucred;
767 	td->td_ucred = job->cred;
768 	cb = &job->uaiocb;
769 	fp = job->fd_file;
770 
771 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
772 	aiov.iov_len = cb->aio_nbytes;
773 
774 	auio.uio_iov = &aiov;
775 	auio.uio_iovcnt = 1;
776 	auio.uio_offset = cb->aio_offset;
777 	auio.uio_resid = cb->aio_nbytes;
778 	cnt = cb->aio_nbytes;
779 	auio.uio_segflg = UIO_USERSPACE;
780 	auio.uio_td = td;
781 
782 	msgrcv_st = td->td_ru.ru_msgrcv;
783 	msgsnd_st = td->td_ru.ru_msgsnd;
784 	inblock_st = td->td_ru.ru_inblock;
785 	oublock_st = td->td_ru.ru_oublock;
786 
787 	/*
788 	 * aio_aqueue() acquires a reference to the file that is
789 	 * released in aio_free_entry().
790 	 */
791 	if (cb->aio_lio_opcode == LIO_READ) {
792 		auio.uio_rw = UIO_READ;
793 		if (auio.uio_resid == 0)
794 			error = 0;
795 		else
796 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
797 	} else {
798 		if (fp->f_type == DTYPE_VNODE)
799 			bwillwrite();
800 		auio.uio_rw = UIO_WRITE;
801 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
802 	}
803 	msgrcv_end = td->td_ru.ru_msgrcv;
804 	msgsnd_end = td->td_ru.ru_msgsnd;
805 	inblock_end = td->td_ru.ru_inblock;
806 	oublock_end = td->td_ru.ru_oublock;
807 
808 	job->msgrcv = msgrcv_end - msgrcv_st;
809 	job->msgsnd = msgsnd_end - msgsnd_st;
810 	job->inblock = inblock_end - inblock_st;
811 	job->outblock = oublock_end - oublock_st;
812 
813 	if ((error) && (auio.uio_resid != cnt)) {
814 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
815 			error = 0;
816 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
817 			PROC_LOCK(job->userproc);
818 			kern_psignal(job->userproc, SIGPIPE);
819 			PROC_UNLOCK(job->userproc);
820 		}
821 	}
822 
823 	cnt -= auio.uio_resid;
824 	td->td_ucred = td_savedcred;
825 	if (error)
826 		aio_complete(job, -1, error);
827 	else
828 		aio_complete(job, cnt, 0);
829 }
830 
831 static void
832 aio_process_sync(struct kaiocb *job)
833 {
834 	struct thread *td = curthread;
835 	struct ucred *td_savedcred = td->td_ucred;
836 	struct file *fp = job->fd_file;
837 	int error = 0;
838 
839 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
840 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
841 
842 	td->td_ucred = job->cred;
843 	if (fp->f_vnode != NULL)
844 		error = aio_fsync_vnode(td, fp->f_vnode);
845 	td->td_ucred = td_savedcred;
846 	if (error)
847 		aio_complete(job, -1, error);
848 	else
849 		aio_complete(job, 0, 0);
850 }
851 
852 static void
853 aio_process_mlock(struct kaiocb *job)
854 {
855 	struct aiocb *cb = &job->uaiocb;
856 	int error;
857 
858 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
859 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
860 
861 	aio_switch_vmspace(job);
862 	error = kern_mlock(job->userproc, job->cred,
863 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
864 	aio_complete(job, error != 0 ? -1 : 0, error);
865 }
866 
867 static void
868 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
869 {
870 	struct aioliojob *lj;
871 	struct kaioinfo *ki;
872 	struct kaiocb *sjob, *sjobn;
873 	int lj_done;
874 	bool schedule_fsync;
875 
876 	ki = userp->p_aioinfo;
877 	AIO_LOCK_ASSERT(ki, MA_OWNED);
878 	lj = job->lio;
879 	lj_done = 0;
880 	if (lj) {
881 		lj->lioj_finished_count++;
882 		if (lj->lioj_count == lj->lioj_finished_count)
883 			lj_done = 1;
884 	}
885 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
886 	MPASS(job->jobflags & KAIOCB_FINISHED);
887 
888 	if (ki->kaio_flags & KAIO_RUNDOWN)
889 		goto notification_done;
890 
891 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
892 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
893 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
894 
895 	KNOTE_LOCKED(&job->klist, 1);
896 
897 	if (lj_done) {
898 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
899 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
900 			KNOTE_LOCKED(&lj->klist, 1);
901 		}
902 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
903 		    == LIOJ_SIGNAL
904 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
905 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
906 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
907 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
908 		}
909 	}
910 
911 notification_done:
912 	if (job->jobflags & KAIOCB_CHECKSYNC) {
913 		schedule_fsync = false;
914 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
915 			if (job->fd_file != sjob->fd_file ||
916 			    job->seqno >= sjob->seqno)
917 				continue;
918 			if (--sjob->pending > 0)
919 				continue;
920 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
921 			if (!aio_clear_cancel_function_locked(sjob))
922 				continue;
923 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
924 			schedule_fsync = true;
925 		}
926 		if (schedule_fsync)
927 			taskqueue_enqueue(taskqueue_aiod_kick,
928 			    &ki->kaio_sync_task);
929 	}
930 	if (ki->kaio_flags & KAIO_WAKEUP) {
931 		ki->kaio_flags &= ~KAIO_WAKEUP;
932 		wakeup(&userp->p_aioinfo);
933 	}
934 }
935 
936 static void
937 aio_schedule_fsync(void *context, int pending)
938 {
939 	struct kaioinfo *ki;
940 	struct kaiocb *job;
941 
942 	ki = context;
943 	AIO_LOCK(ki);
944 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
945 		job = TAILQ_FIRST(&ki->kaio_syncready);
946 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
947 		AIO_UNLOCK(ki);
948 		aio_schedule(job, aio_process_sync);
949 		AIO_LOCK(ki);
950 	}
951 	AIO_UNLOCK(ki);
952 }
953 
954 bool
955 aio_cancel_cleared(struct kaiocb *job)
956 {
957 	struct kaioinfo *ki;
958 
959 	/*
960 	 * The caller should hold the same queue lock held when
961 	 * aio_clear_cancel_function() was called and set this flag
962 	 * ensuring this check sees an up-to-date value.  However,
963 	 * there is no way to assert that.
964 	 */
965 	ki = job->userproc->p_aioinfo;
966 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
967 }
968 
969 static bool
970 aio_clear_cancel_function_locked(struct kaiocb *job)
971 {
972 
973 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
974 	MPASS(job->cancel_fn != NULL);
975 	if (job->jobflags & KAIOCB_CANCELLING) {
976 		job->jobflags |= KAIOCB_CLEARED;
977 		return (false);
978 	}
979 	job->cancel_fn = NULL;
980 	return (true);
981 }
982 
983 bool
984 aio_clear_cancel_function(struct kaiocb *job)
985 {
986 	struct kaioinfo *ki;
987 	bool ret;
988 
989 	ki = job->userproc->p_aioinfo;
990 	AIO_LOCK(ki);
991 	ret = aio_clear_cancel_function_locked(job);
992 	AIO_UNLOCK(ki);
993 	return (ret);
994 }
995 
996 static bool
997 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
998 {
999 
1000 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1001 	if (job->jobflags & KAIOCB_CANCELLED)
1002 		return (false);
1003 	job->cancel_fn = func;
1004 	return (true);
1005 }
1006 
1007 bool
1008 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1009 {
1010 	struct kaioinfo *ki;
1011 	bool ret;
1012 
1013 	ki = job->userproc->p_aioinfo;
1014 	AIO_LOCK(ki);
1015 	ret = aio_set_cancel_function_locked(job, func);
1016 	AIO_UNLOCK(ki);
1017 	return (ret);
1018 }
1019 
1020 void
1021 aio_complete(struct kaiocb *job, long status, int error)
1022 {
1023 	struct kaioinfo *ki;
1024 	struct proc *userp;
1025 
1026 	job->uaiocb._aiocb_private.error = error;
1027 	job->uaiocb._aiocb_private.status = status;
1028 
1029 	userp = job->userproc;
1030 	ki = userp->p_aioinfo;
1031 
1032 	AIO_LOCK(ki);
1033 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1034 	    ("duplicate aio_complete"));
1035 	job->jobflags |= KAIOCB_FINISHED;
1036 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1037 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1038 		aio_bio_done_notify(userp, job);
1039 	}
1040 	AIO_UNLOCK(ki);
1041 }
1042 
1043 void
1044 aio_cancel(struct kaiocb *job)
1045 {
1046 
1047 	aio_complete(job, -1, ECANCELED);
1048 }
1049 
1050 void
1051 aio_switch_vmspace(struct kaiocb *job)
1052 {
1053 
1054 	vmspace_switch_aio(job->userproc->p_vmspace);
1055 }
1056 
1057 /*
1058  * The AIO daemon, most of the actual work is done in aio_process_*,
1059  * but the setup (and address space mgmt) is done in this routine.
1060  */
1061 static void
1062 aio_daemon(void *_id)
1063 {
1064 	struct kaiocb *job;
1065 	struct aioproc *aiop;
1066 	struct kaioinfo *ki;
1067 	struct proc *p;
1068 	struct vmspace *myvm;
1069 	struct thread *td = curthread;
1070 	int id = (intptr_t)_id;
1071 
1072 	/*
1073 	 * Grab an extra reference on the daemon's vmspace so that it
1074 	 * doesn't get freed by jobs that switch to a different
1075 	 * vmspace.
1076 	 */
1077 	p = td->td_proc;
1078 	myvm = vmspace_acquire_ref(p);
1079 
1080 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1081 
1082 	/*
1083 	 * Allocate and ready the aio control info.  There is one aiop structure
1084 	 * per daemon.
1085 	 */
1086 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1087 	aiop->aioproc = p;
1088 	aiop->aioprocflags = 0;
1089 
1090 	/*
1091 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1092 	 * and creating too many daemons.)
1093 	 */
1094 	sema_post(&aio_newproc_sem);
1095 
1096 	mtx_lock(&aio_job_mtx);
1097 	for (;;) {
1098 		/*
1099 		 * Take daemon off of free queue
1100 		 */
1101 		if (aiop->aioprocflags & AIOP_FREE) {
1102 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1103 			aiop->aioprocflags &= ~AIOP_FREE;
1104 		}
1105 
1106 		/*
1107 		 * Check for jobs.
1108 		 */
1109 		while ((job = aio_selectjob(aiop)) != NULL) {
1110 			mtx_unlock(&aio_job_mtx);
1111 
1112 			ki = job->userproc->p_aioinfo;
1113 			job->handle_fn(job);
1114 
1115 			mtx_lock(&aio_job_mtx);
1116 			/* Decrement the active job count. */
1117 			ki->kaio_active_count--;
1118 		}
1119 
1120 		/*
1121 		 * Disconnect from user address space.
1122 		 */
1123 		if (p->p_vmspace != myvm) {
1124 			mtx_unlock(&aio_job_mtx);
1125 			vmspace_switch_aio(myvm);
1126 			mtx_lock(&aio_job_mtx);
1127 			/*
1128 			 * We have to restart to avoid race, we only sleep if
1129 			 * no job can be selected.
1130 			 */
1131 			continue;
1132 		}
1133 
1134 		mtx_assert(&aio_job_mtx, MA_OWNED);
1135 
1136 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1137 		aiop->aioprocflags |= AIOP_FREE;
1138 
1139 		/*
1140 		 * If daemon is inactive for a long time, allow it to exit,
1141 		 * thereby freeing resources.
1142 		 */
1143 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1144 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1145 		    (aiop->aioprocflags & AIOP_FREE) &&
1146 		    num_aio_procs > target_aio_procs)
1147 			break;
1148 	}
1149 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1150 	num_aio_procs--;
1151 	mtx_unlock(&aio_job_mtx);
1152 	uma_zfree(aiop_zone, aiop);
1153 	free_unr(aiod_unr, id);
1154 	vmspace_free(myvm);
1155 
1156 	KASSERT(p->p_vmspace == myvm,
1157 	    ("AIOD: bad vmspace for exiting daemon"));
1158 	KASSERT(myvm->vm_refcnt > 1,
1159 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1160 	kproc_exit(0);
1161 }
1162 
1163 /*
1164  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1165  * AIO daemon modifies its environment itself.
1166  */
1167 static int
1168 aio_newproc(int *start)
1169 {
1170 	int error;
1171 	struct proc *p;
1172 	int id;
1173 
1174 	id = alloc_unr(aiod_unr);
1175 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1176 		RFNOWAIT, 0, "aiod%d", id);
1177 	if (error == 0) {
1178 		/*
1179 		 * Wait until daemon is started.
1180 		 */
1181 		sema_wait(&aio_newproc_sem);
1182 		mtx_lock(&aio_job_mtx);
1183 		num_aio_procs++;
1184 		if (start != NULL)
1185 			(*start)--;
1186 		mtx_unlock(&aio_job_mtx);
1187 	} else {
1188 		free_unr(aiod_unr, id);
1189 	}
1190 	return (error);
1191 }
1192 
1193 /*
1194  * Try the high-performance, low-overhead physio method for eligible
1195  * VCHR devices.  This method doesn't use an aio helper thread, and
1196  * thus has very low overhead.
1197  *
1198  * Assumes that the caller, aio_aqueue(), has incremented the file
1199  * structure's reference count, preventing its deallocation for the
1200  * duration of this call.
1201  */
1202 static int
1203 aio_qphysio(struct proc *p, struct kaiocb *job)
1204 {
1205 	struct aiocb *cb;
1206 	struct file *fp;
1207 	struct bio *bp;
1208 	struct buf *pbuf;
1209 	struct vnode *vp;
1210 	struct cdevsw *csw;
1211 	struct cdev *dev;
1212 	struct kaioinfo *ki;
1213 	int error, ref, poff;
1214 	vm_prot_t prot;
1215 
1216 	cb = &job->uaiocb;
1217 	fp = job->fd_file;
1218 
1219 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1220 		return (-1);
1221 
1222 	vp = fp->f_vnode;
1223 	if (vp->v_type != VCHR)
1224 		return (-1);
1225 	if (vp->v_bufobj.bo_bsize == 0)
1226 		return (-1);
1227 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1228 		return (-1);
1229 
1230 	ref = 0;
1231 	csw = devvn_refthread(vp, &dev, &ref);
1232 	if (csw == NULL)
1233 		return (ENXIO);
1234 
1235 	if ((csw->d_flags & D_DISK) == 0) {
1236 		error = -1;
1237 		goto unref;
1238 	}
1239 	if (cb->aio_nbytes > dev->si_iosize_max) {
1240 		error = -1;
1241 		goto unref;
1242 	}
1243 
1244 	ki = p->p_aioinfo;
1245 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1246 	if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1247 		if (cb->aio_nbytes > MAXPHYS) {
1248 			error = -1;
1249 			goto unref;
1250 		}
1251 
1252 		pbuf = NULL;
1253 	} else {
1254 		if (cb->aio_nbytes > MAXPHYS - poff) {
1255 			error = -1;
1256 			goto unref;
1257 		}
1258 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1259 			error = -1;
1260 			goto unref;
1261 		}
1262 
1263 		job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1264 		BUF_KERNPROC(pbuf);
1265 		AIO_LOCK(ki);
1266 		ki->kaio_buffer_count++;
1267 		AIO_UNLOCK(ki);
1268 	}
1269 	job->bp = bp = g_alloc_bio();
1270 
1271 	bp->bio_length = cb->aio_nbytes;
1272 	bp->bio_bcount = cb->aio_nbytes;
1273 	bp->bio_done = aio_physwakeup;
1274 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1275 	bp->bio_offset = cb->aio_offset;
1276 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1277 	bp->bio_dev = dev;
1278 	bp->bio_caller1 = (void *)job;
1279 
1280 	prot = VM_PROT_READ;
1281 	if (cb->aio_lio_opcode == LIO_READ)
1282 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1283 	job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1284 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1285 	    nitems(job->pages));
1286 	if (job->npages < 0) {
1287 		error = EFAULT;
1288 		goto doerror;
1289 	}
1290 	if (pbuf != NULL) {
1291 		pmap_qenter((vm_offset_t)pbuf->b_data,
1292 		    job->pages, job->npages);
1293 		bp->bio_data = pbuf->b_data + poff;
1294 		atomic_add_int(&num_buf_aio, 1);
1295 	} else {
1296 		bp->bio_ma = job->pages;
1297 		bp->bio_ma_n = job->npages;
1298 		bp->bio_ma_offset = poff;
1299 		bp->bio_data = unmapped_buf;
1300 		bp->bio_flags |= BIO_UNMAPPED;
1301 	}
1302 
1303 	/* Perform transfer. */
1304 	csw->d_strategy(bp);
1305 	dev_relthread(dev, ref);
1306 	return (0);
1307 
1308 doerror:
1309 	if (pbuf != NULL) {
1310 		AIO_LOCK(ki);
1311 		ki->kaio_buffer_count--;
1312 		AIO_UNLOCK(ki);
1313 		relpbuf(pbuf, NULL);
1314 		job->pbuf = NULL;
1315 	}
1316 	g_destroy_bio(bp);
1317 	job->bp = NULL;
1318 unref:
1319 	dev_relthread(dev, ref);
1320 	return (error);
1321 }
1322 
1323 #ifdef COMPAT_FREEBSD6
1324 static int
1325 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1326 {
1327 
1328 	/*
1329 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1330 	 * supported by AIO with the old sigevent structure.
1331 	 */
1332 	nsig->sigev_notify = osig->sigev_notify;
1333 	switch (nsig->sigev_notify) {
1334 	case SIGEV_NONE:
1335 		break;
1336 	case SIGEV_SIGNAL:
1337 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1338 		break;
1339 	case SIGEV_KEVENT:
1340 		nsig->sigev_notify_kqueue =
1341 		    osig->__sigev_u.__sigev_notify_kqueue;
1342 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1343 		break;
1344 	default:
1345 		return (EINVAL);
1346 	}
1347 	return (0);
1348 }
1349 
1350 static int
1351 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1352 {
1353 	struct oaiocb *ojob;
1354 	int error;
1355 
1356 	bzero(kjob, sizeof(struct aiocb));
1357 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1358 	if (error)
1359 		return (error);
1360 	ojob = (struct oaiocb *)kjob;
1361 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1362 }
1363 #endif
1364 
1365 static int
1366 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1367 {
1368 
1369 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1370 }
1371 
1372 static long
1373 aiocb_fetch_status(struct aiocb *ujob)
1374 {
1375 
1376 	return (fuword(&ujob->_aiocb_private.status));
1377 }
1378 
1379 static long
1380 aiocb_fetch_error(struct aiocb *ujob)
1381 {
1382 
1383 	return (fuword(&ujob->_aiocb_private.error));
1384 }
1385 
1386 static int
1387 aiocb_store_status(struct aiocb *ujob, long status)
1388 {
1389 
1390 	return (suword(&ujob->_aiocb_private.status, status));
1391 }
1392 
1393 static int
1394 aiocb_store_error(struct aiocb *ujob, long error)
1395 {
1396 
1397 	return (suword(&ujob->_aiocb_private.error, error));
1398 }
1399 
1400 static int
1401 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1402 {
1403 
1404 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1405 }
1406 
1407 static int
1408 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1409 {
1410 
1411 	return (suword(ujobp, (long)ujob));
1412 }
1413 
1414 static struct aiocb_ops aiocb_ops = {
1415 	.copyin = aiocb_copyin,
1416 	.fetch_status = aiocb_fetch_status,
1417 	.fetch_error = aiocb_fetch_error,
1418 	.store_status = aiocb_store_status,
1419 	.store_error = aiocb_store_error,
1420 	.store_kernelinfo = aiocb_store_kernelinfo,
1421 	.store_aiocb = aiocb_store_aiocb,
1422 };
1423 
1424 #ifdef COMPAT_FREEBSD6
1425 static struct aiocb_ops aiocb_ops_osigevent = {
1426 	.copyin = aiocb_copyin_old_sigevent,
1427 	.fetch_status = aiocb_fetch_status,
1428 	.fetch_error = aiocb_fetch_error,
1429 	.store_status = aiocb_store_status,
1430 	.store_error = aiocb_store_error,
1431 	.store_kernelinfo = aiocb_store_kernelinfo,
1432 	.store_aiocb = aiocb_store_aiocb,
1433 };
1434 #endif
1435 
1436 /*
1437  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1438  * technique is done in this code.
1439  */
1440 int
1441 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1442     int type, struct aiocb_ops *ops)
1443 {
1444 	struct proc *p = td->td_proc;
1445 	cap_rights_t rights;
1446 	struct file *fp;
1447 	struct kaiocb *job;
1448 	struct kaioinfo *ki;
1449 	struct kevent kev;
1450 	int opcode;
1451 	int error;
1452 	int fd, kqfd;
1453 	int jid;
1454 	u_short evflags;
1455 
1456 	if (p->p_aioinfo == NULL)
1457 		aio_init_aioinfo(p);
1458 
1459 	ki = p->p_aioinfo;
1460 
1461 	ops->store_status(ujob, -1);
1462 	ops->store_error(ujob, 0);
1463 	ops->store_kernelinfo(ujob, -1);
1464 
1465 	if (num_queue_count >= max_queue_count ||
1466 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1467 		ops->store_error(ujob, EAGAIN);
1468 		return (EAGAIN);
1469 	}
1470 
1471 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1472 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1473 
1474 	error = ops->copyin(ujob, &job->uaiocb);
1475 	if (error) {
1476 		ops->store_error(ujob, error);
1477 		uma_zfree(aiocb_zone, job);
1478 		return (error);
1479 	}
1480 
1481 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1482 		uma_zfree(aiocb_zone, job);
1483 		return (EINVAL);
1484 	}
1485 
1486 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1487 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1488 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1489 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1490 		ops->store_error(ujob, EINVAL);
1491 		uma_zfree(aiocb_zone, job);
1492 		return (EINVAL);
1493 	}
1494 
1495 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1496 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1497 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1498 		uma_zfree(aiocb_zone, job);
1499 		return (EINVAL);
1500 	}
1501 
1502 	ksiginfo_init(&job->ksi);
1503 
1504 	/* Save userspace address of the job info. */
1505 	job->ujob = ujob;
1506 
1507 	/* Get the opcode. */
1508 	if (type != LIO_NOP)
1509 		job->uaiocb.aio_lio_opcode = type;
1510 	opcode = job->uaiocb.aio_lio_opcode;
1511 
1512 	/*
1513 	 * Validate the opcode and fetch the file object for the specified
1514 	 * file descriptor.
1515 	 *
1516 	 * XXXRW: Moved the opcode validation up here so that we don't
1517 	 * retrieve a file descriptor without knowing what the capabiltity
1518 	 * should be.
1519 	 */
1520 	fd = job->uaiocb.aio_fildes;
1521 	switch (opcode) {
1522 	case LIO_WRITE:
1523 		error = fget_write(td, fd,
1524 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1525 		break;
1526 	case LIO_READ:
1527 		error = fget_read(td, fd,
1528 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1529 		break;
1530 	case LIO_SYNC:
1531 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1532 		break;
1533 	case LIO_MLOCK:
1534 		fp = NULL;
1535 		break;
1536 	case LIO_NOP:
1537 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1538 		break;
1539 	default:
1540 		error = EINVAL;
1541 	}
1542 	if (error) {
1543 		uma_zfree(aiocb_zone, job);
1544 		ops->store_error(ujob, error);
1545 		return (error);
1546 	}
1547 
1548 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1549 		error = EINVAL;
1550 		goto aqueue_fail;
1551 	}
1552 
1553 	if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1554 		error = EINVAL;
1555 		goto aqueue_fail;
1556 	}
1557 
1558 	job->fd_file = fp;
1559 
1560 	mtx_lock(&aio_job_mtx);
1561 	jid = jobrefid++;
1562 	job->seqno = jobseqno++;
1563 	mtx_unlock(&aio_job_mtx);
1564 	error = ops->store_kernelinfo(ujob, jid);
1565 	if (error) {
1566 		error = EINVAL;
1567 		goto aqueue_fail;
1568 	}
1569 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1570 
1571 	if (opcode == LIO_NOP) {
1572 		fdrop(fp, td);
1573 		uma_zfree(aiocb_zone, job);
1574 		return (0);
1575 	}
1576 
1577 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1578 		goto no_kqueue;
1579 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1580 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1581 		error = EINVAL;
1582 		goto aqueue_fail;
1583 	}
1584 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1585 	kev.ident = (uintptr_t)job->ujob;
1586 	kev.filter = EVFILT_AIO;
1587 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1588 	kev.data = (intptr_t)job;
1589 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1590 	error = kqfd_register(kqfd, &kev, td, 1);
1591 	if (error)
1592 		goto aqueue_fail;
1593 
1594 no_kqueue:
1595 
1596 	ops->store_error(ujob, EINPROGRESS);
1597 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1598 	job->userproc = p;
1599 	job->cred = crhold(td->td_ucred);
1600 	job->jobflags = KAIOCB_QUEUEING;
1601 	job->lio = lj;
1602 
1603 	if (opcode == LIO_MLOCK) {
1604 		aio_schedule(job, aio_process_mlock);
1605 		error = 0;
1606 	} else if (fp->f_ops->fo_aio_queue == NULL)
1607 		error = aio_queue_file(fp, job);
1608 	else
1609 		error = fo_aio_queue(fp, job);
1610 	if (error)
1611 		goto aqueue_fail;
1612 
1613 	AIO_LOCK(ki);
1614 	job->jobflags &= ~KAIOCB_QUEUEING;
1615 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1616 	ki->kaio_count++;
1617 	if (lj)
1618 		lj->lioj_count++;
1619 	atomic_add_int(&num_queue_count, 1);
1620 	if (job->jobflags & KAIOCB_FINISHED) {
1621 		/*
1622 		 * The queue callback completed the request synchronously.
1623 		 * The bulk of the completion is deferred in that case
1624 		 * until this point.
1625 		 */
1626 		aio_bio_done_notify(p, job);
1627 	} else
1628 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1629 	AIO_UNLOCK(ki);
1630 	return (0);
1631 
1632 aqueue_fail:
1633 	knlist_delete(&job->klist, curthread, 0);
1634 	if (fp)
1635 		fdrop(fp, td);
1636 	uma_zfree(aiocb_zone, job);
1637 	ops->store_error(ujob, error);
1638 	return (error);
1639 }
1640 
1641 static void
1642 aio_cancel_daemon_job(struct kaiocb *job)
1643 {
1644 
1645 	mtx_lock(&aio_job_mtx);
1646 	if (!aio_cancel_cleared(job))
1647 		TAILQ_REMOVE(&aio_jobs, job, list);
1648 	mtx_unlock(&aio_job_mtx);
1649 	aio_cancel(job);
1650 }
1651 
1652 void
1653 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1654 {
1655 
1656 	mtx_lock(&aio_job_mtx);
1657 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1658 		mtx_unlock(&aio_job_mtx);
1659 		aio_cancel(job);
1660 		return;
1661 	}
1662 	job->handle_fn = func;
1663 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1664 	aio_kick_nowait(job->userproc);
1665 	mtx_unlock(&aio_job_mtx);
1666 }
1667 
1668 static void
1669 aio_cancel_sync(struct kaiocb *job)
1670 {
1671 	struct kaioinfo *ki;
1672 
1673 	ki = job->userproc->p_aioinfo;
1674 	AIO_LOCK(ki);
1675 	if (!aio_cancel_cleared(job))
1676 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1677 	AIO_UNLOCK(ki);
1678 	aio_cancel(job);
1679 }
1680 
1681 int
1682 aio_queue_file(struct file *fp, struct kaiocb *job)
1683 {
1684 	struct aioliojob *lj;
1685 	struct kaioinfo *ki;
1686 	struct kaiocb *job2;
1687 	struct vnode *vp;
1688 	struct mount *mp;
1689 	int error, opcode;
1690 	bool safe;
1691 
1692 	lj = job->lio;
1693 	ki = job->userproc->p_aioinfo;
1694 	opcode = job->uaiocb.aio_lio_opcode;
1695 	if (opcode == LIO_SYNC)
1696 		goto queueit;
1697 
1698 	if ((error = aio_qphysio(job->userproc, job)) == 0)
1699 		goto done;
1700 #if 0
1701 	/*
1702 	 * XXX: This means qphysio() failed with EFAULT.  The current
1703 	 * behavior is to retry the operation via fo_read/fo_write.
1704 	 * Wouldn't it be better to just complete the request with an
1705 	 * error here?
1706 	 */
1707 	if (error > 0)
1708 		goto done;
1709 #endif
1710 queueit:
1711 	safe = false;
1712 	if (fp->f_type == DTYPE_VNODE) {
1713 		vp = fp->f_vnode;
1714 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1715 			mp = fp->f_vnode->v_mount;
1716 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1717 				safe = true;
1718 		}
1719 	}
1720 	if (!(safe || enable_aio_unsafe)) {
1721 		counted_warning(&unsafe_warningcnt,
1722 		    "is attempting to use unsafe AIO requests");
1723 		return (EOPNOTSUPP);
1724 	}
1725 
1726 	if (opcode == LIO_SYNC) {
1727 		AIO_LOCK(ki);
1728 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1729 			if (job2->fd_file == job->fd_file &&
1730 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1731 			    job2->seqno < job->seqno) {
1732 				job2->jobflags |= KAIOCB_CHECKSYNC;
1733 				job->pending++;
1734 			}
1735 		}
1736 		if (job->pending != 0) {
1737 			if (!aio_set_cancel_function_locked(job,
1738 				aio_cancel_sync)) {
1739 				AIO_UNLOCK(ki);
1740 				aio_cancel(job);
1741 				return (0);
1742 			}
1743 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1744 			AIO_UNLOCK(ki);
1745 			return (0);
1746 		}
1747 		AIO_UNLOCK(ki);
1748 	}
1749 
1750 	switch (opcode) {
1751 	case LIO_READ:
1752 	case LIO_WRITE:
1753 		aio_schedule(job, aio_process_rw);
1754 		error = 0;
1755 		break;
1756 	case LIO_SYNC:
1757 		aio_schedule(job, aio_process_sync);
1758 		error = 0;
1759 		break;
1760 	default:
1761 		error = EINVAL;
1762 	}
1763 done:
1764 	return (error);
1765 }
1766 
1767 static void
1768 aio_kick_nowait(struct proc *userp)
1769 {
1770 	struct kaioinfo *ki = userp->p_aioinfo;
1771 	struct aioproc *aiop;
1772 
1773 	mtx_assert(&aio_job_mtx, MA_OWNED);
1774 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1775 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1776 		aiop->aioprocflags &= ~AIOP_FREE;
1777 		wakeup(aiop->aioproc);
1778 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1779 	    ki->kaio_active_count + num_aio_resv_start <
1780 	    ki->kaio_maxactive_count) {
1781 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1782 	}
1783 }
1784 
1785 static int
1786 aio_kick(struct proc *userp)
1787 {
1788 	struct kaioinfo *ki = userp->p_aioinfo;
1789 	struct aioproc *aiop;
1790 	int error, ret = 0;
1791 
1792 	mtx_assert(&aio_job_mtx, MA_OWNED);
1793 retryproc:
1794 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1795 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1796 		aiop->aioprocflags &= ~AIOP_FREE;
1797 		wakeup(aiop->aioproc);
1798 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1799 	    ki->kaio_active_count + num_aio_resv_start <
1800 	    ki->kaio_maxactive_count) {
1801 		num_aio_resv_start++;
1802 		mtx_unlock(&aio_job_mtx);
1803 		error = aio_newproc(&num_aio_resv_start);
1804 		mtx_lock(&aio_job_mtx);
1805 		if (error) {
1806 			num_aio_resv_start--;
1807 			goto retryproc;
1808 		}
1809 	} else {
1810 		ret = -1;
1811 	}
1812 	return (ret);
1813 }
1814 
1815 static void
1816 aio_kick_helper(void *context, int pending)
1817 {
1818 	struct proc *userp = context;
1819 
1820 	mtx_lock(&aio_job_mtx);
1821 	while (--pending >= 0) {
1822 		if (aio_kick(userp))
1823 			break;
1824 	}
1825 	mtx_unlock(&aio_job_mtx);
1826 }
1827 
1828 /*
1829  * Support the aio_return system call, as a side-effect, kernel resources are
1830  * released.
1831  */
1832 static int
1833 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1834 {
1835 	struct proc *p = td->td_proc;
1836 	struct kaiocb *job;
1837 	struct kaioinfo *ki;
1838 	long status, error;
1839 
1840 	ki = p->p_aioinfo;
1841 	if (ki == NULL)
1842 		return (EINVAL);
1843 	AIO_LOCK(ki);
1844 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1845 		if (job->ujob == ujob)
1846 			break;
1847 	}
1848 	if (job != NULL) {
1849 		MPASS(job->jobflags & KAIOCB_FINISHED);
1850 		status = job->uaiocb._aiocb_private.status;
1851 		error = job->uaiocb._aiocb_private.error;
1852 		td->td_retval[0] = status;
1853 		td->td_ru.ru_oublock += job->outblock;
1854 		td->td_ru.ru_inblock += job->inblock;
1855 		td->td_ru.ru_msgsnd += job->msgsnd;
1856 		td->td_ru.ru_msgrcv += job->msgrcv;
1857 		aio_free_entry(job);
1858 		AIO_UNLOCK(ki);
1859 		ops->store_error(ujob, error);
1860 		ops->store_status(ujob, status);
1861 	} else {
1862 		error = EINVAL;
1863 		AIO_UNLOCK(ki);
1864 	}
1865 	return (error);
1866 }
1867 
1868 int
1869 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1870 {
1871 
1872 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1873 }
1874 
1875 /*
1876  * Allow a process to wakeup when any of the I/O requests are completed.
1877  */
1878 static int
1879 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1880     struct timespec *ts)
1881 {
1882 	struct proc *p = td->td_proc;
1883 	struct timeval atv;
1884 	struct kaioinfo *ki;
1885 	struct kaiocb *firstjob, *job;
1886 	int error, i, timo;
1887 
1888 	timo = 0;
1889 	if (ts) {
1890 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1891 			return (EINVAL);
1892 
1893 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1894 		if (itimerfix(&atv))
1895 			return (EINVAL);
1896 		timo = tvtohz(&atv);
1897 	}
1898 
1899 	ki = p->p_aioinfo;
1900 	if (ki == NULL)
1901 		return (EAGAIN);
1902 
1903 	if (njoblist == 0)
1904 		return (0);
1905 
1906 	AIO_LOCK(ki);
1907 	for (;;) {
1908 		firstjob = NULL;
1909 		error = 0;
1910 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1911 			for (i = 0; i < njoblist; i++) {
1912 				if (job->ujob == ujoblist[i]) {
1913 					if (firstjob == NULL)
1914 						firstjob = job;
1915 					if (job->jobflags & KAIOCB_FINISHED)
1916 						goto RETURN;
1917 				}
1918 			}
1919 		}
1920 		/* All tasks were finished. */
1921 		if (firstjob == NULL)
1922 			break;
1923 
1924 		ki->kaio_flags |= KAIO_WAKEUP;
1925 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1926 		    "aiospn", timo);
1927 		if (error == ERESTART)
1928 			error = EINTR;
1929 		if (error)
1930 			break;
1931 	}
1932 RETURN:
1933 	AIO_UNLOCK(ki);
1934 	return (error);
1935 }
1936 
1937 int
1938 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1939 {
1940 	struct timespec ts, *tsp;
1941 	struct aiocb **ujoblist;
1942 	int error;
1943 
1944 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1945 		return (EINVAL);
1946 
1947 	if (uap->timeout) {
1948 		/* Get timespec struct. */
1949 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1950 			return (error);
1951 		tsp = &ts;
1952 	} else
1953 		tsp = NULL;
1954 
1955 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1956 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1957 	if (error == 0)
1958 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1959 	uma_zfree(aiol_zone, ujoblist);
1960 	return (error);
1961 }
1962 
1963 /*
1964  * aio_cancel cancels any non-physio aio operations not currently in
1965  * progress.
1966  */
1967 int
1968 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1969 {
1970 	struct proc *p = td->td_proc;
1971 	struct kaioinfo *ki;
1972 	struct kaiocb *job, *jobn;
1973 	struct file *fp;
1974 	cap_rights_t rights;
1975 	int error;
1976 	int cancelled = 0;
1977 	int notcancelled = 0;
1978 	struct vnode *vp;
1979 
1980 	/* Lookup file object. */
1981 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1982 	if (error)
1983 		return (error);
1984 
1985 	ki = p->p_aioinfo;
1986 	if (ki == NULL)
1987 		goto done;
1988 
1989 	if (fp->f_type == DTYPE_VNODE) {
1990 		vp = fp->f_vnode;
1991 		if (vn_isdisk(vp, &error)) {
1992 			fdrop(fp, td);
1993 			td->td_retval[0] = AIO_NOTCANCELED;
1994 			return (0);
1995 		}
1996 	}
1997 
1998 	AIO_LOCK(ki);
1999 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2000 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2001 		    ((uap->aiocbp == NULL) ||
2002 		     (uap->aiocbp == job->ujob))) {
2003 			if (aio_cancel_job(p, ki, job)) {
2004 				cancelled++;
2005 			} else {
2006 				notcancelled++;
2007 			}
2008 			if (uap->aiocbp != NULL)
2009 				break;
2010 		}
2011 	}
2012 	AIO_UNLOCK(ki);
2013 
2014 done:
2015 	fdrop(fp, td);
2016 
2017 	if (uap->aiocbp != NULL) {
2018 		if (cancelled) {
2019 			td->td_retval[0] = AIO_CANCELED;
2020 			return (0);
2021 		}
2022 	}
2023 
2024 	if (notcancelled) {
2025 		td->td_retval[0] = AIO_NOTCANCELED;
2026 		return (0);
2027 	}
2028 
2029 	if (cancelled) {
2030 		td->td_retval[0] = AIO_CANCELED;
2031 		return (0);
2032 	}
2033 
2034 	td->td_retval[0] = AIO_ALLDONE;
2035 
2036 	return (0);
2037 }
2038 
2039 /*
2040  * aio_error is implemented in the kernel level for compatibility purposes
2041  * only.  For a user mode async implementation, it would be best to do it in
2042  * a userland subroutine.
2043  */
2044 static int
2045 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2046 {
2047 	struct proc *p = td->td_proc;
2048 	struct kaiocb *job;
2049 	struct kaioinfo *ki;
2050 	int status;
2051 
2052 	ki = p->p_aioinfo;
2053 	if (ki == NULL) {
2054 		td->td_retval[0] = EINVAL;
2055 		return (0);
2056 	}
2057 
2058 	AIO_LOCK(ki);
2059 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2060 		if (job->ujob == ujob) {
2061 			if (job->jobflags & KAIOCB_FINISHED)
2062 				td->td_retval[0] =
2063 					job->uaiocb._aiocb_private.error;
2064 			else
2065 				td->td_retval[0] = EINPROGRESS;
2066 			AIO_UNLOCK(ki);
2067 			return (0);
2068 		}
2069 	}
2070 	AIO_UNLOCK(ki);
2071 
2072 	/*
2073 	 * Hack for failure of aio_aqueue.
2074 	 */
2075 	status = ops->fetch_status(ujob);
2076 	if (status == -1) {
2077 		td->td_retval[0] = ops->fetch_error(ujob);
2078 		return (0);
2079 	}
2080 
2081 	td->td_retval[0] = EINVAL;
2082 	return (0);
2083 }
2084 
2085 int
2086 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2087 {
2088 
2089 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2090 }
2091 
2092 /* syscall - asynchronous read from a file (REALTIME) */
2093 #ifdef COMPAT_FREEBSD6
2094 int
2095 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2096 {
2097 
2098 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2099 	    &aiocb_ops_osigevent));
2100 }
2101 #endif
2102 
2103 int
2104 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2105 {
2106 
2107 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2108 }
2109 
2110 /* syscall - asynchronous write to a file (REALTIME) */
2111 #ifdef COMPAT_FREEBSD6
2112 int
2113 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2114 {
2115 
2116 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2117 	    &aiocb_ops_osigevent));
2118 }
2119 #endif
2120 
2121 int
2122 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2123 {
2124 
2125 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2126 }
2127 
2128 int
2129 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2130 {
2131 
2132 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2133 }
2134 
2135 static int
2136 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2137     struct aiocb **acb_list, int nent, struct sigevent *sig,
2138     struct aiocb_ops *ops)
2139 {
2140 	struct proc *p = td->td_proc;
2141 	struct aiocb *job;
2142 	struct kaioinfo *ki;
2143 	struct aioliojob *lj;
2144 	struct kevent kev;
2145 	int error;
2146 	int nerror;
2147 	int i;
2148 
2149 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2150 		return (EINVAL);
2151 
2152 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2153 		return (EINVAL);
2154 
2155 	if (p->p_aioinfo == NULL)
2156 		aio_init_aioinfo(p);
2157 
2158 	ki = p->p_aioinfo;
2159 
2160 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2161 	lj->lioj_flags = 0;
2162 	lj->lioj_count = 0;
2163 	lj->lioj_finished_count = 0;
2164 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2165 	ksiginfo_init(&lj->lioj_ksi);
2166 
2167 	/*
2168 	 * Setup signal.
2169 	 */
2170 	if (sig && (mode == LIO_NOWAIT)) {
2171 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2172 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2173 			/* Assume only new style KEVENT */
2174 			kev.filter = EVFILT_LIO;
2175 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2176 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2177 			kev.data = (intptr_t)lj;
2178 			/* pass user defined sigval data */
2179 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2180 			error = kqfd_register(
2181 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2182 			if (error) {
2183 				uma_zfree(aiolio_zone, lj);
2184 				return (error);
2185 			}
2186 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2187 			;
2188 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2189 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2190 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2191 					uma_zfree(aiolio_zone, lj);
2192 					return EINVAL;
2193 				}
2194 				lj->lioj_flags |= LIOJ_SIGNAL;
2195 		} else {
2196 			uma_zfree(aiolio_zone, lj);
2197 			return EINVAL;
2198 		}
2199 	}
2200 
2201 	AIO_LOCK(ki);
2202 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2203 	/*
2204 	 * Add extra aiocb count to avoid the lio to be freed
2205 	 * by other threads doing aio_waitcomplete or aio_return,
2206 	 * and prevent event from being sent until we have queued
2207 	 * all tasks.
2208 	 */
2209 	lj->lioj_count = 1;
2210 	AIO_UNLOCK(ki);
2211 
2212 	/*
2213 	 * Get pointers to the list of I/O requests.
2214 	 */
2215 	nerror = 0;
2216 	for (i = 0; i < nent; i++) {
2217 		job = acb_list[i];
2218 		if (job != NULL) {
2219 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2220 			if (error != 0)
2221 				nerror++;
2222 		}
2223 	}
2224 
2225 	error = 0;
2226 	AIO_LOCK(ki);
2227 	if (mode == LIO_WAIT) {
2228 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2229 			ki->kaio_flags |= KAIO_WAKEUP;
2230 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2231 			    PRIBIO | PCATCH, "aiospn", 0);
2232 			if (error == ERESTART)
2233 				error = EINTR;
2234 			if (error)
2235 				break;
2236 		}
2237 	} else {
2238 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2239 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2240 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2241 				KNOTE_LOCKED(&lj->klist, 1);
2242 			}
2243 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2244 			    == LIOJ_SIGNAL
2245 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2246 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2247 				aio_sendsig(p, &lj->lioj_signal,
2248 					    &lj->lioj_ksi);
2249 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2250 			}
2251 		}
2252 	}
2253 	lj->lioj_count--;
2254 	if (lj->lioj_count == 0) {
2255 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2256 		knlist_delete(&lj->klist, curthread, 1);
2257 		PROC_LOCK(p);
2258 		sigqueue_take(&lj->lioj_ksi);
2259 		PROC_UNLOCK(p);
2260 		AIO_UNLOCK(ki);
2261 		uma_zfree(aiolio_zone, lj);
2262 	} else
2263 		AIO_UNLOCK(ki);
2264 
2265 	if (nerror)
2266 		return (EIO);
2267 	return (error);
2268 }
2269 
2270 /* syscall - list directed I/O (REALTIME) */
2271 #ifdef COMPAT_FREEBSD6
2272 int
2273 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2274 {
2275 	struct aiocb **acb_list;
2276 	struct sigevent *sigp, sig;
2277 	struct osigevent osig;
2278 	int error, nent;
2279 
2280 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2281 		return (EINVAL);
2282 
2283 	nent = uap->nent;
2284 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2285 		return (EINVAL);
2286 
2287 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2288 		error = copyin(uap->sig, &osig, sizeof(osig));
2289 		if (error)
2290 			return (error);
2291 		error = convert_old_sigevent(&osig, &sig);
2292 		if (error)
2293 			return (error);
2294 		sigp = &sig;
2295 	} else
2296 		sigp = NULL;
2297 
2298 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2299 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2300 	if (error == 0)
2301 		error = kern_lio_listio(td, uap->mode,
2302 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2303 		    &aiocb_ops_osigevent);
2304 	free(acb_list, M_LIO);
2305 	return (error);
2306 }
2307 #endif
2308 
2309 /* syscall - list directed I/O (REALTIME) */
2310 int
2311 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2312 {
2313 	struct aiocb **acb_list;
2314 	struct sigevent *sigp, sig;
2315 	int error, nent;
2316 
2317 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2318 		return (EINVAL);
2319 
2320 	nent = uap->nent;
2321 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2322 		return (EINVAL);
2323 
2324 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2325 		error = copyin(uap->sig, &sig, sizeof(sig));
2326 		if (error)
2327 			return (error);
2328 		sigp = &sig;
2329 	} else
2330 		sigp = NULL;
2331 
2332 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2333 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2334 	if (error == 0)
2335 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2336 		    nent, sigp, &aiocb_ops);
2337 	free(acb_list, M_LIO);
2338 	return (error);
2339 }
2340 
2341 static void
2342 aio_physwakeup(struct bio *bp)
2343 {
2344 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2345 	struct proc *userp;
2346 	struct kaioinfo *ki;
2347 	size_t nbytes;
2348 	int error, nblks;
2349 
2350 	/* Release mapping into kernel space. */
2351 	userp = job->userproc;
2352 	ki = userp->p_aioinfo;
2353 	if (job->pbuf) {
2354 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2355 		relpbuf(job->pbuf, NULL);
2356 		job->pbuf = NULL;
2357 		atomic_subtract_int(&num_buf_aio, 1);
2358 		AIO_LOCK(ki);
2359 		ki->kaio_buffer_count--;
2360 		AIO_UNLOCK(ki);
2361 	}
2362 	vm_page_unhold_pages(job->pages, job->npages);
2363 
2364 	bp = job->bp;
2365 	job->bp = NULL;
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 = (struct kaiocb *)kn->kn_sdata;
2495 
2496 	/*
2497 	 * The job pointer must be validated before using it, so
2498 	 * registration is restricted to the kernel; the user cannot
2499 	 * set EV_FLAG1.
2500 	 */
2501 	if ((kn->kn_flags & EV_FLAG1) == 0)
2502 		return (EPERM);
2503 	kn->kn_ptr.p_aio = job;
2504 	kn->kn_flags &= ~EV_FLAG1;
2505 
2506 	knlist_add(&job->klist, kn, 0);
2507 
2508 	return (0);
2509 }
2510 
2511 /* kqueue detach function */
2512 static void
2513 filt_aiodetach(struct knote *kn)
2514 {
2515 	struct knlist *knl;
2516 
2517 	knl = &kn->kn_ptr.p_aio->klist;
2518 	knl->kl_lock(knl->kl_lockarg);
2519 	if (!knlist_empty(knl))
2520 		knlist_remove(knl, kn, 1);
2521 	knl->kl_unlock(knl->kl_lockarg);
2522 }
2523 
2524 /* kqueue filter function */
2525 /*ARGSUSED*/
2526 static int
2527 filt_aio(struct knote *kn, long hint)
2528 {
2529 	struct kaiocb *job = kn->kn_ptr.p_aio;
2530 
2531 	kn->kn_data = job->uaiocb._aiocb_private.error;
2532 	if (!(job->jobflags & KAIOCB_FINISHED))
2533 		return (0);
2534 	kn->kn_flags |= EV_EOF;
2535 	return (1);
2536 }
2537 
2538 /* kqueue attach function */
2539 static int
2540 filt_lioattach(struct knote *kn)
2541 {
2542 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2543 
2544 	/*
2545 	 * The aioliojob pointer must be validated before using it, so
2546 	 * registration is restricted to the kernel; the user cannot
2547 	 * set EV_FLAG1.
2548 	 */
2549 	if ((kn->kn_flags & EV_FLAG1) == 0)
2550 		return (EPERM);
2551 	kn->kn_ptr.p_lio = lj;
2552 	kn->kn_flags &= ~EV_FLAG1;
2553 
2554 	knlist_add(&lj->klist, kn, 0);
2555 
2556 	return (0);
2557 }
2558 
2559 /* kqueue detach function */
2560 static void
2561 filt_liodetach(struct knote *kn)
2562 {
2563 	struct knlist *knl;
2564 
2565 	knl = &kn->kn_ptr.p_lio->klist;
2566 	knl->kl_lock(knl->kl_lockarg);
2567 	if (!knlist_empty(knl))
2568 		knlist_remove(knl, kn, 1);
2569 	knl->kl_unlock(knl->kl_lockarg);
2570 }
2571 
2572 /* kqueue filter function */
2573 /*ARGSUSED*/
2574 static int
2575 filt_lio(struct knote *kn, long hint)
2576 {
2577 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2578 
2579 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2580 }
2581 
2582 #ifdef COMPAT_FREEBSD32
2583 #include <sys/mount.h>
2584 #include <sys/socket.h>
2585 #include <compat/freebsd32/freebsd32.h>
2586 #include <compat/freebsd32/freebsd32_proto.h>
2587 #include <compat/freebsd32/freebsd32_signal.h>
2588 #include <compat/freebsd32/freebsd32_syscall.h>
2589 #include <compat/freebsd32/freebsd32_util.h>
2590 
2591 struct __aiocb_private32 {
2592 	int32_t	status;
2593 	int32_t	error;
2594 	uint32_t kernelinfo;
2595 };
2596 
2597 #ifdef COMPAT_FREEBSD6
2598 typedef struct oaiocb32 {
2599 	int	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 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2604 	int	aio_lio_opcode;		/* LIO opcode */
2605 	int	aio_reqprio;		/* Request priority -- ignored */
2606 	struct	__aiocb_private32 _aiocb_private;
2607 } oaiocb32_t;
2608 #endif
2609 
2610 typedef struct aiocb32 {
2611 	int32_t	aio_fildes;		/* File descriptor */
2612 	uint64_t aio_offset __packed;	/* File offset for I/O */
2613 	uint32_t aio_buf;		/* I/O buffer in process space */
2614 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2615 	int	__spare__[2];
2616 	uint32_t __spare2__;
2617 	int	aio_lio_opcode;		/* LIO opcode */
2618 	int	aio_reqprio;		/* Request priority -- ignored */
2619 	struct	__aiocb_private32 _aiocb_private;
2620 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2621 } aiocb32_t;
2622 
2623 #ifdef COMPAT_FREEBSD6
2624 static int
2625 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2626 {
2627 
2628 	/*
2629 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2630 	 * supported by AIO with the old sigevent structure.
2631 	 */
2632 	CP(*osig, *nsig, sigev_notify);
2633 	switch (nsig->sigev_notify) {
2634 	case SIGEV_NONE:
2635 		break;
2636 	case SIGEV_SIGNAL:
2637 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2638 		break;
2639 	case SIGEV_KEVENT:
2640 		nsig->sigev_notify_kqueue =
2641 		    osig->__sigev_u.__sigev_notify_kqueue;
2642 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2643 		break;
2644 	default:
2645 		return (EINVAL);
2646 	}
2647 	return (0);
2648 }
2649 
2650 static int
2651 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2652 {
2653 	struct oaiocb32 job32;
2654 	int error;
2655 
2656 	bzero(kjob, sizeof(struct aiocb));
2657 	error = copyin(ujob, &job32, sizeof(job32));
2658 	if (error)
2659 		return (error);
2660 
2661 	CP(job32, *kjob, aio_fildes);
2662 	CP(job32, *kjob, aio_offset);
2663 	PTRIN_CP(job32, *kjob, aio_buf);
2664 	CP(job32, *kjob, aio_nbytes);
2665 	CP(job32, *kjob, aio_lio_opcode);
2666 	CP(job32, *kjob, aio_reqprio);
2667 	CP(job32, *kjob, _aiocb_private.status);
2668 	CP(job32, *kjob, _aiocb_private.error);
2669 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2670 	return (convert_old_sigevent32(&job32.aio_sigevent,
2671 	    &kjob->aio_sigevent));
2672 }
2673 #endif
2674 
2675 static int
2676 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2677 {
2678 	struct aiocb32 job32;
2679 	int error;
2680 
2681 	error = copyin(ujob, &job32, sizeof(job32));
2682 	if (error)
2683 		return (error);
2684 	CP(job32, *kjob, aio_fildes);
2685 	CP(job32, *kjob, aio_offset);
2686 	PTRIN_CP(job32, *kjob, aio_buf);
2687 	CP(job32, *kjob, aio_nbytes);
2688 	CP(job32, *kjob, aio_lio_opcode);
2689 	CP(job32, *kjob, aio_reqprio);
2690 	CP(job32, *kjob, _aiocb_private.status);
2691 	CP(job32, *kjob, _aiocb_private.error);
2692 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2693 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2694 }
2695 
2696 static long
2697 aiocb32_fetch_status(struct aiocb *ujob)
2698 {
2699 	struct aiocb32 *ujob32;
2700 
2701 	ujob32 = (struct aiocb32 *)ujob;
2702 	return (fuword32(&ujob32->_aiocb_private.status));
2703 }
2704 
2705 static long
2706 aiocb32_fetch_error(struct aiocb *ujob)
2707 {
2708 	struct aiocb32 *ujob32;
2709 
2710 	ujob32 = (struct aiocb32 *)ujob;
2711 	return (fuword32(&ujob32->_aiocb_private.error));
2712 }
2713 
2714 static int
2715 aiocb32_store_status(struct aiocb *ujob, long status)
2716 {
2717 	struct aiocb32 *ujob32;
2718 
2719 	ujob32 = (struct aiocb32 *)ujob;
2720 	return (suword32(&ujob32->_aiocb_private.status, status));
2721 }
2722 
2723 static int
2724 aiocb32_store_error(struct aiocb *ujob, long error)
2725 {
2726 	struct aiocb32 *ujob32;
2727 
2728 	ujob32 = (struct aiocb32 *)ujob;
2729 	return (suword32(&ujob32->_aiocb_private.error, error));
2730 }
2731 
2732 static int
2733 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2734 {
2735 	struct aiocb32 *ujob32;
2736 
2737 	ujob32 = (struct aiocb32 *)ujob;
2738 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2739 }
2740 
2741 static int
2742 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2743 {
2744 
2745 	return (suword32(ujobp, (long)ujob));
2746 }
2747 
2748 static struct aiocb_ops aiocb32_ops = {
2749 	.copyin = aiocb32_copyin,
2750 	.fetch_status = aiocb32_fetch_status,
2751 	.fetch_error = aiocb32_fetch_error,
2752 	.store_status = aiocb32_store_status,
2753 	.store_error = aiocb32_store_error,
2754 	.store_kernelinfo = aiocb32_store_kernelinfo,
2755 	.store_aiocb = aiocb32_store_aiocb,
2756 };
2757 
2758 #ifdef COMPAT_FREEBSD6
2759 static struct aiocb_ops aiocb32_ops_osigevent = {
2760 	.copyin = aiocb32_copyin_old_sigevent,
2761 	.fetch_status = aiocb32_fetch_status,
2762 	.fetch_error = aiocb32_fetch_error,
2763 	.store_status = aiocb32_store_status,
2764 	.store_error = aiocb32_store_error,
2765 	.store_kernelinfo = aiocb32_store_kernelinfo,
2766 	.store_aiocb = aiocb32_store_aiocb,
2767 };
2768 #endif
2769 
2770 int
2771 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2772 {
2773 
2774 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2775 }
2776 
2777 int
2778 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2779 {
2780 	struct timespec32 ts32;
2781 	struct timespec ts, *tsp;
2782 	struct aiocb **ujoblist;
2783 	uint32_t *ujoblist32;
2784 	int error, i;
2785 
2786 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2787 		return (EINVAL);
2788 
2789 	if (uap->timeout) {
2790 		/* Get timespec struct. */
2791 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2792 			return (error);
2793 		CP(ts32, ts, tv_sec);
2794 		CP(ts32, ts, tv_nsec);
2795 		tsp = &ts;
2796 	} else
2797 		tsp = NULL;
2798 
2799 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2800 	ujoblist32 = (uint32_t *)ujoblist;
2801 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2802 	    sizeof(ujoblist32[0]));
2803 	if (error == 0) {
2804 		for (i = uap->nent; i > 0; i--)
2805 			ujoblist[i] = PTRIN(ujoblist32[i]);
2806 
2807 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2808 	}
2809 	uma_zfree(aiol_zone, ujoblist);
2810 	return (error);
2811 }
2812 
2813 int
2814 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2815 {
2816 
2817 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2818 }
2819 
2820 #ifdef COMPAT_FREEBSD6
2821 int
2822 freebsd6_freebsd32_aio_read(struct thread *td,
2823     struct freebsd6_freebsd32_aio_read_args *uap)
2824 {
2825 
2826 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2827 	    &aiocb32_ops_osigevent));
2828 }
2829 #endif
2830 
2831 int
2832 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2833 {
2834 
2835 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2836 	    &aiocb32_ops));
2837 }
2838 
2839 #ifdef COMPAT_FREEBSD6
2840 int
2841 freebsd6_freebsd32_aio_write(struct thread *td,
2842     struct freebsd6_freebsd32_aio_write_args *uap)
2843 {
2844 
2845 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2846 	    &aiocb32_ops_osigevent));
2847 }
2848 #endif
2849 
2850 int
2851 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2852 {
2853 
2854 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2855 	    &aiocb32_ops));
2856 }
2857 
2858 int
2859 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2860 {
2861 
2862 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2863 	    &aiocb32_ops));
2864 }
2865 
2866 int
2867 freebsd32_aio_waitcomplete(struct thread *td,
2868     struct freebsd32_aio_waitcomplete_args *uap)
2869 {
2870 	struct timespec32 ts32;
2871 	struct timespec ts, *tsp;
2872 	int error;
2873 
2874 	if (uap->timeout) {
2875 		/* Get timespec struct. */
2876 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2877 		if (error)
2878 			return (error);
2879 		CP(ts32, ts, tv_sec);
2880 		CP(ts32, ts, tv_nsec);
2881 		tsp = &ts;
2882 	} else
2883 		tsp = NULL;
2884 
2885 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2886 	    &aiocb32_ops));
2887 }
2888 
2889 int
2890 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2891 {
2892 
2893 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2894 	    &aiocb32_ops));
2895 }
2896 
2897 #ifdef COMPAT_FREEBSD6
2898 int
2899 freebsd6_freebsd32_lio_listio(struct thread *td,
2900     struct freebsd6_freebsd32_lio_listio_args *uap)
2901 {
2902 	struct aiocb **acb_list;
2903 	struct sigevent *sigp, sig;
2904 	struct osigevent32 osig;
2905 	uint32_t *acb_list32;
2906 	int error, i, nent;
2907 
2908 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2909 		return (EINVAL);
2910 
2911 	nent = uap->nent;
2912 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2913 		return (EINVAL);
2914 
2915 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2916 		error = copyin(uap->sig, &osig, sizeof(osig));
2917 		if (error)
2918 			return (error);
2919 		error = convert_old_sigevent32(&osig, &sig);
2920 		if (error)
2921 			return (error);
2922 		sigp = &sig;
2923 	} else
2924 		sigp = NULL;
2925 
2926 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2927 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2928 	if (error) {
2929 		free(acb_list32, M_LIO);
2930 		return (error);
2931 	}
2932 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2933 	for (i = 0; i < nent; i++)
2934 		acb_list[i] = PTRIN(acb_list32[i]);
2935 	free(acb_list32, M_LIO);
2936 
2937 	error = kern_lio_listio(td, uap->mode,
2938 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2939 	    &aiocb32_ops_osigevent);
2940 	free(acb_list, M_LIO);
2941 	return (error);
2942 }
2943 #endif
2944 
2945 int
2946 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2947 {
2948 	struct aiocb **acb_list;
2949 	struct sigevent *sigp, sig;
2950 	struct sigevent32 sig32;
2951 	uint32_t *acb_list32;
2952 	int error, i, nent;
2953 
2954 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2955 		return (EINVAL);
2956 
2957 	nent = uap->nent;
2958 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2959 		return (EINVAL);
2960 
2961 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2962 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2963 		if (error)
2964 			return (error);
2965 		error = convert_sigevent32(&sig32, &sig);
2966 		if (error)
2967 			return (error);
2968 		sigp = &sig;
2969 	} else
2970 		sigp = NULL;
2971 
2972 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2973 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2974 	if (error) {
2975 		free(acb_list32, M_LIO);
2976 		return (error);
2977 	}
2978 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2979 	for (i = 0; i < nent; i++)
2980 		acb_list[i] = PTRIN(acb_list32[i]);
2981 	free(acb_list32, M_LIO);
2982 
2983 	error = kern_lio_listio(td, uap->mode,
2984 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2985 	    &aiocb32_ops);
2986 	free(acb_list, M_LIO);
2987 	return (error);
2988 }
2989 
2990 #endif
2991